realize sched

Merge branch 'master' into outerworld_work
train loop
2025-10-14 14:44:56 +08:00 · 2025-10-14 14:26:52 +08:00 · 2025-10-10 20:20:19 +08:00 · 2025-10-10 20:07:44 +08:00 · 2025-10-10 19:49:24 +08:00 · 2025-10-10 18:50:16 +08:00
876 changed files with 236177 additions and 264826 deletions
@@ -1,3 +0,0 @@
-[run]
-source = tinygrad
-branch = True
@@ -11,5 +11,5 @@ runs:
        git fetch origin $CURRENT_SHA
        export COMMIT_MESSAGE=$(git show -s --format=%B "$CURRENT_SHA")
        export CURRENT_HEAD=$(git rev-parse HEAD)
-        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && CHECK_OOB=0 PYTHONPATH=. python3 process_replay.py
+        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && IGNORE_OOB=1 PYTHONPATH=. python3 process_replay.py
        git checkout $CURRENT_HEAD  # restore to branch
@@ -41,14 +41,6 @@ inputs:
    description: "Install LLVM?"
    required: false
    default: 'false'
-  mesa:
-    description: "Install mesa"
-    required: false
-    default: 'false'
-  tinydreno:
-    description: "Install tinydreno"
-    required: false
-    default: 'false'
 runs:
  using: "composite"
  steps:
@@ -60,40 +52,32 @@ runs:

    # **** Caching packages ****

-    - name: Cache Python packages (PR)
-      if: github.event_name == 'pull_request'
-      id: restore-venv-pr
-      uses: actions/cache/restore@v4
-      with:
-        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
    - name: Cache Python packages
-      if: github.event_name != 'pull_request'
      id: restore-venv
      uses: actions/cache@v4
      with:
        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ hashFiles('**/setup.py') }}-${{ env.PYTHON_CACHE_VERSION }}

    # **** Caching downloads ****

-    - name: Cache downloads (PR)
-      if: inputs.key != '' && github.event_name == 'pull_request'
-      uses: actions/cache/restore@v4
-      with:
-        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
-    - name: Cache downloads
-      if: inputs.key != '' && github.event_name != 'pull_request'
+    - name: Cache downloads (Linux)
+      if: inputs.key != '' && runner.os == 'Linux'
      uses: actions/cache@v4
      with:
-        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
+        path: ~/.cache/tinygrad/downloads/
+        key: downloads-cache-${{ inputs.key }}-${{ env.DOWNLOAD_CACHE_VERSION }}
+    - name: Cache downloads (macOS)
+      if: inputs.key != '' && runner.os == 'macOS'
+      uses: actions/cache@v4
+      with:
+        path: ~/Library/Caches/tinygrad/downloads/
+        key: osx-downloads-cache-${{ inputs.key }}-${{ env.DOWNLOAD_CACHE_VERSION }}

    # **** Python deps ****

    - name: Install dependencies in venv (with extra)
-      if: inputs.deps != '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps != '' && steps.restore-venv.outputs.cache-hit != 'true'
      shell: bash
      run: |
        python -m venv .venv
@@ -104,7 +88,7 @@ runs:
        fi
        python -m pip install -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --extra-index-url https://download.pytorch.org/whl/cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
    - name: Install dependencies in venv (without extra)
-      if: inputs.deps == '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps == '' && steps.restore-venv.outputs.cache-hit != 'true'
      shell: bash
      run: |
        python -m venv .venv
@@ -149,7 +133,7 @@ runs:
      run: |
        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
        sudo tee /etc/apt/sources.list.d/rocm.list <<EOF
-        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.1 $(lsb_release -cs) main
+        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.2 $(lsb_release -cs) main
        EOF
        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600

@@ -194,18 +178,12 @@ runs:
        echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
        echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"

-    - name: Cache apt (PR)
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name == 'pull_request'
-      uses: actions/cache/restore@v4
-      with:
-        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
    - name: Cache apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name != 'pull_request'
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
      uses: actions/cache@v4
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.APT_CACHE_VERSION }}

    - name: Run apt Update + Install
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -237,9 +215,9 @@ runs:
      shell: bash
      run: |
        sudo mkdir -p /usr/local/lib
-        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/tinygrad/amdcomgr_dylib/releases/latest | \
+        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/amdcomgr_dylib/releases/latest | \
          jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
-          sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib
+          sudo xargs curl -L -o /usr/local/lib/libamd_comgr.dylib
        cargo build --release --manifest-path ./extra/remu/Cargo.toml

    # **** gpuocelot ****
@@ -257,26 +235,17 @@ runs:
        ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
        ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
        ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
-    - name: Cache gpuocelot (PR)
-      if: inputs.ocelot == 'true' && github.event_name == 'pull_request'
-      id: cache-build-pr
-      uses: actions/cache/restore@v4
-      env:
-        cache-name: cache-gpuocelot-build-1
-      with:
-        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
    - name: Cache gpuocelot
-      if: inputs.ocelot == 'true' && github.event_name != 'pull_request'
+      if: inputs.ocelot == 'true'
      id: cache-build
      uses: actions/cache@v4
      env:
        cache-name: cache-gpuocelot-build-1
      with:
        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}
    - name: Clone/compile gpuocelot
-      if: inputs.ocelot == 'true' && steps.cache-build-pr.outputs.cache-hit != 'true' && steps.cache-build.outputs.cache-hit != 'true'
+      if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
      shell: bash
      run: |
        git clone --recurse-submodules https://github.com/gpuocelot/gpuocelot.git ${{ github.workspace }}/gpuocelot
@@ -287,7 +256,6 @@ runs:

        CMAKE_ARGS="-Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5"
        if [[ "${{ runner.os }}" == "macOS" ]]; then
-          sudo xcode-select -s /Applications/Xcode_16.2.app/Contents/Developer
          CMAKE_ARGS="$CMAKE_ARGS -DBoost_INCLUDE_DIR=$(brew --prefix boost)/include -DBoost_LIBRARY_DIR=$(brew --prefix boost)/lib"
        fi

@@ -306,7 +274,7 @@ runs:
      if: inputs.webgpu == 'true' && runner.os == 'Linux'
      shell: bash
      run: |
-        sudo curl -fL https://github.com/wpmed92/pydawn/releases/download/v0.1.6/libwebgpu_dawn.so -o /usr/local/lib/libwebgpu_dawn.so
+        sudo curl -L https://github.com/wpmed92/pydawn/releases/download/v0.1.6/libwebgpu_dawn.so -o /usr/local/lib/libwebgpu_dawn.so
        sudo ldconfig
    - name: Install WebGPU dawn (macOS)
      if: inputs.webgpu == 'true' && runner.os == 'macOS'
@@ -321,19 +289,3 @@ runs:
      if: inputs.llvm == 'true' && runner.os == 'macOS'
      shell: bash
      run: brew install llvm@20
-
-    # **** mesa ****
-    - name: Install mesa (linux)
-      if: inputs.mesa == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinymesa/releases/download/v1/libtinymesa_cpu-mesa-25.2.7-linux-amd64.so -o /usr/lib/libtinymesa_cpu.so
-    - name: Install mesa (macOS)
-      if: inputs.mesa == 'true' && runner.os == 'macOS'
-      shell: bash
-      run: brew install sirhcm/tinymesa/tinymesa_cpu
-
-    # *** tinydreno ***
-    - name: Install tinydreno (linux)
-      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
@@ -1,7 +1,10 @@
 name: Autogen
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '13'
+  DOWNLOAD_CACHE_VERSION: '12'
+  PYTHON_CACHE_VERSION: '3'
+  APT_CACHE_VERSION: '1'
+  BUILD_CACHE_VERSION: '1'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
@@ -11,19 +14,15 @@ on:
    branches:
      - master
  pull_request:
-    paths:
+    paths: 
    - 'tinygrad/runtime/autogen/**/*'
-    - 'tinygrad/runtime/support/autogen.py'
-    - '.github/workflows/autogen.yml'
  workflow_dispatch:
-    paths:
+    paths: 
    - 'tinygrad/runtime/autogen/**/*'
-    - 'tinygrad/runtime/support/autogen.py'
-    - '.github/workflows/autogen.yml'

 jobs:
  autogen:
-    name: In-tree Autogen
+    name: Autogen
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
@@ -32,113 +31,61 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: 'autogen'
        opencl: 'true'
        amd: 'true'
        cuda: 'true'
-        llvm: 'true'
        webgpu: 'true'
-        mesa: 'true'
-        pydeps: 'pyyaml mako'
-    - name: Install autogen support packages
-      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev
-    - name: Regenerate autogen files
-      run: |
-        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "comgr.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
-        python3 -c "from tinygrad.runtime.autogen import opencl"
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv"
-        python3 -c "from tinygrad.runtime.autogen import comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
-        python3 -c "from tinygrad.runtime.autogen.am import am, pm4_soc15, pm4_nv, sdma_4_0_0, sdma_5_0_0, sdma_6_0_0, smu_v13_0_0, smu_v13_0_6, smu_v13_0_12, smu_v14_0_2"
-        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
-        python3 -c "from tinygrad.runtime.autogen import llvm"
-        python3 -c "from tinygrad.runtime.autogen import webgpu"
-        python3 -c "from tinygrad.runtime.autogen import kgsl, qcom_dsp"
-        python3 -c "from tinygrad.runtime.autogen import libusb"
-        python3 -c "from tinygrad.runtime.autogen import mesa"
-        python3 -c "from tinygrad.runtime.autogen import avcodec"
-        python3 -c "from tinygrad.runtime.autogen import llvm_qcom"
-        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
-    - name: Check for differences
-      run: |
-        if ! git diff --quiet; then
-          git diff
-          git diff > autogen-ubuntu.patch
-          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
-          exit 1
-        fi
-    - name: Upload patch artifact
-      if: failure()
-      uses: actions/upload-artifact@v4
-      with:
-        name: autogen-ubuntu-patch
-        path: autogen-ubuntu.patch
-
-  autogen-mac:
-    name: In-tree Autogen (macos)
-    runs-on: macos-14
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: 'autogen-mac'
        llvm: 'true'
-    - name: Regenerate autogen files
-      run: |
-        rm tinygrad/runtime/autogen/metal.py tinygrad/runtime/autogen/iokit.py tinygrad/runtime/autogen/corefoundation.py
-        python3 -c "from tinygrad.runtime.autogen import metal, iokit, corefoundation"
-    - name: Check for differences
-      run: |
-        if ! git diff --quiet; then
-          git diff
-          git diff > autogen-macos.patch
-          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
-          exit 1
-        fi
-    - name: Upload patch artifact
-      if: failure()
-      uses: actions/upload-artifact@v4
-      with:
-        name: autogen-macos-patch
-        path: autogen-macos.patch
-
-  autogen-comgr-2:
-    name: In-tree Autogen (comgr 2)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: 'autogen-comgr'
    - name: Install autogen support packages
+      run: sudo apt-get install -y --no-install-recommends llvm-14-dev libclang-14-dev
+    - name: Verify OpenCL autogen
      run: |
-        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-        sudo tee /etc/apt/sources.list.d/rocm.list <<EOF
-        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.2 $(lsb_release -cs) main
-        EOF
-        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
-        sudo apt -qq update || true
-        sudo apt-get install -y --no-install-recommends libclang-20-dev comgr
-    - name: Regenerate autogen files
+        cp tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
+        ./autogen_stubs.sh opencl
+        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
+    - name: Verify CUDA autogen
      run: |
-        rm tinygrad/runtime/autogen/comgr.py
-        python3 -c "from tinygrad.runtime.autogen import comgr"
-    - name: Check for differences
+        cp tinygrad/runtime/autogen/cuda.py /tmp/cuda.py.bak
+        cp tinygrad/runtime/autogen/nv_gpu.py /tmp/nv_gpu.py.bak
+        ./autogen_stubs.sh cuda
+        ./autogen_stubs.sh nv
+        diff /tmp/cuda.py.bak tinygrad/runtime/autogen/cuda.py
+        diff /tmp/nv_gpu.py.bak tinygrad/runtime/autogen/nv_gpu.py
+    - name: Verify AMD autogen
      run: |
-        if ! git diff --quiet; then
-          git diff
-          git diff > autogen-comgr2.patch
-          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
-          exit 1
-        fi
-    - name: Upload patch artifact
-      if: failure()
-      uses: actions/upload-artifact@v4
-      with:
-        name: autogen-comgr2-patch
-        path: autogen-comgr2.patch
+        cp tinygrad/runtime/autogen/hsa.py /tmp/hsa.py.bak
+        cp tinygrad/runtime/autogen/kfd.py /tmp/kfd.py.bak
+        cp tinygrad/runtime/autogen/comgr.py /tmp/comgr.py.bak
+        cp tinygrad/runtime/autogen/amd_gpu.py /tmp/amd_gpu.py.bak
+        cp tinygrad/runtime/autogen/sqtt.py /tmp/sqtt.py.bak
+        ./autogen_stubs.sh hsa
+        ./autogen_stubs.sh kfd
+        ./autogen_stubs.sh comgr
+        ./autogen_stubs.sh amd
+        ./autogen_stubs.sh sqtt
+        diff /tmp/hsa.py.bak tinygrad/runtime/autogen/hsa.py
+        diff /tmp/kfd.py.bak tinygrad/runtime/autogen/kfd.py
+        diff /tmp/comgr.py.bak tinygrad/runtime/autogen/comgr.py
+        diff /tmp/amd_gpu.py.bak tinygrad/runtime/autogen/amd_gpu.py
+        diff /tmp/sqtt.py.bak tinygrad/runtime/autogen/sqtt.py
+    - name: Verify Linux autogen
+      run: |
+        cp tinygrad/runtime/autogen/libc.py /tmp/libc.py.bak
+        cp tinygrad/runtime/autogen/io_uring.py /tmp/io_uring.py.bak
+        cp tinygrad/runtime/autogen/ib.py /tmp/ib.py.bak
+        ./autogen_stubs.sh libc
+        ./autogen_stubs.sh io_uring
+        ./autogen_stubs.sh ib
+        diff /tmp/libc.py.bak tinygrad/runtime/autogen/libc.py
+        diff /tmp/io_uring.py.bak tinygrad/runtime/autogen/io_uring.py
+        diff /tmp/ib.py.bak tinygrad/runtime/autogen/ib.py
+    - name: Verify WebGPU autogen
+      run: |
+        cp tinygrad/runtime/autogen/webgpu.py /tmp/webgpu.py.bak
+        ./autogen_stubs.sh webgpu
+        diff /tmp/webgpu.py.bak tinygrad/runtime/autogen/webgpu.py
+    - name: Verify LLVM autogen
+      run: |
+        cp tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
+        ./autogen_stubs.sh llvm
+        diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
@@ -14,43 +14,14 @@ on:
      - update_benchmark
      - update_benchmark_staging
  workflow_dispatch:
+    inputs:
+      run_process_replay:
+        description: "Run process replay tests"
+        required: false
+        default: false
+        type: boolean

 jobs:
-  # the goal of this test is to replicate a normal person on a laptop running the test
-  # no process replay, no benchmarks, no CI, just a normal laptop person
-  # the 3 minute timeout should not be raised
-  testmacpytest:
-    name: Mac pytest
-    env:
-      CI: ""
-      CAPTURE_PROCESS_REPLAY: "0"
-    runs-on: [self-hosted, macOS]
-    timeout-minutes: 3
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    # brew install uv
-    - name: setup python environment
-      run: |
-        rm -rf /tmp/tinygrad_pytest_ci
-        uv venv /tmp/tinygrad_pytest_ci
-        source /tmp/tinygrad_pytest_ci/bin/activate
-        uv pip install .[testing]
-    - name: setup staging db
-      run: |
-        echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
-        rm -f /tmp/pytest-db-ci*
-    - name: Run pytest -nauto
-      run: |
-        source /tmp/tinygrad_pytest_ci/bin/activate
-        pytest -nauto --durations=20
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: FLOAT16=1 CL=1 IMAGE=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-
  testmacbenchmark:
    name: Mac Benchmark
    env:
@@ -68,7 +39,6 @@ jobs:
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
-        mkdir -p extra/disassemblers
        ln -s ~/tinygrad/extra/disassemblers/applegpu extra/disassemblers/applegpu
        ln -s ~/tinygrad/weights/sd-v1-4.ckpt weights/sd-v1-4.ckpt
        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
@@ -81,22 +51,21 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: python3.11 test/external/process_replay/reset.py
-    - name: Print macOS version
-      run: sw_vers
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=800 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=900 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
    - name: Run Stable Diffusion v2
      # TODO: very slow step time
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=10000 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
    # process replay can't capture this, the graph is too large
-    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing
+    # TODO: too slow
+    # - name: Run SDXL
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run model inference benchmark
-      run: METAL=1 NOCLANG=1 python3.11 test/external/external_model_benchmark.py
+      run: METAL=1 python3.11 test/external/external_model_benchmark.py
    - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
+      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test tensor cores
      run: METAL=1 python3.11 test/opt/test_tensor_cores.py
    - name: Test AMX tensor cores
@@ -106,97 +75,95 @@ jobs:
        DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
        DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
    - name: Run Tensor Core GEMM (float)
-      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
    - name: Run Tensor Core GEMM (half)
-      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_half.txt
    - name: Run Tensor Core GEMM (bfloat16)
-      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
    - name: Fuzz Padded Tensor Core GEMM
      run: METAL=1 M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
    - name: Run LLaMA
      run: |
-        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    - name: Run quantized LLaMA
      run: |
-        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8
-        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4
+        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
+        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_nf4.txt
    - name: Run quantized LLaMA3
      run: |
-        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8
-        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4
+        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8 | tee llama3_int8.txt
+        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4 | tee llama3_nf4.txt
    #- name: Run LLaMA 7B on 4 (virtual) GPUs
-    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - name: Run OLMoE
      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
    - name: Train MNIST
-      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt

    # NOTE: this is failing in CI. it is not failing on my machine and I don't really have a way to debug it
    # the error is "RuntimeError: Internal Error (0000000e:Internal Error)"
    #- name: Run 10 CIFAR training steps
-    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py
+    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
    #- name: Run 10 CIFAR training steps w HALF
-    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py
+    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt

    #- name: Run 10 CIFAR training steps w BF16
-    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
+    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    - name: UsbGPU boot time
+      run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
+    - name: UsbGPU tiny tests
+      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
+    - name: UsbGPU copy speeds
+      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+    #- name: UsbGPU openpilot test
+    #  run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB NOLOCALS=0 IMAGE=0 GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (Mac)
        path: |
          onnx_inference_speed.csv
+          torch_speed.txt
+          llama_unjitted.txt
+          llama_jitted.txt
+          llama_beam.txt
+          llama_int8.txt
+          llama_nf4.txt
+          llama3_int8.txt
+          llama3_nf4.txt
+          llama_four_gpu.txt
+          gpt2_unjitted.txt
+          gpt2_jitted.txt
+          gpt2_half.txt
+          gpt2_half_beam.txt
+          matmul.txt
+          matmul_half.txt
+          matmul_bfloat16.txt
+          sd.txt
+          sd_no_fp16.txt
+          sdv2.txt
+          sdxl.txt
+          beautiful_mnist.txt
+          train_cifar.txt
+          train_cifar_half.txt
+          train_cifar_bf16.txt
+          train_cifar_wino.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3.11 process_replay.py

-  testusbgpu:
-    name: UsbGPU Benchmark
-    env:
-      PYTHONPYCACHEPREFIX: /tmp/tiny_python_pycache
-    runs-on: [self-hosted, macOS]
-    timeout-minutes: 10
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: Kill stale pids
-      run: |
-        PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-        PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
-    - name: UsbGPU boot time
-      run: sudo -E PYTHONPATH=. GMMU=0 DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
-    - name: UsbGPU tiny tests
-      run: sudo -E PYTHONPATH=. GMMU=0 AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
-    - name: UsbGPU copy speeds
-      run: sudo -E PYTHONPATH=. GMMU=0 AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
-    #- name: UsbGPU openpilot test
-    #  run: sudo -E PYTHONPATH=. GMMU=0 AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
-    - name: UsbGPU (USB4/TB) boot time
-      run: PYTHONPATH=. DEBUG=3 NV=1 NV_IFACE=PCI NV_NAK=1 time python3.11 test/test_tiny.py TestTiny.test_plus
-    - name: UsbGPU (USB4/TB) tiny tests
-      run: PYTHONPATH=. NV=1 NV_IFACE=PCI NV_NAK=1 python3.11 test/test_tiny.py
-
  testnvidiabenchmark:
    name: tinybox green Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
@@ -229,9 +196,9 @@ jobs:
    - name: Run model inference benchmark
      run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
    - name: Test speed vs torch
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test speed vs theoretical
-      run: NV=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      run: NV=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
    - name: Test benchmark allreduce
      run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
    - name: Test tensor cores
@@ -240,58 +207,75 @@ jobs:
        NV=1 NV_PTX=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
    - name: Run Tensor Core GEMM (CUDA)
      run: |
-        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
+        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
+        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
    - name: Run Tensor Core GEMM (PTX)
-      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
    - name: Run Tensor Core GEMM (NV)
-      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_nv.txt
    - name: Test NV=1
      run: DEBUG=2 NV=1 python -m pytest -rA test/test_tiny.py
    - name: Test CUDA=1
      run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    # TODO: too slow
    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA
      run: |
-        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    - name: Run quantized LLaMA3
-      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8
+      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
    # - name: Run LLaMA-2 70B
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA)
        path: |
          onnx_inference_speed.csv
+          torch_speed.txt
+          matmul.txt
+          matmul_bfloat16.txt
+          matmul_tf32.txt
+          matmul_ptx.txt
+          matmul_nv.txt
+          sd.txt
+          sdxl.txt
+          llama_unjitted.txt
+          llama_jitted.txt
+          llama_beam.txt
+          llama3_beam.txt
+          llama3_four_gpu.txt
+          llama3_six_gpu.txt
+          llama_2_70B.txt
+          mixtral.txt
+          gpt2_unjitted.txt
+          gpt2_jitted.txt
+          gpt2_half.txt
+          gpt2_half_beam.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -329,31 +313,44 @@ jobs:
    # TODO: too slow
    # - name: Fuzz Padded Tensor Core GEMM (PTX)
    #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
-    - name: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
    - name: Train MNIST
-      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=270 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=310 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=310 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    - name: Run MLPerf resnet eval on training data
      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (NVIDIA Training)
+        path: |
+          beautiful_mnist.txt
+          train_cifar.txt
+          train_cifar_half.txt
+          train_cifar_bf16.txt
+          train_cifar_wino.txt
+          train_cifar_one_gpu.txt
+          train_cifar_six_gpu.txt
+          train_resnet.txt
+          train_resnet_one_gpu.txt
+          train_bert.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -368,12 +365,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
    #- name: Insert amdgpu
    #  run: sudo modprobe amdgpu
    - name: Symlink models and datasets
@@ -407,18 +402,16 @@ jobs:
    #- name: Test speed vs torch
    #  run: |
    #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test speed vs theoretical
-      run: AMD=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test tensor cores AMD_LLVM=0
-      run: AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
-    # TODO: this is flaky
-    # - name: Test tensor cores AMD_LLVM=1
-    #   run: AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (AMD)
+      run: AMD=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+    - name: Test tensor cores
      run: |
+        AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
+        AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
        AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+    - name: Run Tensor Core GEMM (AMD)
+      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
    - name: Test AMD=1
      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
    #- name: Test HIP=1
@@ -433,39 +426,62 @@ jobs:
    - name: Test AM warm start time
      run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
-    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+    # TODO: too slow
+    # - name: Run SDXL
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA 7B
      run: |
-        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA 7B with BEAM
-      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
    #- name: Restore amdgpu
    #  run: sudo modprobe amdgpu
    # - name: Run LLaMA-2 70B
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AMD)
+        path: |
+          onnx_inference_speed.csv
+          torch_speed.txt
+          llama_unjitted.txt
+          llama_jitted.txt
+          llama_beam.txt
+          llama3_beam.txt
+          llama3_four_gpu.txt
+          llama3_six_gpu.txt
+          llama_2_70B.txt
+          gpt2_unjitted.txt
+          gpt2_jitted.txt
+          gpt2_half.txt
+          gpt2_half_beam.txt
+          matmul.txt
+          matmul_amd.txt
+          sd.txt
+          sdxl.txt
+          mixtral.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -480,12 +496,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -504,23 +518,34 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Train MNIST
-      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=330 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    # - name: Run 10 CIFAR training steps w BF16
-    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # TODO: broken on some of the machines
-    #- name: Test full tinyfs load
-    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+    #- name: Run full CIFAR training steps w 6 GPUS
+    #  run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+    #- name: Run full CIFAR training steps w 6 GPUS (REMOTE)
+    #  run: time BENCHMARK_LOG=cifar_6gpu_remote REMOTE=1 REMOTEDEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu_remote.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AMD Training)
+        path: |
+          beautiful_mnist.txt
+          train_cifar.txt
+          train_cifar_half.txt
+          train_cifar_bf16.txt
+          train_cifar_wino.txt
+          train_cifar_one_gpu.txt
+          train_cifar_six_gpu.txt
+          train_cifar_six_gpu_remote.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -535,12 +560,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -560,13 +583,20 @@ jobs:
      run: test/external/process_replay/reset.py
    - name: Run MLPerf resnet eval
      run: time BENCHMARK_LOG=resnet_eval AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AMD MLPerf)
+        path: |
+          train_resnet.txt
+          train_resnet_one_gpu.txt
+          train_bert.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -588,22 +618,22 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: IR3 openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM QCOM_IR3=1 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.11.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.11.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_11_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
-    - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
-      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: benchmark openpilot 0.9.9 driving_vision
+      run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+    - name: benchmark openpilot 0.9.9 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+    - name: benchmark openpilot 0.9.9 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 0.9.9 driving_vision
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=18 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.9.9 driving_policy
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=7 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.9.9 dmonitoring
+      run: PYTHONPATH="." ASSERT_MIN_STEP_TIME=12 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 Space Lab policy + vision
+      run: |
+        PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
+        PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
    - name: benchmark MobileNetV2 on DSP
      run: |
        # generate quantized weights
@@ -611,30 +641,19 @@ jobs:
        ln -s /data/home/tiny/tinygrad/testsig-*.so .
        PYTHONPATH=. CC=clang-19 CPU=1 CPU_LLVM=0 QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
        # benchmark on DSP with NOOPT=1, the devectorizer has issues
-        PYTHONPATH=. CC=clang-19 DSP=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+        PYTHONPATH=. CC=clang-19 DSP=1 DONT_REALIZE_EXPAND=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
-
-  testcommausbgpubenchmark:
-    name: UsbGPU Benchmark (comma)
-    runs-on: [self-hosted, Linux, comma4]
-    timeout-minutes: 20
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." GMMU=0 DEV=AMD AMD_LLVM=1 AMD_IFACE=USB ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot load_pickle 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_load_pickle PYTHONPATH="." GMMU=0 DEV=AMD AMD_IFACE=USB ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (comma)
+        path: |
+          openpilot_compile_0_9_4.txt
+          openpilot_compile_0_9_7.txt
+          openpilot_0_9_4.txt
+          openpilot_0_9_7.txt
+          openpilot_image_0_9_4.txt
+          openpilot_image_0_9_7.txt

  testreddriverbenchmark:
    name: AM Benchmark
@@ -647,12 +666,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+      run: ./extra/hcq/hcq_smi.py amd rmmod
    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+      run: ./extra/hcq/hcq_smi.py amd kill_pids
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -681,7 +698,7 @@ jobs:
    #     AMD=1 AMD_LLVM=1 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
    #     AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (AMD)
-      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee am_matmul_amd.txt
    - name: Test AMD=1
      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
    - name: Test DISK copy time
@@ -691,20 +708,21 @@ jobs:
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+    # TODO: enable
    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6482 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD=1 AMD_IFACE=PCI python3 test/test_tiny.py
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD=1 AMD_AQL=1 AMD_IFACE=PCI python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
+      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AM Driver)
+        path: |
+          am_matmul_amd.txt
+          am_train_cifar_one_gpu.txt
+          am_train_resnet_one_gpu.txt
+          am_train_bert_one_gpu.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -719,12 +737,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove nv modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
+      run: ./extra/hcq/hcq_smi.py nv rmmod
    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
+      run: ./extra/hcq/hcq_smi.py nv kill_pids
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -753,20 +769,21 @@ jobs:
        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Test LLAMA-3
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6483 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6483 NV=1 python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
+      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (NV Driver)
+        path: |
+          nv_llama3_beam.txt
+          nv_train_cifar_one_gpu.txt
+          nv_train_resnet_one_gpu.txt
+          nv_train_bert_one_gpu.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
@@ -22,13 +22,13 @@ jobs:
    - name: Run SDXL with new search
      # TODO: GCVM_L2_PROTECTION_FAULT_STATUS with llvm19
      run: |
-        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CCACHE=0 python examples/sdxl.py --noshow --timing --seed 0
+        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run SDXL with cached search
      run: |
        BENCHMARK_LOG=search_sdxl_cached PYTHONPATH=. AMD=1 JITBEAM=2 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run winograd cifar with new search
      run: |
-        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 IGNORE_BEAM_CACHE=1 CCACHE=0 BS=1024 STEPS=500 python examples/hlb_cifar10.py
+        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 BS=1024 STEPS=500 python examples/hlb_cifar10.py
    - name: Run winograd cifar with cached search
      run: |
        BENCHMARK_LOG=search_wino_cifar_cached WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 BS=1024 STEPS=500 python examples/hlb_cifar10.py
@@ -12,7 +12,7 @@ jobs:
  run_script_job:
    runs-on: [self-hosted, Linux, tinybox]
    if: github.repository_owner == 'tinygrad'
-    timeout-minutes: 720
+    timeout-minutes: 360

    steps:
    - name: Checkout Code
@@ -27,4 +27,4 @@ jobs:
      run: |
        rm "~/.cache/tinygrad/cache_mlperf.db" || true
        BENCHMARK_LOG=mlpert_train_resnet LOGMLPERF=0 CACHEDB="~/.cache/tinygrad/cache_mlperf.db" examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-        rm "~/.cache/tinygrad/cache_mlperf.db"
+        rm "~/.cache/tinygrad/cache_mlperf.db"
@@ -20,11 +20,11 @@ jobs:
    - name: Install dependencies
      run: |
        python -m pip install --upgrade pip
-        pip install setuptools wheel build twine
+        pip install setuptools wheel twine
    - name: Build and publish
      env:
        TWINE_USERNAME: ${{ secrets.PYPI_USERNAME }}
        TWINE_PASSWORD: ${{ secrets.PYPI_PASSWORD }}
      run: |
-        python -m build
+        python setup.py sdist bdist_wheel
        twine upload dist/*
@@ -56,15 +56,15 @@ jobs:
        uses: actions/checkout@v4
        with:
          path: base
-      - name: Set up Python 3.12
+      - name: Set up Python 3.10
        uses: actions/setup-python@v5
        with:
-          python-version: '3.12'
+          python-version: '3.10'
      - name: Count Line Diff
        run: |
+          pip install tabulate
          BASE="$GITHUB_WORKSPACE/base"
          PR="$GITHUB_WORKSPACE/pr"
-          pip install tabulate $BASE
          cp "$BASE/sz.py" .
          echo "loc_content<<EOF" >> "$GITHUB_ENV"
          python sz.py "$BASE" "$PR" >> "$GITHUB_ENV"
@@ -1,11 +1,13 @@
 name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '18'
+  DOWNLOAD_CACHE_VERSION: '12'
+  PYTHON_CACHE_VERSION: '3'
+  APT_CACHE_VERSION: '1'
+  BUILD_CACHE_VERSION: '1'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
-  CHECK_OOB: 1

 on:
  push:
@@ -26,19 +28,17 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: llvm-speed
-        deps: testing_unit
+        deps: testing_minimal
        llvm: 'true'
    - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py

  docs:
    name: Docs
    runs-on: ubuntu-22.04
    timeout-minutes: 10
-    env:
-      CHECK_OOB: 0
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -70,11 +70,13 @@ jobs:
        source venv/bin/activate
        pip install $GITHUB_WORKSPACE
        cp $GITHUB_WORKSPACE/examples/beautiful_mnist.py .
-        BS=2 STEPS=10 MAX_BUFFER_SIZE=0 python beautiful_mnist.py
+        BS=2 STEPS=10 python beautiful_mnist.py
    - name: Test Docs Build
      run: python -m mkdocs build --strict
    - name: Test Docs
-      run: python docs/abstractions3.py
+      run: |
+        python docs/abstractions2.py
+        python docs/abstractions3.py
    - name: Test README
      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
    - name: Test Quickstart
@@ -98,30 +100,31 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
+        deps: testing_minimal
        pydeps: "pillow torchvision expecttest"
        llvm: 'true'
    - name: Install ninja
      run: |
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
+    - name: Lint with ruff
+      run: |
+        pip3 install --upgrade --force-reinstall ruff==0.11.0
+        python3 -m ruff check extra/torch_backend/backend.py
    - name: Test one op
-      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
    - name: Test ResNet-18
      run: DEBUG=2 python3 extra/torch_backend/example.py
-    - name: custom tests
-      run: python3 -m pytest -n auto extra/torch_backend/test.py --durations=20
+    - name: My (custom) tests
+      run: python3 extra/torch_backend/test.py
    - name: Test one op in torch tests
      run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
    - name: Test Ops with TINY_BACKEND
-      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/backend/test_ops.py --durations=20
+      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
    - name: Test in-place operations on views
      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
    - name: Test multi-gpu
      run: CPU=1 CPU_LLVM=1 GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
-    - name: Test kernel fusion
-      run: python3 extra/torch_backend/test_kernel_fusion.py
-

  torchbackendmore:
    name: Torch Backend Tests More
@@ -134,14 +137,14 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
+        deps: testing_minimal
        llvm: 'true'
    - name: Install ninja
      run: |
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+      run: CPU=1 CPU_LLVM=1 TARGET_EVAL_ACC_PCT=96.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
    - name: Test some torch tests (expect failure)
      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true

@@ -156,27 +159,27 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: be-minimal
-        deps: testing_unit
+        deps: testing_minimal
    - name: Test dtype with Python emulator
-      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py
+      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/test_dtype.py test/test_dtype_alu.py
    - name: Test ops with Python emulator
-      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_ops.py --durations=20
+      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/test_ops.py --durations=20
    - name: Test uops with Python emulator
-      run: PYTHON=1 python3 -m pytest test/backend/test_uops.py --durations=20
+      run: PYTHON=1 python3 -m pytest test/test_uops.py --durations=20
    - name: Test symbolic with Python emulator
-      run: PYTHON=1 python3 test/backend/test_symbolic_ops.py
+      run: PYTHON=1 python3 test/test_symbolic_ops.py
    - name: test_renderer_failures with Python emulator
-      run: PYTHON=1 python3 -m pytest -rA test/backend/test_renderer_failures.py::TestRendererFailures
-    - name: Test IMAGE support
+      run: PYTHON=1 python3 -m pytest -rA test/test_renderer_failures.py::TestRendererFailures
+    - name: Test IMAGE=2 support
      run: |
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_simple_conv2d
+        IMAGE=2 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
+        IMAGE=2 PYTHON=1 python3 test/test_ops.py TestOps.test_simple_conv2d
    - name: Test emulated METAL tensor cores
      run: |
-        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_big_gemm
+        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_big_gemm
        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test emulated AMX tensor cores
-      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
+      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
    - name: Test emulated AMD tensor cores
      run: |
        DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
@@ -197,27 +200,28 @@ jobs:
        DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test emulated CUDA tensor cores
      run: |
-        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
+        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test emulated INTEL OpenCL tensor cores
      run: DEBUG=2 EMULATE=INTEL FORWARD_ONLY=1 PYTHON=1 HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
    - name: Test emulated AMX tensor cores
      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test device flop counts
      run: |
-        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStats.test_simple_matmul
+        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStats.test_simple_matmul

  linter:
    name: Linters
    runs-on: ubuntu-latest
    timeout-minutes: 10

+    # TODO: run the pre-commit hook to replace a lot of this
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -225,55 +229,23 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: linting-only
-        python-version: '3.11'
+        python-version: '3.10'
        deps: linting
    - name: Lint bad-indentation and trailing-whitespace with pylint
      run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
-    - name: Run pre-commit linting hooks
-      run: SKIP=tiny,tests,example pre-commit run --all-files
-    - name: Lint additional files with ruff
+    - name: Lint with ruff
      run: |
+        pip3 install --upgrade --force-reinstall ruff==0.11.0
+        python3 -m ruff check .
        python3 -m ruff check examples/mlperf/ --ignore E501
-        python3 -m ruff check extra/thunder/tiny/ --ignore E501 --ignore F841 --ignore E722
-        python3 -m ruff check extra/torch_backend/backend.py
-    - name: Run mypy with lineprecision report
+    - name: Lint tinygrad with pylint
+      run: python -m pylint tinygrad/
+    - name: Run mypy
      run: |
-        python -m mypy --lineprecision-report .
-        grep -v autogen lineprecision.txt | awk 'NR>2 {lines+=$2; precise+=$3; imprecise+=$4; any+=$5; empty+=$6} END {t=lines-empty; printf "TOTAL: %d lines, %d precise (%.1f%%), %d imprecise (%.1f%%), %d any (%.1f%%)\n", t, precise, 100*precise/t, imprecise, 100*imprecise/t, any, 100*any/t}'
+        python -m mypy --strict-equality --lineprecision-report .
        cat lineprecision.txt
    - name: Run TYPED=1
-      run: CHECK_OOB=0 DEV=CPU TYPED=1 python test/test_tiny.py
-
-  nulltest:
-    name: Null Tests
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
-        deps: testing_unit
-        llvm: 'true'
-        amd: 'true'
-    - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
-    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
-    # TODO: too slow
-    # - name: Run SDXL on NULL backend
-    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
-    - name: Run Clip tests for SD MLPerf on NULL backend
-      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
-    - name: Run AMD emulated BERT training on NULL backend
-      run: EMULATE=AMD_RDNA4 NULL=1 NULL_ALLOW_COPYOUT=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    # TODO: support fake weights
-    #- name: Run LLaMA 7B on 4 fake devices
-    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
+      run: TYPED=1 python -c "import tinygrad"

  unittest:
    name: Unit Tests
@@ -286,19 +258,25 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
+        key: unittest-12
+        pydeps: "pillow numpy ftfy regex"
        deps: testing_unit
-        llvm: 'true'
-        amd: 'true'
-    - name: Run pre-commit test hooks
-      run: SKIP=ruff,mypy pre-commit run --all-files
    - name: Check Device.DEFAULT
      run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
    - name: Run unit tests
-      run: |
-        CPU=1 python test/null/test_device.py TestRunAsModule.test_module_runs
-        CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+      run: CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+    - name: Check SPEC=1
+      run: SPEC=1 python3 test/test_tiny.py
+    - name: Run targetted tests on NULL backend
+      run: NULL=1 python3 -m unittest test.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
+    # TODO: too slow
+    # - name: Run SDXL on NULL backend
+    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+    - name: Run Clip tests for SD MLPerf on NULL backend
+      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
+    # TODO: support fake weights
+    #- name: Run LLaMA 7B on 4 fake devices
+    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
    - name: Run GC tests
      run: python test/external/external_uop_gc.py
    - name: External Benchmark Schedule
@@ -312,28 +290,8 @@ jobs:
        python extra/optimization/extract_dataset.py
        gzip -c /tmp/sops > extra/datasets/sops.gz
        #DEBUG=1 MIN_ASTS=1 python extra/optimization/get_action_space.py
-    - name: Repo line count < 24000 lines
-      run: MAX_LINE_COUNT=24000 python sz.py
-
-  spec:
-    strategy:
-      fail-fast: false
-      matrix:
-        group: [1, 2]
-    name: SPEC=2 (${{ matrix.group }})
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: spec-unit
-        deps: testing_unit
-        python-version: '3.14'
-    - name: Test SPEC=2
-      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+    - name: Repo line count < 18000 lines
+      run: MAX_LINE_COUNT=18000 python sz.py

  fuzzing:
    name: Fuzzing
@@ -349,10 +307,12 @@ jobs:
        deps: testing_unit
    - name: Fuzz Test symbolic
      run: python test/external/fuzz_symbolic.py
-    - name: Fuzz Test symbolic (symbolic divisors)
-      run: python test/external/fuzz_symbolic_symbolic_div.py
    - name: Fuzz Test fast idiv
      run: python test/external/fuzz_fast_idiv.py
+    - name: Fuzz Test shapetracker
+      run: CNT=50 python test/external/fuzz_shapetracker.py
+    - name: Fuzz Test shapetracker math
+      run: CNT=200 python test/external/fuzz_shapetracker_math.py
    - name: Fuzz Test shape ops
      run: python test/external/fuzz_shape_ops.py

@@ -367,13 +327,12 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: gpu-image
-          deps: testing_unit
+          deps: testing_minimal
          opencl: 'true'
-      - name: Test CL IMAGE=1 ops
+      - name: Test CL IMAGE=2 ops + training
        run: |
-          CL=1 IMAGE=1 python -m pytest -n=auto test/backend/test_ops.py --durations=20
-        # TODO: training is broken
-        # CL=1 IMAGE=1 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
+          CL=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
+          CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -388,14 +347,14 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: gen-dataset
-          deps: testing
+          deps: testing_minimal
          opencl: 'true'
      - name: Generate Dataset
        run: CL=1 extra/optimization/generate_dataset.sh
      - name: Run Kernel Count Test
        run: CL=1 python -m pytest -n=auto test/external/external_test_opt.py
      - name: Run fused optimizer tests
-        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/backend/test_optim.py -k "not muon"
+        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py
      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
@@ -418,13 +377,17 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1486 ALLOWED_GATED_READ_IMAGE=17 FLOAT16=1 CL=1 IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
-      - name: Test openpilot CL compile fp16
-        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
-      - name: Test openpilot CL compile fp32 (test correctness)
-        run: CL=1 IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
-      - name: Test openpilot LLVM compile fp16
-        run: IMAGE=1 FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=190 ALLOWED_READ_IMAGE=2081 ALLOWED_GATED_READ_IMAGE=28 FLOAT16=0 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx
+      - name: Test openpilot alt model correctness (float32)
+        run: FLOAT16=0 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/3799fe46b3a629e491d4b8498b8ae83e4c88c304/selfdrive/modeld/models/supercombo.onnx
+      - name: Test openpilot fastvits model correctness (float32)
+        run: FLOAT16=0 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+      # - name: Test openpilot simple_plan vision model correctness (float32)
+      #   run: FLOAT16=0 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/35ff4f4577002f2685e50c8346addae33fe8da27a41dd4d6a0f14d1f4b1af81b
+      - name: Test openpilot LLVM compile
+        run: CPU=1 CPU_LLVM=1 LLVMOPT=1 JIT=2 BEAM=0 IMAGE=0 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+      - name: Test openpilot compile4
+        run: NOLOCALS=1 CL=1 IMAGE=2 FLOAT16=1 DEBUG=2 python3 examples/openpilot/compile4.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -443,7 +406,7 @@ jobs:
        with:
          key: onnxoptc
          deps: testing
-          python-version: '3.12'
+          python-version: '3.11'
          llvm: 'true'
      - name: Test ONNX (CPU)
        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -454,7 +417,7 @@ jobs:
      - name: Test Additional ONNX Ops (CPU)
        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_ops.py
      - name: Test Quantize ONNX
-        run: CPU=1 CPU_LLVM=0 python3 test/backend/test_quantize_onnx.py
+        run: CPU=1 CPU_LLVM=0 python3 test/test_quantize_onnx.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -470,8 +433,8 @@ jobs:
        with:
          key: onnxoptl
          deps: testing
-          pydeps: "tensorflow==2.19"
-          python-version: '3.12'
+          pydeps: "tensorflow==2.15.1 tensorflow_addons"
+          python-version: '3.11'
          opencl: 'true'
      - name: Test ONNX (CL)
        run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -484,11 +447,11 @@ jobs:
      - name: Test MLPerf stuff
        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
      - name: NULL=1 beautiful_mnist_multigpu
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
+        run: NULL=1 python examples/beautiful_mnist_multigpu.py
      - name: Test Bert training
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=8 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -496,8 +459,6 @@ jobs:
    name: Test LLM
    runs-on: ubuntu-24.04
    timeout-minutes: 15
-    env:
-      CHECK_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -505,13 +466,8 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: apps_llm
-      - name: Test 1B LLM (llama)
-        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model llama3.2:1b | tee /dev/stderr | grep -i rooster
-      - name: Test 1B LLM (llama q4)
-        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model llama3.2:1b-q4 | tee /dev/stderr | grep -i rooster
-      - name: Test 1B LLM (qwen)
-        # NOTE: qwen is dumb and only knows about female chickens
-        run: echo "What's a female chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model qwen3:0.6b | tee /dev/stderr | grep -i hen
+      - name: Test 1B LLM
+        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm | grep -i rooster

 # ****** Models Tests ******

@@ -550,7 +506,7 @@ jobs:
      with:
        key: metal
        deps: testing
-        python-version: '3.12'
+        python-version: '3.11'
    - name: Test models (Metal)
      run: METAL=1 python -m pytest -n=auto test/models --durations=20
    - name: Test LLaMA compile speed
@@ -569,15 +525,15 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: devectorize-minimal
-        deps: testing_unit
+        deps: testing_minimal
        pydeps: "pillow"
        llvm: "true"
    - name: Test LLVM=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
+      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
    - name: Test LLVM=1 DEVECTORIZE=0 for model
      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
    - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
+      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"

  testdsp:
    name: Linux (DSP)
@@ -590,8 +546,8 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: dsp-minimal
-        deps: testing_unit
-        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
+        deps: testing_minimal
+        pydeps: "onnx==1.18.0 onnxruntime pillow"
        llvm: "true"
    - name: Set up Docker Buildx
      uses: docker/setup-buildx-action@v3
@@ -603,15 +559,15 @@ jobs:
        load: true
        tags: qemu-hexagon:latest
        cache-from: type=gha
-        cache-to: ${{ github.event_name != 'pull_request' && 'type=gha,mode=min' || '' }}
+        cache-to: type=gha,mode=min
    - name: Set MOCKDSP env
      run: printf "MOCKDSP=1" >> $GITHUB_ENV
    - name: Run test_tiny on DSP
      run: DEBUG=2 DSP=1 python test/test_tiny.py
    - name: Test transcendentals
-      run: CC=clang-20 DEBUG=2 DSP=1 python test/backend/test_transcendental.py TestTranscendentalVectorized
+      run: CC=clang-20 DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
    - name: Test quantize onnx
-      run: DEBUG=2 DSP=1 python3 test/backend/test_quantize_onnx.py
+      run: DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py

  testwebgpu:
    name: Linux (WebGPU)
@@ -624,105 +580,34 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: webgpu-minimal
-        deps: testing_unit
-        python-version: '3.12'
+        deps: testing_minimal
+        python-version: '3.11'
        webgpu: 'true'
    - name: Check Device.DEFAULT (WEBGPU) and print some source
      run: |
        WEBGPU=1 python -c "from tinygrad import Device; assert Device.DEFAULT == 'WEBGPU', Device.DEFAULT"
-        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
    - name: Run selected webgpu tests
      run: |
-          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/backend --durations=20
+          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit \
+          --ignore=test/test_copy_speed.py --ignore=test/test_rearrange_einops.py \
+          --ignore=test/test_fuzz_shape_ops.py --durations=20
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

-  testamdasm:
-    name: AMD ASM IDE
-    runs-on: ubuntu-24.04
-    timeout-minutes: 20
-    env:
-      AMD: 1
-      PYTHON_REMU: 1
-      MOCKGPU: 1
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: rdna3-emu
-          deps: testing_unit
-          amd: 'true'
-          python-version: '3.14'
-      - name: Verify AMD autogen is up to date
-        run: |
-          python -m tinygrad.renderer.amd.generate
-          git diff --exit-code tinygrad/runtime/autogen/amd/
-      - name: Install LLVM 21
-        run: |
-          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
-          echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
-          sudo apt-get update
-          sudo apt-get install llvm-21 llvm-21-tools cloc
-      - name: Install rocprof-trace-decoder
-        run: sudo PYTHONPATH="." ./extra/sqtt/install_rocprof_decoder.py
-      - name: Run AMD renderer tests
-        run: AMD_LLVM=0 python -m pytest -n=auto test/amd/ --durations 20
-      - name: Run AMD renderer tests (AMD_LLVM=1)
-        run: AMD_LLVM=1 python -m pytest -n=auto test/amd/ --durations 20
-      - name: Run SQTT profiling tests
-        run: PROFILE=1 SQTT=1 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
-      - name: Run AMD emulated tests on NULL backend
-        env:
-          AMD: 0
-        run: |
-          PYTHONPATH=. NULL=1 EMULATE=AMD python extra/mmapeak/mmapeak.py
-          PYTHONPATH=. NULL=1 EMULATE=AMD_CDNA4 python3 -m pytest -n=auto test/testextra/test_tk.py test/backend/test_asm_gemm.py
-      - name: Run ASM matmul on MOCKGPU
-        run: PYTHONPATH="." AMD=1 MOCKGPU=1 N=256 python3 extra/gemm/amd_asm_matmul.py
-      - name: Run LLVM test
-        run: AMD_LLVM=1 python test/device/test_amd_llvm.py
-
-  testmockam:
-    name: Linux (am)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    env:
-      AMD: 1
-      MOCKGPU: 1
-      AMD_IFACE: PCI
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: mockam
-          deps: testing_unit
-          amd: 'true'
-      - name: Run test_tiny on MOCKAM
-        run: python test/test_tiny.py
-      - name: Run test_tiny on MOCKAM USB
-        run: GMMU=0 AMD_IFACE=USB python test/test_tiny.py
-      - name: Run test_hcq on MOCKAM
-        run: python -m pytest test/device/test_hcq.py
-
  testamd:
    strategy:
      fail-fast: false
      matrix:
        backend: [amd, amdllvm]
-        arch: [rdna3, rdna4, cdna4]

-    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
+    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
-    timeout-minutes: 15
+    timeout-minutes: 20
    env:
      AMD: 1
      MOCKGPU: 1
-      MOCKGPU_ARCH: ${{ matrix.arch }}
-      SKIP_SLOW_TEST: 1
+      FORWARD_ONLY: 1
      AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
    steps:
      - name: Checkout Code
@@ -731,17 +616,26 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
          amd: 'true'
          llvm: ${{ matrix.backend == 'amdllvm' && 'true' }}
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
+      - name: Run LLVM test
+        if: matrix.backend=='amdllvm'
+        run: python test/device/test_amd_llvm.py
      - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
+        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py --durations=20
+      - name: Run pytest (amd)
+        run: python -m pytest test/external/external_test_am.py --durations=20
      - name: Run TRANSCENDENTAL math
-        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
+        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
+      - name: Run TestOps.test_add with SQTT
+        run: |
+          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -764,7 +658,7 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
          cuda: 'true'
          ocelot: 'true'
      - name: Set env
@@ -772,12 +666,10 @@ jobs:
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CUDA','NV'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (cuda)
        # skip multitensor because it's slow
-        run: python -m pytest -n=auto test/backend --ignore test/backend/test_multitensor.py --durations=20
-      - name: Run TestOps.test_add with PMA
-        run: VIZ=-1 PMA=1 DEBUG=5 python3 test/backend/test_ops.py TestOps.test_add
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -785,7 +677,7 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        backend: [llvm, cpu, opencl, lvp]
+        backend: [llvm, cpu, opencl]

    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
@@ -797,23 +689,87 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
          opencl: ${{ matrix.backend == 'opencl' && 'true' }}
-          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
+          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (${{ matrix.backend }})
-        run: python -m pytest -n=auto test/backend --durations=20
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
      - name: Run TRANSCENDENTAL math
-        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
+        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

+  amdremote:
+    name: Linux (remote)
+    runs-on: ubuntu-22.04
+    timeout-minutes: 20
+    env:
+      REMOTE: 1
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: linux-remote
+          deps: testing_minimal
+          amd: 'true'
+          llvm: 'true'
+          opencl: 'true'
+      - name: Start remote server
+        run: |
+          start_server() {
+            systemd-run --user \
+              --unit="$1" \
+              --setenv=REMOTEDEV="$2" \
+              --setenv=MOCKGPU=1 \
+              --setenv=PYTHONPATH=. \
+              --setenv=PORT="$3" \
+              --working-directory="$(pwd)" \
+              python tinygrad/runtime/ops_remote.py
+          }
+
+          start_server "remote-server-amd-1" "AMD" 6667
+          start_server "remote-server-amd-2" "AMD" 6668
+          start_server "remote-server-gpu" "CL" 7667
+          start_server "remote-server-cpu" "CPU" 8667
+      - name: Check Device.DEFAULT and print some source
+        env:
+          HOST: 127.0.0.1:6667*6,127.0.0.1:6668*6
+        run: |
+          python -c "from tinygrad import Device; assert Device.DEFAULT == 'REMOTE', Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.default.properties.real_device == 'AMD', Device.default.properties.real_device"
+          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
+      - name: Run REMOTE=1 Test (AMD)
+        env:
+          HOST: 127.0.0.1:6667*6,127.0.0.1:6668*6
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_remote.py test/test_tensor_variable.py --durations 20
+      - name: Run REMOTE=1 Test (CL)
+        env:
+          HOST: 127.0.0.1:7667*6
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_image_dtype.py test/test_jit.py --durations 20
+          IMAGE=2 python3 -m pytest test/test_tiny.py test/test_image_dtype.py
+      - name: Run REMOTE=1 Test (CPU)
+        env:
+          HOST: 127.0.0.1:8667*6
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py --durations 20
+      - name: Show remote server logs
+        if: always()
+        run: |
+          journalctl --user -u remote-server-amd-1 --no-pager
+          journalctl --user -u remote-server-amd-2 --no-pager
+          journalctl --user -u remote-server-gpu --no-pager
+          journalctl --user -u remote-server-cpu --no-pager
+
 # ****** OSX Tests ******

  testmetal:
@@ -828,29 +784,25 @@ jobs:
      with:
        key: metal
        deps: testing
-        python-version: '3.12'
+        python-version: '3.11'
        amd: 'true'
        cuda: 'true'
        ocelot: 'true'
        llvm: 'true'
    - name: Run unit tests
      run: METAL=1 python -m pytest -n=auto test/unit/ --durations=20
-    - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
    - name: Run ONNX
      run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
    - name: Test tensor core ops (fake)
-      run: METAL=1 DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
+      run: METAL=1 DEBUG=3 TC=2 python test/test_ops.py TestOps.test_gemm
    - name: Test tensor core ops (real)
-      run: METAL=1 DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
+      run: METAL=1 DEBUG=3 python test/test_ops.py TestOps.test_big_gemm
    - name: Test Beam Search
      run: METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
-    - name: Test Device Specific
-      run: METAL=1 python3 -m pytest test/device/test_metal.py
    #- name: Fuzz Test linearizer
    #  run: METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
    - name: Run TRANSCENDENTAL math
-      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
+      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
    - name: Run pytest (amd)
      env:
        MOCKGPU: 1
@@ -873,8 +825,6 @@ jobs:
        NV_PTX: 1
        NV: 1
        FORWARD_ONLY: 1
-        # TODO: failing due to library loading error
-        CAPTURE_PROCESS_REPLAY: 0
      run: |
        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
    - name: Run process replay tests
@@ -893,14 +843,14 @@ jobs:
        key: osx-webgpu
        deps: testing
        webgpu: 'true'
+    - name: Test infinity math in WGSL
+      run: WEBGPU=1 python -m pytest -n=auto test/test_renderer_failures.py::TestWGSLFailures::test_multiply_infinity --durations=20
    - name: Build WEBGPU Efficientnet
      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m examples.compile_efficientnet
-    - name: Run selected webgpu tests
-      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m pytest -n=auto test/backend --durations=20
-    #- name: Clean npm cache
-    #  run: npm cache clean --force
-    #- name: Install Puppeteer
-    #  run: npm install puppeteer
+    - name: Clean npm cache
+      run: npm cache clean --force
+    - name: Install Puppeteer
+      run: npm install puppeteer
    # this is also flaky
    #- name: Run WEBGPU Efficientnet
    #  run: node test/web/test_webgpu.js
@@ -917,11 +867,35 @@ jobs:
    - name: Test ONNX Runner (WEBGPU)
      run: WEBGPU=1 python3 test/external/external_test_onnx_runner.py

+  osxremote:
+   name: MacOS (remote metal)
+   runs-on: macos-15
+   timeout-minutes: 10
+   env:
+     REMOTE: 1
+     REMOTEDEV: METAL
+   steps:
+     - name: Checkout Code
+       uses: actions/checkout@v4
+     - name: Setup Environment
+       uses: ./.github/actions/setup-tinygrad
+       with:
+         key: macos-remote
+         deps: testing_minimal
+     - name: Check Device.DEFAULT and print some source
+       run: |
+         python -c "from tinygrad import Device; assert Device.DEFAULT == 'REMOTE', Device.DEFAULT"
+         python -c "from tinygrad import Device; assert Device.default.properties.real_device == 'METAL', Device.default.properties.real_device"
+         DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
+     - name: Run REMOTE=1 Test
+       run: |
+         python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_tensor_variable.py
+
  osxtests:
    strategy:
      fail-fast: false
      matrix:
-        backend: [metal, llvm, cpu, lvp]
+        backend: [metal, llvm, cpu]
    name: MacOS (${{ matrix.backend }})
    runs-on: macos-15
    timeout-minutes: 20
@@ -932,17 +906,17 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: macos-${{ matrix.backend }}-minimal
-          deps: testing_unit
-          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
+          deps: testing_minimal
+          pydeps: "capstone"
+          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1'}}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
      - name: Run pytest (${{ matrix.backend }})
-        run: python3 -m pytest -n=auto test/backend --durations=20
+        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
      - name: Run macOS-specific unit test
@@ -975,66 +949,9 @@ jobs:
      - name: Run unit tests
        if: matrix.backend=='llvm'
        # test_newton_schulz hits RecursionError
-        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
-      - name: Run NULL backend tests
-        if: matrix.backend=='llvm'
-        shell: bash
-        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
+        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_elf.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
      - name: Run pytest (${{ matrix.backend }})
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
-          python -m pytest -n=auto test/test_tiny.py test/backend/test_ops.py --durations=20
-
-# ****** Compile-only Tests ******
-
-  compiletests:
-    strategy:
-      fail-fast: false
-      matrix:
-        backend: [ir3, nak]
-    name: Compile-only (${{ matrix.backend }})
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-${{ matrix.backend }}
-          deps: testing_unit
-          mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
-          python-version: '3.12'
-      - name: Set env
-        shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-  qcomclcompiletests:
-    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-qcomcl
-          deps: testing_unit
-          tinydreno: 'true'
-          python-version: '3.12'
-      - name: Set env
-        shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\nNULL_QCOMCL=1" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          python -m pytest -n=auto test/test_tiny.py test/test_ops.py --durations=20
@@ -38,7 +38,6 @@ extra/huggingface_onnx/models/*
 extra/huggingface_onnx/*.yaml
 extra/weights
 venv
-venv_sd_mlperf
 examples/**/net.*[js,json]
 examples/**/*.safetensors
 node_modules
@@ -58,13 +57,8 @@ weights
 *.lprof
 comgr_*
 *.pkl
-!extra/sqtt/examples/**/*.pkl
 site/
 profile_stats
 *.log
 target
 .mypy_cache
-mutants
-.mutmut-cache
-dagre/
-graphlib/
@@ -16,19 +16,25 @@ repos:
        pass_filenames: false
      - id: mypy
        name: mypy
-        entry: python3 -m mypy
-        language: system
-        always_run: true
-        pass_filenames: false
-      - id: example
-        name: test all devices
-        entry: python3 test/external/external_test_example.py
+        entry: python3 -m mypy tinygrad/ --strict-equality
        language: system
        always_run: true
        pass_filenames: false
      - id: tests
-        name: comprehensive test suite
-        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/backend/test_ops.py test/backend/test_schedule.py test/unit/test_assign.py test/backend/test_tensor.py test/backend/test_jit.py test/unit/test_schedule_cache.py test/null/test_pattern_matcher.py test/null/test_uop_symbolic.py test/unit/test_helpers.py
+        name: subset of tests
+        entry: env PYTHONPATH="." python3 -m pytest -n=4 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
        language: system
        always_run: true
        pass_filenames: false
+      - id: example
+        name: multi device tests
+        entry: python3 test/external/external_test_example.py
+        language: system
+        always_run: true
+        pass_filenames: false
+      - id: pylint
+        name: pylint
+        entry: python3 -m pylint tinygrad/
+        language: system
+        always_run: true
+        pass_filenames: false
@@ -0,0 +1,17 @@
+# tinygrad agents
+
+Hello agent. You are one of the most talented programmers of your generation.
+
+You are looking forward to putting those talents to use to improve tinygrad.
+
+## philosophy
+
+tinygrad is a **tensor** library focused on beauty and minimalism, while still matching the functionality of PyTorch and JAX.
+
+Every line must earn its keep. Prefer readability over cleverness. We believe that if carefully designed, 10 lines can have the impact of 1000.
+
+Never mix functionality changes with whitespace changes. All functionality changes must be tested.
+
+## style
+
+Use **2-space indentation**, and keep lines to a maximum of **150 characters**. Match the existing style.
@@ -21,38 +21,17 @@ tinygrad: For something between [PyTorch](https://github.com/pytorch/pytorch) an

 ---

-tinygrad is an end-to-end deep learning stack:
+Despite tinygrad's size, it is a fully featured deep learning framework.

- **Tensor library** with autograd
- **IR and compiler** that fuse and lower kernels
- **JIT + graph execution**
- **nn / optim / datasets** for real training
+Due to its extreme simplicity, it is the easiest framework to add new accelerators to, with support for both inference and training. If XLA is CISC, tinygrad is RISC.

-It’s inspired by PyTorch (ergonomics), JAX (functional transforms and IR-based AD), and TVM (scheduling and codegen), but stays intentionally tiny and hackable.
+tinygrad is now beta software, we [raised some money](https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html) to make it good. Someday, we will tape out chips.

---
+## Features

-## How tinygrad compares
+### LLaMA and Stable Diffusion

-**PyTorch**
-
- ✅ Similar: eager `Tensor` API, autograd, `optim`, basic datasets and layers.
- ✅ You can write familiar training loops.
- 🔁 Unlike PyTorch, the entire compiler and IR are visible and hackable.
-
-**JAX**
-
- ✅ IR-based autodiff over primitives (like JAXPR + XLA).
- ✅ Function-level JIT (`TinyJit`) that captures and replays kernels.
- 🔁 Fewer functional transforms (no full `vmap`/`pmap` yet), but far easier to read.
-
-**TVM**
-
- ✅ Multiple lowering passes, scheduling, and BEAM search over kernels.
- ✅ Device “graphs” for batched execution.
- 🔁 tinygrad also ships the **front-end framework** (tensors, nn, optim), not just the compiler.
-
---
+tinygrad can run [LLaMA](/docs/showcase.md#llama) and [Stable Diffusion](/docs/showcase.md#stable-diffusion)!

 ### Laziness

@@ -192,7 +171,7 @@ For more examples on how to run the full test suite please refer to the [CI work
 Some examples of running tests locally:
 ```sh
 python3 -m pip install -e '.[testing]'  # install extra deps for testing
-python3 test/backend/test_ops.py        # just the ops tests
+python3 test/test_ops.py                # just the ops tests
 python3 -m pytest test/                 # whole test suite
 ```

@@ -0,0 +1,489 @@
+#!/bin/bash -e
+
+# setup instructions for clang2py
+if [[ ! $(clang2py -V) ]]; then
+  pushd .
+  cd /tmp
+  sudo apt-get install -y --no-install-recommends clang
+  pip install --upgrade pip setuptools
+  pip install clang==14.0.6
+  git clone https://github.com/nimlgen/ctypeslib.git
+  cd ctypeslib
+  pip install .
+  clang2py -V
+  popd
+fi
+
+BASE=tinygrad/runtime/autogen/
+
+fixup() {
+  sed -i '1s/^/# mypy: ignore-errors\n/' $1
+  sed -i 's/ *$//' $1
+  grep FIXME_STUB $1 || true
+}
+
+patch_dlopen() {
+  path=$1; shift
+  name=$1; shift
+  cat <<EOF | sed -i "/import ctypes.*/r /dev/stdin" $path
+PATHS_TO_TRY = [
+$(for p in "$@"; do echo "  $p,"; done)
+]
+def _try_dlopen_$name():
+  library = ctypes.util.find_library("$name")
+  if library: return ctypes.CDLL(library)
+  for candidate in PATHS_TO_TRY:
+    try: return ctypes.CDLL(candidate)
+    except OSError: pass
+  return None
+EOF
+}
+
+generate_opencl() {
+  clang2py /usr/include/CL/cl.h -o $BASE/opencl.py -l /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 -k cdefstum
+  fixup $BASE/opencl.py
+  # hot patches
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/opencl.py
+  sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libOpenCL.so.1')\ctypes.CDLL(ctypes.util.find_library('OpenCL'))\g" $BASE/opencl.py
+  python3 -c "import tinygrad.runtime.autogen.opencl"
+}
+
+generate_hip() {
+  clang2py /opt/rocm/include/hip/hip_ext.h /opt/rocm/include/hip/hiprtc.h \
+  /opt/rocm/include/hip/hip_runtime_api.h /opt/rocm/include/hip/driver_types.h \
+  --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/hip.py -l /opt/rocm/lib/libamdhip64.so
+  echo "hipDeviceProp_t = hipDeviceProp_tR0600" >> $BASE/hip.py
+  echo "hipGetDeviceProperties = hipGetDevicePropertiesR0600" >> $BASE/hip.py
+  fixup $BASE/hip.py
+  # we can trust HIP is always at /opt/rocm/lib
+  #sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/hip.py
+  #sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhiprtc.so')\ctypes.CDLL(ctypes.util.find_library('hiprtc'))\g" $BASE/hip.py
+  #sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamdhip64.so')\ctypes.CDLL(ctypes.util.find_library('amdhip64'))\g" $BASE/hip.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/hip.py
+  sed -i "s\'/opt/rocm/\os.getenv('ROCM_PATH', '/opt/rocm/')+'/\g" $BASE/hip.py
+  python3 -c "import tinygrad.runtime.autogen.hip"
+}
+
+generate_comgr() {
+  clang2py /opt/rocm/include/amd_comgr/amd_comgr.h \
+  --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/comgr.py -l /opt/rocm/lib/libamd_comgr.so
+  fixup $BASE/comgr.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/comgr.py
+  patch_dlopen $BASE/comgr.py amd_comgr "'/opt/rocm/lib/libamd_comgr.so'" "os.getenv('ROCM_PATH', '')+'/lib/libamd_comgr.so'" "'/usr/local/lib/libamd_comgr.dylib'" "'/opt/homebrew/lib/libamd_comgr.dylib'"
+  sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamd_comgr.so')\_try_dlopen_amd_comgr()\g" $BASE/comgr.py
+  python3 -c "import tinygrad.runtime.autogen.comgr"
+}
+
+generate_kfd() {
+  clang2py /usr/include/linux/kfd_ioctl.h -o $BASE/kfd.py -k cdefstum
+
+  fixup $BASE/kfd.py
+  sed -i "s/import ctypes/import ctypes, os/g" $BASE/kfd.py
+  sed -i "s/import fcntl, functools/import functools/g" $BASE/kfd.py
+  sed -i "/import functools/a from tinygrad.runtime.support.hcq import FileIOInterface" $BASE/kfd.py
+  sed -i "s/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd, \*\*kwargs):/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd:FileIOInterface, \*\*kwargs):/g" $BASE/kfd.py
+  sed -i "s/fcntl.ioctl(__fd, (__idir<<30)/__fd.ioctl((__idir<<30)/g" $BASE/kfd.py
+  sed -i "s/!!/not not /g" $BASE/kfd.py
+  python3 -c "import tinygrad.runtime.autogen.kfd"
+}
+
+generate_cuda() {
+  clang2py /usr/include/cuda.h --clang-args="-D__CUDA_API_VERSION_INTERNAL" -o $BASE/cuda.py -l /usr/lib/x86_64-linux-gnu/libcuda.so
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/cuda.py
+  sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libcuda.so')\ctypes.CDLL(ctypes.util.find_library('cuda'))\g" $BASE/cuda.py
+  fixup $BASE/cuda.py
+  python3 -c "import tinygrad.runtime.autogen.cuda"
+}
+
+generate_nvrtc() {
+  clang2py /usr/local/cuda/include/nvrtc.h /usr/local/cuda/include/nvJitLink.h -o $BASE/nvrtc.py -l /usr/local/cuda/lib64/libnvrtc.so -l /usr/local/cuda/lib64/libnvJitLink.so
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/nvrtc.py
+  sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvrtc.so')\ctypes.CDLL(ctypes.util.find_library('nvrtc'))\g" $BASE/nvrtc.py
+  sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvJitLink.so')\ctypes.CDLL(ctypes.util.find_library('nvJitLink'))\g" $BASE/nvrtc.py
+  fixup $BASE/nvrtc.py
+  python3 -c "import tinygrad.runtime.autogen.nvrtc"
+}
+
+generate_nv() {
+  NVKERN_COMMIT_HASH=81fe4fb417c8ac3b9bdcc1d56827d116743892a5
+  NVKERN_SRC=/tmp/open-gpu-kernel-modules-$NVKERN_COMMIT_HASH
+  if [ ! -d "$NVKERN_SRC" ]; then
+    git clone https://github.com/NVIDIA/open-gpu-kernel-modules $NVKERN_SRC
+    pushd .
+    cd $NVKERN_SRC
+    git reset --hard $NVKERN_COMMIT_HASH
+    popd
+  fi
+
+  clang2py -k cdefstum \
+    extra/nv_gpu_driver/clc6c0qmd.h \
+    extra/nv_gpu_driver/clcec0qmd.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0000.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0080.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080_notification.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc86f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc96f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc761.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl83de.h \
+    $NVKERN_SRC/src/nvidia/generated/g_allclasses.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc6c0.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clcdc0.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/clc6b5.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/clc9b5.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_ioctl.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_linux_ioctl.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/hwref/ampere/ga100/dev_fault.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv_escape.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numbers.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numa.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-unix-nvos-params-wrappers.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/alloc/alloc_channel.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/nvos.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0000/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0080/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl2080/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl83de/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlc36f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlcb33.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrla06c.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl90f1.h \
+    --clang-args="-include $NVKERN_SRC/src/common/sdk/nvidia/inc/nvtypes.h -I$NVKERN_SRC/src/common/inc -I$NVKERN_SRC/kernel-open/nvidia-uvm -I$NVKERN_SRC/kernel-open/common/inc -I$NVKERN_SRC/src/common/sdk/nvidia/inc -I$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include -I$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl" \
+    -o $BASE/nv_gpu.py
+  fixup $BASE/nv_gpu.py
+  sed -i "s\(0000000001)\1\g" $BASE/nv_gpu.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/nv_gpu.py
+  sed -i 's/#\?\s\([A-Za-z0-9_]\+\) = MW ( \([0-9]\+\) : \([0-9]\+\) )/\1 = (\2 , \3)/' $BASE/nv_gpu.py # NVC6C0_QMDV03_00 processing
+  sed -i 's/#\sdef NVC6C0_QMD\([A-Za-z0-9_()]\+\):/def NVC6C0_QMD\1:/' $BASE/nv_gpu.py
+  sed -i 's/#\sdef NVCEC0_QMD\([A-Za-z0-9_()]\+\):/def NVCEC0_QMD\1:/' $BASE/nv_gpu.py
+  sed -E -i -n '/^def (NVCEC0_QMDV05_00_RELEASE)(_ENABLE)\(i\):/{p;s//\1'"0"'\2=\1\2(0)\n\1'"1"'\2=\1\2(1)/;H;b};p;${x;s/^\n//;p}' "$BASE/nv_gpu.py"
+  sed -i 's/#\s*return MW(\([0-9i()*+]\+\):\([0-9i()*+]\+\))/    return (\1 , \2)/' $BASE/nv_gpu.py
+  sed -i 's/#\?\s*\(.*\)\s*=\s*\(NV\)\?BIT\(32\)\?\s*(\s*\([0-9]\+\)\s*)/\1 = (1 << \4)/' $BASE/nv_gpu.py # name = BIT(x) -> name = (1 << x)
+  sed -i "s/UVM_\([A-Za-z0-9_]\+\) = \['i', '(', '\([0-9]\+\)', ')'\]/UVM_\1 = \2/" $BASE/nv_gpu.py # UVM_name = ['i', '(', '<num>', ')'] -> UVM_name = <num>
+
+  # Parse status codes
+  sed -n '1i\
+nv_status_codes = {}
+/^NV_STATUS_CODE/ { s/^NV_STATUS_CODE(\([^,]*\), *\([^,]*\), *"\([^"]*\)") *.*$/\1 = \2\nnv_status_codes[\1] = "\3"/; p }' $NVKERN_SRC/src/common/sdk/nvidia/inc/nvstatuscodes.h >> $BASE/nv_gpu.py
+  python3 -c "import tinygrad.runtime.autogen.nv_gpu"
+
+  clang2py -k cdefstum \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/fsp/kern_fsp_cot_payload.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gspifpub.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gsp_fw_wpr_meta.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gsp_fw_sr_meta.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/gsp/gsp_init_args.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/gsp/gsp_init_args.h \
+    $NVKERN_SRC/src/common/uproc/os/common/include/libos_init_args.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/rmRiscvUcode.h \
+    $NVKERN_SRC/src/common/shared/msgq/inc/msgq/msgq_priv.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/vgpu/rpc_headers.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/vgpu/rpc_global_enums.h \
+    $NVKERN_SRC/src/nvidia/generated/g_rpc-structures.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/fsp/fsp_nvdm_format.h \
+    extra/nv_gpu_driver/g_rpc-message-header.h \
+    extra/nv_gpu_driver/gsp_static_config.h \
+    extra/nv_gpu_driver/vbios.h \
+    --clang-args="-DRPC_MESSAGE_STRUCTURES -DRPC_STRUCTURES -include $NVKERN_SRC/src/common/sdk/nvidia/inc/nvtypes.h -I$NVKERN_SRC/src/nvidia/generated -I$NVKERN_SRC/src/common/inc -I$NVKERN_SRC/src/nvidia/inc -I$NVKERN_SRC/src/nvidia/interface/ -I$NVKERN_SRC/src/nvidia/inc/kernel -I$NVKERN_SRC/src/nvidia/inc/libraries -I$NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc -I$NVKERN_SRC/kernel-open/nvidia-uvm -I$NVKERN_SRC/kernel-open/common/inc -I$NVKERN_SRC/src/common/sdk/nvidia/inc -I$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include -I$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl" \
+    -o $BASE/nv/nv.py
+
+  fixup $BASE/nv/nv.py
+  python3 -c "import tinygrad.runtime.autogen.nv.nv"
+}
+
+generate_amd() {
+  # clang2py broken when pass -x c++ to prev headers
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    extra/hip_gpu_driver/nvd.h \
+    extra/hip_gpu_driver/gc_11_0_0_offset.h \
+    extra/hip_gpu_driver/sienna_cichlid_ip_offset.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/amd_gpu.py
+
+  fixup $BASE/amd_gpu.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/amd_gpu.py
+  python3 -c "import tinygrad.runtime.autogen.amd_gpu"
+}
+
+generate_hsa() {
+  clang2py \
+    /opt/rocm/include/hsa/hsa.h \
+    /opt/rocm/include/hsa/hsa_ext_amd.h \
+    /opt/rocm/include/hsa/amd_hsa_signal.h \
+    /opt/rocm/include/hsa/amd_hsa_queue.h \
+    /opt/rocm/include/hsa/amd_hsa_kernel_code.h \
+    /opt/rocm/include/hsa/hsa_ext_finalize.h /opt/rocm/include/hsa/hsa_ext_image.h \
+    /opt/rocm/include/hsa/hsa_ven_amd_aqlprofile.h \
+    --clang-args="-I/opt/rocm/include" \
+    -o $BASE/hsa.py -l /opt/rocm/lib/libhsa-runtime64.so
+
+  fixup $BASE/hsa.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/hsa.py
+  sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhsa-runtime64.so')\ctypes.CDLL(os.getenv('ROCM_PATH')+'/lib/libhsa-runtime64.so' if os.getenv('ROCM_PATH') else ctypes.util.find_library('hsa-runtime64'))\g" $BASE/hsa.py
+  python3 -c "import tinygrad.runtime.autogen.hsa"
+}
+
+generate_io_uring() {
+  clang2py -k cdefstum \
+    /usr/include/liburing.h \
+    /usr/include/linux/io_uring.h \
+    -o $BASE/io_uring.py
+
+  sed -r '/^#define __NR_io_uring/ s/^#define __(NR_io_uring[^ ]+) (.*)$/\1 = \2/; t; d' /usr/include/asm-generic/unistd.h >> $BASE/io_uring.py # io_uring syscalls numbers
+  fixup $BASE/io_uring.py
+}
+
+generate_ib() {
+  clang2py -k cdefstum \
+    /usr/include/infiniband/verbs.h \
+    /usr/include/infiniband/verbs_api.h \
+    /usr/include/infiniband/ib_user_ioctl_verbs.h \
+    /usr/include/rdma/ib_user_verbs.h \
+    -o $BASE/ib.py
+
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" "$BASE/ib.py"
+  sed -i "s\FIXME_STUB\libibverbs\g" "$BASE/ib.py"
+  sed -i "s\FunctionFactoryStub()\ctypes.CDLL(ctypes.util.find_library('ibverbs'), use_errno=True)\g" "$BASE/ib.py"
+
+  fixup $BASE/ib.py
+}
+
+generate_libc() {
+  clang2py -k cdefstum \
+    $(dpkg -L libc6-dev | grep sys/mman.h) \
+    $(dpkg -L libc6-dev | grep sys/syscall.h) \
+    /usr/include/string.h \
+    /usr/include/elf.h \
+    /usr/include/unistd.h \
+    /usr/include/asm-generic/mman-common.h \
+    -o $BASE/libc.py
+
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libc.py
+  sed -i "s\FIXME_STUB\libc\g" $BASE/libc.py
+  sed -i "s\FunctionFactoryStub()\None if (libc_path := ctypes.util.find_library('c')) is None else ctypes.CDLL(libc_path, use_errno=True)\g" $BASE/libc.py
+
+  fixup $BASE/libc.py
+}
+
+generate_llvm() {
+  INC="$(llvm-config-14 --includedir)"
+  clang2py -k cdefstum \
+    $(find "$INC/llvm-c/" -type f -name '*.h' | sort) \
+    "$INC/llvm/Config/Targets.def" \
+    "$INC/llvm/Config/AsmPrinters.def" \
+    "$INC/llvm/Config/AsmParsers.def" \
+    "$INC/llvm/Config/Disassemblers.def" \
+    --clang-args="$(llvm-config-14 --cflags)" \
+    -o "$BASE/llvm.py"
+
+  sed -i "s\import ctypes\import ctypes, tinygrad.runtime.support.llvm as llvm_support\g" "$BASE/llvm.py"
+  sed -i "s\FIXME_STUB\llvm\g" "$BASE/llvm.py"
+  sed -i "s\FunctionFactoryStub()\ctypes.CDLL(llvm_support.LLVM_PATH)\g" "$BASE/llvm.py"
+
+  fixup "$BASE/llvm.py"
+}
+
+generate_kgsl() {
+  clang2py extra/qcom_gpu_driver/msm_kgsl.h -o $BASE/kgsl.py -k cdefstum
+  fixup $BASE/kgsl.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/kgsl.py
+  sed -nE 's/#define ([A-Za-z0-9_]+)_SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1_SHIFT) \& \1_MASK/p' extra/qcom_gpu_driver/msm_kgsl.h >> $BASE/kgsl.py
+  sed -i "s\fcntl.ioctl(__fd, (__idir<<30)\__fd.ioctl((__idir<<30)\g" $BASE/kgsl.py
+  python3 -c "import tinygrad.runtime.autogen.kgsl"
+}
+
+generate_adreno() {
+  clang2py extra/qcom_gpu_driver/a6xx.xml.h -o $BASE/adreno.py -k cestum
+  sed -nE 's/#define ([A-Za-z0-9_]+)__SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1__SHIFT) \& \1__MASK/p' extra/qcom_gpu_driver/a6xx.xml.h >> $BASE/adreno.py
+  fixup $BASE/adreno.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/adreno.py
+  python3 -c "import tinygrad.runtime.autogen.adreno"
+}
+
+generate_qcom() {
+  clang2py -k cdefstum \
+    extra/dsp/include/ion.h \
+    extra/dsp/include/msm_ion.h \
+    extra/dsp/include/adsprpc_shared.h \
+    extra/dsp/include/remote_default.h \
+    extra/dsp/include/apps_std.h \
+    -o $BASE/qcom_dsp.py
+
+  fixup $BASE/qcom_dsp.py
+  python3 -c "import tinygrad.runtime.autogen.qcom_dsp"
+}
+
+generate_pci() {
+  clang2py -k cdefstum \
+    /usr/include/linux/pci_regs.h \
+    -o $BASE/pci.py
+  fixup $BASE/pci.py
+}
+
+generate_vfio() {
+  clang2py -k cdefstum \
+    /usr/include/linux/vfio.h \
+    -o $BASE/vfio.py
+  fixup $BASE/vfio.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/vfio.py
+  sed -i "s\import fcntl, functools\import functools" $BASE/vfio.py
+  sed -i "s\import ctypes,os\a from tinygrad.runtime.support import FileIOInterface\g" $BASE/vfio.py
+  sed -i "s\fcntl.ioctl(__fd, (__idir<<30)\return __fd.ioctl((__idir<<30)\g" $BASE/vfio.py
+}
+
+generate_am() {
+  AMKERN_COMMIT_HASH=ceb12c04e2b5b53ec0779362831f5ee40c4921e4
+  AMKERN_SRC=/tmp/ROCK-Kernel-Driver-$AMKERN_COMMIT_HASH
+  if [ ! -d "$AMKERN_SRC" ]; then
+    git clone https://github.com/ROCm/ROCK-Kernel-Driver $AMKERN_SRC --depth 1
+  fi
+  AMKERN_AMD=$AMKERN_SRC/drivers/gpu/drm/amd/
+  AMKERN_INC=$AMKERN_AMD/include/
+
+  clang2py -k cdefstum \
+    extra/amdpci/headers/v11_structs.h \
+    extra/amdpci/headers/v12_structs.h \
+    extra/amdpci/headers/amdgpu_vm.h \
+    extra/amdpci/headers/discovery.h \
+    extra/amdpci/headers/amdgpu_ucode.h \
+    extra/amdpci/headers/psp_gfx_if.h \
+    extra/amdpci/headers/amdgpu_psp.h \
+    extra/amdpci/headers/amdgpu_irq.h \
+    extra/amdpci/headers/amdgpu_doorbell.h \
+    $AMKERN_INC/soc15_ih_clientid.h \
+    --clang-args="-include stdint.h" \
+    -o $BASE/am/am.py
+  fixup $BASE/am/am.py
+  sed -i "s\(int64_t)\ \g" $BASE/am/am.py
+  sed -i "s\AMDGPU_PTE_MTYPE_VG10(2)\AMDGPU_PTE_MTYPE_VG10(0, 2)\g" $BASE/am/am.py # incorrect parsing (TODO: remove when clang2py is gone).
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/amdkfd/kfd_pm4_headers_ai.h \
+    $AMKERN_AMD/amdgpu/soc15d.h \
+    -o $BASE/am/pm4_soc15.py
+  fixup $BASE/am/pm4_soc15.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/amdkfd/kfd_pm4_headers_ai.h \
+    $AMKERN_AMD/amdgpu/nvd.h \
+    -o $BASE/am/pm4_nv.py
+  fixup $BASE/am/pm4_nv.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/vega10_sdma_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_4_0_0.py
+  fixup $BASE/am/sdma_4_0_0.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/navi10_sdma_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_5_0_0.py
+  fixup $BASE/am/sdma_5_0_0.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/sdma_v6_0_0_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_6_0_0.py
+  fixup $BASE/am/sdma_6_0_0.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v13_0_0_ppsmc.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu13_driver_if_v13_0_0.h \
+    extra/amdpci/headers/amdgpu_smu.h \
+    -o $BASE/am/smu_v13_0_0.py
+  fixup $BASE/am/smu_v13_0_0.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v14_0_0_pmfw.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v14_0_2_ppsmc.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu14_driver_if_v14_0.h \
+    extra/amdpci/headers/amdgpu_smu.h \
+    --clang-args="-include stdint.h" \
+    -o $BASE/am/smu_v14_0_2.py
+  fixup $BASE/am/smu_v14_0_2.py
+}
+
+generate_sqtt() {
+  clang2py -k cdefstum \
+    extra/sqtt/sqtt.h \
+    -o $BASE/sqtt.py
+  fixup $BASE/sqtt.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/sqtt.py
+  python3 -c "import tinygrad.runtime.autogen.sqtt"
+
+  ROCPROF_COMMIT_HASH=dd0485100971522cc4cd8ae136bdda431061a04d
+  ROCPROF_SRC=/tmp/rocprof-trace-decoder-$ROCPROF_COMMIT_HASH
+  if [ ! -d "$ROCPROF_SRC" ]; then
+    git clone https://github.com/ROCm/rocprof-trace-decoder $ROCPROF_SRC
+    pushd .
+    cd $ROCPROF_SRC
+    git reset --hard $ROCPROF_COMMIT_HASH
+    popd
+  fi
+
+  clang2py -k cdefstum \
+    $ROCPROF_SRC/include/rocprof_trace_decoder.h \
+    $ROCPROF_SRC/include/trace_decoder_instrument.h \
+    $ROCPROF_SRC/include/trace_decoder_types.h \
+    -o extra/sqtt/rocprof/rocprof.py
+  fixup extra/sqtt/rocprof/rocprof.py
+  sed -i '1s/^/# pylint: skip-file\n/' extra/sqtt/rocprof/rocprof.py
+  sed -i "s/import ctypes/import ctypes, ctypes.util/g" extra/sqtt/rocprof/rocprof.py
+  sed -i "s|FunctionFactoryStub()|ctypes.CDLL(ctypes.util.find_library('rocprof-trace-decoder'))|g" extra/sqtt/rocprof/rocprof.py
+}
+
+generate_webgpu() {
+  clang2py extra/webgpu/webgpu.h -o $BASE/webgpu.py
+  fixup $BASE/webgpu.py
+  sed -i "s/FIXME_STUB/webgpu/g" "$BASE/webgpu.py"
+  sed -i "s/FunctionFactoryStub()/ctypes.CDLL(webgpu_support.WEBGPU_PATH)/g" "$BASE/webgpu.py"
+  sed -i "s/import ctypes/import ctypes, tinygrad.runtime.support.webgpu as webgpu_support/g" "$BASE/webgpu.py"
+  python3 -c "import tinygrad.runtime.autogen.webgpu"
+}
+
+generate_libusb() {
+  clang2py -k cdefstum \
+    /usr/include/libusb-1.0/libusb.h \
+    -o $BASE/libusb.py
+
+  fixup $BASE/libusb.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libusb.py
+  sed -i "s/FIXME_STUB/libusb/g" "$BASE/libusb.py"
+  sed -i "s/libusb_le16_to_cpu = libusb_cpu_to_le16//g" "$BASE/libusb.py"
+  sed -i "s/FunctionFactoryStub()/None if (lib_path:=os.getenv('LIBUSB_PATH', ctypes.util.find_library('usb-1.0'))) is None else ctypes.CDLL(lib_path)/g" "$BASE/libusb.py"
+  python3 -c "import tinygrad.runtime.autogen.libusb"
+}
+
+if [ "$1" == "opencl" ]; then generate_opencl
+elif [ "$1" == "hip" ]; then generate_hip
+elif [ "$1" == "comgr" ]; then generate_comgr
+elif [ "$1" == "cuda" ]; then generate_cuda
+elif [ "$1" == "nvrtc" ]; then generate_nvrtc
+elif [ "$1" == "hsa" ]; then generate_hsa
+elif [ "$1" == "kfd" ]; then generate_kfd
+elif [ "$1" == "nv" ]; then generate_nv
+elif [ "$1" == "amd" ]; then generate_amd
+elif [ "$1" == "am" ]; then generate_am
+elif [ "$1" == "nvdrv" ]; then generate_nvdrv
+elif [ "$1" == "sqtt" ]; then generate_sqtt
+elif [ "$1" == "qcom" ]; then generate_qcom
+elif [ "$1" == "io_uring" ]; then generate_io_uring
+elif [ "$1" == "ib" ]; then generate_ib
+elif [ "$1" == "libc" ]; then generate_libc
+elif [ "$1" == "llvm" ]; then generate_llvm
+elif [ "$1" == "kgsl" ]; then generate_kgsl
+elif [ "$1" == "adreno" ]; then generate_adreno
+elif [ "$1" == "pci" ]; then generate_pci
+elif [ "$1" == "vfio" ]; then generate_vfio
+elif [ "$1" == "webgpu" ]; then generate_webgpu
+elif [ "$1" == "libusb" ]; then generate_libusb
+elif [ "$1" == "all" ]; then generate_opencl; generate_hip; generate_comgr; generate_cuda; generate_nvrtc; generate_hsa; generate_kfd; generate_nv; generate_amd; generate_io_uring; generate_libc; generate_am; generate_webgpu
+else echo "usage: $0 <type>"
+fi
@@ -0,0 +1,137 @@
+# tinygrad is a tensor library, and as a tensor library it has multiple parts
+# 1. a "runtime". this allows buffer management, compilation, and running programs
+# 2. a "Device" that uses the runtime but specifies compute in an abstract way for all
+# 3. a "UOp" that fuses the compute into kernels, using memory only when needed
+# 4. a "Tensor" that provides an easy to use frontend with autograd ".backward()"
+
+
+print("******** first, the runtime ***********")
+
+from tinygrad.runtime.ops_cpu import ClangJITCompiler, CPUDevice, CPUProgram
+
+cpu = CPUDevice()
+
+# allocate some buffers
+out = cpu.allocator.alloc(4)
+a = cpu.allocator.alloc(4)
+b = cpu.allocator.alloc(4)
+
+# load in some values (little endian)
+cpu.allocator._copyin(a, memoryview(bytearray([2,0,0,0])))
+cpu.allocator._copyin(b, memoryview(bytearray([3,0,0,0])))
+
+# compile a program to a binary
+lib = ClangJITCompiler().compile("void add(int *out, int *a, int *b) { out[0] = a[0] + b[0]; }")
+
+# create a runtime for the program
+fxn = cpu.runtime("add", lib)
+
+# run the program
+fxn(out, a, b)
+
+# check the data out
+print(val := cpu.allocator._as_buffer(out).cast("I").tolist()[0])
+assert val == 5
+
+
+print("******** second, the Device ***********")
+
+DEVICE = "CPU"   # NOTE: you can change this!
+
+import struct
+from tinygrad.dtype import dtypes
+from tinygrad.device import Buffer, Device
+from tinygrad.uop.ops import UOp, Ops
+
+# allocate some buffers + load in values
+out = Buffer(DEVICE, 1, dtypes.int32).allocate()
+a = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struct.pack("I", 2))))
+b = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struct.pack("I", 3))))
+# NOTE: a._buf is the same as the return from cpu.allocator.alloc
+
+# describe the computation
+idx = UOp.const(dtypes.index, 0)
+buf_1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 1)
+buf_2 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 2)
+ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1.index(idx),))
+ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2.index(idx),))
+alu = ld_1 + ld_2
+output_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
+st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.index(idx), alu))
+s = UOp(Ops.SINK, dtypes.void, (st_0,))
+
+# convert the computation to a "linearized" format (print the format)
+from tinygrad.engine.realize import get_program, CompiledRunner
+program = get_program(s, Device[DEVICE].renderer)
+
+# compile a program (and print the source)
+fxn = CompiledRunner(program)
+print(fxn.p.src)
+# NOTE: fxn.clprg is the CPUProgram
+
+# run the program
+fxn.exec([out, a, b])
+
+# check the data out
+assert out.as_buffer().cast('I')[0] == 5
+
+
+print("******** third, the UOp ***********")
+
+from tinygrad.engine.realize import run_schedule
+from tinygrad.engine.schedule import create_schedule_with_vars
+from tinygrad.schedule.rangeify import get_rangeify_map
+
+# allocate some values + load in values
+a = UOp.new_buffer(DEVICE, 1, dtypes.int32)
+b = UOp.new_buffer(DEVICE, 1, dtypes.int32)
+a.buffer.allocate().copyin(memoryview(bytearray(struct.pack("I", 2))))
+b.buffer.allocate().copyin(memoryview(bytearray(struct.pack("I", 3))))
+
+# describe the computation
+out = a + b
+s = UOp(Ops.SINK, dtypes.void, (out,))
+
+# group the computation into kernels
+becomes_map = get_rangeify_map(s)
+
+# the compute maps to an assign
+assign = becomes_map[a+b].base
+
+# the first source is the output buffer (data)
+assert assign.src[0].op is Ops.BUFFER
+# the second source is the kernel (compute)
+assert assign.src[1].op is Ops.KERNEL
+
+# schedule the kernel graph in a linear list
+s = UOp(Ops.SINK, dtypes.void, (assign,))
+sched, _ = create_schedule_with_vars(s)
+assert len(sched) == 1
+
+# DEBUGGING: print the compute ast
+print(sched[-1].ast)
+# NOTE: sched[-1].ast is the same as st_0 above
+
+# the output will be stored in a new buffer
+out = assign.buf_uop
+assert out.op is Ops.BUFFER and not out.buffer.is_allocated()
+print(out)
+
+# run that schedule
+run_schedule(sched)
+
+# check the data out
+assert out.is_realized and out.buffer.as_buffer().cast('I')[0] == 5
+
+
+print("******** fourth, the Tensor ***********")
+
+from tinygrad import Tensor
+
+a = Tensor([2], dtype=dtypes.int32, device=DEVICE)
+b = Tensor([3], dtype=dtypes.int32, device=DEVICE)
+out = a + b
+
+# check the data out
+print(val:=out.item())
+assert val == 5
@@ -38,19 +38,25 @@ optim.schedule_step()   # this will step the optimizer without running realize
 # The weight Tensors have been assigned to, but not yet realized. Everything is still lazy at this point
 # l1.uop and l2.uop define a computation graph

-from tinygrad.engine.schedule import ExecItem
-schedule: List[ExecItem] = Tensor.schedule(l1, l2)
+from tinygrad.engine.schedule import ScheduleItem
+schedule: List[ScheduleItem] = Tensor.schedule(l1, l2)

 print(f"The schedule contains {len(schedule)} items.")
 for si in schedule: print(str(si)[:80])

 # *****
-# 4. Lower and run the schedule.
+# 4. Lower a schedule.

-for si in tqdm(schedule): si.run()
+from tinygrad.engine.realize import lower_schedule_item, ExecItem
+lowered: List[ExecItem] = [lower_schedule_item(si) for si in tqdm(schedule)]

 # *****
-# 5. Print the weight change
+# 5. Run the schedule
+
+for ei in tqdm(lowered): ei.run()
+
+# *****
+# 6. Print the weight change

 print("first weight change\n", l1.numpy()-l1n)
 print("second weight change\n", l2.numpy()-l2n)
@@ -13,19 +13,19 @@ There's also a [doc describing speed](../developer/speed.md)

 Everything in [Tensor](../tensor/index.md) is syntactic sugar around constructing a graph of [UOps](../developer/uop.md).

-The `UOp` graph specifies the compute in terms of low level tinygrad ops. Not all UOps will actually become realized. There's two types of UOps, base and view. base contains compute into a contiguous buffer, and view is a view. Inputs to a base can be either base or view, inputs to a view can only be a single base.
+The `UOp` graph specifies the compute in terms of low level tinygrad ops. Not all UOps will actually become realized. There's two types of UOps, base and view. base contains compute into a contiguous buffer, and view is a view (specified by a ShapeTracker). Inputs to a base can be either base or view, inputs to a view can only be a single base.

 ## Scheduling

-The [scheduler](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/schedule.py) converts the graph of UOps into a list of `ExecItem`. One `ExecItem` is one kernel on the GPU, and the scheduler is responsible for breaking the large compute graph into subgraphs that can fit in a kernel. `ast` specifies what compute to run, and `bufs` specifies what buffers to run it on.
+The [scheduler](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/schedule.py) converts the graph of UOps into a list of `ScheduleItem`. One `ScheduleItem` is one kernel on the GPU, and the scheduler is responsible for breaking the large compute graph into subgraphs that can fit in a kernel. `ast` specifies what compute to run, and `bufs` specifies what buffers to run it on.

-::: tinygrad.engine.schedule.ExecItem
+::: tinygrad.engine.schedule.ScheduleItem

 ## Lowering

-The code in [realize](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/realize.py) lowers `ExecItem` by populating its `prg` field with
+The code in [realize](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/realize.py) lowers `ScheduleItem` to `ExecItem` with

-::: tinygrad.engine.realize.run_schedule
+::: tinygrad.engine.realize.lower_schedule

 There's a ton of complexity hidden behind this, see the `codegen/` directory.

@@ -0,0 +1,109 @@
+# Kernel Creation
+
+Tinygrad lazily builds up a graph of Tensor operations. The Tensor graph includes a mix of:
+
+- Buffer and Assignment Ops: `BUFFER`, `BUFFER_VIEW`, `COPY`, `ASSIGN`
+- Movement Ops: `RESHAPE`, `EXPAND`, `PERMUTE`, `PAD`, `SHRINK`, `FLIP`
+- Compute Ops: `ADD`, `MUL`, `REDUCE_AXIS`, ...
+
+`Tensor.kernelize` creates the kernels and buffers needed to realize the output Tensor(s).
+
+## Kernelize flow
+
+Let's see how a multiply add Tensor graph becomes a fused elementwise kernel.
+
+```py
+# initialize 3 input buffers on the device
+a = Tensor([1]).realize()
+b = Tensor([2]).realize()
+c = Tensor([3]).realize()
+
+# create the Tensor graph
+mul = a*b
+out = mul+c
+
+print(mul) # <Tensor <UOp METAL (1,) int (<Ops.MUL: 48>, None)> on METAL with grad None>
+print(out) # <Tensor <UOp METAL (1,) int (<Ops.ADD: 52>, None)> on METAL with grad None>
+
+out.kernelize()
+
+print(mul) # <Tensor <UOp METAL (1,) int (<Ops.MUL: 48>, None)> on METAL with grad None>
+print(out) # <Tensor <UOp METAL (1,) int (<Ops.ASSIGN: 66>, None)> on METAL with grad None>
+```
+
+The multiply Tensor stays the same because it is fused. The output Tensor's UOp becomes a new ASSIGN UOp:
+
+```py
+print(out.uop)
+```
+
+The first source is the output BUFFER:
+
+```
+UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),
+  UOp(Ops.UNIQUE, dtypes.void, arg=6, src=()),))
+```
+
+And the second source is the KERNEL and its 4 buffer edges (output_buffer, a, b, c):
+
+```
+UOp(Ops.KERNEL, dtypes.void, arg=<Kernel 12 SINK(<Ops.STORE: 45>,) (__add__, __mul__)>, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+    x1:=UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),
+    UOp(Ops.UNIQUE, dtypes.void, arg=6, src=()),)),
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+     x1,
+    UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),)),
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+     x1,
+    UOp(Ops.UNIQUE, dtypes.void, arg=3, src=()),)),
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+     x1,
+    UOp(Ops.UNIQUE, dtypes.void, arg=5, src=()),)),))
+```
+
+KERNEL describes the compute AST, metadata and memory dependencies.
+
+BUFFER holds a reference to the device memory where the output will be stored.
+
+Once a Tensor is kernelized, all children will LOAD its BUFFER, instead of fusing it:
+
+```py
+child = out+2
+child.kernelize()
+print(child.uop.src[1].arg.ast)
+```
+
+```
+UOp(Ops.SINK, dtypes.void, arg=None, src=(
+  UOp(Ops.STORE, dtypes.void, arg=None, src=(
+    UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=0, src=()),
+    x2:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1,), strides=(0,), offset=0, mask=None, contiguous=True),)), src=()),
+    UOp(Ops.ADD, dtypes.int, arg=None, src=(
+      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
+        UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=1, src=()),
+         x2,)),
+      UOp(Ops.CONST, dtypes.int, arg=2, src=(
+         x2,)),)),)),))
+```
+
+`Tensor.realize` will execute the kernels and write outputs to memory:
+
+```py
+Tensor.realize(out)
+print(out)        # <Tensor <UOp METAL (1,) int (<Ops.BUFFER: 23>, <buf real:True device:METAL size:1 dtype:dtypes.int offset:0>)> on METAL with grad None>
+print(out.item()) # 5
+```
+
+<hr />
+
+**Summary**
+
+- The large Tensor graph is built from a mix of data, compute and movement Ops.
+
+- `Tensor.kernelize` splits the Tensor graph into data (BUFFER), compute (KERNEL) and links dependencies with ASSIGN.
+
+- `Tensor.realize` executes KERNELs on device and replaces the Tensor graph with just a BUFFER.
+
+- Kernelize can be called multiple times on a Tensor. This allows for incrementally building the kernel fusion layout of a large Tensor graph, without having to call `realize` or `schedule`.
@@ -10,7 +10,7 @@ Directories are listed in order of how they are processed.

 Group UOps into kernels.

-::: tinygrad.schedule.rangeify.get_kernel_graph
+::: tinygrad.schedule.rangeify.get_rangeify_map
    options:
        members: false
        show_labels: false
@@ -26,9 +26,9 @@ Transforms the ast into an optimized ast. This is where BEAM search and heuristi

 ## tinygrad/codegen

-Transform the optimized ast into a linearized and rendered program.
+Transform the optimized ast into a linearized list of UOps.

-::: tinygrad.codegen.get_program
+::: tinygrad.codegen.full_rewrite
    options:
        members: false
        show_labels: false
@@ -41,7 +41,7 @@ BEAM                | [#]        | number of beams in kernel beam search
 DEFAULT_FLOAT       | [HALF, ...]| specify the default float dtype (FLOAT32, HALF, BFLOAT16, FLOAT64, ...), default to FLOAT32
 IMAGE               | [1-2]      | enable 2d specific optimizations
 FLOAT16             | [1]        | use float16 for images instead of float32
-HCQ_VISIBLE_DEVICES | [list[int]]| restricts the HCQ devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
+VISIBLE_DEVICES     | [list[int]]| restricts the NV/AMD devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
 JIT                 | [0-2]      | 0=disabled, 1=[jit enabled](quickstart.md#jit) (default), 2=jit enabled, but graphs are disabled
 VIZ                 | [1]        | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
 ALLOW_TF32          | [1]        | enable TensorFloat-32 tensor cores on Ampere or newer GPUs.
@@ -131,7 +131,7 @@ timeit.repeat(jit_step, repeat=5, number=1)

 1.0 ms is 75x faster! Note that we aren't syncing the GPU, so GPU time may be slower.

-The first two runs of the function execute normally, with the JIT capturing the kernels. Starting from the third run, only the tinygrad operations are replayed, removing the overhead by skipping Python code execution. So be aware that any non-tinygrad Python values affecting the kernels will be "frozen" from the second run. Note that `Tensor` randomness functions work as expected.
+The slowness the first two times is the JIT capturing the kernels. And this JIT will not run any Python in the function, it will just replay the tinygrad kernels that were run, so be aware that non tinygrad Python operations won't work. Randomness functions work as expected.

 Unlike other JITs, we JIT everything, including the optimizer. Think of it as a dumb replay on different data.

@@ -0,0 +1,293 @@
+#!/usr/bin/env python3
+
+# this file is a "ramp" for people new to tinygrad to think about how to approach it
+# it is runnable and editable.
+# whenever you see stuff like DEBUG=2 or CPU=1 discussed, these are environment variables
+# in a unix shell like bash `DEBUG=2 CPU=1 python docs/ramp.py`
+
+# this pip installs tinygrad master for the system
+# the -e allows you to edit the tinygrad folder and update system tinygrad
+# tinygrad is pure Python, so you are encouraged to do this
+# git pull in the tinygrad directory will also get you the latest
+"""
+git clone https://github.com/tinygrad/tinygrad.git
+cd tinygrad
+python3 -m pip install -e .
+"""
+
+# %% ********
+print("******* PART 1 *******")
+
+# we start with a Device.
+# a Device is where Tensors are stored and compute is run
+# tinygrad autodetects the best device on your system and makes it the DEFAULT
+from tinygrad import Device
+print(Device.DEFAULT)  # on Mac, you can see this prints METAL
+
+# now, lets create a Tensor
+from tinygrad import Tensor, dtypes
+t = Tensor([1,2,3,4])
+
+# you can see this Tensor is on the DEFAULT device with int dtype and shape (4,)
+assert t.device == Device.DEFAULT
+assert t.dtype == dtypes.int
+assert t.shape == (4,)
+
+# unlike in torch, if we print it, it doesn't print the contents
+# this is because tinygrad is lazy
+# this Tensor has not been computed yet
+print(t)
+# <Tensor <UOp METAL (4,) int (<Ops.COPY: 7>, None)> on METAL with grad None>
+
+# the ".uop" property on Tensor contains the specification of how to compute it
+print(t.uop)
+"""
+UOp(Ops.COPY, dtypes.int, arg=None, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=0, src=()),
+    UOp(Ops.DEVICE, dtypes.void, arg='PYTHON', src=()),)),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# as you can see, it's specifying a copy from PYTHON device
+# which is where the [1,2,3,4] array lives
+
+# UOps are the specification language in tinygrad
+# they are immutable and form a DAG
+# they have a "Ops", a "dtype", a tuple of srcs (parents), and an arg
+
+t.realize()
+# if we want to "realize" a tensor, we can with the "realize" method
+# now when we look at the uop, it's changed
+print(t.uop)
+"""
+UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+  UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# the copy was actually run, and now the "uop" of the Tensor is just a BUFFER
+# if you run this script with DEBUG=2 in the environment, you can see the copy happen
+# *** METAL      1 copy       16,   METAL <- PYTHON ...
+
+# now let's do some compute
+# we look at the uop to see the specification of the compute
+t_times_2 = t * 2
+print(t_times_2.uop)
+"""
+UOp(Ops.MUL, dtypes.int, arg=None, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+    x2:=UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),)),
+  UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+    UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+      UOp(Ops.CONST, dtypes.int, arg=2, src=(
+        UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(), strides=(), offset=0, mask=None, contiguous=True),)), src=(
+           x2,)),)),)),)),))
+"""
+# the BUFFER from above is being multiplied by a CONST 2
+# it's RESHAPEd and EXPANDed to broadcast the CONST to the BUFFER
+
+# we can check the result with
+assert t_times_2.tolist() == [2, 4, 6, 8]
+
+# UOps are both immutable and globally unique
+# if i multiply the Tensor by 4 twice, these result Tensors will have the same uop specification
+t_times_4_try_1 = t * 4
+t_times_4_try_2 = t * 4
+assert t_times_4_try_1.uop is t_times_4_try_2.uop
+# the specification isn't just the same, it's the exact same Python object
+assert t_times_4_try_1 is not t_times_4_try_2
+# the Tensor is a different Python object
+
+# if we realize `t_times_4_try_1` ...
+t_times_4_try_1.realize()
+print(t_times_4_try_2.uop)
+"""
+UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+  UOp(Ops.UNIQUE, dtypes.void, arg=4, src=()),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# ... `t_times_4_try_2` also becomes the same BUFFER
+assert t_times_4_try_1.uop is t_times_4_try_2.uop
+# so this print doesn't require any computation, just a copy back to the CPU so we can print it
+print("** only the copy start")
+print(t_times_4_try_2.tolist())  # [4, 8, 12, 16]
+print("** only the copy end")
+# you can confirm this with DEBUG=2, seeing what's printed in between the "**" prints
+
+# tinygrad has an auto differentiation engine that operates according to these same principles
+# the derivative of "log(x)" is "1/x", and you can see this on line 20 of gradient.py
+t_float = Tensor([3.0])
+t_log = t_float.log()
+t_log_grad, = t_log.sum().gradient(t_float)
+# due to how log is implemented, this gradient contains a lot of UOps
+print(t_log_grad.uop)
+# ...not shown here...
+# but if you run with DEBUG=4 (CPU=1 used here for simpler code), you can see the generated code
+"""
+void E_(float* restrict data0, float* restrict data1) {
+  float val0 = *(data1+0);
+  *(data0+0) = (1/val0);
+}
+"""
+# the derivative is close to 1/3
+assert (t_log_grad.item() - 1/3) < 1e-6
+
+# %% ********
+print("******* PART 2 *******")
+
+# we redefine the same t here so this cell can run on it's own
+from tinygrad import Tensor
+t = Tensor([1,2,3,4])
+
+# what's above gives you enough of an understanding to go use tinygrad as a library
+# however, a lot of the beauty of tinygrad is in how easy it is to interact with the internals
+# NOTE: the APIs here are subject to change
+
+t_plus_3_plus_4 = t + 3 + 4
+print(t_plus_3_plus_4.uop)
+"""
+UOp(Ops.ADD, dtypes.int, arg=None, src=(
+  UOp(Ops.ADD, dtypes.int, arg=None, src=(
+    UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+      UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+      x3:=UOp(Ops.DEVICE, dtypes.void, arg='CPU', src=()),)),
+    UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+      UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+        UOp(Ops.CONST, dtypes.int, arg=3, src=(
+          x7:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(), strides=(), offset=0, mask=None, contiguous=True),)), src=(
+             x3,)),)),)),)),)),
+  UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+    UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+      UOp(Ops.CONST, dtypes.int, arg=4, src=(
+         x7,)),)),)),))
+"""
+# you can see it's adding both 3 and 4
+
+# but by the time we are actually running the code, it's adding 7
+# `kernelize` will simplify and group the operations in the graph into kernels
+t_plus_3_plus_4.kernelize()
+print(t_plus_3_plus_4.uop)
+"""
+UOp(Ops.ASSIGN, dtypes.int, arg=None, src=(
+  x0:=UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=7, src=()),
+    x2:=UOp(Ops.DEVICE, dtypes.void, arg='CPU', src=()),)),
+  UOp(Ops.KERNEL, dtypes.void, arg=<Kernel 12 SINK(<Ops.STORE: 48>,) (__add__,)>, src=(
+     x0,
+    UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+      UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+       x2,)),)),))
+"""
+# ASSIGN has two srcs, src[0] is the BUFFER that's assigned to, and src[1] is the thing to assign
+# src[1] is the GPU Kernel that's going to be run
+# we can get the ast of the Kernel as follows
+kernel_ast = t_plus_3_plus_4.uop.src[1].arg.ast
+
+# almost everything in tinygrad functions as a rewrite of the UOps
+# the codegen rewrites the ast to a simplified form ready for "rendering"
+from tinygrad.codegen import full_rewrite_to_sink
+rewritten_ast = full_rewrite_to_sink(kernel_ast)
+print(rewritten_ast)
+"""
+UOp(Ops.SINK, dtypes.void, arg=None, src=(
+  UOp(Ops.STORE, dtypes.void, arg=None, src=(
+    UOp(Ops.INDEX, dtypes.int.ptr(4), arg=None, src=(
+      UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(4), arg=0, src=()),
+      x3:=UOp(Ops.SPECIAL, dtypes.int, arg=('gidx0', 4), src=()),)),
+    UOp(Ops.ADD, dtypes.int, arg=None, src=(
+      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
+        UOp(Ops.INDEX, dtypes.int.ptr(4), arg=None, src=(
+          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(4), arg=1, src=()),
+           x3,)),)),
+      UOp(Ops.CONST, dtypes.int, arg=7, src=()),)),)),))
+"""
+# you can see at this point we are adding 7, not 3 and 4
+
+# with DEBUG=4, we can see the code.
+# since optimizations are on, it UPCASTed the operation, explicitly writing out all 4 +7s
+t_plus_3_plus_4.realize()
+"""
+void E_4n2(int* restrict data0, int* restrict data1) {
+  int val0 = *(data1+0);
+  int val1 = *(data1+1);
+  int val2 = *(data1+2);
+  int val3 = *(data1+3);
+  *(data0+0) = (val0+7);
+  *(data0+1) = (val1+7);
+  *(data0+2) = (val2+7);
+  *(data0+3) = (val3+7);
+}
+"""
+# the function name E_4n2 is "E" for elementwise op (as opposed to "r" for reduce op)
+# "4" for the size, and "n2" for name deduping (it's the 3rd function with the same E and 4 in this session)
+# when you print the name with DEBUG=2, you'll see the 4 is yellow, meaning that it's upcasted
+# if you run with NOOPT=1 ...
+"""
+void E_4n2(int* restrict data0, int* restrict data1) {
+  for (int ridx0 = 0; ridx0 < 4; ridx0++) {
+    int val0 = *(data1+ridx0);
+    *(data0+ridx0) = (val0+7);
+  }
+}
+"""
+# ... you get this unoptimized code with a loop and the 4 is blue (for global). the color code is in kernel.py
+
+# %% ********
+print("******* PART 3 *******")
+
+# now, we go even lower and understand UOps better and how the graph rewrite engine works.
+# it's much simpler than what's in LLVM or MLIR
+
+from tinygrad import dtypes
+from tinygrad.uop.ops import UOp, Ops
+
+# first, we'll construct some const UOps
+a = UOp(Ops.CONST, dtypes.int, arg=2)
+b = UOp(Ops.CONST, dtypes.int, arg=2)
+
+# if you have been paying attention, you should know these are the same Python object
+assert a is b
+
+# UOps support normal Python math operations, so a_plus_b expresses the spec for 2 + 2
+a_plus_b = a + b
+print(a_plus_b)
+"""
+UOp(Ops.ADD, dtypes.int, arg=None, src=(
+  x0:=UOp(Ops.CONST, dtypes.int, arg=2, src=()),
+   x0,))
+"""
+
+# we could actually render this 2+2 into a language like c and run it
+# or, we can use tinygrad's graph rewrite engine to "constant fold"
+
+from tinygrad.uop.ops import graph_rewrite, UPat, PatternMatcher
+
+# a `PatternMatcher` is a list of tuples. for each element in the list:
+# [0] is the pattern to match, and [1] is the function to run.
+# this function can return either a UOp to replace the pattern with, or None to not replace
+simple_pm = PatternMatcher([
+  (UPat(Ops.ADD, src=(UPat(Ops.CONST, name="c1"), UPat(Ops.CONST, name="c2"))),
+   lambda c1,c2: UOp(Ops.CONST, dtype=c1.dtype, arg=c1.arg+c2.arg)),
+])
+# this pattern matches the addition of two CONST and rewrites it into a single CONST UOp
+
+# to actually apply the pattern to a_plus_b, we use graph_rewrite
+a_plus_b_simplified = graph_rewrite(a_plus_b, simple_pm)
+print(a_plus_b_simplified)
+"""
+UOp(Ops.CONST, dtypes.int, arg=4, src=())
+"""
+# 2+2 is in fact, 4
+
+# we can also use syntactic sugar to write the pattern nicer
+simpler_pm = PatternMatcher([
+  (UPat.cvar("c1")+UPat.cvar("c2"), lambda c1,c2: c1.const_like(c1.arg+c2.arg))
+])
+assert graph_rewrite(a_plus_b, simple_pm) is graph_rewrite(a_plus_b, simpler_pm)
+# note again the use of is, UOps are immutable and globally unique
+
+# %% ********
+
+# that brings you to an understanding of the most core concepts in tinygrad
+# you can run this with VIZ=1 to use the web based graph rewrite explorer
+# hopefully now you understand it. the nodes in the graph are just UOps
@@ -70,7 +70,7 @@ AMD backend supports several interfaces for communicating with devices:

 * `KFD`: uses the amdgpu driver
 * `PCI`: uses the [AM driver](developer/am.md)
-* `USB`: USB3 interface for asm24xx chips.
+* `USB`: USB3 interafce for asm24xx chips.

 You can force an interface by setting `AMD_IFACE` to one of these values. In the case of `AMD_IFACE=PCI`, this may unbind your GPU from the amdgpu driver.

@@ -6,7 +6,6 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.neg
 ::: tinygrad.Tensor.log
 ::: tinygrad.Tensor.log2
-::: tinygrad.Tensor.log10
 ::: tinygrad.Tensor.exp
 ::: tinygrad.Tensor.exp2
 ::: tinygrad.Tensor.sqrt
@@ -88,8 +87,4 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.float
 ::: tinygrad.Tensor.half
 ::: tinygrad.Tensor.int
-::: tinygrad.Tensor.bool
-::: tinygrad.Tensor.bfloat16
-::: tinygrad.Tensor.double
-::: tinygrad.Tensor.long
-::: tinygrad.Tensor.short
+::: tinygrad.Tensor.bool
@@ -27,6 +27,5 @@
 ::: tinygrad.Tensor.flatten
 ::: tinygrad.Tensor.unflatten
 ::: tinygrad.Tensor.diag
-::: tinygrad.Tensor.diagonal
 ::: tinygrad.Tensor.roll
 ::: tinygrad.Tensor.rearrange
@@ -7,7 +7,6 @@
 ::: tinygrad.Tensor.any
 ::: tinygrad.Tensor.all
 ::: tinygrad.Tensor.isclose
-::: tinygrad.Tensor.allclose
 ::: tinygrad.Tensor.mean
 ::: tinygrad.Tensor.var
 ::: tinygrad.Tensor.var_mean
@@ -31,9 +30,7 @@
 ::: tinygrad.Tensor.matmul
 ::: tinygrad.Tensor.einsum
 ::: tinygrad.Tensor.cumsum
-::: tinygrad.Tensor.cumprod
 ::: tinygrad.Tensor.cummax
-::: tinygrad.Tensor.cummin
 ::: tinygrad.Tensor.triu
 ::: tinygrad.Tensor.tril
 ::: tinygrad.Tensor.interpolate
@@ -41,9 +38,7 @@
 ::: tinygrad.Tensor.scatter_reduce
 ::: tinygrad.Tensor.masked_select
 ::: tinygrad.Tensor.masked_fill
-::: tinygrad.Tensor.nonzero
 ::: tinygrad.Tensor.sort
-::: tinygrad.Tensor.argsort
 ::: tinygrad.Tensor.topk
 ::: tinygrad.Tensor.multinomial

@@ -61,8 +56,3 @@
 ::: tinygrad.Tensor.sparse_categorical_crossentropy
 ::: tinygrad.Tensor.cross_entropy
 ::: tinygrad.Tensor.nll_loss
-
-## Linear Algebra
-
-::: tinygrad.Tensor.qr
-::: tinygrad.Tensor.svd
@@ -41,7 +41,7 @@ The BMC also has a web interface you can use if you find that easier.
 It is recommended that you change the BMC password after setting up the box, as the password on the screen is only the initial password.

 If you do decide to change the BMC password and no longer want the initial password to be displayed, remove the `/root/.bmc_password` file.
-Reboot after making these changes or restart the `tinybox-display.service` service.
+Reboot after making these changes or restart the `displayservice.service` service.

 ## What do I use it for?

@@ -0,0 +1,9 @@
+import globals from "globals";
+import pluginJs from "@eslint/js";
+import pluginHtml from "eslint-plugin-html";
+
+export default [
+  {files: ["**/*.html"], plugins: {html: pluginHtml}, rules:{"max-len": ["error", {"code": 150}]}},
+  {languageOptions: {globals: globals.browser}},
+  pluginJs.configs.recommended,
+];
@@ -1,196 +0,0 @@
-from tinygrad import Tensor, dtypes, Context, getenv, UOp, fetch
-from tinygrad.uop.ops import Ops, PatternMatcher, UPat
-from tinygrad.uop.symbolic import symbolic
-from tinygrad.codegen import Renderer
-from tinygrad.codegen.opt import Opt, OptOps
-
-# ************************* implementation of the problem ************************
-
-def myhash(a: Tensor) -> Tensor:
-  a = (a + 0x7ED55D16) + (a << 12)
-  a = (a ^ 0xC761C23C) ^ (a >> 19)
-  a = (a + 0x165667B1) + (a << 5)
-  a = (a + 0xD3A2646C) ^ (a << 9)
-  a = (a + 0xFD7046C5) + (a << 3)
-  a = (a ^ 0xB55A4F09) ^ (a >> 16)
-  return a
-
-def select_with_where_tree(values: Tensor, relative_idx: Tensor) -> Tensor:
-  n = values.shape[0]
-  if n == 1: return values[0].expand(relative_idx.shape)
-
-  mid = n // 2
-  left = select_with_where_tree(values[:mid], relative_idx)
-  right = select_with_where_tree(values[mid:], relative_idx - mid)
-
-  go_left = relative_idx < mid
-  return go_left.where(left, right)
-
-def tree_traversal(forest: Tensor, val: Tensor, height: int, rounds: int, where_tree_threshold=3) -> Tensor:
-  # All walkers start at idx=0
-  idx = Tensor.zeros(val.shape, device=val.device, dtype=dtypes.uint32)
-
-  for r in range(rounds):
-    level = r % (height + 1)
-    level_start = (1 << level) - 1
-    level_size = 1 << level
-
-    if level == 0:
-      # At root (level 0), all walkers are at idx=0
-      # No gather needed, just broadcast the root value
-      node_val = forest[0].expand(val.shape)
-      idx = idx * 0  # Reset to 0
-    elif level <= where_tree_threshold:
-      # Small level: use where-tree
-      level_values = forest[level_start : level_start + level_size]
-      relative_idx = (idx - level_start)
-      node_val = select_with_where_tree(level_values, relative_idx)
-    else:
-      # Large level: use gather
-      node_val = forest.gather(0, idx)
-
-    val = myhash(val ^ node_val)
-    idx = (idx << 1) + (1 + (val & 1))
-
-    # No wrap check needed! At round 10 (level becomes 0), we reset idx above.
-
-  return val.contiguous(arg=(Opt(OptOps.UPCAST, 0, 8),))
-
-# ************************* renderer for VLIW machine *************************
-
-def loop_unrolling(sink:UOp):
-  rng = [x for x in sink.toposort() if x.op is Ops.RANGE]
-  if len(rng) == 0: return None
-  print(f"unrolling loop with size {rng[0].vmax+1}")
-  unrolled_sinks = [sink.substitute({rng[0]:rng[0].const_like(i)}).src[0] for i in range(rng[0].vmax+1)]
-  return UOp.sink(*unrolled_sinks, arg=sink.arg)
-
-global_addrs = []
-vliw_prepare = PatternMatcher([
-  # loop unrolling (should be a part of tinygrad)
-  (UPat(Ops.SINK, name="sink"), loop_unrolling),
-  # cast is fake
-  (UPat(Ops.CAST, name="c"), lambda c: c.src[0]),
-  # rewrites to hardcode the addresses in memory
-  (UPat(Ops.PARAM, name="dg"), lambda dg: UOp.const(dtypes.uint, global_addrs[dg.arg])),
-  # INDEX is just plus
-  (UPat(Ops.INDEX, name="i"), lambda i: i.src[0]+i.src[1]),
-])+symbolic
-
-class VLIWRenderer(Renderer):
-  has_local = False  # TODO: this should be the default / cleaned up
-  # this says this backend supports MULACC + more. decompositions uses this
-  code_for_op: dict = {Ops.MULACC: None, Ops.ADD: "+", Ops.MUL: "*",
-                       Ops.XOR: "^", Ops.AND: "&", Ops.OR: "|",
-                       Ops.SHL: "<<", Ops.SHR: ">>", Ops.CMPLT: "<"}
-  # this matcher runs while still in graph form
-  pre_matcher = vliw_prepare
-
-  def render(self, uops:list[UOp]):
-
-    # TODO: this is a minimal renderer. for low cycle count, make it good
-    # to get speed, you need to add VLIW packing
-    # to get under 1536 regs, you need to add a register allocator
-    # we left the fun parts to you
-
-    print(f"rendering with {len(uops)} uops")
-    reg, inst = 0, []
-    r: dict[UOp, int] = {}
-    for u in uops:
-      assert u.dtype.count in (1,8), "dtype count must be 1 or 8"
-
-      # dumb register allocator
-      if u.op not in {Ops.STORE, Ops.SINK, Ops.GEP}:
-        r[u] = reg
-        reg += u.dtype.count
-
-      # render UOps to instructions
-      match u.op:
-        case Ops.SINK:
-          inst.append({"flow": [("halt",)]})
-        case Ops.CONST:
-          inst.append({"load": [("const", r[u], u.arg)]})
-        case Ops.GEP:
-          # a GEP is just an alias to a special register in the vector
-          r[u] = r[u.src[0]] + u.arg[0]
-        case Ops.VECTORIZE:
-          if all(s == u.src[0] for s in u.src):
-            # if all sources are the same, we can broadcast
-            inst.append({"valu": [("vbroadcast", r[u], r[u.src[0]])]})
-          else:
-            # this is a copy into a contiguous chunk of registers
-            inst.extend({"flow": [("add_imm", r[u]+i, r[s], 0)]} for i,s in enumerate(u.src) if r[s] != r[u]+i)
-        case Ops.LOAD:
-          op = "vload" if u.dtype.count > 1 else "load"
-          inst.append({"load": [(op, r[u], r[u.src[0]])]})
-        case Ops.STORE:
-          op = "vstore" if u.src[1].dtype.count > 1 else "store"
-          inst.append({"store": [(op, r[u.src[0]], r[u.src[1]])]})
-        case Ops.MULACC:
-          assert u.dtype.count == 8
-          inst.append({"valu": [("multiply_add", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
-        case Ops.WHERE:
-          assert u.dtype.count == 8
-          inst.append({"flow": [("vselect", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
-        case _ if u.op in self.code_for_op:
-          cat = "valu" if u.dtype.count > 1 else "alu"
-          inst.append({cat: [(self.code_for_op[u.op], r[u], r[u.src[0]], r[u.src[1]])]})
-        case _:
-          raise NotImplementedError(f"unhandled op {u.op}")
-    return repr(inst)
-
-# ************************* test and render *************************
-
-import sys, types
-PROBLEM_URL = "https://raw.githubusercontent.com/anthropics/original_performance_takehome/refs/heads/main/tests/frozen_problem.py"
-sys.modules["problem"] = problem = types.ModuleType("problem")
-exec(fetch(PROBLEM_URL).read_text(), problem.__dict__)
-
-if __name__ == "__main__":
-  batch_size = getenv("BS", 256)
-  height = 10
-  rounds = getenv("ROUNDS", 16)
-
-  # build problem
-  tree = problem.Tree.generate(height)
-  inp = problem.Input.generate(tree, batch_size, rounds)
-  mem = problem.build_mem_image(tree, inp)
-  global_addrs.extend([mem[6], mem[6], mem[4]])  # output, input, forest
-
-  # *** verify the kernel in tinygrad compared to reference ***
-
-  forest_t = Tensor(tree.values, dtype=dtypes.uint32)
-  val_t = Tensor(inp.values, dtype=dtypes.uint32)
-
-  if getenv("VERIFY", 1):
-    # verify on normal tinygrad device
-    with Context(PCONTIG=2):
-      out = tree_traversal(forest_t, val_t, height, rounds)
-      val_out = out.tolist()
-    problem.reference_kernel(tree, inp)
-    assert val_out == inp.values
-    print("verification passed")
-
-  # *** render to device ***
-
-  from tinygrad.codegen import get_program
-  with Context(PCONTIG=2, DEVECTORIZE=2, SPEC=0):
-    out = tree_traversal(forest_t, val_t, height, rounds)
-    sink = out.schedule()[-1].ast
-    prg = get_program(sink, VLIWRenderer())
-
-  # *** run on Machine and compare ***
-
-  # NOTE: the scratch size needs to be reduced to 1536 when you have a register allocator
-  src = eval(prg.src)
-  max_regs = max(t[1] for instr in src for v in instr.values() for t in v if len(t) > 1) + 8
-  print(f"{max_regs:5d} regs used" + ("" if max_regs <= 1536 else "       <-- WARNING: TOO MANY REGISTERS, MUST BE <= 1536"))
-  machine = problem.Machine(mem, src, problem.DebugInfo(scratch_map={}), n_cores=1, trace=False, scratch_size=max_regs)
-  machine.run()
-  print(f"ran for {machine.cycle:5d} cycles" + ("" if machine.cycle <= 1363 else "  <-- EVEN CLAUDE GOT 1363"))
-
-  # compare to reference
-  ref_mem = mem.copy()
-  for _ in problem.reference_kernel2(ref_mem, {}): pass
-  assert machine.mem[mem[6]:mem[6]+mem[2]] == ref_mem[mem[6]:mem[6]+mem[2]]
-  print("compare passed!")
@@ -1,79 +0,0 @@
-from typing import Optional
-from tinygrad import Tensor
-from tinygrad.dtype import DTypeLike, dtypes
-import math
-
-# rewritten from numpy
-def rfftfreq(n: int, d: float = 1.0, device=None) -> Tensor:
-  val = 1.0 / (n * d)
-  N = n // 2 + 1
-  results = Tensor.arange(N, device=device)
-  return results * val
-
-# just like in librosa
-def fft_frequencies(sr: float, n_fft: int) -> Tensor:
-  return rfftfreq(n=n_fft, d=1.0 / sr)
-
-def hz_to_mel(freq: Tensor) -> Tensor:
-  # linear part
-  f_min = 0.0
-  f_sp = 200.0 / 3
-  mels = (freq - f_min) / f_sp
-
-  # log-scale part
-  min_log_hz = 1000.0  # beginning of log region (Hz)
-  mask = freq >= min_log_hz
-  return mask.where(((min_log_hz - f_min) / f_sp) + (freq / min_log_hz).log() / (math.log(6.4) / 27.0), mels)
-
-def mel_to_hz(mels: Tensor) -> Tensor:
-  # linear scale
-  f_min = 0.0
-  f_sp = 200.0 / 3
-  freqs = f_min + f_sp * mels
-
-  # nonlinear scale
-  min_log_hz = 1000.0  # beginning of log region (Hz)
-  min_log_mel = (min_log_hz - f_min) / f_sp  # same (Mels)
-  logstep = math.log(6.4) / 27.0  # step size for log region
-
-  log_t = mels >= min_log_mel
-  freqs = log_t.where(min_log_hz * ((logstep * (mels - min_log_mel)).exp()), freqs)
-  return freqs
-
-def mel_frequencies(n_mels: int = 128, *, fmin: float = 0.0, fmax: float = 11025.0) -> Tensor:
-  # center freqs of mel bands - uniformly spaced between limits
-  min_max_mel = hz_to_mel(Tensor([fmin, fmax]))
-
-  mels = Tensor.linspace(min_max_mel[0], min_max_mel[1], n_mels)
-  hz = mel_to_hz(mels)
-  return hz
-
-def mel(
-  *,
-  sr: float,
-  n_fft: int,
-  n_mels: int = 128,
-  fmin: float = 0.0,
-  fmax: Optional[float] = None,
-  dtype: DTypeLike = dtypes.default_float,
-) -> Tensor:
-  if fmax is None:
-    fmax = float(sr) / 2
-
-  n_mels = int(n_mels)
-
-  fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft)  # center freqs of each FFT bin
-  mel_f = mel_frequencies(n_mels + 2, fmin=fmin, fmax=fmax)  # center freqs of mel bands
-
-  fdiff = mel_f[1:] - mel_f[:-1]
-  ramps = mel_f[None].T.expand(-1, fftfreqs.shape[-1]) - fftfreqs
-
-  lower = -ramps[:n_mels] / fdiff[:n_mels][None].T
-  upper = ramps[2 : n_mels + 2] / fdiff[1 : n_mels + 1][None].T
-  weights = lower.minimum(upper).maximum(0)
-
-  # Slaney-style mel is scaled to be approx constant energy per channel
-  enorm = 2.0 / (mel_f[2 : n_mels + 2] - mel_f[:n_mels])
-  weights *= enorm[:, None]
-
-  return weights
@@ -1,6 +1,6 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

@@ -15,31 +15,30 @@ class Model:
      nn.BatchNorm(64), Tensor.max_pool2d,
      lambda x: x.flatten(1), nn.Linear(576, 10)]

-  @function
  def __call__(self, x:Tensor) -> Tensor: return x.sequential(self.layers)

-  @TinyJit
-  @Tensor.train()
-  def train_step(self, X_train:Tensor, Y_train:Tensor) -> Tensor:
-    opt.zero_grad()
-    samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
-    loss = self(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
-    return loss.realize(*opt.schedule_step())
-
-  @TinyJit
-  def get_test_acc(self, X_test:Tensor, Y_test:Tensor) -> Tensor: return (self(X_test).argmax(axis=1) == Y_test).mean()*100
-
 if __name__ == "__main__":
  X_train, Y_train, X_test, Y_test = mnist(fashion=getenv("FASHION"))

  model = Model()
-  opt = (nn.optim.Muon if getenv("MUON") else nn.optim.SGD if getenv("SGD") else nn.optim.Adam)(nn.state.get_parameters(model))
+  opt = (nn.optim.Adam if not getenv("MUON") else nn.optim.Muon)(nn.state.get_parameters(model))
+
+  @TinyJit
+  @Tensor.train()
+  def train_step() -> Tensor:
+    opt.zero_grad()
+    samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
+    loss = model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
+    return loss.realize(*opt.schedule_step())
+
+  @TinyJit
+  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100

  test_acc = float('nan')
  for i in (t:=trange(getenv("STEPS", 70))):
    GlobalCounters.reset()   # NOTE: this makes it nice for DEBUG=2 timing
-    loss = model.train_step(X_train, Y_train)
-    if i%10 == 9: test_acc = model.get_test_acc(X_test, Y_test).item()
+    loss = train_step()
+    if i%10 == 9: test_acc = get_test_acc().item()
    t.set_description(f"loss: {loss.item():6.2f} test_accuracy: {test_acc:5.2f}%")

  # verify eval acc
@@ -5,7 +5,7 @@ from extra.onnx_helpers import get_example_inputs, validate

 def load_onnx_model(onnx_file):
  run_onnx = OnnxRunner(onnx_file)
-  run_onnx_jit = TinyJit(lambda **kwargs: next(iter(run_onnx({k:v.to(None) for k,v in kwargs.items()}).values())), prune=True)
+  run_onnx_jit = TinyJit(lambda **kwargs: next(iter(run_onnx({k:v.to(None) for k,v in kwargs.items()}).values())), prune=True, optimize=True)
  return run_onnx_jit, run_onnx.graph_inputs

 if __name__ == "__main__":
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+import os, sys, traceback
+sys.path.append(os.getcwd())
+
+from io import StringIO
+from contextlib import redirect_stdout
+from tinygrad import Tensor, nn
+from tinygrad.helpers import Timing, colored, getenv, fetch
+from extra.models.llama import Transformer, convert_from_huggingface, fix_bf16
+from sentencepiece import SentencePieceProcessor
+
+def create_fixed_tokenizer(output_file):
+  print("creating fixed tokenizer")
+  import extra.junk.sentencepiece_model_pb2 as spb2
+  mp = spb2.ModelProto()
+  mp.ParseFromString(fetch("https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/resolve/main/tokenizer.model?download=true").read_bytes())
+  mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_end|>", score=0))
+  mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_start|>", score=0))
+  with open(output_file, "wb") as f:
+    f.write(mp.SerializeToString())
+
+# example:
+# echo -en "write 2+2\nwrite hello world\ny\n" | TEMP=0 python3 examples/coder.py
+
+if __name__ == "__main__":
+  # https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/config.json
+  with Timing("create model: "):
+    model = Transformer(4096, 14336, n_heads=32, n_layers=32, norm_eps=1e-5, vocab_size=32002, n_kv_heads=8, max_context=4096, jit=getenv("JIT", 1))
+
+  with Timing("download weights: "):
+    part1 = nn.state.torch_load(fetch("https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/resolve/main/pytorch_model-00001-of-00002.bin?download=true"))
+    part2 = nn.state.torch_load(fetch("https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/resolve/main/pytorch_model-00002-of-00002.bin?download=true"))
+
+  with Timing("weights -> model: "):
+    nn.state.load_state_dict(model, fix_bf16(convert_from_huggingface(part1, 32, 32, 8)), strict=False)
+    nn.state.load_state_dict(model, fix_bf16(convert_from_huggingface(part2, 32, 32, 8)), strict=False)
+
+  if not os.path.isfile("/tmp/tokenizer.model"): create_fixed_tokenizer("/tmp/tokenizer.model")
+  spp = SentencePieceProcessor(model_file="/tmp/tokenizer.model")
+
+  # https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/tokenizer_config.json
+  #   "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+  IM_END = 32000
+  IM_START = 32001
+  def encode_prompt(k, v): return [IM_START]+spp.encode(f"{k}\n{v}")+[IM_END]+spp.encode("\n")
+  def start_prompt(k): return [IM_START]+spp.encode(f"{k}\n")
+  def output(outputted, toks, color):
+    cur = spp.decode(toks)[len(outputted):]
+    sys.stdout.write(colored(cur, color))
+    sys.stdout.flush()
+    outputted += cur
+    return outputted
+
+  # *** app below this line ***
+
+  toks = [spp.bos_id()] + encode_prompt("system", "You are Quentin. Quentin is a useful assistant who writes Python code to answer questions. He keeps the code as short as possible and doesn't read from user input")
+
+  PROMPT = getenv("PROMPT", 1)
+  temperature = getenv("TEMP", 0.7)
+
+  start_pos = 0
+  outputted = output("", toks, "green")
+  turn = True
+  while 1:
+    if PROMPT:
+      toks += encode_prompt("user", input("Q: ")) + start_prompt("assistant")
+    else:
+      toks += start_prompt("user" if turn else "assistant")
+      turn = not turn
+    old_output_len = len(outputted)
+    while 1:
+      tok = model(Tensor([toks[start_pos:]]), start_pos, temperature).item()
+      start_pos = len(toks)
+      toks.append(tok)
+      outputted = output(outputted, toks, "blue" if not turn else "cyan")
+      if tok == IM_END: break
+      if tok == spp.eos_id(): break
+      new_output = outputted[old_output_len:]
+
+      if new_output.endswith("```") and '```python\n' in new_output:
+        python_code = new_output.split('```python\n')[1].split("```")[0]
+        # AI safety. Warning to user. Do not press y if the AI is trying to do unsafe things.
+        if input(colored(f" <-- PYTHON DETECTED, RUN IT? ", "red")).lower() == 'y':
+          my_stdout = StringIO()
+          try:
+            with redirect_stdout(my_stdout): exec(python_code)
+            result = my_stdout.getvalue()
+          except Exception as e:
+            result = ''.join(traceback.format_exception_only(e))
+          toks += spp.encode(f"\nOutput:\n```\n{result}```")
+          outputted = output(outputted, toks, "yellow")
+          old_output_len = len(outputted)
+    print("")
@@ -0,0 +1,341 @@
+import argparse
+import multiprocessing as mp
+import os
+import re
+import sys
+import time
+from contextlib import contextmanager
+from pathlib import Path
+
+import numpy as np
+import pyaudio
+import yaml
+from llama import LLaMa
+from vits import MODELS as VITS_MODELS
+from vits import Y_LENGTH_ESTIMATE_SCALARS, HParams, Synthesizer, TextMapper, get_hparams_from_file, load_model
+from whisper import init_whisper, transcribe_waveform
+from sentencepiece import SentencePieceProcessor
+
+from tinygrad.helpers import Timing, fetch
+from tinygrad import Tensor, dtypes
+
+# Whisper constants
+RATE = 16000
+CHUNK = 1600
+
+# LLaMa constants
+IM_START = 32001
+IM_END = 32002
+
+
+# Functions for encoding prompts to chatml md
+def encode_prompt(spp, k, v): return [IM_START]+spp.encode(f"{k}\n{v}")+[IM_END]+spp.encode("\n")
+def start_prompt(spp, k): return [IM_START]+spp.encode(f"{k}\n")
+
+def chunks(lst, n):
+  for i in range(0, len(lst), n): yield lst[i:i + n]
+
+def create_fixed_tokenizer():
+  """Function needed for extending tokenizer with additional chat tokens"""
+  import extra.junk.sentencepiece_model_pb2 as spb2
+  tokenizer_path = fetch("https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.4/resolve/main/tokenizer.model")
+  if SentencePieceProcessor(model_file=str(tokenizer_path)).vocab_size() != 32003:
+    print("creating fixed tokenizer")
+    mp = spb2.ModelProto()
+    mp.ParseFromString(tokenizer_path.read_bytes())
+    # https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.4/blob/main/added_tokens.json
+    mp.pieces.append(spb2.ModelProto.SentencePiece(piece="[PAD]", score=0))
+    mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_start|>", score=0))
+    mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_end|>", score=0))
+    tokenizer_path.write_bytes(mp.SerializeToString())
+  return tokenizer_path
+
+def llama_prepare(llama: LLaMa, temperature: float, pre_prompt_path: Path) -> tuple[list[int], str, str, str]:
+  """Prepares a llama model from a specified pre-prompt file"""
+  with open(str(pre_prompt_path)) as f:
+    config = yaml.safe_load(f.read())
+  toks = [llama.tokenizer.bos_id()] + encode_prompt(llama.tokenizer, "system", config["pre_prompt"].replace("\n", " "))
+  for i in config["examples"]:
+    toks += encode_prompt(llama.tokenizer, config["user_delim"], i["user_prompt"])
+    toks += encode_prompt(llama.tokenizer, config["resp_delim"], i["resp_prompt"])
+  llama.model(Tensor([toks]), 0, temperature).realize()  # NOTE: outputs are not used
+  return toks, config["user_delim"], config["resp_delim"], len(toks), llama.tokenizer.decode(toks)
+
+def llama_generate(
+  llama: LLaMa,
+  toks: list[int],
+  outputted: str,
+  prompt: str,
+  start_pos: int,
+  user_delim: str,
+  resp_delim: str,
+  temperature=0.7,
+  max_tokens=1000
+):
+  """Generates an output for the specified prompt"""
+  toks += encode_prompt(llama.tokenizer, user_delim, prompt)
+  toks += start_prompt(llama.tokenizer, resp_delim)
+
+  outputted = llama.tokenizer.decode(toks)
+  init_length = len(outputted)
+  for _ in range(max_tokens):
+    token = llama.model(Tensor([toks[start_pos:]]), start_pos, temperature).item()
+    start_pos = len(toks)
+    toks.append(token)
+
+    cur = llama.tokenizer.decode(toks)
+
+    # Print is just for debugging
+    sys.stdout.write(cur[len(outputted):])
+    sys.stdout.flush()
+    outputted = cur
+    if toks[-1] == IM_END: break
+  else:
+    toks.append(IM_END)
+  print() # because the output is flushed
+  return outputted, start_pos, outputted[init_length:].replace("<|im_end|>", "")
+
+def tts(
+  text_to_synthesize: str,
+  synth: Synthesizer,
+  hps: HParams,
+  emotion_embedding: Path,
+  speaker_id: int,
+  model_to_use: str,
+  noise_scale: float,
+  noise_scale_w: float,
+  length_scale: float,
+  estimate_max_y_length: bool,
+  text_mapper: TextMapper,
+  model_has_multiple_speakers: bool,
+  pad_length=600,
+  vits_pad_length=1000
+):
+  if model_to_use == "mmts-tts": text_to_synthesize = text_mapper.filter_oov(text_to_synthesize.lower())
+
+  # Convert the input text to a tensor.
+  stn_tst = text_mapper.get_text(text_to_synthesize, hps.data.add_blank, hps.data.text_cleaners)
+  init_shape = stn_tst.shape
+  assert init_shape[0] < pad_length, "text is too long"
+  x_tst, x_tst_lengths = stn_tst.pad(((0, pad_length - init_shape[0]),), value=1).unsqueeze(0), Tensor([init_shape[0]], dtype=dtypes.int64)
+  sid = Tensor([speaker_id], dtype=dtypes.int64) if model_has_multiple_speakers else None
+
+  # Perform inference.
+  audio_tensor = synth.infer(x_tst, x_tst_lengths, sid, noise_scale, length_scale, noise_scale_w, emotion_embedding=emotion_embedding,
+                             max_y_length_estimate_scale=Y_LENGTH_ESTIMATE_SCALARS[model_to_use] if estimate_max_y_length else None, pad_length=vits_pad_length)[0, 0]
+  # Save the audio output.
+  audio_data = (np.clip(audio_tensor.numpy(), -1.0, 1.0) * 32767).astype(np.int16)
+  return audio_data
+
+def init_vits(
+  model_to_use: str,
+  emotion_path: Path,
+  speaker_id: int,
+  seed: int,
+):
+  model_config = VITS_MODELS[model_to_use]
+
+  # Load the hyperparameters from the config file.
+  hps = get_hparams_from_file(fetch(model_config[0]))
+
+  # If model has multiple speakers, validate speaker id and retrieve name if available.
+  model_has_multiple_speakers = hps.data.n_speakers > 0
+  if model_has_multiple_speakers:
+    if speaker_id >= hps.data.n_speakers: raise ValueError(f"Speaker ID {speaker_id} is invalid for this model.")
+    if hps.__contains__("speakers"): # maps speaker ids to names
+      speakers = hps.speakers
+      if isinstance(speakers, list): speakers = {speaker: i for i, speaker in enumerate(speakers)}
+
+  # Load emotions if any. TODO: find an english model with emotions, this is untested atm.
+  emotion_embedding = None
+  if emotion_path is not None:
+    if emotion_path.endswith(".npy"): emotion_embedding = Tensor(np.load(emotion_path), dtype=dtypes.int64).unsqueeze(0)
+    else: raise ValueError("Emotion path must be a .npy file.")
+
+  # Load symbols, instantiate TextMapper and clean the text.
+  if hps.__contains__("symbols"): symbols = hps.symbols
+  elif model_to_use == "mmts-tts": symbols = [x.replace("\n", "") for x in fetch("https://huggingface.co/facebook/mms-tts/raw/main/full_models/eng/vocab.txt").open(encoding="utf-8").readlines()]
+  else: symbols = ['_'] + list(';:,.!?¡¿—…"«»“” ') + list('ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz') + list("ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ")
+  text_mapper = TextMapper(apply_cleaners=True, symbols=symbols)
+
+  # Load the model.
+  if seed is not None:
+    Tensor.manual_seed(seed)
+    np.random.seed(seed)
+  net_g = load_model(text_mapper.symbols, hps, model_config)
+
+  return net_g, emotion_embedding, text_mapper, hps, model_has_multiple_speakers
+
+@contextmanager
+def output_stream(num_channels: int, sample_rate: int):
+  try:
+    p = pyaudio.PyAudio()
+    stream = p.open(format=pyaudio.paInt16, channels=num_channels, rate=sample_rate, output=True)
+    yield stream
+  except KeyboardInterrupt: pass
+  finally:
+    stream.stop_stream()
+    stream.close()
+    p.terminate()
+
+@contextmanager
+def log_writer():
+  try:
+    logs = []
+    yield logs
+  finally:
+    sep = "="*os.get_terminal_size()[1]
+    print(f"{sep[:-1]}\nCHAT LOG")
+    print(*logs, sep="\n")
+    print(sep)
+
+def listener(q: mp.Queue, event: mp.Event):
+  try:
+    p = pyaudio.PyAudio()
+    stream = p.open(format=pyaudio.paInt16, channels=1, rate=RATE, input=True, frames_per_buffer=CHUNK)
+    did_print = False
+    while True:
+      data = stream.read(CHUNK) # read data to avoid overflow
+      if event.is_set():
+        if not did_print:
+          print("listening")
+          did_print = True
+        q.put(((np.frombuffer(data, np.int16)/32768).astype(np.float32)*3))
+      else:
+        did_print = False
+  finally:
+    stream.stop_stream()
+    stream.close()
+    p.terminate()
+
+def mp_output_stream(q: mp.Queue, counter: mp.Value, num_channels: int, sample_rate: int):
+  with output_stream(num_channels, sample_rate) as stream:
+    while True:
+      try:
+        stream.write(q.get())
+        counter.value += 1
+      except KeyboardInterrupt:
+        break
+
+if __name__ == "__main__":
+  import nltk
+  nltk.download("punkt")
+  # Parse CLI arguments
+  parser = argparse.ArgumentParser("Have a tiny conversation with tinygrad")
+
+  # Whisper args
+  parser.add_argument("--whisper_model_name", type=str, default="tiny.en")
+
+  # LLAMA args
+  parser.add_argument("--llama_pre_prompt_path", type=Path, default=Path(__file__).parent / "conversation_data" / "pre_prompt_stacy.yaml", help="Path to yaml file which contains all pre-prompt data needed. ")
+  parser.add_argument("--llama_count", type=int, default=1000, help="Max number of tokens to generate")
+  parser.add_argument("--llama_temperature", type=float, default=0.7, help="Temperature in the softmax")
+  parser.add_argument("--llama_quantize", type=str, default=None, help="Quantize the weights to int8 or nf4 in memory")
+  parser.add_argument("--llama_model", type=Path, default=None, help="Folder with the original weights to load, or single .index.json, .safetensors or .bin file")
+  parser.add_argument("--llama_gen", type=str, default="tiny", required=False, help="Generation of the model to use")
+  parser.add_argument("--llama_size", type=str, default="1B-Chat", required=False, help="Size of model to use")
+  parser.add_argument("--llama_tokenizer", type=Path, default=None, required=False, help="Path to llama tokenizer.model")
+
+  # vits args
+  parser.add_argument("--vits_model_to_use", default="vctk", help="Specify the model to use. Default is 'vctk'.")
+  parser.add_argument("--vits_speaker_id", type=int, default=12, help="Specify the speaker ID. Default is 6.")
+  parser.add_argument("--vits_noise_scale", type=float, default=0.667, help="Specify the noise scale. Default is 0.667.")
+  parser.add_argument("--vits_noise_scale_w", type=float, default=0.8, help="Specify the noise scale w. Default is 0.8.")
+  parser.add_argument("--vits_length_scale", type=float, default=1, help="Specify the length scale. Default is 1.")
+  parser.add_argument("--vits_seed", type=int, default=None, help="Specify the seed (set to None if no seed). Default is 1337.")
+  parser.add_argument("--vits_num_channels", type=int, default=1, help="Specify the number of audio output channels. Default is 1.")
+  parser.add_argument("--vits_sample_width", type=int, default=2, help="Specify the number of bytes per sample, adjust if necessary. Default is 2.")
+  parser.add_argument("--vits_emotion_path", type=Path, default=None, help="Specify the path to emotion reference.")
+  parser.add_argument("--vits_estimate_max_y_length", type=str, default=False, help="If true, overestimate the output length and then trim it to the correct length, to prevent premature realization, much more performant for larger inputs, for smaller inputs not so much. Default is False.")
+  parser.add_argument("--vits_vocab_path", type=Path, default=None, help="Path to the TTS vocabulary.")
+
+  # conversation args
+  parser.add_argument("--max_sentence_length", type=int, default=20, help="Max words in one sentence to pass to vits")
+
+  args = parser.parse_args()
+
+  # Init models
+  model, enc = init_whisper(args.whisper_model_name)
+  synth, emotion_embedding, text_mapper, hps, model_has_multiple_speakers = init_vits(args.vits_model_to_use, args.vits_emotion_path, args.vits_speaker_id, args.vits_seed)
+
+  # Download tinyllama chat as a default model
+  if args.llama_model is None:
+    args.llama_model = fetch("https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.4/resolve/main/model.safetensors", "tinyllamachat.safetensors")
+    args.llama_gen = "tiny"
+    args.llama_size = "1B-Chat"
+  # Add 3 more tokens to the tokenizer
+  if args.llama_gen == "tiny" and args.llama_size.endswith("Chat"): args.llama_tokenizer = create_fixed_tokenizer()
+  tokenizer_path = args.llama_tokenizer or args.llama_model.parent / "tokenizer.model"
+  llama = LLaMa.build(args.llama_model, tokenizer_path, args.llama_gen, args.llama_size, args.llama_quantize)
+  toks, user_delim, resp_delim, start_pos, outputted = llama_prepare(llama, args.llama_temperature, args.llama_pre_prompt_path)
+
+  # Start child process for mic input
+  q = mp.Queue()
+  is_listening_event = mp.Event()
+  p = mp.Process(target=listener, args=(q, is_listening_event,))
+  p.daemon = True
+  p.start()
+
+  # Start child process for speaker output
+  out_q = mp.Queue()
+  out_counter = mp.Value("i", 0)
+  out_p = mp.Process(target=mp_output_stream, args=(out_q, out_counter, args.vits_num_channels, hps.data.sampling_rate,))
+  out_p.daemon = True
+  out_p.start()
+
+  # JIT tts
+  for i in ["Hello, I'm a chat bot", "I am capable of doing a lot of things"]:
+    tts(
+      i, synth, hps, emotion_embedding,
+      args.vits_speaker_id, args.vits_model_to_use, args.vits_noise_scale,
+      args.vits_noise_scale_w, args.vits_length_scale,
+      args.vits_estimate_max_y_length, text_mapper, model_has_multiple_speakers
+    )
+
+  # Start the pipeline
+  with log_writer() as log:
+    while True:
+      tokens = [enc._special_tokens["<|startoftranscript|>"], enc._special_tokens["<|notimestamps|>"]]
+      total = np.array([])
+      out_counter.value = 0
+
+      s = time.perf_counter()
+      is_listening_event.set()
+      prev_text = None
+      while True:
+        for _ in range(RATE // CHUNK): total = np.concatenate([total, q.get()])
+        txt = transcribe_waveform(model, enc, [total], truncate=True)
+        print(txt, end="\r")
+        if txt == "[BLANK_AUDIO]" or re.match(r"^\([\w+ ]+\)$", txt.strip()): continue
+        if prev_text is not None and prev_text == txt:
+          is_listening_event.clear()
+          break
+        prev_text = txt
+      print() # to avoid llama printing on the same line
+      log.append(f"{user_delim.capitalize()}: {txt}")
+
+      # Generate with llama
+      with Timing("llama generation: "):
+        outputted, start_pos, response = llama_generate(
+          llama, toks, outputted, txt, start_pos,
+          user_delim=user_delim, resp_delim=resp_delim, temperature=args.llama_temperature,
+          max_tokens=args.llama_count
+        )
+        log.append(f"{resp_delim.capitalize()}: {response}")
+
+      # Convert to voice
+      with Timing("tts: "):
+        sentences = nltk.sent_tokenize(response.replace('"', ""))
+        for i in sentences:
+          total = np.array([], dtype=np.int16)
+          for j in chunks(i.split(), args.max_sentence_length):
+            audio_data = tts(
+              " ".join(j), synth, hps, emotion_embedding,
+              args.vits_speaker_id, args.vits_model_to_use, args.vits_noise_scale,
+              args.vits_noise_scale_w, args.vits_length_scale,
+              args.vits_estimate_max_y_length, text_mapper, model_has_multiple_speakers
+            )
+            total = np.concatenate([total, audio_data])
+          out_q.put(total.tobytes())
+      while out_counter.value < len(sentences): continue
+      log.append(f"Total: {time.perf_counter() - s}")
@@ -0,0 +1,89 @@
+# load weights from
+# https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b0-355c32eb.pth
+# a rough copy of
+# https://github.com/lukemelas/EfficientNet-PyTorch/blob/master/efficientnet_pytorch/model.py
+import sys
+import ast
+import time
+import numpy as np
+from PIL import Image
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import getenv, fetch, Timing
+from tinygrad.engine.jit import TinyJit
+from extra.models.efficientnet import EfficientNet
+np.set_printoptions(suppress=True)
+
+# TODO: you should be able to put these in the jitted function
+bias = Tensor([0.485, 0.456, 0.406])
+scale = Tensor([0.229, 0.224, 0.225])
+
+@TinyJit
+def _infer(model, img):
+  img = img.permute((2,0,1))
+  img = img / 255.0
+  img = img - bias.reshape((1,-1,1,1))
+  img = img / scale.reshape((1,-1,1,1))
+  return model.forward(img).realize()
+
+def infer(model, img):
+  # preprocess image
+  aspect_ratio = img.size[0] / img.size[1]
+  img = img.resize((int(224*max(aspect_ratio,1.0)), int(224*max(1.0/aspect_ratio,1.0))))
+
+  img = np.array(img)
+  y0,x0=(np.asarray(img.shape)[:2]-224)//2
+  retimg = img = img[y0:y0+224, x0:x0+224]
+
+  # if you want to look at the image
+  """
+  import matplotlib.pyplot as plt
+  plt.imshow(img)
+  plt.show()
+  """
+
+  # run the net
+  out = _infer(model, Tensor(img.astype("float32"))).numpy()
+
+  # if you want to look at the outputs
+  """
+  import matplotlib.pyplot as plt
+  plt.plot(out[0])
+  plt.show()
+  """
+  return out, retimg
+
+if __name__ == "__main__":
+  # instantiate my net
+  model = EfficientNet(getenv("NUM", 0))
+  model.load_from_pretrained()
+
+  # category labels
+  lbls = ast.literal_eval(fetch("https://gist.githubusercontent.com/yrevar/942d3a0ac09ec9e5eb3a/raw/238f720ff059c1f82f368259d1ca4ffa5dd8f9f5/imagenet1000_clsidx_to_labels.txt").read_text())
+
+  # load image and preprocess
+  url = sys.argv[1] if len(sys.argv) >= 2 else "https://raw.githubusercontent.com/tinygrad/tinygrad/master/docs/showcase/stable_diffusion_by_tinygrad.jpg"
+  if url == 'webcam':
+    import cv2
+    cap = cv2.VideoCapture(0)
+    cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
+    while 1:
+      _ = cap.grab() # discard one frame to circumvent capture buffering
+      ret, frame = cap.read()
+      img = Image.fromarray(frame[:, :, [2,1,0]])
+      lt = time.monotonic_ns()
+      out, retimg = infer(model, img)
+      print(f"{(time.monotonic_ns()-lt)*1e-6:7.2f} ms", np.argmax(out), np.max(out), lbls[np.argmax(out)])
+      SCALE = 3
+      simg = cv2.resize(retimg, (224*SCALE, 224*SCALE))
+      retimg = cv2.cvtColor(simg, cv2.COLOR_RGB2BGR)
+      cv2.imshow('capture', retimg)
+      if cv2.waitKey(1) & 0xFF == ord('q'):
+        break
+    cap.release()
+    cv2.destroyAllWindows()
+  else:
+    img = Image.open(fetch(url))
+    for i in range(getenv("CNT", 1)):
+      with Timing("did inference in "):
+        out, _ = infer(model, img)
+        print(np.argmax(out), np.max(out), lbls[np.argmax(out)])
@@ -0,0 +1,498 @@
+# pip3 install sentencepiece
+
+# This file incorporates code from the following:
+# Github Name                    | License | Link
+# black-forest-labs/flux         | Apache  | https://github.com/black-forest-labs/flux/tree/main/model_licenses
+
+from tinygrad import Tensor, nn, dtypes, TinyJit
+from tinygrad.nn.state import safe_load, load_state_dict
+from tinygrad.helpers import fetch, tqdm, colored
+from sdxl import FirstStage
+from extra.models.clip import FrozenClosedClipEmbedder
+from extra.models.t5 import T5Embedder
+import numpy as np
+
+import math, time, argparse, tempfile
+from typing import List, Dict, Optional, Union, Tuple, Callable
+from dataclasses import dataclass
+from pathlib import Path
+from PIL import Image
+
+urls:dict = {
+  "flux-schnell": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/flux1-schnell.safetensors",
+  "flux-dev": "https://huggingface.co/camenduru/FLUX.1-dev/resolve/main/flux1-dev.sft",
+  "ae": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/ae.safetensors",
+  "T5_1_of_2": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/text_encoder_2/model-00001-of-00002.safetensors",
+  "T5_2_of_2": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/text_encoder_2/model-00002-of-00002.safetensors",
+  "T5_tokenizer": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/tokenizer_2/spiece.model",
+  "clip": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/text_encoder/model.safetensors"
+}
+
+def tensor_identity(x:Tensor) -> Tensor: return x
+
+class AutoEncoder:
+  def __init__(self, scale_factor:float, shift_factor:float):
+    self.decoder = FirstStage.Decoder(128, 3, 3, 16, [1, 2, 4, 4], 2, 256)
+    self.scale_factor = scale_factor
+    self.shift_factor = shift_factor
+
+  def decode(self, z:Tensor) -> Tensor:
+    z = z / self.scale_factor + self.shift_factor
+    return self.decoder(z)
+
+# Conditioner
+class ClipEmbedder(FrozenClosedClipEmbedder):
+  def __call__(self, texts:Union[str, List[str], Tensor]) -> Tensor:
+    if isinstance(texts, str): texts = [texts]
+    assert isinstance(texts, (list,tuple)), f"expected list of strings, got {type(texts).__name__}"
+    tokens = Tensor.cat(*[Tensor(self.tokenizer.encode(text)) for text in texts], dim=0)
+    return self.transformer.text_model(tokens.reshape(len(texts),-1))[:, tokens.argmax(-1)]
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/math.py
+def attention(q:Tensor, k:Tensor, v:Tensor, pe:Tensor) -> Tensor:
+  q, k = apply_rope(q, k, pe)
+  x = Tensor.scaled_dot_product_attention(q, k, v)
+  return x.rearrange("B H L D -> B L (H D)")
+
+def rope(pos:Tensor, dim:int, theta:int) -> Tensor:
+  assert dim % 2 == 0
+  scale = Tensor.arange(0, dim, 2, dtype=dtypes.float32, device=pos.device) / dim # NOTE: this is torch.float64 in reference implementation
+  omega = 1.0 / (theta**scale)
+  out = Tensor.einsum("...n,d->...nd", pos, omega)
+  out = Tensor.stack(Tensor.cos(out), -Tensor.sin(out), Tensor.sin(out), Tensor.cos(out), dim=-1)
+  out = out.rearrange("b n d (i j) -> b n d i j", i=2, j=2)
+  return out.float()
+
+def apply_rope(xq:Tensor, xk:Tensor, freqs_cis:Tensor) -> Tuple[Tensor, Tensor]:
+  xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
+  xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+  xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
+  xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
+  return xq_out.reshape(*xq.shape).cast(xq.dtype), xk_out.reshape(*xk.shape).cast(xk.dtype)
+
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/modules/layers.py
+class EmbedND:
+  def __init__(self, dim:int, theta:int, axes_dim:List[int]):
+    self.dim = dim
+    self.theta = theta
+    self.axes_dim = axes_dim
+
+  def __call__(self, ids:Tensor) -> Tensor:
+    n_axes = ids.shape[-1]
+    emb = Tensor.cat(*[rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)], dim=-3)
+    return emb.unsqueeze(1)
+
+class MLPEmbedder:
+  def __init__(self, in_dim:int, hidden_dim:int):
+    self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
+    self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)
+
+  def __call__(self, x:Tensor) -> Tensor:
+    return self.out_layer(self.in_layer(x).silu())
+
+class QKNorm:
+  def __init__(self, dim:int):
+    self.query_norm = nn.RMSNorm(dim)
+    self.key_norm = nn.RMSNorm(dim)
+
+  def __call__(self, q:Tensor, k:Tensor) -> Tuple[Tensor, Tensor]:
+    return self.query_norm(q), self.key_norm(k)
+
+class SelfAttention:
+  def __init__(self, dim:int, num_heads:int = 8, qkv_bias:bool = False):
+    self.num_heads = num_heads
+    head_dim = dim // num_heads
+
+    self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+    self.norm = QKNorm(head_dim)
+    self.proj = nn.Linear(dim, dim)
+
+  def __call__(self, x:Tensor, pe:Tensor) -> Tensor:
+    qkv = self.qkv(x)
+    q, k, v = qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    q, k = self.norm(q, k)
+    x = attention(q, k, v, pe=pe)
+    return self.proj(x)
+
+@dataclass
+class ModulationOut:
+  shift:Tensor
+  scale:Tensor
+  gate:Tensor
+
+class Modulation:
+  def __init__(self, dim:int, double:bool):
+    self.is_double = double
+    self.multiplier = 6 if double else 3
+    self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
+
+  def __call__(self, vec:Tensor) -> Tuple[ModulationOut, Optional[ModulationOut]]:
+    out = self.lin(vec.silu())[:, None, :].chunk(self.multiplier, dim=-1)
+    return ModulationOut(*out[:3]), ModulationOut(*out[3:]) if self.is_double else None
+
+class DoubleStreamBlock:
+  def __init__(self, hidden_size:int, num_heads:int, mlp_ratio:float, qkv_bias:bool = False):
+    mlp_hidden_dim = int(hidden_size * mlp_ratio)
+    self.num_heads = num_heads
+    self.hidden_size = hidden_size
+    self.img_mod = Modulation(hidden_size, double=True)
+    self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+    self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.img_mlp = [nn.Linear(hidden_size, mlp_hidden_dim, bias=True), Tensor.gelu, nn.Linear(mlp_hidden_dim, hidden_size, bias=True)]
+
+    self.txt_mod = Modulation(hidden_size, double=True)
+    self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+    self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.txt_mlp = [nn.Linear(hidden_size, mlp_hidden_dim, bias=True), Tensor.gelu, nn.Linear(mlp_hidden_dim, hidden_size, bias=True)]
+
+  def __call__(self, img:Tensor, txt:Tensor, vec:Tensor, pe:Tensor) -> tuple[Tensor, Tensor]:
+    img_mod1, img_mod2 = self.img_mod(vec)
+    txt_mod1, txt_mod2 = self.txt_mod(vec)
+    assert img_mod2 is not None and txt_mod2 is not None
+    # prepare image for attention
+    img_modulated = self.img_norm1(img)
+    img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+    img_qkv = self.img_attn.qkv(img_modulated)
+    img_q, img_k, img_v = img_qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    img_q, img_k = self.img_attn.norm(img_q, img_k)
+
+    # prepare txt for attention
+    txt_modulated = self.txt_norm1(txt)
+    txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+    txt_qkv = self.txt_attn.qkv(txt_modulated)
+    txt_q, txt_k, txt_v = txt_qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k)
+
+    # run actual attention
+    q = Tensor.cat(txt_q, img_q, dim=2)
+    k = Tensor.cat(txt_k, img_k, dim=2)
+    v = Tensor.cat(txt_v, img_v, dim=2)
+
+    attn = attention(q, k, v, pe=pe)
+    txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+    # calculate the img bloks
+    img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+    img = img + img_mod2.gate * ((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift).sequential(self.img_mlp)
+
+    # calculate the txt bloks
+    txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
+    txt = txt + txt_mod2.gate * ((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift).sequential(self.txt_mlp)
+    return img, txt
+
+
+class SingleStreamBlock:
+  """
+  A DiT block with parallel linear layers as described in
+  https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+  """
+
+  def __init__(self,hidden_size:int, num_heads:int, mlp_ratio:float=4.0, qk_scale:Optional[float]=None):
+    self.hidden_dim = hidden_size
+    self.num_heads = num_heads
+    head_dim = hidden_size // num_heads
+    self.scale = qk_scale or head_dim**-0.5
+
+    self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+    # qkv and mlp_in
+    self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
+    # proj and mlp_out
+    self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
+
+    self.norm = QKNorm(head_dim)
+
+    self.hidden_size = hidden_size
+    self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+    self.mlp_act = Tensor.gelu
+    self.modulation = Modulation(hidden_size, double=False)
+
+  def __call__(self, x:Tensor, vec:Tensor, pe:Tensor) -> Tensor:
+    mod, _ = self.modulation(vec)
+    x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+    qkv, mlp = Tensor.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+    q, k, v = qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    q, k = self.norm(q, k)
+
+    # compute attention
+    attn = attention(q, k, v, pe=pe)
+    # compute activation in mlp stream, cat again and run second linear layer
+    output = self.linear2(Tensor.cat(attn, self.mlp_act(mlp), dim=2))
+    return x + mod.gate * output
+
+
+class LastLayer:
+  def __init__(self, hidden_size:int, patch_size:int, out_channels:int):
+    self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+    self.adaLN_modulation:List[Callable[[Tensor], Tensor]] = [Tensor.silu, nn.Linear(hidden_size, 2 * hidden_size, bias=True)]
+
+  def __call__(self, x:Tensor, vec:Tensor) -> Tensor:
+    shift, scale = vec.sequential(self.adaLN_modulation).chunk(2, dim=1)
+    x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+    return self.linear(x)
+
+def timestep_embedding(t:Tensor, dim:int, max_period:int=10000, time_factor:float=1000.0) -> Tensor:
+  """
+  Create sinusoidal timestep embeddings.
+  :param t: a 1-D Tensor of N indices, one per batch element.
+                    These may be fractional.
+  :param dim: the dimension of the output.
+  :param max_period: controls the minimum frequency of the embeddings.
+  :return: an (N, D) Tensor of positional embeddings.
+  """
+  t = time_factor * t
+  half = dim // 2
+  freqs = Tensor.exp(-math.log(max_period) * Tensor.arange(0, stop=half, dtype=dtypes.float32) / half).to(t.device)
+
+  args = t[:, None].float() * freqs[None]
+  embedding = Tensor.cat(Tensor.cos(args), Tensor.sin(args), dim=-1)
+  if dim % 2:  embedding = Tensor.cat(*[embedding, Tensor.zeros_like(embedding[:, :1])], dim=-1)
+  if Tensor.is_floating_point(t):  embedding = embedding.cast(t.dtype)
+  return embedding
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/model.py
+class Flux:
+  """
+  Transformer model for flow matching on sequences.
+  """
+
+  def __init__(
+      self,
+      guidance_embed:bool,
+      in_channels:int = 64,
+      vec_in_dim:int = 768,
+      context_in_dim:int = 4096,
+      hidden_size:int = 3072,
+      mlp_ratio:float = 4.0,
+      num_heads:int = 24,
+      depth:int = 19,
+      depth_single_blocks:int = 38,
+      axes_dim:Optional[List[int]] = None,
+      theta:int = 10_000,
+      qkv_bias:bool = True,
+      ):
+
+    axes_dim = axes_dim or [16, 56, 56]
+    self.guidance_embed = guidance_embed
+    self.in_channels = in_channels
+    self.out_channels = self.in_channels
+    if hidden_size % num_heads != 0:
+      raise ValueError(f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}")
+    pe_dim = hidden_size // num_heads
+    if sum(axes_dim) != pe_dim:
+      raise ValueError(f"Got {axes_dim} but expected positional dim {pe_dim}")
+    self.hidden_size = hidden_size
+    self.num_heads = num_heads
+    self.pe_embedder = EmbedND(dim=pe_dim, theta=theta, axes_dim=axes_dim)
+    self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=True)
+    self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
+    self.vector_in = MLPEmbedder(vec_in_dim, self.hidden_size)
+    self.guidance_in:Callable[[Tensor], Tensor] = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if guidance_embed else tensor_identity
+    self.txt_in = nn.Linear(context_in_dim, self.hidden_size)
+
+    self.double_blocks = [DoubleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias) for _ in range(depth)]
+    self.single_blocks = [SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=mlp_ratio) for _ in range(depth_single_blocks)]
+    self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels)
+
+  def __call__(self, img:Tensor, img_ids:Tensor, txt:Tensor, txt_ids:Tensor, timesteps:Tensor, y:Tensor, guidance:Optional[Tensor] = None) -> Tensor:
+    if img.ndim != 3 or txt.ndim != 3:
+      raise ValueError("Input img and txt tensors must have 3 dimensions.")
+    # running on sequences img
+    img = self.img_in(img)
+    vec = self.time_in(timestep_embedding(timesteps, 256))
+    if self.guidance_embed:
+      if guidance is None:
+        raise ValueError("Didn't get guidance strength for guidance distilled model.")
+      vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+    vec = vec + self.vector_in(y)
+    txt = self.txt_in(txt)
+    ids = Tensor.cat(txt_ids, img_ids, dim=1)
+    pe = self.pe_embedder(ids)
+    for double_block in self.double_blocks:
+      img, txt = double_block(img=img, txt=txt, vec=vec, pe=pe)
+
+    img = Tensor.cat(txt, img, dim=1)
+    for single_block in self.single_blocks:
+      img = single_block(img, vec=vec, pe=pe)
+
+    img = img[:, txt.shape[1] :, ...]
+
+    return self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/util.py
+def load_flow_model(name:str, model_path:str):
+  # Loading Flux
+  print("Init model")
+  model = Flux(guidance_embed=(name != "flux-schnell"))
+  if not model_path: model_path = fetch(urls[name])
+  state_dict = {k.replace("scale", "weight"): v for k, v in safe_load(model_path).items()}
+  load_state_dict(model, state_dict)
+  return model
+
+def load_T5(max_length:int=512):
+  # max length 64, 128, 256 and 512 should work (if your sequence is short enough)
+  print("Init T5")
+  T5 = T5Embedder(max_length, fetch(urls["T5_tokenizer"]))
+  pt_1 = fetch(urls["T5_1_of_2"])
+  pt_2 = fetch(urls["T5_2_of_2"])
+  load_state_dict(T5.encoder, safe_load(pt_1) | safe_load(pt_2), strict=False)
+  return T5
+
+def load_clip():
+  print("Init Clip")
+  clip = ClipEmbedder()
+  load_state_dict(clip.transformer, safe_load(fetch(urls["clip"])))
+  return clip
+
+def load_ae() -> AutoEncoder:
+  # Loading the autoencoder
+  print("Init AE")
+  ae = AutoEncoder(0.3611, 0.1159)
+  load_state_dict(ae, safe_load(fetch(urls["ae"])))
+  return ae
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/sampling.py
+def prepare(T5:T5Embedder, clip:ClipEmbedder, img:Tensor, prompt:Union[str, List[str]]) -> Dict[str, Tensor]:
+  bs, _, h, w = img.shape
+  if bs == 1 and not isinstance(prompt, str):
+    bs = len(prompt)
+
+  img = img.rearrange("b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+  if img.shape[0] == 1 and bs > 1:
+    img = img.expand((bs, *img.shape[1:]))
+
+  img_ids = Tensor.zeros(h // 2, w // 2, 3).contiguous()
+  img_ids[..., 1] = img_ids[..., 1] + Tensor.arange(h // 2)[:, None]
+  img_ids[..., 2] = img_ids[..., 2] + Tensor.arange(w // 2)[None, :]
+  img_ids = img_ids.rearrange("h w c -> 1 (h w) c")
+  img_ids = img_ids.expand((bs, *img_ids.shape[1:]))
+
+  if isinstance(prompt, str):
+    prompt = [prompt]
+  txt = T5(prompt).realize()
+  if txt.shape[0] == 1 and bs > 1:
+    txt = txt.expand((bs, *txt.shape[1:]))
+  txt_ids = Tensor.zeros(bs, txt.shape[1], 3)
+
+  vec = clip(prompt).realize()
+  if vec.shape[0] == 1 and bs > 1:
+    vec = vec.expand((bs, *vec.shape[1:]))
+
+  return {"img": img, "img_ids": img_ids.to(img.device), "txt": txt.to(img.device), "txt_ids": txt_ids.to(img.device), "vec": vec.to(img.device)}
+
+
+def get_schedule(num_steps:int, image_seq_len:int, base_shift:float=0.5, max_shift:float=1.15, shift:bool=True) -> List[float]:
+  # extra step for zero
+  step_size = -1.0 / num_steps
+  timesteps = Tensor.arange(1, 0 + step_size, step_size)
+
+  # shifting the schedule to favor high timesteps for higher signal images
+  if shift:
+    # estimate mu based on linear estimation between two points
+    mu = 0.5 + (max_shift - base_shift) * (image_seq_len - 256) / (4096 - 256)
+    timesteps = math.exp(mu) / (math.exp(mu) + (1 / timesteps - 1))
+  return timesteps.tolist()
+
+@TinyJit
+def run(model, *args): return model(*args).realize()
+
+def denoise(model, img:Tensor, img_ids:Tensor, txt:Tensor, txt_ids:Tensor, vec:Tensor, timesteps:List[float], guidance:float=4.0) -> Tensor:
+  # this is ignored for schnell
+  guidance_vec = Tensor((guidance,), device=img.device, dtype=img.dtype).expand((img.shape[0],))
+  for t_curr, t_prev in tqdm(list(zip(timesteps[:-1], timesteps[1:])), "Denoising"):
+    t_vec = Tensor((t_curr,), device=img.device, dtype=img.dtype).expand((img.shape[0],))
+    pred = run(model, img, img_ids, txt, txt_ids, t_vec, vec, guidance_vec)
+    img = img + (t_prev - t_curr) * pred
+
+  return img
+
+def unpack(x:Tensor, height:int, width:int) -> Tensor:
+  return x.rearrange("b (h w) (c ph pw) -> b c (h ph) (w pw)", h=math.ceil(height / 16), w=math.ceil(width / 16), ph=2, pw=2)
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/cli.py
+if __name__ == "__main__":
+  default_prompt = "bananas and a can of coke"
+  parser = argparse.ArgumentParser(description="Run Flux.1", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+
+  parser.add_argument("--name",       type=str,   default="flux-schnell", help="Name of the model to load")
+  parser.add_argument("--model_path", type=str,   default="",             help="path of the model file")
+  parser.add_argument("--width",      type=int,   default=512,            help="width of the sample in pixels (should be a multiple of 16)")
+  parser.add_argument("--height",     type=int,   default=512,            help="height of the sample in pixels (should be a multiple of 16)")
+  parser.add_argument("--seed",       type=int,   default=None,           help="Set a seed for sampling")
+  parser.add_argument("--prompt",     type=str,   default=default_prompt, help="Prompt used for sampling")
+  parser.add_argument('--out',        type=str,   default=Path(tempfile.gettempdir()) / "rendered.png", help="Output filename")
+  parser.add_argument("--num_steps",  type=int,   default=None,           help="number of sampling steps (default 4 for schnell, 50 for guidance distilled)") #noqa:E501
+  parser.add_argument("--guidance",   type=float, default=3.5,            help="guidance value used for guidance distillation")
+  parser.add_argument("--output_dir", type=str,   default="output",       help="output directory")
+  args = parser.parse_args()
+
+  if args.name not in ["flux-schnell", "flux-dev"]:
+    raise ValueError(f"Got unknown model name: {args.name}, chose from flux-schnell and flux-dev")
+
+  if args.num_steps is None:
+    args.num_steps = 4 if args.name == "flux-schnell" else 50
+
+  # allow for packing and conversion to latent space
+  height = 16 * (args.height // 16)
+  width = 16 * (args.width // 16)
+
+  if args.seed is None: args.seed = Tensor._seed
+  else: Tensor.manual_seed(args.seed)
+
+  print(f"Generating with seed {args.seed}:\n{args.prompt}")
+  t0 = time.perf_counter()
+
+  # prepare input noise
+  x = Tensor.randn(1, 16, 2 * math.ceil(height / 16), 2 * math.ceil(width / 16), dtype="bfloat16")
+
+  # load text embedders
+  T5 = load_T5(max_length=256 if args.name == "flux-schnell" else 512)
+  clip = load_clip()
+
+  # embed text to get inputs for model
+  inp = prepare(T5, clip, x, prompt=args.prompt)
+  timesteps = get_schedule(args.num_steps, inp["img"].shape[1], shift=(args.name != "flux-schnell"))
+
+  # done with text embedders
+  del T5, clip
+
+  # load model
+  model = load_flow_model(args.name, args.model_path)
+
+  # denoise initial noise
+  x = denoise(model, **inp, timesteps=timesteps, guidance=args.guidance)
+
+  # done with model
+  del model, run
+
+  # load autoencoder
+  ae = load_ae()
+
+  # decode latents to pixel space
+  x = unpack(x.float(), height, width)
+  x = ae.decode(x).realize()
+
+  t1 = time.perf_counter()
+  print(f"Done in {t1 - t0:.1f}s. Saving {args.out}")
+
+  # bring into PIL format and save
+  x = x.clamp(-1, 1)
+  x = x[0].rearrange("c h w -> h w c")
+  x = (127.5 * (x + 1.0)).cast("uint8")
+
+  img = Image.fromarray(x.numpy())
+
+  img.save(args.out)
+
+  # validation!
+  if args.prompt == default_prompt and args.name=="flux-schnell" and args.seed == 0 and args.width == args.height == 512:
+    ref_image = Tensor(np.array(Image.open("examples/flux1_seed0.png")))
+    distance = (((x.cast(dtypes.float) - ref_image.cast(dtypes.float)) / ref_image.max())**2).mean().item()
+    assert distance < 4e-3, colored(f"validation failed with {distance=}", "red")
+    print(colored(f"output validated with {distance=}", "green"))
@@ -232,7 +232,7 @@ if __name__ == "__main__":
  gpt2 = GPT2.build_gguf(args.model_size) if args.model_size.startswith("gpt2_gguf_") else GPT2.build(args.model_size)

  if args.benchmark != -1:
-    gpt2.model(Tensor.randint(args.batch_size, args.benchmark), Variable("a", 0, MAX_CONTEXT).bind(0)).realize()
+    gpt2.model(Tensor.rand(args.batch_size, args.benchmark), Variable("a", 0, MAX_CONTEXT).bind(0)).realize()
  else:
    texts = gpt2.generate(args.prompt, args.count, args.temperature, timing=args.timing, batch_size=args.batch_size)
    if not args.noshow:
@@ -1,108 +0,0 @@
-import itertools
-from typing import Callable
-from tinygrad import nn, Tensor, dtypes, Device, TinyJit
-from tinygrad.helpers import getenv, trange, partition
-
-class Model:
-  def __init__(self):
-    self.layers: list[Callable[[Tensor], Tensor]] = [
-      nn.Conv2d(1, 32, 5), Tensor.relu,
-      nn.Conv2d(32, 32, 5), Tensor.relu,
-      nn.BatchNorm(32), Tensor.max_pool2d,
-      nn.Conv2d(32, 64, 3), Tensor.relu,
-      nn.Conv2d(64, 64, 3), Tensor.relu,
-      nn.BatchNorm(64), Tensor.max_pool2d,
-      lambda x: x.flatten(1), nn.Linear(576, 10)]
-
-  def __call__(self, x:Tensor) -> Tensor: return x.sequential(self.layers)
-
-# TODO: refactor this into optim/onnx
-def functional_adam(g:Tensor, m:Tensor, v:Tensor, b1_t:Tensor, b2_t:Tensor, lr=0.001, b1=0.9, b2=0.999, eps=1e-6) -> Tensor:
-  b1_t *= b1
-  b2_t *= b2
-  m.assign(b1 * m + (1.0 - b1) * g)
-  v.assign(b2 * v + (1.0 - b2) * (g * g))
-  m_hat = m / (1.0 - b1_t)
-  v_hat = v / (1.0 - b2_t)
-  return lr * (m_hat / (v_hat.sqrt() + eps))
-
-if __name__ == "__main__":
-  BS = getenv("BS", 512)
-  ACC_STEPS = getenv("ACC_STEPS", 8)
-
-  X_train, Y_train, X_test, Y_test = nn.datasets.mnist()
-  model = Model()
-
-  params = nn.state.get_parameters(model)
-
-  # init params, set requires grad on the ones we need gradients of
-  for x in params:
-    if x.requires_grad is None: x.requires_grad_()
-    x.replace(x.contiguous())
-  Tensor.realize(*params)
-
-  # split params (with grads) and buffers (without)
-  params, buffers = partition(params, lambda x: x.requires_grad)
-  print(f"params: {len(params)} buffers: {len(buffers)}")
-
-  # optim params
-  pos_params = list(itertools.accumulate(params, lambda x,y: x+y.numel(), initial=0))
-  adam_m = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
-  adam_v = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
-  adam_b1_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False).contiguous()
-  adam_b2_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False).contiguous()
-  adam_params = [adam_m, adam_v, adam_b1_t, adam_b2_t]
-
-  # create loss and grads. init all state so the JIT works on microbatch
-  for x in params: x.assign(x.detach())
-  loss = Tensor.zeros(tuple()).contiguous()
-  grads = Tensor.zeros(pos_params[-1]).contiguous()
-  Tensor.realize(*params, *buffers, *adam_params, loss, grads)
-
-  @TinyJit
-  @Tensor.train()
-  def microbatch():
-    samples = Tensor.randint(BS // ACC_STEPS, high=X_train.shape[0])
-    for t in params: t.grad = None
-    # divide by ACC_STEPS at the loss
-    uloss = (model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]) / ACC_STEPS).backward()
-    ugrads = Tensor.cat(*[t.grad.contiguous().flatten() for t in params], dim=0)
-    for t in params: t.grad = None
-    # concat the grads and assign them
-    loss.assign(loss + uloss)
-    grads.assign(grads + ugrads)
-    Tensor.realize(*params, *buffers, loss, grads)
-
-  @TinyJit
-  def optimizer():
-    # run optimizer (on CPU, where adam params live)
-    delta = functional_adam(grads.to("CPU"), adam_m, adam_v, adam_b1_t, adam_b2_t)
-
-    # update the params, copying back the delta one at a time to avoid OOM
-    # NOTE: the scheduler is ordering things poorly, all the copies are happening before the adds
-    for j,tt in enumerate(params):
-      tt.assign(tt.detach() - delta[pos_params[j]:pos_params[j+1]].reshape(tt.shape).to(Device.DEFAULT))
-
-    # realize everything, zero out loss and grads
-    loss.assign(Tensor.zeros_like(loss))
-    grads.assign(Tensor.zeros_like(grads))
-    Tensor.realize(*params, *adam_params, loss, grads)
-
-  @TinyJit
-  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100
-
-  test_acc = float('nan')
-  for i in (t:=trange(getenv("STEPS", 70))):
-    # microbatch sets the gradients
-    for _ in range(ACC_STEPS): microbatch()
-
-    # get the loss before the optimizer clears it
-    # this is already realized so this isn't a schedule
-    loss_item = loss.item()
-
-    # run the optimizer
-    optimizer()
-
-    # eval
-    if i%10 == 9: test_acc = get_test_acc().item()
-    t.set_description(f"loss: {loss_item:6.2f} test_accuracy: {test_acc:5.2f}%")
@@ -19,8 +19,8 @@ cifar_std = [0.24703225141799082, 0.24348516474564, 0.26158783926049628]
 BS, STEPS = getenv("BS", 512), getenv("STEPS", 1000)
 EVAL_BS = getenv("EVAL_BS", BS)
 GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]
-assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}"
-assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}"
+assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"
+assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"

 class UnsyncedBatchNorm:
  def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1, num_devices=len(GPUS)):
@@ -1,6 +1,8 @@
 from pathlib import Path
 from typing import List
 import json, argparse, random, time, os
+import tiktoken
+from tiktoken.load import load_tiktoken_bpe
 from extra.models.llama import Transformer, convert_from_huggingface, convert_from_gguf, fix_bf16
 from tinygrad.nn.state import safe_load, torch_load, load_state_dict, get_parameters, gguf_load
 from tinygrad import Tensor, dtypes, nn, Context, Device, GlobalCounters
@@ -10,8 +12,6 @@ from extra.bench_log import BenchEvent, WallTimeEvent
 class Tokenizer:
  pat_str = r"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"
  def __init__(self, model_path: str):
-    import tiktoken
-    from tiktoken.load import load_tiktoken_bpe
    mergeable_ranks = load_tiktoken_bpe(model_path)
    self.num_base_tokens = len(mergeable_ranks)
    special_tokens = [
@@ -145,41 +145,6 @@ def NF4Linear(block_size):
      return new_state_dict
  return _NF4Linear

-def quantize_to_fp8(x: Tensor, dtype=dtypes.fp8e4m3):
-  fp8_min = -448.0 if dtype == dtypes.fp8e4m3 else -57344.0
-  fp8_max = 448.0 if dtype == dtypes.fp8e4m3 else 57344.0
-  scale = fp8_max / x.abs().max()
-  x_scl_sat = (x * scale).clamp(fp8_min, fp8_max)
-  return x_scl_sat.cast(dtype), scale.float().reciprocal()
-
-class FP8Linear:
-  def __init__(self, in_features, out_features, bias=True):
-    self.weight = Tensor.empty(out_features, in_features, dtype=dtypes.fp8e4m3)
-    self.bias = Tensor.empty(out_features, dtype=dtypes.float16) if bias else None
-    self.weight_scale = Tensor.empty((), dtype=dtypes.float16)
-
-  def __call__(self, x:Tensor):
-    y = x.dot(self.weight.T.cast(dtypes.float32)) * self.weight_scale
-    if self.bias is not None: y = y + self.bias.cast(y.dtype)
-    return y.cast(x.dtype)
-
-  @staticmethod
-  def quantize(tensors, device, scale_dtype=dtypes.float16, quantize_embeds=False):
-    assert not quantize_embeds
-    new_tensors = {}
-    for name,v in tensors.items():
-      if "feed_forward" in name or "attention.w" in name:
-        assert "weight" in name, name
-        fp8_weight, scale = quantize_to_fp8(v)
-        new_tensors[name] = fp8_weight
-        new_tensors[name.replace('weight', 'weight_scale')] = scale.cast(scale_dtype)
-        if isinstance(device, tuple):
-          new_tensors[name].shard_(device, axis=-1)
-          new_tensors[name.replace('weight', 'weight_scale')].shard_(device, axis=None)
-      else:
-        new_tensors[name] = v
-    return new_tensors
-
 MODEL_PARAMS = {
  "1B": {
    "args": {"dim": 2048, "n_heads": 32, "n_kv_heads": 8, "n_layers": 16, "norm_eps": 1e-5, "rope_theta": 500000, "vocab_size": 128256, "hidden_dim": 8192},
@@ -202,7 +167,6 @@ def build_transformer(model_path: Path, model_size="8B", quantize=None, scale_dt
  # build model
  if quantize == "int8": linear, embedding, quantize_embeds = Int8Linear, Int8Embedding, True
  elif quantize == "nf4": linear, embedding, quantize_embeds = NF4Linear(64), nn.Embedding, False
-  elif quantize == "fp8": linear, embedding, quantize_embeds = FP8Linear, nn.Embedding, False
  else: linear, embedding, quantize_embeds = nn.Linear, nn.Embedding, False
  model = Transformer(**MODEL_PARAMS[model_size]["args"], linear=linear, embedding=embedding, max_context=max_context, jit=True)

@@ -278,7 +242,7 @@ if __name__ == "__main__":
  parser.add_argument("--model", type=Path, help="Model path")
  parser.add_argument("--size", choices=["1B", "8B", "70B", "405B"], default="1B", help="Model size")
  parser.add_argument("--shard", type=int, default=1, help="Shard the model across multiple devices")
-  parser.add_argument("--quantize", choices=["int8", "nf4", "float16", "fp8"], help="Quantization method")
+  parser.add_argument("--quantize", choices=["int8", "nf4", "float16"], help="Quantization method")
  parser.add_argument("--no_api", action="store_true", help="Disable the api and run a cli test interface")
  parser.add_argument("--host", type=str, default="0.0.0.0", help="Web server bind address")
  parser.add_argument("--port", type=int, default=7776, help="Web server port")
@@ -0,0 +1,299 @@
+from extra.models.mask_rcnn import MaskRCNN
+from extra.models.resnet import ResNet
+from extra.models.mask_rcnn import BoxList
+from torch.nn import functional as F
+from torchvision import transforms as T
+from torchvision.transforms import functional as Ft
+import random
+from tinygrad.tensor import Tensor
+from PIL import Image
+import numpy as np
+import torch
+import argparse
+import cv2
+
+
+class Resize:
+  def __init__(self, min_size, max_size):
+    if not isinstance(min_size, (list, tuple)):
+      min_size = (min_size,)
+    self.min_size = min_size
+    self.max_size = max_size
+
+  # modified from torchvision to add support for max size
+  def get_size(self, image_size):
+    w, h = image_size
+    size = random.choice(self.min_size)
+    max_size = self.max_size
+    if max_size is not None:
+      min_original_size = float(min((w, h)))
+      max_original_size = float(max((w, h)))
+      if max_original_size / min_original_size * size > max_size:
+        size = int(round(max_size * min_original_size / max_original_size))
+
+      if (w <= h and w == size) or (h <= w and h == size):
+        return (h, w)
+
+      if w < h:
+        ow = size
+        oh = int(size * h / w)
+      else:
+        oh = size
+        ow = int(size * w / h)
+
+      return (oh, ow)
+
+  def __call__(self, image):
+    size = self.get_size(image.size)
+    image = Ft.resize(image, size)
+    return image
+
+
+class Normalize:
+  def __init__(self, mean, std, to_bgr255=True):
+    self.mean = mean
+    self.std = std
+    self.to_bgr255 = to_bgr255
+
+  def __call__(self, image):
+    if self.to_bgr255:
+      image = image[[2, 1, 0]] * 255
+    else:
+      image = image[[0, 1, 2]] * 255
+    image = Ft.normalize(image, mean=self.mean, std=self.std)
+    return image
+
+transforms = lambda size_scale: T.Compose(
+  [
+    Resize(int(800*size_scale), int(1333*size_scale)),
+    T.ToTensor(),
+    Normalize(
+      mean=[102.9801, 115.9465, 122.7717], std=[1., 1., 1.], to_bgr255=True
+    ),
+  ]
+)
+
+def expand_boxes(boxes, scale):
+  w_half = (boxes[:, 2] - boxes[:, 0]) * .5
+  h_half = (boxes[:, 3] - boxes[:, 1]) * .5
+  x_c = (boxes[:, 2] + boxes[:, 0]) * .5
+  y_c = (boxes[:, 3] + boxes[:, 1]) * .5
+
+  w_half *= scale
+  h_half *= scale
+
+  boxes_exp = torch.zeros_like(boxes)
+  boxes_exp[:, 0] = x_c - w_half
+  boxes_exp[:, 2] = x_c + w_half
+  boxes_exp[:, 1] = y_c - h_half
+  boxes_exp[:, 3] = y_c + h_half
+  return boxes_exp
+
+
+def expand_masks(mask, padding):
+  N = mask.shape[0]
+  M = mask.shape[-1]
+  pad2 = 2 * padding
+  scale = float(M + pad2) / M
+  padded_mask = mask.new_zeros((N, 1, M + pad2, M + pad2))
+  padded_mask[:, :, padding:-padding, padding:-padding] = mask
+  return padded_mask, scale
+
+
+def paste_mask_in_image(mask, box, im_h, im_w, thresh=0.5, padding=1):
+  # TODO: remove torch
+  mask = torch.tensor(mask.numpy())
+  box = torch.tensor(box.numpy())
+  padded_mask, scale = expand_masks(mask[None], padding=padding)
+  mask = padded_mask[0, 0]
+  box = expand_boxes(box[None], scale)[0]
+  box = box.to(dtype=torch.int32)
+
+  TO_REMOVE = 1
+  w = int(box[2] - box[0] + TO_REMOVE)
+  h = int(box[3] - box[1] + TO_REMOVE)
+  w = max(w, 1)
+  h = max(h, 1)
+
+  mask = mask.expand((1, 1, -1, -1))
+
+  mask = mask.to(torch.float32)
+  mask = F.interpolate(mask, size=(h, w), mode='bilinear', align_corners=False)
+  mask = mask[0][0]
+
+  if thresh >= 0:
+    mask = mask > thresh
+  else:
+    mask = (mask * 255).to(torch.uint8)
+
+  im_mask = torch.zeros((im_h, im_w), dtype=torch.uint8)
+  x_0 = max(box[0], 0)
+  x_1 = min(box[2] + 1, im_w)
+  y_0 = max(box[1], 0)
+  y_1 = min(box[3] + 1, im_h)
+
+  im_mask[y_0:y_1, x_0:x_1] = mask[
+                              (y_0 - box[1]): (y_1 - box[1]), (x_0 - box[0]): (x_1 - box[0])
+                              ]
+  return im_mask
+
+
+class Masker:
+  def __init__(self, threshold=0.5, padding=1):
+    self.threshold = threshold
+    self.padding = padding
+
+  def forward_single_image(self, masks, boxes):
+    boxes = boxes.convert("xyxy")
+    im_w, im_h = boxes.size
+    res = [
+      paste_mask_in_image(mask[0], box, im_h, im_w, self.threshold, self.padding)
+      for mask, box in zip(masks, boxes.bbox)
+    ]
+    if len(res) > 0:
+      res = torch.stack(*res, dim=0)[:, None]
+    else:
+      res = masks.new_empty((0, 1, masks.shape[-2], masks.shape[-1]))
+    return Tensor(res.numpy())
+
+  def __call__(self, masks, boxes):
+    if isinstance(boxes, BoxList):
+      boxes = [boxes]
+
+    results = []
+    for mask, box in zip(masks, boxes):
+      result = self.forward_single_image(mask, box)
+      results.append(result)
+    return results
+
+
+masker = Masker(threshold=0.5, padding=1)
+
+def select_top_predictions(predictions, confidence_threshold=0.9):
+  scores = predictions.get_field("scores").numpy()
+  keep = [idx for idx, score in enumerate(scores) if score > confidence_threshold]
+  return predictions[keep]
+
+def compute_prediction(original_image, model, confidence_threshold, size_scale=1.0):
+  image = transforms(size_scale)(original_image).numpy()
+  image = Tensor(image, requires_grad=False)
+  predictions = model(image)
+  prediction = predictions[0]
+  prediction = select_top_predictions(prediction, confidence_threshold)
+  width, height = original_image.size
+  prediction = prediction.resize((width, height))
+
+  if prediction.has_field("mask"):
+    masks = prediction.get_field("mask")
+    masks = masker([masks], [prediction])[0]
+    prediction.add_field("mask", masks)
+  return prediction
+
+def compute_prediction_batched(batch, model, size_scale=1.0):
+  imgs = []
+  for img in batch:
+    imgs.append(transforms(size_scale)(img).numpy())
+  image = [Tensor(image, requires_grad=False) for image in imgs]
+  predictions = model(image)
+  del image
+  return predictions
+
+palette = np.array([2 ** 25 - 1, 2 ** 15 - 1, 2 ** 21 - 1])
+
+def findContours(*args, **kwargs):
+  if cv2.__version__.startswith('4'):
+    contours, hierarchy = cv2.findContours(*args, **kwargs)
+  elif cv2.__version__.startswith('3'):
+    _, contours, hierarchy = cv2.findContours(*args, **kwargs)
+  return contours, hierarchy
+
+def compute_colors_for_labels(labels):
+  l = labels[:, None]
+  colors = l * palette
+  colors = (colors % 255).astype("uint8")
+  return colors
+
+def overlay_mask(image, predictions):
+  image = np.asarray(image)
+  masks = predictions.get_field("mask").numpy()
+  labels = predictions.get_field("labels").numpy()
+
+  colors = compute_colors_for_labels(labels).tolist()
+
+  for mask, color in zip(masks, colors):
+    thresh = mask[0, :, :, None]
+    contours, hierarchy = findContours(
+        thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE
+    )
+    image = cv2.drawContours(image, contours, -1, color, 3)
+
+  composite = image
+
+  return composite
+
+CATEGORIES = [
+    "__background", "person", "bicycle", "car", "motorcycle", "airplane", "bus", "train", "truck", "boat", "traffic light",
+    "fire hydrant", "stop sign", "parking meter", "bench", "bird", "cat", "dog", "horse", "sheep", "cow", "elephant",
+    "bear", "zebra", "giraffe", "backpack", "umbrella", "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard",
+    "sports ball", "kite", "baseball bat", "baseball glove", "skateboard", "surfboard", "tennis racket", "bottle",
+    "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple", "sandwich", "orange", "broccoli",
+    "carrot", "hot dog", "pizza", "donut", "cake", "chair", "couch", "potted plant", "bed", "dining table",
+    "toilet", "tv", "laptop", "mouse", "remote", "keyboard", "cell phone", "microwave", "oven", "toaster",
+    "sink", "refrigerator", "book", "clock", "vase", "scissors", "teddy bear", "hair drier", "toothbrush",
+]
+
+def overlay_boxes(image, predictions):
+  labels = predictions.get_field("labels").numpy()
+  boxes = predictions.bbox
+  image = np.asarray(image)
+  colors = compute_colors_for_labels(labels).tolist()
+
+  for box, color in zip(boxes, colors):
+    box = torch.tensor(box.numpy())
+    box = box.to(torch.int64)
+    top_left, bottom_right = box[:2].tolist(), box[2:].tolist()
+    image = cv2.rectangle(
+        image, tuple(top_left), tuple(bottom_right), tuple(color), 1
+    )
+
+  return image
+
+def overlay_class_names(image, predictions):
+  scores = predictions.get_field("scores").numpy().tolist()
+  labels = predictions.get_field("labels").numpy().tolist()
+  labels = [CATEGORIES[int(i)] for i in labels]
+  boxes = predictions.bbox.numpy()
+  image = np.asarray(image)
+  template = "{}: {:.2f}"
+  for box, score, label in zip(boxes, scores, labels):
+    x, y = box[:2]
+    s = template.format(label, score)
+    x, y = int(x), int(y)
+    cv2.putText(
+        image, s, (x, y), cv2.FONT_HERSHEY_SIMPLEX, .5, (255, 255, 255), 1
+    )
+
+  return image
+
+
+if __name__ == '__main__':
+  parser = argparse.ArgumentParser(description='Run MaskRCNN', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+  parser.add_argument('--image', type=str, help="Path of the image to run")
+  parser.add_argument('--threshold', type=float, default=0.7, help="Detector threshold")
+  parser.add_argument('--size_scale', type=float, default=1.0, help="Image resize multiplier")
+  parser.add_argument('--out', type=str, default="/tmp/rendered.png", help="Output filename")
+  args = parser.parse_args()
+
+  resnet = ResNet(50, num_classes=None, stride_in_1x1=True)
+  model_tiny = MaskRCNN(resnet)
+  model_tiny.load_from_pretrained()
+  img = Image.open(args.image)
+  top_result_tiny = compute_prediction(img, model_tiny, confidence_threshold=args.threshold, size_scale=args.size_scale)
+  bbox_image = overlay_boxes(img, top_result_tiny)
+  mask_image = overlay_mask(bbox_image, top_result_tiny)
+  final_image = overlay_class_names(mask_image, top_result_tiny)
+
+  im = Image.fromarray(final_image)
+  print(f"saving {args.out}")
+  im.save(args.out)
+  im.show()
@@ -65,7 +65,17 @@ def loader_process(q_in, q_out, X:Tensor, seed):
      else:
        # pad data with training mean
        img = np.tile(np.array([[[123.68, 116.78, 103.94]]], dtype=np.uint8), (224, 224, 1))
-      X[idx].flatten().assign(img.tobytes())
+
+      # broken out
+      #img_tensor = Tensor(img.tobytes(), device='CPU')
+      #storage_tensor = X[idx].contiguous().realize().lazydata.base.realized
+      #storage_tensor._copyin(img_tensor.numpy())
+
+      # faster
+      X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+
+      # ideal
+      #X[idx].assign(img.tobytes())   # NOTE: this is slow!
      q_out.put(idx)
    q_out.put(None)

@@ -203,13 +213,12 @@ class InterleavedDataset:
    self.queues[queue_index].queue.extend(load_file(file))

 # Reference: https://github.com/mlcommons/training/blob/1c8a098ae3e70962a4f7422c0b0bd35ae639e357/language_model/tensorflow/bert/run_pretraining.py, Line 394
-def batch_load_train_bert(BS:int, seed:int|None=None):
+def batch_load_train_bert(BS:int):
  from extra.datasets.wikipedia import get_wiki_train_files
-  rng = random.Random(seed)
  fs = sorted(get_wiki_train_files())
  train_files = []
  while fs: # TF shuffle
-    rng.shuffle(fs)
+    random.shuffle(fs)
    train_files.append(fs.pop(0))

  cycle_length = min(getenv("NUM_CPU_THREADS", min(os.cpu_count(), 8)), len(train_files))
@@ -254,8 +263,8 @@ def load_unet3d_data(preprocessed_dataset_dir, seed, queue_in, queue_out, X:Tens
      x = random_brightness_augmentation(x)
      x = gaussian_noise(x)

-    X[idx].flatten().assign(x.tobytes())
-    Y[idx].flatten().assign(y.tobytes())
+    X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = x.tobytes()
+    Y[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = y.tobytes()

    queue_out.put(idx)
  queue_out.put(None)
@@ -369,12 +378,12 @@ def load_retinanet_data(base_dir:Path, val:bool, queue_in:Queue, queue_out:Queue
      clipped_match_idxs = np.clip(match_idxs, 0, None)
      clipped_boxes, clipped_labels = tgt["boxes"][clipped_match_idxs], tgt["labels"][clipped_match_idxs]

-      boxes[idx].flatten().assign(clipped_boxes.tobytes())
-      labels[idx].flatten().assign(clipped_labels.tobytes())
-      matches[idx].flatten().assign(match_idxs.tobytes())
-      anchors[idx].flatten().assign(anchor.tobytes())
+      boxes[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_boxes.tobytes()
+      labels[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = clipped_labels.tobytes()
+      matches[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = match_idxs.tobytes()
+      anchors[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = anchor.tobytes()

-    imgs[idx].flatten().assign(img.tobytes())
+    imgs[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()

    queue_out.put(idx)
  queue_out.put(None)
@@ -396,7 +405,6 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
      queue_in.put((idx, img, tgt))

  def _setup_shared_mem(shm_name:str, size:tuple[int, ...], dtype:dtypes) -> tuple[shared_memory.SharedMemory, Tensor]:
-    shm_name = f"{shm_name}_{os.getpid()}"
    if os.path.exists(f"/dev/shm/{shm_name}"): os.unlink(f"/dev/shm/{shm_name}")
    shm = shared_memory.SharedMemory(name=shm_name, create=True, size=prod(size))
    shm_tensor = Tensor.empty(*size, dtype=dtype, device=f"disk:/dev/shm/{shm_name}")
@@ -543,7 +551,7 @@ class BinIdxDataset:
    version, = struct.unpack("<Q", self.idx.read(8))
    assert version == 1, "unsupported index version"
    dtype_code, = struct.unpack("<B", self.idx.read(1))
-    self.dtype = {1:np.dtype(np.uint8), 2:np.dtype(np.int8), 3:np.dtype(np.int16), 4:np.dtype(np.int32), 5:np.dtype(np.int64), 6:np.dtype(np.float64), 7:np.dtype(np.double), 8:np.dtype(np.uint16)}[dtype_code]
+    self.dtype = {1:dtypes.uint8, 2:dtypes.int8, 3:dtypes.int16, 4:dtypes.int32, 5:dtypes.int64, 6:dtypes.float64, 7:dtypes.double, 8:dtypes.uint16}[dtype_code]
    self.count, = struct.unpack("<Q", self.idx.read(8))
    doc_count, = struct.unpack("<Q", self.idx.read(8))

@@ -560,7 +568,7 @@ class BinIdxDataset:
    self.doc_idx = self.idx_t[start:end].bitcast(dtypes.int64).numpy()

    # bin file
-    self.bin_t = Tensor(base_path.with_name(f"{base_path.name}.bin")).numpy()
+    self.bin_t = Tensor(base_path.with_name(f"{base_path.name}.bin"))

  def _index(self, idx) -> tuple[int, int]:
    return int(self.pointers[idx]), int(self.sizes[idx])
@@ -569,7 +577,7 @@ class BinIdxDataset:
    ptr, size = self._index(idx)
    if length is None: length = size - offset
    ptr += offset * self.dtype.itemsize
-    return self.bin_t[ptr:ptr+length*self.dtype.itemsize].view(self.dtype)
+    return self.bin_t[ptr:ptr+length*self.dtype.itemsize].bitcast(self.dtype).to(None)

 # https://docs.nvidia.com/megatron-core/developer-guide/latest/api-guide/datasets.html
 class GPTDataset:
@@ -628,7 +636,7 @@ class GPTDataset:
        sample_parts.append(self.indexed_dataset.get(int(self.doc_idx[i]), offset=int(offset), length=length))

    # concat all parts
-    text = np.concatenate(sample_parts, axis=0)
+    text = Tensor.cat(*sample_parts)

    return text

@@ -755,27 +763,48 @@ class BlendedGPTDataset:

    return dataset_idx, dataset_sample_idx

-def get_llama3_dataset(samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True, small:bool=False) -> BlendedGPTDataset:
-  if small:
-    if val:
-      return BlendedGPTDataset(
-        [base_dir / "c4-validation-91205-samples.en_text_document"], [1.0], samples, seqlen, seed, shuffle=False)
-    return BlendedGPTDataset(
-      [base_dir / "c4-train.en_6_text_document"], [1.0], samples, seqlen, seed, shuffle=True)
+def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
  if val:
-    return BlendedGPTDataset(
-      [base_dir / "validation" / "c4-validationn-91205-samples.en_text_document"], [1.0], samples, seqlen, seed, shuffle=False)
-  return BlendedGPTDataset(
-    [base_dir / "c4-train.en_6_text_document", base_dir / "c4-train.en_7_text_document"], [1.0, 1.0], samples, seqlen, seed, shuffle=True)
+    dataset = BlendedGPTDataset([
+      base_dir / "validation" / "c4-validationn-91205-samples.en_text_document",
+    ], [
+      1.0
+    ], samples, seqlen, seed, False)
+  else:
+    dataset = BlendedGPTDataset([
+      base_dir / "c4-train.en_6_text_document",
+      base_dir / "c4-train.en_7_text_document",
+    ], [
+      1.0, 1.0
+    ], samples, seqlen, seed, True)

-def iterate_llama3_dataset(dataset:BlendedGPTDataset, bs:int):
-  for b in range(math.ceil(dataset.samples / bs)):
-    batch = [dataset.get(b * bs + i) for i in range(bs)]
-    stacked = np.stack(batch, axis=0)
-    yield Tensor(stacked, device="NPY")
+  for b in range(math.ceil(samples / bs)):
+    batch = []
+    for i in range(bs):
+      tokens = dataset.get(b * bs + i)
+      batch.append(tokens)
+    yield Tensor.stack(batch, dim=0)

-def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True, small:bool=False):
-  return iterate_llama3_dataset(get_llama3_dataset(samples, seqlen, base_dir, seed, val, small), bs)
+def batch_load_llama3_small(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
+  if val:
+    dataset = BlendedGPTDataset([
+      base_dir / "c4-validation-91205-samples.en_text_document",
+    ], [
+      1.0
+    ], samples, seqlen, seed, False)
+  else:
+    dataset = BlendedGPTDataset([
+      base_dir / "c4-train.en_6_text_document",
+    ], [
+      1.0
+    ], samples, seqlen, seed, True)
+
+  for b in range(math.ceil(samples / bs)):
+    batch = []
+    for i in range(bs):
+      tokens = dataset.get(b * bs + i)
+      batch.append(tokens)
+    yield Tensor.stack(batch, dim=0)

 if __name__ == "__main__":
  def load_unet3d(val):
@@ -219,28 +219,17 @@ def get_mlperf_bert_model():
  config = get_mlperf_bert_config()
  if getenv("DISABLE_DROPOUT", 0):
    config["hidden_dropout_prob"] = config["attention_probs_dropout_prob"] = 0.0
-  model = BertForPretraining(**config)
-  if getenv("FP8_TRAIN"):
-    from extra.fp8.fp8_linear import convert_to_float8_training
-    def module_filter_fn(mod, fqn):
-      if isinstance(mod, LinearBert):
-        skip_layers = [] if (ln:=config["num_hidden_layers"]) <= 2 else ["bert.encoder.layer.0.", f"bert.encoder.layer.{ln-1}"]
-        if mod.weight.shape[-1] >= 1024 and "encoder" in fqn and not any(name in fqn for name in skip_layers):
-          print(f"replacing linear with fp8: {fqn} {mod.weight.shape}")
-          return True
-      return False
-    convert_to_float8_training(model, module_filter_fn)
-  return model
+  return BertForPretraining(**config)

 def get_fake_data_bert(BS:int):
  return {
-    "input_ids": Tensor.zeros((BS, 512), dtype=dtypes.int32, device="CPU").contiguous(),
-    "input_mask": Tensor.zeros((BS, 512), dtype=dtypes.int32, device="CPU").contiguous(),
-    "segment_ids": Tensor.zeros((BS, 512), dtype=dtypes.int32, device="CPU").contiguous(),
-    "masked_lm_positions": Tensor.zeros((BS, 76), dtype=dtypes.int32, device="CPU").contiguous(),
-    "masked_lm_ids": Tensor.zeros((BS, 76), dtype=dtypes.int32, device="CPU").contiguous(),
-    "masked_lm_weights": Tensor.zeros((BS, 76), dtype=dtypes.float32, device="CPU").contiguous(),
-    "next_sentence_labels": Tensor.zeros((BS, 1), dtype=dtypes.int32, device="CPU").contiguous(),
+    "input_ids": Tensor.empty((BS, 512), dtype=dtypes.int32, device="CPU"),
+    "input_mask": Tensor.empty((BS, 512), dtype=dtypes.int32, device="CPU"),
+    "segment_ids": Tensor.empty((BS, 512), dtype=dtypes.int32, device="CPU"),
+    "masked_lm_positions": Tensor.empty((BS, 76), dtype=dtypes.int32, device="CPU"),
+    "masked_lm_ids": Tensor.empty((BS, 76), dtype=dtypes.int32, device="CPU"),
+    "masked_lm_weights": Tensor.empty((BS, 76), dtype=dtypes.float32, device="CPU"),
+    "next_sentence_labels": Tensor.empty((BS, 1), dtype=dtypes.int32, device="CPU"),
  }

 def find_matches(match_quality_matrix:np.ndarray, high_threshold:float=0.5, low_threshold:float=0.4, allow_low_quality_matches:bool=False) -> np.ndarray:
@@ -59,7 +59,9 @@ class EmbeddingBert(nn.Embedding):
    arange_shp, weight_shp, big_shp = (1, 1, self.vocab_sz, 1), (1, 1, self.vocab_sz, self.embed_sz), idx.shape+(self.vocab_sz, self.embed_sz,)
    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz, requires_grad=False, device=self.weight.device).reshape(arange_shp)
    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1,)).expand(big_shp), self.weight.cast(dtypes.default_float).reshape(weight_shp).expand(big_shp)
-    return (arange == idx).where(vals, 0).sum(2, dtype=vals.dtype)
+    # TODO: contiguous() here because the embedding dropout creates different asts on each device, and search becomes very slow.
+    # Should fix with fixing random ast on multi device, and fuse arange to make embedding fast.
+    return (arange == idx).mul(vals).sum(2, dtype=vals.dtype).contiguous()

 class LayerNormBert:
  def __init__(self, normalized_shape:Union[int, tuple[int, ...]], eps:float=1e-12, elementwise_affine:bool=True):
@@ -204,6 +204,43 @@ def eval_bert():

    st = time.perf_counter()

+def eval_mrcnn():
+  from tqdm import tqdm
+  from extra.models.mask_rcnn import MaskRCNN
+  from extra.models.resnet import ResNet
+  from extra.datasets.coco import BASEDIR, images, convert_prediction_to_coco_bbox, convert_prediction_to_coco_mask, accumulate_predictions_for_coco, evaluate_predictions_on_coco, iterate
+  from examples.mask_rcnn import compute_prediction_batched, Image
+  mdl = MaskRCNN(ResNet(50, num_classes=None, stride_in_1x1=True))
+  mdl.load_from_pretrained()
+
+  bbox_output = '/tmp/results_bbox.json'
+  mask_output = '/tmp/results_mask.json'
+
+  accumulate_predictions_for_coco([], bbox_output, rm=True)
+  accumulate_predictions_for_coco([], mask_output, rm=True)
+
+  #TODO: bs > 1 not as accurate
+  bs = 1
+
+  for batch in tqdm(iterate(images, bs=bs), total=len(images)//bs):
+    batch_imgs = []
+    for image_row in batch:
+      image_name = image_row['file_name']
+      img = Image.open(BASEDIR/f'val2017/{image_name}').convert("RGB")
+      batch_imgs.append(img)
+    batch_result = compute_prediction_batched(batch_imgs, mdl)
+    for image_row, result in zip(batch, batch_result):
+      image_name = image_row['file_name']
+      box_pred = convert_prediction_to_coco_bbox(image_name, result)
+      mask_pred = convert_prediction_to_coco_mask(image_name, result)
+      accumulate_predictions_for_coco(box_pred, bbox_output)
+      accumulate_predictions_for_coco(mask_pred, mask_output)
+    del batch_imgs
+    del batch_result
+
+  evaluate_predictions_on_coco(bbox_output, iou_type='bbox')
+  evaluate_predictions_on_coco(mask_output, iou_type='segm')
+
 def eval_llama3():
  from extra.models.llama import Transformer
  from examples.llama3 import MODEL_PARAMS, load, convert_from_huggingface
@@ -234,9 +271,12 @@ def eval_llama3():
    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
    return loss.flatten().float()

-  from examples.mlperf.dataloader import get_llama3_dataset, iterate_llama3_dataset
-  eval_dataset = get_llama3_dataset(5760, SEQLEN, BASEDIR, val=True, small=bool(SMALL))
-  iter = iterate_llama3_dataset(eval_dataset, BS)
+  if SMALL:
+    from examples.mlperf.dataloader import batch_load_llama3_small
+    iter = batch_load_llama3_small(BS, 5760, SEQLEN, BASEDIR, val=True)
+  else:
+    from examples.mlperf.dataloader import batch_load_llama3
+    iter = batch_load_llama3(BS, 5760, SEQLEN, BASEDIR, val=True)

  losses = []
  for tokens in tqdm(iter, total=5760//BS):
@@ -501,7 +541,7 @@ if __name__ == "__main__":
  # inference only
  Tensor.training = False

-  models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert").split(",")
+  models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(",")
  for m in models:
    nm = f"eval_{m}"
    if nm in globals():
@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing

 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
-from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling, profile_marker, DEBUG
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
 from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW

@@ -13,8 +13,6 @@ from extra.bench_log import BenchEvent, WallTimeEvent
 # TODO: fix benchmark logging and use tinygrad tqdm
 from tqdm import tqdm

-from tinygrad.uop.ops import UOp
-
 def train_resnet():
  from extra.models import resnet
  from examples.mlperf.dataloader import batch_load_resnet
@@ -920,6 +918,40 @@ def train_rnnt():
  # TODO: RNN-T
  pass

+@TinyJit
+def train_step_bert(model, optimizer, scheduler, loss_scaler:float, GPUS, grad_acc:int, **kwargs):
+  optimizer.zero_grad()
+
+  for i in range(grad_acc):
+    input_ids, segment_ids = kwargs[f"input_ids{i}"], kwargs[f"segment_ids{i}"]
+    # NOTE: these two have different names
+    attention_mask, masked_positions = kwargs[f"input_mask{i}"], kwargs[f"masked_lm_positions{i}"]
+    masked_lm_ids, masked_lm_weights, next_sentence_labels = kwargs[f"masked_lm_ids{i}"], kwargs[f"masked_lm_weights{i}"], kwargs[f"next_sentence_labels{i}"]
+
+    for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
+      if len(GPUS) > 1: t.shard_(GPUS, axis=0)
+      else: t.to_(GPUS[0])
+
+    lm_logits, seq_relationship_logits = model(input_ids, attention_mask, masked_positions, segment_ids)
+    loss = model.loss(lm_logits, seq_relationship_logits, masked_lm_ids, masked_lm_weights, next_sentence_labels)
+    (loss * loss_scaler).backward()
+    # TODO: OOM without this realize with large grad_acc
+    Tensor.realize(*[p.grad for p in optimizer.params])
+
+  global_norm = Tensor(0.0, dtype=dtypes.float32, device=optimizer[0].device)
+  for p in optimizer.params:
+    p.grad = p.grad / loss_scaler
+    global_norm += p.grad.float().square().sum()
+  global_norm = global_norm.sqrt().contiguous()
+  for p in optimizer.params:
+    p.grad = (global_norm > 1.0).where((p.grad/global_norm).cast(p.grad.dtype), p.grad)
+
+  optimizer.step()
+  scheduler.step()
+  # TODO: no to("CPU") here because it blocks and messes the python time
+  Tensor.realize(loss, global_norm, optimizer.optimizers[0].lr)
+  return loss, global_norm, optimizer.optimizers[0].lr
+
@TinyJit
 def eval_step_bert(model, input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor, masked_positions:Tensor, masked_lm_ids:Tensor,
                   masked_lm_weights:Tensor, next_sentence_labels:Tensor, GPUS):
@@ -982,8 +1014,7 @@ def train_bert():
  # ** hyperparameters **
  BS                 = config["BS"]                     = getenv("BS", 11 * len(GPUS) if dtypes.default_float in (dtypes.float16, dtypes.bfloat16) else 8 * len(GPUS))
  grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
-  # TODO: implement grad accumulation + mlperf logging
-  assert grad_acc == 1
+  # TODO: mlperf logging
  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 1 * len(GPUS))
  max_lr             = config["OPT_BASE_LEARNING_RATE"] = getenv("OPT_BASE_LEARNING_RATE", 0.000175 * math.sqrt(GBS/96))
@@ -1010,7 +1041,6 @@ def train_bert():
  config["DISABLE_DROPOUT"] = getenv("DISABLE_DROPOUT", 0)
  config["TRAIN_BEAM"]    = TRAIN_BEAM = getenv("TRAIN_BEAM", BEAM.value)
  config["EVAL_BEAM"]     = EVAL_BEAM  = getenv("EVAL_BEAM", BEAM.value)
-  config["FP8_TRAIN"]     = getenv("FP8_TRAIN", 0)

  Tensor.manual_seed(seed)  # seed for weight initialization

@@ -1043,8 +1073,8 @@ def train_bert():

  # ** Optimizer **
  parameters_no_wd = [v for k, v in get_state_dict(model).items() if "bias" in k or "LayerNorm" in k]
-  parameters_wd = [x for x in parameters if x not in set(parameters_no_wd)]
-  optimizer_wd = LAMB(parameters_wd, lr=max_lr, b1=opt_lamb_beta_1, b2=opt_lamb_beta_2, eps=epsilon, weight_decay=decay, adam=False)
+  parameters = [x for x in parameters if x not in set(parameters_no_wd)]
+  optimizer_wd = LAMB(parameters, lr=max_lr, b1=opt_lamb_beta_1, b2=opt_lamb_beta_2, eps=epsilon, weight_decay=decay, adam=False)
  optimizer_no_wd = LAMB(parameters_no_wd, lr=max_lr, b1=opt_lamb_beta_1, b2=opt_lamb_beta_2, eps=epsilon, weight_decay=0.0, adam=False)
  optimizer_group = OptimizerGroup(optimizer_wd, optimizer_no_wd)

@@ -1088,7 +1118,7 @@ def train_bert():
  if RUNMLPERF:
    # only load real data with RUNMLPERF
    eval_it = iter(batch_load_val_bert(EVAL_BS))
-    train_it = iter(tqdm(batch_load_train_bert(BS, seed=seed), total=train_steps, disable=BENCHMARK))
+    train_it = iter(tqdm(batch_load_train_bert(BS), total=train_steps, disable=BENCHMARK))
    for _ in range(start_step): next(train_it) # Fast forward
  else:
    # repeat fake data
@@ -1101,38 +1131,12 @@ def train_bert():
  # ** train loop **
  wc_start = time.perf_counter()

-  i, train_data = start_step, next(train_it)
+  i, train_data = start_step, [next(train_it) for _ in range(grad_acc)]

  if RUNMLPERF:
    if MLLOGGER:
      MLLOGGER.start(key=mllog_constants.EPOCH_START, value=i*GBS, metadata={"epoch_num": i*GBS})

-  @TinyJit
-  def train_step_bert(input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor,
-                      masked_positions:Tensor, masked_lm_ids:Tensor, masked_lm_weights:Tensor, next_sentence_labels:Tensor):
-    for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
-      if len(GPUS) > 1: t.shard_(GPUS, axis=0)
-      else: t.to_(GPUS[0])
-    optimizer_group.zero_grad()
-
-    lm_logits, seq_relationship_logits = model(input_ids, attention_mask, masked_positions, segment_ids)
-    loss = model.loss(lm_logits, seq_relationship_logits, masked_lm_ids, masked_lm_weights, next_sentence_labels)
-    (loss * loss_scaler).backward()
-
-    global_norm = Tensor(0.0, dtype=dtypes.float32, device=optimizer_group[0].device)
-    for p in optimizer_group.params:
-      p.grad = p.grad / loss_scaler
-      global_norm += p.grad.float().square().sum()
-    global_norm = global_norm.sqrt().contiguous()
-    for p in optimizer_group.params:
-      p.grad = (global_norm > 1.0).where((p.grad/global_norm).cast(p.grad.dtype), p.grad)
-
-    optimizer_group.step()
-    scheduler_group.step()
-    # TODO: no to("CPU") here because it blocks and messes the python time
-    Tensor.realize(loss, global_norm, optimizer_group.optimizers[0].lr)
-    return loss, global_norm, optimizer_group.optimizers[0].lr
-
  while train_data is not None and i < train_steps and not achieved:
    if getenv("TRAIN", 1):
      Tensor.training = True
@@ -1140,17 +1144,21 @@ def train_bert():
      st = time.perf_counter()
      GlobalCounters.reset()
      with WallTimeEvent(BenchEvent.STEP):
-        loss, global_norm, lr = train_step_bert(
-          train_data["input_ids"], train_data["segment_ids"], train_data["input_mask"], train_data["masked_lm_positions"], \
-          train_data["masked_lm_ids"], train_data["masked_lm_weights"], train_data["next_sentence_labels"])
+        data = {f"{k}{i}":v for i,d in enumerate(train_data) for k,v in d.items()}
+        loss, global_norm, lr = train_step_bert(model, optimizer_group, scheduler_group, loss_scaler, GPUS, grad_acc, **data)

        pt = time.perf_counter()
-        next_data = next(train_it)
+
+        try:
+          next_data = [next(train_it) for _ in range(grad_acc)]
+        except StopIteration:
+          next_data = None
+
        dt = time.perf_counter()

        device_str = parameters[0].device if isinstance(parameters[0].device, str) else f"{parameters[0].device[0]} * {len(parameters[0].device)}"
        loss = loss.item()
-        if not getenv("FP8_TRAIN"): assert not math.isnan(loss)
+        assert not math.isnan(loss)
        lr = lr.item()

      cl = time.perf_counter()
@@ -1163,7 +1171,7 @@ def train_bert():
      if WANDB:
        wandb.log({"lr": lr, "train/loss": loss, "train/global_norm": global_norm.item(), "train/step_time": cl - st,
                    "train/python_time": pt - st, "train/data_time": dt - pt, "train/cl_time": cl - dt,
-                    "train/mem":GlobalCounters.mem_used / 1e9, "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
+                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})

      train_data, next_data = next_data, None
      i += 1
@@ -1180,8 +1188,8 @@ def train_bert():
      if MLLOGGER and RUNMLPERF:
        MLLOGGER.start(key=mllog_constants.EVAL_START, value=None, metadata={"epoch_num": i*GBS, "step_num": i})
      if getenv("RESET_STEP"): train_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE") and train_step_bert.captured is not None:
-        # TODO: this hangs on tiny green after 90 minutes of training
+      elif getenv("FREE_INTERMEDIATE", 0) and train_step_bert.captured is not None:
+        # TODO: FREE_INTERMEDIATE nan'ed after jit step 2
        train_step_bert.captured.free_intermediates()
      eval_lm_losses = []
      eval_clsf_losses = []
@@ -1216,7 +1224,7 @@ def train_bert():
          return

      if getenv("RESET_STEP"): eval_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE") and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()
+      elif getenv("FREE_INTERMEDIATE", 0) and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()

      del eval_data
      avg_lm_loss = sum(eval_lm_losses) / len(eval_lm_losses)
@@ -1284,12 +1292,9 @@ def train_bert():
        previous_step = i

 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad
+  from extra.models.llama import Transformer
  from examples.llama3 import MODEL_PARAMS
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
-  from examples.mlperf.optim import GradAccClipAdamW
-
-  BENCHMARK = getenv("BENCHMARK")

  config = {}
  BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
@@ -1297,16 +1302,10 @@ def train_llama3():
  grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
  SEED               = config["SEED"]                   = getenv("SEED", 5760)
-  DATA_SEED          = config["DATA_SEED"]              = getenv("DATA_SEED", SEED)
  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
  TRAIN_ON_VAL       = config["TRAIN_ON_VAL"]           = getenv("TRAIN_ON_VAL", 0)
  SMALL              = config["SMALL"]                  = getenv("SMALL", 0)
  SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000 * 1152)
-  EVAL_SAMPLES       = config["EVAL_SAMPLES"]           = getenv("EVAL_SAMPLES", 5760 if not SMALL else 1024)
-  MAX_STEPS          = config["MAX_STEPS"]              = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS))
-  WARMUP_STEPS       = config["WARMUP_STEPS"]           = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
-  LR                 = config["LR"]                     = getenv("LR", 8e-5 * GBS / 1152)
-  END_LR             = config["END_LR"]                 = getenv("END_LR", 8e-7)
  EVAL_FREQ          = config["EVAL_FREQ"]              = getenv("EVAL_FREQ", 46080)
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)
@@ -1320,61 +1319,48 @@ def train_llama3():
  opt_adamw_weight_decay = 0.1

  opt_gradient_clip_norm = 1.0
-  opt_learning_rate_warmup_steps = WARMUP_STEPS
-  opt_learning_rate_decay_steps = MAX_STEPS - opt_learning_rate_warmup_steps
-  opt_base_learning_rate = LR
-  opt_end_learning_rate = END_LR
+  opt_learning_rate_warmup_steps = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
+  opt_learning_rate_decay_steps = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS)) - opt_learning_rate_warmup_steps
+  opt_base_learning_rate = getenv("LR", 8e-5 * GBS / 1152)  # NOTE: cannot change for benchmark
+  opt_end_learning_rate = getenv("END_LR", 8e-7)

-  Tensor.manual_seed(SEED)  # seed for weight initialization
-
-  # ** init wandb **
-  WANDB = getenv("WANDB")
-  if WANDB:
-    import wandb
-    wandb_args = {"id": wandb_id, "resume": "must"} if (wandb_id := getenv("WANDB_RESUME", "")) else {}
-    wandb.init(config=config, **wandb_args, project="MLPerf-LLaMA3")
-
-  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
+  # TODO: confirm weights are in bf16
  # vocab_size from the mixtral tokenizer
-  if not SMALL: model_params |= {"vocab_size": 32000}
-  real_vocab_size = model_params['vocab_size']
-  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
-  print(f"model parameters: {model_params}")
-
-  # pad vocab
-  if (MP := getenv("MP", 1)) > 1: model_params['vocab_size'] = round_up(model_params['vocab_size'], 256 * MP)
-  vocab_mask:Tensor = Tensor.arange(model_params['vocab_size']).reshape(1, 1, -1) >= real_vocab_size
-
-  model = FlatTransformer(**model_params, max_context=SEQLEN)
-
-  params = get_parameters(model)
-  # weights are all bfloat16 for now
-  assert params and all(p.dtype == dtypes.bfloat16 for p in params)
+  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
+  params = params | {"vocab_size": 32000} if not SMALL else params
+  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: params['n_layers'] = llama_layers
+  model = Transformer(**params, max_context=SEQLEN, jit=False, disable_kv_cache=True)

  if getenv("FAKEDATA"):
    for v in get_parameters(model):
      v = v.assign(Tensor.empty(v.shape))

-  is_dp = (DP := getenv("DP", 1)) > 1
-  is_mp = (MP := getenv("MP", 1)) > 1
-  is_sharding = is_dp or is_mp
-  device_count = max(DP, MP)
-  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(device_count))
+  if (DP := getenv("DP", 1)) > 1:
+    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
+    for v in get_parameters(model):
+      v.shard_(device, axis=None)

-  model.shard(device, is_mp)
-
-  if is_dp: vocab_mask.shard_(device, axis=None).realize()
-  if is_mp: vocab_mask.shard_(device, axis=2).realize()
-
-  is_offload_optim = bool(getenv("OFFLOAD_OPTIM"))
-  is_fake_offload = Device.DEFAULT == "NULL"
-  optim_device = ("CPU" if not is_fake_offload else "NULL:99") if is_offload_optim else None
-  optim = GradAccClipAdamW(params, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2,
-                           eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device)
-
-  # init grads
-  grads = [Tensor.zeros_like(p).contiguous() for p in optim.params]
+  if (MP := getenv("MP", 1)) > 1:
+    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
+    for k,v in get_state_dict(model).items():
+      if 'scale' in k: v.shard_(device, axis=None)  # from quantized
+      elif '.attention.wq' in k: v.shard_(device, axis=0)
+      elif '.attention.wk' in k: v.shard_(device, axis=0)
+      elif '.attention.wv' in k: v.shard_(device, axis=0)
+      elif '.attention.wo' in k: v.shard_(device, axis=1)
+      elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
+      elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
+      elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
+      elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
+      elif 'output.weight' in k: v.shard_(device, axis=0)
+      else:
+        # attention_norm, ffn_norm, norm
+        v.shard_(device, axis=None)
+      # prevents memory spike on device 0
+      v.realize()

+  optim = AdamW(get_parameters(model), lr=0.0,
+                b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay)
  scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)

  if resume_ckpt := getenv("RESUME_CKPT"):
@@ -1387,172 +1373,119 @@ def train_llama3():
    load_state_dict(scheduler, safe_load(fn), realize=False)

  @TinyJit
-  def minibatch(tokens:Tensor):
-    if is_dp: tokens = tokens.to(None).shard(device, 0)
-    if is_mp: tokens = tokens.shard(device)
-    if not is_sharding: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1])
-    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
+  @Tensor.train()
+  def train_step(model, tokens:Tensor, grad_acc:int):
+    optim.zero_grad()
+    # grad acc
+    for batch in tokens.split(tokens.shape[0]//grad_acc):
+      if (DP := getenv("DP", 1)) > 1:
+        device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
+        batch = batch.shard(device, 0)
+      if (MP := getenv("MP", 1)) > 1:
+        device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
+        batch = batch.shard(device)
+      logits:Tensor = model(batch[:, :-1], start_pos=0, temperature=math.nan)
+      loss = logits.sparse_categorical_crossentropy(batch[:, 1:])
+      loss.backward()
+      Tensor.realize(*[p.grad for p in optim.params])
+    # L2 norm grad clip
+    # https://github.com/NVIDIA/NeMo/blob/3368c3fc0b4a186ab33a1d68a504315100c0b2a6/nemo/collections/nlp/modules/common/megatron/clip_grads.py#L57
+    # https://docs.pytorch.org/docs/stable/generated/torch.nn.utils.clip_grad_norm_.html
+    if not getenv("DISABLE_GRAD_CLIP_NORM"):
+      total_norm = Tensor(0.0, dtype=dtypes.float32, device=optim.params[0].device)
+      for p in optim.params:
+        total_norm += p.grad.float().square().sum()
+      total_norm = total_norm.sqrt().contiguous()
+      for p in optim.params:
+        p.grad = p.grad * (opt_gradient_clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)

-    for i,(t,g) in enumerate(zip(optim.params, loss.gradient(*optim.params))):
-      grads[i].replace(Tensor(grads[i].uop.after(UOp.group(*apply_grad(grads[i].uop, g.uop))), device=t.device))
-
-    loss_cpu = loss.flatten().float().to("CPU")
-    return loss_cpu.realize(*grads)
-
-  @TinyJit
-  def optim_step():
-    grad_norm = optim.fstep(grads)
+    optim.step()
    scheduler.step()

-    for g in grads: g.assign(g.zeros_like())
-
-    lr_cpu = optim.lr.float().to("CPU")
-    grad_norm_cpu = grad_norm.float().to("CPU")
-    Tensor.realize(lr_cpu, grad_norm_cpu, *grads)
-
-    return lr_cpu, grad_norm_cpu
+    lr = optim.lr
+    loss.realize(lr)
+    return loss, lr

  @TinyJit
  @Tensor.train(False)
-  def eval_step(tokens:Tensor):
-    if is_dp: tokens = tokens.to(None).shard(device, 0)
-    if is_mp: tokens = tokens.shard(device)
-    if not is_sharding: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1])
-    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
-    return loss.flatten().float().to("CPU")
+  def eval_step(model, tokens:Tensor):
+    if (DP := getenv("DP", 1)) > 1:
+      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
+      tokens = tokens.shard(device, 0)
+    if (MP := getenv("MP", 1)) > 1:
+      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
+      tokens = tokens.shard(device)
+    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
+    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
+    return loss.flatten().float()

  # ** data iters **
  def fake_data(bs, samples):
-    import numpy as np
    for _ in range(samples // bs):
-      fake_data_np = np.random.randint(0, model_params["vocab_size"], size=(bs, SEQLEN + 1), dtype=np.int32)
-      yield Tensor(fake_data_np, device="NPY")
+      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)

  def get_train_iter():
    if getenv("FAKEDATA", 0):
-      return fake_data(BS, SAMPLES)
+      return fake_data(GBS, SAMPLES)
    else:
-      from examples.mlperf.dataloader import batch_load_llama3
-      return batch_load_llama3(BS, SAMPLES, SEQLEN, BASEDIR, seed=DATA_SEED, val=bool(TRAIN_ON_VAL), small=bool(SMALL))
-
-  if getenv("FAKEDATA", 0):
-    eval_dataset = None
-  else:
-    from examples.mlperf.dataloader import get_llama3_dataset
-    eval_dataset = get_llama3_dataset(EVAL_SAMPLES, SEQLEN, BASEDIR, val=True, small=bool(SMALL))
+      if SMALL:
+        from examples.mlperf.dataloader import batch_load_llama3_small
+        return batch_load_llama3_small(GBS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
+      else:
+        from examples.mlperf.dataloader import batch_load_llama3
+        return batch_load_llama3(GBS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))

  def get_eval_iter():
-    if eval_dataset is None:
-      return fake_data(EVAL_BS, EVAL_SAMPLES)
-    from examples.mlperf.dataloader import iterate_llama3_dataset
-    return iterate_llama3_dataset(eval_dataset, EVAL_BS)
+    if getenv("FAKEDATA", 0):
+      return fake_data(EVAL_BS, 5760)
+    else:
+      if SMALL:
+        from examples.mlperf.dataloader import batch_load_llama3_small
+        return batch_load_llama3_small(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)
+      else:
+        from examples.mlperf.dataloader import batch_load_llama3
+        return batch_load_llama3(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)

-  num_params = sum(p.numel() for p in params) - model_params["vocab_size"]*model_params["dim"]
-  train_iter = get_train_iter()
+  iter = get_train_iter()
  i, sequences_seen = resume_ckpt, 0
-  step_times = []
-  while i < MAX_STEPS:
+  for tokens in tqdm(iter, total=SAMPLES//GBS):
+    t = time.perf_counter()
    GlobalCounters.reset()
-    actual_gbs = GBS if i >= 2 else BS
-    if getenv("TRAIN", 1):
-      profile_marker(f"train @ {i}")
-      st = time.perf_counter()
+    loss, lr = train_step(model, tokens, grad_acc)
+    loss = loss.float().item()

-      stopped = False
-      losses, data_time, dev_time = [], 0, 0
-      for _ in range(grad_acc if i >= 2 else 1):
-        ist = time.perf_counter()
-        try: tokens = next(train_iter)
-        except StopIteration:
-          stopped = True
-          break
-        mst = time.perf_counter()
-        data_time += mst - ist
-        losses.append(minibatch(tokens).item())
-        dev_time += time.perf_counter() - mst
-      if stopped: break
+    i += 1
+    sequences_seen += tokens.shape[0]

-      gt = time.perf_counter()
-      ret = optim_step()
-      lr, grad_norm = ret[0].item(), ret[1].item()
-      et = time.perf_counter()
+    tqdm.write(f"{loss:.4f} loss, {lr.item():.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, {time.perf_counter()-t:.2f} s")
+    if (fname:=getenv("LOSS_FILE", "")):
+      with open(fname, "a") as f:
+        f.write(f"{i} {loss:.4f} {lr.item():.12f} {GlobalCounters.mem_used / 1e9:.2f}\n")

-      loss = sum(losses) / len(losses)
-      optim_time = et - gt
-      dev_time += optim_time
-      step_time = et - st
-      gbs_time = gt - st
-      if BENCHMARK: step_times.append(step_time)
+    if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
+      tqdm.write("saving checkpoint")
+      if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
+      fn = f"{ckpt_dir}/llama3_{i}.safe"
+      safe_save(get_state_dict(model), fn)

-      i += 1
-      sequences_seen += actual_gbs
+      tqdm.write("saving optim checkpoint")
+      fn = f"{ckpt_dir}/llama3_{i}_optim.safe"
+      safe_save(get_state_dict(scheduler), fn)

-      mem_gb = GlobalCounters.mem_used / 1e9
-      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * max(getenv("DP", 1), getenv("MP", 1)) * 2.3e15)) * 100
-      tqdm.write(
-          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
-          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
-      if DEBUG >= 1: tqdm.write("  mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
-
-      if WANDB:
-        wandb.log({
-          "train/loss": loss,
-          "train/lr": lr,
-          "train/grad_norm": grad_norm,
-          "train/step_time": step_time,
-          "train/gbs_time": gbs_time,
-          "train/optim_time": optim_time,
-          "train/dev_time": dev_time,
-          "train/data_time": data_time,
-          "train/mem": mem_gb,
-          "train/GFLOPS": gflops,
-          "train/MFU": mfu,
-          "train/sequences_seen": sequences_seen
-        })
-
-      if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
-        tqdm.write("saving checkpoint")
-        if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
-        fn = f"{ckpt_dir}/llama3_{i}.safe"
-        safe_save(get_state_dict(model), fn)
-
-        tqdm.write("saving optim checkpoint")
-        fn = f"{ckpt_dir}/llama3_{i}_optim.safe"
-        safe_save(get_state_dict(scheduler), fn)
-
-      if i == BENCHMARK:
-        median_step_time = sorted(step_times)[(BENCHMARK + 1) // 2]
-        estimated_total_minutes = int(median_step_time * (SAMPLES // GBS) / 60)
-        print(f"Estimated training time: {estimated_total_minutes // 60}h{estimated_total_minutes % 60}m")
-        print(f"epoch global_ops: {GlobalCounters.global_ops:_}, "
-              f"epoch global_mem: {GlobalCounters.global_mem:_}")
-
-    if (sequences_seen // EVAL_FREQ != (sequences_seen - actual_gbs) // EVAL_FREQ and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
-      if EVAL_BS == 0: return
+    if sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1):
      tqdm.write(f"evaluating after {sequences_seen} sequences")
-      profile_marker(f"eval @ {i}")

      # run eval
      eval_losses = []
      eval_iter = get_eval_iter()
-      tqdm.write(f"evaluating {EVAL_SAMPLES//EVAL_BS} batches of {EVAL_BS} sequences")
-
-      for j,tokens in tqdm(enumerate(eval_iter), total=EVAL_SAMPLES//EVAL_BS):
-        eval_losses += eval_step(tokens).tolist()
-
-        if BENCHMARK and (j+1) == min(BENCHMARK, EVAL_SAMPLES//EVAL_BS):
-          return
+      tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")

+      for tokens in tqdm(eval_iter, total=5760//EVAL_BS):
+        eval_losses += eval_step(model, tokens).tolist()
      log_perplexity = Tensor(eval_losses).mean().float().item()

      tqdm.write(f"eval log perplexity: {log_perplexity:.4f}")

-      if WANDB:
-        wandb.log({"eval/log_perplexity": log_perplexity, "eval/sequences_seen": sequences_seen})
-
      if log_perplexity < EVAL_TARGET:
        tqdm.write(f"target achieved after {sequences_seen} sequences")
        if getenv("CKPT"):
@@ -1631,7 +1564,7 @@ def train_stable_diffusion():
    loss, out_lr = loss.detach().to("CPU"), optimizer.lr.to("CPU")
    Tensor.realize(loss, out_lr)
    return loss, out_lr
-
+    
  # checkpointing takes ~9 minutes without this, and ~1 minute with this
  @TinyJit
  def ckpt_to_cpu():
@@ -1670,7 +1603,7 @@ def train_stable_diffusion():
    if i == 3:
      for _ in range(3): ckpt_to_cpu() # do this at the beginning of run to prevent OOM surprises when checkpointing
      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-
+      
    total_train_time = time.perf_counter() - train_start_time
    if WANDB:
      wandb.log({"train/loss": loss_item, "train/lr": lr_item, "train/loop_time_prev": loop_time, "train/dl_time": dl_time, "train/step": i,
@@ -1,193 +0,0 @@
-import math, os
-if __name__ == "__main__":
-  os.environ["DEFAULT_FLOAT"] = "bfloat16"
-  os.environ["OPTIM_DTYPE"] = "bfloat16"
-  if "DEV" not in os.environ: os.environ["DEV"] = "NULL"
-  # CDNA
-  os.environ["EMULATE"] = "AMD_CDNA4"
-  os.environ["DEVICE_IN_FUNCTION_BUG"] = "1"
-  os.environ["ALL2ALL"] = "1"
-  os.environ["USE_ATOMICS"] = "1"
-  if "HK_FLASH_ATTENTION" not in os.environ:
-    os.environ["HK_FLASH_ATTENTION"] = "1"
-    if "ASM_GEMM" not in os.environ:
-      os.environ["ASM_GEMM"] = "1"
-from tinygrad import Tensor, nn, function, getenv, dtypes, TinyJit
-from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
-from tinygrad.uop.ops import Ops, UOp
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
-
-FP8 = getenv("FP8", 0)
-
-FP8_DTYPE = dtypes.fp8e4m3
-FP8_MAX = 448.0
-
-def quantize_fp8(x:Tensor):
-  scale = FP8_MAX / (x.abs().max().detach() + 1e-8)
-  x_scaled = x * scale
-  x_clamped = x_scaled + (x_scaled.detach().clamp(-FP8_MAX, FP8_MAX) - x_scaled.detach())  # STE
-  return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal()
-
-def matmul(x:Tensor, w:Tensor) -> Tensor:
-  if not FP8: return x @ w.T
-  # weights are already FP8, just quantize activations
-  x_fp8, x_scale = quantize_fp8(x)
-  return x_fp8.dot(w.T, dtype=dtypes.float) * x_scale
-
-def rmsnorm(x_in:Tensor, eps:float):
-  x = x_in.float()
-  x = x * (x.square().mean(-1, keepdim=True) + eps).rsqrt()
-  return x.cast(x_in.dtype)
-
-class FlatTransformer:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
-               rope_theta:int=10000, max_context:int=1024):
-    self.vocab_size = vocab_size
-    self.n_layers = n_layers
-    self.n_heads = n_heads
-    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
-    self.head_dim = dim // n_heads
-    self.n_rep = self.n_heads // self.n_kv_heads
-
-    # Attention
-    self.wqkv = self.lin_per_layer(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2)
-    self.wo = self.lin_per_layer(self.n_heads * self.head_dim, dim)
-
-    # FeedForward
-    self.w1 = self.lin_per_layer(dim, hidden_dim)
-    self.w2 = self.lin_per_layer(hidden_dim, dim)
-    self.w3 = self.lin_per_layer(dim, hidden_dim)
-
-    self.norm_eps = norm_eps
-    self.attention_norm = Tensor.ones(n_layers, dim).contiguous()
-    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
-
-    # output
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = nn.Embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
-
-  def lin_per_layer(self, in_features:int, out_features:int):
-    bound = 1 / math.sqrt(in_features)
-    dt = FP8_DTYPE if FP8 else None
-    if getenv("ZEROS"): return Tensor.zeros(self.n_layers, out_features, in_features, dtype=dt)
-    return Tensor.uniform(self.n_layers, out_features, in_features, low=-bound, high=bound, dtype=dt)
-
-  def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wqkv:Tensor, wo:Tensor):
-    x = rmsnorm(x, self.norm_eps) * attention_norm
-    xqkv = matmul(x, wqkv)
-
-    bsz, seqlen, _ = xqkv.shape
-    # interleaved layout: each kv group has [n_rep q heads, 1 k head, 1 v head] for clean MP sharding
-    xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
-    xq = xqkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
-    xk = xqkv[:, :, :, self.n_rep].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-    xv = xqkv[:, :, :, self.n_rep+1].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
-    attn = attn.reshape(bsz, seqlen, -1)
-    return matmul(attn, wo)
-
-  def feed_forward(self, x:Tensor, ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
-    x = rmsnorm(x, self.norm_eps) * ffn_norm
-    x_w1 = matmul(x, w1).silu()
-    x_w3 = matmul(x.contiguous_backward(), w3)
-    return matmul(x_w1 * x_w3, w2)
-
-  @function(precompile=True, precompile_backward=True)
-  def run_layer(self, x:Tensor, freqs_cis:Tensor,
-                attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
-    h = x + self.attention(x, freqs_cis, attention_norm, wqkv, wo)
-    return h + self.feed_forward(h, ffn_norm, w1, w2, w3)
-
-  def shard(self, device:tuple[str, ...], mp:bool=False):
-    from tinygrad.nn.state import get_parameters
-    if not mp:
-      for v in get_parameters(self): v.shard_(device, axis=None)
-    else:
-      # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
-      self.wqkv.shard_(device, axis=1).realize()          # (n_layers, out, dim) shard out
-      self.wo.shard_(device, axis=2).realize()             # (n_layers, dim, in) shard in
-      self.w1.shard_(device, axis=1).realize()             # (n_layers, hidden, dim) shard out
-      self.w2.shard_(device, axis=2).realize()             # (n_layers, dim, hidden) shard in
-      self.w3.shard_(device, axis=1).realize()             # (n_layers, hidden, dim) shard out
-      self.attention_norm.shard_(device, axis=None).realize()
-      self.ffn_norm.shard_(device, axis=None).realize()
-      self.norm.weight.shard_(device, axis=None).realize()
-      self.tok_embeddings.weight.shard_(device, axis=0).realize()
-      self.output.weight.shard_(device, axis=0).realize()
-      self.freqs_cis.shard_(device, axis=None).realize()
-
-  def __call__(self, tokens:Tensor):
-    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    for i in range(self.n_layers):
-      h = self.run_layer(h, freqs_cis,
-                         self.attention_norm[i], self.wqkv[i], self.wo[i],
-                         self.ffn_norm[i], self.w1[i], self.w2[i], self.w3[i])
-    logits = self.output(self.norm(h))
-    return logits
-
-# TODO: this shouldn't be needed, but it prevents a copy of the grads. CAT can help
-def apply_grad(old_grad:UOp, new_grad:UOp) -> list[UOp]:
-  if new_grad.op == Ops.ADD:
-    return apply_grad(old_grad, new_grad.src[0])+apply_grad(old_grad, new_grad.src[1])
-  elif new_grad.op == Ops.PAD:
-    grad_shrink = tuple([(p[0], s+p[0]) for s,p in zip(new_grad.src[0].shape, new_grad.marg)])
-    return apply_grad(old_grad.shrink(grad_shrink), new_grad.src[0])
-  else:
-    return [old_grad.store(old_grad + new_grad)]
-
-if __name__ == "__main__":
-  config = {}
-  BS                 = config["BS"]                     = getenv("BS", 16)
-  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
-
-  from examples.llama3 import MODEL_PARAMS
-  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
-  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
-  model = FlatTransformer(**model_params, max_context=SEQLEN)
-  state = nn.state.get_state_dict(model)
-  print("tensor count:", len(state))
-
-  # shard the model
-  from tinygrad import Device
-  if (DP := getenv("DP", 1)) > 1:
-    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)))
-  if (MP := getenv("MP", 1)) > 1:
-    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)), mp=True)
-
-  # preallocate all the grad buffers and zero them out
-  grads = {x:Tensor.zeros_like(x).contiguous() for x in state.values() if x.requires_grad is None}
-
-  # print model size
-  sz = 0
-  for k,v in state.items():
-    print(f"{colored(k, 'green' if v in grads else 'white'):30s} {str(v.shape):30s} {str(v.dtype):20s} {v.device}  {v.nbytes()/1e9:.2f} GB")
-    sz += v.nbytes()
-  print(f"total sz: {sz/1e9:.2f} GB")
-
-  with Timing("fake data: "): tokens = Tensor.randint(BS, SEQLEN+1, low=0, high=model.vocab_size, dtype=dtypes.int)
-  with Timing("realize weights/grads/data: "): Tensor.realize(*state.values(), *grads.values(), tokens)
-  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
-  if DP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)), axis=0)
-  if MP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)))
-
-  @TinyJit
-  def jit_step(tokens:Tensor):
-    with Timing("python forward: "): loss = model(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    with Timing("python backward: "):
-      for t,g in zip(grads, loss.gradient(*grads)):
-        grads[t] = Tensor(grads[t].uop.after(UOp.group(*apply_grad(grads[t].uop, g.uop))), device=t.device)
-    with Timing("run step: "): loss.realize(*grads.values())
-
-  for i in range(6):
-    GlobalCounters.reset()
-    profile_marker(f"step {i}")
-    with Timing(colored(f"*** step {i}: ", "red")):
-      jit_step(tokens)
-  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
@@ -1,80 +0,0 @@
-from tinygrad import Tensor, nn
-from tinygrad.helpers import getenv
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
-
-class Attention:
-  def __init__(self, dim:int, n_heads:int, n_kv_heads:int|None=None, linear=nn.Linear):
-    self.n_heads = n_heads
-    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
-    self.head_dim = dim // n_heads
-    self.n_rep = self.n_heads // self.n_kv_heads
-
-    if getenv("WQKV"):
-      self.wqkv = linear(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2, bias=False)
-    else:
-      self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
-      self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-      self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-
-    self.wo = linear(self.n_heads * self.head_dim, dim, bias=False)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor) -> Tensor:
-    if getenv("WQKV"):
-      xqkv = self.wqkv(x)
-      xqkv = xqkv.reshape(xqkv.shape[0], xqkv.shape[1], self.n_kv_heads, self.n_rep + 2, self.head_dim)
-      xq = xqkv[:, :, :, :self.n_rep].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xk = xqkv[:, :, :, self.n_rep:self.n_rep+1].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xv = xqkv[:, :, :, self.n_rep+1:self.n_rep+2].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-    else:
-      xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
-
-    xq = xq.reshape(xq.shape[0], xq.shape[1], self.n_heads, self.head_dim)
-    xk = xk.reshape(xk.shape[0], xk.shape[1], self.n_kv_heads, self.head_dim)
-    xv = xv.reshape(xv.shape[0], xv.shape[1], self.n_kv_heads, self.head_dim)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    bsz, seqlen, _, _ = xq.shape
-
-    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
-
-    attn = attn.reshape(bsz, seqlen, -1)
-    return self.wo(attn)
-
-class FeedForward:
-  def __init__(self, dim:int, hidden_dim:int, linear=nn.Linear):
-    self.w1 = linear(dim, hidden_dim, bias=False)
-    self.w2 = linear(hidden_dim, dim, bias=False)
-    self.w3 = linear(dim, hidden_dim, bias=False) # the gate in Gated Linear Unit
-
-  def __call__(self, x:Tensor) -> Tensor:
-    w1 = self.w1(x).silu()
-    w3 = self.w3(x)
-    return self.w2(w1 * w3)
-
-class TransformerBlock:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int|None, norm_eps:float, linear=nn.Linear):
-    self.attention = Attention(dim, n_heads, n_kv_heads, linear)
-    self.feed_forward = FeedForward(dim, hidden_dim, linear)
-    self.attention_norm = nn.RMSNorm(dim, norm_eps)
-    self.ffn_norm = nn.RMSNorm(dim, norm_eps)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor):
-    h = x + self.attention(self.attention_norm(x), freqs_cis)
-    return h + self.feed_forward(self.ffn_norm(h))
-
-class Transformer:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
-               rope_theta:int=10000, max_context:int=1024, linear=nn.Linear, embedding=nn.Embedding):
-    self.layers = [TransformerBlock(dim, hidden_dim, n_heads, n_kv_heads, norm_eps, linear) for _ in range(n_layers)]
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False) if embedding == nn.Embedding else linear(dim, vocab_size, bias=False)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
-
-  def __call__(self, tokens:Tensor):
-    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    for layer in self.layers: h = layer(h, freqs_cis)
-    logits = self.output(self.norm(h))
-    return logits
@@ -1,140 +0,0 @@
-import os
-os.environ["WQKV"] = "1"
-import unittest
-import numpy as np
-from tinygrad import Tensor, nn, dtypes
-from tinygrad.nn.state import get_parameters
-from tinygrad.device import is_dtype_supported
-from examples.mlperf.models.llama import Transformer
-from examples.mlperf.models.flat_llama import FlatTransformer
-
-def copy_weights(flat:FlatTransformer, ref:Transformer):
-  n_layers = flat.n_layers
-  Tensor.realize(*nn.state.get_state_dict(ref).values())
-  flat.wqkv.assign(Tensor(np.stack([ref.layers[i].attention.wqkv.weight.numpy() for i in range(n_layers)])))
-  flat.wo.assign(Tensor(np.stack([ref.layers[i].attention.wo.weight.numpy() for i in range(n_layers)])))
-  flat.w1.assign(Tensor(np.stack([ref.layers[i].feed_forward.w1.weight.numpy() for i in range(n_layers)])))
-  flat.w2.assign(Tensor(np.stack([ref.layers[i].feed_forward.w2.weight.numpy() for i in range(n_layers)])))
-  flat.w3.assign(Tensor(np.stack([ref.layers[i].feed_forward.w3.weight.numpy() for i in range(n_layers)])))
-  flat.attention_norm.assign(Tensor(np.stack([ref.layers[i].attention_norm.weight.numpy() for i in range(n_layers)])))
-  flat.ffn_norm.assign(Tensor(np.stack([ref.layers[i].ffn_norm.weight.numpy() for i in range(n_layers)])))
-  flat.norm.weight.assign(Tensor(ref.norm.weight.numpy()))
-  flat.tok_embeddings.weight.assign(Tensor(ref.tok_embeddings.weight.numpy()))
-  flat.output.weight.assign(Tensor(ref.output.weight.numpy()))
-
-class TestFlatLlama(unittest.TestCase):
-  def test_forward_match(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-    tokens = Tensor([[1, 50, 100, 999, 2]])
-    ref_logits = ref(tokens).realize()
-    flat_logits = flat(tokens).realize()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    diff = (ref_logits - flat_logits).abs().max().item()
-    self.assertLess(diff, 1e-5, f"forward mismatch: max abs diff {diff}")
-
-  def test_backward_match(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-
-    for p in get_parameters(ref): p.requires_grad_(True)
-    for p in get_parameters(flat): p.requires_grad_(True)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-    tokens = Tensor([[1, 50, 100, 999, 2, 10]])
-
-    ref_loss = ref(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    ref_loss.backward()
-    ref_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(ref).items() if v.grad is not None}
-
-    flat_loss = flat(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    flat_loss.backward()
-    flat_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(flat).items() if v.grad is not None}
-
-    # check loss matches
-    self.assertAlmostEqual(ref_loss.item(), flat_loss.item(), places=4)
-
-    # check output weight grad matches
-    diff = abs(ref_grads["output.weight"] - flat_grads["output.weight"]).max()
-    self.assertLess(diff, 1e-4, f"output.weight grad mismatch: max abs diff {diff}")
-
-    # check per-layer weight grads match
-    for i in range(params["n_layers"]):
-      for flat_key, ref_key in [
-        ("wqkv", f"layers.{i}.attention.wqkv.weight"),
-        ("wo", f"layers.{i}.attention.wo.weight"),
-        ("w1", f"layers.{i}.feed_forward.w1.weight"),
-        ("w2", f"layers.{i}.feed_forward.w2.weight"),
-        ("w3", f"layers.{i}.feed_forward.w3.weight"),
-      ]:
-        diff = abs(ref_grads[ref_key] - flat_grads[flat_key][i]).max()
-        self.assertLess(diff, 1e-4, f"layer {i} {flat_key} grad mismatch: max abs diff {diff}")
-
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
-  def test_forward_match_mp(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    from tinygrad import Device
-    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-    flat.shard(devices, mp=True)
-
-    tokens = Tensor([[1, 50, 100, 999, 2]], device=devices[0])
-    ref_logits = ref(tokens.to(devices[0])).numpy()
-    flat_logits = flat(tokens.shard(devices)).numpy()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
-
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
-  def test_forward_match_dp(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    from tinygrad import Device
-    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-    flat.shard(devices)
-
-    tokens = Tensor([[1, 50, 100, 999, 2], [2, 100, 50, 1, 999]], device=devices[0])
-    ref_logits = ref(tokens.to(devices[0])).numpy()
-    flat_logits = flat(tokens.shard(devices, axis=0)).numpy()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
-
-  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3), "fp8 not supported on this device")
-  def test_forward_fp8(self):
-    import examples.mlperf.models.flat_llama as flat_llama_mod
-    old_fp8 = flat_llama_mod.FP8
-    try:
-      flat_llama_mod.FP8 = 1
-      Tensor.manual_seed(42)
-      params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-      ref = Transformer(**params)
-      flat = FlatTransformer(**params)
-      copy_weights(flat, ref)
-      Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-      tokens = Tensor([[1, 50, 100, 999, 2]])
-      ref_logits = ref(tokens).numpy()
-      flat_logits = flat(tokens).numpy()
-      self.assertEqual(ref_logits.shape, flat_logits.shape)
-      # FP8 has lower precision, allow larger tolerance
-      np.testing.assert_allclose(flat_logits, ref_logits, atol=1.0, rtol=0.1)
-    finally:
-      flat_llama_mod.FP8 = old_fp8
-
-if __name__ == "__main__":
-  unittest.main()
@@ -1,59 +0,0 @@
-from tinygrad.tensor import Tensor
-from tinygrad.dtype import dtypes
-from tinygrad.nn.optim import Optimizer
-from tinygrad.helpers import FUSE_OPTIM
-
-class GradAccClipAdamW(Optimizer):
-  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, device=None, fused=FUSE_OPTIM):
-    super().__init__(params, lr, device, fused)
-    self.b1, self.b2, self.eps, self.wd = b1, b2, eps, weight_decay
-    self.b1_t, self.b2_t = (Tensor.ones((1,), dtype=dtypes.float32, device=self.device, requires_grad=False) for _ in [b1, b2])
-    self.m = self._new_optim_param()
-    self.v = self._new_optim_param()
-    self.grad_acc, self.clip_norm = grad_acc, clip_norm
-
-  def fstep(self, grads:list[Tensor]):
-    if self.fused:
-      out, extra = self._step([], grads)
-      updates = [out[0][self.pos_params[i]:self.pos_params[i+1]].reshape(tt.shape) for i, tt in enumerate(self.params)]
-    else:
-      updates, extra = self._step([], grads)
-    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i]))
-    to_realize = extra+self.params+self.buffers
-
-    Tensor.realize(*to_realize)
-    return extra[-1]
-
-  def _step(self, params:list[Tensor], grads:list[Tensor]) -> tuple[list[Tensor], list[Tensor]]:
-    grads = list(grads)
-
-    for i in range(len(grads)):
-      if grads[i].device != self.m[i].device: grads[i] = grads[i].to(self.m[i].device)
-
-    if self.fused:
-      grads[0].assign(grads[0] / self.grad_acc)
-      total_norm = grads[0].float().square().sum().sqrt()
-      grads[0].assign((grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype))
-    else:
-      for i in range(len(grads)):
-        grads[i].assign(grads[i] / self.grad_acc)
-      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
-      for i in range(len(grads)):
-        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype))
-
-    ret = []
-    self.b1_t *= self.b1
-    self.b2_t *= self.b2
-    for i, g in enumerate(grads):
-      self.m[i].assign((self.b1 * self.m[i] + (1.0 - self.b1) * g).cast(self.m[i].dtype))
-      self.v[i].assign((self.b2 * self.v[i] + (1.0 - self.b2) * (g * g)).cast(self.v[i].dtype))
-      m_hat = (self.m[i] / (1.0 - self.b1_t)).cast(self.m[i].dtype)
-      v_hat = (self.v[i] / (1.0 - self.b2_t)).cast(self.v[i].dtype)
-      up = m_hat / (v_hat.sqrt() + self.eps)
-      ret.append((self.lr * up).cast(g.dtype))
-    return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]
-
-  def _apply_update(self, t:Tensor, up:Tensor) -> Tensor:
-    wd = self.wd if t.ndim >= 3 else 0.0
-    up = up.shard_like(t) + self.lr.to(t.device) * wd * t.detach()
-    return t.detach() - up.cast(t.dtype)
@@ -4,7 +4,7 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128

-export CHECK_OOB=0
+export IGNORE_OOB=1

 export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1
@@ -5,7 +5,7 @@ export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -8,7 +8,7 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -11,7 +11,7 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -1,20 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1 DEBUG=0 JIT=1 FLASH_ATTENTION=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=0 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,24 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-export BEAM_TIMEOUT_SEC=15
-export FP8_TRAIN=1
-# search
-IGNORE_BEAM_CACHE=1 BENCHMARK=10 BERT_LAYERS=2 RUNMLPERF=0 python3 examples/mlperf/model_train.py
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_8xMI350X"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_8xMI350x_${DATETIME}_${SEED}.log"
-
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -2,9 +2,9 @@

 export PYTHONPATH="." NV=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -2,9 +2,9 @@

 export PYTHONPATH="." NV=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -5,9 +5,9 @@ set -o pipefail  # Make pipeline fail if any command fails
 export PYTHONPATH="." NV=1
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -2,9 +2,9 @@

 export PYTHONPATH="." AMD=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -2,9 +2,9 @@

 export PYTHONPATH="." AMD=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -5,9 +5,9 @@ set -o pipefail  # Make pipeline fail if any command fails
 export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_red"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=500000

 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -15,7 +15,7 @@ export IGNORE_JIT_FIRST_BEAM=1
 export BASEDIR="/raid/datasets/wiki"

 # pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
+export LOGMLPERF=1

 export SEED=$RANDOM
 DATETIME=$(date "+%m%d%H%M")
@@ -1,38 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,32 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1152}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -1,42 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,43 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,37 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-4}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -1,38 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-32}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -1,5 +0,0 @@
-#!/bin/bash
-export BENCHMARK=5
-export EVAL_BS=0
-VIZ=${VIZ:--1} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
-extra/viz/cli.py --profile --device "AMD" --limit 20
@@ -1,10 +0,0 @@
-#!/bin/bash
-export BENCHMARK=5
-export EVAL_BS=0
-export FAKEDATA=1
-export NULL_ALLOW_COPYOUT=1
-export HIP_VISIBLE_DEVICES=""
-export DEV=NULL
-export JITBEAM=0
-export LLAMA_LAYERS=${LLAMA_LAYERS:-"2"}
-time examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
@@ -1,11 +1,12 @@
 from pathlib import Path
+import numpy as np
 import torch
 from torchvision.utils import make_grid, save_image
 from tinygrad.nn.state import get_parameters
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import trange
 from tinygrad.nn import optim
-from tinygrad.nn.datasets import mnist
+from extra.datasets import fetch_mnist

 class LinearGen:
  def __init__(self):
@@ -37,14 +38,14 @@ class LinearDisc:
    return x

 def make_batch(images):
-  sample = Tensor.randint(batch_size, low=0, high=images.shape[0])
-  return images[sample].reshape(batch_size, 28*28).cast('float').div(127.5).sub(1.0)
+  sample = np.random.randint(0, len(images), size=(batch_size))
+  image_b = images[sample].reshape(-1, 28*28).astype(np.float32) / 127.5 - 1.0
+  return Tensor(image_b)

 def make_labels(bs, col, val=-2.0):
-  y = Tensor.zeros(bs, 2)
-  if col == 0: y = y + Tensor([val, 0.0])
-  else: y = y + Tensor([0.0, val])
-  return y
+  y = np.zeros((bs, 2), np.float32)
+  y[range(bs), [col] * bs] = val  # Can we do label smoothing? i.e -2.0 changed to -1.98789.
+  return Tensor(y)

 def train_discriminator(optimizer, data_real, data_fake):
  real_labels = make_labels(batch_size, 1)
@@ -70,12 +71,12 @@ def train_generator(optimizer, data_fake):

 if __name__ == "__main__":
  # data for training and validation
-  X_train, _, _, _ = mnist()
+  images_real = np.vstack(fetch_mnist()[::2])
  ds_noise = Tensor.randn(64, 128, requires_grad=False)
  # parameters
  epochs, batch_size, k = 300, 512, 1
  sample_interval = epochs // 10
-  n_steps = X_train.shape[0] // batch_size
+  n_steps = len(images_real) // batch_size
  # models and optimizer
  generator = LinearGen()
  discriminator = LinearDisc()
@@ -83,24 +84,24 @@ if __name__ == "__main__":
  output_dir = Path(".").resolve() / "outputs"
  output_dir.mkdir(exist_ok=True)
  # optimizers
-  optim_g = optim.Adam(get_parameters(generator), lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
-  optim_d = optim.Adam(get_parameters(discriminator), lr=0.0002, b1=0.5)
+  optim_g = optim.Adam(get_parameters(generator),lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
+  optim_d = optim.Adam(get_parameters(discriminator),lr=0.0002, b1=0.5)
  # training loop
-  with Tensor.train():
-    for epoch in (t := trange(epochs)):
-      loss_g, loss_d = 0.0, 0.0
-      for _ in range(n_steps):
-        data_real = make_batch(X_train)
-        for step in range(k):  # Try with k = 5 or 7.
-          noise = Tensor.randn(batch_size, 128)
-          data_fake = generator.forward(noise).detach()
-          loss_d += train_discriminator(optim_d, data_real, data_fake)
+  Tensor.training = True
+  for epoch in (t := trange(epochs)):
+    loss_g, loss_d = 0.0, 0.0
+    for _ in range(n_steps):
+      data_real = make_batch(images_real)
+      for step in range(k):  # Try with k = 5 or 7.
        noise = Tensor.randn(batch_size, 128)
-        data_fake = generator.forward(noise)
-        loss_g += train_generator(optim_g, data_fake)
-      if (epoch + 1) % sample_interval == 0:
-        fake_images = generator.forward(ds_noise).detach().numpy()
-        fake_images = (fake_images.reshape(-1, 1, 28, 28) + 1) / 2  # 0 - 1 range.
-        save_image(make_grid(torch.tensor(fake_images)), output_dir / f"image_{epoch+1}.jpg")
-      t.set_description(f"Generator loss: {loss_g/n_steps}, Discriminator loss: {loss_d/n_steps}")
+        data_fake = generator.forward(noise).detach()
+        loss_d += train_discriminator(optim_d, data_real, data_fake)
+      noise = Tensor.randn(batch_size, 128)
+      data_fake = generator.forward(noise)
+      loss_g += train_generator(optim_g, data_fake)
+    if (epoch + 1) % sample_interval == 0:
+      fake_images = generator.forward(ds_noise).detach().numpy()
+      fake_images = (fake_images.reshape(-1, 1, 28, 28) + 1) / 2  # 0 - 1 range.
+      save_image(make_grid(torch.tensor(fake_images)), output_dir / f"image_{epoch+1}.jpg")
+    t.set_description(f"Generator loss: {loss_g/n_steps}, Discriminator loss: {loss_d/n_steps}")
  print("Training Completed!")
@@ -0,0 +1,118 @@
+import json, pprint
+from tinygrad import fetch, nn, Tensor
+from tinygrad.helpers import DEBUG
+
+class FeedForward:
+  def __init__(self, model_dim, intermediate_dim):
+    self.proj_1 = nn.Linear(model_dim, 2*intermediate_dim, bias=False)
+    self.proj_2 = nn.Linear(intermediate_dim, model_dim, bias=False)
+
+  def __call__(self, x):
+    y_12 = self.proj_1(x)
+    y_1, y_2 = y_12.chunk(2, dim=-1)
+    return self.proj_2(y_1.silu() * y_2)
+
+# NOTE: this RoPE doesn't match LLaMA's?
+def _rotate_half(x: Tensor) -> Tensor:
+  x1, x2 = x.chunk(2, dim=-1)
+  return Tensor.cat(-x2, x1, dim=-1)
+
+def _apply_rotary_pos_emb(x: Tensor, pos_sin: Tensor, pos_cos: Tensor) -> Tensor:
+  return (x * pos_cos) + (_rotate_half(x) * pos_sin)
+
+class Attention:
+  def __init__(self, model_dim, num_query_heads, num_kv_heads, head_dim):
+    self.qkv_proj = nn.Linear(model_dim, (num_query_heads + num_kv_heads*2) * head_dim, bias=False)
+    self.num_query_heads, self.num_kv_heads = num_query_heads, num_kv_heads
+    self.head_dim = head_dim
+    self.q_norm = nn.RMSNorm(head_dim)
+    self.k_norm = nn.RMSNorm(head_dim)
+    self.out_proj = nn.Linear(num_query_heads * head_dim, model_dim, bias=False)
+
+  def __call__(self, x:Tensor) -> Tensor:
+    batch_size, seq_len, embed_dim = x.shape
+    qkv = self.qkv_proj(x)
+    qkv = qkv.reshape(batch_size, seq_len, self.num_query_heads+self.num_kv_heads*2, self.head_dim).transpose(1, 2)
+    xq,xk,xv = qkv.split([self.num_query_heads, self.num_kv_heads, self.num_kv_heads], dim=1)
+    xq = self.q_norm(xq)
+    xk = self.k_norm(xk)
+
+    # add positional embedding (how many kernels is this?)
+    freq_constant = 10000
+    inv_freq = 1.0 / (freq_constant ** (Tensor.arange(0, self.head_dim, 2) / self.head_dim))
+    pos_index_theta = Tensor.einsum("i,j->ij", Tensor.arange(seq_len), inv_freq)
+    emb = Tensor.cat(pos_index_theta, pos_index_theta, dim=-1)
+    cos_emb, sin_emb = emb.cos()[None, None, :, :], emb.sin()[None, None, :, :]
+    xq = _apply_rotary_pos_emb(xq, sin_emb, cos_emb)
+    xk = _apply_rotary_pos_emb(xk, sin_emb, cos_emb)
+
+    # grouped-query attention
+    num_groups = self.num_query_heads // self.num_kv_heads
+    xk = xk.repeat_interleave(num_groups, dim=1)
+    xv = xv.repeat_interleave(num_groups, dim=1)
+
+    # masked attention
+    #start_pos = 0
+    #mask = Tensor.full((1, 1, seq_len, start_pos+seq_len), float("-inf"), dtype=xq.dtype, device=xq.device).triu(start_pos+1)
+    #attn_output = xq.scaled_dot_product_attention(xk, xv, mask).transpose(1, 2)
+
+    # causal is fine, no mask needed
+    attn_output = xq.scaled_dot_product_attention(xk, xv, is_causal=True).transpose(1, 2)
+    return self.out_proj(attn_output.reshape(batch_size, seq_len, self.num_query_heads * self.head_dim))
+
+class Layer:
+  def __init__(self, model_dim, intermediate_dim, num_query_heads, num_kv_heads, head_dim):
+    self.ffn = FeedForward(model_dim, intermediate_dim)
+    self.attn = Attention(model_dim, num_query_heads, num_kv_heads, head_dim)
+    self.ffn_norm = nn.RMSNorm(model_dim)
+    self.attn_norm = nn.RMSNorm(model_dim)
+
+  def __call__(self, x:Tensor) -> Tensor: # (batch, seq_len, embed_dim)
+    x = x + self.attn(self.attn_norm(x))
+    x = x + self.ffn(self.ffn_norm(x))
+    return x
+
+# stupidly complex
+def make_divisible(v, divisor):
+  new_v = max(divisor, int(v + divisor / 2) // divisor * divisor)
+  if new_v < 0.9 * v: new_v += divisor
+  return new_v
+
+class Transformer:
+  def __init__(self, cfg):
+    if DEBUG >= 3: pprint.pp(cfg)
+    self.layers = [Layer(cfg['model_dim'], make_divisible(int(cfg["model_dim"] * cfg['ffn_multipliers'][i]), cfg['ffn_dim_divisor']),
+                         cfg['num_query_heads'][i], cfg['num_kv_heads'][i], cfg['head_dim']) for i in range(cfg['num_transformer_layers'])]
+    self.norm = nn.RMSNorm(cfg['model_dim'])
+    self.token_embeddings = nn.Embedding(cfg['vocab_size'], cfg['model_dim'])
+
+  def __call__(self, tokens:Tensor):
+    # _bsz, seqlen = tokens.shape
+    x = self.token_embeddings(tokens)
+    for l in self.layers: x = l(x)
+    return self.norm(x) @ self.token_embeddings.weight.T
+
+if __name__ == "__main__":
+  #model_name = "OpenELM-270M-Instruct"
+  model_name = "OpenELM-270M"  # this is fp32
+  model = Transformer(json.loads(fetch(f"https://huggingface.co/apple/{model_name}/resolve/main/config.json?download=true").read_bytes()))
+  weights = nn.state.safe_load(fetch(f"https://huggingface.co/apple/{model_name}/resolve/main/model.safetensors?download=true"))
+  if DEBUG >= 3:
+    for k, v in weights.items(): print(k, v.shape)
+  nn.state.load_state_dict(model, {k.removeprefix("transformer."):v for k,v in weights.items()})
+
+  from sentencepiece import SentencePieceProcessor
+  tokenizer = SentencePieceProcessor(fetch("https://github.com/karpathy/llama2.c/raw/master/tokenizer.model").as_posix())
+  toks = [tokenizer.bos_id()] + tokenizer.encode("Some car brands include")
+  for i in range(100):
+    ttoks = Tensor([toks])
+    out = model(ttoks).realize()
+    t0 = out[0].argmax(axis=-1).tolist()
+    toks.append(t0[-1])
+    # hmmm...passthrough still doesn't match (it shouldn't, it outputs the most likely)
+    print(tokenizer.decode(toks))
+    #print(toks)
+    #print(tokenizer.decode(t0))
+    #print(t0)
+
+
@@ -1,10 +1,15 @@
 import os, sys, pickle, time, re
 import numpy as np
+if "FLOAT16" not in os.environ: os.environ["FLOAT16"] = "1"
+if "IMAGE" not in os.environ: os.environ["IMAGE"] = "2"
+if "NOLOCALS" not in os.environ: os.environ["NOLOCALS"] = "1"
 if "JIT_BATCH_SIZE" not in os.environ: os.environ["JIT_BATCH_SIZE"] = "0"

 from tinygrad import fetch, Tensor, TinyJit, Context, GlobalCounters, Device, dtypes
 from tinygrad.helpers import DEBUG, getenv
 from tinygrad.engine.realize import CompiledRunner
+
+import onnx
 from tinygrad.nn.onnx import OnnxRunner

 OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
@@ -16,14 +21,11 @@ def compile(onnx_file):

  input_shapes = {name: spec.shape for name, spec in run_onnx.graph_inputs.items()}
  input_types = {name: spec.dtype for name, spec in run_onnx.graph_inputs.items()}
-
  # Float inputs and outputs to tinyjits for openpilot are always float32
-  # TODO this seems dumb
  input_types = {k:(dtypes.float32 if v is dtypes.float16 else v) for k,v in input_types.items()}
  Tensor.manual_seed(100)
-  inputs = {k:Tensor(Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize().numpy(), device='NPY') for k,shp in sorted(input_shapes.items())}
-  if not getenv("NPY_IMG"):
-    inputs = {k:Tensor(v.numpy(), device=Device.DEFAULT).realize() if 'img' in k else v for k,v in inputs.items()}
+  new_inputs = {k:Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize() for k,shp in sorted(input_shapes.items())}
+  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
  print("created tensors")

  run_onnx_jit = TinyJit(lambda **kwargs:
@@ -31,7 +33,9 @@ def compile(onnx_file):
  for i in range(3):
    GlobalCounters.reset()
    print(f"run {i}")
-    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1), OPENPILOT_HACKS=1):
+    inputs = {**{k:v.clone() for k,v in new_inputs.items() if 'img' in k},
+              **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}
+    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1)):
      ret = run_onnx_jit(**inputs).numpy()
    # copy i == 1 so use of JITBEAM is okay
    if i == 1: test_val = np.copy(ret)
@@ -39,7 +43,7 @@ def compile(onnx_file):
  np.testing.assert_equal(test_val, ret, "JIT run failed")
  print("jit run validated")

-  # check gated read_image usage
+  # checks from compile2
  kernel_count = 0
  read_image_count = 0
  gated_read_image_count = 0
@@ -65,9 +69,14 @@ def compile(onnx_file):
  print(f"mdl size is {mdl_sz/1e6:.2f}M")
  print(f"pkl size is {pkl_sz/1e6:.2f}M")
  print("**** compile done ****")
-  return inputs, test_val
+  return test_val

-def test_vs_compile(run, inputs, test_val=None):
+def test_vs_compile(run, new_inputs, test_val=None):
+  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
+
+  # create fake "from_blob" tensors for the inputs, and wrapped NPY tensors for the numpy inputs (these have the same underlying memory)
+  inputs = {**{k:v for k,v in new_inputs.items() if 'img' in k},
+            **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}

  # run 20 times
  step_times = []
@@ -84,58 +93,68 @@ def test_vs_compile(run, inputs, test_val=None):
    min_time = min(step_times)
    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"

+  print(out, val.shape, val.dtype)
  if test_val is not None: np.testing.assert_equal(test_val, val)
  print("**** test done ****")

  # test that changing the numpy changes the model outputs
-  inputs_2x = {k: Tensor(v.numpy()*2, device=v.device) for k,v in inputs.items()}
-  out = run(**inputs_2x)
-  changed_val = out.numpy()
-  np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, val, changed_val)
+  if any([x.device == 'NPY' for x in inputs.values()]):
+    for v in new_inputs_numpy.values(): v *= 2
+    out = run(**inputs)
+    changed_val = out.numpy()
+    np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, val, changed_val)
  return val

-def test_vs_onnx(new_inputs, test_val, onnx_file, tol):
-  import onnx
-  import onnxruntime as ort
-  
-  onnx_inputs = {k:v.numpy() for k,v in new_inputs.items()}
+def test_vs_onnx(new_inputs, test_val, onnx_file, ort=False):
+  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
  onnx_model = onnx.load(onnx_file)

-  ORT_TO_NP_DTYPES: dict[str, np.dtype] = {
-    'tensor(float)': np.dtype('float32'),
-    'tensor(float16)': np.dtype('float16'),
-    'tensor(uint8)': np.dtype('uint8'),
-  }
-
  timings = []
-  onnx_session = ort.InferenceSession(onnx_file)
-  onnx_types = {x.name: ORT_TO_NP_DTYPES[x.type] for x in onnx_session.get_inputs()}
-  onnx_inputs = {k:onnx_inputs[k].astype(onnx_types[k]) for k in onnx_inputs}
+  if ort:
+    # test with onnxruntime
+    import onnxruntime as ort
+    onnx_session = ort.InferenceSession(onnx_file)
+    for _ in range(1 if test_val is not None else 5):
+      st = time.perf_counter()
+      onnx_output = onnx_session.run([onnx_model.graph.output[0].name], {k:v.astype(np.float16) for k,v in new_inputs_numpy.items()})
+      timings.append(time.perf_counter() - st)
+    new_torch_out = onnx_output[0]
+  else:
+    # test with torch
+    import torch
+    from onnx2torch import convert
+    inputs = {k.name:new_inputs_numpy[k.name] for k in onnx_model.graph.input}
+    torch_model = convert(onnx_model).float()
+    with torch.no_grad():
+      for _ in range(1 if test_val is not None else 5):
+        st = time.perf_counter()
+        torch_out = torch_model(*[torch.tensor(x) for x in inputs.values()])
+        timings.append(time.perf_counter() - st)
+      new_torch_out = torch_out.numpy()

-  for _ in range(1 if test_val is not None else 5):
-    st = time.perf_counter()
-    onnx_output = onnx_session.run([onnx_model.graph.output[0].name], onnx_inputs)
-    timings.append(time.perf_counter() - st)
-
-  np.testing.assert_allclose(onnx_output[0].reshape(test_val.shape), test_val, atol=tol, rtol=tol)
-  print("test vs onnx passed")
+  if test_val is not None:
+    np.testing.assert_allclose(new_torch_out.reshape(test_val.shape), test_val, atol=1e-4, rtol=1e-2)
+    print("test vs onnx passed")
  return timings

-def bench(run, inputs):
-  from extra.bench_log import WallTimeEvent, BenchEvent
-  for _ in range(10):
-    with WallTimeEvent(BenchEvent.STEP):
-      run(**inputs).numpy()
-
 if __name__ == "__main__":
  onnx_file = fetch(OPENPILOT_MODEL)
-  inputs, outputs = compile(onnx_file)
+  test_val = compile(onnx_file) if not getenv("RUN") else None

  with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)

-  test_vs_compile(pickle_loaded, inputs, outputs)
-  if getenv("SELFTEST"):
-    test_vs_onnx(inputs, outputs, onnx_file, 1e-4)
+  # same randomness as compile
+  Tensor.manual_seed(100)
+  new_inputs = {nm:Tensor.randn(*st.shape, dtype=dtype).mul(8).realize() for nm, (st, _, dtype, _) in
+                sorted(zip(pickle_loaded.captured.expected_names, pickle_loaded.captured.expected_st_vars_dtype_device))}
+
+  test_val = test_vs_compile(pickle_loaded, new_inputs, test_val)
+  if getenv("BENCHMARK"):
+    for be in ["torch", "ort"]:
+      try:
+        timings = test_vs_onnx(new_inputs, None, onnx_file, be=="ort")
+        print(f"timing {be}: {min(timings)*1000:.2f} ms")
+      except Exception as e:
+        print(f"{be} fail with {e}")
+  if not getenv("FLOAT16"): test_vs_onnx(new_inputs, test_val, onnx_file, getenv("ORT"))

-  if getenv("BENCHMARK_LOG", ""):
-    bench(pickle_loaded, inputs)
@@ -0,0 +1,47 @@
+import sys
+from tinygrad import Tensor, fetch, GlobalCounters, dtypes
+from tinygrad.uop.ops import UOp
+from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.schedule.rangeify import get_rangeify_map
+from tinygrad.engine.schedule import create_schedule_with_vars
+from tinygrad.engine.realize import run_schedule
+
+# NOLOCALS=1 CL=1 IMAGE=2 FLOAT16=1 VIZ=1 DEBUG=2 python3 examples/openpilot/compile4.py
+
+OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
+OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
+
+if __name__ == "__main__":
+  onnx_file = fetch(OPENPILOT_MODEL)
+  run_onnx = OnnxRunner(onnx_file)
+
+  inputs = run_onnx.get_empty_input_data("npy", dtypes.float32)
+  out: Tensor = next(iter(run_onnx({k:v.to(None) for k,v in inputs.items()}).values())).to('cpu')
+  root = out.uop
+  targets = [x.uop for x in inputs.values()]
+  print(targets)
+
+  # TODO: abstract this from gradient?
+
+  # compute the target path (top down)
+  in_target_path: dict[UOp, bool] = {}
+  for u in root.toposort(): in_target_path[u] = any(x in targets or in_target_path[x] for x in u.src)
+  independent_set = {}
+  for u in root.toposort():
+    if in_target_path[u]:
+      for s in u.src:
+        if not in_target_path[s]:
+          independent_set[s] = None
+  independent = UOp.sink(*independent_set.keys())
+  kernelized = get_rangeify_map(independent)
+  independent = independent.substitute(kernelized)
+  schedule, var_vals = create_schedule_with_vars(independent)
+  run_schedule(schedule)
+
+  print("**** real ****")
+  GlobalCounters.reset()
+  out.uop = root.substitute(kernelized)
+  out.kernelize()
+
+  # realize
+  out.realize()
@@ -1,16 +0,0 @@
-import sys, pickle
-from extra.bench_log import WallTimeEvent, BenchEvent
-from tinygrad.helpers import getenv
-
-PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
-
-load_times = []
-
-for _ in range(10):
-  with WallTimeEvent(BenchEvent.STEP) as wte: pickle.load(open(PKL, 'rb'))
-  load_times.append(wte.time)
-  print(f"pickle load: {wte.time:6.2f} s")
-
-if (assert_time:=getenv("ASSERT_MIN_LOAD_TIME")):
-  min_time = min(load_times)
-  assert min_time < assert_time, f"Speed regression, expected min load time of < {assert_time} s but took: {min_time} s"
@@ -0,0 +1,55 @@
+from tinygrad.helpers import trange
+from tinygrad.nn.datasets import mnist
+import mlx.core as mx
+import mlx.nn as nn
+import mlx.optimizers as optim
+from functools import partial
+
+class Model(nn.Module):
+  def __init__(self):
+    super().__init__()
+    self.c1 = nn.Conv2d(1, 32, 5)
+    self.c2 = nn.Conv2d(32, 32, 5)
+    self.bn1 = nn.BatchNorm(32)
+    self.m1 = nn.MaxPool2d(2)
+    self.c3 = nn.Conv2d(32, 64, 3)
+    self.c4 = nn.Conv2d(64, 64, 3)
+    self.bn2 = nn.BatchNorm(64)
+    self.m2 = nn.MaxPool2d(2)
+    self.lin = nn.Linear(576, 10)
+  def __call__(self, x):
+    x = mx.maximum(self.c1(x), 0)
+    x = mx.maximum(self.c2(x), 0)
+    x = self.m1(self.bn1(x))
+    x = mx.maximum(self.c3(x), 0)
+    x = mx.maximum(self.c4(x), 0)
+    x = self.m2(self.bn2(x))
+    return self.lin(mx.flatten(x, 1))
+
+if __name__ == "__main__":
+  X_train, Y_train, X_test, Y_test = mnist()
+  X_train = mx.array(X_train.float().permute((0,2,3,1)).numpy())
+  Y_train = mx.array(Y_train.numpy())
+  X_test = mx.array(X_test.float().permute((0,2,3,1)).numpy())
+  Y_test = mx.array(Y_test.numpy())
+
+  model = Model()
+  optimizer = optim.Adam(1e-3)
+  def loss_fn(model, x, y): return nn.losses.cross_entropy(model(x), y).mean()
+
+  state = [model.state, optimizer.state]
+  @partial(mx.compile, inputs=state, outputs=state)
+  def step(samples):
+    # Compiled functions will also treat any inputs not in the parameter list as constants.
+    X,Y = X_train[samples], Y_train[samples]
+    loss_and_grad_fn = nn.value_and_grad(model, loss_fn)
+    loss, grads = loss_and_grad_fn(model, X, Y)
+    optimizer.update(model, grads)
+    return loss
+
+  test_acc = float('nan')
+  for i in (t:=trange(70)):
+    samples = mx.random.randint(0, X_train.shape[0], (512,))  # putting this in JIT didn't work well
+    loss = step(samples)
+    if i%10 == 9: test_acc = ((model(X_test).argmax(axis=-1) == Y_test).sum() * 100 / X_test.shape[0]).item()
+    t.set_description(f"loss: {loss.item():6.2f} test_accuracy: {test_acc:5.2f}%")
@@ -0,0 +1,45 @@
+import gymnasium as gym
+import numpy as np
+from gymnasium.envs.registration import register
+
+# a very simple game
+# one of <size> lights will light up
+# take the action of the lit up light
+# in <hard_mode>, you act differently based on the step number and need to track this
+
+class PressTheLightUpButton(gym.Env):
+  metadata = {"render_modes": []}
+  def __init__(self, render_mode=None, size=2, game_length=10, hard_mode=False):
+    self.size, self.game_length = size, game_length
+    self.observation_space = gym.spaces.Box(0, 1, shape=(self.size,), dtype=np.float32)
+    self.action_space = gym.spaces.Discrete(self.size)
+    self.step_num = 0
+    self.done = True
+    self.hard_mode = hard_mode
+
+  def _get_obs(self):
+    obs = [0]*self.size
+    if self.step_num < len(self.state):
+      obs[self.state[self.step_num]] = 1
+    return np.array(obs, dtype=np.float32)
+
+  def reset(self, seed=None, options=None):
+    super().reset(seed=seed)
+    self.state = np.random.randint(0, self.size, size=self.game_length)
+    self.step_num = 0
+    self.done = False
+    return self._get_obs(), {}
+
+  def step(self, action):
+    target = ((action + self.step_num) % self.size) if self.hard_mode else action
+    reward = int(target == self.state[self.step_num])
+    self.step_num += 1
+    if not reward:
+      self.done = True
+    return self._get_obs(), reward, self.done, self.step_num >= self.game_length, {}
+
+register(
+  id="PressTheLightUpButton-v0",
+  entry_point="examples.rl.lightupbutton:PressTheLightUpButton",
+  max_episode_steps=None,
+)
@@ -99,7 +99,6 @@ if __name__ == "__main__":
  parser.add_argument('--timing',      action='store_true', help="Print timing per step")
  parser.add_argument('--noshow',      action='store_true', help="Don't show the image")
  parser.add_argument('--fp16',        action='store_true', help="Cast the weights to float16")
-  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
  args = parser.parse_args()

  N = 1
@@ -113,22 +112,19 @@ if __name__ == "__main__":

  model = StableDiffusionV2(**params)

-  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    if not args.fakeweights:
-      default_weights_url = 'https://huggingface.co/sd2-community/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
-      weights_fn = args.weights_fn
-      if not weights_fn:
-        weights_url = args.weights_url if args.weights_url else default_weights_url
-        weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))
+  default_weights_url = 'https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
+  weights_fn = args.weights_fn
+  if not weights_fn:
+    weights_url = args.weights_url if args.weights_url else default_weights_url
+    weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))

-        load_state_dict(model, safe_load(weights_fn), strict=False)
+  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+    load_state_dict(model, safe_load(weights_fn), strict=False)

    if args.fp16:
      for k,v in get_state_dict(model).items():
        if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16))
-
-    Tensor.realize(*get_state_dict(model).values())
+          v.replace(v.cast(dtypes.float16).realize())

  c  = { "crossattn": model.cond_stage_model(args.prompt) }
  uc = { "crossattn": model.cond_stage_model("") }
@@ -0,0 +1,48 @@
+import os, pathlib, argparse
+from examples.llama3 import Tokenizer
+from tabulate import tabulate
+from tinygrad import fetch
+from tinygrad.helpers import flatten
+
+# llama 3 tokenizer
+tokenizer = Tokenizer(fetch("https://huggingface.co/bofenghuang/Meta-Llama-3-8B/resolve/main/original/tokenizer.model").as_posix())
+
+def read_code(base_path):
+  ret = []
+  for path, _, files in os.walk(os.path.join(base_path, "tinygrad")):
+    for name in files:
+      if not name.endswith(".py"): continue
+      if 'tinygrad/runtime/autogen' in path.replace('\\', '/'): continue
+      fullpath = os.path.join(path, name)
+      code = pathlib.Path(fullpath).read_text()
+      ret.append(("### " + fullpath.split("tinygrad/", 1)[1], code))
+  return ret
+
+def write_code_to_file(filename, code_list):
+  """Writes the combined code to a specified file."""
+  with open(filename, 'w') as f:
+    f.write('\n'.join(flatten(code_list)))
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(description="Analyze and optionally save tinygrad code.")
+  parser.add_argument("--output", help="Output file to write the combined code to.")
+  args = parser.parse_args()
+
+  ret = read_code(".")
+
+  table = []
+  for name,code in ret:
+    table.append([name, len(tokenizer.encode(name+"\x00"+code))])
+  print(tabulate([["name", "llm tokens"]]+sorted(table, key=lambda x: -x[1]), headers="firstrow"))
+
+  code_str = '\x00'.join(flatten(ret))
+  print(f"code has {len(code_str)} chars")
+  newline_count = code_str.count('\n')
+  print(f"code has {newline_count} newlines")
+
+  encoded = tokenizer.encode(code_str)
+  print(f"code has {len(encoded)} tokens")
+
+  if args.output:
+    write_code_to_file(args.output, ret)
+    print(f"Combined code written to {args.output}")
@@ -0,0 +1,136 @@
+#!/usr/bin/env python
+#inspired by https://github.com/Matuzas77/MNIST-0.17/blob/master/MNIST_final_solution.ipynb
+import sys
+import numpy as np
+from tinygrad.nn.state import get_parameters
+from tinygrad.tensor import Tensor
+from tinygrad.nn import BatchNorm2d, optim
+from tinygrad.helpers import getenv
+from extra.datasets import fetch_mnist
+from extra.augment import augment_img
+from extra.training import train, evaluate
+GPU = getenv("GPU")
+QUICK = getenv("QUICK")
+DEBUG = getenv("DEBUG")
+
+class SqueezeExciteBlock2D:
+  def __init__(self, filters):
+    self.filters = filters
+    self.weight1 = Tensor.scaled_uniform(self.filters, self.filters//32)
+    self.bias1 = Tensor.scaled_uniform(1,self.filters//32)
+    self.weight2 = Tensor.scaled_uniform(self.filters//32, self.filters)
+    self.bias2 = Tensor.scaled_uniform(1, self.filters)
+
+  def __call__(self, input):
+    se = input.avg_pool2d(kernel_size=(input.shape[2], input.shape[3])) #GlobalAveragePool2D
+    se = se.reshape(shape=(-1, self.filters))
+    se = se.dot(self.weight1) + self.bias1
+    se = se.relu()
+    se = se.dot(self.weight2) + self.bias2
+    se = se.sigmoid().reshape(shape=(-1,self.filters,1,1)) #for broadcasting
+    se = input.mul(se)
+    return se
+
+class ConvBlock:
+  def __init__(self, h, w, inp, filters=128, conv=3):
+    self.h, self.w = h, w
+    self.inp = inp
+    #init weights
+    self.cweights = [Tensor.scaled_uniform(filters, inp if i==0 else filters, conv, conv) for i in range(3)]
+    self.cbiases = [Tensor.scaled_uniform(1, filters, 1, 1) for i in range(3)]
+    #init layers
+    self._bn = BatchNorm2d(128)
+    self._seb = SqueezeExciteBlock2D(filters)
+
+  def __call__(self, input):
+    x = input.reshape(shape=(-1, self.inp, self.w, self.h))
+    for cweight, cbias in zip(self.cweights, self.cbiases):
+      x = x.pad(padding=[1,1,1,1]).conv2d(cweight).add(cbias).relu()
+    x = self._bn(x)
+    x = self._seb(x)
+    return x
+
+class BigConvNet:
+  def __init__(self):
+    self.conv = [ConvBlock(28,28,1), ConvBlock(28,28,128), ConvBlock(14,14,128)]
+    self.weight1 = Tensor.scaled_uniform(128,10)
+    self.weight2 = Tensor.scaled_uniform(128,10)
+
+  def parameters(self):
+    if DEBUG: #keeping this for a moment
+      pars = [par for par in get_parameters(self) if par.requires_grad]
+      no_pars = 0
+      for par in pars:
+        print(par.shape)
+        no_pars += np.prod(par.shape)
+      print('no of parameters', no_pars)
+      return pars
+    else:
+      return get_parameters(self)
+
+  def save(self, filename):
+    with open(filename+'.npy', 'wb') as f:
+      for par in get_parameters(self):
+        #if par.requires_grad:
+        np.save(f, par.numpy())
+
+  def load(self, filename):
+    with open(filename+'.npy', 'rb') as f:
+      for par in get_parameters(self):
+        #if par.requires_grad:
+        try:
+          par.numpy()[:] = np.load(f)
+          if GPU:
+            par.gpu()
+        except:
+          print('Could not load parameter')
+
+  def forward(self, x):
+    x = self.conv[0](x)
+    x = self.conv[1](x)
+    x = x.avg_pool2d(kernel_size=(2,2))
+    x = self.conv[2](x)
+    x1 = x.avg_pool2d(kernel_size=(14,14)).reshape(shape=(-1,128)) #global
+    x2 = x.max_pool2d(kernel_size=(14,14)).reshape(shape=(-1,128)) #global
+    xo = x1.dot(self.weight1) + x2.dot(self.weight2)
+    return xo
+
+
+if __name__ == "__main__":
+  lrs = [1e-4, 1e-5] if QUICK else [1e-3, 1e-4, 1e-5, 1e-5]
+  epochss = [2, 1] if QUICK else [13, 3, 3, 1]
+  BS = 32
+
+  lmbd = 0.00025
+  lossfn = lambda out,y: out.sparse_categorical_crossentropy(y) + lmbd*(model.weight1.abs() + model.weight2.abs()).sum()
+  X_train, Y_train, X_test, Y_test = fetch_mnist()
+  X_train = X_train.reshape(-1, 28, 28).astype(np.uint8)
+  X_test = X_test.reshape(-1, 28, 28).astype(np.uint8)
+  steps = len(X_train)//BS
+  np.random.seed(1337)
+  if QUICK:
+    steps = 1
+    X_test, Y_test = X_test[:BS], Y_test[:BS]
+
+  model = BigConvNet()
+
+  if len(sys.argv) > 1:
+    try:
+      model.load(sys.argv[1])
+      print('Loaded weights "'+sys.argv[1]+'", evaluating...')
+      evaluate(model, X_test, Y_test, BS=BS)
+    except:
+      print('could not load weights "'+sys.argv[1]+'".')
+
+  if GPU:
+    params = get_parameters(model)
+    [x.gpu_() for x in params]
+
+  for lr, epochs in zip(lrs, epochss):
+    optimizer = optim.Adam(model.parameters(), lr=lr)
+    for epoch in range(1,epochs+1):
+      #first epoch without augmentation
+      X_aug = X_train if epoch == 1 else augment_img(X_train)
+      train(model, X_aug, Y_train, optimizer, steps=steps, lossfn=lossfn, BS=BS)
+      accuracy = evaluate(model, X_test, Y_test, BS=BS)
+      model.save(f'examples/checkpoint{accuracy * 1e6:.0f}')
@@ -0,0 +1,17 @@
+from tinygrad.tensor import Tensor
+from tinygrad.nn import Conv2d, BatchNorm2d
+from tinygrad.nn.state import get_parameters
+
+if __name__ == "__main__":
+  with Tensor.train():
+
+    BS, C1, H, W = 4, 16, 224, 224
+    C2, K, S, P = 64, 7, 2, 1
+
+    x = Tensor.uniform(BS, C1, H, W)
+    conv = Conv2d(C1, C2, kernel_size=K, stride=S, padding=P)
+    bn = BatchNorm2d(C2, track_running_stats=False)
+    for t in get_parameters([x, conv, bn]): t.realize()
+
+    print("running network")
+    x.sequential([conv, bn]).numpy()
@@ -0,0 +1,669 @@
+# original implementation: https://github.com/svc-develop-team/so-vits-svc
+from __future__ import annotations
+import sys, logging, time, io, math, argparse, operator, numpy as np
+from functools import partial, reduce
+from pathlib import Path
+from typing import Tuple, Optional, Type
+from tinygrad import nn, dtypes, Tensor
+from tinygrad.helpers import getenv, fetch
+from tinygrad.nn.state import torch_load
+from examples.vits import ResidualCouplingBlock, PosteriorEncoder, Encoder, ResBlock1, ResBlock2, LRELU_SLOPE, sequence_mask, split, get_hparams_from_file, load_checkpoint, weight_norm, HParams
+from examples.sovits_helpers import preprocess
+import soundfile
+
+DEBUG = getenv("DEBUG")
+
+F0_BIN = 256
+F0_MAX = 1100.0
+F0_MIN = 50.0
+F0_MEL_MIN = 1127 * np.log(1 + F0_MIN / 700)
+F0_MEL_MAX = 1127 * np.log(1 + F0_MAX / 700)
+
+class SpeechEncoder:
+  def __init__(self, hidden_dim, model:ContentVec): self.hidden_dim, self.model = hidden_dim, model
+  def encode(self, ): raise NotImplementedError("implement me")
+  @classmethod
+  def load_from_pretrained(cls, checkpoint_path:str, checkpoint_url:str) -> ContentVec:
+    contentvec = ContentVec.load_from_pretrained(checkpoint_path, checkpoint_url)
+    return cls(contentvec)
+
+class ContentVec256L9(SpeechEncoder):
+  def __init__(self, model:ContentVec): super().__init__(hidden_dim=256, model=model)
+  def encode(self, wav: Tensor):
+    feats = wav
+    if len(feats.shape) == 2:  # double channels
+      feats = feats.mean(-1)
+    assert len(feats.shape) == 1, feats.dim()
+    feats = feats.reshape(1, -1)
+    padding_mask = Tensor.zeros_like(feats).cast(dtypes.bool)
+    logits = self.model.extract_features(feats.to(wav.device), padding_mask=padding_mask.to(wav.device), output_layer=9)
+    feats = self.model.final_proj(logits[0])
+    return feats.transpose(1,2)
+
+class ContentVec768L12(SpeechEncoder):
+  def __init__(self, model:ContentVec): super().__init__(hidden_dim=768, model=model)
+  def encode(self, wav: Tensor):
+    feats = wav
+    if len(feats.shape) == 2:  # double channels
+      feats = feats.mean(-1)
+    assert len(feats.shape) == 1, feats.dim()
+    feats = feats.reshape(1, -1)
+    padding_mask = Tensor.zeros_like(feats).cast(dtypes.bool)
+    logits = self.model.extract_features(feats.to(wav.device), padding_mask=padding_mask.to(wav.device), output_layer=12)
+    return logits[0].transpose(1,2)
+
+# original code for contentvec: https://github.com/auspicious3000/contentvec/
+class ContentVec:
+  # self.final_proj dims are hardcoded and depend on fairseq.data.dictionary Dictionary in the checkpoint. This param can't yet be loaded since there is no pickle for it. See with DEBUG=2.
+  # This means that the ContentVec only works with the hubert weights used in all SVC models
+  def __init__(self, cfg: HParams):
+    self.feature_grad_mult, self.untie_final_proj = cfg.feature_grad_mult, cfg.untie_final_proj
+    feature_enc_layers = eval(cfg.conv_feature_layers)
+    self.embed = feature_enc_layers[-1][0]
+    final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim
+    self.feature_extractor = ConvFeatureExtractionModel(conv_layers=feature_enc_layers, dropout=0.0, mode=cfg.extractor_mode, conv_bias=cfg.conv_bias)
+    self.post_extract_proj = nn.Linear(self.embed, cfg.encoder_embed_dim) if self.embed != cfg.encoder_embed_dim else None
+    self.encoder = TransformerEncoder(cfg)
+    self.layer_norm = nn.LayerNorm(self.embed)
+    self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim * 1) if self.untie_final_proj else nn.Linear(cfg.encoder_embed_dim, final_dim)
+    self.mask_emb = Tensor.uniform(cfg.encoder_embed_dim, dtype=dtypes.float32)
+    self.label_embs_concat = Tensor.uniform(504, final_dim, dtype=dtypes.float32)
+  def forward_features(self, source, padding_mask):
+    if self.feature_grad_mult > 0:
+      features = self.feature_extractor(source, padding_mask)
+      if self.feature_grad_mult != 1.0: pass  # training: GradMultiply.forward(features, self.feature_grad_mult)
+    else:
+      features = self.feature_extractor(source, padding_mask)
+    return features
+  def forward_padding_mask(self, features, padding_mask):  # replaces original forward_padding_mask for batch inference
+    lengths_org = tilde(padding_mask.cast(dtypes.bool)).cast(dtypes.int64).sum(1)  # ensure its bool for tilde
+    lengths = (lengths_org - 400).float().div(320).floor().cast(dtypes.int64) + 1  # intermediate float to divide
+    padding_mask = lengths_to_padding_mask(lengths)
+    return padding_mask
+  def extract_features(self, source: Tensor, spk_emb:Tensor=None, padding_mask=None, ret_conv=False, output_layer=None, tap=False):
+    features = self.forward_features(source, padding_mask)
+    if padding_mask is not None:
+      padding_mask = self.forward_padding_mask(features, padding_mask)
+    features = features.transpose(1, 2)
+    features = self.layer_norm(features)
+    if self.post_extract_proj is not None:
+      features = self.post_extract_proj(features)
+    x, _ = self.encoder(features, spk_emb, padding_mask=padding_mask, layer=(None if output_layer is None else output_layer - 1), tap=tap)
+    res = features if ret_conv else x
+    return res, padding_mask
+  @classmethod
+  def load_from_pretrained(cls, checkpoint_path:str, checkpoint_url:str) -> ContentVec:
+    fetch(checkpoint_url, checkpoint_path)
+    cfg = load_fairseq_cfg(checkpoint_path)
+    enc = cls(cfg.model)
+    _ = load_checkpoint_enc(checkpoint_path, enc, None)
+    logging.debug(f"{cls.__name__}: Loaded model with cfg={cfg}")
+    return enc
+
+class TransformerEncoder:
+  def __init__(self, cfg: HParams):
+    def make_conv() -> nn.Conv1d:
+      layer = nn.Conv1d(self.embedding_dim, self.embedding_dim, kernel_size=cfg.conv_pos, padding=cfg.conv_pos // 2, groups=cfg.conv_pos_groups)
+      std = std = math.sqrt(4 / (cfg.conv_pos * self.embedding_dim))
+      layer.weight, layer.bias = (Tensor.normal(*layer.weight.shape, std=std)), (Tensor.zeros(*layer.bias.shape))
+      # for training: layer.weights need to be weight_normed
+      return layer
+    self.dropout, self.embedding_dim, self.layer_norm_first, self.layerdrop, self.num_layers, self.num_layers_1 = cfg.dropout, cfg.encoder_embed_dim, cfg.layer_norm_first, cfg.encoder_layerdrop, cfg.encoder_layers, cfg.encoder_layers_1
+    self.pos_conv, self.pos_conv_remove = [make_conv()], (1 if cfg.conv_pos % 2 == 0 else 0)
+    self.layers = [
+      TransformerEncoderLayer(self.embedding_dim, cfg.encoder_ffn_embed_dim, cfg.encoder_attention_heads, self.dropout, cfg.attention_dropout, cfg.activation_dropout, cfg.activation_fn, self.layer_norm_first, cond_layer_norm=(i >= cfg.encoder_layers))
+      for i in range(cfg.encoder_layers + cfg.encoder_layers_1)
+      ]
+    self.layer_norm = nn.LayerNorm(self.embedding_dim)
+    self.cond_layer_norm = CondLayerNorm(self.embedding_dim) if cfg.encoder_layers_1 > 0 else None
+    # training: apply init_bert_params
+  def __call__(self, x, spk_emb, padding_mask=None, layer=None, tap=False):
+    x, layer_results = self.extract_features(x, spk_emb, padding_mask, layer, tap)
+    if self.layer_norm_first and layer is None:
+      x = self.cond_layer_norm(x, spk_emb) if (self.num_layers_1 > 0) else self.layer_norm(x)
+    return x, layer_results
+  def extract_features(self, x: Tensor, spk_emb: Tensor, padding_mask=None, tgt_layer=None, tap=False):
+    if tgt_layer is not None:  # and not self.training
+      assert tgt_layer >= 0 and tgt_layer < len(self.layers)
+    if padding_mask is not None:
+      # x[padding_mask] = 0
+      assert padding_mask.shape == x.shape[:len(padding_mask.shape)]  # first few dims of x must match padding_mask
+      tmp_mask = padding_mask.unsqueeze(-1).repeat((1, 1, x.shape[-1]))
+      tmp_mask = tilde(tmp_mask.cast(dtypes.bool))
+      x = tmp_mask.where(x, 0)
+    x_conv = self.pos_conv[0](x.transpose(1,2))
+    if self.pos_conv_remove > 0: x_conv = x_conv[:, :, : -self.pos_conv_remove]
+    x_conv = x_conv.gelu().transpose(1, 2)
+    x = (x + x_conv).transpose(0, 1)  # B x T x C -> T x B x C
+    if not self.layer_norm_first: x = self.layer_norm(x)
+    x = x.dropout(p=self.dropout)
+    layer_results = []
+    r = None
+    for i, layer in enumerate(self.layers):
+      if i < self.num_layers:  # if (not self.training or (dropout_probability > self.layerdrop)) and (i < self.num_layers):
+        assert layer.cond_layer_norm == False
+        x = layer(x, self_attn_padding_mask=padding_mask, need_weights=False)
+        if tgt_layer is not None or tap:
+          layer_results.append(x.transpose(0, 1))
+      if i>= self.num_layers:
+        assert layer.cond_layer_norm == True
+        x = layer(x, emb=spk_emb, self_attn_padding_mask=padding_mask, need_weights=False)
+      if i == tgt_layer:
+        r = x
+        break
+    if r is not None:
+      x = r
+    x = x.transpose(0, 1)  # T x B x C -> B x T x C
+    return x, layer_results
+
+class TransformerEncoderLayer:
+  def __init__(self, embedding_dim=768.0, ffn_embedding_dim=3072.0, num_attention_heads=8.0, dropout=0.1, attention_dropout=0.1, activation_dropout=0.1, activation_fn="relu", layer_norm_first=False, cond_layer_norm=False):
+    def get_activation_fn(activation):
+      if activation == "relu": return Tensor.relu
+      if activation == "gelu": return Tensor.gelu
+      else: raise RuntimeError(f"activation function={activation} is not forseen")
+    self.embedding_dim, self.dropout, self.activation_dropout, self.layer_norm_first, self.num_attention_heads, self.cond_layer_norm, self.activation_fn = embedding_dim, dropout, activation_dropout, layer_norm_first, num_attention_heads, cond_layer_norm, get_activation_fn(activation_fn)
+    self.self_attn = MultiHeadAttention(self.embedding_dim, self.num_attention_heads)
+    self.self_attn_layer_norm = nn.LayerNorm(self.embedding_dim) if not cond_layer_norm else CondLayerNorm(self.embedding_dim)
+    self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
+    self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
+    self.final_layer_norm = nn.LayerNorm(self.embedding_dim) if not cond_layer_norm else CondLayerNorm(self.embedding_dim)
+  def __call__(self, x:Tensor, self_attn_mask:Tensor=None, self_attn_padding_mask:Tensor=None, emb:Tensor=None, need_weights=False):
+    #self_attn_padding_mask = self_attn_padding_mask.reshape(x.shape[0], 1, 1, self_attn_padding_mask.shape[1]).expand(-1, self.num_attention_heads, -1, -1).reshape(x.shape[0] * self.num_attention_heads, 1, self_attn_padding_mask.shape[1]) if self_attn_padding_mask is not None else None
+    assert self_attn_mask is None and self_attn_padding_mask is not None
+    residual = x
+    if self.layer_norm_first:
+      x = self.self_attn_layer_norm(x) if not self.cond_layer_norm else self.self_attn_layer_norm(x, emb)
+      x = self.self_attn(x=x, mask=self_attn_padding_mask)
+      x = x.dropout(self.dropout)
+      x = residual + x
+      x = self.final_layer_norm(x) if not self.cond_layer_norm else self.final_layer_norm(x, emb)
+      x = self.activation_fn(self.fc1(x))
+      x = x.dropout(self.activation_dropout)
+      x = self.fc2(x)
+      x = x.dropout(self.dropout)
+      x = residual + x
+    else:
+      x = self.self_attn(x=x, mask=self_attn_padding_mask)
+      x = x.dropout(self.dropout)
+      x = residual + x
+      x = self.self_attn_layer_norm(x) if not self.cond_layer_norm else self.self_attn_layer_norm(x, emb)
+      residual = x
+      x = self.activation_fn(self.fc1(x))
+      x = x.dropout(self.activation_dropout)
+      x = self.fc2(x)
+      x = x.dropout(self.dropout)
+      x = residual + x
+      x = self.final_layer_norm(x) if not self.cond_layer_norm else self.final_layer_norm(x, emb)
+    return x
+
+class MultiHeadAttention:
+  def __init__(self, n_state, n_head):
+    self.n_state, self.n_head = n_state, n_head
+    self.q_proj, self.k_proj, self.v_proj, self.out_proj = [nn.Linear(n_state, n_state) for _ in range(4)]
+  def __call__(self, x:Tensor, xa:Optional[Tensor]=None, mask:Optional[Tensor]=None):
+    x = x.transpose(0,1)  # TxBxC -> BxTxC
+    q, k, v = self.q_proj(x), self.k_proj(xa or x), self.v_proj(xa or x)
+    q, k, v = [x.reshape(*q.shape[:2], self.n_head, -1) for x in (q, k, v)]
+    wv = Tensor.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), None).transpose(1, 2).reshape(*x.shape[:2], -1)
+    ret =  self.out_proj(wv).transpose(0,1)  # BxTxC -> TxBxC
+    return ret
+
+class ConvFeatureExtractionModel:
+  def __init__(self, conv_layers, dropout=.0, mode="default", conv_bias=False):
+    assert mode in {"default", "group_norm_masked", "layer_norm"}
+    def block(n_in, n_out, k, stride, is_layer_norm=False, is_group_norm=False, conv_bias=False):
+      def make_conv():
+        conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias)
+        conv.weight = Tensor.kaiming_normal(*conv.weight.shape)
+        return conv
+      assert (is_layer_norm and is_group_norm) == False, "layer norm and group norm are exclusive"
+      if is_layer_norm:
+        return [make_conv(), partial(Tensor.dropout, p=dropout),[partial(Tensor.transpose, dim0=-2, dim1=-1), nn.LayerNorm(dim, elementwise_affine=True), partial(Tensor.transpose, dim0=-2, dim1=-1)], Tensor.gelu]
+      elif is_group_norm and mode == "default":
+        return [make_conv(), partial(Tensor.dropout, p=dropout), nn.GroupNorm(dim, dim, affine=True), Tensor.gelu]
+      elif is_group_norm and mode == "group_norm_masked":
+        return [make_conv(), partial(Tensor.dropout, p=dropout), GroupNormMasked(dim, dim, affine=True), Tensor.gelu]
+      else:
+        return [make_conv(), partial(Tensor.dropout, p=dropout), Tensor.gelu]
+    in_d, self.conv_layers, self.mode = 1, [], mode
+    for i, cl in enumerate(conv_layers):
+      assert len(cl) == 3, "invalid conv definition: " + str(cl)
+      (dim, k, stride) = cl
+      if i == 0: self.cl = cl
+      self.conv_layers.append(block(in_d, dim, k, stride, is_layer_norm=(mode == "layer_norm"), is_group_norm=((mode == "default" or mode == "group_norm_masked") and i == 0), conv_bias=conv_bias))
+      in_d = dim
+  def __call__(self, x:Tensor, padding_mask:Tensor):
+    x = x.unsqueeze(1)  # BxT -> BxCxT
+    if self.mode == "group_norm_masked":
+      if padding_mask is not None:
+        _, k, stride = self.cl
+        lengths_org = tilde(padding_mask.cast(dtypes.bool)).cast(dtypes.int64).sum(1)  # ensure padding_mask is bool for tilde
+        lengths = (((lengths_org - k) / stride) + 1).floor().cast(dtypes.int64)
+        padding_mask = tilde(lengths_to_padding_mask(lengths)).cast(dtypes.int64)  # lengths_to_padding_mask returns bool tensor
+      x = self.conv_layers[0][0](x)  # padding_mask is numeric
+      x = self.conv_layers[0][1](x)
+      x = self.conv_layers[0][2](x, padding_mask)
+      x = self.conv_layers[0][3](x)
+    else:
+      x = x.sequential(self.conv_layers[0])  # default
+    for _, conv in enumerate(self.conv_layers[1:], start=1):
+      conv = reduce(lambda a,b: operator.iconcat(a,b if isinstance(b, list) else [b]), conv, [])  # flatten
+      x = x.sequential(conv)
+    return x
+
+class CondLayerNorm:  # https://github.com/auspicious3000/contentvec/blob/main/contentvec/modules/cond_layer_norm.py#L10
+  def __init__(self, dim_last, eps=1e-5, dim_spk=256, elementwise_affine=True):
+    self.dim_last, self.eps, self.dim_spk, self.elementwise_affine = dim_last, eps, dim_spk, elementwise_affine
+    if self.elementwise_affine:
+      self.weight_ln = nn.Linear(self.dim_spk, self.dim_last, bias=False)
+      self.bias_ln = nn.Linear(self.dim_spk, self.dim_last, bias=False)
+      self.weight_ln.weight, self.bias_ln.weight = (Tensor.ones(*self.weight_ln.weight.shape)), (Tensor.zeros(*self.bias_ln.weight.shape))
+  def __call__(self, x: Tensor, spk_emb: Tensor):
+    axis = tuple(-1-i for i in range(len(x.shape[1:])))
+    x = x.layernorm(axis=axis, eps=self.eps)
+    if not self.elementwise_affine: return x
+    weights, bias = self.weight_ln(spk_emb), self.bias_ln(spk_emb)
+    return weights * x + bias
+
+class GroupNormMasked:  # https://github.com/auspicious3000/contentvec/blob/d746688a32940f4bee410ed7c87ec9cf8ff04f74/contentvec/modules/fp32_group_norm.py#L16
+  def __init__(self, num_groups, num_channels, eps=1e-5, affine=True):
+    self.num_groups, self.num_channels, self.eps, self.affine = num_groups, num_channels, eps, affine
+    self.weight, self.bias = (Tensor.ones(num_channels)), (Tensor.zeros(num_channels)) if self.affine else (None, None)
+  def __call__(self, x:Tensor, mask:Tensor):
+    bsz, n_c, length = x.shape
+    assert n_c % self.num_groups == 0
+    x = x.reshape(bsz, self.num_groups, n_c // self.num_groups, length)
+    if mask is None: mask = Tensor.ones_like(x)
+    else: mask = mask.reshape(bsz, 1, 1, length)
+    x = x * mask
+    lengths = mask.sum(axis=3, keepdim=True)
+    assert x.shape[2] == 1
+    mean_ = x.mean(dim=3, keepdim=True)
+    mean = mean_ * length / lengths
+    var = (((x.std(axis=3, keepdim=True) ** 2) + mean_**2) * length / lengths - mean**2) + self.eps
+    return x.add(-mean).div(var.sqrt()).reshape(bsz, n_c, length).mul(self.weight.reshape(1,-1,1)).add(self.bias.reshape(1,-1,1))
+
+class Synthesizer:
+  def __init__(self, spec_channels, segment_size, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels, ssl_dim, n_speakers, sampling_rate=44100, vol_embedding=False, n_flow_layer=4, **kwargs):
+    self.spec_channels, self.inter_channels, self.hidden_channels, self.filter_channels, self.n_heads, self.n_layers, self.kernel_size, self.p_dropout, self.resblock, self.resblock_kernel_sizes, self.resblock_dilation_sizes, self.upsample_rates, self.upsample_initial_channel, self.upsample_kernel_sizes, self.segment_size, self.n_speakers, self.gin_channels, self.vol_embedding = spec_channels, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, segment_size, n_speakers, gin_channels, vol_embedding
+    self.emb_g = nn.Embedding(n_speakers, gin_channels)
+    if vol_embedding: self.emb_vol = nn.Linear(1, hidden_channels)
+    self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
+    self.enc_p = TextEncoder(inter_channels, hidden_channels, kernel_size, n_layers, filter_channels=filter_channels, n_heads=n_heads, p_dropout=p_dropout)
+    self.dec = Generator(sampling_rate, inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels)
+    self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
+    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, n_flow_layer, gin_channels=gin_channels)
+    self.emb_uv = nn.Embedding(vocab_size=2, embed_size=hidden_channels)
+  def infer(self, c:Tensor, f0:Tensor, uv:Tensor, g:Tensor=None, noise_scale=0.35, seed=52468, vol=None) -> Tuple[Tensor, Tensor]:
+    Tensor.manual_seed(getenv('SEED', seed))
+    c_lengths = (Tensor.ones([c.shape[0]]) * c.shape[-1]).to(c.device)
+    if len(g.shape) == 1: g = g.unsqueeze(0)
+    g = self.emb_g(g).transpose(1, 2)
+    x_mask = sequence_mask(c_lengths, c.shape[2]).unsqueeze(1).cast(c.dtype)
+    vol = self.emb_vol(vol[:,:,None]).transpose(1,2) if vol is not None and self.vol_embedding else 0
+    x = self.pre(c) * x_mask + self.emb_uv(uv.cast(dtypes.int64)).transpose(1, 2) + vol
+    z_p, _, _, c_mask = self.enc_p.forward(x, x_mask, f0=self._f0_to_coarse(f0), noise_scale=noise_scale)
+    z = self.flow.forward(z_p, c_mask, g=g, reverse=True)
+    o = self.dec.forward(z * c_mask, g=g, f0=f0)
+    return o,f0
+  def _f0_to_coarse(self, f0 : Tensor):
+    f0_mel = 1127 * (1 + f0 / 700).log()
+    a = (F0_BIN - 2) / (F0_MEL_MAX - F0_MEL_MIN)
+    b = F0_MEL_MIN * a - 1.
+    f0_mel = (f0_mel > 0).where(f0_mel * a - b, f0_mel)
+    f0_coarse = f0_mel.ceil().cast(dtype=dtypes.int64)
+    f0_coarse = f0_coarse * (f0_coarse > 0)
+    f0_coarse = f0_coarse + ((f0_coarse < 1) * 1)
+    f0_coarse = f0_coarse * (f0_coarse < F0_BIN)
+    f0_coarse = f0_coarse + ((f0_coarse >= F0_BIN) * (F0_BIN - 1))
+    return f0_coarse
+  @classmethod
+  def load_from_pretrained(cls, config_path:str, config_url:str, weights_path:str, weights_url:str) -> Synthesizer:
+    fetch(config_url, config_path)
+    hps = get_hparams_from_file(config_path)
+    fetch(weights_url, weights_path)
+    net_g = cls(hps.data.filter_length // 2 + 1, hps.train.segment_size // hps.data.hop_length, **hps.model)
+    _ = load_checkpoint(weights_path, net_g, None, skip_list=["f0_decoder"])
+    logging.debug(f"{cls.__name__}:Loaded model with hps: {hps}")
+    return net_g, hps
+
+class TextEncoder:
+  def __init__(self, out_channels, hidden_channels, kernel_size, n_layers, gin_channels=0, filter_channels=None, n_heads=None, p_dropout=None):
+    self.out_channels, self.hidden_channels, self.kernel_size, self.n_layers, self.gin_channels = out_channels, hidden_channels, kernel_size, n_layers, gin_channels
+    self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+    self.f0_emb = nn.Embedding(256, hidden_channels)  # n_vocab = 256
+    self.enc_ = Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
+  def forward(self, x, x_mask, f0=None, noise_scale=1):
+    x = x + self.f0_emb(f0).transpose(1, 2)
+    x = self.enc_.forward(x * x_mask, x_mask)
+    stats = self.proj(x) * x_mask
+    m, logs = split(stats, self.out_channels, dim=1)
+    z = (m + randn_like(m) * logs.exp() * noise_scale) * x_mask
+    return z, m, logs, x_mask
+
+class Upsample:
+  def __init__(self, scale_factor):
+    assert scale_factor % 1 == 0, "Only integer scale factor allowed."
+    self.scale = int(scale_factor)
+  def forward(self, x:Tensor):
+    repeats = tuple([1] * len(x.shape) + [self.scale])
+    new_shape = (*x.shape[:-1], x.shape[-1] * self.scale)
+    return x.unsqueeze(-1).repeat(repeats).reshape(new_shape)
+
+class SineGen:
+  def __init__(self, samp_rate, harmonic_num=0, sine_amp=0.1, noise_std=0.003, voice_threshold=0, flag_for_pulse=False):
+    self.sine_amp, self.noise_std, self.harmonic_num, self.sampling_rate, self.voiced_threshold, self.flag_for_pulse = sine_amp, noise_std, harmonic_num, samp_rate, voice_threshold, flag_for_pulse
+    self.dim = self.harmonic_num + 1
+  def _f02uv(self, f0): return (f0 > self.voiced_threshold).float()  #generate uv signal
+  def _f02sine(self, f0_values):
+    def padDiff(x : Tensor): return (x.pad((0,0,-1,1)) - x).pad((0,0,0,-1))
+    def mod(x: Tensor, n: int) -> Tensor: return x - n * x.div(n).floor()  # this is what the % operator does in pytorch.
+    rad_values = mod((f0_values / self.sampling_rate) , 1)  # convert to F0 in rad
+    rand_ini = Tensor.rand(f0_values.shape[0], f0_values.shape[2], device=f0_values.device)  # initial phase noise
+
+    #rand_ini[:, 0] = 0
+    m = Tensor.ones(f0_values.shape[0]).unsqueeze(1).pad((0,f0_values.shape[2]-1,0,0)).cast(dtypes.bool)
+    m = tilde(m)
+    rand_ini = m.where(rand_ini, 0)
+
+    #rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+    tmp = rad_values[:, 0, :] + rand_ini
+    m = Tensor.ones(tmp.shape).pad((0,0,0,rad_values.shape[1]-1,0)).cast(dtypes.bool)
+    m = tilde(m)
+    tmp = tmp.unsqueeze(1).pad((0,0,0,rad_values.shape[1]-1,0))
+    rad_values = m.where(rad_values, tmp)
+
+    tmp_over_one = mod(rad_values.cumsum(1), 1)
+    tmp_over_one_idx = padDiff(tmp_over_one) < 0
+    cumsum_shift = Tensor.zeros_like(rad_values)
+
+    #cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+    tmp_over_one_idx = (tmp_over_one_idx * -1.0).pad((0,0,1,0))
+    cumsum_shift = tmp_over_one_idx
+
+    sines = ((rad_values + cumsum_shift).cumsum(1) * 2 * np.pi).sin()
+    return sines
+  def forward(self, f0, upp=None):
+    fn = f0.mul(Tensor([[range(1, self.harmonic_num + 2)]], dtype=dtypes.float32).to(f0.device))
+    sine_waves = self._f02sine(fn) * self.sine_amp  #generate sine waveforms
+    uv = self._f02uv(f0)  # generate uv signal
+    noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+    noise = noise_amp * randn_like(sine_waves)
+    sine_waves = sine_waves * uv + noise
+    return sine_waves, uv, noise
+
+class SourceHnNSF:
+  def __init__(self, sampling_rate, harmonic_num=0, sine_amp=0.1, add_noise_std=0.003, voiced_threshold=0):
+    self.sine_amp, self.noise_std = sine_amp, add_noise_std
+    self.l_sin_gen = SineGen(sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshold)
+    self.l_linear = nn.Linear(harmonic_num + 1, 1)
+  def forward(self, x, upp=None):
+    sine_waves, uv, _ = self.l_sin_gen.forward(x, upp)
+    sine_merge = self.l_linear(sine_waves.cast(self.l_linear.weight.dtype)).tanh()
+    noise = randn_like(uv) * self.sine_amp / 3
+    return sine_merge, noise, uv
+
+# most of the hifigan in standard vits is reused here, but need to upsample and construct harmonic source from f0
+class Generator:
+  def __init__(self, sampling_rate, inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels):
+    self.sampling_rate, self.inter_channels, self.resblock, self.resblock_kernel_sizes, self.resblock_dilation_sizes, self.upsample_rates, self.upsample_initial_channel, self.upsample_kernel_sizes, self.gin_channels = sampling_rate, inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels
+    self.num_kernels, self.num_upsamples = len(resblock_kernel_sizes), len(upsample_rates)
+    self.conv_pre = nn.Conv1d(inter_channels, upsample_initial_channel, 7, 1, padding=3)
+    self.f0_upsamp = Upsample(scale_factor=np.prod(upsample_rates))
+    self.m_source = SourceHnNSF(sampling_rate, harmonic_num=8)
+    resblock = ResBlock1 if resblock == '1' else ResBlock2
+    self.ups, self.noise_convs, self.resblocks = [], [], []
+    for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+      c_cur = upsample_initial_channel//(2**(i+1))
+      self.ups.append(nn.ConvTranspose1d(upsample_initial_channel//(2**i), c_cur, k, u, padding=(k-u)//2))
+      stride_f0 = int(np.prod(upsample_rates[i + 1:]))
+      self.noise_convs.append(nn.Conv1d(1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2) if (i + 1 < len(upsample_rates)) else nn.Conv1d(1, c_cur, kernel_size=1))
+    for i in range(len(self.ups)):
+      ch = upsample_initial_channel // (2 ** (i + 1))
+      for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+        self.resblocks.append(resblock(ch, k, d))
+    self.conv_post = nn.Conv1d(ch, 1, 7, 1, padding=3)
+    if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+    self.upp = np.prod(upsample_rates)
+  def forward(self, x, f0, g=None):
+    f0 = self.f0_upsamp.forward(f0[:, None]).transpose(1, 2)  # bs,n,t
+    har_source, _, _ = self.m_source.forward(f0, self.upp)
+    har_source = har_source.transpose(1, 2)
+    x = self.conv_pre(x)
+    if g is not None:  x = x + self.cond(g)
+    for i in range(self.num_upsamples):
+      x, xs = self.ups[i](x.leaky_relu(LRELU_SLOPE)), None
+      x_source = self.noise_convs[i](har_source)
+      x = x + x_source
+      for j in range(self.num_kernels):
+        if xs is None: xs = self.resblocks[i * self.num_kernels + j].forward(x)
+        else: xs += self.resblocks[i * self.num_kernels + j].forward(x)
+      x = xs / self.num_kernels
+    return self.conv_post(x.leaky_relu()).tanh()
+
+# **** helpers ****
+
+def randn_like(x:Tensor) -> Tensor: return Tensor.randn(*x.shape, dtype=x.dtype).to(device=x.device)
+
+def tilde(x: Tensor) -> Tensor:
+  if x.dtype == dtypes.bool: return (1 - x).cast(dtypes.bool)
+  return (x + 1) * -1  # this seems to be what the ~ operator does in pytorch for non bool
+
+def lengths_to_padding_mask(lens:Tensor) -> Tensor:
+  bsz, max_lens = lens.shape[0], lens.max().numpy().item()
+  mask = Tensor.arange(max_lens).to(lens.device).reshape(1, max_lens)
+  mask = mask.expand(bsz, -1) >= lens.reshape(bsz, 1).expand(-1, max_lens)
+  return mask.cast(dtypes.bool)
+
+def repeat_expand_2d_left(content, target_len): # content : [h, t]
+  src_len = content.shape[-1]
+  temp = np.arange(src_len+1) * target_len / src_len
+  current_pos, cols = 0, []
+  for i in range(target_len):
+    if i >= temp[current_pos+1]:
+      current_pos += 1
+    cols.append(content[:, current_pos])
+  return Tensor.stack(*cols).transpose(0, 1)
+
+def load_fairseq_cfg(checkpoint_path):
+  assert Path(checkpoint_path).is_file()
+  state = torch_load(checkpoint_path)
+  cfg = state["cfg"] if ("cfg" in state and state["cfg"] is not None) else None
+  if cfg is None: raise RuntimeError(f"No cfg exist in state keys = {state.keys()}")
+  return HParams(**cfg)
+
+def load_checkpoint_enc(checkpoint_path, model: ContentVec, optimizer=None, skip_list=[]):
+  assert Path(checkpoint_path).is_file()
+  start_time = time.time()
+  checkpoint_dict = torch_load(checkpoint_path)
+  saved_state_dict = checkpoint_dict['model']
+  weight_g, weight_v, parent = None, None, None
+  for key, v in saved_state_dict.items():
+    if any(layer in key for layer in skip_list): continue
+    try:
+      obj, skip = model, False
+      for k in key.split('.'):
+        if k.isnumeric(): obj = obj[int(k)]
+        elif isinstance(obj, dict): obj = obj[k]
+        else:
+          if k in ["weight_g", "weight_v"]:
+            parent, skip = obj, True
+            if k == "weight_g": weight_g = v
+            else: weight_v = v
+          if not skip:
+            parent = obj
+            obj = getattr(obj, k)
+      if weight_g and weight_v:
+        setattr(obj, "weight_g", weight_g.numpy())
+        setattr(obj, "weight_v", weight_v.numpy())
+        obj, v = getattr(parent, "weight"), weight_norm(weight_v, weight_g, 0)
+        weight_g, weight_v, parent, skip = None, None, None, False
+      if not skip and obj.shape == v.shape:
+        if "feature_extractor" in key and (isinstance(parent, (nn.GroupNorm, nn.LayerNorm))):  # cast
+          obj.assign(v.to(obj.device).float())
+        else:
+          obj.assign(v.to(obj.device))
+      elif not skip: logging.error(f"MISMATCH SHAPE IN {key}, {obj.shape} {v.shape}")
+    except Exception as e: raise e
+  logging.info(f"Loaded checkpoint '{checkpoint_path}' in {time.time() - start_time:.4f}s")
+  return model, optimizer
+
+def pad_array(arr, target_length):
+  current_length = arr.shape[0]
+  if current_length >= target_length: return arr
+  pad_width = target_length - current_length
+  pad_left = pad_width // 2
+  pad_right = pad_width - pad_left
+  padded_arr = np.pad(arr, (pad_left, pad_right), 'constant', constant_values=(0, 0))
+  return padded_arr
+
+def split_list_by_n(list_collection, n, pre=0):
+  for i in range(0, len(list_collection), n):
+    yield list_collection[i-pre if i-pre>=0 else i: i + n]
+
+def get_sid(spk2id:HParams, speaker:str) -> Tensor:
+  speaker_id = spk2id[speaker]
+  if not speaker_id and type(speaker) is int:
+    if len(spk2id.__dict__) >= speaker: speaker_id = speaker
+  if speaker_id is None: raise RuntimeError(f"speaker={speaker} not in the speaker list")
+  return Tensor([int(speaker_id)], dtype=dtypes.int64).unsqueeze(0)
+
+def get_encoder(ssl_dim) -> Type[SpeechEncoder]:
+  if ssl_dim == 256: return ContentVec256L9
+  if ssl_dim == 768: return ContentVec768L12
+
+#########################################################################################
+# CODE: https://github.com/svc-develop-team/so-vits-svc
+#########################################################################################
+# CONTENTVEC:
+#   CODE: https://github.com/auspicious3000/contentvec
+#   PAPER: https://arxiv.org/abs/2204.09224
+#########################################################################################
+# INSTALLATION: dependencies are for preprocessing and loading/saving audio.
+# pip3 install soundfile librosa praat-parselmouth
+#########################################################################################
+# EXAMPLE USAGE:
+# python3 examples/so_vits_svc.py --model tf2spy --file ~/recording.wav
+#########################################################################################
+# DEMO USAGE (uses audio sample from LJ-Speech):
+# python3 examples/so_vits_svc.py --model saul_goodman
+#########################################################################################
+SO_VITS_SVC_PATH = Path(__file__).parents[1] / "weights/So-VITS-SVC"
+VITS_MODELS = { # config_path, weights_path, config_url, weights_url
+  "saul_goodman" : (SO_VITS_SVC_PATH / "config_saul_gman.json", SO_VITS_SVC_PATH / "pretrained_saul_gman.pth", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/Saul_Goodman_80000/config.json", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/Saul_Goodman_80000/G_80000.pth"),
+  "drake" : (SO_VITS_SVC_PATH / "config_drake.json", SO_VITS_SVC_PATH / "pretrained_drake.pth", "https://huggingface.co/jaspa/so-vits-svc/resolve/main/aubrey/config_aubrey.json", "https://huggingface.co/jaspa/so-vits-svc/resolve/main/aubrey/pretrained_aubrey.pth"),
+  "cartman" : (SO_VITS_SVC_PATH / "config_cartman.json", SO_VITS_SVC_PATH / "pretrained_cartman.pth", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/EricCartman/config.json", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/EricCartman/G_10200.pth"),
+  "tf2spy" : (SO_VITS_SVC_PATH / "config_tf2spy.json", SO_VITS_SVC_PATH / "pretrained_tf2spy.pth", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_spy_60k/config.json", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_spy_60k/G_60000.pth"),
+  "tf2heavy" : (SO_VITS_SVC_PATH / "config_tf2heavy.json", SO_VITS_SVC_PATH / "pretrained_tf2heavy.pth", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_heavy_100k/config.json", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_heavy_100k/G_100000.pth"),
+  "lady_gaga" : (SO_VITS_SVC_PATH / "config_gaga.json", SO_VITS_SVC_PATH / "pretrained_gaga.pth", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/LadyGaga/config.json", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/LadyGaga/G_14400.pth")
+}
+ENCODER_MODELS = { # weights_path, weights_url
+  "contentvec": (SO_VITS_SVC_PATH / "contentvec_checkpoint.pt", "https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt")
+}
+ENCODER_MODEL = "contentvec"
+DEMO_PATH, DEMO_URL = Path(__file__).parents[1] / "temp/LJ037-0171.wav", "https://keithito.com/LJ-Speech-Dataset/LJ037-0171.wav"
+if __name__=="__main__":
+  logging.basicConfig(stream=sys.stdout, level=(logging.INFO if DEBUG < 1 else logging.DEBUG))
+  parser = argparse.ArgumentParser()
+  parser.add_argument("-m", "--model", default=None, help=f"Specify the model to use. All supported models: {VITS_MODELS.keys()}", required=True)
+  parser.add_argument("-f", "--file", default=DEMO_PATH, help=f"Specify the path of the input file")
+  parser.add_argument("--out_dir", default=str(Path(__file__).parents[1] / "temp"), help="Specify the output path.")
+  parser.add_argument("--out_path", default=None, help="Specify the full output path. Overrides the --out_dir and --name parameter.")
+  parser.add_argument("--base_name", default="test", help="Specify the base of the output file name. Default is 'test'.")
+  parser.add_argument("--speaker", default=None, help="If not specified, the first available speaker is chosen. Usually there is only one speaker per model.")
+  parser.add_argument("--noise_scale", default=0.4)
+  parser.add_argument("--tran", default=0.0, help="Pitch shift, supports positive and negative (semitone) values. Default 0.0")
+  parser.add_argument("--pad_seconds", default=0.5)
+  parser.add_argument("--lg_num", default=0.0)
+  parser.add_argument("--clip_seconds", default=0.0)
+  parser.add_argument("--slice_db", default=-40)
+  args = parser.parse_args()
+
+  vits_model = args.model
+  encoder_location, vits_location = ENCODER_MODELS[ENCODER_MODEL], VITS_MODELS[vits_model]
+
+  Tensor.training = False
+  # Get Synthesizer and ContentVec
+  net_g, hps = Synthesizer.load_from_pretrained(vits_location[0], vits_location[2], vits_location[1], vits_location[3])
+  Encoder = get_encoder(hps.model.ssl_dim)
+  encoder = Encoder.load_from_pretrained(encoder_location[0], encoder_location[1])
+
+  # model config args
+  target_sample, spk2id, hop_length, target_sample = hps.data.sampling_rate, hps.spk, hps.data.hop_length, hps.data.sampling_rate
+  vol_embedding = hps.model.vol_embedding if hasattr(hps.data, "vol_embedding") and hps.model.vol_embedding is not None else False
+
+  # args
+  slice_db, clip_seconds, lg_num, pad_seconds, tran, noise_scale, audio_path = args.slice_db, args.clip_seconds, args.lg_num, args.pad_seconds, args.tran, args.noise_scale, args.file
+  speaker = args.speaker if args.speaker is not None else list(hps.spk.__dict__.keys())[0]
+
+  ### Loading audio and slicing ###
+  if audio_path == DEMO_PATH: fetch(DEMO_URL, DEMO_PATH)
+  assert Path(audio_path).is_file() and Path(audio_path).suffix == ".wav"
+  chunks = preprocess.cut(audio_path, db_thresh=slice_db)
+  audio_data, audio_sr = preprocess.chunks2audio(audio_path, chunks)
+
+  per_size = int(clip_seconds * audio_sr)
+  lg_size = int(lg_num * audio_sr)
+
+  ### Infer per slice ###
+  global_frame = 0
+  audio = []
+  for (slice_tag, data) in audio_data:
+    print(f"\n====segment start, {round(len(data) / audio_sr, 3)}s====")
+    length = int(np.ceil(len(data) / audio_sr * target_sample))
+
+    if slice_tag:
+      print("empty segment")
+      _audio = np.zeros(length)
+      audio.extend(list(pad_array(_audio, length)))
+      global_frame += length // hop_length
+      continue
+
+    datas = [data] if per_size == 0 else split_list_by_n(data, per_size, lg_size)
+
+    for k, dat in enumerate(datas):
+      per_length = int(np.ceil(len(dat) / audio_sr * target_sample)) if clip_seconds!=0 else length
+      pad_len = int(audio_sr * pad_seconds)
+      dat = np.concatenate([np.zeros([pad_len]), dat, np.zeros([pad_len])])
+      raw_path = io.BytesIO()
+      soundfile.write(raw_path, dat, audio_sr, format="wav")
+      raw_path.seek(0)
+
+      ### Infer START ###
+      wav, sr = preprocess.load_audiofile(raw_path)
+      wav = preprocess.sinc_interp_resample(wav, sr, target_sample)[0]
+      wav16k, f0, uv = preprocess.get_unit_f0(wav, tran, hop_length, target_sample)
+      sid = get_sid(spk2id, speaker)
+      n_frames = f0.shape[1]
+
+      # ContentVec infer
+      start = time.time()
+      c = encoder.encode(wav16k)
+      c = repeat_expand_2d_left(c.squeeze(0).realize(), f0.shape[1])  # interpolate speech encoding to match f0
+      c = c.unsqueeze(0).realize()
+      enc_time = time.time() - start
+
+      # VITS infer
+      vits_start = time.time()
+      out_audio, f0 = net_g.infer(c, f0=f0, uv=uv, g=sid, noise_scale=noise_scale, vol=None)
+      out_audio = out_audio[0,0].float().realize()
+      vits_time = time.time() - vits_start
+
+      infer_time = time.time() - start
+      logging.info("total infer time:{:.2f}s, speech_enc time:{:.2f}s, vits time:{:.2f}s".format(infer_time, enc_time, vits_time))
+      ### Infer END ###
+
+      out_sr, out_frame = out_audio.shape[-1], n_frames
+      global_frame += out_frame
+      _audio = out_audio.numpy()
+      pad_len = int(target_sample * pad_seconds)
+      _audio = _audio[pad_len:-pad_len]
+      _audio = pad_array(_audio, per_length)
+      audio.extend(list(_audio))
+
+  audio = np.array(audio)
+  out_path = Path(args.out_path or Path(args.out_dir)/f"{args.model}{f'_spk_{speaker}'}_{args.base_name}.wav")
+  out_path.parent.mkdir(parents=True, exist_ok=True)
+  soundfile.write(out_path, audio, target_sample, format="flac")
+  logging.info(f"Saved audio output to {out_path}")
@@ -0,0 +1,204 @@
+import math
+from typing import Optional, Tuple
+from tinygrad import Tensor, dtypes
+import librosa
+import soundfile
+import numpy as np
+import parselmouth
+
+class PMF0Predictor:  # from https://github.com/svc-develop-team/so-vits-svc/
+  def __init__(self,hop_length=512,f0_min=50,f0_max=1100,sampling_rate=44100):
+    self.hop_length, self.f0_min, self.f0_max, self.sampling_rate, self.name = hop_length, f0_min, f0_max, sampling_rate, "pm"
+  def interpolate_f0(self,f0):
+    vuv_vector = np.zeros_like(f0, dtype=np.float32)
+    vuv_vector[f0 > 0.0] = 1.0
+    vuv_vector[f0 <= 0.0] = 0.0
+    nzindex = np.nonzero(f0)[0]
+    data = f0[nzindex]
+    nzindex = nzindex.astype(np.float32)
+    time_org = self.hop_length / self.sampling_rate * nzindex
+    time_frame = np.arange(f0.shape[0]) * self.hop_length / self.sampling_rate
+    if data.shape[0] <= 0: return np.zeros(f0.shape[0], dtype=np.float32),vuv_vector
+    if data.shape[0] == 1: return np.ones(f0.shape[0], dtype=np.float32) * f0[0],vuv_vector
+    f0 = np.interp(time_frame, time_org, data, left=data[0], right=data[-1])
+    return f0,vuv_vector
+  def compute_f0(self,wav,p_len=None):
+    x = wav
+    if p_len is None: p_len = x.shape[0]//self.hop_length
+    else: assert abs(p_len-x.shape[0]//self.hop_length) < 4, "pad length error"
+    time_step = self.hop_length / self.sampling_rate * 1000
+    f0 = parselmouth.Sound(x, self.sampling_rate) \
+                    .to_pitch_ac(time_step=time_step / 1000, voicing_threshold=0.6,pitch_floor=self.f0_min, pitch_ceiling=self.f0_max) \
+                    .selected_array['frequency']
+    pad_size=(p_len - len(f0) + 1) // 2
+    if(pad_size>0 or p_len - len(f0) - pad_size>0):
+      f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
+    f0,uv = self.interpolate_f0(f0)
+    return f0
+  def compute_f0_uv(self,wav,p_len=None):
+    x = wav
+    if p_len is None: p_len = x.shape[0]//self.hop_length
+    else: assert abs(p_len-x.shape[0]//self.hop_length) < 4, "pad length error"
+    time_step = self.hop_length / self.sampling_rate * 1000
+    f0 = parselmouth.Sound(x, self.sampling_rate).to_pitch_ac(
+      time_step=time_step / 1000, voicing_threshold=0.6,
+      pitch_floor=self.f0_min, pitch_ceiling=self.f0_max).selected_array['frequency']
+    pad_size=(p_len - len(f0) + 1) // 2
+    if(pad_size>0 or p_len - len(f0) - pad_size>0):
+      f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
+    f0,uv = self.interpolate_f0(f0)
+    return f0,uv
+
+class Slicer:  # from https://github.com/svc-develop-team/so-vits-svc/
+  def __init__(self, sr: int, threshold: float = -40., min_length: int = 5000, min_interval: int = 300, hop_size: int = 20, max_sil_kept: int = 5000):
+    if not min_length >= min_interval >= hop_size:
+      raise ValueError('The following condition must be satisfied: min_length >= min_interval >= hop_size')
+    if not max_sil_kept >= hop_size:
+      raise ValueError('The following condition must be satisfied: max_sil_kept >= hop_size')
+    min_interval = sr * min_interval / 1000
+    self.threshold = 10 ** (threshold / 20.)
+    self.hop_size = round(sr * hop_size / 1000)
+    self.win_size = min(round(min_interval), 4 * self.hop_size)
+    self.min_length = round(sr * min_length / 1000 / self.hop_size)
+    self.min_interval = round(min_interval / self.hop_size)
+    self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+  def _apply_slice(self, waveform, begin, end):
+    if len(waveform.shape) > 1: return waveform[:, begin * self.hop_size: min(waveform.shape[1], end * self.hop_size)]
+    else: return waveform[begin * self.hop_size: min(waveform.shape[0], end * self.hop_size)]
+  def slice(self, waveform):
+    samples = librosa.to_mono(waveform) if len(waveform.shape) > 1 else waveform
+    if samples.shape[0] <= self.min_length: return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
+    rms_list = librosa.feature.rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
+    sil_tags, silence_start, clip_start = [], None, 0
+    for i, rms in enumerate(rms_list):
+      if rms < self.threshold:  # Keep looping while frame is silent.
+        if silence_start is None:  # Record start of silent frames.
+          silence_start = i
+        continue
+      if silence_start is None: continue  # Keep looping while frame is not silent and silence start has not been recorded.
+      # Clear recorded silence start if interval is not enough or clip is too short
+      is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+      need_slice_middle = i - silence_start >= self.min_interval and i - clip_start >= self.min_length
+      if not is_leading_silence and not need_slice_middle:
+        silence_start = None
+        continue
+      if i - silence_start <= self.max_sil_kept:  # Need slicing. Record the range of silent frames to be removed.
+        pos = rms_list[silence_start: i + 1].argmin() + silence_start
+        sil_tags.append((0, pos) if silence_start == 0 else (pos, pos))
+        clip_start = pos
+      elif i - silence_start <= self.max_sil_kept * 2:
+        pos = rms_list[i - self.max_sil_kept: silence_start + self.max_sil_kept + 1].argmin()
+        pos += i - self.max_sil_kept
+        pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
+        pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
+        if silence_start == 0:
+          sil_tags.append((0, pos_r))
+          clip_start = pos_r
+        else:
+          sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+          clip_start = max(pos_r, pos)
+      else:
+        pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
+        pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
+        sil_tags.append((0, pos_r) if silence_start == 0 else (pos_l, pos_r))
+        clip_start = pos_r
+      silence_start = None
+    total_frames = rms_list.shape[0]
+    if silence_start is not None and total_frames - silence_start >= self.min_interval:  # Deal with trailing silence.
+      silence_end = min(total_frames, silence_start + self.max_sil_kept)
+      pos = rms_list[silence_start: silence_end + 1].argmin() + silence_start
+      sil_tags.append((pos, total_frames + 1))
+    if len(sil_tags) == 0: return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}  # Apply and return slices.
+    chunks = []
+    if sil_tags[0][0]:
+      chunks.append({"slice": False, "split_time": f"0,{min(waveform.shape[0], sil_tags[0][0] * self.hop_size)}"})
+    for i in range(0, len(sil_tags)):
+      if i: chunks.append({"slice": False, "split_time": f"{sil_tags[i - 1][1] * self.hop_size},{min(waveform.shape[0], sil_tags[i][0] * self.hop_size)}"})
+      chunks.append({"slice": True, "split_time": f"{sil_tags[i][0] * self.hop_size},{min(waveform.shape[0], sil_tags[i][1] * self.hop_size)}"})
+    if sil_tags[-1][1] * self.hop_size < len(waveform):
+      chunks.append({"slice": False, "split_time": f"{sil_tags[-1][1] * self.hop_size},{len(waveform)}"})
+    chunk_dict = {}
+    for i in range(len(chunks)): chunk_dict[str(i)] = chunks[i]
+    return chunk_dict
+
+# sinc_interp_hann audio resampling
+class Resample:
+  def __init__(self, orig_freq:int=16000, new_freq:int=16000, lowpass_filter_width:int=6, rolloff:float=0.99, beta:Optional[float]=None, dtype:Optional[dtypes]=None):
+    self.orig_freq, self.new_freq, self.lowpass_filter_width, self.rolloff, self.beta = orig_freq, new_freq, lowpass_filter_width, rolloff, beta
+    self.gcd = math.gcd(int(self.orig_freq), int(self.new_freq))
+    self.kernel, self.width = self._get_sinc_resample_kernel(dtype) if self.orig_freq != self.new_freq else (None, None)
+  def __call__(self, waveform:Tensor) -> Tensor:
+    if self.orig_freq == self.new_freq: return waveform
+    return self._apply_sinc_resample_kernel(waveform)
+  def _apply_sinc_resample_kernel(self, waveform:Tensor):
+    if not waveform.is_floating_point(): raise TypeError(f"Waveform tensor expected to be of type float, but received {waveform.dtype}.")
+    orig_freq, new_freq = (int(self.orig_freq) // self.gcd), (int(self.new_freq) // self.gcd)
+    shape = waveform.shape
+    waveform = waveform.reshape(-1, shape[-1])  # pack batch
+    num_wavs, length = waveform.shape
+    target_length = int(math.ceil(new_freq * length / orig_freq))
+    waveform = waveform.pad((self.width, self.width + orig_freq))
+    resampled = waveform[:, None].conv2d(self.kernel, stride=orig_freq)
+    resampled = resampled.transpose(1, 2).reshape(num_wavs, -1)
+    resampled = resampled[..., :target_length]
+    resampled = resampled.reshape(shape[:-1] + resampled.shape[-1:])  # unpack batch
+    return resampled
+  def _get_sinc_resample_kernel(self, dtype=None):
+    orig_freq, new_freq = (int(self.orig_freq) // self.gcd), (int(self.new_freq) // self.gcd)
+    if self.lowpass_filter_width <= 0: raise ValueError("Low pass filter width should be positive.")
+    base_freq = min(orig_freq, new_freq)
+    base_freq *= self.rolloff
+    width = math.ceil(self.lowpass_filter_width * orig_freq / base_freq)
+    idx = Tensor.arange(-width, width + orig_freq, dtype=(dtype if dtype is not None else dtypes.float32))[None, None] / orig_freq
+    t = Tensor.arange(0, -new_freq, -1, dtype=dtype)[:, None, None] / new_freq + idx
+    t *= base_freq
+    t = t.clip(-self.lowpass_filter_width, self.lowpass_filter_width)
+    window = (t * math.pi / self.lowpass_filter_width / 2).cos() ** 2
+    t *= math.pi
+    scale = base_freq / orig_freq
+    kernels = Tensor.where(t == 0, Tensor(1.0, dtype=t.dtype).to(t.device), t.sin() / t)
+    kernels *= window * scale
+    if dtype is None: kernels = kernels.cast(dtype=dtypes.float32)
+    return kernels, width
+
+def sinc_interp_resample(x:Tensor, orig_freq:int=16000, new_freq:int=1600, lowpass_filter_width:int=6, rolloff:float=0.99, beta:Optional[float]=None):
+  resamp = Resample(orig_freq, new_freq, lowpass_filter_width, rolloff, beta, x.dtype)
+  return resamp(x)
+
+def cut(audio_path, db_thresh=-30, min_len=5000):
+  audio, sr = librosa.load(audio_path, sr=None)
+  slicer = Slicer(sr=sr, threshold=db_thresh, min_length=min_len)
+  chunks = slicer.slice(audio)
+  return chunks
+
+def chunks2audio(audio_path, chunks):
+  chunks = dict(chunks)
+  audio, sr = load_audiofile(audio_path)
+  if len(audio.shape) == 2 and audio.shape[1] >= 2:
+    audio = audio.mean(0).unsqueeze(0)
+  audio = audio.numpy()[0]
+  result = []
+  for k, v in chunks.items():
+    tag = v["split_time"].split(",")
+    if tag[0] != tag[1]:
+      result.append((v["slice"], audio[int(tag[0]):int(tag[1])]))
+  return result, sr
+
+def load_audiofile(filepath:str, frame_offset:int=0, num_frames:int=-1, channels_first:bool=True):
+  with soundfile.SoundFile(filepath, "r") as file_:
+    frames = file_._prepare_read(frame_offset, None, num_frames)
+    waveform = file_.read(frames, "float32", always_2d=True)
+    sample_rate = file_.samplerate
+  waveform = Tensor(waveform)
+  if channels_first: waveform = waveform.transpose(0, 1)
+  return waveform, sample_rate
+
+def get_unit_f0(wav:Tensor, tran, hop_length, target_sample, f0_filter=False) -> Tuple[Tensor,Tensor,Tensor]:
+  f0_predictor = PMF0Predictor(hop_length, sampling_rate=target_sample)
+  f0, uv = f0_predictor.compute_f0_uv(wav.numpy())
+  if f0_filter and sum(f0) == 0: raise RuntimeError("No voice detected")
+  f0 = Tensor(f0.astype(np.float32)).float()
+  f0 = (f0 * 2 ** (tran / 12)).unsqueeze(0)
+  uv = Tensor(uv.astype(np.float32)).float().unsqueeze(0)
+  wav16k = sinc_interp_resample(wav[None,:], target_sample, 16000)[0]
+  return wav16k.realize(), f0.realize(), uv.realize()
@@ -6,9 +6,10 @@ import argparse, time
 from collections import namedtuple
 from typing import Dict, Any

+from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
-from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten, profile_marker
+from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten
 from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import torch_load, load_state_dict, get_state_dict
 from extra.models.clip import Closed, Tokenizer, FrozenOpenClipEmbedder
@@ -262,19 +263,13 @@ if __name__ == "__main__":
  parser.add_argument('--timing', action='store_true', help="Print timing per step")
  parser.add_argument('--seed', type=int, help="Set the random latent seed")
  parser.add_argument('--guidance', type=float, default=7.5, help="Prompt strength")
-  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
  args = parser.parse_args()

-  profile_marker("create model")
  model = StableDiffusion()

-  profile_marker("load in weights")
+  # load in weights
  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    if not args.fakeweights:
-      model_bin = fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt')
-      state_dict = torch_load(model_bin)['state_dict']
-      profile_marker("state dict loaded")
-      load_state_dict(model, state_dict, verbose=False, strict=False, realize=False)
+    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], verbose=False, strict=False, realize=False)

    if args.fp16:
      for k,v in get_state_dict(model).items():
@@ -283,13 +278,12 @@ if __name__ == "__main__":

    Tensor.realize(*get_state_dict(model).values())

-  profile_marker("run clip (conditional)")
+  # run through CLIP to get context
  tokenizer = Tokenizer.ClipTokenizer()
  prompt = Tensor([tokenizer.encode(args.prompt)])
  context = model.cond_stage_model.transformer.text_model(prompt).realize()
  print("got CLIP context", context.shape)

-  profile_marker("run clip (unconditional)")
  prompt = Tensor([tokenizer.encode("")])
  unconditional_context = model.cond_stage_model.transformer.text_model(prompt).realize()
  print("got unconditional CLIP context", unconditional_context.shape)
@@ -313,7 +307,6 @@ if __name__ == "__main__":
  step_times = []
  with Context(BEAM=getenv("LATEBEAM")):
    for index, timestep in (t:=tqdm(list(enumerate(timesteps))[::-1])):
-      profile_marker(f"step {len(timesteps)-index-1}")
      GlobalCounters.reset()
      st = time.perf_counter_ns()
      t.set_description("%3d %3d" % (index, timestep))
@@ -323,27 +316,24 @@ if __name__ == "__main__":
          latent = run(model, unconditional_context, context, latent, Tensor([timestep]), alphas[tid], alphas_prev[tid], Tensor([args.guidance]))
          if args.timing: Device[Device.DEFAULT].synchronize()
      step_times.append((time.perf_counter_ns() - st)*1e-6)
-    # done with diffusion model
    del run
-    del model.model

  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
    min_time = min(step_times)
    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
-  profile_marker("run decoder") # upsample latent space to image with autoencoder
-  x = model.decode(latent).realize()
+  # upsample latent space to image with autoencoder
+  x = model.decode(latent)
  print(x.shape)

-  profile_marker("save image")
-  from PIL import Image
+  # save image
  im = Image.fromarray(x.numpy())
  print(f"saving {args.out}")
  im.save(args.out)
  # Open image.
  if not args.noshow: im.show()

+  # validation!
  if args.prompt == default_prompt and args.steps == 6 and args.seed == 0 and args.guidance == 7.5:
-    profile_marker("validate")
    ref_image = Tensor(np.array(Image.open(Path(__file__).parent / "stable_diffusion_seed0.png")))
    distance = (((x.cast(dtypes.float) - ref_image.cast(dtypes.float)) / ref_image.max())**2).mean().item()
    assert distance < 3e-3, colored(f"validation failed with {distance=}", "red")  # higher distance with WINO
@@ -19,8 +19,8 @@ from tinygrad.helpers import fetch, getenv

 # QUANT=1 python3 examples/test_onnx_imagenet.py
 #   https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx
-# python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
-# VIZ=1 python3 examples/benchmark_onnx.py /tmp/model.quant.onnx
+# DONT_REALIZE_EXPAND=1 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+# VIZ=1 DONT_REALIZE_EXPAND=1 python3 examples/benchmark_onnx.py /tmp/model.quant.onnx

 def imagenet_dataloader(cnt=0):
  input_mean = Tensor([0.485, 0.456, 0.406]).reshape(1, -1, 1, 1)
@@ -7,7 +7,7 @@ if __name__ == "__main__":
  with open(fetch(sys.argv[1]), "rb") as f:
    run_onnx_jit = pickle.load(f)
  input_name = run_onnx_jit.captured.expected_names[0]
-  device = run_onnx_jit.captured.expected_input_info[0][-1]
+  device = run_onnx_jit.captured.expected_st_vars_dtype_device[0][-1]
  print(f"input goes into {input_name=} on {device=}")
  hit = 0
  for i,(img,y) in enumerate(imagenet_dataloader(cnt=getenv("CNT", 100))):
@@ -48,7 +48,7 @@ def prepare_browser_chunks(model):
        weight_metadata = metadata.get(name, default)
        weight_metadata["parts"][part_num] = {"file": i, "file_start_pos": cursor, "size": size}
        metadata[name] = weight_metadata
-        data = bytes(state_dict[name].uop.base.realized.as_memoryview())
+        data = bytes(state_dict[name].uop.base.realized.as_buffer())
        data = data if not offsets else data[offsets[0]:offsets[1]]
        writer.write(data)
        cursor += size
@@ -1,34 +0,0 @@
-#!/usr/bin/env python3
-from tinygrad import Tensor, Device, GlobalCounters, Context, dtypes
-from tinygrad.helpers import getenv, colored
-
-SZ = 8_000_000_000
-GPUS = getenv("GPUS", 4) # TODO: expose a way in tinygrad to access this
-
-if __name__ == "__main__":
-  # create tensors
-  tens = [Tensor.ones(SZ, dtype=dtypes.uint8, device=f"{Device.DEFAULT}:{i}").contiguous() for i in range(GPUS)]
-  Tensor.realize(*tens)
-
-  bw = [[0.0]*GPUS for _ in range(GPUS)]
-  for i in range(GPUS):
-    for j in range(GPUS):
-      GlobalCounters.reset()
-      with Context(DEBUG=2):
-        if i == j:
-          # this copy would be optimized out, just add 1
-          (tens[i]+1).realize()
-        else:
-          tens[i].to(f"{Device.DEFAULT}:{j}").realize()
-      t = max(GlobalCounters.time_sum_s, 1e-9)
-      bw[i][j] = SZ / t / 1e9  # GB/s
-
-  def fmt(x):
-    c = "green" if x > 50 else "yellow" if x > 20 else "red"
-    return colored(f"{x:6.1f}", c)
-
-  # header
-  print(" " * 8 + " ".join(f"{'d'+str(j):>6}" for j in range(GPUS)))
-  # rows
-  for i in range(GPUS):
-    print(f"{'s'+str(i):>6} -> " + " ".join(fmt(x) for x in bw[i]))
@@ -1,16 +0,0 @@
-from tinygrad import Tensor, Device, TinyJit, dtypes
-from tinygrad.helpers import getenv
-
-GPUS = getenv("GPUS", 4) # TODO: expose a way in tinygrad to access this
-N = 6144
-
-@TinyJit
-def many_matmul(A, B):
-  out = A
-  for _ in range(8): out = out@B
-  return out
-
-if __name__ == "__main__":
-  A = Tensor.ones(GPUS, N, N, dtype=dtypes.half).shard(devices=tuple([f"{Device.DEFAULT}:{i}" for i in range(GPUS)]), axis=0).contiguous()
-  B = Tensor.ones(GPUS, N, N, dtype=dtypes.half).shard(devices=tuple([f"{Device.DEFAULT}:{i}" for i in range(GPUS)]), axis=0).contiguous()
-  while 1: many_matmul(A, B)
@@ -0,0 +1,104 @@
+import traceback
+import time
+from multiprocessing import Process, Queue
+import numpy as np
+from tinygrad.nn.state import get_parameters
+from tinygrad.nn import optim
+from tinygrad.helpers import getenv, trange
+from tinygrad.tensor import Tensor
+from extra.datasets import fetch_cifar
+from extra.models.efficientnet import EfficientNet
+
+class TinyConvNet:
+  def __init__(self, classes=10):
+    conv = 3
+    inter_chan, out_chan = 8, 16   # for speed
+    self.c1 = Tensor.uniform(inter_chan,3,conv,conv)
+    self.c2 = Tensor.uniform(out_chan,inter_chan,conv,conv)
+    self.l1 = Tensor.uniform(out_chan*6*6, classes)
+
+  def forward(self, x):
+    x = x.conv2d(self.c1).relu().max_pool2d()
+    x = x.conv2d(self.c2).relu().max_pool2d()
+    x = x.reshape(shape=[x.shape[0], -1])
+    return x.dot(self.l1)
+
+if __name__ == "__main__":
+  IMAGENET = getenv("IMAGENET")
+  classes = 1000 if IMAGENET else 10
+
+  TINY = getenv("TINY")
+  TRANSFER = getenv("TRANSFER")
+  if TINY:
+    model = TinyConvNet(classes)
+  elif TRANSFER:
+    model = EfficientNet(getenv("NUM", 0), classes, has_se=True)
+    model.load_from_pretrained()
+  else:
+    model = EfficientNet(getenv("NUM", 0), classes, has_se=False)
+
+  parameters = get_parameters(model)
+  print("parameter count", len(parameters))
+  optimizer = optim.Adam(parameters, lr=0.001)
+
+  BS, steps = getenv("BS", 64 if TINY else 16), getenv("STEPS", 2048)
+  print(f"training with batch size {BS} for {steps} steps")
+
+  if IMAGENET:
+    from extra.datasets.imagenet import fetch_batch
+    def loader(q):
+      while 1:
+        try:
+          q.put(fetch_batch(BS))
+        except Exception:
+          traceback.print_exc()
+    q = Queue(16)
+    for i in range(2):
+      p = Process(target=loader, args=(q,))
+      p.daemon = True
+      p.start()
+  else:
+    X_train, Y_train, _, _ = fetch_cifar()
+    X_train = X_train.reshape((-1, 3, 32, 32))
+    Y_train = Y_train.reshape((-1,))
+
+  with Tensor.train():
+    for i in (t := trange(steps)):
+      if IMAGENET:
+        X, Y = q.get(True)
+      else:
+        samp = np.random.randint(0, X_train.shape[0], size=(BS))
+        X, Y = X_train.numpy()[samp], Y_train.numpy()[samp]
+
+      st = time.time()
+      out = model.forward(Tensor(X.astype(np.float32), requires_grad=False))
+      fp_time = (time.time()-st)*1000.0
+
+      y = np.zeros((BS,classes), np.float32)
+      y[range(y.shape[0]),Y] = -classes
+      y = Tensor(y, requires_grad=False)
+      loss = out.log_softmax().mul(y).mean()
+
+      optimizer.zero_grad()
+
+      st = time.time()
+      loss.backward()
+      bp_time = (time.time()-st)*1000.0
+
+      st = time.time()
+      optimizer.step()
+      opt_time = (time.time()-st)*1000.0
+
+      st = time.time()
+      loss = loss.numpy()
+      cat = out.argmax(axis=1).numpy()
+      accuracy = (cat == Y).mean()
+      finish_time = (time.time()-st)*1000.0
+
+      # printing
+      t.set_description("loss %.2f accuracy %.2f -- %.2f + %.2f + %.2f + %.2f = %.2f" %
+        (loss, accuracy,
+        fp_time, bp_time, opt_time, finish_time,
+        fp_time + bp_time + opt_time + finish_time))
+
+      del out, y, loss
@@ -0,0 +1,46 @@
+import ast
+import numpy as np
+from PIL import Image
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import getenv, fetch
+from extra.models.vit import ViT
+"""
+fn = "gs://vit_models/augreg/Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0.npz"
+import tensorflow as tf
+with tf.io.gfile.GFile(fn, "rb") as f:
+  dat = f.read()
+  with open("cache/"+ fn.rsplit("/", 1)[1], "wb") as g:
+    g.write(dat)
+"""
+
+Tensor.training = False
+if getenv("LARGE", 0) == 1:
+  m = ViT(embed_dim=768, num_heads=12)
+else:
+  # tiny
+  m = ViT(embed_dim=192, num_heads=3)
+m.load_from_pretrained()
+
+# category labels
+lbls = ast.literal_eval(fetch("https://gist.githubusercontent.com/yrevar/942d3a0ac09ec9e5eb3a/raw/238f720ff059c1f82f368259d1ca4ffa5dd8f9f5/imagenet1000_clsidx_to_labels.txt").read_text())
+
+#url = "https://upload.wikimedia.org/wikipedia/commons/4/41/Chicken.jpg"
+url = "https://repository-images.githubusercontent.com/296744635/39ba6700-082d-11eb-98b8-cb29fb7369c0"
+
+# junk
+img = Image.open(fetch(url))
+aspect_ratio = img.size[0] / img.size[1]
+img = img.resize((int(224*max(aspect_ratio,1.0)), int(224*max(1.0/aspect_ratio,1.0))))
+img = np.array(img)
+y0,x0=(np.asarray(img.shape)[:2]-224)//2
+img = img[y0:y0+224, x0:x0+224]
+img = np.moveaxis(img, [2,0,1], [0,1,2])
+img = img.astype(np.float32)[:3].reshape(1,3,224,224)
+img /= 255.0
+img -= 0.5
+img /= 0.5
+
+out = m.forward(Tensor(img))
+outnp = out.numpy().ravel()
+choice = outnp.argmax()
+print(out.shape, choice, outnp[choice], lbls[choice])
@@ -0,0 +1,740 @@
+import json, logging, math, re, sys, time, wave, argparse, numpy as np
+from phonemizer.phonemize import default_separator, _phonemize
+from phonemizer.backend import EspeakBackend
+from phonemizer.punctuation import Punctuation
+from functools import reduce
+from pathlib import Path
+from typing import List
+from tinygrad import nn, dtypes
+from tinygrad.helpers import fetch
+from tinygrad.nn.state import torch_load
+from tinygrad.tensor import Tensor
+from tinygrad.engine.jit import TinyJit
+from unidecode import unidecode
+
+LRELU_SLOPE = 0.1
+
+class Synthesizer:
+  def __init__(self, n_vocab, spec_channels, segment_size, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, n_speakers=0, gin_channels=0, use_sdp=True, emotion_embedding=False, **kwargs):
+    self.n_vocab, self.spec_channels, self.inter_channels, self.hidden_channels, self.filter_channels, self.n_heads, self.n_layers, self.kernel_size, self.p_dropout, self.resblock, self.resblock_kernel_sizes, self.resblock_dilation_sizes, self.upsample_rates, self.upsample_initial_channel, self.upsample_kernel_sizes, self.segment_size, self.n_speakers, self.gin_channels, self.use_sdp = n_vocab, spec_channels, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, segment_size, n_speakers, gin_channels, use_sdp
+    self.enc_p = TextEncoder(n_vocab, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, emotion_embedding)
+    self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
+    self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
+    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
+    self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels) if use_sdp else DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
+    if n_speakers > 1: self.emb_g = nn.Embedding(n_speakers, gin_channels)
+  def infer(self, x, x_lengths, sid=None, noise_scale=1.0, length_scale=1, noise_scale_w=1., max_len=None, emotion_embedding=None, max_y_length_estimate_scale=None, pad_length=-1):
+    x, m_p, logs_p, x_mask = self.enc_p.forward(x.realize(), x_lengths.realize(), emotion_embedding.realize() if emotion_embedding is not None else emotion_embedding)
+    g = self.emb_g(sid.reshape(1, 1)).squeeze(1).unsqueeze(-1) if self.n_speakers > 0 else None
+    logw = self.dp.forward(x, x_mask.realize(), g=g.realize(), reverse=self.use_sdp, noise_scale=noise_scale_w if self.use_sdp else 1.0)
+    w_ceil = Tensor.ceil(logw.exp() * x_mask * length_scale)
+    y_lengths = Tensor.maximum(w_ceil.sum([1, 2]), 1).cast(dtypes.int64)
+    return self.generate(g, logs_p, m_p, max_len, max_y_length_estimate_scale, noise_scale, w_ceil, x, x_mask, y_lengths, pad_length)
+  def generate(self, g, logs_p, m_p, max_len, max_y_length_estimate_scale, noise_scale, w_ceil, x, x_mask, y_lengths, pad_length):
+    max_y_length = y_lengths.max().item() if max_y_length_estimate_scale is None else max(15, x.shape[-1]) * max_y_length_estimate_scale
+    y_mask = sequence_mask(y_lengths, max_y_length).unsqueeze(1).cast(x_mask.dtype)
+    attn_mask = x_mask.unsqueeze(2) * y_mask.unsqueeze(-1)
+    attn = generate_path(w_ceil, attn_mask)
+    m_p_2 = attn.squeeze(1).matmul(m_p.transpose(1, 2)).transpose(1, 2)        # [b, t', t], [b, t, d] -> [b, d, t']
+    logs_p_2 = attn.squeeze(1).matmul(logs_p.transpose(1, 2)).transpose(1, 2)  # [b, t', t], [b, t, d] -> [b, d, t']
+    z_p = m_p_2 + Tensor.randn(*m_p_2.shape, dtype=m_p_2.dtype) * logs_p_2.exp() * noise_scale
+    row_len = y_mask.shape[2]
+    if pad_length > -1:
+      # Pad flow forward inputs to enable JIT
+      assert pad_length > row_len, "pad length is too small"
+      y_mask = y_mask.pad(((0, 0), (0, 0), (0, pad_length - row_len))).cast(z_p.dtype)
+      # New y_mask tensor to remove sts mask
+      y_mask = Tensor(y_mask.numpy(), device=y_mask.device, dtype=y_mask.dtype, requires_grad=y_mask.requires_grad)
+      z_p = z_p.squeeze(0).pad(((0, 0), (0, pad_length - z_p.shape[2])), value=1).unsqueeze(0)
+    z = self.flow.forward(z_p.realize(), y_mask.realize(), g=g.realize(), reverse=True)
+    result_length = reduce(lambda x, y: x * y, self.dec.upsample_rates, row_len)
+    o = self.dec.forward((z * y_mask)[:, :, :max_len], g=g)[:, :, :result_length]
+    if max_y_length_estimate_scale is not None:
+      length_scaler = o.shape[-1] / max_y_length
+      o.realize()
+      real_max_y_length = y_lengths.max().numpy()
+      if real_max_y_length > max_y_length:
+        logging.warning(f"Underestimated max length by {(((real_max_y_length / max_y_length) * 100) - 100):.2f}%, recomputing inference without estimate...")
+        return self.generate(g, logs_p, m_p, max_len, None, noise_scale, w_ceil, x, x_mask, y_lengths)
+      if real_max_y_length < max_y_length:
+        overestimation = ((max_y_length / real_max_y_length) * 100) - 100
+        logging.info(f"Overestimated max length by {overestimation:.2f}%")
+        if overestimation > 10: logging.warning("Warning: max length overestimated by more than 10%")
+      o = o[:, :, :(real_max_y_length * length_scaler).astype(np.int32)]
+    return o
+
+class StochasticDurationPredictor:
+  def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
+    filter_channels = in_channels # it needs to be removed from future version.
+    self.in_channels, self.filter_channels, self.kernel_size, self.p_dropout, self.n_flows, self.gin_channels = in_channels, filter_channels, kernel_size, p_dropout, n_flows, gin_channels
+    self.log_flow, self.flows = Log(), [ElementwiseAffine(2)]
+    for _ in range(n_flows):
+      self.flows.append(ConvFlow(2, filter_channels, kernel_size, n_layers=3))
+      self.flows.append(Flip())
+    self.post_pre, self.post_proj = nn.Conv1d(1, filter_channels, 1), nn.Conv1d(filter_channels, filter_channels, 1)
+    self.post_convs = DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
+    self.post_flows = [ElementwiseAffine(2)]
+    for _ in range(4):
+      self.post_flows.append(ConvFlow(2, filter_channels, kernel_size, n_layers=3))
+      self.post_flows.append(Flip())
+    self.pre, self.proj = nn.Conv1d(in_channels, filter_channels, 1), nn.Conv1d(filter_channels, filter_channels, 1)
+    self.convs = DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
+    if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
+  @TinyJit
+  def forward(self, x: Tensor, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
+    x = self.pre(x.detach())
+    if g is not None: x = x + self.cond(g.detach())
+    x = self.convs.forward(x, x_mask)
+    x = self.proj(x) * x_mask
+    if not reverse:
+      flows = self.flows
+      assert w is not None
+      log_det_tot_q = 0
+      h_w = self.post_proj(self.post_convs.forward(self.post_pre(w), x_mask)) * x_mask
+      e_q = Tensor.randn(w.size(0), 2, w.size(2), dtype=x.dtype).to(device=x.device) * x_mask
+      z_q = e_q
+      for flow in self.post_flows:
+        z_q, log_det_q = flow.forward(z_q, x_mask, g=(x + h_w))
+        log_det_tot_q += log_det_q
+      z_u, z1 = z_q.split([1, 1], 1)
+      u = z_u.sigmoid() * x_mask
+      z0 = (w - u) * x_mask
+      log_det_tot_q += Tensor.sum((z_u.logsigmoid() + (-z_u).logsigmoid()) * x_mask, [1,2])
+      log_q = Tensor.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - log_det_tot_q
+      log_det_tot = 0
+      z0, log_det = self.log_flow.forward(z0, x_mask)
+      log_det_tot += log_det
+      z = z0.cat(z1, 1)
+      for flow in flows:
+        z, log_det = flow.forward(z, x_mask, g=x, reverse=reverse)
+        log_det_tot = log_det_tot + log_det
+      nll = Tensor.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - log_det_tot
+      return (nll + log_q).realize() # [b]
+    flows = list(reversed(self.flows))
+    flows = flows[:-2] + [flows[-1]] # remove a useless vflow
+    z = Tensor.randn(x.shape[0], 2, x.shape[2], dtype=x.dtype).to(device=x.device) * noise_scale
+    for flow in flows: z = flow.forward(z, x_mask, g=x, reverse=reverse)
+    z0, z1 = z.split([1, 1], 1)
+    return z0.realize()
+
+class DurationPredictor:
+  def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
+    self.in_channels, self.filter_channels, self.kernel_size, self.p_dropout, self.gin_channels = in_channels, filter_channels, kernel_size, p_dropout, gin_channels
+    self.conv_1, self.norm_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2), LayerNorm(filter_channels)
+    self.conv_2, self.norm_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2), LayerNorm(filter_channels)
+    self.proj = nn.Conv1d(filter_channels, 1, 1)
+    if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+  def forward(self, x: Tensor, x_mask, g=None):
+    x = x.detach()
+    if g is not None: x = x + self.cond(g.detach())
+    x = self.conv_1(x * x_mask).relu()
+    x = self.norm_1(x).dropout(self.p_dropout)
+    x = self.conv_2(x * x_mask).relu(x)
+    x = self.norm_2(x).dropout(self.p_dropout)
+    return self.proj(x * x_mask) * x_mask
+
+class TextEncoder:
+  def __init__(self, n_vocab, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, emotion_embedding):
+    self.n_vocab, self.out_channels, self.hidden_channels, self.filter_channels, self.n_heads, self.n_layers, self.kernel_size, self.p_dropout = n_vocab, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+    if n_vocab!=0:self.emb = nn.Embedding(n_vocab, hidden_channels)
+    if emotion_embedding: self.emo_proj = nn.Linear(1024, hidden_channels)
+    self.encoder = Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
+    self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+  @TinyJit
+  def forward(self, x: Tensor, x_lengths: Tensor, emotion_embedding=None):
+    if self.n_vocab!=0: x = (self.emb(x) * math.sqrt(self.hidden_channels))
+    if emotion_embedding: x = x + self.emo_proj(emotion_embedding).unsqueeze(1)
+    x = x.transpose(1, -1)  # [b, t, h] -transpose-> [b, h, t]
+    x_mask = sequence_mask(x_lengths, x.shape[2]).unsqueeze(1).cast(x.dtype)
+    x = self.encoder.forward(x * x_mask, x_mask)
+    m, logs = (self.proj(x) * x_mask).split(self.out_channels, dim=1)
+    return x.realize(), m.realize(), logs.realize(), x_mask.realize()
+
+class ResidualCouplingBlock:
+  def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, n_flows=4, gin_channels=0):
+    self.channels, self.hidden_channels, self.kernel_size, self.dilation_rate, self.n_layers, self.n_flows, self.gin_channels = channels, hidden_channels, kernel_size, dilation_rate, n_layers, n_flows, gin_channels
+    self.flows = []
+    for _ in range(n_flows):
+      self.flows.append(ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
+      self.flows.append(Flip())
+  @TinyJit
+  def forward(self, x, x_mask, g=None, reverse=False):
+    for flow in reversed(self.flows) if reverse else self.flows: x = flow.forward(x, x_mask, g=g, reverse=reverse)
+    return x.realize()
+
+class PosteriorEncoder:
+  def __init__(self, in_channels, out_channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0):
+    self.in_channels, self.out_channels, self.hidden_channels, self.kernel_size, self.dilation_rate, self.n_layers, self.gin_channels = in_channels, out_channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels
+    self.pre, self.proj = nn.Conv1d(in_channels, hidden_channels, 1), nn.Conv1d(hidden_channels, out_channels * 2, 1)
+    self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
+  def forward(self, x, x_lengths, g=None):
+    x_mask = sequence_mask(x_lengths, x.size(2)).unsqueeze(1).cast(x.dtype)
+    stats = self.proj(self.enc.forward(self.pre(x) * x_mask, x_mask, g=g)) * x_mask
+    m, logs = stats.split(self.out_channels, dim=1)
+    z = (m + Tensor.randn(m.shape, m.dtype) * logs.exp()) * x_mask
+    return z, m, logs, x_mask
+
+class Generator:
+  def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
+    self.num_kernels, self.num_upsamples = len(resblock_kernel_sizes), len(upsample_rates)
+    self.conv_pre = nn.Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
+    resblock = ResBlock1 if resblock == '1' else ResBlock2
+    self.ups = [nn.ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)), k, u, padding=(k-u)//2) for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes))]
+    self.resblocks = []
+    self.upsample_rates = upsample_rates
+    for i in range(len(self.ups)):
+      ch = upsample_initial_channel // (2 ** (i + 1))
+      for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+        self.resblocks.append(resblock(ch, k, d))
+    self.conv_post = nn.Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+    if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+  @TinyJit
+  def forward(self, x: Tensor, g=None):
+    x = self.conv_pre(x)
+    if g is not None:  x = x + self.cond(g)
+    for i in range(self.num_upsamples):
+      x = self.ups[i](x.leaky_relu(LRELU_SLOPE))
+      xs = sum(self.resblocks[i * self.num_kernels + j].forward(x) for j in range(self.num_kernels))
+      x = (xs / self.num_kernels).realize()
+    res = self.conv_post(x.leaky_relu()).tanh().realize()
+    return res
+
+class LayerNorm(nn.LayerNorm):
+  def __init__(self, channels, eps=1e-5): super().__init__(channels, eps, elementwise_affine=True)
+  def forward(self, x: Tensor): return self.__call__(x.transpose(1, -1)).transpose(1, -1)
+
+class WN:
+  def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
+    assert (kernel_size % 2 == 1)
+    self.hidden_channels, self.kernel_size, self.dilation_rate, self.n_layers, self.gin_channels, self.p_dropout = hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels, p_dropout
+    self.in_layers, self.res_skip_layers = [], []
+    if gin_channels != 0: self.cond_layer = nn.Conv1d(gin_channels, 2 * hidden_channels * n_layers, 1)
+    for i in range(n_layers):
+      dilation = dilation_rate ** i
+      self.in_layers.append(nn.Conv1d(hidden_channels, 2 * hidden_channels, kernel_size, dilation=dilation, padding=int((kernel_size * dilation - dilation) / 2)))
+      self.res_skip_layers.append(nn.Conv1d(hidden_channels, 2 * hidden_channels if i < n_layers - 1 else hidden_channels, 1))
+  def forward(self, x, x_mask, g=None, **kwargs):
+    output = Tensor.zeros_like(x)
+    if g is not None: g = self.cond_layer(g)
+    for i in range(self.n_layers):
+      x_in = self.in_layers[i](x)
+      if g is not None:
+        cond_offset = i * 2 * self.hidden_channels
+        g_l = g[:, cond_offset:cond_offset + 2 * self.hidden_channels, :]
+      else:
+        g_l = Tensor.zeros_like(x_in)
+      acts = fused_add_tanh_sigmoid_multiply(x_in, g_l, self.hidden_channels)
+      res_skip_acts = self.res_skip_layers[i](acts)
+      if i < self.n_layers - 1:
+        x = (x + res_skip_acts[:, :self.hidden_channels, :]) * x_mask
+        output = output + res_skip_acts[:, self.hidden_channels:, :]
+      else:
+        output = output + res_skip_acts
+    return output * x_mask
+
+class ResBlock1:
+  def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+    self.convs1 = [nn.Conv1d(channels, channels, kernel_size, 1, dilation=dilation[i], padding=get_padding(kernel_size, dilation[i])) for i in range(3)]
+    self.convs2 = [nn.Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1)) for _ in range(3)]
+  def forward(self, x: Tensor, x_mask=None):
+    for c1, c2 in zip(self.convs1, self.convs2):
+      xt = x.leaky_relu(LRELU_SLOPE)
+      xt = c1(xt if x_mask is None else xt * x_mask).leaky_relu(LRELU_SLOPE)
+      x = c2(xt if x_mask is None else xt * x_mask) + x
+    return x if x_mask is None else x * x_mask
+
+class ResBlock2:
+  def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+    self.convs = [nn.Conv1d(channels, channels, kernel_size, 1, dilation=dilation[i], padding=get_padding(kernel_size, dilation[i])) for i in range(2)]
+  def forward(self, x, x_mask=None):
+    for c in self.convs:
+      xt = x.leaky_relu(LRELU_SLOPE)
+      xt = c(xt if x_mask is None else xt * x_mask)
+      x = xt + x
+    return x if x_mask is None else x * x_mask
+
+class DDSConv: # Dilated and Depth-Separable Convolution
+  def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
+    self.channels, self.kernel_size, self.n_layers, self.p_dropout = channels, kernel_size, n_layers, p_dropout
+    self.convs_sep, self.convs_1x1, self.norms_1, self.norms_2 = [], [], [], []
+    for i in range(n_layers):
+      dilation = kernel_size ** i
+      padding = (kernel_size * dilation - dilation) // 2
+      self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size, groups=channels, dilation=dilation, padding=padding))
+      self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+      self.norms_1.append(LayerNorm(channels))
+      self.norms_2.append(LayerNorm(channels))
+  def forward(self, x, x_mask, g=None):
+    if g is not None: x = x + g
+    for i in range(self.n_layers):
+      y = self.convs_sep[i](x * x_mask)
+      y = self.norms_1[i].forward(y).gelu()
+      y = self.convs_1x1[i](y)
+      y = self.norms_2[i].forward(y).gelu()
+      x = x + y.dropout(self.p_dropout)
+    return x * x_mask
+
+class ConvFlow:
+  def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
+    self.in_channels, self.filter_channels, self.kernel_size, self.n_layers, self.num_bins, self.tail_bound = in_channels, filter_channels, kernel_size, n_layers, num_bins, tail_bound
+    self.half_channels = in_channels // 2
+    self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+    self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
+    self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
+  def forward(self, x, x_mask, g=None, reverse=False):
+    x0, x1 = x.split([self.half_channels] * 2, 1)
+    h = self.proj(self.convs.forward(self.pre(x0), x_mask, g=g)) * x_mask
+    b, c, t = x0.shape
+    h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
+    un_normalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
+    un_normalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
+    un_normalized_derivatives = h[..., 2 * self.num_bins:]
+    x1, log_abs_det = piecewise_rational_quadratic_transform(x1, un_normalized_widths, un_normalized_heights, un_normalized_derivatives, inverse=reverse, tails='linear', tail_bound=self.tail_bound)
+    x = x0.cat(x1, dim=1) * x_mask
+    return x if reverse else (x, Tensor.sum(log_abs_det * x_mask, [1,2]))
+
+class ResidualCouplingLayer:
+  def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=0, gin_channels=0, mean_only=False):
+    assert channels % 2 == 0, "channels should be divisible by 2"
+    self.channels, self.hidden_channels, self.kernel_size, self.dilation_rate, self.n_layers, self.mean_only = channels, hidden_channels, kernel_size, dilation_rate, n_layers, mean_only
+    self.half_channels = channels // 2
+    self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+    self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
+    self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+  def forward(self, x, x_mask, g=None, reverse=False):
+    x0, x1 = x.split([self.half_channels] * 2, 1)
+    stats = self.post(self.enc.forward(self.pre(x0) * x_mask, x_mask, g=g)) * x_mask
+    if not self.mean_only:
+      m, logs = stats.split([self.half_channels] * 2, 1)
+    else:
+      m = stats
+      logs = Tensor.zeros_like(m)
+    if not reverse: return x0.cat((m + x1 * logs.exp() * x_mask), dim=1)
+    return x0.cat(((x1 - m) * (-logs).exp() * x_mask), dim=1)
+
+class Log:
+  def forward(self, x : Tensor, x_mask, reverse=False):
+    if not reverse:
+      y = x.maximum(1e-5).log() * x_mask
+      return y, (-y).sum([1, 2])
+    return x.exp() * x_mask
+
+class Flip:
+  def forward(self, x: Tensor, *args, reverse=False, **kwargs):
+    return x.flip([1]) if reverse else (x.flip([1]), Tensor.zeros(x.shape[0], dtype=x.dtype).to(device=x.device))
+
+class ElementwiseAffine:
+  def __init__(self, channels): self.m, self.logs = Tensor.zeros(channels, 1), Tensor.zeros(channels, 1)
+  def forward(self, x, x_mask, reverse=False, **kwargs): # x if reverse else y, logdet
+    return (x - self.m) * Tensor.exp(-self.logs) * x_mask if reverse \
+      else ((self.m + Tensor.exp(self.logs) * x) * x_mask, Tensor.sum(self.logs * x_mask, [1, 2]))
+
+class MultiHeadAttention:
+  def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
+    assert channels % n_heads == 0
+    self.channels, self.out_channels, self.n_heads, self.p_dropout, self.window_size, self.heads_share, self.block_length, self.proximal_bias, self.proximal_init = channels, out_channels, n_heads, p_dropout, window_size, heads_share, block_length, proximal_bias, proximal_init
+    self.attn, self.k_channels  = None, channels // n_heads
+    self.conv_q, self.conv_k, self.conv_v = [nn.Conv1d(channels, channels, 1) for _ in range(3)]
+    self.conv_o = nn.Conv1d(channels, out_channels, 1)
+    if window_size is not None: self.emb_rel_k, self.emb_rel_v = [Tensor.randn(1 if heads_share else n_heads, window_size * 2 + 1, self.k_channels) * (self.k_channels ** -0.5) for _ in range(2)]
+  def forward(self, x, c, attn_mask=None):
+    q, k, v = self.conv_q(x), self.conv_k(c), self.conv_v(c)
+    x, self.attn = self.attention(q, k, v, mask=attn_mask)
+    return self.conv_o(x)
+  def attention(self, query: Tensor, key: Tensor, value: Tensor, mask=None):# reshape [b, d, t] -> [b, n_h, t, d_k]
+    b, d, t_s, t_t = key.shape[0], key.shape[1], key.shape[2], query.shape[2]
+    query = query.reshape(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+    key = key.reshape(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+    value = value.reshape(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+    scores = (query / math.sqrt(self.k_channels)) @ key.transpose(-2, -1)
+    if self.window_size is not None:
+      assert t_s == t_t, "Relative attention is only available for self-attention."
+      key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+      rel_logits = self._matmul_with_relative_keys(query / math.sqrt(self.k_channels), key_relative_embeddings)
+      scores = scores + self._relative_position_to_absolute_position(rel_logits)
+    if mask is not None:
+      scores = Tensor.where(mask, scores, -1e4)
+      if self.block_length is not None:
+        assert t_s == t_t, "Local attention is only available for self-attention."
+        scores = Tensor.where(Tensor.ones_like(scores).triu(-self.block_length).tril(self.block_length), scores, -1e4)
+    p_attn = scores.softmax(axis=-1)  # [b, n_h, t_t, t_s]
+    output = p_attn.matmul(value)
+    if self.window_size is not None:
+      relative_weights = self._absolute_position_to_relative_position(p_attn)
+      value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
+      output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
+    output = output.transpose(2, 3).contiguous().reshape(b, d, t_t)  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+    return output, p_attn
+  def _matmul_with_relative_values(self, x, y): return x.matmul(y.unsqueeze(0))                 # x: [b, h, l, m], y: [h or 1, m, d], ret: [b, h, l, d]
+  def _matmul_with_relative_keys(self, x, y): return x.matmul(y.unsqueeze(0).transpose(-2, -1)) # x: [b, h, l, d], y: [h or 1, m, d], re, : [b, h, l, m]
+  def _get_relative_embeddings(self, relative_embeddings, length):
+    pad_length, slice_start_position = max(length - (self.window_size + 1), 0), max((self.window_size + 1) - length, 0)
+    padded_relative_embeddings = relative_embeddings if pad_length <= 0\
+      else relative_embeddings.pad(convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
+    return padded_relative_embeddings[:, slice_start_position:(slice_start_position + 2 * length - 1)] #used_relative_embeddings
+  def _relative_position_to_absolute_position(self, x: Tensor): # x: [b, h, l, 2*l-1] -> [b, h, l, l]
+    batch, heads, length, _ = x.shape
+    x = x.pad(convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
+    x_flat = x.reshape([batch, heads, length * 2 * length]).pad(convert_pad_shape([[0,0],[0,0],[0,length-1]]))
+    return x_flat.reshape([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
+  def _absolute_position_to_relative_position(self, x: Tensor): # x: [b, h, l, l] -> [b, h, l, 2*l-1]
+    batch, heads, length, _ = x.shape
+    x = x.pad(convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
+    x_flat = x.reshape([batch, heads, length**2 + length*(length -1)]).pad(convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+    return x_flat.reshape([batch, heads, length, 2*length])[:,:,:,1:]
+
+class FFN:
+  def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
+    self.in_channels, self.out_channels, self.filter_channels, self.kernel_size, self.p_dropout, self.activation, self.causal = in_channels, out_channels, filter_channels, kernel_size, p_dropout, activation, causal
+    self.padding = self._causal_padding if causal else self._same_padding
+    self.conv_1, self.conv_2 = nn.Conv1d(in_channels, filter_channels, kernel_size), nn.Conv1d(filter_channels, out_channels, kernel_size)
+  def forward(self, x, x_mask):
+    x = self.conv_1(self.padding(x * x_mask))
+    x = x * (1.702 * x).sigmoid() if self.activation == "gelu" else x.relu()
+    return self.conv_2(self.padding(x.dropout(self.p_dropout) * x_mask)) * x_mask
+  def _causal_padding(self, x):return x if self.kernel_size == 1 else x.pad(convert_pad_shape([[0, 0], [0, 0], [self.kernel_size - 1, 0]]))
+  def _same_padding(self, x): return x if self.kernel_size == 1 else x.pad(convert_pad_shape([[0, 0], [0, 0], [(self.kernel_size - 1) // 2, self.kernel_size // 2]]))
+
+class Encoder:
+  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
+    self.hidden_channels, self.filter_channels, self.n_heads, self.n_layers, self.kernel_size, self.p_dropout, self.window_size = hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, window_size
+    self.attn_layers, self.norm_layers_1, self.ffn_layers, self.norm_layers_2 = [], [], [], []
+    for _ in range(n_layers):
+      self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
+      self.norm_layers_2.append(LayerNorm(hidden_channels))
+  def forward(self, x, x_mask):
+    attn_mask, x = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1), x * x_mask
+    for i in range(self.n_layers):
+      y = self.attn_layers[i].forward(x, x, attn_mask).dropout(self.p_dropout)
+      x = self.norm_layers_1[i].forward(x + y)
+      y = self.ffn_layers[i].forward(x, x_mask).dropout(self.p_dropout)
+      x = self.norm_layers_2[i].forward(x + y)
+    return x * x_mask
+
+DEFAULT_MIN_BIN_WIDTH, DEFAULT_MIN_BIN_HEIGHT, DEFAULT_MIN_DERIVATIVE = 1e-3, 1e-3, 1e-3
+def piecewise_rational_quadratic_transform(inputs, un_normalized_widths, un_normalized_heights, un_normalized_derivatives, inverse=False, tails=None, tail_bound=1., min_bin_width=DEFAULT_MIN_BIN_WIDTH, min_bin_height=DEFAULT_MIN_BIN_HEIGHT, min_derivative=DEFAULT_MIN_DERIVATIVE):
+  if tails is None: spline_fn, spline_kwargs = rational_quadratic_spline, {}
+  else: spline_fn, spline_kwargs = unconstrained_rational_quadratic_spline, {'tails': tails, 'tail_bound': tail_bound}
+  return spline_fn(inputs=inputs, un_normalized_widths=un_normalized_widths, un_normalized_heights=un_normalized_heights, un_normalized_derivatives=un_normalized_derivatives, inverse=inverse, min_bin_width=min_bin_width, min_bin_height=min_bin_height, min_derivative=min_derivative, **spline_kwargs)
+def unconstrained_rational_quadratic_spline(inputs, un_normalized_widths, un_normalized_heights, un_normalized_derivatives, inverse=False, tails='linear', tail_bound=1., min_bin_width=DEFAULT_MIN_BIN_WIDTH, min_bin_height=DEFAULT_MIN_BIN_HEIGHT, min_derivative=DEFAULT_MIN_DERIVATIVE):
+  if not tails == 'linear': raise RuntimeError('{} tails are not implemented.'.format(tails))
+  constant = np.log(np.exp(1 - min_derivative) - 1).item()
+  un_normalized_derivatives = cat_lr(un_normalized_derivatives, constant, constant)
+  output, log_abs_det = rational_quadratic_spline(inputs=inputs.squeeze(dim=0).squeeze(dim=0), unnormalized_widths=un_normalized_widths.squeeze(dim=0).squeeze(dim=0), unnormalized_heights=un_normalized_heights.squeeze(dim=0).squeeze(dim=0), unnormalized_derivatives=un_normalized_derivatives.squeeze(dim=0).squeeze(dim=0), inverse=inverse, left=-tail_bound, right=tail_bound, bottom=-tail_bound, top=tail_bound, min_bin_width=min_bin_width, min_bin_height=min_bin_height, min_derivative=min_derivative)
+  return output.unsqueeze(dim=0).unsqueeze(dim=0), log_abs_det.unsqueeze(dim=0).unsqueeze(dim=0)
+def rational_quadratic_spline(inputs: Tensor, unnormalized_widths: Tensor, unnormalized_heights: Tensor, unnormalized_derivatives: Tensor, inverse=False, left=0., right=1., bottom=0., top=1., min_bin_width=DEFAULT_MIN_BIN_WIDTH, min_bin_height=DEFAULT_MIN_BIN_HEIGHT, min_derivative=DEFAULT_MIN_DERIVATIVE):
+  num_bins = unnormalized_widths.shape[-1]
+  if min_bin_width * num_bins > 1.0: raise ValueError('Minimal bin width too large for the number of bins')
+  if min_bin_height * num_bins > 1.0: raise ValueError('Minimal bin height too large for the number of bins')
+  widths = min_bin_width + (1 - min_bin_width * num_bins) * unnormalized_widths.softmax(axis=-1)
+  cum_widths = cat_lr(((right - left) * widths[..., :-1].cumsum(axis=1) + left), left, right + 1e-6 if not inverse else right)
+  widths = cum_widths[..., 1:] - cum_widths[..., :-1]
+  derivatives = min_derivative + (unnormalized_derivatives.exp()+1).log()
+  heights = min_bin_height + (1 - min_bin_height * num_bins) * unnormalized_heights.softmax(axis=-1)
+  cum_heights = cat_lr(((top - bottom) * heights[..., :-1].cumsum(axis=1) + bottom), bottom, top + 1e-6 if inverse else top)
+  heights = cum_heights[..., 1:] - cum_heights[..., :-1]
+  bin_idx = ((inputs[..., None] >= (cum_heights if inverse else cum_widths)).sum(axis=-1) - 1)[..., None]
+  input_cum_widths = gather(cum_widths, bin_idx, axis=-1)[..., 0]
+  input_bin_widths = gather(widths, bin_idx, axis=-1)[..., 0]
+  input_cum_heights = gather(cum_heights, bin_idx, axis=-1)[..., 0]
+  input_delta = gather(heights / widths, bin_idx, axis=-1)[..., 0]
+  input_derivatives = gather(derivatives, bin_idx, axis=-1)[..., 0]
+  input_derivatives_plus_one = gather(derivatives[..., 1:], bin_idx, axis=-1)[..., 0]
+  input_heights = gather(heights, bin_idx, axis=-1)[..., 0]
+  if inverse:
+    a = ((inputs - input_cum_heights) * (input_derivatives + input_derivatives_plus_one - 2 * input_delta) + input_heights * (input_delta - input_derivatives))
+    b = (input_heights * input_derivatives - (inputs - input_cum_heights) * (input_derivatives + input_derivatives_plus_one - 2 * input_delta))
+    c = - input_delta * (inputs - input_cum_heights)
+    discriminant = b.square() - 4 * a * c
+    # assert (discriminant.numpy() >= 0).all()
+    root = (2 * c) / (-b - discriminant.sqrt())
+    theta_one_minus_theta = root * (1 - root)
+    denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta) * theta_one_minus_theta)
+    derivative_numerator = input_delta.square() * (input_derivatives_plus_one * root.square() + 2 * input_delta * theta_one_minus_theta + input_derivatives * (1 - root).square())
+    return root * input_bin_widths + input_cum_widths, -(derivative_numerator.log() - 2 * denominator.log())
+  theta = (inputs - input_cum_widths) / input_bin_widths
+  theta_one_minus_theta = theta * (1 - theta)
+  numerator = input_heights * (input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta)
+  denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta) * theta_one_minus_theta)
+  derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * theta.pow(2) + 2 * input_delta * theta_one_minus_theta + input_derivatives * (1 - theta).pow(2))
+  return input_cum_heights + numerator / denominator, derivative_numerator.log() - 2 * denominator.log()
+
+def sequence_mask(length: Tensor, max_length): return Tensor.arange(max_length, dtype=length.dtype, device=length.device).unsqueeze(0) < length.unsqueeze(1)
+def generate_path(duration: Tensor, mask: Tensor):  # duration: [b, 1, t_x], mask: [b, 1, t_y, t_x]
+  b, _, t_y, t_x = mask.shape
+  path = sequence_mask(duration.cumsum(axis=2).reshape(b * t_x), t_y).cast(mask.dtype).reshape(b, t_x, t_y)
+  path = path - path.pad(convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+  return path.unsqueeze(1).transpose(2, 3) * mask
+def fused_add_tanh_sigmoid_multiply(input_a: Tensor, input_b: Tensor, n_channels: int):
+  n_channels_int, in_act = n_channels, input_a + input_b
+  t_act, s_act = in_act[:, :n_channels_int, :].tanh(), in_act[:, n_channels_int:, :].sigmoid()
+  return t_act * s_act
+
+def cat_lr(t, left, right): return Tensor.full(get_shape(t), left).cat(t, dim=-1).cat(Tensor.full(get_shape(t), right), dim=-1)
+def get_shape(tensor):
+  (shape := list(tensor.shape))[-1] = 1
+  return tuple(shape)
+def convert_pad_shape(pad_shape): return tuple(tuple(x) for x in pad_shape)
+def get_padding(kernel_size, dilation=1): return int((kernel_size*dilation - dilation)/2)
+
+def gather(x, indices, axis):
+  indices = (indices < 0).where(indices + x.shape[axis], indices).transpose(0, axis)
+  permute_args = list(range(x.ndim))
+  permute_args[0], permute_args[axis] = permute_args[axis], permute_args[0]
+  permute_args.append(permute_args.pop(0))
+  x = x.permute(*permute_args)
+  reshape_arg = [1] * x.ndim + [x.shape[-1]]
+  return ((indices.unsqueeze(indices.ndim).expand(*indices.shape, x.shape[-1]) ==
+           Tensor.arange(x.shape[-1]).reshape(*reshape_arg).expand(*indices.shape, x.shape[-1])) * x).sum(indices.ndim).transpose(0, axis)
+
+def norm_except_dim(v, dim):
+  if dim == -1: return np.linalg.norm(v)
+  if dim == 0:
+    (output_shape := [1] * v.ndim)[0] = v.shape[0]
+    return np.linalg.norm(v.reshape(v.shape[0], -1), axis=1).reshape(output_shape)
+  if dim == v.ndim - 1:
+    (output_shape := [1] * v.ndim)[-1] = v.shape[-1]
+    return np.linalg.norm(v.reshape(-1, v.shape[-1]), axis=0).reshape(output_shape)
+  transposed_v = np.transpose(v, (dim,) + tuple(i for i in range(v.ndim) if i != dim))
+  return np.transpose(norm_except_dim(transposed_v, 0), (dim,) + tuple(i for i in range(v.ndim) if i != dim))
+def weight_norm(v: Tensor, g: Tensor, dim):
+  v, g = v.numpy(), g.numpy()
+  return Tensor(v * (g / norm_except_dim(v, dim)))
+
+# HPARAMS LOADING
+def get_hparams_from_file(path):
+  with open(path, "r") as f:
+    data = f.read()
+  return HParams(**json.loads(data))
+class HParams:
+  def __init__(self, **kwargs):
+    for k, v in kwargs.items(): self[k] = v if type(v) != dict else HParams(**v)
+  def keys(self): return self.__dict__.keys()
+  def items(self): return self.__dict__.items()
+  def values(self): return self.__dict__.values()
+  def __len__(self): return len(self.__dict__)
+  def __getitem__(self, key): return getattr(self, key)
+  def __setitem__(self, key, value): return setattr(self, key, value)
+  def __contains__(self, key): return key in self.__dict__
+  def __repr__(self): return self.__dict__.__repr__()
+
+# MODEL LOADING
+def load_model(symbols, hps, model) -> Synthesizer:
+  net_g = Synthesizer(len(symbols), hps.data.filter_length // 2 + 1, hps.train.segment_size // hps.data.hop_length, n_speakers = hps.data.n_speakers, **hps.model)
+  _ = load_checkpoint(fetch(model[1]), net_g, None)
+  return net_g
+def load_checkpoint(checkpoint_path, model: Synthesizer, optimizer=None, skip_list=[]):
+  assert Path(checkpoint_path).is_file()
+  start_time = time.time()
+  checkpoint_dict = torch_load(checkpoint_path)
+  iteration, learning_rate = checkpoint_dict['iteration'], checkpoint_dict['learning_rate']
+  if optimizer: optimizer.load_state_dict(checkpoint_dict['optimizer'])
+  saved_state_dict = checkpoint_dict['model']
+  weight_g, weight_v, parent = None, None, None
+  for key, v in saved_state_dict.items():
+    if any(layer in key for layer in skip_list): continue
+    try:
+      obj, skip = model, False
+      for k in key.split('.'):
+        if k.isnumeric(): obj = obj[int(k)]
+        elif isinstance(obj, dict): obj = obj[k]
+        else:
+          if isinstance(obj, (LayerNorm, nn.LayerNorm)) and k in ["gamma", "beta"]:
+            k = "weight" if k == "gamma" else "bias"
+          elif k in ["weight_g", "weight_v"]:
+            parent, skip = obj, True
+            if k == "weight_g": weight_g = v
+            else: weight_v = v
+          if not skip: obj = getattr(obj, k)
+      if weight_g is not None and weight_v is not None:
+        setattr(obj, "weight_g", weight_g.numpy())
+        setattr(obj, "weight_v", weight_v.numpy())
+        obj, v = getattr(parent, "weight"), weight_norm(weight_v, weight_g, 0)
+        weight_g, weight_v, parent, skip = None, None, None, False
+      if not skip and obj.shape == v.shape: obj.assign(v.to(obj.device))
+      elif not skip: logging.error(f"MISMATCH SHAPE IN {key}, {obj.shape} {v.shape}")
+    except Exception as e: raise e
+  logging.info(f"Loaded checkpoint '{checkpoint_path}' (iteration {iteration}) in {time.time() - start_time:.4f}s")
+  return model, optimizer, learning_rate, iteration
+
+# Used for cleaning input text and mapping to symbols
+class TextMapper: # Based on https://github.com/keithito/tacotron
+  def __init__(self, symbols, apply_cleaners=True):
+    self.apply_cleaners, self.symbols, self._inflect = apply_cleaners, symbols, None
+    self._symbol_to_id, _id_to_symbol = {s: i for i, s in enumerate(symbols)}, {i: s for i, s in enumerate(symbols)}
+    self._whitespace_re, self._abbreviations = re.compile(r'\s+'), [(re.compile('\\b%s\\.' % x[0], re.IGNORECASE), x[1]) for x in [('mrs', 'misess'), ('mr', 'mister'), ('dr', 'doctor'), ('st', 'saint'), ('co', 'company'), ('jr', 'junior'), ('maj', 'major'), ('gen', 'general'), ('drs', 'doctors'), ('rev', 'reverend'), ('lt', 'lieutenant'), ('hon', 'honorable'), ('sgt', 'sergeant'), ('capt', 'captain'), ('esq', 'esquire'), ('ltd', 'limited'), ('col', 'colonel'), ('ft', 'fort'), ]]
+    self.phonemizer = EspeakBackend(
+        language="en-us", punctuation_marks=Punctuation.default_marks(), preserve_punctuation=True, with_stress=True,
+    )
+  def text_to_sequence(self, text, cleaner_names):
+    if self.apply_cleaners:
+      for name in cleaner_names:
+        cleaner = getattr(self, name)
+        if not cleaner: raise ModuleNotFoundError('Unknown cleaner: %s' % name)
+        text = cleaner(text)
+    else: text = text.strip()
+    return [self._symbol_to_id[symbol] for symbol in text]
+  def get_text(self, text, add_blank=False, cleaners=('english_cleaners2',)):
+    text_norm = self.text_to_sequence(text, cleaners)
+    return Tensor(self.intersperse(text_norm, 0) if add_blank else text_norm, dtype=dtypes.int64)
+  def intersperse(self, lst, item):
+    (result := [item] * (len(lst) * 2 + 1))[1::2] = lst
+    return result
+  def phonemize(self, text, strip=True): return _phonemize(self.phonemizer, text, default_separator, strip, 1, False, False)
+  def filter_oov(self, text): return "".join(list(filter(lambda x: x in self._symbol_to_id, text)))
+  def base_english_cleaners(self, text): return self.collapse_whitespace(self.phonemize(self.expand_abbreviations(unidecode(text.lower()))))
+  def english_cleaners2(self, text): return self.base_english_cleaners(text)
+  def transliteration_cleaners(self, text): return self.collapse_whitespace(unidecode(text.lower()))
+  def cjke_cleaners(self, text): return re.sub(r'([^\.,!\?\-…~])$', r'\1.', re.sub(r'\s+$', '', self.english_to_ipa2(text).replace('ɑ', 'a').replace('ɔ', 'o').replace('ɛ', 'e').replace('ɪ', 'i').replace('ʊ', 'u')))
+  def cjke_cleaners2(self, text): return re.sub(r'([^\.,!\?\-…~])$', r'\1.', re.sub(r'\s+$', '', self.english_to_ipa2(text)))
+  def cjks_cleaners(self, text): return re.sub(r'([^\.,!\?\-…~])$', r'\1.', re.sub(r'\s+$', '', self.english_to_lazy_ipa(text)))
+  def english_to_ipa2(self, text):
+    _ipa_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [ ('r', 'ɹ'), ('ʤ', 'dʒ'), ('ʧ', 'tʃ')]]
+    return reduce(lambda t, rx: re.sub(rx[0], rx[1], t), _ipa_to_ipa2, self.mark_dark_l(self.english_to_ipa(text))).replace('...', '…')
+  def mark_dark_l(self, text): return re.sub(r'l([^aeiouæɑɔəɛɪʊ ]*(?: |$))', lambda x: 'ɫ' + x.group(1), text)
+  def english_to_ipa(self, text):
+    import eng_to_ipa as ipa
+    return self.collapse_whitespace(ipa.convert(self.normalize_numbers(self.expand_abbreviations(unidecode(text).lower()))))
+  def english_to_lazy_ipa(self, text):
+    _lazy_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [('r', 'ɹ'), ('æ', 'e'), ('ɑ', 'a'), ('ɔ', 'o'), ('ð', 'z'), ('θ', 's'), ('ɛ', 'e'), ('ɪ', 'i'), ('ʊ', 'u'), ('ʒ', 'ʥ'), ('ʤ', 'ʥ'), ('ˈ', '↓')]]
+    return reduce(lambda t, rx: re.sub(rx[0], rx[1], t), _lazy_ipa, self.english_to_ipa(text))
+  def expand_abbreviations(self, text): return reduce(lambda t, abbr: re.sub(abbr[0], abbr[1], t), self._abbreviations, text)
+  def collapse_whitespace(self, text): return re.sub(self._whitespace_re, ' ', text)
+  def normalize_numbers(self, text):
+    import inflect
+    self._inflect = inflect.engine()
+    text = re.sub(re.compile(r'([0-9][0-9\,]+[0-9])'), self._remove_commas, text)
+    text = re.sub(re.compile(r'£([0-9\,]*[0-9]+)'), r'\1 pounds', text)
+    text = re.sub(re.compile(r'\$([0-9\.\,]*[0-9]+)'), self._expand_dollars, text)
+    text = re.sub(re.compile(r'([0-9]+\.[0-9]+)'), self._expand_decimal_point, text)
+    text = re.sub(re.compile(r'[0-9]+(st|nd|rd|th)'), self._expand_ordinal, text)
+    text = re.sub(re.compile(r'[0-9]+'), self._expand_number, text)
+    return text
+  def _remove_commas(self, m): return m.group(1).replace(',', '') # george won't like this
+  def _expand_dollars(self, m):
+    match = m.group(1)
+    parts = match.split('.')
+    if len(parts) > 2: return match + ' dollars'  # Unexpected format
+    dollars, cents = int(parts[0]) if parts[0] else 0, int(parts[1]) if len(parts) > 1 and parts[1] else 0
+    if dollars and cents: return '%s %s, %s %s' % (dollars, 'dollar' if dollars == 1 else 'dollars', cents, 'cent' if cents == 1 else 'cents')
+    if dollars: return '%s %s' % (dollars, 'dollar' if dollars == 1 else 'dollars')
+    if cents: return '%s %s' % (cents, 'cent' if cents == 1 else 'cents')
+    return 'zero dollars'
+  def _expand_decimal_point(self, m): return m.group(1).replace('.', ' point ')
+  def _expand_ordinal(self, m): return self._inflect.number_to_words(m.group(0))
+  def _expand_number(self, _inflect, m):
+    num = int(m.group(0))
+    if 1000 < num < 3000:
+      if num == 2000: return 'two thousand'
+      if 2000 < num < 2010: return 'two thousand ' + self._inflect.number_to_words(num % 100)
+      if num % 100 == 0: return self._inflect.number_to_words(num // 100) + ' hundred'
+      return _inflect.number_to_words(num, andword='', zero='oh', group=2).replace(', ', ' ')
+    return self._inflect.number_to_words(num, andword='')
+
+#########################################################################################
+# PAPER: https://arxiv.org/abs/2106.06103
+# CODE: https://github.com/jaywalnut310/vits/tree/main
+#########################################################################################
+# INSTALLATION: this is based on default config, dependencies are for preprocessing.
+# vctk, ljs                      | pip3 install unidecode phonemizer          | phonemizer requires [eSpeak](https://espeak.sourceforge.net) backend to be installed on your system
+# mmts-tts                       | pip3 install unidecode                     |
+# uma_trilingual, cjks, voistock | pip3 install unidecode inflect eng_to_ipa  |
+#########################################################################################
+# Some good speakers to try out, there may be much better ones, I only tried out a few:
+# male vctk 1  | --model_to_use vctk --speaker_id 2
+# male vctk 2  | --model_to_use vctk --speaker_id 6
+# anime lady 1 | --model_to_use uma_trilingual --speaker_id 36
+# anime lady 2 | --model_to_use uma_trilingual --speaker_id 121
+#########################################################################################
+VITS_PATH = Path(__file__).parents[1] / "weights/VITS/"
+MODELS = { # config_url, weights_url
+  "ljs": ("https://raw.githubusercontent.com/jaywalnut310/vits/main/configs/ljs_base.json", "https://drive.google.com/uc?export=download&id=1q86w74Ygw2hNzYP9cWkeClGT5X25PvBT&confirm=t"),
+  "vctk": ("https://huggingface.co/csukuangfj/vits-vctk/resolve/main/vctk_base.json", "https://huggingface.co/csukuangfj/vits-vctk/resolve/main/pretrained_vctk.pth"),
+  "mmts-tts": ("https://huggingface.co/facebook/mms-tts/raw/main/full_models/eng/config.json", "https://huggingface.co/facebook/mms-tts/resolve/main/full_models/eng/G_100000.pth"),
+  "uma_trilingual": ("https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/raw/main/configs/uma_trilingual.json", "https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/resolve/main/pretrained_models/G_trilingual.pth"),
+  "cjks": ("https://huggingface.co/spaces/skytnt/moe-tts/resolve/main/saved_model/14/config.json", "https://huggingface.co/spaces/skytnt/moe-tts/resolve/main/saved_model/14/model.pth"),
+  "voistock": ("https://huggingface.co/spaces/skytnt/moe-tts/resolve/main/saved_model/15/config.json", "https://huggingface.co/spaces/skytnt/moe-tts/resolve/main/saved_model/15/model.pth"),
+}
+Y_LENGTH_ESTIMATE_SCALARS = {"ljs": 2.8, "vctk": 1.74, "mmts-tts": 1.9, "uma_trilingual": 2.3, "cjks": 3.3, "voistock": 3.1}
+if __name__ == '__main__':
+  logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--model_to_use", default="vctk", help="Specify the model to use. Default is 'vctk'.")
+  parser.add_argument("--speaker_id", type=int, default=6, help="Specify the speaker ID. Default is 6.")
+  parser.add_argument("--out_path", default=None, help="Specify the full output path. Overrides the --out_dir and --name parameter.")
+  parser.add_argument("--out_dir", default=str(Path(__file__).parents[1] / "temp"), help="Specify the output path.")
+  parser.add_argument("--base_name", default="test", help="Specify the base of the output file name. Default is 'test'.")
+  parser.add_argument("--text_to_synthesize", default="""Hello person. If the code you are contributing isn't some of the highest quality code you've written in your life, either put in the effort to make it great, or don't bother.""", help="Specify the text to synthesize. Default is a greeting message.")
+  parser.add_argument("--noise_scale", type=float, default=0.667, help="Specify the noise scale. Default is 0.667.")
+  parser.add_argument("--noise_scale_w", type=float, default=0.8, help="Specify the noise scale w. Default is 0.8.")
+  parser.add_argument("--length_scale", type=float, default=1, help="Specify the length scale. Default is 1.")
+  parser.add_argument("--seed", type=int, default=1337, help="Specify the seed (set to None if no seed). Default is 1337.")
+  parser.add_argument("--num_channels", type=int, default=1, help="Specify the number of audio output channels. Default is 1.")
+  parser.add_argument("--sample_width", type=int, default=2, help="Specify the number of bytes per sample, adjust if necessary. Default is 2.")
+  parser.add_argument("--emotion_path", type=str, default=None, help="Specify the path to emotion reference.")
+  parser.add_argument("--estimate_max_y_length", type=str, default=False, help="If true, overestimate the output length and then trim it to the correct length, to prevent premature realization, much more performant for larger inputs, for smaller inputs not so much. Default is False.")
+  args = parser.parse_args()
+
+  model_config = MODELS[args.model_to_use]
+
+  # Load the hyperparameters from the config file.
+  hps = get_hparams_from_file(fetch(model_config[0]))
+
+  # If model has multiple speakers, validate speaker id and retrieve name if available.
+  model_has_multiple_speakers = hps.data.n_speakers > 0
+  if model_has_multiple_speakers:
+    logging.info(f"Model has {hps.data.n_speakers} speakers")
+    if args.speaker_id >= hps.data.n_speakers: raise ValueError(f"Speaker ID {args.speaker_id} is invalid for this model.")
+    speaker_name = "?"
+    if hps.__contains__("speakers"): # maps speaker ids to names
+      speakers = hps.speakers
+      if isinstance(speakers, List): speakers = {speaker: i for i, speaker in enumerate(speakers)}
+      speaker_name = next((key for key, value in speakers.items() if value == args.speaker_id), None)
+    logging.info(f"You selected speaker {args.speaker_id} (name: {speaker_name})")
+
+  # Load emotions if any. TODO: find an english model with emotions, this is untested atm.
+  emotion_embedding = None
+  if args.emotion_path is not None:
+    if args.emotion_path.endswith(".npy"): emotion_embedding = Tensor(np.load(args.emotion_path), dtype=dtypes.int64).unsqueeze(0)
+    else: raise ValueError("Emotion path must be a .npy file.")
+
+  # Load symbols, instantiate TextMapper and clean the text.
+  if hps.__contains__("symbols"): symbols = hps.symbols
+  elif args.model_to_use == "mmts-tts": symbols = [x.replace("\n", "") for x in fetch("https://huggingface.co/facebook/mms-tts/raw/main/full_models/eng/vocab.txt").open(encoding="utf-8").readlines()]
+  else: symbols = ['_'] + list(';:,.!?¡¿—…"«»“” ') + list('ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz') + list("ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ")
+  text_mapper = TextMapper(apply_cleaners=True, symbols=symbols)
+
+  # Load the model.
+  if args.seed is not None:
+    Tensor.manual_seed(args.seed)
+    np.random.seed(args.seed)
+  net_g = load_model(text_mapper.symbols, hps, model_config)
+  logging.debug(f"Loaded model with hps: {hps}")
+
+  # Convert the input text to a tensor.
+  text_to_synthesize = args.text_to_synthesize
+  if args.model_to_use == "mmts-tts": text_to_synthesize = text_mapper.filter_oov(text_to_synthesize.lower())
+  stn_tst = text_mapper.get_text(text_to_synthesize, hps.data.add_blank, hps.data.text_cleaners)
+  logging.debug(f"Converted input text to tensor \"{text_to_synthesize}\" -> Tensor({stn_tst.shape}): {stn_tst.numpy()}")
+  x_tst, x_tst_lengths = stn_tst.unsqueeze(0), Tensor([stn_tst.shape[0]], dtype=dtypes.int64)
+  sid = Tensor([args.speaker_id], dtype=dtypes.int64) if model_has_multiple_speakers else None
+
+  # Perform inference.
+  start_time = time.time()
+  audio_tensor = net_g.infer(x_tst, x_tst_lengths, sid, args.noise_scale, args.length_scale, args.noise_scale_w, emotion_embedding=emotion_embedding,
+                             max_y_length_estimate_scale=Y_LENGTH_ESTIMATE_SCALARS[args.model_to_use] if args.estimate_max_y_length else None)[0, 0].realize()
+  logging.info(f"Inference took {(time.time() - start_time):.2f}s")
+
+  # Save the audio output.
+  audio_data = (np.clip(audio_tensor.numpy(), -1.0, 1.0) * 32767).astype(np.int16)
+  out_path = Path(args.out_path or Path(args.out_dir)/f"{args.model_to_use}{f'_sid_{args.speaker_id}' if model_has_multiple_speakers else ''}_{args.base_name}.wav")
+  out_path.parent.mkdir(parents=True, exist_ok=True)
+  with wave.open(str(out_path), 'wb') as wav_file:
+    wav_file.setnchannels(args.num_channels)
+    wav_file.setsampwidth(args.sample_width)
+    wav_file.setframerate(hps.data.sampling_rate)
+    wav_file.setnframes(len(audio_data))
+    wav_file.writeframes(audio_data.tobytes())
+  logging.info(f"Saved audio output to {out_path}")
@@ -93,7 +93,7 @@ if __name__ == "__main__":
    forward: Any = None

  sub_steps = [
-    Step(name = "textModel", input = [Tensor.randint(1, 77, low=0, high=49408, dtype=dtypes.int32)], forward = model.cond_stage_model.transformer.text_model),
+    Step(name = "textModel", input = [Tensor.randn(1, 77)], forward = model.cond_stage_model.transformer.text_model),
    Step(name = "diffusor", input = [Tensor.randn(1, 77, 768), Tensor.randn(1, 77, 768), Tensor.randn(1,4,64,64), Tensor.rand(1), Tensor.randn(1), Tensor.randn(1), Tensor.randn(1)], forward = model),
    Step(name = "decoder", input = [Tensor.randn(1,4,64,64)], forward = model.decode),
    Step(name = "f16tof32", input = [Tensor.randn(2097120, dtype=dtypes.uint32)], forward = u32_to_f16)
@@ -3,11 +3,10 @@
 import sys, base64, multiprocessing, itertools, collections
 from typing import Optional, Union, Literal, List

-from tinygrad import Tensor, TinyJit, Variable, nn, dtypes
+from tinygrad import Tensor, TinyJit, Variable, nn
 from tinygrad.nn.state import torch_load, load_state_dict
 from tinygrad.helpers import getenv, fetch

-from examples.audio_helpers import mel
 import numpy as np
 import librosa

@@ -160,7 +159,7 @@ def prep_audio(waveforms: List[np.ndarray], batch_size: int, truncate=False) ->

  stft = librosa.stft(waveforms, n_fft=N_FFT, hop_length=HOP_LENGTH, window='hann', dtype=np.csingle)
  magnitudes = np.absolute(stft[..., :-1]) ** 2
-  mel_spec = mel(sr=RATE, n_fft=N_FFT, n_mels=N_MELS).numpy() @ magnitudes
+  mel_spec = librosa.filters.mel(sr=RATE, n_fft=N_FFT, n_mels=N_MELS) @ magnitudes

  log_spec = np.log10(np.clip(mel_spec, 1e-10, None))
  log_spec = np.maximum(log_spec, log_spec.max((1,2), keepdims=True) - 8.0)
@@ -245,16 +244,15 @@ def transcribe_waveform(model: Whisper, enc, waveforms, truncate=False):

  log_spec = prep_audio(waveforms, model.batch_size, truncate)
  nsample = model.decoder.max_tokens_to_sample
-  nctx = model.decoder.max_self_attn_cache_len

  def inferloop(ctx: Union[np.ndarray, List[np.ndarray]], encoded_audio):
    pos, next_tokens = 0, ctx
-    for i in range(nsample):
-      next_tokens = model.decoder(Tensor(next_tokens, dtype=dtypes.int32), pos, encoded_audio)[:, -1].argmax(axis=-1).numpy().astype(np.int32).reshape(-1, 1)
+    for i in range((nsample-len(start_tokens))*2):
+      next_tokens = model.decoder(Tensor(next_tokens), pos, encoded_audio)[:, -1].argmax(axis=-1).numpy().astype(np.int32).reshape(-1, 1)
      next_tokens[ctx[:, -1] == eot] = eot
      ctx = np.concatenate((ctx, next_tokens), axis=1)
      pos = ctx.shape[-1] - 1
-      if (next_tokens == eot).all() or pos == nctx: break
+      if (next_tokens == eot).all(): break
    return ctx

  def gettexttoks(line): return [tok for tok in line if tok < eot or tok > enc._special_tokens["<|notimestamps|>"]][-nsample+len(start_tokens):]
@@ -26,13 +26,11 @@ def color_temp(temp):
 def color_voltage(voltage): return colored(f"{voltage/1000:>5.3f}V", "cyan")

 def draw_bar(percentage, width=40, fill='|', empty=' ', opt_text='', color='cyan'):
-  percentage = 0.0 if percentage != percentage else percentage  # NaN guard
-  percentage = max(0.0, min(1.0, float(percentage)))
  filled_width = int(width * percentage)
  if not opt_text: opt_text = f'{percentage*100:.1f}%'

  bar = fill * filled_width + empty * (width - filled_width)
-  if opt_text and len(opt_text) <= len(bar): bar = (bar[:-len(opt_text)] + opt_text)
+  bar = (bar[:-len(opt_text)] + opt_text) if opt_text else bar
  bar = colored(bar[:filled_width], color) + bar[filled_width:]
  return f'[{bar}]'

@@ -65,7 +63,7 @@ def get_bar0_size(pcibus):
 class AMSMI(AMDev):
  def __init__(self, pcibus, vram_bar:MMIOInterface, doorbell_bar:MMIOInterface, mmio_bar:MMIOInterface):
    self.pcibus = pcibus
-    self.vram, self.doorbell64, self.mmio = vram_bar, doorbell_bar, mmio_bar
+    self.vram, self.doorbell64, self.mmio, self.dma_regions = vram_bar, doorbell_bar, mmio_bar, None
    self.pci_state = self.read_pci_state()
    if self.pci_state == "D0": self._init_from_d0()

@@ -90,28 +88,13 @@ class SMICtx:
    self.opened_pci_resources = {}
    self.prev_lines_cnt = 0
    self.prev_terminal_width = 0
-    self.prev_terminal_height = 0

-    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:", "Processing accelerators:"]
+    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:"]
    lspci = subprocess.check_output(["lspci"]).decode("utf-8").splitlines()
    self.lspci = {l.split()[0]: l.split(" ", 1)[1] for l in lspci}
    for k,v in self.lspci.items():
      for part in remove_parts: self.lspci[k] = self.lspci[k].replace(part, "").strip().rstrip()

-  def _smuq10_round(self, v:int) -> int:
-    v = int(v)
-    return (v + 512) >> 10  # SMUQ10_ROUND
-
-  def _fmt_kb(self, kb:int) -> str:
-    kb = int(kb)
-    if kb < 1024: return f"{kb}KB"
-    mb = kb / 1024.0
-    if mb < 1024: return f"{mb:.1f}MB"
-    gb = mb / 1024.0
-    if gb < 1024: return f"{gb:.2f}GB"
-    tb = gb / 1024.0
-    return f"{tb:.2f}TB"
-
  def _open_am_device(self, pcibus):
    if pcibus not in self.opened_pci_resources:
      bar_fds = {bar: os.open(f"/sys/bus/pci/devices/{pcibus}/resource{bar}", os.O_RDWR | os.O_SYNC) for bar in [0, 2, 5]}
@@ -133,7 +116,6 @@ class SMICtx:
  def rescan_devs(self):
    pattern = os.path.join('/tmp', 'am_*.lock')
    for d in [f[8:-5] for f in glob.glob(pattern)]:
-      if d.startswith("usb"): continue
      if d not in self.opened_pcidevs:
        self._open_am_device(d)

@@ -149,53 +131,21 @@ class SMICtx:
        os.system('clear')
        if DEBUG >= 2: print(f"Removed AM device {d.pcibus}")

-  def collect(self):
-    tables = {}
-    for dev in self.devs:
-      match dev.ip_ver[am.MP1_HWIP]:
-        case (13,0,6): table_t = dev.smu.smu_mod.MetricsTableV0_t
-        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTable_t
-        case _: table_t = dev.smu.smu_mod.SmuMetricsExternal_t
-      tables[dev] = dev.smu.read_table(table_t, dev.smu.smu_mod.SMU_TABLE_SMU_METRICS) if dev.pci_state == "D0" else None
-    return tables
+  def collect(self): return {d: d.smu.read_metrics() if d.pci_state == "D0" else None for d in self.devs}

-  def _pick_nonzero_avg(self, vals) -> int:
-    xs = [x for x in vals if x > 0]
-    return int(sum(xs) / len(xs)) if xs else 0
-
-  def get_gfx_activity(self, dev, metrics):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
-      case _: return metrics.SmuMetrics.AverageGfxActivity
-
-  def get_mem_activity(self, dev, metrics):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
-      case _: return metrics.SmuMetrics.AverageUclkActivity
+  def get_gfx_activity(self, dev, metrics): return metrics.SmuMetrics.AverageGfxActivity
+  def get_mem_activity(self, dev, metrics): return metrics.SmuMetrics.AverageUclkActivity

  def get_temps(self, dev, metrics, compact=False):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12):
-        temps = {
-          "Hotspot": self._smuq10_round(metrics.MaxSocketTemperature),
-          "HBM": self._smuq10_round(metrics.MaxHbmTemperature),
-          "VR": self._smuq10_round(metrics.MaxVrTemperature),
-        }
-        if compact: return {k: temps[k] for k in ("Hotspot", "HBM") if temps.get(k, 0) != 0}
-        return {k: v for k, v in temps.items() if v != 0}
-      case _:
-        temps_keys = [(k, name) for k, name in dev.smu.smu_mod.TEMP_e.items()
-                      if k < dev.smu.smu_mod.TEMP_COUNT and metrics.SmuMetrics.AvgTemperature[k] != 0]
-        if compact: temps_keys = [(k, name) for k, name in temps_keys if k in (dev.smu.smu_mod.TEMP_HOTSPOT, dev.smu.smu_mod.TEMP_MEM)]
-        return {name: metrics.SmuMetrics.AvgTemperature[k] for k, name in temps_keys}
+    temps_keys = [(k, name) for k, name in dev.smu.smu_mod.c__EA_TEMP_e__enumvalues.items()
+                  if k < dev.smu.smu_mod.TEMP_COUNT and metrics.SmuMetrics.AvgTemperature[k] != 0]
+    if compact: temps_keys = [(k, name) for k, name in temps_keys if k in (dev.smu.smu_mod.TEMP_HOTSPOT, dev.smu.smu_mod.TEMP_MEM)]
+    return {name: metrics.SmuMetrics.AvgTemperature[k] for k, name in temps_keys}

  def get_voltage(self, dev, metrics, compact=False):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return {}
-      case _:
-        voltage_keys = [(k, name) for k, name in dev.smu.smu_mod.SVI_PLANE_e.items()
+    voltage_keys = [(k, name) for k, name in dev.smu.smu_mod.c__EA_SVI_PLANE_e__enumvalues.items()
                        if k < dev.smu.smu_mod.SVI_PLANE_COUNT and metrics.SmuMetrics.AvgVoltage[k] != 0]
-        return {name: metrics.SmuMetrics.AvgVoltage[k] for k, name in voltage_keys}
+    return {name: metrics.SmuMetrics.AvgVoltage[k] for k, name in voltage_keys}

  def get_busy_threshold(self, dev):
    match dev.ip_ver[am.MP1_HWIP]:
@@ -203,37 +153,20 @@ class SMICtx:
      case _: return 15

  def get_gfx_freq(self, dev, metrics):
-    if metrics is None: return 0
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.GfxclkFrequency[0])
-      case _:
-        return metrics.SmuMetrics.AverageGfxclkFrequencyPostDs if self.get_gfx_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
-               metrics.SmuMetrics.AverageGfxclkFrequencyPreDs
+    return metrics.SmuMetrics.AverageGfxclkFrequencyPostDs if self.get_gfx_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
+          metrics.SmuMetrics.AverageGfxclkFrequencyPreDs

  def get_mem_freq(self, dev, metrics):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.UclkFrequency)
-      case _:
-        return metrics.SmuMetrics.AverageMemclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
-               metrics.SmuMetrics.AverageMemclkFrequencyPreDs
+    return metrics.SmuMetrics.AverageMemclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
+           metrics.SmuMetrics.AverageMemclkFrequencyPreDs

  def get_fckl_freq(self, dev, metrics):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.FclkFrequency)
-      case _:
-        return metrics.SmuMetrics.AverageFclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
-               metrics.SmuMetrics.AverageFclkFrequencyPreDs
+    return metrics.SmuMetrics.AverageFclkFrequencyPostDs if self.get_mem_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
+           metrics.SmuMetrics.AverageFclkFrequencyPreDs

-  def get_fan_rpm_pwm(self, dev, metrics):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return None, None
-      case _: return metrics.SmuMetrics.AvgFanRpm, metrics.SmuMetrics.AvgFanPwm
+  def get_fan_rpm_pwm(self, dev, metrics): return metrics.SmuMetrics.AvgFanRpm, metrics.SmuMetrics.AvgFanPwm

-  def get_power(self, dev, metrics):
-    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6): return self._smuq10_round(metrics.SocketPower), self._smuq10_round(metrics.MaxSocketPowerLimit)
-      case (13,0,12): return self._smuq10_round(metrics.SocketPower), self._smuq10_round(metrics.SocketPowerLimit)
-      case _: return metrics.SmuMetrics.AverageSocketPower, metrics.SmuMetrics.dGPU_W_MAX
+  def get_power(self, dev, metrics): return metrics.SmuMetrics.AverageSocketPower, metrics.SmuMetrics.dGPU_W_MAX

  def get_mem_usage(self, dev):
    usage = 0
@@ -244,8 +177,8 @@ class SMICtx:
        entry = pt.entries[i]

        if (entry & am.AMDGPU_PTE_VALID) == 0: continue
-        if pt.lv < am.AMDGPU_VM_PDB0 and not dev.gmc.is_pte_huge_page(pt.lv, entry):
-          pt_stack.append(AMPageTableEntry(dev, dev.xgmi2paddr(entry & 0x0000FFFFFFFFF000), lv=pt.lv+1))
+        if pt.lv!=am.AMDGPU_VM_PTB and not dev.gmc.is_pte_huge_page(entry):
+          pt_stack.append(AMPageTableEntry(dev, entry & 0x0000FFFFFFFFF000, lv=pt.lv+1))
          continue
        if (entry & am.AMDGPU_PTE_SYSTEM) != 0: continue
        usage += (1 << ((9 * (3-pt.lv)) + 12))
@@ -279,35 +212,30 @@ class SMICtx:
      device_line = [f"{bold(dev.pcibus)} {trim(self.lspci[dev.pcibus[5:]], col_size - 20)}"] + [pad("", col_size)]
      activity_line = [f"GFX Activity {draw_bar(self.get_gfx_activity(dev, metrics) / 100, activity_line_width)}"] \
                    + [f"MEM Activity {draw_bar(self.get_mem_activity(dev, metrics) / 100, activity_line_width)}"] \
-                    + [f"MEM Usage    {draw_bar(mem_used / mem_total, activity_line_width, opt_text=mem_fmt)}"] \
+                    + [f"MEM Usage    {draw_bar((mem_used / mem_total) / 100, activity_line_width, opt_text=mem_fmt)}"] \

      temps_data, temps_data_compact = self.get_temps(dev, metrics), self.get_temps(dev, metrics, compact=True)
      temps_table = ["=== Temps (°C) ==="] + [f"{name:<16}: {color_temp(val)}" for name, val in temps_data.items()]
      temps_table_compact = ["Temps (°C):" + '/'.join([f"{color_temp(val)} {name}" for name, val in temps_data_compact.items()])]

      fan_rpm, fan_pwm = self.get_fan_rpm_pwm(dev, metrics)
-      power_table = ["=== Power ==="]
-      power_table += ["Fan: N/A"] if fan_rpm is None or fan_pwm is None else [f"Fan Speed: {fan_rpm} RPM", f"Fan Power: {fan_pwm}%"]
+      power_table = ["=== Power ==="] + [f"Fan Speed: {fan_rpm} RPM"] + [f"Fan Power: {fan_pwm}%"]

      total_power, max_power = self.get_power(dev, metrics)
-      if max_power > 0:
-        power_line = [f"Power: " + draw_bar(total_power / max_power, 16, opt_text=f"{total_power}/{max_power}W")]
-        power_line_compact = [f"Power:       " + draw_bar(total_power / max_power, activity_line_width, opt_text=f"{total_power}/{max_power}W")]
-      else:
-        power_line = ["Power: N/A"]
-        power_line_compact = ["Power: N/A"]
+      power_line = [f"Power: " + draw_bar(total_power / max_power, 16, opt_text=f"{total_power}/{max_power}W")]
+      power_line_compact = [f"Power:       " + draw_bar(total_power / max_power, activity_line_width, opt_text=f"{total_power}/{max_power}W")]

      voltage_data = self.get_voltage(dev, metrics)
-      voltage_table = None if not voltage_data else (["=== Voltages ==="] + [f"{name:<20}: {color_voltage(voltage)}" for name, voltage in voltage_data.items()])
+      voltage_table = ["=== Voltages ==="] + [f"{name:<20}: {color_voltage(voltage)}" for name, voltage in voltage_data.items()]

      gfx_freq = self.get_gfx_freq(dev, metrics)
      mclk_freq = self.get_mem_freq(dev, metrics)
      fclk_freq = self.get_fckl_freq(dev, metrics)
+
      frequency_table = ["=== Frequencies ===", f"GFXCLK: {gfx_freq:>4} MHz", f"FCLK  : {fclk_freq:>4} MHz", f"MCLK  : {mclk_freq:>4} MHz"]

      if self.prev_terminal_width >= 231:
-        power_table += power_line
-        if voltage_table is not None: power_table += [""] + voltage_table
+        power_table += power_line + [""] + voltage_table
        activity_line += [""]
      elif self.prev_terminal_width >= 171:
        power_table += power_line + [""] + frequency_table
@@ -379,5 +307,4 @@ if __name__ == "__main__":
      smi_ctx.draw(args.list)
      if args.list: break
      time.sleep(1)
-  except KeyboardInterrupt:
-    print("Exiting...")
+  except KeyboardInterrupt: print("Exiting...")
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	30ff87eab4	realize sched	2025-10-14 14:44:56 +08:00
George HotzandGitHub	fe683bafa6	Merge branch 'master' into outerworld_work	2025-10-14 14:26:52 +08:00
geohot	ab9064c411	train loop	2025-10-10 20:20:19 +08:00
George HotzandGitHub	8832f08af3	Merge branch 'master' into outerworld_work	2025-10-10 20:07:44 +08:00
geohot	402e1cf48f	work	2025-10-10 19:49:24 +08:00
geohot	b2490b6e31	test assign/reduce	2025-10-10 18:50:16 +08:00
George HotzandGitHub	33e8babdd8	Merge branch 'master' into outerworld_work	2025-10-10 18:25:08 +08:00
geohot	67a409343d	work	2025-10-10 18:10:42 +08:00
geohot	5b24999a36	work on outerworld	2025-10-10 14:48:05 +08:00