can't allocate this at once, need parts back

No NamedTuple in compile.py
More WIP
2024-12-13 13:51:45 +01:00 · 2024-12-13 12:38:59 +01:00 · 2024-12-13 12:23:02 +01:00 · 2024-12-13 11:54:14 +01:00 · 2024-12-13 11:54:14 +01:00
1169 changed files with 221305 additions and 284071 deletions
@@ -1,3 +0,0 @@
-[run]
-source = tinygrad
-branch = True
@@ -1,15 +0,0 @@
-name: Run process replay tests
-description: Verify process replay compared to master
-runs:
-  using: "composite"
-  steps:
-    - name: Run process replay tests
-      shell: bash
-      run: |
-        export PR_TITLE=$(jq -r .pull_request.title "$GITHUB_EVENT_PATH")
-        export CURRENT_SHA=${{ github.event.pull_request && github.event.pull_request.head.sha || github.sha }}
-        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 && IGNORE_OOB=1 PYTHONPATH=. python3 process_replay.py
-        git checkout $CURRENT_HEAD  # restore to branch
@@ -1,305 +0,0 @@
-name: Setup Python & Install
-description: Sets up Python and installs project dependencies.
-inputs:
-  python-version:
-    description: 'Python version to use'
-    required: false
-    default: '3.12'
-  key:
-    description: 'Key for the python cache'
-    required: false
-    default: ''  # if you don't set a key, it doesn't cache
-  deps:
-    description: 'Extra dependency groups (comma separated)'
-    required: false
-    default: ''
-  pydeps:
-    description: 'Extra Python dependency groups (space separated)'
-    required: false
-    default: ''
-  opencl:
-    description: "Install OpenCL?"
-    required: false
-    default: 'false'
-  amd:
-    description: "Install AMD?"
-    required: false
-    default: 'false'
-  cuda:
-    description: "Install CUDA?"
-    required: false
-    default: 'false'
-  ocelot:
-    description: "Install gpuocelot?"
-    required: false
-    default: 'false'
-  webgpu:
-    description: "Install webgpu?"
-    required: false
-    default: 'false'
-  llvm:
-    description: "Install LLVM?"
-    required: false
-    default: 'false'
-  mesa:
-    description: "Install mesa"
-    required: false
-    default: 'false'
-runs:
-  using: "composite"
-  steps:
-    - name: Set up Python ${{ inputs.python-version }}
-      id: setup-python
-      uses: actions/setup-python@v5
-      with:
-        python-version: ${{ inputs.python-version }}
-
-    # **** Caching packages ****
-
-    - name: Cache Python packages
-      id: restore-venv
-      uses: actions/cache@v4
-      with:
-        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
-
-    # **** Caching downloads ****
-
-    - name: Cache downloads (Linux)
-      if: inputs.key != '' && runner.os == 'Linux'
-      uses: actions/cache@v4
-      with:
-        path: ~/.cache/tinygrad/downloads/
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
-    - name: Cache downloads (macOS)
-      if: inputs.key != '' && runner.os == 'macOS'
-      uses: actions/cache@v4
-      with:
-        path: ~/Library/Caches/tinygrad/downloads/
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
-
-    # **** Python deps ****
-
-    - name: Install dependencies in venv (with extra)
-      if: inputs.deps != '' && steps.restore-venv.outputs.cache-hit != 'true'
-      shell: bash
-      run: |
-        python -m venv .venv
-        if [[ "$RUNNER_OS" == "Windows" ]]; then
-          source .venv/Scripts/activate
-        else
-          . .venv/bin/activate
-        fi
-        python -m pip install -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --extra-index-url https://download.pytorch.org/whl/cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
-    - name: Install dependencies in venv (without extra)
-      if: inputs.deps == '' && steps.restore-venv.outputs.cache-hit != 'true'
-      shell: bash
-      run: |
-        python -m venv .venv
-        if [[ "$RUNNER_OS" == "Windows" ]]; then
-          source .venv/Scripts/activate
-        else
-          . .venv/bin/activate
-        fi
-        python -m pip install -e . ${{ inputs.pydeps }}
-    - name: Set up venv environment
-      shell: bash
-      run: |
-        echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
-        echo "OMP_NUM_THREADS=1" >> "$GITHUB_ENV"
-        # no buffers should be over 300MB in CI
-        echo "MAX_BUFFER_SIZE=300000000" >> "$GITHUB_ENV"
-        if [[ "$RUNNER_OS" == "Windows" ]]; then
-          echo "${{ github.workspace }}/.venv/Scripts" >> "$GITHUB_PATH"
-        else
-          echo "${{ github.workspace }}/.venv/bin" >> "$GITHUB_PATH"
-        fi
-
-    # ******************* apt *******************
-    - name: Setup apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
-      shell: bash
-      run: |
-        sudo chown -R $USER:$USER /var/cache/apt/archives
-
-        echo 'Acquire::GzipIndexes "true";' | sudo tee /etc/apt/apt.conf.d/gzip
-        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
-        echo 'Binary::apt::APT::Keep-Downloaded-Packages "true";' | sudo tee -a /etc/apt/apt.conf.d/99keep-debs
-
-    - name: Add OpenCL Repo
-      if: inputs.opencl == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: echo "deb [ allow-insecure=yes ] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list
-
-    - name: Add AMD Repo (Linux)
-      if: inputs.amd == 'true' && runner.os == 'Linux'
-      shell: bash
-      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 $(lsb_release -cs) main
-        EOF
-        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
-
-    - name: Add LLVM Repo (Linux)
-      if: inputs.llvm == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: |
-        wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
-        echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list
-
-    - name: Compute Package List + Hash
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
-      id: apt-pkgs
-      shell: bash
-      run: |
-        pkgs=""
-        # **** OpenCL ****
-        if [[ "${{ inputs.opencl }}" == "true" ]]; then
-          pkgs+=" opencl-headers \
-            intel-oneapi-runtime-openmp=2023.2.1-16 intel-oneapi-runtime-compilers-common=2023.2.1-16 intel-oneapi-runtime-compilers=2023.2.1-16 \
-            intel-oneapi-runtime-dpcpp-sycl-opencl-cpu=2023.2.1-16 intel-oneapi-runtime-tbb-common=2021.10.0-49541 \
-            intel-oneapi-runtime-tbb=2021.10.0-49541 intel-oneapi-runtime-opencl=2023.2.1-16"
-        fi
-        # **** AMD ****
-        if [[ "${{ inputs.amd }}" == "true" ]]; then
-          pkgs+=" hsa-rocr comgr hsa-rocr-dev liburing-dev libibverbs-dev libc6-dev"
-        fi
-        # **** CUDA ****
-        if [[ "${{ inputs.cuda }}" == "true" ]]; then
-          pkgs+=" git g++ cmake ninja-build llvm-15-dev zlib1g-dev libglew-dev \
-            flex bison libfl-dev libboost-thread-dev libboost-filesystem-dev nvidia-cuda-toolkit-gcc libzstd-dev"
-        fi
-        # **** WebGPU (dependencies for software-based vulkan) ****
-        if [[ "${{ inputs.webgpu }}" == "true" ]]; then
-          pkgs+=" libgl1 libglx-mesa0 libgl1-mesa-dri libxcb-xfixes0-dev mesa-vulkan-drivers"
-        fi
-        # **** LLVM ****
-        if [[ "${{ inputs.llvm }}" == "true" ]]; then
-          pkgs+=" libllvm20 clang-20 lld-20"
-        fi
-
-        echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
-        echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"
-
-    - name: Cache apt
-      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: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
-
-    - name: Run apt Update + Install
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
-      shell: bash
-      run: |
-        sudo apt -qq update || true
-
-        # ******** do install ********
-        if [[ -n "${{ steps.apt-pkgs.outputs.pkgs }}" ]]; then
-          sudo apt-get -y --allow-unauthenticated --no-install-recommends install ${{ steps.apt-pkgs.outputs.pkgs }}
-        fi
-
-        sudo chown -R $USER:$USER /var/cache/apt/archives/
-
-    # **** AMD ****
-    - name: Setup AMD (Linux)
-      if: inputs.amd == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: |
-        cargo build --release --manifest-path ./extra/remu/Cargo.toml
-        sudo ln -sf ${{ github.workspace }}/extra/remu/target/release/libremu.so /usr/local/lib/libremu.so
-        sudo tee --append /etc/ld.so.conf.d/rocm.conf <<'EOF'
-          /opt/rocm/lib
-          /opt/rocm/lib64
-        EOF
-        sudo ldconfig
-    - name: Setup AMD comgr+remu (macOS)
-      if: inputs.amd == 'true' && runner.os == 'macOS'
-      shell: bash
-      run: |
-        sudo mkdir -p /usr/local/lib
-        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/amdcomgr_dylib/releases/latest | \
-          jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
-          sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib
-        cargo build --release --manifest-path ./extra/remu/Cargo.toml
-
-    # **** gpuocelot ****
-
-    - name: Install gpuocelot dependencies (MacOS)
-      if: inputs.ocelot == 'true' && runner.os == 'macOS'
-      shell: bash
-      run: |
-        pkgs=(cmake ninja llvm@15 zlib glew flex bison [email protected] zstd ncurses)
-        for f in "${pkgs[@]}"; do
-          brew ls --versions "$f" >/dev/null 2>&1 || brew install --quiet "$f"
-        done
-
-        # Fix boost 1.85 for gpuocelot
-        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
-      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 }}
-    - name: Clone/compile gpuocelot
-      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
-        cd ${{ github.workspace }}/gpuocelot/ocelot
-        git checkout b16039dc940dc6bc4ea0a98380495769ff35ed99
-        mkdir build
-        cd build
-
-        CMAKE_ARGS="-Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5"
-        if [[ "${{ runner.os }}" == "macOS" ]]; then
-          CMAKE_ARGS="$CMAKE_ARGS -DBoost_INCLUDE_DIR=$(brew --prefix boost)/include -DBoost_LIBRARY_DIR=$(brew --prefix boost)/lib"
-        fi
-
-        cmake .. $CMAKE_ARGS
-        ninja
-    - name: Install gpuocelot
-      if: inputs.ocelot == 'true'
-      shell: bash
-      run: |
-        cd ${{ github.workspace }}/gpuocelot/ocelot/build
-        sudo cp libgpuocelot.${{ runner.os == 'macOS' && 'dylib' || 'so' }} /usr/${{ runner.os == 'macOS' && 'local/' || '' }}lib/
-
-    # **** WebGPU ****
-
-    - name: Install WebGPU dawn (Linux)
-      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 ldconfig
-    - name: Install WebGPU dawn (macOS)
-      if: inputs.webgpu == 'true' && runner.os == 'macOS'
-      shell: bash
-      run: |
-        brew tap wpmed92/dawn
-        brew install dawn
-
-    # **** LLVM ****
-
-    - name: Install LLVM (macOS)
-      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
@@ -1,176 +0,0 @@
-name: Autogen
-env:
-  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '13'
-  CAPTURE_PROCESS_REPLAY: 1
-  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-  PYTHONPATH: ${{ github.workspace }}
-
-on:
-  push:
-    branches:
-      - master
-  pull_request:
-    paths:
-    - 'tinygrad/runtime/autogen/**/*'
-    - 'tinygrad/runtime/support/autogen.py'
-  workflow_dispatch:
-    paths:
-    - 'tinygrad/runtime/autogen/**/*'
-    - 'tinygrad/runtime/support/autogen.py'
-
-jobs:
-  autogen:
-    name: In-tree Autogen
-    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:
-        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
-    - name: Verify OpenCL autogen
-      run: |
-        mv tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
-        python3 -c "from tinygrad.runtime.autogen import opencl"
-        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
-    - name: Verify CUDA autogen
-      run: |
-        mv tinygrad/runtime/autogen/cuda.py /tmp/cuda.py.bak
-        mv tinygrad/runtime/autogen/nvrtc.py /tmp/nvrtc.py.bak
-        mv tinygrad/runtime/autogen/nvjitlink.py /tmp/nvjitlink.py.bak
-        mv tinygrad/runtime/autogen/nv_570.py /tmp/nv_570.py.bak
-        mv tinygrad/runtime/autogen/nv.py /tmp/nv.py.bak
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv"
-        diff /tmp/cuda.py.bak tinygrad/runtime/autogen/cuda.py
-        diff /tmp/nvrtc.py.bak tinygrad/runtime/autogen/nvrtc.py
-        diff /tmp/nvjitlink.py.bak tinygrad/runtime/autogen/nvjitlink.py
-        diff /tmp/nv_570.py.bak tinygrad/runtime/autogen/nv_570.py
-        diff /tmp/nv.py.bak tinygrad/runtime/autogen/nv.py
-    - name: Verify AMD autogen
-      run: |
-        mv tinygrad/runtime/autogen/comgr.py /tmp/comgr.py.bak
-        mv tinygrad/runtime/autogen/hsa.py /tmp/hsa.py.bak
-        mv tinygrad/runtime/autogen/hip.py /tmp/hip.py.bak
-        mv tinygrad/runtime/autogen/amd_gpu.py /tmp/amd_gpu.py.bak
-        mv tinygrad/runtime/autogen/sqtt.py /tmp/sqtt.py.bak
-        mv tinygrad/runtime/autogen/rocprof.py /tmp/rocprof.py.bak
-        mv tinygrad/runtime/autogen/am/am.py /tmp/am_am.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_soc15.py /tmp/am_pm4_soc15.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_nv.py /tmp/am_pm4_nv.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_4_0_0.py /tmp/am_sdma_4_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_5_0_0.py /tmp/am_sdma_5_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_6_0_0.py /tmp/am_sdma_6_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v13_0_0.py /tmp/am_smu_v13_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v14_0_2.py /tmp/am_smu_v14_0_2.py.bak
-        python3 -c "from tinygrad.runtime.autogen import comgr, hsa, hip, amd_gpu, sqtt, rocprof; 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_v14_0_2"
-        diff /tmp/comgr.py.bak tinygrad/runtime/autogen/comgr.py
-        diff /tmp/hsa.py.bak tinygrad/runtime/autogen/hsa.py
-        diff /tmp/hip.py.bak tinygrad/runtime/autogen/hip.py
-        diff /tmp/amd_gpu.py.bak tinygrad/runtime/autogen/amd_gpu.py
-        diff /tmp/sqtt.py.bak tinygrad/runtime/autogen/sqtt.py
-        diff /tmp/rocprof.py.bak tinygrad/runtime/autogen/rocprof.py
-        diff /tmp/am_am.py.bak tinygrad/runtime/autogen/am/am.py
-        diff /tmp/am_pm4_soc15.py.bak tinygrad/runtime/autogen/am/pm4_soc15.py
-        diff /tmp/am_pm4_nv.py.bak tinygrad/runtime/autogen/am/pm4_nv.py
-        diff /tmp/am_sdma_4_0_0.py.bak tinygrad/runtime/autogen/am/sdma_4_0_0.py
-        diff /tmp/am_sdma_5_0_0.py.bak tinygrad/runtime/autogen/am/sdma_5_0_0.py
-        diff /tmp/am_sdma_6_0_0.py.bak tinygrad/runtime/autogen/am/sdma_6_0_0.py
-        diff /tmp/am_smu_v13_0_0.py.bak tinygrad/runtime/autogen/am/smu_v13_0_0.py
-        diff /tmp/am_smu_v14_0_2.py.bak tinygrad/runtime/autogen/am/smu_v14_0_2.py
-    - name: Verify Linux autogen
-      run: |
-        mv tinygrad/runtime/autogen/libc.py /tmp/libc.py.bak
-        mv tinygrad/runtime/autogen/kfd.py /tmp/kfd.py.bak
-        mv tinygrad/runtime/autogen/io_uring.py /tmp/io_uring.py.bak
-        mv tinygrad/runtime/autogen/ib.py /tmp/ib.py.bak
-        mv tinygrad/runtime/autogen/pci.py /tmp/pci.py.bak
-        mv tinygrad/runtime/autogen/vfio.py /tmp/vfio.py.bak
-        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
-        diff /tmp/libc.py.bak tinygrad/runtime/autogen/libc.py
-        diff /tmp/kfd.py.bak tinygrad/runtime/autogen/kfd.py
-        diff /tmp/io_uring.py.bak tinygrad/runtime/autogen/io_uring.py
-        diff /tmp/ib.py.bak tinygrad/runtime/autogen/ib.py
-        diff /tmp/pci.py.bak tinygrad/runtime/autogen/pci.py
-        diff /tmp/vfio.py.bak tinygrad/runtime/autogen/vfio.py
-    - name: Verify LLVM autogen
-      run: |
-        mv tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
-        python3 -c "from tinygrad.runtime.autogen import llvm"
-        diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
-    - name: Verify WebGPU autogen
-      run: |
-        mv tinygrad/runtime/autogen/webgpu.py /tmp/webgpu.py.bak
-        python3 -c "from tinygrad.runtime.autogen import webgpu"
-        diff /tmp/webgpu.py.bak tinygrad/runtime/autogen/webgpu.py
-    - name: Verify Qualcomm autogen
-      run: |
-        mv tinygrad/runtime/autogen/kgsl.py /tmp/kgsl.py.bak
-        mv tinygrad/runtime/autogen/qcom_dsp.py /tmp/qcom_dsp.py.bak
-        python3 -c "from tinygrad.runtime.autogen import kgsl, qcom_dsp"
-        diff /tmp/kgsl.py.bak tinygrad/runtime/autogen/kgsl.py
-        diff /tmp/qcom_dsp.py.bak tinygrad/runtime/autogen/qcom_dsp.py
-    - name: Verify libusb autogen
-      run: |
-        mv tinygrad/runtime/autogen/libusb.py /tmp/libusb.py.bak
-        python3 -c "from tinygrad.runtime.autogen import libusb"
-        diff /tmp/libusb.py.bak tinygrad/runtime/autogen/libusb.py
-    - name: Verify mesa autogen
-      run: |
-        mv tinygrad/runtime/autogen/mesa.py /tmp/mesa.py.bak
-        python3 -c "from tinygrad.runtime.autogen import mesa"
-        diff /tmp/mesa.py.bak tinygrad/runtime/autogen/mesa.py
-    - name: Verify libclang autogen
-      run: |
-        cp tinygrad/runtime/autogen/libclang.py /tmp/libclang.py.bak
-        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
-        diff /tmp/libclang.py.bak tinygrad/runtime/autogen/libclang.py
-  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:
-        llvm: 'true'
-    - name: Verify macos autogen
-      run: |
-        mv tinygrad/runtime/autogen/metal.py /tmp/metal.py.bak
-        LIBCLANG_PATH=/opt/homebrew/opt/llvm@20/lib/libclang.dylib python3 -c "from tinygrad.runtime.autogen import metal"
-        diff /tmp/metal.py.bak tinygrad/runtime/autogen/metal.py
-  autogen-comgr-3:
-    name: In-tree Autogen (comgr 3)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-    - name: Install autogen support packages
-      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/6.4 $(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: Verify comgr (3) autogen
-      run: |
-        mv tinygrad/runtime/autogen/comgr_3.py /tmp/comgr_3.py.bak
-        python3 -c "from tinygrad.runtime.autogen import comgr_3"
-        diff /tmp/comgr_3.py.bak tinygrad/runtime/autogen/comgr_3.py
@@ -2,7 +2,7 @@ name: Benchmarks
 env:
  # TODO: this rescheduling makes gpt2, mixtral and llama unjitted slower
  # TODO: very slow for llama 70B and resnet training 6 GPU
-  CAPTURE_PROCESS_REPLAY: "1"
+  RUN_PROCESS_REPLAY: "1"
  ASSERT_PROCESS_REPLAY: "0"
  PYTHONPATH: .
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -14,18 +14,21 @@ on:
      - update_benchmark
      - update_benchmark_staging
  workflow_dispatch:
+    inputs:
+      run_process_replay:
+        description: "Run process replay tests"
+        required: false
+        default: false
+        type: boolean

 jobs:
  testmacbenchmark:
    name: Mac Benchmark
-    env:
-      # since sudo is required for usbgpu on macos, move the cache to a new location, as some of the files are owned by root
-      PYTHONPYCACHEPREFIX: /tmp/tiny_python_pycache
    runs-on: [self-hosted, macOS]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -e -o pipefail {0}
+        shell: bash -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -33,7 +36,6 @@ 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
@@ -46,79 +48,58 @@ 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 | tee sd.txt
+      run: JIT=1 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 | tee sd_no_fp16.txt
+      run: JIT=1 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 | tee sdv2.txt
-    # process replay can't capture this, the graph is too large
+      run: JIT=1 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
    - 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
+      run: 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 | tee torch_speed.txt
+      run: BIG=2 MPS=1 python3.11 test/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
-      run: |
-        DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_tensor_cores.py
-        DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_tensor_cores.py
-        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
+      run: METAL=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
    - name: Run Tensor Core GEMM (float)
-      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
+      run: DEBUG=2 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 | tee matmul_half.txt
+      run: DEBUG=2 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 | tee matmul_bfloat16.txt
+      run: DEBUG=2 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 | 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
+        JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        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 | tee llama_beam.txt
+      run: 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 | 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 | 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 | tee llama_four_gpu.txt
+        python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
+        python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_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 | 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 | 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
+        JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        JIT=1 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 | tee gpt2_half.txt
+      run: 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 | tee gpt2_half_beam.txt
-    - name: Run OLMoE
-      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
+      run: HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAST_BEFORE_VIEW=0 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - name: Train MNIST
      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 | 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 | tee train_cifar_half.txt
-
+    - name: Run 10 CIFAR training steps
+      run: JIT=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+    - name: Run 10 CIFAR training steps w HALF
+      run: JIT=2 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 | 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 | tee train_cifar_wino.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: JIT=1 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (Mac)
@@ -130,8 +111,6 @@ jobs:
          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
@@ -152,44 +131,13 @@ jobs:
    - 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: 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 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]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -e -o pipefail {0}
+        shell: bash -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -213,66 +161,59 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Run model inference benchmark
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
+      run: NV=1 RUN_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 | tee torch_speed.txt
+      run: NV=1 RUN_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/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
-    - name: Test benchmark allreduce
-      run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
+      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 tensor cores
      run: |
-        NV=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-        NV=1 NV_PTX=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+        NV=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
+        PTX=1 NV=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
    - name: Run Tensor Core GEMM (CUDA)
      run: |
-        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
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_fp8.txt
+        CUDA=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
+        CUDA=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.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 | tee matmul_ptx.txt
+      run: NV=1 PTX=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 | tee matmul_nv.txt
+      run: NV=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 | 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 | tee sdxl.txt
+      run: NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+    - name: Run SDXL
+      run: NV=1 RUN_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 | 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
+        NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        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 | tee llama_beam.txt
+      run: 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 | tee llama_four_gpu.txt
+    #   run: NV=1 RUN_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 | tee llama_six_gpu.txt
+    #   run: NV=1 RUN_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 | 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 | tee llama3_four_gpu.txt
-    - 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 | tee llama3_fp8.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 | 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 | tee llama_2_70B.txt
+      run: 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
+      run: NV=1 RUN_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 RUN_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 RUN_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 | tee mixtral.txt
+      run: time NV=1 RUN_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 | 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
+        NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        NV=1 JIT=1 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 | tee gpt2_half.txt
+      run: NV=1 HALF=1 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 | tee gpt2_half_beam.txt
+      run: NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAST_BEFORE_VIEW=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA)
@@ -281,7 +222,6 @@ jobs:
          torch_speed.txt
          matmul.txt
          matmul_bfloat16.txt
-          matmul_tf32.txt
          matmul_ptx.txt
          matmul_nv.txt
          sd.txt
@@ -292,7 +232,6 @@ jobs:
          llama3_beam.txt
          llama3_four_gpu.txt
          llama3_six_gpu.txt
-          llama3_fp8.txt
          llama_2_70B.txt
          mixtral.txt
          gpt2_unjitted.txt
@@ -305,10 +244,10 @@ jobs:
  testmorenvidiabenchmark:
    name: tinybox green Training Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -e -o pipefail {0}
+        shell: bash -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -330,37 +269,30 @@ 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
-    # TODO: too slow
-    # - name: Fuzz Padded Tensor Core GEMM (NV)
-    #   run: NV=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
-    # 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 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+    - name: Fuzz Padded Tensor Core GEMM (NV)
+      run: NV=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: Fuzz Padded Tensor Core GEMM (PTX)
+      run: NV=1 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: Train MNIST
      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 | tee train_cifar.txt
+      run: 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=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: 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 | tee train_cifar_bf16.txt
-    # - name: Run 10 CIFAR training steps w winograd
-    #   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
+      run: NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: NV=1 RUN_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 | tee train_cifar_one_gpu.txt
+      run: time NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 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 | tee train_cifar_six_gpu.txt
+      run: time RUN_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 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
+      run: time 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 | tee train_resnet_one_gpu.txt
+      run: 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 | tee train_bert.txt
+      run: NV=1 RUN_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA Training)
@@ -371,30 +303,23 @@ jobs:
          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
+          train_cifar_six_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

  testamdbenchmark:
    name: tinybox red Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -e -o pipefail {0}
+        shell: bash -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - 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
      run: |
        mkdir -p weights
@@ -413,12 +338,12 @@ 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: setup perflevel
-    #  run: |
-    #    examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
-    #    rocm-smi
-    #- name: Show off tinybox
-    #  run: /opt/rocm/bin/rocm-bandwidth-test
+    - name: setup perflevel
+      run: |
+        examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
+        rocm-smi
+    - name: Show off tinybox
+      run: /opt/rocm/bin/rocm-bandwidth-test
    # TODO: unstable on AMD
    #- name: Run model inference benchmark
    #  run: LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 NOCLANG=1 python3 test/external/external_model_benchmark.py
@@ -426,63 +351,52 @@ 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 | tee torch_speed.txt
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/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
+      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
+      run: AMD=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded
    - 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
+      run: AMD=1 HALF=1 DEBUG=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
-    #  run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
+    - name: Test HIP=1
+      run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
    # TODO: AMD compiler bug causes this to fail
    #- name: Fuzz Padded Tensor Core GEMM
    #  run: HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
-    #- name: Remove amdgpu
-    #  run: sleep 10 && sudo rmmod amdgpu # sleep a bit to let the driver unload the prev pid.
-    - name: Test AM cold start time
-      run: time AMD=1 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
-    - 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 | tee sd.txt
+      run: AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - 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
+      run: 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 | 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
+        AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        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 | tee llama_beam.txt
+      run: 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 | tee llama_four_gpu.txt
+    #   run: AMD=1 RUN_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 | tee llama_six_gpu.txt
+    #   run: AMD=1 RUN_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 | 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 | 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 | 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 | tee llama_2_70B.txt
+      run: 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
+      run: AMD=1 RUN_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 RUN_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: AMD=1 RUN_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 | tee mixtral.txt
+      run: time 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 | 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
+        AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        AMD=1 JIT=1 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 | tee gpt2_half.txt
+      run: AMD=1 HALF=1 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 | tee gpt2_half_beam.txt
+      run: AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAST_BEFORE_VIEW=0 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD)
@@ -511,18 +425,14 @@ jobs:
  testmoreamdbenchmark:
    name: tinybox red Training Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -e -o pipefail {0}
+        shell: bash -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - 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
@@ -540,21 +450,30 @@ 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: setup perflevel
+      run: |
+        examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
+        rocm-smi
    - name: Train MNIST
      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 | tee train_cifar.txt
+      run: 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=200 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 | 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 | tee train_cifar_wino.txt
+      run: 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: AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: 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 | tee train_cifar_one_gpu.txt
+      run: time AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 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
+      run: time AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+    - name: Run MLPerf resnet eval
+      run: time AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      run: 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: AMD=1 RUN_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD Training)
@@ -565,58 +484,9 @@ jobs:
          train_cifar_bf16.txt
          train_cifar_wino.txt
          train_cifar_one_gpu.txt
-          train_cifar_six_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
-
-  testmlperfamdbenchmark:
-    name: tinybox red MLPerf Benchmark
-    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - 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
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - 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: reset process replay
-      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 | 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 | 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
+          train_cifar_six_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

@@ -626,7 +496,7 @@ jobs:
    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -e -o pipefail {0}
+        shell: bash -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -638,159 +508,31 @@ 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 0.10.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.10.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/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=2 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: DEBUG=2 IMAGE=1 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=2 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=4 DEV=QCOM FLOAT16=1 IMAGE=2 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=10 DEV=QCOM FLOAT16=1 IMAGE=2 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 MobileNetV2 on DSP
-      run: |
-        # generate quantized weights
-        ln -s /data/home/tiny/tinygrad/extra/datasets/imagenet extra/datasets/imagenet
-        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
+    - name: validate openpilot 0.9.7
+      run: PYTHONPATH=. FLOAT16=0 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
+    - name: benchmark openpilot 0.9.4
+      run: PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_4.txt
+    - name: benchmark openpilot 0.9.7
+      run: PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_7.txt
+    - name: benchmark openpilot w IMAGE=2 0.9.4
+      run: 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/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_4.txt
+    - name: benchmark openpilot w IMAGE=2 0.9.7
+      run: 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/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
+    - name: openpilot compile3 0.9.7
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx
+    - name: openpilot compile3 0.9.7+ tomb raider
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/e8bea2c78ffa92685ece511e9b554122aaf1a79d/selfdrive/modeld/models/supercombo.onnx
+    - name: openpilot dmonitoring compile3 0.9.7
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/dmonitoring_model.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
-
-  testreddriverbenchmark:
-    name: AM Benchmark
-    runs-on: [self-hosted, Linux, tinyboxrandom]
-    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: Remove amd modules
-      run: ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - 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: reset process replay
-      run: test/external/process_replay/reset.py
-    - name: Test driver cold start time
-      run: time DEBUG=3 AMD=1 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
-    - name: Test driver warm start time
-      run: time DEBUG=3 AMD=1 python3 test/test_tiny.py TestTiny.test_plus
-    # Fails on 9070
-    # - name: Test tensor cores
-    #   run: |
-    #     AMD=1 AMD_LLVM=0 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     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 | 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
-      run: AMD=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
-    - name: Test CPU copy time
-      run: |
-        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 | tee am_train_cifar_one_gpu.txt
-    # - 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 | 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 | tee am_train_bert_one_gpu.txt
    - uses: actions/upload-artifact@v4
      with:
-        name: Speed (AM Driver)
+        name: Speed (comma)
        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
-
-  testgreendriverbenchmark:
-    name: NV Benchmark
-    runs-on: [self-hosted, Linux, tinyboxrandom]
-    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: Remove nv modules
-      run: ./extra/hcq/hcq_smi.py nv rmmod
-    - name: Kill stale pids
-      run: ./extra/hcq/hcq_smi.py nv kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - 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: reset process replay
-      run: test/external/process_replay/reset.py
-    - name: Test driver start time
-      run: time DEBUG=3 NV=1 python3 test/test_tiny.py TestTiny.test_plus
-    - name: Test tensor cores
-      run: NV=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-    - name: Test DISK copy time
-      run: NV=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
-    - name: Test CPU copy time
-      run: |
-        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 | 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 | 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 | 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
+          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
@@ -1,34 +0,0 @@
-name: Benchmark with kernel search
-
-on:
-  push:
-    branches:
-      - update_benchmark_search
-  workflow_dispatch:
-
-jobs:
-  run_script_job:
-    runs-on: [self-hosted, Linux, tinybox]
-    if: github.repository_owner == 'tinygrad'
-    timeout-minutes: 100
-
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Remove amdgpu
-      run: sudo rmmod amdgpu || true
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
-    - 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
-    - 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
-    - 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
@@ -1,30 +0,0 @@
-name: Run MLPerf Training
-
-on:
-  schedule:
-    - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
-  push:
-    branches:
-      - update_mlperf
-  workflow_dispatch:
-
-jobs:
-  run_script_job:
-    runs-on: [self-hosted, Linux, tinybox]
-    if: github.repository_owner == 'tinygrad'
-    timeout-minutes: 720
-
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Cleanup running AM processes
-      run: python extra/amdpci/am_smi.py --pids --kill
-    - name: Symlink datasets
-      run: |
-        mkdir -p extra/datasets
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: Run resnet
-      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"
@@ -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"
@@ -10,7 +10,6 @@ notebooks
 *.so
 *.txt
 build
-!examples/tinychat/assets/cdn.jsdelivr.net/npm/[email protected]/build/
 /dist
 *.egg-info
 /env
@@ -34,13 +33,11 @@ extra/datasets/open-images-v6-mlperf
 extra/datasets/kits/
 extra/datasets/COCO/
 extra/datasets/audio*
-extra/huggingface_onnx/models/*
-extra/huggingface_onnx/*.yaml
 extra/weights
 venv
-venv_sd_mlperf
-examples/**/net.*[js,json]
+examples/**/**/net*.*[js,json]
 examples/**/*.safetensors
+examples/webgpu/stable_diffusion/*.*[js,mjs]
 node_modules
 package.json
 package-lock.json
@@ -59,9 +56,4 @@ weights
 comgr_*
 *.pkl
 site/
-profile_stats
-*.log
-target
-.mypy_cache
-mutants
-.mutmut-cache
+master_schedule.py
@@ -20,15 +20,27 @@ repos:
        language: system
        always_run: true
        pass_filenames: false
-      - id: example
-        name: test all devices
-        entry: python3 test/external/external_test_example.py
+      - id: devicetests
+        name: select GPU tests
+        entry: env GPU=1 PYTHONPATH="." python3 -m pytest test/test_uops.py test/test_search.py
        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/test_ops.py test/test_schedule.py test/test_assign.py test/test_tensor.py test/test_jit.py test/unit/test_schedule_cache.py test/unit/test_pattern_matcher.py test/unit/test_uop_symbolic.py test/unit/test_helpers.py
+        name: subset of tests
+        entry: env PYTHONPATH="." python3 -m pytest -n=4 test/unit/ test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py test/test_symbolic_shapetracker.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
@@ -7,7 +7,7 @@ extension-pkg-whitelist=scipy,cereal.messaging.messaging_pyx,PyQt5,av

 # Add files or directories to the blacklist. They should be base names, not
 # paths.
-ignore=CVS,autogen,msm_kgsl.py,runtime,.venv
+ignore=CVS,autogen,msm_kgsl.py

 # Add files or directories matching the regex patterns to the blacklist. The
 # regex matches against base names, not paths.
@@ -30,6 +30,10 @@ persistent=yes
 # Specify a configuration file.
 #rcfile=

+# When enabled, pylint would attempt to guess common misconfiguration and emit
+# user-friendly hints instead of false-positive error messages
+suggestion-mode=yes
+
 # Allow loading of arbitrary C extensions. Extensions are imported into the
 # active Python interpreter and may run arbitrary code.
 unsafe-load-any-extension=no
@@ -50,12 +54,11 @@ confidence=
 # --enable=similarities". If you want to run only the classes checker, but have
 # no Warning level messages displayed, use"--disable=all --enable=classes
 # --disable=W"
-disable=C,R,W0613,W0511,W0212,W0201,W0106,W0603,W0621,W0703,W1201,W1203,E1136,W1514,E1101,W0221,W0105,E0401,abstract-method,W0707
+disable=C,R,W0613,W0511,W0212,W0201,W0106,W0603,W0621,W0703,W1201,W1203,E1136,W1514,E1101,W0221,W0105,E0401,abstract-method
 # E1101 for function binding
 # W0221 for Function class
 # W0105 for comment strings
 # E0401 for missing imports
-# W0707 for not reraising

 # Enable the message, report, category or checker with the given id(s). You can
 # either give multiple identifier separated by comma (,) or put this option
@@ -1,17 +0,0 @@
-# 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.
@@ -1,227 +0,0 @@
-# Claude Code Guide for tinygrad
-
-## Architecture Overview
-
-tinygrad compiles tensor operations into optimized kernels. The pipeline:
-
-1. **Tensor** (`tensor.py`) - User-facing API, creates UOp graph
-2. **UOp** (`uop/ops.py`) - Unified IR for all operations (both tensor and kernel level)
-3. **Schedule** (`engine/schedule.py`, `schedule/`) - Converts tensor UOps to kernel UOps
-4. **Codegen** (`codegen/`) - Converts kernel UOps to device code
-5. **Runtime** (`runtime/`) - Device-specific execution
-
-## Key Concepts
-
-### UOp (Universal Operation)
-Everything is a UOp - tensors, operations, buffers, kernels. Key properties:
- `op`: The operation type (Ops enum)
- `dtype`: Data type
- `src`: Tuple of source UOps
- `arg`: Operation-specific argument
- `tag`: Optional tag for graph transformations
-
-UOps are **immutable and cached** - creating the same UOp twice returns the same object (ucache).
-
-### PatternMatcher
-Used extensively for graph transformations:
-```python
-pm = PatternMatcher([
-  (UPat(Ops.ADD, src=(UPat.cvar("x"), UPat.cvar("x"))), lambda x: x * 2),
-])
-result = graph_rewrite(uop, pm)
-```
-
-### Schedule Cache
-Schedules are cached by graph structure. BIND nodes (variables with bound values) are unbound before cache key computation so different values hit the same cache.
-
-## Directory Structure
-
-```
-tinygrad/
-├── tensor.py          # Tensor class, user API
-├── device.py          # Buffer, device management
-├── dtype.py           # Data types
-├── helpers.py         # Utilities, environment vars
-├── uop/
-│   ├── ops.py         # UOp class, Ops enum, PatternMatcher
-│   ├── spec.py        # UOp type verification
-│   └── symbolic.py    # Symbolic math simplification
-├── engine/
-│   ├── schedule.py    # Schedule creation, caching
-│   ├── realize.py     # Tensor realization
-│   ├── jit.py         # JIT compilation
-│   └── memory.py      # Memory planning
-├── schedule/
-│   ├── rangeify.py    # Convert movements to ranges
-│   └── indexing.py    # Index calculations
-├── codegen/
-│   ├── kernel.py      # Kernel optimization
-│   └── uopgraph.py    # UOp graph transformations
-├── renderer/          # Code generation (CUDA, Metal, etc.)
-└── runtime/           # Device backends
-```
-
-## Testing
-
-```bash
-# Run specific test
-python -m pytest test/unit/test_schedule_cache.py -xvs
-
-# Run with timeout
-python -m pytest test/test_symbolic_ops.py -x --timeout=60
-
-# Debug with print
-DEBUG=2 python -m pytest test/test_schedule.py::test_name -xvs
-
-# Visualize UOp graphs
-VIZ=1 python -c "from tinygrad import Tensor; Tensor.ones(10).sum().realize()"
-```
-
-## Common Environment Variables
-
- `DEBUG=1-4` - Increasing verbosity
- `VIZ=1` - Enable graph visualization
- `SPEC=1` - Enable UOp spec verification
- `NOOPT=1` - Disable optimizations
- `DEVICE=CPU/CUDA/AMD/METAL` - Set default device
-
-## Debugging Tips
-
-1. **Print UOp graphs**: `print(tensor.uop)` or `print(tensor.uop.sink())`
-2. **Check schedule**: `tensor.schedule()` returns list of ExecItems
-3. **Trace graph rewrites**: Use `VIZ=1` or add print in PatternMatcher callbacks
-4. **Find UOps by type**: `[u for u in uop.toposort() if u.op is Ops.SOMETHING]`
-
-## Workflow Rules
-
- **NEVER commit without explicit user approval** - always show the diff and wait for approval
- **NEVER amend commits** - always create a new commit instead
- Run `pre-commit run --all-files` before committing to catch linting/type errors
- Run tests before proposing commits
- Test with `SPEC=2` when modifying UOp-related code
-
-## Style Notes
-
- 2-space indentation, 150 char line limit
- PatternMatchers should be defined at module level (slow to construct)
- Prefer `graph_rewrite` over manual graph traversal
- UOp methods like `.replace()` preserve tags unless explicitly changed
- Use `.rtag(value)` to add tags to UOps
-
-## Lessons Learned
-
-### UOp ucache Behavior
-UOps are cached by their contents - creating a UOp with identical (op, dtype, src, arg) returns the **same object**. This means:
- `uop.replace(tag=None)` on a tagged UOp returns the original untagged UOp if it exists in cache
- Two UOps with same structure are identical (`is` comparison works)
-
-### Spec Validation
-When adding new UOp patterns, update `tinygrad/uop/spec.py`. Test with:
-```bash
-SPEC=2 python3 test/unit/test_something.py
-```
-Spec issues appear as `RuntimeError: SPEC ISSUE None: UOp(...)`.
-
-### Schedule Cache Key Normalization
-The schedule cache strips values from BIND nodes so different bound values (e.g., KV cache positions) hit the same cache entry:
- `pm_pre_sched_cache`: BIND(DEFINE_VAR, CONST) → BIND(DEFINE_VAR) for cache key
- `pm_post_sched_cache`: restores original BIND from context
- When accessing `bind.src[1]`, check `len(bind.src) > 1` first (might be stripped)
- Extract var_vals from `input_buffers` dict after graph_rewrite (avoids extra toposort)
-
-### Avoiding Extra Work
- Use ctx dict from graph_rewrite to collect info during traversal instead of separate toposort
- Only extract var_vals when schedule is non-empty (no kernels = no vars needed)
- PatternMatchers are slow to construct - define at module level, not in functions
-
-### Readability Over Speed
-Don't add complexity for marginal performance gains. Simpler code that's slightly slower is often better:
-```python
-# BAD: "optimized" with extra complexity
-if has_afters:  # skip toposort if no AFTERs
-  after_map = [(u, u.buf_uop) for u in big_sink.toposort() if u.op is Ops.AFTER]
-
-# GOOD: simple, always works
-after_map = [(u, u.buf_uop) for u in big_sink.toposort() if u.op is Ops.AFTER]
-```
-The conditional check adds complexity, potential bugs, and often negligible speedup. Only optimize when profiling shows a real bottleneck.
-
-### Testing LLM Changes
-```bash
-# Quick smoke test
-echo "Hello" | DEBUG=1 python tinygrad/apps/llm.py --model "llama3.2:1b"
-
-# Check cache hits (should see "cache hit" after warmup)
-echo "Hello world" | DEBUG=1 python tinygrad/apps/llm.py --model "llama3.2:1b" 2>&1 | grep cache
-
-# Test with beam search
-echo "Hello" | BEAM=2 python tinygrad/apps/llm.py --model "llama3.2:1b"
-```
-
-## Common Patterns
-
-### Graph Transformation
-```python
-def my_transform(ctx, x):
-  # Return new UOp or None to skip
-  return x.replace(arg=new_arg)
-
-pm = PatternMatcher([
-  (UPat(Ops.SOMETHING, name="x"), my_transform),
-])
-result = graph_rewrite(input_uop, pm, ctx={})
-```
-
-### Finding Variables
-```python
-# Get all variables in a UOp graph
-variables = uop.variables()
-
-# Get bound variable values
-var, val = bind_uop.unbind()
-```
-
-### Shape Handling
-```python
-# Shapes can be symbolic (contain UOps)
-shape = tensor.shape  # tuple[sint, ...] where sint = int | UOp
-```
-
-## Performance Optimization
-
-When optimizing tinygrad internals:
-
-1. **Measure wall time, not just call counts** - Reducing `graph_rewrite` calls doesn't always improve wall time. The overhead of conditional checks can exceed the cost of the operation being skipped.
-
-2. **Profile each optimization individually** - Run benchmarks with and without each change to measure actual impact. Use `test/external/external_benchmark_schedule.py` for schedule/rewrite timing.
-
-3. **Early exits in hot paths are effective** - Simple checks like `if self.op is Ops.CONST: return self` in `simplify()` can eliminate many unnecessary `graph_rewrite` calls.
-
-4. **`graph_rewrite` is expensive** - Each call has overhead even for small graphs. Avoid calling it when the result is trivially known (e.g., simplifying a CONST returns itself).
-
-5. **Beware iterator overhead** - Checks like `all(x.op is Ops.CONST for x in self.src)` can be slower than just running the operation, especially for small sequences.
-
-6. **Verify cache hit rates before adding/keeping caches** - Measure actual hit rates with real workloads. A cache with 0% hit rate is pure overhead (e.g., `pm_cache` was removed because the algorithm guarantees each UOp is only passed to `pm_rewrite` once).
-
-7. **Use `TRACK_MATCH_STATS=2` to profile pattern matching** - This shows match rates and time per pattern. Look for patterns with 0% match rate that still cost significant time - these are pure overhead for that workload.
-
-8. **Cached properties beat manual traversal** - `backward_slice` uses `@functools.cached_property`. A DFS with early-exit sounds faster but is actually slower because it doesn't benefit from caching. The cache hit benefit often outweighs algorithmic improvements.
-
-9. **Avoid creating intermediate objects in hot paths** - For example, `any(x.op in ops for x in self.backward_slice)` is faster than `any(x.op in ops for x in {self:None, **self.backward_slice})` because it avoids dict creation.
-
-## Pattern Matching Profiling
-
-Use `TRACK_MATCH_STATS=2` to identify expensive patterns:
-
-```bash
-TRACK_MATCH_STATS=2 PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
-```
-
-Output format: `matches / attempts -- match_time / total_time ms -- location`
-
-Key patterns to watch (from ResNet50 benchmark):
- `split_load_store`: ~146ms, 31% match rate - does real work
- `simplify_valid`: ~75ms, 0% match rate in this workload - checks AND ops for INDEX in backward slice
- `vmin==vmax folding`: ~55ms, 0.33% match rate - checks 52K ops but rarely matches
-
-Patterns with 0% match rate are workload-specific overhead. They may be useful in other workloads, so don't remove them without understanding their purpose.
@@ -21,38 +21,17 @@ tinygrad: For something between [PyTorch](https://github.com/pytorch/pytorch) an

 ---

-tinygrad is an end-to-end deep learning stack:
+This may not be the best deep learning framework, but it is a 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 aims to be 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 still alpha software, but 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

@@ -60,8 +39,9 @@ Try a matmul. See how, despite the style, it is fused into one kernel with the p

 ```sh
 DEBUG=3 python3 -c "from tinygrad import Tensor;
-N = 1024; a, b = Tensor.empty(N, N), Tensor.empty(N, N);
-(a.reshape(N, 1, N) * b.T.reshape(1, N, N)).sum(axis=2).realize()"
+N = 1024; a, b = Tensor.rand(N, N), Tensor.rand(N, N);
+c = (a.reshape(N, 1, N) * b.T.reshape(1, N, N)).sum(axis=2);
+print((c.numpy() - (a.numpy() @ b.numpy())).mean())"
 ```

 And we can change `DEBUG` to `4` to see the generated code.
@@ -100,8 +80,9 @@ See [examples/beautiful_mnist.py](examples/beautiful_mnist.py) for the full vers

 tinygrad already supports numerous accelerators, including:

- [x] [OpenCL](tinygrad/runtime/ops_cl.py)
- [x] [CPU](tinygrad/runtime/ops_cpu.py)
+- [x] [GPU (OpenCL)](tinygrad/runtime/ops_gpu.py)
+- [x] [CLANG (C Code)](tinygrad/runtime/ops_clang.py)
+- [x] [LLVM](tinygrad/runtime/ops_llvm.py)
 - [x] [METAL](tinygrad/runtime/ops_metal.py)
 - [x] [CUDA](tinygrad/runtime/ops_cuda.py)
 - [x] [AMD](tinygrad/runtime/ops_amd.py)
@@ -170,7 +151,7 @@ We'll start with what will get your PR closed with a pointer to this section:

 - No code golf! While low line count is a guiding light of this project, anything that remotely looks like code golf will be closed. The true goal is reducing complexity and increasing readability, and deleting `\n`s does nothing to help with that.
 - All docs and whitespace changes will be closed unless you are a well-known contributor. The people writing the docs should be those who know the codebase the absolute best. People who have not demonstrated that shouldn't be messing with docs. Whitespace changes are both useless *and* carry a risk of introducing bugs.
- Anything you claim is a "speedup" must be benchmarked. In general, the goal is simplicity, so even if your PR makes things marginally faster, you have to consider the tradeoff with maintainability and readability.
+- Anything you claim is a "speedup" must be benchmarked. In general, the goal is simplicity, so even if your PR makes things marginally faster, you have to consider the tradeoff with maintainablity and readablity.
 - In general, the code outside the core `tinygrad/` folder is not well tested, so unless the current code there is broken, you shouldn't be changing it.
 - If your PR looks "complex", is a big diff, or adds lots of lines, it won't be reviewed or merged. Consider breaking it up into smaller PRs that are individually clear wins. A common pattern I see is prerequisite refactors before adding new functionality. If you can (cleanly) refactor to the point that the feature is a 3 line change, this is great, and something easy for us to review.

@@ -0,0 +1,266 @@
+#!/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 --user .
+  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
+  raise RuntimeError("library $name not found")
+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'"
+  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
+  python3 -c "import tinygrad.runtime.autogen.kfd"
+}
+
+generate_cuda() {
+  clang2py /usr/include/cuda.h -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 $BASE/nvrtc.py
+  python3 -c "import tinygrad.runtime.autogen.nvrtc"
+}
+
+generate_nv() {
+  NVKERN_COMMIT_HASH=d6b75a34094b0f56c2ccadf14e5d0bd515ed1ab6
+  NVKERN_SRC=/tmp/open-gpu-kernel-modules-$NVKERN_COMMIT_HASH
+  if [ ! -d "$NVKERN_SRC" ]; then
+    git clone https://github.com/tinygrad/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 \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0080.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/clc56f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.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/kernel-open/nvidia-uvm/clc6b5.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 \
+    --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/#\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"
+}
+
+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/kfd_pm4_headers_ai.h \
+    extra/hip_gpu_driver/soc21_enum.h \
+    extra/hip_gpu_driver/sdma_v6_0_0_pkt_open.h \
+    extra/hip_gpu_driver/gc_11_0_0_offset.h \
+    extra/hip_gpu_driver/gc_10_3_0_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_libc() {
+  clang2py -k cdefstum \
+    $(dpkg -L libc6-dev | grep sys/mman.h) \
+    $(dpkg -L libc6-dev | grep sys/syscall.h) \
+    /usr/include/elf.h \
+    /usr/include/unistd.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)\g" $BASE/libc.py
+
+  fixup $BASE/libc.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
+  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"
+}
+
+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" == "qcom" ]; then generate_qcom
+elif [ "$1" == "io_uring" ]; then generate_io_uring
+elif [ "$1" == "libc" ]; then generate_libc
+elif [ "$1" == "kgsl" ]; then generate_kgsl
+elif [ "$1" == "adreno" ]; then generate_adreno
+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
+else echo "usage: $0 <type>"
+fi
@@ -0,0 +1,118 @@
+# 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 "LazyBuffer" 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_clang import ClangProgram, ClangCompiler, MallocAllocator
+
+# allocate some buffers
+out = MallocAllocator.alloc(4)
+a = MallocAllocator.alloc(4)
+b = MallocAllocator.alloc(4)
+
+# load in some values (little endian)
+MallocAllocator._copyin(a, memoryview(bytearray([2,0,0,0])))
+MallocAllocator._copyin(b, memoryview(bytearray([3,0,0,0])))
+
+# compile a program to a binary
+lib = ClangCompiler().compile("void add(int *out, int *a, int *b) { out[0] = a[0] + b[0]; }")
+
+# create a runtime for the program (ctypes.CDLL)
+fxn = ClangProgram("add", lib)
+
+# run the program
+fxn(out, a, b)
+
+# check the data out
+print(val := MallocAllocator._as_buffer(out).cast("I").tolist()[0])
+assert val == 5
+
+
+print("******** second, the Device ***********")
+
+DEVICE = "CLANG"   # NOTE: you can change this!
+
+import struct
+from tinygrad.dtype import dtypes
+from tinygrad.device import Buffer, Device
+from tinygrad.ops import UOp, Ops
+from tinygrad.shape.shapetracker import ShapeTracker
+
+# 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 MallocAllocator.alloc
+
+# describe the computation
+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, ShapeTracker.from_shape((1,)).to_uop()))
+ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2, ShapeTracker.from_shape((1,)).to_uop()))
+alu = ld_1 + ld_2
+output_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
+st_0 = UOp(Ops.STORE, dtypes.void, (output_buf, ShapeTracker.from_shape((1,)).to_uop(), 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_kernel, CompiledRunner
+kernel = get_kernel(Device[DEVICE].renderer, s).linearize()
+
+# compile a program (and print the source)
+fxn = CompiledRunner(kernel.to_program())
+print(fxn.p.src)
+# NOTE: fxn.clprg is the ClangProgram
+
+# run the program
+fxn.exec([out, a, b])
+
+# check the data out
+assert out.as_buffer().cast('I')[0] == 5
+
+
+print("******** third, the LazyBuffer ***********")
+
+from tinygrad.engine.realize import run_schedule
+from tinygrad.engine.schedule import create_schedule
+
+# allocate some values + load in values
+a = UOp.metaop(Ops.EMPTY, (1,), dtypes.int32, DEVICE)
+b = UOp.metaop(Ops.EMPTY, (1,), dtypes.int32, DEVICE)
+a.buffer.allocate().copyin(memoryview(bytearray(struct.pack("I", 2))))
+b.buffer.allocate().copyin(memoryview(bytearray(struct.pack("I", 3))))
+del a.srcs
+del b.srcs
+
+# describe the computation
+out = a.alu(Ops.ADD, b)
+
+# schedule the computation as a list of kernels
+sched = create_schedule([out])
+for si in sched: print(si.ast.op)  # NOTE: the first two convert it to CLANG
+
+# DEBUGGING: print the compute ast
+print(sched[-1].ast)
+# NOTE: sched[-1].ast is the same as st_0 above
+
+# run that schedule
+run_schedule(sched)
+
+# check the data out
+assert out.realized is not None and out.realized.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
@@ -26,31 +26,36 @@ l1n, l2n = l1.numpy(), l2.numpy()
 from tinygrad.nn.optim import SGD
 optim = SGD([l1, l2])

-Tensor.training = True
 X, Y = X_train[(samples:=Tensor.randint(128, high=X_train.shape[0]))], Y_train[samples]
 optim.zero_grad()
 model(X).sparse_categorical_crossentropy(Y).backward()
-optim.schedule_step()   # this will step the optimizer without running realize
+optim._step()   # this will step the optimizer without running realize

 # *****
 # 3. Create a schedule.

 # 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
+# l1.lazydata and l2.lazydata 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] = [ExecItem(lower_schedule_item(si).prg, list(si.bufs)) 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)
@@ -1,39 +0,0 @@
-# AM Driver
-
-AM driver is a userspace driver targeting AMD's RDNA3/RDNA4. You only need tinygrad to send compute tasks to your GPU!
-
-## How to run?
-Make sure that amdgpu module is unloaded and just run tinygrad with `AMD=1`!
-
-Optional requirements:
-
-* System without IOMMU for P2P / SDMA support
-* vfio-pci module for IRQ handling
-
-## Environment Variables
-
-| Variable | Possible Value(s) | Description |
-|----------|------------------|-------------|
-| AM_RESET | [1] | Performs a full GPU reset (reloading all firmware and IP blocks) |
-| AM_DEBUG | [0-4] | Sets the level of additional debugging information |
-
-## AM Driver Details
-
-### Compute & SDMA Queues
-
-AM binds compute queues directly to MEC (bypassing MES). Tinygrad uses only one compute queue, which is bound at `pipe=0 queue=0`. Similarly, the single SDMA queue is bound at `engine=0 queue=0`.
-
-### Boot
-
-The GPU being passed can be in one of several states:
-1. Not initialized
-2. Initialized by amdgpu
-3. Initialized by AM
-
-The first and second states require a full GPU setup since their states are unknown. The second state also requires a mode1 reset to reinitialize all components.
-
-The third state can be set up partially to optimize boot time. In this case, only the GFX and SDMA IPs need to be initialized. To enable this, AM uses a separate boot memory that is guaranteed not to be overwritten. This physical memory is utilized for all blocks that are initialized only during the initial AM boot. To determine if the GPU is in the third state, AM uses `regSCRATCH_REG7` as a flag.
-
-### VM Management
-
-Each AM device sets up only a single `VMID=0` and one page directory. The page directory used is 3-level and thus supports up to 512GB of virtual addresses. All AM devices are located in one virtual address space.
@@ -7,25 +7,23 @@ The tinygrad framework has four pieces

 There is a good [bunch of tutorials](https://mesozoic-egg.github.io/tinygrad-notes/) by Di Zhu that go over tinygrad internals.

-There's also a [doc describing speed](../developer/speed.md)
-
 ## Frontend

-Everything in [Tensor](../tensor/index.md) is syntactic sugar around constructing a graph of [UOps](../developer/uop.md).
+Everything in [Tensor](../tensor/index.md) is syntactic sugar around [function.py](function.md), where the forwards and backwards passes are implemented for the different functions. There's about 25 of them, implemented using about 20 basic ops. Those basic ops go on to construct 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 +1,33 @@
+::: tinygrad.function
+    options:
+        members: [
+            "Contiguous",
+            "ContiguousBackward",
+            "Cast",
+            "Neg",
+            "Reciprocal",
+            "Sin",
+            "Relu",
+            "Log",
+            "Exp",
+            "Sqrt",
+            "Sigmoid",
+            "Sign",
+            "Less",
+            "Eq",
+            "Xor",
+            "Add",
+            "Sub",
+            "Mul",
+            "Div",
+            "Where",
+            "Sum",
+            "Max",
+            "Expand",
+            "Reshape",
+            "Permute",
+            "Pad",
+            "Shrink",
+            "Flip",
+        ]
+        show_source: false
@@ -52,7 +52,7 @@ Signals are device-dependent structures used for synchronization and timing in H
 The following Python code demonstrates the usage of signals:

 ```python
-signal = your_device.new_signal(value=0)
+signal = your_device.signal_t()

 HWQueue().timestamp(signal) \
         .signal(signal, value_to_fire) \
@@ -115,8 +115,9 @@ HCQ-compatible devices use a global timeline signal for synchronizing all operat
 ```python
 HWQueue().wait(your_device.timeline_signal, your_device.timeline_value - 1) \
         .exec(...)
-         .signal(your_device.timeline_signal, your_device.next_timeline()) \
+         .signal(your_device.timeline_signal, your_device.timeline_value) \
         .submit(your_device)
+your_device.timeline_value += 1

 # Optionally wait for execution
 your_device.timeline_signal.wait(your_device.timeline_value - 1)
@@ -1,60 +0,0 @@
-# tinygrad directory layout
-
-This explains the flow of a big graph down to programs.
-
-Directories are listed in order of how they are processed.
-
---
-
-## tinygrad/schedule
-
-Group UOps into kernels.
-
-::: tinygrad.schedule.rangeify.get_rangeify_map
-    options:
-        members: false
-        show_labels: false
-        show_source: false
-
---
-
-## tinygrad/codegen/opt
-
-Transforms the ast into an optimized ast. This is where BEAM search and heuristics live.
-
---
-
-## tinygrad/codegen
-
-Transform the optimized ast into a linearized and rendered program.
-
-::: tinygrad.codegen.full_rewrite_to_program
-    options:
-        members: false
-        show_labels: false
-        show_source: false
-
---
-
-## tinygrad/renderer
-
-Transform the linearized list of UOps into a program, represented as a string.
-
-::: tinygrad.renderer.Renderer
-    options:
-        members:
-            - render
-        show_labels: false
-        show_source: false
-
---
-
-## tinygrad/engine
-
-Abstracted high level interface to the runtimes.
-
-::: tinygrad.engine.realize.get_program
-    options:
-        members: false
-        show_labels: false
-        show_source: false
@@ -36,9 +36,9 @@ The `Allocator` class is responsible for managing memory on the device. There is

 ### Program

-The `Program` class is created for each loaded program. It is responsible for executing the program on the device. As an example, here is a `CPUProgram` implementation which loads program and runs it.
+The `Program` class is created for each loaded program. It is responsible for compiling and executing the program on the device. As an example, here is a `ClangProgram` implementation which loads program and runs it.

-::: tinygrad.runtime.ops_cpu.CPUProgram
+::: tinygrad.runtime.ops_clang.ClangProgram
    options:
        members: true

@@ -1,71 +0,0 @@
-# speed in tinygrad
-
-## Overview
-
-Speed refers to many different things. To break it down to four, there's:
-
- Compile Speed (Python)
- Execution Speed (driver)
- Model Speed (scheduler)
- Kernel Speed (codegen)
-
-## Compile Speed (Python)
-
-This is how long the first run of your model takes. It's limited largely by the runtime of the Python doing UOp rewrites. Currently it's a bit slow, but on par with torch.compile. It gets even slower if you are using BEAM, since that's compiling many variants of each kernel.
-
-This will be improved by writing faster graph_rewrite, doing less graph_rewrite, and better parallelization.
-
-## Execution Speed (driver)
-
-After your model is compiled, you are often using the `TinyJIT`. tinygrad has the best execution speed of any framework because it usually bypasses the GPU driver and prebuilds the command queue. It's tons faster than normal CUDA, and often even faster than CUDA Graph.
-
-There's very little to improve here, as this is almost never the bottleneck.
-
-## Model Speed (scheduler)
-
-The scheduler determines how operations are grouped into kernels and which Tensors are written to memory. This is currently a big bottleneck of training speed.
-
-The decisions are often not obvious. For example, when is it worth recomputing an arithmetic operation instead of storing and loading from memory? Example:
-
-```python
-from tinygrad import Tensor
-a = Tensor.rand(100)
-b = Tensor.rand(100)
-c = Tensor.rand(100)
-d = Tensor.rand(100)
-out1 = a+b+c
-out2 = a+b+d
-Tensor.realize(out1, out2)
-```
-
-The real answer is obvious, compute both `out1` and `out2` in the same kernel. But you can't always do that. If you can't, should `a+b` first be saved to a subbuffer? Or should both the `out1` and `out2` kernels recompute `a+b`?
-
-In this case: with recompute (6 reads + 2 writes), no recompute (6 reads + 3 writes), so we should probably recompute. However, once you add movement ops and casts this is even harder to figure out. tinygrad doesn't yet have a systematic way to do it.
-
-## Kernel Speed (codegen)
-
-Given that you have decided how the model ops will be grouped and what will be written to memory, kernel speed determines how fast that operation is done. This is what BEAM changes, it searches over a set of equivalent kernels which all perform the same operation and finds the one which performs the task the fastest.
-
-In `kernel.py` we have a set of `OptOps`, these control the parameters of the speed optimizations applied to the kernel.
-
-### Memory
-
-The main bottleneck in most kernels is accessing memory. In a freshman algorithms class, you'll learn about cache aware matrix multiplication, and this is all forms of that. While the same math is run, the order in which you run it can have large impacts on the speed depending on if the data you are loading. OptOps will change this order.
-
-Memory, even cache, is often much slower than accessing the register file. The amount of times data is used in math is called the "arithmetic intensity". For operations like BS=1 GEMV, the arithmetic intensity is 1, but for GEMMs and convs it can be much higher. OptOps like UPCAST and UNROLL can increase this, but be careful of making them too large, as if there's too much register pressure on the GPU the warp scheduler may not be able to fit many warps, or even worse, it could be spilling to local memory.
-
-4090s have 1 TB/s of ram bandwidth and ~160 TFLOPS of compute, so you need to use each loaded value ~100 times. The L1 cache has around 40 TB/s of bandwidth, so in order to get full compute utilization you need to use each value ~4 times.
-
-A lot of work can still be done here. For example, we never copy the inputs to on chip SRAM, but this is often quite helpful for kernel speed. Also, we aren't doing a good job with L2 cache awareness (the locals handle L1 quite well)
-
-### Tensor Cores
-
-Many accelerators have Tensor Cores / MAC arrays / systolic arrays. The main value of these is that, since they are 2-D, they create an n^2 ratio between the compute and the input data.
-
-GPUs use Tensor Cores instead of MAC arrays to fit better in the GPU warp paradigm. This is because the output of Tensor Cores is O(n) wrt the input, while the output of MAC arrays like the AMX is O(n^2)
-
-We have a simple framework in tinygrad for adding these ALU blocks and achieving good performance from them.
-
-### Indexing
-
-Indexing determines the address of the memory we need to load. GPUs often have less integer math resources than floating point math, so this can sometimes be the bottleneck. We have a symbolic math engine in our rewrite rules to simplify indexing before it's emitted to the kernel. Newer NVIDIA GPUs have a "Tensor Memory Accelerator" to assist with fast indexing, however, this is not supported in tinygrad yet.
@@ -1,10 +1,10 @@
-::: tinygrad.uop.ops.UOp
+::: tinygrad.ops.UOp
    options:
        members: false
        members_order: source
        show_labels: false

-::: tinygrad.uop.ops.Ops
+::: tinygrad.ops.Ops
    options:
        members: true
        members_order: source
@@ -3,7 +3,7 @@
 This is a list of environment variable that control the runtime behavior of tinygrad and its examples.
 Most of these are self-explanatory, and are usually used to set an option at runtime.

-Example: `CL=1 DEBUG=4 python3 -m pytest`
+Example: `GPU=1 DEBUG=4 python3 -m pytest`

 However you can also decorate a function to set a value only inside that function.

@@ -30,32 +30,21 @@ These control the behavior of core tinygrad even when used as a library.

 Variable | Possible Value(s) | Description
 ---|---|---
-DEBUG               | [1-7]      | enable debugging output (operations, timings, speed, generated code and more)
-CL                  | [1]        | enable OpenCL backend
+DEBUG               | [1-6]      | enable debugging output, with 4 you get operations, timings, speed, generated code and more
+GPU                 | [1]        | enable the GPU backend
 CUDA                | [1]        | enable CUDA backend
 AMD                 | [1]        | enable AMD backend
 NV                  | [1]        | enable NV backend
 METAL               | [1]        | enable Metal backend (for Mac M1 and after)
-CPU                 | [1]        | enable CPU backend
+METAL_XCODE         | [1]        | enable Metal using macOS Xcode SDK
+CLANG               | [1]        | enable Clang backend
+LLVM                | [1]        | enable LLVM backend
 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).
+PTX                 | [1]        | enable the specialized [PTX](https://docs.nvidia.com/cuda/parallel-thread-execution/) assembler for Nvidia GPUs. If not set, defaults to generic CUDA codegen backend.
+PROFILE             | [1]        | enable output of [perfetto](https://ui.perfetto.dev/) compatible profile. This feature is supported in NV and AMD backends.
+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.
-WEBGPU_BACKEND      | [WGPUBackendType_Metal, ...]          | Force select a backend for WebGPU (Metal, DirectX, OpenGL, Vulkan...)
-CUDA_PATH           | str        | Use `CUDA_PATH/include` for CUDA headers for CUDA and NV backends. If not set, TinyGrad will use `/usr/local/cuda/include`, `/usr/include` and `/opt/cuda/include`.
-
-## Debug breakdown
-
-Variable | Value | Description
---|---|---
-DEBUG               | >= 1       | Enables debugging and lists devices being used
-DEBUG               | >= 2       | Provides performance metrics for operations, including timing, memory usage, bandwidth for each kernel execution
-DEBUG               | >= 3       | Outputs buffers used for each kernel (shape, dtype and strides) and the applied optimizations at a kernel level
-DEBUG               | >= 4       | Outputs the generated kernel code
-DEBUG               | >= 5       | Displays the intermediate representation of the computation UOps (AST)
-DEBUG               | >= 6       | Displays the intermediate representation of the computation UOps in a linearized manner, detailing the operation sequence
-DEBUG               | >= 7       | Outputs the assembly code generated for the target hardware
+VIZ                 | [1]        | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
@@ -42,7 +42,7 @@ There's nothing special about a "Module" class in tinygrad, it's just a normal c

 ### tinygrad is functional

-In tinygrad, you can do [`x.conv2d(w, b)`](tensor/ops.md/#tinygrad.Tensor.conv2d) or [`x.sparse_categorical_crossentropy(y)`](tensor/ops.md/#tinygrad.Tensor.sparse_categorical_crossentropy). We do also have a [`Conv2D`](nn.md/#tinygrad.nn.Conv2d) class like PyTorch if you want a place to keep the state, but all stateless operations don't have classes.
+In tinygrad, you can do [`x.conv2d(w, b)`](tensor/ops.md/#tinygrad.Tensor.conv2d) or [`x.sparse_categorical_cross_entropy(y)`](tensor/ops.md/#tinygrad.Tensor.sparse_categorical_crossentropy). We do also have a [`Conv2D`](nn.md/#tinygrad.nn.Conv2d) class like PyTorch if you want a place to keep the state, but all stateless operations don't have classes.

 ### tinygrad is lazy

@@ -17,7 +17,7 @@ from tinygrad import Device
 print(Device.DEFAULT)
 ```

-You will see `CUDA` here on a GPU instance, or `CPU` here on a CPU instance.
+You will see `CUDA` here on a GPU instance, or `CLANG` here on a CPU instance.

 ## A simple model

@@ -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.

@@ -29,12 +29,4 @@
 ::: tinygrad.nn.state.get_state_dict
 ::: tinygrad.nn.state.get_parameters
 ::: tinygrad.nn.state.load_state_dict
-::: tinygrad.nn.state.tar_extract
-    options:
-        show_signature: false
-        separate_signature: false
 ::: tinygrad.nn.state.torch_load
-    options:
-        show_signature: false
-        separate_signature: false
-::: tinygrad.nn.state.gguf_load
@@ -110,7 +110,7 @@ class TinyNet:

  def __call__(self, x):
    x = self.l1(x)
-    x = x.leaky_relu()
+    x = x.leakyrelu()
    x = self.l2(x)
    return x

@@ -118,7 +118,7 @@ net = TinyNet()
 ```

 We can see that the forward pass of our neural network is just the sequence of operations performed on the input tensor `x`.
-We can also see that functional operations like `leaky_relu` are not defined as classes and instead are just methods we can just call.
+We can also see that functional operations like `leakyrelu` are not defined as classes and instead are just methods we can just call.
 Finally, we just initialize an instance of our neural network, and we are ready to start training it.

 ## Training
@@ -1,81 +1,14 @@
 # Runtimes

-tinygrad supports various runtimes, enabling your code to scale across a wide range of devices. The default runtime can be automatically selected based on the available hardware, or you can force a specific runtime to be default using environment variables (e.g., `CPU=1`).
+tinygrad supports various runtimes, enabling your code to scale across a wide range of devices. The default runtime can be automatically selected based on the available hardware, or you can force a specific runtime to be default using environment variables (e.g., `CLANG=1`).

-| Runtime | Description | Compiler Options | Requirements |
-|---------|-------------|------------------|--------------|
-| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`NV_PTX=1`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via `NV_IFACE=(NVK\|PCI)`. See [NV interfaces](#nv-interfaces) for details. |
-| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`AMD_LLVM=1`)<br>HIP/COMGR (`AMD_HIP=1`) | RDNA2 or newer GPUs.<br>You can select an interface via `AMD_IFACE=(KFD\|PCI\|USB)`. See [AMD interfaces](#amd-interfaces) for details. |
-| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
-| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for `bfloat` support |
-| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`CUDA_PTX=1`) | NVIDIA GPU with CUDA support |
-| [CL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | - | OpenCL 2.0 compatible device |
-| [CPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang or llvm compiler | Clang JIT (default)<br>LLVM IR (`CPU_LLVM=1`) | `clang` compiler in system `PATH` |
-| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | - | Dawn library installed and discoverable. Binaries: [pydawn v0.3.0](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0) |
-
-
-## Interoperability
-
-tinygrad provides interoperability with OpenCL and PyTorch, allowing efficient tensor data sharing between frameworks through the `Tensor.from_blob` API. This enables zero-copy operations by working directly with external memory pointers.
-
-**Important**: When using external memory pointers with tinygrad tensors, you must ensure these pointers remain valid throughout the entire lifetime of the tinygrad tensor to prevent memory corruption.
-
-### `CUDA`/`METAL` PyTorch Interoperability
-
-You can seamlessly work with CUDA/MPS tensors between PyTorch and tinygrad without data copying:
-```python
-from tinygrad.dtype import _from_torch_dtype
-tensor1 = torch.tensor([1.0, 2.0, 3.0], device=torch.device("cuda"))
-tiny_tensor1 = Tensor.from_blob(tensor1.data_ptr(), tensor1.shape, dtype=_from_torch_dtype(tensor1.dtype), device='CUDA')
-
-# Before tinygrad calculations, mps needs to be synchronized to make sure data is valid.
-if data.device.type == "mps": torch.mps.synchronize()
-else: torch.cuda.synchronize()
-
-x = (tiny_tensor1 + 1).realize()
-```
-
-### `QCOM` OpenCL Interoperability
-
-tinygrad supports OpenCL interoperability on `QCOM` backend.
-
-Buffer interop allows direct access to OpenCL memory buffers:
-```python
-# create raw opencl buffer.
-cl_buf = cl.clCreateBuffer(cl_context, cl.CL_MEM_READ_WRITE, 0x100, None, status := ctypes.c_int32())
-
-# extract pointers
-cl_buf_desc_ptr = to_mv(ctypes.addressof(cl_buf), 8).cast('Q')[0]
-rawbuf_ptr = to_mv(cl_buf_desc_ptr, 0x100).cast('Q')[20] # offset 0xA0 is a raw gpu pointer.
-
-# create tiny tensor
-tiny = Tensor.from_blob(rawbuf_ptr, (8, 8), dtype=dtypes.int, device='QCOM')
-```
-
-And the same for the images:
-```python
-# create cl image.
-cl_img = cl.clCreateImage2D(cl_context, cl.CL_MEM_READ_WRITE, cl.cl_image_format(cl.CL_RGBA, cl.CL_FLOAT), w, h, 0, None, status := ctypes.c_int32())
-
-# extract pointers
-cl_buf_desc_ptr = to_mv(ctypes.addressof(cl_img), 8).cast('Q')[0]
-rawbuf_ptr = to_mv(cl_buf_desc_ptr, 0x100).cast('Q')[20] # offset 0xA0 is a raw gpu pointer.
-
-# create tiny tensor
-tiny = Tensor.from_blob(rawbuf_ptr, (h*w*4,), dtype=dtypes.imagef((h,w)), device='QCOM')
-```
-
-## AMD Interfaces
-AMD backend supports several interfaces for communicating with devices:
-
-* `KFD`: uses the amdgpu driver
-* `PCI`: uses the [AM driver](developer/am.md)
-* `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.
-
-## NV Interfaces
-NV backend supports several interfaces for communicating with devices:
-
-* `NVK`: uses the nvidia driver
-* `PCI`: uses the [NV driver](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/support/nv/nvdev.py)
+| Runtime | Description | Requirements |
+|---------|-------------|--------------|
+| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | Ampere/Ada series GPUs |
+| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | RDNA2/RDNA3 series GPUs |
+| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | 6xx series GPUs |
+| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | M1+ Macs; Metal 3.0+ for `bfloat` support |
+| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | NVIDIA GPU with CUDA support |
+| [GPU (OpenCL)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_gpu.py) | Accelerates computations using OpenCL on GPUs | OpenCL 2.0 compatible device |
+| [CLANG (C Code)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_clang.py) | Runs on CPU using the clang compiler | `clang` compiler in system `PATH` |
+| [LLVM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_llvm.py) | Runs on CPU using the LLVM compiler infrastructure | `llvmlite` package installed |
@@ -20,11 +20,8 @@

 ::: tinygrad.Tensor.manual_seed
 ::: tinygrad.Tensor.rand
-::: tinygrad.Tensor.rand_like
 ::: tinygrad.Tensor.randn
-::: tinygrad.Tensor.randn_like
 ::: tinygrad.Tensor.randint
-::: tinygrad.Tensor.randperm
 ::: tinygrad.Tensor.normal
 ::: tinygrad.Tensor.uniform
 ::: tinygrad.Tensor.scaled_uniform
@@ -22,7 +22,6 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.round
 ::: tinygrad.Tensor.isinf
 ::: tinygrad.Tensor.isnan
-::: tinygrad.Tensor.isfinite
 ::: tinygrad.Tensor.lerp
 ::: tinygrad.Tensor.square
 ::: tinygrad.Tensor.clamp
@@ -35,7 +34,6 @@ Elementwise ops operate on a per element basis. They don't change the shape of t

 ::: tinygrad.Tensor.relu
 ::: tinygrad.Tensor.sigmoid
-::: tinygrad.Tensor.logsigmoid
 ::: tinygrad.Tensor.hardsigmoid
 ::: tinygrad.Tensor.elu
 ::: tinygrad.Tensor.celu
@@ -54,7 +52,7 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.erf
 ::: tinygrad.Tensor.gelu
 ::: tinygrad.Tensor.quick_gelu
-::: tinygrad.Tensor.leaky_relu
+::: tinygrad.Tensor.leakyrelu
 ::: tinygrad.Tensor.mish
 ::: tinygrad.Tensor.softplus
 ::: tinygrad.Tensor.softsign
@@ -65,20 +63,13 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.sub
 ::: tinygrad.Tensor.mul
 ::: tinygrad.Tensor.div
-::: tinygrad.Tensor.idiv
-::: tinygrad.Tensor.mod
-::: tinygrad.Tensor.bitwise_xor
-::: tinygrad.Tensor.bitwise_and
-::: tinygrad.Tensor.bitwise_or
-::: tinygrad.Tensor.bitwise_not
+::: tinygrad.Tensor.xor
 ::: tinygrad.Tensor.lshift
 ::: tinygrad.Tensor.rshift
 ::: tinygrad.Tensor.pow
 ::: tinygrad.Tensor.maximum
 ::: tinygrad.Tensor.minimum
 ::: tinygrad.Tensor.where
-::: tinygrad.Tensor.copysign
-::: tinygrad.Tensor.logaddexp

 ## Casting Ops

@@ -10,7 +10,6 @@

 ## Movement (high level)

-::: tinygrad.Tensor.__getitem__
 ::: tinygrad.Tensor.gather
 ::: tinygrad.Tensor.cat
 ::: tinygrad.Tensor.stack
@@ -18,7 +17,6 @@
 ::: tinygrad.Tensor.repeat_interleave
 ::: tinygrad.Tensor.split
 ::: tinygrad.Tensor.chunk
-::: tinygrad.Tensor.unfold
 ::: tinygrad.Tensor.meshgrid
 ::: tinygrad.Tensor.squeeze
 ::: tinygrad.Tensor.unsqueeze
@@ -26,6 +24,3 @@
 ::: tinygrad.Tensor.transpose
 ::: tinygrad.Tensor.flatten
 ::: tinygrad.Tensor.unflatten
-::: tinygrad.Tensor.diag
-::: tinygrad.Tensor.roll
-::: tinygrad.Tensor.rearrange
@@ -6,10 +6,8 @@
 ::: tinygrad.Tensor.min
 ::: tinygrad.Tensor.any
 ::: tinygrad.Tensor.all
-::: tinygrad.Tensor.isclose
 ::: tinygrad.Tensor.mean
 ::: tinygrad.Tensor.var
-::: tinygrad.Tensor.var_mean
 ::: tinygrad.Tensor.std
 ::: tinygrad.Tensor.std_mean
 ::: tinygrad.Tensor.softmax
@@ -23,7 +21,6 @@

 ::: tinygrad.Tensor.avg_pool2d
 ::: tinygrad.Tensor.max_pool2d
-::: tinygrad.Tensor.max_unpool2d
 ::: tinygrad.Tensor.conv2d
 ::: tinygrad.Tensor.conv_transpose2d
 ::: tinygrad.Tensor.dot
@@ -35,12 +32,6 @@
 ::: tinygrad.Tensor.tril
 ::: tinygrad.Tensor.interpolate
 ::: tinygrad.Tensor.scatter
-::: tinygrad.Tensor.scatter_reduce
-::: tinygrad.Tensor.masked_select
-::: tinygrad.Tensor.masked_fill
-::: tinygrad.Tensor.sort
-::: tinygrad.Tensor.topk
-::: tinygrad.Tensor.multinomial

 ## Neural Network (functional)

@@ -25,15 +25,10 @@
 ::: tinygrad.Tensor.replace
 ::: tinygrad.Tensor.assign
 ::: tinygrad.Tensor.detach
-::: tinygrad.Tensor.clone
 ::: tinygrad.Tensor.to
 ::: tinygrad.Tensor.to_
 ::: tinygrad.Tensor.shard
 ::: tinygrad.Tensor.shard_
 ::: tinygrad.Tensor.contiguous
 ::: tinygrad.Tensor.contiguous_backward
-
-## Gradient
-
-::: tinygrad.Tensor.gradient
 ::: tinygrad.Tensor.backward
@@ -6,7 +6,7 @@ If you don't have a tinybox and you want one, see [tinygrad.org](https://tinygra

 ## Welcome

-Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, the green box includes six 4090 GPUs, and the green v2 box includes four 5090 GPUs. Whether you bought a red one or a green one, we want you to love it.
+Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, and the green box includes six 4090 GPUs. Whether you bought a red one or a green one, we want you to love it.

 We don't have a stupid cloud service, you don't have to create a tiny account to set it up, and we aren't tracking how you use the box. We're just happy you bought one. This petaflop is your petaflop.

@@ -41,14 +41,14 @@ 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?

 The [default tinybox image](https://github.com/tinygrad/tinyos) ships with tinygrad and PyTorch. While we develop tinygrad, the box is universal hardware. Use whatever framework you desire, run notebooks, download demos, install more things, train, inference, live, laugh, love, you aren't paying per hour for this box so the only limit is your imagination.

-## Building the OS image
+## tinychat

-The OS image is built using `ubuntu-image` from <https://github.com/tinygrad/tinyos>.
+Since LLMs are so popular, we ship with a built in tinygrad based chatbot using a LLaMA-3 finetune. Visit the IP (not the BMC IP) of your tinybox in a web browser on your computer or phone, and you'll find a friendly looking chat interface. This chatbot also provides an OpenAI compatible LLM API on that port, so you can script it.

-After cloning, run `make green` or `make red` to build a tinybox green or tinybox red image respectively.
+The conversations you have with this chatbot are between you and your tinybox. Also, the history in the web app is saved on the client, not the tinybox.
@@ -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,
+];
@@ -78,7 +78,10 @@ if __name__ == "__main__":

  @TinyJit
  def get_action(obs:Tensor) -> Tensor:
+    # TODO: with no_grad
+    Tensor.no_grad = True
    ret = model(obs)[0].exp().multinomial().realize()
+    Tensor.no_grad = False
    return ret

  st, steps = time.perf_counter(), 0
@@ -2,19 +2,15 @@ import time
 start_tm = time.perf_counter()
 import math
 from typing import Tuple, cast
-from tinygrad import Tensor, nn, GlobalCounters, TinyJit, dtypes, Device
+import numpy as np
+from tinygrad import Tensor, nn, GlobalCounters, TinyJit, dtypes
 from tinygrad.helpers import partition, trange, getenv, Context
 from extra.lr_scheduler import OneCycleLR

-GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]
-
-# override tinygrad defaults
 dtypes.default_float = dtypes.half
-Context(FUSE_OPTIM=1).__enter__()

 # from https://github.com/tysam-code/hlb-CIFAR10/blob/main/main.py
 batchsize = getenv("BS", 1024)
-assert batchsize % len(GPUS) == 0, f"{batchsize=} is not a multiple of {len(GPUS)=}"
 bias_scaler = 64
 hyp = {
  'opt': {
@@ -71,7 +67,7 @@ class ConvGroup:
    cast(Tensor, self.norm2.weight).requires_grad = False
  def __call__(self, x:Tensor) -> Tensor:
    x =    self.norm1(self.conv1(x).max_pool2d().float()).cast(dtypes.default_float).quick_gelu()
-    return self.norm2(self.conv2(x).float()).cast(dtypes.default_float).quick_gelu() + x
+    return self.norm2(self.conv2(x).float()).cast(dtypes.default_float).quick_gelu()

 class SpeedyConvNet:
  def __init__(self):
@@ -82,25 +78,23 @@ class SpeedyConvNet:
    self.linear = nn.Linear(depths['block3'], depths['num_classes'], bias=False)
  def __call__(self, x:Tensor) -> Tensor:
    x = self.whiten(x).quick_gelu()
-    # ************* HACKS *************
-    x = x.pad((1,0,0,1)) # TODO: this pad should not be here! copied from hlb_cifar10 for speed
-    # ************* HACKS *************
    x = x.sequential([self.conv_group_1, self.conv_group_2, self.conv_group_3])
    return self.linear(x.max(axis=(2,3))) * hyp['opt']['scaling_factor']

 if __name__ == "__main__":
  # *** dataset ***
  X_train, Y_train, X_test, Y_test = nn.datasets.cifar()
+  # TODO: without this line indexing doesn't fuse!
+  X_train, Y_train, X_test, Y_test = [x.contiguous() for x in [X_train, Y_train, X_test, Y_test]]
  cifar10_std, cifar10_mean = X_train.float().std_mean(axis=(0, 2, 3))
-  def preprocess(X:Tensor) -> Tensor: return ((X - cifar10_mean.view(1, -1, 1, 1)) / cifar10_std.view(1, -1, 1, 1)).cast(dtypes.default_float)
+  def preprocess(X:Tensor, Y:Tensor) -> Tuple[Tensor, Tensor]:
+    return ((X - cifar10_mean.view(1, -1, 1, 1)) / cifar10_std.view(1, -1, 1, 1)).cast(dtypes.default_float), Y.one_hot(depths['num_classes'])

  # *** model ***
  model = SpeedyConvNet()
  state_dict = nn.state.get_state_dict(model)
-  if len(GPUS) > 1:
-    cifar10_std.to_(GPUS)
-    cifar10_mean.to_(GPUS)
-    for x in state_dict.values(): x.to_(GPUS)
+
+  #for k,v in nn.state.torch_load("/tmp/cifar_net.pt").items(): print(k)

  params_bias, params_non_bias = partition(state_dict.items(), lambda x: 'bias' in x[0])
  opt_bias     = nn.optim.SGD([x[1] for x in params_bias],     lr=0.01, momentum=.85, nesterov=True, weight_decay=hyp['opt']['bias_decay'])
@@ -117,44 +111,48 @@ if __name__ == "__main__":
  lr_sched_bias     = OneCycleLR(opt_bias,     max_lr=hyp['opt']['bias_lr'],     pct_start=pct_start, div_factor=initial_div_factor, final_div_factor=1./(initial_div_factor*final_lr_ratio), total_steps=total_train_steps)
  lr_sched_non_bias = OneCycleLR(opt_non_bias, max_lr=hyp['opt']['non_bias_lr'], pct_start=pct_start, div_factor=initial_div_factor, final_div_factor=1./(initial_div_factor*final_lr_ratio), total_steps=total_train_steps)

-  def loss_fn(out:Tensor, Y:Tensor) -> Tensor:
-    ret = out.sparse_categorical_crossentropy(Y, reduction='none', label_smoothing=0.2)
-    return ret.mul(hyp['opt']['loss_scale_scaler']*loss_batchsize_scaler).sum().div(hyp['opt']['loss_scale_scaler'])
+  def loss_fn(out, Y):
+    return out.cross_entropy(Y, reduction='none', label_smoothing=0.2).mul(hyp['opt']['loss_scale_scaler']*loss_batchsize_scaler).sum().div(hyp['opt']['loss_scale_scaler'])

  @TinyJit
  @Tensor.train()
  def train_step(idxs:Tensor) -> Tensor:
-    X, Y = X_train[idxs], Y_train[idxs]
-    if len(GPUS) > 1:
-      X.shard_(GPUS, axis=0)
-      Y.shard_(GPUS, axis=0)
-    out = model(preprocess(X))
+    with Context(SPLIT_REDUCEOP=0, FUSE_ARANGE=1):
+      X = X_train[idxs]
+      Y = Y_train[idxs].realize(X)
+    X, Y = preprocess(X, Y)
+    out = model(X)
    loss = loss_fn(out, Y)
    opt.zero_grad()
    loss.backward()
-    return (loss / (batchsize*loss_batchsize_scaler)).realize(*opt.schedule_step(),
-                                                              *lr_sched_bias.schedule_step(), *lr_sched_non_bias.schedule_step())
+    opt.step()
+    lr_sched_bias.step()
+    lr_sched_non_bias.step()
+    return loss / (batchsize*loss_batchsize_scaler)

  eval_batchsize = 2500
  @TinyJit
+  @Tensor.test()
  def val_step() -> Tuple[Tensor, Tensor]:
+    # TODO with Tensor.no_grad()
+    Tensor.no_grad = True
    loss, acc = [], []
    for i in range(0, X_test.size(0), eval_batchsize):
-      X, Y = X_test[i:i+eval_batchsize], Y_test[i:i+eval_batchsize]
-      if len(GPUS) > 1:
-        X.shard_(GPUS, axis=0)
-        Y.shard_(GPUS, axis=0)
-      out = model(preprocess(X))
+      X, Y = preprocess(X_test[i:i+eval_batchsize], Y_test[i:i+eval_batchsize])
+      out = model(X)
      loss.append(loss_fn(out, Y))
-      acc.append((out.argmax(-1) == Y).sum() / eval_batchsize)
-    return Tensor.stack(*loss).mean() / (batchsize*loss_batchsize_scaler), Tensor.stack(*acc).mean()
-
-  Tensor.manual_seed(1337)
-  num_train_samples = X_train.shape[0]
+      acc.append((out.argmax(-1).one_hot(depths['num_classes']) * Y).sum() / eval_batchsize)
+    ret = Tensor.stack(*loss).mean() / (batchsize*loss_batchsize_scaler), Tensor.stack(*acc).mean()
+    Tensor.no_grad = False
+    return ret

+  np.random.seed(1337)
  for epoch in range(math.ceil(hyp['misc']['train_epochs'])):
+    # TODO: move to tinygrad
    gst = time.perf_counter()
-    tidxs = Tensor.randperm(num_train_samples, dtype='int')[:num_steps_per_epoch*batchsize].reshape(num_steps_per_epoch, batchsize)
+    idxs = np.arange(X_train.shape[0])
+    np.random.shuffle(idxs)
+    tidxs = Tensor(idxs, dtype='int')[:num_steps_per_epoch*batchsize].reshape(num_steps_per_epoch, batchsize)  # NOTE: long doesn't fold
    train_loss:float = 0
    for epoch_step in (t:=trange(num_steps_per_epoch)):
      st = time.perf_counter()
@@ -1,12 +1,12 @@
-# model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
-from typing import Callable
+# model based off https://towardsdatascience.com/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
+from typing import List, Callable
 from tinygrad import Tensor, TinyJit, nn, GlobalCounters
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

 class Model:
  def __init__(self):
-    self.layers: list[Callable[[Tensor], Tensor]] = [
+    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,
@@ -21,17 +21,20 @@ 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(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])
+    # TODO: this "gather" of samples is very slow. will be under 5s when this is fixed
    loss = model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
-    return loss.realize(*opt.schedule_step())
+    opt.step()
+    return loss

  @TinyJit
+  @Tensor.test()
  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100

  test_acc = float('nan')
@@ -1,35 +0,0 @@
-import sys, time
-from tinygrad import TinyJit, GlobalCounters, fetch, getenv
-from tinygrad.nn.onnx import OnnxRunner
-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, optimize=True)
-  return run_onnx_jit, run_onnx.graph_inputs
-
-if __name__ == "__main__":
-  onnx_file = fetch(sys.argv[1])
-  run_onnx_jit, input_specs = load_onnx_model(onnx_file)
-  print("loaded model")
-
-  for i in range(3):
-    new_inputs = get_example_inputs(input_specs)
-    GlobalCounters.reset()
-    print(f"run {i}")
-    run_onnx_jit(**new_inputs)
-
-  # run 20 times
-  for _ in range(20):
-    new_inputs = get_example_inputs(input_specs)
-    GlobalCounters.reset()
-    st = time.perf_counter()
-    out = run_onnx_jit(**new_inputs)
-    mt = time.perf_counter()
-    val = out.numpy()
-    et = time.perf_counter()
-    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {(et-st)*1e3:6.2f} ms")
-
-  if getenv("ORT"):
-    validate(onnx_file, new_inputs, rtol=1e-3, atol=1e-3)
-    print("model validated")
@@ -0,0 +1,95 @@
+#!/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, Device, dtypes
+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__":
+  Tensor.no_grad = True
+
+  # 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, model, 32, 8)), strict=False)
+    nn.state.load_state_dict(model, fix_bf16(convert_from_huggingface(part2, model, 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("")
@@ -15,9 +15,9 @@ if __name__ == "__main__":
  if getenv("WEBGPU"):
    safe_save(get_state_dict(model), (dirname / "net.safetensors").as_posix())
    load_state_dict(model, safe_load(str(dirname / "net.safetensors")))
-  mode = "clang" if getenv("CPU", "") != "" else "webgpu" if getenv("WEBGPU", "") != "" else ""
+  mode = "clang" if getenv("CLANG", "") != "" else "webgpu" if getenv("WEBGPU", "") != "" else ""
  prg, inp_sizes, out_sizes, state = export_model(model, mode, Tensor.randn(1,3,224,224))
-  if getenv("CPU", "") == "":
+  if getenv("CLANG", "") == "":
    ext = "js" if getenv("WEBGPU", "") != "" else "json"
    with open(dirname / f"net.{ext}", "w") as text_file:
      text_file.write(prg)
@@ -68,6 +68,6 @@ if __name__ == "__main__":
    else printf("%s\\n", lbls[best_idx]);
  }""")

-    # CPU=1 python3 examples/compile_efficientnet.py | clang -O2 -lm -x c - -o recognize && DEBUG=1 time ./recognize docs/showcase/stable_diffusion_by_tinygrad.jpg
+    # CLANG=1 python3 examples/compile_efficientnet.py | clang -O2 -lm -x c - -o recognize && DEBUG=1 time ./recognize docs/showcase/stable_diffusion_by_tinygrad.jpg
    # category : 281 (tabby, tabby cat) with 9.452788
    print('\n'.join(cprog))
@@ -1,16 +1,15 @@
 # An example to compile a small Tensorflow model to extremely portable C code

 import os, sys
-os.environ["CPU"] = '1'
+os.environ["CLANG"] = '1'
 os.environ["JIT"] = '2'

 import numpy as np
 import subprocess
 import tensorflow as tf
 import tf2onnx
-from tinygrad.nn.onnx import OnnxRunner
+from extra.onnx import get_run_onnx
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import to_mv
 from extra.export_model import export_model_clang, compile_net, jit_model

 def get_uncompiled_model2(dataset_size=32, output_size=4):
@@ -26,7 +25,7 @@ class TinyOnnx:
  def __init__(self, keras_model):
    input_signature = [tf.TensorSpec([1,32], tf.float32, name='x')]
    onnx_model, _ = tf2onnx.convert.from_keras(keras_model, input_signature, opset=13)
-    self.run_onnx = OnnxRunner(Tensor(onnx_model.SerializeToString(), device="PYTHON"))
+    self.run_onnx = get_run_onnx(onnx_model)

  def forward(self, x):
    return self.run_onnx({"x": x}, debug=False)['predictions']
@@ -48,8 +47,8 @@ def compile_onnx_model(onnx_model):
  cprog.append("void initialize(float *weights) {")
  weights = bytes()
  for name,cl in bufs_to_save.items():
-    cprog.append(f"memcpy({name}, weights + {len(weights)//4}, {cl._buf.size});")
-    weights += bytes(to_mv(cl._buf.va_addr, cl._buf.size))
+    cprog.append(f"memcpy({name}, weights + {len(weights)//4}, {len(cl._buf)*4});")
+    weights += bytes(cl._buf)
  cprog.append("}")

  # write the weights to disk
@@ -0,0 +1,343 @@
+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]),), 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.
+  Tensor.no_grad = True
+  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")
+  Tensor.no_grad = True
+  # 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"))
@@ -1,13 +1,12 @@
 #!/usr/bin/env python3
-import os, argparse, contextlib
+import os, argparse
 from typing import Optional, Union
-with contextlib.suppress(ImportError): import tiktoken
+import tiktoken
 from tinygrad import Tensor, TinyJit, Device, GlobalCounters, Variable, dtypes
-from tinygrad.uop.ops import UOp
+from tinygrad.ops import UOp
 from tinygrad.helpers import Timing, DEBUG, JIT, getenv, fetch, colored, trange
 from tinygrad.nn import Embedding, Linear, LayerNorm
 from tinygrad.nn.state import gguf_load, torch_load, load_state_dict, get_state_dict
-from extra.bench_log import BenchEvent, WallTimeEvent

 MAX_CONTEXT = getenv("MAX_CONTEXT", 128)
 HALF = getenv("HALF")
@@ -26,8 +25,8 @@ class Attention:
      start_pos = start_pos.val

    if HALF: x = x.half()
-    xqkv = self.c_attn(x).reshape(None, None, 3, self.n_heads, self.head_dim)
-    xq, xk, xv = [xqkv[:, :, i, :, :] for i in range(3)]
+    xqkv = self.c_attn(x)
+    xq, xk, xv = [xqkv.shrink((None, None, (i*self.dim, (i+1)*self.dim))).reshape(None, None, self.n_heads, self.head_dim) for i in range(3)]
    bsz, seqlen, _, _ = xq.shape

    # create kv cache
@@ -35,11 +34,11 @@ class Attention:
      self.cache_kv = Tensor.zeros(2, bsz, MAX_CONTEXT, self.n_heads, self.head_dim, dtype=x.dtype).contiguous().realize()

    # update the cache
-    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()
+    self.cache_kv.shrink((None, None,(start_pos,start_pos+seqlen),None,None)).assign(Tensor.stack(xk, xv)).realize()

    if start_pos > 0:
-      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
-      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
+      keys = self.cache_kv[0].shrink((None, (0, start_pos+seqlen), None, None))
+      values = self.cache_kv[1].shrink((None, (0, start_pos+seqlen), None, None))
    else:
      keys = xk
      values = xv
@@ -64,7 +63,7 @@ class TransformerBlock:

  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]):
    h = x + self.attn(self.ln_1(x), start_pos, mask).float()
-    return (h + self.mlp(self.ln_2(h))).contiguous()
+    return (h + self.mlp(self.ln_2(h)))

 class Transformer:
  def __init__(self, dim, n_heads, n_layers, norm_eps, vocab_size, max_seq_len=1024):
@@ -85,10 +84,7 @@ class Transformer:
      seqlen = tokens.shape[1]
      tok_emb = self.wte(tokens)

-    # not symbolic when consuming the prompt
-    selected_pos = (0, seqlen) if start_pos.val == 0 else (start_pos, start_pos+1)
-    pos_emb = self.wpe(self.allpos.shrink((None, selected_pos)))
-
+    pos_emb = self.wpe(self.allpos.shrink((None, (start_pos, start_pos+seqlen))))
    h = tok_emb + pos_emb

    if HALF: h = h.half()
@@ -138,12 +134,11 @@ class GPT2:
    # lm head and wte are tied
    weights['lm_head.weight'] = weights['wte.weight']

-    with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-      load_state_dict(model, weights)
+    load_state_dict(model, weights)

-      if HALF:
-        for l in get_state_dict(model).values():
-          l.replace(l.half().realize())
+    if HALF:
+      for l in get_state_dict(model).values():
+        l.replace(l.half().realize())

    return GPT2(model, tokenizer)

@@ -172,8 +167,7 @@ class GPT2:
      return key
    state_dict = { _remap_gguf_key(k): v for k, v in state_dict.items() }
    model = Transformer(**gpt2_params)
-    with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-      load_state_dict(model, state_dict)
+    load_state_dict(model, state_dict)
    return GPT2(model, tiktoken.get_encoding("gpt2"))

  def __init__(self, model, tokenizer):
@@ -181,7 +175,6 @@ class GPT2:
    self.tokenizer = tokenizer

  def generate(self, prompt:str, max_length:int, temperature:float, timing:bool=False, batch_size:int=1):
-    step_times = []
    prompt_tokens = self.tokenizer.encode(prompt, allowed_special={"<|endoftext|>"})
    toks = [prompt_tokens[:] for _ in range(batch_size)]
    start_pos = 0
@@ -189,27 +182,22 @@ class GPT2:
      GlobalCounters.reset()
      if timing: print("")
      st = GlobalCounters.time_sum_s
-      with Timing("ran model in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
+      with Timing("ran model in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
                  f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
                  (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=timing):
-        with WallTimeEvent(BenchEvent.STEP):
-          if batch_size == 1 and len(toks[0][start_pos:]) == 1:
-            tokens = Variable("tokens", 0, VOCAB_SIZE-1).bind(toks[0][start_pos])
-          else:
-            tokens = Tensor([x[start_pos:] for x in toks])
-          tok = self.model(tokens, Variable("start_pos", 1 if start_pos else 0, MAX_CONTEXT-1).bind(start_pos), temperature).tolist()
-      step_times.append((GlobalCounters.time_sum_s-st)*1e3)
+        if batch_size == 1 and len(toks[0][start_pos:]) == 1:
+          tokens = Variable("tokens", 0, VOCAB_SIZE).bind(toks[0][start_pos])
+        else:
+          tokens = Tensor([x[start_pos:] for x in toks])
+        tok = self.model(tokens, Variable("start_pos", 1 if start_pos else 0, MAX_CONTEXT).bind(start_pos), temperature).tolist()
      start_pos = len(toks[0])
      for i,t in enumerate(tok): toks[i].append(t)
-
-    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"
    return [self.tokenizer.decode(x) for x in toks]

 # **** main code ****

 if __name__ == "__main__":
+  Tensor.no_grad = True
  print(f"using {Device.DEFAULT} backend")
  default_prompt = "What is the answer to life, the universe, and everything?"

@@ -232,7 +220,7 @@ 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}%")
@@ -0,0 +1,134 @@
+from typing import List, Tuple
+from extra.models.resnet import ResNet50
+from extra.mcts_search import mcts_search
+from examples.mlperf.helpers import get_mlperf_bert_model
+from tinygrad import Tensor, Device, dtypes, nn
+from tinygrad.codegen.kernel import Kernel
+from tinygrad.ops import Ops, sym_infer
+from tinygrad.device import Compiled
+from tinygrad.engine.schedule import create_schedule
+from tinygrad.engine.search import time_linearizer, beam_search, bufs_from_lin
+from tinygrad.helpers import DEBUG, ansilen, getenv, colored, TRACEMETA
+
+def get_sched_resnet():
+  mdl = ResNet50()
+  optim = (nn.optim.LARS if getenv("LARS") else nn.optim.SGD)(nn.state.get_parameters(mdl))
+  BS = getenv("BS", 64)
+
+  # run model twice to get only what changes, these are the kernels of the model
+  for _ in range(2):
+    out = mdl(Tensor.empty(BS, 3, 224, 224))
+    targets = [out.lazydata]
+    if getenv("BACKWARD"):
+      optim.zero_grad()
+      out.sparse_categorical_crossentropy(Tensor.empty(BS, dtype=dtypes.int)).backward()
+      targets += [x.lazydata for x in optim.schedule_step()]
+    sched = create_schedule(targets)
+    print(f"schedule length {len(sched)}")
+  return sched
+
+def get_sched_bert():
+  mdl = get_mlperf_bert_model()
+  optim = nn.optim.LAMB(nn.state.get_parameters(mdl))
+
+  # fake data
+  BS = getenv("BS", 9)
+  input_ids = Tensor.empty((BS, 512), dtype=dtypes.float32)
+  segment_ids = Tensor.empty((BS, 512), dtype=dtypes.float32)
+  attention_mask = Tensor.empty((BS, 512), dtype=dtypes.default_float)
+  masked_positions = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  masked_lm_ids = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  masked_lm_weights = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  next_sentence_labels = Tensor.empty((BS, 1), dtype=dtypes.float32)
+
+  # run model twice to get only what changes, these are the kernels of the model
+  seen = set()
+  for _ in range(2):
+    lm_logits, seq_relationship_logits = mdl(input_ids, attention_mask, masked_positions, segment_ids)
+    targets = [lm_logits.lazydata, seq_relationship_logits.lazydata]
+    if getenv("BACKWARD"):
+      optim.zero_grad()
+      loss = mdl.loss(lm_logits, seq_relationship_logits, masked_lm_ids, masked_lm_weights, next_sentence_labels)
+      # ignore grad norm and loss scaler for now
+      loss.backward()
+      targets += [x.lazydata for x in optim.schedule_step()]
+    sched = create_schedule(targets)
+    print(f"schedule length {len(sched)}")
+  return sched
+
+if __name__ == "__main__":
+  if getenv("HALF", 1):
+    dtypes.default_float = dtypes.half
+
+  # the device we are optimizing for
+  device: Compiled = Device[Device.DEFAULT]
+  if getenv("BACKWARD"): Tensor.training = True
+  print(f"optimizing for {Device.DEFAULT}")
+
+  sched = globals()[f"get_sched_{getenv('MODEL', 'resnet')}"]()
+  sched = [x for x in sched if x.ast.op is Ops.SINK]
+
+  # focus on one kernel
+  if getenv("KERNEL", -1) >= 0: sched = sched[getenv("KERNEL", -1):getenv("KERNEL", -1)+1]
+
+  # work with the schedule
+  total_tm = 0
+  running_gflops = 0
+  usage = {}
+  for i,si in enumerate(sched):
+    if DEBUG >= 3: print(si.ast)
+
+    rawbufs = bufs_from_lin(Kernel(si.ast))
+
+    # "linearize" the op into uops in different ways
+    lins: List[Tuple[Kernel, str]] = []
+
+    # always try hand coded opt
+    lin = Kernel(si.ast, opts=device.renderer)
+    lin.hand_coded_optimizations()
+    lins.append((lin, "HC"))
+
+    # maybe try tensor cores
+    lin = Kernel(si.ast, opts=device.renderer)
+    if lin.apply_tensor_cores():
+      lins.append((lin, "TC"))
+
+    # try a beam search
+    if beam:=getenv("BEAM"):
+      lin = Kernel(si.ast, opts=device.renderer)
+      lin = beam_search(lin, rawbufs, beam, bool(getenv("BEAM_ESTIMATE", 1)))
+      lins.append((lin, "BEAM"))
+
+    # try MCTS
+    if mcts:=getenv("MCTS"):
+      lin = Kernel(si.ast, opts=device.renderer)
+      lin = mcts_search(lin, rawbufs, mcts)
+      lins.append((lin, "MCTS"))
+
+    # benchmark the programs
+    choices = []
+    for lin, nm in lins:
+      tm = time_linearizer(lin, rawbufs, allow_test_size=False, cnt=10, disable_cache=True)
+      ops = (prg:=lin.to_program()).op_estimate
+      gflops = sym_infer(ops, {k:k.min for k in lin.ast.variables()})*1e-9/tm
+      choices.append((tm, gflops, lin, prg, nm))
+
+    sorted_choices = sorted(choices, key=lambda x: x[0])
+    if DEBUG >= 1: # print all kernels
+      for tm, gflops, lin, prg, nm in choices:
+        print(f"                 kernel {i:2d} {lin.name+' '*(37-ansilen(lin.name))} {str(prg.global_size):18s} {str(prg.local_size):12s} takes {tm*1000:7.2f} ms, {gflops:6.0f} GFLOPS -- {colored(nm, 'green') if lin is sorted_choices[0][2] else nm}")
+
+    tm, gflops, lin, prg, nm = sorted_choices[0]
+    if getenv("SRC"):
+      print(si.ast)
+      print(lin.applied_opts)
+      print(lin.to_program().src)
+    total_tm += tm
+    running_gflops += gflops * tm
+    if (key := str([str(m) for m in si.metadata])) not in usage: usage[key] = (0, 0)
+    usage[key] = (usage[key][0] + tm, usage[key][1] + 1)
+    print(f"*** {total_tm*1000:7.2f} ms : kernel {i:2d} {lin.name+' '*(37-ansilen(lin.name))} {str(prg.global_size):18s} {str(prg.local_size):12s} takes {tm*1000:7.2f} ms, {gflops:6.0f} GFLOPS {[repr(m) if TRACEMETA >= 2 else str(m) for m in si.metadata]}")
+  print(f"******* total {total_tm*1000:.2f} ms, {running_gflops/total_tm:6.0f} GFLOPS")
+  print("usage:")
+  for k in sorted(usage, key=lambda x: -usage[x][0])[:10]:
+    print(f"{usage[k][0]*1000:.2f} ms: {k} ({usage[k][1]} times)")
@@ -7,11 +7,11 @@ import random, time
 import numpy as np
 from typing import Optional
 from extra.lr_scheduler import OneCycleLR
-from tinygrad import nn, dtypes, Tensor, Device, GlobalCounters, TinyJit, Variable
-from tinygrad.nn.state import get_state_dict
+from tinygrad import nn, dtypes, Tensor, Device, GlobalCounters, TinyJit
+from tinygrad.nn.state import get_state_dict, get_parameters
 from tinygrad.nn import optim
 from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod
-from extra.bench_log import BenchEvent, WallTimeEvent
+from tinygrad.multi import MultiLazyBuffer

 cifar_mean = [0.4913997551666284, 0.48215855929893703, 0.4465309133731618]
 cifar_std = [0.24703225141799082, 0.24348516474564, 0.26158783926049628]
@@ -35,6 +35,8 @@ class UnsyncedBatchNorm:
    self.num_batches_tracked = Tensor.zeros(1, dtype=dtypes.int, requires_grad=False)

  def __call__(self, x:Tensor):
+    if isinstance(x.lazydata, MultiLazyBuffer): assert x.lazydata.axis is None or x.lazydata.axis == 0 and len(x.lazydata.lbs) == self.num_devices
+
    xr = x.reshape(self.num_devices, -1, *x.shape[1:]).cast(dtypes.float32)
    batch_mean, batch_invstd = self.calc_stats(xr)
    ret = xr.batchnorm(
@@ -118,7 +120,7 @@ class SpeedyResNet:
 # hyper-parameters were exactly the same as the original repo
 bias_scaler = 58
 hyp = {
-  'seed' : 201,
+  'seed' : 209,
  'opt': {
    'bias_lr':            1.76 * bias_scaler/512,
    'non_bias_lr':        1.76 / 512,
@@ -201,38 +203,24 @@ def train_cifar():
    idx_y = Tensor.arange(H, dtype=dtypes.int32).reshape((1,1,H,1))
    return (idx_x >= low_x) * (idx_x < (low_x + mask_size)) * (idx_y >= low_y) * (idx_y < (low_y + mask_size))

-  # Similar, but different enough.
-  def make_random_crop_indices(shape, mask_size) -> Tensor:
-    BS, _, H, W = shape
-    low_x = Tensor.randint(BS, low=0, high=W-mask_size).reshape(BS,1,1,1)
-    low_y = Tensor.randint(BS, low=0, high=H-mask_size).reshape(BS,1,1,1)
-    idx_x = Tensor.arange(mask_size, dtype=dtypes.int32).reshape((1,1,1,mask_size))
-    idx_y = Tensor.arange(mask_size, dtype=dtypes.int32).reshape((1,1,mask_size,1))
-    return low_x, low_y, idx_x, idx_y
-
  def random_crop(X:Tensor, crop_size=32):
-    Xs, Ys, Xi, Yi = make_random_crop_indices(X.shape, crop_size)
-    return X.gather(-1, (Xs + Xi).expand(-1, 3, X.shape[2], -1)).gather(-2, ((Ys+Yi).expand(-1, 3, crop_size, crop_size)))
+    mask = make_square_mask(X.shape, crop_size)
+    mask = mask.expand((-1,3,-1,-1))
+    X_cropped = Tensor(X.numpy()[mask.numpy()])
+    return X_cropped.reshape((-1, 3, crop_size, crop_size))

-  def cutmix(X, Y, order, mask_size=3):
+  def cutmix(X:Tensor, Y:Tensor, mask_size=3):
+    # fill the square with randomly selected images from the same batch
    mask = make_square_mask(X.shape, mask_size)
-    X_patch, Y_patch = X[order], Y[order]
+    order = list(range(0, X.shape[0]))
+    random.shuffle(order)
+    X_patch = Tensor(X.numpy()[order], device=X.device, dtype=X.dtype)
+    Y_patch = Tensor(Y.numpy()[order], device=Y.device, dtype=Y.dtype)
    X_cutmix = mask.where(X_patch, X)
    mix_portion = float(mask_size**2)/(X.shape[-2]*X.shape[-1])
    Y_cutmix = mix_portion * Y_patch + (1. - mix_portion) * Y
    return X_cutmix, Y_cutmix

-  @TinyJit
-  def augmentations(X:Tensor, Y:Tensor):
-    perms = Tensor.randperm(X.shape[0], device=X.device) # We reuse perms for cutmix, because they are expensivne to generate
-    if getenv("RANDOM_CROP", 1):
-      X = random_crop(X, crop_size=32)
-    if getenv("RANDOM_FLIP", 1):
-      # NOTE: RANGEIFY=1 needs this contiguous or the X[perms] is very slow
-      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X).contiguous() # flip LR
-    X, Y = X[perms], Y[perms]
-    return X, Y, *cutmix(X, Y, perms, mask_size=hyp['net']['cutmix_size'])
-
  # the operations that remain inside batch fetcher is the ones that involves random operations
  def fetch_batches(X_in:Tensor, Y_in:Tensor, BS:int, is_train:bool):
    step, epoch = 0, 0
@@ -240,16 +228,28 @@ def train_cifar():
      st = time.monotonic()
      X, Y = X_in, Y_in
      if is_train:
-        X, Y, X_cm, Y_cm = augmentations(X, Y)
-        if getenv("CUTMIX", 1) and step >= hyp['net']['cutmix_steps']: X, Y = X_cm, Y_cm
+        # TODO: these are not jitted
+        if getenv("RANDOM_CROP", 1):
+          X = random_crop(X, crop_size=32)
+        if getenv("RANDOM_FLIP", 1):
+          X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X) # flip LR
+        if getenv("CUTMIX", 1):
+          if step >= hyp['net']['cutmix_steps']:
+            X, Y = cutmix(X, Y, mask_size=hyp['net']['cutmix_size'])
+        order = list(range(0, X.shape[0]))
+        random.shuffle(order)
+        X, Y = X.numpy()[order], Y.numpy()[order]
+      else:
+        X, Y = X.numpy(), Y.numpy()
      et = time.monotonic()
      print(f"shuffling {'training' if is_train else 'test'} dataset in {(et-st)*1e3:.2f} ms ({epoch=})")
-
-      vi = Variable("i", 0, (full_batches := (X.shape[0] // BS) * BS) - BS)
-      for i in range(0, full_batches, BS):
+      for i in range(0, X.shape[0], BS):
+        # pad the last batch  # TODO: not correct for test
+        batch_end = min(i+BS, Y.shape[0])
+        x = Tensor(X[batch_end-BS:batch_end], device=X_in.device, dtype=X_in.dtype)
+        y = Tensor(Y[batch_end-BS:batch_end], device=Y_in.device, dtype=Y_in.dtype)
        step += 1
-        vib = vi.bind(i)
-        yield X[vib:vib+BS], Y[vib:vib+BS]
+        yield x, y
      epoch += 1
      if not is_train: break

@@ -269,10 +269,13 @@ def train_cifar():

    @TinyJit
    def update(self, net, decay):
+      # TODO with Tensor.no_grad()
+      Tensor.no_grad = True
      for net_ema_param, (param_name, net_param) in zip(get_state_dict(self.net_ema).values(), get_state_dict(net).items()):
        # batchnorm currently is not being tracked
        if not ("num_batches_tracked" in param_name) and not ("running" in param_name):
          net_ema_param.assign(net_ema_param.detach()*decay + net_param.detach()*(1.-decay)).realize()
+      Tensor.no_grad = False

  set_seed(getenv('SEED', hyp['seed']))

@@ -355,7 +358,7 @@ def train_cifar():

  # https://www.anandtech.com/show/16727/nvidia-announces-geforce-rtx-3080-ti-3070-ti-upgraded-cards-coming-in-june
  # 136 TFLOPS is the theoretical max w float16 on 3080 Ti
-  step_times = []
+
  model_ema: Optional[modelEMA] = None
  projected_ema_decay_val = hyp['ema']['decay_base'] ** hyp['ema']['every_n_steps']
  i = 0
@@ -395,35 +398,27 @@ def train_cifar():
      if STEPS == 0 or i == STEPS: break

      GlobalCounters.reset()
+      X, Y = next(batcher)
+      if len(GPUS) > 1:
+        X.shard_(GPUS, axis=0)
+        Y.shard_(GPUS, axis=0)

-      with WallTimeEvent(BenchEvent.STEP):
-        X, Y = next(batcher)
-        if len(GPUS) > 1:
-          X.shard_(GPUS, axis=0)
-          Y.shard_(GPUS, axis=0)
-
-        with Context(BEAM=getenv("LATEBEAM", BEAM.value), WINO=getenv("LATEWINO", WINO.value)):
-          loss = train_step_jitted(model, optim.OptimizerGroup(opt_bias, opt_non_bias), [lr_sched_bias, lr_sched_non_bias], X, Y)
-          et = time.monotonic()
-          loss_cpu = loss.numpy()
-        # EMA for network weights
-        if getenv("EMA") and i > hyp['ema']['steps'] and (i+1) % hyp['ema']['every_n_steps'] == 0:
-          if model_ema is None:
-            model_ema = modelEMA(W, model)
-          model_ema.update(model, Tensor([projected_ema_decay_val*(i/STEPS)**hyp['ema']['decay_pow']]))
-
+      with Context(BEAM=getenv("LATEBEAM", BEAM.value), WINO=getenv("LATEWINO", WINO.value)):
+        loss = train_step_jitted(model, optim.OptimizerGroup(opt_bias, opt_non_bias), [lr_sched_bias, lr_sched_non_bias], X, Y)
+        et = time.monotonic()
+        loss_cpu = loss.numpy()
+      # EMA for network weights
+      if getenv("EMA") and i > hyp['ema']['steps'] and (i+1) % hyp['ema']['every_n_steps'] == 0:
+        if model_ema is None:
+          model_ema = modelEMA(W, model)
+        model_ema.update(model, Tensor([projected_ema_decay_val*(i/STEPS)**hyp['ema']['decay_pow']]))
      cl = time.monotonic()
-      step_times.append((cl-st)*1000.0)
      device_str = loss.device if isinstance(loss.device, str) else f"{loss.device[0]} * {len(loss.device)}"
      #  53  221.74 ms run,    2.22 ms python,  219.52 ms CL,  803.39 loss, 0.000807 LR, 4.66 GB used,   3042.49 GFLOPS,    674.65 GOPS
      print(f"{i:3d} {(cl-st)*1000.0:7.2f} ms run, {(et-st)*1000.0:7.2f} ms python, {(cl-et)*1000.0:7.2f} ms {device_str}, {loss_cpu:7.2f} loss, {opt_non_bias.lr.numpy()[0]:.6f} LR, {GlobalCounters.mem_used/1e9:.2f} GB used, {GlobalCounters.global_ops*1e-9/(cl-st):9.2f} GFLOPS, {GlobalCounters.global_ops*1e-9:9.2f} GOPS")
      st = cl
      i += 1

-  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"
-
  # verify eval acc
  if target := getenv("TARGET_EVAL_ACC_PCT", 0.0):
    if eval_acc_pct >= target:
@@ -432,5 +427,4 @@ def train_cifar():
      raise ValueError(colored(f"{eval_acc_pct=} < {target}", "red"))

 if __name__ == "__main__":
-  with WallTimeEvent(BenchEvent.FULL):
-    train_cifar()
+  train_cifar()
@@ -13,7 +13,6 @@ from extra.models.llama import Transformer, convert_from_huggingface, fix_bf16
 from sentencepiece import SentencePieceProcessor
 import tiktoken, sys
 from tiktoken.load import load_tiktoken_bpe
-from extra.bench_log import BenchEvent, WallTimeEvent

 MAX_CONTEXT = getenv("MAX_CONTEXT", 4096)

@@ -207,42 +206,40 @@ class LLaMa:

    model = Transformer(**params["args"], linear=linear, max_context=MAX_CONTEXT, jit=bool(JIT))

-    with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-      if model_path.is_dir():
-        weights = concat_weights([load(filename) for filename in [f"{model_path}/consolidated.{i:02d}.pth" for i in range(params["files"])]], device[0] if isinstance(device, tuple) else device)
-      else:
-        weights = load(str(model_path))
-      if "model.embed_tokens.weight" in weights:
-        weights = convert_from_huggingface(weights, params["args"]["n_layers"], params["args"]["n_heads"], params["args"].get("n_kv_heads", params["args"]["n_heads"]))
+    if model_path.is_dir():
+      weights = concat_weights([load(filename) for filename in [f"{model_path}/consolidated.{i:02d}.pth" for i in range(params["files"])]], device[0] if isinstance(device, tuple) else device)
+    else:
+      weights = load(str(model_path))
+    if "model.embed_tokens.weight" in weights:
+      weights = convert_from_huggingface(weights, model, params["args"]["n_heads"], params["args"].get("n_kv_heads", params["args"]["n_heads"]))

-      weights = fix_bf16(weights)
+    weights = fix_bf16(weights)

-      # prevent tracking model weights
-      # this is a part of a larger problem with BUFFER UOps and gc in TRACK_MATCH_STATS=2
-      with Context(BEAM=0, TRACK_MATCH_STATS=0):
-        # quantize
-        if quantize is not None:
-          weights = linear.quantize(weights, device)
-          for _,v in weights.items(): v.realize()
+    with Context(BEAM=0):
+      # quantize
+      if quantize is not None:
+        weights = linear.quantize(weights, device)
+        for _,v in weights.items(): v.realize()

-        # shard
-        if isinstance(device, tuple):
-          for k,v in nn.state.get_state_dict(model).items():
-            if 'scale' in k: v.shard_(device, axis=None)  # from quantized
-            elif '.attention.' in k:
-              if getenv("SHARD_KVCACHE") and ('.wq.' in k or '.wk.' in k or '.wv.' in k): v.shard_(device, axis=0)
-              else: v.shard_(device, axis=-1)
-            elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
-            elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
-            elif '.feed_forward.' in k: v.shard_(device, axis=-1)
-            elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
-            elif 'output.weight' in k: v.shard_(device, axis=-1)
-            #elif k.endswith('.weight'): v.shard_(device, axis=-1)
-            #elif 'norm.' in k: v.shard_(device, axis=-1)
-            else: v.shard_(device, axis=None)
+      # shard
+      if isinstance(device, tuple):
+        for k,v in nn.state.get_state_dict(model).items():
+          if 'scale' in k: v.shard_(device, axis=None)  # from quantized
+          elif '.attention.' in k:
+            if getenv("SHARD_KVCACHE") and ('.wq.' in k or '.wk.' in k or '.wv.' in k): v.shard_(device, axis=0)
+            else: v.shard_(device, axis=-1)
+          elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
+          elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
+          elif '.feed_forward.' in k: v.shard_(device, axis=-1)
+          elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
+          elif 'output.weight' in k: v.shard_(device, axis=-1)
+          #elif k.endswith('.weight'): v.shard_(device, axis=-1)
+          #elif 'norm.' in k: v.shard_(device, axis=-1)
+          else: v.shard_(device, axis=None)
+          #print(k, v.shape, v.lazydata.axis)

-        # replace weights in model
-        load_state_dict(model, weights, strict=False, consume=True)
+      # replace weights in model
+      load_state_dict(model, weights, strict=False, consume=True)

    return LLaMa(model, tokenizer)

@@ -330,6 +327,7 @@ int main()
 \end{code}
 """
 if __name__ == "__main__":
+  Tensor.no_grad = True
  print(f"using {Device.DEFAULT} backend")

  parser = argparse.ArgumentParser(description="Run LLaMA in tinygrad", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
@@ -445,7 +443,7 @@ After you are done speaking, output [EOS]. You are not Chad.
  print(f"using LLaMA{LLAMA_SUFFIX}-{args.size} model")
  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(args.shard)) if args.shard > 1 else Device.DEFAULT
  llama = LLaMa.build(MODEL_PATH, TOKENIZER_PATH, model_gen=args.gen, model_size=args.size, quantize=args.quantize, device=device)
-  param_bytes = sum(x.uop.size * x.dtype.itemsize for x in get_parameters(llama.model))
+  param_bytes = sum(x.lazydata.size * x.dtype.itemsize for x in get_parameters(llama.model))

  outputted = pre_prompt if chatbot else args.prompt
  start_pos, toks = 0, [llama.tokenizer.bos_id()] + llama.tokenizer.encode(outputted)
@@ -477,12 +475,11 @@ After you are done speaking, output [EOS]. You are not Chad.
      next_tok = Tensor([toks[start_pos:]], device=device) if tok_tensor is None or (len(toks)-start_pos) > 1 else tok_tensor.reshape(1, 1)
      with Profiling(enabled=args.profile):
        with Timing("total ", enabled=args.timing, on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
-          with WallTimeEvent(BenchEvent.STEP):
-            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
-                        f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
-                        (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=args.timing):
-              tok_tensor = llama.model(next_tok, start_pos, args.temperature)
-            tok = tok_tensor.item()
+          with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
+                      f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
+                      (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=args.timing):
+            tok_tensor = llama.model(next_tok, start_pos, args.temperature)
+          tok = tok_tensor.item()

      # use the kv cache
      start_pos = len(toks)
@@ -1,17 +1,16 @@
 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
 from tinygrad.helpers import Profiling, Timing, DEBUG, colored, fetch, tqdm
-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 = [
@@ -48,7 +47,7 @@ def concat_weights(models, device=None):
    disk_tensors: List[Tensor] = [model[name] for model in models]
    if len(disk_tensors) == 1 or len(disk_tensors[0].shape) == 1:
      return disk_tensors[0].to(device=device)
-    axis = 1 if name.endswith((".attention.wo.weight", ".feed_forward.w2.weight")) else 0
+    axis = 1 if name.endswith(".attention.wo.weight") or name.endswith(".feed_forward.w2.weight") else 0
    lazy_tensors = [data.to(device=device) for data in disk_tensors]
    return lazy_tensors[0].cat(*lazy_tensors[1:], dim=axis)
  return {name: convert(name) for name in {name: None for model in models for name in model}}
@@ -74,17 +73,16 @@ class Int8Linear:
    self.scale = Tensor.ones(out_features, dtype=dtypes.half)

  def __call__(self, x):
-    return x.dot(self.weight.cast(self.scale.dtype).T*self.scale)
+    return x.dot(self.weight.cast(dtype=dtypes.half).T*self.scale)

  @staticmethod
-  def quantize(tensors, device, scale_dtype=dtypes.float16, quantize_embeds=False):
+  def quantize(tensors, device):
    new_tensors = {}
    for name,v in tensors.items():
-      if "feed_forward" in name or "attention.w" in name or (quantize_embeds and "tok_embeddings.weight" in name):
+      if "feed_forward" in name or "attention.w" in name:
        assert "weight" in name, name
-        v = v.cast(scale_dtype)
        scale = v.abs().max(axis=1) / 127.0
-        int8_weight = (v.T/scale).T.round().cast(dtype=dtypes.int8) # without round(), cast truncates -34.9 to -34
+        int8_weight = (v.T/scale).T.cast(dtype=dtypes.int8)
        new_tensors[name] = int8_weight
        new_tensors[name.replace('weight', 'scale')] = scale
        if isinstance(device, tuple):
@@ -92,20 +90,8 @@ class Int8Linear:
          new_tensors[name.replace('weight', 'scale')].shard_(device, axis=None)
      else:
        new_tensors[name] = v
-    if quantize_embeds: new_tensors.update({"output.weight": new_tensors["tok_embeddings.weight"], "output.scale": new_tensors["tok_embeddings.scale"]})
    return new_tensors

-class Int8Embedding:
-  def __init__(self, vocab_size:int, embed_size:int):
-    self.vocab_sz, self.embed_sz = vocab_size, embed_size
-    self.weight, self.scale = Tensor.ones(vocab_size, embed_size, dtype=dtypes.int8), Tensor.ones(vocab_size, dtype=dtypes.half)
-
-  def __call__(self, idx:Tensor) -> Tensor:
-    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz, requires_grad=False, device=self.weight.device).unsqueeze(-1)
-    big_shp = idx.shape+(self.vocab_sz, self.embed_sz)
-    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1)).expand(big_shp), (self.weight.cast(self.scale.dtype).T*self.scale).T
-    return (arange == idx).mul(vals).sum(-2, dtype=vals.dtype)
-
 def NF4Linear(block_size):
  _CODE = [
    -1.0, -0.6961928009986877, -0.5250730514526367, -0.39491748809814453, -0.28444138169288635, -0.18477343022823334, -0.09105003625154495, 0.0,
@@ -127,8 +113,7 @@ def NF4Linear(block_size):
      return x.linear(unscaled.reshape(self.out_features, self.in_features).T)

    @staticmethod
-    def quantize(state_dict: dict[str, Tensor], device, scale_dtype=dtypes.float16, quantize_embeds=False) -> dict[str, Tensor]:
-      assert not quantize_embeds  # TODO: support this?
+    def quantize(state_dict: dict[str, Tensor], device) -> dict[str, Tensor]:
      new_state_dict = {}
      for k, v in state_dict.items():
        if "feed_forward" in k or "attention.w" in k:
@@ -136,7 +121,7 @@ def NF4Linear(block_size):
          scale = (grouped.abs().max(axis=1, keepdim=True))
          coded = ((grouped / scale).unsqueeze(-1) - CODE.to(v.device)).abs().argmin(axis=-1).cast(dtypes.uint8).flatten()
          new_state_dict[k] = coded[::2] * 2 ** 4 + coded[1::2]
-          new_state_dict[k.replace(".weight", ".scale")] = scale.cast(scale_dtype)
+          new_state_dict[k.replace(".weight", ".scale")] = scale.cast(dtypes.float16)
          if isinstance(device, tuple):
            new_state_dict[k].shard_(device, axis=-1)
            new_state_dict[k.replace('weight', 'scale')].shard_(device, axis=None)
@@ -145,41 +130,6 @@ 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},
@@ -192,57 +142,49 @@ MODEL_PARAMS = {
  "70B": {
    "args": {"dim": 8192, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-5, "rope_theta": 500000, "vocab_size": 128256,  "hidden_dim": 28672},
    "files": 8
-  },
-  "405B": {
-    "args": {"dim": 16384, "n_heads": 128, "n_kv_heads": 8, "n_layers": 126, "norm_eps": 1e-5, "rope_theta": 500000, "vocab_size": 128256,  "hidden_dim": 53248},
-    "files": 191
-  },
+  }
 }
-def build_transformer(model_path: Path, model_size="8B", quantize=None, scale_dtype=dtypes.float16, device=None, max_context=8192, load_weights=True):
+def build_transformer(model_path: Path, model_size="8B", quantize=None, device=None):
  # 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)
-
-  if not load_weights: return model
+  if quantize == "int8": linear = Int8Linear
+  elif quantize == "nf4": linear = NF4Linear(64)
+  else: linear = nn.Linear
+  model = Transformer(**MODEL_PARAMS[model_size]["args"], linear=linear, max_context=8192, jit=True)

  # load weights
-  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    if model_path.is_dir():
-      if (model_path / "model.safetensors.index.json").exists(): weights = load(str(model_path / "model.safetensors.index.json"))
-      elif (model_path / "model.safetensors").exists(): weights = load(str(model_path / "model.safetensors"))
-      else: weights = concat_weights([load(str(model_path / f"consolidated.{i:02d}.pth")) for i in range(MODEL_PARAMS[model_size]["files"])], device[0] if isinstance(device, tuple) else device)
-    else:
-      weights = load(str(model_path))
-    if "model.embed_tokens.weight" in weights:
-      weights = convert_from_huggingface(weights, MODEL_PARAMS[model_size]["args"]["n_layers"], MODEL_PARAMS[model_size]["args"]["n_heads"], MODEL_PARAMS[model_size]["args"]["n_kv_heads"])
-    elif "token_embd.weight" in weights:
-      weights = convert_from_gguf(weights, MODEL_PARAMS[model_size]["args"]["n_layers"])
-    weights = fix_bf16(weights)
+  if model_path.is_dir():
+    if (model_path / "model.safetensors.index.json").exists(): weights = load(str(model_path / "model.safetensors.index.json"))
+    elif (model_path / "model.safetensors").exists(): weights = load(str(model_path / "model.safetensors"))
+    else: weights = concat_weights([load(str(model_path / f"consolidated.{i:02d}.pth")) for i in range(MODEL_PARAMS[model_size]["files"])], device[0] if isinstance(device, tuple) else device)
+  else:
+    weights = load(str(model_path))
+  if "model.embed_tokens.weight" in weights:
+    weights = convert_from_huggingface(weights, model, MODEL_PARAMS[model_size]["args"]["n_heads"], MODEL_PARAMS[model_size]["args"]["n_kv_heads"])
+  elif "token_embd.weight" in weights:
+    weights = convert_from_gguf(weights, model)
+  weights = fix_bf16(weights)

-    with Context(BEAM=0):
-      # quantize
-      if quantize == "float16": weights = {k:v.cast(quantize).contiguous() for k,v in weights.items()}
-      elif quantize is not None:
-        weights = linear.quantize(weights, device, scale_dtype, quantize_embeds)
-        for _,v in weights.items(): v.realize()
+  with Context(BEAM=0):
+    # quantize
+    if quantize == "float16": weights = {k:v.cast(quantize).contiguous() for k,v in weights.items()}
+    elif quantize is not None:
+      weights = linear.quantize(weights, device)
+      for _,v in weights.items(): v.realize()

-      # shard
-      if isinstance(device, tuple):
-        for k,v in nn.state.get_state_dict(model).items():
-          if 'scale' in k: v.shard_(device, axis=None)  # from quantized
-          elif '.attention.' in k: v.shard_(device, axis=-1)
-          elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
-          elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
-          elif '.feed_forward.' in k: v.shard_(device, axis=-1)
-          elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
-          elif 'output.weight' in k: v.shard_(device, axis=0)
-          else: v.shard_(device, axis=None)
+    # shard
+    if isinstance(device, tuple):
+      for k,v in nn.state.get_state_dict(model).items():
+        if 'scale' in k: v.shard_(device, axis=None)  # from quantized
+        elif '.attention.' in k: v.shard_(device, axis=-1)
+        elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
+        elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
+        elif '.feed_forward.' in k: v.shard_(device, axis=-1)
+        elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
+        elif 'output.weight' in k: v.shard_(device, axis=0)
+        else: v.shard_(device, axis=None)

-      # replace weights in model
-      load_state_dict(model, weights, strict=False, consume=True)
+    # replace weights in model
+    load_state_dict(model, weights, strict=False, consume=True)
  return model

 # default settings
@@ -273,18 +215,20 @@ def prefill(model, toks, start_pos=0):
  return start_pos

 if __name__ == "__main__":
+  Tensor.no_grad = True
+
  parser = argparse.ArgumentParser()
  parser.add_argument("--download_model", action="store_true", help="Download a model")
  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("--size", choices=["1B", "8B", "70B"], 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")
  parser.add_argument("--debug", action="store_true", help="Enable debug mode")
  parser.add_argument("--seed", type=int, help="Random seed")
-  parser.add_argument("--temperature", type=float, default=0.85, help="Temperature")
+  parser.add_argument("--temperature", type=int, default=0.85, help="Temperature")
  parser.add_argument("--benchmark", action="store_true", help="Run a benchmark")
  parser.add_argument("--timing", action="store_true", help="Print timing per token")
  parser.add_argument("--profile", action="store_true", help="Output profile data")
@@ -303,11 +247,11 @@ if __name__ == "__main__":
      fetch("https://huggingface.co/TriAiExperiments/SFR-Iterative-DPO-LLaMA-3-8B-R/resolve/main/model-00004-of-00004.safetensors", "model-00004-of-00004.safetensors", subdir="llama3-8b-sfr")
      args.model = fetch("https://huggingface.co/TriAiExperiments/SFR-Iterative-DPO-LLaMA-3-8B-R/raw/main/model.safetensors.index.json", "model.safetensors.index.json", subdir="llama3-8b-sfr")
    elif args.size == "70B":
-      subdir = "DeepSeek-R1-Distill-Llama-70B"
-      args.model = fetch("https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/resolve/main/model.safetensors.index.json?download=true", "model.safetensors.index.json", subdir=subdir)
+      subdir = "Llama-3.1-Nemotron-70B-Instruct-HF"
+      args.model = fetch("https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF/resolve/main/model.safetensors.index.json?download=true", "model.safetensors.index.json", subdir=subdir)
      fetch("https://huggingface.co/bofenghuang/Meta-Llama-3-8B/resolve/main/original/tokenizer.model", "tokenizer.model", subdir=subdir)
-      for i in range(17):
-        fetch(f"https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/resolve/main/model-{i+1:05d}-of-000017.safetensors?download=true", f"model-{i+1:05d}-of-000017.safetensors", subdir=subdir)
+      for i in range(30):
+        fetch(f"https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF/resolve/main/model-{i+1:05d}-of-00030.safetensors?download=true", f"model-{i+1:05d}-of-00030.safetensors", subdir=subdir)

  assert args.model is not None, "please provide --model option"

@@ -324,7 +268,7 @@ if __name__ == "__main__":

  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(args.shard)) if args.shard > 1 else Device.DEFAULT
  model = build_transformer(args.model, model_size=args.size, quantize=args.quantize, device=device)
-  param_bytes = sum(x.uop.size * x.dtype.itemsize for x in get_parameters(model))
+  param_bytes = sum(x.lazydata.size * x.dtype.itemsize for x in get_parameters(model))

  if not args.no_api and not args.benchmark:
    from bottle import Bottle, request, response, HTTPResponse, abort, static_file
@@ -476,12 +420,11 @@ if __name__ == "__main__":
      st = GlobalCounters.time_sum_s
      with Profiling(enabled=args.profile):
        with Timing("total ", on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
-          with WallTimeEvent(BenchEvent.STEP):
-            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
-                        f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
-                        (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None):
-              tok = model(Tensor([[last_tok]], device=device), start_pos, TEMPERATURE, TOP_K, TOP_P, ALPHA_F, ALPHA_P)
-            tok = tok.item()
+          with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
+                      f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
+                      (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None):
+            tok = model(Tensor([[last_tok]], device=device), start_pos, TEMPERATURE, TOP_K, TOP_P, ALPHA_F, ALPHA_P)
+          tok = tok.item()
      start_pos += 1
      last_tok = tok
      generated += tokenizer.decode([tok])
@@ -515,7 +458,7 @@ if __name__ == "__main__":
        st = GlobalCounters.time_sum_s
        with Profiling(enabled=args.profile):
          with Timing("total ", enabled=args.timing, on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
-            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
                        f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
                        (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=args.timing):

@@ -1,13 +1,14 @@
 #!/usr/bin/env python3
 import os
 if "NOOPT" not in os.environ: os.environ["NOOPT"] = "1"
-from tinygrad import Device, nn, Tensor, dtypes
-Device.DEFAULT = "CPU"
+from tinygrad import Device, nn, Tensor, dtypes, Variable
+Device.DEFAULT = "CLANG"
 from train_gpt2 import GPT, GPTConfig
-from tinygrad.helpers import dedup, flatten, getenv, GlobalCounters, to_function_name
-from tinygrad.engine.realize import get_kernel
+from tinygrad.helpers import dedup, to_function_name, flatten, getenv, GlobalCounters, ansilen, to_function_name
+from tinygrad.engine.schedule import create_schedule
+from tinygrad.engine.realize import get_kernel, run_schedule
 from tinygrad.engine.memory import memory_planner
-from tinygrad.uop.ops import Ops
+from tinygrad.ops import Ops

 TIMING = getenv("TIMING")

@@ -16,7 +17,7 @@ if __name__ == "__main__":
  #model.load_pretrained()
  for p in nn.state.get_parameters(model): p.replace(Tensor.empty(p.shape, dtype=p.dtype)) # fake load pretrained

-  #early_sched = create_schedule([x.uop for x in nn.state.get_parameters(model)])
+  #early_sched = create_schedule([x.lazydata for x in nn.state.get_parameters(model)])
  #print(f"built model {len(early_sched)}")

  #B, T = Variable("B", 1, 128).bind(4), 64 #Variable("T", 1, 1024).bind(64)
@@ -25,7 +26,7 @@ if __name__ == "__main__":
  Tensor.training = True
  optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=1e-4)
  warmup_count = getenv("WARMUP", 3)
-  for i in range(warmup_count):  # TODO: why does it take three and not two to stabilize
+  for i in range(warmup_count):  # TODO: why does it take three and not two to stablize
    GlobalCounters.reset()
    X = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
    Y = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
@@ -36,16 +37,16 @@ if __name__ == "__main__":
      tensors = optimizer.schedule_step()
    else:
      tensors = []
-    sched = loss.schedule(*tensors)
+    sched = create_schedule([loss.lazydata] + [x.lazydata for x in tensors])
    print(f"calls {i}:", len(sched))
    #run_schedule(sched[:])
  sched = memory_planner(sched)
  ast_dedup = dedup([si.ast for si in sched if si.ast.op is Ops.SINK])
  srcs = {}
  for ast in ast_dedup:
-    k = get_kernel(Device["CPU"].renderer, ast)
+    k = get_kernel(Device["CLANG"].renderer, ast)
    k.linearize()
-    src = Device["CPU"].renderer.render(to_function_name(k.name), k.uops)
+    src = Device["CLANG"].renderer.render(to_function_name(k.name), k.uops)
    srcs[ast] = (k.name, src)
  print("functions:", len(srcs))
  used_buffers = dedup(flatten([si.bufs for si in sched]))
@@ -56,7 +57,7 @@ if __name__ == "__main__":
  state_dict.update({'X': X, 'Y': Y, 'loss': loss})
  grad_state_dict = {}
  for k,v in state_dict.items():
-    if v.uop.base.buffer not in used_buffers: print(f"UNUSED: {k}")
+    if v.lazydata.base.buffer not in used_buffers: print(f"UNUSED: {k}")
    if v.grad is not None: grad_state_dict['grad_'+k] = v.grad
  state_dict.update(grad_state_dict)
  state_dict.update({'adam_b1_t': optimizer.b1_t, 'adam_b2_t': optimizer.b2_t, 'adam_lr': optimizer.lr})
@@ -65,7 +66,7 @@ if __name__ == "__main__":
    nm = inverse_state_dict[p]
    state_dict["adam_m_"+nm] = m
    state_dict["adam_v_"+nm] = v
-  named_buffers = {v.uop.base.buffer:k.replace(".", "_") for k,v in state_dict.items()}
+  named_buffers = {v.lazydata.base.buffer:k.replace(".", "_") for k,v in state_dict.items()}

  c_code = ["#include <stdlib.h>", "#include <tgmath.h>", "#include <stdbool.h>"]
  if TIMING: c_code += ["#include <stdio.h>", "#include <time.h>"]
@@ -99,7 +99,7 @@ class GPT:

  def __call__(self, idx:Tensor, targets=None):
    b, t = idx.shape
-    pos = Tensor.arange(0, t, device=idx.device)
+    pos = Tensor.arange(0, t)

    tok_emb = self.wte(idx) # token embeddings of shape (b, t, n_embd)
    pos_emb = self.wpe(pos) # position embeddings of shape (t, n_embd)
@@ -124,7 +124,6 @@ if __name__ == "__main__":
  parser.add_argument("--batch_size", type=int, default=4, help="batch size")
  parser.add_argument("--sequence_length", type=int, default=64, help="sequence length")
  parser.add_argument("--skip_test", action="store_true", help="skip test")
-  parser.add_argument("--gpus", type=int, default=1, help="sequence length")
  args = parser.parse_args()
  B, T = args.batch_size, args.sequence_length
  assert 1 <= T <= 1024
@@ -132,10 +131,6 @@ if __name__ == "__main__":
  model = GPT(GPTConfig(n_layer=12, n_head=12, n_embd=768))
  model.load_pretrained()

-  if args.gpus > 1:
-    GPUS = tuple(f'{Device.DEFAULT}:{i}' for i in range(args.gpus))
-    for x in nn.state.get_parameters(model): x.to_(GPUS)  # we put a copy of the model on every GPU
-
  # init the tokenizer
  enc = tiktoken.get_encoding("gpt2")
  encode = lambda s: enc.encode(s, allowed_special={"<|endoftext|>"})
@@ -170,32 +165,23 @@ if __name__ == "__main__":
  x, y = next(data_iter) # we'll overfit this batch below
  optimizer = nn.optim.AdamW(nn.state.get_parameters(model), lr=1e-4, weight_decay=0)

-  print(f"model state:     {sum(x.nbytes() for x in nn.state.get_parameters(model))/1e9:.2f} GB")
-  print(f"optimizer state: {sum(x.nbytes() for x in nn.state.get_parameters(optimizer))/1e9:.2f} GB")
-
-  # shard the data on axis 0
-  if args.gpus > 1: x, y = x.shard(GPUS, axis=0), y.shard(GPUS, axis=0)
-
  @TinyJit
-  @Tensor.train()
-  def step(x:Tensor, y:Tensor) -> Tensor:
+  def step(x, y):
    _, loss = model(x, y)
    optimizer.zero_grad()
    loss.backward()
    return loss.realize(*optimizer.schedule_step())

-  for i in range(args.num_iterations):
-    GlobalCounters.reset()
-    t0 = time.perf_counter()
-    loss = step(x.contiguous(), y.contiguous())
-    Device[Device.DEFAULT].synchronize()
-    t1 = time.perf_counter()
-    print(f"iteration {i}, loss: {loss.item():.6f}, time: {(t1-t0)*1000:.3f}ms, {int(B*T/(t1-t0))} tok/s, {GlobalCounters.global_mem/1e9:.2f} GB")
+  with Tensor.train():
+    for i in range(args.num_iterations):
+      GlobalCounters.reset()
+      t0 = time.time()
+      loss = step(x.contiguous(), y.contiguous())
+      Device[Device.DEFAULT].synchronize()
+      t1 = time.time()
+      print(f"iteration {i}, loss: {loss.item():.6f}, time: {(t1-t0)*1000:.3f}ms, {int(B*T/(t1-t0))} tok/s")

  if not args.skip_test:
-    # copy back to single gpu for test
-    if args.gpus > 1:
-      for x in nn.state.get_parameters(model): x.to_(Device.DEFAULT)
    start = "<|endoftext|>"
    start_ids = encode(start)
    x = (Tensor(start_ids)[None, ...])
@@ -279,15 +279,9 @@ def generate(model, tokenizer, prompt: str, n_tokens_to_gen: int = 10, temp: boo
  # Loading in the prompt tokens
  logits = model.forward(Tensor([tks]))[:, -1, :]
  for _ in tqdm(range(n_tokens_to_gen), desc="Speed Gen"):
+    # TODO: topk
    if sample:
-      scaled_logits = logits / temp
-      if top_k is not None:
-        topk_values, topk_indices = scaled_logits.topk(top_k)
-        filtered_logits = Tensor.full_like(scaled_logits, -float("inf"))
-        filtered_logits = filtered_logits.scatter(dim=-1, index=topk_indices, src=topk_values)
-        tok_Tens = filtered_logits.softmax().multinomial()
-      else:
-        tok_Tens = scaled_logits.softmax().multinomial()
+      tok_Tens = (logits/temp).softmax().multinomial()
    else:
      tok_Tens = logits.argmax(axis=-1).unsqueeze(0)
    tok = tok_Tens.item()
@@ -304,7 +298,6 @@ if __name__ == "__main__":
  parser.add_argument("--size", type=str, default="370m",
                      help=f"Size of model to use [{', '.join([k for k in MODELS.keys()])}]")
  parser.add_argument("--n_tokens", type=int, default=10, help="Number of tokens to generate")
-  parser.add_argument("--top_k", type=int, help="Limit sampling to the top k most likely tokens")
  parser.add_argument("--sample", dest="sample", action="store_true", help="Sample flag")
  parser.add_argument("--temp", type=float, default=1.0, help="Sampling temp has to be <=1.0")
  args = parser.parse_args()
@@ -315,9 +308,8 @@ if __name__ == "__main__":
  num_toks = args.n_tokens
  sample = args.sample
  temp = args.temp
-  top_k = args.top_k
  s = time.time()
-  tinyoutput = generate(model, tokenizer, prompt, n_tokens_to_gen=num_toks, sample=sample, temp=temp, top_k=top_k)
+  tinyoutput = generate(model, tokenizer, prompt, n_tokens_to_gen=num_toks, sample=sample, temp=temp)
  print(tinyoutput)
  print('TIME: ', time.time() - s)
  TORCHOUTPUT = "Why is gravity \nso important?\nBecause it's the only"
@@ -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()
@@ -1,156 +0,0 @@
-# much taken from https://github.com/cloneofsimo/minRF
-from tinygrad import Tensor, nn, GlobalCounters, TinyJit
-from tinygrad.helpers import getenv, trange
-from extra.models.llama import Attention, FeedForward, precompute_freqs_cis
-
-def modulate(x:Tensor, shift:Tensor, scale:Tensor) -> Tensor: return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
-
-# TODO: why doesn't the TimestepEmbedder from minRF work?
-class TimestepEmbedder:
-  def __init__(self, hidden_size): self.mlp = [nn.Linear(1, hidden_size), Tensor.silu, nn.Linear(hidden_size, hidden_size)]
-  def __call__(self, t:Tensor): return t.reshape(-1, 1).sequential(self.mlp)
-
-class TransformerBlock:
-  def __init__(self, dim, n_heads, norm_eps=1e-5):
-    self.attention = Attention(dim, n_heads)
-    self.feed_forward = FeedForward(dim, 4*dim)
-    self.attention_norm = nn.LayerNorm(dim, eps=norm_eps)
-    self.ffn_norm = nn.LayerNorm(dim, eps=norm_eps)
-    self.adaLN_modulation = nn.Linear(dim, 6 * dim, bias=True)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor, adaln_input:Tensor):
-    shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input.silu()).chunk(6, dim=1)
-    x = x + gate_msa.unsqueeze(1) * self.attention(modulate(self.attention_norm(x), shift_msa, scale_msa), 0, freqs_cis)
-    x = x + gate_mlp.unsqueeze(1) * self.feed_forward(modulate(self.ffn_norm(x), shift_mlp, scale_mlp))
-    return x.contiguous().contiguous_backward()
-
-class FinalLayer:
-  def __init__(self, dim, patch_size, out_channels):
-    self.norm_final = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
-    self.linear = nn.Linear(dim, patch_size*patch_size*out_channels, bias=True)
-    self.adaLN_modulation = nn.Linear(dim, 2 * dim, bias=True)
-
-    # init weights/bias to 0
-    self.linear.weight.replace(self.linear.weight.zeros_like().contiguous())
-    self.linear.bias.replace(self.linear.bias.zeros_like().contiguous())
-
-  def __call__(self, x:Tensor, c:Tensor):
-    shift, scale = self.adaLN_modulation(c.silu()).chunk(2, dim=1)
-    x = modulate(self.norm_final(x), shift, scale)
-    return self.linear(x)
-
-# channels=1, input_size=32, dim=64, n_layers=6, n_heads=4, num_classes=10
-class DiT_Llama:
-  def __init__(self, in_channels=1, dim=64, n_layers=6, n_heads=4, num_classes=10, patch_size=2):
-    self.patch_size = patch_size
-    self.out_channels = in_channels
-    self.num_classes = num_classes
-
-    self.init_conv_seq = [
-      nn.Conv2d(in_channels, dim // 2, kernel_size=5, padding=2, stride=1), Tensor.silu, nn.GroupNorm(32, dim//2),
-      nn.Conv2d(dim //2, dim // 2, kernel_size=5, padding=2, stride=1), Tensor.silu, nn.GroupNorm(32, dim//2),
-    ]
-
-    self.x_embedder = nn.Linear(self.patch_size * self.patch_size * dim // 2, dim, bias=True)
-    self.t_embedder = TimestepEmbedder(dim)
-    self.y_embedder = nn.Embedding(num_classes+1, dim)
-    self.final_layer = FinalLayer(dim, self.patch_size, self.out_channels)
-
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, 4096)
-    self.layers = [TransformerBlock(dim, n_heads) for _ in range(n_layers)]
-
-  def unpatchify(self, x:Tensor):
-    c, p = self.out_channels, self.patch_size
-    h = w = int(x.shape[1] ** 0.5)
-    x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
-    x = x.rearrange("n h w p q c -> n c h p w q")
-    return x.reshape(shape=(x.shape[0], c, h * p, h * p))
-
-  def patchify(self, x:Tensor):
-    B, C, H, W = x.shape
-    x = x.reshape(B, C, H // self.patch_size, self.patch_size, W // self.patch_size, self.patch_size)
-    x = x.permute(0, 2, 4, 1, 3, 5).flatten(-3).flatten(1, 2)
-    return x  # B <H*W ish> <C*patch_size*patch_size>
-
-  def __call__(self, x:Tensor, t:Tensor, y:Tensor) -> Tensor:
-    x = x.sequential(self.init_conv_seq)
-    x = self.patchify(x)
-    x = self.x_embedder(x)
-    adaln_input = self.t_embedder(t) + self.y_embedder(y)
-    adaln_input = adaln_input.contiguous()
-    for layer in self.layers:
-      x = layer(x, self.freqs_cis[:, :x.size(1)], adaln_input=adaln_input)
-    x = self.final_layer(x, adaln_input)
-    return self.unpatchify(x)
-
-  def rf(self, x:Tensor, cond:Tensor):
-    b = x.shape[0]
-    # self.ln is True
-    t = Tensor.randn((b,)).sigmoid()
-    texp = t.view([b, *([1] * len(x.shape[1:]))])
-
-    # conditional dropout
-    dropout_prob = 0.1
-    cond = (Tensor.rand(cond.shape[0]) < dropout_prob).where(cond.full_like(self.num_classes), cond)
-
-    # this is rectified flow
-    z1 = x.randn_like()
-    zt = (1 - texp) * x + texp * z1
-    vtheta = self(zt, t, cond)
-
-    # MSE loss
-    return ((z1 - x) - vtheta).square().mean()
-
-  def sample(self, z, cond, null_cond, sample_steps=50, cfg=2.0):
-    b = z.size(0)
-    dt = Tensor.full((b,)+(1,)*len(z.shape[1:]), fill_value=1.0/sample_steps).contiguous()
-    images = [z]
-    for i in range(sample_steps, 0, -1):
-      t = Tensor.full((b,), fill_value=i/sample_steps).contiguous()
-      vc = self(z, t, cond)
-      vu = self(z, t, null_cond)
-      vc = vu + cfg * (vc - vu)
-      z = z - dt * vc
-      z = z.contiguous()
-      images.append(z)
-    return images
-
-def mviz(t:Tensor):
-  assert len(t.shape) == 4 and t.shape[1] == 1
-  ft = t.permute(1,2,0,3).reshape(32, -1)
-  assert ft.shape[-1]%32 == 0
-  print("")
-  for y in ((ft+1)/2).clamp(0,1).tolist():
-    ln = [f"\033[38;5;{232+int(x*23)}m██" for x in y]
-    print(''.join(ln) + "\033[0m")
-
-if __name__ == "__main__":
-  X_train, Y_train, X_test, Y_test = nn.datasets.mnist()
-  X_train = X_train.pad((2,2,2,2))
-  X_train = ((X_train.float()/255)-0.5)/0.5
-  Y_train = Y_train.int()
-
-  model = DiT_Llama(patch_size=getenv("PATCH_SIZE", 2))
-  for r in nn.state.get_parameters(model): r.realize()
-  optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=5e-4)
-
-  @TinyJit
-  @Tensor.train()
-  def train_step():
-    if getenv("OVERFIT"): samples = Tensor.zeros(getenv("BS", 256), dtype='int')
-    else: samples = Tensor.randint(getenv("BS", 256), high=X_train.shape[0])
-    optimizer.zero_grad()
-    loss = model.rf(X_train[samples], Y_train[samples])
-    loss.backward()
-    optimizer.step()
-    return loss
-
-  @TinyJit
-  def sample(z:Tensor, cond:Tensor) -> Tensor:
-    return model.sample(z, cond, Tensor.full_like(cond, 10), sample_steps=getenv("SAMPLE_STEPS", 20))[-1]
-
-  for steps in (t:=trange(getenv("STEPS", 5000))):
-    if steps%10 == 0: mviz(sample(Tensor.randn(3, 1, 32, 32), Tensor([5,0,4], dtype='int')))
-    GlobalCounters.reset()
-    loss = train_step()
-    t.set_description(f"loss: {loss.item():9.2f}")
@@ -3,7 +3,6 @@ from tinygrad import Tensor, nn, Device, GlobalCounters, Variable
 from tinygrad.helpers import Timing, Profiling, CI, tqdm
 from tinygrad.nn.state import torch_load, get_state_dict
 from extra.models.llama import FeedForward, Transformer
-from extra.bench_log import BenchEvent, WallTimeEvent

 class MixtureFeedForward:
  def __init__(self, num_experts:int, dim:int, hidden_dim:int, linear=nn.Linear):
@@ -31,19 +30,18 @@ if __name__ == "__main__":
                      help="Path to the downloaded weights")
  args = parser.parse_args()

-  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    state = torch_load(args.weights + "/consolidated.00.pth.b")
-    model = Transformer(n_layers=32, dim=4096, hidden_dim=14336, n_heads=32, n_kv_heads=8, norm_eps=1e-5, vocab_size=32000, feed_forward=functools.partial(MixtureFeedForward, 8), jit=False)
-    model_state_dict = get_state_dict(model)
+  state = torch_load(args.weights + "/consolidated.00.pth.b")
+  model = Transformer(n_layers=32, dim=4096, hidden_dim=14336, n_heads=32, n_kv_heads=8, norm_eps=1e-5, vocab_size=32000, feed_forward=functools.partial(MixtureFeedForward, 8), jit=False)
+  model_state_dict = get_state_dict(model)

-    for k in (t := tqdm(state, disable=CI)):
-      if 'feed_forward.experts.' in k:
-        expert_no = int(k.split('feed_forward.experts.')[1].split('.')[0])
-        device = Device.DEFAULT + ":" + str((expert_no//2)+1)
-      else:
-        device = Device.DEFAULT
-      t.set_description(f"ram used: {GlobalCounters.mem_used/1e9:5.2f} GB, loading {k} to {device}")
-      model_state_dict[k].replace(state[k].to(device).half()).realize()
+  for k in (t := tqdm(state, disable=CI)):
+    if 'feed_forward.experts.' in k:
+      expert_no = int(k.split('feed_forward.experts.')[1].split('.')[0])
+      device = Device.DEFAULT + ":" + str((expert_no//2)+1)
+    else:
+      device = Device.DEFAULT
+    t.set_description(f"ram used: {GlobalCounters.mem_used/1e9:5.2f} GB, loading {k} to {device}")
+    model_state_dict[k].replace(state[k].to(device).half()).realize()
  if CI: print(f"ram used: {GlobalCounters.mem_used/1e9:5.2f} GB")

  from sentencepiece import SentencePieceProcessor
@@ -55,8 +53,7 @@ if __name__ == "__main__":
    GlobalCounters.reset()
    with Profiling(sort="time", frac=0.1, enabled=args.profile):
      with Timing("total ", enabled=args.timing, on_exit=lambda x: f", {1e9/x:.2f} tok/sec"):
-        with WallTimeEvent(BenchEvent.STEP):
-          tok = model(Tensor([toks[start_pos:]]), 0 if start_pos == 0 else Variable("start_pos", 1, 1024-1).bind(start_pos), args.temperature).item()
+        tok = model(Tensor([toks[start_pos:]]), 0 if start_pos == 0 else Variable("start_pos", 1, 1024).bind(start_pos), args.temperature).item()
    toks.append(tok)
    start_pos += 1
    print(spp.decode(toks))
@@ -1,12 +1,11 @@
-import os, random, pickle, queue, struct, math, functools, hashlib, time
+import os, random, pickle, queue
 from typing import List
 from pathlib import Path
 from multiprocessing import Queue, Process, shared_memory, connection, Lock, cpu_count

 import numpy as np
 from tinygrad import dtypes, Tensor
-from tinygrad.helpers import getenv, prod, Context, round_up, tqdm, OSX
-from tinygrad.nn.state import TensorIO
+from tinygrad.helpers import getenv, prod, Context, round_up, tqdm

 ### ResNet

@@ -72,7 +71,7 @@ def loader_process(q_in, q_out, X:Tensor, seed):
      #storage_tensor._copyin(img_tensor.numpy())

      # faster
-      X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+      X[idx].contiguous().realize().lazydata.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()

      # ideal
      #X[idx].assign(img.tobytes())   # NOTE: this is slow!
@@ -130,15 +129,14 @@ def batch_load_resnet(batch_size=64, val=False, shuffle=True, seed=None, pad_fir
  q_in, q_out = Queue(), Queue()

  sz = (batch_size*BATCH_COUNT, 224, 224, 3)
-  shm_name = "resnet_X_val" if val else "resnet_X_train"
-  if not OSX and 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(sz))
+  if os.path.exists("/dev/shm/resnet_X"): os.unlink("/dev/shm/resnet_X")
+  shm = shared_memory.SharedMemory(name="resnet_X", create=True, size=prod(sz))
  procs = []

  try:
    # disk:shm is slower
-    if OSX: X = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:shm:{shm.name}")
-    else: X = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/{shm_name}")
+    #X = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:shm:{shm.name}")
+    X = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/resnet_X")
    Y = [None] * (batch_size*BATCH_COUNT)

    for _ in range(cpu_count()):
@@ -172,13 +170,13 @@ def batch_load_resnet(batch_size=64, val=False, shuffle=True, seed=None, pad_fir

 def process_batch_bert(data: List[dict]) -> dict[str, Tensor]:
  return {
-    "input_ids": Tensor(np.concatenate([s["input_ids"] for s in data], axis=0), dtype=dtypes.int32, device="CPU"),
-    "input_mask": Tensor(np.concatenate([s["input_mask"] for s in data], axis=0), dtype=dtypes.int32, device="CPU"),
-    "segment_ids": Tensor(np.concatenate([s["segment_ids"] for s in data], axis=0), dtype=dtypes.int32, device="CPU"),
-    "masked_lm_positions": Tensor(np.concatenate([s["masked_lm_positions"] for s in data], axis=0), dtype=dtypes.int32, device="CPU"),
-    "masked_lm_ids": Tensor(np.concatenate([s["masked_lm_ids"] for s in data], axis=0), dtype=dtypes.int32, device="CPU"),
-    "masked_lm_weights": Tensor(np.concatenate([s["masked_lm_weights"] for s in data], axis=0), dtype=dtypes.float32, device="CPU"),
-    "next_sentence_labels": Tensor(np.concatenate([s["next_sentence_labels"] for s in data], axis=0), dtype=dtypes.int32, device="CPU"),
+    "input_ids": Tensor(np.concatenate([s["input_ids"] for s in data], axis=0), dtype=dtypes.float32),
+    "input_mask": Tensor(np.concatenate([s["input_mask"] for s in data], axis=0), dtype=dtypes.default_float),
+    "segment_ids": Tensor(np.concatenate([s["segment_ids"] for s in data], axis=0), dtype=dtypes.float32),
+    "masked_lm_positions": Tensor(np.concatenate([s["masked_lm_positions"] for s in data], axis=0), dtype=dtypes.float32),
+    "masked_lm_ids": Tensor(np.concatenate([s["masked_lm_ids"] for s in data], axis=0), dtype=dtypes.float32),
+    "masked_lm_weights": Tensor(np.concatenate([s["masked_lm_weights"] for s in data], axis=0), dtype=dtypes.float32),
+    "next_sentence_labels": Tensor(np.concatenate([s["next_sentence_labels"] for s in data], axis=0), dtype=dtypes.float32),
  }

 def load_file(file: str):
@@ -225,8 +223,14 @@ def batch_load_train_bert(BS:int):
  assert cycle_length > 0, "cycle_length must be greater than 0"

  dataset = InterleavedDataset(train_files, cycle_length)
+  buffer = [dataset.get() for _ in range(1000)]
  while True:
-    yield process_batch_bert([dataset.get() for _ in range(BS)])
+    batch = []
+    for _ in range(BS):
+      index = random.randint(0, 999)
+      batch.append(buffer[index])
+      buffer[index] = dataset.get()
+    yield process_batch_bert(batch)

 # Reference: https://github.com/mlcommons/training/blob/1c8a098ae3e70962a4f7422c0b0bd35ae639e357/language_model/tensorflow/bert/run_pretraining.py, Line 416
 def batch_load_val_bert(BS:int):
@@ -263,8 +267,8 @@ 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].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()
+    X[idx].contiguous().realize().lazydata.base.realized.as_buffer(force_zero_copy=True)[:] = x.tobytes()
+    Y[idx].contiguous().realize().lazydata.base.realized.as_buffer(force_zero_copy=True)[:] = y.tobytes()

    queue_out.put(idx)
  queue_out.put(None)
@@ -314,7 +318,7 @@ def batch_load_unet3d(preprocessed_dataset_dir:Path, batch_size:int=6, val:bool=
      proc = Process(target=load_unet3d_data, args=(preprocessed_dataset_dir, seed, queue_in, queue_out, X, Y))
      proc.daemon = True
      proc.start()
-
+      
      procs.append(proc)

    for bc in range(batch_count):
@@ -350,462 +354,6 @@ def batch_load_unet3d(preprocessed_dataset_dir:Path, batch_size:int=6, val:bool=
      # happens with BENCHMARK set
      pass

-### RetinaNet
-
-def load_retinanet_data(base_dir:Path, val:bool, queue_in:Queue, queue_out:Queue,
-                        imgs:Tensor, boxes:Tensor, labels:Tensor, matches:Tensor|None=None,
-                        anchors:Tensor|None=None, seed:int|None=None):
-  from extra.datasets.openimages import image_load, random_horizontal_flip, resize
-  from examples.mlperf.helpers import box_iou, find_matches, generate_anchors
-  import torch
-
-  while (data:=queue_in.get()) is not None:
-    idx, img, tgt = data
-    img = image_load(base_dir, img["subset"], img["file_name"])
-
-    if val:
-      img = resize(img)[0]
-    else:
-      if seed is not None:
-        np.random.seed(seed)
-        random.seed(seed)
-        torch.manual_seed(seed)
-
-      img, tgt = random_horizontal_flip(img, tgt)
-      img, tgt, _ = resize(img, tgt=tgt)
-      match_quality_matrix = box_iou(tgt["boxes"], (anchor := np.concatenate(generate_anchors((800, 800)))))
-      match_idxs = find_matches(match_quality_matrix, allow_low_quality_matches=True)
-      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].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].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
-
-    queue_out.put(idx)
-  queue_out.put(None)
-
-def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, shuffle:bool=True, seed:int|None=None):
-  def _enqueue_batch(bc):
-    from extra.datasets.openimages import prepare_target
-    for idx in range(bc * batch_size, (bc+1) * batch_size):
-      img = dataset.loadImgs(next(dataset_iter))[0]
-      ann = dataset.loadAnns(dataset.getAnnIds(img_id:=img["id"]))
-      tgt = prepare_target(ann, img_id, (img["height"], img["width"]))
-
-      if img_ids is not None:
-        img_ids[idx] = img_id
-
-      if img_sizes is not None:
-        img_sizes[idx] = tgt["image_size"]
-
-      queue_in.put((idx, img, tgt))
-
-  def _setup_shared_mem(shm_name:str, size:tuple[int, ...], dtype:dtypes) -> tuple[shared_memory.SharedMemory, Tensor]:
-    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}")
-    return shm, shm_tensor
-
-  image_ids = sorted(dataset.imgs.keys())
-  batch_count = min(32, len(image_ids) // batch_size)
-
-  queue_in, queue_out = Queue(), Queue()
-  procs, data_out_count = [], [0] * batch_count
-
-  shm_imgs, imgs = _setup_shared_mem("retinanet_imgs", (batch_size * batch_count, 800, 800, 3), dtypes.uint8)
-
-  if val:
-    boxes, labels, matches, anchors = None, None, None, None
-    img_ids, img_sizes = [None] * (batch_size * batch_count), [None] * (batch_size * batch_count)
-  else:
-    img_ids, img_sizes = None, None
-    shm_boxes, boxes = _setup_shared_mem("retinanet_boxes", (batch_size * batch_count, 120087, 4), dtypes.float32)
-    shm_labels, labels = _setup_shared_mem("retinanet_labels", (batch_size * batch_count, 120087), dtypes.int64)
-    shm_matches, matches = _setup_shared_mem("retinanet_matches", (batch_size * batch_count, 120087), dtypes.int64)
-    shm_anchors, anchors = _setup_shared_mem("retinanet_anchors", (batch_size * batch_count, 120087, 4), dtypes.float64)
-
-  shutdown = False
-  class Cookie:
-    def __init__(self, bc):
-      self.bc = bc
-    def __del__(self):
-      if not shutdown:
-        try: _enqueue_batch(self.bc)
-        except StopIteration: pass
-
-  def shuffle_indices(indices, seed):
-    rng = random.Random(seed)
-    rng.shuffle(indices)
-
-  if shuffle: shuffle_indices(image_ids, seed=seed)
-  dataset_iter = iter(image_ids)
-
-  try:
-    for _ in range(cpu_count()):
-      proc = Process(
-        target=load_retinanet_data,
-        args=(base_dir, val, queue_in, queue_out, imgs, boxes, labels),
-        kwargs={"matches": matches, "anchors": anchors, "seed": seed}
-      )
-      proc.daemon = True
-      proc.start()
-      procs.append(proc)
-
-    for bc in range(batch_count):
-      _enqueue_batch(bc)
-
-    for _ in range(len(image_ids) // batch_size):
-      while True:
-        bc = queue_out.get() // batch_size
-        data_out_count[bc] += 1
-        if data_out_count[bc] == batch_size: break
-
-      data_out_count[bc] = 0
-
-      if val:
-        yield (imgs[bc * batch_size:(bc + 1) * batch_size],
-               img_ids[bc * batch_size:(bc + 1) * batch_size],
-               img_sizes[bc * batch_size:(bc + 1) * batch_size],
-               Cookie(bc))
-      else:
-        yield (imgs[bc * batch_size:(bc + 1) * batch_size],
-               boxes[bc * batch_size:(bc + 1) * batch_size],
-               labels[bc * batch_size:(bc + 1) * batch_size],
-               matches[bc * batch_size:(bc + 1) * batch_size],
-               anchors[bc * batch_size:(bc + 1) * batch_size],
-               Cookie(bc))
-  finally:
-    shutdown = True
-
-    for _ in procs: queue_in.put(None)
-    queue_in.close()
-
-    for _ in procs:
-      while queue_out.get() is not None: pass
-    queue_out.close()
-
-    # shutdown processes
-    for proc in procs: proc.join()
-
-    shm_imgs.close()
-
-    if not val:
-      shm_boxes.close()
-      shm_labels.close()
-      shm_matches.close()
-      shm_anchors.close()
-
-    try:
-      shm_imgs.unlink()
-
-      if not val:
-        shm_boxes.unlink()
-        shm_labels.unlink()
-        shm_matches.unlink()
-        shm_anchors.unlink()
-    except FileNotFoundError:
-      # happens with BENCHMARK set
-      pass
-
-# stable diffusion callbacks to match mlperf ref; declared here because they're pickled
-def filter_dataset(sample:dict): return {k:v for k,v in sample.items() if k in {'npy', 'txt'}}
-def collate(batch:list[dict]):
-  ret = {"npy": [], "txt": [], "__key__": []}
-  for sample in batch:
-    for k,v in sample.items():
-      ret[k].append(v)
-  return ret
-def collate_fn(batch): return batch
-
-# Reference (code): https://github.com/mlcommons/training/blob/2f4a93fb4888180755a8ef55f4b977ef8f60a89e/stable_diffusion/ldm/data/webdatasets.py, Line 55
-# Reference (params): https://github.com/mlcommons/training/blob/ab4ae1ca718d7fe62c369710a316dff18768d04b/stable_diffusion/configs/train_01x08x08.yaml, Line 107
-def batch_load_train_stable_diffusion(urls:str, BS:int):
-  import webdataset
-  dataset = webdataset.WebDataset(urls=urls, resampled=True, cache_size=-1, cache_dir=None)
-  dataset = dataset.shuffle(size=1000)
-  dataset = dataset.decode()
-  dataset = dataset.map(filter_dataset)
-  dataset = dataset.batched(BS, partial=False, collation_fn=collate)
-  dataset = webdataset.WebLoader(dataset, batch_size=None, shuffle=False, num_workers=1, persistent_workers=True, collate_fn=collate_fn)
-
-  for x in dataset:
-    assert isinstance(x, dict) and all(isinstance(k, str) for k in x.keys()) and all(isinstance(v, list) for v in x.values())
-    assert all(isinstance(moment_mean_logvar, np.ndarray) and moment_mean_logvar.shape==(1,8,64,64) for moment_mean_logvar in x["npy"])
-    assert all(isinstance(caption, str) for caption in x["txt"])
-    yield x
-
-# llama3
-
-class BinIdxDataset:
-  def __init__(self, base_path:Path):
-    self.idx_t = Tensor(base_path.with_name(f"{base_path.name}.idx"))
-    self.idx = TensorIO(self.idx_t)
-
-    # parse idx file
-    magic = self.idx.read(9)
-    assert magic == b"MMIDIDX\x00\x00", "invalid index file format"
-    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: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))
-
-    start = self.idx.tell()
-    end = start + self.count * dtypes.int32.itemsize
-    self.sizes = self.idx_t[start:end].bitcast(dtypes.int32).numpy()
-
-    start = end
-    end = start + self.count * dtypes.int64.itemsize
-    self.pointers = self.idx_t[start:end].bitcast(dtypes.int64).numpy()
-
-    start = end
-    end = start + doc_count * dtypes.int64.itemsize
-    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"))
-
-  def _index(self, idx) -> tuple[int, int]:
-    return int(self.pointers[idx]), int(self.sizes[idx])
-
-  def get(self, idx, offset:int=0, length:int|None=None):
-    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].bitcast(self.dtype).to(None)
-
-# https://docs.nvidia.com/megatron-core/developer-guide/latest/api-guide/datasets.html
-class GPTDataset:
-  def __init__(self, base_path:Path, samples:int, seqlen:int, seed:int, shuffle:bool):
-    self.samples, self.seqlen = samples, seqlen
-    self.shuffle = shuffle
-    self.rng = np.random.RandomState(seed)
-
-    self.indexed_dataset = BinIdxDataset(base_path)
-
-    # check for cache
-    cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
-    cache_path = base_path.with_name(f"{base_path.name}.{cache_hash}.index_cache")
-    print(f"try loading GPTDataset from {cache_path}...")
-    if cache_path.exists():
-      print("cache found, loading...")
-      with open(cache_path, "rb") as f:
-        self.doc_idx, self.sample_idx, self.shuffle_idx = pickle.load(f)
-    else:
-      print("cache not found, building index...")
-      self.doc_idx = self._build_doc_idx()
-      self.sample_idx = self._build_sample_idx()
-      self.shuffle_idx = self._build_shuffle_idx()
-      # save cache
-      with open(cache_path, "wb") as f:
-        pickle.dump((self.doc_idx, self.sample_idx, self.shuffle_idx), f)
-
-  def __getitem__(self, idx):
-    if idx is None:
-      text = self._get(0)
-    else:
-      text = self._get(idx)
-
-    return text
-
-  def _get(self, idx):
-    idx = self.shuffle_idx[idx]
-
-    doc_idx_beg, doc_idx_beg_offset = self.sample_idx[idx]
-    doc_idx_end, doc_idx_end_offset = self.sample_idx[idx + 1]
-
-    doc_ids, sample_parts = [], []
-
-    if doc_idx_beg == doc_idx_end:
-      doc_ids.append(self.doc_idx[doc_idx_beg])
-
-      sample_parts.append(
-          self.indexed_dataset.get(
-            int(self.doc_idx[doc_idx_beg]), offset=int(doc_idx_beg_offset), length=int(doc_idx_end_offset - doc_idx_beg_offset + 1)))
-    else:
-      for i in range(doc_idx_beg, doc_idx_end + 1):
-        doc_ids.append(self.doc_idx[i])
-
-        offset = 0 if i > doc_idx_beg else doc_idx_beg_offset
-        length = None if i < doc_idx_end else int(doc_idx_end_offset + 1)
-        sample_parts.append(self.indexed_dataset.get(int(self.doc_idx[i]), offset=int(offset), length=length))
-
-    # concat all parts
-    text = Tensor.cat(*sample_parts)
-
-    return text
-
-  @functools.cached_property
-  def tokens_per_epoch(self) -> int:
-    return sum(self.indexed_dataset.sizes.tolist())
-
-  @functools.cached_property
-  def num_epochs(self) -> int:
-    # we need enough epochs to cover the requested amount of tokens
-    num_epochs = 1
-    num_tokens = self.tokens_per_epoch
-    while num_tokens < self.samples * self.seqlen:
-      num_epochs += 1
-      num_tokens += self.tokens_per_epoch
-    return num_epochs
-
-  # https://github.com/NVIDIA/Megatron-LM/blob/94bd476bd840c2fd4c3ebfc7448c2af220f4832b/megatron/core/datasets/gpt_dataset.py#L558
-  def _build_doc_idx(self):
-    print(f"building doc_idx for {self.num_epochs=}, {self.indexed_dataset.count=}")
-    st = time.perf_counter()
-    # doc_idx = np.mgrid[:self.num_epochs, :self.indexed_dataset.count][1]
-    doc_idx = np.arange(self.indexed_dataset.count).reshape(1, -1).repeat(self.num_epochs, axis=0).flatten()
-    doc_idx = doc_idx.astype(np.int32)
-    at = time.perf_counter()
-    if self.shuffle: self.rng.shuffle(doc_idx)
-    print(f"doc_idx built in {at - st:.3f}s, shuffled in {time.perf_counter() - at:.3f}s")
-    return doc_idx
-
-  def _build_sample_idx(self):
-    print(f"building sample_idx for {self.samples=}, {self.seqlen=}, {self.doc_idx.shape[0]=}")
-    sample_idx_max = max(self.doc_idx.shape[0], self.indexed_dataset.sizes.max())
-    sample_idx = np.empty((self.samples + 1, 2), dtype=np.int64 if sample_idx_max > dtypes.int32.max else np.int32)
-
-    sample_idx_idx, doc_idx_idx, doc_offset = 0, 0, 0
-    sample_idx[sample_idx_idx, 0], sample_idx[sample_idx_idx, 1] = doc_idx_idx, doc_offset
-    sample_idx_idx += 1
-
-    for _ in tqdm(range(1, self.samples + 1)):
-      remaining_seqlen = self.seqlen + 1
-      while remaining_seqlen > 0:
-        doc_idx = int(self.doc_idx[doc_idx_idx])
-        doc_len = int(self.indexed_dataset.sizes[doc_idx]) - doc_offset
-        remaining_seqlen -= doc_len
-        if remaining_seqlen <= 0:
-          doc_offset += remaining_seqlen + doc_len - 1
-          remaining_seqlen = 0
-        else:
-          if doc_idx_idx == len(self.doc_idx) - 1:
-            assert sample_idx_idx == self.samples
-            doc_idx = int(self.doc_idx[doc_idx_idx])
-            doc_offset = int(self.indexed_dataset.sizes[doc_idx]) - 1
-            break
-          doc_idx_idx += 1
-          doc_offset = 0
-
-      sample_idx[sample_idx_idx, 0], sample_idx[sample_idx_idx, 1] = doc_idx_idx, doc_offset
-      sample_idx_idx += 1
-
-    return sample_idx
-
-  def _build_shuffle_idx(self):
-    print(f"building shuffle_idx for {self.samples=}")
-    st = time.perf_counter()
-    shuffle_idx = np.arange(self.samples, dtype=np.int32)
-    at = time.perf_counter()
-    if self.shuffle: self.rng.shuffle(shuffle_idx)
-    print(f"shuffle_idx built in {at - st:.3f}s, shuffled in {time.perf_counter() - at:.3f}s")
-    return shuffle_idx
-
-class BlendedGPTDataset:
-  def __init__(self, paths:list[Path], weights:list[float], samples:int, seqlen:int, seed:int, shuffle:bool):
-    self.shuffle = shuffle
-    self.rng = np.random.RandomState(seed)
-
-    # normalize weights
-    total_weight = sum(weights)
-    self.weights = [w / total_weight for w in weights]
-
-    self.samples = samples
-    surplus = 0.005
-    samples_per_blend = [math.ceil(math.ceil(self.samples * w) * (1 + surplus)) for w in self.weights]
-
-    self.datasets = [GPTDataset(path, samples_per_blend[i], seqlen, seed + i, shuffle) for i,path in enumerate(paths)]
-
-    # check for cache
-    cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
-    cache_path = paths[0].with_name(f"{paths[0].name}.{cache_hash}.blend_cache")
-    print(f"try loading BlendedGPTDataset from {cache_path}...")
-    if cache_path.exists():
-      print("cache found, loading...")
-      with open(cache_path, "rb") as f:
-        self.dataset_idx, self.dataset_sample_idx = pickle.load(f)
-    else:
-      print("cache not found, building index...")
-      self.dataset_idx, self.dataset_sample_idx = self._build_blend_idx()
-      # save cache
-      with open(cache_path, "wb") as f:
-        pickle.dump((self.dataset_idx, self.dataset_sample_idx), f)
-
-  def get(self, idx:int):
-    tokens = self.datasets[self.dataset_idx[idx]][self.dataset_sample_idx[idx]]
-    return tokens
-
-  def _build_blend_idx(self):
-    dataset_idx = np.zeros(self.samples, dtype=np.int16)
-    dataset_sample_idx = np.zeros(self.samples, dtype=np.int64)
-
-    unspent_datasets = set(range(len(self.datasets)))
-    dataset_sample_counts = [0] * len(self.datasets)
-
-    for i in tqdm(range(self.samples)):
-      error_argmax, error_max = 0, 0.0
-      for di in unspent_datasets:
-        error = self.weights[di] * max(i, 1) - dataset_sample_counts[di]
-        if error > error_max:
-          error_max = error
-          error_argmax = di
-
-      dataset_idx[i] = error_argmax
-      dataset_sample_idx[i] = dataset_sample_counts[error_argmax]
-
-      dataset_sample_counts[error_argmax] += 1
-
-    return dataset_idx, dataset_sample_idx
-
-def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
-  if val:
-    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)
-
-  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_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):
    assert not val, "validation set is not supported due to different sizes on inputs"
@@ -826,26 +374,6 @@ if __name__ == "__main__":
      for x,y,c in batch_load_resnet(val=val):
        pbar.update(x.shape[0])

-  def load_retinanet(val):
-    from extra.datasets.openimages import BASEDIR, download_dataset
-    from pycocotools.coco import COCO
-    dataset = COCO(download_dataset(base_dir:=getenv("BASEDIR", BASEDIR), "validation" if val else "train"))
-    with tqdm(total=len(dataset.imgs.keys())) as pbar:
-      for x in batch_load_retinanet(dataset, val, base_dir):
-        pbar.update(x[0].shape[0])
-
-  def load_llama3(val):
-    bs = 24
-    samples = 5760 if val else 1_200_000 * 1152
-    seqlen = 8192
-
-    max_, min_ = 0, math.inf
-    for tokens in tqdm(batch_load_llama3(bs, samples, seqlen, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=5760, val=bool(val)), total=samples//bs):
-      max_ = max(max_, tokens.shape[1])
-      min_ = min(min_, tokens.shape[1])
-    print(f"max seq length: {max_}")
-    print(f"min seq length: {min_}")
-
  load_fn_name = f"load_{getenv('MODEL', 'resnet')}"
  if load_fn_name in globals():
    globals()[load_fn_name](getenv("VAL", 1))
@@ -1,7 +1,6 @@
 from collections import OrderedDict
 import unicodedata
 from typing import Optional
-import math
 import numpy as np
 from tinygrad.nn import state
 from tinygrad.tensor import Tensor, dtypes
@@ -196,161 +195,46 @@ def get_bert_qa_prediction(features, example, start_end_logits):
  return "empty"

 def get_mlperf_bert_config():
-  """benchmark is BERT-large"""
-  ret = {"attention_probs_dropout_prob": 0.1, "hidden_dropout_prob": 0.1, "vocab_size": 30522, "type_vocab_size": 2, "max_position_embeddings": 512}
+  """Config is BERT-large"""
+  return {
+    "attention_probs_dropout_prob": 0.1,
+    "hidden_dropout_prob": 0.1,
+    "hidden_size": 1024,
+    "intermediate_size": 4096,
+    "max_position_embeddings": 512,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 24,
+    "type_vocab_size": 2,
+    "vocab_size": 30522
+  }

-  match (bert_size:=getenv("BERT_SIZE", "large")):
-    case "large": ret.update({"hidden_size": 1024, "intermediate_size": 4096, "num_attention_heads": 16, "num_hidden_layers": 24})
-    case "tiny": ret.update({"hidden_size": 128, "intermediate_size": 512, "num_attention_heads": 2, "num_hidden_layers": 2})
-    case _: raise RuntimeError(f"unhandled {bert_size=}")
-
-  if (bert_layers:=getenv("BERT_LAYERS")): ret["num_hidden_layers"] = bert_layers
-  return ret
-
-def get_mlperf_bert_model():
+def get_mlperf_bert_model(checkpoint_path:Optional[str]=None):
  from extra.models import bert
  from examples.mlperf.initializers import LinearBert, EmbeddingBert, LayerNormBert

  bert.Linear = LinearBert
-  bert.Embedding = EmbeddingBert
+  bert.Embedding = EmbeddingBert 
  bert.LayerNorm = LayerNormBert

  from extra.models.bert import BertForPretraining
  config = get_mlperf_bert_config()
  if getenv("DISABLE_DROPOUT", 0):
    config["hidden_dropout_prob"] = config["attention_probs_dropout_prob"] = 0.0
-  return BertForPretraining(**config)
+  model = BertForPretraining(**config)
+  return model.load_from_pretrained(checkpoint_path) if checkpoint_path else model

-def get_fake_data_bert(BS:int):
+def get_data_bert(GPUS:list[str], it):
+  data: dict[str, Tensor] = next(it)
+  for key in data.keys(): data[key].shard_(GPUS, axis=0)
+  return data
+
+def get_fake_data_bert(GPUS:list[str], 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.float32).contiguous().shard_(GPUS, axis=0),
+    "input_mask": Tensor.empty((BS, 512), dtype=dtypes.default_float).contiguous().shard_(GPUS, axis=0),
+    "segment_ids": Tensor.empty((BS, 512), dtype=dtypes.float32).contiguous().shard_(GPUS, axis=0),
+    "masked_lm_positions": Tensor.empty((BS, 76), dtype=dtypes.float32).contiguous().shard_(GPUS, axis=0),
+    "masked_lm_ids": Tensor.empty((BS, 76), dtype=dtypes.float32).contiguous().shard_(GPUS, axis=0),
+    "masked_lm_weights": Tensor.empty((BS, 76), dtype=dtypes.float32).contiguous().shard_(GPUS, axis=0),
+    "next_sentence_labels": Tensor.empty((BS, 1), dtype=dtypes.float32).contiguous().shard_(GPUS, axis=0),
  }
-
-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:
-  BELOW_LOW_THRESHOLD, BETWEEN_THRESHOLDS = -1, -2
-
-  def _set_low_quality_matches_(matches:np.ndarray, all_matches:np.ndarray, match_quality_matrix:np.ndarray):
-    highest_quality_foreach_gt = np.max(match_quality_matrix, axis=1)
-    pred_inds_to_update = np.nonzero(match_quality_matrix == highest_quality_foreach_gt[:, None])[1]
-    matches[pred_inds_to_update] = all_matches[pred_inds_to_update]
-
-  assert low_threshold <= high_threshold
-
-  matched_vals, matches = match_quality_matrix.max(axis=0), match_quality_matrix.argmax(axis=0)
-  all_matches = np.copy(matches) if allow_low_quality_matches else None
-  below_low_threshold = matched_vals < low_threshold
-  between_thresholds = (matched_vals >= low_threshold) & (matched_vals < high_threshold)
-  matches[below_low_threshold] = BELOW_LOW_THRESHOLD
-  matches[between_thresholds] = BETWEEN_THRESHOLDS
-
-  if allow_low_quality_matches:
-    assert all_matches is not None
-    _set_low_quality_matches_(matches, all_matches, match_quality_matrix)
-
-  return matches
-
-def box_iou(boxes1:np.ndarray, boxes2:np.ndarray) -> np.ndarray:
-  def _box_area(boxes:np.ndarray) -> np.ndarray: return (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
-
-  def _box_inter_union(boxes1:np.ndarray, boxes2:np.ndarray) -> tuple[np.ndarray, np.ndarray]:
-    area1, area2 = _box_area(boxes1), _box_area(boxes2)
-    lt, rb = np.maximum(boxes1[:, None, :2], boxes2[:, :2]), np.minimum(boxes1[:, None, 2:], boxes2[:, 2:])
-    wh = np.clip(rb - lt, a_min=0, a_max=None)
-    inter = wh[:, :, 0] * wh[:, :, 1]
-    union = area1[:, None] + area2 - inter
-    return inter, union
-
-  inter, union = _box_inter_union(boxes1, boxes2)
-  return inter / union
-
-def generate_anchors(input_size:tuple[int, int], scales:Optional[tuple[Tensor, ...]]=None, aspect_ratios:Optional[tuple[Tensor, ...]]=None) -> list[np.ndarray]:
-  def _compute_grid_sizes(input_size:tuple[int, int]) -> np.ndarray:
-    return np.ceil(np.array(input_size)[None, :] / 2 ** np.arange(3, 8)[:, None])
-
-  scales = tuple((i, int(i * 2 ** (1/3)), int(i * 2 ** (2/3))) for i in 2 ** np.arange(5, 10)) if scales is None else scales
-  aspect_ratios = ((0.5, 1.0, 2.0),) * len(scales) if aspect_ratios is None else aspect_ratios
-  aspect_ratios = tuple(ar for ar in aspect_ratios)
-  grid_sizes = _compute_grid_sizes(input_size)
-
-  assert len(scales) == len(aspect_ratios) == len(grid_sizes), "scales, aspect_ratios, and grid_sizes must have the same length"
-
-  anchors = []
-  for s, ar, gs in zip(scales, aspect_ratios, grid_sizes):
-    s, ar = np.array(s), np.array(ar)
-    h_ratios = np.sqrt(ar)
-    w_ratios = 1 / h_ratios
-    ws = (w_ratios[:, None] * s[None, :]).reshape(-1)
-    hs = (h_ratios[:, None] * s[None, :]).reshape(-1)
-    base_anchors = (np.stack([-ws, -hs, ws, hs], axis=1) / 2).round()
-    stride_h, stride_w = input_size[0] // gs[0], input_size[1] // gs[1]
-    shifts_x, shifts_y = np.meshgrid(np.arange(gs[1]) * stride_w, np.arange(gs[0]) * stride_h)
-    shifts_x, shifts_y = shifts_x.reshape(-1), shifts_y.reshape(-1)
-    shifts = np.stack([shifts_x, shifts_y, shifts_x, shifts_y], axis=1, dtype=np.float32)
-    anchors.append((shifts[:, None] + base_anchors[None, :]).reshape(-1, 4))
-
-  return anchors
-
-
-class BoxCoder(object):
-  def __init__(self, weights, bbox_xform_clip=math.log(1000. / 16), apply_to_remove=True):
-    self.weights = weights
-    self.bbox_xform_clip = bbox_xform_clip
-    self.apply_to_remove = apply_to_remove
-
-  def encode(self, reference_boxes, proposals):
-    TO_REMOVE = self.apply_to_remove  # TODO remove
-    ex_widths = proposals[..., 2] - proposals[..., 0] + TO_REMOVE
-    ex_heights = proposals[..., 3] - proposals[..., 1] + TO_REMOVE
-    ex_ctr_x = proposals[..., 0] + 0.5 * ex_widths
-    ex_ctr_y = proposals[..., 1] + 0.5 * ex_heights
-
-    gt_widths = reference_boxes[..., 2] - reference_boxes[..., 0] + TO_REMOVE
-    gt_heights = reference_boxes[..., 3] - reference_boxes[..., 1] + TO_REMOVE
-    gt_ctr_x = reference_boxes[..., 0] + 0.5 * gt_widths
-    gt_ctr_y = reference_boxes[..., 1] + 0.5 * gt_heights
-
-    wx, wy, ww, wh = self.weights
-    targets_dx = wx * (gt_ctr_x - ex_ctr_x) / ex_widths
-    targets_dy = wy * (gt_ctr_y - ex_ctr_y) / ex_heights
-    targets_dw = ww * Tensor.log(gt_widths / ex_widths)
-    targets_dh = wh * Tensor.log(gt_heights / ex_heights)
-
-    targets = Tensor.stack(targets_dx, targets_dy, targets_dw, targets_dh, dim=-1)
-    return targets
-
-  def decode(self, rel_codes, boxes):
-    boxes = boxes.cast(rel_codes.dtype)
-    rel_codes = rel_codes
-
-    TO_REMOVE = self.apply_to_remove  # TODO remove
-    widths = boxes[:, 2] - boxes[:, 0] + TO_REMOVE
-    heights = boxes[:, 3] - boxes[:, 1] + TO_REMOVE
-    ctr_x = boxes[:, 0] + 0.5 * widths
-    ctr_y = boxes[:, 1] + 0.5 * heights
-
-    wx, wy, ww, wh = self.weights
-    dx = rel_codes[:, 0::4] / wx
-    dy = rel_codes[:, 1::4] / wy
-    dw = rel_codes[:, 2::4] / ww
-    dh = rel_codes[:, 3::4] / wh
-
-    # Prevent sending too large values into Tensor.exp()
-    dw = dw.clip(min_=dw.min(), max_=self.bbox_xform_clip)
-    dh = dh.clip(min_=dh.min(), max_=self.bbox_xform_clip)
-
-    pred_ctr_x = dx * widths[:, None] + ctr_x[:, None]
-    pred_ctr_y = dy * heights[:, None] + ctr_y[:, None]
-    pred_w = dw.exp() * widths[:, None]
-    pred_h = dh.exp() * heights[:, None]
-    x = pred_ctr_x - 0.5 * pred_w
-    y = pred_ctr_y - 0.5 * pred_h
-    w = pred_ctr_x + 0.5 * pred_w - 1
-    h = pred_ctr_y + 0.5 * pred_h - 1
-    pred_boxes = Tensor.stack(x, y, w, h).permute(1,2,0).reshape(rel_codes.shape[0], rel_codes.shape[1])
-    return pred_boxes
@@ -1,10 +1,8 @@
 import math
-from typing import Union
+from typing import Union, Tuple

 from tinygrad import Tensor, nn, dtypes
-from tinygrad.helpers import prod, argfix, Context
-from tinygrad.nn.state import get_parameters
-from extra.models.unet import UNetModel
+from tinygrad.helpers import prod, argfix

 # rejection sampling truncated randn
 def rand_truncn(*shape, dtype=None, truncstds=2, **kwargs) -> Tensor:
@@ -19,10 +17,6 @@ def he_normal(*shape, a: float = 0.00, **kwargs) -> Tensor:
  std = math.sqrt(2.0 / (1 + a ** 2)) / math.sqrt(prod(argfix(*shape)[1:])) / 0.87962566103423978
  return std * rand_truncn(*shape, **kwargs)

-# Stable Diffusion v2 training uses default torch gelu, which doesn't use tanh approximation
-def gelu_erf(x:Tensor) -> Tensor:
-  return 0.5 * x * (1.0 + (x / 1.4142135623730951).erf())
-
 class Conv2dHeNormal(nn.Conv2d):
  def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
    super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
@@ -45,7 +39,7 @@ class LinearBert(nn.Linear):
  def __init__(self, in_features, out_features, bias=True, std=0.02):
    self.weight = std * rand_truncn(out_features, in_features, dtype=dtypes.float32)
    self.bias = Tensor.zeros(out_features, dtype=dtypes.float32) if bias else None
-
+  
  def __call__(self, x:Tensor):
    return x.cast(dtypes.default_float).linear(self.weight.cast(dtypes.default_float).transpose(), self.bias.cast(dtypes.default_float) if self.bias is not None else None)

@@ -59,10 +53,10 @@ 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)
+    return (arange == idx).mul(vals).sum(2, acc_dtype=vals.dtype)

 class LayerNormBert:
-  def __init__(self, normalized_shape:Union[int, tuple[int, ...]], eps:float=1e-12, elementwise_affine:bool=True):
+  def __init__(self, normalized_shape:Union[int, Tuple[int, ...]], eps:float=1e-12, elementwise_affine:bool=True):
    self.normalized_shape = (normalized_shape,) if isinstance(normalized_shape, int) else tuple(normalized_shape)
    self.axis, self.eps, self.elementwise_affine = tuple(-1-i for i in range(len(self.normalized_shape))), eps, elementwise_affine
    self.weight, self.bias = (Tensor.ones(*self.normalized_shape, dtype=dtypes.float32), Tensor.zeros(*self.normalized_shape, dtype=dtypes.float32)) if elementwise_affine else (None, None)
@@ -72,118 +66,3 @@ class LayerNormBert:
    xn = x.cast(dtypes.float32).layernorm(eps=self.eps, axis=self.axis).cast(x.dtype)
    if not self.elementwise_affine: return xn
    return (xn * self.weight.cast(dtypes.default_float) + self.bias.cast(dtypes.default_float))
-
-class FrozenBatchNorm2dRetinaNet(nn.BatchNorm2d):
-  def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1):
-    self.eps, self.track_running_stats, self.momentum = eps, track_running_stats, momentum
-
-    self.weight = Tensor.ones(sz, dtype=dtypes.float32, requires_grad=False) if affine else None
-    self.bias = Tensor.zeros(sz, dtype=dtypes.float32, requires_grad=False) if affine else None
-
-    if track_running_stats: self.running_mean, self.running_var = Tensor.zeros(sz, dtype=dtypes.float32, requires_grad=False), Tensor.ones(sz, dtype=dtypes.float32, requires_grad=False)
-    self.num_batches_tracked = Tensor.zeros(1, dtype=dtypes.long, requires_grad=False)
-
-  def __call__(self, x:Tensor) -> Tensor:
-    batch_mean, batch_var = super().calc_stats(x.cast(dtypes.float32))
-    if self.track_running_stats and Tensor.training:
-      self.running_mean.assign((1-self.momentum) * self.running_mean + self.momentum * batch_mean.detach().cast(self.running_mean.dtype))
-      self.running_var.assign((1-self.momentum) * self.running_var + self.momentum * x.numel()/(x.numel()-x.shape[1]) * batch_var.detach().cast(self.running_var.dtype))
-      self.num_batches_tracked += 1
-    return x.cast(dtypes.float32).batchnorm(self.weight, self.bias, batch_mean, batch_var.add(self.eps).rsqrt()).cast(x.dtype)
-
-class Conv2dNormalRetinaNet(nn.Conv2d):
-  def __init__(self, in_channels:int, out_channels:int, kernel_size:int|tuple[int, ...],
-               stride:int=1, padding:int|tuple[int, ...]|str=0, dilation:int=1, groups:int=1,
-               bias:bool=True, prior_prob:float|None=None):
-    super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
-    self.weight = Tensor.normal(*self.weight.shape, std=0.01, dtype=dtypes.float32)
-    if bias:
-      if prior_prob:
-        prior_prob = Tensor(prior_prob, device=self.bias.device, dtype=dtypes.float32).expand(*self.bias.shape)
-        self.bias = -(((1 - prior_prob) / prior_prob).log())
-      else: self.bias = Tensor.zeros_like(self.bias, dtype=dtypes.float32)
-
-  def __call__(self, x:Tensor) -> Tensor:
-    return x.conv2d(self.weight.cast(dtypes.default_float), self.bias.cast(dtypes.default_float) if self.bias is not None else None,
-                    groups=self.groups, stride=self.stride, padding=self.padding)
-
-class Conv2dKaimingUniformRetinaNet(nn.Conv2d):
-  def __init__(self, in_channels:int, out_channels:int, kernel_size:int|tuple[int, ...],
-               stride:int=1, padding:int|tuple[int, ...]|str=0, dilation:int=1, groups:int=1,
-               bias:bool=True):
-    super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
-    self.weight = Tensor.kaiming_uniform(*self.weight.shape, a=1, dtype=dtypes.float32)
-    if bias: self.bias = Tensor.zeros_like(self.bias, dtype=dtypes.float32)
-
-  def __call__(self, x:Tensor) -> Tensor:
-    return x.conv2d(self.weight.cast(dtypes.default_float), self.bias.cast(dtypes.default_float) if self.bias is not None else None,
-                    groups=self.groups, stride=self.stride, padding=self.padding)
-
-class Conv2dRetinaNet(nn.Conv2d):
-  def __init__(self, in_channels:int, out_channels:int, kernel_size:int|tuple[int, ...],
-               stride:int=1, padding:int|tuple[int, ...]|str=0, dilation:int=1, groups:int=1,
-               bias:bool=True):
-    super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
-    scale = 1 / math.sqrt(in_channels * prod(self.kernel_size))
-    self.weight = Tensor.uniform(out_channels, in_channels//groups, *self.kernel_size, low=-scale, high=scale, dtype=dtypes.float32)
-    self.bias: Tensor|None = Tensor.uniform(out_channels, low=-scale, high=scale, dtype=dtypes.float32) if bias else None
-
-  def __call__(self, x:Tensor) -> Tensor:
-    return x.conv2d(self.weight.cast(dtypes.default_float), self.bias.cast(dtypes.default_float) if self.bias is not None else None,
-                    groups=self.groups, stride=self.stride, dilation=self.dilation, padding=self.padding)
-
-# copy torch AMP: isolate mixed precision to just the below autocast ops, instead of using dtypes.default_float which affects all new Tensors
-class AutocastLinear(nn.Linear):
-  cast_dtype=dtypes.bfloat16 # enable monkeypatching of the mixed precision dtype
-  def __call__(self, x:Tensor) -> Tensor:
-    dtype = type(self).cast_dtype
-    return x.cast(dtype).linear(self.weight.cast(dtype).transpose(), self.bias.cast(dtype) if self.bias is not None else None)
-
-class AutocastConv2d(nn.Conv2d):
-  cast_dtype=dtypes.bfloat16
-  def __call__(self, x:Tensor) -> Tensor:
-    dtype = type(self).cast_dtype
-    return x.cast(dtype).conv2d(self.weight.cast(dtype), self.bias.cast(dtype), self.groups, self.stride, self.dilation, self.padding)
-
-# copy torch AMP: upcast to float32 before GroupNorm and LayerNorm
-class AutocastGroupNorm(nn.GroupNorm):
-  def __call__(self, x:Tensor) -> Tensor:
-    return super().__call__(x.cast(dtypes.float32))
-
-class AutocastLayerNorm(nn.LayerNorm):
-  def __call__(self, x:Tensor) -> Tensor:
-    return super().__call__(x.cast(dtypes.float32))
-
-def zero_module(module):
-  for p in get_parameters(module): p.assign(Tensor.zeros_like(p).contiguous())
-
-# Stable Diffusion mlperf reference doesn't call scaled_dot_product_attention
-# copy torch AMP: upcast to float32 before softmax on CUDA
-def attn_f32_softmax(q:Tensor, k:Tensor, v:Tensor) -> Tensor:
-  return (q.matmul(k.transpose(-2,-1), dtype=dtypes.float32) / math.sqrt(q.shape[-1])).softmax(-1).cast(q.dtype) @ v
-
-def init_stable_diffusion(version:str, pretrained:str, devices:list[str]):
-  from examples.stable_diffusion import StableDiffusion
-  from tinygrad.nn.state import safe_load, safe_save, load_state_dict, get_state_dict
-  from tempfile import TemporaryDirectory
-  model = StableDiffusion(version=version, pretrained=pretrained)
-  unet:UNetModel = model.model.diffusion_model
-
-  # this prevents extra consumption of memory, enabling much larger BS
-  Tensor.realize(*get_parameters(unet))
-  with TemporaryDirectory(prefix="unet_init") as tmp:
-    safe_save(get_state_dict(unet), init_fn:=f"{tmp}/init_model.safetensors")
-    load_state_dict(unet, safe_load(init_fn))
-
-  sqrt_alphas_cumprod = model.alphas_cumprod.sqrt().realize()
-  sqrt_one_minus_alphas_cumprod = (1 - model.alphas_cumprod).sqrt().realize()
-
-  if len(devices) > 1:
-    to_move = [sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod]
-    if version == "v2-mlperf-train": to_move += get_parameters(unet) + get_parameters(model.cond_stage_model)
-    for p in to_move:
-      p.to_(devices)
-    with Context(BEAM=0):
-      Tensor.realize(*to_move)
-
-  return model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod
@@ -1,29 +1,6 @@
 from examples.mlperf.metrics import dice_score
-from tinygrad import Tensor

 def dice_ce_loss(pred, tgt):
  ce = pred.permute(0, 2, 3, 4, 1).sparse_categorical_crossentropy(tgt.squeeze(1))
  dice = (1.0 - dice_score(pred, tgt, argmax=False, to_one_hot_x=False)).mean()
  return (dice + ce) / 2
-
-def sigmoid_focal_loss(pred:Tensor, tgt:Tensor, alpha:float=0.25, gamma:float=2.0, reduction:str="none") -> Tensor:
-  assert reduction in ["mean", "sum", "none"], f"unsupported reduction {reduction}"
-  p, ce_loss = pred.sigmoid(), pred.binary_crossentropy_logits(tgt, reduction="none")
-  p_t = p * tgt + (1 - p) * (1 - tgt)
-  loss = ce_loss * ((1 - p_t) ** gamma)
-
-  if alpha >= 0:
-    alpha_t = alpha * tgt + (1 - alpha) * (1 - tgt)
-    loss = loss * alpha_t
-
-  if reduction == "mean": loss = loss.mean()
-  elif reduction == "sum": loss = loss.sum()
-  return loss
-
-def l1_loss(pred:Tensor, tgt:Tensor, reduction:str="none") -> Tensor:
-  assert reduction in ["mean", "sum", "none"], f"unsupported reduction {reduction}"
-  loss = (pred - tgt).abs()
-
-  if reduction == "mean": loss = loss.mean()
-  elif reduction == "sum": loss = loss.sum()
-  return loss
@@ -1,9 +1,7 @@
-import math
-from tinygrad import dtypes, Tensor
+from tinygrad import Tensor, dtypes
 from tinygrad.nn.optim import Optimizer

 from extra.lr_scheduler import LR_Scheduler
-from typing import Callable

 # https://github.com/mlcommons/training/blob/e237206991d10449d9675d95606459a3cb6c21ad/image_classification/tensorflow2/lars_util.py
 class PolynomialDecayWithWarmup(LR_Scheduler):
@@ -22,39 +20,3 @@ class PolynomialDecayWithWarmup(LR_Scheduler):
    warmup_lr = (self.epoch_counter * (1.0 / self.warmup)) * self.initial_lr
    x = (1 - (self.epoch_counter - self.warmup) / (self.epochs - self.warmup + 1))
    return (self.epoch_counter <= self.warmup).where(warmup_lr, (self.initial_lr - self.end_lr) * x ** self.power + self.end_lr).cast(self.optimizer.lr.dtype)
-
-class CosineAnnealingLRWithWarmup(LR_Scheduler):
-  def __init__(self, optimizer:Optimizer, base_lr, end_lr, warmup_steps:int, decay_steps:int):
-    assert warmup_steps > 0 and decay_steps > 0
-    super().__init__(optimizer)
-    self.base_lr = base_lr
-    self.end_lr = end_lr
-    self.warmup_steps = warmup_steps
-    self.decay_steps = decay_steps
-    # set lr for first warmup step
-    self.optimizer.lr.assign(self.get_lr()).realize()
-
-  def get_lr(self):
-    warmup_lr = ((self.epoch_counter+1) / self.warmup_steps) * self.base_lr
-    decay_lr = self.end_lr + 0.5 * (self.base_lr-self.end_lr) * (1 + (((self.epoch_counter+1-self.warmup_steps)/self.decay_steps) * math.pi).cos())
-    return (self.epoch_counter < self.warmup_steps).where(warmup_lr, decay_lr).cast(self.optimizer.lr.dtype)
-
-# Reference: https://github.com/mlcommons/training/blob/64b14a9abc74e08779a175abca7d291f8c957632/stable_diffusion/ldm/lr_scheduler.py, Lines 36-97
-class LambdaLinearScheduler:
-  def __init__(self, warm_up_steps:int, f_min:float, f_max:float, f_start:float, cycle_lengths:int):
-    self.lr_warm_up_steps, self.f_min, self.f_max, self.f_start, self.cycle_lengths = warm_up_steps, f_min, f_max, f_start, cycle_lengths
-
-  def schedule(self, n:Tensor) -> Tensor:
-    warm_up = (n < self.lr_warm_up_steps)
-    f_warm_up = (self.f_max - self.f_start) / self.lr_warm_up_steps * n + self.f_start
-    return warm_up.where(f_warm_up, self.f_min + (self.f_max - self.f_min) * (self.cycle_lengths - n) / (self.cycle_lengths))
-
-# based on torch.optim.lr_scheduler.LambdaLR
-class LambdaLR(LR_Scheduler):
-  def __init__(self, optimizer:Optimizer, base_lr:Tensor, lr_lambda:Callable):
-    super().__init__(optimizer)
-    self.base_lr, self.lr_lambda = base_lr, lr_lambda
-    self.step()
-
-  def get_lr(self):
-    return self.base_lr * self.lr_lambda(self.epoch_counter - 1)
@@ -1,6 +1,6 @@
-import re, string
+import re
+import string
 from collections import Counter
-from tinygrad import Tensor

 def levenshtein(a, b):
  n, m = len(a), len(b)
@@ -59,11 +59,3 @@ def f1_score(x, y):
  p = ns / len(xt)
  r = ns / len(yt)
  return 2 * p * r / (p + r)
-
-def log_perplexity(logit:Tensor, target:Tensor, ignore_index:int|None=None):
-  # logit has shape (n_samples, seq_len, vocab_size), target has shape (n_samples, seq_len)
-  assert logit.ndim == 3, logit.ndim
-  assert target.ndim == 2, target.ndim
-  assert logit.shape[:2] == target.shape, f"{logit.shape[:2]=}, {target.shape=}"
-  log_prob = logit.log_softmax(axis=-1)
-  return log_prob.transpose(1, 2).nll_loss(target, ignore_index=ignore_index)
@@ -1,66 +1,65 @@
-import time, math, os
+import time
 start = time.perf_counter()
 from pathlib import Path
 import numpy as np
 from tinygrad import Tensor, Device, dtypes, GlobalCounters, TinyJit
 from tinygrad.nn.state import get_parameters, load_state_dict, safe_load
-from tinygrad.helpers import getenv, Context, prod
-from extra.bench_log import BenchEvent, WallTimeEvent
+from tinygrad.helpers import getenv
 def tlog(x): print(f"{x:25s}  @ {time.perf_counter()-start:5.2f}s")

 def eval_resnet():
-  with WallTimeEvent(BenchEvent.FULL):
-    # Resnet50-v1.5
-    from extra.models.resnet import ResNet50
-    tlog("imports")
-    GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 6))]
-    for x in GPUS: Device[x]
-    tlog("got devices")    # NOTE: this is faster with rocm-smi running
+  Tensor.no_grad = True
+  # Resnet50-v1.5
+  from extra.models.resnet import ResNet50
+  tlog("imports")
+  GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 6))]
+  for x in GPUS: Device[x]
+  tlog("got devices")    # NOTE: this is faster with rocm-smi running

-    class ResnetRunner:
-      def __init__(self, device=None):
-        self.mdl = ResNet50()
-        for x in get_parameters(self.mdl) if device else []: x.to_(device)
-        if (fn:=getenv("RESNET_MODEL", "")): load_state_dict(self.mdl, safe_load(fn))
-        else: self.mdl.load_from_pretrained()
-        self.input_mean = Tensor([0.485, 0.456, 0.406], device=device).reshape(1, -1, 1, 1)
-        self.input_std = Tensor([0.229, 0.224, 0.225], device=device).reshape(1, -1, 1, 1)
-      def __call__(self, x:Tensor) -> Tensor:
-        x = x.permute([0,3,1,2]).cast(dtypes.float32) / 255.0
-        x -= self.input_mean
-        x /= self.input_std
-        return self.mdl(x).log_softmax().argmax(axis=1).realize()
+  class ResnetRunner:
+    def __init__(self, device=None):
+      self.mdl = ResNet50()
+      for x in get_parameters(self.mdl) if device else []: x.to_(device)
+      if (fn:=getenv("RESNET_MODEL", "")): load_state_dict(self.mdl, safe_load(fn))
+      else: self.mdl.load_from_pretrained()
+      self.input_mean = Tensor([0.485, 0.456, 0.406], device=device).reshape(1, -1, 1, 1)
+      self.input_std = Tensor([0.229, 0.224, 0.225], device=device).reshape(1, -1, 1, 1)
+    def __call__(self, x:Tensor) -> Tensor:
+      x = x.permute([0,3,1,2]).cast(dtypes.float32) / 255.0
+      x -= self.input_mean
+      x /= self.input_std
+      return self.mdl(x).log_softmax().argmax(axis=1).realize()

-    mdl = TinyJit(ResnetRunner(GPUS))
-    tlog("loaded models")
+  mdl = TinyJit(ResnetRunner(GPUS))
+  tlog("loaded models")

-    # evaluation on the mlperf classes of the validation set from imagenet
-    from examples.mlperf.dataloader import batch_load_resnet
-    iterator = batch_load_resnet(getenv("BS", 128*6), val=getenv("VAL", 1), shuffle=False, pad_first_batch=True)
-    def data_get():
-      x,y,cookie = next(iterator)
-      return x.shard(GPUS, axis=0).realize(), y, cookie
-    n,d = 0,0
-    proc = data_get()
-    tlog("loaded initial data")
-    st = time.perf_counter()
-    while proc is not None:
-      GlobalCounters.reset()
-      proc = (mdl(proc[0]), proc[1], proc[2])  # this frees the images
-      run = time.perf_counter()
-      # load the next data here
-      try: next_proc = data_get()
-      except StopIteration: next_proc = None
-      nd = time.perf_counter()
-      y = np.array(proc[1])
-      proc = (proc[0].numpy() == y) & (y != -1)  # this realizes the models and frees the cookies
-      n += proc.sum()
-      d += (y != -1).sum()
-      et = time.perf_counter()
-      tlog(f"****** {n:5d}/{d:5d}  {n*100.0/d:.2f}% -- {(run-st)*1000:7.2f} ms to enqueue, {(et-run)*1000:7.2f} ms to realize ({(nd-run)*1000:7.2f} ms fetching). {(len(proc))/(et-st):8.2f} examples/sec. {GlobalCounters.global_ops*1e-12/(et-st):5.2f} TFLOPS")
-      st = et
-      proc, next_proc = next_proc, None
-    tlog("done")
+  # evaluation on the mlperf classes of the validation set from imagenet
+  from examples.mlperf.dataloader import batch_load_resnet
+  iterator = batch_load_resnet(getenv("BS", 128*6), val=getenv("VAL", 1), shuffle=False, pad_first_batch=True)
+  def data_get():
+    x,y,cookie = next(iterator)
+    return x.shard(GPUS, axis=0).realize(), y, cookie
+  n,d = 0,0
+  proc = data_get()
+  tlog("loaded initial data")
+  st = time.perf_counter()
+  while proc is not None:
+    GlobalCounters.reset()
+    proc = (mdl(proc[0]), proc[1], proc[2])  # this frees the images
+    run = time.perf_counter()
+    # load the next data here
+    try: next_proc = data_get()
+    except StopIteration: next_proc = None
+    nd = time.perf_counter()
+    y = np.array(proc[1])
+    proc = (proc[0].numpy() == y) & (y != -1)  # this realizes the models and frees the cookies
+    n += proc.sum()
+    d += (y != -1).sum()
+    et = time.perf_counter()
+    tlog(f"****** {n:5d}/{d:5d}  {n*100.0/d:.2f}% -- {(run-st)*1000:7.2f} ms to enqueue, {(et-run)*1000:7.2f} ms to realize ({(nd-run)*1000:7.2f} ms fetching). {(len(proc))/(et-st):8.2f} examples/sec. {GlobalCounters.global_ops*1e-12/(et-st):5.2f} TFLOPS")
+    st = et
+    proc, next_proc = next_proc, None
+  tlog("done")

 def eval_unet3d():
  # UNet3D
@@ -82,43 +81,47 @@ def eval_unet3d():

 def eval_retinanet():
  # RetinaNet with ResNeXt50_32X4D
-  from examples.mlperf.dataloader import batch_load_retinanet
-  from extra.datasets.openimages import normalize, download_dataset, BASEDIR
  from extra.models.resnet import ResNeXt50_32X4D
  from extra.models.retinanet import RetinaNet
+  mdl = RetinaNet(ResNeXt50_32X4D())
+  mdl.load_from_pretrained()
+
+  input_mean = Tensor([0.485, 0.456, 0.406]).reshape(1, -1, 1, 1)
+  input_std = Tensor([0.229, 0.224, 0.225]).reshape(1, -1, 1, 1)
+  def input_fixup(x):
+    x = x.permute([0,3,1,2]) / 255.0
+    x -= input_mean
+    x /= input_std
+    return x
+
+  from extra.datasets.openimages import download_dataset, iterate, BASEDIR
  from pycocotools.coco import COCO
  from pycocotools.cocoeval import COCOeval
  from contextlib import redirect_stdout
-  tlog("imports")
-
-  mdl = RetinaNet(ResNeXt50_32X4D())
-  mdl.load_from_pretrained()
-  tlog("loaded models")
-
-  coco = COCO(download_dataset(base_dir:=getenv("BASEDIR", BASEDIR), 'validation'))
+  coco = COCO(download_dataset(base_dir:=getenv("BASE_DIR", BASEDIR), 'validation'))
  coco_eval = COCOeval(coco, iouType="bbox")
  coco_evalimgs, evaluated_imgs, ncats, narea = [], [], len(coco_eval.params.catIds), len(coco_eval.params.areaRng)
-  tlog("loaded dataset")

-  iterator = batch_load_retinanet(coco, True, Path(base_dir), getenv("BS", 8), shuffle=False)
-  def data_get():
-    x, img_ids, img_sizes, cookie = next(iterator)
-    return x.to(Device.DEFAULT).realize(), img_ids, img_sizes, cookie
-  n = 0
-  proc = data_get()
-  tlog("loaded initial data")
+  from tinygrad.engine.jit import TinyJit
+  mdlrun = TinyJit(lambda x: mdl(input_fixup(x)).realize())
+
+  n, bs = 0, 8
  st = time.perf_counter()
-  while proc is not None:
-    GlobalCounters.reset()
-    proc = (mdl(normalize(proc[0])), proc[1], proc[2], proc[3])
-    run = time.perf_counter()
-    # load the next data here
-    try: next_proc = data_get()
-    except StopIteration: next_proc = None
-    nd = time.perf_counter()
-    predictions, img_ids = mdl.postprocess_detections(proc[0].numpy(), orig_image_sizes=proc[2]), proc[1]
-    pd = time.perf_counter()
-    coco_results  = [{"image_id": img_ids[i], "category_id": label, "bbox": box.tolist(), "score": score}
+  for x, targets in iterate(coco, base_dir, bs):
+    dat = Tensor(x.astype(np.float32))
+    mt = time.perf_counter()
+    if dat.shape[0] == bs:
+      outs = mdlrun(dat).numpy()
+    else:
+      mdlrun._jit_cache = []
+      outs =  mdl(input_fixup(dat)).numpy()
+    et = time.perf_counter()
+    predictions = mdl.postprocess_detections(outs, input_size=dat.shape[1:3], orig_image_sizes=[t["image_size"] for t in targets])
+    ext = time.perf_counter()
+    n += len(targets)
+    print(f"[{n}/{len(coco.imgs)}] == {(mt-st)*1000:.2f} ms loading data, {(et-mt)*1000:.2f} ms to run model, {(ext-et)*1000:.2f} ms for postprocessing")
+    img_ids = [t["image_id"] for t in targets]
+    coco_results  = [{"image_id": targets[i]["image_id"], "category_id": label, "bbox": box.tolist(), "score": score}
      for i, prediction in enumerate(predictions) for box, score, label in zip(*prediction.values())]
    with redirect_stdout(None):
      coco_eval.cocoDt = coco.loadRes(coco_results)
@@ -126,18 +129,13 @@ def eval_retinanet():
      coco_eval.evaluate()
    evaluated_imgs.extend(img_ids)
    coco_evalimgs.append(np.array(coco_eval.evalImgs).reshape(ncats, narea, len(img_ids)))
-    n += len(proc[0])
-    et = time.perf_counter()
-    tlog(f"****** {(run-st)*1000:7.2f} ms to enqueue, {(et-run)*1000:7.2f} ms to realize ({(nd-run)*1000:7.2f} ms fetching, {(pd-run)*1000:4.2f} ms postprocess_detections). {(len(proc))/(et-st):8.2f} examples/sec. {GlobalCounters.global_ops*1e-12/(et-st):5.2f} TFLOPS")
-    st = et
-    proc, next_proc = next_proc, None
+    st = time.perf_counter()

  coco_eval.params.imgIds = evaluated_imgs
  coco_eval._paramsEval.imgIds = evaluated_imgs
  coco_eval.evalImgs = list(np.concatenate(coco_evalimgs, -1).flatten())
  coco_eval.accumulate()
  coco_eval.summarize()
-  tlog("done")

 def eval_rnnt():
  # RNN-T
@@ -204,307 +202,49 @@ def eval_bert():

    st = time.perf_counter()

-def eval_llama3():
-  from extra.models.llama import Transformer
-  from examples.llama3 import MODEL_PARAMS, load, convert_from_huggingface
-  from tinygrad.helpers import tqdm
-
-  BASEDIR = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
-  BS = getenv("BS", 4)
-  SMALL = getenv("SMALL", 0)
-  SEQLEN = getenv("SEQLEN", 8192)
-  MODEL_PATH = Path(getenv("MODEL_PATH", "/raid/weights/llama31_8b/"))
-
-  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)
-
-  # load weights
-  weights = load(str(MODEL_PATH / "model.safetensors.index.json"))
-  if "model.embed_tokens.weight" in weights:
-    print("converting from huggingface format")
-    weights = convert_from_huggingface(weights, params["n_layers"], params["n_heads"], params["n_kv_heads"])
-
-  load_state_dict(model, weights, strict=False, consume=True)
-
-  @TinyJit
-  def eval_step(model, tokens):
-    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
-    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    return loss.flatten().float()
-
-  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):
-    GlobalCounters.reset()
-    losses += eval_step(model, tokens).tolist()
-    tqdm.write(f"loss: {np.mean(losses)}")
-
-  log_perplexity = np.mean(losses)
-  print(f"Log Perplexity: {log_perplexity}")
-
-# NOTE: BEAM hangs on 8xmi300x with DECODE_BS=384 in final realize below; function is declared here for external testing
-@TinyJit
-def vae_decode(x:Tensor, vae, disable_beam=False) -> Tensor:
-  from examples.stable_diffusion import AutoencoderKL
-  assert isinstance(vae, AutoencoderKL)
-  x = vae.post_quant_conv(1./0.18215 * x)
-
-  x = vae.decoder.conv_in(x)
-  x = vae.decoder.mid(x)
-  for i, l in enumerate(vae.decoder.up[::-1]):
-    print("decode", x.shape)
-    for b in l['block']: x = b(x)
-    if 'upsample' in l:
-      bs,c,py,px = x.shape
-      x = x.reshape(bs, c, py, 1, px, 1).expand(bs, c, py, 2, px, 2).reshape(bs, c, py*2, px*2)
-      x = l['upsample']['conv'](x)
-    if i == len(vae.decoder.up) - 1 and disable_beam:
-      with Context(BEAM=0): x.realize()
-    else: x.realize()
-  x = vae.decoder.conv_out(vae.decoder.norm_out(x).swish())
-
-  x = ((x + 1.0) / 2.0).clip(0.0, 1.0)
-  return x
-
-def eval_stable_diffusion():
-  import csv, PIL, sys
+def eval_mrcnn():
  from tqdm import tqdm
-  from examples.mlperf.initializers import init_stable_diffusion, gelu_erf
-  from examples.stable_diffusion import AutoencoderKL
-  from extra.models.unet import UNetModel
-  from tinygrad.nn.state import load_state_dict, torch_load
-  from tinygrad.helpers import BEAM
-  from extra.models import clip
-  from extra.models.clip import FrozenOpenClipEmbedder
-  from extra.models.clip import OpenClipEncoder
-  from extra.models.inception import FidInceptionV3
+  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()

-  config = {}
-  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
-  for x in GPUS: Device[x]
-  print(f"running eval on {GPUS}")
-  seed               = config["seed"]                   = getenv("SEED", 12345)
-  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
-  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
-  CONTEXT_BS         = config["CONTEXT_BS"]             = getenv("CONTEXT_BS", 1 * len(GPUS))
-  DENOISE_BS         = config["DENOISE_BS"]             = getenv("DENOISE_BS", 1 * len(GPUS))
-  DECODE_BS          = config["DECODE_BS"]              = getenv("DECODE_BS", 1 * len(GPUS))
-  INCEPTION_BS       = config["INCEPTION_BS"]           = getenv("INCEPTION_BS", 1 * len(GPUS))
-  CLIP_BS            = config["CLIP_BS"]                = getenv("CLIP_BS", 1 * len(GPUS))
-  EVAL_CKPT_DIR      = config["EVAL_CKPT_DIR"]          = getenv("EVAL_CKPT_DIR", "")
-  STOP_IF_CONVERGED  = config["STOP_IF_CONVERGED"]      = getenv("STOP_IF_CONVERGED", 0)
+  bbox_output = '/tmp/results_bbox.json'
+  mask_output = '/tmp/results_mask.json'

-  if (WANDB := getenv("WANDB", "")):
-    import wandb
-    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
+  accumulate_predictions_for_coco([], bbox_output, rm=True)
+  accumulate_predictions_for_coco([], mask_output, rm=True)

-  assert EVAL_CKPT_DIR != "", "provide a directory with checkpoints to be evaluated"
-  print(f"running eval on checkpoints in {EVAL_CKPT_DIR}\nSEED={seed}")
-  eval_queue:list[tuple[int, Path]] = []
-  for p in Path(EVAL_CKPT_DIR).iterdir():
-    if p.name.endswith(".safetensors"):
-      ckpt_iteration = p.name.split(".safetensors")[0]
-      assert ckpt_iteration.isdigit(), f"invalid checkpoint name: {p.name}, expected <digits>.safetensors"
-      eval_queue.append((int(ckpt_iteration), p))
-  assert len(eval_queue), f'no files ending with ".safetensors" were found in {EVAL_CKPT_DIR}'
-  print(sorted(eval_queue, reverse=True))
+  #TODO: bs > 1 not as accurate
+  bs = 1

-  Tensor.manual_seed(seed)  # seed for weight initialization
-  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-eval", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
+  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

-  # load prompts for generating images for validation; 2 MB of data total
-  with open(DATADIR / "coco2014" / "val2014_30k.tsv") as f:
-    reader = csv.DictReader(f, delimiter="\t")
-    eval_inputs:list[dict] = [{"image_id": int(row["image_id"]), "id": int(row["id"]), "caption": row["caption"]} for row in reader]
-  assert len(eval_inputs) == 30_000
-  # NOTE: the clip weights are the same between model.cond_stage_model and clip_encoder
-  eval_timesteps = list(reversed(range(1, 1000, 20)))
-
-  original_device, Device.DEFAULT = Device.DEFAULT, "CPU"
-  # The choice of alphas_prev[0] = alphas_cumprod[0] seems arbitrary, but it's how the mlperf ref does it:
-  #   alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
-  eval_alphas_prev = model.alphas_cumprod[0:1].cat(model.alphas_cumprod[list(range(1, 1000, 20))[:-1]]).to(GPUS).realize()
-  inception = FidInceptionV3().load_from_pretrained(CKPTDIR / "inception" / "pt_inception-2015-12-05-6726825d.pth")
-  vision_cfg = {'width': 1280, 'layers': 32, 'd_head': 80, 'image_size': 224, 'patch_size': 14}
-  text_cfg = {'width': 1024, 'n_heads': 16, 'layers': 24, 'vocab_size': 49408, 'ctx_length': 77}
-  clip.gelu = gelu_erf
-  clip_encoder = OpenClipEncoder(1024, text_cfg, vision_cfg)
-  loaded = torch_load(CKPTDIR / "clip" / "open_clip_pytorch_model.bin")
-  loaded.update({"attn_mask": clip_encoder.attn_mask, "mean": clip_encoder.mean, "std": clip_encoder.std})
-  load_state_dict(clip_encoder, loaded)
-  Device.DEFAULT=original_device
-
-  @TinyJit
-  def denoise_step(x:Tensor, x_x:Tensor, t_t:Tensor, uc_c:Tensor, sqrt_alphas_cumprod_t:Tensor, sqrt_one_minus_alphas_cumprod_t:Tensor,
-                    alpha_prev:Tensor, unet:UNetModel, GPUS) -> Tensor:
-    out_uncond, out = unet(x_x, t_t, uc_c).to("CPU").reshape(-1, 2, 4, 64, 64).chunk(2, dim=1)
-    out_uncond = out_uncond.squeeze(1).shard(GPUS,axis=0)
-    out = out.squeeze(1).shard(GPUS,axis=0)
-    v_t = out_uncond + 8.0 * (out - out_uncond)
-    e_t = sqrt_alphas_cumprod_t * v_t + sqrt_one_minus_alphas_cumprod_t * x
-    pred_x0 = sqrt_alphas_cumprod_t * x - sqrt_one_minus_alphas_cumprod_t * v_t
-    dir_xt = (1. - alpha_prev).sqrt() * e_t
-    x_prev = alpha_prev.sqrt() * pred_x0 + dir_xt
-    return x_prev.realize()
-
-  def shard_tensor(t:Tensor) -> Tensor: return t.shard(GPUS, axis=0) if len(GPUS) > 1 else t.to(GPUS[0])
-  def get_batch(whole:Tensor, i:int, bs:int) -> tuple[Tensor, int]:
-    batch = whole[i: i + bs].to("CPU")
-    if (unpadded_bs:=batch.shape[0]) < bs:
-      batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
-    return batch, unpadded_bs 
-
-  @Tensor.train(mode=False)
-  def eval_unet(eval_inputs:list[dict], unet:UNetModel, cond_stage:FrozenOpenClipEmbedder, first_stage:AutoencoderKL,
-                inception:FidInceptionV3, clip:OpenClipEncoder) -> tuple[float, float]:
-    # Eval is divided into 5 jits, one per model
-    # It doesn't make sense to merge these jits, e.g. unet repeats 50 times in isolation; images fork to separate inception/clip
-    # We're generating and scoring 30,000 images per eval, and all the data can flow through one jit at a time
-    # To maximize throughput for each jit, we have only one model/jit on the GPU at a time, and pool outputs from each jit off-GPU
-    for model in (unet, first_stage, inception, clip):
-      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
-
-    uc_written = False
-    models = (cond_stage, unet, first_stage, inception, clip)
-    jits = (jit_context:=TinyJit(cond_stage.embed_tokens), denoise_step, vae_decode, jit_inception:=TinyJit(inception),
-            jit_clip:=TinyJit(clip.get_clip_score))
-    all_bs = (CONTEXT_BS, DENOISE_BS, DECODE_BS, INCEPTION_BS, CLIP_BS)
-    if (EVAL_SAMPLES:=getenv("EVAL_SAMPLES", 0)) and EVAL_SAMPLES > 0:
-      eval_inputs = eval_inputs[0:EVAL_SAMPLES]
-    output_shapes = [(ns:=len(eval_inputs),77), (ns,77,1024), (ns,4,64,64), (ns,3,512,512), (ns,2048), (ns,)]
-    # Writing progress to disk lets us resume eval if we crash
-    stages = ["tokens", "embeds", "latents", "imgs", "inception", "clip"]
-    disk_tensor_names, disk_tensor_shapes = stages + ["end", "uc"], output_shapes + [(6,), (1,77,1024)]
-    if not all(os.path.exists(f"{EVAL_CKPT_DIR}/{name}.bytes") for name in disk_tensor_names):
-      for name, shape in zip(disk_tensor_names, disk_tensor_shapes):
-        file = Path(f"{EVAL_CKPT_DIR}/{name}.bytes")
-        file.unlink(missing_ok=True)
-        with file.open("wb") as f: f.truncate(prod(shape) * 4)
-    progress = {name: Tensor.empty(*shape, device=f"disk:{EVAL_CKPT_DIR}/{name}.bytes", dtype=dtypes.int if name in {"tokens", "end"} else dtypes.float)
-                for name, shape in zip(disk_tensor_names, disk_tensor_shapes)}
-
-    def embed_tokens(tokens:Tensor) -> Tensor:
-      nonlocal uc_written
-      if not uc_written:
-        with Context(BEAM=0): progress["uc"].assign(cond_stage.embed_tokens(cond_stage.tokenize("").to(GPUS)).to("CPU").realize()).realize()
-        uc_written = True
-      return jit_context(shard_tensor(tokens))
-
-    def generate_latents(embeds:Tensor) -> Tensor:
-      uc_c = Tensor.stack(progress["uc"].to("CPU").expand(bs, 77, 1024), embeds, dim=1).reshape(-1, 77, 1024)
-      uc_c = shard_tensor(uc_c)
-      x = shard_tensor(Tensor.randn(bs,4,64,64))
-      for step_idx, timestep in enumerate(tqdm(eval_timesteps)):
-        reversed_idx = Tensor([50 - step_idx - 1], device=GPUS)
-        alpha_prev = eval_alphas_prev[reversed_idx]
-        ts = Tensor.full(bs, fill_value=timestep, dtype=dtypes.int, device="CPU")
-        ts_ts = shard_tensor(ts.cat(ts))
-        ts = shard_tensor(ts)
-        sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
-        sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
-        x_x = shard_tensor(Tensor.stack(x.to("CPU"), x.to("CPU"), dim=1).reshape(-1, 4, 64, 64))
-        x.assign(denoise_step(x, x_x, ts_ts, uc_c, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t, alpha_prev, unet, GPUS)).realize()
-      return x
-
-    def decode_latents(latents:Tensor) -> Tensor: return vae_decode(shard_tensor(latents), first_stage, disable_beam=True)
-    def generate_inception(imgs:Tensor) -> Tensor: return jit_inception(shard_tensor(imgs))[:,:,0,0]
-
-    def calc_clip_scores(batch:Tensor, batch_tokens:Tensor) -> Tensor:
-      # Tensor.interpolate does not yet support bicubic, so we use PIL
-      batch = (batch.to(GPUS[0]).permute(0,2,3,1) * 255).clip(0, 255).cast(dtypes.uint8).numpy()
-      batch = [np.array(PIL.Image.fromarray(batch[i]).resize((224,224), PIL.Image.BICUBIC)) for i in range(bs)]
-      batch = shard_tensor(Tensor(np.stack(batch, axis=0).transpose(0,3,1,2), device="CPU").realize())
-      batch = batch.cast(dtypes.float) / 255
-      batch = (batch - model.mean) / model.std
-      batch = jit_clip(shard_tensor(batch_tokens), batch)
-      return batch
-
-    callbacks = (embed_tokens, generate_latents, decode_latents, generate_inception, calc_clip_scores)
-
-    # save every forward pass output to disk; NOTE: this needs ~100 GB disk space because 30k images are large
-    def stage_progress(stage_idx:int) -> int: return progress["end"].to("CPU")[stage_idx].item()
-    if stage_progress(0) < len(eval_inputs):
-      tokens = []
-      for i in tqdm(range(0, len(eval_inputs), CONTEXT_BS)):
-        subset = [cond_stage.tokenize(row["caption"], device="CPU") for row in eval_inputs[i: i+CONTEXT_BS]]
-        tokens.append(Tensor.cat(*subset, dim=0).realize())
-      progress["tokens"].assign(Tensor.cat(*tokens, dim=0).realize()).realize()
-      progress["end"][0:1].assign(Tensor([len(eval_inputs)], dtype=dtypes.int)).realize()
-    prev_stage = "tokens"
-    tokens = progress["tokens"]
-
-    # wrapper code for every model
-    for stage_idx, model, jit, bs, callback in zip(range(1,6), models, jits, all_bs, callbacks):
-      stage = stages[stage_idx]
-      if stage_progress(stage_idx) >= len(eval_inputs):
-        prev_stage = stage
-        continue # use cache
-      t0 = time.perf_counter()
-      print(f"starting eval with model: {model}")
-      if stage_idx == 1: inputs = tokens
-      elif stage_idx == 5: inputs = progress["imgs"]
-      else: inputs = progress[prev_stage]
-
-      Tensor.realize(*[p.to_(GPUS) for p in get_parameters(model)])
-      for batch_idx in tqdm(range(stage_progress(stage_idx), inputs.shape[0], bs)):
-        t1 = time.perf_counter()
-        batch, unpadded_bs = get_batch(inputs, batch_idx, bs)
-        if isinstance(model, OpenClipEncoder): batch = callback(batch, get_batch(tokens, batch_idx, bs)[0].realize())
-        else: batch = callback(batch)
-        # to(GPUS[0]) is necessary for this to work, without that the result is still on GPUS, probably due to a bug
-        batch = batch.to(GPUS[0]).to("CPU")[0:unpadded_bs].realize()
-        progress[stage][batch_idx: batch_idx + bs].assign(batch).realize()
-        # keep track of what our last output was, so we can resume from there if we crash in this loop
-        progress["end"][stage_idx: stage_idx + 1].assign(Tensor([batch_idx + bs], dtype=dtypes.int)).realize()
-        print(f"model: {model}, batch_idx: {batch_idx}, elapsed: {(time.perf_counter() - t1):.2f}")
-      del batch
-        
-      jit.reset()
-      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
-      print(f"done with model: {model}, elapsed: {(time.perf_counter() - t0):.2f}")
-      prev_stage = stage
-
-    inception_stats_fn = str(DATADIR / "coco2014" / "val2014_30k_stats.npz")
-    fid_score = inception.compute_score(progress["inception"].to("CPU"), inception_stats_fn)
-    clip_score = progress["clip"].to(GPUS[0]).mean().item()
-    for name in disk_tensor_names:
-      Path(f"{EVAL_CKPT_DIR}/{name}.bytes").unlink(missing_ok=True)
-    
-    if EVAL_SAMPLES and BEAM:
-      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-      sys.exit() # Don't eval additional models; we don't care about clip/fid scores when running BEAM on eval sample subset
-
-    return clip_score, fid_score
-
-  # evaluate checkpoints in reverse chronological order
-  for ckpt_iteration, p in sorted(eval_queue, reverse=True):
-    unet_ckpt = safe_load(p)
-    load_state_dict(unet, unet_ckpt)
-    clip_score, fid_score = eval_unet(eval_inputs, unet, model.cond_stage_model, model.first_stage_model, inception, clip_encoder)
-    converged = True if clip_score >= 0.15 and fid_score <= 90 else False
-    print(f"eval results for {EVAL_CKPT_DIR}/{p.name}: clip={clip_score}, fid={fid_score}, converged={converged}")
-    if WANDB:
-      wandb.log({"eval/ckpt_iteration": ckpt_iteration, "eval/clip_score": clip_score, "eval/fid_score": fid_score})
-    if converged and STOP_IF_CONVERGED:
-      print(f"Convergence detected, exiting early before evaluating other checkpoints due to STOP_IF_CONVERGED={STOP_IF_CONVERGED}")
-      sys.exit()
-
-  # for testing
-  return clip_score, fid_score, ckpt_iteration
+  evaluate_predictions_on_coco(bbox_output, iou_type='bbox')
+  evaluate_predictions_on_coco(mask_output, iou_type='segm')

 if __name__ == "__main__":
  # inference only
  Tensor.training = False
+  Tensor.no_grad = True

-  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():
@@ -60,6 +60,7 @@ def spec_mrcnn():
 if __name__ == "__main__":
  # inference only for now
  Tensor.training = False
+  Tensor.no_grad = True

  for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
    nm = f"spec_{m}"
@@ -1,57 +0,0 @@
-#!/usr/bin/env bash
-# adapted from https://github.com/mlcommons/training/blob/4bdf5c8ed218ad76565a2ba1ac27c919ccc6d689/stable_diffusion/README.md
-
-# setup dirs
-
-DATA=/raid/datasets/stable_diffusion
-
-LAION=$DATA/laion-400m/webdataset-moments-filtered 
-COCO=$DATA/coco2014
-mkdir -p $LAION $COCO
-
-CKPT=/raid/weights/stable_diffusion
-mkdir -p $CKPT/clip $CKPT/sd $CKPT/inception
-
-# download data
-
-# if rclone isn't installed system-wide / in your PATH, put the executable path in quotes below
-#RCLONE=""
-RCLONE="rclone"
-
-## VAE-encoded image latents, from 6.1M image subset of laion-400m
-## about 1 TB for whole download
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/ ${LAION} --include="*.tar" -P
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/sha512sums.txt ${LAION} -P
-cd $LAION && grep -E '\.tar$' sha512sums.txt | sha512sum -c --quiet - && \
-  echo "All .tar files verified" || { echo "Checksum failure when validating downloaded Laion moments"; exit 1; }
-
-## prompts and FID statistics from 30k image subset of coco2014
-## 33 MB
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k.tsv ${COCO} -P
-
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k_stats.npz ${COCO} -P
-
-# download checkpoints
-
-## clip (needed for text and vision encoders for validation)
-CLIP_WEIGHTS_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin"
-CLIP_WEIGHTS_SHA256="9a78ef8e8c73fd0df621682e7a8e8eb36c6916cb3c16b291a082ecd52ab79cc4"
-CLIP_CONFIG_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/raw/main/open_clip_config.json"
-wget -N -P ${CKPT}/clip ${CLIP_WEIGHTS_URL}
-wget -N -P ${CKPT}/clip ${CLIP_CONFIG_URL}
-echo "${CLIP_WEIGHTS_SHA256}  ${CKPT}/clip/open_clip_pytorch_model.bin"                    | sha256sum -c
-
-## sd (needed for latent->image decoder for validation, also has clip text encoder for training)
-SD_WEIGHTS_URL='https://huggingface.co/stabilityai/stable-diffusion-2-base/resolve/main/512-base-ema.ckpt'
-SD_WEIGHTS_SHA256="d635794c1fedfdfa261e065370bea59c651fc9bfa65dc6d67ad29e11869a1824"
-wget -N -P ${CKPT}/sd ${SD_WEIGHTS_URL}
-echo "${SD_WEIGHTS_SHA256}  ${CKPT}/sd/512-base-ema.ckpt"                    | sha256sum -c
-
-## inception (needed for validation)
-FID_WEIGHTS_URL='https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth'
-FID_WEIGHTS_SHA1="bd836944fd6db519dfd8d924aa457f5b3c8357ff"
-wget -N -P ${CKPT}/inception ${FID_WEIGHTS_URL}
-echo "${FID_WEIGHTS_SHA1}  ${CKPT}/inception/pt_inception-2015-12-05-6726825d.pth"                    | sha1sum -c
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
-
-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
-# export BEAM_LOG_SURPASS_MAX=1
-# export BASEDIR="/raid/datasets/wiki"
-
-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-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 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 BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,17 +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 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 WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,28 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_8xMI300X"
-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 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=$(date "+%m%d%H%M")
-LOGFILE="bert_8xMI300x_${DATETIME}_${SEED}.log"
-
-# init  # TODO: without DEBUG=2 it hangs
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 DEBUG=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -4,20 +4,24 @@ This problem uses BERT for NLP.

 ## Requirements

-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
+Install tinygrad and mlperf-logging from master.
 ```
 git clone https://github.com/tinygrad/tinygrad.git
 python3 -m pip install -e ".[mlperf]"
 ```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
+Also install tqdm and tensorflow.
 ```
-pip install gdown numpy tqdm tensorflow
+pip install tqdm tensorflow
 ```

 ### tinybox_green
 Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
 This is the default on production tinybox green.

+### tinybox_red
+Disable cwsr + increase mes timeout.
+Install the custom amdgpu driver per [README](https://github.com/nimlgen/amdgpu_ubuntu_22_04/blob/v6.1.3/readme.md)
+
 # 2. Directions

 ## Steps to download and verify data
@@ -52,18 +56,18 @@ BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval

 #### Steps to run benchmark
 ```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
+examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
 ```

 ### tinybox_red

-#### Steps to run benchmark
+#### One time setup
+
 ```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
+examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
 ```
-### tinybox_8xMI300X

 #### Steps to run benchmark
 ```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
+examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
 ```
@@ -1,16 +1,13 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="."
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=72 EVAL_BS=6

-export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export BEAM=4 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=512
 export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
 export BASEDIR="/raid/datasets/wiki"

-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
+export BENCHMARK=10 DEBUG=2

 python3 examples/mlperf/model_train.py
@@ -1,12 +1,10 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="."
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=72 EVAL_BS=6

-export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export BEAM=4 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=512
 export IGNORE_JIT_FIRST_BEAM=1
 export BASEDIR="/raid/datasets/wiki"

@@ -1,14 +1,11 @@
 #!/bin/bash
-set -e  # Exit on any error

-export PYTHONPATH="." NV=1
+export PYTHONPATH="."
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=72 EVAL_BS=6

-export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export BEAM=4 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=512
 export IGNORE_JIT_FIRST_BEAM=1
 export BASEDIR="/raid/datasets/wiki"

@@ -20,7 +17,7 @@ DATETIME=$(date "+%m%d%H%M")
 LOGFILE="bert_green_${DATETIME}_${SEED}.log"

 # init
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
+BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE

 # run
 PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -4,20 +4,24 @@ This problem uses BERT for NLP.

 ## Requirements

-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
+Install tinygrad and mlperf-logging from master.
 ```
 git clone https://github.com/tinygrad/tinygrad.git
 python3 -m pip install -e ".[mlperf]"
 ```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
+Also install tqdm and tensorflow.
 ```
-pip install gdown numpy tqdm tensorflow
+pip install tqdm tensorflow
 ```

 ### tinybox_green
 Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
 This is the default on production tinybox green.

+### tinybox_red
+Disable cwsr + increase mes timeout.
+Install the custom amdgpu driver per [README](https://github.com/nimlgen/amdgpu_ubuntu_22_04/blob/v6.1.3/readme.md)
+
 # 2. Directions

 ## Steps to download and verify data
@@ -52,18 +56,18 @@ BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval

 #### Steps to run benchmark
 ```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
+examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
 ```

 ### tinybox_red

-#### Steps to run benchmark
+#### One time setup
+
 ```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
+examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
 ```
-### tinybox_8xMI300X

 #### Steps to run benchmark
 ```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
+examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
 ```
@@ -1,17 +1,13 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="."
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=72 EVAL_BS=6

-export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export BEAM=3
 export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
 export BASEDIR="/raid/datasets/wiki"

-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
+export BENCHMARK=10 DEBUG=2

 python3 examples/mlperf/model_train.py
@@ -1,12 +1,10 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="."
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=72 EVAL_BS=6

-export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export BEAM=3
 export IGNORE_JIT_FIRST_BEAM=1
 export BASEDIR="/raid/datasets/wiki"

@@ -1,14 +1,11 @@
 #!/bin/bash
-set -e  # Exit on any error

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="."
 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" GPUS=6 BS=72 EVAL_BS=6

-export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export BEAM=3
 export IGNORE_JIT_FIRST_BEAM=1
 export BASEDIR="/raid/datasets/wiki"

@@ -19,13 +16,8 @@ export SEED=$RANDOM
 DATETIME=$(date "+%m%d%H%M")
 LOGFILE="bert_red_${DATETIME}_${SEED}.log"

-export HCQDEV_WAIT_TIMEOUT_MS=100000  # prevents hang?
-
 # init
-sleep 5 && sudo rmmod amdgpu || true
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
+BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE

 # run
-# TODO: AM driver resulted in nan
-sudo modprobe amdgpu
 PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -5,4 +5,4 @@ rocm-smi --setmclk 3
 rocm-smi --setperflevel high

 # power cap to 350W
-echo "350000000" | sudo tee /sys/class/drm/card{1..6}/device/hwmon/hwmon*/power1_cap
+# echo "350000000" | sudo tee /sys/class/drm/card{1..6}/device/hwmon/hwmon*/power1_cap
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="."
 export MODEL="resnet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192

@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="."
 export MODEL="resnet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192

@@ -1,7 +1,6 @@
 #!/bin/bash
-set -e  # Exit on any error

-export PYTHONPATH="." NV=1
+export PYTHONPATH="."
 export MODEL="resnet"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
@@ -11,7 +10,7 @@ export RESET_STEP=0
 export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0

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

 export SEED=$RANDOM
 DATETIME=$(date "+%m%d%H%M")
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="."
 export MODEL="resnet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192

@@ -8,6 +8,6 @@ export RESET_STEP=0

 export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0

-export BENCHMARK=10 DEBUG=${DEBUG:-2}
+export BENCHMARK=10 DEBUG=2

 python3 examples/mlperf/model_train.py
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="."
 export MODEL="resnet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192

@@ -1,7 +1,6 @@
 #!/bin/bash
-set -e  # Exit on any error

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="."
 export MODEL="resnet"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
@@ -11,14 +10,13 @@ export RESET_STEP=0
 export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0

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

 export SEED=$RANDOM
 DATETIME=$(date "+%m%d%H%M")
 LOGFILE="resnet_red_${DATETIME}_${SEED}.log"

 # init
-sleep 5 && sudo rmmod amdgpu || true
 BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE

 # run
@@ -1,38 +0,0 @@
-# 1. Problem
-
-This problem uses RetinaNet for SSD.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-Also install the following dependencies:
-```
-pip install tqdm numpy pycocotools boto3 pandas torch torchvision
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download data
-
-Run the following:
-```
-BASEDIR=/raid/datasets/openimages python3 extra/datasets/openimages.py
-```
-
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh
-```
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -1,24 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export TRAIN_BEAM=2 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/openimages"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="retinanet_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -1,72 +0,0 @@
-#!/usr/bin/env bash
-
-DATETIME=${2:-$(date "+%m%d%H%M")}
-LOGFILE="${HOME}/logs/sd_mi300x_${DATETIME}.log"
-# UNET_CKPTDIR must be set: training saves checkpoints to this path, then a separate eval process scans this path to know which checkpoints to eval
-export UNET_CKPTDIR="${HOME}/stable_diffusion/training_checkpoints/${DATETIME}"
-mkdir -p "${HOME}/logs" "$UNET_CKPTDIR"
-
-# run this script in isolation when using the --bg flag
-if [[ "${1:-}" == "--bg" ]]; then
-  echo "logging output to $LOGFILE"
-  echo "saving UNet checkpoints to $UNET_CKPTDIR"
-  script_path="$(readlink -f "${BASH_SOURCE[0]}")"
-  nohup bash "$script_path" run "$DATETIME" >"$LOGFILE" 2>&1 & disown $!
-  exit 0
-fi
-
-# venv management
-if [[ -d .venv-sd-mlperf ]]; then
-  . .venv-sd-mlperf/bin/activate
-else
-  python3 -m venv .venv-sd-mlperf && . .venv-sd-mlperf/bin/activate
-  pip install --index-url https://download.pytorch.org/whl/cpu torch && pip install tqdm numpy ftfy regex pillow scipy wandb webdataset
-fi
-pip list
-apt list --installed | grep amdgpu
-rocm-smi --version
-modinfo amdgpu | grep version
-
-export BEAM=2 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 IGNORE_JIT_FIRST_BEAM=1 HCQDEV_WAIT_TIMEOUT_MS=300000
-export AMD_LLVM=0 # bf16 seems to require this
-export DATADIR="/raid/datasets/stable_diffusion"
-export CKPTDIR="/raid/weights/stable_diffusion"
-export EVAL_CKPT_DIR=$UNET_CKPTDIR
-export MODEL="stable_diffusion" PYTHONPATH="."
-export GPUS=8 BS=304
-export CONTEXT_BS=816 DENOISE_BS=600 DECODE_BS=384 INCEPTION_BS=560 CLIP_BS=240
-export WANDB=1
-export PARALLEL=4
-export PYTHONUNBUFFERED=1
-sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setperfdeterminism 1500 || exit 1
-
-# Retry BEAM search if script fails before BEAM COMPLETE is printed, but don't retry after that
-run_retry(){ local try=0 max=5 code tmp py pgid kids
-  while :; do
-    tmp=$(mktemp)
-    setsid bash -c 'exec env "$@"' _ "$@" > >(tee -a "$LOGFILE" | tee "$tmp") 2>&1 &
-    py=$!; pgid=$(ps -o pgid= -p "$py" | tr -d ' ')
-    wait "$py"; code=$?
-    [[ -n "$pgid" ]] && { kill -TERM -"$pgid" 2>/dev/null; sleep 1; kill -KILL -"$pgid" 2>/dev/null; }
-    kids=$(pgrep -P "$py" || true)
-    while [[ -n "$kids" ]]; do
-      kill -TERM $kids 2>/dev/null; sleep 0.5
-      kids=$(for k in $kids; do pgrep -P "$k" || true; done)
-    done
-    grep -q 'BEAM COMPLETE' "$tmp" && { rm -f "$tmp"; return 1; }
-    rm -f "$tmp"
-    ((code==0)) && return 0
-    ((try>=max)) && return 2
-    ((try++)); sleep 90; echo "try = ${try}"
-  done
-}
-
-# Power limiting to 400W is only needed if GPUs fall out of sync (causing 2.2x increased train time) at higher power, which has been observed at 450W
-sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setpoweroverdrive 750 && \
-run_retry TOTAL_CKPTS=7 python3 examples/mlperf/model_train.py; (( $? == 2 )) && { echo "training failed before BEAM completion"; exit 2; }
-sleep 90
-
-run_retry EVAL_SAMPLES=600 python3 examples/mlperf/model_eval.py; (( $? == 2 )) && { echo "eval failed before BEAM completion"; exit 2; }
-# Checkpoints will be evaluated in reverse chronological order, even if above training crashed early
-# STOP_IF_CONVERGED=1: Stop the eval after the first time convergence is detected; no more checkpoints will be evaluated after that.
-STOP_IF_CONVERGED=1 python3 examples/mlperf/model_eval.py
@@ -1,38 +0,0 @@
-{
-    "submitter": "tinycorp",
-    "division": "closed",
-    "status": "Available on-premise",
-    "system_name": "tinybox 8xMI300X",
-    "number_of_nodes": "1",
-    "host_processors_per_node": "2",
-    "host_processor_model_name": "AMD EPYC 9354",
-    "host_processor_core_count": "32",
-    "host_processor_vcpu_count": "64",
-    "host_processor_frequency": "",
-    "host_processor_caches": "",
-    "host_processor_interconnect": "",
-    "host_memory_capacity": "2304GB",
-    "host_storage_type": "NVMe SSD",
-    "host_storage_capacity": "3x 4TB raid array",
-    "host_networking": "",
-    "host_networking_topology": "",
-    "host_memory_configuration": "24x 96GB DDR5",
-    "accelerators_per_node": "8",
-    "accelerator_model_name": "AMD Instinct MI300X 192GB HBM3",
-    "accelerator_host_interconnect": "PCIe 5.0 x16",
-    "accelerator_frequency": "",
-    "accelerator_on-chip_memories": "",
-    "accelerator_memory_configuration": "HBM3",
-    "accelerator_memory_capacity": "192GB",
-    "accelerator_interconnect": "",
-    "accelerator_interconnect_topology": "",
-    "cooling": "air",
-    "hw_notes": "",
-    "framework": "tinygrad, branch mlperf_training_v5.0",
-    "other_software_stack": {
-        "python": "3.10.16",
-        "ROCm": "3.0.0+94441cb"
-    },
-    "operating_system": "Ubuntu 24.04.1 LTS",
-    "sw_notes": ""
-  }
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Ahmed Harmouche	d4bce5be8c	can't allocate this at once, need parts back	2024-12-13 13:51:45 +01:00
Ahmed Harmouche	4396ac96b5	No NamedTuple in compile.py	2024-12-13 12:38:59 +01:00
Ahmed Harmouche	04041b1342	More WIP	2024-12-13 12:23:02 +01:00
Ahmed Harmouche	3d4bbc3a35	WIP	2024-12-13 11:54:14 +01:00
Ahmed Harmouche	52f4547303	WIP	2024-12-13 11:54:14 +01:00