clean

more
chain
2025-12-21 21:39:17 -04:00 · 2025-12-21 15:48:59 -04:00 · 2025-12-21 12:52:20 -04:00 · 2025-12-21 12:34:00 -04:00
708 changed files with 147197 additions and 229910 deletions
@@ -11,5 +11,5 @@ runs:
        git fetch origin $CURRENT_SHA
        export COMMIT_MESSAGE=$(git show -s --format=%B "$CURRENT_SHA")
        export CURRENT_HEAD=$(git rev-parse HEAD)
-        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && CHECK_OOB=0 PYTHONPATH=. python3 process_replay.py
+        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && IGNORE_OOB=1 PYTHONPATH=. python3 process_replay.py
        git checkout $CURRENT_HEAD  # restore to branch
@@ -45,10 +45,6 @@ inputs:
    description: "Install mesa"
    required: false
    default: 'false'
-  tinydreno:
-    description: "Install tinydreno"
-    required: false
-    default: 'false'
 runs:
  using: "composite"
  steps:
@@ -60,40 +56,32 @@ runs:

    # **** Caching packages ****

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

    # **** Caching downloads ****

-    - name: Cache downloads (PR)
-      if: inputs.key != '' && github.event_name == 'pull_request'
-      uses: actions/cache/restore@v4
-      with:
-        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
-    - name: Cache downloads
-      if: inputs.key != '' && github.event_name != 'pull_request'
+    - name: Cache downloads (Linux)
+      if: inputs.key != '' && runner.os == 'Linux'
      uses: actions/cache@v4
      with:
-        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
+        path: ~/.cache/tinygrad/downloads/
+        key: downloads-${{ 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-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps != '' && steps.restore-venv.outputs.cache-hit != 'true'
      shell: bash
      run: |
        python -m venv .venv
@@ -104,7 +92,7 @@ runs:
        fi
        python -m pip install -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --extra-index-url https://download.pytorch.org/whl/cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
    - name: Install dependencies in venv (without extra)
-      if: inputs.deps == '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
+      if: inputs.deps == '' && steps.restore-venv.outputs.cache-hit != 'true'
      shell: bash
      run: |
        python -m venv .venv
@@ -149,7 +137,7 @@ runs:
      run: |
        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
        sudo tee /etc/apt/sources.list.d/rocm.list <<EOF
-        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.1 $(lsb_release -cs) main
+        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.2 $(lsb_release -cs) main
        EOF
        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600

@@ -194,18 +182,12 @@ runs:
        echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
        echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"

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

    - name: Run apt Update + Install
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -237,7 +219,7 @@ runs:
      shell: bash
      run: |
        sudo mkdir -p /usr/local/lib
-        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/tinygrad/amdcomgr_dylib/releases/latest | \
+        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/amdcomgr_dylib/releases/latest | \
          jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
          sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib
        cargo build --release --manifest-path ./extra/remu/Cargo.toml
@@ -257,17 +239,8 @@ runs:
        ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
        ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
        ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
-    - name: Cache gpuocelot (PR)
-      if: inputs.ocelot == 'true' && github.event_name == 'pull_request'
-      id: cache-build-pr
-      uses: actions/cache/restore@v4
-      env:
-        cache-name: cache-gpuocelot-build-1
-      with:
-        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
    - name: Cache gpuocelot
-      if: inputs.ocelot == 'true' && github.event_name != 'pull_request'
+      if: inputs.ocelot == 'true'
      id: cache-build
      uses: actions/cache@v4
      env:
@@ -276,7 +249,7 @@ runs:
        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-pr.outputs.cache-hit != 'true' && steps.cache-build.outputs.cache-hit != 'true'
+      if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
      shell: bash
      run: |
        git clone --recurse-submodules https://github.com/gpuocelot/gpuocelot.git ${{ github.workspace }}/gpuocelot
@@ -287,7 +260,6 @@ runs:

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

@@ -331,9 +303,3 @@ runs:
      if: inputs.mesa == 'true' && runner.os == 'macOS'
      shell: bash
      run: brew install sirhcm/tinymesa/tinymesa_cpu
-
-    # *** tinydreno ***
-    - name: Install tinydreno (linux)
-      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
@@ -14,12 +14,10 @@ on:
    paths:
    - 'tinygrad/runtime/autogen/**/*'
    - 'tinygrad/runtime/support/autogen.py'
-    - '.github/workflows/autogen.yml'
  workflow_dispatch:
    paths:
    - 'tinygrad/runtime/autogen/**/*'
    - 'tinygrad/runtime/support/autogen.py'
-    - '.github/workflows/autogen.yml'

 jobs:
  autogen:
@@ -32,7 +30,6 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: 'autogen'
        opencl: 'true'
        amd: 'true'
        cuda: 'true'
@@ -41,38 +38,103 @@ jobs:
        mesa: 'true'
        pydeps: 'pyyaml mako'
    - name: Install autogen support packages
-      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev
-    - name: Regenerate autogen files
+      run: 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: |
-        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "comgr.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
+        mv tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
        python3 -c "from tinygrad.runtime.autogen import opencl"
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv"
-        python3 -c "from tinygrad.runtime.autogen import comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
-        python3 -c "from tinygrad.runtime.autogen.am import am, pm4_soc15, pm4_nv, sdma_4_0_0, sdma_5_0_0, sdma_6_0_0, smu_v13_0_0, smu_v13_0_6, smu_v13_0_12, smu_v14_0_2"
-        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
-        python3 -c "from tinygrad.runtime.autogen import llvm"
-        python3 -c "from tinygrad.runtime.autogen import webgpu"
-        python3 -c "from tinygrad.runtime.autogen import kgsl, qcom_dsp"
-        python3 -c "from tinygrad.runtime.autogen import libusb"
-        python3 -c "from tinygrad.runtime.autogen import mesa"
-        python3 -c "from tinygrad.runtime.autogen import avcodec"
-        python3 -c "from tinygrad.runtime.autogen import llvm_qcom"
-        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
-    - name: Check for differences
+        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
+    - name: Verify CUDA autogen
      run: |
-        if ! git diff --quiet; then
-          git diff
-          git diff > autogen-ubuntu.patch
-          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
-          exit 1
-        fi
-    - name: Upload patch artifact
-      if: failure()
-      uses: actions/upload-artifact@v4
-      with:
-        name: autogen-ubuntu-patch
-        path: autogen-ubuntu.patch
-
+        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
@@ -83,29 +145,14 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: 'autogen-mac'
        llvm: 'true'
-    - name: Regenerate autogen files
+    - name: Verify macos autogen
      run: |
-        rm tinygrad/runtime/autogen/metal.py tinygrad/runtime/autogen/iokit.py tinygrad/runtime/autogen/corefoundation.py
-        python3 -c "from tinygrad.runtime.autogen import metal, iokit, corefoundation"
-    - name: Check for differences
-      run: |
-        if ! git diff --quiet; then
-          git diff
-          git diff > autogen-macos.patch
-          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
-          exit 1
-        fi
-    - name: Upload patch artifact
-      if: failure()
-      uses: actions/upload-artifact@v4
-      with:
-        name: autogen-macos-patch
-        path: autogen-macos.patch
-
-  autogen-comgr-2:
-    name: In-tree Autogen (comgr 2)
+        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:
@@ -113,32 +160,17 @@ jobs:
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: 'autogen-comgr'
    - 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/6.2 $(lsb_release -cs) main
+        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: Regenerate autogen files
+    - name: Verify comgr (3) autogen
      run: |
-        rm tinygrad/runtime/autogen/comgr.py
-        python3 -c "from tinygrad.runtime.autogen import comgr"
-    - name: Check for differences
-      run: |
-        if ! git diff --quiet; then
-          git diff
-          git diff > autogen-comgr2.patch
-          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
-          exit 1
-        fi
-    - name: Upload patch artifact
-      if: failure()
-      uses: actions/upload-artifact@v4
-      with:
-        name: autogen-comgr2-patch
-        path: autogen-comgr2.patch
+        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
@@ -16,41 +16,6 @@ on:
  workflow_dispatch:

 jobs:
-  # the goal of this test is to replicate a normal person on a laptop running the test
-  # no process replay, no benchmarks, no CI, just a normal laptop person
-  # the 3 minute timeout should not be raised
-  testmacpytest:
-    name: Mac pytest
-    env:
-      CI: ""
-      CAPTURE_PROCESS_REPLAY: "0"
-    runs-on: [self-hosted, macOS]
-    timeout-minutes: 3
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    # brew install uv
-    - name: setup python environment
-      run: |
-        rm -rf /tmp/tinygrad_pytest_ci
-        uv venv /tmp/tinygrad_pytest_ci
-        source /tmp/tinygrad_pytest_ci/bin/activate
-        uv pip install .[testing]
-    - name: setup staging db
-      run: |
-        echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
-        rm -f /tmp/pytest-db-ci*
-    - name: Run pytest -nauto
-      run: |
-        source /tmp/tinygrad_pytest_ci/bin/activate
-        pytest -nauto --durations=20
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: FLOAT16=1 CL=1 IMAGE=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-
  testmacbenchmark:
    name: Mac Benchmark
    env:
@@ -84,19 +49,19 @@ jobs:
    - name: Print macOS version
      run: sw_vers
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
    - name: Run Stable Diffusion v2
      # TODO: very slow step time
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=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
    - 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
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run model inference benchmark
      run: METAL=1 NOCLANG=1 python3.11 test/external/external_model_benchmark.py
    - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
+      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test tensor cores
      run: METAL=1 python3.11 test/opt/test_tensor_cores.py
    - name: Test AMX tensor cores
@@ -106,59 +71,84 @@ jobs:
        DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
        DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
    - name: Run Tensor Core GEMM (float)
-      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
    - name: Run Tensor Core GEMM (half)
-      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_half.txt
    - name: Run Tensor Core GEMM (bfloat16)
-      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
+      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
    - name: Fuzz Padded Tensor Core GEMM
      run: METAL=1 M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
    - name: Run LLaMA
      run: |
-        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    - name: Run quantized LLaMA
      run: |
-        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8
-        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4
+        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
+        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_nf4.txt
    - name: Run quantized LLaMA3
      run: |
-        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8
-        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4
+        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8 | tee llama3_int8.txt
+        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4 | tee llama3_nf4.txt
    #- name: Run LLaMA 7B on 4 (virtual) GPUs
-    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - name: Run OLMoE
      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
    - name: Train MNIST
-      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt

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

    #- name: Run 10 CIFAR training steps w BF16
-    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
+    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (Mac)
        path: |
          onnx_inference_speed.csv
+          torch_speed.txt
+          llama_unjitted.txt
+          llama_jitted.txt
+          llama_beam.txt
+          llama_int8.txt
+          llama_nf4.txt
+          llama3_int8.txt
+          llama3_nf4.txt
+          llama_four_gpu.txt
+          gpt2_unjitted.txt
+          gpt2_jitted.txt
+          gpt2_half.txt
+          gpt2_half_beam.txt
+          matmul.txt
+          matmul_half.txt
+          matmul_bfloat16.txt
+          sd.txt
+          sd_no_fp16.txt
+          sdv2.txt
+          sdxl.txt
+          beautiful_mnist.txt
+          train_cifar.txt
+          train_cifar_half.txt
+          train_cifar_bf16.txt
+          train_cifar_wino.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3.11 process_replay.py

@@ -180,18 +170,14 @@ jobs:
      run: |
        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: Kill stale pids
-      run: |
-        PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-        PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
    - name: UsbGPU boot time
-      run: sudo -E PYTHONPATH=. GMMU=0 DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
+      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=. GMMU=0 AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
+      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
    - name: UsbGPU copy speeds
-      run: sudo -E PYTHONPATH=. GMMU=0 AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+      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=. GMMU=0 AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+    #  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
@@ -229,7 +215,7 @@ jobs:
    - name: Run model inference benchmark
      run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
    - name: Test speed vs torch
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test speed vs theoretical
      run: NV=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
    - name: Test benchmark allreduce
@@ -240,58 +226,79 @@ jobs:
        NV=1 NV_PTX=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
    - name: Run Tensor Core GEMM (CUDA)
      run: |
-        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
+        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
+        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
+        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_fp8.txt
    - name: Run Tensor Core GEMM (PTX)
-      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
    - name: Run Tensor Core GEMM (NV)
-      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_nv.txt
    - name: Test NV=1
      run: DEBUG=2 NV=1 python -m pytest -rA test/test_tiny.py
    - name: Test CUDA=1
      run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    # TODO: too slow
    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA
      run: |
-        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      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
+      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
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
    # - name: Run LLaMA-2 70B
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA)
        path: |
          onnx_inference_speed.csv
+          torch_speed.txt
+          matmul.txt
+          matmul_bfloat16.txt
+          matmul_tf32.txt
+          matmul_ptx.txt
+          matmul_nv.txt
+          sd.txt
+          sdxl.txt
+          llama_unjitted.txt
+          llama_jitted.txt
+          llama_beam.txt
+          llama3_beam.txt
+          llama3_four_gpu.txt
+          llama3_six_gpu.txt
+          llama3_fp8.txt
+          llama_2_70B.txt
+          mixtral.txt
+          gpt2_unjitted.txt
+          gpt2_jitted.txt
+          gpt2_half.txt
+          gpt2_half_beam.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -330,30 +337,44 @@ jobs:
    # - name: Fuzz Padded Tensor Core GEMM (PTX)
    #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    - name: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+      run: VALIDATE=1 MAX_FRAMES=100 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
    - name: Train MNIST
-      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=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 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    - name: Run MLPerf resnet eval on training data
      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=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 (NVIDIA Training)
+        path: |
+          beautiful_mnist.txt
+          train_cifar.txt
+          train_cifar_half.txt
+          train_cifar_bf16.txt
+          train_cifar_wino.txt
+          train_cifar_one_gpu.txt
+          train_cifar_six_gpu.txt
+          train_resnet.txt
+          train_resnet_one_gpu.txt
+          train_bert.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -368,12 +389,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
    #- name: Insert amdgpu
    #  run: sudo modprobe amdgpu
    - name: Symlink models and datasets
@@ -407,18 +426,16 @@ jobs:
    #- name: Test speed vs torch
    #  run: |
    #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test speed vs theoretical
      run: AMD=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test tensor cores AMD_LLVM=0
-      run: AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
-    # TODO: this is flaky
-    # - name: Test tensor cores AMD_LLVM=1
-    #   run: AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (AMD)
+    - name: Test tensor cores
      run: |
+        AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
+        AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
        AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+    - name: Run Tensor Core GEMM (AMD)
+      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
    - name: Test AMD=1
      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
    #- name: Test HIP=1
@@ -433,39 +450,61 @@ jobs:
    - name: Test AM warm start time
      run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+      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
    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA 7B
      run: |
-        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA 7B with BEAM
-      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
    #- name: Restore amdgpu
    #  run: sudo modprobe amdgpu
    # - name: Run LLaMA-2 70B
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AMD)
+        path: |
+          onnx_inference_speed.csv
+          torch_speed.txt
+          llama_unjitted.txt
+          llama_jitted.txt
+          llama_beam.txt
+          llama3_beam.txt
+          llama3_four_gpu.txt
+          llama3_six_gpu.txt
+          llama_2_70B.txt
+          gpt2_unjitted.txt
+          gpt2_jitted.txt
+          gpt2_half.txt
+          gpt2_half_beam.txt
+          matmul.txt
+          matmul_amd.txt
+          sd.txt
+          sdxl.txt
+          mixtral.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -480,12 +519,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -504,23 +541,31 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Train MNIST
-      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=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
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # TODO: broken on some of the machines
-    #- name: Test full tinyfs load
-    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
+      run: time BENCHMARK_LOG=cifar_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
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AMD Training)
+        path: |
+          beautiful_mnist.txt
+          train_cifar.txt
+          train_cifar_half.txt
+          train_cifar_bf16.txt
+          train_cifar_wino.txt
+          train_cifar_one_gpu.txt
+          train_cifar_six_gpu.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -535,12 +580,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -561,12 +604,19 @@ jobs:
    - 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
+      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
+      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=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
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -588,22 +638,20 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: IR3 openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM QCOM_IR3=1 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.11.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.11.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_11_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: 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=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
+      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=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
+      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=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
+      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
@@ -615,27 +663,6 @@ 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 process_replay.py

-  testcommausbgpubenchmark:
-    name: UsbGPU Benchmark (comma)
-    runs-on: [self-hosted, Linux, comma4]
-    timeout-minutes: 20
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." GMMU=0 DEV=AMD AMD_LLVM=1 AMD_IFACE=USB ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot load_pickle 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_load_pickle PYTHONPATH="." GMMU=0 DEV=AMD AMD_IFACE=USB ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
-
  testreddriverbenchmark:
    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
@@ -647,12 +674,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
+      run: ./extra/hcq/hcq_smi.py amd rmmod
    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+      run: ./extra/hcq/hcq_smi.py amd kill_pids
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -681,7 +706,7 @@ jobs:
    #     AMD=1 AMD_LLVM=1 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
    #     AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (AMD)
-      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee am_matmul_amd.txt
    - name: Test AMD=1
      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
    - name: Test DISK copy time
@@ -691,20 +716,20 @@ jobs:
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6482 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD=1 AMD_IFACE=PCI python3 test/test_tiny.py
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD=1 AMD_AQL=1 AMD_IFACE=PCI python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
+      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AM Driver)
+        path: |
+          am_matmul_amd.txt
+          am_train_cifar_one_gpu.txt
+          am_train_resnet_one_gpu.txt
+          am_train_bert_one_gpu.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -719,12 +744,10 @@ jobs:
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove nv modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
+      run: ./extra/hcq/hcq_smi.py nv rmmod
    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
+      run: ./extra/hcq/hcq_smi.py nv kill_pids
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -753,20 +776,21 @@ jobs:
        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Test LLAMA-3
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      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
+      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6483 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6483 NV=1 python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
+      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (NV Driver)
+        path: |
+          nv_llama3_beam.txt
+          nv_train_cifar_one_gpu.txt
+          nv_train_resnet_one_gpu.txt
+          nv_train_bert_one_gpu.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
@@ -1,11 +1,10 @@
 name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '18'
+  CACHE_VERSION: '15'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
-  CHECK_OOB: 1

 on:
  push:
@@ -26,19 +25,17 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: llvm-speed
-        deps: testing_unit
+        deps: testing_minimal
        llvm: 'true'
    - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py

  docs:
    name: Docs
    runs-on: ubuntu-22.04
    timeout-minutes: 10
-    env:
-      CHECK_OOB: 0
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -70,7 +67,7 @@ jobs:
        source venv/bin/activate
        pip install $GITHUB_WORKSPACE
        cp $GITHUB_WORKSPACE/examples/beautiful_mnist.py .
-        BS=2 STEPS=10 MAX_BUFFER_SIZE=0 python beautiful_mnist.py
+        BS=2 STEPS=10 python beautiful_mnist.py
    - name: Test Docs Build
      run: python -m mkdocs build --strict
    - name: Test Docs
@@ -98,23 +95,27 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
+        deps: testing_minimal
        pydeps: "pillow torchvision expecttest"
        llvm: 'true'
    - name: Install ninja
      run: |
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
+    - name: Lint with ruff
+      run: |
+        pip3 install --upgrade --force-reinstall ruff==0.11.0
+        python3 -m ruff check extra/torch_backend/backend.py
    - name: Test one op
-      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
    - name: Test ResNet-18
      run: DEBUG=2 python3 extra/torch_backend/example.py
-    - name: custom tests
-      run: python3 -m pytest -n auto extra/torch_backend/test.py --durations=20
+    - name: My (custom) tests
+      run: python3 extra/torch_backend/test.py
    - name: Test one op in torch tests
      run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
    - name: Test Ops with TINY_BACKEND
-      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/backend/test_ops.py --durations=20
+      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
    - name: Test in-place operations on views
      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
    - name: Test multi-gpu
@@ -134,14 +135,14 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
+        deps: testing_minimal
        llvm: 'true'
    - name: Install ninja
      run: |
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
    - name: Test some torch tests (expect failure)
      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true

@@ -156,27 +157,27 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: be-minimal
-        deps: testing_unit
+        deps: testing_minimal
    - name: Test dtype with Python emulator
-      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py
+      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/test_dtype.py test/test_dtype_alu.py
    - name: Test ops with Python emulator
-      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_ops.py --durations=20
+      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/test_ops.py --durations=20
    - name: Test uops with Python emulator
-      run: PYTHON=1 python3 -m pytest test/backend/test_uops.py --durations=20
+      run: PYTHON=1 python3 -m pytest test/test_uops.py --durations=20
    - name: Test symbolic with Python emulator
-      run: PYTHON=1 python3 test/backend/test_symbolic_ops.py
+      run: PYTHON=1 python3 test/test_symbolic_ops.py
    - name: test_renderer_failures with Python emulator
-      run: PYTHON=1 python3 -m pytest -rA test/backend/test_renderer_failures.py::TestRendererFailures
-    - name: Test IMAGE support
+      run: PYTHON=1 python3 -m pytest -rA test/test_renderer_failures.py::TestRendererFailures
+    - name: Test IMAGE=2 support
      run: |
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_simple_conv2d
+        IMAGE=2 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
+        IMAGE=2 PYTHON=1 python3 test/test_ops.py TestOps.test_simple_conv2d
    - name: Test emulated METAL tensor cores
      run: |
-        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_big_gemm
+        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_big_gemm
        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test emulated AMX tensor cores
-      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
+      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
    - name: Test emulated AMD tensor cores
      run: |
        DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
@@ -197,9 +198,9 @@ jobs:
        DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test emulated CUDA tensor cores
      run: |
-        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
+        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
        DEBUG=2 EMULATE=CUDA_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test emulated INTEL OpenCL tensor cores
      run: DEBUG=2 EMULATE=INTEL FORWARD_ONLY=1 PYTHON=1 HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
@@ -207,17 +208,18 @@ jobs:
      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test device flop counts
      run: |
-        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStats.test_simple_matmul
+        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStats.test_simple_matmul

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

+    # TODO: run the pre-commit hook to replace a lot of this
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -229,51 +231,19 @@ jobs:
        deps: linting
    - name: Lint bad-indentation and trailing-whitespace with pylint
      run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
-    - name: Run pre-commit linting hooks
-      run: SKIP=tiny,tests,example pre-commit run --all-files
-    - name: Lint additional files with ruff
+    - name: Lint with ruff
      run: |
+        pip3 install --upgrade --force-reinstall ruff==0.11.0
+        python3 -m ruff check .
        python3 -m ruff check examples/mlperf/ --ignore E501
        python3 -m ruff check extra/thunder/tiny/ --ignore E501 --ignore F841 --ignore E722
-        python3 -m ruff check extra/torch_backend/backend.py
-    - name: Run mypy with lineprecision report
+    - name: Run mypy
      run: |
-        python -m mypy --lineprecision-report .
-        grep -v autogen lineprecision.txt | awk 'NR>2 {lines+=$2; precise+=$3; imprecise+=$4; any+=$5; empty+=$6} END {t=lines-empty; printf "TOTAL: %d lines, %d precise (%.1f%%), %d imprecise (%.1f%%), %d any (%.1f%%)\n", t, precise, 100*precise/t, imprecise, 100*imprecise/t, any, 100*any/t}'
+        python -m mypy --strict-equality --lineprecision-report .
        cat lineprecision.txt
-    - name: Run TYPED=1
-      run: CHECK_OOB=0 DEV=CPU TYPED=1 python test/test_tiny.py
-
-  nulltest:
-    name: Null Tests
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
-        deps: testing_unit
-        llvm: 'true'
-        amd: 'true'
-    - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
-    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
-    # TODO: too slow
-    # - name: Run SDXL on NULL backend
-    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
-    - name: Run Clip tests for SD MLPerf on NULL backend
-      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
-    - name: Run AMD emulated BERT training on NULL backend
-      run: EMULATE=AMD_RDNA4 NULL=1 NULL_ALLOW_COPYOUT=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    # TODO: support fake weights
-    #- name: Run LLaMA 7B on 4 fake devices
-    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
+    # broken because of UPatAny
+    #- name: Run TYPED=1
+    #  run: TYPED=1 python -c "import tinygrad"

  unittest:
    name: Unit Tests
@@ -286,19 +256,27 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
+        key: unittest-12
+        pydeps: "pillow numpy ftfy regex"
        deps: testing_unit
-        llvm: 'true'
-        amd: 'true'
-    - name: Run pre-commit test hooks
-      run: SKIP=ruff,mypy pre-commit run --all-files
    - name: Check Device.DEFAULT
      run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
    - name: Run unit tests
      run: |
-        CPU=1 python test/null/test_device.py TestRunAsModule.test_module_runs
-        CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+        CPU=1 python test/unit/test_device.py TestRunAsModule.test_module_runs
+        CPU=1 python -m pytest -n=auto test/unit/ --durations=20 --deselect=test/unit/test_device.py::TestRunAsModule::test_module_runs
+    - name: Run targetted tests on NULL backend
+      run: NULL=1 python3 -m unittest test.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
+    # TODO: too slow
+    # - name: Run SDXL on NULL backend
+    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+    - name: Run Clip tests for SD MLPerf on NULL backend
+      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
+    - name: Run AMD emulated BERT training on NULL backend
+      run: EMULATE=AMD_RDNA4 NULL=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    # TODO: support fake weights
+    #- name: Run LLaMA 7B on 4 fake devices
+    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
    - name: Run GC tests
      run: python test/external/external_uop_gc.py
    - name: External Benchmark Schedule
@@ -312,8 +290,8 @@ jobs:
        python extra/optimization/extract_dataset.py
        gzip -c /tmp/sops > extra/datasets/sops.gz
        #DEBUG=1 MIN_ASTS=1 python extra/optimization/get_action_space.py
-    - name: Repo line count < 24000 lines
-      run: MAX_LINE_COUNT=24000 python sz.py
+    - name: Repo line count < 20000 lines
+      run: MAX_LINE_COUNT=20000 python sz.py

  spec:
    strategy:
@@ -333,7 +311,7 @@ jobs:
        deps: testing_unit
        python-version: '3.14'
    - name: Test SPEC=2
-      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}

  fuzzing:
    name: Fuzzing
@@ -367,13 +345,13 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: gpu-image
-          deps: testing_unit
+          deps: testing_minimal
          opencl: 'true'
-      - name: Test CL IMAGE=1 ops
+      - name: Test CL IMAGE=2 ops
        run: |
-          CL=1 IMAGE=1 python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          CL=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
        # TODO: training is broken
-        # CL=1 IMAGE=1 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
+        # CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -388,14 +366,14 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: gen-dataset
-          deps: testing
+          deps: testing_minimal
          opencl: 'true'
      - name: Generate Dataset
        run: CL=1 extra/optimization/generate_dataset.sh
      - name: Run Kernel Count Test
        run: CL=1 python -m pytest -n=auto test/external/external_test_opt.py
      - name: Run fused optimizer tests
-        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/backend/test_optim.py -k "not muon"
+        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/test_optim.py -k "not muon"
      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
@@ -418,13 +396,13 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1486 ALLOWED_GATED_READ_IMAGE=17 FLOAT16=1 CL=1 IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1397 ALLOWED_GATED_READ_IMAGE=94 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp16
-        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp32 (test correctness)
-        run: CL=1 IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
+        run: DEBUGCL=1 CL=1 IMAGE=2 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
      - name: Test openpilot LLVM compile fp16
-        run: IMAGE=1 FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+        run: FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -443,7 +421,7 @@ jobs:
        with:
          key: onnxoptc
          deps: testing
-          python-version: '3.12'
+          python-version: '3.11'
          llvm: 'true'
      - name: Test ONNX (CPU)
        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -454,7 +432,7 @@ jobs:
      - name: Test Additional ONNX Ops (CPU)
        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_ops.py
      - name: Test Quantize ONNX
-        run: CPU=1 CPU_LLVM=0 python3 test/backend/test_quantize_onnx.py
+        run: CPU=1 CPU_LLVM=0 python3 test/test_quantize_onnx.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -471,7 +449,7 @@ jobs:
          key: onnxoptl
          deps: testing
          pydeps: "tensorflow==2.19"
-          python-version: '3.12'
+          python-version: '3.11'
          opencl: 'true'
      - name: Test ONNX (CL)
        run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
@@ -484,11 +462,11 @@ jobs:
      - name: Test MLPerf stuff
        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
      - name: NULL=1 beautiful_mnist_multigpu
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
+        run: NULL=1 python examples/beautiful_mnist_multigpu.py
      - name: Test Bert training
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -496,8 +474,6 @@ jobs:
    name: Test LLM
    runs-on: ubuntu-24.04
    timeout-minutes: 15
-    env:
-      CHECK_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -505,13 +481,8 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: apps_llm
-      - name: Test 1B LLM (llama)
-        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model llama3.2:1b | tee /dev/stderr | grep -i rooster
-      - name: Test 1B LLM (llama q4)
-        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model llama3.2:1b-q4 | tee /dev/stderr | grep -i rooster
-      - name: Test 1B LLM (qwen)
-        # NOTE: qwen is dumb and only knows about female chickens
-        run: echo "What's a female chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm --model qwen3:0.6b | tee /dev/stderr | grep -i hen
+      - name: Test 1B LLM
+        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm | grep -i rooster

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

@@ -550,7 +521,7 @@ jobs:
      with:
        key: metal
        deps: testing
-        python-version: '3.12'
+        python-version: '3.11'
    - name: Test models (Metal)
      run: METAL=1 python -m pytest -n=auto test/models --durations=20
    - name: Test LLaMA compile speed
@@ -569,15 +540,15 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: devectorize-minimal
-        deps: testing_unit
+        deps: testing_minimal
        pydeps: "pillow"
        llvm: "true"
    - name: Test LLVM=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
+      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
    - name: Test LLVM=1 DEVECTORIZE=0 for model
      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
    - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
+      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py

  testdsp:
    name: Linux (DSP)
@@ -590,8 +561,8 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: dsp-minimal
-        deps: testing_unit
-        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
+        deps: testing_minimal
+        pydeps: "onnx==1.18.0 onnxruntime pillow"
        llvm: "true"
    - name: Set up Docker Buildx
      uses: docker/setup-buildx-action@v3
@@ -603,15 +574,15 @@ jobs:
        load: true
        tags: qemu-hexagon:latest
        cache-from: type=gha
-        cache-to: ${{ github.event_name != 'pull_request' && 'type=gha,mode=min' || '' }}
+        cache-to: type=gha,mode=min
    - name: Set MOCKDSP env
      run: printf "MOCKDSP=1" >> $GITHUB_ENV
    - name: Run test_tiny on DSP
      run: DEBUG=2 DSP=1 python test/test_tiny.py
    - name: Test transcendentals
-      run: CC=clang-20 DEBUG=2 DSP=1 python test/backend/test_transcendental.py TestTranscendentalVectorized
+      run: CC=clang-20 DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
    - name: Test quantize onnx
-      run: DEBUG=2 DSP=1 python3 test/backend/test_quantize_onnx.py
+      run: DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py

  testwebgpu:
    name: Linux (WebGPU)
@@ -624,105 +595,34 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: webgpu-minimal
-        deps: testing_unit
-        python-version: '3.12'
+        deps: testing_minimal
+        python-version: '3.11'
        webgpu: 'true'
    - name: Check Device.DEFAULT (WEBGPU) and print some source
      run: |
        WEBGPU=1 python -c "from tinygrad import Device; assert Device.DEFAULT == 'WEBGPU', Device.DEFAULT"
-        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
    - name: Run selected webgpu tests
      run: |
-          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/backend --durations=20
+          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit \
+          --ignore=test/test_copy_speed.py --ignore=test/test_rearrange_einops.py \
+          --ignore=test/test_fuzz_shape_ops.py --durations=20
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

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

-    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
+    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
-    timeout-minutes: 15
+    timeout-minutes: 20
    env:
      AMD: 1
      MOCKGPU: 1
-      MOCKGPU_ARCH: ${{ matrix.arch }}
-      SKIP_SLOW_TEST: 1
+      FORWARD_ONLY: 1
      AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
    steps:
      - name: Checkout Code
@@ -731,17 +631,27 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
          amd: 'true'
          llvm: ${{ matrix.backend == 'amdllvm' && 'true' }}
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
+      - name: Run LLVM test
+        if: matrix.backend=='amdllvm'
+        run: python test/device/test_amd_llvm.py
      - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
+        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py --durations=20
+      - name: Run pytest (amd)
+        run: python -m pytest test/external/external_test_am.py --durations=20
      - name: Run TRANSCENDENTAL math
-        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
+        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
+      - name: Run TestOps.test_add with SQTT
+        run: |
+          VIZ=1 PMC=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -764,7 +674,7 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
          cuda: 'true'
          ocelot: 'true'
      - name: Set env
@@ -772,12 +682,10 @@ jobs:
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CUDA','NV'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (cuda)
        # skip multitensor because it's slow
-        run: python -m pytest -n=auto test/backend --ignore test/backend/test_multitensor.py --durations=20
-      - name: Run TestOps.test_add with PMA
-        run: VIZ=-1 PMA=1 DEBUG=5 python3 test/backend/test_ops.py TestOps.test_add
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -797,7 +705,7 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
          opencl: ${{ matrix.backend == 'opencl' && 'true' }}
          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
@@ -806,11 +714,11 @@ jobs:
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (${{ matrix.backend }})
-        run: python -m pytest -n=auto test/backend --durations=20
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
      - name: Run TRANSCENDENTAL math
-        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
+        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -828,29 +736,25 @@ jobs:
      with:
        key: metal
        deps: testing
-        python-version: '3.12'
+        python-version: '3.11'
        amd: 'true'
        cuda: 'true'
        ocelot: 'true'
        llvm: 'true'
    - name: Run unit tests
      run: METAL=1 python -m pytest -n=auto test/unit/ --durations=20
-    - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
    - name: Run ONNX
      run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
    - name: Test tensor core ops (fake)
-      run: METAL=1 DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
+      run: METAL=1 DEBUG=3 TC=2 python test/test_ops.py TestOps.test_gemm
    - name: Test tensor core ops (real)
-      run: METAL=1 DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
+      run: METAL=1 DEBUG=3 python test/test_ops.py TestOps.test_big_gemm
    - name: Test Beam Search
      run: METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
-    - name: Test Device Specific
-      run: METAL=1 python3 -m pytest test/device/test_metal.py
    #- name: Fuzz Test linearizer
    #  run: METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
    - name: Run TRANSCENDENTAL math
-      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
+      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
    - name: Run pytest (amd)
      env:
        MOCKGPU: 1
@@ -873,8 +777,6 @@ jobs:
        NV_PTX: 1
        NV: 1
        FORWARD_ONLY: 1
-        # TODO: failing due to library loading error
-        CAPTURE_PROCESS_REPLAY: 0
      run: |
        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
    - name: Run process replay tests
@@ -893,14 +795,14 @@ jobs:
        key: osx-webgpu
        deps: testing
        webgpu: 'true'
+    - name: Test infinity math in WGSL
+      run: WEBGPU=1 python -m pytest -n=auto test/test_renderer_failures.py::TestWGSLFailures::test_multiply_infinity --durations=20
    - name: Build WEBGPU Efficientnet
      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m examples.compile_efficientnet
-    - name: Run selected webgpu tests
-      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m pytest -n=auto test/backend --durations=20
-    #- name: Clean npm cache
-    #  run: npm cache clean --force
-    #- name: Install Puppeteer
-    #  run: npm install puppeteer
+    - name: Clean npm cache
+      run: npm cache clean --force
+    - name: Install Puppeteer
+      run: npm install puppeteer
    # this is also flaky
    #- name: Run WEBGPU Efficientnet
    #  run: node test/web/test_webgpu.js
@@ -932,7 +834,8 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: macos-${{ matrix.backend }}-minimal
-          deps: testing_unit
+          deps: testing_minimal
+          pydeps: "capstone"
          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
      - name: Set env
@@ -942,7 +845,7 @@ jobs:
          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
      - name: Run pytest (${{ matrix.backend }})
-        run: python3 -m pytest -n=auto test/backend --durations=20
+        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
      - name: Run macOS-specific unit test
@@ -975,16 +878,12 @@ jobs:
      - name: Run unit tests
        if: matrix.backend=='llvm'
        # test_newton_schulz hits RecursionError
-        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
-      - name: Run NULL backend tests
-        if: matrix.backend=='llvm'
-        shell: bash
-        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
+        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_elf.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
      - name: Run pytest (${{ matrix.backend }})
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
-          python -m pytest -n=auto test/test_tiny.py test/backend/test_ops.py --durations=20
+          python -m pytest -n=auto test/test_tiny.py test/test_ops.py --durations=20

 # ****** Compile-only Tests ******

@@ -1003,38 +902,15 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: compile-${{ matrix.backend }}
-          deps: testing_unit
+          deps: testing_minimal
          mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
-          python-version: '3.12'
+          python-version: '3.14'
      - name: Set env
        shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
+        run: printf "NULL=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
      - name: Run test_ops
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-  qcomclcompiletests:
-    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-qcomcl
-          deps: testing_unit
-          tinydreno: 'true'
-          python-version: '3.12'
-      - name: Set env
-        shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\nNULL_QCOMCL=1" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          DEBUG=4 python3 test/test_ops.py TestOps.test_add
+          python -m pytest -n=auto test/test_ops.py --durations=20
@@ -58,13 +58,10 @@ weights
 *.lprof
 comgr_*
 *.pkl
-!extra/sqtt/examples/**/*.pkl
 site/
 profile_stats
 *.log
 target
 .mypy_cache
 mutants
-.mutmut-cache
-dagre/
-graphlib/
+.mutmut-cache
@@ -16,7 +16,7 @@ repos:
        pass_filenames: false
      - id: mypy
        name: mypy
-        entry: python3 -m mypy
+        entry: python3 -m mypy tinygrad/ --strict-equality
        language: system
        always_run: true
        pass_filenames: false
@@ -28,7 +28,7 @@ repos:
        pass_filenames: false
      - id: tests
        name: comprehensive test suite
-        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/backend/test_ops.py test/backend/test_schedule.py test/unit/test_assign.py test/backend/test_tensor.py test/backend/test_jit.py test/unit/test_schedule_cache.py test/null/test_pattern_matcher.py test/null/test_uop_symbolic.py test/unit/test_helpers.py
+        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
        language: system
        always_run: true
        pass_filenames: false
@@ -0,0 +1,17 @@
+# tinygrad agents
+
+Hello agent. You are one of the most talented programmers of your generation.
+
+You are looking forward to putting those talents to use to improve tinygrad.
+
+## philosophy
+
+tinygrad is a **tensor** library focused on beauty and minimalism, while still matching the functionality of PyTorch and JAX.
+
+Every line must earn its keep. Prefer readability over cleverness. We believe that if carefully designed, 10 lines can have the impact of 1000.
+
+Never mix functionality changes with whitespace changes. All functionality changes must be tested.
+
+## style
+
+Use **2-space indentation**, and keep lines to a maximum of **150 characters**. Match the existing style.
@@ -0,0 +1,227 @@
+# 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.
@@ -192,7 +192,7 @@ For more examples on how to run the full test suite please refer to the [CI work
 Some examples of running tests locally:
 ```sh
 python3 -m pip install -e '.[testing]'  # install extra deps for testing
-python3 test/backend/test_ops.py        # just the ops tests
+python3 test/test_ops.py                # just the ops tests
 python3 -m pytest test/                 # whole test suite
 ```

@@ -45,9 +45,10 @@ print(f"The schedule contains {len(schedule)} items.")
 for si in schedule: print(str(si)[:80])

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

-for si in tqdm(schedule): si.run()
+from tinygrad.engine.realize import run_schedule
+run_schedule(schedule)

 # *****
 # 5. Print the weight change
@@ -13,7 +13,7 @@ There's also a [doc describing speed](../developer/speed.md)

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

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

 ## Scheduling

@@ -10,7 +10,7 @@ Directories are listed in order of how they are processed.

 Group UOps into kernels.

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

 ## tinygrad/codegen

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

-::: tinygrad.codegen.get_program
+::: tinygrad.codegen.full_rewrite
    options:
        members: false
        show_labels: false
@@ -70,7 +70,7 @@ AMD backend supports several interfaces for communicating with devices:

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

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

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

@@ -61,8 +56,3 @@
 ::: tinygrad.Tensor.sparse_categorical_crossentropy
 ::: tinygrad.Tensor.cross_entropy
 ::: tinygrad.Tensor.nll_loss
-
-## Linear Algebra
-
-::: tinygrad.Tensor.qr
-::: tinygrad.Tensor.svd
@@ -0,0 +1,9 @@
+import globals from "globals";
+import pluginJs from "@eslint/js";
+import pluginHtml from "eslint-plugin-html";
+
+export default [
+  {files: ["**/*.html"], plugins: {html: pluginHtml}, rules:{"max-len": ["error", {"code": 150}]}},
+  {languageOptions: {globals: globals.browser}},
+  pluginJs.configs.recommended,
+];
@@ -1,196 +0,0 @@
-from tinygrad import Tensor, dtypes, Context, getenv, UOp, fetch
-from tinygrad.uop.ops import Ops, PatternMatcher, UPat
-from tinygrad.uop.symbolic import symbolic
-from tinygrad.codegen import Renderer
-from tinygrad.codegen.opt import Opt, OptOps
-
-# ************************* implementation of the problem ************************
-
-def myhash(a: Tensor) -> Tensor:
-  a = (a + 0x7ED55D16) + (a << 12)
-  a = (a ^ 0xC761C23C) ^ (a >> 19)
-  a = (a + 0x165667B1) + (a << 5)
-  a = (a + 0xD3A2646C) ^ (a << 9)
-  a = (a + 0xFD7046C5) + (a << 3)
-  a = (a ^ 0xB55A4F09) ^ (a >> 16)
-  return a
-
-def select_with_where_tree(values: Tensor, relative_idx: Tensor) -> Tensor:
-  n = values.shape[0]
-  if n == 1: return values[0].expand(relative_idx.shape)
-
-  mid = n // 2
-  left = select_with_where_tree(values[:mid], relative_idx)
-  right = select_with_where_tree(values[mid:], relative_idx - mid)
-
-  go_left = relative_idx < mid
-  return go_left.where(left, right)
-
-def tree_traversal(forest: Tensor, val: Tensor, height: int, rounds: int, where_tree_threshold=3) -> Tensor:
-  # All walkers start at idx=0
-  idx = Tensor.zeros(val.shape, device=val.device, dtype=dtypes.uint32)
-
-  for r in range(rounds):
-    level = r % (height + 1)
-    level_start = (1 << level) - 1
-    level_size = 1 << level
-
-    if level == 0:
-      # At root (level 0), all walkers are at idx=0
-      # No gather needed, just broadcast the root value
-      node_val = forest[0].expand(val.shape)
-      idx = idx * 0  # Reset to 0
-    elif level <= where_tree_threshold:
-      # Small level: use where-tree
-      level_values = forest[level_start : level_start + level_size]
-      relative_idx = (idx - level_start)
-      node_val = select_with_where_tree(level_values, relative_idx)
-    else:
-      # Large level: use gather
-      node_val = forest.gather(0, idx)
-
-    val = myhash(val ^ node_val)
-    idx = (idx << 1) + (1 + (val & 1))
-
-    # No wrap check needed! At round 10 (level becomes 0), we reset idx above.
-
-  return val.contiguous(arg=(Opt(OptOps.UPCAST, 0, 8),))
-
-# ************************* renderer for VLIW machine *************************
-
-def loop_unrolling(sink:UOp):
-  rng = [x for x in sink.toposort() if x.op is Ops.RANGE]
-  if len(rng) == 0: return None
-  print(f"unrolling loop with size {rng[0].vmax+1}")
-  unrolled_sinks = [sink.substitute({rng[0]:rng[0].const_like(i)}).src[0] for i in range(rng[0].vmax+1)]
-  return UOp.sink(*unrolled_sinks, arg=sink.arg)
-
-global_addrs = []
-vliw_prepare = PatternMatcher([
-  # loop unrolling (should be a part of tinygrad)
-  (UPat(Ops.SINK, name="sink"), loop_unrolling),
-  # cast is fake
-  (UPat(Ops.CAST, name="c"), lambda c: c.src[0]),
-  # rewrites to hardcode the addresses in memory
-  (UPat(Ops.PARAM, name="dg"), lambda dg: UOp.const(dtypes.uint, global_addrs[dg.arg])),
-  # INDEX is just plus
-  (UPat(Ops.INDEX, name="i"), lambda i: i.src[0]+i.src[1]),
-])+symbolic
-
-class VLIWRenderer(Renderer):
-  has_local = False  # TODO: this should be the default / cleaned up
-  # this says this backend supports MULACC + more. decompositions uses this
-  code_for_op: dict = {Ops.MULACC: None, Ops.ADD: "+", Ops.MUL: "*",
-                       Ops.XOR: "^", Ops.AND: "&", Ops.OR: "|",
-                       Ops.SHL: "<<", Ops.SHR: ">>", Ops.CMPLT: "<"}
-  # this matcher runs while still in graph form
-  pre_matcher = vliw_prepare
-
-  def render(self, uops:list[UOp]):
-
-    # TODO: this is a minimal renderer. for low cycle count, make it good
-    # to get speed, you need to add VLIW packing
-    # to get under 1536 regs, you need to add a register allocator
-    # we left the fun parts to you
-
-    print(f"rendering with {len(uops)} uops")
-    reg, inst = 0, []
-    r: dict[UOp, int] = {}
-    for u in uops:
-      assert u.dtype.count in (1,8), "dtype count must be 1 or 8"
-
-      # dumb register allocator
-      if u.op not in {Ops.STORE, Ops.SINK, Ops.GEP}:
-        r[u] = reg
-        reg += u.dtype.count
-
-      # render UOps to instructions
-      match u.op:
-        case Ops.SINK:
-          inst.append({"flow": [("halt",)]})
-        case Ops.CONST:
-          inst.append({"load": [("const", r[u], u.arg)]})
-        case Ops.GEP:
-          # a GEP is just an alias to a special register in the vector
-          r[u] = r[u.src[0]] + u.arg[0]
-        case Ops.VECTORIZE:
-          if all(s == u.src[0] for s in u.src):
-            # if all sources are the same, we can broadcast
-            inst.append({"valu": [("vbroadcast", r[u], r[u.src[0]])]})
-          else:
-            # this is a copy into a contiguous chunk of registers
-            inst.extend({"flow": [("add_imm", r[u]+i, r[s], 0)]} for i,s in enumerate(u.src) if r[s] != r[u]+i)
-        case Ops.LOAD:
-          op = "vload" if u.dtype.count > 1 else "load"
-          inst.append({"load": [(op, r[u], r[u.src[0]])]})
-        case Ops.STORE:
-          op = "vstore" if u.src[1].dtype.count > 1 else "store"
-          inst.append({"store": [(op, r[u.src[0]], r[u.src[1]])]})
-        case Ops.MULACC:
-          assert u.dtype.count == 8
-          inst.append({"valu": [("multiply_add", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
-        case Ops.WHERE:
-          assert u.dtype.count == 8
-          inst.append({"flow": [("vselect", r[u], r[u.src[0]], r[u.src[1]], r[u.src[2]])]})
-        case _ if u.op in self.code_for_op:
-          cat = "valu" if u.dtype.count > 1 else "alu"
-          inst.append({cat: [(self.code_for_op[u.op], r[u], r[u.src[0]], r[u.src[1]])]})
-        case _:
-          raise NotImplementedError(f"unhandled op {u.op}")
-    return repr(inst)
-
-# ************************* test and render *************************
-
-import sys, types
-PROBLEM_URL = "https://raw.githubusercontent.com/anthropics/original_performance_takehome/refs/heads/main/tests/frozen_problem.py"
-sys.modules["problem"] = problem = types.ModuleType("problem")
-exec(fetch(PROBLEM_URL).read_text(), problem.__dict__)
-
-if __name__ == "__main__":
-  batch_size = getenv("BS", 256)
-  height = 10
-  rounds = getenv("ROUNDS", 16)
-
-  # build problem
-  tree = problem.Tree.generate(height)
-  inp = problem.Input.generate(tree, batch_size, rounds)
-  mem = problem.build_mem_image(tree, inp)
-  global_addrs.extend([mem[6], mem[6], mem[4]])  # output, input, forest
-
-  # *** verify the kernel in tinygrad compared to reference ***
-
-  forest_t = Tensor(tree.values, dtype=dtypes.uint32)
-  val_t = Tensor(inp.values, dtype=dtypes.uint32)
-
-  if getenv("VERIFY", 1):
-    # verify on normal tinygrad device
-    with Context(PCONTIG=2):
-      out = tree_traversal(forest_t, val_t, height, rounds)
-      val_out = out.tolist()
-    problem.reference_kernel(tree, inp)
-    assert val_out == inp.values
-    print("verification passed")
-
-  # *** render to device ***
-
-  from tinygrad.codegen import get_program
-  with Context(PCONTIG=2, DEVECTORIZE=2, SPEC=0):
-    out = tree_traversal(forest_t, val_t, height, rounds)
-    sink = out.schedule()[-1].ast
-    prg = get_program(sink, VLIWRenderer())
-
-  # *** run on Machine and compare ***
-
-  # NOTE: the scratch size needs to be reduced to 1536 when you have a register allocator
-  src = eval(prg.src)
-  max_regs = max(t[1] for instr in src for v in instr.values() for t in v if len(t) > 1) + 8
-  print(f"{max_regs:5d} regs used" + ("" if max_regs <= 1536 else "       <-- WARNING: TOO MANY REGISTERS, MUST BE <= 1536"))
-  machine = problem.Machine(mem, src, problem.DebugInfo(scratch_map={}), n_cores=1, trace=False, scratch_size=max_regs)
-  machine.run()
-  print(f"ran for {machine.cycle:5d} cycles" + ("" if machine.cycle <= 1363 else "  <-- EVEN CLAUDE GOT 1363"))
-
-  # compare to reference
-  ref_mem = mem.copy()
-  for _ in problem.reference_kernel2(ref_mem, {}): pass
-  assert machine.mem[mem[6]:mem[6]+mem[2]] == ref_mem[mem[6]:mem[6]+mem[2]]
-  print("compare passed!")
@@ -1,79 +0,0 @@
-from typing import Optional
-from tinygrad import Tensor
-from tinygrad.dtype import DTypeLike, dtypes
-import math
-
-# rewritten from numpy
-def rfftfreq(n: int, d: float = 1.0, device=None) -> Tensor:
-  val = 1.0 / (n * d)
-  N = n // 2 + 1
-  results = Tensor.arange(N, device=device)
-  return results * val
-
-# just like in librosa
-def fft_frequencies(sr: float, n_fft: int) -> Tensor:
-  return rfftfreq(n=n_fft, d=1.0 / sr)
-
-def hz_to_mel(freq: Tensor) -> Tensor:
-  # linear part
-  f_min = 0.0
-  f_sp = 200.0 / 3
-  mels = (freq - f_min) / f_sp
-
-  # log-scale part
-  min_log_hz = 1000.0  # beginning of log region (Hz)
-  mask = freq >= min_log_hz
-  return mask.where(((min_log_hz - f_min) / f_sp) + (freq / min_log_hz).log() / (math.log(6.4) / 27.0), mels)
-
-def mel_to_hz(mels: Tensor) -> Tensor:
-  # linear scale
-  f_min = 0.0
-  f_sp = 200.0 / 3
-  freqs = f_min + f_sp * mels
-
-  # nonlinear scale
-  min_log_hz = 1000.0  # beginning of log region (Hz)
-  min_log_mel = (min_log_hz - f_min) / f_sp  # same (Mels)
-  logstep = math.log(6.4) / 27.0  # step size for log region
-
-  log_t = mels >= min_log_mel
-  freqs = log_t.where(min_log_hz * ((logstep * (mels - min_log_mel)).exp()), freqs)
-  return freqs
-
-def mel_frequencies(n_mels: int = 128, *, fmin: float = 0.0, fmax: float = 11025.0) -> Tensor:
-  # center freqs of mel bands - uniformly spaced between limits
-  min_max_mel = hz_to_mel(Tensor([fmin, fmax]))
-
-  mels = Tensor.linspace(min_max_mel[0], min_max_mel[1], n_mels)
-  hz = mel_to_hz(mels)
-  return hz
-
-def mel(
-  *,
-  sr: float,
-  n_fft: int,
-  n_mels: int = 128,
-  fmin: float = 0.0,
-  fmax: Optional[float] = None,
-  dtype: DTypeLike = dtypes.default_float,
-) -> Tensor:
-  if fmax is None:
-    fmax = float(sr) / 2
-
-  n_mels = int(n_mels)
-
-  fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft)  # center freqs of each FFT bin
-  mel_f = mel_frequencies(n_mels + 2, fmin=fmin, fmax=fmax)  # center freqs of mel bands
-
-  fdiff = mel_f[1:] - mel_f[:-1]
-  ramps = mel_f[None].T.expand(-1, fftfreqs.shape[-1]) - fftfreqs
-
-  lower = -ramps[:n_mels] / fdiff[:n_mels][None].T
-  upper = ramps[2 : n_mels + 2] / fdiff[1 : n_mels + 1][None].T
-  weights = lower.minimum(upper).maximum(0)
-
-  # Slaney-style mel is scaled to be approx constant energy per channel
-  enorm = 2.0 / (mel_f[2 : n_mels + 2] - mel_f[:n_mels])
-  weights *= enorm[:, None]
-
-  return weights
@@ -1,6 +1,6 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

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

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

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

  model = Model()
  opt = (nn.optim.Muon if getenv("MUON") else nn.optim.SGD if getenv("SGD") else nn.optim.Adam)(nn.state.get_parameters(model))

+  @TinyJit
+  @Tensor.train()
+  def train_step() -> Tensor:
+    opt.zero_grad()
+    samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
+    loss = model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
+    return loss.realize(*opt.schedule_step())
+
+  @TinyJit
+  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100
+
  test_acc = float('nan')
  for i in (t:=trange(getenv("STEPS", 70))):
    GlobalCounters.reset()   # NOTE: this makes it nice for DEBUG=2 timing
-    loss = model.train_step(X_train, Y_train)
-    if i%10 == 9: test_acc = model.get_test_acc(X_test, Y_test).item()
+    loss = train_step()
+    if i%10 == 9: test_acc = get_test_acc().item()
    t.set_description(f"loss: {loss.item():6.2f} test_accuracy: {test_acc:5.2f}%")

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

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

 if __name__ == "__main__":
@@ -19,8 +19,8 @@ cifar_std = [0.24703225141799082, 0.24348516474564, 0.26158783926049628]
 BS, STEPS = getenv("BS", 512), getenv("STEPS", 1000)
 EVAL_BS = getenv("EVAL_BS", BS)
 GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]
-assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}"
-assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}"
+assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"
+assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"

 class UnsyncedBatchNorm:
  def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1, num_devices=len(GPUS)):
@@ -65,7 +65,17 @@ def loader_process(q_in, q_out, X:Tensor, seed):
      else:
        # pad data with training mean
        img = np.tile(np.array([[[123.68, 116.78, 103.94]]], dtype=np.uint8), (224, 224, 1))
-      X[idx].flatten().assign(img.tobytes())
+
+      # broken out
+      #img_tensor = Tensor(img.tobytes(), device='CPU')
+      #storage_tensor = X[idx].contiguous().realize().lazydata.base.realized
+      #storage_tensor._copyin(img_tensor.numpy())
+
+      # faster
+      X[idx].contiguous().realize().uop.base.realized.as_buffer(force_zero_copy=True)[:] = img.tobytes()
+
+      # ideal
+      #X[idx].assign(img.tobytes())   # NOTE: this is slow!
      q_out.put(idx)
    q_out.put(None)

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

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

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

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

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

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

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

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

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

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

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

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

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

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

    return text

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

    return dataset_idx, dataset_sample_idx

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

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

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

 if __name__ == "__main__":
  def load_unet3d(val):
@@ -219,18 +219,7 @@ def get_mlperf_bert_model():
  config = get_mlperf_bert_config()
  if getenv("DISABLE_DROPOUT", 0):
    config["hidden_dropout_prob"] = config["attention_probs_dropout_prob"] = 0.0
-  model = BertForPretraining(**config)
-  if getenv("FP8_TRAIN"):
-    from extra.fp8.fp8_linear import convert_to_float8_training
-    def module_filter_fn(mod, fqn):
-      if isinstance(mod, LinearBert):
-        skip_layers = [] if (ln:=config["num_hidden_layers"]) <= 2 else ["bert.encoder.layer.0.", f"bert.encoder.layer.{ln-1}"]
-        if mod.weight.shape[-1] >= 1024 and "encoder" in fqn and not any(name in fqn for name in skip_layers):
-          print(f"replacing linear with fp8: {fqn} {mod.weight.shape}")
-          return True
-      return False
-    convert_to_float8_training(model, module_filter_fn)
-  return model
+  return BertForPretraining(**config)

 def get_fake_data_bert(BS:int):
  return {
@@ -234,9 +234,12 @@ def eval_llama3():
    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
    return loss.flatten().float()

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

  losses = []
  for tokens in tqdm(iter, total=5760//BS):
@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing

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

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

-from tinygrad.uop.ops import UOp
-
 def train_resnet():
  from extra.models import resnet
  from examples.mlperf.dataloader import batch_load_resnet
@@ -1010,7 +1008,6 @@ def train_bert():
  config["DISABLE_DROPOUT"] = getenv("DISABLE_DROPOUT", 0)
  config["TRAIN_BEAM"]    = TRAIN_BEAM = getenv("TRAIN_BEAM", BEAM.value)
  config["EVAL_BEAM"]     = EVAL_BEAM  = getenv("EVAL_BEAM", BEAM.value)
-  config["FP8_TRAIN"]     = getenv("FP8_TRAIN", 0)

  Tensor.manual_seed(seed)  # seed for weight initialization

@@ -1088,7 +1085,7 @@ def train_bert():
  if RUNMLPERF:
    # only load real data with RUNMLPERF
    eval_it = iter(batch_load_val_bert(EVAL_BS))
-    train_it = iter(tqdm(batch_load_train_bert(BS, seed=seed), total=train_steps, disable=BENCHMARK))
+    train_it = iter(tqdm(batch_load_train_bert(BS), total=train_steps, disable=BENCHMARK))
    for _ in range(start_step): next(train_it) # Fast forward
  else:
    # repeat fake data
@@ -1150,7 +1147,7 @@ def train_bert():

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

      cl = time.perf_counter()
@@ -1163,7 +1160,7 @@ def train_bert():
      if WANDB:
        wandb.log({"lr": lr, "train/loss": loss, "train/global_norm": global_norm.item(), "train/step_time": cl - st,
                    "train/python_time": pt - st, "train/data_time": dt - pt, "train/cl_time": cl - dt,
-                    "train/mem":GlobalCounters.mem_used / 1e9, "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})
+                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (cl - st), "epoch": (i+1)*GBS})

      train_data, next_data = next_data, None
      i += 1
@@ -1284,29 +1281,21 @@ def train_bert():
        previous_step = i

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

  config = {}
  BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
  BS                 = config["BS"]                     = getenv("BS", 16)
  grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
+  assert grad_acc == 1, f"{grad_acc=} is not supported"
  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
  SEED               = config["SEED"]                   = getenv("SEED", 5760)
-  DATA_SEED          = config["DATA_SEED"]              = getenv("DATA_SEED", SEED)
  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
  TRAIN_ON_VAL       = config["TRAIN_ON_VAL"]           = getenv("TRAIN_ON_VAL", 0)
  SMALL              = config["SMALL"]                  = getenv("SMALL", 0)
  SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000 * 1152)
-  EVAL_SAMPLES       = config["EVAL_SAMPLES"]           = getenv("EVAL_SAMPLES", 5760 if not SMALL else 1024)
-  MAX_STEPS          = config["MAX_STEPS"]              = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS))
-  WARMUP_STEPS       = config["WARMUP_STEPS"]           = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
-  LR                 = config["LR"]                     = getenv("LR", 8e-5 * GBS / 1152)
-  END_LR             = config["END_LR"]                 = getenv("END_LR", 8e-7)
  EVAL_FREQ          = config["EVAL_FREQ"]              = getenv("EVAL_FREQ", 46080)
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)
@@ -1320,33 +1309,16 @@ def train_llama3():
  opt_adamw_weight_decay = 0.1

  opt_gradient_clip_norm = 1.0
-  opt_learning_rate_warmup_steps = WARMUP_STEPS
-  opt_learning_rate_decay_steps = MAX_STEPS - opt_learning_rate_warmup_steps
-  opt_base_learning_rate = LR
-  opt_end_learning_rate = END_LR
-
-  Tensor.manual_seed(SEED)  # seed for weight initialization
-
-  # ** init wandb **
-  WANDB = getenv("WANDB")
-  if WANDB:
-    import wandb
-    wandb_args = {"id": wandb_id, "resume": "must"} if (wandb_id := getenv("WANDB_RESUME", "")) else {}
-    wandb.init(config=config, **wandb_args, project="MLPerf-LLaMA3")
+  opt_learning_rate_warmup_steps = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
+  opt_learning_rate_decay_steps = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS)) - opt_learning_rate_warmup_steps
+  opt_base_learning_rate = getenv("LR", 8e-5 * GBS / 1152)  # NOTE: cannot change for benchmark
+  opt_end_learning_rate = getenv("END_LR", 8e-7)

  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
  # vocab_size from the mixtral tokenizer
  if not SMALL: model_params |= {"vocab_size": 32000}
-  real_vocab_size = model_params['vocab_size']
  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
-  print(f"model parameters: {model_params}")
-
-  # pad vocab
-  if (MP := getenv("MP", 1)) > 1: model_params['vocab_size'] = round_up(model_params['vocab_size'], 256 * MP)
-  vocab_mask:Tensor = Tensor.arange(model_params['vocab_size']).reshape(1, 1, -1) >= real_vocab_size
-
-  model = FlatTransformer(**model_params, max_context=SEQLEN)
-
+  model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
  params = get_parameters(model)
  # weights are all bfloat16 for now
  assert params and all(p.dtype == dtypes.bfloat16 for p in params)
@@ -1355,26 +1327,32 @@ def train_llama3():
    for v in get_parameters(model):
      v = v.assign(Tensor.empty(v.shape))

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

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

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

  if resume_ckpt := getenv("RESUME_CKPT"):
@@ -1387,172 +1365,116 @@ def train_llama3():
    load_state_dict(scheduler, safe_load(fn), realize=False)

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

-      if WANDB:
-        wandb.log({"eval/log_perplexity": log_perplexity, "eval/sequences_seen": sequences_seen})
-
      if log_perplexity < EVAL_TARGET:
        tqdm.write(f"target achieved after {sequences_seen} sequences")
        if getenv("CKPT"):
@@ -1631,7 +1553,7 @@ def train_stable_diffusion():
    loss, out_lr = loss.detach().to("CPU"), optimizer.lr.to("CPU")
    Tensor.realize(loss, out_lr)
    return loss, out_lr
-
+    
  # checkpointing takes ~9 minutes without this, and ~1 minute with this
  @TinyJit
  def ckpt_to_cpu():
@@ -1670,7 +1592,7 @@ def train_stable_diffusion():
    if i == 3:
      for _ in range(3): ckpt_to_cpu() # do this at the beginning of run to prevent OOM surprises when checkpointing
      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-
+      
    total_train_time = time.perf_counter() - train_start_time
    if WANDB:
      wandb.log({"train/loss": loss_item, "train/lr": lr_item, "train/loop_time_prev": loop_time, "train/dl_time": dl_time, "train/step": i,
@@ -1,193 +0,0 @@
-import math, os
-if __name__ == "__main__":
-  os.environ["DEFAULT_FLOAT"] = "bfloat16"
-  os.environ["OPTIM_DTYPE"] = "bfloat16"
-  if "DEV" not in os.environ: os.environ["DEV"] = "NULL"
-  # CDNA
-  os.environ["EMULATE"] = "AMD_CDNA4"
-  os.environ["DEVICE_IN_FUNCTION_BUG"] = "1"
-  os.environ["ALL2ALL"] = "1"
-  os.environ["USE_ATOMICS"] = "1"
-  if "HK_FLASH_ATTENTION" not in os.environ:
-    os.environ["HK_FLASH_ATTENTION"] = "1"
-    if "ASM_GEMM" not in os.environ:
-      os.environ["ASM_GEMM"] = "1"
-from tinygrad import Tensor, nn, function, getenv, dtypes, TinyJit
-from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
-from tinygrad.uop.ops import Ops, UOp
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
-
-FP8 = getenv("FP8", 0)
-
-FP8_DTYPE = dtypes.fp8e4m3
-FP8_MAX = 448.0
-
-def quantize_fp8(x:Tensor):
-  scale = FP8_MAX / (x.abs().max().detach() + 1e-8)
-  x_scaled = x * scale
-  x_clamped = x_scaled + (x_scaled.detach().clamp(-FP8_MAX, FP8_MAX) - x_scaled.detach())  # STE
-  return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal()
-
-def matmul(x:Tensor, w:Tensor) -> Tensor:
-  if not FP8: return x @ w.T
-  # weights are already FP8, just quantize activations
-  x_fp8, x_scale = quantize_fp8(x)
-  return x_fp8.dot(w.T, dtype=dtypes.float) * x_scale
-
-def rmsnorm(x_in:Tensor, eps:float):
-  x = x_in.float()
-  x = x * (x.square().mean(-1, keepdim=True) + eps).rsqrt()
-  return x.cast(x_in.dtype)
-
-class FlatTransformer:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
-               rope_theta:int=10000, max_context:int=1024):
-    self.vocab_size = vocab_size
-    self.n_layers = n_layers
-    self.n_heads = n_heads
-    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
-    self.head_dim = dim // n_heads
-    self.n_rep = self.n_heads // self.n_kv_heads
-
-    # Attention
-    self.wqkv = self.lin_per_layer(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2)
-    self.wo = self.lin_per_layer(self.n_heads * self.head_dim, dim)
-
-    # FeedForward
-    self.w1 = self.lin_per_layer(dim, hidden_dim)
-    self.w2 = self.lin_per_layer(hidden_dim, dim)
-    self.w3 = self.lin_per_layer(dim, hidden_dim)
-
-    self.norm_eps = norm_eps
-    self.attention_norm = Tensor.ones(n_layers, dim).contiguous()
-    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
-
-    # output
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = nn.Embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
-
-  def lin_per_layer(self, in_features:int, out_features:int):
-    bound = 1 / math.sqrt(in_features)
-    dt = FP8_DTYPE if FP8 else None
-    if getenv("ZEROS"): return Tensor.zeros(self.n_layers, out_features, in_features, dtype=dt)
-    return Tensor.uniform(self.n_layers, out_features, in_features, low=-bound, high=bound, dtype=dt)
-
-  def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wqkv:Tensor, wo:Tensor):
-    x = rmsnorm(x, self.norm_eps) * attention_norm
-    xqkv = matmul(x, wqkv)
-
-    bsz, seqlen, _ = xqkv.shape
-    # interleaved layout: each kv group has [n_rep q heads, 1 k head, 1 v head] for clean MP sharding
-    xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
-    xq = xqkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
-    xk = xqkv[:, :, :, self.n_rep].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-    xv = xqkv[:, :, :, self.n_rep+1].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
-    attn = attn.reshape(bsz, seqlen, -1)
-    return matmul(attn, wo)
-
-  def feed_forward(self, x:Tensor, ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
-    x = rmsnorm(x, self.norm_eps) * ffn_norm
-    x_w1 = matmul(x, w1).silu()
-    x_w3 = matmul(x.contiguous_backward(), w3)
-    return matmul(x_w1 * x_w3, w2)
-
-  @function(precompile=True, precompile_backward=True)
-  def run_layer(self, x:Tensor, freqs_cis:Tensor,
-                attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
-    h = x + self.attention(x, freqs_cis, attention_norm, wqkv, wo)
-    return h + self.feed_forward(h, ffn_norm, w1, w2, w3)
-
-  def shard(self, device:tuple[str, ...], mp:bool=False):
-    from tinygrad.nn.state import get_parameters
-    if not mp:
-      for v in get_parameters(self): v.shard_(device, axis=None)
-    else:
-      # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
-      self.wqkv.shard_(device, axis=1).realize()          # (n_layers, out, dim) shard out
-      self.wo.shard_(device, axis=2).realize()             # (n_layers, dim, in) shard in
-      self.w1.shard_(device, axis=1).realize()             # (n_layers, hidden, dim) shard out
-      self.w2.shard_(device, axis=2).realize()             # (n_layers, dim, hidden) shard in
-      self.w3.shard_(device, axis=1).realize()             # (n_layers, hidden, dim) shard out
-      self.attention_norm.shard_(device, axis=None).realize()
-      self.ffn_norm.shard_(device, axis=None).realize()
-      self.norm.weight.shard_(device, axis=None).realize()
-      self.tok_embeddings.weight.shard_(device, axis=0).realize()
-      self.output.weight.shard_(device, axis=0).realize()
-      self.freqs_cis.shard_(device, axis=None).realize()
-
-  def __call__(self, tokens:Tensor):
-    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    for i in range(self.n_layers):
-      h = self.run_layer(h, freqs_cis,
-                         self.attention_norm[i], self.wqkv[i], self.wo[i],
-                         self.ffn_norm[i], self.w1[i], self.w2[i], self.w3[i])
-    logits = self.output(self.norm(h))
-    return logits
-
-# TODO: this shouldn't be needed, but it prevents a copy of the grads. CAT can help
-def apply_grad(old_grad:UOp, new_grad:UOp) -> list[UOp]:
-  if new_grad.op == Ops.ADD:
-    return apply_grad(old_grad, new_grad.src[0])+apply_grad(old_grad, new_grad.src[1])
-  elif new_grad.op == Ops.PAD:
-    grad_shrink = tuple([(p[0], s+p[0]) for s,p in zip(new_grad.src[0].shape, new_grad.marg)])
-    return apply_grad(old_grad.shrink(grad_shrink), new_grad.src[0])
-  else:
-    return [old_grad.store(old_grad + new_grad)]
-
-if __name__ == "__main__":
-  config = {}
-  BS                 = config["BS"]                     = getenv("BS", 16)
-  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
-
-  from examples.llama3 import MODEL_PARAMS
-  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
-  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
-  model = FlatTransformer(**model_params, max_context=SEQLEN)
-  state = nn.state.get_state_dict(model)
-  print("tensor count:", len(state))
-
-  # shard the model
-  from tinygrad import Device
-  if (DP := getenv("DP", 1)) > 1:
-    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)))
-  if (MP := getenv("MP", 1)) > 1:
-    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)), mp=True)
-
-  # preallocate all the grad buffers and zero them out
-  grads = {x:Tensor.zeros_like(x).contiguous() for x in state.values() if x.requires_grad is None}
-
-  # print model size
-  sz = 0
-  for k,v in state.items():
-    print(f"{colored(k, 'green' if v in grads else 'white'):30s} {str(v.shape):30s} {str(v.dtype):20s} {v.device}  {v.nbytes()/1e9:.2f} GB")
-    sz += v.nbytes()
-  print(f"total sz: {sz/1e9:.2f} GB")
-
-  with Timing("fake data: "): tokens = Tensor.randint(BS, SEQLEN+1, low=0, high=model.vocab_size, dtype=dtypes.int)
-  with Timing("realize weights/grads/data: "): Tensor.realize(*state.values(), *grads.values(), tokens)
-  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
-  if DP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)), axis=0)
-  if MP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)))
-
-  @TinyJit
-  def jit_step(tokens:Tensor):
-    with Timing("python forward: "): loss = model(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    with Timing("python backward: "):
-      for t,g in zip(grads, loss.gradient(*grads)):
-        grads[t] = Tensor(grads[t].uop.after(UOp.group(*apply_grad(grads[t].uop, g.uop))), device=t.device)
-    with Timing("run step: "): loss.realize(*grads.values())
-
-  for i in range(6):
-    GlobalCounters.reset()
-    profile_marker(f"step {i}")
-    with Timing(colored(f"*** step {i}: ", "red")):
-      jit_step(tokens)
-  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
@@ -1,80 +0,0 @@
-from tinygrad import Tensor, nn
-from tinygrad.helpers import getenv
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
-
-class Attention:
-  def __init__(self, dim:int, n_heads:int, n_kv_heads:int|None=None, linear=nn.Linear):
-    self.n_heads = n_heads
-    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
-    self.head_dim = dim // n_heads
-    self.n_rep = self.n_heads // self.n_kv_heads
-
-    if getenv("WQKV"):
-      self.wqkv = linear(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2, bias=False)
-    else:
-      self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
-      self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-      self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-
-    self.wo = linear(self.n_heads * self.head_dim, dim, bias=False)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor) -> Tensor:
-    if getenv("WQKV"):
-      xqkv = self.wqkv(x)
-      xqkv = xqkv.reshape(xqkv.shape[0], xqkv.shape[1], self.n_kv_heads, self.n_rep + 2, self.head_dim)
-      xq = xqkv[:, :, :, :self.n_rep].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xk = xqkv[:, :, :, self.n_rep:self.n_rep+1].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xv = xqkv[:, :, :, self.n_rep+1:self.n_rep+2].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-    else:
-      xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
-
-    xq = xq.reshape(xq.shape[0], xq.shape[1], self.n_heads, self.head_dim)
-    xk = xk.reshape(xk.shape[0], xk.shape[1], self.n_kv_heads, self.head_dim)
-    xv = xv.reshape(xv.shape[0], xv.shape[1], self.n_kv_heads, self.head_dim)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    bsz, seqlen, _, _ = xq.shape
-
-    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
-
-    attn = attn.reshape(bsz, seqlen, -1)
-    return self.wo(attn)
-
-class FeedForward:
-  def __init__(self, dim:int, hidden_dim:int, linear=nn.Linear):
-    self.w1 = linear(dim, hidden_dim, bias=False)
-    self.w2 = linear(hidden_dim, dim, bias=False)
-    self.w3 = linear(dim, hidden_dim, bias=False) # the gate in Gated Linear Unit
-
-  def __call__(self, x:Tensor) -> Tensor:
-    w1 = self.w1(x).silu()
-    w3 = self.w3(x)
-    return self.w2(w1 * w3)
-
-class TransformerBlock:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int|None, norm_eps:float, linear=nn.Linear):
-    self.attention = Attention(dim, n_heads, n_kv_heads, linear)
-    self.feed_forward = FeedForward(dim, hidden_dim, linear)
-    self.attention_norm = nn.RMSNorm(dim, norm_eps)
-    self.ffn_norm = nn.RMSNorm(dim, norm_eps)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor):
-    h = x + self.attention(self.attention_norm(x), freqs_cis)
-    return h + self.feed_forward(self.ffn_norm(h))
-
-class Transformer:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
-               rope_theta:int=10000, max_context:int=1024, linear=nn.Linear, embedding=nn.Embedding):
-    self.layers = [TransformerBlock(dim, hidden_dim, n_heads, n_kv_heads, norm_eps, linear) for _ in range(n_layers)]
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False) if embedding == nn.Embedding else linear(dim, vocab_size, bias=False)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
-
-  def __call__(self, tokens:Tensor):
-    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    for layer in self.layers: h = layer(h, freqs_cis)
-    logits = self.output(self.norm(h))
-    return logits
@@ -1,140 +0,0 @@
-import os
-os.environ["WQKV"] = "1"
-import unittest
-import numpy as np
-from tinygrad import Tensor, nn, dtypes
-from tinygrad.nn.state import get_parameters
-from tinygrad.device import is_dtype_supported
-from examples.mlperf.models.llama import Transformer
-from examples.mlperf.models.flat_llama import FlatTransformer
-
-def copy_weights(flat:FlatTransformer, ref:Transformer):
-  n_layers = flat.n_layers
-  Tensor.realize(*nn.state.get_state_dict(ref).values())
-  flat.wqkv.assign(Tensor(np.stack([ref.layers[i].attention.wqkv.weight.numpy() for i in range(n_layers)])))
-  flat.wo.assign(Tensor(np.stack([ref.layers[i].attention.wo.weight.numpy() for i in range(n_layers)])))
-  flat.w1.assign(Tensor(np.stack([ref.layers[i].feed_forward.w1.weight.numpy() for i in range(n_layers)])))
-  flat.w2.assign(Tensor(np.stack([ref.layers[i].feed_forward.w2.weight.numpy() for i in range(n_layers)])))
-  flat.w3.assign(Tensor(np.stack([ref.layers[i].feed_forward.w3.weight.numpy() for i in range(n_layers)])))
-  flat.attention_norm.assign(Tensor(np.stack([ref.layers[i].attention_norm.weight.numpy() for i in range(n_layers)])))
-  flat.ffn_norm.assign(Tensor(np.stack([ref.layers[i].ffn_norm.weight.numpy() for i in range(n_layers)])))
-  flat.norm.weight.assign(Tensor(ref.norm.weight.numpy()))
-  flat.tok_embeddings.weight.assign(Tensor(ref.tok_embeddings.weight.numpy()))
-  flat.output.weight.assign(Tensor(ref.output.weight.numpy()))
-
-class TestFlatLlama(unittest.TestCase):
-  def test_forward_match(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-    tokens = Tensor([[1, 50, 100, 999, 2]])
-    ref_logits = ref(tokens).realize()
-    flat_logits = flat(tokens).realize()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    diff = (ref_logits - flat_logits).abs().max().item()
-    self.assertLess(diff, 1e-5, f"forward mismatch: max abs diff {diff}")
-
-  def test_backward_match(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-
-    for p in get_parameters(ref): p.requires_grad_(True)
-    for p in get_parameters(flat): p.requires_grad_(True)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-    tokens = Tensor([[1, 50, 100, 999, 2, 10]])
-
-    ref_loss = ref(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    ref_loss.backward()
-    ref_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(ref).items() if v.grad is not None}
-
-    flat_loss = flat(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
-    flat_loss.backward()
-    flat_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(flat).items() if v.grad is not None}
-
-    # check loss matches
-    self.assertAlmostEqual(ref_loss.item(), flat_loss.item(), places=4)
-
-    # check output weight grad matches
-    diff = abs(ref_grads["output.weight"] - flat_grads["output.weight"]).max()
-    self.assertLess(diff, 1e-4, f"output.weight grad mismatch: max abs diff {diff}")
-
-    # check per-layer weight grads match
-    for i in range(params["n_layers"]):
-      for flat_key, ref_key in [
-        ("wqkv", f"layers.{i}.attention.wqkv.weight"),
-        ("wo", f"layers.{i}.attention.wo.weight"),
-        ("w1", f"layers.{i}.feed_forward.w1.weight"),
-        ("w2", f"layers.{i}.feed_forward.w2.weight"),
-        ("w3", f"layers.{i}.feed_forward.w3.weight"),
-      ]:
-        diff = abs(ref_grads[ref_key] - flat_grads[flat_key][i]).max()
-        self.assertLess(diff, 1e-4, f"layer {i} {flat_key} grad mismatch: max abs diff {diff}")
-
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
-  def test_forward_match_mp(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    from tinygrad import Device
-    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-    flat.shard(devices, mp=True)
-
-    tokens = Tensor([[1, 50, 100, 999, 2]], device=devices[0])
-    ref_logits = ref(tokens.to(devices[0])).numpy()
-    flat_logits = flat(tokens.shard(devices)).numpy()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
-
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
-  def test_forward_match_dp(self):
-    Tensor.manual_seed(42)
-    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-    from tinygrad import Device
-    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
-    ref = Transformer(**params)
-    flat = FlatTransformer(**params)
-    copy_weights(flat, ref)
-    Tensor.realize(*nn.state.get_state_dict(flat).values())
-    flat.shard(devices)
-
-    tokens = Tensor([[1, 50, 100, 999, 2], [2, 100, 50, 1, 999]], device=devices[0])
-    ref_logits = ref(tokens.to(devices[0])).numpy()
-    flat_logits = flat(tokens.shard(devices, axis=0)).numpy()
-    self.assertEqual(ref_logits.shape, flat_logits.shape)
-    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
-
-  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3), "fp8 not supported on this device")
-  def test_forward_fp8(self):
-    import examples.mlperf.models.flat_llama as flat_llama_mod
-    old_fp8 = flat_llama_mod.FP8
-    try:
-      flat_llama_mod.FP8 = 1
-      Tensor.manual_seed(42)
-      params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
-      ref = Transformer(**params)
-      flat = FlatTransformer(**params)
-      copy_weights(flat, ref)
-      Tensor.realize(*nn.state.get_state_dict(flat).values())
-
-      tokens = Tensor([[1, 50, 100, 999, 2]])
-      ref_logits = ref(tokens).numpy()
-      flat_logits = flat(tokens).numpy()
-      self.assertEqual(ref_logits.shape, flat_logits.shape)
-      # FP8 has lower precision, allow larger tolerance
-      np.testing.assert_allclose(flat_logits, ref_logits, atol=1.0, rtol=0.1)
-    finally:
-      flat_llama_mod.FP8 = old_fp8
-
-if __name__ == "__main__":
-  unittest.main()
@@ -1,59 +0,0 @@
-from tinygrad.tensor import Tensor
-from tinygrad.dtype import dtypes
-from tinygrad.nn.optim import Optimizer
-from tinygrad.helpers import FUSE_OPTIM
-
-class GradAccClipAdamW(Optimizer):
-  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, device=None, fused=FUSE_OPTIM):
-    super().__init__(params, lr, device, fused)
-    self.b1, self.b2, self.eps, self.wd = b1, b2, eps, weight_decay
-    self.b1_t, self.b2_t = (Tensor.ones((1,), dtype=dtypes.float32, device=self.device, requires_grad=False) for _ in [b1, b2])
-    self.m = self._new_optim_param()
-    self.v = self._new_optim_param()
-    self.grad_acc, self.clip_norm = grad_acc, clip_norm
-
-  def fstep(self, grads:list[Tensor]):
-    if self.fused:
-      out, extra = self._step([], grads)
-      updates = [out[0][self.pos_params[i]:self.pos_params[i+1]].reshape(tt.shape) for i, tt in enumerate(self.params)]
-    else:
-      updates, extra = self._step([], grads)
-    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i]))
-    to_realize = extra+self.params+self.buffers
-
-    Tensor.realize(*to_realize)
-    return extra[-1]
-
-  def _step(self, params:list[Tensor], grads:list[Tensor]) -> tuple[list[Tensor], list[Tensor]]:
-    grads = list(grads)
-
-    for i in range(len(grads)):
-      if grads[i].device != self.m[i].device: grads[i] = grads[i].to(self.m[i].device)
-
-    if self.fused:
-      grads[0].assign(grads[0] / self.grad_acc)
-      total_norm = grads[0].float().square().sum().sqrt()
-      grads[0].assign((grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype))
-    else:
-      for i in range(len(grads)):
-        grads[i].assign(grads[i] / self.grad_acc)
-      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
-      for i in range(len(grads)):
-        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype))
-
-    ret = []
-    self.b1_t *= self.b1
-    self.b2_t *= self.b2
-    for i, g in enumerate(grads):
-      self.m[i].assign((self.b1 * self.m[i] + (1.0 - self.b1) * g).cast(self.m[i].dtype))
-      self.v[i].assign((self.b2 * self.v[i] + (1.0 - self.b2) * (g * g)).cast(self.v[i].dtype))
-      m_hat = (self.m[i] / (1.0 - self.b1_t)).cast(self.m[i].dtype)
-      v_hat = (self.v[i] / (1.0 - self.b2_t)).cast(self.v[i].dtype)
-      up = m_hat / (v_hat.sqrt() + self.eps)
-      ret.append((self.lr * up).cast(g.dtype))
-    return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]
-
-  def _apply_update(self, t:Tensor, up:Tensor) -> Tensor:
-    wd = self.wd if t.ndim >= 3 else 0.0
-    up = up.shard_like(t) + self.lr.to(t.device) * wd * t.detach()
-    return t.detach() - up.cast(t.dtype)
@@ -4,7 +4,7 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128

-export CHECK_OOB=0
+export IGNORE_OOB=1

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

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

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

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

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

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

 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -1,20 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1 DEBUG=0 JIT=1 FLASH_ATTENTION=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=0 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,24 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-export BEAM_TIMEOUT_SEC=15
-export FP8_TRAIN=1
-# search
-IGNORE_BEAM_CACHE=1 BENCHMARK=10 BERT_LAYERS=2 RUNMLPERF=0 python3 examples/mlperf/model_train.py
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -11,7 +11,7 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900

-export CHECK_OOB=0
+export IGNORE_OOB=1
 export REWRITE_STACK_LIMIT=5000000

 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -4,7 +4,7 @@ export PYTHONPATH="." NV=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72

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

 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -4,7 +4,7 @@ export PYTHONPATH="." NV=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72

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

 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -7,7 +7,7 @@ export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72

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

 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -4,7 +4,7 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

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

 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -4,7 +4,7 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

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

 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -7,7 +7,7 @@ export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

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

 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
@@ -1,38 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,32 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1152}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -1,42 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,43 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,37 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-4}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -1,38 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-32}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -1,5 +0,0 @@
-#!/bin/bash
-export BENCHMARK=5
-export EVAL_BS=0
-VIZ=${VIZ:--1} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
-extra/viz/cli.py --profile --device "AMD" --limit 20
@@ -1,10 +0,0 @@
-#!/bin/bash
-export BENCHMARK=5
-export EVAL_BS=0
-export FAKEDATA=1
-export NULL_ALLOW_COPYOUT=1
-export HIP_VISIBLE_DEVICES=""
-export DEV=NULL
-export JITBEAM=0
-export LLAMA_LAYERS=${LLAMA_LAYERS:-"2"}
-time examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
@@ -1,11 +1,12 @@
 from pathlib import Path
+import numpy as np
 import torch
 from torchvision.utils import make_grid, save_image
 from tinygrad.nn.state import get_parameters
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import trange
 from tinygrad.nn import optim
-from tinygrad.nn.datasets import mnist
+from extra.datasets import fetch_mnist

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

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

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

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

 if __name__ == "__main__":
  # data for training and validation
-  X_train, _, _, _ = mnist()
+  images_real = np.vstack(fetch_mnist()[::2])
  ds_noise = Tensor.randn(64, 128, requires_grad=False)
  # parameters
  epochs, batch_size, k = 300, 512, 1
  sample_interval = epochs // 10
-  n_steps = X_train.shape[0] // batch_size
+  n_steps = len(images_real) // batch_size
  # models and optimizer
  generator = LinearGen()
  discriminator = LinearDisc()
@@ -83,24 +84,24 @@ if __name__ == "__main__":
  output_dir = Path(".").resolve() / "outputs"
  output_dir.mkdir(exist_ok=True)
  # optimizers
-  optim_g = optim.Adam(get_parameters(generator), lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
-  optim_d = optim.Adam(get_parameters(discriminator), lr=0.0002, b1=0.5)
+  optim_g = optim.Adam(get_parameters(generator),lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
+  optim_d = optim.Adam(get_parameters(discriminator),lr=0.0002, b1=0.5)
  # training loop
-  with Tensor.train():
-    for epoch in (t := trange(epochs)):
-      loss_g, loss_d = 0.0, 0.0
-      for _ in range(n_steps):
-        data_real = make_batch(X_train)
-        for step in range(k):  # Try with k = 5 or 7.
-          noise = Tensor.randn(batch_size, 128)
-          data_fake = generator.forward(noise).detach()
-          loss_d += train_discriminator(optim_d, data_real, data_fake)
+  Tensor.training = True
+  for epoch in (t := trange(epochs)):
+    loss_g, loss_d = 0.0, 0.0
+    for _ in range(n_steps):
+      data_real = make_batch(images_real)
+      for step in range(k):  # Try with k = 5 or 7.
        noise = Tensor.randn(batch_size, 128)
-        data_fake = generator.forward(noise)
-        loss_g += train_generator(optim_g, data_fake)
-      if (epoch + 1) % sample_interval == 0:
-        fake_images = generator.forward(ds_noise).detach().numpy()
-        fake_images = (fake_images.reshape(-1, 1, 28, 28) + 1) / 2  # 0 - 1 range.
-        save_image(make_grid(torch.tensor(fake_images)), output_dir / f"image_{epoch+1}.jpg")
-      t.set_description(f"Generator loss: {loss_g/n_steps}, Discriminator loss: {loss_d/n_steps}")
+        data_fake = generator.forward(noise).detach()
+        loss_d += train_discriminator(optim_d, data_real, data_fake)
+      noise = Tensor.randn(batch_size, 128)
+      data_fake = generator.forward(noise)
+      loss_g += train_generator(optim_g, data_fake)
+    if (epoch + 1) % sample_interval == 0:
+      fake_images = generator.forward(ds_noise).detach().numpy()
+      fake_images = (fake_images.reshape(-1, 1, 28, 28) + 1) / 2  # 0 - 1 range.
+      save_image(make_grid(torch.tensor(fake_images)), output_dir / f"image_{epoch+1}.jpg")
+    t.set_description(f"Generator loss: {loss_g/n_steps}, Discriminator loss: {loss_d/n_steps}")
  print("Training Completed!")
@@ -31,7 +31,7 @@ def compile(onnx_file):
  for i in range(3):
    GlobalCounters.reset()
    print(f"run {i}")
-    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1), OPENPILOT_HACKS=1):
+    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1)):
      ret = run_onnx_jit(**inputs).numpy()
    # copy i == 1 so use of JITBEAM is okay
    if i == 1: test_val = np.copy(ret)
@@ -1,16 +0,0 @@
-import sys, pickle
-from extra.bench_log import WallTimeEvent, BenchEvent
-from tinygrad.helpers import getenv
-
-PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
-
-load_times = []
-
-for _ in range(10):
-  with WallTimeEvent(BenchEvent.STEP) as wte: pickle.load(open(PKL, 'rb'))
-  load_times.append(wte.time)
-  print(f"pickle load: {wte.time:6.2f} s")
-
-if (assert_time:=getenv("ASSERT_MIN_LOAD_TIME")):
-  min_time = min(load_times)
-  assert min_time < assert_time, f"Speed regression, expected min load time of < {assert_time} s but took: {min_time} s"
@@ -6,6 +6,7 @@ import argparse, time
 from collections import namedtuple
 from typing import Dict, Any

+from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
 from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten, profile_marker
@@ -335,7 +336,6 @@ if __name__ == "__main__":
  print(x.shape)

  profile_marker("save image")
-  from PIL import Image
  im = Image.fromarray(x.numpy())
  print(f"saving {args.out}")
  im.save(args.out)
@@ -7,7 +7,7 @@ if __name__ == "__main__":
  with open(fetch(sys.argv[1]), "rb") as f:
    run_onnx_jit = pickle.load(f)
  input_name = run_onnx_jit.captured.expected_names[0]
-  device = run_onnx_jit.captured.expected_input_info[0][-1]
+  device = run_onnx_jit.captured.expected_st_vars_dtype_device[0][-1]
  print(f"input goes into {input_name=} on {device=}")
  hit = 0
  for i,(img,y) in enumerate(imagenet_dataloader(cnt=getenv("CNT", 100))):
@@ -48,7 +48,7 @@ def prepare_browser_chunks(model):
        weight_metadata = metadata.get(name, default)
        weight_metadata["parts"][part_num] = {"file": i, "file_start_pos": cursor, "size": size}
        metadata[name] = weight_metadata
-        data = bytes(state_dict[name].uop.base.realized.as_memoryview())
+        data = bytes(state_dict[name].uop.base.realized.as_buffer())
        data = data if not offsets else data[offsets[0]:offsets[1]]
        writer.write(data)
        cursor += size
@@ -93,7 +93,7 @@ if __name__ == "__main__":
    forward: Any = None

  sub_steps = [
-    Step(name = "textModel", input = [Tensor.randint(1, 77, low=0, high=49408, dtype=dtypes.int32)], forward = model.cond_stage_model.transformer.text_model),
+    Step(name = "textModel", input = [Tensor.randn(1, 77)], forward = model.cond_stage_model.transformer.text_model),
    Step(name = "diffusor", input = [Tensor.randn(1, 77, 768), Tensor.randn(1, 77, 768), Tensor.randn(1,4,64,64), Tensor.rand(1), Tensor.randn(1), Tensor.randn(1), Tensor.randn(1)], forward = model),
    Step(name = "decoder", input = [Tensor.randn(1,4,64,64)], forward = model.decode),
    Step(name = "f16tof32", input = [Tensor.randn(2097120, dtype=dtypes.uint32)], forward = u32_to_f16)
@@ -7,7 +7,6 @@ from tinygrad import Tensor, TinyJit, Variable, nn, dtypes
 from tinygrad.nn.state import torch_load, load_state_dict
 from tinygrad.helpers import getenv, fetch

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

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

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

  log_spec = np.log10(np.clip(mel_spec, 1e-10, None))
  log_spec = np.maximum(log_spec, log_spec.max((1,2), keepdims=True) - 8.0)
@@ -65,7 +65,7 @@ def get_bar0_size(pcibus):
 class AMSMI(AMDev):
  def __init__(self, pcibus, vram_bar:MMIOInterface, doorbell_bar:MMIOInterface, mmio_bar:MMIOInterface):
    self.pcibus = pcibus
-    self.vram, self.doorbell64, self.mmio = vram_bar, doorbell_bar, mmio_bar
+    self.vram, self.doorbell64, self.mmio, self.dma_regions = vram_bar, doorbell_bar, mmio_bar, None
    self.pci_state = self.read_pci_state()
    if self.pci_state == "D0": self._init_from_d0()

@@ -92,7 +92,7 @@ class SMICtx:
    self.prev_terminal_width = 0
    self.prev_terminal_height = 0

-    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:", "Processing accelerators:"]
+    remove_parts = ["Advanced Micro Devices, Inc. [AMD/ATI]", "VGA compatible controller:"]
    lspci = subprocess.check_output(["lspci"]).decode("utf-8").splitlines()
    self.lspci = {l.split()[0]: l.split(" ", 1)[1] for l in lspci}
    for k,v in self.lspci.items():
@@ -153,8 +153,8 @@ class SMICtx:
    tables = {}
    for dev in self.devs:
      match dev.ip_ver[am.MP1_HWIP]:
-        case (13,0,6): table_t = dev.smu.smu_mod.MetricsTableV0_t
-        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTable_t
+        case (13,0,6): table_t = dev.smu.smu_mod.MetricsTableX_t
+        case (13,0,12): table_t = dev.smu.smu_mod.MetricsTableV2_t
        case _: table_t = dev.smu.smu_mod.SmuMetricsExternal_t
      tables[dev] = dev.smu.read_table(table_t, dev.smu.smu_mod.SMU_TABLE_SMU_METRICS) if dev.pci_state == "D0" else None
    return tables
@@ -165,17 +165,17 @@ class SMICtx:

  def get_gfx_activity(self, dev, metrics):
    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
+      case (13,0,6): return max(0, min(100, self._smuq10_round(metrics.SocketGfxBusy)))
      case _: return metrics.SmuMetrics.AverageGfxActivity

  def get_mem_activity(self, dev, metrics):
    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
+      case (13,0,6): return max(0, min(100, self._smuq10_round(metrics.DramBandwidthUtilization)))
      case _: return metrics.SmuMetrics.AverageUclkActivity

  def get_temps(self, dev, metrics, compact=False):
    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12):
+      case (13,0,6):
        temps = {
          "Hotspot": self._smuq10_round(metrics.MaxSocketTemperature),
          "HBM": self._smuq10_round(metrics.MaxHbmTemperature),
@@ -184,16 +184,16 @@ class SMICtx:
        if compact: return {k: temps[k] for k in ("Hotspot", "HBM") if temps.get(k, 0) != 0}
        return {k: v for k, v in temps.items() if v != 0}
      case _:
-        temps_keys = [(k, name) for k, name in dev.smu.smu_mod.TEMP_e.items()
+        temps_keys = [(k, name) for k, name in dev.smu.smu_mod.c__EA_TEMP_e__enumvalues.items()
                      if k < dev.smu.smu_mod.TEMP_COUNT and metrics.SmuMetrics.AvgTemperature[k] != 0]
        if compact: temps_keys = [(k, name) for k, name in temps_keys if k in (dev.smu.smu_mod.TEMP_HOTSPOT, dev.smu.smu_mod.TEMP_MEM)]
        return {name: metrics.SmuMetrics.AvgTemperature[k] for k, name in temps_keys}

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

@@ -205,37 +205,38 @@ class SMICtx:
  def get_gfx_freq(self, dev, metrics):
    if metrics is None: return 0
    match dev.ip_ver[am.MP1_HWIP]:
-      case (13,0,6)|(13,0,12): return self._smuq10_round(metrics.GfxclkFrequency[0])
+      case (13,0,6): return self._smuq10_round(metrics.GfxclkFrequency[0])
      case _:
        return metrics.SmuMetrics.AverageGfxclkFrequencyPostDs if self.get_gfx_activity(dev, metrics) <= self.get_busy_threshold(dev) else \
               metrics.SmuMetrics.AverageGfxclkFrequencyPreDs

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

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

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

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

  def get_mem_usage(self, dev):
+    return 0
+
    usage = 0
    pt_stack = [dev.mm.root_page_table]
    while len(pt_stack) > 0:
@@ -244,8 +245,8 @@ class SMICtx:
        entry = pt.entries[i]

        if (entry & am.AMDGPU_PTE_VALID) == 0: continue
-        if pt.lv < am.AMDGPU_VM_PDB0 and not dev.gmc.is_pte_huge_page(pt.lv, entry):
-          pt_stack.append(AMPageTableEntry(dev, dev.xgmi2paddr(entry & 0x0000FFFFFFFFF000), lv=pt.lv+1))
+        if pt.lv!=am.AMDGPU_VM_PTB and not dev.gmc.is_pte_huge_page(pt.lv, entry):
+          pt_stack.append(AMPageTableEntry(dev, entry & 0x0000FFFFFFFFF000, lv=pt.lv+1))
          continue
        if (entry & am.AMDGPU_PTE_SYSTEM) != 0: continue
        usage += (1 << ((9 * (3-pt.lv)) + 12))
@@ -279,7 +280,7 @@ class SMICtx:
      device_line = [f"{bold(dev.pcibus)} {trim(self.lspci[dev.pcibus[5:]], col_size - 20)}"] + [pad("", col_size)]
      activity_line = [f"GFX Activity {draw_bar(self.get_gfx_activity(dev, metrics) / 100, activity_line_width)}"] \
                    + [f"MEM Activity {draw_bar(self.get_mem_activity(dev, metrics) / 100, activity_line_width)}"] \
-                    + [f"MEM Usage    {draw_bar(mem_used / mem_total, activity_line_width, opt_text=mem_fmt)}"] \
+                    + [f"MEM Usage    {draw_bar((mem_used / mem_total) / 100, activity_line_width, opt_text=mem_fmt)}"] \

      temps_data, temps_data_compact = self.get_temps(dev, metrics), self.get_temps(dev, metrics, compact=True)
      temps_table = ["=== Temps (°C) ==="] + [f"{name:<16}: {color_temp(val)}" for name, val in temps_data.items()]
@@ -1,18 +1,12 @@
 #!/usr/bin/env python3

-import os
 from tinygrad.helpers import Context
-from tinygrad.runtime.support.system import System, PCIDevice
-from tinygrad.runtime.support.hcq import FileIOInterface
+from tinygrad.runtime.support.system import System, PCIDevice, PCIDevImplBase
 from tinygrad.runtime.support.am.amdev import AMDev

 if __name__ == "__main__":
  gpus = System.pci_scan_bus(0x1002, [(0xffff, [0x74a1, 0x75a0])])
-  for gpu in gpus:
-    drv_path = f"/sys/bus/pci/devices/{gpu}/driver"
-    if FileIOInterface.exists(drv_path) and os.path.basename(os.readlink(drv_path)) == "amdgpu":
-      raise RuntimeError(f"amdgpu is bound to {gpu}. Stopping...")
-  pcidevs = [PCIDevice("AM", gpu) for gpu in gpus]
+  pcidevs = [PCIDevice(f"reset:{gpu}", gpu, bars=[0, 2, 5]) for gpu in gpus]
  amdevs = []
  with Context(DEBUG=2):
    for pcidev in pcidevs:
@@ -7,8 +7,8 @@ class GFXFake:
  def __init__(self): self.xccs = 8

 class AMDFake(AMDev):
-  def __init__(self, pci_dev):
-    self.pci_dev, self.devfmt = pci_dev, pci_dev.pcibus
+  def __init__(self, pci_dev, dma_regions=None):
+    self.pci_dev, self.devfmt, self.dma_regions = pci_dev, pci_dev.pcibus, dma_regions
    self.vram, self.doorbell64, self.mmio = self.pci_dev.map_bar(0), self.pci_dev.map_bar(2, fmt='Q'), self.pci_dev.map_bar(5, fmt='I')
    self._run_discovery()
    self._build_regs()
@@ -19,9 +19,8 @@ amdev = importlib.import_module("tinygrad.runtime.support.am.amdev")
 amdev.AMDev = AMDFake
 from tinygrad.runtime.ops_amd import PCIIface

-def parse_amdgpu_logs(log_content, register_names=None, register_objects=None, *, only_xcc0: bool = False):
+def parse_amdgpu_logs(log_content, register_names=None, *, only_xcc0: bool = False):
  register_map = register_names or {}
-  register_objs = register_objects or {}

  def replace_register(match):
    reg = match.group(1)
@@ -38,28 +37,6 @@ def parse_amdgpu_logs(log_content, register_names=None, register_objects=None, *
  # remove timing prefix
  processed_log = re.sub(r'^\[\s*\d+(?:\.\d+)?\]\s*', '', processed_log, flags=re.MULTILINE)

-  # decode register values into field dicts
-  def decode_value(match):
-    reg_name = match.group(1)
-    xcc_part = match.group(2)  # "xcc=0 " or ""
-    val_str = match.group(3)
-    val = int(val_str, 16)
-
-    reg_obj = register_objs.get(reg_name)
-    if reg_obj is not None and reg_obj.fields:
-      fields = reg_obj.decode(val)
-      # show raw for unaccounted bits
-      accounted = 0
-      for name, (start, end) in reg_obj.fields.items():
-        accounted |= (((1 << (end - start + 1)) - 1) << start)
-      unaccounted = val & ~accounted
-      parts = {k: v for k, v in fields.items() if v != 0}
-      if unaccounted: parts['_raw_unaccounted'] = hex(unaccounted)
-      return f"register {reg_name}, {xcc_part}with value {val_str} {parts}"
-    return match.group(0)
-
-  processed_log = re.sub(r'register (reg\w+), ((?:xcc=\d+ )?)with value (0x[0-9a-fA-F]+)', decode_value, processed_log)
-
  # keep only xcc=0 lines (but keep lines with no xcc at all)
  if only_xcc0:
    kept = []
@@ -73,18 +50,16 @@ def main():
  only_xcc0 = bool(getenv("ONLY_XCC0", 0))

  reg_names = {}
-  reg_objs = {}
  dev = PCIIface(None, 0)
  for x, y in dev.dev_impl.__dict__.items():
    if isinstance(y, AMRegister):
      for xcc, addr in y.addr.items():
        reg_names[addr] = f"{x}, xcc={xcc}"
-      reg_objs[x] = y

  with open(sys.argv[1], 'r') as f:
    log_content = f.read()

-  processed_log = parse_amdgpu_logs(log_content, reg_names, reg_objs, only_xcc0=only_xcc0)
+  processed_log = parse_amdgpu_logs(log_content, reg_names, only_xcc0=only_xcc0)

  with open(sys.argv[2], 'w') as f:
    f.write(processed_log)
@@ -0,0 +1,136 @@
+import os, sys, struct
+sys.path.append(os.getcwd())
+# PROFILE=1 to use
+#os.environ["PROFILE"] = "1"
+os.environ["SQTT"] = "1"
+os.environ["SQTT_ITRACE_SE_MASK"] = "1"
+os.environ["SQTT_LIMIT_SE"] = "1"
+import xml.etree.ElementTree as ET
+
+from tinygrad import nn, Tensor, Device
+from tinygrad.helpers import get_single_element
+from tinygrad.runtime.support.elf import elf_loader
+from tinygrad.runtime.ops_amd import ProfileSQTTEvent
+from extra.sqtt.attempt_sqtt_parse import parse_sqtt_print_packets
+
+def disassemble(text, root:ET.Element):
+  i = 0
+  while i < len(text):
+    ins = struct.unpack("I", text[i:i+4])[0]
+
+    # 1. Get the encoding
+    did_match = False
+    for enc_el in root.findall("./ISA/Encodings/Encoding"):
+      mask = enc_el.findtext("EncodingIdentifierMask")
+      assert len(mask)%32 == 0
+      bit_mask = int(mask, 2)
+      iden = [int(x.text, 2) for x in enc_el.find("EncodingIdentifiers").findall("EncodingIdentifier")]
+      for ide in iden:
+        if ins&bit_mask == ide:
+          did_match = True
+          break
+      if did_match: break
+    if not did_match: raise RuntimeError(f"unknown instruction {ins:08X}")
+    if len(mask) >= 64: ins = (struct.unpack("I", text[i+4:i+8])[0]<<32) | ins
+    if len(mask) >= 96: ins = (struct.unpack("I", text[i+8:i+12])[0]<<64) | ins
+    encoding_name = enc_el.findtext("EncodingName")
+
+    #print(ET.tostring(enc_el).decode())
+
+    # 2. Parse the Fields for this Encoding
+    field_data = {}
+    for field in enc_el.findall("MicrocodeFormat/BitMap/Field"):
+      # Fields can be split into multiple ranges (RangeCount > 1)
+      ranges = sorted(field.findall("BitLayout/Range"), key=lambda x: int(x.attrib.get('Order')))
+      val = 0
+      current_shift = 0
+      for rng in ranges:
+        width = int(rng.find("BitCount").text)
+        chunk = (ins >> int(rng.find("BitOffset").text)) & ((1 << width) - 1)
+        val |= (chunk << current_shift)
+        current_shift += width
+      field_data[field.find("FieldName").text] = val
+    # this is already used
+    del field_data["ENCODING"]
+
+    # 3. Extract the instruction
+    did_match = False
+    for ins_el in root.findall("./ISA/Instructions/Instruction"):
+      ins_name = ins_el.findtext("InstructionName")
+      for ins_enc in ins_el.findall("InstructionEncodings/InstructionEncoding"):
+        if ins_enc.findtext("EncodingName") == encoding_name:
+          opcode = int(ins_enc.findtext("Opcode"))
+          if "OP" in field_data and opcode == field_data["OP"]:
+            did_match = True
+            del field_data["OP"]
+            break
+        if did_match: break
+      if did_match: break
+
+    #print(ET.tostring(ins_enc).decode())
+    #print()
+    #print(field_data)
+    if not did_match:
+      print(f"{i:4X} : {ins:16x} -- {encoding_name}")
+    elif did_match:
+      params = []
+      #print(ET.tostring(ins_el).decode())
+
+      # 4. Extract the opcodes
+      for op_ins in ins_enc.findall("Operands/Operand"):
+        op_type = op_ins.findtext("OperandType")
+        op_size = op_ins.findtext("OperandSize")
+        op_fmt = op_ins.findtext("DataFormatName")
+        op_field_name = op_ins.findtext("FieldName")
+        if op_field_name is None: continue
+        assert op_field_name in field_data
+        # loop through operands for compare
+        for op_el in root.findall("./ISA/OperandTypes/OperandType"):
+          test_op_type = op_el.findtext("OperandTypeName")
+          val_dict = {}
+          for op_val in op_el.findall("OperandPredefinedValues/PredefinedValue"):
+            val_dict[int(op_val.findtext("Value"))] = op_val.findtext("Name")
+          if op_type == test_op_type:
+            if field_data[op_field_name] in val_dict:
+              print(op_type, op_size, op_fmt)
+              params.append(val_dict[field_data[op_field_name]])
+            else:
+              params.append(f"{op_type}({field_data[op_field_name]})")
+            del field_data[op_field_name]
+            #print(op_type, op_size, op_fmt, op_el, op_field_name,
+            #      field_data[op_field_name],
+            #      val_dict.get(field_data[op_field_name], "<UNK>"))
+            #print(ET.tostring(op_el).decode())
+
+      print(f"{i:4X} : {ins:16x} -- {ins_name.lower()} {', '.join(params)}", field_data)
+
+    # advance
+    i += len(mask) // 8
+
+  #print(ET.tostring(root).decode())
+
+if __name__ == "__main__":
+  # human readable manual at https://docs.amd.com/v/u/en-US/rdna35_instruction_set_architecture
+  fns = nn.state.zip_extract(Tensor.from_url("https://gpuopen.com/download/machine-readable-isa/latest/"))
+  xml_str = fns['amdgpu_isa_rdna3_5.xml'].to("CPU").data()
+  with open("/tmp/rdna35.xml", "wb") as f: f.write(bytes(xml_str))
+  root = ET.fromstring(xml_str)
+
+  a = Tensor.empty(16)+1
+  for ei in a.schedule():
+    ei.lower()
+    # get text
+    _, hdr, _ = elf_loader(ei.prg.lib)
+    text = get_single_element([x for x in hdr if x.name==".text"]).content
+
+    # llvm disassembler
+    Device["AMD"].compiler.disassemble(ei.prg.lib)
+
+    # run program
+    ei.run()
+
+  sqtt_events = [e for e in Device["AMD"].profile_events if isinstance(e, ProfileSQTTEvent)]
+  for e in sqtt_events[0:1]: # only the first SE
+    parse_sqtt_print_packets(e.blob)
+
+  disassemble(text[:0x40], root)
@@ -0,0 +1,15 @@
+from tinygrad import Tensor, nn
+import xml.etree.ElementTree as ET
+
+if __name__ == "__main__":
+  # human readable manual at https://docs.amd.com/v/u/en-US/rdna35_instruction_set_architecture
+  fns = nn.state.zip_extract(Tensor.from_url("https://gpuopen.com/download/machine-readable-isa/latest/"))
+  xml_str = fns['amdgpu_isa_rdna3_5.xml'].to("CPU").data()
+  root = ET.fromstring(xml_str)
+
+  for op_el in root.findall("./ISA/OperandTypes/OperandType"):
+    op_name = op_el.findtext("OperandTypeName")
+    val_dict = {}
+    for op_val in op_el.findall("OperandPredefinedValues/PredefinedValue"):
+      val_dict[int(op_val.findtext("Value"))] = op_val.findtext("Name")
+    print(op_name, val_dict)
@@ -0,0 +1,4 @@
+*.deb
+build
+src
+sniffer/sniff.so
@@ -0,0 +1,20 @@
+Built ROCT-Thunk-Interface (hsakmt)
+  hsakmt-roct-dev_5.4.4.99999-local_amd64.deb
+  note: installs to /opt/rocm
+Built ROCm-Device-Libs
+  Works with ROCM_PATH=/home/tiny/build/ROCm-Device-Libs/build/dist
+  rocm-device-libs_1.0.0.99999-local_amd64.deb
+Built ROCm-CompilerSupport (amd_comgr)
+  no deb, sudo make install to /usr/local
+Built ROCR-Runtime
+  hsa-rocr_1.8.0-local_amd64.deb
+  hsa-rocr-dev_1.8.0-local_amd64.deb
+Built ROCm-OpenCL-Runtime
+  rocm-ocl-icd_2.0.0-local_amd64.deb
+  ISSUE: these depend on "comgr"
+  rocm-opencl_2.0.0-local_amd64.deb
+  rocm-opencl-dev_2.0.0-local_amd64.deb
+  Did sudo make install
+
+
+
@@ -0,0 +1,41 @@
+# run two "rocm-bandwidth-test" in a loop
+# amdgpu-6.0.5-1581431.20.04
+# fixed in kernel 6.2.14
+
+[   72.153646] RIP: 0010:pm_send_runlist+0x4a/0x630 [amdgpu]
+[   72.153815] Code: 30 65 48 8b 04 25 28 00 00 00 48 89 45 d0 31 c0 80 fb 01 0f 87 aa 9d 49 00 83 e3 01 0f 85 1c 05 00 00 49 8b 3f b8 01 00 00 00 <48> 8b 97 30 01 00 00 44 8b b7 6c 01 00 00 8b 9f 70 01 00 00 8b 8a
+[   72.153900] RSP: 0018:ffffb48445c03c30 EFLAGS: 00010246
+[   72.153928] RAX: 0000000000000001 RBX: 0000000000000000 RCX: 0000000000000000
+[   72.153962] RDX: 000000000000007b RSI: ffff9395e1562558 RDI: 0000000000000000
+[   72.153996] RBP: ffffb48445c03cb8 R08: 0000000000000000 R09: 0000000000000001
+[   72.154030] R10: ffff9395c900d840 R11: 0000000000000000 R12: 0000000000000000
+[   72.154065] R13: ffff9395c9e00400 R14: 0000000000000001 R15: ffff9395e15624e0
+[   72.154099] FS:  00007f345c6463c0(0000) GS:ffff93a4aee80000(0000) knlGS:0000000000000000
+[   72.154137] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
+[   72.154165] CR2: 0000000000000130 CR3: 0000000112840000 CR4: 0000000000750ee0
+[   72.154201] PKRU: 55555554
+[   72.154215] Call Trace:
+[   72.154230]  <TASK>
+[   72.154244]  map_queues_cpsch+0x75/0xc0 [amdgpu]
+[   72.154365]  debug_map_and_unlock+0x51/0x90 [amdgpu]
+[   72.154480]  debug_refresh_runlist+0x1f/0x30 [amdgpu]
+[   72.154591]  kfd_dbg_runtime_disable+0x13c/0x240 [amdgpu]
+[   72.154705]  kfd_ioctl_dbg_set_debug_trap+0x69d/0x8b0 [amdgpu]
+[   72.154820]  kfd_ioctl+0x24a/0x5b0 [amdgpu]
+[   72.154925]  ? kfd_ioctl_create_queue+0x770/0x770 [amdgpu]
+[   72.155035]  ? syscall_exit_to_user_mode+0x27/0x50
+[   72.155061]  ? exit_to_user_mode_prepare+0x3d/0x1c0
+[   72.155088]  __x64_sys_ioctl+0x95/0xd0
+[   72.155109]  do_syscall_64+0x5c/0xc0
+[   72.155128]  ? syscall_exit_to_user_mode+0x27/0x50
+[   72.155151]  ? do_syscall_64+0x69/0xc0
+[   72.155172]  entry_SYSCALL_64_after_hwframe+0x61/0xcb
+[   72.155198] RIP: 0033:0x7f345c7f63ab
+[   72.155218] Code: 0f 1e fa 48 8b 05 e5 7a 0d 00 64 c7 00 26 00 00 00 48 c7 c0 ff ff ff ff c3 66 0f 1f 44 00 00 f3 0f 1e fa b8 10 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d b5 7a 0d 00 f7 d8 64 89 01 48
+[   72.155301] RSP: 002b:00007ffc97cc89f8 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
+[   72.155339] RAX: ffffffffffffffda RBX: 00007ffc97cc8a30 RCX: 00007f345c7f63ab
+[   72.155375] RDX: 00007ffc97cc8a30 RSI: 00000000c0284b82 RDI: 0000000000000003
+[   72.155411] RBP: 00000000c0284b82 R08: 0000000000000000 R09: 0000000000000000
+[   72.155447] R10: 00007f345cd4ddb0 R11: 0000000000000246 R12: 00007ffc97cc8a30
+[   72.155481] R13: 0000000000000003 R14: 00007ffc97cc8d20 R15: 0000000000000000
+[   72.155517]  </TASK>
@@ -0,0 +1,41 @@
+# run two tinygrad matrix example in a loop
+# amdgpu-6.0.5-1581431.20.04
+# NOT fixed in kernel 6.2.14
+
+[  553.016624] gmc_v11_0_process_interrupt: 30 callbacks suppressed
+[  553.016631] amdgpu 0000:0b:00.0: amdgpu: [gfxhub] page fault (src_id:0 ring:24 vmid:9 pasid:32770, for process python3 pid 10001 thread python3 pid 10001)
+[  553.016790] amdgpu 0000:0b:00.0: amdgpu:   in page starting at address 0x00007f0000000000 from client 10
+[  553.016892] amdgpu 0000:0b:00.0: amdgpu: GCVM_L2_PROTECTION_FAULT_STATUS:0x00901A30
+[  553.016974] amdgpu 0000:0b:00.0: amdgpu:      Faulty UTCL2 client ID: SDMA0 (0xd)
+[  553.017051] amdgpu 0000:0b:00.0: amdgpu:      MORE_FAULTS: 0x0
+[  553.017111] amdgpu 0000:0b:00.0: amdgpu:      WALKER_ERROR: 0x0
+[  553.017173] amdgpu 0000:0b:00.0: amdgpu:      PERMISSION_FAULTS: 0x3
+[  553.017238] amdgpu 0000:0b:00.0: amdgpu:      MAPPING_ERROR: 0x0
+[  553.017300] amdgpu 0000:0b:00.0: amdgpu:      RW: 0x0
+[  553.123921] [drm:mes_v11_0_submit_pkt_and_poll_completion.constprop.0 [amdgpu]] *ERROR* MES failed to response msg=2
+[  553.124153] amdgpu: failed to add hardware queue to MES, doorbell=0x1a16
+[  553.124195] amdgpu: MES might be in unrecoverable state, issue a GPU reset
+[  553.124237] amdgpu: Failed to restore queue 2
+[  553.124266] amdgpu: Failed to restore process queues
+[  553.124270] amdgpu: Failed to evict queue 3
+[  553.124297] amdgpu: amdgpu_amdkfd_restore_userptr_worker: Failed to resume KFD
+
+# alternative crash in kernel 6.2.14
+
+[  151.097948] gmc_v11_0_process_interrupt: 30 callbacks suppressed
+[  151.097953] amdgpu 0000:0b:00.0: amdgpu: [gfxhub] page fault (src_id:0 ring:24 vmid:8 pasid:32771, for process python3 pid 7525 thread python3 pid 7525)
+[  151.097993] amdgpu 0000:0b:00.0: amdgpu:   in page starting at address 0x00007f0000000000 from client 10
+[  151.098008] amdgpu 0000:0b:00.0: amdgpu: GCVM_L2_PROTECTION_FAULT_STATUS:0x00801A30
+[  151.098020] amdgpu 0000:0b:00.0: amdgpu:      Faulty UTCL2 client ID: SDMA0 (0xd)
+[  151.098032] amdgpu 0000:0b:00.0: amdgpu:      MORE_FAULTS: 0x0
+[  151.098042] amdgpu 0000:0b:00.0: amdgpu:      WALKER_ERROR: 0x0
+[  151.098052] amdgpu 0000:0b:00.0: amdgpu:      PERMISSION_FAULTS: 0x3
+[  151.098062] amdgpu 0000:0b:00.0: amdgpu:      MAPPING_ERROR: 0x0
+[  151.098071] amdgpu 0000:0b:00.0: amdgpu:      RW: 0x0
+[  151.209517] [drm:mes_v11_0_submit_pkt_and_poll_completion.constprop.0 [amdgpu]] *ERROR* MES failed to response msg=2
+[  151.209724] amdgpu: failed to add hardware queue to MES, doorbell=0x1002
+[  151.209734] amdgpu: MES might be in unrecoverable state, issue a GPU reset
+[  151.209743] amdgpu: Failed to restore queue 1
+[  151.209751] amdgpu: Failed to restore process queues
+[  151.209759] amdgpu: amdgpu_amdkfd_restore_userptr_worker: Failed to resume KFD
+[  151.209858] amdgpu 0000:0b:00.0: amdgpu: GPU reset begin!
@@ -0,0 +1,20 @@
+# two tinygrad + two bandwidth test
+# RDNA2, driver 6.0.5
+# recovered from this!
+
+[  136.971209] gmc_v10_0_process_interrupt: 39 callbacks suppressed
+[  136.971218] amdgpu 0000:0b:00.0: amdgpu: [gfxhub] page fault (src_id:0 ring:24 vmid:11 pasid:32773, for process rocm-bandwidth- pid 20281 thread rocm-bandwidth- pid 20281)
+[  136.971228] amdgpu 0000:0b:00.0: amdgpu:   in page starting at address 0x00007f5c2b800000 from client 0x1b (UTCL2)
+[  136.971232] amdgpu 0000:0b:00.0: amdgpu: GCVM_L2_PROTECTION_FAULT_STATUS:0x00B01A31
+[  136.971233] amdgpu 0000:0b:00.0: amdgpu:      Faulty UTCL2 client ID: SDMA0 (0xd)
+[  136.971235] amdgpu 0000:0b:00.0: amdgpu:      MORE_FAULTS: 0x1
+[  136.971236] amdgpu 0000:0b:00.0: amdgpu:      WALKER_ERROR: 0x0
+[  136.971236] amdgpu 0000:0b:00.0: amdgpu:      PERMISSION_FAULTS: 0x3
+[  136.971237] amdgpu 0000:0b:00.0: amdgpu:      MAPPING_ERROR: 0x0
+[  136.971238] amdgpu 0000:0b:00.0: amdgpu:      RW: 0x0
+...
+[  136.993979] amdgpu 0000:0b:00.0: amdgpu: IH ring buffer overflow (0x000BE5A0, 0x0003C480, 0x0003E5C0)
+[  138.209072] amdgpu 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x001a address=0x7c00004000 flags=0x0000]
+[  138.209078] amdgpu 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x001a address=0x7c00004d80 flags=0x0000]
+[  138.209081] amdgpu 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x001a address=0x7c00005000 flags=0x0000]
+[  138.209084] amdgpu 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x001a address=0x7c00005d80 flags=0x0000]
@@ -0,0 +1,33 @@
+# ROCK-Kernel-Driver 0b579de9622f5c93021dcb7927d13926313740a2
+# non fatal "crash"
+
+[  127.418045] ------------[ cut here ]------------
+[  127.418046] User pages unexpectedly invalid
+[  127.418056] WARNING: CPU: 16 PID: 260 at drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c:3000 amdgpu_amdkfd_restore_userptr_worker+0x4d9/0x500 [amdgpu]
+[  127.418235] Modules linked in: rfcomm cmac algif_hash algif_skcipher af_alg bnep nls_iso8859_1 iwlmvm mac80211 intel_rapl_msr intel_rapl_common edac_mce_amd snd_hda_codec_realtek snd_hda_codec_generic snd_hda_codec_hdmi kvm_amd binfmt_misc snd_hda_intel snd_intel_dspcfg kvm libarc4 snd_intel_sdw_acpi snd_hda_codec btusb iwlwifi btrtl snd_hda_core btbcm btintel irqbypass btmtk snd_hwdep crct10dif_pclmul snd_pcm polyval_clmulni bluetooth snd_seq_midi snd_seq_midi_event snd_rawmidi snd_seq polyval_generic cfg80211 ghash_clmulni_intel eeepc_wmi snd_seq_device snd_timer aesni_intel asus_wmi ecdh_generic snd platform_profile crypto_simd ledtrig_audio cryptd ecc ccp soundcore sparse_keymap rapl k10temp wmi_bmof mac_hid sch_fq_codel msr parport_pc ppdev lp parport ramoops pstore_blk efi_pstore reed_solomon pstore_zone ip_tables x_tables autofs4 amdgpu hid_generic usbhid hid i2c_algo_bit drm_ttm_helper ttm video iommu_v2 drm_buddy gpu_sched drm_display_helper drm_kms_helper syscopyarea
+[  127.418276]  sysfillrect sysimgblt fb_sys_fops drm nvme nvme_core cec r8169 ahci crc32_pclmul rc_core i2c_piix4 xhci_pci libahci nvme_common xhci_pci_renesas realtek wmi
+[  127.418284] CPU: 16 PID: 260 Comm: kworker/16:1 Tainted: G        W          6.0.0 #4
+[  127.418286] Hardware name: System manufacturer System Product Name/TUF GAMING X570-PLUS (WI-FI), BIOS 3603 03/20/2021
+[  127.418287] Workqueue: events amdgpu_amdkfd_restore_userptr_worker [amdgpu]
+[  127.418455] RIP: 0010:amdgpu_amdkfd_restore_userptr_worker+0x4d9/0x500 [amdgpu]
+[  127.418601] Code: ff e8 2b 8a 96 d1 e9 66 fe ff ff 48 c7 c7 40 4f f5 c0 e8 56 7b 8a d1 0f 0b e9 2e ff ff ff 48 c7 c7 d8 d0 ed c0 e8 43 7b 8a d1 <0f> 0b e9 0a fe ff ff 4c 89 ef e8 f8 89 96 d1 e9 cb fd ff ff e8 ce
+[  127.418603] RSP: 0018:ffffb36740a83dc8 EFLAGS: 00010282
+[  127.418604] RAX: 0000000000000000 RBX: ffff9d159ee9df30 RCX: 0000000000000027
+[  127.418605] RDX: 0000000000000027 RSI: ffffb36740a83c88 RDI: ffff9d242a220568
+[  127.418606] RBP: ffffb36740a83e58 R08: ffff9d242a220560 R09: 0000000000000001
+[  127.418607] R10: 0000000000000001 R11: 0000000000000020 R12: ffff9d159ee9df98
+[  127.418607] R13: ffff9d159ee9df70 R14: ffff9d159ee9dee0 R15: ffff9d159ee9dee0
+[  127.418608] FS:  0000000000000000(0000) GS:ffff9d242a200000(0000) knlGS:0000000000000000
+[  127.418609] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
+[  127.418610] CR2: 00007fd5d4715000 CR3: 0000000120ffe000 CR4: 0000000000750ee0
+[  127.418611] PKRU: 55555554
+[  127.418611] Call Trace:
+[  127.418612]  <TASK>
+[  127.418613]  process_one_work+0x21f/0x3f0
+[  127.418615]  worker_thread+0x4a/0x3c0
+[  127.418617]  ? process_one_work+0x3f0/0x3f0
+[  127.418618]  kthread+0xf0/0x120
+[  127.418619]  ? kthread_complete_and_exit+0x20/0x20
+[  127.418620]  ret_from_fork+0x22/0x30
+[  127.418622]  </TASK>
+[  127.418623] ---[ end trace 0000000000000000 ]---
@@ -0,0 +1,80 @@
+import numpy as np
+import pathlib
+from hexdump import hexdump
+from tinygrad.helpers import colored
+from extra.helpers import enable_early_exec
+early_exec = enable_early_exec()
+
+from tinygrad.runtime.ops_cl import CLProgram, CLBuffer, ROCM_LLVM_PATH
+
+ENABLE_NON_ASM = False
+
+WMMA = True
+DUAL_ALU = True
+F32 = True
+
+if ENABLE_NON_ASM:
+  buf = CLBuffer.fromCPU(np.zeros(10, np.float32))
+  prg_empty = CLProgram("code", "__kernel void code(__global float *a) { a[0] = 1; }")
+  asm_real = prg_empty.binary()
+  with open("/tmp/cc.elf", "wb") as f:
+    f.write(asm_real)
+  prg_empty([1], [1], buf, wait=True)
+  print(buf.toCPU())
+
+print(colored("creating CLBuffer", "green"))
+buf = CLBuffer.fromCPU(np.zeros(10, np.float32))
+code = open(pathlib.Path(__file__).parent / "prog.s", "r").read()
+
+gen = []
+FLOPS = 0
+MAX_REG = 251
+for j in range(1):
+  if WMMA:
+    KY, KX = 4, 4
+    for y in range(KY):
+      for x in range(KX):
+        c = (y*KX+x)*8
+        a = (KY*KX*8) + y*8
+        b = (KY*KX*8) + (KY*8) + x*8
+        gen.append(f"v_wmma_f32_16x16x16_f16 v[{c}:{c+7}], v[{a}:{a+7}], v[{b}:{b+7}], v[{c}:{c+7}]")
+        FLOPS += 16*8*2
+  else:
+    for i in range(0, MAX_REG, 6):
+      if DUAL_ALU:
+        if F32:
+          gen.append(f"v_dual_fmac_f32 v{i+0}, v{i+1}, v{i+2} :: v_dual_fmac_f32 v{i+3}, v{i+4}, v{i+5}")
+          FLOPS += 4
+        else:
+          gen.append(f"v_dual_dot2acc_f32_f16 v{i+0}, v{i+1}, v{i+2} :: v_dual_dot2acc_f32_f16 v{i+3}, v{i+4}, v{i+5}")
+          FLOPS += 8
+      else:
+        assert F32
+        gen.append(f"v_fmac_f32 v{i+0}, v{i+1}, v{i+2}")
+        gen.append(f"v_fmac_f32 v{i+3}, v{i+4}, v{i+5}")
+code = code.replace("// FLOPS", '\n'.join(gen))
+print(code)
+
+
+# fix: COMGR failed to get code object ISA name. set triple to 'amdgcn-amd-amdhsa'
+
+object = early_exec(([ROCM_LLVM_PATH / "llvm-mc", '--arch=amdgcn', '--mcpu=gfx1100', '--triple=amdgcn-amd-amdhsa', '--filetype=obj', '-'], code.encode("utf-8")))
+asm = early_exec(([ROCM_LLVM_PATH / "ld.lld", "/dev/stdin", "-o", "/dev/stdout", "--pie"], object))
+
+with open("/tmp/cc2.o", "wb") as f:
+  f.write(object)
+with open("/tmp/cc2.elf", "wb") as f:
+  f.write(asm)
+
+print(colored("creating CLProgram", "green"))
+prg = CLProgram("code", asm)
+
+print(colored("running program", "green"))
+G = 512
+FLOPS *= 100000*G*G  # loop * global_size
+for i in range(3):
+  tm = prg(buf, global_size=[G//256, G, 1], local_size=[256, 1, 1], wait=True)
+  print(f"ran in {tm*1e3:.2f} ms, {FLOPS/(tm*1e9):.2f} GFLOPS")
+
+print(colored("transferring buffer", "green"))
+print(buf.toCPU())
@@ -0,0 +1,80 @@
+.global _start
+_start:
+.rodata
+.align 0x10
+.global code.kd
+.type code.kd,STT_OBJECT
+# amd_kernel_code_t (must be at 0x440 for kernel_code_entry_byte_offset to be right)
+code.kd:
+# amd_kernel_..., amd_machine_...
+.long 0,0,0,0
+# kernel_code_entry_byte_offset, kernel_code_prefetch_byte_offset
+.long 0x00000bc0,0x00000000,0x00000000,0x00000000
+# kernel_code_prefetch_byte_size, max_scratch_backing_memory_byte_size
+.long 0,0,0,0
+# compute_pgm_rsrc1, compute_pgm_rsrc2, kernel_code_properties, workitem_private_segment_byte_size
+.long 0x60af0000,0x0000009e,0x00000408,0x00000000
+# compute_pgm_rsrc1 |= AMD_COMPUTE_PGM_RSRC_ONE_FLOAT_DENORM_MODE_32 | AMD_COMPUTE_PGM_RSRC_ONE_FLOAT_DENORM_MODE_16_64
+# compute_pgm_rsrc1 |= AMD_COMPUTE_PGM_RSRC_ONE_ENABLE_DX10_CLAMP | AMD_COMPUTE_PGM_RSRC_ONE_ENABLE_IEEE_MODE
+# compute_pgm_rsrc2 |= AMD_COMPUTE_PGM_RSRC_TWO_USER_SGPR_COUNT = 0xF
+# compute_pgm_rsrc2 |= AMD_COMPUTE_PGM_RSRC_TWO_ENABLE_SGPR_WORKGROUP_ID_X
+# kernel_code_properties |= AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_KERNARG_SEGMENT_PTR = 1
+# kernel_code_properties |= AMD_KERNEL_CODE_PROPERTIES_RESERVED1 = 1
+.text
+.global code
+.type code,STT_FUNC
+code:
+# https://llvm.org/docs/AMDGPUUsage.html#initial-kernel-execution-state
+# s[0:1] contains the kernarg_address
+# TODO: can we use s[2:3] if this was really a wave since we only alloced 2 SGPRs?
+s_load_b64 s[2:3], s[0:1], null
+
+s_mov_b32 s8, 0
+loop:
+s_addk_i32 s8, 1
+s_cmp_eq_u32 s8, 100000
+// FLOPS
+s_cbranch_scc0 loop
+
+# wait for the s_load_b64
+s_waitcnt lgkmcnt(0)
+
+v_dual_mov_b32 v0, 4 :: v_dual_mov_b32 v1, 2.0
+global_store_b32 v0, v1, s[2:3]
+
+# Deallocate all VGPRs for this wave. Use only when next instruction is S_ENDPGM.
+s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+s_endpgm
+s_code_end
+
+.amdgpu_metadata
+amdhsa.kernels:
+  - .args:
+      - .address_space:  global
+        .name:           a
+        .offset:         0
+        .size:           8
+        .type_name:      'float*'
+        .value_kind:     global_buffer
+    .group_segment_fixed_size: 0
+    .kernarg_segment_align: 8
+    .kernarg_segment_size: 8
+    .language:       OpenCL C
+    .language_version:
+      - 1
+      - 2
+    .max_flat_workgroup_size: 256
+    .name:           code
+    .private_segment_fixed_size: 0
+    .sgpr_count:     2
+    .sgpr_spill_count: 0
+    .symbol:         code.kd
+    .uses_dynamic_stack: false
+    .vgpr_count:     256
+    .vgpr_spill_count: 0
+    .wavefront_size: 32
+amdhsa.target:   amdgcn-amd-amdhsa--gfx1100
+amdhsa.version:
+  - 1
+  - 2
+.end_amdgpu_metadata
@@ -0,0 +1,11 @@
+#!/bin/bash
+mkdir -p src
+cd src
+git clone https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface.git -b rocm-5.5.0
+git clone https://github.com/RadeonOpenCompute/ROCm-Device-Libs.git -b rocm-5.5.0
+git clone https://github.com/RadeonOpenCompute/llvm-project.git -b rocm-5.5.0 --depth 1
+git clone https://github.com/RadeonOpenCompute/ROCR-Runtime.git -b rocm-5.5.0
+git clone https://github.com/ROCm-Developer-Tools/ROCclr.git -b rocm-5.5.0
+git clone https://github.com/RadeonOpenCompute/ROCm-CompilerSupport.git -b rocm-5.5.0
+git clone https://github.com/RadeonOpenCompute/ROCm-OpenCL-Runtime.git -b rocm-5.5.0
+cd ../
@@ -0,0 +1,69 @@
+#!/bin/bash
+mkdir -p build/debs
+cd build
+
+# ROCT-Thunk-Interface (hsakmt)
+if [ ! -f debs/hsakmt-roct-dev_5.5.0.99999-local_amd64.deb ]
+then
+  mkdir -p ROCT-Thunk-Interface
+  cd ROCT-Thunk-Interface
+  cmake ../../src/ROCT-Thunk-Interface
+  make -j32 package
+  cp hsakmt-roct-dev_5.5.0.99999-local_amd64.deb ../debs
+  cd ../
+fi
+
+
+# build custom LLVM
+if [ ! -f llvm-project/bin/clang ]
+then
+  mkdir -p llvm-project
+  cd llvm-project
+  cmake -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_PROJECTS="llvm;clang;lld" -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" ../../src/llvm-project/llvm
+  make -j32
+  cd ..
+fi
+
+# use custom LLVM
+export PATH="$PWD/llvm-project/bin:$PATH"
+
+# ROCm-Device-Libs
+if [ ! -f debs/rocm-device-libs_1.0.0.99999-local_amd64.deb ]
+then
+  mkdir -p ROCm-Device-Libs
+  cd ROCm-Device-Libs
+  cmake ../../src/ROCm-Device-Libs
+  make -j32 package
+  cp rocm-device-libs_1.0.0.99999-local_amd64.deb ../debs
+  cd ../
+fi
+
+# ROCR-Runtime
+if [ ! -f debs/hsa-rocr_1.8.0-local_amd64.deb ]
+then
+  mkdir -p ROCR-Runtime
+  cd ROCR-Runtime
+  cmake ../../src/ROCR-Runtime/src
+  make -j32 package
+  cp hsa-rocr_1.8.0-local_amd64.deb ../debs
+  cp hsa-rocr-dev_1.8.0-local_amd64.deb ../debs
+  cd ../
+fi
+
+# ROCm-OpenCL-Runtime (needs ROCclr)
+if [ ! -f debs/rocm-opencl_2.0.0-local_amd64.deb ]
+then
+  mkdir -p ROCm-OpenCL-Runtime
+  cd ROCm-OpenCL-Runtime
+  cmake ../../src/ROCm-OpenCL-Runtime
+  make -j32 package
+  cp rocm-opencl_2.0.0-local_amd64.deb ../debs
+  cp rocm-opencl-dev_2.0.0-local_amd64.deb ../debs
+  cp rocm-ocl-icd_2.0.0-local_amd64.deb ../debs
+fi
+
+# ROCm-CompilerSupport (broken)
+#mkdir -p ROCm-CompilerSupport
+#cd ROCm-CompilerSupport
+#cmake ../../src/ROCm-CompilerSupport/lib/comgr
+#make -j32
@@ -0,0 +1,14 @@
+#!/bin/bash
+rm amdgpu-install_5.5.50500-1_all.deb
+wget https://repo.radeon.com/amdgpu-install/5.5/ubuntu/$(lsb_release -cs)/amdgpu-install_5.5.50500-1_all.deb
+sudo dpkg -i amdgpu-install_5.5.50500-1_all.deb
+sudo apt-get update
+
+# kernel driver
+sudo apt-get install amdgpu-dkms
+
+# for opencl
+sudo apt-get install rocm-opencl-runtime
+
+# for HIP
+sudo apt-get install hip-runtime-amd rocm-device-libs hip-dev
@@ -0,0 +1,11 @@
+#!/bin/bash -e
+clang sniff.cc -Werror -shared -fPIC -I../src/ -I../src/ROCT-Thunk-Interface/include -I../src/ROCm-Device-Libs/ockl/inc -o sniff.so -lstdc++
+#AMD_LOG_LEVEL=4 HSAKMT_DEBUG_LEVEL=7 LD_PRELOAD=$PWD/sniff.so /home/tiny/build/HIP-Examples/HIP-Examples-Applications/HelloWorld/HelloWorld
+#AMD_LOG_LEVEL=4 LD_PRELOAD=$PWD/sniff.so $HOME/build/HIP-Examples/HIP-Examples-Applications/HelloWorld/HelloWorld
+#AMD_LOG_LEVEL=5 LD_PRELOAD=$PWD/sniff.so python3 ../rdna3/asm.py
+DEBUG=5 LD_PRELOAD=$PWD/sniff.so python3 ../rdna3/asm.py
+#AMD_LOG_LEVEL=5 HSAKMT_DEBUG_LEVEL=7 DEBUG=5 LD_PRELOAD=$PWD/sniff.so strace -F python3 ../rdna3/asm.py
+#LD_PRELOAD=$PWD/sniff.so python3 ../rdna3/asm.py
+#AMD_LOG_LEVEL=4 LD_PRELOAD=$PWD/sniff.so FORWARD_ONLY=1 DEBUG=2 python3 ../../../test/test_ops.py TestOps.test_add
+#AMD_LOG_LEVEL=4 HSAKMT_DEBUG_LEVEL=7 LD_PRELOAD=$PWD/sniff.so rocm-bandwidth-test -s 0 -d 1 -m 1
+#AMD_LOG_LEVEL=4 HSAKMT_DEBUG_LEVEL=7 LD_PRELOAD=$PWD/sniff.so rocm-bandwidth-test -s 1 -d 2 -m 1
@@ -0,0 +1,282 @@
+// template copied from https://github.com/geohot/cuda_ioctl_sniffer/blob/master/sniff.cc
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <dlfcn.h>
+#include <signal.h>
+#include <ucontext.h>
+
+#include <sys/mman.h>
+
+// includes from the ROCm sources
+#include <linux/kfd_ioctl.h>
+#include <hsa.h>
+#include <amd_hsa_kernel_code.h>
+#include <ROCR-Runtime/src/core/inc/sdma_registers.h>
+using namespace rocr::AMD;
+
+#include <string>
+#include <map>
+std::map<int, std::string> files;
+std::map<uint64_t, uint64_t> ring_base_addresses;
+
+#define D(args...) fprintf(stderr, args)
+
+uint64_t doorbell_offset = -1;
+std::map<uint64_t, int> queue_types;
+
+void hexdump(void *d, int l) {
+  for (int i = 0; i < l; i++) {
+    if (i%0x10 == 0 && i != 0) printf("\n");
+    if (i%0x10 == 8) printf(" ");
+    if (i%0x10 == 0) printf("%8X: ", i);
+    printf("%2.2X ", ((uint8_t*)d)[i]);
+  }
+  printf("\n");
+}
+
+extern "C" {
+
+// https://defuse.ca/online-x86-assembler.htm#disassembly2
+static void handler(int sig, siginfo_t *si, void *unused) {
+  ucontext_t *u = (ucontext_t *)unused;
+  uint8_t *rip = (uint8_t*)u->uc_mcontext.gregs[REG_RIP];
+
+  int store_size = 0;
+  uint64_t value;
+  if (rip[0] == 0x48 && rip[1] == 0x89 && rip[2] == 0x30) {
+    // 0:  48 89 30                mov    QWORD PTR [rax],rsi
+    store_size = 8;
+    value = u->uc_mcontext.gregs[REG_RSI];
+    u->uc_mcontext.gregs[REG_RIP] += 3;
+  } else if (rip[0] == 0x4c && rip[1] == 0x89 && rip[2] == 0x28) {
+    // 0:  4c 89 28                mov    QWORD PTR [rax],r13
+    store_size = 8;
+    value = u->uc_mcontext.gregs[REG_R13];
+    u->uc_mcontext.gregs[REG_RIP] += 3;
+  } else {
+    D("segfault %02X %02X %02X %02X %02X %02X %02X %02X rip: %p addr: %p\n", rip[0], rip[1], rip[2], rip[3], rip[4], rip[5], rip[6], rip[7], rip, si->si_addr);
+    D("rax: %llx rcx: %llx rdx: %llx rsi: %llx rbx: %llx\n", u->uc_mcontext.gregs[REG_RAX], u->uc_mcontext.gregs[REG_RCX], u->uc_mcontext.gregs[REG_RDX], u->uc_mcontext.gregs[REG_RSI], u->uc_mcontext.gregs[REG_RBX]);
+    exit(-1);
+  }
+
+  uint64_t ring_base_address = ring_base_addresses[((uint64_t)si->si_addr)&0xFFF];
+  int queue_type = queue_types[((uint64_t)si->si_addr)&0xFFF];
+  D("%16p: \u001b[31mDING DONG\u001b[0m (queue_type %d) store(%d): 0x%8lx -> %p ring_base_address:0x%lx\n", rip, queue_type, store_size, value, si->si_addr, ring_base_address);
+
+  if (queue_type == KFD_IOC_QUEUE_TYPE_SDMA) {
+    uint8_t *sdma_ptr = (uint8_t*)(ring_base_address);
+    while (sdma_ptr < ((uint8_t*)(ring_base_address)+value)) {
+      D("0x%3lx: ", sdma_ptr-(uint8_t*)(ring_base_address));
+      if (sdma_ptr[0] == SDMA_OP_TIMESTAMP) {
+        D("SDMA_PKT_TIMESTAMP\n");
+        sdma_ptr += sizeof(SDMA_PKT_TIMESTAMP);
+      } else if (sdma_ptr[0] == SDMA_OP_GCR) {
+        D("SDMA_PKT_GCR\n");
+        sdma_ptr += sizeof(SDMA_PKT_GCR);
+      } else if (sdma_ptr[0] == SDMA_OP_ATOMIC) {
+        D("SDMA_PKT_ATOMIC\n");
+        sdma_ptr += sizeof(SDMA_PKT_ATOMIC);
+      } else if (sdma_ptr[0] == SDMA_OP_FENCE) {
+        D("SDMA_PKT_FENCE\n");
+        sdma_ptr += sizeof(SDMA_PKT_FENCE);
+      } else if (sdma_ptr[0] == SDMA_OP_TRAP) {
+        D("SDMA_PKT_TRAP\n");
+        sdma_ptr += sizeof(SDMA_PKT_TRAP);
+      } else if (sdma_ptr[0] == SDMA_OP_COPY && sdma_ptr[1] == SDMA_SUBOP_COPY_LINEAR) {
+        SDMA_PKT_COPY_LINEAR *pkt = (SDMA_PKT_COPY_LINEAR *)sdma_ptr;
+        D("SDMA_PKT_COPY_LINEAR: count:0x%x src:0x%lx dst:0x%lx\n", pkt->COUNT_UNION.count+1,
+          (uint64_t)pkt->SRC_ADDR_LO_UNION.src_addr_31_0 | ((uint64_t)pkt->SRC_ADDR_HI_UNION.src_addr_63_32 << 32),
+          (uint64_t)pkt->DST_ADDR_LO_UNION.dst_addr_31_0 | ((uint64_t)pkt->DST_ADDR_HI_UNION.dst_addr_63_32 << 32)
+        );
+        sdma_ptr += sizeof(SDMA_PKT_COPY_LINEAR);
+      } else {
+        D("unhandled packet type %d %d, exiting\n", sdma_ptr[0], sdma_ptr[1]);
+        break;
+      }
+    }
+
+    //hexdump((void*)(ring_base_address), 0x100);
+  } else if (queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL) {
+    hsa_kernel_dispatch_packet_t *pkt = (hsa_kernel_dispatch_packet_t *)(ring_base_address+value*0x40);
+    if ((pkt->header&0xFF) == HSA_PACKET_TYPE_KERNEL_DISPATCH) {
+      D("HSA_PACKET_TYPE_KERNEL_DISPATCH -- setup:%d workgroup[%d, %d, %d] grid[%d, %d, %d] kernel_object:0x%lx kernarg_address:%p\n", pkt->setup, pkt->workgroup_size_x, pkt->workgroup_size_y, pkt->workgroup_size_z, pkt->grid_size_x, pkt->grid_size_y, pkt->grid_size_z, pkt->kernel_object, pkt->kernarg_address);
+      amd_kernel_code_t *code = (amd_kernel_code_t *)pkt->kernel_object;
+      D("kernel_code_entry_byte_offset:%lx\n", code->kernel_code_entry_byte_offset);
+      uint32_t *kernel_code = (uint32_t*)(pkt->kernel_object + code->kernel_code_entry_byte_offset);
+      int code_len = 0;
+      while (kernel_code[code_len] != 0xbf9f0000 && kernel_code[code_len] != 0) code_len++;
+      hexdump(kernel_code, code_len*4);
+      /*FILE *f = fopen("/tmp/kernel_code", "wb");
+      fwrite(kernel_code, 4, code_len, f);
+      fclose(f);
+      system("python -c 'print(\" \".join([(\"0x%02X\"%x) for x in open(\"/tmp/kernel_code\", \"rb\").read()]))' | ../build/llvm-project/bin/llvm-mc --disassemble --arch=amdgcn --mcpu=gfx1100 --show-encoding");*/
+      D("kernargs (kernarg_segment_byte_size:0x%lx)\n", code->kernarg_segment_byte_size);
+      // get length
+      int i;
+      for (i = 0; i < 0x400; i+=0x10) {
+        if (memcmp((void*)((uint64_t)pkt->kernarg_address+i), "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00", 0x10) == 0) break;
+      }
+      hexdump((void*)pkt->kernarg_address, i+0x10);
+    } else if ((pkt->header&0xFF) == HSA_PACKET_TYPE_BARRIER_AND) {
+      hsa_barrier_and_packet_t *pkt_and = (hsa_barrier_and_packet_t *)(ring_base_address+value*0x40);
+      D("HSA_PACKET_TYPE_BARRIER_AND completion_signal:0x%lx\n", pkt_and->completion_signal.handle);
+      //hexdump((void*)(ring_base_address+value*0x40), 0x40);
+    } else if ((pkt->header&0xFF) == HSA_PACKET_TYPE_VENDOR_SPECIFIC) {
+      D("HSA_PACKET_TYPE_VENDOR_SPECIFIC\n");
+      hexdump((void*)(ring_base_address+value*0x40), 0x40);
+    } else {
+      hexdump((void*)(ring_base_address+value*0x40), 0x40);
+    }
+  }
+
+  mprotect((void *)((uint64_t)si->si_addr & ~0xFFF), 0x2000, PROT_READ | PROT_WRITE);
+  if (store_size == 8) {
+    *(volatile uint64_t*)(si->si_addr) = value;
+  } else if (store_size == 4) {
+    *(volatile uint32_t*)(si->si_addr) = value;
+  } else if (store_size == 2) {
+    *(volatile uint16_t*)(si->si_addr) = value;
+  } else {
+    D("store size not supported\n");
+    exit(-1);
+  }
+  mprotect((void *)((uint64_t)si->si_addr & ~0xFFF), 0x2000, PROT_NONE);
+}
+
+void register_sigsegv_handler() {
+  struct sigaction sa = {0};
+  sa.sa_flags = SA_SIGINFO;
+  sigemptyset(&sa.sa_mask);
+  sa.sa_sigaction = handler;
+  if (sigaction(SIGSEGV, &sa, NULL) == -1) {
+    D("ERROR: failed to register sigsegv handler");
+    exit(-1);
+  }
+  // NOTE: python (or ocl runtime?) blocks the SIGSEGV signal
+  sigset_t x;
+  sigemptyset(&x);
+  sigaddset(&x, SIGSEGV);
+  sigprocmask(SIG_UNBLOCK, &x, NULL);
+}
+
+int (*my_open)(const char *pathname, int flags, mode_t mode);
+#undef open
+int open(const char *pathname, int flags, mode_t mode) {
+  if (my_open == NULL) my_open = reinterpret_cast<decltype(my_open)>(dlsym(RTLD_NEXT, "open"));
+  int ret = my_open(pathname, flags, mode);
+  //D("open %s (0o%o) = %d\n", pathname, flags, ret);
+  files[ret] = pathname;
+  return ret;
+}
+
+
+int (*my_open64)(const char *pathname, int flags, mode_t mode);
+#undef open
+int open64(const char *pathname, int flags, mode_t mode) {
+  if (my_open64 == NULL) my_open64 = reinterpret_cast<decltype(my_open64)>(dlsym(RTLD_NEXT, "open64"));
+  int ret = my_open64(pathname, flags, mode);
+  //D("open %s (0o%o) = %d\n", pathname, flags, ret);
+  files[ret] = pathname;
+  return ret;
+}
+
+void *(*my_mmap)(void *addr, size_t length, int prot, int flags, int fd, off_t offset);
+#undef mmap
+void *mmap(void *addr, size_t length, int prot, int flags, int fd, off_t offset) {
+  if (my_mmap == NULL) my_mmap = reinterpret_cast<decltype(my_mmap)>(dlsym(RTLD_NEXT, "mmap"));
+  void *ret = my_mmap(addr, length, prot, flags, fd, offset);
+
+  if (doorbell_offset != -1 && offset == doorbell_offset) {
+    D("HIDDEN DOORBELL %p, handled by %p\n", addr, handler);
+    register_sigsegv_handler();
+    mprotect(addr, length, PROT_NONE);
+  }
+
+  if (fd != -1) D("mmapped %p (target %p) with flags 0x%x length 0x%zx fd %d %s offset 0x%lx\n", ret, addr, flags, length, fd, files[fd].c_str(), offset);
+  return ret;
+}
+
+void *(*my_mmap64)(void *addr, size_t length, int prot, int flags, int fd, off_t offset);
+#undef mmap64
+void *mmap64(void *addr, size_t length, int prot, int flags, int fd, off_t offset) { return mmap(addr, length, prot, flags, fd, offset); }
+
+int ioctl_num = 1;
+int (*my_ioctl)(int filedes, unsigned long request, void *argp) = NULL;
+#undef ioctl
+int ioctl(int filedes, unsigned long request, void *argp) {
+  if (my_ioctl == NULL) my_ioctl = reinterpret_cast<decltype(my_ioctl)>(dlsym(RTLD_NEXT, "ioctl"));
+  int ret = 0;
+  ret = my_ioctl(filedes, request, argp);
+  if (!files.count(filedes)) return ret;
+
+  uint8_t type = (request >> 8) & 0xFF;
+  uint8_t nr = (request >> 0) & 0xFF;
+  uint16_t size = (request >> 16) & 0xFFF;
+
+  D("%3d: %d = %3d(%20s) 0x%3x ", ioctl_num, ret, filedes, files[filedes].c_str(), size);
+
+  if (request == AMDKFD_IOC_SET_EVENT) {
+    kfd_ioctl_set_event_args *args = (kfd_ioctl_set_event_args *)argp;
+    D("AMDKFD_IOC_SET_EVENT event_id:%d", args->event_id);
+  } else if (request == AMDKFD_IOC_ALLOC_MEMORY_OF_GPU) {
+    kfd_ioctl_alloc_memory_of_gpu_args *args = (kfd_ioctl_alloc_memory_of_gpu_args *)argp;
+    D("AMDKFD_IOC_ALLOC_MEMORY_OF_GPU va_addr:0x%llx size:0x%llx handle:%llX gpu_id:0x%x", args->va_addr, args->size, args->handle, args->gpu_id);
+  } else if (request == AMDKFD_IOC_MAP_MEMORY_TO_GPU) {
+    kfd_ioctl_map_memory_to_gpu_args *args = (kfd_ioctl_map_memory_to_gpu_args *)argp;
+    D("AMDKFD_IOC_MAP_MEMORY_TO_GPU handle:%llX", args->handle);
+  } else if (request == AMDKFD_IOC_CREATE_EVENT) {
+    kfd_ioctl_create_event_args *args = (kfd_ioctl_create_event_args *)argp;
+    D("AMDKFD_IOC_CREATE_EVENT event_page_offset:0x%llx event_type:%d event_id:%d", args->event_page_offset, args->event_type, args->event_id);
+  } else if (request == AMDKFD_IOC_WAIT_EVENTS) {
+    D("AMDKFD_IOC_WAIT_EVENTS");
+  } else if (request == AMDKFD_IOC_SET_XNACK_MODE) {
+    D("AMDKFD_IOC_SET_XNACK_MODE");
+  } else if (request == AMDKFD_IOC_SVM || (type == 0x4b && nr == 0x20)) {
+    // NOTE: this one is variable length
+    kfd_ioctl_svm_args *args = (kfd_ioctl_svm_args *)argp;
+    D("AMDKFD_IOC_SVM start_addr:0x%llx size:0x%llx op:%d", args->start_addr, args->size, args->op);
+  } else if (request == AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU) {
+    kfd_ioctl_unmap_memory_from_gpu_args *args = (kfd_ioctl_unmap_memory_from_gpu_args *)argp;
+    D("AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU handle:%llX", args->handle);
+  } else if (request == AMDKFD_IOC_FREE_MEMORY_OF_GPU) {
+    D("AMDKFD_IOC_FREE_MEMORY_OF_GPU");
+  } else if (request == AMDKFD_IOC_SET_SCRATCH_BACKING_VA) {
+    D("AMDKFD_IOC_SET_SCRATCH_BACKING_VA");
+  } else if (request == AMDKFD_IOC_GET_TILE_CONFIG) {
+    D("AMDKFD_IOC_GET_TILE_CONFIG");
+  } else if (request == AMDKFD_IOC_SET_TRAP_HANDLER) {
+    D("AMDKFD_IOC_SET_TRAP_HANDLER");
+  } else if (request == AMDKFD_IOC_GET_VERSION) {
+    kfd_ioctl_get_version_args *args = (kfd_ioctl_get_version_args *)argp;
+    D("AMDKFD_IOC_GET_VERSION major_version:%d minor_version:%d", args->major_version, args->minor_version);
+  } else if (request == AMDKFD_IOC_GET_PROCESS_APERTURES_NEW) {
+    D("AMDKFD_IOC_GET_PROCESS_APERTURES_NEW");
+  } else if (request == AMDKFD_IOC_ACQUIRE_VM) {
+    D("AMDKFD_IOC_ACQUIRE_VM");
+  } else if (request == AMDKFD_IOC_SET_MEMORY_POLICY) {
+    D("AMDKFD_IOC_SET_MEMORY_POLICY");
+  } else if (request == AMDKFD_IOC_GET_CLOCK_COUNTERS) {
+    D("AMDKFD_IOC_GET_CLOCK_COUNTERS");
+  } else if (request == AMDKFD_IOC_CREATE_QUEUE) {
+    kfd_ioctl_create_queue_args *args = (kfd_ioctl_create_queue_args *)argp;
+    D("AMDKFD_IOC_CREATE_QUEUE\n");
+    D("queue_type:%d ring_base_address:0x%llx\n", args->queue_type, args->ring_base_address);
+    D("eop_buffer_address:0x%llx ctx_save_restore_address:0x%llx\n", args->eop_buffer_address, args->ctx_save_restore_address);
+    D("ring_size:0x%x queue_priority:%d\n", args->ring_size, args->queue_priority);
+    D("RETURNS write_pointer_address:0x%llx read_pointer_address:0x%llx doorbell_offset:0x%llx queue_id:%d\n", args->write_pointer_address, args->read_pointer_address, args->doorbell_offset, args->queue_id);
+    //D("RETURNS *write_pointer_address:0x%llx *read_pointer_address:0x%llx\n", *(uint64_t*)args->write_pointer_address, *(uint64_t*)args->read_pointer_address);
+    ring_base_addresses[args->doorbell_offset&0xFFF] = args->ring_base_address;
+    queue_types[args->doorbell_offset&0xFFF] = args->queue_type;
+    doorbell_offset = args->doorbell_offset&~0xFFF;
+  } else {
+    D("type:0x%x nr:0x%x size:0x%x", type, nr, size);
+  }
+
+  D("\n");
+  ioctl_num++;
+  return ret;
+}
+
+}
@@ -34,8 +34,7 @@ class WallTimeEvent:
    self.start = time.monotonic()
    return self
  def __exit__(self, *_):
-    self.time = time.monotonic() - self.start
-    _events[self.event]["wall"].append(self.time)
+    _events[self.event]["wall"].append(time.monotonic() - self.start)
    return False

 class KernelTimeEvent:
@@ -1,11 +1,11 @@
 from typing import Tuple, Dict, List, Optional
-from tinygrad.dtype import DType, dtypes
+from tinygrad.dtype import DType
 from tinygrad.renderer import ProgramSpec
-from tinygrad.tensor import Tensor
-from tinygrad.device import Device
+from tinygrad.tensor import Device, Tensor
 from tinygrad.engine.jit import TinyJit
 from tinygrad.nn.state import get_state_dict
 from tinygrad.helpers import Context, to_mv
+from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import Ops
 import json
 from collections import OrderedDict
@@ -13,20 +13,12 @@ from collections import OrderedDict
 EXPORT_SUPPORTED_DEVICE = ["WEBGPU", "CPU", "CUDA", "CL"]

 def compile_net(run:TinyJit, special_names:Dict[int,str]) -> Tuple[Dict[str,str],List[Tuple[str,List[str],List[int]]],Dict[str,Tuple[int,DType,int]],Dict[str,Tensor]]:
-  # memory-planned subbuffers can have multiple Buffer objects for the same memory region
-  canon, _seen = {}, {}
-  for ji in run.jit_cache:
-    for b in ji.bufs:
-      if b is not None: canon[id(b)] = _seen.setdefault((id(b.base._buf), b.offset, b.size, b.dtype), b)
-  special_names = {id(canon[k]): v for k, v in special_names.items() if k in canon}
-
  functions, bufs, bufs_to_save, statements, bufnum = {}, {}, {}, [], 0
  for ji in run.jit_cache:
    fxn: ProgramSpec = ji.prg.p
    functions[fxn.function_name] = fxn.src   # NOTE: this assumes all with the same name are the same
    cargs = []
    for i,arg in enumerate(ji.bufs):
-      arg = canon[id(arg)]
      key = id(arg)
      if key not in bufs:
        if key in special_names:
@@ -75,11 +67,12 @@ def export_model_clang(functions:Dict[str,str], statements:Dict[str,Tuple[str,in
  forward_args = ",".join(f"{dtype}{'*' if name not in symbolic_vars.values() else ''} {name}" for name,dtype,_ in (outputs+inputs if wasm else inputs+outputs))

  if not wasm:
+    thread_id = 0 # NOTE: export does not support threading, thread_id is always 0
    for name,cl in bufs_to_save.items():
      weight = ''.join(["\\x%02X"%x for x in bytes(to_mv(cl._buf.va_addr, cl._buf.size))])
      cprog.append(f"unsigned char {name}_data[] = \"{weight}\";")
    cprog += [f"{dtype_map[dtype]} {name}[{len}];" if name not in bufs_to_save else f"{dtype_map[dtype]} *{name} = ({dtype_map[dtype]} *){name}_data;" for name,(len,dtype,_key) in bufs.items() if name not in input_names+output_names]
-    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)});" for (name, args, _global_size, _local_size) in statements] + ["}"]
+    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)}, {thread_id});" for (name, args, _global_size, _local_size) in statements] + ["}"]
    return '\n'.join(headers + cprog)
  else:
    if bufs_to_save:
@@ -1,101 +0,0 @@
-from typing import Callable, Any
-from tinygrad import Tensor, dtypes, nn, UOp
-from tinygrad.uop.ops import KernelInfo, AxisType, Ops
-
-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
-  x_abs_max = x.abs().max().detach()
-  scale = fp8_max / (x_abs_max + 1e-8)
-  x_scaled = x * scale
-  x_det = x_scaled.detach()
-  x_clamped = x_det.clamp(fp8_min, fp8_max)
-  x_clamped_ste = x_scaled + (x_clamped - x_det)
-  res = x_clamped_ste.cast(dtype)
-  return res, scale.float().reciprocal()
-
-def custom_matmul(output: UOp, inp: UOp, weight: UOp) -> UOp:
-  SEQ = inp.shape[1]
-  OUT = weight.shape[0]
-  IN = weight.shape[-1]
-  seq_idx = UOp.range(SEQ, 2, AxisType.LOOP)
-  out_idx = UOp.range(OUT, 3, AxisType.LOOP)
-  batch_idx = UOp.range(output.size//SEQ//OUT, 1, AxisType.LOOP)
-  reduce_idx = UOp.range(IN, 0, AxisType.REDUCE)
-  product = (inp.index((seq_idx*IN+reduce_idx+batch_idx*IN*SEQ)) * weight.index((out_idx*IN+reduce_idx))).cast(dtypes.float)
-  reduced = product.reduce(reduce_idx, arg=Ops.ADD)
-  store_op = output.index((seq_idx*OUT+out_idx+batch_idx*OUT*SEQ), ptr=True).store(reduced).end(batch_idx, seq_idx, out_idx)
-  return store_op.sink(arg=KernelInfo(name=f"fp8_matmul_{inp.shape}x{weight.shape}"))
-
-def custom_matmul_backward(gradient: UOp, kernel: UOp) -> tuple[UOp, UOp]:
-  _, input_uop, weight_uop = kernel.src[1:]
-  input_tensor = Tensor(input_uop, device=input_uop.device)
-  grad_tensor = Tensor(gradient, device=gradient.device)
-  weight_tensor = Tensor(weight_uop, device=weight_uop.device)
-  grad_quantized, scale = quantize_to_fp8(grad_tensor)
-  scale_scalar = scale.reshape(())
-  grad_weight = Tensor.einsum("bso,bsi->oi", grad_quantized, input_tensor, dtype=dtypes.float)
-  grad_weight = grad_weight * scale_scalar
-  grad_2d = grad_quantized.reshape(grad_tensor.shape[0] * grad_tensor.shape[1], grad_tensor.shape[-1])
-  grad_input = (grad_2d.dot(weight_tensor, dtype=dtypes.float)).contiguous().reshape(input_tensor.shape) * scale
-  return (None, grad_input.uop, grad_weight.uop)
-
-class FP8Linear:
-  def __init__(self, in_features:int, out_features:int, bias:bool=True):
-    self.weight = Tensor.empty(out_features, in_features, dtype=dtypes.float32)
-    self.bias = Tensor.empty(out_features, dtype=dtypes.float32) if bias else None
-
-  def __call__(self, x: Tensor) -> Tensor:
-    original_ndim = len(x.shape)
-    if original_ndim == 2: x = x.reshape(x.shape[0], 1, x.shape[1])
-    batch, seq, _ = x.shape
-    w_fp8, w_scale = quantize_to_fp8(self.weight)
-    x_fp8, x_scale = quantize_to_fp8(x)
-    GPUS = self.weight.device
-    if isinstance(GPUS, tuple) and len(GPUS) > 1:
-      y = Tensor(Tensor.empty((batch//len(GPUS), seq, self.weight.shape[0]), dtype=dtypes.float, device=GPUS).uop.multi(0), device=GPUS)
-    else:
-      y = Tensor.empty((batch, seq, self.weight.shape[0]), dtype=dtypes.float)
-    y = Tensor.custom_kernel(y, x_fp8, w_fp8, fxn=custom_matmul, grad_fxn=custom_matmul_backward)[0]
-    y = y * w_scale * x_scale
-    if self.bias is not None: y = y + self.bias
-    if original_ndim == 2: y = y.reshape(batch, self.weight.shape[0])
-    return y.cast(x.dtype)
-
-def _replace_linear(layer: nn.Linear):
-  fp8_linear = FP8Linear(layer.weight.shape[1], layer.weight.shape[0], layer.bias is not None)
-  fp8_linear.weight = layer.weight
-  if layer.bias is not None: fp8_linear.bias = layer.bias
-  return fp8_linear
-
-def _swap_linear_with_fp8(model, module_filter_fn:Callable[[Any, str],bool]|None=None, fqn:str="", parent:Any|None=None,
-                          attr_name:str="", visited:set|None=None):
-  if visited is None: visited = set()
-  if id(model) in visited: return
-  visited.add(id(model))
-  if isinstance(model, (str, int, float, bool, type(None), Tensor, UOp)): return
-  elif isinstance(model, nn.Linear):
-    if module_filter_fn is not None and not module_filter_fn(model, fqn): return
-    fp8_linear = _replace_linear(model)
-    if parent is not None and attr_name:
-      setattr(parent, attr_name, fp8_linear)
-  elif isinstance(model, list):
-    for i, item in enumerate(model):
-      child_fqn = f"{fqn}.{i}" if fqn else str(i)
-      if isinstance(item, nn.Linear) and (module_filter_fn is None or module_filter_fn(item, child_fqn)): model[i] = _replace_linear(item)
-      else: _swap_linear_with_fp8(item, module_filter_fn, child_fqn, None, "", visited)
-  elif isinstance(model, dict):
-    for key, item in list(model.items()):
-      child_fqn = f"{fqn}.{key}" if fqn else str(key)
-      if isinstance(item, nn.Linear) and (module_filter_fn is None or module_filter_fn(item, child_fqn)): model[key] = _replace_linear(item)
-      else: _swap_linear_with_fp8(item, module_filter_fn, child_fqn, None, "", visited)
-  elif hasattr(model, "__dict__"):
-    for attr_key in list(vars(model).keys()):
-      try: attr = getattr(model, attr_key)
-      except Exception: continue
-      child_fqn = f"{fqn}.{attr_key}" if fqn else attr_key
-      _swap_linear_with_fp8(attr, module_filter_fn, child_fqn, model, attr_key, visited)
-
-def convert_to_float8_training(model, module_filter_fn:Callable[[Any,str],bool]|None=None):
-  _swap_linear_with_fp8(model, module_filter_fn, "", None, "")
-  return model
@@ -1,2 +1,3 @@
+*.s
 *.ll
 fp32_sgemm_amd
@@ -1,497 +0,0 @@
-# RDNA3 128x128 tiled GEMM kernel - DSL version
-# Computes C = A @ B for NxN float32 matrices using 128x128 tiles
-#
-# Architecture: RDNA3 (gfx1100)
-# Tile size: 128x128 (each workgroup computes one tile of C)
-# Workgroup: 128 threads (arranged as 32x4 for coalesced memory access)
-# Inner loop: 8 iterations per K-block, processing 8 columns of A and 8 rows of B
-#
-# Accumulators: 128 vgprs (v[2-129])
-
-import numpy as np
-from tinygrad import Tensor, Device, Context, GlobalCounters
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.helpers import getenv, colored
-from tinygrad.dtype import dtypes, AddrSpace
-from tinygrad.engine.realize import Estimates
-from tinygrad.renderer.amd.dsl import s, v, VCC_LO, NULL
-from tinygrad.runtime.autogen.amd.rdna3.ins import *
-
-# =============================================================================
-# Kernel constants
-# =============================================================================
-LDS_SIZE = 8320       # Local data share size in bytes
-LDS_A_STRIDE = 0x210  # LDS stride for A tile (528 bytes)
-LDS_B_STRIDE = 0x200  # LDS stride for B tile (512 bytes)
-LDS_BASE_OFFSET = 0x1080  # Base LDS offset for tiles
-ADDR_MASK = 0x3fffff80    # Address alignment mask
-
-# =============================================================================
-# Named register assignments (VGPRs)
-# =============================================================================
-V_LANE_ID = 0             # lane_id set on startup
-# Use tile gaps (v146-159) for named regs to minimize max VGPR
-V_LANE_ID_MOD8 = 146      # lane_id & 7
-V_LANE_MOD8_X4 = 147      # (lane_id & 7) << 2
-V_LANE_DIV8_X4 = 150      # ((lane_id >> 3) & 3) << 2
-V_LDS_B_BASE = 151        # LDS B-tile base address for inner loop
-V_LDS_A_BASE = 154        # LDS A-tile base address for inner loop
-V_GLOBAL_A_ADDR = 155     # global memory A prefetch address
-V_GLOBAL_B_ADDR = 158     # global memory B prefetch address
-V_LDS_A_ADDR = 159        # single base register for A stores
-V_LDS_B_ADDR = 162        # single base register for B stores
-
-# LDS tile register destinations - SEPARATE from DATA to avoid overlap
-# A on banks 2-3, B on banks 0-1 to avoid bank conflicts in VOPD
-V_A_TILE_REGS = [130, 134, 138, 142]  # A tile: banks 2,2,2,2 (130%4=2, etc.)
-V_B_TILE_REGS = [132, 136, 140, 144, 148, 152, 156, 160]  # B tile: banks 0,0,0,0,0,0,0,0
-
-# =============================================================================
-# Named register assignments (SGPRs)
-# =============================================================================
-S_OUT_PTR = (0, 1)        # output C matrix base pointer
-S_WORKGROUP_X = 2         # workgroup_id_x (system SGPR, follows user SGPRs)
-S_WORKGROUP_Y = 3         # workgroup_id_y (system SGPR)
-S_DIM_N = 4               # matrix dimension N
-S_LOOP_BOUND = 7          # K-8 (loop termination bound)
-S_LOOP_CTR = 12           # loop counter (increments by 8)
-S_PREFETCH_FLAG = 13      # prefetch condition flag / row stride in epilogue
-S_TILE_X = 14             # workgroup_x << 7
-S_TILE_Y = 15             # workgroup_y << 7
-# Kernarg load destinations
-S_KERNARG_A = (20, 21)    # A pointer from kernarg
-S_KERNARG_B = (22, 23)    # B pointer from kernarg
-# Prefetch base pointers (8 pairs each, B: N*4 bytes apart, A: N*64 bytes apart)
-S_PREFETCH_B = 24         # s[24:39] - 8 B tile pointers
-S_PREFETCH_A = 40         # s[40:55] - 8 A tile pointers
-
-# =============================================================================
-# Data tables
-# =============================================================================
-
-# Accumulator grid: ACC_GRID[a_idx][b_idx] = vgpr for C[a,b]
-# a_idx: which A value (0-7), b_idx: which B value (0-15)
-# Scattered due to VOPD bank constraints (vdst_x % 4 != vdst_y % 4)
-# Range is from v2 - v129
-ACC_GRID = [
-  [  5,  3,  9,  8,   37, 35, 41, 40,   69, 67, 73, 72,  101, 99,105,104],  # a0
-  [  4,  2,  7,  6,   36, 34, 39, 38,   68, 66, 71, 70,  100, 98,103,102],  # a1
-  [ 17, 16, 13, 11,   49, 48, 45, 43,   81, 80, 77, 75,  113,112,109,107],  # a2
-  [ 15, 14, 12, 10,   47, 46, 44, 42,   79, 78, 76, 74,  111,110,108,106],  # a3
-  [ 21, 19, 25, 24,   53, 51, 57, 56,   85, 83, 89, 88,  117,115,121,120],  # a4
-  [ 20, 18, 23, 22,   52, 50, 55, 54,   84, 82, 87, 86,  116,114,123,122],  # a5
-  [125,128, 29, 27,   33, 32, 61, 59,   65, 64, 93, 91,   97, 96,129,127],  # a6
-  [119,118, 28, 26,   31, 30, 60, 58,   63, 62, 92, 90,   95, 94,124,126],  # a7
-]
-
-# Optimized (a_pair, b_pair) iteration order for better GPU scheduling
-# Interleaves A and B pairs to maximize instruction-level parallelism
-FMAC_PAIR_ORDER = [
-  (0,0),(0,1),(1,1),(1,0), (2,0),(2,1),(3,1),(3,2), (0,2),(0,3),(1,3),(1,2), (2,2),(2,3),(3,3),(3,4),
-  (0,4),(0,5),(1,5),(1,4), (2,4),(2,5),(3,5),(3,6), (0,6),(0,7),(1,7),(1,6), (2,6),(2,7),(3,7),(3,0),
-]
-
-def derive_fmac_pattern(acc_grid, a_tile_regs=None, b_tile_regs=None):
-  """Generate 64 dual FMAC ops from accumulator grid with optimized iteration order."""
-  pattern = []
-  for idx, (a_pair, b_pair) in enumerate(FMAC_PAIR_ORDER):
-    a_even, a_odd = a_pair * 2, a_pair * 2 + 1
-    b_even, b_odd = b_pair * 2, b_pair * 2 + 1
-    a_base, b_base = a_tile_regs[a_pair], b_tile_regs[b_pair]
-    # Op 1: normal order -> C[a_even, b_even] + C[a_odd, b_odd]
-    pattern.append((acc_grid[a_even][b_even], acc_grid[a_odd][b_odd],
-                   a_base, b_base, a_base+1, b_base+1))
-    # Op 2: alternate swapping A vs B to vary register banks
-    if idx % 2 == 0:  # swap B
-      pattern.append((acc_grid[a_even][b_odd], acc_grid[a_odd][b_even],
-                     a_base, b_base+1, a_base+1, b_base))
-    else:  # swap A
-      pattern.append((acc_grid[a_odd][b_even], acc_grid[a_even][b_odd],
-                     a_base+1, b_base, a_base, b_base+1))
-  return pattern
-
-# Derived: 64 dual FMAC operations
-FMAC_PATTERN = derive_fmac_pattern(ACC_GRID, V_A_TILE_REGS, V_B_TILE_REGS)
-
-def derive_permute_swaps(acc_grid, out_regs):
-  """Derive swap sequence to permute accumulators from FMAC layout to output order.
-
-  After FMAC loop: acc_grid[a][b] holds C[a,b]
-  Output order: for row_half in 0,1; col_group in 0-3; row_in_group in 0-3; b_off in 0-3
-    -> need C[row_half*4 + row_in_group, col_group*4 + b_off] in specified reg order
-  """
-  def target_ab(i):
-    row_half, col_group = i // 64, (i // 16) % 4
-    row_in_group, b_off = (i // 4) % 4, i % 4
-    return (row_half * 4 + row_in_group, col_group * 4 + b_off)
-
-  reg_contents = {acc_grid[a][b]: (a, b) for a in range(8) for b in range(16)}
-  ab_location = {ab: r for r, ab in reg_contents.items()}
-
-  swaps = []
-  for i in range(128):
-    target_reg, needed_ab = out_regs[i], target_ab(i)
-    current_reg = ab_location[needed_ab]
-    if current_reg != target_reg:
-      swaps.append((current_reg, target_reg))
-      ab_at_target = reg_contents.get(target_reg)
-      reg_contents[target_reg], ab_location[needed_ab] = needed_ab, target_reg
-      if ab_at_target is not None:
-        reg_contents[current_reg], ab_location[ab_at_target] = ab_at_target, current_reg
-  return swaps
-
-# Derived: swap sequence to arrange accumulators for output
-# Each group of 4 registers is ascending for direct global_store_b128
-OUT_REGS = [r for i in range(32) for r in range(126 - i*4, 130 - i*4)]
-PERMUTE_SWAPS = derive_permute_swaps(ACC_GRID, OUT_REGS)
-
-# =============================================================================
-# LDS tile staging registers
-# =============================================================================
-# DATA regs receive contiguous global prefetch, then write to LDS
-# TILE regs receive scattered LDS loads (ds_load_b64 pairs), then feed FMACs
-# Contiguous layout with mod4=[3,0,1,2,3,0,1,2] for bank conflict avoidance
-V_LDS_A_DATA = [163, 164, 165, 166, 167, 168, 169, 170]
-V_LDS_B_DATA = [171, 172, 173, 174, 175, 176, 177, 178]
-
-# Initial tile prefetch: (vdst, saddr_lo) - load into A data regs using B prefetch pointers (s[24:31])
-INIT_PREFETCH = [(V_LDS_A_DATA[i], S_PREFETCH_B+2*i) for i in range(4)]
-
-# Global memory prefetch schedule: (vdst1, vdst2, addr_vreg, saddr_lo1, saddr_lo2)
-# First 2 pairs from B prefetch pointers (s[32:39]), next 4 pairs from A prefetch pointers (s[40:55])
-PREFETCH_LOADS = [(V_LDS_A_DATA[4+2*i], V_LDS_A_DATA[4+2*i+1], V_GLOBAL_B_ADDR, S_PREFETCH_B+8+4*i, S_PREFETCH_B+10+4*i) for i in range(2)] + \
-                 [(V_LDS_B_DATA[2*(i-2)], V_LDS_B_DATA[2*(i-2)+1], V_GLOBAL_A_ADDR, S_PREFETCH_A+4*(i-2), S_PREFETCH_A+2+4*(i-2)) for i in range(2, 6)]
-
-# =============================================================================
-# Kernel class
-# =============================================================================
-
-class Kernel:
-  def __init__(self, arch='gfx1100'): self.instructions, self.labels, self.pos, self.arch = [], {}, 0, arch
-  def label(self, name): self.labels[name] = self.pos
-
-  def emit(self, inst, target=None):
-    self.instructions.append(inst)
-    inst._target, inst._pos = target, self.pos
-    self.pos += inst.size()
-    return inst
-
-  def waitcnt(self, lgkm=None, vm=None):
-    """Wait for memory operations. lgkm=N waits until N lgkm ops remain, vm=N waits until N vmem ops remain."""
-    vmcnt, lgkmcnt, expcnt = vm if vm is not None else 63, lgkm if lgkm is not None else 63, 7
-    waitcnt = (expcnt & 0x7) | ((lgkmcnt & 0x3f) << 4) | ((vmcnt & 0x3f) << 10)
-    self.emit(s_waitcnt(simm16=waitcnt))
-
-  def finalize(self):
-    """Patch branch offsets and return the finalized instruction list."""
-    for inst in self.instructions:
-      if inst._target is None: continue
-      offset_dwords = (self.labels[inst._target] - inst._pos - inst.size()) // 4
-      if not -32768 <= offset_dwords <= 32767: raise ValueError(f"branch to '{inst._target}' offset {offset_dwords} exceeds simm16 range")
-      inst.simm16 = offset_dwords
-    return self.instructions
-
-
-# =============================================================================
-# Kernel builder
-# =============================================================================
-
-def build_kernel(N, arch='gfx1100'):
-  assert N % 128 == 0, f"N must be a multiple of 128 (tile size), got {N}"
-  assert N >= 256, f"N must be >= 256 (prefetch pipeline requires at least 2 K-blocks), got {N}"
-  k = Kernel(arch)
-
-  # ===========================================================================
-  # PROLOGUE: Load kernel arguments, compute tile coordinates and addresses
-  # ===========================================================================
-  k.emit(s_load_b128(sdata=s[S_KERNARG_A[0]:S_KERNARG_B[1]], sbase=s[0:1], offset=0x0, soffset=NULL))
-  k.emit(s_load_b64(sdata=s[S_OUT_PTR[0]:S_OUT_PTR[1]], sbase=s[0:1], offset=0x10, soffset=NULL))
-  k.emit(s_mov_b32(s[S_DIM_N], N))
-  k.emit(s_mov_b32(s[S_LOOP_CTR], 0))  # used by LDS swizzle, always 0 for valid workgroups
-  k.emit(s_lshl_b32(s[S_TILE_X], s[S_WORKGROUP_X], 7))
-  k.emit(s_lshl_b32(s[S_TILE_Y], s[S_WORKGROUP_Y], 7))
-
-  # Lane-derived values
-  k.emit(v_and_b32_e32(v[V_LANE_ID_MOD8], 7, v[V_LANE_ID]))
-  k.emit(v_lshrrev_b32_e32(v[4], 3, v[V_LANE_ID]))
-  k.emit(v_or_b32_e32(v[1], s[S_TILE_X], v[V_LANE_ID]))
-  k.emit(v_or_b32_e32(v[22], s[S_TILE_Y], v[4]))
-  k.emit(v_lshlrev_b32_e32(v[V_LANE_MOD8_X4], 2, v[V_LANE_ID_MOD8]))
-  k.waitcnt(lgkm=0)
-
-  # Compute 8 A and B matrix tile base pointers for prefetch
-  k.emit(s_mov_b64(s[S_PREFETCH_B:S_PREFETCH_B+1], s[S_KERNARG_B[0]:S_KERNARG_B[1]]))  # B[0]: no offset
-  for i in range(1, 8):  # B: each pointer 1 row of B apart (N*4 bytes)
-    k.emit(s_add_u32(s[S_PREFETCH_B+i*2], s[S_KERNARG_B[0]], i * N * 4))
-    k.emit(s_addc_u32(s[S_PREFETCH_B+i*2+1], s[S_KERNARG_B[1]], 0))
-  k.emit(s_mov_b64(s[S_PREFETCH_A:S_PREFETCH_A+1], s[S_KERNARG_A[0]:S_KERNARG_A[1]]))  # A[0]: no offset
-  for i in range(1, 8):  # A: each pointer 16 rows of A apart (16*N*4 bytes)
-    k.emit(s_add_u32(s[S_PREFETCH_A+i*2], s[S_KERNARG_A[0]], i * N * 64))
-    k.emit(s_addc_u32(s[S_PREFETCH_A+i*2+1], s[S_KERNARG_A[1]], 0))
-
-  # Global prefetch addresses: B = (tile_x + lane_id) * 4, A = (tile_y*N + (lane_id/8)*N + lane_id%8) * 4
-  k.emit(v_add_nc_u32_e32(v[V_GLOBAL_B_ADDR], s[S_TILE_X], v[V_LANE_ID]))
-  k.emit(v_lshlrev_b32_e32(v[V_GLOBAL_B_ADDR], 2, v[V_GLOBAL_B_ADDR]))
-  k.emit(s_mul_i32(s[19], s[S_TILE_Y], N))
-  k.emit(v_mul_lo_u32(v[V_GLOBAL_A_ADDR], v[4], N))                          # (lane_id/8)*N
-  k.emit(v_add_nc_u32_e32(v[V_GLOBAL_A_ADDR], v[V_LANE_ID_MOD8], v[V_GLOBAL_A_ADDR]))  # + lane_id%8
-  k.emit(v_add_nc_u32_e32(v[V_GLOBAL_A_ADDR], s[19], v[V_GLOBAL_A_ADDR]))
-  k.emit(v_lshlrev_b32_e32(v[V_GLOBAL_A_ADDR], 2, v[V_GLOBAL_A_ADDR]))
-
-  # Do initial loads
-  for vdst, saddr_lo in INIT_PREFETCH:
-    k.emit(global_load_b32(vdst=v[vdst], addr=v[V_GLOBAL_B_ADDR], saddr=s[saddr_lo:saddr_lo+1]))
-  for iter in range(6):
-    vdst1, vdst2, addr, slo1, slo2 = PREFETCH_LOADS[iter]
-    k.emit(global_load_b32(vdst=v[vdst1], addr=v[addr], saddr=s[slo1:slo1+1]))
-    k.emit(global_load_b32(vdst=v[vdst2], addr=v[addr], saddr=s[slo2:slo2+1]))
-
-  # ===========================================================================
-  # LDS store address computation (bank-conflict-avoiding swizzle)
-  # ===========================================================================
-  # This section computes LDS store addresses with a swizzle pattern to avoid bank conflicts.
-  # The swizzle ensures that threads in the same wavefront write to different LDS banks.
-  # Formula: swizzled_addr = base + (lane_id & 7) * LDS_A_STRIDE + swizzle_offset
-  # where swizzle_offset depends on (lane_id >> 3) to distribute across banks.
-  k.emit(v_add_nc_u32_e32(v[9], s[S_LOOP_CTR], v[22]))  # row 0 base
-  k.emit(v_and_b32_e32(v[9], ADDR_MASK, v[9]))
-  k.emit(v_sub_nc_u32_e32(v[9], v[22], v[9]))  # row 0 swizzle offset
-  k.emit(v_lshlrev_b32_e32(v[9], 2, v[9]))  # * 4
-  k.emit(v_mad_u32_u24(v[V_LDS_B_ADDR], LDS_A_STRIDE, v[V_LANE_ID_MOD8], v[9]))
-
-  # For V_LDS_A_BASE and epilogue
-  k.emit(v_bfe_u32(v[2], v[V_LANE_ID], 3, 2))  # v[2] = (lane_id >> 3) & 3
-  k.emit(v_lshlrev_b32_e32(v[V_LANE_DIV8_X4], 2, v[2]))
-
-  # Compute LDS load/store base addresses for inner loop
-  k.emit(v_lshlrev_b32_e32(v[2], 4, v[2]))
-  k.emit(v_and_b32_e32(v[3], 0x7F, v[1]))  # simplified from 3 lines
-  k.emit(v_lshl_or_b32(v[V_LDS_B_BASE], v[V_LANE_ID_MOD8], 4, LDS_BASE_OFFSET))
-  k.emit(v_lshl_add_u32(v[V_LDS_A_ADDR], v[3], 2, LDS_BASE_OFFSET))
-  k.emit(v_lshlrev_b32_e32(v[3], 2, v[V_LANE_ID]))
-  k.emit(v_and_or_b32(v[V_LDS_A_BASE], 0x180, v[3], v[2]))
-
-  # Do initial stores
-  k.waitcnt(vm=0)
-  for i in range(4):  # A tile: 8 values via 4 stride64 stores
-    k.emit(ds_store_2addr_stride64_b32(addr=v[V_LDS_A_ADDR], data0=v[V_LDS_A_DATA[i*2]], data1=v[V_LDS_A_DATA[i*2+1]], offset0=i*4, offset1=i*4+2))
-  for i in range(8):  # B tile: 8 values via 8 scalar stores with 64-byte spacing
-    offset = i * 64
-    k.emit(ds_store_b32(addr=v[V_LDS_B_ADDR], data0=v[V_LDS_B_DATA[i]], offset0=offset & 0xFF, offset1=offset >> 8))
-
-  # Zero all 128 accumulators using VOPD dual moves (64 instructions instead of 128)
-  for i in range(0, len(OUT_REGS), 2):
-    k.emit(VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, vdstx=v[OUT_REGS[i]], vdsty=v[OUT_REGS[i+1]], srcx0=0, srcy0=0))
-  k.emit(s_add_i32(s[S_LOOP_BOUND], s[S_DIM_N], -8))
-
-  # S_LOOP_CTR is already 0 from prologue initialization
-  k.emit(s_branch(), target='LOOP_ENTRY')
-
-  # ===========================================================================
-  # MAIN GEMM LOOP
-  # ===========================================================================
-
-  NO_ALU, NO_DS, NO_GLOBAL = getenv("NO_ALU", 0), getenv("NO_DS", 0), getenv("NO_GLOBAL", 0)
-
-  k.label('LOOP_INC')
-  k.emit(s_add_i32(s[S_LOOP_CTR], s[S_LOOP_CTR], 8))
-  k.emit(s_cmp_ge_i32(s[S_LOOP_CTR], s[S_DIM_N]))
-  k.emit(s_cbranch_scc1(), target='EPILOGUE')
-
-  k.label('LOOP_ENTRY')
-  k.emit(s_cmp_lt_i32(s[S_LOOP_CTR], s[S_LOOP_BOUND]))
-  k.emit(s_cselect_b32(s[S_PREFETCH_FLAG], -1, 0))  # s_cselect doesn't modify SCC
-  k.emit(s_cbranch_scc0(), target='SKIP_PREFETCH')  # branch if loop_ctr >= loop_bound
-
-  if not NO_GLOBAL:
-    # Advance prefetch pointers (VGPR)
-    #k.emit(v_add_nc_u32_e32(v[V_GLOBAL_B_ADDR], N * 32, v[V_GLOBAL_B_ADDR]))
-    #k.emit(v_add_nc_u32_e32(v[V_GLOBAL_A_ADDR], 0x20, v[V_GLOBAL_A_ADDR]))
-
-    # Advance prefetch pointers (64-bit adds): B advances 8 rows (8*N*4 bytes), A advances 8 cols (8*4 bytes)
-    k.emit(s_clause(simm16=31))
-    for i in range(8):
-      k.emit(s_add_u32(s[S_PREFETCH_B+i*2], s[S_PREFETCH_B+i*2], N * 32))
-      k.emit(s_addc_u32(s[S_PREFETCH_B+i*2+1], s[S_PREFETCH_B+i*2+1], 0))
-    for i in range(8):
-      k.emit(s_add_u32(s[S_PREFETCH_A+i*2], s[S_PREFETCH_A+i*2], 0x20))
-      k.emit(s_addc_u32(s[S_PREFETCH_A+i*2+1], s[S_PREFETCH_A+i*2+1], 0))
-
-    # do the fetch
-    for vdst, saddr_lo in INIT_PREFETCH:
-      k.emit(global_load_b32(vdst=v[vdst], addr=v[V_GLOBAL_B_ADDR], saddr=s[saddr_lo:saddr_lo+1]))
-
-  k.label('SKIP_PREFETCH')
-
-  # wait for local stores to finish (either initial or loop)
-  # then sync the warp so it's safe to load local
-  k.waitcnt(lgkm=0)
-  k.emit(s_barrier())
-
-  # 8 inner loop iterations
-  for iter in range(8):
-    # Load A tile (4 pairs) and B tile (8 pairs) from LDS
-    if not NO_DS:
-      k.emit(s_clause(simm16=len(V_A_TILE_REGS) + len(V_B_TILE_REGS) - 1))  # 12 loads total: 4 A + 8 B
-      # A tile: 4 ds_load_b64
-      for i, vdst in enumerate(V_A_TILE_REGS):
-        a_off = (i & 1) * 8 + (i >> 1) * 64 + iter * LDS_A_STRIDE
-        k.emit(ds_load_b64(vdst=v[vdst:vdst+1], addr=v[V_LDS_A_BASE], offset0=a_off & 0xFF, offset1=a_off >> 8))
-      # B tile: 8 ds_load_b64
-      for i, vdst in enumerate(V_B_TILE_REGS):
-        b_off = (i & 1) * 8 + (i & 2) * 64 + (i >> 2) * 256 + iter * LDS_B_STRIDE
-        k.emit(ds_load_b64(vdst=v[vdst:vdst+1], addr=v[V_LDS_B_BASE], offset0=b_off & 0xFF, offset1=b_off >> 8))
-
-    # Issue global prefetch (first 6 iterations only)
-    if iter < 6 and not NO_GLOBAL:
-      vdst1, vdst2, addr, slo1, slo2 = PREFETCH_LOADS[iter]
-      k.emit(global_load_b32(vdst=v[vdst1], addr=v[addr], saddr=s[slo1:slo1+1]))
-      k.emit(global_load_b32(vdst=v[vdst2], addr=v[addr], saddr=s[slo2:slo2+1]))
-
-    # 64 dual FMACs
-    k.waitcnt(lgkm=0)
-    if not NO_ALU:
-      k.emit(s_clause(simm16=len(FMAC_PATTERN)-1))
-      for i, (vdst_x, vdst_y, ax, bx, ay, by) in enumerate(FMAC_PATTERN):
-        k.emit(VOPD(VOPDOp.V_DUAL_FMAC_F32, VOPDOp.V_DUAL_FMAC_F32,
-                    vdstx=v[vdst_x], vdsty=v[vdst_y], srcx0=v[ax], vsrcx1=v[bx], srcy0=v[ay], vsrcy1=v[by]))
-
-  # wait for all global loads to finish
-  # then sync the warp so it's safe to store local
-  k.waitcnt(vm=0)
-  k.emit(s_barrier())
-
-  # Store prefetched data to LDS
-  # NOTE: Register naming reflects LDS tile organization, not source matrix:
-  #   V_LDS_A_DATA (v155-162) holds data that goes to LDS A-tile region
-  #   V_LDS_B_DATA (v163-170) holds data that goes to LDS B-tile region
-  # The data sources are swapped: A-tile receives B matrix rows, B-tile receives A matrix columns
-  if not NO_DS:
-    for i in range(4):  # A tile: 8 values via 4 stride64 stores
-      k.emit(ds_store_2addr_stride64_b32(addr=v[V_LDS_A_ADDR], data0=v[V_LDS_A_DATA[i*2]], data1=v[V_LDS_A_DATA[i*2+1]], offset0=i*4, offset1=i*4+2))
-    for i in range(8):  # B tile: 8 values via 8 scalar stores with 64-byte spacing
-      offset = i * 64
-      k.emit(ds_store_b32(addr=v[V_LDS_B_ADDR], data0=v[V_LDS_B_DATA[i]], offset0=offset & 0xFF, offset1=offset >> 8))
-
-  k.emit(s_branch(), target='LOOP_INC')
-
-  # ===========================================================================
-  # EPILOGUE: Permute and store results
-  # ===========================================================================
-  k.label('EPILOGUE')
-
-  # Rearrange accumulators from FMAC layout to contiguous output order
-  for a, b in PERMUTE_SWAPS:
-    k.emit(v_swap_b32_e32(v[a], v[b]))
-
-  # Compute output base coordinates
-  # v[130] = col_base = tile_x + (lane_id & 7) * 4
-  # v[131] = row_base = tile_y + (lane_id & 0x60) + ((lane_id >> 3) & 3) * 4
-  # v[132] = 0 (for 64-bit address high part)
-  k.emit(v_add_nc_u32_e32(v[130], s[S_TILE_X], v[V_LANE_MOD8_X4]))
-  k.emit(v_and_b32_e32(v[131], 0x60, v[V_LANE_ID]))
-  k.emit(v_add_nc_u32_e32(v[131], s[S_TILE_Y], v[131]))
-  k.emit(v_add_nc_u32_e32(v[131], v[V_LANE_DIV8_X4], v[131]))
-  k.emit(v_mov_b32_e32(v[132], 0))
-
-  # Precompute row offsets: v[133-136] for rows 0-3, v[137-140] for rows 16-19
-  for base, row_off in [(133, 0), (137, 16)]:
-    if row_off: k.emit(v_add_nc_u32_e32(v[141], row_off, v[131]))
-    k.emit(v_mul_lo_u32(v[base], v[141] if row_off else v[131], s[S_DIM_N]))
-    for j in range(3): k.emit(v_add_nc_u32_e32(v[base + 1 + j], s[S_DIM_N], v[base + j]))
-
-  # s[S_PREFETCH_FLAG] = row stride in bytes (N * 4)
-  k.emit(s_lshl_b32(s[S_PREFETCH_FLAG], s[S_DIM_N], 2))
-
-  # Store 128 output values as 32 groups of 4 (128-bit stores)
-  # Layout: 2 row halves (0-3, 16-19) x 4 col groups x 4 rows = 32 stores of 4 floats
-  for i, (row_half, col_off, row_in_group) in enumerate([(rh, co, ri)
-      for rh in range(2) for co in [0, 32, 64, 96] for ri in range(4)]):
-    row = row_half * 16 + row_in_group
-    src = OUT_REGS[i*4]  # first reg of ascending group of 4
-
-    if row_in_group == 0:
-      # First row of group: compute full address
-      if col_off == 0: k.emit(v_mov_b32_e32(v[141], v[130]))
-      else: k.emit(v_add_nc_u32_e32(v[141], col_off, v[130]))
-      row_base = 133 + row if row < 4 else 137 + row - 16
-      k.emit(v_add_nc_u32_e32(v[141], v[row_base], v[141]))
-      k.emit(v_lshlrev_b32_e32(v[141], 2, v[141]))
-      k.emit(v_add_co_u32(v[141], VCC_LO, s[S_OUT_PTR[0]], v[141]))
-      k.emit(v_add_co_ci_u32_e32(v[142], s[S_OUT_PTR[1]], v[132]))
-    else:
-      # Subsequent rows: add stride
-      k.emit(v_add_co_u32(v[141], VCC_LO, s[S_PREFETCH_FLAG], v[141]))
-      k.emit(v_add_co_ci_u32_e32(v[142], v[142], v[132]))
-
-    k.emit(global_store_b128(addr=v[141:142], data=v[src:src+3], saddr=NULL))
-
-  k.emit(s_sendmsg(simm16=3))  # DEALLOC_VGPRS
-  k.emit(s_endpgm())
-
-  return k.finalize()
-
-# =============================================================================
-# Test harness
-# =============================================================================
-
-N = getenv("N", 4096)
-BLOCK_M, BLOCK_N = 128, 128
-THREADS = 128
-
-def test_matmul():
-  dev = Device[Device.DEFAULT]
-  print(f"Device arch: {dev.renderer.arch}")
-
-  insts = build_kernel(N, dev.renderer.arch)
-
-  rng = np.random.default_rng(42)
-  a = Tensor(rng.random((N, N), dtype=np.float32) - 0.5)
-  b = Tensor(rng.random((N, N), dtype=np.float32) - 0.5)
-  c = Tensor.empty(N, N)
-  Tensor.realize(a, b, c)
-
-  grid, local = (N // BLOCK_N, N // BLOCK_M, 1), (THREADS, 1, 1)
-  print(f"Grid: {grid}, Local: {local}")
-
-  dname:str = Device.DEFAULT
-  def asm_kernel(A:UOp, B:UOp, C:UOp) -> UOp:
-    gidxs = [UOp.special(n, f"gidx{i}") for i,n in enumerate(grid)]
-    lidxs = [UOp.special(n, f"lidx{i}") for i,n in enumerate(local)]
-    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=max(LDS_SIZE, 65536//getenv("LIMIT_OCC", 65536)), addrspace=AddrSpace.LOCAL), (), 'lds')
-    sink = UOp.sink(A.base, B.base, C.base, lds, *gidxs, *lidxs, arg=KernelInfo(name=colored("kernel", "cyan"),
-                                                                                  estimates=Estimates(ops=N*N*N*2, mem=N*N*4*3)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
-  c = Tensor.custom_kernel(a, b, c, fxn=asm_kernel)[2]
-  ei = c.schedule()[0].lower()
-
-  ets = []
-  with Context(DEBUG=2):
-    for _ in range(getenv("CNT", 5)): ets.append(ei.run(wait=True))
-  print(f"REAL TFLOPS {N * N * N * 2 / min(ets) * 1e-12:.2f}")
-
-  if getenv("VERIFY", 1):
-    GlobalCounters.reset()
-    with Context(DEBUG=2): tc = (a @ b).realize()
-    with Context(DEBUG=0): err = (c - tc).square().mean().item()
-    print(f"mean squared error {err}")
-    if err != err or err > 1e-06:
-      c_np, tc_np = c.numpy(), tc.numpy()
-      for bi in range(N // 128):
-        for bj in range(N // 128):
-          blk_c = c_np[bi*128:(bi+1)*128, bj*128:(bj+1)*128]
-          blk_ref = tc_np[bi*128:(bi+1)*128, bj*128:(bj+1)*128]
-          blk_diff = blk_c - blk_ref
-          zero_rows = [i for i in range(128) if np.all(np.abs(blk_c[i,:]) < 1e-10)]
-          nz_rows = [i for i in range(128) if i not in zero_rows]
-          nz_mse = float(np.mean(blk_diff[nz_rows,:]**2)) if nz_rows else 0
-          print(f"Block ({bi},{bj}): zero_rows={zero_rows}, nz_rows_mse={nz_mse:.2e}")
-          # show first few non-zero row comparisons
-          if nz_rows and nz_mse > 1e-6:
-            for r in nz_rows[:3]:
-              print(f"  row {r} asm[0:8]:  {blk_c[r,:8]}")
-              print(f"  row {r} ref[0:8]:  {blk_ref[r,:8]}")
-      raise RuntimeError("matmul is wrong!")
-
-if __name__ == "__main__":
-  test_matmul()
@@ -1,110 +0,0 @@
-from tinygrad import UOp, getenv
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-from tinygrad.dtype import AddrSpace, dtypes
-
-N = getenv("N", 4096)
-M = getenv("M", N)
-K = getenv("K", N)
-
-WARP_SIZE = 32
-BLOCK_M, BLOCK_N = 128, 128
-BLOCK_K = getenv("BK", 16)
-assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
-
-use_wmma = getenv("WMMA")
-if use_wmma:
-  WAVES_M, WAVES_N = 2, 2
-  LANES_PER_WAVE_M, LANES_PER_WAVE_N = 2, 16
-  UNROLL_M, UNROLL_N = 1, 1
-
-  # wmma params
-  WMMA_M, WMMA_N, WMMA_K = 16, 16, 16
-  WMMA_ACC = WMMA_M // LANES_PER_WAVE_M
-else:
-  WAVES_M, WAVES_N = 4, 1
-  LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
-  UNROLL_M, UNROLL_N = 4, 4
-
-# WARP_SIZE * total waves
-THREADS_PER_BLOCK = WARP_SIZE * WAVES_M * WAVES_N
-
-# accumulator size
-TM = BLOCK_M // (WAVES_M * LANES_PER_WAVE_M)
-TN = BLOCK_N // (WAVES_N * LANES_PER_WAVE_N)
-
-def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
-  wave_m = UOp.range(WAVES_M, 2, AxisType.LOCAL)
-  wave_n = UOp.range(WAVES_N, 3, AxisType.LOCAL)
-  lane = UOp.range(WARP_SIZE, -1, AxisType.WARP)
-  tid = (wave_m * WAVES_N + wave_n) * WARP_SIZE + lane
-
-  # -- GLOBAL -> LOCAL --
-  # wmma: spatial outer, k inner (k contiguous for vectorized WMMA tile loads)
-  # gemm: k outer, spatial inner
-  A_local = UOp.placeholder((BLOCK_M, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_M), a.dtype.base, slot=0, addrspace=AddrSpace.LOCAL)
-  B_local = UOp.placeholder((BLOCK_N, BLOCK_K) if use_wmma else (BLOCK_K, BLOCK_N), b.dtype.base, slot=1, addrspace=AddrSpace.LOCAL)
-
-  a = a.reshape(K // BLOCK_K, BLOCK_K, BLOCK_M)
-  b = b.reshape(K // BLOCK_K, BLOCK_K, BLOCK_N)
-  k_tile = UOp.range(K // BLOCK_K, 100, AxisType.REDUCE)
-
-  # copy with transpose for wmma (input is k×spatial, LDS is spatial×k)
-  A_copy = A_local.permute((1,0)) if use_wmma else A_local
-  B_copy = B_local.permute((1,0)) if use_wmma else B_local
-  A_store = A_copy.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(a[k_tile].reshape(-1, THREADS_PER_BLOCK)[:, tid])
-  B_store = B_copy.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(b[k_tile].reshape(-1, THREADS_PER_BLOCK)[:, tid])
-  barrier = UOp.barrier(A_store, B_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
-
-  # -- COMPUTE --
-  lane_m, lane_n = lane // LANES_PER_WAVE_N, lane % LANES_PER_WAVE_N
-
-  # accumulator (unified: both paths use (TM, TN) with scalar dtypes.float)
-  acc = UOp.placeholder((TM, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(UOp.const(dtypes.float, 0).reshape((1,)*len(acc.shape)).expand(acc.shape)))
-
-  if use_wmma:
-    k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)
-    tile_m = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-    tile_n = UOp.range(TN, 201, AxisType.LOOP)
-
-    acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0,2,1)[tile_m, tile_n]
-    a_frag = A_local.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_K // WMMA_K, WMMA_K)[wave_m, tile_m, lane_n, k]
-    b_frag = B_local.reshape(WAVES_N, TN, WMMA_N, BLOCK_K // WMMA_K, WMMA_K)[wave_n, tile_n, lane_n, k]
-
-    wmma = UOp(Ops.SHAPED_WMMA, dtypes.float, (a_frag, b_frag, acc_frag.after(k)), arg=((16, 16, 16), 'AMD', 32))
-    acc_store = acc_frag.store(wmma).end(tile_m, tile_n)
-  else:
-    # registers for LOCAL -> REG
-    a_frag = UOp.placeholder((TM//UNROLL_M, UNROLL_M), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-    b_frag = UOp.placeholder((TN//UNROLL_N, UNROLL_N), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-
-    k = UOp.range(BLOCK_K, 101, AxisType.REDUCE)
-    a_frag = a_frag.after(a_frag.store(A_local[k].reshape(WAVES_M, TM//UNROLL_M, LANES_PER_WAVE_M, UNROLL_M)[wave_m, :, lane_m, :]))
-    b_frag = b_frag.after(b_frag.store(B_local[k].reshape(WAVES_N, TN//UNROLL_N, LANES_PER_WAVE_N, UNROLL_N)[wave_n, :, lane_n, :]))
-
-    # FMA
-    a_frag = a_frag.reshape(TM, 1).expand(TM, TN)
-    b_frag = b_frag.reshape(1, TN).expand(TM, TN)
-    acc_store = acc.store(acc.after(k) + (a_frag * b_frag))
-
-  # store accumulator and loop
-  acc = acc.after(acc_store.end(k).barrier().end(k_tile))
-
-  # store accumulator to output (unified)
-  c = c.reshape(WAVES_M, TM//UNROLL_M, LANES_PER_WAVE_M, UNROLL_M,
-                WAVES_N, TN//UNROLL_N, LANES_PER_WAVE_N, UNROLL_N)
-  c = c.permute((0,4,2,6, 1,3,5,7)).reshape(THREADS_PER_BLOCK, TM, TN)
-  return c[tid].store(acc).end(wave_m, wave_n, lane)
-
-def amd_copy_matmul(c:UOp, a:UOp, b:UOp) -> UOp:
-  block_id_m = UOp.range(M // BLOCK_M, 0, AxisType.GLOBAL)
-  block_id_n = UOp.range(N // BLOCK_N, 1, AxisType.GLOBAL)
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
-  a = a.T.reshape(K, M // BLOCK_M, BLOCK_M)[:, block_id_m, :]
-  b = b.reshape(K, N // BLOCK_N, BLOCK_N)[:, block_id_n, :]
-  return block_128x128_gemm(c, a, b).end(block_id_n, block_id_m).sink(arg=KernelInfo(opts_to_apply=()))
-
-if __name__ == "__main__":
-  from amd_uop_matmul import eval_custom_matmul
-  eval_custom_matmul(amd_copy_matmul, dtypes.half if use_wmma else dtypes.float)
@@ -1,205 +0,0 @@
-from tinygrad import Tensor, UOp, getenv
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-from tinygrad.dtype import AddrSpace, dtypes
-from tinygrad.helpers import DEBUG, GlobalCounters, Context
-import math
-
-BLOCK_M, BLOCK_N = 64, 64
-WARP_SIZE = 32
-WMMA_M, WMMA_N, WMMA_K = 16, 16, 16
-WAVES_M, WAVES_N = 4, 1
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 2, 16
-WMMA_ACC = WMMA_M // LANES_PER_WAVE_M
-THREADS_PER_BLOCK = WARP_SIZE * WAVES_M * WAVES_N
-LDS_PAD = 4  # pad LDS rows to reduce bank conflicts
-
-WMMA_ARG = ((WMMA_M, WMMA_N, WMMA_K), 'AMD', 32)
-LOG2E = math.log2(math.e)
-
-def warp_shfl_xor(val, offset, lane):
-  """Read val from lane ^ offset using ds_bpermute."""
-  idx = ((lane ^ offset) * 4).cast(dtypes.int)
-  return UOp(Ops.CUSTOM, dtypes.float, (idx, val),
-             arg="__builtin_bit_cast(float, __builtin_amdgcn_ds_bpermute({0}, __builtin_bit_cast(int, {1})))")
-
-def warp_reduce_max(val, lane):
-  """Tree reduce MAX across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = UOp(Ops.MAX, dtypes.float, (val, warp_shfl_xor(val, offset, lane)))
-  return val
-
-def warp_reduce_sum(val, lane):
-  """Tree reduce SUM across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = val + warp_shfl_xor(val, offset, lane)
-  return val
-
-def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
-  # inputs are (B*H, N, D)
-  BH, N, D = q.shape
-  assert N % BLOCK_M == 0 and N % BLOCK_N == 0, f"N={N} must be divisible by BLOCK_M={BLOCK_M} and BLOCK_N={BLOCK_N}"
-  assert D % WMMA_K == 0 and D % LANES_PER_WAVE_N == 0, f"D={D} must be divisible by WMMA_K={WMMA_K} and LANES_PER_WAVE_N={LANES_PER_WAVE_N}"
-  assert BLOCK_M % (WAVES_M * WMMA_M) == 0 and BLOCK_N % LANES_PER_WAVE_N == 0
-  TM = BLOCK_M // (WAVES_M * LANES_PER_WAVE_M)
-  TN = BLOCK_N // (WAVES_N * LANES_PER_WAVE_N)
-  TD = D // (WAVES_N * LANES_PER_WAVE_N)
-  SCALE = 1.0 / math.sqrt(D)
-
-  block_bh = UOp.range(BH, 0, AxisType.GLOBAL)
-  block_m = UOp.range(N // BLOCK_M, 1, AxisType.GLOBAL)
-
-  q = q.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-  k = k.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  v = v.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  o = o.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-
-  wave_m = UOp.range(WAVES_M, 2, AxisType.LOCAL)
-  wave_n = UOp.range(WAVES_N, 3, AxisType.LOCAL)
-  lane = UOp.range(WARP_SIZE, -1, AxisType.WARP)
-  tid = (wave_m * WAVES_N + wave_n) * WARP_SIZE + lane
-  lane_m = lane // LANES_PER_WAVE_N
-  lane_n = lane % LANES_PER_WAVE_N
-
-  # LDS allocation: slot 0 = Q then P (shared), slot 1 = K then V
-  # TODO: the memory planner should be able to find this reuse
-  ELEMS_PER_THREAD = BLOCK_M * D // THREADS_PER_BLOCK
-  QP_lds = UOp.placeholder((BLOCK_M, D + LDS_PAD), dtypes.half, slot=0, addrspace=AddrSpace.LOCAL)
-  KV_lds = UOp.placeholder((BLOCK_N, D + LDS_PAD), dtypes.half, slot=1, addrspace=AddrSpace.LOCAL)[:, :D]
-
-  # register state
-  acc = UOp.placeholder((TM, TD), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  m_i = UOp.placeholder((TM,), dtypes.float, slot=3, addrspace=AddrSpace.REG)
-  l_i = UOp.placeholder((TM,), dtypes.float, slot=4, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(acc.const_like(0)))
-  m_i = m_i.after(m_i.store(m_i.const_like(-math.inf)))
-  l_i = l_i.after(l_i.store(l_i.const_like(0)))
-
-  # ====== KV tile loop ======
-  n_tile = UOp.range(N // BLOCK_N, 100, AxisType.REDUCE)
-
-  # load Q + K into LDS (Q reloaded each iteration since P overwrites slot 0)
-  Q_lds = QP_lds[:, :D]
-  Q_store = Q_lds.after(n_tile).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    q.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  K_store = KV_lds.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    k[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  qk_load_barrier = UOp.barrier(UOp.group(Q_store, K_store))
-  Q_lds = Q_lds.after(qk_load_barrier)
-  KV_lds_k = KV_lds.after(qk_load_barrier)
-
-  # -- S = Q @ K^T via WMMA (re-init each n_tile) --
-  S_reg = UOp.placeholder((TM, TN), dtypes.float, slot=6, addrspace=AddrSpace.REG)
-  S_reg = S_reg.after(S_reg.after(n_tile).store(S_reg.const_like(0)))
-  k_qk = UOp.range(D // WMMA_K, 101, AxisType.REDUCE)
-  tm1 = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-  tn1 = UOp.range(TN, 201, AxisType.LOOP)
-  S_frag = S_reg.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0, 2, 1)[tm1, tn1]
-  q_frag = Q_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, D // WMMA_K, WMMA_K)[wave_m, tm1, lane_n, k_qk]
-  k_frag = KV_lds_k.reshape(WAVES_N, TN, WMMA_N, D // WMMA_K, WMMA_K)[wave_n, tn1, lane_n, k_qk]
-  qk = UOp(Ops.SHAPED_WMMA, dtypes.float, (q_frag, k_frag, S_frag.after(k_qk)), arg=WMMA_ARG)
-  qk_done = S_frag.store(qk).end(tm1, tn1).end(k_qk)
-  S_reg = S_reg.after(qk_done)
-
-  # -- softmax in registers with warp shuffles --
-  S_reg = S_reg.after(S_reg.store(S_reg * SCALE))
-
-  # per-thread local row max over TN=4 elements, then warp reduce across 16 lanes
-  m_ij = UOp.placeholder((TM,), dtypes.float, slot=7, addrspace=AddrSpace.REG)
-  m_ij = m_ij.after(m_ij.after(n_tile).store(m_ij.const_like(-math.inf)))
-  rm2 = UOp.range(TN, 261, AxisType.REDUCE)
-  m_ij = m_ij.after(m_ij.store(m_ij.after(rm2).maximum(S_reg[:, rm2])).end(rm2))
-  # warp reduce max (in-place)
-  ri_w = UOp.range(TM, 270, AxisType.LOOP)
-  m_ij = m_ij.after(m_ij[ri_w].store(warp_reduce_max(m_ij[ri_w], lane)).end(ri_w))
-
-  # compute P = exp(S - m_ij) in S_reg
-  S_reg = S_reg.after(S_reg.store(((S_reg - m_ij.reshape(TM, 1).expand(TM, TN)) * LOG2E).exp2()))
-
-  p_local = UOp.placeholder((TM,), dtypes.float, slot=8, addrspace=AddrSpace.REG)
-  p_local = p_local.after(p_local.after(n_tile).store(p_local.const_like(0)))
-  rp2 = UOp.range(TN, 291, AxisType.REDUCE)
-  p_local = p_local.after(p_local.store(p_local.after(rp2) + S_reg[:, rp2]).end(rp2))
-  ri_ws = UOp.range(TM, 295, AxisType.LOOP)
-  p_sum = p_local.after(p_local[ri_ws].store(warp_reduce_sum(p_local[ri_ws], lane)).end(ri_ws))
-
-  # write P = exp(S - m_ij) to P_lds (reuses slot 0, Q no longer needed)
-  P_lds = QP_lds[:, :BLOCK_N]
-  P_write = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TN, LANES_PER_WAVE_N)
-  P_write = P_write.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TN)
-  # TODO: P_write[tid].store(S_reg.cast(dtypes.half)) — shaped store fails due to RESHAPE(DEFINE_LOCAL) surviving linearization
-  rw1 = UOp.range(TM, 296, AxisType.LOOP)
-  rw2 = UOp.range(TN, 297, AxisType.LOOP)
-  P_store = P_write[tid, rw1, rw2].store(S_reg[rw1, rw2].cast(dtypes.half)).end(rw1, rw2)
-
-  # -- online softmax correction --
-  ri4 = UOp.range(TM, 330, AxisType.LOOP)
-  m_new_val = m_i[ri4].maximum(m_ij[ri4])
-  alpha_val = ((m_i[ri4] - m_new_val) * LOG2E).exp2()
-  beta_val = ((m_ij[ri4] - m_new_val) * LOG2E).exp2()
-  rj4 = UOp.range(TD, 331, AxisType.LOOP)
-  correction = UOp.group(
-    acc[ri4, rj4].store(alpha_val * acc[ri4, rj4]).end(rj4),
-    l_i[ri4].store(alpha_val * l_i[ri4] + beta_val * p_sum[ri4]),
-    m_i[ri4].store(m_new_val),
-  ).end(ri4)
-  acc = acc.after(correction)
-  l_i = l_i.after(correction)
-  m_i = m_i.after(correction)
-
-  # load V into KV_lds (must wait for QK WMMA to finish reading K from KV_lds)
-  V_store = KV_lds.after(qk_done).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    v[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  pv_barrier = UOp.barrier(UOp.group(P_store, V_store))
-  P_lds = P_lds.after(pv_barrier)
-  KV_lds_v = KV_lds.after(pv_barrier)
-
-  # -- acc += P @ V via WMMA --
-  k_pv = UOp.range(BLOCK_N // WMMA_K, 400, AxisType.REDUCE)
-  tm2 = UOp.range(TM // WMMA_ACC, 401, AxisType.LOOP)
-  tn2 = UOp.range(TD, 402, AxisType.LOOP)
-  acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TD).permute(0, 2, 1)[tm2, tn2]
-  p_frag = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_N // WMMA_K, WMMA_K)[wave_m, tm2, lane_n, k_pv]
-  v_frag = KV_lds_v.reshape(WAVES_N, TD, WMMA_N, BLOCK_N // WMMA_K, WMMA_K)[wave_n, tn2, lane_n, k_pv]
-  pv = UOp(Ops.SHAPED_WMMA, dtypes.float, (p_frag, v_frag, acc_frag.after(k_pv)), arg=WMMA_ARG)
-
-  # end KV tile loop
-  n_tile_end = acc_frag.store(pv).end(tm2, tn2).end(k_pv).barrier().end(n_tile)
-  acc = acc.after(n_tile_end)
-  l_i = l_i.after(n_tile_end)
-  m_i = m_i.after(n_tile_end)
-
-  # normalize: acc /= l_i
-  acc = acc.after(acc.store(acc * (1 / l_i).reshape(TM, 1).expand(TM, TD)))
-
-  # store output
-  o = o.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TD, LANES_PER_WAVE_N)
-  o = o.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TD)
-  return o[tid].store(acc).end(wave_m, wave_n, lane).end(block_m, block_bh).sink(arg=KernelInfo(opts_to_apply=()))
-
-if __name__ == "__main__":
-  B, H, N, D = getenv("B", 1), getenv("H", 32), getenv("N", 1024), getenv("D", 64)
-  q = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  k = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  v = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  o = Tensor.empty(B, H, N, D, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(q, k, v)
-
-  q_flat, k_flat, v_flat, o_flat = q.reshape(B*H, N, D), k.reshape(B*H, N, D), v.reshape(B*H, N, D), o.reshape(B*H, N, D)
-  NUM_RUNS = getenv("CNT", 5)
-  ets = []
-  with Context(DEBUG=2):
-    for _ in range(NUM_RUNS):
-      GlobalCounters.reset()
-      tst = Tensor.custom_kernel(o_flat, q_flat, k_flat, v_flat, fxn=amd_flash_attention)[0].realize()
-      ets.append(GlobalCounters.time_sum_s)
-  print(f"best time: {min(ets)*1e3:.2f}ms")
-
-  if getenv("VERIFY", 1):
-    with Context(DEBUG=0):
-      ref = q.float().scaled_dot_product_attention(k.float(), v.float()).reshape(B*H, N, D).realize()
-      err = (ref - tst).square().mean().item()
-    print(f"mean squared error {err}")
-    if err > 1e-2:
-      raise RuntimeError("flash attention is wrong!")
-    else:
-      print("flash attention is correct!")
@@ -1,74 +1,98 @@
-from tinygrad import Tensor, Context, GlobalCounters, dtypes
+import numpy as np
+from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
 from tinygrad.uop.ops import UOp, KernelInfo, sint, AxisType
+from tinygrad.engine.realize import ExecItem, get_runner
 from tinygrad.dtype import AddrSpace
-from tinygrad.helpers import DEBUG, getenv
+from tinygrad.helpers import getenv

 N = getenv("N", 4096)
-M = getenv("M", N)
-K = getenv("K", N)
-NUM_RUNS = getenv("CNT", 5)
+M = K = N
+run_count = getenv("CNT", 5)

 # ---------------------------
 # launch/config constants
 # ---------------------------

 WARP_SIZE = 32
-BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 8
-TM, TN = 4, 4
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
-assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
+
+# Threadblock tile sizes (block-level tile of C that a block computes)
+BLOCK_N = 128   # columns of C (N-dim) per block
+BLOCK_M = 128   # rows of C (M-dim) per block
+BLOCK_K = 8     # K-slice per block iteration
+
+# Register tile sizes (per-thread accumulator tile of C)
+TN = 4     # columns per thread
+TM = 4     # rows per thread

 is_kernel5 = getenv("K5", 0)
 THREADS_PER_BLOCK = 128 if is_kernel5 else 256
-WAVES_PER_BLOCK_N = 1 if is_kernel5 else 2
-WAVES_PER_BLOCK_M = THREADS_PER_BLOCK // WARP_SIZE // WAVES_PER_BLOCK_N
-REG_TILES_PER_WAVE_N = BLOCK_N // (WAVES_PER_BLOCK_N * LANES_PER_WAVE_N * TN)
-REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
+assert THREADS_PER_BLOCK % BLOCK_N == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_N"
+assert THREADS_PER_BLOCK % BLOCK_K == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_K"
+assert (BLOCK_N * BLOCK_K) % THREADS_PER_BLOCK == 0
+assert (BLOCK_M * BLOCK_K) % THREADS_PER_BLOCK == 0

-assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
-assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"
+WARPS_PER_BLOCK = THREADS_PER_BLOCK // WARP_SIZE
+WAVE_TILE_N = 128 if is_kernel5 else 64
+WAVE_TILE_M = BLOCK_N * BLOCK_M // WARPS_PER_BLOCK // WAVE_TILE_N
+assert BLOCK_N % WAVE_TILE_N == 0, "BN must be a multiple of WN"
+assert BLOCK_M % WAVE_TILE_M == 0, "BM must be a multiple of WM"
+WAVES_IN_BLOCK_X = BLOCK_N // WAVE_TILE_N
+WAVES_IN_BLOCK_Y = BLOCK_M // WAVE_TILE_M
+assert WAVES_IN_BLOCK_X * WAVES_IN_BLOCK_Y == WARPS_PER_BLOCK, "wave grid must match warps/block"
+
+LANES_PER_WAVE_X = 8
+LANES_PER_WAVE_Y = 4
+ITERS_PER_WAVE_N = WAVE_TILE_N // (LANES_PER_WAVE_X * TN)
+ITERS_PER_WAVE_M = WAVE_TILE_M // (LANES_PER_WAVE_Y * TM)
+assert WAVE_TILE_N % (LANES_PER_WAVE_X * TN) == 0, "WAVE_TILE_N must be divisible by LANES_PER_WAVE_X*TN"
+assert WAVE_TILE_M % (LANES_PER_WAVE_Y * TM) == 0, "WAVE_TILE_M must be divisible by LANES_PER_WAVE_Y*TM"

 def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
-def copy(dest:UOp, src:UOp, rng:int, upcast=False):
+def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=False):
  assert dest.shape == src.shape
  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
-  return dest[*rngs].store(src[*rngs]).end(*rngs)
+  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
+  return dest.after(copy) if set else copy

-def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
+def hand_spec_kernel3():
  # ---------------------------
-  # block indices
+  # block indices & placeholders
  # ---------------------------
-  block_id_n = UOp.special(N // BLOCK_N, "gidx0")
-  block_id_m = UOp.special(M // BLOCK_M, "gidx1")
+  blockIdx_x = UOp.special(N // BLOCK_N, "gidx0")
+  blockIdx_y = UOp.special(N // BLOCK_M, "gidx1")
+
+  a = UOp.placeholder((N, N), dtypes.float, slot=1)
+  b = UOp.placeholder((N, N), dtypes.float, slot=2)
+  c = UOp.placeholder((N, N), dtypes.float, slot=0)

  # index the output with the globals
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
+  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[blockIdx_y, :, blockIdx_x, :]

  # open the main reduction range
-  k_tile_range = UOp.range(K // BLOCK_K, 0, AxisType.REDUCE)
-  a = a.reshape(M // BLOCK_M, BLOCK_M, K // BLOCK_K, BLOCK_K)[block_id_m, :, k_tile_range, :]
-  b = b.reshape(K // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, block_id_n, :]
+  k_tile_range = UOp.range(N // BLOCK_K, 0, AxisType.REDUCE)
+  a = a.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_K, BLOCK_K)[blockIdx_y, :, k_tile_range, :]
+  b = b.reshape(N // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, blockIdx_x, :]

  # globals are no longer used, they are already in the indexes
-  del block_id_m, block_id_n
+  del blockIdx_y, blockIdx_x

  # ---------------------------
-  # GLOBAL -> LOCAL (A_local, B_local)
+  # GLOBAL -> LOCAL (As, Bs)
  # ---------------------------
  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")

  # A: read BM x BK tiles (permute on store into locals)
-  BM_A_local_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
-  A_local = UOp.placeholder((BLOCK_K, BM_A_local_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
-  A_local_store = copy(A_local.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
+  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
+  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
+  As_store = copy(As.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)

  # B: read BK x BN tiles
-  B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
-  B_local_store = copy(B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
+  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+  Bs_store = copy(Bs.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)

  # TODO: can we automate barrier?
-  barrier = UOp.barrier(A_local_store, B_local_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
+  barrier = UOp.barrier(As_store, Bs_store)
+  As, Bs = As.after(barrier), Bs.after(barrier)

  # open inner k range
  k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
@@ -76,30 +100,31 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
  # ---------------------------
  # LOCAL -> REG (per-wave tiles)
  # ---------------------------
-  warp, lane = tid // WARP_SIZE, tid % WARP_SIZE
-  waveIdx, waveIdy = warp % WAVES_PER_BLOCK_N, warp // WAVES_PER_BLOCK_N
-  laneIdx, laneIdy = lane % LANES_PER_WAVE_N, lane // LANES_PER_WAVE_N
-  assert waveIdy.vmax+1 == WAVES_PER_BLOCK_M and laneIdy.vmax+1 == LANES_PER_WAVE_M
+  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
+  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
+  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y

-  A_col = UOp.placeholder((REG_TILES_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  A_local_slice = A_local[k, :].reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM)[waveIdy, :, laneIdy, :]
-  A_col = A_col.after(copy(A_col, A_local_slice, 300, upcast=True))
+  laneIdx = (tid % WARP_SIZE) % LANES_PER_WAVE_X
+  laneIdy = (tid % WARP_SIZE) // LANES_PER_WAVE_X
+  assert laneIdy.vmax+1 == LANES_PER_WAVE_Y

-  B_row = UOp.placeholder((REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-  B_local_slice = B_local[k, :].reshape(WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)[waveIdx, :, laneIdx, :]
-  B_row = B_row.after(copy(B_row, B_local_slice, 400, upcast=True))
+  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  A_col = copy(A_col, As[k, :].reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)[waveIdy, :, laneIdy, :], 300, set=True, upcast=True)
+
+  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+  B_row = copy(B_row, Bs[k, :].reshape(WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)[waveIdx, :, laneIdx, :], 400, set=True, upcast=True)

  # ---------------------------
  # FMA: c_regs += A_col * B_row
  # ---------------------------
-  c_regs = UOp.placeholder((REG_TILES_PER_WAVE_M, TM, REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
  i = UOp.range(c_regs.size, 16)
  c_regs = c_regs.after(c_regs.flatten()[i].store(0.0).end(i))

  # TODO: why don't these work as upcast?
  # why if the ranges merge is it slow?!? (if you change the order on end, they will merge. big slowdown on METAL)
-  iter_m, t_m, iter_n, t_n = rngs = rngs_for_shape(c_regs.shape, 500)
-  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iter_m, t_m] * B_row[iter_n, t_n]).end(iter_m, iter_n, t_m, t_n)
+  iterWaveM, yt, iterWaveN, xt = rngs = rngs_for_shape(c_regs.shape, 500)
+  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)

  # Close k, sync, and close K tiles
  sink = sink.end(k).barrier().end(k_tile_range)
@@ -107,37 +132,38 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
  # ---------------------------
  # REG -> GLOBAL (epilogue)
  # ---------------------------
-  c = c.reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM,
-                WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)
+  c = c.reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
+                WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
  c = c[waveIdy, :, laneIdy, :,
        waveIdx, :, laneIdx, :]
  sink = copy(c, c_regs.after(sink), rng=600)

  return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()

-def eval_custom_matmul(fxn, dt=dtypes.float):
-  a = Tensor.randn(M, K, dtype=dt)
-  b = Tensor.randn(K, N, dtype=dt)
-  c = Tensor.empty(M, N, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(a, b)
+def test_matmul(sink:UOp, N=N):
+  rng = np.random.default_rng()
+  a = Tensor(rng.random((N, N), dtype=np.float32)-0.5)
+  b = Tensor(rng.random((N, N), dtype=np.float32)-0.5)
+  hc = Tensor.empty(N, N)
+  Tensor.realize(a, b, hc)
+
+  ei = ExecItem(sink, [t.uop.buffer for t in [hc, a, b]], prg=get_runner(Device.DEFAULT, sink))

  ets = []
-  with Context(DEBUG=max(2, DEBUG.value), DEVECTORIZE=2 if dt == dtypes.half else 0):
-    for _ in range(NUM_RUNS):
-      GlobalCounters.reset()
-      tst = Tensor.custom_kernel(c, a, b, fxn=fxn)[0].realize()
-      ets.append(GlobalCounters.time_sum_s)
-  print(f"REAL TFLOPS {M * N * K * 2 / min(ets) * 1e-12:.2f}")
+  with Context(DEBUG=2):
+    for _ in range(run_count):
+      ets.append(ei.run(wait=True))
+  print(f"REAL TFLOPS {N * N * N * 2 / min(ets) * 1e-12:.2f}")

  if getenv("VERIFY", 1):
    GlobalCounters.reset()
    with Context(DEBUG=2):
-      tc = (a.float() @ b.float()).realize()
+      tc = (a @ b).realize()
    with Context(DEBUG=0):
-      err = (tc - tst).square().mean().item()
+      err = (hc - tc).square().mean().item()
    print(f"mean squared error {err}")
-    if err > (1e-2 if dt == dtypes.half else 1e-6):
+    if err > 1e-06:
      raise RuntimeError("matmul is wrong!")

 if __name__ == "__main__":
-  eval_custom_matmul(hand_spec_kernel3)
+  test_matmul(hand_spec_kernel3(), N=N)
@@ -37,7 +37,7 @@ b.copyin(row.data)
 c.copyin(mat.data)
 ret = prog(a._buf, b._buf, c._buf, global_size=[1,1,1], local_size=[8,1,1], wait=True)
 print(ret)
-out = np.frombuffer(a.as_memoryview(), np.float32)
+out = np.frombuffer(a.as_buffer(), np.float32)
 real = row.astype(np.float32)@mat.T.astype(np.float32)
 print("out:", out)
 print("real", real)
@@ -98,10 +98,10 @@ if __name__ == "__main__":
  # check correctness
  if getenv("VERIFY"):
    from tinygrad.engine.realize import run_schedule
-    triton_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
+    triton_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
    print(triton_buf)
    run_schedule(sched)
-    tinygrad_buf = np.frombuffer(si.bufs[0].as_memoryview(), np.float16).reshape(M,N)
+    tinygrad_buf = np.frombuffer(si.bufs[0].as_buffer(), np.float16).reshape(M,N)
    print(tinygrad_buf)
    np.testing.assert_allclose(triton_buf, tinygrad_buf)
    print("correct!")
@@ -1,15 +1,14 @@
 #!/usr/bin/env python3

 import argparse, glob, os, time, subprocess, sys
-from tinygrad.helpers import temp

 def scan_devs_based_on_lock(prefix:str, args) -> list[str]:
  target_dev = args.pci_bus if 'pci_bus' in args.__dir__() else ""

  devs = []
-  for dev in glob.glob(temp(f'{prefix}_*.lock')):
-    dev_id = dev.split('/')[-1][len(prefix)+1:-5]
-    if dev_id.startswith(target_dev): devs.append(dev_id)
+  for dev in glob.glob(f'/tmp/{prefix}_*.lock'):
+    dev_id = dev[8:-5]
+    if os.path.exists(f"/sys/bus/pci/devices/{dev_id}") and dev_id.startswith(target_dev): devs.append(dev_id)
  return devs

 def _do_reset_device(pci_bus): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/devices/{pci_bus}/reset'")
@@ -54,7 +53,16 @@ def cmd_show_pids(args):

  for dev in devs:
    try:
-      pid = subprocess.check_output(['sudo', 'lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
+      pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
+      print(f"{dev}: {pid}")
+    except subprocess.CalledProcessError: print(f"{dev}: No processes found using this device")
+
+def cmd_kill_pids(args):
+  devs = scan_devs_based_on_lock(prefix:={"amd":"am", "nv":"nv"}[args.backend], args)
+
+  for dev in devs:
+    try:
+      pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
      print(f"{dev}: {pid}")
    except subprocess.CalledProcessError: print(f"{dev}: No processes found using this device")

@@ -66,7 +74,7 @@ def cmd_kill_pids(args):
      if i > 0: time.sleep(0.2)

      try:
-        try: pid = subprocess.check_output(['sudo', 'lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
+        try: pid = subprocess.check_output(['sudo', 'lsof', f'/tmp/{prefix}_{dev}.lock']).decode('utf-8').strip().split('\n')[1].split()[1]
        except subprocess.CalledProcessError: break

        print(f"Killing process {pid} (which uses {dev})")
@@ -1,42 +1,8 @@
-import argparse, os, hashlib, functools
-from typing import Iterator, Callable
-from tinygrad.helpers import getenv, DEBUG, round_up, Timing, tqdm, fetch, ceildiv
+import argparse, os, hashlib
+from tinygrad.helpers import getenv, DEBUG, round_up, Timing, tqdm, fetch
 from extra.hevc.hevc import parse_hevc_file_headers, untile_nv12, to_bgr, nv_gpu
 from tinygrad import Tensor, dtypes, Device, Variable, TinyJit

-# rounds up hevc input data to 32 bytes, so more optimal kernels can be generated
-HEVC_ROUNDUP = getenv("DATA_ROUNDUP", 32)
-
-@functools.cache
-def _hevc_jitted_decoder(out_image_size:tuple[int, int], max_hist:int, inplace:bool):
-  def hevc_decode_frame(pos:Variable, hevc_tensor:Tensor, offset:Variable, sz:Variable, opaque:Tensor, i:Variable, *hist:Tensor, outbuf:Tensor|None=None):
-    x = hevc_tensor[offset:offset+sz*HEVC_ROUNDUP].decode_hevc_frame(pos, out_image_size, opaque[i], hist).realize()
-    if outbuf is not None: outbuf.assign(x).realize()
-    return x
-  return TinyJit(hevc_decode_frame)
-
-def hevc_decode(hevc_tensor:Tensor, opaque:Tensor, frame_info:list, luma_h:int, luma_w:int,
-                history:list[Tensor]|None=None, preallocated_outputs:list[Tensor]|None=None, warmup=False) -> Iterator[Tensor]:
-  out_image_size = luma_h + (luma_h + 1) // 2, round_up(luma_w, 64)
-  max_hist = max((hs for _, _, _, hs, _ in frame_info), default=0)
-
-  v_pos = Variable("pos", 0, max_hist + 1)
-  v_offset = Variable("offset", 0, hevc_tensor.numel()-1)
-  v_sz = Variable("sz", 1, ceildiv(hevc_tensor.numel(), HEVC_ROUNDUP))
-  v_i = Variable("i", 0, len(frame_info)-1)
-
-  decode_jit = _hevc_jitted_decoder(out_image_size, max_hist, preallocated_outputs is not None)
-  history = history or [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV").contiguous().realize() for _ in range(max_hist)]
-  assert len(history) == max_hist, f"history length {len(history)} does not match max_hist {max_hist}"
-
-  for i, (offset, sz, frame_pos, _, is_hist) in enumerate(frame_info):
-    history = history[-max_hist:] if max_hist > 0 else []
-    img = decode_jit(v_pos.bind(frame_pos), hevc_tensor, v_offset.bind(offset), v_sz.bind(ceildiv(sz, HEVC_ROUNDUP)),
-                     opaque, v_i.bind(i), *history, outbuf=preallocated_outputs[i] if preallocated_outputs else None)
-    res = preallocated_outputs[i] if preallocated_outputs else img.clone().realize()
-    if is_hist: history.append(res)
-    yield res
-
 if __name__ == "__main__":
  parser = argparse.ArgumentParser()
  parser.add_argument("--input_file", type=str, default="")
@@ -53,36 +19,49 @@ if __name__ == "__main__":
  dat_hash = hashlib.md5(dat).hexdigest()

  with Timing("prep infos: "):
+    dat_nv = hevc_tensor.to("NV")
    opaque, frame_info, w, h, luma_w, luma_h, chroma_off = parse_hevc_file_headers(dat)

  frame_info = frame_info[:getenv("MAX_FRAMES", len(frame_info))]

  # move all needed data to gpu
+  #all_slices = []
  with Timing("copy to gpu: "):
    opaque_nv = opaque.to("NV").contiguous().realize()
    hevc_tensor = hevc_tensor.to("NV")

  out_image_size = luma_h + (luma_h + 1) // 2, round_up(luma_w, 64)
+  max_hist = max(history_sz for _, _, _, history_sz, _ in frame_info)

-  # preallocate output/hist buffers
-  max_hist = max((hs for _, _, _, hs, _ in frame_info), default=0)
-  hist = [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV").contiguous().realize() for _ in range(max_hist)]
-  out_images = [Tensor.zeros(*out_image_size, dtype=dtypes.uint8, device="NV").contiguous().realize() for _ in range(len(frame_info))]
+  # define variables
+  v_pos = Variable("pos", 0, max_hist + 1)
+  v_offset = Variable("offset", 0, hevc_tensor.numel()-1)
+  v_sz = Variable("sz", 0, hevc_tensor.numel())
+  v_i = Variable("i", 0, len(frame_info)-1)

-  # warmup decode
-  _ = list(hevc_decode(hevc_tensor, opaque_nv, frame_info[:3], luma_h, luma_w, history=hist, preallocated_outputs=out_images))
-  Device.default.synchronize()
+  @TinyJit
+  def decode_jit(pos:Variable, src:Tensor, data:Tensor, *hist:Tensor):
+    return src.decode_hevc_frame(pos, out_image_size, data, hist).realize()
+
+  # warm up
+  history = [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV") for _ in range(max_hist)]
+  for i in range(3):
+    hevc_frame = hevc_tensor.shrink((((bound_offset:=v_offset.bind(frame_info[0][0])), bound_offset+v_sz.bind(frame_info[0][1])),))
+    decode_jit(v_pos.bind(0), hevc_frame, opaque_nv[v_i.bind(0)], *history)
+
+  out_images = []
+  with Timing("decoding whole file: ", on_exit=(lambda et: f", {len(frame_info)} frames, {len(frame_info)/(et/1e9):.2f} fps")):
+    for i, (offset, sz, frame_pos, history_sz, is_hist) in enumerate(frame_info):
+      history = history[-max_hist:] if max_hist > 0 else []
+      # TODO: this shrink should work as a slice
+      hevc_frame = hevc_tensor.shrink((((bound_offset:=v_offset.bind(offset)), bound_offset+v_sz.bind(sz)),))
+
+      outimg = decode_jit(v_pos.bind(frame_pos), hevc_frame, opaque_nv[v_i.bind(i)], *history).clone()
+      out_images.append(outimg)
+      if is_hist: history.append(outimg)

-  # decode all frames using the iterator
-  tm = Timing("decoding whole file: ", on_exit=(lambda et: f", {len(frame_info)} frames, {len(frame_info)/(et/1e9):.2f} fps"))
-  with tm:
-    images = list(hevc_decode(hevc_tensor, opaque_nv, frame_info, luma_h, luma_w, history=hist, preallocated_outputs=out_images))
    Device.default.synchronize()

-  fps = len(frame_info)/(tm.et/1e9)
-  assert fps >= getenv("ASSERT_FPS", 0), f"HEVC decode too slow: {fps:.2f} fps"
-
-  # validation
  if getenv("VALIDATE", 0):
    import pickle
    if dat_hash == "b813bfdbec194fd17fdf0e3ceb8cea1c":
@@ -91,7 +70,7 @@ if __name__ == "__main__":
    else: decoded_frames = pickle.load(open(f"extra/hevc/decoded_frames_{dat_hash}.pkl", "rb"))
  else: import cv2

-  for i, img in tqdm(enumerate(images)):
+  for i, img in tqdm(enumerate(out_images)):
    if getenv("VALIDATE", 0):
      if i < len(decoded_frames) and len(decoded_frames[i]) > 0:
        img = untile_nv12(img, h, w, luma_w, chroma_off).realize()
@@ -13,7 +13,7 @@ def get_struct(argp, stype):

 def format_struct(s):
  sdats = []
-  for field_name, *_ in s._real_fields_:
+  for field_name, field_type in s._fields_:
    dat = getattr(s, field_name)
    if isinstance(dat, int): sdats.append(f"{field_name}:0x{dat:X}")
    else: sdats.append(f"{field_name}:{dat}")
@@ -46,22 +46,6 @@ def install_hook(c_function, python_function):
 # *** ioctl lib end ***

 import tinygrad.runtime.autogen.kfd as kfd_ioctl
-import tinygrad.runtime.autogen.hsa as hsa
-
-def print_aql_queue(read_pointer_address):
-  rptr_offset = getattr(hsa.amd_queue_v2_t, 'read_dispatch_id').offset
-  queue_base = read_pointer_address - rptr_offset
-  queue = hsa.amd_queue_v2_t.from_address(queue_base)
-  print(f"  AQL Queue @ 0x{queue_base:X}:")
-  for field_name, *_ in hsa.amd_queue_v2_t._real_fields_:
-    val = getattr(queue, field_name)
-    if isinstance(val, int): print(f"    {field_name}: 0x{val:X}")
-    elif hasattr(val, '_length_'):
-      arr_vals = [f"{format_struct(v)}" if hasattr(v, '_real_fields_') else f"{v:#X}" for v in val]
-      print(f"    {field_name}: [{', '.join(arr_vals)}]")
-    elif hasattr(val, '_real_fields_'): print(f"    {field_name}: {format_struct(val)}")
-    else: print(f"    {field_name}: {val}")
-
 def ioctls_from_header():
  hdr = (pathlib.Path(__file__).parent / "kfd_ioctl.h").read_text().replace("\\\n", "")
  pattern = r'#define\s+(AMDKFD_IOC_[A-Z0-9_]+)\s+AMDKFD_IOW?R?\((0x[0-9a-fA-F]+),\s+struct\s([A-Za-z0-9_]+)\)'
@@ -83,7 +67,6 @@ def ioctl(fd, request, argp):
    if name == "AMDKFD_IOC_SVM":
      out = ctypes.cast(s.attrs, ctypes.POINTER(kfd_ioctl.struct_kfd_ioctl_svm_attribute))
      for i in range(s.nattr): print(f"{i}: {kfd_ioctl.enum_kfd_ioctl_svm_attr_type.get(out[i].type):40s}: {out[i].value:#x}")
-    if name == "AMDKFD_IOC_CREATE_QUEUE" and s.queue_type == kfd_ioctl.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL: print_aql_queue(s.read_pointer_address)
  else:
    print(f"{(st-start)*1000:7.2f} ms +{et*1000.:7.2f} ms : ioctl",
          f"{idir=} {size=} {itype=} {nr=} {fd=} {ret=}", os.readlink(f"/proc/self/fd/{fd}") if fd >= 0 else "")
@@ -18,7 +18,7 @@ prg = dev.runtime("write_ones", mbin)
 prg(buf0._buf, global_size=(1,65537,1), local_size=(1,1,1), wait=True)

 import numpy as np
-def to_np(buf): return np.frombuffer(buf.as_memoryview().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))
+def to_np(buf): return np.frombuffer(buf.as_buffer().cast(buf.dtype.base.fmt), dtype=_to_np_dtype(buf.dtype.base))

 big = to_np(buf0)
 print(big)
@@ -8,14 +8,14 @@ from tinygrad.helpers import _ensure_downloads_dir
 DOWNLOADS_DIR = _ensure_downloads_dir() / "models"
 from tinygrad.helpers import tqdm

-def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, local_dir: str|Path|None=None,
+def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, cache_dir: str|Path|None=None,
                                 tries: int=2, **kwargs) -> Path:
  for attempt in range(tries):
    try:
      return Path(snapshot_download(
        repo_id=repo_id,
        allow_patterns=allow_patterns,
-        local_dir=str(local_dir) if local_dir is not None else None,
+        cache_dir=str(cache_dir) if cache_dir is not None else None,
        **kwargs
      ))
    except Exception as e:
@@ -144,14 +144,14 @@ class HuggingFaceONNXManager:
        root_path = snapshot_download_with_retry(
          repo_id=model_id,
          allow_patterns=allow_patterns,
-          local_dir=str(self.models_dir / model_id)
+          cache_dir=str(self.models_dir)
        )

        # Download config files (usually small)
        snapshot_download_with_retry(
          repo_id=model_id,
          allow_patterns=["*config.json"],
-          local_dir=str(self.models_dir / model_id)
+          cache_dir=str(self.models_dir)
        )

        model_data["download_path"] = str(root_path)
@@ -88,8 +88,8 @@ if __name__ == "__main__":
      # repo id
      # validates all onnx models inside repo
      repo_id = "/".join(path)
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], local_dir=DOWNLOADS_DIR / repo_id)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=DOWNLOADS_DIR)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
      config = get_config(root_path)
      for onnx_model in root_path.rglob("*.onnx"):
        rtol, atol = get_tolerances(onnx_model.name)
@@ -101,8 +101,8 @@ if __name__ == "__main__":
      onnx_model = path[-1]
      assert path[-1].endswith(".onnx")
      repo_id, relative_path = "/".join(path[:2]), "/".join(path[2:])
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], local_dir=DOWNLOADS_DIR / repo_id)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], local_dir=DOWNLOADS_DIR / repo_id)
+      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=DOWNLOADS_DIR)
+      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
      config = get_config(root_path)
      rtol, atol = get_tolerances(onnx_model)
      print(f"validating {relative_path} with truncate={args.truncate}, {rtol=}, {atol=}")
@@ -1,105 +1,99 @@
-import os
+import os, pathlib

 # TODO: there is a timing bug without this
 os.environ["AMD_AQL"] = "1"

-from tinygrad import Tensor, Device
-from tinygrad.helpers import getenv
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.renderer import Estimates
-from tinygrad.renderer.amd.dsl import Reg, Inst, s, v
+from tinygrad.device import Device
+from tinygrad.runtime.ops_amd import AMDProgram, HIPCompiler

 NUM_WORKGROUPS = 96
 WAVE_SIZE = 32
-NUM_WAVES = 4
+NUM_WAVES = 2
 FLOPS_PER_MATMUL = 16*16*16*2
-INTERNAL_LOOP = getenv("LOOP", 10_000)
+INTERNAL_LOOP = 1_000_00
 INSTRUCTIONS_PER_LOOP = 200
+DIRECTIVE = ".amdhsa_wavefront_size32 1"

-def repeat(insts:list[Inst], n:int, counter_sreg:Reg) -> list[Inst]:
-  insts_bytes = b"".join([inst.to_bytes() for inst in insts])
-  sub_inst, cmp_inst = s_sub_u32(counter_sreg, counter_sreg, 1), s_cmp_lg_i32(counter_sreg, 0)
-  loop_sz = len(insts_bytes) + sub_inst.size() + cmp_inst.size()
-  branch_inst = s_cbranch_scc1(simm16=-((loop_sz // 4) + 1) & 0xFFFF)
-  return [s_mov_b32(counter_sreg, n)] + insts + [sub_inst, cmp_inst, branch_inst, s_endpgm()]
+assemblyTemplate = (pathlib.Path(__file__).parent / "template.s").read_text()

-def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, **kwargs):
+def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, extra=""):
  if accum:
-    inst = instruction(v[0:vgprIndices[0]], v[vgprIndices[1]:vgprIndices[2]], v[vgprIndices[1]:vgprIndices[2]], 1, acc_cd=1, **kwargs)
+    instructions = "{} a[0:{}], v[{}:{}], v[{}:{}], 1{}\n".format(instruction, vgprIndices[0],
+                                                                vgprIndices[1], vgprIndices[2],
+                                                                vgprIndices[1], vgprIndices[2], extra)
  elif dense:
-    inst = instruction(v[0:vgprIndices[0]], v[vgprIndices[1]:vgprIndices[2]], v[vgprIndices[1]:vgprIndices[2]], 1)
+    instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], 1\n".format(instruction, vgprIndices[0],
+                                                                vgprIndices[1], vgprIndices[2],
+                                                                vgprIndices[1], vgprIndices[2])
  else:
-    inst = instruction(v[0:vgprIndices[0]], v[vgprIndices[1]:vgprIndices[2]], v[vgprIndices[3]:vgprIndices[4]], v[vgprIndices[5]])
-  insts = repeat([inst for _ in range(INSTRUCTIONS_PER_LOOP)], n=INTERNAL_LOOP, counter_sreg=s[1])
-  def fxn(A:UOp) -> UOp:
-    threads = UOp.special(WAVE_SIZE * NUM_WAVES, "lidx0")
-    gidx = UOp.special(NUM_WORKGROUPS, "gidx0")
-    FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
-    sink = UOp.sink(A.base, threads, gidx, arg=KernelInfo(inst.op.name.lower(), estimates=Estimates(ops=FLOPs, mem=0)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
-  dummy = Tensor.zeros(1).contiguous().realize()
-  out = Tensor.custom_kernel(dummy, fxn=fxn)[0]
-  ei = out.schedule()[-1].lower()
-  elapsed = min([ei.run(wait=True) for _ in range(2)])
+    instructions = "{} v[0:{}], v[{}:{}], v[{}:{}], v{}\n".format(instruction, vgprIndices[0],
+                                                                  vgprIndices[1], vgprIndices[2],
+                                                                  vgprIndices[3], vgprIndices[4],
+                                                                  vgprIndices[5])
+  src = assemblyTemplate.replace("INTERNAL_LOOP", str(INTERNAL_LOOP)).replace("INSTRUCTION", instructions*INSTRUCTIONS_PER_LOOP)
+  src = src.replace("DIRECTIVE", DIRECTIVE)
+  lib = COMPILER.compile(src)
+  fxn = AMDProgram(DEV, "matmul", lib)
+  elapsed = min([fxn(global_size=(NUM_WORKGROUPS,1,1), local_size=(WAVE_SIZE*NUM_WAVES,1,1), wait=True) for _ in range(2)])
  FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
-  print(f"{inst.op_name.lower():<29} : {FLOPs/elapsed/10**12:.2f} T(FL)OPS")
+  print(f"{instruction:<29} : {FLOPs/elapsed/10**12:.2f} T(FL)OPS")

 if __name__=="__main__":
-  DEV = Device[Device.DEFAULT]
-  arch = DEV.renderer.arch
+  DEVICENUM = os.getenv("DEVICENUM", "0")
+  try:
+    DEV = Device['AMD:' + DEVICENUM]
+  except:
+    raise RuntimeError("Error while initiating AMD device")

-  if arch in {'gfx1100', 'gfx1103', 'gfx1151'}:
-    from tinygrad.runtime.autogen.amd.rdna3.ins import *
-    if arch == 'gfx1103': NUM_WORKGROUPS = 8
-    if arch == 'gfx1151': NUM_WORKGROUPS = 32
-    launchBenchmark(v_wmma_bf16_16x16x16_bf16, (7,8,15))
-    launchBenchmark(v_wmma_f16_16x16x16_f16, (7,8,15))
-    launchBenchmark(v_wmma_f32_16x16x16_bf16, (7,8,15))
-    launchBenchmark(v_wmma_f32_16x16x16_f16, (7,8,15))
-    launchBenchmark(v_wmma_i32_16x16x16_iu4, (7,8,9))
-    launchBenchmark(v_wmma_i32_16x16x16_iu8, (7,8,11))
-  elif arch in {'gfx1200', 'gfx1201'}:
-    from tinygrad.runtime.autogen.amd.rdna4.ins import *
-    # this instruction does not exist in the rdna4 isa, use the co version
-    s_sub_u32 = s_sub_co_u32
+  COMPILER = HIPCompiler(DEV.arch)
+  if DEV.arch in {'gfx1100', 'gfx1103', 'gfx1151'}:
+    if DEV.arch == 'gfx1103': NUM_WORKGROUPS = 8
+    if DEV.arch == 'gfx1151': NUM_WORKGROUPS = 32
+    launchBenchmark("v_wmma_bf16_16x16x16_bf16", (7,8,15))
+    launchBenchmark("v_wmma_f16_16x16x16_f16", (7,8,15))
+    launchBenchmark("v_wmma_f32_16x16x16_bf16", (7,8,15))
+    launchBenchmark("v_wmma_f32_16x16x16_f16", (7,8,15))
+    launchBenchmark("v_wmma_i32_16x16x16_iu4", (7,8,9))
+    launchBenchmark("v_wmma_i32_16x16x16_iu8", (7,8,11))
+  elif DEV.arch == 'gfx1201':
    NUM_WORKGROUPS = 64
-    launchBenchmark(v_wmma_bf16_16x16x16_bf16, (3,4,7))
-    launchBenchmark(v_wmma_f16_16x16x16_f16, (3,4,7))
-    launchBenchmark(v_wmma_f32_16x16x16_bf16, (7,8,11))
-    launchBenchmark(v_wmma_f32_16x16x16_f16, (7,8,11))
-    launchBenchmark(v_wmma_i32_16x16x16_iu4, (7,8,8))
-    launchBenchmark(v_wmma_i32_16x16x16_iu8, (7,8,9))
-    launchBenchmark(v_wmma_f32_16x16x16_fp8_fp8, (7,8,9))
-    launchBenchmark(v_wmma_f32_16x16x16_fp8_bf8, (7,8,9))
-    launchBenchmark(v_wmma_f32_16x16x16_bf8_fp8, (7,8,9))
-    launchBenchmark(v_wmma_f32_16x16x16_bf8_bf8, (7,8,9))
+    launchBenchmark("v_wmma_bf16_16x16x16_bf16", (3,4,7))
+    launchBenchmark("v_wmma_f16_16x16x16_f16", (3,4,7))
+    launchBenchmark("v_wmma_f32_16x16x16_bf16", (7,8,11))
+    launchBenchmark("v_wmma_f32_16x16x16_f16", (7,8,11))
+    launchBenchmark("v_wmma_i32_16x16x16_iu4", (7,8,8))
+    launchBenchmark("v_wmma_i32_16x16x16_iu8", (7,8,9))
+    launchBenchmark("v_wmma_f32_16x16x16_fp8_fp8", (7,8,9))
+    launchBenchmark("v_wmma_f32_16x16x16_fp8_bf8", (7,8,9))
+    launchBenchmark("v_wmma_f32_16x16x16_bf8_fp8", (7,8,9))
+    launchBenchmark("v_wmma_f32_16x16x16_bf8_bf8", (7,8,9))
    FLOPS_PER_MATMUL = 16*16*32*2
-    launchBenchmark(v_wmma_i32_16x16x32_iu4, (7,8,9))
-    launchBenchmark(v_swmmac_f32_16x16x32_f16, (7,8,11,12,19,20), False)
-    launchBenchmark(v_swmmac_f32_16x16x32_bf16, (7,8,11,12,19,20), False)
-    launchBenchmark(v_swmmac_f16_16x16x32_f16, (3,4,7,8,15,16), False)
-    launchBenchmark(v_swmmac_bf16_16x16x32_bf16, (3,4,7,8,15,16), False)
-    launchBenchmark(v_swmmac_i32_16x16x32_iu8, (7,8,9,10,13,14), False)
-    launchBenchmark(v_swmmac_i32_16x16x32_iu4, (7,8,8,9,10,11), False)
-    launchBenchmark(v_swmmac_f32_16x16x32_fp8_fp8, (7,8,9,10,13,14), False)
-    launchBenchmark(v_swmmac_f32_16x16x32_fp8_bf8, (7,8,9,10,13,14), False)
-    launchBenchmark(v_swmmac_f32_16x16x32_bf8_fp8, (7,8,9,10,13,14), False)
-    launchBenchmark(v_swmmac_f32_16x16x32_bf8_bf8, (7,8,9,10,13,14), False)
+    launchBenchmark("v_wmma_i32_16X16X32_iu4", (7,8,9))
+    launchBenchmark("v_swmmac_f32_16x16x32_f16", (7,8,11,12,19,20), False)
+    launchBenchmark("v_swmmac_f32_16x16x32_bf16", (7,8,11,12,19,20), False)
+    launchBenchmark("v_swmmac_f16_16x16x32_f16", (3,4,7,8,15,16), False)
+    launchBenchmark("v_swmmac_bf16_16x16x32_bf16", (3,4,7,8,15,16), False)
+    launchBenchmark("v_swmmac_i32_16x16x32_iu8", (7,8,9,10,13,14), False)
+    launchBenchmark("v_swmmac_i32_16x16x32_iu4", (7,8,8,9,10,11), False)
+    launchBenchmark("v_swmmac_f32_16x16x32_fp8_fp8", (7,8,9,10,13,14), False)
+    launchBenchmark("v_swmmac_f32_16x16x32_fp8_bf8", (7,8,9,10,13,14), False)
+    launchBenchmark("v_swmmac_f32_16x16x32_bf8_fp8", (7,8,9,10,13,14), False)
+    launchBenchmark("v_swmmac_f32_16x16x32_bf8_bf8", (7,8,9,10,13,14), False)
    FLOPS_PER_MATMUL = 16*16*64*2
-    launchBenchmark(v_swmmac_i32_16x16x64_iu4, (7,8,9,10,13,14), False)
-  elif arch == 'gfx950':
-    from tinygrad.runtime.autogen.amd.cdna.ins import *
+    launchBenchmark("v_swmmac_i32_16x16x64_iu4", (7,8,9,10,13,14), False)
+  elif DEV.arch == 'gfx950':
+    DIRECTIVE = ".amdhsa_accum_offset 4"
    NUM_WORKGROUPS = 256
    WAVE_SIZE = 64
    NUM_WAVES = 4
-    launchBenchmark(v_mfma_f32_16x16x16_f16, (3,0,1), accum=True)
-    launchBenchmark(v_mfma_f32_16x16x16_bf16, (3,0,1), accum=True)
+    launchBenchmark("v_mfma_f32_16x16x16_f16", (3,0,1), accum=True)
+    launchBenchmark("v_mfma_f32_16x16x16_bf16", (3,0,1), accum=True)
    FLOPS_PER_MATMUL = 16*16*32*2
-    launchBenchmark(v_mfma_f32_16x16x32_f16, (3,0,3), accum=True)
-    launchBenchmark(v_mfma_f32_16x16x32_bf16, (3,0,3), accum=True)
+    launchBenchmark("v_mfma_f32_16x16x32_f16", (3,0,3), accum=True)
+    launchBenchmark("v_mfma_f32_16x16x32_bf16", (3,0,3), accum=True)
    FLOPS_PER_MATMUL = 16*16*128*2
-    launchBenchmark(v_mfma_f32_16x16x128_f8f6f4, (3,0,7), accum=True) # fp8
-    launchBenchmark(v_mfma_f32_16x16x128_f8f6f4, (3,0,5), accum=True, cbsz=2, blgp=2) # fp6
-    launchBenchmark(v_mfma_f32_16x16x128_f8f6f4, (3,0,3), accum=True, cbsz=4, blgp=4) # fp4
+    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,7), accum=True) # fp8
+    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,5), accum=True, extra=", cbsz:2 blgp:2") # fp6
+    launchBenchmark("v_mfma_f32_16x16x128_f8f6f4", (3,0,3), accum=True, extra=", cbsz:4 blgp:4") # fp4
  else:
-    raise RuntimeError(f"arch {arch} not supported.")
+    raise RuntimeError(f"arch {DEV.arch} not supported.")
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	43c033e5e5	clean	2025-12-21 21:39:17 -04:00
geohot	76f2f14233	more	2025-12-21 15:48:59 -04:00
geohot	5515de6553	chain	2025-12-21 12:52:20 -04:00
geohot	439c4319ec	phase 1	2025-12-21 12:34:00 -04:00