feat: results

feat: training 6.0
2026-05-11 17:56:29 +00:00 · 2026-05-11 16:22:47 +00:00
757 changed files with 29334 additions and 90722 deletions
@@ -5,7 +5,6 @@ runs:
  steps:
    - name: Run process replay tests
      shell: bash
-      if: env.CAPTURE_PROCESS_REPLAY == '1'
      run: |
        export PR_TITLE=$(jq -r .pull_request.title "$GITHUB_EVENT_PATH")
        export CURRENT_SHA=${{ github.event.pull_request && github.event.pull_request.head.sha || github.sha }}
@@ -4,13 +4,13 @@ inputs:
  python-version:
    description: 'Python version to use'
    required: false
-    default: '3.14'
+    default: '3.12'
  key:
    description: 'Key for the python cache'
    required: false
    default: ''  # if you don't set a key, it doesn't cache
  deps:
-    description: 'Extra dependency groups (space separated)'
+    description: 'Extra dependency groups (comma separated)'
    required: false
    default: ''
  pydeps:
@@ -41,58 +41,45 @@ inputs:
    description: "Install LLVM?"
    required: false
    default: 'false'
-  qemu:
-    description: "Install qemu?"
+  mesa:
+    description: "Install mesa"
    required: false
    default: 'false'
-  ninja:
-    description: "Install ninja?"
+  tinydreno:
+    description: "Install tinydreno"
    required: false
    default: 'false'
 runs:
  using: "composite"
  steps:
-    - name: Setup environment
-      shell: bash
-      run: |
-        echo "UV_CACHE_DIR=/tmp/.uv-cache" >> "$GITHUB_ENV"
-        echo "OMP_NUM_THREADS=1" >> "$GITHUB_ENV"
-        # no buffers should be over 300MB in CI
-        echo "MAX_BUFFER_SIZE=300000000" >> "$GITHUB_ENV"
-        if [[ "$RUNNER_OS" == "Linux" ]]; then
-          echo "VIRTUAL_ENV=/opt/venv/${{ inputs.python-version }}" >> "$GITHUB_ENV"
-          echo "UV_PYTHON_INSTALL_DIR=/opt/python" >> "$GITHUB_ENV"
-        else
-          echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
-        fi
-
-    - name: Set up uv
-      uses: astral-sh/setup-uv@08807647e7069bb48b6ef5acd8ec9567f424441b
+    - name: Set up Python ${{ inputs.python-version }}
+      id: setup-python
+      uses: actions/setup-python@v6
      with:
-        enable-cache: 'false' # see below for manual caching
+        python-version: ${{ inputs.python-version }}

    # **** Caching packages ****

    - name: Cache Python packages (PR)
      if: github.event_name == 'pull_request'
      id: restore-venv-pr
-      uses: actions/cache/restore@v5
+      uses: actions/cache/restore@v4
      with:
-        path: /tmp/.uv-cache
-        key: uv-${{ runner.os }}-${{ runner.arch }}-python-${{ inputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        path: ${{ github.workspace }}/.venv
+        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
    - name: Cache Python packages
      if: github.event_name != 'pull_request'
      id: restore-venv
      uses: actions/cache@v5
      with:
-        path: /tmp/.uv-cache
-        key: uv-${{ runner.os }}-${{ runner.arch }}-python-${{ inputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        path: ${{ github.workspace }}/.venv
+        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ 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@v5
+      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 }}
@@ -106,43 +93,56 @@ runs:
    # **** Python deps ****

    - name: Install dependencies in venv (with extra)
-      if: inputs.deps != ''
-      shell: bash
-      run: |
-        uv venv --allow-existing --python ${{ inputs.python-version }} "$VIRTUAL_ENV"
-        DEPS="${{ inputs.deps }}"
-        uv pip install --python "$VIRTUAL_ENV" -e ".[${DEPS// /,}]" ${{ inputs.pydeps }} --torch-backend 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 == ''
-      shell: bash
-      run: |
-        uv venv --allow-existing --python ${{ inputs.python-version }} "$VIRTUAL_ENV"
-        uv pip install --python "$VIRTUAL_ENV" -e . ${{ inputs.pydeps }}
-    - name: Prune uv cache
-      if: github.event_name != 'pull_request'
-      shell: bash
-      run: uv cache prune --ci
-    - name: Configure venv
+      if: inputs.deps != '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
      shell: bash
      run: |
+        python -m venv .venv
        if [[ "$RUNNER_OS" == "Windows" ]]; then
-          echo "$VIRTUAL_ENV/Scripts" >> "$GITHUB_PATH"
+          source .venv/Scripts/activate
        else
-          echo "$VIRTUAL_ENV/bin" >> "$GITHUB_PATH"
+          . .venv/bin/activate
+        fi
+        python -m pip install -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --extra-index-url https://download.pytorch.org/whl/cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
+    - name: Install dependencies in venv (without extra)
+      if: inputs.deps == '' && steps.restore-venv-pr.outputs.cache-hit != 'true' && steps.restore-venv.outputs.cache-hit != 'true'
+      shell: bash
+      run: |
+        python -m venv .venv
+        if [[ "$RUNNER_OS" == "Windows" ]]; then
+          source .venv/Scripts/activate
+        else
+          . .venv/bin/activate
+        fi
+        python -m pip install -e . ${{ inputs.pydeps }}
+    - name: Set up venv environment
+      shell: bash
+      run: |
+        echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
+        echo "OMP_NUM_THREADS=1" >> "$GITHUB_ENV"
+        # no buffers should be over 300MB in CI
+        echo "MAX_BUFFER_SIZE=300000000" >> "$GITHUB_ENV"
+        if [[ "$RUNNER_OS" == "Windows" ]]; then
+          echo "${{ github.workspace }}/.venv/Scripts" >> "$GITHUB_PATH"
+        else
+          echo "${{ github.workspace }}/.venv/bin" >> "$GITHUB_PATH"
        fi

    # ******************* apt *******************
    - name: Setup apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
      shell: bash
      run: |
-        sudo mkdir -p /var/cache/apt/archives
        sudo chown -R $USER:$USER /var/cache/apt/archives

        echo 'Acquire::GzipIndexes "true";' | sudo tee /etc/apt/apt.conf.d/gzip
        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
        echo 'Binary::apt::APT::Keep-Downloaded-Packages "true";' | sudo tee -a /etc/apt/apt.conf.d/99keep-debs

+    - name: Add OpenCL Repo
+      if: inputs.opencl == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: echo "deb [ allow-insecure=yes ] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list
+
    - name: Add AMD Repo (Linux)
      if: inputs.amd == 'true' && runner.os == 'Linux'
      shell: bash
@@ -161,62 +161,54 @@ runs:
        echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list

    - name: Compute Package List + Hash
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
      id: apt-pkgs
      shell: bash
      run: |
        pkgs=""
        # **** OpenCL ****
        if [[ "${{ inputs.opencl }}" == "true" ]]; then
-          pkgs+=" ocl-icd-opencl-dev"
+          pkgs+=" opencl-headers \
+            intel-oneapi-runtime-openmp=2023.2.1-16 intel-oneapi-runtime-compilers-common=2023.2.1-16 intel-oneapi-runtime-compilers=2023.2.1-16 \
+            intel-oneapi-runtime-dpcpp-sycl-opencl-cpu=2023.2.1-16 intel-oneapi-runtime-tbb-common=2021.10.0-49541 \
+            intel-oneapi-runtime-tbb=2021.10.0-49541 intel-oneapi-runtime-opencl=2023.2.1-16"
        fi
        # **** AMD ****
        if [[ "${{ inputs.amd }}" == "true" ]]; then
-          pkgs+=" comgr"
+          pkgs+=" hsa-rocr comgr hsa-rocr-dev liburing-dev libibverbs-dev libc6-dev"
+        fi
+        # **** CUDA ****
+        if [[ "${{ inputs.cuda }}" == "true" ]]; then
+          pkgs+=" git g++ cmake ninja-build llvm-15-dev zlib1g-dev libglew-dev \
+            flex bison libfl-dev libboost-thread-dev libboost-filesystem-dev nvidia-cuda-toolkit-gcc libzstd-dev"
        fi
        # **** WebGPU (dependencies for software-based vulkan) ****
        if [[ "${{ inputs.webgpu }}" == "true" ]]; then
-          pkgs+=" mesa-vulkan-drivers"
+          pkgs+=" libgl1 libglx-mesa0 libgl1-mesa-dri libxcb-xfixes0-dev mesa-vulkan-drivers"
        fi
        # **** LLVM ****
        if [[ "${{ inputs.llvm }}" == "true" ]]; then
          pkgs+=" libllvm20 clang-20 lld-20"
        fi
-        # **** QEMU ****
-        if [[ "${{ inputs.qemu }}" == "true" ]]; then
-          pkgs+=" qemu-user-static"
-        fi
-        # **** ninja ****
-        if [[ "${{ inputs.ninja }}" == "true" ]]; then
-          pkgs+=" ninja-build"
-        fi

        echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
        echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"

-        installed=true
-        for pkg in $pkgs; do
-          info=$(dpkg-query -W -f='${db:Status-Abbrev} ${Version}' "$pkg" 2> /dev/null || true)
-          echo "${pkg}: ${info:-not in dpkg database}"
-          [[ "$info" == ii* ]] || installed=false
-        done
-        echo "installed=$installed" >> "$GITHUB_OUTPUT"
-
    - name: Cache apt (PR)
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && github.event_name == 'pull_request' && steps.apt-pkgs.outputs.installed == 'false'
-      uses: actions/cache/restore@v5
+      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.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && github.event_name != 'pull_request' && steps.apt-pkgs.outputs.installed == 'false'
+      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@v5
      with:
        path: /var/cache/apt/archives/
        key: ${{ runner.os }}-${{ runner.arch }}-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.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && steps.apt-pkgs.outputs.installed == 'false'
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
      shell: bash
      run: |
        sudo apt -qq update || true
@@ -226,14 +218,8 @@ runs:
          sudo apt-get -y --allow-unauthenticated --no-install-recommends install ${{ steps.apt-pkgs.outputs.pkgs }}
        fi

-        sudo mkdir -p /var/cache/apt/archives
        sudo chown -R $USER:$USER /var/cache/apt/archives/

-    - name: Add clang to PATH (Linux)
-      if: inputs.llvm == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: echo "/usr/lib/llvm-20/bin" >> "$GITHUB_PATH"
-
    # **** AMD ****
    - name: Setup AMD (Linux)
      if: inputs.amd == 'true' && runner.os == 'Linux'
@@ -248,44 +234,83 @@ runs:
      if: inputs.amd == 'true' && runner.os == 'macOS'
      shell: bash
      run: |
-        sudo "$VIRTUAL_ENV/bin/python" -c "
-          from tinygrad.helpers import fetch
-          fetch('https://github.com/tinygrad/amdcomgr_dylib/releases/download/v7.2.0/libamd_comgr.dylib', name='/usr/local/lib/libamd_comgr.dylib',
-                sha256='7712fbe4fcb9fcdea49aeac989876448df975ce0a8ce7c9b15b55c15e7a05935').chmod(0o644)"
-
-    # **** CUDA ****
-    - name: Install CUDA
-      if: inputs.cuda == 'true'
-      shell: bash
-      run: |
-        sudo mkdir -p /usr/local/cuda/targets/x86_64-linux
-        curl -fL https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/linux-x86_64/cuda_nvrtc-linux-x86_64-11.5.119-archive.tar.xz \
-          | sudo tar -xJ -C /usr/local/cuda/targets/x86_64-linux --strip-components=1
-        echo /usr/local/cuda/targets/x86_64-linux/lib | sudo tee /etc/ld.so.conf.d/cuda-nvrtc.conf
-        sudo ldconfig
+        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 | \
+          jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
+          sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib

    # **** gpuocelot ****
+
+    - name: Install gpuocelot dependencies (MacOS)
+      if: inputs.ocelot == 'true' && runner.os == 'macOS'
+      shell: bash
+      run: |
+        pkgs=(cmake ninja llvm@15 zlib glew flex bison [email protected] zstd ncurses)
+        for f in "${pkgs[@]}"; do
+          brew ls --versions "$f" >/dev/null 2>&1 || brew install --quiet "$f"
+        done
+
+        # Fix boost 1.85 for gpuocelot
+        ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
+        ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
+        ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
+    - name: Cache gpuocelot (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'
+      id: cache-build
+      uses: actions/cache@v5
+      env:
+        cache-name: cache-gpuocelot-build-1
+      with:
+        path: ${{ github.workspace }}/gpuocelot/ocelot
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
+    - name: Clone/compile gpuocelot
+      if: inputs.ocelot == 'true' && steps.cache-build-pr.outputs.cache-hit != 'true' && steps.cache-build.outputs.cache-hit != 'true'
+      shell: bash
+      run: |
+        git clone --recurse-submodules https://github.com/gpuocelot/gpuocelot.git ${{ github.workspace }}/gpuocelot
+        cd ${{ github.workspace }}/gpuocelot/ocelot
+        git checkout b16039dc940dc6bc4ea0a98380495769ff35ed99
+        mkdir build
+        cd build
+
+        CMAKE_ARGS="-Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5"
+        if [[ "${{ runner.os }}" == "macOS" ]]; then
+          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
+
+        cmake .. $CMAKE_ARGS
+        ninja
    - name: Install gpuocelot
      if: inputs.ocelot == 'true'
      shell: bash
      run: |
-        sudo "$VIRTUAL_ENV/bin/python" -c "
-          from tinygrad.helpers import fetch
-          fetch('https://github.com/tinygrad/gpuocelot/releases/download/v0.1.0/libgpuocelot.${{ runner.os == 'Linux' && 'so' || 'dylib' }}',
-                name='/usr/local/lib/libgpuocelot.${{ runner.os == 'Linux' && 'so' || 'dylib' }}',
-                sha256='${{ runner.os == 'Linux' && 'a24705276a9a187111371465987b3258f8836ef512a34266e3075bc4714e125a' || '5106c998c795a36dec79eb7b2aae324a93d1338236d36eeaae232649ec457663' }}').chmod(0o644)"
+        cd ${{ github.workspace }}/gpuocelot/ocelot/build
+        sudo cp libgpuocelot.${{ runner.os == 'macOS' && 'dylib' || 'so' }} /usr/${{ runner.os == 'macOS' && 'local/' || '' }}lib/

    # **** WebGPU ****

-    - name: Install WebGPU dawn
-      if: inputs.webgpu == 'true'
+    - name: Install WebGPU dawn (Linux)
+      if: inputs.webgpu == 'true' && runner.os == 'Linux'
      shell: bash
      run: |
-        sudo "$VIRTUAL_ENV/bin/python" -c "
-          from tinygrad.helpers import fetch
-          fetch('https://github.com/wpmed92/pydawn/releases/download/v0.1.6/libwebgpu_dawn.${{ runner.os == 'Linux' && 'so' || 'dylib' }}',
-                name='/usr/local/lib/libwebgpu_dawn.${{ runner.os == 'Linux' && 'so' || 'dylib' }}',
-                sha256='${{ runner.os == 'Linux' && 'cf36091d266a32c9d5080f14662de44cece241987939713282ea0ff558db81c6' || '7e87c7acefda8b6af1a1c5debfedcf62958311284b8fd8d9bcf93e312e6636e3' }}').chmod(0o644)"
+        sudo curl -fL https://github.com/wpmed92/pydawn/releases/download/v0.1.6/libwebgpu_dawn.so -o /usr/local/lib/libwebgpu_dawn.so
+        sudo ldconfig
+    - name: Install WebGPU dawn (macOS)
+      if: inputs.webgpu == 'true' && runner.os == 'macOS'
+      shell: bash
+      run: |
+        brew tap wpmed92/dawn
+        brew install dawn

    # **** LLVM ****

@@ -294,15 +319,18 @@ runs:
      shell: bash
      run: brew install llvm@20

-    # *** OpenCL ***
-    - name: Install rusticl
-      if: inputs.opencl == 'true'
+    # **** mesa ****
+    - name: Install mesa (linux)
+      if: inputs.mesa == 'true' && runner.os == 'Linux'
      shell: bash
-      run: |
-        sudo "$VIRTUAL_ENV/bin/python" -c "
-          from tinygrad.helpers import fetch
-          fetch('https://github.com/sirhcm/tinymesa/releases/download/rusticl-v1/libRusticlOpenCL.so.1.0.0', name='/usr/lib/libRusticlOpenCL.so',
-                sha256='d4f48566d8fd33f6cdd8ef6de35a71966e8a8517e6f68ff3c52dbb43765a2513').chmod(0o644)"
-        sudo mkdir -p /etc/OpenCL/vendors
-        echo "/usr/lib/libRusticlOpenCL.so" | sudo tee /etc/OpenCL/vendors/rusticl.icd
-        echo "RUSTICL_ENABLE=llvmpipe" >> "$GITHUB_ENV"
+      run: sudo curl -fL https://github.com/sirhcm/tinymesa/releases/download/v1/libtinymesa_cpu-mesa-25.2.7-linux-amd64.so -o /usr/lib/libtinymesa_cpu.so
+    - name: Install mesa (macOS)
+      if: inputs.mesa == 'true' && runner.os == 'macOS'
+      shell: bash
+      run: brew install sirhcm/tinymesa/tinymesa_cpu
+
+    # *** tinydreno ***
+    - name: Install tinydreno (linux)
+      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
@@ -35,18 +35,17 @@ jobs:
        key: 'autogen'
        amd: 'true'
        llvm: 'true'
-        deps: 'autogen'
+        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 liburing-dev
+      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev
    - name: Regenerate autogen files
      run: |
-        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
+        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "comgr.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
        python3 -c "from tinygrad.runtime.autogen import opencl"
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv_610, nv"
-        python3 -c "from tinygrad.runtime.autogen import comgr, comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
+        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 *"
-        python3 -c "from tinygrad.runtime.autogen.nv_regs import *"
-        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, pci, vfio"
+        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"
@@ -54,7 +53,7 @@ jobs:
        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"
-        python3 -c "from tinygrad.runtime.autogen import mlx5, bnxt"
+        python3 -c "from tinygrad.runtime.autogen import mlx5"
        python3 -c "from tinygrad.runtime.autogen import ggml_common"
        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
    - name: Check for differences
@@ -102,3 +101,42 @@ jobs:
      with:
        name: autogen-macos-patch
        path: autogen-macos.patch
+
+  autogen-comgr-2:
+    name: In-tree Autogen (comgr 2)
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - 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
+        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
+      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@v7
+      with:
+        name: autogen-comgr2-patch
+        path: autogen-comgr2.patch
@@ -8,7 +8,7 @@ permissions:
  contents: write
 jobs:
  deploy:
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
      - name: Configure Git Credentials
@@ -1,8 +1,8 @@
 name: Run MLPerf Training

 on:
-  #schedule:
-  #  - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
+  schedule:
+    - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
  push:
    branches:
      - update_mlperf
@@ -1,181 +0,0 @@
-name: Platform Tests
-env:
-  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '19'
-  CAPTURE_PROCESS_REPLAY: ${{ github.event_name == 'pull_request' && contains(github.event.pull_request.title, '[pr]') && '1' || '0' }}
-  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-  PYTHONPATH: ${{ github.workspace }}
-  CHECK_OOB: 1
-
-on:
-  push:
-    branches:
-      - master
-  pull_request:
-  workflow_dispatch:
-
-concurrency:
-  group: platform-${{ github.event_name }}-${{ github.event_name == 'pull_request' && github.event.pull_request.number || github.run_id }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
-
-jobs:
-
-# ****** OSX Tests ******
-
-  unittestmacos:
-    name: MacOS (unit)
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-macos
-        deps: testing_unit
-    - name: Run unit tests
-      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
-    - name: Test tensor core ops (fake)
-      run: DEV=METAL DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
-    - name: Test tensor core ops (real)
-      run: DEV=METAL DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
-    - name: Test Beam Search
-      run: DEV=METAL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
-    - name: Test Device Specific
-      run: DEV=METAL python3 -m pytest test/device/test_metal.py
-    #- name: Fuzz Test linearizer
-    #  run: DEV=METAL DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  unittestmacosmock:
-    name: MacOS (unit, mock)
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-macos-mock
-        deps: testing_unit
-        amd: 'true'
-        ocelot: 'true'
-    - name: Run NULL backend tests
-      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
-    - name: Run pytest (amd)
-      env:
-        DEV: MOCKKFD+AMD
-        FORWARD_ONLY: 1
-      run: |
-        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
-    - name: Run pytest (ptx)
-      env:
-        DEV: "MOCK+NV:PTX"
-        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 \
-          test/testextra/test_hevc.py::TestHevc::test_hevc_decode_compile --durations=20
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmetal:
-    strategy:
-      fail-fast: false
-      matrix:
-        group: [1, 2]
-    name: MacOS (DEV=METAL) (${{ matrix.group }})
-    runs-on: macos-26
-    timeout-minutes: 20
-    env:
-      DEV: METAL
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: macos-metal
-        deps: testing_unit
-    - name: Check Device.DEFAULT and print some source
-      run: |
-        python -c "from tinygrad import Device; assert Device.DEFAULT == 'METAL'"
-        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run backend tests
-      run: python -m pytest -n=auto test/backend --durations=20 --splits 2 --group ${{ matrix.group }}
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmacos:
-    strategy:
-      fail-fast: false
-      matrix:
-        dev:
-          - 'CPU:CLANG'
-          - 'CPU:LLVM'
-          - 'CPU:LVP'
-          - 'WEBGPU'
-
-    name: MacOS (DEV=${{ matrix.dev }})
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: macos-${{ matrix.dev }}
-        deps: "testing_unit${{ contains(matrix.dev, 'LVP') && ' mesa' || '' }}"
-        llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') }}
-        webgpu: ${{ matrix.dev == 'WEBGPU' }}
-    - name: Set env
-      run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
-    - name: Check Device.DEFAULT and print some source
-      run: |
-        python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
-        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run test_tiny
-      run: python -m pytest -n=auto test/test_tiny.py --durations=20
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-# ****** Windows Tests ******
-
-  testwindows:
-    strategy:
-      fail-fast: false
-      matrix:
-        dev:
-          - 'CPU:CLANG'
-          - 'CPU:LLVM'
-          - 'CPU:X86'
-          - 'WEBGPU'
-
-    name: Windows (DEV=${{ matrix.dev }})
-    runs-on: windows-2025
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: windows-${{ matrix.dev }}-minimal
-          deps: testing_minimal
-          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
-      - name: Set env
-        shell: bash
-        run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
-      - name: Check Device.DEFAULT and print some source
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
-          DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-      - name: Run test_tiny
-        shell: bash
-        run: python -m pytest -n=auto test/test_tiny.py --durations=20
@@ -10,7 +10,7 @@ on:

 jobs:
  deploy:
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    steps:
    - uses: actions/checkout@v6
    - name: Set up Python
@@ -10,19 +10,16 @@ concurrency:
 jobs:
  checkbranch:
    name: Check PR Branch status
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    outputs:
      branchstat: ${{ steps.brstat.outputs.stat}}
    steps:
-      - name: Check code from PR branch
+      - name: Check code from PR branch 
        uses: actions/checkout@v6
        with:
          repository: ${{ github.event.pull_request.head.repo.full_name }}
          ref: ${{ github.event.pull_request.head.sha }}
          fetch-depth: 0
-          # PR code is only inspected with git rev-list, never executed
-          allow-unsafe-pr-checkout: true
-          persist-credentials: false
      - name: Check whether branch is up-to-date
        id: brstat
        run: |
@@ -44,7 +41,7 @@ jobs:
    permissions:
      contents: read
      pull-requests: write
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    needs: checkbranch
    if: needs.checkbranch.outputs.branchstat == 'false'
    steps:
@@ -54,9 +51,6 @@ jobs:
          repository: ${{ github.event.pull_request.head.repo.full_name }}
          ref: ${{ github.event.pull_request.head.sha }}
          path: pr
-          # PR code is only line-counted by master's sz.py, never executed
-          allow-unsafe-pr-checkout: true
-          persist-credentials: false
        # the base default to tinygrad master and cannot be other fork branch for security purpose
      - name: Checkout code from tinygrad master
        uses: actions/checkout@v6
@@ -87,7 +81,7 @@ jobs:
    name: Core Library Line Difference
    permissions:
      pull-requests: write
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    needs: checkbranch
    if: needs.checkbranch.outputs.branchstat == 'true'
    steps:
@@ -69,4 +69,3 @@ mutants
 dagre/
 graphlib/
 uv.lock
-pi_session_window0.jsonl
@@ -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/backend/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/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
        language: system
        always_run: true
        pass_filenames: false
@@ -1,7 +0,0 @@
-# Notes
-
- Run tests with `-n12` for speed (e.g. `python -m pytest test/null/test_dtype.py -x -q -n12`)
- Run `python -m mypy tinygrad/` to typecheck
- Run `python -m ruff check .` to lint
- Read `./tinygrad/viz/README.md` for profiling and debugging rewrite rules
- Do not do amend commits. Always do a new commit if a force push to origin would be required.
@@ -72,7 +72,7 @@ As it turns out, 90% of what you need for neural networks are a decent autograd/
 Throw in an optimizer, a data loader, and some compute, and you have all you need.

 ```python
-from tinygrad import Tensor, nn, Context
+from tinygrad import Tensor, nn

 class LinearNet:
  def __init__(self):
@@ -86,7 +86,7 @@ optim = nn.optim.Adam([model.l1, model.l2], lr=0.001)

 x, y = Tensor.rand(4, 1, 28, 28), Tensor([2,4,3,7])  # replace with real mnist dataloader

-with Context(TRAINING=1):
+with Tensor.train():
  for i in range(10):
    optim.zero_grad()
    loss = model(x).sparse_categorical_crossentropy(y).backward()
@@ -140,8 +140,8 @@ Documentation along with a quick start guide can be found on the [docs website](
 ```python
 from tinygrad import Tensor

-x = Tensor.eye(3).clone()  # clone to make it a buffer
-y = Tensor([[2.0,0,-2.0]])
+x = Tensor.eye(3, requires_grad=True)
+y = Tensor([[2.0,0,-2.0]], requires_grad=True)
 z = y.matmul(x).sum()
 z.backward()

@@ -164,9 +164,7 @@ print(y.grad.tolist())  # dz/dy

 ## Contributing

-There has been a lot of interest in tinygrad lately. Following these guidelines will help your PR get accepted. If you do submit a PR, please include a sentence or two about why you want this merged and why you think it will improve the project.
-
-If you are a new contributor with something that looks even close to AI written, it will be closed without feedback and you may be banned from our GitHub. No human should waste time reading AI slop. And for everyone, if you used AI, disclose what you used it for.
+There has been a lot of interest in tinygrad lately. Following these guidelines will help your PR get accepted.

 We'll start with what will get your PR closed with a pointer to this section:

@@ -198,8 +196,6 @@ python3 test/backend/test_ops.py        # just the ops tests
 python3 -m pytest test/                 # whole test suite
 ```

-For agents, always run tests with `-n12` for speed.
-
 #### Process replay tests

 [Process replay](https://github.com/tinygrad/tinygrad/blob/master/test/external/process_replay/README.md) compares your PR's generated kernels against master. If your PR is a refactor or speedup without any expected behavior change, It should include [pr] in the pull request title.
@@ -1,10 +0,0 @@
-import os, pytest, signal, threading
-
-@pytest.hookimpl(wrapper=True)
-def pytest_runtest_call(item):
-  t = threading.Timer(int(os.getenv("TEST_TIMEOUT", 90)), os.kill, args=(os.getpid(), signal.SIGABRT))
-  t.start()
-  try: yield
-  finally:
-    t.cancel()
-    t.join()
@@ -11,7 +11,7 @@ X_train -= X_train.mean()
 # *****
 # 1. Define an MNIST model.

-from tinygrad import Tensor, Context
+from tinygrad import Tensor

 l1 = Tensor.kaiming_uniform(128, 784)
 l2 = Tensor.kaiming_uniform(10, 128)
@@ -24,11 +24,11 @@ l1n, l2n = l1.numpy(), l2.numpy()
 from tinygrad.nn.optim import SGD
 optim = SGD([l1, l2])

-with Context(TRAINING=1):
-  X, Y = X_train[(samples:=Tensor.randint(128, high=X_train.shape[0]))], Y_train[samples]
-  optim.zero_grad()
-  model(X).sparse_categorical_crossentropy(Y).backward()
-  optim.schedule_step()   # this will step the optimizer without running realize
+Tensor.training = True
+X, Y = X_train[(samples:=Tensor.randint(128, high=X_train.shape[0]))], Y_train[samples]
+optim.zero_grad()
+model(X).sparse_categorical_crossentropy(Y).backward()
+optim.schedule_step()   # this will step the optimizer without running realize

 # *****
 # 3. Create a schedule (linear uop).
@@ -67,7 +67,8 @@ def example_2_hip(a:Tensor, correct):
    # the sink specifies the GLOBAL and LOCAL sizes, along with the input buffers and name
    sink = UOp.sink(UOp.special(GLOBALS, 'gidx0'), UOp.special(THREADS, 'lidx0'), out, buf,
                    arg=KernelInfo(name="hip_reduce_sum_kernel"))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
+                UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
  eval_harness("HIP kernel", a, lambda x: Tensor.empty(GLOBALS).custom_kernel(x, fxn=hip_reduce_sum)[0].sum(), check=correct)

 def example_3_custom_uop(a:Tensor, correct):
@@ -88,7 +89,7 @@ def example_3_custom_uop(a:Tensor, correct):

    # store all the per lane accumulators to LOCAL
    local_accs = UOp.placeholder((LCLS,), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL)
-    local_accs = local_accs.after(local_accs[lane].store(acc[0]))
+    local_accs = local_accs.after(local_accs[lane].store(acc[0]).barrier())

    # accumulate LOCALs into a single per CU accumulator
    late_reduce_loop = UOp.range(LCLS, 3, AxisType.REDUCE)
@@ -122,7 +123,8 @@ def example_5_custom_assembly(a:Tensor, correct):
        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 UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=(x, dtypes.void)) for x in self.instructions]))))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
+                                   UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in self.instructions]))))

  CU_COUNT = 32
  LANES = 64
@@ -62,7 +62,7 @@ A lot of work can still be done here. For example, we never copy the inputs to o

 Many accelerators have Tensor Cores / MAC arrays / systolic arrays. The main value of these is that, since they are 2-D, they create an n^2 ratio between the compute and the input data.

-GPUs use Tensor Cores instead of MAC arrays to fit better in the GPU warp paradigm. This is because the output of Tensor Cores is O(n) wrt the input, while the output of MAC arrays is O(n^2)
+GPUs use Tensor Cores instead of MAC arrays to fit better in the GPU warp paradigm. This is because the output of Tensor Cores is O(n) wrt the input, while the output of MAC arrays like the AMX is O(n^2)

 We have a simple framework in tinygrad for adding these ALU blocks and achieving good performance from them.

@@ -1,8 +1,7 @@
 ::: tinygrad.dtype.DType

-::: tinygrad.dtype.DTypes
+::: tinygrad.dtype.dtypes
    options:
-        heading: dtypes
        members: true
        members_order: source
        show_labels: false
@@ -24,7 +24,7 @@ You will see `CUDA` here on a GPU instance, or `CPU` here on a CPU instance.
 We'll use the model from [the Keras tutorial](https://keras.io/examples/vision/mnist_convnet/).

 ```python
-from tinygrad import Tensor, nn, Context
+from tinygrad import Tensor, nn

 class Model:
  def __init__(self):
@@ -74,8 +74,8 @@ We'll use the Adam optimizer. The `nn.state.get_parameters` will walk the model
 ```python
 optim = nn.optim.Adam(nn.state.get_parameters(model))
 batch_size = 128
-@Context(TRAINING=1)
 def step():
+  Tensor.training = True  # makes dropout work
  samples = Tensor.randint(batch_size, high=X_train.shape[0])
  X, Y = X_train[samples], Y_train[samples]
  optim.zero_grad()
@@ -143,6 +143,7 @@ Since we are just randomly sampling from the dataset, there's no real concept of
 for step in range(7000):
  loss = jit_step()
  if step%100 == 0:
+    Tensor.training = False
    acc = (model(X_test).argmax(axis=1) == Y_test).mean().item()
    print(f"step {step:4d}, loss {loss.item():.2f}, acc {acc*100.:.2f}%")
 ```
@@ -133,7 +133,7 @@ For our loss function we will be using sparse categorical cross entropy loss. Th
 ```python
 def sparse_categorical_crossentropy(self, Y, ignore_index=-1) -> Tensor:
    loss_mask = Y != ignore_index
-    y_counter = Tensor.arange(self.shape[-1], dtype=dtypes.int32).unsqueeze(0).expand(Y.numel(), self.shape[-1])
+    y_counter = Tensor.arange(self.shape[-1], dtype=dtypes.int32, requires_grad=False, device=self.device).unsqueeze(0).expand(Y.numel(), self.shape[-1])
    y = ((y_counter == Y.flatten().reshape(-1, 1)).where(-1.0, 0) * loss_mask.reshape(-1, 1)).reshape(*Y.shape, self.shape[-1])
    return self.log_softmax().mul(y).sum() / loss_mask.sum()
 ```
@@ -165,18 +165,17 @@ from extra.datasets import fetch_mnist
 Now we have everything we need to start training our neural network.
 We will be training for 1000 steps with a batch size of 64.

-We use `with Context(TRAINING=1)` to enable training mode.
+We use `with Tensor.train()` to set the internal flag `Tensor.training` to `True` during training.
 Upon exit, the flag is restored to its previous value by the context manager.

 ```python
-from tinygrad import Context
 X_train, Y_train, X_test, Y_test = fetch_mnist()

-with Context(TRAINING=1):
+with Tensor.train():
  for step in range(1000):
    # random sample a batch
    samp = np.random.randint(0, X_train.shape[0], size=(64))
-    batch = Tensor(X_train[samp])
+    batch = Tensor(X_train[samp], requires_grad=False)
    # get the corresponding labels
    labels = Tensor(Y_train[samp])

@@ -214,7 +213,7 @@ with Timing("Time: "):
  for step in range(1000):
    # random sample a batch
    samp = np.random.randint(0, X_test.shape[0], size=(64))
-    batch = Tensor(X_test[samp])
+    batch = Tensor(X_test[samp], requires_grad=False)
    # get the corresponding labels
    labels = Y_test[samp]

@@ -258,7 +257,7 @@ with Timing("Time: "):
  for step in range(1000):
    # random sample a batch
    samp = np.random.randint(0, X_test.shape[0], size=(64))
-    batch = Tensor(X_test[samp])
+    batch = Tensor(X_test[samp], requires_grad=False)
    # get the corresponding labels
    labels = Y_test[samp]

@@ -83,5 +83,9 @@ NV backend supports several interfaces for communicating with devices:
 ## CPU Arch
 The CPU renderers may be additionally configured using the arch component of [the `DEV` environment variable](env_vars.md#dev-variable).
 CPU arch should be specified as a comma-separated list of parameters, and must contain at least two values: the architecture family (ie. x86_64, arm64, or riscv64) and the cpu type (as accepted by `clang`'s `-march`).
-If native is specified as the cpu type, tinygrad (or delegate compiler) will query the host cpu type. Additional comma-separated values are interpreted as cpu feature flags. When a value is preceded by a `-` character, the corresponding feature flag will be disabled, otherwise the flag will be enabled.
+If native is specified as the cpu type, tinygrad (or delegate compiler) will query the host cpu type. Additional comma-separated values may be specified as follows:
+
+* `AMX`: emit Apple silicon AMX instructions
+
+All other additional values are interpreted as cpu feature flags. When a value is preceded by a `-` character, the corresponding feature flag will be disabled, otherwise the flag will be enabled.
 Note that enabled feature flags should not be preceded by a `+`.
@@ -0,0 +1,196 @@
+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.STACK:
+          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 to_program
+  with Context(PCONTIG=2, DEVECTORIZE=2, SPEC=0):
+    out = tree_traversal(forest_t, val_t, height, rounds)
+    sink = out.schedule_linear().src[-1].src[0]
+    prg = to_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[3].arg)
+  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!")
@@ -4,10 +4,10 @@ from tinygrad.dtype import DTypeLike, dtypes
 import math

 # rewritten from numpy
-def rfftfreq(n: int, d: float = 1.0) -> Tensor:
+def rfftfreq(n: int, d: float = 1.0, device=None) -> Tensor:
  val = 1.0 / (n * d)
  N = n // 2 + 1
-  results = Tensor.arange(N)
+  results = Tensor.arange(N, device=device)
  return results * val

 # just like in librosa
@@ -1,6 +1,6 @@
 from typing import Tuple
 import time
-from tinygrad import Tensor, TinyJit, nn, Context
+from tinygrad import Tensor, TinyJit, nn
 import gymnasium as gym
 from tinygrad.helpers import trange
 import numpy as np  # TODO: remove numpy import
@@ -55,7 +55,7 @@ if __name__ == "__main__":

  @TinyJit
  def train_step(x:Tensor, selected_action:Tensor, reward:Tensor, old_log_dist:Tensor) -> Tuple[Tensor, Tensor, Tensor]:
-    with Context(TRAINING=1):
+    with Tensor.train():
      log_dist, value = model(x)
      action_mask = (selected_action.reshape(-1, 1) == Tensor.arange(log_dist.shape[1]).reshape(1, -1).expand(selected_action.shape[0], -1)).float()

@@ -9,7 +9,8 @@ from extra.lr_scheduler import OneCycleLR
 GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]

 # override tinygrad defaults
-Context(DEFAULT_FLOAT=dtypes.half, FUSE_OPTIM=1).__enter__()
+dtypes.default_float = dtypes.half
+Context(FUSE_OPTIM=1).__enter__()

 # from https://github.com/tysam-code/hlb-CIFAR10/blob/main/main.py
 batchsize = getenv("BS", 1024)
@@ -66,8 +67,8 @@ class ConvGroup:
    self.conv2 = nn.Conv2d(channels_out, channels_out, kernel_size=3, padding=1, bias=False)
    self.norm1 = nn.BatchNorm(channels_out, track_running_stats=False, eps=1e-12, momentum=hyp['net']['batch_norm_momentum'])
    self.norm2 = nn.BatchNorm(channels_out, track_running_stats=False, eps=1e-12, momentum=hyp['net']['batch_norm_momentum'])
-    cast(Tensor, self.norm1.weight).is_param_(False)
-    cast(Tensor, self.norm2.weight).is_param_(False)
+    cast(Tensor, self.norm1.weight).requires_grad = False
+    cast(Tensor, self.norm2.weight).requires_grad = False
  def __call__(self, x:Tensor) -> Tensor:
    x =    self.norm1(self.conv1(x).max_pool2d().float()).cast(dtypes.default_float).quick_gelu()
    return self.norm2(self.conv2(x).float()).cast(dtypes.default_float).quick_gelu() + x
@@ -121,7 +122,7 @@ if __name__ == "__main__":
    return ret.mul(hyp['opt']['loss_scale_scaler']*loss_batchsize_scaler).sum().div(hyp['opt']['loss_scale_scaler'])

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def train_step(idxs:Tensor) -> Tensor:
    X, Y = X_train[idxs], Y_train[idxs]
    if len(GPUS) > 1:
@@ -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, Context
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

@@ -19,7 +19,7 @@ class Model:
  def __call__(self, x:Tensor) -> Tensor: return x.sequential(self.layers)

  @TinyJit
-  @Context(TRAINING=1)
+  @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])
@@ -1,6 +1,6 @@
 # model based off https://towardsdatascience.com/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import List, Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, Device, Context
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, Device
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

@@ -31,7 +31,7 @@ if __name__ == "__main__":

  @TinyJit
  def train_step() -> Tensor:
-    with Context(TRAINING=1):
+    with Tensor.train():
      opt.zero_grad()
      samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
      Xt, Yt = X_train[samples].shard_(GPUS, axis=0), Y_train[samples].shard_(GPUS, axis=0)  # we shard the data on axis 0
@@ -22,6 +22,10 @@ class Attention:
    self.head_dim = dim // n_heads

  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]) -> Tensor:
+    if mask is not None or start_pos.val == 0:
+      # no symbolic shape qkv when consuming prompts
+      start_pos = start_pos.val
+
    if HALF: x = x.half()
    xqkv = self.c_attn(x).reshape(None, None, 3, self.n_heads, self.head_dim)
    xq, xk, xv = [xqkv[:, :, i, :, :] for i in range(3)]
@@ -34,8 +38,12 @@ class Attention:
    # update the cache
    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()

-    keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
-    values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
+    if start_pos > 0:
+      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
+      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
+    else:
+      keys = xk
+      values = xv

    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
    return self.c_proj(xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2).reshape(bsz, seqlen, self.dim))
@@ -78,14 +86,15 @@ class Transformer:
      seqlen = tokens.shape[1]
      tok_emb = self.wte(tokens)

-    # start_pos is a bound Variable, so everything below it stays symbolic
-    pos_emb = self.wpe(self.allpos.shrink((None, (start_pos, start_pos+seqlen))))
+    # not symbolic when consuming the prompt
+    selected_pos = (0, seqlen) if start_pos.val == 0 else (start_pos, start_pos+1)
+    pos_emb = self.wpe(self.allpos.shrink((None, selected_pos)))

    h = tok_emb + pos_emb

    if HALF: h = h.half()

-    mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype).triu(start_pos+1) if seqlen > 1 else None
+    mask = Tensor.full((1, 1, seqlen, start_pos.val+seqlen), float("-inf"), dtype=h.dtype).triu(start_pos.val+1) if seqlen > 1 else None

    for hi in self.h: h = hi(h, start_pos, mask)

@@ -1,6 +1,6 @@
 import itertools
 from typing import Callable
-from tinygrad import nn, Tensor, dtypes, Device, TinyJit, Context
+from tinygrad import nn, Tensor, dtypes, Device, TinyJit
 from tinygrad.helpers import getenv, trange, partition

 class Model:
@@ -35,21 +35,22 @@ if __name__ == "__main__":

  params = nn.state.get_parameters(model)

-  # init params
+  # init params, set requires grad on the ones we need gradients of
  for x in params:
+    if x.requires_grad is None: x.requires_grad_()
    x.replace(x.contiguous())
  Tensor.realize(*params)

  # split params (with grads) and buffers (without)
-  params, buffers = partition(params, lambda x: x.is_param)
+  params, buffers = partition(params, lambda x: x.requires_grad)
  print(f"params: {len(params)} buffers: {len(buffers)}")

  # optim params
  pos_params = list(itertools.accumulate(params, lambda x,y: x+y.numel(), initial=0))
  adam_m = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
  adam_v = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
-  adam_b1_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU").contiguous()
-  adam_b2_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU").contiguous()
+  adam_b1_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False).contiguous()
+  adam_b2_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False).contiguous()
  adam_params = [adam_m, adam_v, adam_b1_t, adam_b2_t]

  # create loss and grads. init all state so the JIT works on microbatch
@@ -59,7 +60,7 @@ if __name__ == "__main__":
  Tensor.realize(*params, *buffers, *adam_params, loss, grads)

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def microbatch():
    samples = Tensor.randint(BS // ACC_STEPS, high=X_train.shape[0])
    for t in params: t.grad = None
@@ -10,7 +10,7 @@ from extra.lr_scheduler import OneCycleLR
 from tinygrad import nn, dtypes, Tensor, Device, GlobalCounters, TinyJit, Variable
 from tinygrad.nn.state import get_state_dict
 from tinygrad.nn import optim
-from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod, TRAINING
+from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod
 from extra.bench_log import BenchEvent, WallTimeEvent

 cifar_mean = [0.4913997551666284, 0.48215855929893703, 0.4465309133731618]
@@ -30,9 +30,9 @@ class UnsyncedBatchNorm:
    if affine: self.weight, self.bias = Tensor.ones(sz, dtype=dtypes.float32), Tensor.zeros(sz, dtype=dtypes.float32)
    else: self.weight, self.bias = None, None

-    self.running_mean = Tensor.zeros(num_devices, sz, dtype=dtypes.float32).is_param_(False)
-    self.running_var = Tensor.ones(num_devices, sz, dtype=dtypes.float32).is_param_(False)
-    self.num_batches_tracked = Tensor.zeros(1, dtype=dtypes.int).is_param_(False)
+    self.running_mean = Tensor.zeros(num_devices, sz, dtype=dtypes.float32, requires_grad=False)
+    self.running_var = Tensor.ones(num_devices, sz, dtype=dtypes.float32, requires_grad=False)
+    self.num_batches_tracked = Tensor.zeros(1, dtype=dtypes.int, requires_grad=False)

  def __call__(self, x:Tensor):
    xr = x.reshape(self.num_devices, -1, *x.shape[1:]).cast(dtypes.float32)
@@ -44,7 +44,7 @@ class UnsyncedBatchNorm:
    return ret.reshape(x.shape).cast(x.dtype)

  def calc_stats(self, x:Tensor):
-    if TRAINING:
+    if Tensor.training:
      # This requires two full memory accesses to x
      # https://github.com/pytorch/pytorch/blob/c618dc13d2aa23625cb0d7ada694137532a4fa33/aten/src/ATen/native/cuda/Normalization.cuh
      # There's "online" algorithms that fix this, like https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Welford's_Online_algorithm
@@ -68,7 +68,8 @@ class UnsyncedBatchNorm:
 class BatchNorm(nn.BatchNorm2d if getenv("SYNCBN") else UnsyncedBatchNorm):
  def __init__(self, num_features):
    super().__init__(num_features, track_running_stats=False, eps=1e-12, momentum=0.85, affine=True)
-    self.weight.is_param_(False)
+    self.weight.requires_grad = False
+    self.bias.requires_grad = True

 class ConvGroup:
  def __init__(self, channels_in, channels_out):
@@ -152,21 +153,26 @@ def train_cifar():

  # ========== Model ==========
  def whitening(X, kernel_size=hyp['net']['kernel_size']):
-    def _patches(data:Tensor, patch_size=(kernel_size,kernel_size)):
+    def _cov(X):
+      return (X.T @ X) / (X.shape[0] - 1)
+
+    def _patches(data, patch_size=(kernel_size,kernel_size)):
      h, w = patch_size
-      _, c, _, _ = data.shape
-      return data._pool((h, w)).permute(1, 4, 5, 0, 3, 2).reshape(c*h*w, -1)
+      c = data.shape[1]
+      axis = (2, 3)
+      return np.lib.stride_tricks.sliding_window_view(data, window_shape=(h,w), axis=axis).transpose((0,3,2,1,4,5)).reshape((-1,c,h,w))

    def _eigens(patches):
-      cov = ((patches @ patches.T) / (patches.shape[1] - 1)).numpy()
-      eigvals, eigvecs = np.linalg.eigh(cov, UPLO='U')
-      return np.flip(eigvals, 0), np.flip(eigvecs.T.reshape(patches.shape[0], X.shape[1], kernel_size, kernel_size), 0)
+      n,c,h,w = patches.shape
+      Σ = _cov(patches.reshape(n, c*h*w))
+      Λ, V = np.linalg.eigh(Σ, UPLO='U')
+      return np.flip(Λ, 0), np.flip(V.T.reshape(c*h*w, c, h, w), 0)

    # NOTE: np.linalg.eigh only supports float32 so the whitening layer weights need to be converted to float16 manually
-    eigvals, eigvecs = _eigens(_patches(X.float()))
-    W = eigvecs/np.sqrt(eigvals+1e-2)[:,None,None,None]
+    Λ, V = _eigens(_patches(X.float().numpy()))
+    W = V/np.sqrt(Λ+1e-2)[:,None,None,None]

-    return Tensor(W.astype(np.float32)).cast(dtypes.default_float).is_param_(False)
+    return Tensor(W.astype(np.float32), requires_grad=False).cast(dtypes.default_float)

  # ========== Loss ==========
  def cross_entropy(x:Tensor, y:Tensor, reduction:str='mean', label_smoothing:float=0.0) -> Tensor:
@@ -218,7 +224,7 @@ def train_cifar():

  @TinyJit
  def augmentations(X:Tensor, Y:Tensor):
-    perms = Tensor.randperm(X.shape[0], device=X.device) # We reuse perms for cutmix, because they are expensive to generate
+    perms = Tensor.randperm(X.shape[0], device=X.device) # We reuse perms for cutmix, because they are expensivne to generate
    if getenv("RANDOM_CROP", 1):
      X = random_crop(X, crop_size=32)
    if getenv("RANDOM_FLIP", 1):
@@ -258,6 +264,7 @@ def train_cifar():
      # self.model_ema = copy.deepcopy(net) # won't work for opencl due to unpickeable pyopencl._cl.Buffer
      self.net_ema = SpeedyResNet(w)
      for net_ema_param, net_param in zip(get_state_dict(self.net_ema).values(), get_state_dict(net).values()):
+        net_ema_param.requires_grad = False
        net_ema_param.assign(net_param.numpy())

    @TinyJit
@@ -300,7 +307,7 @@ def train_cifar():
  params_bias = []
  params_non_bias = []
  for params in params_dict:
-    if params_dict[params].is_param:
+    if params_dict[params].requires_grad is not False:
      if 'bias' in params:
        params_bias.append(params_dict[params])
      else:
@@ -309,9 +316,6 @@ def train_cifar():
  opt_bias     = optim.SGD(params_bias,     lr=0.01, momentum=hyp['opt']['momentum'], nesterov=True, weight_decay=hyp['opt']['bias_decay'])
  opt_non_bias = optim.SGD(params_non_bias, lr=0.01, momentum=hyp['opt']['momentum'], nesterov=True, weight_decay=hyp['opt']['non_bias_decay'])

-  # realize model params and optimizer state before JIT to avoid cache misses
-  Tensor.realize(*params_dict.values(), *opt_bias.b, *opt_non_bias.b)
-
  # NOTE taken from the hlb_CIFAR repository, might need to be tuned
  initial_div_factor = hyp['opt']['initial_div_factor']
  final_lr_ratio = hyp['opt']['final_lr_ratio']
@@ -328,7 +332,9 @@ def train_cifar():
      # index 0 for bias and 1 for non-bias
      optimizer.zero_grad()
      loss.backward()
-      return loss.realize(*optimizer.schedule_step(), *lr_scheduler[0].schedule_step(), *lr_scheduler[1].schedule_step())
+      optimizer.step()
+      lr_scheduler[0].step()
+      lr_scheduler[1].step()
    return loss.realize()

  train_step_jitted = TinyJit(train_step)
@@ -355,11 +361,11 @@ def train_cifar():
  i = 0
  eval_acc_pct = 0.0
  batcher = fetch_batches(X_train, Y_train, BS=BS, is_train=True)
-  with Context(TRAINING=1):
+  with Tensor.train():
    st = time.monotonic()
    while i <= STEPS:
      if i % getenv("EVAL_STEPS", STEPS) == 0 and i > 1 and not getenv("DISABLE_BACKWARD"):
-        # Using Context(TRAINING=0) here actually bricks batchnorm, even with track_running_stats=True
+        # Use Tensor.training = False here actually bricks batchnorm, even with track_running_stats=True
        corrects = []
        corrects_ema = []
        losses = []
@@ -102,7 +102,7 @@ class Int8Embedding:
    self.weight, self.scale = Tensor.ones(vocab_size, embed_size, dtype=dtypes.int8), Tensor.ones(vocab_size, dtype=dtypes.half)

  def __call__(self, idx:Tensor) -> Tensor:
-    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz).unsqueeze(-1)
+    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz, requires_grad=False, device=self.weight.device).unsqueeze(-1)
    big_shp = idx.shape+(self.vocab_sz, self.embed_sz)
    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1)).expand(big_shp), (self.weight.cast(self.scale.dtype).T*self.scale).T
    return (arange == idx).mul(vals).sum(-2, dtype=vals.dtype)
@@ -3,7 +3,7 @@ import os
 if "NOOPT" not in os.environ: os.environ["NOOPT"] = "1"
 from tinygrad import Device, nn, Tensor, dtypes
 from train_gpt2 import GPT, GPTConfig
-from tinygrad.helpers import DEV, dedup, flatten, getenv, GlobalCounters, to_function_name, Context
+from tinygrad.helpers import DEV, dedup, flatten, getenv, GlobalCounters, to_function_name
 from tinygrad.engine.realize import get_kernel
 from tinygrad.schedule.memory import memory_planner
 from tinygrad.uop.ops import Ops
@@ -23,23 +23,23 @@ if __name__ == "__main__":
  #B, T = Variable("B", 1, 128).bind(4), 64 #Variable("T", 1, 1024).bind(64)
  B, T = 4, 64

+  Tensor.training = True
  optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=1e-4)
  warmup_count = getenv("WARMUP", 3)
-  with Context(TRAINING=1):
-    for i in range(warmup_count):  # TODO: why does it take three and not two to stabilize
-      GlobalCounters.reset()
-      X = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
-      Y = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
-      _, loss = model(X, Y)
-      optimizer.zero_grad()
-      if getenv("BACKWARD", 1):
-        loss.backward()
-        tensors = optimizer.schedule_step()
-      else:
-        tensors = []
-      sched = loss.schedule(*tensors)
-      print(f"calls {i}:", len(sched))
-      #run_schedule(sched[:])
+  for i in range(warmup_count):  # TODO: why does it take three and not two to stabilize
+    GlobalCounters.reset()
+    X = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
+    Y = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
+    _, loss = model(X, Y)
+    optimizer.zero_grad()
+    if getenv("BACKWARD", 1):
+      loss.backward()
+      tensors = optimizer.schedule_step()
+    else:
+      tensors = []
+    sched = loss.schedule(*tensors)
+    print(f"calls {i}:", len(sched))
+    #run_schedule(sched[:])
  sched = memory_planner(sched)
  ast_dedup = dedup([si.ast for si in sched if si.ast.op is Ops.SINK])
  srcs = {}
@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 import os, math, time
 import numpy as np
-from tinygrad import Tensor, nn, fetch, Device, TinyJit, GlobalCounters, Context
+from tinygrad import Tensor, nn, fetch, Device, TinyJit, GlobalCounters
 from dataclasses import dataclass

@dataclass
@@ -25,7 +25,7 @@ class CausalSelfAttention:
    self.n_embd = config.n_embd
    # not really a 'bias', more of a mask, but following the OpenAI/HF naming though
    self.bias = Tensor.ones(1, 1, config.block_size, config.block_size).tril()
-    self.bias.is_param_(False)
+    self.bias.requires_grad = False

  def __call__(self, x:Tensor):
    B, T, C = x.shape
@@ -99,7 +99,7 @@ class GPT:

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

    tok_emb = self.wte(idx) # token embeddings of shape (b, t, n_embd)
    pos_emb = self.wpe(pos) # position embeddings of shape (t, n_embd)
@@ -177,7 +177,7 @@ if __name__ == "__main__":
  if args.gpus > 1: x, y = x.shard(GPUS, axis=0), y.shard(GPUS, axis=0)

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def step(x:Tensor, y:Tensor) -> Tensor:
    _, loss = model(x, y)
    optimizer.zero_grad()
@@ -204,3 +204,4 @@ if __name__ == "__main__":
    top_k = 40
    y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
    print(decode(y[0].tolist()))
+
@@ -1,5 +1,5 @@
 # much taken from https://github.com/cloneofsimo/minRF
-from tinygrad import Tensor, nn, GlobalCounters, TinyJit, Context
+from tinygrad import Tensor, nn, GlobalCounters, TinyJit
 from tinygrad.helpers import getenv, trange
 from extra.models.llama import Attention, FeedForward, precompute_freqs_cis

@@ -135,7 +135,7 @@ if __name__ == "__main__":
  optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=5e-4)

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def train_step():
    if getenv("OVERFIT"): samples = Tensor.zeros(getenv("BS", 256), dtype='int')
    else: samples = Tensor.randint(getenv("BS", 256), high=X_train.shape[0])
@@ -1,6 +1,6 @@
 import functools, argparse, pathlib
 from tinygrad import Tensor, nn, Device, GlobalCounters, Variable
-from tinygrad.helpers import Timing, Profiling, tqdm
+from tinygrad.helpers import Timing, Profiling, CI, tqdm
 from tinygrad.nn.state import torch_load, get_state_dict
 from extra.models.llama import FeedForward, Transformer
 from extra.bench_log import BenchEvent, WallTimeEvent
@@ -36,7 +36,7 @@ if __name__ == "__main__":
    model = Transformer(n_layers=32, dim=4096, hidden_dim=14336, n_heads=32, n_kv_heads=8, norm_eps=1e-5, vocab_size=32000, feed_forward=functools.partial(MixtureFeedForward, 8), jit=False)
    model_state_dict = get_state_dict(model)

-    for k in (t := tqdm(state, disable=None)):
+    for k in (t := tqdm(state, disable=CI)):
      if 'feed_forward.experts.' in k:
        expert_no = int(k.split('feed_forward.experts.')[1].split('.')[0])
        device = Device.DEFAULT + ":" + str((expert_no//2)+1)
@@ -44,7 +44,7 @@ if __name__ == "__main__":
        device = Device.DEFAULT
      t.set_description(f"ram used: {GlobalCounters.mem_used/1e9:5.2f} GB, loading {k} to {device}")
      model_state_dict[k].replace(state[k].to(device).half()).realize()
-  if t.disable: print(f"ram used: {GlobalCounters.mem_used/1e9:5.2f} GB")
+  if CI: print(f"ram used: {GlobalCounters.mem_used/1e9:5.2f} GB")

  from sentencepiece import SentencePieceProcessor
  spp = SentencePieceProcessor(model_file=args.weights + "/tokenizer.model")
@@ -1,11 +1,11 @@
 import os, random, pickle, queue, struct, math, functools, hashlib, time
 from typing import List
 from pathlib import Path
-from multiprocessing import Queue, Process, shared_memory, connection, Lock
+from multiprocessing import Queue, Process, shared_memory, connection, Lock, cpu_count

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

 ### ResNet
@@ -131,7 +131,7 @@ def batch_load_resnet(batch_size=64, val=False, shuffle=True, seed=None, pad_fir
    else: X = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/{shm_name}")
    Y = [None] * (batch_size*BATCH_COUNT)

-    for _ in range(NUM_CPU_THREADS.value):
+    for _ in range(cpu_count()):
      p = Process(target=loader_process, args=(q_in, q_out, X, seed))
      p.daemon = True
      p.start()
@@ -212,7 +212,7 @@ def batch_load_train_bert(BS:int, seed:int|None=None):
    rng.shuffle(fs)
    train_files.append(fs.pop(0))

-  cycle_length = min(NUM_CPU_THREADS.value, len(train_files))
+  cycle_length = min(getenv("NUM_CPU_THREADS", min(os.cpu_count(), 8)), len(train_files))
  assert cycle_length > 0, "cycle_length must be greater than 0"

  dataset = InterleavedDataset(train_files, cycle_length)
@@ -301,7 +301,7 @@ def batch_load_unet3d(preprocessed_dataset_dir:Path, batch_size:int=6, val:bool=
    X = Tensor.empty(*sz, dtype=dtypes.float32, device=f"disk:/dev/shm/{shm_name_x}")
    Y = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/{shm_name_y}")

-    for _ in range(NUM_CPU_THREADS.value):
+    for _ in range(cpu_count()):
      proc = Process(target=load_unet3d_data, args=(preprocessed_dataset_dir, seed, queue_in, queue_out, X, Y))
      proc.daemon = True
      proc.start()
@@ -437,7 +437,7 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
  dataset_iter = iter(image_ids)

  try:
-    for _ in range(NUM_CPU_THREADS.value):
+    for _ in range(cpu_count()):
      proc = Process(
        target=load_retinanet_data,
        args=(base_dir, val, queue_in, queue_out, imgs, boxes, labels),
@@ -2,7 +2,7 @@ import math
 from typing import Union

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

@@ -57,7 +57,7 @@ class EmbeddingBert(nn.Embedding):
  def __call__(self, idx:Tensor) -> Tensor:
    if idx.numel() == 0: return Tensor.empty(idx.shape+(self.embed_sz,), dtype=self.weight.dtype, device=self.weight.device)
    arange_shp, weight_shp, big_shp = (1, 1, self.vocab_sz, 1), (1, 1, self.vocab_sz, self.embed_sz), idx.shape+(self.vocab_sz, self.embed_sz,)
-    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz).reshape(arange_shp)
+    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz, requires_grad=False, device=self.weight.device).reshape(arange_shp)
    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1,)).expand(big_shp), self.weight.cast(dtypes.default_float).reshape(weight_shp).expand(big_shp)
    return (arange == idx).where(vals, 0).sum(2, dtype=vals.dtype)

@@ -77,15 +77,15 @@ class FrozenBatchNorm2dRetinaNet(nn.BatchNorm2d):
  def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1):
    self.eps, self.track_running_stats, self.momentum = eps, track_running_stats, momentum

-    self.weight = Tensor.ones(sz, dtype=dtypes.float32).is_param_(False) if affine else None
-    self.bias = Tensor.zeros(sz, dtype=dtypes.float32).is_param_(False) if affine else None
+    self.weight = Tensor.ones(sz, dtype=dtypes.float32, requires_grad=False) if affine else None
+    self.bias = Tensor.zeros(sz, dtype=dtypes.float32, requires_grad=False) if affine else None

-    if track_running_stats: self.running_mean, self.running_var = Tensor.zeros(sz, dtype=dtypes.float32).is_param_(False), Tensor.ones(sz, dtype=dtypes.float32).is_param_(False)
-    self.num_batches_tracked = Tensor.zeros(1, dtype=dtypes.long).is_param_(False)
+    if track_running_stats: self.running_mean, self.running_var = Tensor.zeros(sz, dtype=dtypes.float32, requires_grad=False), Tensor.ones(sz, dtype=dtypes.float32, requires_grad=False)
+    self.num_batches_tracked = Tensor.zeros(1, dtype=dtypes.long, requires_grad=False)

  def __call__(self, x:Tensor) -> Tensor:
    batch_mean, batch_var = super().calc_stats(x.cast(dtypes.float32))
-    if self.track_running_stats and TRAINING:
+    if self.track_running_stats and Tensor.training:
      self.running_mean.assign((1-self.momentum) * self.running_mean + self.momentum * batch_mean.detach().cast(self.running_mean.dtype))
      self.running_var.assign((1-self.momentum) * self.running_var + self.momentum * x.numel()/(x.numel()-x.shape[1]) * batch_var.detach().cast(self.running_var.dtype))
      self.num_batches_tracked += 1
@@ -358,7 +358,7 @@ def eval_stable_diffusion():
      batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
    return batch, unpadded_bs 

-  @Context(TRAINING=0)
+  @Tensor.train(mode=False)
  def eval_unet(eval_inputs:list[dict], unet:UNetModel, cond_stage:FrozenOpenClipEmbedder, first_stage:AutoencoderKL,
                inception:FidInceptionV3, clip:OpenClipEncoder) -> tuple[float, float]:
    # Eval is divided into 5 jits, one per model
@@ -498,10 +498,11 @@ def eval_stable_diffusion():

 if __name__ == "__main__":
  # inference only
+  Tensor.training = False
+
  models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert").split(",")
-  with Context(TRAINING=0):
-    for m in models:
-      nm = f"eval_{m}"
-      if nm in globals():
-        print(f"eval {m}")
-        globals()[nm]()
+  for m in models:
+    nm = f"eval_{m}"
+    if nm in globals():
+      print(f"eval {m}")
+      globals()[nm]()
@@ -1,6 +1,6 @@
 # load each model here, quick benchmark
 from tinygrad import Tensor, GlobalCounters
-from tinygrad.helpers import getenv, Context
+from tinygrad.helpers import getenv
 import numpy as np

 def test_model(model, *inputs):
@@ -59,10 +59,11 @@ def spec_mrcnn():

 if __name__ == "__main__":
  # inference only for now
-  with Context(TRAINING=0):
-    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
-      nm = f"spec_{m}"
-      if nm in globals():
-        print(f"testing {m}")
-        globals()[nm]()
+  Tensor.training = False
+
+  for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
+    nm = f"spec_{m}"
+    if nm in globals():
+      print(f"testing {m}")
+      globals()[nm]()

@@ -2,7 +2,7 @@ import os, time, math, functools, random, contextlib
 from pathlib import Path
 import multiprocessing

-from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes, Context
+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.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
@@ -157,7 +157,6 @@ def train_resnet():
  # input_std = Tensor([0.229, 0.224, 0.225], device=GPUS, dtype=dtypes.float32).reshape(1, -1, 1, 1)
  def normalize(x): return (x.permute([0, 3, 1, 2]) - input_mean).cast(dtypes.default_float)
  @TinyJit
-  @Context(TRAINING=1)
  def train_step(X, Y):
    optimizer_group.zero_grad()
    X = normalize(X)
@@ -171,7 +170,6 @@ def train_resnet():
    return loss.realize(), top_1.realize()

  @TinyJit
-  @Context(TRAINING=0)
  def eval_step(X, Y):
    X = normalize(X)
    out = model.forward(X)
@@ -182,11 +180,11 @@ def train_resnet():
  def fake_data_get(batch_size):
    x = Tensor.zeros(batch_size, 224, 224, 3, dtype=dtypes.uchar).contiguous()
    y = [0] * batch_size
-    return x.shard(GPUS, axis=0).realize(), Tensor(y).shard(GPUS, axis=0), y, None
+    return x.shard(GPUS, axis=0).realize(), Tensor(y, requires_grad=False).shard(GPUS, axis=0), y, None

  def data_get(it):
    x, y, cookie = next(it)
-    return x.shard(GPUS, axis=0).realize(), Tensor(y).shard(GPUS, axis=0), y, cookie
+    return x.shard(GPUS, axis=0).realize(), Tensor(y, requires_grad=False).shard(GPUS, axis=0), y, cookie

  # ** epoch loop **
  step_times = []
@@ -194,6 +192,7 @@ def train_resnet():
    # ** train loop **
    if MLLOGGER and RUNMLPERF:
      MLLOGGER.start(key=mllog_constants.EPOCH_START, value=e+1, metadata=dict(epoch_num=e+1))
+    Tensor.training = True
    BEAM.value = TRAIN_BEAM

    if INITMLPERF:
@@ -272,6 +271,7 @@ def train_resnet():
      eval_loss = 0.0
      eval_top_1 = 0
      eval_num_samples = 0
+      Tensor.training = False
      BEAM.value = EVAL_BEAM

      if INITMLPERF:
@@ -413,7 +413,7 @@ def train_retinanet():
    layers_to_train = ["layer4", "layer3", "layer2", "layer1", "conv1"][:trainable_layers]
    for k, v in get_state_dict(backbone).items():
      if all([not k.startswith(layer) for layer in layers_to_train]):
-        v.is_param_(False)
+        v.requires_grad = False

  def _data_get(it:Iterator[tuple[Tensor, ...]], val:bool=False):
    if val:
@@ -614,7 +614,7 @@ def train_retinanet():

      if getenv("RESET_STEP", 1): _train_step.reset()

-      with Context(TRAINING=0):
+      with Tensor.train(mode=False):
        if not RUNMLPERF:
          i, proc = 0, _fake_data_get(EVAL_BS, val=(val:=True))
        else:
@@ -784,7 +784,7 @@ def train_unet3d():
    return x.shard(GPUS, axis=0).realize(), y.shard(GPUS, axis=0), cookie

  @TinyJit
-  @Context(TRAINING=1)
+  @Tensor.train()
  def train_step(model, x, y):
    optim.zero_grad()

@@ -795,10 +795,10 @@ def train_unet3d():
    optim.step()
    return loss.realize()

-  @Context(TRAINING=0)
+  @Tensor.train(mode=False)
  def eval_step(model, x, y):
    y_hat, y = sliding_window_inference(model, x, y, gpus=GPUS)
-    y_hat, y = Tensor(y_hat), Tensor(y)
+    y_hat, y = Tensor(y_hat), Tensor(y, requires_grad=False)
    loss = dice_ce_loss(y_hat, y)
    score = dice_score(y_hat, y)
    return loss.realize(), score.realize()
@@ -919,7 +919,6 @@ def train_rnnt():
  pass

@TinyJit
-@Context(TRAINING=0)
 def eval_step_bert(model, input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor, masked_positions:Tensor, masked_lm_ids:Tensor,
                   masked_lm_weights:Tensor, next_sentence_labels:Tensor, GPUS):
  for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
@@ -1107,7 +1106,6 @@ def train_bert():
      MLLOGGER.start(key=mllog_constants.EPOCH_START, value=i*GBS, metadata={"epoch_num": i*GBS})

  @TinyJit
-  @Context(TRAINING=1)
  def train_step_bert(input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor,
                      masked_positions:Tensor, masked_lm_ids:Tensor, masked_lm_weights:Tensor, next_sentence_labels:Tensor):
    for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
@@ -1135,6 +1133,7 @@ def train_bert():

  while train_data is not None and i < train_steps and not achieved:
    if getenv("TRAIN", 1):
+      Tensor.training = True
      BEAM.value = TRAIN_BEAM
      st = time.perf_counter()
      GlobalCounters.reset()
@@ -1187,6 +1186,7 @@ def train_bert():
      eval_lm_accs = []
      eval_clsf_accs = []
      eval_times = []
+      Tensor.training = False
      BEAM.value = EVAL_BEAM

      for j in tqdm(range(max_eval_steps), desc="Evaluating", total=max_eval_steps, disable=BENCHMARK):
@@ -1282,10 +1282,10 @@ def train_bert():
        previous_step = i

 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8, MXFP4
+  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE
  from examples.llama3 import MODEL_PARAMS
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
-  from examples.mlperf.optim import GradAccClipAdamW, clip_grads
+  from examples.mlperf.optim import GradAccClipAdamW

  INITMLPERF = getenv("INITMLPERF")
  RUNMLPERF = getenv("RUNMLPERF")
@@ -1419,7 +1419,10 @@ def train_llama3():

  for p in optim.params:
    grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
-    p.grad = p.zeros_like(dtype=grad_dtype).contiguous()
+    if isinstance(p.device, tuple) and p.uop.axis is not None:
+      p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device[0]).shard_(p.device, axis=p.uop.axis).contiguous()
+    else:
+      p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device).contiguous()
  grads = [p.grad for p in optim.params]

  scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)
@@ -1434,40 +1437,28 @@ def train_llama3():
    load_state_dict(scheduler, safe_load(fn), realize=False)

  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
-  fp8_next_amax = [t for ts in model._fp8_next_amax.values() for t in ts]
-  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts]
-  fp8_next_grad_amax = [t for ts in model._fp8_next_grad_amax.values() for t in ts]
-  fp8_inv_scales = list(model._fp8_inv_scale.values()) + list(model._fp8_next_inv_scale.values())
+  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts] if hasattr(model, "_fp8_grad_amax") else []
+  fp8_inv_scales = list(model._fp8_inv_scale.values())

  from tinygrad.nn.state import get_state_dict
  model_state = get_state_dict(model)
-  for wname in model._fp8_inv_scale:
+  for wname in ["wqkv", "wo", "w13", "w2"]:
    w = model_state[wname]
    w._inv_scale = model._fp8_inv_scale[wname]
-    w._next_inv_scale = model._fp8_next_inv_scale[wname]
    if optim.master_params:
      idx = next(j for j, p in enumerate(optim.params) if p is w)
-      master = optim.master_params[idx]
-      inv = w._inv_scale if w._inv_scale.device == master.device else w._inv_scale.to(master.device)
-      if MXFP8:
-        from extra.gemm.cdna_asm_gemm import _mx_block_scale
-        bs = _mx_block_scale(inv.reshape(-1, inv.shape[-1])).reshape(w.shape)
-        master.assign((master * bs).contiguous())
-      else:
-        master.assign((master * inv.reshape(*inv.shape, *([1]*(w.ndim-inv.ndim)))).contiguous())
+      optim.master_params[idx].assign((optim.master_params[idx] * w._inv_scale.reshape(-1, *([1]*(w.ndim-1)))).contiguous())

  # realize everything here
  if optim.master_params: Tensor.realize(*optim.master_params)
-  loss_acc = Tensor.zeros(1, dtype=dtypes.float32, device=device)
-  Tensor.realize(loss_acc, *optim.params, *fp8_inv_scales, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
+  Tensor.realize(*optim.params, *fp8_inv_scales, *fp8_amax, *fp8_grad_amax)

  @TinyJit
  def minibatch(tokens:Tensor):
-    model.reset_amax()
    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], save=bool(SMALL))
+    logits:Tensor = model(tokens[:, :-1])
    if getenv("FAST_CE", 0):
      from extra.llama_kernels.fused_ce import fused_ce_loss
      loss = fused_ce_loss(logits.cast(dtypes.bfloat16), tokens[:, 1:], label_smoothing=0.0)
@@ -1477,27 +1468,24 @@ def train_llama3():
    for g, new_g in zip(grads, loss.gradient(*optim.params)):
      apply_grad(g, new_g.uop)

-    loss_acc.assign(loss_acc + loss.flatten().float())
-    return loss_acc.realize(*grads, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
+    loss_cpu = loss.flatten().float().to("CPU")
+    return loss_cpu.realize(*grads, *fp8_amax, *fp8_grad_amax)

  @TinyJit
  def optim_step():
-    grad_norm = clip_grads(grads, grad_acc, 1.0)
-    optim.fstep(grads, grad_norm)
+    grad_norm = optim.fstep(grads)
    scheduler.step()

-    for g in grads: g.assign(0)
-    model.update_amax()
+    for g in grads: g.assign(g.zeros_like())

    lr_cpu = optim.lr.float().to("CPU")
    grad_norm_cpu = grad_norm.float().to("CPU")
-    loss_cpu = loss_acc.to("CPU")
-    Tensor.realize(lr_cpu, grad_norm_cpu, loss_cpu, loss_acc.assign(0), *grads, *fp8_inv_scales, *fp8_amax, *fp8_grad_amax)
+    Tensor.realize(lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales)

-    return lr_cpu, grad_norm_cpu, loss_cpu
+    return lr_cpu, grad_norm_cpu

  @TinyJit
-  @Context(TRAINING=0)
+  @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)
@@ -1510,7 +1498,7 @@ def train_llama3():
  def fake_data(bs, samples):
    import numpy as np
    for _ in range(samples // bs):
-      fake_data_np = np.random.randint(0, real_vocab_size, size=(bs, SEQLEN + 1), dtype=np.int32)
+      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")

  def get_train_iter():
@@ -1549,8 +1537,8 @@ def train_llama3():
      st = time.perf_counter()

      stopped = False
-      data_time, dev_time = 0, 0
-      for _ in range(accum_steps:=grad_acc if i >= 2 else 1):
+      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:
@@ -1558,15 +1546,16 @@ def train_llama3():
          break
        mst = time.perf_counter()
        data_time += mst - ist
-        minibatch(tokens)
+        losses.append(minibatch(tokens).item())
        dev_time += time.perf_counter() - mst
      if stopped: break

      gt = time.perf_counter()
      ret = optim_step()
-      lr, grad_norm, loss = ret[0].item(), ret[1].item(), ret[2].item() / accum_steps
+      lr, grad_norm = ret[0].item(), ret[1].item()
      et = time.perf_counter()

+      loss = sum(losses) / len(losses)
      optim_time = et - gt
      dev_time += optim_time
      step_time = et - st
@@ -1578,7 +1567,7 @@ def train_llama3():

      mem_gb = GlobalCounters.mem_used / 1e9
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * (9.2e15 if MXFP4 else 4.6e15))) * 100
+      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 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")
@@ -1664,275 +1653,6 @@ def train_llama3():
      if MLLOGGER and RUNMLPERF:
        MLLOGGER.start(key=mllog_constants.BLOCK_START, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})

-def train_gptoss():
-  from examples.mlperf.models.gpt_oss import GPTOSS, GPT_OSS_20B, apply_grad, FP8_DTYPE
-  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
-  from examples.mlperf.optim import GradAccClipAdamW, GradAccClipAdamWGroup, fclip_grads
-
-  BENCHMARK = getenv("BENCHMARK")
-
-  config = {}
-  BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4-8b/"))
-  BS                 = config["BS"]                     = getenv("BS", 16)
-  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS
-  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)
-  MAX_STEPS          = config["MAX_STEPS"]              = getenv("MAX_STEPS", 1_200_000)
-  SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else MAX_STEPS * GBS)
-  EVAL_SAMPLES       = config["EVAL_SAMPLES"]           = getenv("EVAL_SAMPLES", 1024)
-  WARMUP_STEPS       = config["WARMUP_STEPS"]           = getenv("WARMUP_STEPS", 128)
-  LR                 = config["LR"]                     = getenv("LR", 4e-4 * GBS / 16)
-  END_LR             = config["END_LR"]                 = getenv("END_LR", 4e-5)
-  EVAL_FREQ          = config["EVAL_FREQ"]              = getenv("EVAL_FREQ", 12288)
-  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
-  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 3.34)
-
-  opt_adamw_beta_1 = 0.9
-  opt_adamw_beta_2 = 0.95
-  opt_adamw_epsilon = 1e-5
-  opt_adamw_weight_decay = 0.1
-
-  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-gpt-oss")
-
-  model_params = GPT_OSS_20B
-  model_params['vocab_size'] = getenv("VOCAB_SIZE", 128256)
-  real_vocab_size = model_params['vocab_size']
-  if (layers:=getenv("LAYERS")) != 0: model_params['n_layers'] = layers
-  if (experts:=getenv("EXPERTS")) != 0: model_params['n_experts'] = experts
-  print(f"model parameters: {model_params}")
-
-  model = GPTOSS(**model_params, max_context=SEQLEN)
-
-  params = get_parameters(model)
-
-  if getenv("EMPTYWEIGHT"):
-    for v in get_parameters(model):
-      v = v.assign(Tensor.empty(v.shape, dtype=v.dtype))
-
-  is_dp = (DP := getenv("DP", 1)) > 1
-  is_sharding = is_dp
-  device_count = DP
-  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(device_count))
-
-  model.shard(device, False)
-
-  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
-  params_wd = [p for p in params if p.ndim >= 3]
-  params_no_wd = [p for p in params if p.ndim < 3]
-  optim = GradAccClipAdamWGroup(
-    GradAccClipAdamW(params_wd, 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=1, device=optim_device),
-    GradAccClipAdamW(params_no_wd, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=0.0, grad_acc=1, device=optim_device),
-  )
-
-  for p in optim.params:
-    p.grad = p.zeros_like(dtype=dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype).contiguous()
-    if getattr(p, "_zero2", False): p.grad = optim.optimizers[0]._zero_shard(p.grad)
-  grads = [p.grad for p in optim.params]
-
-  from extra.gemm.cdna_asm_gemm import _mx_block_scale
-  model_state = get_state_dict(model)
-  def _scale_key(n):
-    if "." in n and (c:=f"{(b:=n.rsplit('.',1))[0]}_scale.{b[1]}") in model_state: return c
-    return f"{n}_scale"
-  fp8_scale_names = {n: _scale_key(n) for n, t in model_state.items() if t.dtype == FP8_DTYPE}
-  fp8_inv_scales = [model_state[sname] for sname in fp8_scale_names.values()]
-  for wname, sname in fp8_scale_names.items():
-    w, scale = model_state[wname], model_state[sname]
-    w._inv_scale = scale
-    if optim.master_params:
-      master = optim.master_params[next(j for j, p in enumerate(optim.params) if p is w)]
-      inv = scale if scale.device == master.device else scale.to(master.device)
-      bs = _mx_block_scale(inv.reshape(-1, inv.shape[-1])).reshape(w.shape)
-      master.assign((master * bs).contiguous())
-
-  scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)
-
-  if optim.master_params:
-    for m in optim.master_params: m.realize()
-  Tensor.realize(*optim.params, *fp8_inv_scales)
-
-  @TinyJit
-  @Context(TRAINING=1)
-  def step(tokens:Tensor):
-    if is_dp: tokens = tokens.to(None).shard(device, 0)
-    if not is_sharding: tokens = tokens.to(None)
-
-    logits:Tensor = model(tokens[:, :-1], save=True)
-    if getenv("FUSED_CE", 0):
-      from extra.llama_kernels.fused_ce import fused_ce_loss
-      loss = fused_ce_loss(logits.cast(dtypes.bfloat16), tokens[:, 1:], label_smoothing=0.0)
-    else:
-      loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-
-    for g, new_g in zip(grads, loss.gradient(*optim.params)):
-      apply_grad(g, new_g.uop)
-
-    Tensor.realize(loss, *grads)
-
-    clipped_grads, grad_norm = fclip_grads(grads, 1.0)
-    optim.fstep(clipped_grads, grad_norm)
-    scheduler.step()
-
-    loss_cpu = loss.flatten().float().to("CPU")
-    lr_cpu = optim.lr.float().to("CPU")
-    grad_norm_cpu = grad_norm.float().to("CPU")
-    Tensor.realize(loss_cpu, lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales)
-
-    return loss_cpu, lr_cpu, grad_norm_cpu
-
-  @TinyJit
-  @Context(TRAINING=0)
-  def eval_step(tokens:Tensor):
-    if is_dp: tokens = tokens.to(None).shard(device, 0)
-    if not is_sharding: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1])
-    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    return loss.flatten().float().to("CPU")
-
-  # ** data iters **
-  def fake_data(bs, samples):
-    import numpy as np
-    for _ in range(samples // bs):
-      fake_data_np = np.random.randint(0, real_vocab_size, size=(bs, SEQLEN + 1), dtype=np.int32)
-      yield Tensor(fake_data_np, device="NPY")
-
-  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=True)
-
-  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=True)
-
-  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)
-
-  num_params = sum(p.numel() for p in params) - model_params["vocab_size"]*model_params["dim"]
-  train_iter = get_train_iter()
-  i, sequences_seen = 0, 0
-  step_times = []
-
-  while i < MAX_STEPS:
-    GlobalCounters.reset()
-    actual_gbs = GBS if i >= 2 else BS
-    if getenv("TRAIN", 1):
-      profile_marker(f"train @ {i}")
-      st = time.perf_counter()
-
-      ist = time.perf_counter()
-
-      try: tokens = next(train_iter)
-      except StopIteration: break
-      mst = time.perf_counter()
-      data_time = mst - ist
-
-      ret = step(tokens)
-      dev_time = time.perf_counter() - mst
-
-      loss, lr, grad_norm = ret[0].item(), ret[1].item(), ret[2].item()
-      et = time.perf_counter()
-
-      step_time = et - st
-      if BENCHMARK: step_times.append(step_time)
-
-      i += 1
-      sequences_seen += actual_gbs
-
-      mem_gb = GlobalCounters.mem_used / 1e9
-      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 100
-      tqdm.write(
-          f"{i:5} {step_time:.3f} s step, {dev_time:.3f} s dev, {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/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}/gptoss_{i}.safe"
-        safe_save(get_state_dict(model), fn)
-
-        tqdm.write("saving optim checkpoint")
-        fn = f"{ckpt_dir}/gptoss_{i}_optim.safe"
-        safe_save(get_state_dict(scheduler), fn)
-
-      if i == BENCHMARK:
-        median_step_time = sorted(step_times)[BENCHMARK // 2]
-        estimated_steps = MAX_STEPS
-        estimated_total_minutes = int(median_step_time * estimated_steps / 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
-      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
-
-      log_perplexity = sum(eval_losses) / len(eval_losses)
-
-      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"):
-          if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
-          fn = f"{ckpt_dir}/gptoss.safe"
-          safe_save(get_state_dict(model), fn)
-        break
-
 def train_stable_diffusion():
  from extra.models.unet import UNetModel
  from examples.mlperf.dataloader import batch_load_train_stable_diffusion
@@ -2011,7 +1731,7 @@ def train_stable_diffusion():
    # move to CPU first so more GPU bufs aren't created (can trigger OOM)
    for k,v in ckpt.items(): ckpt[k] = v.detach().to("CPU")
    Tensor.realize(*[v for v in ckpt.values()])
-    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype).contiguous()
+    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype.base).contiguous()
    Tensor.realize(*[v for v in ckpt.values()])
    return ckpt

@@ -2078,7 +1798,7 @@ if __name__ == "__main__":
  elif getenv("RUNMLPERF"): bench_log_manager = WallTimeEvent(BenchEvent.MLPERF_RUN)
  else: bench_log_manager = contextlib.nullcontext()

-  with Context(TRAINING=1):
+  with Tensor.train():
    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
      nm = f"train_{m}"
      if nm in globals():
@@ -2,8 +2,9 @@ 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::gfx950"
+  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"
@@ -12,7 +13,7 @@ if __name__ == "__main__":
    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, round_up
+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
 from extra.llama_kernels.rmsnorm import rmsnorm
@@ -22,10 +23,6 @@ ASM_GEMM = getenv("ASM_GEMM", 0)
 FUSED_INPUT_QUANTIZE = getenv("FUSED_INPUT_QUANTIZE", 0)
 FUSED_ADD_NORM_MUL_QUANTIZE = getenv("FUSED_ADD_NORM_MUL_QUANTIZE", 0)
 FUSED_SILU_W13 = getenv("FUSED_SILU_W13", 0)
-SPLIT_W13 = getenv("SPLIT_W13", 0)
-COLUMNWISE_WEIGHT_SCALE = getenv("COLUMNWISE_WEIGHT_SCALE", 0)
-MXFP8 = getenv("MXFP8", 0)
-MXFP4 = getenv("MXFP4", 0)

 FP8_DTYPE = dtypes.fp8e4m3
 FP8_GRAD_DTYPE = dtypes.fp8e5m2
@@ -38,98 +35,58 @@ def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
  return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal(), new_amax

 def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
-           x_fp8:Tensor|None=None, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None, x_prequant_mx:tuple|None=None,
-           next_amax_x:Tensor|None=None) -> tuple[Tensor,...]:
+           x_fp8:Tensor|None=None, x_scale:Tensor|None=None, x_new_amax:Tensor|None=None,
+           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
  if not fp8:
    if ASM_GEMM:
      from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
      if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T),)
    return (x @ w.T,)
-  if MXFP4:
-    assert x is not None, "MXFP4 matmul requires an unquantized input"
-    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm
-    if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T, mxfp4=True),)
-    return (x @ w.T,)
  assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
-  if MXFP8:
-    from extra.gemm.cdna_asm_gemm import asm_gemm, quantize_mxfp8, mx_pack, can_use_asm_gemm, _mx_block_scale
-    if x_prequant_mx is not None: x_q, x_e8, x_si = x_prequant_mx       # fused producer already quantized (2d)
-    else: x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
-    l_shape = x.shape[:-1] if x is not None else x_q.shape[:-1]
-    if can_use_asm_gemm(x_q, w.T):
-      out = asm_gemm(x_q, w.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(w_inv_scale), w_inv_scale),
-                     mx_w_stored=True).reshape(*l_shape, w.shape[0])
-    else:
-      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*l_shape, x_q.shape[-1])
-      out = x_phys @ (w.cast(dtypes.bfloat16) * _mx_block_scale(w_inv_scale)).T
-    return out, x_q
  if x_fp8 is None:
-    if FUSED_INPUT_QUANTIZE:
+    if FUSED_INPUT_QUANTIZE and amax_x is not None:
      from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed
-      x_fp8, _ = quantize_fp8_delayed(x, amax_x, next_amax_x, FP8_DTYPE)
+      x_fp8, x_scale, x_new_amax, _ = quantize_fp8_delayed(x, amax_x, FP8_DTYPE)
    else:
-      x_fp8, _, new_amax_x = quantize_fp8(x, amax_state=amax_x)
-      next_amax_x.assign(new_amax_x)
+      x_fp8, x_scale, x_new_amax = quantize_fp8(x, amax_state=amax_x)
  if ASM_GEMM:
    from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
    if can_use_asm_gemm(x_fp8, w.T):
-      assert amax_x is not None
-      if COLUMNWISE_WEIGHT_SCALE:
-        out = asm_gemm(x_fp8, w.T, x_scale=amax_x, grad_amax_state=grad_amax_state,
-                       next_grad_amax_state=next_grad_amax_state, w_post_scale=w_inv_scale)
-      else:
-        out = asm_gemm(x_fp8, w.T, x_scale=amax_x, w_scale=w_inv_scale, grad_amax_state=grad_amax_state,
-                       next_grad_amax_state=next_grad_amax_state)
-      return out, x_fp8
-  return (x_fp8.dot(w.T, dtype=dtypes.float) * ((amax_x.float() + 1e-8) / FP8_MAX) * w_inv_scale).cast(dtypes.bfloat16), x_fp8
+      return asm_gemm(x_fp8, w.T, x_scale=x_scale, w_scale=w_inv_scale, grad_amax_state=grad_amax_state), x_new_amax, x_fp8, w
+  return x_fp8.dot(w.T, dtype=dtypes.float) * x_scale * w_inv_scale, x_new_amax, x_fp8, w

-def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
-                         next_amax_x:Tensor|None, grad_amax_state:Tensor|None, next_grad_amax_state:Tensor|None):
-  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
+def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor, grad_amax_state:Tensor):
+  if FUSED_ADD_NORM_MUL_QUANTIZE:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_rmsnorm_mul_quantize_fp8
-    x_fp8, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
-    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x,
-                       grad_amax_state=grad_amax_state, next_grad_amax_state=next_grad_amax_state)
+    x_fp8, x_inv_scale, new_amax, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE)
+    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, x_scale=x_inv_scale, x_new_amax=new_amax, grad_amax_state=grad_amax_state)
    return out, x_normed, rrms, ret
  x_normed, rrms = rmsnorm(x, eps)
-  out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale, grad_amax_state=grad_amax_state,
-                     next_grad_amax_state=next_grad_amax_state, next_amax_x=next_amax_x)
+  out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale, grad_amax_state=grad_amax_state)
  return out, x_normed, rrms, ret

-def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
-                             next_amax_x:Tensor|None, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
-  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
+def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor):
+  if FUSED_ADD_NORM_MUL_QUANTIZE:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
-    x_fp8, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
-    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x,
-                       grad_amax_state=grad_amax_state, next_grad_amax_state=next_grad_amax_state)
+    x_fp8, x_inv_scale, new_amax, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE)
+    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, x_scale=x_inv_scale, x_new_amax=new_amax)
    return out, h, x_normed, rrms, ret
  h = x + residual
  x_normed, rrms = rmsnorm(h, eps)
-  out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale, grad_amax_state=grad_amax_state,
-                     next_grad_amax_state=next_grad_amax_state, next_amax_x=next_amax_x)
+  out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale)
  return out, h, x_normed, rrms, ret

 def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
-                             amax_x2:Tensor|None, next_amax_x2:Tensor|None,
-                             grad_amax_xw13:Tensor|None, next_grad_amax_xw13:Tensor|None,
-                             grad_amax_xout:Tensor|None, next_grad_amax_xout:Tensor|None):
-  if FUSED_SILU_W13 and MXFP4:
-    from extra.llama_kernels.swiglu import swiglu
-    out, *ret = matmul(swiglu(x_w13), w2, amax_x=amax_x2, w_inv_scale=s_2, grad_amax_state=grad_amax_xout,
-                       next_grad_amax_state=next_grad_amax_xout, next_amax_x=next_amax_x2)
-    return out, ret
-  if FUSED_SILU_W13 and not MXFP4:
+                             amax_x2:Tensor,
+                             grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
+  if FUSED_SILU_W13:
    from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
-    x2_fp8 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13,
-                                                 next_grad_amax_state=next_grad_amax_xw13, amax_out=next_amax_x2)
-    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, amax_x=amax_x2,
-                       grad_amax_state=grad_amax_xout, next_grad_amax_state=next_grad_amax_xout)
+    x2_fp8, x2_inv_scale, new_amax_x2 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13)
+    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, x_scale=x2_inv_scale, x_new_amax=new_amax_x2, grad_amax_state=grad_amax_xout)
    return out, ret
  hidden = x_w13.shape[-1] // 2
  x_w1, x_w3 = x_w13[..., :hidden], x_w13[..., hidden:]
-  out, *ret = matmul(x_w1.silu() * x_w3, w2, amax_x=amax_x2, w_inv_scale=s_2, grad_amax_state=grad_amax_xout,
-                     next_grad_amax_state=next_grad_amax_xout, next_amax_x=next_amax_x2)
+  out, *ret = matmul(x_w1.silu() * x_w3, w2, amax_x=amax_x2, w_inv_scale=s_2, grad_amax_state=grad_amax_xout)
  return out, ret

 class FlatTransformer:
@@ -146,16 +103,13 @@ class FlatTransformer:
    scaled_std = 0.02 / math.sqrt(2 * n_layers)

    # Attention
-    self.wqkv, s_qkv = self.lin_per_layer(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2)
-    self.wo, s_o = self.lin_per_layer(self.n_heads * self.head_dim, dim, std=scaled_std)
+    self._init_inv_scales = []  # populated by lin_per_layer
+    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, std=scaled_std)

    # FeedForward
-    if SPLIT_W13:
-      self.w1, s_1 = self.lin_per_layer(dim, hidden_dim)
-      self.w3, s_3 = self.lin_per_layer(dim, hidden_dim)
-    else:
-      self.w13, s_13 = self.lin_per_layer(dim, hidden_dim * 2)
-    self.w2, s_2 = self.lin_per_layer(hidden_dim, dim, std=scaled_std)
+    self.w13 = self.lin_per_layer(dim, hidden_dim * 2)
+    self.w2 = self.lin_per_layer(hidden_dim, dim, std=scaled_std)

    self.norm_eps = norm_eps
    self.attention_norm = Tensor.ones(n_layers, dim).contiguous()
@@ -166,124 +120,93 @@ class FlatTransformer:
    self.tok_embeddings = nn.Embedding(vocab_size, dim)
    self.tok_embeddings.weight = Tensor.normal(vocab_size, dim, mean=0.0, std=0.02, dtype=dtypes.bfloat16)
    self.output = Tensor.normal(1, vocab_size, dim, mean=0.0, std=0.02, dtype=dtypes.bfloat16)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).clone().is_param_(False)
+    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)

-    def _amax(): return Tensor.full((), FP8_MAX, dtype=dtypes.float32).contiguous().is_param_(False)
-    n_amax = 0 if MXFP4 else n_layers
-    names = ["xqkv", "xo", "x2"]
-    names += ["x1", "x3"] if SPLIT_W13 else ["x13"]
-    self._fp8_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
-    self._fp8_next_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
-    grad_names = ["xqkv", "xo", "xout"]
-    grad_names += ["xw1", "xw3"] if SPLIT_W13 else ["xw13"]
-    self._fp8_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
-    self._fp8_next_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
-    w_scales = [("wqkv", s_qkv), ("wo", s_o), ("w2", s_2)]
-    w_scales += [("w1", s_1), ("w3", s_3)] if SPLIT_W13 else [("w13", s_13)]
-    self._fp8_inv_scale = {name: (s if MXFP8 else s.float()).contiguous().is_param_(False) for name, s in w_scales}
-    self._fp8_next_inv_scale = {name: (s if MXFP8 else s.float()).contiguous().is_param_(False) for name, s in w_scales}
+    def _amax(): return Tensor.full((), FP8_MAX, dtype=dtypes.float32).contiguous().requires_grad_(False)
+    names = ["xqkv", "xo", "x13", "x2"]
+    self._fp8_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
+    grad_names = ["xqkv", "xo", "xw13", "xout"]
+    self._fp8_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
+    w_names = ["wqkv", "wo", "w13", "w2"]
+    self._fp8_inv_scale = {wname: inv_scales.float().contiguous().requires_grad_(False)
+                           for wname, inv_scales in zip(w_names, self._init_inv_scales)}
+    del self._init_inv_scales

-  def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02, w:Tensor|None=None):
-    if w is None:
-      if getenv("ZEROS"): w = Tensor.zeros(self.n_layers, out_features, in_features)
-      else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std)
-    if MXFP8:
-      from extra.gemm.cdna_asm_gemm import quantize_mxfp8
-      w_q, w_e8, _ = quantize_mxfp8(w.reshape(self.n_layers * out_features, in_features))
-      return w_q.reshape(self.n_layers, out_features, in_features), w_e8.reshape(self.n_layers, out_features, in_features // 32)
-    if MXFP4:
-      # FP4 is produced dynamically so optimizer updates always start from the current BF16 weight.
-      return w.cast(dtypes.bfloat16), Tensor.ones(self.n_layers)
-    amax = (w.abs().max(axis=2) if COLUMNWISE_WEIGHT_SCALE else w.abs().flatten(1).max(1)).detach()
+  def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02):
+    if getenv("ZEROS"): w = Tensor.zeros(self.n_layers, out_features, in_features)
+    else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std)
+    amax = w.abs().flatten(1).max(1).detach()
    scale = FP8_MAX / (amax + 1e-8)
-    inv_scale = (amax + 1e-8) / FP8_MAX
-    scale_b = scale.reshape(self.n_layers, out_features, 1) if COLUMNWISE_WEIGHT_SCALE else scale.reshape(-1, 1, 1)
-    return (w * scale_b).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE), inv_scale
+    self._init_inv_scales.append((amax + 1e-8) / FP8_MAX)
+    return (w * scale.reshape(-1, 1, 1)).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE)

-  def attention(self, x:Tensor, freqs_cis:Tensor, *, attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                amax_xqkv:Tensor|None, amax_xo:Tensor|None, s_qkv:Tensor, s_o:Tensor,
-                next_amax_xqkv:Tensor|None, next_amax_xo:Tensor|None,
-                grad_amax_xqkv:Tensor|None, grad_amax_xo:Tensor|None,
-                next_grad_amax_xqkv:Tensor|None, next_grad_amax_xo:Tensor|None):
+  def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
+                amax_xqkv:Tensor, amax_xo:Tensor, s_qkv:Tensor, s_o:Tensor,
+                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor):
    bsz, seqlen, _ = x.shape
-    saves = []
+    new_amaxs, saves = [], []

-    xqkv, x_normed, rrms, s = norm_quantize_matmul(x, attention_norm, wqkv, s_qkv, self.norm_eps,
-                                                                  amax_x=amax_xqkv, grad_amax_state=grad_amax_xqkv,
-                                                                  next_grad_amax_state=next_grad_amax_xqkv, next_amax_x=next_amax_xqkv)
-    saves.extend([x_normed, rrms, *s, xqkv])
+    xqkv, x_normed, rrms, ret = norm_quantize_matmul(x, attention_norm, wqkv, s_qkv, self.norm_eps,
+                                                     amax_x=amax_xqkv, grad_amax_state=grad_amax_xqkv)
+    saves.extend([x_normed, rrms])
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [xqkv])
+    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.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)
    if getenv("HK_FLASH_ATTENTION"):
-      from extra.thunder.amd.fa import flash_attention, fused_qkv_rope
-      xq, xk, xv = fused_qkv_rope(xqkv, freqs_cis, self.n_heads, self.n_kv_heads, self.head_dim)
-      attn, *save = flash_attention(xq, xk, xv, is_causal=True, write_flat=True)
+      from extra.thunder.amd.fa import flash_attention
+      attn, *save = flash_attention(xq, xk, xv, is_causal=True)
      saves.extend(save)
    else:
-      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.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)
      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)

-    out, *s = matmul(attn, wo, amax_x=amax_xo, w_inv_scale=s_o, grad_amax_state=grad_amax_xo,
-                               next_grad_amax_state=next_grad_amax_xo, next_amax_x=next_amax_xo)
-    saves.extend([*s, out])
-    return out, saves
+    out, *ret = matmul(attn, wo, amax_x=amax_xo, w_inv_scale=s_o, grad_amax_state=grad_amax_xo)
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [out])
+    return (out, *new_amaxs, *saves)

-  def feed_forward(self, x:Tensor, residual:Tensor, **kwargs):
-    saves = []
+  def feed_forward(self, x:Tensor, residual:Tensor, ffn_norm:Tensor, w13:Tensor, w2:Tensor,
+                   amax_x13:Tensor, amax_x2:Tensor, s_13:Tensor, s_2:Tensor,
+                   grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
+    new_amaxs, saves = [], []

-    if SPLIT_W13:
-      h = x + residual
-      x_normed, rrms = rmsnorm(h, self.norm_eps)
-      saves.extend([x_normed, rrms])
-      inp = x_normed * kwargs["ffn_norm"]
-      x_w1, *s = matmul(inp, kwargs["w1"], amax_x=kwargs["amax_x1"], w_inv_scale=kwargs["s_1"],
-                                  grad_amax_state=kwargs["grad_amax_xw1"], next_grad_amax_state=kwargs["next_grad_amax_xw1"],
-                                  next_amax_x=kwargs["next_amax_x1"])
-      saves.extend([*s, x_w1])
-      x_w3, *s = matmul(inp, kwargs["w3"], amax_x=kwargs["amax_x3"], w_inv_scale=kwargs["s_3"],
-                                  grad_amax_state=kwargs["grad_amax_xw3"], next_grad_amax_state=kwargs["next_grad_amax_xw3"],
-                                  next_amax_x=kwargs["next_amax_x3"])
-      saves.extend([*s, x_w3])
-      if FUSED_SILU_W13 and MXFP8:
-        from extra.llama_kernels.fused_silu_mul_quantize_mxfp8 import fused_silu_mul_quantize_mxfp8
-        aq, ae8, asi = fused_silu_mul_quantize_mxfp8(x_w1.reshape(-1, x_w1.shape[-1]), x_w3.reshape(-1, x_w3.shape[-1]))
-        out, *s = matmul(None, kwargs["w2"], x_prequant_mx=(aq, ae8, asi), amax_x=kwargs["amax_x2"],
-                         w_inv_scale=kwargs["s_2"], grad_amax_state=kwargs["grad_amax_xout"],
-                         next_grad_amax_state=kwargs["next_grad_amax_xout"], next_amax_x=kwargs["next_amax_x2"])
-        out = out.reshape(*x_w1.shape[:-1], kwargs["w2"].shape[0])
-      else:
-        out, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
-                         grad_amax_state=kwargs["grad_amax_xout"], next_grad_amax_state=kwargs["next_grad_amax_xout"],
-                         next_amax_x=kwargs["next_amax_x2"])
-      saves.extend([*s, out])
-    else:
-      x_w13, h, x_normed, rrms, s = add_norm_quantize_matmul(x, residual, kwargs["ffn_norm"], kwargs["w13"], kwargs["s_13"],
-                                                                          self.norm_eps, amax_x=kwargs["amax_x13"],
-                                                                          next_amax_x=kwargs["next_amax_x13"],
-                                                                          grad_amax_state=kwargs["grad_amax_xw13"],
-                                                                          next_grad_amax_state=kwargs["next_grad_amax_xw13"])
-      saves.extend([x_normed, rrms, *s, x_w13])
-      out, s = silu_w13_quantize_matmul(x_w13, kwargs["w2"], kwargs["s_2"], amax_x2=kwargs["amax_x2"],
-                                                     next_amax_x2=kwargs["next_amax_x2"],
-                                                     grad_amax_xw13=kwargs["grad_amax_xw13"],
-                                                     next_grad_amax_xw13=kwargs["next_grad_amax_xw13"],
-                                                     grad_amax_xout=kwargs["grad_amax_xout"],
-                                                     next_grad_amax_xout=kwargs["next_grad_amax_xout"])
-      saves.extend([*s, out])
-    return out, h, saves
+    x_w13, h, x_normed, rrms, ret = add_norm_quantize_matmul(x, residual, ffn_norm, w13, s_13, self.norm_eps,
+                                                             amax_x=amax_x13)
+    saves.extend([x_normed, rrms])
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [x_w13])
+
+    out, ret = silu_w13_quantize_matmul(x_w13, w2, s_2, amax_x2=amax_x2, grad_amax_xw13=grad_amax_xw13, grad_amax_xout=grad_amax_xout)
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [out])
+    return (out, h, *new_amaxs, *saves)

  @function(precompile=True, precompile_backward=True)
-  def run_layer(self, x:Tensor, freqs_cis:Tensor, attn_kwargs:dict, ffn_kwargs:dict, save:bool=True):
-    attn, attn_saves = self.attention(x, freqs_cis, **attn_kwargs)
-    ffn, h, ffn_saves = self.feed_forward(x, attn, **ffn_kwargs)
+  def run_layer(self, x:Tensor, freqs_cis:Tensor,
+                attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
+                ffn_norm:Tensor, w13:Tensor, w2:Tensor,
+                amax_xqkv:Tensor, amax_xo:Tensor,
+                amax_x13:Tensor, amax_x2:Tensor,
+                s_qkv:Tensor, s_o:Tensor, s_13:Tensor, s_2:Tensor,
+                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor,
+                grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
+    attn, *attn_ret = self.attention(x, freqs_cis, attention_norm, wqkv, wo,
+                                     amax_xqkv=amax_xqkv, amax_xo=amax_xo, s_qkv=s_qkv, s_o=s_o,
+                                     grad_amax_xqkv=grad_amax_xqkv, grad_amax_xo=grad_amax_xo)
+    attn_amaxs, attn_saves = attn_ret[:2], attn_ret[2:]
+    ffn, h, *ffn_ret = self.feed_forward(x, attn, ffn_norm, w13, w2,
+                                                   amax_x13=amax_x13, amax_x2=amax_x2, s_13=s_13, s_2=s_2,
+                                                   grad_amax_xw13=grad_amax_xw13, grad_amax_xout=grad_amax_xout)
+    ffn_amaxs, ffn_saves = ffn_ret[:2], ffn_ret[2:]
    h = h + ffn
-    if save: return (h, *attn_saves, *ffn_saves)
-    else: return (h,)
+    return (h, *attn_amaxs, *ffn_amaxs, *attn_saves, *ffn_saves)

  def shard(self, device:tuple[str, ...], mp:bool=False):
    from tinygrad.nn.state import get_parameters
@@ -291,69 +214,39 @@ class FlatTransformer:
      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
-      def _shard_fp8(name:str, axis:int, std:float=0.02):
-        w = getattr(self, name)
-        if MXFP8:
-          from extra.gemm.cdna_asm_gemm import quantize_mxfp8
-          w_bf16 = Tensor.empty(self.n_layers, w.shape[1], w.shape[2], dtype=dtypes.bfloat16).shard(device, axis=axis).randn_like() * std
-          w_q, w_e8, _ = quantize_mxfp8(w_bf16)
-          w.replace(w_q)
-          self._fp8_inv_scale[name].replace(w_e8.contiguous()).is_param_(False)
-          self._fp8_next_inv_scale[name].replace(w_e8.contiguous()).is_param_(False)
-        else:
-          w.shard_(device, axis=axis)
-          scale_axis = (1 if axis == 1 else None) if COLUMNWISE_WEIGHT_SCALE else None
-          self._fp8_inv_scale[name] = self._fp8_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
-          self._fp8_next_inv_scale[name] = self._fp8_next_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
-        Tensor.realize(w, self._fp8_inv_scale[name], self._fp8_next_inv_scale[name])
-      sstd = 0.02 / math.sqrt(2 * self.n_layers)
-      _shard_fp8("wqkv", 1)          # (n_layers, out, dim) shard out
-      _shard_fp8("wo", 2, sstd)      # (n_layers, dim, in) shard in
-      if SPLIT_W13:
-        _shard_fp8("w1", 1)
-        _shard_fp8("w3", 1)
-      else:
-        _shard_fp8("w13", 1)         # (n_layers, hidden*2, dim) shard out
-      _shard_fp8("w2", 2, sstd)      # (n_layers, dim, hidden) shard in
+      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.w13.shard_(device, axis=1).realize()           # (n_layers, hidden*2, dim) shard out
+      self.w2.shard_(device, axis=2).realize()            # (n_layers, dim, hidden) shard in
      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.shard_(device, axis=1).realize()
      self.freqs_cis.shard_(device, axis=None).realize()
-      for amax_dict in (self._fp8_amax, self._fp8_next_amax, self._fp8_grad_amax, self._fp8_next_grad_amax):
+      for amax_dict in (self._fp8_amax, self._fp8_grad_amax):
        for name in amax_dict:
          for i in range(len(amax_dict[name])):
-            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().is_param_(False)
+            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().requires_grad_(False)
+      for name in self._fp8_inv_scale:
+        self._fp8_inv_scale[name] = self._fp8_inv_scale[name].to(device).contiguous().requires_grad_(False)

-  def reset_amax(self):
-    for st in (self._fp8_next_amax, self._fp8_next_grad_amax):
-      for ts in st.values():
-        for t in ts: t.assign(0)
-
-  def update_amax(self):
-    for cur, nxt in ((self._fp8_amax, self._fp8_next_amax), (self._fp8_grad_amax, self._fp8_next_grad_amax)):
-      for name in cur:
-        for c, n in zip(cur[name], nxt[name]): c.assign(n)
-
-  def __call__(self, tokens:Tensor, save:bool=True):
+  def __call__(self, tokens:Tensor):
    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)
-    if not getenv("HK_FLASH_ATTENTION"): freqs_cis = freqs_cis[:, :tokens.shape[1], :, :, :]
-    a, na, ga, nga, s = self._fp8_amax, self._fp8_next_amax, self._fp8_grad_amax, self._fp8_next_grad_amax, self._fp8_inv_scale
-    def amax_kwargs(i:int, act_names:tuple[str, ...], grad_names:tuple[str, ...]) -> dict[str, Tensor|None]:
-      specs = (("amax_", a, act_names), ("next_amax_", na, act_names), ("grad_amax_", ga, grad_names), ("next_grad_amax_", nga, grad_names))
-      if MXFP4: return dict.fromkeys(f"{prefix}{name}" for prefix, _, names in specs for name in names)
-      return {f"{prefix}{name}":val[name][i] for prefix, val, names in specs for name in names}
+    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
+    a, ga, s = self._fp8_amax, self._fp8_grad_amax, self._fp8_inv_scale
    for i in range(self.n_layers):
-      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wo=self.wo[i], s_qkv=s["wqkv"][i], s_o=s["wo"][i],
-                         **amax_kwargs(i, ("xqkv", "xo"), ("xqkv", "xo")))
-      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], w2=self.w2[i], s_2=s["w2"][i], **amax_kwargs(i, ("x2",), ("xout",)))
-      if SPLIT_W13:
-        ffn_kwargs.update(w1=self.w1[i], w3=self.w3[i], s_1=s["w1"][i], s_3=s["w3"][i], **amax_kwargs(i, ("x1", "x3"), ("xw1", "xw3")))
-      else:
-        ffn_kwargs.update(w13=self.w13[i], s_13=s["w13"][i], **amax_kwargs(i, ("x13",), ("xw13",)))
-      h, *_ = self.run_layer(h, freqs_cis, attn_kwargs, ffn_kwargs, save=save)
+      h, *ret = self.run_layer(h, freqs_cis,
+                               self.attention_norm[i], self.wqkv[i], self.wo[i],
+                               self.ffn_norm[i], self.w13[i], self.w2[i],
+                               amax_xqkv=a["xqkv"][i], amax_xo=a["xo"][i],
+                               amax_x13=a["x13"][i], amax_x2=a["x2"][i],
+                               s_qkv=s["wqkv"][i], s_o=s["wo"][i],
+                               s_13=s["w13"][i], s_2=s["w2"][i],
+                               grad_amax_xqkv=ga["xqkv"][i], grad_amax_xo=ga["xo"][i],
+                               grad_amax_xw13=ga["xw13"][i], grad_amax_xout=ga["xout"][i])
+      for name, new_val in zip(["xqkv", "xo", "x13", "x2"], ret[:5]):
+        a[name][i].assign(new_val)

    logits = matmul(self.norm(h), self.output[0], fp8=False)[0]
    return logits
@@ -366,59 +259,41 @@ def apply_grad(grad_buf:Tensor, new_grad:UOp):
  pads = _get_pads(new_grad)
  if len(pads) <= 1:
    new_grad = new_grad.cast(grad_buf.dtype)
-    grad_buf.uop = grad_buf.uop.after(grad_buf.uop.store(grad_buf.uop + new_grad))
+    store = grad_buf.uop.store(grad_buf.uop + new_grad)
+    grad_buf.uop = grad_buf.uop.after(store)
    return
-  cur = grad_buf.uop
-  for pad in sorted(pads, key=lambda p: p.marg[0][0] if p.op == Ops.PAD else 0, reverse=True):
-    if pad.op == Ops.PAD:
-      grad_shrink = tuple([(p[0], s+p[0]) for s,p in zip(pad.src[0].shape, pad.marg)])
-      buf_slice = cur.shrink(grad_shrink)
-      cur = cur.after(buf_slice.store(buf_slice + pad.src[0].cast(cur.dtype)))
-    else:
-      cur = cur.after(cur.store(cur + pad.cast(cur.dtype)))
-  grad_buf.uop = cur
+  sorted_pads = sorted(pads, key=lambda p: p.marg[0][0] if p.op == Ops.PAD else 0)
+  inners_raw = [Tensor(p.src[0] if p.op == Ops.PAD else p, device=grad_buf.device) for p in sorted_pads]
+  if getenv("FUSED_PAD_GRAD_ACCUM", 0):
+    from extra.llama_kernels.fused_pad_grad_accum import fused_pad_grad_accum, can_fused_pad_grad_accum
+    if can_fused_pad_grad_accum(grad_buf, inners_raw):
+      grad_buf.uop = fused_pad_grad_accum(grad_buf, inners_raw).uop
+      return
+  inners = [t.cast(grad_buf.dtype) for t in inners_raw]
+  grad_buf.assign(grad_buf + inners[0].cat(*inners[1:], dim=0))

 if __name__ == "__main__":
  config = {}
  BS                 = config["BS"]                     = getenv("BS", 16)
  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
-  SMALL              = config["SMALL"]                  = getenv("SMALL", 0)

  from examples.llama3 import MODEL_PARAMS
-  model_params = MODEL_PARAMS[llama_size:=getenv("LLAMA3_SIZE", "8B")]["args"]
-  # vocab_size from 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
-
-  # 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_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
-  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))
-
-  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()
+  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
-  grad_dtype = lambda x: dtypes.bfloat16 if x.dtype in dtypes.fp8s else x.dtype
-  grads = {x:x.zeros_like(dtype=grad_dtype(x)).contiguous() for x in state.values() if x.is_param}
-
-  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
-  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts]
+  grads = {x:Tensor.zeros(x.shape, dtype=x.dtype, device=x.device).contiguous()
+           for x in state.values() if x.requires_grad is None}

  # print model size
  sz = 0
@@ -427,32 +302,23 @@ if __name__ == "__main__":
    sz += v.nbytes()
  print(f"total sz: {sz/1e9:.2f} GB")

-  with Timing("fake data: "): tokens = Tensor.randint(BS, SEQLEN+1, low=0, high=real_vocab_size, dtype=dtypes.int)
+  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 fwd_bwd(tokens:Tensor):
-    with Timing("python forward: "):
-      model.reset_amax()
-      logits = model(tokens[:, :-1], save=llama_size=="8B")
-      loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
+  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)):
        apply_grad(grads[t], g.uop)
-    with Timing("run fwd_bwd: "): loss.realize(*grads.values(), *fp8_amax, *fp8_grad_amax)
-
-  @TinyJit
-  def optim_step():
-    for g in grads.values(): g.assign(g.zeros_like())
-    Tensor.realize(*grads.values())
+    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")):
-      fwd_bwd(tokens)
-      optim_step()
+      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())))
@@ -1,359 +0,0 @@
-import math, os, functools
-if __name__ == "__main__":
-  os.environ["DEFAULT_FLOAT"] = "bfloat16"
-  os.environ["OPTIM_DTYPE"] = "bfloat16"
-  if "DEV" not in os.environ: os.environ["DEV"] = "NULL::gfx950"
-  # CDNA
-  os.environ["DEVICE_IN_FUNCTION_BUG"] = "1"
-  os.environ["ALL2ALL"] = "1"
-  os.environ["USE_ATOMICS"] = "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
-from extra.llama_kernels.rmsnorm import rmsnorm
-from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8, asm_gemm, can_use_asm_gemm
-from extra.gemm.moe_gemm import grouped_mx_gemm
-from extra.gemm.moe_routing import route, dispatch, combine
-
-FP8_DTYPE = dtypes.fp8e4m3
-FP8_MAX = 448.0
-INIT_STD = 0.02
-ASM_GEMM = getenv("ASM_GEMM", 0)
-
-
-def _quant_dequant_fwd(x:Tensor) -> Tensor:
-  # x (2d bf16) -> bf16 value after an mxfp8 round-trip (1x32 block scaling on the last axis)
-  M, K = x.shape
-  scale_K = K // 32
-  amax = x.float().reshape(M, scale_K, 32).abs().max(axis=-1)
-  e8 = (amax.maximum(1e-38).log2().floor() + 127).clamp(0, 254).cast(dtypes.uint8)
-  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(M, scale_K, 1).expand(M, scale_K, 32).reshape(M, K)
-  x_fp8 = (x.float() * qscale).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE).cast(dtypes.float32)
-  return (x_fp8 * _mx_block_scale(e8)).cast(dtypes.bfloat16)
-
-@functools.cache
-def _quant_dequant_fwd_fxn(x_p, device):
-  return _quant_dequant_fwd(Tensor(x_p, device=device))
-
-def _quant_dequant_bwd(grad:UOp, call:UOp) -> tuple:
-  return (Tensor(grad).cast(dtypes.bfloat16).uop,)
-
-def quant_dequant_mx(x:Tensor) -> Tensor:
-  fxn = _quant_dequant_fwd_fxn(x.as_param(0).uop, x.device)
-  return Tensor(UOp.maketuple(fxn.uop).call(x.uop, grad_fxn=_quant_dequant_bwd).gettuple(0))
-
-def _mx_scale(e8:Tensor) -> Tensor:
-  return _mx_block_scale(e8) if e8.ndim == 2 else _mx_block_scale_3d(e8)
-
-def _dequant_fwd(w_q:Tensor, w_scale:Tensor) -> Tensor:
-  return w_q.cast(dtypes.bfloat16) * _mx_scale(w_scale)
-
-@functools.cache
-def _dequant_fwd_fxn(wq_p, ws_p, device):
-  return _dequant_fwd(Tensor(wq_p, device=device), Tensor(ws_p, device=device))
-
-def _dequant_bwd(grad:UOp, call:UOp) -> tuple:
-  return (Tensor(grad).cast(dtypes.bfloat16).uop, None)
-
-def dequant_weight(w_q:Tensor, w_scale:Tensor) -> Tensor:
-  fxn = _dequant_fwd_fxn(w_q.as_param(0).uop, w_scale.as_param(1).uop, w_q.device)
-  call = UOp.maketuple(fxn.uop).call(w_q.uop, w_scale.uop, grad_fxn=_dequant_bwd)
-  return Tensor(call.gettuple(0))
-
-def matmul_mx(x:Tensor, w_q:Tensor, w_scale:Tensor) -> Tensor:
-  l_shape = x.shape[:-1]
-  if ASM_GEMM:
-    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm, mx_pack
-    x2, K, N = x.reshape(-1, x.shape[-1]), x.shape[-1], w_q.shape[0]
-    wq, ws = w_q, w_scale
-    if (pad := (-K) % 256):
-      x2 = x2.pad(((0, 0), (0, pad)))
-      wq = wq.pad(((0, 0), (0, pad)))
-      ws = ws.pad(((0, 0), (0, pad // 32)), value=127).cast(dtypes.uint8)
-    if (npad := (-N) % 256):
-      wq = wq.pad(((0, npad), (0, 0)))
-      ws = ws.pad(((0, npad), (0, 0)), value=127).cast(dtypes.uint8)
-    x_q, x_e8, x_si = quantize_mxfp8(x2)
-    if x_si is not None and can_use_asm_gemm(x_q, wq.T):
-      out = asm_gemm(x_q, wq.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(ws), ws), mx_w_stored=True)
-      return (out[:, :N] if npad else out).reshape(*l_shape, N).cast(dtypes.bfloat16)
-  x_phys = quant_dequant_mx(x.reshape(-1, x.shape[-1])).reshape(*l_shape, x.shape[-1])
-  w_phys = dequant_weight(w_q, w_scale)
-  return (x_phys @ w_phys.T).cast(dtypes.bfloat16)
-
-def _pad_to_mult(t:Tensor, axis:int, mult:int=256) -> Tensor:
-  if (r := (-t.shape[axis]) % mult) == 0: return t
-  pads = [(0, 0)] * t.ndim
-  pads[axis] = (0, r)
-  return t.pad(tuple(pads))
-
-def _pad_cols(t:Tensor) -> Tensor: return _pad_to_mult(t, -1)
-def _pad_rows(t:Tensor) -> Tensor: return _pad_to_mult(t, -2)
-
-def swiglu(x:Tensor, limit:float=7.0, alpha:float=1.702) -> Tensor:
-  x_glu, x_linear = x[..., ::2], x[..., 1::2]
-  x_glu = x_glu.clamp(max_=limit)
-  x_linear = x_linear.clamp(-limit, limit)
-  return (x_glu * (alpha * x_glu).sigmoid()) * (x_linear + 1)
-
-def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0) -> Tensor:
-  freqs = 1.0 / (theta ** (Tensor.arange(0, dim, 2, dtype=dtypes.float32)[:(dim // 2)] / dim))
-  freqs = Tensor.arange(end, dtype=dtypes.float32).unsqueeze(dim=1) * freqs.unsqueeze(dim=0)
-  return Tensor.stack(freqs.cos(), freqs.sin(), dim=-1).cast(dtypes.default_float).reshape(1, end, 1, dim//2, 2)
-
-class GPTOSS:
-  def __init__(self, dim:int, n_layers:int, n_heads:int, n_kv_heads:int, head_dim:int, n_experts:int, experts_per_tok:int,
-               intermediate_size:int, vocab_size:int, norm_eps:float=1e-5, rope_theta:int=150000, sliding_window:int=128,
-               swiglu_limit:float=7.0, max_context:int=8192):
-    self.dim, self.n_layers, self.n_heads, self.n_kv_heads, self.head_dim = dim, n_layers, n_heads, n_kv_heads, head_dim
-    self.n_rep = n_heads // n_kv_heads
-    self.n_experts, self.experts_per_tok, self.intermediate_size = n_experts, experts_per_tok, intermediate_size
-    self.vocab_size, self.norm_eps, self.sliding_window, self.swiglu_limit = vocab_size, norm_eps, sliding_window, swiglu_limit
-    self.sm_scale = 1.0 / math.sqrt(head_dim)
-
-    scaled_std = INIT_STD / math.sqrt(2 * n_layers)
-    q_dim, qkv_dim = n_heads * head_dim, head_dim * (n_heads + 2 * n_kv_heads)
-
-    # attn
-    self.wqkv, self.wqkv_scale = self._quant_weight(n_layers, qkv_dim, dim)
-    self.wqkv_bias = Tensor.zeros(n_layers, qkv_dim, dtype=dtypes.bfloat16).contiguous()
-    self.wo, self.wo_scale = self._quant_weight(n_layers, dim, q_dim, std=scaled_std)
-    self.wo_bias = Tensor.zeros(n_layers, dim, dtype=dtypes.bfloat16).contiguous()
-    self.sinks = Tensor.zeros(n_layers, n_heads, dtype=dtypes.bfloat16).contiguous()
-    self.attention_norm = Tensor.ones(n_layers, dim).contiguous()
-
-    # moe ffn
-    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
-    self.gate = Tensor.normal(n_layers, n_experts, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
-    self.gate_bias = Tensor.zeros(n_layers, n_experts, dtype=dtypes.bfloat16).contiguous()
-    self.w_gate_up, self.w_gate_up_scale = self._quant_weight(n_layers, n_experts, intermediate_size * 2, dim, moe=True)
-    self.w_gate_up_bias = Tensor.zeros(n_layers, n_experts, intermediate_size * 2, dtype=dtypes.bfloat16).contiguous()
-    self.w_down, self.w_down_scale = self._quant_weight(n_layers, n_experts, dim, intermediate_size, std=scaled_std, moe=True)
-    self.w_down_bias = Tensor.zeros(n_layers, n_experts, dim, dtype=dtypes.bfloat16).contiguous()
-
-    # output
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = nn.Embedding(vocab_size, dim)
-    self.tok_embeddings.weight = Tensor.normal(vocab_size, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
-    self.output = Tensor.normal(vocab_size, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
-    self.freqs_cis = precompute_freqs_cis(head_dim, max_context * 2, rope_theta).contiguous().is_param_(False)
-
-  def _quant_weight(self, *shape:int, std:float=INIT_STD, moe:bool=False):
-    def _one(*s:int):
-      w = Tensor.zeros(*s) if getenv("ZEROS") else Tensor.normal(*s, mean=0.0, std=std)
-      w_q, w_e8, _ = quantize_mxfp8(_pad_cols(_pad_rows(w)) if moe else w)
-      return w_q, w_e8.is_param_(False)
-    if moe:
-      qs = [_one(*shape[1:]) for _ in range(shape[0])]
-      for q in qs: q[0]._zero2 = True  # grad arrives sharded on the expert axis under ZeRO-2 (moe_gemm)
-      return [q[0] for q in qs], [q[1] for q in qs]
-    return _one(*shape)
-
-  def _attn_mask(self, seqlen:int, dtype) -> Tensor:
-    i, j = Tensor.arange(seqlen).reshape(seqlen, 1), Tensor.arange(seqlen).reshape(1, seqlen)
-    return (j <= i).where(0.0, -1e30).cast(dtype).contiguous()
-
-  def _sliding_attention(self, xq:Tensor, xk:Tensor, xv:Tensor, sinks:Tensor) -> Tensor:
-    bsz, seqlen, H, hd = xq.shape
-    KV, R, W = self.n_kv_heads, self.n_rep, self.sliding_window
-    assert seqlen % W == 0, f"seqlen {seqlen} must be a multiple of sliding_window {W} for banded attention"
-    nb = seqlen // W
-    q = xq.reshape(bsz, seqlen, KV, R, hd).permute(0, 2, 3, 1, 4).reshape(bsz, KV, R, nb, W, hd).float()
-    k, v = (x.permute(0, 2, 1, 3).reshape(bsz, KV, 1, nb, W, hd).float() for x in (xk, xv))
-    kk, vv = (x.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb].cat(x, dim=-2) for x in (k, v))
-    sc = (q @ kk.transpose(-1, -2)) * self.sm_scale   # (B,KV,R,nb,W,2W)
-    i, j, pv = Tensor.arange(W).reshape(W, 1), Tensor.arange(2 * W).reshape(1, 2 * W), Tensor.arange(nb).reshape(nb, 1, 1) >= 1
-    sc = ((j > i) & (j <= i + W) & (pv | (j >= W))).where(sc, -float("inf"))
-    sink = sinks.reshape(1, KV, R, 1, 1, 1).float()
-    m = sc.max(-1, keepdim=True).maximum(sink)
-    e = (sc - m).exp()
-    p = (e / (e.sum(-1, keepdim=True) + (sink - m).exp())).cast(dtypes.bfloat16)
-    attn = p @ vv.cast(dtypes.bfloat16)
-    return attn.reshape(bsz, KV, R, seqlen, hd).permute(0, 3, 1, 2, 4).reshape(bsz, seqlen, H * hd)
-
-  def attention(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, sliding:bool, *, attention_norm:Tensor, wqkv:Tensor,
-                wqkv_scale:Tensor, wqkv_bias:Tensor, wo:Tensor, wo_scale:Tensor, wo_bias:Tensor, sinks:Tensor):
-    bsz, seqlen, _ = x.shape
-    x_normed, rrms = rmsnorm(x, self.norm_eps)
-    qkv = matmul_mx(x_normed * attention_norm, wqkv, wqkv_scale) + wqkv_bias
-    qkv = qkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
-    xq = qkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
-    xk, xv = qkv[:, :, :, self.n_rep], qkv[:, :, :, self.n_rep + 1]
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)  # (B,N,H,D)/(B,N,KV,D)
-
-    fa_saves = []
-    if getenv("HK_FLASH_ATTENTION"):
-      from extra.thunder.amd.fa import flash_attention
-      attn, _, l_vec = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks, window=self.sliding_window if sliding else 0)
-      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
-      fa_saves = [xq, xk, xv, l_vec]
-    elif sliding:
-      attn = self._sliding_attention(xq, xk, xv, sinks)
-    else:
-      xqm = xq.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
-      xkm, xvm = xk.permute(0, 2, 1, 3).unsqueeze(2), xv.permute(0, 2, 1, 3).unsqueeze(2)
-      scores = (xqm @ xkm.transpose(-2, -1)).float() * self.sm_scale + mask
-      sink = sinks.reshape(1, self.n_kv_heads, self.n_rep, 1, 1).float()
-      m = scores.max(-1, keepdim=True).maximum(sink)
-      e = (scores - m).exp()
-      w = (e / (e.sum(-1, keepdim=True) + (sink - m).exp())).cast(dtypes.bfloat16)
-      attn = (w @ xvm).permute(0, 3, 1, 2, 4).reshape(bsz, seqlen, self.n_heads * self.head_dim)
-
-    out = matmul_mx(attn, wo, wo_scale) + wo_bias
-    return out, [x_normed, rrms, attn] + fa_saves
-
-  def feed_forward(self, x:Tensor, *, ffn_norm:Tensor, gate:Tensor, gate_bias:Tensor,
-                   w_gate_up:Tensor, w_gate_up_scale:Tensor, w_gate_up_bias:Tensor,
-                   w_down:Tensor, w_down_scale:Tensor, w_down_bias:Tensor):
-    x_normed, rrms = rmsnorm(x, self.norm_eps)
-    inp = x_normed * ffn_norm
-    logits = inp.float() @ gate.float().T + gate_bias.float()
-    dim, inter = self.dim, self.intermediate_size
-
-    if getenv("GROUPED_MOE", 0):
-      bsz, seqlen = x.shape[:2]
-      inp, logits = inp.reshape(-1, dim), logits.reshape(-1, self.n_experts)
-      r = route(logits, self.experts_per_tok, self.n_experts)
-      onehot = r.rows_e.one_hot(self.n_experts).float()
-      xg = dispatch(_pad_cols(inp.cast(dtypes.bfloat16)), r)
-      h = grouped_mx_gemm(xg, (w_gate_up, w_gate_up_scale), r.off)[:, :2*inter] + (onehot @ w_gate_up_bias.float()).cast(dtypes.bfloat16)
-      y = swiglu(h, self.swiglu_limit)
-      z = grouped_mx_gemm(_pad_cols(y.cast(dtypes.bfloat16)), (w_down, w_down_scale), r.off)[:, :dim] \
-          + (onehot @ w_down_bias.float()).cast(dtypes.bfloat16)
-      out = combine(z, r, inp.shape[0], self.experts_per_tok).reshape(bsz, seqlen, dim)
-      return out, [x_normed, rrms, xg, h, y, z, r.weights, r.dest_row, r.off]
-    else:
-      thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
-      weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
-
-      out = None
-      for e in range(self.n_experts):
-        gu_q, gu_s = w_gate_up[e][:2*inter, :dim].contiguous(), w_gate_up_scale[e][:2*inter, :dim//32].contiguous()
-        dn_q, dn_s = w_down[e][:dim, :inter].contiguous(), w_down_scale[e][:dim, :inter//32].contiguous()
-        gate_up = matmul_mx(inp, gu_q, gu_s) + w_gate_up_bias[e]
-        y = (matmul_mx(swiglu(gate_up, self.swiglu_limit), dn_q, dn_s) + w_down_bias[e]).contiguous()
-        contrib = weights[..., e:e+1].cast(y.dtype) * y
-        out = contrib if out is None else out + contrib
-    return out, [x_normed, rrms]
-
-  @function(precompile=True, precompile_backward=True)
-  def run_layer(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, sliding:bool, attn_kwargs:dict, ffn_kwargs:dict, save:bool=True):
-    attn, attn_saves = self.attention(x, freqs_cis, mask, sliding, **attn_kwargs)
-    h = x + attn
-    ffn, ffn_saves = self.feed_forward(h, **ffn_kwargs)
-    h = h + ffn
-    if save: return (h, *attn_saves, *ffn_saves)
-    return (h,)
-
-  def shard(self, device:tuple[str, ...], mp:bool=False):
-    assert not mp, "MP not supported"
-    from tinygrad.nn.state import get_parameters
-    for v in get_parameters(self): v.shard_(device, axis=None)
-    Tensor.realize(*get_parameters(self))
-
-  def __call__(self, tokens:Tensor, save:bool=True):
-    h = self.tok_embeddings(tokens)
-    bsz, seqlen = tokens.shape
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :seqlen, :, :, :]
-    mask_full = None if getenv("HK_FLASH_ATTENTION") else self._attn_mask(seqlen, dtypes.float32)
-    for i in range(self.n_layers):
-      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wqkv_scale=self.wqkv_scale[i],
-                         wqkv_bias=self.wqkv_bias[i], wo=self.wo[i], wo_scale=self.wo_scale[i], wo_bias=self.wo_bias[i],
-                         sinks=self.sinks[i])
-      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], gate=self.gate[i], gate_bias=self.gate_bias[i],
-                        w_gate_up=self.w_gate_up[i], w_gate_up_scale=self.w_gate_up_scale[i], w_gate_up_bias=self.w_gate_up_bias[i],
-                        w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
-      h, *_ = self.run_layer(h, freqs_cis, mask_full, i % 2 == 0, attn_kwargs, ffn_kwargs, save=save)
-
-    h_normed = self.norm(h)
-    pad = (-self.dim) % 256
-    h_padded, w_padded = h_normed.pad((None, None, (0, pad))), self.output.pad(((0, 0), (0, pad)))
-    if ASM_GEMM and can_use_asm_gemm(h_padded, w_padded.T): logits = asm_gemm(h_padded, w_padded.T)
-    else: logits = h_normed @ self.output.T
-    return logits
-
-def _get_pads(uop:UOp) -> list[UOp]:
-  if uop.op == Ops.ADD: return _get_pads(uop.src[0]) + _get_pads(uop.src[1])
-  return [uop]
-
-def apply_grad(grad_buf:Tensor, new_grad:UOp):
-  pads = _get_pads(new_grad)
-  if len(pads) <= 1:
-    new_grad = new_grad.cast(grad_buf.dtype)
-    grad_buf.uop = grad_buf.uop.after(grad_buf.uop.store(new_grad))
-    return
-  cur = grad_buf.uop
-  for pad in sorted(pads, key=lambda p: p.marg[0][0] if p.op == Ops.PAD else 0, reverse=True):
-    if pad.op == Ops.PAD:
-      grad_shrink = tuple((p[0], s+p[0]) for s,p in zip(pad.src[0].shape, pad.marg))
-      buf_slice = cur.shrink(grad_shrink)
-      cur = cur.after(buf_slice.store(pad.src[0].cast(cur.dtype)))
-    else:
-      cur = cur.after(cur.store(cur + pad.cast(cur.dtype)))
-  grad_buf.uop = cur
-
-GPT_OSS_20B = dict(dim=2880, n_layers=24, n_heads=64, n_kv_heads=8, head_dim=64, n_experts=32, experts_per_tok=4,
-                   intermediate_size=2880, vocab_size=128256, norm_eps=1e-5, rope_theta=150000, sliding_window=128,
-                   swiglu_limit=7.0)
-
-if __name__ == "__main__":
-  config = {}
-  BS      = config["BS"]      = getenv("BS", 16)
-  SEQLEN  = config["SEQLEN"]  = getenv("SEQLEN", 8192)
-
-  model_params = GPT_OSS_20B
-  real_vocab_size = model_params["vocab_size"]
-  if (layers := getenv("LAYERS")) != 0: model_params["n_layers"] = layers
-
-  model = GPTOSS(**model_params, max_context=SEQLEN)
-
-  state = nn.state.get_state_dict(model)
-  print("tensor count:", len(state))
-
-  from tinygrad import Device
-  is_dp = (DP := getenv("DP", 1)) > 1
-  device_count = DP
-  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(device_count))
-
-  if is_dp: model.shard(device)
-
-  # preallocate all the grad buffers and zero them out
-  grad_dtype = lambda x: dtypes.bfloat16 if x.dtype in dtypes.fp8s else x.dtype
-  grads = {x:x.zeros_like(dtype=grad_dtype(x)).contiguous() for x in state.values() if x.is_param}
-
-  # 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=real_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 is_dp: tokens = tokens.shard(device, axis=0)
-
-  @TinyJit
-  def fwd_bwd(tokens:Tensor):
-    with Timing("python forward: "):
-      logits = model(tokens[:, :-1], save=True)
-      loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    with Timing("python backward: "):
-      for t,g in zip(grads, loss.gradient(*grads)):
-        apply_grad(grads[t], g.uop)
-    with Timing("run fwd_bwd: "): loss.realize(*grads.values())
-
-  @TinyJit
-  def optim_step():
-    for g in grads.values(): g.assign(g.zeros_like())
-    Tensor.realize(*grads.values())
-
-  for i in range(6):
-    GlobalCounters.reset()
-    profile_marker(f"step {i}")
-    with Timing(colored(f"*** step {i}: ", "red")):
-      fwd_bwd(tokens)
-      optim_step()
-  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
@@ -1,68 +0,0 @@
-import unittest
-from tinygrad import Tensor, TinyJit
-from tinygrad.nn.state import get_parameters
-from examples.mlperf.models.flat_llama import apply_grad
-
-class FlatModel:
-  def __init__(self, n_layers:int, dim:int, hidden:int):
-    self.n_layers = n_layers
-    self.w1 = Tensor.uniform(n_layers, dim, hidden, low=-0.1, high=0.1)
-    self.w2 = Tensor.uniform(n_layers, hidden, dim, low=-0.1, high=0.1)
-    self.scale = Tensor.uniform(dim, low=0.9, high=1.1)
-    self.bias = Tensor.zeros(dim).contiguous()
-
-  def __call__(self, x:Tensor) -> Tensor:
-    h = x
-    for i in range(self.n_layers):
-      h = (h @ self.w1[i]).relu() @ self.w2[i] + h
-    return (h * self.scale + self.bias).sum()
-
-class TestApplyGradE2E(unittest.TestCase):
-  def _run_with_apply_grad(self, model, xs):
-    grads = {p: Tensor.zeros(p.shape, dtype=p.dtype).contiguous().realize() for p in get_parameters(model)}
-    for x in xs:
-      loss = model(x)
-      for p, g in zip(grads, loss.gradient(*grads)):
-        apply_grad(grads[p], g.uop)
-      Tensor.realize(loss, *grads.values())
-    return [grads[p] for p in get_parameters(model)]
-
-  def _run_reference(self, model, xs):
-    for x in xs: model(x).backward()
-    return [p.grad for p in get_parameters(model)]
-
-  def _assert_close(self, got, expected, atol, rtol):
-    for g, e in zip(got, expected):
-      self.assertTrue(g.allclose(e, atol=atol, rtol=rtol).item(), f"grad mismatch (max abs diff {(g - e).abs().max().item()})")
-
-  def _assert_match(self, model, xs, atol, rtol):
-    self._assert_close(self._run_with_apply_grad(model, xs), self._run_reference(model, xs), atol, rtol)
-
-  def test_e2e_single_step(self):
-    model = FlatModel(n_layers=3, dim=8, hidden=16)
-    Tensor.realize(*get_parameters(model))
-    self._assert_match(model, [Tensor.randn(2, 8).realize()], atol=1e-4, rtol=1e-4)
-
-  def test_e2e_multi_step_accumulation(self):
-    model = FlatModel(n_layers=4, dim=8, hidden=16)
-    Tensor.realize(*get_parameters(model))
-    self._assert_match(model, [Tensor.randn(2, 8).realize() for _ in range(3)], atol=1e-4, rtol=1e-4)
-
-  def test_e2e_jit(self):
-    model = FlatModel(n_layers=3, dim=8, hidden=16)
-    Tensor.realize(*get_parameters(model))
-    grads = {p: Tensor.zeros(p.shape, dtype=p.dtype).contiguous().realize() for p in get_parameters(model)}
-
-    @TinyJit
-    def fwd_bwd(x:Tensor):
-      loss = model(x)
-      for p, g in zip(grads, loss.gradient(*grads)): apply_grad(grads[p], g.uop)
-      Tensor.realize(loss, *grads.values())
-
-    xs = [Tensor.randn(2, 8).realize() for _ in range(3)]
-    for x in xs: fwd_bwd(x)
-    self._assert_close([grads[p] for p in get_parameters(model)], self._run_reference(model, xs), atol=1e-3, rtol=1e-3)
-
-
-if __name__ == "__main__":
-  unittest.main()
@@ -3,7 +3,8 @@ os.environ["WQKV"] = "1"
 import unittest
 import numpy as np
 from tinygrad import Tensor, nn, dtypes
-from tinygrad.device import Device
+from tinygrad.nn.state import get_parameters
+from tinygrad.device import is_dtype_supported, Device
 from examples.mlperf.models.llama import Transformer
 from examples.mlperf.models.flat_llama import FlatTransformer

@@ -44,6 +45,8 @@ class TestFlatLlama(unittest.TestCase):
    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]])
@@ -111,7 +114,7 @@ class TestFlatLlama(unittest.TestCase):
    self.assertEqual(ref_logits.shape, flat_logits.shape)
    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)

-  @unittest.skipUnless(dtypes.fp8e4m3 in Device[Device.DEFAULT].renderer.supported_dtypes(), "fp8 not supported on this device")
+  @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
@@ -1,75 +1,63 @@
 from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes
-from tinygrad.nn.optim import Optimizer, OptimizerGroup
+from tinygrad.nn.optim import Optimizer
 from tinygrad.helpers import FUSE_OPTIM, getenv
-from tinygrad.uop.ops import UOp, Ops, AxisType
+from tinygrad.uop.ops import UOp, Ops

 STOCHASTIC_ROUND = getenv("STOCHASTIC_ROUND", 0)
 MASTER_WEIGHTS = getenv("MASTER_WEIGHTS", 0)
-ZERO_OPTIM = getenv("ZERO_OPTIM", 0)
-FP8_AMAX_MARGIN = getenv("FP8_AMAX_MARGIN", 1.1)
-IMMEDIATE_SCALE = getenv("IMMEDIATE_SCALE", 0)
-MXFP8 = getenv("MXFP8", 0)

 def stochastic_round_bf16(x:Tensor) -> Tensor:
  bits = x.bitcast(dtypes.uint32)
  if isinstance(x.device, tuple):
    shape = x.uop.shard_shape if x.uop.axis is not None else x.shape
-    noise = Tensor(UOp(Ops.MSTACK, src=tuple(Tensor.rand(*shape, device=d).uop for d in x.device)))
+    noise = Tensor(UOp(Ops.MSTACK, dtypes.default_float, tuple(Tensor.rand(*shape, device=d).uop for d in x.device)))
  else:
    noise = x.rand_like()
  noise = (noise * 0xFFFF).cast(dtypes.uint32)
  return ((bits + noise) & 0xFFFF0000).bitcast(dtypes.float32).cast(dtypes.bfloat16)

-def clip_grads(grads:list[Tensor], grad_acc, clip_norm) -> Tensor:
-  for g in grads: g.assign(g / grad_acc)
-  total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
-  for g in grads: g.assign((g * (clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(g.dtype))
-  return total_norm
-
-def fclip_grads(grads:list[Tensor], clip_norm) -> Tensor:
-  total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
-  scale = (clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)
-  return [(g * scale).cast(g.dtype) for g in grads], total_norm
-
 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) for _ in [b1, b2])
-    self.zero = bool(ZERO_OPTIM) and isinstance(self.device, tuple) and not self.fused
-    self.m = [self._zero_shard(x) for x in self._new_optim_param()]
-    self.v = [self._zero_shard(x) for x in self._new_optim_param()]
+    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
-    if MASTER_WEIGHTS and self.params[0].dtype != dtypes.float32:
-      self.master_params:list[Tensor]|None = [self._zero_shard(p.to(self.device).float().contiguous()) for p in self.params]
+    self.master_params:list[Tensor]|None = [p.float().contiguous() for p in self.params] if MASTER_WEIGHTS and self.params[0].dtype != dtypes.float32 else None
+
+  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:
-      self.master_params = None
-
-  def _zero_shard(self, t:Tensor) -> Tensor:
-    if not self.zero or t.ndim < 2 or (t.shape[0] % len(self.device)) != 0: return t
-    return Tensor(t.uop._shard(0, UOp.range(len(self.device), -1, AxisType.DEVICE)).unshard(0)).clone()
-
-  def _zero_gather(self, t:Tensor) -> Tensor:
-    if not isinstance(t.device, tuple) or t.uop.axis != 0: return t
-    n, sz = len(t.device), t.shape[0] // len(t.device)
-    return Tensor.cat(*[t[p*sz:(p+1)*sz] for p in range(n)], dim=0)
-
-  def fschedule_step(self, grads:list[Tensor]) -> list[Tensor]:
-    updates, extra = self._step([], grads)
+      updates, extra = self._step([], grads)
    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i], self.master_params[i] if self.master_params else None))
+    # collect inv_scale tensors attached to fp8 params (set by _apply_update)
    fp8_inv_scales = [tt._inv_scale for tt in self.params if hasattr(tt, '_inv_scale')]
-    fp8_next_inv_scales = [tt._next_inv_scale for tt in self.params if hasattr(tt, '_next_inv_scale')]
-    return extra + self.params + self.buffers + (self.master_params or []) + fp8_inv_scales + fp8_next_inv_scales
+    to_realize = extra+self.params+self.buffers+(self.master_params or [])+fp8_inv_scales

-  def fstep(self, grads:list[Tensor], grad_norm:Tensor|None=None):
-    Tensor.realize(*([grad_norm] if grad_norm is not None else []), *self.fschedule_step(grads))
+    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
@@ -82,7 +70,7 @@ class GradAccClipAdamW(Optimizer):
      v_hat = v_new / (1.0 - self.b2_t)
      up = m_hat / (v_hat.sqrt() + self.eps)
      ret.append(self.lr * up)
-    return ret, [self.b1_t, self.b2_t] + self.m + self.v
+    return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]

  def _apply_update(self, t:Tensor, up:Tensor, master:Tensor|None=None) -> Tensor:
    w = master if master is not None else t
@@ -90,60 +78,13 @@ class GradAccClipAdamW(Optimizer):
    up = up.float().shard_like(w) + self.lr.to(w.device) * wd * w.detach()
    new_w = w.detach() - up
    if master is not None: master.assign(new_w)
-    if self.zero and not (MXFP8 and t.dtype in dtypes.fp8s): new_w = self._zero_gather(new_w)
-    # when master is offloaded to a different device than the param, results are resharded back onto the param's (sharded) device
-    offloaded = master is not None and master.device != t.device
-    if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16:
-      out = stochastic_round_bf16(new_w)
-      return out.shard_like(t) if offloaded else out
+    if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16: return stochastic_round_bf16(new_w)
    if t.dtype in dtypes.fp8s:
-      if MXFP8:
-        from extra.gemm.cdna_asm_gemm import quantize_mxfp8
-        w_q, w_e8, _ = quantize_mxfp8(new_w.reshape(-1, new_w.shape[-1]))
-        if self.zero: w_q, w_e8 = self._zero_gather(w_q), self._zero_gather(w_e8)
-        new_e8 = w_e8.reshape(t._inv_scale.shape)
-        t._inv_scale.assign(new_e8.shard_like(t._inv_scale) if offloaded else new_e8)
-        ret = w_q.reshape(t.shape)
-        return ret.shard_like(t) if offloaded else ret
      from examples.mlperf.models.flat_llama import FP8_MAX
-      if IMMEDIATE_SCALE:
-        amax_axis = tuple(range(t._inv_scale.ndim, new_w.ndim))
-        new_inv = ((new_w.float().abs().max(axis=amax_axis).detach() + 1e-8) / FP8_MAX).cast(t._inv_scale.dtype)
-        t._inv_scale.assign(new_inv.shard_like(t._inv_scale) if offloaded else new_inv)
-        scale = new_inv.reciprocal().reshape(*new_inv.shape, *([1]*(new_w.ndim-new_inv.ndim)))
-        ret = (new_w * scale).clamp(-FP8_MAX, FP8_MAX).cast(t.dtype)
-        return ret.shard_like(t) if offloaded else ret
-      # delayed scaling: reuse previous step's inv_scale
-      t._inv_scale.assign(t._next_inv_scale)
-      inv_scale = t._inv_scale.to(new_w.device) if offloaded else t._inv_scale
-      scale = inv_scale.reciprocal().reshape(*inv_scale.shape, *([1]*(new_w.ndim-inv_scale.ndim)))
-      scaled = (new_w * scale).clamp(-FP8_MAX, FP8_MAX)
-      ret = scaled.cast(t.dtype)
-      # update inv_scale for next step from quantized result
-      new_amax = (ret.float().abs().max(axis=tuple(range(inv_scale.ndim, ret.ndim))) * inv_scale * FP8_AMAX_MARGIN).detach()
-      new_inv = ((new_amax + 1e-8) / FP8_MAX).cast(t._inv_scale.dtype)
-      t._next_inv_scale.assign(new_inv.shard_like(t._next_inv_scale) if offloaded else new_inv)
-      return ret.shard_like(t) if offloaded else ret
-    out = new_w.cast(t.dtype)
-    return out.shard_like(t) if offloaded else out
-
-class GradAccClipAdamWGroup(OptimizerGroup):
-  def __init__(self, *optimizers:GradAccClipAdamW):
-    super().__init__(*optimizers)
-    for o in self.optimizers[1:]: o.lr = self.optimizers[0].lr
-  def fstep(self, grads:list[Tensor], grad_norm:Tensor|None=None):
-    offset = 0
-    to_realize = []
-    for o in self.optimizers:
-      n = len(o.params)
-      to_realize += o.fschedule_step(grads[offset:offset+n])
-      offset += n
-    Tensor.realize(*to_realize, *([grad_norm] if grad_norm is not None else []))
-  @property
-  def lr(self): return self.optimizers[0].lr
-  @property
-  def device(self): return self.optimizers[0].device
-  @property
-  def master_params(self):
-    mp = [mp for o in self.optimizers for mp in (o.master_params or [])]
-    return mp if mp else None
+      amax = new_w.float().abs().max(axis=tuple(range(1, new_w.ndim))).detach()  # per-layer amax for (n_layers, out, in)
+      scale = FP8_MAX / (amax + 1e-8)
+      fp8_w = (new_w * scale.reshape(-1, *([1]*(new_w.ndim-1)))).clamp(-FP8_MAX, FP8_MAX).cast(t.dtype)
+      if hasattr(t, '_inv_scale'):
+        t._inv_scale.assign(((amax + 1e-8) / FP8_MAX).cast(t._inv_scale.dtype))
+      return fp8_w
+    return new_w.cast(t.dtype)
@@ -0,0 +1 @@
+!*.txt
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
-
-export CHECK_OOB=0
-
-export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-# export BEAM_LOG_SURPASS_MAX=1
-# export BASEDIR="/raid/datasets/wiki"
-
-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-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 REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export BENCHMARK=10 BERT_LAYERS=2
-
-python3 examples/mlperf/model_train.py
@@ -1,20 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-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=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_8xMI300X"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_8xMI300x_${DATETIME}_${SEED}.log"
-
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,20 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD 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="." DEV=AMD
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-export BEAM_TIMEOUT_SEC=15
-export FP8_TRAIN=1
-# search
-IGNORE_BEAM_CACHE=1 BENCHMARK=10 BERT_LAYERS=2 RUNMLPERF=0 python3 examples/mlperf/model_train.py
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_8xMI350X"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_8xMI350x_${DATETIME}_${SEED}.log"
-
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=NV
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
-export BASEDIR="/raid/datasets/wiki"
-
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,16 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=NV
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,28 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." DEV=NV
-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 REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,18 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
-export BASEDIR="/raid/datasets/wiki"
-
-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,16 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." DEV=AMD
-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 REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_red_${DATETIME}_${SEED}.log"
-
-export HCQDEV_WAIT_TIMEOUT_MS=100000  # prevents hang?
-
-# init
-sleep 5 && sudo rmmod amdgpu || true
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,44 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
-export DEV=${DEV:-AMD}
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
-export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export MXFP8=${MXFP8:-1}
-export ZERO_OPTIM=${ZERO_OPTIM:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="gptoss"
-export BASEDIR="/raid/datasets/c4-8b/"
-export EVAL_TARGET=3.34 EVAL_FREQ=12288
-export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
-if [ -z "$FULL_LAYERS" ]; then
-  export LAYERS=${LAYERS:-2}
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,39 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
-export DEV=${DEV:-AMD}
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
-export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export MXFP8=${MXFP8:-1}
-export ZERO_OPTIM=${ZERO_OPTIM:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="gptoss"
-export BASEDIR="/raid/datasets/c4-8b/"
-export EVAL_TARGET=3.34 EVAL_FREQ=12288
-export END_LR="4e-5" WARMUP_STEPS=128 MAX_STEPS=1200000
-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=0
-
-python3 examples/mlperf/model_train.py
@@ -1,49 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
-export DEV=${DEV:-AMD}
-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=${ALL2ALL:-1}
-export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export FP8=${FP8:-1}
-export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export FAST_CE=${FAST_CE:-0}
-export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-0}
-export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-0}
-export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-0}
-export FUSED_SILU_W13=${FUSED_SILU_W13:-0}
-export SPLIT_W13=${SPLIT_W13:-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=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${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=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,44 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
-export DEV=${DEV:-AMD}
-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=${ALL2ALL:-1}
-export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export FP8=${FP8:-1}
-export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export FAST_CE=${FAST_CE:-0}
-export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-0}
-export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-0}
-export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-0}
-export FUSED_SILU_W13=${FUSED_SILU_W13:-0}
-export SPLIT_W13=${SPLIT_W13:-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:-1152}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${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
@@ -0,0 +1,28 @@
+# 1. Problem
+
+small llm pretraining: llama 3.1 8b on c4.
+
+## Requirements
+
+Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v6.0.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+
+# 2. Directions
+
+## Steps to download and verify data
+
+### 1. Download raw data
+
+follow mlperf steps to download the preprocessed c4 dataset.
+
+## Running
+
+### tinybox_8xMI350X
+
+#### Steps to run benchmark
+```
+examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/run_and_time.sh
+```
@@ -1,8 +1,6 @@
 #!/usr/bin/env bash

 export PYTHONPATH="."
-export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
-export PATH="$ROCM_PATH/bin:$PATH"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -16,17 +14,15 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export MXFP4=${MXFP4:-1}
+export FP8=${FP8:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
 export FAST_CE=${FAST_CE:-1}
 export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
-export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-1}
 export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
 export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
-export SPLIT_W13=${SPLIT_W13:-0}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -35,7 +31,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=2048 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -44,11 +40,11 @@ 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=0
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1

 export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
+  export LLAMA_LAYERS=2
 fi

 python3 examples/mlperf/model_train.py
@@ -1,54 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
-export DEV=${DEV:-AMD}
-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=${ALL2ALL:-1}
-export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export FP8=${FP8:-1}
-export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export FAST_CE=${FAST_CE:-0}
-export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-0}
-export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-0}
-export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-0}
-export FUSED_SILU_W13=${FUSED_SILU_W13:-0}
-export SPLIT_W13=${SPLIT_W13:-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=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,8 +1,6 @@
 #!/usr/bin/env bash

 export PYTHONPATH="."
-export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
-export PATH="$ROCM_PATH/bin:$PATH"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -16,17 +14,15 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export MXFP4=${MXFP4:-1}
+export FP8=${FP8:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
 export FAST_CE=${FAST_CE:-1}
 export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
-export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-1}
 export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
 export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
-export SPLIT_W13=${SPLIT_W13:-0}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -35,7 +31,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=2048 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -44,6 +40,6 @@ 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=0
+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
@@ -1,49 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
-export DEV=${DEV:-AMD}
-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=${ALL2ALL:-1}
-export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
-export USE_ATOMICS=${USE_ATOMICS:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export FP8=${FP8:-1}
-export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
-export FAST_CE=${FAST_CE:-0}
-export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-0}
-export FUSED_GRAD_QUANTIZE=${FUSED_GRAD_QUANTIZE:-0}
-export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-0}
-export FUSED_SILU_W13=${FUSED_SILU_W13:-0}
-export SPLIT_W13=${SPLIT_W13:-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=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -1,7 +1,6 @@
 #!/bin/bash
-set -e
-
-export BENCHMARK=${BENCHMARK:-5}
+export BENCHMARK=5
 export EVAL_BS=0
-VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
-[ "$BENCHMARK" -le 3 ] || [[ $DEV == NULL* ]] || python -m tinygrad.viz.cli -s AMD -t --interval "train @ 2" "train @ 3"
+VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=0 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam.sh
+SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
+python -m tinygrad.viz.cli -s "$SRC" -t
@@ -3,8 +3,6 @@ set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

 export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
 export DEV=AMD
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -17,16 +15,15 @@ export USE_ATOMICS=1
 export ASM_GEMM=1
 export WQKV=1
 export MASTER_WEIGHTS=1
-export MXFP4=1
+export FP8=1
 export ALLREDUCE_CAST=1
 export FAST_CE=1
 export FUSED_INPUT_QUANTIZE=1
-export FUSED_GRAD_QUANTIZE=1
 export FUSED_ADD_NORM_MUL_QUANTIZE=1
 export FUSED_SILU_W13=1
-export SPLIT_W13=0
+export FUSED_PAD_GRAD_ACCUM=1

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=8 MP=1 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=2
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -44,7 +41,7 @@ export SEED=$RANDOM
 export DATA_SEED=$SEED

 export JITBEAM=3
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1

 export LOGMLPERF=1

@@ -4,7 +4,7 @@ export EVAL_BS=0
 export FAKEDATA=1
 export NULL_ALLOW_COPYOUT=1
 export HIP_VISIBLE_DEVICES=""
-export DEV=NULL:HIP:gfx950
+export DEV=NULL
 export JITBEAM=0
 export LLAMA_LAYERS=${LLAMA_LAYERS:-"2"}
-time examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh
+time examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
@@ -1,50 +0,0 @@
-# 1. Problem
-
-This problem uses the ResNet-50 CNN to do image classification.
-
-## Requirements
-
-Install tinygrad and mlperf-logging from master.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-### tinybox_red
-Disable cwsr
-This is the default on production tinybox red.
-```
-sudo vi /etc/modprobe.d/amdgpu.conf
-cat <<EOF > /etc/modprobe.d/amdgpu.conf
-options amdgpu cwsr_enable=0
-EOF
-sudo update-initramfs -u
-sudo reboot
-
-# validate
-sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
-```
-
-# 2. Directions
-
-## Steps to download and verify data
-
-```
-IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
-```
-
-## Steps for one time setup
-
-### tinybox_red
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
-```
-
-## Steps to run benchmark
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-```
@@ -1,13 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=NV
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-export BENCHMARK=10 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=NV
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-export EVAL_START_EPOCH=3 EVAL_FREQ=4
-
-export WANDB=1 PARALLEL=0
-
-python3 examples/mlperf/model_train.py
@@ -1,25 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." DEV=NV
-export MODEL="resnet"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="resnet_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,50 +0,0 @@
-# 1. Problem
-
-This problem uses the ResNet-50 CNN to do image classification.
-
-## Requirements
-
-Install tinygrad and mlperf-logging from master.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-### tinybox_red
-Disable cwsr
-This is the default on production tinybox red.
-```
-sudo vi /etc/modprobe.d/amdgpu.conf
-cat <<EOF > /etc/modprobe.d/amdgpu.conf
-options amdgpu cwsr_enable=0
-EOF
-sudo update-initramfs -u
-sudo reboot
-
-# validate
-sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
-```
-
-# 2. Directions
-
-## Steps to download and verify data
-
-```
-IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
-```
-
-## Steps for one time setup
-
-### tinybox_red
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
-```
-
-## Steps to run benchmark
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-```
@@ -1,13 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=10 DEBUG=${DEBUG:-2}
-
-python3 examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export EVAL_START_EPOCH=3 EVAL_FREQ=4
-
-export WANDB=1 PARALLEL=0
-
-python3 examples/mlperf/model_train.py
@@ -1,26 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="resnet"
-export SUBMISSION_PLATFORM="tinybox_red"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="resnet_red_${DATETIME}_${SEED}.log"
-
-# init
-sleep 5 && sudo rmmod amdgpu || true
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,8 +0,0 @@
-#!/bin/bash
-
-rocm-smi --setprofile compute
-rocm-smi --setmclk 3
-rocm-smi --setperflevel high
-
-# power cap to 350W
-echo "350000000" | sudo tee /sys/class/drm/card{1..6}/device/hwmon/hwmon*/power1_cap
@@ -1,38 +0,0 @@
-# 1. Problem
-
-This problem uses RetinaNet for SSD.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-Also install the following dependencies:
-```
-pip install tqdm numpy pycocotools boto3 pandas torch torchvision
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download data
-
-Run the following:
-```
-BASEDIR=/raid/datasets/openimages python3 extra/datasets/openimages.py
-```
-
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh
-```
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=NV
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=NV
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -1,25 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." DEV=NV
-export MODEL="retinanet"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export TRAIN_BEAM=2 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/openimages"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="retinanet_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." DEV=AMD
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -0,0 +1,106 @@
+:::MLLOG {"namespace": "", "time_ms": 1778207373785, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778207373789, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778207373790, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778207373790, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778207373790, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778207373791, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778207373791, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778207734506, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747904, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "seed", "value": 25580, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747908, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778207747909, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778208080716, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778208080717, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778208901302, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778208901303, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778208952059, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.705078125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778208952060, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778208952060, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778209608282, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778209608282, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778209637796, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.552001953125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778209637796, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778209637797, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778210294879, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778210294879, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778210324584, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.1011962890625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778210324584, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778210324585, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778210980564, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778210980565, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778211010225, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.8807373046875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778211010225, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778211010226, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778211667184, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778211667185, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778211696784, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.7498779296875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778211696785, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778211696786, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778212356059, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778212356060, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778212385775, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.65478515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778212385776, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778212385776, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778213044774, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778213044775, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778213074311, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5731201171875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778213074312, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778213074313, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778213732225, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778213732225, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778213761806, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5137939453125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778213761806, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778213761807, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778214419768, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778214419769, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778214449443, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.46630859375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778214449444, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778214449445, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778215112018, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778215112019, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778215141586, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.428955078125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778215141586, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778215141587, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778215794970, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778215794970, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778215824346, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.390869140625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778215824346, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778215824347, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778216475810, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778216475810, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778216505269, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.361328125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778216505269, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778216505270, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778217157389, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778217157390, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778217186831, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.346923828125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778217186832, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778217186832, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778217846265, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778217846266, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778217876013, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3133544921875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778217876014, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778217876014, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778218532377, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778218532378, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778218561863, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.2989501953125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778218561863, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778218561864, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778218561864, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,111 @@
+:::MLLOG {"namespace": "", "time_ms": 1778218577779, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778218577783, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778218577784, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778218577784, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778218577784, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778218578371, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778218578371, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778218957180, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971058, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "seed", "value": 356, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971063, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971064, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971064, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971064, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971064, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971064, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778218971064, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778219289653, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778219289654, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778220097041, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778220097042, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778220141757, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.743896484375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778220141758, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778220141758, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778220795772, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778220795773, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778220825439, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.58349609375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778220825440, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778220825440, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778221480609, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778221480610, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778221510284, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.1131591796875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778221510285, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778221510286, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778222164664, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778222164665, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778222194290, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.8935546875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778222194291, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778222194291, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778222848846, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778222848847, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778222878557, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.7567138671875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778222878558, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778222878558, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778223532447, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778223532447, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778223562036, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.658203125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778223562037, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778223562037, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778224215343, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778224215344, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778224244924, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5860595703125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778224244925, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778224244925, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778224898378, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778224898379, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778224928021, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.51708984375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778224928021, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778224928022, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778225581424, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778225581425, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778225611002, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.471923828125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778225611003, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778225611003, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778226265043, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778226265044, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778226294659, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.43701171875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778226294660, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778226294661, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778226949577, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778226949577, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778226979238, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5406494140625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778226979239, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778226979239, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778227635352, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778227635352, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778227664978, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3836669921875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778227664978, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778227664979, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778228323150, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778228323151, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778228352865, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.355712890625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778228352865, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778228352866, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778229010307, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778229010307, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778229040142, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3319091796875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778229040143, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778229040143, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778229696378, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778229696379, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778229726195, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.30615234375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778229726195, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778229726196, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778230383239, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778230383240, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778230412831, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.29052734375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778230412832, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778230412832, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778230412833, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,111 @@
+:::MLLOG {"namespace": "", "time_ms": 1778230427283, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778230427287, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778230427287, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778230427287, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778230427287, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778230427939, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778230427939, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778230779581, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792886, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792890, "event_type": "POINT_IN_TIME", "key": "seed", "value": 2774, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792891, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792892, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792892, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792892, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792892, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792892, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778230792892, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778231115792, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778231115793, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778232030906, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778232030907, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778232075494, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.812255859375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778232075494, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778232075495, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778232729579, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778232729580, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778232759140, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.582275390625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778232759141, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778232759142, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778233413630, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778233413631, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778233443219, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.11767578125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778233443220, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778233443220, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778234097427, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778234097428, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778234127034, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.9005126953125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778234127034, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778234127035, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778234780955, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778234780956, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778234810558, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.7586669921875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778234810558, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778234810559, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778235463904, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778235463905, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778235493473, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.657958984375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778235493474, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778235493475, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778236147005, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778236147005, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778236176551, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.585693359375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778236176552, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778236176552, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778236830530, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778236830530, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778236860107, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.521484375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778236860108, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778236860108, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778237514002, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778237514003, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778237543592, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4742431640625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778237543592, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778237543593, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778238197935, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778238197936, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778238227501, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.428955078125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778238227502, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778238227503, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778238882036, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778238882037, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778238911645, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4019775390625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778238911645, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778238911646, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778239565129, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778239565130, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778239594721, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.37890625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778239594722, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778239594722, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778240248763, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778240248764, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778240278335, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3448486328125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778240278336, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778240278337, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778240933651, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778240933651, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778240963429, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.325439453125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778240963430, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778240963431, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778241626264, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778241626265, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778241656303, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3072509765625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778241656304, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778241656304, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778242315322, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778242315323, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778242345178, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.2781982421875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778242345178, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778242345179, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778242345179, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,106 @@
+:::MLLOG {"namespace": "", "time_ms": 1778242359541, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778242359545, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778242359545, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778242359545, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778242359545, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778242360117, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778242360118, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778242702158, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715949, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715953, "event_type": "POINT_IN_TIME", "key": "seed", "value": 1261, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715953, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715953, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715954, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715955, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715955, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715955, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778242715955, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778243033805, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778243033806, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778243851371, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778243851372, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778243896651, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.7802734375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778243896652, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778243896652, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778244555628, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778244555629, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778244585531, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.574951171875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778244585532, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778244585533, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778245246511, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778245246512, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778245276502, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.1171875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778245276503, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778245276503, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778245937187, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778245937187, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778245967058, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.8995361328125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778245967059, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778245967059, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778246626117, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778246626117, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778246656019, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.762451171875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778246656019, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778246656020, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778247315255, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778247315256, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778247345128, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.6572265625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778247345128, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778247345129, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778248003582, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778248003582, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778248033442, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.58740234375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778248033443, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778248033443, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778248692764, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778248692764, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778248722726, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5286865234375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778248722727, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778248722727, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778249383186, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778249383186, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778249413099, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.475830078125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778249413099, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778249413100, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778250072852, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778250072852, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778250102740, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4278564453125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778250102741, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778250102741, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778250762230, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778250762230, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778250792198, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.400146484375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778250792199, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778250792199, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778251455492, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778251455492, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778251485544, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3818359375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778251485545, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778251485545, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778252146772, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778252146772, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778252176776, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.345458984375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778252176776, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778252176777, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778252836585, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778252836586, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778252866442, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.322265625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778252866443, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778252866443, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778253526422, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778253526422, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778253556343, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.299072265625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778253556343, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778253556344, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778253556344, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,111 @@
+:::MLLOG {"namespace": "", "time_ms": 1778253570454, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778253570459, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778253570459, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778253570459, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778253570459, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778253571045, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778253571045, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778253944036, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957691, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "seed", "value": 14711, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957695, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778253957696, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778254276545, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778254276546, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778255100535, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778255100536, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778255143977, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.77978515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778255143977, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778255143978, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778255806844, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778255806845, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778255836518, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.578857421875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778255836519, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778255836520, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778256495933, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778256495933, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778256525443, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.1239013671875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778256525443, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778256525444, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778257180826, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778257180827, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778257210282, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.906494140625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778257210283, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778257210283, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778257866434, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778257866435, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778257895945, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.75244140625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778257895945, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778257895946, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778258550818, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778258550819, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778258580369, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.6553955078125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778258580369, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778258580370, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778259234200, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778259234201, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778259263770, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5762939453125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778259263771, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778259263772, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778259917494, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778259917495, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778259947011, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.52197265625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778259947012, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778259947013, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778260600453, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778260600454, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778260629950, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4765625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778260629951, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778260629951, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778261285126, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778261285127, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778261314809, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4378662109375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778261314810, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778261314810, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778261971632, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778261971632, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778262001260, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3968505859375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778262001261, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778262001261, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778262657393, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778262657394, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778262686962, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.365966796875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778262686962, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778262686963, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778263342665, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778263342666, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778263372176, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3365478515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778263372176, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778263372177, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778264027427, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778264027428, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778264056993, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3363037109375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778264056993, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778264056994, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778264710992, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778264710993, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778264740486, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3016357421875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778264740486, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778264740487, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778265396989, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778265396989, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778265426521, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.2861328125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778265426522, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778265426522, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778265426522, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,111 @@
+:::MLLOG {"namespace": "", "time_ms": 1778265440911, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778265440915, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778265440915, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778265440916, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778265440916, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778265441493, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778265441493, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778265779467, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792765, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "seed", "value": 27754, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792769, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778265792770, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778266108942, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778266108943, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778266913943, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778266913944, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778266957471, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.74072265625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778266957472, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778266957472, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778267616663, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778267616663, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778267648052, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.612060546875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778267648053, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778267648053, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778268306168, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778268306168, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778268335863, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.16552734375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778268335864, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778268335864, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778268998030, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778268998030, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778269027991, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.915283203125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778269027992, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778269027992, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778269689514, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778269689515, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778269719312, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.7637939453125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778269719313, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778269719313, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778270378319, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778270378320, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778270408037, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.6695556640625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778270408038, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778270408038, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778271066429, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778271066430, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778271096134, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.583251953125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778271096135, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778271096135, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778271754376, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778271754377, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778271784142, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.525146484375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778271784142, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778271784143, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778272442458, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778272442459, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778272472257, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4774169921875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778272472257, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778272472258, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778273129575, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778273129576, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778273159231, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.443359375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778273159231, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778273159232, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778273816098, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778273816099, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778273845769, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4072265625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778273845770, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778273845770, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778274505683, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778274505684, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778274535540, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3677978515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778274535541, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778274535541, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778275195662, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778275195662, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778275225396, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4146728515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778275225397, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778275225397, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778275884245, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778275884246, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778275913924, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3697509765625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778275913925, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778275913925, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778276570930, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778276570931, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778276600619, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.321533203125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778276600620, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778276600620, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778277262406, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778277262407, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778277292466, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.287353515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778277292467, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778277292467, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778277292468, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,111 @@
+:::MLLOG {"namespace": "", "time_ms": 1778277306868, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778277306872, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778277306872, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778277306873, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778277306873, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778277307428, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778277307429, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778277671564, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685153, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685157, "event_type": "POINT_IN_TIME", "key": "seed", "value": 17816, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685157, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685157, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685158, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685159, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685159, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778277685159, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778278007248, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778278007260, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778278810368, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778278810369, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778278855284, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.768798828125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778278855285, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778278855285, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778279519460, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778279519461, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778279549391, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.568603515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778279549392, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778279549392, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778280214562, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778280214563, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778280244495, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.151123046875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778280244496, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778280244496, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778280909906, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778280909906, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778280939913, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.9197998046875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778280939913, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778280939914, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778281607749, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778281607750, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778281637814, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.7734375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778281637815, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778281637815, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778282306223, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778282306224, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778282336322, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.673583984375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778282336323, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778282336323, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778283007699, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778283007700, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778283037808, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.6011962890625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778283037808, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778283037809, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778283706598, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778283706598, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778283736748, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.526123046875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778283736748, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778283736749, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778284408590, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778284408590, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778284438316, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.475341796875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778284438317, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778284438317, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778285098897, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778285098898, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778285128703, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.432861328125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778285128703, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778285128704, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778285786660, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778285786660, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778285816222, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4031982421875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778285816222, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778285816223, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778286473781, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778286473782, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778286503417, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3638916015625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778286503418, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778286503418, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778287160556, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778287160556, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778287190213, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.341796875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778287190214, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778287190215, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778287846424, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778287846424, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778287876044, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.32177734375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778287876045, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778287876046, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778288531947, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778288531947, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778288561549, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5465087890625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778288561550, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778288561550, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778289220442, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778289220442, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778289250127, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.2855224609375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778289250128, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778289250128, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778289250129, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,106 @@
+:::MLLOG {"namespace": "", "time_ms": 1778289264340, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778289264344, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778289264344, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778289264344, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778289264344, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778289264911, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778289264912, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778289599730, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613197, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613200, "event_type": "POINT_IN_TIME", "key": "seed", "value": 16781, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613201, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613202, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613202, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613202, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613202, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613202, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778289613202, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778289929875, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778289929878, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778290756967, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778290756968, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778290801735, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.758544921875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778290801736, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778290801736, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778291460896, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778291460896, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778291490685, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.683349609375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778291490685, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778291490686, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778292152773, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778292152774, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778292182518, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.1280517578125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778292182519, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778292182519, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778292842100, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778292842101, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778292871768, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.90185546875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778292871769, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778292871769, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778293529314, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778293529315, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778293559042, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.757080078125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778293559043, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778293559043, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778294218188, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778294218189, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778294247880, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.6575927734375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778294247880, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778294247881, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778294908017, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778294908018, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778294937688, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.586181640625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778294937689, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778294937690, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778295595710, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778295595710, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778295625392, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5230712890625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778295625393, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778295625394, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778296283795, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778296283795, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778296313518, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.467529296875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778296313519, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778296313519, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778296973892, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778296973893, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778297003579, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4351806640625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778297003580, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778297003580, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778297661577, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778297661578, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778297691130, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.406982421875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778297691130, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778297691131, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778298348217, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778298348218, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778298377837, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3848876953125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778298377837, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778298377838, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778299035939, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778299035940, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778299065575, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3480224609375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778299065576, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778299065576, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778299724382, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778299724383, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778299754023, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3209228515625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778299754023, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778299754024, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778300412415, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778300412415, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778300442058, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.2950439453125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778300442059, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778300442060, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778300442060, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
@@ -0,0 +1,111 @@
+:::MLLOG {"namespace": "", "time_ms": 1778300456451, "event_type": "POINT_IN_TIME", "key": "submission_org", "value": "tinycorp", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1328}}
+:::MLLOG {"namespace": "", "time_ms": 1778300456455, "event_type": "POINT_IN_TIME", "key": "submission_platform", "value": "tinybox", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1329}}
+:::MLLOG {"namespace": "", "time_ms": 1778300456455, "event_type": "POINT_IN_TIME", "key": "submission_division", "value": "closed", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1330}}
+:::MLLOG {"namespace": "", "time_ms": 1778300456455, "event_type": "POINT_IN_TIME", "key": "submission_status", "value": "onprem", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1331}}
+:::MLLOG {"namespace": "", "time_ms": 1778300456455, "event_type": "POINT_IN_TIME", "key": "submission_benchmark", "value": "llama31_8b", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1333}}
+:::MLLOG {"namespace": "", "time_ms": 1778300457011, "event_type": "POINT_IN_TIME", "key": "cache_clear", "value": true, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1336}}
+:::MLLOG {"namespace": "", "time_ms": 1778300457012, "event_type": "INTERVAL_START", "key": "init_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1337}}
+:::MLLOG {"namespace": "", "time_ms": 1778300803665, "event_type": "INTERVAL_END", "key": "init_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1629}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817390, "event_type": "INTERVAL_START", "key": "run_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1340}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "seed", "value": 4729, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1341}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "global_batch_size", "value": 32, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1343}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "max_sequence_length", "value": 8192, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1344}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "max_steps", "value": 1200000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1345}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "gradient_accumulation_steps", "value": 2, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1346}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "eval_samples", "value": 1024, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1347}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "train_samples", "value": 38400000, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1348}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "opt_name", "value": "adamw", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1350}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "opt_base_learning_rate", "value": 0.001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1351}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817395, "event_type": "POINT_IN_TIME", "key": "opt_end_learning_rate", "value": 0.0001, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1352}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_1", "value": 0.9, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1353}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_adamw_beta_2", "value": 0.95, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1354}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_adamw_epsilon", "value": 1e-05, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1355}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_adamw_weight_decay", "value": 0.1, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1356}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1357}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "num_warmup_steps", "value": 128, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1358}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_steps", "value": 1199872, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1359}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_learning_rate_decay_schedule", "value": "cosine with linear warmup", "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1360}}
+:::MLLOG {"namespace": "", "time_ms": 1778300817396, "event_type": "POINT_IN_TIME", "key": "opt_gradient_clip_norm", "value": 1.0, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1361}}
+:::MLLOG {"namespace": "", "time_ms": 1778301145773, "event_type": "INTERVAL_START", "key": "epoch_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1529, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778301145774, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1530, "samples_count": 0}}
+:::MLLOG {"namespace": "", "time_ms": 1778301985088, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778301985089, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778302030319, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 5.865966796875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778302030319, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778302030320, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 12288}}
+:::MLLOG {"namespace": "", "time_ms": 1778302687526, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778302687527, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778302717259, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.615966796875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778302717260, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778302717261, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 24576}}
+:::MLLOG {"namespace": "", "time_ms": 1778303376036, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778303376037, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778303406044, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 4.154296875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778303406045, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778303406045, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 36864}}
+:::MLLOG {"namespace": "", "time_ms": 1778304071224, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778304071225, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778304101168, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.9095458984375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778304101169, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778304101170, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 49152}}
+:::MLLOG {"namespace": "", "time_ms": 1778304762172, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778304762173, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778304792161, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.775634765625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778304792162, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778304792162, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 61440}}
+:::MLLOG {"namespace": "", "time_ms": 1778305452836, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778305452836, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778305482708, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.676513671875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778305482708, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778305482709, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 73728}}
+:::MLLOG {"namespace": "", "time_ms": 1778306140246, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778306140246, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778306169947, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5947265625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778306169947, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778306169948, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 86016}}
+:::MLLOG {"namespace": "", "time_ms": 1778306828284, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778306828285, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778306858077, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.5255126953125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778306858077, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778306858078, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 98304}}
+:::MLLOG {"namespace": "", "time_ms": 1778307519609, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778307519610, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778307549531, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4757080078125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778307549532, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778307549532, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 110592}}
+:::MLLOG {"namespace": "", "time_ms": 1778308208151, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778308208152, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778308237856, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.4312744140625, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778308237857, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778308237857, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 122880}}
+:::MLLOG {"namespace": "", "time_ms": 1778308896397, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778308896398, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778308926271, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.402099609375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778308926271, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778308926272, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 135168}}
+:::MLLOG {"namespace": "", "time_ms": 1778309586346, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778309586347, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778309616134, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.37060546875, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778309616134, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778309616135, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 147456}}
+:::MLLOG {"namespace": "", "time_ms": 1778310273337, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778310273338, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778310303090, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3968505859375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778310303091, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778310303092, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 159744}}
+:::MLLOG {"namespace": "", "time_ms": 1778310958883, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778310958883, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778310988541, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.3284912109375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778310988542, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778310988542, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 172032}}
+:::MLLOG {"namespace": "", "time_ms": 1778311645004, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778311645004, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778311674742, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.302001953125, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778311674743, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778311674744, "event_type": "INTERVAL_START", "key": "block_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1654, "samples_count": 184320}}
+:::MLLOG {"namespace": "", "time_ms": 1778312331845, "event_type": "INTERVAL_END", "key": "block_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1616, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778312331846, "event_type": "INTERVAL_START", "key": "eval_start", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1617, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778312361570, "event_type": "POINT_IN_TIME", "key": "eval_accuracy", "value": 3.2777099609375, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1637, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778312361571, "event_type": "INTERVAL_END", "key": "eval_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1638, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778312361571, "event_type": "INTERVAL_END", "key": "epoch_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1646, "samples_count": 196608}}
+:::MLLOG {"namespace": "", "time_ms": 1778312361572, "event_type": "INTERVAL_END", "key": "run_stop", "value": null, "metadata": {"file": "tinygrad3/examples/mlperf/model_train.py", "lineno": 1647, "status": "success"}}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Woze Parrot	01848d1e17	feat: results	2026-05-11 17:56:29 +00:00
Woze Parrot	9938b5da8b	feat: training 6.0	2026-05-11 16:22:47 +00:00