remove AMD_LLVM=0 in mlperf and search ci

tinybox updated to llvm 20
feat: bump to llvm20 (#10784 )
2026-08-27 03:46:07 +00:00 · 2025-06-11 16:16:30 -07:00 · 2025-06-11 16:04:18 -07:00 · 2025-06-11 16:28:40 -04:00 · 2025-06-11 23:20:41 +03:00 · 2025-06-11 16:00:38 -04:00
441 changed files with 21632 additions and 1432744 deletions
@@ -11,5 +11,5 @@ runs:
        git fetch origin $CURRENT_SHA
        export COMMIT_MESSAGE=$(git show -s --format=%B "$CURRENT_SHA")
        export CURRENT_HEAD=$(git rev-parse HEAD)
-        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
+        cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && IGNORE_OOB=1 PYTHONPATH=. python3 process_replay.py
        git checkout $CURRENT_HEAD  # restore to branch
@@ -41,34 +41,19 @@ runs:
  using: "composite"
  steps:
    - name: Set up Python ${{ inputs.python-version }}
+      id: setup-python
      uses: actions/setup-python@v5
      with:
        python-version: ${{ inputs.python-version }}
-    - name: Upgrade pip
-      shell: bash
-      run: python -m pip install --upgrade pip

    # **** Caching packages ****
-    # TODO: key should include input.deps, but it can't since it can't contain commas

-    - name: Cache Python packages (Linux)
-      if: inputs.key != '' && runner.os == 'Linux'
+    - name: Cache Python packages
+      id: restore-venv
      uses: actions/cache@v4
      with:
-        path: ${{ env.Python3_ROOT_DIR }}/lib/python${{ inputs.python-version }}/site-packages
-        key: python-package-${{ inputs.key }}-${{ hashFiles('**/setup.py') }}
-    - name: Cache Python packages (macOS)
-      if: inputs.key != '' && runner.os == 'macOS'
-      uses: actions/cache@v4
-      with:
-        path: /Users/runner/Library/Python/${{ inputs.python-version }}/lib/python/site-packages
-        key: osx-python-package-${{ inputs.key }}-${{ hashFiles('**/setup.py') }}
-    - name: Cache Python packages (Windows)
-      if: inputs.key != '' && runner.os == 'Windows'
-      uses: actions/cache@v4
-      with:
-        path: ${{ env.Python3_ROOT_DIR }}\Lib\site-packages
-        key: windows-python-package-${{ inputs.key }}-${{ hashFiles('**/setup.py') }}
+        path: ${{ github.workspace }}/.venv
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ hashFiles('**/setup.py') }}-${{ env.PYTHON_CACHE_VERSION }}

    # **** Caching downloads ****

@@ -87,43 +72,110 @@ runs:

    # **** Python deps ****

-    - name: Install dependencies (with extra)
-      if: inputs.deps != ''
-      shell: bash
-      run: pip install ${{ (runner.os == 'macOS' && '--user') || (runner.os != 'macOS' && '') }} -e ".[${{ 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 (without extra)
-      if: inputs.deps == ''
-      shell: bash
-      run: pip install ${{ (runner.os == 'macOS' && '--user') || (runner.os != 'macOS' && '') }} -e . ${{ inputs.pydeps }}
-
-    # **** OpenCL ****
-
-    - name: Install OpenCL
-      if: inputs.opencl == 'true'
+    - name: Install dependencies in venv (with extra)
+      if: inputs.deps != '' && steps.restore-venv.outputs.cache-hit != 'true'
      shell: bash
      run: |
-        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
-        echo "deb [ allow-insecure=yes ] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list
-        sudo apt update || true
-        sudo apt install --allow-unauthenticated -y --no-install-recommends opencl-headers \
-          intel-oneapi-runtime-openmp=2023.2.1-16 intel-oneapi-runtime-compilers-common=2023.2.1-16 intel-oneapi-runtime-compilers=2023.2.1-16 \
-          intel-oneapi-runtime-dpcpp-sycl-opencl-cpu=2023.2.1-16 intel-oneapi-runtime-tbb-common=2021.10.0-49541 \
-          intel-oneapi-runtime-tbb=2021.10.0-49541 intel-oneapi-runtime-opencl=2023.2.1-16
+        python -m venv .venv
+        if [[ "$RUNNER_OS" == "Windows" ]]; then
+          source .venv/Scripts/activate
+        else
+          . .venv/bin/activate
+        fi
+        python -m pip install -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --extra-index-url https://download.pytorch.org/whl/cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
+    - name: Install dependencies in venv (without extra)
+      if: inputs.deps == '' && steps.restore-venv.outputs.cache-hit != 'true'
+      shell: bash
+      run: |
+        python -m venv .venv
+        if [[ "$RUNNER_OS" == "Windows" ]]; then
+          source .venv/Scripts/activate
+        else
+          . .venv/bin/activate
+        fi
+        python -m pip install -e . ${{ inputs.pydeps }}
+    - name: Set up venv environment
+      shell: bash
+      run: |
+        echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
+        echo "OMP_NUM_THREADS=1" >> "$GITHUB_ENV"
+        # no buffers should be over 300MB in CI
+        echo "MAX_BUFFER_SIZE=300000000" >> "$GITHUB_ENV"
+        if [[ "$RUNNER_OS" == "Windows" ]]; then
+          echo "${{ github.workspace }}/.venv/Scripts" >> "$GITHUB_PATH"
+        else
+          echo "${{ github.workspace }}/.venv/bin" >> "$GITHUB_PATH"
+        fi

-    # **** AMD ****
+    # ******************* apt *******************

-    - name: Install AMD (Linux)
+    - name: Add OpenCL Repo
+      if: inputs.opencl == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: echo "deb [ allow-insecure=yes ] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list
+
+    - name: Add AMD Repo (Linux)
      if: inputs.amd == 'true' && runner.os == 'Linux'
      shell: bash
      run: |
-        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
        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/6.1.2 jammy main
+        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 $(lsb_release -cs) main
        EOF
        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
-        sudo apt update || true
-        sudo apt install --no-install-recommends --allow-unauthenticated -y hsa-rocr comgr hsa-rocr-dev liburing-dev libc6-dev
+
+    - name: Add LLVM Repo (Linux)
+      if: inputs.llvm == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: |
+        wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
+        echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list
+
+    - name: apt-get update + install
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
+      shell: bash
+      run: |
+        echo 'Acquire::GzipIndexes "true";' | sudo tee /etc/apt/apt.conf.d/gzip
+        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
+        sudo apt -qq update || true
+
+        pkgs=""
+        # **** OpenCL ****
+        if [[ "${{ inputs.opencl }}" == "true" ]]; then
+          pkgs+=" opencl-headers \
+            intel-oneapi-runtime-openmp=2023.2.1-16 intel-oneapi-runtime-compilers-common=2023.2.1-16 intel-oneapi-runtime-compilers=2023.2.1-16 \
+            intel-oneapi-runtime-dpcpp-sycl-opencl-cpu=2023.2.1-16 intel-oneapi-runtime-tbb-common=2021.10.0-49541 \
+            intel-oneapi-runtime-tbb=2021.10.0-49541 intel-oneapi-runtime-opencl=2023.2.1-16"
+        fi
+        # **** AMD ****
+        if [[ "${{ inputs.amd }}" == "true" ]]; then
+          pkgs+=" hsa-rocr comgr hsa-rocr-dev liburing-dev libc6-dev"
+        fi
+        # **** CUDA ****
+        if [[ "${{ inputs.cuda }}" == "true" ]]; then
+          pkgs+=" git g++ cmake ninja-build llvm-15-dev zlib1g-dev libglew-dev \
+            flex bison libfl-dev libboost-thread-dev libboost-filesystem-dev nvidia-cuda-toolkit-gcc libzstd-dev"
+        fi
+        # **** WebGPU (dependencies for software-based vulkan) ****
+        if [[ "${{ inputs.webgpu }}" == "true" ]]; then
+          pkgs+=" libgl1 libglx-mesa0 libgl1-mesa-dri libxcb-xfixes0-dev mesa-vulkan-drivers"
+        fi
+        # **** LLVM ****
+        if [[ "${{ inputs.llvm }}" == "true" ]]; then
+          pkgs+=" libllvm20 clang-20 lld-20"
+        fi
+
+        # ******** do install ********
+        if [[ -n "$pkgs" ]]; then
+          sudo apt-get -y --allow-unauthenticated --no-install-recommends install $pkgs
+        fi
+
+    # **** AMD ****
+
+    - name: Setup AMD (Linux)
+      if: inputs.amd == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: |
        cargo build --release --manifest-path ./extra/remu/Cargo.toml
        sudo ln -sf ${{ github.workspace }}/extra/remu/target/release/libremu.so /usr/local/lib/libremu.so
        sudo tee --append /etc/ld.so.conf.d/rocm.conf <<'EOF'
@@ -131,7 +183,7 @@ runs:
          /opt/rocm/lib64
        EOF
        sudo ldconfig
-    - name: Install AMD comgr+remu (macOS)
+    - name: Setup AMD comgr+remu (macOS)
      if: inputs.amd == 'true' && runner.os == 'macOS'
      shell: bash
      run: |
@@ -143,20 +195,10 @@ runs:

    # **** CUDA ****

-    - name: Install cuda packages (Linux)
-      if: inputs.cuda == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: |
-        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
-        sudo apt update -y || true
-        sudo apt install -y --no-install-recommends 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
    - name: Install gpuocelot dependencies (MacOS)
      if: inputs.cuda == 'true' && runner.os == 'macOS'
      shell: bash
-      run: |
-        brew update
-        brew install cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses
+      run: brew install --quiet cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses
    - name: Cache gpuocelot
      if: inputs.cuda == 'true'
      id: cache-build
@@ -191,13 +233,7 @@ runs:
      shell: bash
      run: |
        sudo curl -L https://github.com/wpmed92/pydawn/releases/download/v0.1.6/libwebgpu_dawn.so -o /usr/local/lib/libwebgpu_dawn.so
-    - name: Install dependencies for software-based vulkan
-      if: inputs.webgpu == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: |
-        sudo apt update -y || true
-        sudo apt install -y libegl1-mesa libgl1-mesa-dri libxcb-xfixes0-dev mesa-vulkan-drivers
-
+        sudo ldconfig
    - name: Install WebGPU dawn (macOS)
      if: inputs.webgpu == 'true' && runner.os == 'macOS'
      shell: bash
@@ -207,18 +243,7 @@ runs:

    # **** LLVM ****

-    - name: Install LLVM (Linux)
-      if: inputs.llvm == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: |
-        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
-        wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
-        echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-19 main" | sudo tee /etc/apt/sources.list.d/llvm.list
-        sudo apt update -y || true
-        sudo apt install -y --no-install-recommends libllvm19 clang-19 lld-19
-
    - name: Install LLVM (macOS)
      if: inputs.llvm == 'true' && runner.os == 'macOS'
      shell: bash
-      run: |
-        brew install llvm@19
+      run: brew install llvm@20
@@ -31,7 +31,7 @@ jobs:
    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -o pipefail {0}
+        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -52,16 +52,18 @@ jobs:
    - name: reset process replay
      run: python3.11 test/external/process_replay/reset.py
    - name: Run Stable Diffusion
-      run: JIT=1 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run Stable Diffusion without fp16
-      run: JIT=1 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
    - name: Run Stable Diffusion v2
-      run: JIT=1 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
    # process replay can't capture this, the graph is too large
    - name: Run SDXL
-      run: CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run model inference benchmark
      run: METAL=1 python3.11 test/external/external_model_benchmark.py
+    - name: Run huggingface_onnx test
+      run: METAL=1 python3.11 extra/huggingface_onnx/run_models.py test --debug FacebookAI/xlm-roberta-large
    - name: Test speed vs torch
      run: BIG=2 MPS=1 python3.11 test/test_speed_v_torch.py | tee torch_speed.txt
    - name: Test tensor cores
@@ -80,40 +82,40 @@ jobs:
      run: METAL=1 M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
    - name: Run LLaMA
      run: |
-        JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA with BEAM
-      run: JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    - name: Run quantized LLaMA
      run: |
-        python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
-        python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_nf4.txt
+        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8 | tee llama_int8.txt
+        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4 | tee llama_nf4.txt
    - name: Run quantized LLaMA3
      run: |
-        python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8 | tee llama3_int8.txt
-        python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4 | tee llama3_nf4.txt
+        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8 | tee llama3_int8.txt
+        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4 | tee llama3_nf4.txt
    #- name: Run LLaMA 7B on 4 (virtual) GPUs
    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    - name: Run GPT2
      run: |
-        JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        JIT=1 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 JIT=1 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - name: Run OLMoE
-      run: python3.11 examples/olmoe.py
+      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
    - name: Train MNIST
      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: JIT=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps JIT=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: JIT=2 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half JIT=2 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
    #- name: Run 10 CIFAR training steps w BF16
    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    - name: Run 10 CIFAR training steps w winograd
-      run: JIT=1 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: UsbGPU boot time
      run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU tiny tests
@@ -161,7 +163,7 @@ jobs:
    timeout-minutes: 30
    defaults:
      run:
-        shell: bash -o pipefail {0}
+        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -210,37 +212,37 @@ jobs:
    - name: Test CUDA=1
      run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
    - name: Run Stable Diffusion
-      run: NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run SDXL
-      run: CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA
      run: |
-        NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA with BEAM
-      run: NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    # - name: Run LLaMA 7B on 4 GPUs
    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    # - name: Run LLaMA 7B on 6 GPUs
    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
    - name: Run LLaMA-3 8B BEAM
-      run: NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
+      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
    # - name: Run LLaMA-3 8B on 6 GPUs
    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
    # - name: Run LLaMA-2 70B
    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
    - name: Run Mixtral 8x7B
-      run: time NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
+      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
    - name: Run GPT2
      run: |
-        NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        NV=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: NV=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA)
@@ -275,7 +277,7 @@ jobs:
    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -o pipefail {0}
+        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -304,26 +306,26 @@ jobs:
    - name: Train MNIST
      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    - name: Run 10 CIFAR training steps w BF16
-      run: NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    - name: Run 10 CIFAR training steps w winograd
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_half_wino NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    - name: Run MLPerf resnet eval on training data
-      run: time NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
+      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA Training)
@@ -347,13 +349,17 @@ jobs:
    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -o pipefail {0}
+        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
-    - name: Insert amdgpu
-      run: sudo modprobe amdgpu
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
+    #- name: Insert amdgpu
+    #  run: sudo modprobe amdgpu
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -372,12 +378,12 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: setup perflevel
-      run: |
-        examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
-        rocm-smi
-    - name: Show off tinybox
-      run: /opt/rocm/bin/rocm-bandwidth-test
+    #- name: setup perflevel
+    #  run: |
+    #    examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
+    #    rocm-smi
+    #- name: Show off tinybox
+    #  run: /opt/rocm/bin/rocm-bandwidth-test
    # TODO: unstable on AMD
    #- name: Run model inference benchmark
    #  run: LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 NOCLANG=1 python3 test/external/external_model_benchmark.py
@@ -397,51 +403,51 @@ jobs:
      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
    - name: Test AMD=1
      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
-    - name: Test HIP=1
-      run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
+    #- name: Test HIP=1
+    #  run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
    # TODO: AMD compiler bug causes this to fail
    #- name: Fuzz Padded Tensor Core GEMM
    #  run: HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
-    - name: Remove amdgpu
-      run: sleep 10 && sudo rmmod amdgpu # sleep a bit to let the driver unload the prev pid.
+    #- name: Remove amdgpu
+    #  run: sleep 10 && sudo rmmod amdgpu # sleep a bit to let the driver unload the prev pid.
    - name: Test AM cold start time
      run: time AMD=1 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Test AM warm start time
      run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Run Stable Diffusion
-      run: AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run SDXL
-      run: CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA 7B
      run: |
-        AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
-        AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
+        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
+        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_jitted.txt
    - name: Run LLaMA 7B with BEAM
-      run: AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
+      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_beam.txt
    # - name: Run LLaMA 7B on 4 GPUs
    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_four_gpu.txt
    # - name: Run LLaMA 7B on 6 GPUs
    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_six_gpu.txt
    - name: Run LLaMA-3 8B BEAM
-      run: AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
+      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
+      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_four_gpu.txt
    # - name: Run LLaMA-3 8B on 6 GPUs
    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
-    - name: Restore amdgpu
-      run: sudo modprobe amdgpu
+    #- name: Restore amdgpu
+    #  run: sudo modprobe amdgpu
    # - name: Run LLaMA-2 70B
    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing | tee llama_2_70B.txt
    - name: Run Mixtral 8x7B
-      run: time AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
+      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing | tee mixtral.txt
    - name: Run GPT2
      run: |
-        AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        AMD=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
+        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: AMD=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
-      run: AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
+      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD)
@@ -473,13 +479,15 @@ jobs:
    timeout-minutes: 30
    defaults:
      run:
-        shell: bash -o pipefail {0}
+        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Remove amdgpu
      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -500,26 +508,19 @@ jobs:
    - name: Train MNIST
      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    - name: Run 10 CIFAR training steps w BF16
-      run: AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
    - name: Run 10 CIFAR training steps w winograd
-      run: AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_half_wino AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
-    - name: Run MLPerf resnet eval
-      run: time AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
-    - name: Run 10 MLPerf Bert training steps (6 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+    - name: Run full CIFAR training steps w 6 GPUS (REMOTE)
+      run: time BENCHMARK_LOG=cifar_6gpu_remote REMOTE=1 REMOTEDEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu_remote.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD Training)
@@ -531,6 +532,55 @@ jobs:
          train_cifar_wino.txt
          train_cifar_one_gpu.txt
          train_cifar_six_gpu.txt
+          train_cifar_six_gpu_remote.txt
+    - name: Run process replay tests
+      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
+
+  testmlperfamdbenchmark:
+    name: tinybox red MLPerf Benchmark
+    runs-on: [self-hosted, Linux, tinybox]
+    timeout-minutes: 30
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
+        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
+        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
+        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: Run MLPerf resnet eval
+      run: time BENCHMARK_LOG=resnet_eval AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+    - name: Run 10 MLPerf Bert training steps (6 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (AMD MLPerf)
+        path: |
          train_resnet.txt
          train_resnet_one_gpu.txt
          train_bert.txt
@@ -543,7 +593,7 @@ jobs:
    timeout-minutes: 20
    defaults:
      run:
-        shell: bash -o pipefail {0}
+        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -558,19 +608,27 @@ jobs:
    - name: validate openpilot 0.9.7
      run: PYTHONPATH=. FLOAT16=0 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
    - name: benchmark openpilot 0.9.4
-      run: PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_4.txt
+      run: BENCHMARK_LOG=openpilot_0_9_4 PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_4.txt
    - name: benchmark openpilot 0.9.7
-      run: PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_7.txt
+      run: BENCHMARK_LOG=openpilot_0_9_7 PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_7.txt
    - name: benchmark openpilot w IMAGE=2 0.9.4
-      run: PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_4.txt
+      run: BENCHMARK_LOG=openpilot_0_9_4_image PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_4.txt
    - name: benchmark openpilot w IMAGE=2 0.9.7
-      run: PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
+      run: BENCHMARK_LOG=openpilot_0_9_7_image PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
    - name: openpilot compile3 0.9.7
      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx
    - name: openpilot compile3 0.9.7+ tomb raider
      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/e8bea2c78ffa92685ece511e9b554122aaf1a79d/selfdrive/modeld/models/supercombo.onnx
    - name: openpilot dmonitoring compile3 0.9.7
      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: benchmark MobileNetV2 on DSP
+      run: |
+        # generate quantized weights
+        ln -s /data/home/tiny/tinygrad/extra/datasets/imagenet extra/datasets/imagenet
+        ln -s /data/home/tiny/tinygrad/testsig-*.so .
+        PYTHONPATH=. CC=clang-19 CPU=1 QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
+        # benchmark on DSP with NOOPT=1, the devectorizer has issues
+        PYTHONPATH=. CC=clang-19 DSP=1 DONT_REALIZE_EXPAND=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
    - uses: actions/upload-artifact@v4
@@ -0,0 +1,33 @@
+name: Benchmark with kernel search
+
+on:
+  push:
+    branches:
+      - update_benchmark_search
+  workflow_dispatch:
+
+jobs:
+  run_script_job:
+    runs-on: [self-hosted, Linux, tinybox]
+    if: github.repository_owner == 'tinygrad'
+
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Remove amdgpu
+      run: sudo rmmod amdgpu || true
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Run SDXL with new search
+      # TODO: GCVM_L2_PROTECTION_FAULT_STATUS with llvm19
+      run: |
+        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 python examples/sdxl.py --noshow --timing --seed 0
+    - name: Run SDXL with cached search
+      run: |
+        BENCHMARK_LOG=search_sdxl_cached PYTHONPATH=. AMD=1 JITBEAM=2 python examples/sdxl.py --noshow --timing --seed 0
+    - name: Run winograd cifar with new search
+      run: |
+        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF FUSE_ARANGE=1 JITBEAM=4 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 BS=1024 STEPS=500 python examples/hlb_cifar10.py
+    - name: Run winograd cifar with cached search
+      run: |
+        BENCHMARK_LOG=search_wino_cifar_cached WINO=1 DEFAULT_FLOAT=HALF FUSE_ARANGE=1 JITBEAM=4 BS=1024 STEPS=500 python examples/hlb_cifar10.py
@@ -0,0 +1,29 @@
+name: Run MLPerf Training
+
+on:
+  schedule:
+    - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
+  push:
+    branches:
+      - update_mlperf
+  workflow_dispatch:
+
+jobs:
+  run_script_job:
+    runs-on: [self-hosted, Linux, tinybox]
+    if: github.repository_owner == 'tinygrad'
+
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Cleanup running AM processes
+      run: python extra/amdpci/am_smi.py --pids --kill
+    - name: Symlink datasets
+      run: |
+        mkdir -p extra/datasets
+        ln -s /raid/datasets/imagenet extra/datasets/imagenet
+    - name: Run resnet
+      run: |
+        rm "~/.cache/tinygrad/cache_mlperf.db" || true
+        BENCHMARK_LOG=mlpert_train_resnet LOGMLPERF=0 CACHEDB="~/.cache/tinygrad/cache_mlperf.db" examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
+        rm "~/.cache/tinygrad/cache_mlperf.db"
@@ -1,7 +1,8 @@
 name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
-  DOWNLOAD_CACHE_VERSION: '9'
+  DOWNLOAD_CACHE_VERSION: '10'
+  PYTHON_CACHE_VERSION: '2'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

@@ -53,11 +54,6 @@ jobs:
        source venv/bin/activate
        pip install $GITHUB_WORKSPACE
        python -c "from tinygrad.tensor import Tensor; print(Tensor([1,2,3,4,5]))"
-        # Test using VIZ=1 after package installation
-        VIZ=1 python -c "from tinygrad.tensor import Tensor; Tensor([1,2,3,4,5]).realize()" & VIZ_PID=$!
-        echo "started VIZ server at: $(ps -p "$VIZ_PID" -o pid=)"
-        i=0; while ((i++ < 10)); do curl -sSf localhost:8000 > /dev/null && break || { echo "VIZ verification attempt $i/10"; sleep 1; }; done; ((i > 10)) && echo "Could not verify VIZ server" && exit 1
-        kill $VIZ_PID
        pip install mypy
        mypy -c "from tinygrad.tensor import Tensor; print(Tensor([1,2,3,4,5]))"
    - name: Run beautiful_mnist with tinygrad only
@@ -87,7 +83,7 @@ jobs:

  autogen:
    name: Autogen
-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -100,6 +96,8 @@ jobs:
        cuda: 'true'
        webgpu: 'true'
        llvm: 'true'
+    - name: Install autogen support packages
+      run: sudo apt-get install -y --no-install-recommends llvm-14-dev libclang-14-dev
    - name: Verify OpenCL autogen
      run: |
        cp tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
@@ -356,20 +354,25 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-12
+        pydeps: "pillow"
        deps: testing_unit
    - name: Test README
      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  PYTHONPATH=. python README.py
    - name: Run unit tests
-      run: PYTHONPATH="." python -m pytest -n=auto test/unit/
+      run: PYTHONPATH="." python -m pytest -n=auto test/unit/ --durations=20
    - name: Run targetted tests on NULL backend
      run: PYTHONPATH="." NULL=1 python3 test/test_multitensor.py TestMultiTensor.test_data_parallel_resnet_train_step
+    - name: Run SDXL on NULL backend
+      run: MAX_BUFFER_SIZE=0 PYTHONPATH="." NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
    # TODO: support fake weights
    #- name: Run LLaMA 7B on 4 fake devices
    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
    - name: Run GC tests
      run: PYTHONPATH="." python test/external/external_uop_gc.py
-    - name: Repo line count < 13000 lines
-      run: MAX_LINE_COUNT=13000 python sz.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+    - name: Repo line count < 14000 lines
+      run: MAX_LINE_COUNT=14000 python sz.py

  fuzzing:
    name: Fuzzing
@@ -381,8 +384,8 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: fuzzing-minimal
-        deps: testing_minimal
+        key: fuzzing-unit
+        deps: testing_unit
    - name: Fuzz Test symbolic
      run: python test/external/fuzz_symbolic.py
    - name: Fuzz Test fast idiv
@@ -391,6 +394,8 @@ jobs:
      run: |
        PYTHONPATH="." python test/external/fuzz_shapetracker.py
        PYTHONPATH="." python test/external/fuzz_shapetracker_math.py
+    - name: Fuzz Test shape ops
+      run: python test/external/fuzz_shape_ops.py

  testgpuimage:
    name: 'GPU IMAGE Tests'
@@ -415,6 +420,8 @@ jobs:
        run: |
          PYTHONPATH="." GPU=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
          PYTHONPATH="." GPU=1 IMAGE=2 python3 test/models/test_end2end.py TestEnd2End.test_linear_mnist
+      - name: Run fused optimizer tests
+        run: PYTHONPATH="." GPU=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -436,7 +443,7 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          PYTHONPATH="." ALLOWED_KERNEL_COUNT=209 ALLOWED_READ_IMAGE=2138 ALLOWED_GATED_READ_IMAGE=29 FLOAT16=0 GPU=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx
+          PYTHONPATH="." ALLOWED_KERNEL_COUNT=209 ALLOWED_READ_IMAGE=2137 ALLOWED_GATED_READ_IMAGE=29 FLOAT16=0 GPU=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx
      - name: Test openpilot alt model correctness (float32)
        run: PYTHONPATH="." FLOAT16=0 DEBUGCL=1 GPU=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/3799fe46b3a629e491d4b8498b8ae83e4c88c304/selfdrive/modeld/models/supercombo.onnx
      - name: Test openpilot fastvits model correctness (float32)
@@ -446,8 +453,39 @@ jobs:
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

+  testonnxcpu:
+    name: 'ONNX (CPU) Tests'
+    runs-on: ubuntu-22.04
+    timeout-minutes: 20
+    env:
+      IGNORE_OOB: 0
+
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: onnxoptc
+          deps: testing
+          python-version: '3.11'
+          llvm: 'true'
+      - name: Test ONNX (CPU)
+        run: CPU=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+      - name: Test ONNX (LLVM)
+        run: LLVM=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+      - name: Test Additional ONNX Ops (CPU)
+        run: CPU=1 PYTHONPATH=. python3 test/external/external_test_onnx_ops.py
+      - name: Test Quantize ONNX
+        run: CPU=1 PYTHONPATH=. python3 test/test_quantize_onnx.py
+      - name: Run REMOTE=1 Test
+        run: |
+          REMOTEDEV=CPU REMOTE=1 python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py
+      - name: Run process replay tests
+        uses: ./.github/actions/process-replay
+
  testopencl:
-    name: 'ONNX+Optimization Tests'
+    name: 'ONNX (GPU)+Optimization Tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    env:
@@ -460,23 +498,14 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: onnxoptl
-          deps: testing,testing_tf
+          deps: testing
+          pydeps: "tensorflow==2.15.1 tensorflow_addons"
          python-version: '3.11'
          opencl: 'true'
-          llvm: 'true'
      - name: Test ONNX (GPU)
        run: GPU=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
-      - name: Test ONNX (CPU)
-        run: CPU=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
-      - name: Test ONNX (LLVM)
-        run: LLVM=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
-      - name: Test Additional ONNX Ops (CPU)
-        run: CPU=1 PYTHONPATH=. python3 test/external/external_test_onnx_ops.py
-      - name: Test Quantize ONNX
-        run: CPU=1 PYTHONPATH=. python3 test/test_quantize_onnx.py
      - name: Run REMOTE=1 Test
        run: |
-          REMOTEDEV=CPU REMOTE=1 python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py
          REMOTEDEV=GPU REMOTE=1 python3 -m pytest test/test_tiny.py test/test_image_dtype.py test/test_jit.py
          REMOTEDEV=GPU IMAGE=2 REMOTE=1 python3 -m pytest test/test_tiny.py test/test_image_dtype.py
      - name: Test Optimization Helpers
@@ -495,7 +524,7 @@ jobs:
  testmodels:
    name: Models (llvm+cpu+gpu)
    runs-on: ubuntu-22.04
-    timeout-minutes: 10
+    timeout-minutes: 15
    env:
      IGNORE_OOB: 0
    steps:
@@ -544,10 +573,12 @@ jobs:
        tags: qemu-hexagon:latest
        cache-from: type=gha
        cache-to: type=gha,mode=min
+    - name: Set MOCKDSP env
+      run: printf "MOCKDSP=1" >> $GITHUB_ENV
    - name: Run test_tiny on DSP
      run: DEBUG=2 DSP=1 python test/test_tiny.py
    - name: Test transcendentals
-      run: CC=clang-19 PYTHONPATH="." DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
+      run: CC=clang-20 PYTHONPATH="." DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
    - name: Test quantize onnx
      run: PYTHONPATH="." DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py
    - name: Test LLVM=1 DEVECTORIZE=0
@@ -598,6 +629,7 @@ jobs:
      MOCKGPU: 1
      FORWARD_ONLY: 1
      AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
+      PYTHONPATH: ${{ github.workspace }}
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -610,15 +642,19 @@ jobs:
          llvm: ${{ matrix.backend == 'amdllvm' && 'true' }}
      - name: Check Device.DEFAULT and print some source
        run: |
-          PYTHONPATH=${{ github.workspace }} python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
-          DEBUG=5 PYTHONPATH=${{ github.workspace }} FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
+          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
+          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run LLVM test
        if: matrix.backend=='amdllvm'
-        run: PYTHONPATH="." python test/test_amd_llvm.py
+        run: python test/test_amd_llvm.py
      - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/imported/test_indexing.py test/test_hcq.py test/external/external_test_am.py --durations=20
+        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/test_hcq.py test/external/external_test_am.py --durations=20
      - name: Run TRANSCENDENTAL math
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
+      - name: Run TestOps.test_add with SQTT
+        run: |
+          PROFILE=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -648,7 +684,8 @@ jobs:
          PYTHONPATH=${{ github.workspace }} python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CUDA','NV'], Device.DEFAULT"
          DEBUG=5 PYTHONPATH=${{ github.workspace }} FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (cuda)
-        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore test/test_gc.py --durations=20
+        # skip multitensor because it's slow
+        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -711,7 +748,7 @@ jobs:
    - name: Run real world test
      run: METAL=1 python -m pytest -n=auto test/models/test_real_world.py --durations=20
    - name: Test models (Metal)
-      run: METAL=1 python -m pytest test/models -v --durations=20
+      run: METAL=1 python -m pytest -n=auto test/models -v --durations=20
    - name: Run ONNX
      run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
    - name: Test tensor core ops (fake)
@@ -775,9 +812,18 @@ jobs:
      run: npm install puppeteer
    - name: Run WEBGPU Efficientnet
      run: node test/web/test_webgpu.js
+    - name: Run VIZ tests as external package
+      run: |
+        mkdir $GITHUB_WORKSPACE/test_dir
+        cd $GITHUB_WORKSPACE/test_dir
+        python -m venv venv
+        source venv/bin/activate
+        pip install $GITHUB_WORKSPACE
+        cp $GITHUB_WORKSPACE/test/web/test_viz.js .
+        node test_viz.js

  osxremote:
-    name: MacOS (remote)
+    name: MacOS (remote metal)
    runs-on: macos-15
    timeout-minutes: 10
    env:
@@ -798,7 +844,35 @@ jobs:
          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
      - name: Run REMOTE=1 Test
        run: |
-          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_tensor_variable.py
+
+  amdremote:
+    name: Linux (remote amd)
+    runs-on: ubuntu-22.04
+    timeout-minutes: 20
+    env:
+      REMOTE: 1
+      REMOTEDEV: 'AMD'
+      PYTHONPATH: ${{ github.workspace }}
+      MOCKGPU: 1
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: linux-remote-amd
+          deps: testing_minimal
+          amd: 'true'
+          llvm: 'true'
+      - name: Check Device.DEFAULT and print some source
+        run: |
+          python -c "from tinygrad import Device; assert Device.DEFAULT == 'REMOTE', Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.default.properties.real_device == 'AMD', Device.default.properties.real_device"
+          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
+      - name: Run REMOTE=1 Test
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_tensor_variable.py

  osxtests:
    strategy:
@@ -28,7 +28,7 @@ repos:
        pass_filenames: false
      - id: tests
        name: subset of tests
-        entry: env PYTHONPATH="." python3 -m pytest -n=4 test/unit/ test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py test/test_symbolic_shapetracker.py
+        entry: env PYTHONPATH="." python3 -m pytest -n=4 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
        language: system
        always_run: true
        pass_filenames: false
@@ -7,7 +7,7 @@ extension-pkg-whitelist=scipy,cereal.messaging.messaging_pyx,PyQt5,av

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

 # Add files or directories matching the regex patterns to the blacklist. The
 # regex matches against base names, not paths.
@@ -0,0 +1,17 @@
+# tinygrad agents
+
+Hello agent. You are one of the most talented programmers of your generation.
+
+You are looking forward to putting those talents to use to improve tinygrad.
+
+## philosophy
+
+tinygrad is a **tensor** library focused on beauty and minimalism, while still matching the functionality of PyTorch and JAX.
+
+Every line must earn its keep. Prefer readability over cleverness. We believe that if carefully designed, 10 lines can have the impact of 1000.
+
+Never mix functionality changes with whitespace changes. All functionality changes must be tested.
+
+## style
+
+Use **2-space indentation**, and keep lines to a maximum of **150 characters**. Match the existing style.
@@ -21,11 +21,11 @@ tinygrad: For something between [PyTorch](https://github.com/pytorch/pytorch) an

 ---

-This may not be the best deep learning framework, but it is a deep learning framework.
+Despite tinygrad's size, it is a fully featured deep learning framework.

-Due to its extreme simplicity, it aims to be the easiest framework to add new accelerators to, with support for both inference and training. If XLA is CISC, tinygrad is RISC.
+Due to its extreme simplicity, it is the easiest framework to add new accelerators to, with support for both inference and training. If XLA is CISC, tinygrad is RISC.

-tinygrad is still alpha software, but we [raised some money](https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html) to make it good. Someday, we will tape out chips.
+tinygrad is now beta software, we [raised some money](https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html) to make it good. Someday, we will tape out chips.

 ## Features

@@ -9,7 +9,7 @@ if [[ ! $(clang2py -V) ]]; then
  pip install clang==14.0.6
  git clone https://github.com/nimlgen/ctypeslib.git
  cd ctypeslib
-  pip install --user .
+  pip install .
  clang2py -V
  popd
 fi
@@ -83,11 +83,12 @@ generate_kfd() {
  sed -i "/import functools/a from tinygrad.runtime.support.hcq import FileIOInterface" $BASE/kfd.py
  sed -i "s/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd, \*\*kwargs):/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd:FileIOInterface, \*\*kwargs):/g" $BASE/kfd.py
  sed -i "s/fcntl.ioctl(__fd, (__idir<<30)/__fd.ioctl((__idir<<30)/g" $BASE/kfd.py
+  sed -i "s/!!/not not /g" $BASE/kfd.py
  python3 -c "import tinygrad.runtime.autogen.kfd"
 }

 generate_cuda() {
-  clang2py /usr/include/cuda.h -o $BASE/cuda.py -l /usr/lib/x86_64-linux-gnu/libcuda.so
+  clang2py /usr/include/cuda.h --clang-args="-D__CUDA_API_VERSION_INTERNAL" -o $BASE/cuda.py -l /usr/lib/x86_64-linux-gnu/libcuda.so
  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/cuda.py
  sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libcuda.so')\ctypes.CDLL(ctypes.util.find_library('cuda'))\g" $BASE/cuda.py
  fixup $BASE/cuda.py
@@ -104,10 +105,10 @@ generate_nvrtc() {
 }

 generate_nv() {
-  NVKERN_COMMIT_HASH=d6b75a34094b0f56c2ccadf14e5d0bd515ed1ab6
+  NVKERN_COMMIT_HASH=81fe4fb417c8ac3b9bdcc1d56827d116743892a5
  NVKERN_SRC=/tmp/open-gpu-kernel-modules-$NVKERN_COMMIT_HASH
  if [ ! -d "$NVKERN_SRC" ]; then
-    git clone https://github.com/tinygrad/open-gpu-kernel-modules $NVKERN_SRC
+    git clone https://github.com/NVIDIA/open-gpu-kernel-modules $NVKERN_SRC
    pushd .
    cd $NVKERN_SRC
    git reset --hard $NVKERN_COMMIT_HASH
@@ -116,15 +117,19 @@ generate_nv() {

  clang2py -k cdefstum \
    extra/nv_gpu_driver/clc6c0qmd.h \
+    extra/nv_gpu_driver/clcec0qmd.h \
    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0080.h \
    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080_notification.h \
    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \
-    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \
-    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc86f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc96f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc761.h \
    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl83de.h \
    $NVKERN_SRC/src/nvidia/generated/g_allclasses.h \
    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc6c0.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clcdc0.h \
    $NVKERN_SRC/kernel-open/nvidia-uvm/clc6b5.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/clc9b5.h \
    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_ioctl.h \
    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_linux_ioctl.h \
    $NVKERN_SRC/kernel-open/nvidia-uvm/hwref/ampere/ga100/dev_fault.h \
@@ -149,6 +154,8 @@ generate_nv() {
  sed -i "s\import ctypes\import ctypes, os\g" $BASE/nv_gpu.py
  sed -i 's/#\?\s\([A-Za-z0-9_]\+\) = MW ( \([0-9]\+\) : \([0-9]\+\) )/\1 = (\2 , \3)/' $BASE/nv_gpu.py # NVC6C0_QMDV03_00 processing
  sed -i 's/#\sdef NVC6C0_QMD\([A-Za-z0-9_()]\+\):/def NVC6C0_QMD\1:/' $BASE/nv_gpu.py
+  sed -i 's/#\sdef NVCEC0_QMD\([A-Za-z0-9_()]\+\):/def NVCEC0_QMD\1:/' $BASE/nv_gpu.py
+  sed -E -i -n '/^def (NVCEC0_QMDV05_00_RELEASE)(_ENABLE)\(i\):/{p;s//\1'"0"'\2=\1\2(0)\n\1'"1"'\2=\1\2(1)/;H;b};p;${x;s/^\n//;p}' "$BASE/nv_gpu.py"
  sed -i 's/#\s*return MW(\([0-9i()*+]\+\):\([0-9i()*+]\+\))/    return (\1 , \2)/' $BASE/nv_gpu.py
  sed -i 's/#\?\s*\(.*\)\s*=\s*\(NV\)\?BIT\(32\)\?\s*(\s*\([0-9]\+\)\s*)/\1 = (1 << \4)/' $BASE/nv_gpu.py # name = BIT(x) -> name = (1 << x)
  sed -i "s/UVM_\([A-Za-z0-9_]\+\) = \['i', '(', '\([0-9]\+\)', ')'\]/UVM_\1 = \2/" $BASE/nv_gpu.py # UVM_name = ['i', '(', '<num>', ')'] -> UVM_name = <num>
@@ -220,7 +227,7 @@ generate_libc() {

  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libc.py
  sed -i "s\FIXME_STUB\libc\g" $BASE/libc.py
-  sed -i "s\FunctionFactoryStub()\None if (libc_path := ctypes.util.find_library('c')) is None else ctypes.CDLL(libc_path)\g" $BASE/libc.py
+  sed -i "s\FunctionFactoryStub()\None if (libc_path := ctypes.util.find_library('c')) is None else ctypes.CDLL(libc_path, use_errno=True)\g" $BASE/libc.py

  fixup $BASE/libc.py
 }
@@ -349,50 +356,6 @@ generate_am() {
    -o $BASE/am/soc24.py
  fixup $BASE/am/soc24.py

-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/mp/mp_13_0_0_offset.h \
-    $AMKERN_INC/asic_reg/mp/mp_13_0_0_sh_mask.h \
-    -o $BASE/am/mp_13_0_0.py
-  fixup $BASE/am/mp_13_0_0.py
-
-  # 14_0_3 reuses 14_0_2
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/mp/mp_14_0_2_offset.h \
-    $AMKERN_INC/asic_reg/mp/mp_14_0_2_sh_mask.h \
-    -o $BASE/am/mp_14_0_3.py
-  fixup $BASE/am/mp_14_0_3.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/mp/mp_11_0_offset.h \
-    $AMKERN_INC/asic_reg/mp/mp_11_0_sh_mask.h \
-    -o $BASE/am/mp_11_0.py
-  fixup $BASE/am/mp_11_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/gc/gc_9_4_3_offset.h \
-    $AMKERN_INC/asic_reg/gc/gc_9_4_3_sh_mask.h \
-    extra/amdpci/overlay/gc_9_4_3.h \
-    -o $BASE/am/gc_9_4_3.py
-  fixup $BASE/am/gc_9_4_3.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/gc/gc_10_3_0_offset.h \
-    $AMKERN_INC/asic_reg/gc/gc_10_3_0_sh_mask.h \
-    -o $BASE/am/gc_10_3_0.py
-  fixup $BASE/am/gc_10_3_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/gc/gc_11_0_0_offset.h \
-    $AMKERN_INC/asic_reg/gc/gc_11_0_0_sh_mask.h \
-    -o $BASE/am/gc_11_0_0.py
-  fixup $BASE/am/gc_11_0_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/gc/gc_12_0_0_offset.h \
-    $AMKERN_INC/asic_reg/gc/gc_12_0_0_sh_mask.h \
-    -o $BASE/am/gc_12_0_0.py
-  fixup $BASE/am/gc_12_0_0.py
-
  clang2py -k cdefstum \
    extra/hip_gpu_driver/sdma_registers.h \
    $AMKERN_AMD/amdgpu/vega10_sdma_pkt_open.h \
@@ -414,72 +377,6 @@ generate_am() {
    -o $BASE/am/sdma_6_0_0.py
  fixup $BASE/am/sdma_6_0_0.py

-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/mmhub/mmhub_3_0_0_offset.h \
-    $AMKERN_INC/asic_reg/mmhub/mmhub_3_0_0_sh_mask.h \
-    -o $BASE/am/mmhub_3_0_0.py
-  fixup $BASE/am/mmhub_3_0_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/mmhub/mmhub_3_0_2_offset.h \
-    $AMKERN_INC/asic_reg/mmhub/mmhub_3_0_2_sh_mask.h \
-    -o $BASE/am/mmhub_3_0_2.py
-  fixup $BASE/am/mmhub_3_0_2.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/nbio/nbio_2_3_offset.h \
-    $AMKERN_INC/asic_reg/nbio/nbio_2_3_sh_mask.h \
-    -o $BASE/am/nbio_2_3_0.py
-  fixup $BASE/am/nbio_2_3_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/nbio/nbio_7_2_0_offset.h \
-    $AMKERN_INC/asic_reg/nbio/nbio_7_2_0_sh_mask.h \
-    -o $BASE/am/nbio_7_2_0.py
-  fixup $BASE/am/nbio_7_2_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/mmhub/mmhub_4_1_0_offset.h \
-    $AMKERN_INC/asic_reg/mmhub/mmhub_4_1_0_sh_mask.h \
-    -o $BASE/am/mmhub_4_1_0.py
-  fixup $BASE/am/mmhub_4_1_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/nbio/nbio_4_3_0_offset.h \
-    $AMKERN_INC/asic_reg/nbio/nbio_4_3_0_sh_mask.h \
-    -o $BASE/am/nbio_4_3_0.py
-  fixup $BASE/am/nbio_4_3_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/nbif/nbif_6_3_1_offset.h \
-    $AMKERN_INC/asic_reg/nbif/nbif_6_3_1_sh_mask.h \
-    -o $BASE/am/nbif_6_3_1.py
-  fixup $BASE/am/nbif_6_3_1.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/nbio/nbio_7_9_0_offset.h \
-    $AMKERN_INC/asic_reg/nbio/nbio_7_9_0_sh_mask.h \
-    -o $BASE/am/nbio_7_9_0.py
-  fixup $BASE/am/nbio_7_9_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/nbio/nbio_7_11_0_offset.h \
-    $AMKERN_INC/asic_reg/nbio/nbio_7_11_0_sh_mask.h \
-    -o $BASE/am/nbio_7_11_0.py
-  fixup $BASE/am/nbio_7_11_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/oss/osssys_6_0_0_offset.h \
-    $AMKERN_INC/asic_reg/oss/osssys_6_0_0_sh_mask.h \
-    -o $BASE/am/osssys_6_0_0.py
-  fixup $BASE/am/osssys_6_0_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/oss/osssys_7_0_0_offset.h \
-    $AMKERN_INC/asic_reg/oss/osssys_7_0_0_sh_mask.h \
-    -o $BASE/am/osssys_7_0_0.py
-  fixup $BASE/am/osssys_7_0_0.py
-
  clang2py -k cdefstum \
    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v13_0_0_ppsmc.h \
    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu13_driver_if_v13_0_0.h \
@@ -495,18 +392,6 @@ generate_am() {
    --clang-args="-include stdint.h" \
    -o $BASE/am/smu_v14_0_3.py
  fixup $BASE/am/smu_v14_0_3.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/hdp/hdp_6_0_0_offset.h \
-    $AMKERN_INC/asic_reg/hdp/hdp_6_0_0_sh_mask.h \
-    -o $BASE/am/hdp_6_0_0.py
-  fixup $BASE/am/hdp_6_0_0.py
-
-  clang2py -k cdefstum \
-    $AMKERN_INC/asic_reg/hdp/hdp_7_0_0_offset.h \
-    $AMKERN_INC/asic_reg/hdp/hdp_7_0_0_sh_mask.h \
-    -o $BASE/am/hdp_7_0_0.py
-  fixup $BASE/am/hdp_7_0_0.py
 }

 generate_sqtt() {
@@ -39,7 +39,7 @@ DEVICE = "CPU"   # NOTE: you can change this!
 import struct
 from tinygrad.dtype import dtypes
 from tinygrad.device import Buffer, Device
-from tinygrad.ops import UOp, Ops
+from tinygrad.uop.ops import UOp, Ops
 from tinygrad.shape.shapetracker import ShapeTracker

 # allocate some buffers + load in values
@@ -51,11 +51,11 @@ b = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struc
 # describe the computation
 buf_1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 1)
 buf_2 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 2)
-ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1, ShapeTracker.from_shape((1,)).to_uop()))
-ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2, ShapeTracker.from_shape((1,)).to_uop()))
+ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1.view(ShapeTracker.from_shape((1,))),))
+ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2.view(ShapeTracker.from_shape((1,))),))
 alu = ld_1 + ld_2
 output_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
-st_0 = UOp(Ops.STORE, dtypes.void, (output_buf, ShapeTracker.from_shape((1,)).to_uop(), alu))
+st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.view(ShapeTracker.from_shape((1,))), alu))
 s = UOp(Ops.SINK, dtypes.void, (st_0,))

 # convert the computation to a "linearized" format (print the format)
@@ -78,7 +78,7 @@ print("******** third, the UOp ***********")

 from tinygrad.engine.realize import run_schedule
 from tinygrad.engine.schedule import create_schedule_with_vars
-from tinygrad.engine.grouper import get_becomes_map
+from tinygrad.engine.grouper import get_kernelize_map

 # allocate some values + load in values
 a = UOp.new_buffer(DEVICE, 1, dtypes.int32)
@@ -91,7 +91,7 @@ out = a + b
 s = UOp(Ops.SINK, dtypes.void, (out,))

 # group the computation into kernels
-becomes_map = get_becomes_map(s)
+becomes_map = get_kernelize_map(s)

 # the compute maps to an assign
 assign = becomes_map[a+b]
@@ -103,7 +103,7 @@ assert assign.src[1].op is Ops.KERNEL

 # schedule the kernel graph in a linear list
 s = UOp(Ops.SINK, dtypes.void, (assign,))
-sched, _, becomes_map = create_schedule_with_vars(s)
+sched, _ = create_schedule_with_vars(s)
 assert len(sched) == 1

 # DEBUGGING: print the compute ast
@@ -111,7 +111,7 @@ print(sched[-1].ast)
 # NOTE: sched[-1].ast is the same as st_0 above

 # the output will be stored in a new buffer
-out = becomes_map[assign]
+out = assign.buf_uop
 assert out.op is Ops.BUFFER and not out.buffer.is_allocated()
 print(out)

@@ -36,7 +36,7 @@ optim.schedule_step()   # this will step the optimizer without running realize
 # 3. Create a schedule.

 # The weight Tensors have been assigned to, but not yet realized. Everything is still lazy at this point
-# l1.lazydata and l2.lazydata define a computation graph
+# l1.uop and l2.uop define a computation graph

 from tinygrad.engine.schedule import ScheduleItem
 schedule: List[ScheduleItem] = Tensor.schedule(l1, l2)
@@ -34,7 +34,7 @@ print(out) # <Tensor <UOp METAL (1,) int (<Ops.ASSIGN: 66>, None)> on METAL with
 The multiply Tensor stays the same because it is fused. The output Tensor's UOp becomes a new ASSIGN UOp:

 ```py
-print(out.lazydata)
+print(out.uop)
 ```

 The first source is the output BUFFER:
@@ -72,7 +72,7 @@ Once a Tensor is kernelized, all children will LOAD its BUFFER, instead of fusin
 ```py
 child = out+2
 child.kernelize()
-print(child.lazydata.src[1].arg.ast)
+print(child.uop.src[1].arg.ast)
 ```

 ```
@@ -68,4 +68,4 @@ We have a simple framework in tinygrad for adding these ALU blocks and achieving

 ### Indexing

-Indexing determines the address of the memory we need to load. GPUs often have less integer math resources than floating point math, so this can sometimes be the bottleneck. We have a symbolic math engine in our rewrite rules to simplifiy indexing before it's emitted to the kernel. Newer NVIDIA GPUs have a "Tensor Memory Accelerator" to assist with fast indexing, however, this is not supported in tinygrad yet.
+Indexing determines the address of the memory we need to load. GPUs often have less integer math resources than floating point math, so this can sometimes be the bottleneck. We have a symbolic math engine in our rewrite rules to simplify indexing before it's emitted to the kernel. Newer NVIDIA GPUs have a "Tensor Memory Accelerator" to assist with fast indexing, however, this is not supported in tinygrad yet.
@@ -1,10 +1,10 @@
-::: tinygrad.ops.UOp
+::: tinygrad.uop.ops.UOp
    options:
        members: false
        members_order: source
        show_labels: false

-::: tinygrad.ops.Ops
+::: tinygrad.uop.ops.Ops
    options:
        members: true
        members_order: source
@@ -36,7 +36,6 @@ CUDA                | [1]        | enable CUDA backend
 AMD                 | [1]        | enable AMD backend
 NV                  | [1]        | enable NV backend
 METAL               | [1]        | enable Metal backend (for Mac M1 and after)
-METAL_XCODE         | [1]        | enable Metal using macOS Xcode SDK
 CPU                 | [1]        | enable CPU (Clang) backend
 LLVM                | [1]        | enable LLVM backend
 BEAM                | [#]        | number of beams in kernel beam search
@@ -42,7 +42,7 @@ There's nothing special about a "Module" class in tinygrad, it's just a normal c

 ### tinygrad is functional

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

 ### tinygrad is lazy

@@ -0,0 +1,293 @@
+#!/usr/bin/env python3
+
+# this file is a "ramp" for people new to tinygrad to think about how to approach it
+# it is runnable and editable.
+# whenever you see stuff like DEBUG=2 or CPU=1 discussed, these are environment variables
+# in a unix shell like bash `DEBUG=2 CPU=1 python docs/ramp.py`
+
+# this pip installs tinygrad master for the system
+# the -e allows you to edit the tinygrad folder and update system tinygrad
+# tinygrad is pure Python, so you are encouraged to do this
+# git pull in the tinygrad directory will also get you the latest
+"""
+git clone https://github.com/tinygrad/tinygrad.git
+cd tinygrad
+python3 -m pip install -e .
+"""
+
+# %% ********
+print("******* PART 1 *******")
+
+# we start with a Device.
+# a Device is where Tensors are stored and compute is run
+# tinygrad autodetects the best device on your system and makes it the DEFAULT
+from tinygrad import Device
+print(Device.DEFAULT)  # on Mac, you can see this prints METAL
+
+# now, lets create a Tensor
+from tinygrad import Tensor, dtypes
+t = Tensor([1,2,3,4])
+
+# you can see this Tensor is on the DEFAULT device with int dtype and shape (4,)
+assert t.device == Device.DEFAULT
+assert t.dtype == dtypes.int
+assert t.shape == (4,)
+
+# unlike in torch, if we print it, it doesn't print the contents
+# this is because tinygrad is lazy
+# this Tensor has not been computed yet
+print(t)
+# <Tensor <UOp METAL (4,) int (<Ops.COPY: 7>, None)> on METAL with grad None>
+
+# the ".uop" property on Tensor contains the specification of how to compute it
+print(t.uop)
+"""
+UOp(Ops.COPY, dtypes.int, arg=None, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=0, src=()),
+    UOp(Ops.DEVICE, dtypes.void, arg='PYTHON', src=()),)),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# as you can see, it's specifying a copy from PYTHON device
+# which is where the [1,2,3,4] array lives
+
+# UOps are the specification language in tinygrad
+# they are immutable and form a DAG
+# they have a "Ops", a "dtype", a tuple of srcs (parents), and an arg
+
+t.realize()
+# if we want to "realize" a tensor, we can with the "realize" method
+# now when we look at the uop, it's changed
+print(t.uop)
+"""
+UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+  UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# the copy was actually run, and now the "uop" of the Tensor is just a BUFFER
+# if you run this script with DEBUG=2 in the environment, you can see the copy happen
+# *** METAL      1 copy       16,   METAL <- PYTHON ...
+
+# now let's do some compute
+# we look at the uop to see the specification of the compute
+t_times_2 = t * 2
+print(t_times_2.uop)
+"""
+UOp(Ops.MUL, dtypes.int, arg=None, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+    x2:=UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),)),
+  UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+    UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+      UOp(Ops.CONST, dtypes.int, arg=2, src=(
+        UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(), strides=(), offset=0, mask=None, contiguous=True),)), src=(
+           x2,)),)),)),)),))
+"""
+# the BUFFER from above is being multiplied by a CONST 2
+# it's RESHAPEd and EXPANDed to broadcast the CONST to the BUFFER
+
+# we can check the result with
+assert t_times_2.tolist() == [2, 4, 6, 8]
+
+# UOps are both immutable and globally unique
+# if i multiply the Tensor by 4 twice, these result Tensors will have the same uop specification
+t_times_4_try_1 = t * 4
+t_times_4_try_2 = t * 4
+assert t_times_4_try_1.uop is t_times_4_try_2.uop
+# the specification isn't just the same, it's the exact same Python object
+assert t_times_4_try_1 is not t_times_4_try_2
+# the Tensor is a different Python object
+
+# if we realize `t_times_4_try_1` ...
+t_times_4_try_1.realize()
+print(t_times_4_try_2.uop)
+"""
+UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+  UOp(Ops.UNIQUE, dtypes.void, arg=4, src=()),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# ... `t_times_4_try_2` also becomes the same BUFFER
+assert t_times_4_try_1.uop is t_times_4_try_2.uop
+# so this print doesn't require any computation, just a copy back to the CPU so we can print it
+print("** only the copy start")
+print(t_times_4_try_2.tolist())  # [4, 8, 12, 16]
+print("** only the copy end")
+# you can confirm this with DEBUG=2, seeing what's printed in between the "**" prints
+
+# tinygrad has an auto differentiation engine that operates according to these same principles
+# the derivative of "log(x)" is "1/x", and you can see this on line 20 of gradient.py
+t_float = Tensor([3.0])
+t_log = t_float.log()
+t_log_grad, = t_log.sum().gradient(t_float)
+# due to how log is implemented, this gradient contains a lot of UOps
+print(t_log_grad.uop)
+# ...not shown here...
+# but if you run with DEBUG=4 (CPU=1 used here for simpler code), you can see the generated code
+"""
+void E_(float* restrict data0, float* restrict data1) {
+  float val0 = *(data1+0);
+  *(data0+0) = (0.6931471805599453f*(1/(val0*0.6931471805599453f)));
+}
+"""
+# the derivative is close to 1/3
+assert (t_log_grad.item() - 1/3) < 1e-6
+
+# %% ********
+print("******* PART 2 *******")
+
+# we redefine the same t here so this cell can run on it's own
+from tinygrad import Tensor
+t = Tensor([1,2,3,4])
+
+# what's above gives you enough of an understanding to go use tinygrad as a library
+# however, a lot of the beauty of tinygrad is in how easy it is to interact with the internals
+# NOTE: the APIs here are subject to change
+
+t_plus_3_plus_4 = t + 3 + 4
+print(t_plus_3_plus_4.uop)
+"""
+UOp(Ops.ADD, dtypes.int, arg=None, src=(
+  UOp(Ops.ADD, dtypes.int, arg=None, src=(
+    UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+      UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+      x3:=UOp(Ops.DEVICE, dtypes.void, arg='CPU', src=()),)),
+    UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+      UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+        UOp(Ops.CONST, dtypes.int, arg=3, src=(
+          x7:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(), strides=(), offset=0, mask=None, contiguous=True),)), src=(
+             x3,)),)),)),)),)),
+  UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+    UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+      UOp(Ops.CONST, dtypes.int, arg=4, src=(
+         x7,)),)),)),))
+"""
+# you can see it's adding both 3 and 4
+
+# but by the time we are actually running the code, it's adding 7
+# `kernelize` will simplify and group the operations in the graph into kernels
+t_plus_3_plus_4.kernelize()
+print(t_plus_3_plus_4.uop)
+"""
+UOp(Ops.ASSIGN, dtypes.int, arg=None, src=(
+  x0:=UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=7, src=()),
+    x2:=UOp(Ops.DEVICE, dtypes.void, arg='CPU', src=()),)),
+  UOp(Ops.KERNEL, dtypes.void, arg=<Kernel 12 SINK(<Ops.STORE: 48>,) (__add__,)>, src=(
+     x0,
+    UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+      UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+       x2,)),)),))
+"""
+# ASSIGN has two srcs, src[0] is the BUFFER that's assigned to, and src[1] is the thing to assign
+# src[1] is the GPU Kernel that's going to be run
+# we can get the ast of the Kernel as follows
+kernel_ast = t_plus_3_plus_4.uop.src[1].arg.ast
+
+# almost everything in tinygrad functions as a rewrite of the UOps
+# the codegen rewrites the ast to a simplified form ready for "rendering"
+from tinygrad.codegen import full_rewrite_to_sink
+rewritten_ast = full_rewrite_to_sink(kernel_ast)
+print(rewritten_ast)
+"""
+UOp(Ops.SINK, dtypes.void, arg=None, src=(
+  UOp(Ops.STORE, dtypes.void, arg=None, src=(
+    UOp(Ops.INDEX, dtypes.int.ptr(4), arg=None, src=(
+      UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(4), arg=0, src=()),
+      x3:=UOp(Ops.SPECIAL, dtypes.int, arg=('gidx0', 4), src=()),)),
+    UOp(Ops.ADD, dtypes.int, arg=None, src=(
+      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
+        UOp(Ops.INDEX, dtypes.int.ptr(4), arg=None, src=(
+          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(4), arg=1, src=()),
+           x3,)),)),
+      UOp(Ops.CONST, dtypes.int, arg=7, src=()),)),)),))
+"""
+# you can see at this point we are adding 7, not 3 and 4
+
+# with DEBUG=4, we can see the code.
+# since optimizations are on, it UPCASTed the operation, explicitly writing out all 4 +7s
+t_plus_3_plus_4.realize()
+"""
+void E_4n2(int* restrict data0, int* restrict data1) {
+  int val0 = *(data1+0);
+  int val1 = *(data1+1);
+  int val2 = *(data1+2);
+  int val3 = *(data1+3);
+  *(data0+0) = (val0+7);
+  *(data0+1) = (val1+7);
+  *(data0+2) = (val2+7);
+  *(data0+3) = (val3+7);
+}
+"""
+# the function name E_4n2 is "E" for elementwise op (as opposed to "r" for reduce op)
+# "4" for the size, and "n2" for name deduping (it's the 3rd function with the same E and 4 in this session)
+# when you print the name with DEBUG=2, you'll see the 4 is yellow, meaning that it's upcasted
+# if you run with NOOPT=1 ...
+"""
+void E_4n2(int* restrict data0, int* restrict data1) {
+  for (int ridx0 = 0; ridx0 < 4; ridx0++) {
+    int val0 = *(data1+ridx0);
+    *(data0+ridx0) = (val0+7);
+  }
+}
+"""
+# ... you get this unoptimized code with a loop and the 4 is blue (for global). the color code is in kernel.py
+
+# %% ********
+print("******* PART 3 *******")
+
+# now, we go even lower and understand UOps better and how the graph rewrite engine works.
+# it's much simpler than what's in LLVM or MLIR
+
+from tinygrad import dtypes
+from tinygrad.uop import UOp, Ops
+
+# first, we'll construct some const UOps
+a = UOp(Ops.CONST, dtypes.int, arg=2)
+b = UOp(Ops.CONST, dtypes.int, arg=2)
+
+# if you have been paying attention, you should know these are the same Python object
+assert a is b
+
+# UOps support normal Python math operations, so a_plus_b expresses the spec for 2 + 2
+a_plus_b = a + b
+print(a_plus_b)
+"""
+UOp(Ops.ADD, dtypes.int, arg=None, src=(
+  x0:=UOp(Ops.CONST, dtypes.int, arg=2, src=()),
+   x0,))
+"""
+
+# we could actually render this 2+2 into a language like c and run it
+# or, we can use tinygrad's graph rewrite engine to "constant fold"
+
+from tinygrad.uop.ops import graph_rewrite, UPat, PatternMatcher
+
+# a `PatternMatcher` is a list of tuples. for each element in the list:
+# [0] is the pattern to match, and [1] is the function to run.
+# this function can return either a UOp to replace the pattern with, or None to not replace
+simple_pm = PatternMatcher([
+  (UPat(Ops.ADD, src=(UPat(Ops.CONST, name="c1"), UPat(Ops.CONST, name="c2"))),
+   lambda c1,c2: UOp(Ops.CONST, dtype=c1.dtype, arg=c1.arg+c2.arg)),
+])
+# this pattern matches the addition of two CONST and rewrites it into a single CONST UOp
+
+# to actually apply the pattern to a_plus_b, we use graph_rewrite
+a_plus_b_simplified = graph_rewrite(a_plus_b, simple_pm)
+print(a_plus_b_simplified)
+"""
+UOp(Ops.CONST, dtypes.int, arg=4, src=())
+"""
+# 2+2 is in fact, 4
+
+# we can also use syntactic sugar to write the pattern nicer
+simpler_pm = PatternMatcher([
+  (UPat.cvar("c1")+UPat.cvar("c2"), lambda c1,c2: c1.const_like(c1.arg+c2.arg))
+])
+assert graph_rewrite(a_plus_b, simple_pm) is graph_rewrite(a_plus_b, simpler_pm)
+# note again the use of is, UOps are immutable and globally unique
+
+# %% ********
+
+# that brings you to an understanding of the most core concepts in tinygrad
+# you can run this with VIZ=1 to use the web based graph rewrite explorer
+# hopefully now you understand it. the nodes in the graph are just UOps
@@ -24,6 +24,7 @@
 ::: tinygrad.Tensor.randn
 ::: tinygrad.Tensor.randn_like
 ::: tinygrad.Tensor.randint
+::: tinygrad.Tensor.randperm
 ::: tinygrad.Tensor.normal
 ::: tinygrad.Tensor.uniform
 ::: tinygrad.Tensor.scaled_uniform
@@ -18,6 +18,7 @@
 ::: tinygrad.Tensor.repeat_interleave
 ::: tinygrad.Tensor.split
 ::: tinygrad.Tensor.chunk
+::: tinygrad.Tensor.unfold
 ::: tinygrad.Tensor.meshgrid
 ::: tinygrad.Tensor.squeeze
 ::: tinygrad.Tensor.unsqueeze
@@ -37,8 +37,10 @@
 ::: tinygrad.Tensor.scatter
 ::: tinygrad.Tensor.scatter_reduce
 ::: tinygrad.Tensor.masked_select
+::: tinygrad.Tensor.masked_fill
 ::: tinygrad.Tensor.sort
 ::: tinygrad.Tensor.topk
+::: tinygrad.Tensor.multinomial

 ## Neural Network (functional)

@@ -78,10 +78,7 @@ if __name__ == "__main__":

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

  st, steps = time.perf_counter(), 0
@@ -3,14 +3,19 @@ start_tm = time.perf_counter()
 import math
 from typing import Tuple, cast
 import numpy as np
-from tinygrad import Tensor, nn, GlobalCounters, TinyJit, dtypes
+from tinygrad import Tensor, nn, GlobalCounters, TinyJit, dtypes, Device
 from tinygrad.helpers import partition, trange, getenv, Context
 from extra.lr_scheduler import OneCycleLR

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

 # from https://github.com/tysam-code/hlb-CIFAR10/blob/main/main.py
 batchsize = getenv("BS", 1024)
+assert batchsize % len(GPUS) == 0, f"{batchsize=} is not a multiple of {len(GPUS)=}"
 bias_scaler = 64
 hyp = {
  'opt': {
@@ -67,7 +72,7 @@ class ConvGroup:
    cast(Tensor, self.norm2.weight).requires_grad = False
  def __call__(self, x:Tensor) -> Tensor:
    x =    self.norm1(self.conv1(x).max_pool2d().float()).cast(dtypes.default_float).quick_gelu()
-    return self.norm2(self.conv2(x).float()).cast(dtypes.default_float).quick_gelu()
+    return self.norm2(self.conv2(x).float()).cast(dtypes.default_float).quick_gelu() + x

 class SpeedyConvNet:
  def __init__(self):
@@ -78,23 +83,25 @@ class SpeedyConvNet:
    self.linear = nn.Linear(depths['block3'], depths['num_classes'], bias=False)
  def __call__(self, x:Tensor) -> Tensor:
    x = self.whiten(x).quick_gelu()
+    # ************* HACKS *************
+    x = x.pad((1,0,0,1)) # TODO: this pad should not be here! copied from hlb_cifar10 for speed
+    # ************* HACKS *************
    x = x.sequential([self.conv_group_1, self.conv_group_2, self.conv_group_3])
    return self.linear(x.max(axis=(2,3))) * hyp['opt']['scaling_factor']

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

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

  params_bias, params_non_bias = partition(state_dict.items(), lambda x: 'bias' in x[0])
  opt_bias     = nn.optim.SGD([x[1] for x in params_bias],     lr=0.01, momentum=.85, nesterov=True, weight_decay=hyp['opt']['bias_decay'])
@@ -111,40 +118,37 @@ if __name__ == "__main__":
  lr_sched_bias     = OneCycleLR(opt_bias,     max_lr=hyp['opt']['bias_lr'],     pct_start=pct_start, div_factor=initial_div_factor, final_div_factor=1./(initial_div_factor*final_lr_ratio), total_steps=total_train_steps)
  lr_sched_non_bias = OneCycleLR(opt_non_bias, max_lr=hyp['opt']['non_bias_lr'], pct_start=pct_start, div_factor=initial_div_factor, final_div_factor=1./(initial_div_factor*final_lr_ratio), total_steps=total_train_steps)

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

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

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

  np.random.seed(1337)
  for epoch in range(math.ceil(hyp['misc']['train_epochs'])):
@@ -34,7 +34,6 @@ if __name__ == "__main__":
    return loss

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

  test_acc = float('nan')
@@ -1,10 +1,10 @@
-import sys, onnx, time, pickle
+import sys, time, pickle
 from tinygrad import TinyJit, GlobalCounters, fetch, getenv
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner, onnx_load
 from extra.onnx_helpers import get_example_inputs, validate

 def load_onnx_model(onnx_file):
-  onnx_model = onnx.load(onnx_file)
+  onnx_model = onnx_load(onnx_file)
  run_onnx = OnnxRunner(onnx_model)
  run_onnx_jit = TinyJit(lambda **kwargs: next(iter(run_onnx({k:v.to(None) for k,v in kwargs.items()}).values())), prune=True, optimize=True)
  return run_onnx_jit, run_onnx.graph_inputs
@@ -23,8 +23,6 @@ def create_fixed_tokenizer(output_file):
 # echo -en "write 2+2\nwrite hello world\ny\n" | TEMP=0 python3 examples/coder.py

 if __name__ == "__main__":
-  Tensor.no_grad = True
-
  # https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/config.json
  with Timing("create model: "):
    model = Transformer(4096, 14336, n_heads=32, n_layers=32, norm_eps=1e-5, vocab_size=32002, n_kv_heads=8, max_context=4096, jit=getenv("JIT", 1))
@@ -159,7 +159,6 @@ def init_vits(
  text_mapper = TextMapper(apply_cleaners=True, symbols=symbols)

  # Load the model.
-  Tensor.no_grad = True
  if seed is not None:
    Tensor.manual_seed(seed)
    np.random.seed(seed)
@@ -221,7 +220,6 @@ def mp_output_stream(q: mp.Queue, counter: mp.Value, num_channels: int, sample_r
 if __name__ == "__main__":
  import nltk
  nltk.download("punkt")
-  Tensor.no_grad = True
  # Parse CLI arguments
  parser = argparse.ArgumentParser("Have a tiny conversation with tinygrad")

@@ -3,10 +3,11 @@ import os, argparse, contextlib
 from typing import Optional, Union
 with contextlib.suppress(ImportError): import tiktoken
 from tinygrad import Tensor, TinyJit, Device, GlobalCounters, Variable, dtypes
-from tinygrad.ops import UOp
+from tinygrad.uop.ops import UOp
 from tinygrad.helpers import Timing, DEBUG, JIT, getenv, fetch, colored, trange
 from tinygrad.nn import Embedding, Linear, LayerNorm
 from tinygrad.nn.state import gguf_load, torch_load, load_state_dict, get_state_dict
+from extra.bench_log import BenchEvent, WallTimeEvent

 MAX_CONTEXT = getenv("MAX_CONTEXT", 128)
 HALF = getenv("HALF")
@@ -84,7 +85,10 @@ class Transformer:
      seqlen = tokens.shape[1]
      tok_emb = self.wte(tokens)

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

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

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

    return GPT2(model, tokenizer)

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

  def __init__(self, model, tokenizer):
@@ -185,11 +191,12 @@ class GPT2:
      with Timing("ran model in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
                  f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
                  (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=timing):
-        if batch_size == 1 and len(toks[0][start_pos:]) == 1:
-          tokens = Variable("tokens", 0, VOCAB_SIZE).bind(toks[0][start_pos])
-        else:
-          tokens = Tensor([x[start_pos:] for x in toks])
-        tok = self.model(tokens, Variable("start_pos", 1 if start_pos else 0, MAX_CONTEXT-1).bind(start_pos), temperature).tolist()
+        with WallTimeEvent(BenchEvent.STEP):
+          if batch_size == 1 and len(toks[0][start_pos:]) == 1:
+            tokens = Variable("tokens", 0, VOCAB_SIZE-1).bind(toks[0][start_pos])
+          else:
+            tokens = Tensor([x[start_pos:] for x in toks])
+          tok = self.model(tokens, Variable("start_pos", 1 if start_pos else 0, MAX_CONTEXT-1).bind(start_pos), temperature).tolist()
      start_pos = len(toks[0])
      for i,t in enumerate(tok): toks[i].append(t)
    return [self.tokenizer.decode(x) for x in toks]
@@ -197,7 +204,6 @@ class GPT2:
 # **** main code ****

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

@@ -4,7 +4,7 @@ from examples.mlperf.helpers import get_mlperf_bert_model
 from tinygrad import Tensor, Device, dtypes, nn
 from tinygrad.codegen.kernel import Kernel
 from tinygrad.codegen.heuristic import hand_coded_optimizations
-from tinygrad.ops import Ops, sym_infer
+from tinygrad.uop.ops import Ops, sym_infer
 from tinygrad.device import Compiled
 from tinygrad.engine.search import beam_search, bufs_from_lin
 from tinygrad.helpers import DEBUG, ansilen, getenv, colored, TRACEMETA
@@ -11,6 +11,7 @@ from tinygrad import nn, dtypes, Tensor, Device, GlobalCounters, TinyJit
 from tinygrad.nn.state import get_state_dict, get_parameters
 from tinygrad.nn import optim
 from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod
+from extra.bench_log import BenchEvent, WallTimeEvent

 cifar_mean = [0.4913997551666284, 0.48215855929893703, 0.4465309133731618]
 cifar_std = [0.24703225141799082, 0.24348516474564, 0.26158783926049628]
@@ -117,7 +118,7 @@ class SpeedyResNet:
 # hyper-parameters were exactly the same as the original repo
 bias_scaler = 58
 hyp = {
-  'seed' : 209,
+  'seed' : 200,
  'opt': {
    'bias_lr':            1.76 * bias_scaler/512,
    'non_bias_lr':        1.76 / 512,
@@ -266,13 +267,10 @@ def train_cifar():

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

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

@@ -395,20 +393,23 @@ def train_cifar():
      if STEPS == 0 or i == STEPS: break

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

-      with Context(BEAM=getenv("LATEBEAM", BEAM.value), WINO=getenv("LATEWINO", WINO.value)):
-        loss = train_step_jitted(model, optim.OptimizerGroup(opt_bias, opt_non_bias), [lr_sched_bias, lr_sched_non_bias], X, Y)
-        et = time.monotonic()
-        loss_cpu = loss.numpy()
-      # EMA for network weights
-      if getenv("EMA") and i > hyp['ema']['steps'] and (i+1) % hyp['ema']['every_n_steps'] == 0:
-        if model_ema is None:
-          model_ema = modelEMA(W, model)
-        model_ema.update(model, Tensor([projected_ema_decay_val*(i/STEPS)**hyp['ema']['decay_pow']]))
+      with WallTimeEvent(BenchEvent.STEP):
+        X, Y = next(batcher)
+        if len(GPUS) > 1:
+          X.shard_(GPUS, axis=0)
+          Y.shard_(GPUS, axis=0)
+
+        with Context(BEAM=getenv("LATEBEAM", BEAM.value), WINO=getenv("LATEWINO", WINO.value)):
+          loss = train_step_jitted(model, optim.OptimizerGroup(opt_bias, opt_non_bias), [lr_sched_bias, lr_sched_non_bias], X, Y)
+          et = time.monotonic()
+          loss_cpu = loss.numpy()
+        # EMA for network weights
+        if getenv("EMA") and i > hyp['ema']['steps'] and (i+1) % hyp['ema']['every_n_steps'] == 0:
+          if model_ema is None:
+            model_ema = modelEMA(W, model)
+          model_ema.update(model, Tensor([projected_ema_decay_val*(i/STEPS)**hyp['ema']['decay_pow']]))
+
      cl = time.monotonic()
      device_str = loss.device if isinstance(loss.device, str) else f"{loss.device[0]} * {len(loss.device)}"
      #  53  221.74 ms run,    2.22 ms python,  219.52 ms CL,  803.39 loss, 0.000807 LR, 4.66 GB used,   3042.49 GFLOPS,    674.65 GOPS
@@ -424,4 +425,5 @@ def train_cifar():
      raise ValueError(colored(f"{eval_acc_pct=} < {target}", "red"))

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

 MAX_CONTEXT = getenv("MAX_CONTEXT", 4096)

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

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

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

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

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

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

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

    return LLaMa(model, tokenizer)

@@ -329,7 +330,6 @@ int main()
 \end{code}
 """
 if __name__ == "__main__":
-  Tensor.no_grad = True
  print(f"using {Device.DEFAULT} backend")

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

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

      # use the kv cache
      start_pos = len(toks)
@@ -7,6 +7,7 @@ from extra.models.llama import Transformer, convert_from_huggingface, convert_fr
 from tinygrad.nn.state import safe_load, torch_load, load_state_dict, get_parameters, gguf_load
 from tinygrad import Tensor, dtypes, nn, Context, Device, GlobalCounters
 from tinygrad.helpers import Profiling, Timing, DEBUG, colored, fetch, tqdm
+from extra.bench_log import BenchEvent, WallTimeEvent

 class Tokenizer:
  pat_str = r"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"
@@ -166,40 +167,42 @@ def build_transformer(model_path: Path, model_size="8B", quantize=None, scale_dt
  model = Transformer(**MODEL_PARAMS[model_size]["args"], linear=linear, embedding=embedding, max_context=max_context, jit=True)

  if not load_weights: return model
+
  # load weights
-  if model_path.is_dir():
-    if (model_path / "model.safetensors.index.json").exists(): weights = load(str(model_path / "model.safetensors.index.json"))
-    elif (model_path / "model.safetensors").exists(): weights = load(str(model_path / "model.safetensors"))
-    else: weights = concat_weights([load(str(model_path / f"consolidated.{i:02d}.pth")) for i in range(MODEL_PARAMS[model_size]["files"])], device[0] if isinstance(device, tuple) else device)
-  else:
-    weights = load(str(model_path))
-  if "model.embed_tokens.weight" in weights:
-    weights = convert_from_huggingface(weights, MODEL_PARAMS[model_size]["args"]["n_layers"], MODEL_PARAMS[model_size]["args"]["n_heads"], MODEL_PARAMS[model_size]["args"]["n_kv_heads"])
-  elif "token_embd.weight" in weights:
-    weights = convert_from_gguf(weights, MODEL_PARAMS[model_size]["args"]["n_layers"])
-  weights = fix_bf16(weights)
+  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+    if model_path.is_dir():
+      if (model_path / "model.safetensors.index.json").exists(): weights = load(str(model_path / "model.safetensors.index.json"))
+      elif (model_path / "model.safetensors").exists(): weights = load(str(model_path / "model.safetensors"))
+      else: weights = concat_weights([load(str(model_path / f"consolidated.{i:02d}.pth")) for i in range(MODEL_PARAMS[model_size]["files"])], device[0] if isinstance(device, tuple) else device)
+    else:
+      weights = load(str(model_path))
+    if "model.embed_tokens.weight" in weights:
+      weights = convert_from_huggingface(weights, MODEL_PARAMS[model_size]["args"]["n_layers"], MODEL_PARAMS[model_size]["args"]["n_heads"], MODEL_PARAMS[model_size]["args"]["n_kv_heads"])
+    elif "token_embd.weight" in weights:
+      weights = convert_from_gguf(weights, MODEL_PARAMS[model_size]["args"]["n_layers"])
+    weights = fix_bf16(weights)

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

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

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

 # default settings
@@ -230,8 +233,6 @@ def prefill(model, toks, start_pos=0):
  return start_pos

 if __name__ == "__main__":
-  Tensor.no_grad = True
-
  parser = argparse.ArgumentParser()
  parser.add_argument("--download_model", action="store_true", help="Download a model")
  parser.add_argument("--model", type=Path, help="Model path")
@@ -283,7 +284,7 @@ if __name__ == "__main__":

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

  if not args.no_api and not args.benchmark:
    from bottle import Bottle, request, response, HTTPResponse, abort, static_file
@@ -435,11 +436,12 @@ if __name__ == "__main__":
      st = GlobalCounters.time_sum_s
      with Profiling(enabled=args.profile):
        with Timing("total ", on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
-          with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
-                      f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
-                      (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None):
-            tok = model(Tensor([[last_tok]], device=device), start_pos, TEMPERATURE, TOP_K, TOP_P, ALPHA_F, ALPHA_P)
-          tok = tok.item()
+          with WallTimeEvent(BenchEvent.STEP):
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
+                        f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
+                        (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None):
+              tok = model(Tensor([[last_tok]], device=device), start_pos, TEMPERATURE, TOP_K, TOP_P, ALPHA_F, ALPHA_P)
+            tok = tok.item()
      start_pos += 1
      last_tok = tok
      generated += tokenizer.decode([tok])
@@ -7,7 +7,7 @@ from train_gpt2 import GPT, GPTConfig
 from tinygrad.helpers import dedup, to_function_name, flatten, getenv, GlobalCounters, ansilen, to_function_name
 from tinygrad.engine.realize import get_kernel, run_schedule
 from tinygrad.engine.memory import memory_planner
-from tinygrad.ops import Ops
+from tinygrad.uop.ops import Ops

 TIMING = getenv("TIMING")

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

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

  #B, T = Variable("B", 1, 128).bind(4), 64 #Variable("T", 1, 1024).bind(64)
@@ -56,7 +56,7 @@ if __name__ == "__main__":
  state_dict.update({'X': X, 'Y': Y, 'loss': loss})
  grad_state_dict = {}
  for k,v in state_dict.items():
-    if v.lazydata.base.buffer not in used_buffers: print(f"UNUSED: {k}")
+    if v.uop.base.buffer not in used_buffers: print(f"UNUSED: {k}")
    if v.grad is not None: grad_state_dict['grad_'+k] = v.grad
  state_dict.update(grad_state_dict)
  state_dict.update({'adam_b1_t': optimizer.b1_t, 'adam_b2_t': optimizer.b2_t, 'adam_lr': optimizer.lr})
@@ -65,7 +65,7 @@ if __name__ == "__main__":
    nm = inverse_state_dict[p]
    state_dict["adam_m_"+nm] = m
    state_dict["adam_v_"+nm] = v
-  named_buffers = {v.lazydata.base.buffer:k.replace(".", "_") for k,v in state_dict.items()}
+  named_buffers = {v.uop.base.buffer:k.replace(".", "_") for k,v in state_dict.items()}

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

  def __call__(self, idx:Tensor, targets=None):
    b, t = idx.shape
-    pos = Tensor.arange(0, t)
+    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)
@@ -124,6 +124,7 @@ if __name__ == "__main__":
  parser.add_argument("--batch_size", type=int, default=4, help="batch size")
  parser.add_argument("--sequence_length", type=int, default=64, help="sequence length")
  parser.add_argument("--skip_test", action="store_true", help="skip test")
+  parser.add_argument("--gpus", type=int, default=1, help="sequence length")
  args = parser.parse_args()
  B, T = args.batch_size, args.sequence_length
  assert 1 <= T <= 1024
@@ -131,6 +132,10 @@ if __name__ == "__main__":
  model = GPT(GPTConfig(n_layer=12, n_head=12, n_embd=768))
  model.load_pretrained()

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

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

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

  if not args.skip_test:
+    # copy back to single gpu for test
+    if args.gpus > 1:
+      for x in nn.state.get_parameters(model): x.to_(Device.DEFAULT)
    start = "<|endoftext|>"
    start_ids = encode(start)
    x = (Tensor(start_ids)[None, ...])
@@ -146,7 +146,6 @@ if __name__ == "__main__":
    return loss

  @TinyJit
-  @Tensor.test()
  def sample(z:Tensor, cond:Tensor) -> Tensor:
    return model.sample(z, cond, Tensor.full_like(cond, 10), sample_steps=getenv("SAMPLE_STEPS", 20))[-1]

@@ -3,6 +3,7 @@ from tinygrad import Tensor, nn, Device, GlobalCounters, Variable
 from tinygrad.helpers import Timing, Profiling, CI, tqdm
 from tinygrad.nn.state import torch_load, get_state_dict
 from extra.models.llama import FeedForward, Transformer
+from extra.bench_log import BenchEvent, WallTimeEvent

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

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

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

  from sentencepiece import SentencePieceProcessor
@@ -53,7 +55,8 @@ if __name__ == "__main__":
    GlobalCounters.reset()
    with Profiling(sort="time", frac=0.1, enabled=args.profile):
      with Timing("total ", enabled=args.timing, on_exit=lambda x: f", {1e9/x:.2f} tok/sec"):
-        tok = model(Tensor([toks[start_pos:]]), 0 if start_pos == 0 else Variable("start_pos", 1, 1024).bind(start_pos), args.temperature).item()
+        with WallTimeEvent(BenchEvent.STEP):
+          tok = model(Tensor([toks[start_pos:]]), 0 if start_pos == 0 else Variable("start_pos", 1, 1024-1).bind(start_pos), args.temperature).item()
    toks.append(tok)
    start_pos += 1
    print(spp.decode(toks))
@@ -5,7 +5,7 @@ from multiprocessing import Queue, Process, shared_memory, connection, Lock, cpu

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

 ### ResNet

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

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

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

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

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

    for _ in range(cpu_count()):
@@ -261,8 +262,8 @@ def load_unet3d_data(preprocessed_dataset_dir, seed, queue_in, queue_out, X:Tens
      x = random_brightness_augmentation(x)
      x = gaussian_noise(x)

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

    queue_out.put(idx)
  queue_out.put(None)
@@ -312,7 +313,7 @@ def batch_load_unet3d(preprocessed_dataset_dir:Path, batch_size:int=6, val:bool=
      proc = Process(target=load_unet3d_data, args=(preprocessed_dataset_dir, seed, queue_in, queue_out, X, Y))
      proc.daemon = True
      proc.start()
-      
+
      procs.append(proc)

    for bc in range(batch_count):
@@ -376,12 +377,12 @@ def load_retinanet_data(base_dir:Path, val:bool, queue_in:Queue, queue_out:Queue
      clipped_match_idxs = np.clip(match_idxs, 0, None)
      clipped_boxes, clipped_labels = tgt["boxes"][clipped_match_idxs], tgt["labels"][clipped_match_idxs]

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

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

    queue_out.put(idx)
  queue_out.put(None)
@@ -5,61 +5,62 @@ import numpy as np
 from tinygrad import Tensor, Device, dtypes, GlobalCounters, TinyJit
 from tinygrad.nn.state import get_parameters, load_state_dict, safe_load
 from tinygrad.helpers import getenv
+from extra.bench_log import BenchEvent, WallTimeEvent
 def tlog(x): print(f"{x:25s}  @ {time.perf_counter()-start:5.2f}s")

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

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

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

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

 def eval_unet3d():
  # UNet3D
@@ -243,7 +244,6 @@ def eval_mrcnn():
 if __name__ == "__main__":
  # inference only
  Tensor.training = False
-  Tensor.no_grad = True

  models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(",")
  for m in models:
@@ -60,7 +60,6 @@ def spec_mrcnn():
 if __name__ == "__main__":
  # inference only for now
  Tensor.training = False
-  Tensor.no_grad = True

  for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
    nm = f"spec_{m}"
@@ -1,4 +1,4 @@
-import os, time, math, functools, random
+import os, time, math, functools, random, contextlib
 from pathlib import Path
 import multiprocessing

@@ -9,6 +9,7 @@ from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam

 from extra.lr_scheduler import LRSchedulerGroup
 from examples.mlperf.helpers import get_training_state, load_training_state
+from extra.bench_log import BenchEvent, WallTimeEvent
 # TODO: fix benchmark logging and use tinygrad tqdm
 from tqdm import tqdm

@@ -205,24 +206,25 @@ def train_resnet():
    st = time.perf_counter()
    while proc is not None:
      GlobalCounters.reset()
-      (loss, top_1), y, proc = train_step(proc[0], proc[1]), proc[2], proc[3]
+      with WallTimeEvent(BenchEvent.STEP):
+        (loss, top_1), y, proc = train_step(proc[0], proc[1]), proc[2], proc[3]

-      pt = time.perf_counter()
+        pt = time.perf_counter()

-      if len(prev_cookies) == getenv("STORE_COOKIES", 1): prev_cookies = []  # free previous cookies after gpu work has been enqueued
-      try:
-        if INITMLPERF:
-          next_proc = fake_data_get(BS)
-        else:
-          next_proc = data_get(it)
-      except StopIteration:
-        next_proc = None
+        if len(prev_cookies) == getenv("STORE_COOKIES", 1): prev_cookies = []  # free previous cookies after gpu work has been enqueued
+        try:
+          if INITMLPERF:
+            next_proc = fake_data_get(BS)
+          else:
+            next_proc = data_get(it)
+        except StopIteration:
+          next_proc = None

-      dt = time.perf_counter()
+        dt = time.perf_counter()

-      device_str = loss.device if isinstance(loss.device, str) else f"{loss.device[0]} * {len(loss.device)}"
-      loss, top_1 = loss.numpy().item(), top_1.numpy().item()
-      top_1_acc = top_1 / sum(yi != -1 for yi in y)
+        device_str = loss.device if isinstance(loss.device, str) else f"{loss.device[0]} * {len(loss.device)}"
+        loss, top_1 = loss.numpy().item(), top_1.numpy().item()
+        top_1_acc = top_1 / sum(yi != -1 for yi in y)

      cl = time.perf_counter()
      if BENCHMARK:
@@ -366,6 +368,9 @@ def train_retinanet():
  INITMLPERF = getenv("INITMLPERF")
  RUNMLPERF = getenv("RUNMLPERF")

+  if INITMLPERF:
+    diskcache_clear()
+
  if getenv("LOGMLPERF"):
    from mlperf_logging import mllog
    import mlperf_logging.mllog.constants as mllog_constants
@@ -384,7 +389,6 @@ def train_retinanet():

      MLLOGGER.event(key=mllog_constants.SUBMISSION_BENCHMARK, value=mllog_constants.RETINANET)

-      diskcache_clear()
      MLLOGGER.event(key=mllog_constants.CACHE_CLEAR, value=True)
      MLLOGGER.start(key=mllog_constants.INIT_START)

@@ -604,7 +608,7 @@ def train_retinanet():

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

-      with Tensor.train(mode=False), Tensor.test():
+      with Tensor.train(mode=False):
        if not RUNMLPERF:
          i, proc = 0, _fake_data_get(EVAL_BS, val=(val:=True))
        else:
@@ -787,7 +791,6 @@ def train_unet3d():
    return loss.realize()

  @Tensor.train(mode=False)
-  @Tensor.test()
  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, requires_grad=False)
@@ -1124,23 +1127,24 @@ def train_bert():
      BEAM.value = TRAIN_BEAM
      st = time.perf_counter()
      GlobalCounters.reset()
-      loss, global_norm, lr = train_step_bert(model, optimizer_group, scheduler_group, loss_scaler,
-        train_data["input_ids"], train_data["segment_ids"], train_data["input_mask"], train_data["masked_lm_positions"], \
-        train_data["masked_lm_ids"], train_data["masked_lm_weights"], train_data["next_sentence_labels"], GPUS)
+      with WallTimeEvent(BenchEvent.STEP):
+        loss, global_norm, lr = train_step_bert(model, optimizer_group, scheduler_group, loss_scaler,
+          train_data["input_ids"], train_data["segment_ids"], train_data["input_mask"], train_data["masked_lm_positions"], \
+          train_data["masked_lm_ids"], train_data["masked_lm_weights"], train_data["next_sentence_labels"], GPUS)

-      pt = time.perf_counter()
+        pt = time.perf_counter()

-      try:
-        next_data = next(train_it)
-      except StopIteration:
-        next_data = None
+        try:
+          next_data = next(train_it)
+        except StopIteration:
+          next_data = None

-      dt = time.perf_counter()
+        dt = time.perf_counter()

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

      cl = time.perf_counter()
      if BENCHMARK: step_times.append(cl - st)
@@ -1276,9 +1280,15 @@ def train_maskrcnn():

 if __name__ == "__main__":
  multiprocessing.set_start_method('spawn')
+
+  if getenv("INITMLPERF"): bench_log_manager = WallTimeEvent(BenchEvent.MLPERF_INIT)
+  elif getenv("RUNMLPERF"): bench_log_manager = WallTimeEvent(BenchEvent.MLPERF_RUN)
+  else: bench_log_manager = contextlib.nullcontext()
+
  with Tensor.train():
    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn").split(","):
      nm = f"train_{m}"
      if nm in globals():
        print(f"training {m}")
-        with Profiling(enabled=getenv("PYPROFILE")): globals()[nm]()
+        with bench_log_manager:
+          with Profiling(enabled=getenv("PYPROFILE")): globals()[nm]()
@@ -1,4 +1,5 @@
 #!/bin/bash
+set -e  # Exit on any error

 export PYTHONPATH="." AMD=1
 export MODEL="bert"
@@ -1,4 +1,5 @@
 #!/bin/bash
+set -e  # Exit on any error

 export PYTHONPATH="." NV=1
 export MODEL="bert"
@@ -1,4 +1,5 @@
 #!/bin/bash
+set -e  # Exit on any error

 export PYTHONPATH="." AMD=1
 export MODEL="bert"
@@ -1,4 +1,5 @@
 #!/bin/bash
+set -e  # Exit on any error

 export PYTHONPATH="." NV=1
 export MODEL="resnet"
@@ -10,7 +11,7 @@ export RESET_STEP=0
 export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0

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

 export SEED=$RANDOM
 DATETIME=$(date "+%m%d%H%M")
@@ -8,6 +8,6 @@ export RESET_STEP=0

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

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

 python3 examples/mlperf/model_train.py
@@ -1,4 +1,5 @@
 #!/bin/bash
+set -e  # Exit on any error

 export PYTHONPATH="." AMD=1
 export MODEL="resnet"
@@ -10,13 +11,14 @@ 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=1
+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
@@ -1,4 +1,5 @@
 #!/bin/bash
+set -e  # Exit on any error

 export PYTHONPATH="." NV=1
 export MODEL="retinanet"
@@ -5,7 +5,7 @@
    "system_name": "tinybox 8xMI300X",
    "number_of_nodes": "1",
    "host_processors_per_node": "2",
-    "host_processor_model_name": "AMD EPYC 9354 32-Core Processor",
+    "host_processor_model_name": "AMD EPYC 9354",
    "host_processor_core_count": "32",
    "host_processor_vcpu_count": "64",
    "host_processor_frequency": "",
@@ -18,7 +18,7 @@
    "host_networking_topology": "",
    "host_memory_configuration": "24x 96GB DDR5",
    "accelerators_per_node": "8",
-    "accelerator_model_name": "AMD Instinct MI300X",
+    "accelerator_model_name": "AMD Instinct MI300X 192GB HBM3",
    "accelerator_host_interconnect": "PCIe 5.0 x16",
    "accelerator_frequency": "",
    "accelerator_on-chip_memories": "",
@@ -30,10 +30,9 @@
    "hw_notes": "",
    "framework": "tinygrad, branch mlperf_training_v5.0",
    "other_software_stack": {
-      "python": "3.10.16",
-      "ROCm": "3.0.0+94441cb"
+        "python": "3.10.16",
+        "ROCm": "3.0.0+94441cb"
    },
    "operating_system": "Ubuntu 24.04.1 LTS",
    "sw_notes": ""
-  }
-  
+  }
@@ -5,7 +5,7 @@
  "system_name": "tinybox green",
  "number_of_nodes": "1",
  "host_processors_per_node": "1",
-  "host_processor_model_name": "AMD EPYC 7532 32-Core Processor",
+  "host_processor_model_name": "AMD EPYC 7532",
  "host_processor_core_count": "32",
  "host_processor_vcpu_count": "64",
  "host_processor_frequency": "",
@@ -35,4 +35,4 @@
  },
  "operating_system": "Ubuntu 22.04.4",
  "sw_notes": ""
-}
+}
@@ -5,7 +5,7 @@
  "system_name": "tinybox red",
  "number_of_nodes": "1",
  "host_processors_per_node": "1",
-  "host_processor_model_name": "AMD EPYC 7532 32-Core Processor",
+  "host_processor_model_name": "AMD EPYC 7532",
  "host_processor_core_count": "32",
  "host_processor_vcpu_count": "64",
  "host_processor_frequency": "",
@@ -34,4 +34,4 @@
  },
  "operating_system": "Ubuntu 22.04.4",
  "sw_notes": ""
-}
+}
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
+
+export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+# export BEAM_LOG_SURPASS_MAX=1
+# export BASEDIR="/raid/datasets/wiki"
+
+export RESET_STEP=1
+export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,69 @@
+# 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
+```
@@ -0,0 +1,14 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+
+export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+
+export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+
+# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+export TRAIN_STEPS=3900
+
+export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -0,0 +1,29 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export SUBMISSION_PLATFORM="tinybox_8xMI300X"
+export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
+
+# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
+export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
+export TRAIN_STEPS=3900
+
+export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
+export BASEDIR="/raid/datasets/wiki"
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=1
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="bert_8xMI300x_${DATETIME}_${SEED}.log"
+
+# init  # TODO: without DEBUG=2 it hangs
+BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 DEBUG=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,69 @@
+# 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
+```
@@ -0,0 +1,16 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
+
+export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export 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
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
+
+export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/wiki"
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -0,0 +1,27 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." NV=1
+export MODEL="bert"
+export SUBMISSION_PLATFORM="tinybox_green"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
+
+export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export 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
@@ -0,0 +1,69 @@
+# 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
+```
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
+
+export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export 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
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
+
+export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/wiki"
+
+export WANDB=1 PARALLEL=0
+
+RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -0,0 +1,32 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." AMD=1
+export MODEL="bert"
+export SUBMISSION_PLATFORM="tinybox_red"
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
+
+export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/wiki"
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=1
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="bert_red_${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
+# TODO: AM driver resulted in nan
+sudo modprobe amdgpu
+PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,50 @@
+# 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
+```
@@ -0,0 +1,13 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+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
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+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
@@ -0,0 +1,25 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." NV=1
+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 $LOGFILE
@@ -0,0 +1,50 @@
+# 1. Problem
+
+This problem uses the ResNet-50 CNN to do image classification.
+
+## Requirements
+
+Install tinygrad and mlperf-logging from master.
+```
+git clone https://github.com/tinygrad/tinygrad.git
+python3 -m pip install -e ".[mlperf]"
+```
+
+### tinybox_green
+Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
+This is the default on production tinybox green.
+
+### tinybox_red
+Disable cwsr
+This is the default on production tinybox red.
+```
+sudo vi /etc/modprobe.d/amdgpu.conf
+cat <<EOF > /etc/modprobe.d/amdgpu.conf
+options amdgpu cwsr_enable=0
+EOF
+sudo update-initramfs -u
+sudo reboot
+
+# validate
+sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
+```
+
+# 2. Directions
+
+## Steps to download and verify data
+
+```
+IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
+```
+
+## Steps for one time setup
+
+### tinybox_red
+```
+examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
+```
+
+## Steps to run benchmark
+```
+examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
+```
@@ -0,0 +1,13 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="resnet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
+
+export RESET_STEP=0
+
+export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export BENCHMARK=10 DEBUG=${DEBUG:-2}
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+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
@@ -0,0 +1,26 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." AMD=1
+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
@@ -0,0 +1,8 @@
+#!/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
@@ -0,0 +1,38 @@
+# 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
+```
@@ -0,0 +1,14 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export BENCHMARK=5 DEBUG=2
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." NV=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export WANDB=1 PARALLEL=0
+export RUNMLPERF=1
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,25 @@
+#!/bin/bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="." NV=1
+export MODEL="retinanet"
+export SUBMISSION_PLATFORM="tinybox_green"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+
+export TRAIN_BEAM=2 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+export IGNORE_JIT_FIRST_BEAM=1
+export BASEDIR="/raid/datasets/openimages"
+
+# pip install -e ".[mlperf]"
+export LOGMLPERF=1
+
+export SEED=$RANDOM
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="retinanet_green_${DATETIME}_${SEED}.log"
+
+# init
+BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
+
+# run
+PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -0,0 +1,14 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export BENCHMARK=5 DEBUG=2
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,15 @@
+#!/bin/bash
+
+export PYTHONPATH="." AMD=1
+export MODEL="retinanet"
+export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
+export BASEDIR="/raid/datasets/openimages"
+
+# export RESET_STEP=0
+
+export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+
+export WANDB=1 PARALLEL=0
+export RUNMLPERF=1
+
+python examples/mlperf/model_train.py
@@ -0,0 +1,38 @@
+{
+    "submitter": "tinycorp",
+    "division": "closed",
+    "status": "Available on-premise",
+    "system_name": "tinybox 8xMI300X",
+    "number_of_nodes": "1",
+    "host_processors_per_node": "2",
+    "host_processor_model_name": "AMD EPYC 9354",
+    "host_processor_core_count": "32",
+    "host_processor_vcpu_count": "64",
+    "host_processor_frequency": "",
+    "host_processor_caches": "",
+    "host_processor_interconnect": "",
+    "host_memory_capacity": "2304GB",
+    "host_storage_type": "NVMe SSD",
+    "host_storage_capacity": "3x 4TB raid array",
+    "host_networking": "",
+    "host_networking_topology": "",
+    "host_memory_configuration": "24x 96GB DDR5",
+    "accelerators_per_node": "8",
+    "accelerator_model_name": "AMD Instinct MI300X 192GB HBM3",
+    "accelerator_host_interconnect": "PCIe 5.0 x16",
+    "accelerator_frequency": "",
+    "accelerator_on-chip_memories": "",
+    "accelerator_memory_configuration": "HBM3",
+    "accelerator_memory_capacity": "192GB",
+    "accelerator_interconnect": "",
+    "accelerator_interconnect_topology": "",
+    "cooling": "air",
+    "hw_notes": "",
+    "framework": "tinygrad, branch mlperf_training_v5.0",
+    "other_software_stack": {
+        "python": "3.10.16",
+        "ROCm": "3.0.0+94441cb"
+    },
+    "operating_system": "Ubuntu 24.04.1 LTS",
+    "sw_notes": ""
+  }
@@ -0,0 +1,38 @@
+{
+  "submitter": "tinycorp",
+  "division": "closed",
+  "status": "Available on-premise",
+  "system_name": "tinybox green",
+  "number_of_nodes": "1",
+  "host_processors_per_node": "1",
+  "host_processor_model_name": "AMD EPYC 7532",
+  "host_processor_core_count": "32",
+  "host_processor_vcpu_count": "64",
+  "host_processor_frequency": "",
+  "host_processor_caches": "",
+  "host_processor_interconnect": "",
+  "host_memory_capacity": "128GB",
+  "host_storage_type": "NVMe SSD",
+  "host_storage_capacity": "4 TB raid array + 1 TB boot",
+  "host_networking": "",
+  "host_networking_topology": "",
+  "host_memory_configuration": "8x 16GB DDR4",
+  "accelerators_per_node": "6",
+  "accelerator_model_name": "NVIDIA GeForce RTX 4090",
+  "accelerator_host_interconnect": "PCIe 4.0 x16",
+  "accelerator_frequency": "",
+  "accelerator_on-chip_memories": "",
+  "accelerator_memory_configuration": "GDDR6X",
+  "accelerator_memory_capacity": "24GB",
+  "accelerator_interconnect": "",
+  "accelerator_interconnect_topology": "",
+  "cooling": "air",
+  "hw_notes": "",
+  "framework": "tinygrad, branch mlperf_training_v5.0",
+  "other_software_stack": {
+    "python": "3.10.12",
+    "CUDA": "12.4"
+  },
+  "operating_system": "Ubuntu 22.04.4",
+  "sw_notes": ""
+}
@@ -0,0 +1,37 @@
+{
+  "submitter": "tinycorp",
+  "division": "closed",
+  "status": "Available on-premise",
+  "system_name": "tinybox red",
+  "number_of_nodes": "1",
+  "host_processors_per_node": "1",
+  "host_processor_model_name": "AMD EPYC 7532",
+  "host_processor_core_count": "32",
+  "host_processor_vcpu_count": "64",
+  "host_processor_frequency": "",
+  "host_processor_caches": "",
+  "host_processor_interconnect": "",
+  "host_memory_capacity": "128GB",
+  "host_storage_type": "NVMe SSD",
+  "host_storage_capacity": "4 TB raid array + 1 TB boot",
+  "host_networking": "",
+  "host_networking_topology": "",
+  "host_memory_configuration": "8x 16GB DDR4",
+  "accelerators_per_node": "6",
+  "accelerator_model_name": "AMD Radeon RX 7900 XTX",
+  "accelerator_host_interconnect": "PCIe 4.0 x16",
+  "accelerator_frequency": "",
+  "accelerator_on-chip_memories": "",
+  "accelerator_memory_configuration": "GDDR6",
+  "accelerator_memory_capacity": "24GB",
+  "accelerator_interconnect": "",
+  "accelerator_interconnect_topology": "",
+  "cooling": "air",
+  "hw_notes": "",
+  "framework": "tinygrad, branch mlperf_training_v5.0",
+  "other_software_stack": {
+    "python": "3.10.12"
+  },
+  "operating_system": "Ubuntu 22.04.4",
+  "sw_notes": ""
+}
@@ -12,13 +12,13 @@ from tinygrad.engine.realize import CompiledRunner

 import onnx
 from onnx.helper import tensor_dtype_to_np_dtype
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner, onnx_load

 OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
 OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"

 def compile(onnx_file):
-  onnx_model = onnx.load(onnx_file)
+  onnx_model = onnx_load(onnx_file)
  run_onnx = OnnxRunner(onnx_model)
  print("loaded model")

@@ -0,0 +1,49 @@
+import sys, onnx
+from tinygrad import Tensor, fetch, GlobalCounters
+from tinygrad.uop import UOp
+from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.engine.grouper import get_kernelize_map
+from tinygrad.engine.schedule import create_schedule_with_vars
+from tinygrad.engine.realize import run_schedule
+
+# NOLOCALS=1 GPU=1 IMAGE=2 FLOAT16=1 VIZ=1 DEBUG=2 python3 examples/openpilot/compile4.py
+
+OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
+OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
+
+if __name__ == "__main__":
+  fn = fetch(OPENPILOT_MODEL)
+  onnx_file = fetch(OPENPILOT_MODEL)
+  onnx_model = onnx.load(onnx_file)
+  run_onnx = OnnxRunner(onnx_model)
+
+  inputs = run_onnx.get_empty_input_data("npy")
+  out: Tensor = next(iter(run_onnx({k:v.to(None) for k,v in inputs.items()}).values())).to('cpu')
+  root = out.uop
+  targets = [x.uop for x in inputs.values()]
+  print(targets)
+
+  # TODO: abstract this from gradient?
+
+  # compute the target path (top down)
+  in_target_path: dict[UOp, bool] = {}
+  for u in root.toposort(): in_target_path[u] = any(x in targets or in_target_path[x] for x in u.src)
+  independent_set = {}
+  for u in root.toposort():
+    if in_target_path[u]:
+      for s in u.src:
+        if not in_target_path[s]:
+          independent_set[s] = None
+  independent = UOp.sink(*independent_set.keys())
+  kernelized = get_kernelize_map(independent)
+  independent = independent.substitute(kernelized)
+  schedule, var_vals, becomes_map = create_schedule_with_vars(independent)
+  run_schedule(schedule)
+
+  print("**** real ****")
+  GlobalCounters.reset()
+  out.uop = root.substitute(kernelized).substitute(becomes_map)
+  out.kernelize()
+
+  # realize
+  out.realize()
@@ -52,8 +52,6 @@ def load_model(model_path:Path, model_params:Dict[str, Union[int, float]]) -> Tr


 if __name__ == "__main__":
-  Tensor.no_grad = True
-
  parser = argparse.ArgumentParser(description="Run QwQ in tinygrad", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
  parser.add_argument("--size", choices=["32B"], default="32B", help="Model size")
  parser.add_argument("--count", type=int, default=30, help="Max number of tokens to generate")
@@ -68,7 +66,7 @@ if __name__ == "__main__":
  model_path = Path(args.weights) if args.weights else download_weights(model_info["total_num_weights"])
  transformer = load_model(model_path, model_info["model_params"])
  tokenizer = AutoTokenizer.from_pretrained(model_info["tokenizer"])
-  param_bytes = sum(x.lazydata.size * x.dtype.itemsize for x in get_parameters(transformer))
+  param_bytes = sum(x.uop.size * x.dtype.itemsize for x in get_parameters(transformer))

  outputted = args.prompt
  start_pos, toks = 0, tokenizer(outputted)["input_ids"]
@@ -5,6 +5,7 @@ from examples.stable_diffusion import AutoencoderKL, get_alphas_cumprod
 from examples.sdxl import DPMPP2MSampler, append_dims, LegacyDDPMDiscretization
 from extra.models.unet import UNetModel
 from extra.models.clip import FrozenOpenClipEmbedder
+from extra.bench_log import BenchEvent, WallTimeEvent

 from typing import Dict
 import argparse, tempfile, os
@@ -106,7 +107,6 @@ if __name__ == "__main__":
  assert args.width  % F == 0, f"img_width must be multiple of {F}, got {args.width}"
  assert args.height % F == 0, f"img_height must be multiple of {F}, got {args.height}"

-  Tensor.no_grad = True
  if args.seed is not None:
    Tensor.manual_seed(args.seed)

@@ -117,12 +117,14 @@ if __name__ == "__main__":
  if not weights_fn:
    weights_url = args.weights_url if args.weights_url else default_weights_url
    weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))
-  load_state_dict(model, safe_load(weights_fn), strict=False)

-  if args.fp16:
-    for k,v in get_state_dict(model).items():
-      if k.startswith("model"):
-        v.replace(v.cast(dtypes.float16).realize())
+  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+    load_state_dict(model, safe_load(weights_fn), strict=False)
+
+    if args.fp16:
+      for k,v in get_state_dict(model).items():
+        if k.startswith("model"):
+          v.replace(v.cast(dtypes.float16).realize())

  c  = { "crossattn": model.cond_stage_model(args.prompt) }
  uc = { "crossattn": model.cond_stage_model("") }
@@ -9,6 +9,7 @@ from tinygrad.nn.state import safe_load, load_state_dict, get_state_dict
 from tinygrad.helpers import fetch, trange, colored, Timing
 from extra.models.clip import Embedder, FrozenClosedClipEmbedder, FrozenOpenClipEmbedder
 from extra.models.unet import UNetModel, Upsample, Downsample, timestep_embedding
+from extra.bench_log import BenchEvent, WallTimeEvent
 from examples.stable_diffusion import ResnetBlock, Mid
 import numpy as np

@@ -346,17 +347,18 @@ class DPMPP2MSampler:
    for i in trange(num_sigmas - 1):
      with Timing("step in ", enabled=timing, on_exit=lambda _: f", using {GlobalCounters.mem_used/1e9:.2f} GB"):
        GlobalCounters.reset()
-        x, old_denoised = self.sampler_step(
-          old_denoised=old_denoised,
-          prev_sigma=(None if i==0 else sigmas[i-1].expand(x.shape[0])),
-          sigma=sigmas[i].expand(x.shape[0]),
-          next_sigma=sigmas[i+1].expand(x.shape[0]),
-          denoiser=denoiser,
-          x=x,
-          c=c,
-          uc=uc,
-        )
-        x.realize(old_denoised)
+        with WallTimeEvent(BenchEvent.STEP):
+          x, old_denoised = self.sampler_step(
+            old_denoised=old_denoised,
+            prev_sigma=(None if i==0 else sigmas[i-1].expand(x.shape[0])),
+            sigma=sigmas[i].expand(x.shape[0]),
+            next_sigma=sigmas[i+1].expand(x.shape[0]),
+            denoiser=denoiser,
+            x=x,
+            c=c,
+            uc=uc,
+          )
+          x.realize(old_denoised)

    return x

@@ -374,22 +376,24 @@ if __name__ == "__main__":
  parser.add_argument('--weights',  type=str,   help="Custom path to weights")
  parser.add_argument('--timing',   action='store_true', help="Print timing per step")
  parser.add_argument('--noshow',   action='store_true', help="Don't show the image")
+  parser.add_argument('--fakeweights',  action='store_true', help="Load fake weights")
  args = parser.parse_args()

-  Tensor.no_grad = True
  if args.seed is not None:
    Tensor.manual_seed(args.seed)

  model = SDXL(configs["SDXL_Base"])

-  default_weight_url = 'https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors'
-  weights = args.weights if args.weights else fetch(default_weight_url, 'sd_xl_base_1.0.safetensors')
-  loaded_weights = load_state_dict(model, safe_load(weights), strict=False, verbose=False, realize=False)
+  if not args.fakeweights:
+    default_weight_url = 'https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors'
+    weights = args.weights if args.weights else fetch(default_weight_url, 'sd_xl_base_1.0.safetensors')
+    loaded_weights = load_state_dict(model, safe_load(weights), strict=False, verbose=False, realize=False)

-  start_mem_used = GlobalCounters.mem_used
-  with Timing("loaded weights in ", lambda et_ns: f", {(B:=(GlobalCounters.mem_used-start_mem_used))/1e9:.2f} GB loaded at {B/et_ns:.2f} GB/s"):
-    Tensor.realize(*loaded_weights)
-    del loaded_weights
+    start_mem_used = GlobalCounters.mem_used
+    with Timing("loaded weights in ", lambda et_ns: f", {(B:=(GlobalCounters.mem_used-start_mem_used))/1e9:.2f} GB loaded at {B/et_ns:.2f} GB/s"):
+      with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+        Tensor.realize(*loaded_weights)
+      del loaded_weights

  N = 1
  C = 4
@@ -587,7 +587,7 @@ if __name__=="__main__":
  vits_model = args.model
  encoder_location, vits_location = ENCODER_MODELS[ENCODER_MODEL], VITS_MODELS[vits_model]

-  Tensor.no_grad, Tensor.training = True, False
+  Tensor.training = False
  # Get Synthesizer and ContentVec
  net_g, hps = Synthesizer.load_from_pretrained(vits_location[0], vits_location[2], vits_location[1], vits_location[3])
  Encoder = get_encoder(hps.model.ssl_dim)
@@ -14,6 +14,7 @@ from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import torch_load, load_state_dict, get_state_dict
 from extra.models.clip import Closed, Tokenizer
 from extra.models.unet import UNetModel
+from extra.bench_log import BenchEvent, WallTimeEvent

 class AttnBlock:
  def __init__(self, in_channels):
@@ -228,16 +229,16 @@ if __name__ == "__main__":
  parser.add_argument('--guidance', type=float, default=7.5, help="Prompt strength")
  args = parser.parse_args()

-  Tensor.no_grad = True
  model = StableDiffusion()

  # load in weights
-  load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], strict=False)
+  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], strict=False)

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

  # run through CLIP to get context
  tokenizer = Tokenizer.ClipTokenizer()
@@ -270,9 +271,10 @@ if __name__ == "__main__":
      GlobalCounters.reset()
      t.set_description("%3d %3d" % (index, timestep))
      with Timing("step in ", enabled=args.timing, on_exit=lambda _: f", using {GlobalCounters.mem_used/1e9:.2f} GB"):
-        tid = Tensor([index])
-        latent = run(model, unconditional_context, context, latent, Tensor([timestep]), alphas[tid], alphas_prev[tid], Tensor([args.guidance]))
-        if args.timing: Device[Device.DEFAULT].synchronize()
+        with WallTimeEvent(BenchEvent.STEP):
+          tid = Tensor([index])
+          latent = run(model, unconditional_context, context, latent, Tensor([timestep]), alphas[tid], alphas_prev[tid], Tensor([args.guidance]))
+          if args.timing: Device[Device.DEFAULT].synchronize()
    del run

  # upsample latent space to image with autoencoder
@@ -45,8 +45,7 @@ if __name__ == "__main__":
  print("*** scheduled training")

  # evaluate the model
-  with Tensor.test():
-    test_acc = ((model(X_test).argmax(axis=1) == Y_test).mean()*100)
+  test_acc = ((model(X_test).argmax(axis=1) == Y_test).mean()*100)
  print("*** scheduled eval")

  # NOTE: there's no kernels run in the scheduling phase
@@ -27,7 +27,7 @@ def imagenet_dataloader(cnt=0):
  input_std = Tensor([0.229, 0.224, 0.225]).reshape(1, -1, 1, 1)
  files = get_val_files()
  random.shuffle(files)
-  if cnt != 0: files = files[:cnt]
+  files = files[:cnt]
  cir = get_imagenet_categories()
  for fn in files:
    img = Image.open(fn)
@@ -66,7 +66,7 @@ if __name__ == "__main__":
  assert t_spec.shape[1:] == (3,224,224), f"shape is {t_spec.shape}"

  hit = 0
-  for i,(img,y) in enumerate(imagenet_dataloader(cnt=getenv("CNT", 100))):
+  for i,(img,y) in enumerate(imagenet_dataloader(cnt:=getenv("CNT", 100))):
    GlobalCounters.reset()
    p = run_onnx_jit(**{t_name:img})
    assert p.shape == (1,1000)
@@ -77,5 +77,6 @@ if __name__ == "__main__":
  MS_TARGET = 13.4
  print(f"need {GlobalCounters.global_ops/1e9*(1000/MS_TARGET):.2f} GFLOPS for {MS_TARGET:.2f} ms")

-  import pickle
-  with open("/tmp/im.pkl", "wb") as f: pickle.dump(run_onnx_jit, f)
+  if cnt >= 2:
+    import pickle
+    with open("/tmp/im.pkl", "wb") as f: pickle.dump(run_onnx_jit, f)
@@ -13,8 +13,8 @@ def prepare_browser_chunks(model):
  chunk_size = 16 * 1024 * 1024 # small chunks based on iphone browser constraints
  metadata = {}
  # We won't export cache_kv bytes (because we start inference on client at start_pos=0), but we will tell the client how big cache_kv needs to be
-  t_infos = [(v.lazydata.base.realized.nbytes, k, v.dtype) for k,v in state_dict.items() if "cache_kv" not in k]
-  empty_t_infos = [(v.lazydata.base.realized.nbytes, k, v.dtype) for k,v in state_dict.items() if "cache_kv" in k]
+  t_infos = [(v.uop.base.realized.nbytes, k, v.dtype) for k,v in state_dict.items() if "cache_kv" not in k]
+  empty_t_infos = [(v.uop.base.realized.nbytes, k, v.dtype) for k,v in state_dict.items() if "cache_kv" in k]

  split_t_infos = []
  for size, name, dtype in t_infos:
@@ -48,7 +48,7 @@ def prepare_browser_chunks(model):
        weight_metadata = metadata.get(name, default)
        weight_metadata["parts"][part_num] = {"file": i, "file_start_pos": cursor, "size": size}
        metadata[name] = weight_metadata
-        data = bytes(state_dict[name].lazydata.base.realized.as_buffer())
+        data = bytes(state_dict[name].uop.base.realized.as_buffer())
        data = data if not offsets else data[offsets[0]:offsets[1]]
        writer.write(data)
        cursor += size
@@ -109,7 +109,6 @@ if __name__=="__main__":
  tokenizer = Tokenizer(str(tokenizer_path))

  model_path = fetch("https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-f16.gguf", "Llama-3.2-1B-Instruct-f16.gguf", subdir="llama3-1b-instruct")
-  Tensor.no_grad = True
  max_context=1024
  tok = 128000
  TEMPERATURE, TOP_K, TOP_P, ALPHA_F, ALPHA_P = 0.95, 0, 0.0, 0.0, 0.0
@@ -651,7 +651,7 @@ class TextMapper: # Based on https://github.com/keithito/tacotron
 VITS_PATH = Path(__file__).parents[1] / "weights/VITS/"
 MODELS = { # config_url, weights_url
  "ljs": ("https://raw.githubusercontent.com/jaywalnut310/vits/main/configs/ljs_base.json", "https://drive.google.com/uc?export=download&id=1q86w74Ygw2hNzYP9cWkeClGT5X25PvBT&confirm=t"),
-  "vctk": ("https://raw.githubusercontent.com/jaywalnut310/vits/main/configs/vctk_base.json", "https://drive.google.com/uc?export=download&id=11aHOlhnxzjpdWDpsz1vFDCzbeEfoIxru&confirm=t"),
+  "vctk": ("https://huggingface.co/csukuangfj/vits-vctk/resolve/main/vctk_base.json", "https://huggingface.co/csukuangfj/vits-vctk/resolve/main/pretrained_vctk.pth"),
  "mmts-tts": ("https://huggingface.co/facebook/mms-tts/raw/main/full_models/eng/config.json", "https://huggingface.co/facebook/mms-tts/resolve/main/full_models/eng/G_100000.pth"),
  "uma_trilingual": ("https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/raw/main/configs/uma_trilingual.json", "https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer/resolve/main/pretrained_models/G_trilingual.pth"),
  "cjks": ("https://huggingface.co/spaces/skytnt/moe-tts/resolve/main/saved_model/14/config.json", "https://huggingface.co/spaces/skytnt/moe-tts/resolve/main/saved_model/14/model.pth"),
@@ -707,7 +707,6 @@ if __name__ == '__main__':
  text_mapper = TextMapper(apply_cleaners=True, symbols=symbols)

  # Load the model.
-  Tensor.no_grad = True
  if args.seed is not None:
    Tensor.manual_seed(args.seed)
    np.random.seed(args.seed)
@@ -82,7 +82,6 @@ if __name__ == "__main__":
  args = parser.parse_args()
  Device.DEFAULT = "WEBGPU"

-  Tensor.no_grad = True
  model = StableDiffusion()

  # load in weights
@@ -115,7 +114,7 @@ if __name__ == "__main__":
    run, special_names = jit_model(step, *step.input)
    functions, statements, bufs, _ = compile_net(run, special_names)
    state = get_state_dict(model)
-    weights = {id(x.lazydata.base.realized): name for name, x in state.items()}
+    weights = {id(x.uop.base.realized): name for name, x in state.items()}
    kernel_code = '\n\n'.join([f"const {key} = `{fixup_code(code, key)}`;" for key, code in functions.items()])
    kernel_names = ', '.join([name for (name, _, _, _) in statements])
    input_names = [name for _,name in special_names.items() if "input" in name]
@@ -94,7 +94,7 @@ class AudioEncoder:
 class TextDecoder:
  def __init__(self, n_vocab, n_text_ctx, n_text_state, n_text_head, n_text_layer, **_):
    self.max_tokens_to_sample = n_text_ctx // 2
-    self.max_self_attn_cache_len = self.max_tokens_to_sample * 2 + 5  # roughly prompt + start toks + max_tokens_to_sample
+    self.max_self_attn_cache_len = n_text_ctx

    self.token_embedding = nn.Embedding(n_vocab, n_text_state)
    self.positional_embedding = Tensor.empty(n_text_ctx, n_text_state)
@@ -104,7 +104,7 @@ class TextDecoder:
    self.getjitted = collections.defaultdict(lambda: TinyJit(self.forward))

  def __call__(self, x: Tensor, pos: int, encoded_audio: Tensor):
-    pos = Variable("self_attn_cache_len", 1, self.max_self_attn_cache_len).bind(pos) if pos else 0
+    pos = Variable("self_attn_cache_len", 1, self.max_self_attn_cache_len-1).bind(pos) if pos else 0
    return self.getjitted[x.shape](x, pos, encoded_audio)

  def forward(self, x:Tensor, pos:Union[Variable, Literal[0]], encoded_audio:Tensor):
@@ -1,9 +1,8 @@
 #!/usr/bin/env python3
 import os
 from ultralytics import YOLO
-import onnx
 from pathlib import Path
-from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner, onnx_load
 from extra.onnx_helpers import get_example_inputs
 from tinygrad.tensor import Tensor

@@ -11,6 +10,6 @@ os.chdir("/tmp")
 if not Path("yolov8n-seg.onnx").is_file():
  model = YOLO("yolov8n-seg.pt")
  model.export(format="onnx", imgsz=[480,640])
-onnx_model = onnx.load(open("yolov8n-seg.onnx", "rb"))
+onnx_model = onnx_load(open("yolov8n-seg.onnx", "rb"))
 run_onnx = OnnxRunner(onnx_model)
 run_onnx(get_example_inputs(run_onnx.graph_inputs), debug=True)
@@ -1,52 +0,0 @@
-We have to figure out how to make the tinygrad ops match to hw.
-Generic folded reduce may not work.
-
-
-GPUs:
-
-
-AMD:
-
-RDNA2: https://developer.amd.com/wp-content/resources/RDNA2_Shader_ISA_November2020.pdf
-
-We have RX6900XT with 80 CU, 40 WGP, and 1 "processor"
-@ 1.825 GHz, there's 18,688 FP32 GFLOPS of compute. 10240 FLOPS/cycle, 128 per CU (32 FMAs per vALU, 2 per compute unit)
-
-286 GFLOP for ENET=2 BS=64. At theoretical max, (286/18688)*1000 = 15.3 ms.
-We observe about 10x factor off with pytorch.
-
-We will focus on speed for AMD, since we have complete docs for that GPU.
-Each "processor" has an "ultra threaded dispatch processor"
-
-Each SIMD unit has 256 vector registers (or 1024?), and operates on 32 at once.
-Ahh, I think there's 1024 total, but only 256 per wavefront
-
-
-M1:
-
-On M1 GPU, theoretical is 2.275 TFLOPS. https://www.notebookcheck.net/Apple-M1-GPU-Benchmarks-and-Specs.503610.0.html
-
-We observe 2000ms for BS=8 (37 GFLOP). 37/2275 = 11.9 ms. tinygrad is over a factor of 100x off (similar on AMD GPU)
-
-NOTE: the timer in the M1 OpenCL doesn't seem to be anywhere close to wall time.
-
-
-Adreno:
-
-TBD, no comma three here. Image > Buffer because the L1 cache is used. Would UBWC help on weights?
-
-We have a good bit of work on this in hyperthneed. Let's get the disassembler out and make this fast.
-
-
-
-TPUs:
-
-These use really big systolic arrays and have a lot less flexibility.
-
-IIRC, their vector math unit is similar to the GPU.
-
-
-
-
-
-
@@ -1,5 +0,0 @@
-This is where we scope out adding accelerators to tinygrad
-
-ane -- Apple Neural Engine, in the M1 + newer iPhones
-tpu -- Google's TPU, available for rent in Google Cloud
-
@@ -1 +0,0 @@
-run
--- a/Show More
+++ b/Show More