Merge branch 'master' into remove_vec_3

viz: fix cfg for emulated amd on the null device (#15976 )
* simple failing when i test it end to end * pass * linter * assemble
2026-04-29 13:31:51 -07:00 · 2026-04-30 05:18:09 +09:00 · 2026-04-29 22:45:48 +03:00 · 2026-04-29 12:37:15 -07:00 · 2026-04-30 03:39:52 +09:00 · 2026-04-29 14:02:21 -04:00
589 changed files with 139378 additions and 144431 deletions
@@ -45,12 +45,16 @@ inputs:
    description: "Install mesa"
    required: false
    default: 'false'
+  tinydreno:
+    description: "Install tinydreno"
+    required: false
+    default: 'false'
 runs:
  using: "composite"
  steps:
    - name: Set up Python ${{ inputs.python-version }}
      id: setup-python
-      uses: actions/setup-python@v5
+      uses: actions/setup-python@v6
      with:
        python-version: ${{ inputs.python-version }}

@@ -62,14 +66,14 @@ runs:
      uses: actions/cache/restore@v4
      with:
        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
    - name: Cache Python packages
      if: github.event_name != 'pull_request'
      id: restore-venv
-      uses: actions/cache@v4
+      uses: actions/cache@v5
      with:
        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}

    # **** Caching downloads ****

@@ -81,7 +85,7 @@ runs:
        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
    - name: Cache downloads
      if: inputs.key != '' && github.event_name != 'pull_request'
-      uses: actions/cache@v4
+      uses: actions/cache@v5
      with:
        path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
@@ -195,13 +199,13 @@ runs:
      uses: actions/cache/restore@v4
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
    - name: Cache apt
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name != 'pull_request'
-      uses: actions/cache@v4
+      uses: actions/cache@v5
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}

    - name: Run apt Update + Install
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -221,22 +225,19 @@ runs:
      if: inputs.amd == 'true' && runner.os == 'Linux'
      shell: bash
      run: |
-        cargo build --release --manifest-path ./extra/remu/Cargo.toml
-        sudo ln -sf ${{ github.workspace }}/extra/remu/target/release/libremu.so /usr/local/lib/libremu.so
        sudo tee --append /etc/ld.so.conf.d/rocm.conf <<'EOF'
          /opt/rocm/lib
          /opt/rocm/lib64
        EOF
        sudo ldconfig
-    - name: Setup AMD comgr+remu (macOS)
+    - name: Setup AMD comgr (macOS)
      if: inputs.amd == 'true' && runner.os == 'macOS'
      shell: bash
      run: |
        sudo mkdir -p /usr/local/lib
-        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/amdcomgr_dylib/releases/latest | \
+        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/tinygrad/amdcomgr_dylib/releases/latest | \
          jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
          sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib
-        cargo build --release --manifest-path ./extra/remu/Cargo.toml

    # **** gpuocelot ****

@@ -265,7 +266,7 @@ runs:
    - name: Cache gpuocelot
      if: inputs.ocelot == 'true' && github.event_name != 'pull_request'
      id: cache-build
-      uses: actions/cache@v4
+      uses: actions/cache@v5
      env:
        cache-name: cache-gpuocelot-build-1
      with:
@@ -283,6 +284,7 @@ runs:

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

@@ -326,3 +328,9 @@ runs:
      if: inputs.mesa == 'true' && runner.os == 'macOS'
      shell: bash
      run: brew install sirhcm/tinymesa/tinymesa_cpu
+
+    # *** tinydreno ***
+    - name: Install tinydreno (linux)
+      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
@@ -28,10 +28,11 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
+        key: 'autogen'
        opencl: 'true'
        amd: 'true'
        cuda: 'true'
@@ -55,6 +56,9 @@ jobs:
        python3 -c "from tinygrad.runtime.autogen import libusb"
        python3 -c "from tinygrad.runtime.autogen import mesa"
        python3 -c "from tinygrad.runtime.autogen import avcodec"
+        python3 -c "from tinygrad.runtime.autogen import llvm_qcom"
+        python3 -c "from tinygrad.runtime.autogen import mlx5"
+        python3 -c "from tinygrad.runtime.autogen import ggml_common"
        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
    - name: Check for differences
      run: |
@@ -66,7 +70,7 @@ jobs:
        fi
    - name: Upload patch artifact
      if: failure()
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
      with:
        name: autogen-ubuntu-patch
        path: autogen-ubuntu.patch
@@ -77,10 +81,11 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
+        key: 'autogen-mac'
        llvm: 'true'
    - name: Regenerate autogen files
      run: |
@@ -96,7 +101,7 @@ jobs:
        fi
    - name: Upload patch artifact
      if: failure()
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
      with:
        name: autogen-macos-patch
        path: autogen-macos.patch
@@ -107,9 +112,11 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: 'autogen-comgr'
    - name: Install autogen support packages
      run: |
        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
@@ -133,7 +140,7 @@ jobs:
        fi
    - name: Upload patch artifact
      if: failure()
-      uses: actions/upload-artifact@v4
+      uses: actions/upload-artifact@v7
      with:
        name: autogen-comgr2-patch
        path: autogen-comgr2.patch
@@ -21,6 +21,9 @@ jobs:
  # the 3 minute timeout should not be raised
  testmacpytest:
    name: Mac pytest
+    env:
+      CI: ""
+      CAPTURE_PROCESS_REPLAY: "0"
    runs-on: [self-hosted, macOS]
    timeout-minutes: 3
    defaults:
@@ -29,7 +32,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    # brew install uv
    - name: setup python environment
      run: |
@@ -41,18 +44,37 @@ jobs:
      run: |
        echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
        rm -f /tmp/pytest-db-ci*
-    # TODO: remove this step once all old caches are migrated
-    - name: Migrate old huggingface cache (symlinks break onnxruntime 1.24+)
+    - name: Run pytest -nauto
      run: |
-        cd ~/Library/Caches/tinygrad/downloads/models 2>/dev/null || exit 0
-        for old_dir in models--*; do
-          [ -d "$old_dir" ] || continue
-          repo_id=$(echo "$old_dir" | sed 's/models--//; s/--/\//g')
-          snapshot=$(ls -1 "$old_dir/snapshots" 2>/dev/null | head -1)
-          [ -n "$snapshot" ] || continue
-          mkdir -p "$repo_id"
-          cp -RLn "$old_dir/snapshots/$snapshot/"* "$repo_id/" 2>/dev/null || true
-        done
+        source /tmp/tinygrad_pytest_ci/bin/activate
+        pytest -nauto --durations=20
+    - name: openpilot compile3 0.10.1 driving_vision
+      run: FLOAT16=1 DEV=CL IMAGE=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+
+  testframeworkpytest:
+    name: framework pytest
+    env:
+      CI: ""
+      CAPTURE_PROCESS_REPLAY: "0"
+    runs-on: [self-hosted, framework]
+    timeout-minutes: 10
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: setup python environment
+      run: |
+        rm -rf /tmp/tinygrad_pytest_ci
+        uv venv /tmp/tinygrad_pytest_ci
+        source /tmp/tinygrad_pytest_ci/bin/activate
+        uv pip install .[testing]
+    - name: setup staging db
+      run: |
+        echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
+        rm -f /tmp/pytest-db-ci*
    - name: Run pytest -nauto
      run: |
        source /tmp/tinygrad_pytest_ci/bin/activate
@@ -71,7 +93,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -101,17 +123,17 @@ jobs:
    - name: Run SDXL
      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing
    - name: Run model inference benchmark
-      run: METAL=1 NOCLANG=1 python3.11 test/external/external_model_benchmark.py
+      run: DEV=METAL NOCLANG=1 python3.11 test/external/external_model_benchmark.py
    - name: Test speed vs torch
      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
    - name: Test tensor cores
-      run: METAL=1 python3.11 test/opt/test_tensor_cores.py
+      run: DEV=METAL python3.11 test/opt/test_tensor_cores.py
    - name: Test AMX tensor cores
      run: |
-        DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_tensor_cores.py
-        DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_tensor_cores.py
-        DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
-        DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
+        DEBUG=2 DEV=CPU AMX=1 python3.11 test/opt/test_tensor_cores.py
+        DEBUG=2 DEV=CPU:LLVM AMX=1 python3.11 test/opt/test_tensor_cores.py
+        DEBUG=2 DEV=CPU AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
+        DEBUG=2 DEV=CPU:LLVM AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
    - name: Run Tensor Core GEMM (float)
      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (half)
@@ -119,7 +141,7 @@ jobs:
    - name: Run Tensor Core GEMM (bfloat16)
      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
    - name: Fuzz Padded Tensor Core GEMM
-      run: METAL=1 M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
+      run: DEV=METAL M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
    - name: Run LLaMA
      run: |
        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
@@ -161,7 +183,7 @@ jobs:
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
-    - uses: actions/upload-artifact@v4
+    - uses: actions/upload-artifact@v7
      with:
        name: Speed (Mac)
        path: |
@@ -181,7 +203,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -192,17 +214,19 @@ jobs:
        PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
        PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
    - name: UsbGPU boot time
-      run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
+      run: sudo -E PYTHONPATH=. GMMU=0 DEBUG=2 AM_RESET=1 DEV=USB+AMD time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU tiny tests
-      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
+      run: sudo -E PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/test_tiny.py
    - name: UsbGPU copy speeds
-      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+      run: sudo -E PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
    #- name: UsbGPU openpilot test
-    #  run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+    #  run: sudo -E PYTHONPATH=. GMMU=0 DEV=USB+AMD GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+    - name: UsbGPU (USB4/TB) install script
+      run: PYTHONPATH=. sh extra/setup_tinygpu_osx.sh
    - name: UsbGPU (USB4/TB) boot time
-      run: PYTHONPATH=. DEBUG=3 NV=1 NV_IFACE=PCI NV_NAK=1 time python3.11 test/test_tiny.py TestTiny.test_plus
+      run: PYTHONPATH=. DEBUG=3 DEV=PCI+NV:NAK time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU (USB4/TB) tiny tests
-      run: PYTHONPATH=. NV=1 NV_IFACE=PCI NV_NAK=1 python3.11 test/test_tiny.py
+      run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py

  testnvidiabenchmark:
    name: tinybox green Benchmark
@@ -214,7 +238,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Print nvidia-smi
      run: nvidia-smi
    - name: Symlink models and datasets
@@ -234,67 +258,67 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Run model inference benchmark
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
+      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
    - name: Test speed vs torch
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
+      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Test speed vs theoretical
-      run: NV=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      run: DEV=NV IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
    - name: Test benchmark allreduce
-      run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
+      run: DEV=NV python test/external/external_benchmark_multitensor_allreduce.py
    - name: Test tensor cores
      run: |
-        NV=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-        NV=1 NV_PTX=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+        DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+        DEV=NV:PTX ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
    - name: Run Tensor Core GEMM (CUDA)
      run: |
-        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        DEV=CUDA SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        DEV=CUDA SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        DEV=CUDA SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+        DEV=CUDA SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (PTX)
-      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+      run: DEV=NV:PTX SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (NV)
-      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Test NV=1
-      run: DEBUG=2 NV=1 python -m pytest -rA test/test_tiny.py
-    - name: Test CUDA=1
-      run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
+      run: DEV=NV SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+    - name: Test DEV=NV
+      run: DEBUG=2 DEV=NV python -m pytest -rA test/test_tiny.py
+    - name: Test DEV=CUDA
+      run: DEBUG=2 DEV=CUDA python -m pytest -rA test/test_tiny.py
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion DEV=NV python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
    # TODO: too slow
    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
+    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
    - name: Run LLaMA
      run: |
-        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama NV=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit DEV=NV JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama DEV=NV JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    - name: Run quantized LLaMA3
      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8
    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    # - name: Run LLaMA-2 70B
-    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit DEV=NV JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 DEV=NV JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half DEV=NV HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
-    - uses: actions/upload-artifact@v4
+      run: BENCHMARK_LOG=gpt2_half_beam DEV=NV HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+    - uses: actions/upload-artifact@v7
      with:
        name: Speed (NVIDIA)
        path: |
@@ -312,7 +336,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -332,35 +356,35 @@ jobs:
      run: test/external/process_replay/reset.py
    # TODO: too slow
    # - name: Fuzz Padded Tensor Core GEMM (NV)
-    #   run: NV=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    #   run: DEV=NV M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    # TODO: too slow
    # - name: Fuzz Padded Tensor Core GEMM (PTX)
-    #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+    #   run: DEV=NV:PTX M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    - name: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 DEV=NV PYTHONPATH=. python3 extra/hevc/decode.py
    - name: Train MNIST
-      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. DEV=NV TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 python3 examples/hlb_cifar10.py
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 DEV=NV WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar DEV=NV DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEV=NV DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    - name: Run MLPerf resnet eval on training data
-      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
+      run: time BENCHMARK_LOG=resnet_eval DEV=NV MODEL=resnet python3 examples/mlperf/model_eval.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps DEV=NV DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps_6gpu DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps_6gpu DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    - 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

@@ -374,7 +398,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setcap to python
      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove amd modules
@@ -416,18 +440,18 @@ jobs:
    #    python3 -c "import torch; print(torch.__version__)"
    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Test speed vs theoretical
-      run: AMD=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test tensor cores AMD_LLVM=0
-      run: AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
+      run: DEV=AMD IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+    - name: Test tensor cores (no LLVM)
+      run: DEV=AMD python3 test/opt/test_tensor_cores.py
    # TODO: this is flaky
-    # - name: Test tensor cores AMD_LLVM=1
-    #   run: AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
+    # - name: Test tensor cores AMD:LLVM
+    #   run: DEV=AMD:LLVM python3 test/opt/test_tensor_cores.py
    - name: Run Tensor Core GEMM (AMD)
      run: |
-        AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-    - name: Test AMD=1
-      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
+        DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+        DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+    - name: Test DEV=AMD
+      run: DEBUG=2 DEV=AMD 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
@@ -436,43 +460,43 @@ jobs:
    #- 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
+      run: time DEV=AMD 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
+      run: time DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 DEV=AMD python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing
+      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 DEV=AMD python3 examples/sdxl.py --seed 0 --noshow --timing
    - name: Run LLaMA 7B
      run: |
-        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama AMD=1 JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama_nojit DEV=AMD JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=llama DEV=AMD JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
    - name: Run LLaMA 7B with BEAM
-      run: BENCHMARK_LOG=llama_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=llama_beam DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    #- name: Restore amdgpu
    #  run: sudo modprobe amdgpu
    # - name: Run LLaMA-2 70B
-    #   run: AMD=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
+    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral AMD=1 python3 examples/mixtral.py --temperature 0 --count 10 --timing
+      run: time BENCHMARK_LOG=mixtral DEV=AMD python3 examples/mixtral.py --temperature 0 --count 10 --timing
    - name: Run GPT2
      run: |
-        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2_nojit DEV=AMD JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
+        BENCHMARK_LOG=gpt2 DEV=AMD JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half DEV=AMD HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
+      run: BENCHMARK_LOG=gpt2_half_beam DEV=AMD HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
    - 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

@@ -486,7 +510,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setcap to python
      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove amd modules
@@ -510,23 +534,26 @@ 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: Test GPU crash recovery
+      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
    - name: Train MNIST
-      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. DEV=AMD TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 DEV=AMD STEPS=10 python3 examples/hlb_cifar10.py
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 DEV=AMD STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
    # - name: Run 10 CIFAR training steps w BF16
-    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 DEV=AMD STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
    # TODO: too slow
    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 DEV=AMD WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar DEV=AMD DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Test full tinyfs load
-      run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
+      run: time BENCHMARK_LOG=cifar_6gpu DEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    # TODO: broken on some of the machines
+    #- name: Test full tinyfs load
+    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
    - 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

@@ -540,7 +567,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setcap to python
      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove amd modules
@@ -565,14 +592,14 @@ jobs:
    - 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
+      run: time BENCHMARK_LOG=resnet_eval DEV=AMD MODEL=resnet python3 examples/mlperf/model_eval.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps DEV=AMD DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps_6gpu DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps_6gpu DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    - 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

@@ -586,7 +613,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -594,33 +621,58 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: openpilot compile3 0.10.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.10.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 0.11.0 driving_vision
+      run: BENCHMARK_LOG=openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.11.0 driving_vision (from pickle)
+      run: BENCHMARK_LOG=openpilot_0_11_0_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM taskset -c 4-7 python3 examples/openpilot/compile3.py
+    - name: IR3 openpilot compile3 0.11.0 driving_vision
+      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.11.0 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.11.0 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_11_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
-      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: DEBUG=2 IMAGE=1 openpilot compile3 0.10.1 driving_vision
      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: IMAGE=1 openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=image_1_openpilot_0_10_1_vision PYTHONPATH="." DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
    - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
    - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: benchmark MobileNetV2 on DSP
      run: |
        # generate quantized weights
        ln -s /data/home/tiny/tinygrad/extra/datasets/imagenet extra/datasets/imagenet
        ln -s /data/home/tiny/tinygrad/testsig-*.so .
-        PYTHONPATH=. CC=clang-19 CPU=1 CPU_LLVM=0 QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
+        PYTHONPATH=. CC=clang-19 DEV=CPU QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
        # benchmark on DSP with NOOPT=1, the devectorizer has issues
-        PYTHONPATH=. CC=clang-19 DSP=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+        PYTHONPATH=. CC=clang-19 DEV=DSP NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

+  testcommausbgpubenchmark:
+    name: UsbGPU Benchmark (comma)
+    runs-on: [self-hosted, Linux, comma4]
+    timeout-minutes: 20
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: openpilot compile3 0.10.1 driving_vision
+      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." GMMU=0 DEV=USB+AMD:LLVM ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot load_pickle 0.10.1 driving_vision
+      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_load_pickle PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
+    - name: openpilot run_pickle 0.10.1 driving_vision
+      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py
+
  testreddriverbenchmark:
    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
@@ -631,7 +683,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setcap to python
      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove amd modules
@@ -656,32 +708,42 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Test driver cold start time
-      run: time DEBUG=3 AMD=1 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: time DEBUG=3 DEV=AMD AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Test driver warm start time
-      run: time DEBUG=3 AMD=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: time DEBUG=3 DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
+    - name: Test GPU crash recovery
+      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
    # Fails on 9070
    # - name: Test tensor cores
    #   run: |
-    #     AMD=1 AMD_LLVM=0 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     AMD=1 AMD_LLVM=1 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+    #     DEV=AMD python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+    #     DEV=AMD:LLVM python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+    #     DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (AMD)
-      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-    - name: Test AMD=1
-      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
+      run: DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+    - name: Test DEV=AMD
+      run: DEBUG=2 DEV=AMD python -m pytest -rA test/test_tiny.py
    - name: Test DISK copy time
-      run: AMD=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+      run: DEV=AMD TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
    - name: Test CPU copy time
      run: |
-        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
+        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
+        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar DEV=AMD DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    #   run: BENCHMARK_LOG=resnet_10steps DEV=AMD MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Remote
+      run: |
+        pkill -f 'extra/remote/serve.py' || true
+        PYTHONPATH=. python3 extra/remote/serve.py 6482 &
+        sleep 1
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD python3 test/test_tiny.py
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD AMD_AQL=1 python3 test/test_tiny.py
+        pkill -f 'extra/remote/serve.py' || true
    - 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

@@ -695,7 +757,7 @@ jobs:
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setcap to python
      run: ./extra/amdpci/setup_python_cap.sh
    - name: Remove nv modules
@@ -720,23 +782,30 @@ jobs:
    - name: reset process replay
      run: test/external/process_replay/reset.py
    - name: Test driver start time
-      run: time DEBUG=3 NV=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: time DEBUG=3 DEV=NV python3 test/test_tiny.py TestTiny.test_plus
    - name: Test tensor cores
-      run: NV=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+      run: DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
    - name: Test DISK copy time
-      run: NV=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+      run: DEV=NV TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
    - name: Test CPU copy time
      run: |
-        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
+        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
+        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Test LLAMA-3
-      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
+      run: BENCHMARK_LOG=llama3_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+      run: time BENCHMARK_LOG=cifar DEV=NV DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=resnet_10steps DEV=NV MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Remote
+      run: |
+        pkill -f 'extra/remote/serve.py' || true
+        PYTHONPATH=. python3 extra/remote/serve.py 6483 &
+        sleep 1
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6483 DEV=NV python3 test/test_tiny.py
+        pkill -f 'extra/remote/serve.py' || true
    - 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
@@ -14,7 +14,7 @@ jobs:

    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Remove amdgpu
      run: sudo rmmod amdgpu || true
    - name: Cleanup running AM processes
@@ -22,10 +22,10 @@ jobs:
    - name: Run SDXL with new search
      # TODO: GCVM_L2_PROTECTION_FAULT_STATUS with llvm19
      run: |
-        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CCACHE=0 python examples/sdxl.py --noshow --timing --seed 0
+        BENCHMARK_LOG=search_sdxl PYTHONPATH=. DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 CCACHE=0 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run SDXL with cached search
      run: |
-        BENCHMARK_LOG=search_sdxl_cached PYTHONPATH=. AMD=1 JITBEAM=2 python examples/sdxl.py --noshow --timing --seed 0
+        BENCHMARK_LOG=search_sdxl_cached PYTHONPATH=. DEV=AMD JITBEAM=2 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run winograd cifar with new search
      run: |
        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 IGNORE_BEAM_CACHE=1 CCACHE=0 BS=1024 STEPS=500 python examples/hlb_cifar10.py
@@ -10,16 +10,16 @@ jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
      - name: Configure Git Credentials
        run: |
          git config user.name github-actions[bot]
          git config user.email 41898282+github-actions[bot]@users.noreply.github.com
-      - uses: actions/setup-python@v5
+      - uses: actions/setup-python@v6
        with:
          python-version: 3.x
      - run: echo "cache_id=$(date --utc '+%V')" >> $GITHUB_ENV
-      - uses: actions/cache@v4
+      - uses: actions/cache@v5
        with:
          key: mkdocs-material-${{ env.cache_id }}
          path: .cache
@@ -16,7 +16,7 @@ jobs:

    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Cleanup running AM processes
      run: python extra/amdpci/am_smi.py --pids --kill
    - name: Symlink datasets
@@ -12,9 +12,9 @@ jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
    - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v6
      with:
        python-version: '3.x'
    - name: Install dependencies
@@ -15,7 +15,7 @@ jobs:
      branchstat: ${{ steps.brstat.outputs.stat}}
    steps:
      - name: Check code from PR branch 
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          repository: ${{ github.event.pull_request.head.repo.full_name }}
          ref: ${{ github.event.pull_request.head.sha }}
@@ -46,18 +46,18 @@ jobs:
    if: needs.checkbranch.outputs.branchstat == 'false'
    steps:
      - name: Checkout code from PR branch
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          repository: ${{ github.event.pull_request.head.repo.full_name }}
          ref: ${{ github.event.pull_request.head.sha }}
          path: pr
        # the base default to tinygrad master and cannot be other fork branch for security purpose
      - name: Checkout code from tinygrad master
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          path: base
      - name: Set up Python 3.12
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
        with:
          python-version: '3.12'
      - name: Count Line Diff
@@ -66,18 +66,16 @@ jobs:
          PR="$GITHUB_WORKSPACE/pr"
          pip install tabulate $BASE
          cp "$BASE/sz.py" .
-          echo "loc_content<<EOF" >> "$GITHUB_ENV"
-          python sz.py "$BASE" "$PR" >> "$GITHUB_ENV"
-          echo "EOF" >> "$GITHUB_ENV"
+          python sz.py "$BASE" "$PR" > loc_content.txt
      - name: Comment Code Line Diff
        continue-on-error: false
-        uses: marocchino/sticky-pull-request-comment@v2
+        uses: marocchino/sticky-pull-request-comment@v3
        with:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          ignore_empty: true
          skip_unchanged: true
          recreate: true
-          message: ${{ env.loc_content }}
+          path: loc_content.txt

  rebase:
    name: Core Library Line Difference
@@ -89,7 +87,7 @@ jobs:
    steps:
      - name: Comment Rebase
        continue-on-error: false
-        uses: marocchino/sticky-pull-request-comment@v2
+        uses: marocchino/sticky-pull-request-comment@v3
        with:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          skip_unchanged: true
@@ -1,7 +1,7 @@
 name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '16'
+  CACHE_VERSION: '19'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
@@ -21,7 +21,7 @@ jobs:
    timeout-minutes: 20
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -29,9 +29,9 @@ jobs:
        deps: testing_unit
        llvm: 'true'
    - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py

  docs:
    name: Docs
@@ -41,7 +41,7 @@ jobs:
      CHECK_OOB: 0
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -70,7 +70,7 @@ jobs:
        source venv/bin/activate
        pip install $GITHUB_WORKSPACE
        cp $GITHUB_WORKSPACE/examples/beautiful_mnist.py .
-        BS=2 STEPS=10 python beautiful_mnist.py
+        BS=2 STEPS=10 MAX_BUFFER_SIZE=0 python beautiful_mnist.py
    - name: Test Docs Build
      run: python -m mkdocs build --strict
    - name: Test Docs
@@ -83,7 +83,7 @@ jobs:
      run: DEBUG=100 python3 -c "from tinygrad import Tensor; N = 1024; a, b = Tensor.rand(N, N), Tensor.rand(N, N); c = (a.reshape(N, 1, N) * b.T.reshape(1, N, N)).sum(axis=2); print((c.numpy() - (a.numpy() @ b.numpy())).mean())"
    - name: Compile EfficientNet to C and test it
      run: |
-        CPU=1 CPU_LLVM=0 python examples/compile_efficientnet.py > recognize.c
+        DEV=CPU python examples/compile_efficientnet.py > recognize.c
        clang -O2 recognize.c -lm -o recognize
        cat test/models/efficientnet/Chicken.jpg | ./recognize | grep cock

@@ -93,7 +93,7 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -114,11 +114,11 @@ jobs:
    - name: Test one op in torch tests
      run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
    - name: Test Ops with TINY_BACKEND
-      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/backend/test_ops.py --durations=20
+      run: DEV=CPU:LLVM LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/backend/test_ops.py --durations=20
    - name: Test in-place operations on views
      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
    - name: Test multi-gpu
-      run: CPU=1 CPU_LLVM=1 GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
+      run: DEV=CPU:LLVM GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
    - name: Test kernel fusion
      run: python3 extra/torch_backend/test_kernel_fusion.py

@@ -129,7 +129,7 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -141,7 +141,7 @@ jobs:
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+      run: STEPS=20 DEV=CPU TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
    - name: Test some torch tests (expect failure)
      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true

@@ -151,67 +151,67 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: be-minimal
        deps: testing_unit
    - name: Test dtype with Python emulator
-      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py
+      run: DEBUG=1 DEV=PYTHON python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py
    - name: Test ops with Python emulator
-      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/backend/test_ops.py --durations=20
+      run: DEBUG=2 SKIP_SLOW_TEST=1 DEV=PYTHON python3 -m pytest -n=auto test/backend/test_ops.py --durations=20
    - name: Test uops with Python emulator
-      run: PYTHON=1 python3 -m pytest test/backend/test_uops.py --durations=20
+      run: DEV=PYTHON python3 -m pytest test/backend/test_uops.py --durations=20
    - name: Test symbolic with Python emulator
-      run: PYTHON=1 python3 test/backend/test_symbolic_ops.py
+      run: DEV=PYTHON python3 test/backend/test_symbolic_ops.py
    - name: test_renderer_failures with Python emulator
-      run: PYTHON=1 python3 -m pytest -rA test/backend/test_renderer_failures.py::TestRendererFailures
-    - name: Test IMAGE=2 support
+      run: DEV=PYTHON python3 -m pytest -rA test/backend/test_renderer_failures.py::TestRendererFailures
+    - name: Test IMAGE support
      run: |
-        IMAGE=2 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        IMAGE=2 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_simple_conv2d
+        IMAGE=1 DEV=PYTHON python3 test/backend/test_ops.py TestOps.test_gemm
+        IMAGE=1 DEV=PYTHON python3 test/backend/test_ops.py TestOps.test_simple_conv2d
    - name: Test emulated METAL tensor cores
      run: |
-        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_big_gemm
-        DEBUG=2 EMULATE=METAL FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::METAL python3 test/backend/test_ops.py TestOps.test_big_gemm
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::METAL python3 test/opt/test_tensor_cores.py
    - name: Test emulated AMX tensor cores
-      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
+      run: DEBUG=2 AMX=1 FORWARD_ONLY=1 DEV=PYTHON::AMX python3 test/backend/test_ops.py TestOps.test_gemm
    - name: Test emulated AMD tensor cores
      run: |
-        DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1100 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1100 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1100 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1100 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1100 python3 test/opt/test_tensor_cores.py
    - name: Test emulated AMD MFMA tensor cores
      run: |
-        DEBUG=2 EMULATE=AMD_MFMA FORWARD_ONLY=1 PYTHON=1 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD_MFMA FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx950 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx950 python3 test/opt/test_tensor_cores.py
    - name: Test emulated AMD RDNA4 tensor cores
      run: |
-        DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 EMULATE=AMD_RDNA4 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1201 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1201 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1201 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1201 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::gfx1201 python3 test/opt/test_tensor_cores.py
    - name: Test emulated CUDA tensor cores
      run: |
-        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::sm_80 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 ALLOW_TF32=1 FORWARD_ONLY=1 DEV=PYTHON::sm_80 python3 test/backend/test_ops.py TestOps.test_gemm
+        DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::sm_75 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
+        DEBUG=2 ALLOW_TF32=1 FORWARD_ONLY=1 DEV=PYTHON::sm_89 python3 test/opt/test_tensor_cores.py
    - name: Test emulated INTEL OpenCL tensor cores
-      run: DEBUG=2 EMULATE=INTEL FORWARD_ONLY=1 PYTHON=1 HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
+      run: DEBUG=2 FORWARD_ONLY=1 DEV=PYTHON::INTEL HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
    - name: Test emulated AMX tensor cores
-      run: DEBUG=2 AMX=1 EMULATE=AMX FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+      run: DEBUG=2 AMX=1 FORWARD_ONLY=1 DEV=PYTHON::AMX python3 test/opt/test_tensor_cores.py
    - name: Test device flop counts
      run: |
-        DEBUG=2 EMULATE=METAL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=AMD PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=CUDA PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/null/test_uops_stats.py TestUOpsStats.test_simple_matmul
+        DEBUG=2 DEV=PYTHON::METAL python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 DEV=PYTHON::gfx1100 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 DEV=PYTHON::sm_80 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 DEV=PYTHON::INTEL python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
+        DEBUG=2 AMX=1 DEV=PYTHON::AMX python3 ./test/null/test_uops_stats.py TestUOpsStats.test_simple_matmul

  linter:
    name: Linters
@@ -220,7 +220,7 @@ jobs:

    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -244,6 +244,37 @@ jobs:
    - name: Run TYPED=1
      run: CHECK_OOB=0 DEV=CPU TYPED=1 python test/test_tiny.py

+  nulltest:
+    name: Null Tests
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: unittest-13
+        pydeps: "pillow ftfy regex pre-commit"
+        deps: testing_unit
+        llvm: 'true'
+        amd: 'true'
+    - name: Run NULL backend tests
+      run: DEV=NULL python -m pytest -n=auto test/null/ --durations=20
+    - name: Run targeted tests on NULL backend
+      run: DEV=NULL python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+    # TODO: too slow
+    # - name: Run SDXL on NULL backend
+    #   run: DEV=NULL DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+    - name: Run Clip tests for SD MLPerf on NULL backend
+      run: DEV=NULL python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
+    - name: Run AMD emulated BERT training on NULL backend
+      run: DEV=NULL::gfx1201 NULL_ALLOW_COPYOUT=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    # TODO: support fake weights
+    #- name: Run LLaMA 7B on 4 fake devices
+    #  run: DEV=NULL python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
+
  unittest:
    name: Unit Tests
    runs-on: ubuntu-latest
@@ -251,7 +282,7 @@ jobs:

    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -266,22 +297,8 @@ jobs:
      run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
    - name: Run unit tests
      run: |
-        CPU=1 python test/null/test_device.py TestRunAsModule.test_module_runs
-        CPU=1 python -m pytest -n=auto test/unit/ --durations=20
-    - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
-    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
-    # TODO: too slow
-    # - name: Run SDXL on NULL backend
-    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
-    - name: Run Clip tests for SD MLPerf on NULL backend
-      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
-    - name: Run AMD emulated BERT training on NULL backend
-      run: EMULATE=AMD_RDNA4 NULL=1 NULL_ALLOW_COPYOUT=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    # TODO: support fake weights
-    #- name: Run LLaMA 7B on 4 fake devices
-    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
+        DEV=CPU python test/null/test_device.py TestRunAsModule.test_module_runs
+        DEV=CPU python -m pytest -n=auto test/unit/ --durations=20
    - name: Run GC tests
      run: python test/external/external_uop_gc.py
    - name: External Benchmark Schedule
@@ -308,7 +325,7 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -324,7 +341,7 @@ jobs:
    timeout-minutes: 10
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -345,18 +362,17 @@ jobs:
    timeout-minutes: 15
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: gpu-image
          deps: testing_unit
          opencl: 'true'
-      - name: Test CL IMAGE=2 ops
+      - name: Test CL IMAGE=1 ops
        run: |
-          CL=1 IMAGE=2 python -m pytest -n=auto test/backend/test_ops.py --durations=20
-        # TODO: training is broken
-        # CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
+          DEV=CL IMAGE=1 python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          DEV=CL IMAGE=1 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -366,7 +382,7 @@ jobs:
    timeout-minutes: 10
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -374,13 +390,13 @@ jobs:
          deps: testing
          opencl: 'true'
      - name: Generate Dataset
-        run: CL=1 extra/optimization/generate_dataset.sh
+        run: DEV=CL extra/optimization/generate_dataset.sh
      - name: Run Kernel Count Test
-        run: CL=1 python -m pytest -n=auto test/external/external_test_opt.py
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_opt.py
      - name: Run fused optimizer tests
-        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/backend/test_optim.py -k "not muon"
+        run: DEV=CL FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/backend/test_optim.py -k "not muon"
      - name: Upload artifact
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v7
        with:
          name: sops.gz
          path: /tmp/sops.gz
@@ -391,7 +407,7 @@ jobs:
    timeout-minutes: 15
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -401,13 +417,13 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1397 ALLOWED_GATED_READ_IMAGE=94 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1486 ALLOWED_GATED_READ_IMAGE=17 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp16
-        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+        run: FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp32 (test correctness)
-        run: DEBUGCL=1 CL=1 IMAGE=2 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
+        run: DEV=CL IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
      - name: Test openpilot LLVM compile fp16
-        run: FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+        run: IMAGE=1 FLOAT16=1 DEV=CPU:LLVM python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -420,7 +436,7 @@ jobs:

    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -429,15 +445,15 @@ jobs:
          python-version: '3.12'
          llvm: 'true'
      - name: Test ONNX (CPU)
-        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+        run: DEV=CPU python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      - name: Test ONNX (LLVM)
-        run: CPU=1 CPU_LLVM=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+        run: DEV=CPU:LLVM python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      - name: Test ONNX Runner (CPU)
-        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_runner.py
+        run: DEV=CPU python3 test/external/external_test_onnx_runner.py
      - name: Test Additional ONNX Ops (CPU)
-        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_ops.py
+        run: DEV=CPU python3 test/external/external_test_onnx_ops.py
      - name: Test Quantize ONNX
-        run: CPU=1 CPU_LLVM=0 python3 test/backend/test_quantize_onnx.py
+        run: DEV=CPU python3 test/backend/test_quantize_onnx.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -447,7 +463,7 @@ jobs:
    timeout-minutes: 20
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -457,21 +473,21 @@ jobs:
          python-version: '3.12'
          opencl: 'true'
      - name: Test ONNX (CL)
-        run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      #- name: Test Optimization Helpers
      #  run: DEBUG=1 python3 extra/optimization/test_helpers.py
      #- name: Test Action Space
-      #  run: DEBUG=1 CL=1 python3 extra/optimization/get_action_space.py
+      #  run: DEBUG=1 DEV=CL python3 extra/optimization/get_action_space.py
      - name: Test Beam Search
-        run: CL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
+        run: DEV=CL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
-        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
-      - name: NULL=1 beautiful_mnist_multigpu
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
+      - name: DEV=NULL beautiful_mnist_multigpu
+        run: DEV=NULL NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
      - name: Test Bert training
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+        run: DEV=NULL NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
-        run: NULL=1 NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: DEV=NULL NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -483,13 +499,20 @@ jobs:
      CHECK_OOB: 0
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: apps_llm
-      - name: Test 1B LLM
-        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm | grep -i rooster
+      - name: Test 1B LLM (llama)
+        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.llm --model llama3.2:1b | tee /dev/stderr | grep -i rooster
+      - name: Test 1B LLM (llama q4)
+        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.llm --model llama3.2:1b-q4 | tee /dev/stderr | grep -i rooster
+      - name: Test 1B LLM (qwen3.5)
+        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.llm --model qwen3.5:0.8b | tee /dev/stderr | grep -i rooster
+      - name: Test 1B LLM (qwen)
+        # NOTE: qwen is dumb and only knows about female chickens
+        run: echo "What's a female chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.llm --model qwen3:0.6b | tee /dev/stderr | grep -i hen

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

@@ -499,7 +522,7 @@ jobs:
    timeout-minutes: 15
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -508,11 +531,11 @@ jobs:
          opencl: 'true'
          llvm: 'true'
      - name: Test models (llvm)
-        run: CPU=1 CPU_LLVM=1 python -m pytest -n=auto test/models --durations=20
+        run: DEV=CPU:LLVM python -m pytest -n=auto test/models --durations=20
      - name: Test models (opencl)
-        run: CL=1 python -m pytest -n=auto test/models --durations=20
+        run: DEV=CL python -m pytest -n=auto test/models --durations=20
      - name: Test models (cpu)
-        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/models --durations=20
+        run: DEV=CPU python -m pytest -n=auto test/models --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -522,7 +545,7 @@ jobs:
    timeout-minutes: 20
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -530,9 +553,9 @@ jobs:
        deps: testing
        python-version: '3.12'
    - name: Test models (Metal)
-      run: METAL=1 python -m pytest -n=auto test/models --durations=20
+      run: DEV=METAL python -m pytest -n=auto test/models --durations=20
    - name: Test LLaMA compile speed
-      run: METAL=1 python test/external/external_test_speed_llama.py
+      run: DEV=METAL python test/external/external_test_speed_llama.py

 # ****** Feature Tests ******

@@ -542,7 +565,7 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -551,11 +574,11 @@ jobs:
        pydeps: "pillow"
        llvm: "true"
    - name: Test LLVM=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
+      run: DEV=CPU:LLVM DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
    - name: Test LLVM=1 DEVECTORIZE=0 for model
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
-    - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py
+      run: DEV=CPU:LLVM DEVECTORIZE=0 python3 test/models/test_efficientnet.py
+    - name: Test DEV=CPU DEVECTORIZE=0
+      run: DEV=CPU DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/backend/test_ops.py

  testdsp:
    name: Linux (DSP)
@@ -563,7 +586,7 @@ jobs:
    timeout-minutes: 15
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -572,9 +595,9 @@ jobs:
        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
        llvm: "true"
    - name: Set up Docker Buildx
-      uses: docker/setup-buildx-action@v3
+      uses: docker/setup-buildx-action@v4
    - name: Build QEMU Docker with cache
-      uses: docker/build-push-action@v4
+      uses: docker/build-push-action@v7
      with:
        file: extra/dsp/Dockerfile
        push: false
@@ -585,11 +608,11 @@ jobs:
    - 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
+      run: DEBUG=2 DEV=DSP python test/test_tiny.py
    - name: Test transcendentals
-      run: CC=clang-20 DEBUG=2 DSP=1 python test/backend/test_transcendental.py TestTranscendentalVectorized
+      run: CC=clang-20 DEBUG=2 DEV=DSP python test/backend/test_transcendental.py TestTranscendentalVectorized
    - name: Test quantize onnx
-      run: DEBUG=2 DSP=1 python3 test/backend/test_quantize_onnx.py
+      run: DEBUG=2 DEV=DSP python3 test/backend/test_quantize_onnx.py

  testwebgpu:
    name: Linux (WebGPU)
@@ -597,7 +620,7 @@ jobs:
    timeout-minutes: 20
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -607,11 +630,11 @@ jobs:
        webgpu: 'true'
    - name: Check Device.DEFAULT (WEBGPU) and print some source
      run: |
-        WEBGPU=1 python -c "from tinygrad import Device; assert Device.DEFAULT == 'WEBGPU', Device.DEFAULT"
-        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+        DEV=WEBGPU python -c "from tinygrad import Device; assert Device.DEFAULT == 'WEBGPU', Device.DEFAULT"
+        DEV=WEBGPU DEBUG=4 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Run selected webgpu tests
      run: |
-          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/backend --durations=20
+          DEV=WEBGPU WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/backend --durations=20
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -620,12 +643,10 @@ jobs:
    runs-on: ubuntu-24.04
    timeout-minutes: 20
    env:
-      AMD: 1
-      PYTHON_REMU: 1
-      MOCKGPU: 1
+      DEV: MOCKKFD+AMD
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -644,46 +665,72 @@ jobs:
          sudo apt-get update
          sudo apt-get install llvm-21 llvm-21-tools cloc
      - name: Install rocprof-trace-decoder
-        run: sudo PYTHONPATH="." ./extra/sqtt/install_sqtt_decoder.py
+        run: sudo PYTHONPATH="." ./extra/sqtt/install_rocprof_decoder.py
      - name: Run AMD renderer tests
-        run: AMD_LLVM=0 python -m pytest -n=auto test/amd/ --durations 20
-      - name: Run AMD renderer tests (AMD_LLVM=1)
-        run: AMD_LLVM=1 python -m pytest -n=auto test/amd/ --durations 20
-      - name: Run TestOps.test_add with SQTT
-        run: |
-          VIZ=-2 DEBUG=5 python3 test/backend/test_ops.py TestOps.test_add
-          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
+        run: python -m pytest -n=auto test/amd/ --durations 20
+      - name: Run AMD renderer tests (AMD:LLVM)
+        run: DEV=MOCKKFD+AMD:LLVM python -m pytest -n=auto test/amd/ --durations 20
+      - name: Run SQTT profiling tests
+        run: PROFILE=1 SQTT=1 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
      - name: Run AMD emulated tests on NULL backend
        env:
          AMD: 0
        run: |
-          PYTHONPATH=. NULL=1 EMULATE=AMD python extra/mmapeak/mmapeak.py
-          PYTHONPATH=. NULL=1 EMULATE=AMD_CDNA4 python3 -m pytest -n=auto test/testextra/test_tk.py test/backend/test_asm_gemm.py
-      - name: Run ASM matmul on MOCKGPU
-        run: PYTHONPATH="." AMD=1 MOCKGPU=1 N=256 python3 extra/gemm/amd_asm_matmul.py
+          PYTHONPATH=. DEV=NULL:HIP:gfx1100 python extra/mmapeak/mmapeak.py
+          PYTHONPATH=. DEV=NULL:HIP:gfx950 python3 -m pytest -n=auto test/testextra/test_tk.py test/backend/test_asm_gemm.py
+      - name: Run matmul on MOCKKFD
+        run: |
+          PYTHONPATH="." DEV=MOCKKFD+AMD N=256 python3 extra/gemm/amd_asm_matmul.py
+          PYTHONPATH="." DEV=MOCKKFD+AMD N=256 python3 extra/gemm/amd_copy_matmul.py
      - name: Run LLVM test
-        run: AMD_LLVM=1 python test/device/test_amd_llvm.py
+        run: DEV=MOCKKFD+AMD:LLVM python test/device/test_amd_llvm.py
+
+  testmockam:
+    name: Linux (am)
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    env:
+      DEV: MOCKPCI+AMD
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v6
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: mockam
+          deps: testing_unit
+          amd: 'true'
+      - name: Run test_tiny on MOCKAM
+        run: python test/test_tiny.py
+      - name: Run test_tiny on MOCKUSB
+        run: GMMU=0 DEV=MOCKUSB+AMD python test/test_tiny.py
+      - name: Run test_hcq on MOCKPCI
+        run: python -m pytest test/device/test_hcq.py
+      - name: Run disk copy tests on MOCKPCI
+        run: python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk
+      - name: Run test_tiny on MOCKPCI Remote
+        run: |
+          python extra/remote/serve.py 6667 &
+          sleep 2
+          REMOTE=127.0.0.1:6667 python test/test_tiny.py
+          REMOTE=127.0.0.1:6667 python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk; kill %1

  testamd:
    strategy:
      fail-fast: false
      matrix:
        backend: [amd, amdllvm]
-        arch: [rdna3, rdna4]
-        #arch: [rdna3, rdna4, cdna4]
+        arch: [gfx1100, gfx1201, gfx950]

    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
    runs-on: ubuntu-22.04
    timeout-minutes: 15
    env:
-      AMD: 1
-      MOCKGPU: 1
-      MOCKGPU_ARCH: ${{ matrix.arch }}
+      DEV: MOCKKFD+AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
      SKIP_SLOW_TEST: 1
-      AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -696,7 +743,9 @@ jobs:
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
      - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/testextra/test_cfg_viz.py test/external/external_test_am.py --durations=20
+        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
+      - name: Run disk copy tests
+        run: python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk
      - name: Run TRANSCENDENTAL math
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
      - name: Run process replay tests
@@ -712,11 +761,10 @@ jobs:
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    env:
-      MOCKGPU: 1
      FORWARD_ONLY: 1
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -725,7 +773,7 @@ jobs:
          cuda: 'true'
          ocelot: 'true'
      - name: Set env
-        run: printf "${{ matrix.backend == 'PTX' && 'CUDA=1\nCUDA_PTX=1' || matrix.backend == 'nv' && 'NV=1\nSKIP_SLOW_TEST=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'ptx' && 'DEV=MOCK+CUDA:PTX' || matrix.backend == 'nv' && 'DEV=MOCK+NV\nSKIP_SLOW_TEST=1' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CUDA','NV'], Device.DEFAULT"
@@ -749,7 +797,7 @@ jobs:
    timeout-minutes: 20
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -759,7 +807,7 @@ jobs:
          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'DEV=CPU:LLVM' || matrix.backend == 'cpu' && 'DEV=CPU\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'DEV=CL' || matrix.backend == 'lvp' && 'DEV=CPU:LVP' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
@@ -779,7 +827,7 @@ jobs:
    timeout-minutes: 20
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -791,42 +839,38 @@ jobs:
        ocelot: 'true'
        llvm: 'true'
    - name: Run unit tests
-      run: METAL=1 python -m pytest -n=auto test/unit/ --durations=20
+      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
    - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
+      run: DEV=NULL python -m pytest -n=auto test/null/ --durations=20
    - name: Run ONNX
-      run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+      run: DEV=METAL python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
    - name: Test tensor core ops (fake)
-      run: METAL=1 DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
+      run: DEV=METAL DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
    - name: Test tensor core ops (real)
-      run: METAL=1 DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
+      run: DEV=METAL DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
    - name: Test Beam Search
-      run: METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
+      run: DEV=METAL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
+    - name: Test Device Specific
+      run: DEV=METAL python3 -m pytest test/device/test_metal.py
    #- name: Fuzz Test linearizer
-    #  run: METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
+    #  run: DEV=METAL DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
    - name: Run TRANSCENDENTAL math
-      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
+      run: DEV=METAL TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
    - name: Run pytest (amd)
      env:
-        MOCKGPU: 1
-        AMD: 1
-        AMD_LLVM: 0
+        DEV: MOCKKFD+AMD
        FORWARD_ONLY: 1
      run: |
        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
    - name: Run pytest (amd with llvm backend)
      env:
-        MOCKGPU: 1
-        AMD: 1
-        AMD_LLVM: 1
+        DEV: "MOCKKFD+AMD:LLVM"
        FORWARD_ONLY: 1
      run: |
        python -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py test/device/test_amd_llvm.py --durations=20
    - name: Run pytest (ptx)
      env:
-        MOCKGPU: 1
-        NV_PTX: 1
-        NV: 1
+        DEV: "MOCK+NV:PTX"
        FORWARD_ONLY: 1
        # TODO: failing due to library loading error
        CAPTURE_PROCESS_REPLAY: 0
@@ -841,7 +885,7 @@ jobs:
    timeout-minutes: 10
    steps:
    - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
@@ -849,9 +893,9 @@ jobs:
        deps: testing
        webgpu: 'true'
    - name: Build WEBGPU Efficientnet
-      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m examples.compile_efficientnet
+      run: DEV=WEBGPU WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m examples.compile_efficientnet
    - name: Run selected webgpu tests
-      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m pytest -n=auto test/backend --durations=20
+      run: DEV=WEBGPU WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m pytest -n=auto test/backend --durations=20
    #- name: Clean npm cache
    #  run: npm cache clean --force
    #- name: Install Puppeteer
@@ -870,7 +914,7 @@ jobs:
    #    cp $GITHUB_WORKSPACE/test/web/test_viz.js .
    #    node test_viz.js
    - name: Test ONNX Runner (WEBGPU)
-      run: WEBGPU=1 python3 test/external/external_test_onnx_runner.py
+      run: DEV=WEBGPU python3 test/external/external_test_onnx_runner.py

  osxtests:
    strategy:
@@ -882,7 +926,7 @@ jobs:
    timeout-minutes: 20
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -891,7 +935,7 @@ jobs:
          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'DEV=CPU:LLVM' || matrix.backend == 'cpu' && 'DEV=CPU\nCPU_COUNT=2' || matrix.backend == 'metal' && 'DEV=METAL' || matrix.backend == 'lvp' && 'DEV=CPU:LVP' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
@@ -901,8 +945,8 @@ jobs:
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
      - name: Run macOS-specific unit test
-        if: matrix.backend == 'cpu'
-        run: python3 -m pytest test/unit/test_disk_tensor.py::TestDiskTensor::test_copy_to_cpu_not_truncated
+        if: matrix.backend == 'llvm'
+        run: python3 -m pytest test/unit/test_disk_tensor.py::TestDiskTensor::test_copy_to_cpu_not_truncated test/unit/test_cpu.py

 # ****** Windows Tests ******

@@ -917,7 +961,7 @@ jobs:
    timeout-minutes: 15
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -926,7 +970,7 @@ jobs:
          pydeps: ${{ matrix.backend == 'webgpu' && 'dawn-python' || '' }}
      - name: Set env
        shell: bash
-        run:  printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'webgpu' && 'WEBGPU=1'}}" >> $GITHUB_ENV
+        run:  printf "${{ matrix.backend == 'llvm' && 'DEV=CPU:LLVM' || matrix.backend == 'cpu' && 'DEV=CPU\nCPU_COUNT=2' || matrix.backend == 'webgpu' && 'DEV=WEBGPU'}}" >> $GITHUB_ENV
      - name: Run unit tests
        if: matrix.backend=='llvm'
        # test_newton_schulz hits RecursionError
@@ -934,7 +978,7 @@ jobs:
      - name: Run NULL backend tests
        if: matrix.backend=='llvm'
        shell: bash
-        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
+        run: DEV=NULL python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
      - name: Run pytest (${{ matrix.backend }})
        shell: bash
        run: |
@@ -953,7 +997,7 @@ jobs:
    timeout-minutes: 15
    steps:
      - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
@@ -963,7 +1007,30 @@ jobs:
          python-version: '3.12'
      - name: Set env
        shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
+        run: printf "NULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'DEV=NULL:IR3:a630' || matrix.backend == 'nak' && 'DEV=NULL:NAK:sm_120' }}" >> $GITHUB_ENV
+      - name: Run test_ops
+        shell: bash
+        run: |
+          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
+          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+  qcomclcompiletests:
+    name: Compile-only (QCOM CL)
+    runs-on: ubuntu-24.04-arm
+    timeout-minutes: 15
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v6
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: compile-qcomcl
+          deps: testing_unit
+          tinydreno: 'true'
+          python-version: '3.12'
+      - name: Set env
+        shell: bash
+        run: printf "DEV=NULL:QCOMCL:a630\nNULL_ALLOW_COPYOUT=1" >> $GITHUB_ENV
      - name: Run test_ops
        shell: bash
        run: |
@@ -66,3 +66,6 @@ target
 .mypy_cache
 mutants
 .mutmut-cache
+dagre/
+graphlib/
+uv.lock
@@ -1,6 +1,4 @@
 # abstractions2 goes from back to front, here we will go from front to back
-from typing import List
-from tinygrad.helpers import tqdm

 # *****
 # 0. Load mnist on the device
@@ -33,21 +31,21 @@ model(X).sparse_categorical_crossentropy(Y).backward()
 optim.schedule_step()   # this will step the optimizer without running realize

 # *****
-# 3. Create a schedule.
+# 3. Create a schedule (linear uop).

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

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

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

 # *****
-# 4. Lower and run the schedule.
+# 4. Lower and run the schedule (linear uop).

-for si in tqdm(schedule): si.run()
+run_linear(linear)

 # *****
 # 5. Print the weight change
@@ -0,0 +1,253 @@
+# tinygrad allows you to write kernels at many different abstractions levels.
+# This is for RDNA3, but if you don't have one you can run with the emulator
+# PYTHONPATH="." DEV=MOCKPCI+AMD
+
+from tinygrad import Tensor, Context, GlobalCounters, UOp, Device
+from tinygrad.helpers import DEV, DEBUG, getenv
+from tinygrad.uop.ops import AxisType, KernelInfo, Ops
+from tinygrad.dtype import AddrSpace, dtypes
+from tinygrad.runtime.autogen.amd.rdna3.ins import *
+
+def eval_harness(name, tensor, fxn, check=None):
+  print(f"***** {name}")
+  GlobalCounters.reset()
+  with Context(DEBUG=max(DEBUG.value, 2)): out = fxn(tensor).item()
+  assert check is None or abs(out - check) < abs(check) * 1e-3, f"out was wrong {out}, expected {check}, off by {out/check}x"
+  print(f"computed in {GlobalCounters.time_sum_s*1000:.2f} ms, {(a.nbytes()/1e9)/GlobalCounters.time_sum_s:.2f} GB/s")
+  return out
+
+SZ = 256*1024 if DEV.interface.startswith("MOCK") else 1024*1024*1024
+
+def example_2_hip(a:Tensor, correct):
+  GLOBALS = 1024
+  THREADS = 256
+  def hip_reduce_sum(out:UOp, buf:UOp) -> UOp:
+    assert SZ % (GLOBALS * THREADS) == 0
+    CHUNK = SZ // (GLOBALS * THREADS)
+    # NOTE: tinygrad doesn't populate HIP hidden kernargs, so blockDim.x/gridDim.x read as 0.
+    # We hardcode block/grid sizes as constexpr to avoid any dependency on those builtins.
+    code = f"""
+    #include <hip/hip_runtime.h>
+    constexpr unsigned int BLOCK = {THREADS};
+    constexpr unsigned int CHUNK = {CHUNK};
+    extern "C" __global__ void hip_reduce_sum_kernel(float* __restrict__ block_sums, const float* __restrict__ x) {{
+      __shared__ float sdata[BLOCK];
+
+      unsigned int tid = threadIdx.x;
+      unsigned int gid = blockIdx.x * BLOCK + tid;
+
+      // Each thread sums CHUNK consecutive elements from its own region
+      float sum = 0.0f;
+      const float* base = x + gid * CHUNK;
+      #pragma unroll 16
+      for (unsigned int k = 0; k < CHUNK; k++) {{
+        sum += base[k];
+      }}
+
+      sdata[tid] = sum;
+      __syncthreads();
+
+      // Block reduction in shared memory
+      for (unsigned int s = BLOCK / 2; s > 0; s >>= 1) {{
+        if (tid < s) {{
+          sdata[tid] += sdata[tid + s];
+        }}
+        __syncthreads();
+      }}
+
+      // One partial sum per block
+      if (tid == 0) {{
+        block_sums[blockIdx.x] = sdata[0];
+      }}
+    }}"""
+
+    # TODO: remove the need for the compiler here, you should just be able to remove Ops.BINARY
+    from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
+    lib = HIPCCCompiler(Device[Device.DEFAULT].renderer.target.arch, []).compile_cached(code)
+    # the sink specifies the GLOBAL and LOCAL sizes, along with the input buffers and name
+    sink = UOp.sink(UOp.special(GLOBALS, 'gidx0'), UOp.special(THREADS, 'lidx0'), out, buf,
+                    arg=KernelInfo(name="hip_reduce_sum_kernel"))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
+                UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+  eval_harness("HIP kernel", a, lambda x: Tensor.empty(GLOBALS).custom_kernel(x, fxn=hip_reduce_sum)[0].sum(), check=correct)
+
+def example_3_custom_uop(a:Tensor, correct):
+  # This GPU has 32 CUs, keep them all busy
+  CU_COUNT = 32
+  def custom_sum(out:UOp, buf:UOp) -> UOp:
+    LCLS = 256
+    buf = buf.reshape(CU_COUNT, -1, LCLS)
+
+    glbl = UOp.range(CU_COUNT, 0, AxisType.GLOBAL)
+    lane = UOp.range(LCLS, 1, AxisType.LOCAL)
+
+    # accumulate the globals into a per lane accumulator
+    reduce_loop = UOp.range(buf.shape[1], 2, AxisType.REDUCE)
+    acc = UOp.placeholder((1,), dtypes.float, slot=6, addrspace=AddrSpace.REG)
+    acc = acc.after(acc.store(0))
+    acc = acc.after(acc[0].store(acc.after(reduce_loop)[0] + buf[glbl, reduce_loop, lane]).end(reduce_loop))
+
+    # store all the per lane accumulators to LOCAL
+    local_accs = UOp.placeholder((LCLS,), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL)
+    local_accs = local_accs.after(local_accs[lane].store(acc[0]).barrier())
+
+    # accumulate LOCALs into a single per CU accumulator
+    late_reduce_loop = UOp.range(LCLS, 3, AxisType.REDUCE)
+    acc2 = UOp.placeholder((1,), dtypes.float, slot=7, addrspace=AddrSpace.REG)
+    acc2 = acc2.after(acc2.store(0))
+    acc2 = acc2.after(acc2[0].store(acc2.after(late_reduce_loop)[0] + local_accs[late_reduce_loop]).end(late_reduce_loop))[0]
+
+    # store (NOTE: since the address doesn't depend on the warp, this will be automatically gated)
+    return out[glbl].store(acc2).end(lane, glbl).sink(arg=KernelInfo(opts_to_apply=()))
+
+  eval_harness("custom UOp kernel", a, lambda x: Tensor.empty(CU_COUNT).custom_kernel(x, fxn=custom_sum)[0].sum(), check=correct)
+
+def example_5_custom_assembly(a:Tensor, correct):
+  # Kernel class copied from amd_asm_matmul
+  class Kernel:
+    def __init__(self): self.instructions, self.labels, self.pos = [], {}, 0
+    def label(self, name): self.labels[name] = self.pos
+    def emit(self, inst, target=None):
+      self.instructions.append(inst)
+      inst._target, inst._pos = target, self.pos
+      self.pos += inst.size()
+      return inst
+    def waitcnt(self, lgkm=None, vm=None):
+      # Wait for memory operations. lgkm=N waits until N lgkm ops remain, vm=N waits until N vmem ops remain.
+      vmcnt, lgkmcnt, expcnt = vm if vm is not None else 63, lgkm if lgkm is not None else 63, 7
+      waitcnt = (expcnt & 0x7) | ((lgkmcnt & 0x3f) << 4) | ((vmcnt & 0x3f) << 10)
+      self.emit(s_waitcnt(simm16=waitcnt))
+    def finalize(self, sink:UOp) -> UOp:
+      for inst in self.instructions:
+        if inst._target is None: continue
+        offset_dwords = (self.labels[inst._target] - inst._pos - inst.size()) // 4
+        if not -32768 <= offset_dwords <= 32767: raise ValueError(f"branch to '{inst._target}' offset {offset_dwords} exceeds simm16 range")
+        inst.simm16 = offset_dwords
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
+                                   UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in self.instructions]))))
+
+  CU_COUNT = 32
+  LANES = 64
+  def asm_sum(out:UOp, buf:UOp) -> UOp:
+    V_LANE_ID = 0             # lane_id set on startup
+    S_WORKGROUP_X = 2         # workgroup_id_x
+    S_LOOP_CTR = 3
+    k = Kernel()
+    # mul lane id by 16 for offsets (4 for float, 4 for b128)
+    k.emit(v_mul_lo_u32(v[0], v[V_LANE_ID], 16))
+    k.emit(v_add_nc_u32_e32(v[1], 4096, v[0]))
+    k.emit(v_add_nc_u32_e32(v[2], 4096, v[1]))
+    k.emit(v_add_nc_u32_e32(v[3], 4096, v[2]))
+    # load both addresses
+    k.emit(s_load_b128(sdata=s[4:7], sbase=s[0:1], offset=0x0, soffset=NULL))
+    k.waitcnt(lgkm=0)
+    # offset buffer pointer by workgroup_id_x * chunk_size_bytes
+    k.emit(s_mul_i32(s[S_LOOP_CTR], s[S_WORKGROUP_X], buf.numel()*4//CU_COUNT))
+    k.emit(s_add_u32(s[6], s[6], s[S_LOOP_CTR]))
+    k.emit(s_addc_u32(s[7], s[7], 0))
+    # zero the accumulators
+    k.emit(VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, vdstx=v[4], vdsty=v[5], srcx0=0, srcy0=0))
+    k.emit(VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, vdstx=v[6], vdsty=v[7], srcx0=0, srcy0=0))
+
+    def emit_loads(base_vreg, reg_len):
+      assert reg_len%4 == 0
+      k.emit(s_clause(simm16=(reg_len//4)-1))
+      for i in range(reg_len//4):
+        offset = i*LANES*16
+        assert offset < 16384
+        k.emit(global_load_b128(vdst=v[base_vreg+i*4:base_vreg+i*4+3], addr=v[offset//4096], saddr=s[6:7], offset=offset%4096))
+      k.emit(s_add_u32(s[6], s[6], reg_len * LANES * 4))
+      k.emit(s_addc_u32(s[7], s[7], 0))
+
+    def tree_reduce_to_4567(base_vreg, reg_len):
+      assert reg_len%4 == 0
+      reg_len //= 4
+      while reg_len > 1:
+        half = reg_len // 2
+        for j in range(half):
+          a, b = base_vreg + j*4, base_vreg + (j+half)*4
+          # v[a+0](bank0) += v[b+2](bank2), v[a+1](bank1) += v[b+3](bank3) — src0 and src1 on different banks
+          k.emit(VOPD(VOPDOp.V_DUAL_ADD_F32, VOPDOp.V_DUAL_ADD_F32, vdstx=v[a], vdsty=v[a+1], srcx0=v[a], vsrcx1=v[b+2], srcy0=v[a+1], vsrcy1=v[b+3]))
+          # v[a+2](bank2) += v[b+0](bank0), v[a+3](bank3) += v[b+1](bank1) — src0 and src1 on different banks
+          k.emit(VOPD(VOPDOp.V_DUAL_ADD_F32, VOPDOp.V_DUAL_ADD_F32, vdstx=v[a+2], vdsty=v[a+3], srcx0=v[a+2], vsrcx1=v[b], srcy0=v[a+3], vsrcy1=v[b+1]))
+        reg_len = half
+      k.emit(VOPD(VOPDOp.V_DUAL_ADD_F32, VOPDOp.V_DUAL_ADD_F32, vdstx=v[4], vdsty=v[5], srcx0=v[4], vsrcx1=v[base_vreg], srcy0=v[5], vsrcy1=v[base_vreg+1]))
+      k.emit(VOPD(VOPDOp.V_DUAL_ADD_F32, VOPDOp.V_DUAL_ADD_F32, vdstx=v[6], vdsty=v[7], srcx0=v[6], vsrcx1=v[base_vreg+2], srcy0=v[7], vsrcy1=v[base_vreg+3]))
+
+    BASE_REG = 8
+    LOAD_UNROLL = 64
+    INNER_UNROLL = 2
+
+    assert buf.numel() % (CU_COUNT*LANES*LOAD_UNROLL*INNER_UNROLL) == 0
+    total_batches = buf.numel()//(CU_COUNT*LANES*LOAD_UNROLL*INNER_UNROLL)
+    k.emit(s_mov_b32(s[S_LOOP_CTR], total_batches-1))
+
+    k.label('LOOP')
+    for _ in range(INNER_UNROLL):
+      emit_loads(BASE_REG, reg_len=LOAD_UNROLL)
+      k.waitcnt(vm=0)
+      tree_reduce_to_4567(BASE_REG, reg_len=LOAD_UNROLL)
+    k.emit(s_sub_u32(s[S_LOOP_CTR], s[S_LOOP_CTR], 1))
+    k.emit(s_cbranch_scc0(), target='LOOP')
+
+    # add into v[4]
+    k.emit(v_add_f32_e32(v[4], v[4], v[5]))
+    k.emit(v_add_f32_e32(v[6], v[6], v[7]))
+    k.emit(v_add_f32_e32(v[4], v[4], v[6]))
+
+    # warp shuffle into v[4] on lane 0 using DPP row_shl within each 16-lane row
+    for shift in [1, 2, 4, 8]:
+      k.emit(v_add_f32_e32(v[4], DPP, v[4], vsrc0=v[4], dpp=0x100 | shift, row_mask=0xf, bank_mask=0xf, bc=1))
+    # combine rows: get lane 16's value to lane 0 via permlanex16
+    k.emit(v_permlanex16_b32(v[5], v[4], 0, 0))
+    k.emit(v_add_f32_e32(v[4], v[4], v[5]))
+
+    # atomic store (only on lane 0)
+    k.emit(s_mov_b32(EXEC_LO, 1))
+    k.emit(v_mov_b32_e32(v[0], 0))
+    k.emit(global_atomic_add_f32(addr=v[0], saddr=s[4:5], data=v[4]))
+
+    k.emit(s_sendmsg(simm16=3))  # DEALLOC_VGPRS
+    k.emit(s_endpgm())
+    return k.finalize(UOp.sink(UOp.special(CU_COUNT, 'gidx0'), UOp.special(LANES, 'lidx0'), out, buf, arg=KernelInfo(name="asm_reduce")))
+
+  out = Tensor.zeros(1,).contiguous().realize()
+  eval_harness("RDNA3 assembly kernel", a, lambda x: out.custom_kernel(x, fxn=asm_sum)[0], check=correct)
+
+if __name__ == "__main__":
+  examples = [int(x) for x in getenv("EXAMPLES", "1,2,3,4,5").split(",")]
+
+  correct = None
+  # First define a Tensor and realize it. We will focus on a 1GB sum kernel on RDNA3
+  a = (Tensor.randn(SZ) if getenv("RAND") else Tensor.ones(SZ)).contiguous().realize()
+
+  if 1 in examples:
+    # *****
+    # This is the high level tinygrad way.
+    # Note that this is split into multiple kernels for speed.
+    correct = eval_harness("basic kernel", a, lambda x: x.sum())
+
+  if 2 in examples:
+    # *****
+    # You can import kernels from CUDA/HIP/Metal.
+    # ChatGPT is great at writing these Kernel
+    example_2_hip(a, correct)
+
+  if 3 in examples:
+    # *****
+    # Now we get to the lower abstraction layers of tinygrad.
+    # You can write a kernel in UOps, and it's 2.5x faster than normal.
+    example_3_custom_uop(a, correct)
+
+  if 4 in examples:
+    # *****
+    # You can also BEAM search stock tinygrad for a faster kernel.
+    # This does even better than all the kernels to date in this simple case.
+    with Context(BEAM=2):
+      eval_harness("BEAMed kernel", a, lambda x: x.sum(), check=correct)
+
+  if 5 in examples:
+    # *****
+    # If you really want to go crazy with speed, you can code in assembly.
+    # There's not too much to gain here over BEAM, but it's a few percent faster.
+    example_5_custom_assembly(a, correct)
@@ -3,7 +3,7 @@
 AM driver is a userspace driver targeting AMD's RDNA3/RDNA4. You only need tinygrad to send compute tasks to your GPU!

 ## How to run?
-Make sure that amdgpu module is unloaded and just run tinygrad with `AMD=1`!
+Make sure that amdgpu module is unloaded and just run tinygrad with `DEV=AMD`!

 Optional requirements:

@@ -17,15 +17,13 @@ The `UOp` graph specifies the compute in terms of low level tinygrad ops. Not al

 ## Scheduling

-The [scheduler](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/schedule.py) converts the graph of UOps into a list of `ExecItem`. One `ExecItem` is one kernel on the GPU, and the scheduler is responsible for breaking the large compute graph into subgraphs that can fit in a kernel. `ast` specifies what compute to run, and `bufs` specifies what buffers to run it on.
-
-::: tinygrad.engine.schedule.ExecItem
+The [scheduler](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/schedule/__init__.py) converts the graph of UOps into a `LINEAR` UOp whose `src` is a list of `CALL` UOps. One `CALL` is one kernel on the GPU, and the scheduler is responsible for breaking the large compute graph into subgraphs that can fit in a kernel. The `CALL`'s `src[0]` (a `SINK` ast) specifies what compute to run, and the remaining `src` are the buffers to run it on.

 ## Lowering

-The code in [realize](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/realize.py) lowers `ExecItem` by populating its `prg` field with
+The code in [realize](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/realize.py) lowers each `CALL` by compiling its ast into a `PROGRAM` and running it.

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

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

@@ -35,13 +33,7 @@ Then we render the UOps into code with a `Renderer`, then we compile the code to

 ## Execution

-Creating `ExecItem`, which has a run method
-
-::: tinygrad.engine.realize.ExecItem
-    options:
-        members: true
-
-Lists of `ExecItem` can be condensed into a single ExecItem with the Graph API (rename to Queue?)
+`run_linear` walks the `LINEAR` UOp, dispatching each `CALL` to a runner (kernel, copy, view, encdec, or graph).

 ## Runtime

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

 Group UOps into kernels.

-::: tinygrad.schedule.rangeify.get_rangeify_map
+::: tinygrad.schedule.rangeify.get_kernel_graph
    options:
        members: false
        show_labels: false
@@ -28,7 +28,7 @@ Transforms the ast into an optimized ast. This is where BEAM search and heuristi

 Transform the optimized ast into a linearized and rendered program.

-::: tinygrad.codegen.get_program
+::: tinygrad.codegen.to_program
    options:
        members: false
        show_labels: false
@@ -53,7 +53,7 @@ Transform the linearized list of UOps into a program, represented as a string.

 Abstracted high level interface to the runtimes.

-::: tinygrad.engine.realize.get_program
+::: tinygrad.engine.realize.to_program
    options:
        members: false
        show_labels: false
@@ -3,7 +3,7 @@
 This is a list of environment variable that control the runtime behavior of tinygrad and its examples.
 Most of these are self-explanatory, and are usually used to set an option at runtime.

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

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

@@ -31,31 +31,43 @@ These control the behavior of core tinygrad even when used as a library.
 Variable | Possible Value(s) | Description
 ---|---|---
 DEBUG               | [1-7]      | enable debugging output (operations, timings, speed, generated code and more)
-CL                  | [1]        | enable OpenCL backend
-CUDA                | [1]        | enable CUDA backend
-AMD                 | [1]        | enable AMD backend
-NV                  | [1]        | enable NV backend
-METAL               | [1]        | enable Metal backend (for Mac M1 and after)
-CPU                 | [1]        | enable CPU backend
+DEV                 | [AMD, NV, ...] | enable a specific backend, see [below](#dev-variable)
 BEAM                | [#]        | number of beams in kernel beam search
 DEFAULT_FLOAT       | [HALF, ...]| specify the default float dtype (FLOAT32, HALF, BFLOAT16, FLOAT64, ...), default to FLOAT32
-IMAGE               | [1-2]      | enable 2d specific optimizations
+IMAGE               | [1]        | enable 2d specific optimizations
 FLOAT16             | [1]        | use float16 for images instead of float32
-HCQ_VISIBLE_DEVICES | [list[int]]| restricts the HCQ devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
 JIT                 | [0-2]      | 0=disabled, 1=[jit enabled](quickstart.md#jit) (default), 2=jit enabled, but graphs are disabled
 VIZ                 | [1]        | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
 ALLOW_TF32          | [1]        | enable TensorFloat-32 tensor cores on Ampere or newer GPUs.
 WEBGPU_BACKEND      | [WGPUBackendType_Metal, ...]          | Force select a backend for WebGPU (Metal, DirectX, OpenGL, Vulkan...)
 CUDA_PATH           | str        | Use `CUDA_PATH/include` for CUDA headers for CUDA and NV backends. If not set, TinyGrad will use `/usr/local/cuda/include`, `/usr/include` and `/opt/cuda/include`.

-## Debug breakdown
+### DEV variable
+
+The `DEV` variable deserves special note due to its more nuanced syntax.
+`DEV` is used to specify the target device, target renderer and target architecture for said device, separated by colons.
+Specifying the renderer and architecture is optional, omitting a preference will cause tinygrad to automatically determine a suitable setting.
+The `DEV` variable may also be used to specify the interface through which to access the device (eg. `PCI`, `USB`). Interfaces may be specified preceding the target triple,
+separated by a plus (eg. `DEV=USB+AMD:LLVM`). Similarly as above, the interface may be omitted. Example usage follows:
+
+`DEV` contents | Interpretation
+--- | ---
+AMD           | use the AMD device
+AMD:LLVM      | use the AMD device with the LLVM renderer
+NV:CUDA:sm_70 | use the NV device with the CUDA renderer targetting sm_70
+AMD::gfx950   | use the AMD device targetting gfx950
+USB+AMD       | use the AMD device over the USB interface
+CPU:LLVM      | use the CPU device with the LLVM renderer
+CPU:LLVM:x86_64,znver2,avx2,-avx512f | use the CPU device with the LLVM renderer, with [additional arch flags](runtime.md#cpu-arch)
+
+### Debug breakdown

 Variable | Value | Description
 ---|---|---
 DEBUG               | >= 1       | Enables debugging and lists devices being used
 DEBUG               | >= 2       | Provides performance metrics for operations, including timing, memory usage, bandwidth for each kernel execution
-DEBUG               | >= 3       | Outputs buffers used for each kernel (shape, dtype and strides) and the applied optimizations at a kernel level
+DEBUG               | >= 3       | Outputs the applied optimizations at a kernel level
 DEBUG               | >= 4       | Outputs the generated kernel code
-DEBUG               | >= 5       | Displays the intermediate representation of the computation UOps (AST)
+DEBUG               | >= 5       | Displays the intermediate representation of the computation UOps
 DEBUG               | >= 6       | Displays the intermediate representation of the computation UOps in a linearized manner, detailing the operation sequence
 DEBUG               | >= 7       | Outputs the assembly code generated for the target hardware
@@ -37,4 +37,4 @@
    options:
        show_signature: false
        separate_signature: false
-::: tinygrad.nn.state.gguf_load
+::: tinygrad.llm.gguf.gguf_load
@@ -1,16 +1,16 @@
 # Runtimes

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

 | Runtime | Description | Compiler Options | Requirements |
 |---------|-------------|------------------|--------------|
-| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`NV_PTX=1`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via `NV_IFACE=(NVK\|PCI)`. See [NV interfaces](#nv-interfaces) for details. |
-| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`AMD_LLVM=1`)<br>HIP/COMGR (`AMD_HIP=1`) | RDNA2 or newer GPUs.<br>You can select an interface via `AMD_IFACE=(KFD\|PCI\|USB)`. See [AMD interfaces](#amd-interfaces) for details. |
+| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`DEV=NV:PTX`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via [the `DEV` variable](env_vars.md#dev-variable). See [NV interfaces](#nv-interfaces) for details. |
+| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`DEV=AMD:LLVM`)<br>HIP/COMGR (`DEV=AMD:HIP`) | RDNA2 or newer GPUs.<br>You can select an interface via [the `DEV` variable](env_vars.md#dev-variable). See [AMD interfaces](#amd-interfaces) for details. |
 | [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
 | [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for `bfloat` support |
-| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`CUDA_PTX=1`) | NVIDIA GPU with CUDA support |
+| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`DEV=CUDA:PTX`) | NVIDIA GPU with CUDA support |
 | [CL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | - | OpenCL 2.0 compatible device |
-| [CPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang or llvm compiler | Clang JIT (default)<br>LLVM IR (`CPU_LLVM=1`) | `clang` compiler in system `PATH` |
+| [CPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang or llvm compiler | Clang JIT (default)<br>LLVM IR (`DEV=CPU:LLVM`) | `clang` compiler in system `PATH`<br>You can specify additional arch parameters via [the `DEV` variable](env_vars.md#dev-variable). See [CPU arch](#cpu-arch) for details. |
 | [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | - | Dawn library installed and discoverable. Binaries: [pydawn v0.3.0](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0) |


@@ -72,10 +72,20 @@ AMD backend supports several interfaces for communicating with devices:
 * `PCI`: uses the [AM driver](developer/am.md)
 * `USB`: USB3 interface for asm24xx chips.

-You can force an interface by setting `AMD_IFACE` to one of these values. In the case of `AMD_IFACE=PCI`, this may unbind your GPU from the amdgpu driver.
+You can force an interface by setting the interface component of [the `DEV` environment variable](env_vars.md#dev-variable) to one of these values. When set to `PCI`, this may unbind your GPU from the amdgpu driver.

 ## NV Interfaces
 NV backend supports several interfaces for communicating with devices:

 * `NVK`: uses the nvidia driver
 * `PCI`: uses the [NV driver](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/support/nv/nvdev.py)
+
+## CPU Arch
+The CPU renderers may be additionally configured using the arch component of [the `DEV` environment variable](env_vars.md#dev-variable).
+CPU arch should be specified as a comma-separated list of parameters, and must contain at least two values: the architecture family (ie. x86_64, arm64, or riscv64) and the cpu type (as accepted by `clang`'s `-march`).
+If native is specified as the cpu type, tinygrad (or delegate compiler) will query the host cpu type. Additional comma-separated values may be specified as follows:
+
+* `AMX`: emit Apple silicon AMX instructions
+
+All other additional values are interpreted as cpu feature flags. When a value is preceded by a `-` character, the corresponding feature flag will be disabled, otherwise the flag will be enabled.
+Note that enabled feature flags should not be preceded by a `+`.
@@ -19,8 +19,8 @@

 ## tinygrad ops

-::: tinygrad.Tensor.schedule_with_vars
-::: tinygrad.Tensor.schedule
+::: tinygrad.Tensor.linear_with_vars
+::: tinygrad.Tensor.schedule_linear
 ::: tinygrad.Tensor.realize
 ::: tinygrad.Tensor.replace
 ::: tinygrad.Tensor.assign
@@ -0,0 +1,61 @@
+# TinyGPU
+
+TinyGPU app lets you use AMD and NVIDIA GPUs on macOS over USB4/Thunderbolt with tinygrad.
+
+## Requirements
+
+- macOS (12.1+)
+- USB4/Thunderbolt port
+- A supported GPU (AMD RDNA3+ or NVIDIA Ampere+)
+
+## Setup
+
+### 1. Connect your GPU
+
+Plug the supported GPU into your Mac over USB4/Thunderbolt.
+
+### 2. Initiate the driver install
+
+> **Note:** If tinygrad is cloned but not installed, run commands with `PYTHONPATH=.`
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/tinygrad/tinygrad/master/extra/setup_tinygpu_osx.sh | sh
+```
+
+This downloads TinyGPU.app and triggers a system prompt to install the driver extension.
+
+### 3. Enable the driver
+
+You should see a system prompt: **"TinyGPU" would like to use a new driver extension**. Click **Open System Settings** and toggle TinyGPU on.
+
+If you missed the prompt, go to **System Settings > General > Login Items & Extensions > Driver Extensions** and toggle TinyGPU on.
+
+### 4. Compiler Setup
+
+#### AMD
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/tinygrad/tinygrad/master/extra/setup_hipcomgr_osx.sh | sh
+```
+
+#### NV
+
+Install [Docker Desktop](https://www.docker.com/products/docker-desktop/) if you don't have it.
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/tinygrad/tinygrad/master/extra/setup_nvcc_osx.sh | sh
+```
+
+Make sure `~/.local/bin` is on your `PATH`:
+
+```bash
+export PATH="$HOME/.local/bin:$PATH"
+```
+
+### 5. Use it!
+
+```bash
+DEV={AMD|NV} python3 -m tinygrad.llm
+```
+
+**Note:** Use `JITBEAM=2` to search for faster kernels (one-time search cost, results cached).
@@ -113,7 +113,7 @@ class VLIWRenderer(Renderer):
        case Ops.GEP:
          # a GEP is just an alias to a special register in the vector
          r[u] = r[u.src[0]] + u.arg[0]
-        case Ops.VECTORIZE:
+        case Ops.STACK:
          if all(s == u.src[0] for s in u.src):
            # if all sources are the same, we can broadcast
            inst.append({"valu": [("vbroadcast", r[u], r[u.src[0]])]})
@@ -173,16 +173,16 @@ if __name__ == "__main__":

  # *** render to device ***

-  from tinygrad.codegen import get_program
+  from tinygrad.codegen import to_program
  with Context(PCONTIG=2, DEVECTORIZE=2, SPEC=0):
    out = tree_traversal(forest_t, val_t, height, rounds)
-    sink = out.schedule()[-1].ast
-    prg = get_program(sink, VLIWRenderer())
+    sink = out.schedule_linear().src[-1].src[0]
+    prg = to_program(sink, VLIWRenderer())

  # *** run on Machine and compare ***

  # NOTE: the scratch size needs to be reduced to 1536 when you have a register allocator
-  src = eval(prg.src)
+  src = eval(prg.src[3].arg)
  max_regs = max(t[1] for instr in src for v in instr.values() for t in v if len(t) > 1) + 8
  print(f"{max_regs:5d} regs used" + ("" if max_regs <= 1536 else "       <-- WARNING: TOO MANY REGISTERS, MUST BE <= 1536"))
  machine = problem.Machine(mem, src, problem.DebugInfo(scratch_map={}), n_cores=1, trace=False, scratch_size=max_regs)
@@ -1,6 +1,6 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

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

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

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

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

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

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

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

 if __name__ == "__main__":
@@ -1,9 +1,10 @@
 from pathlib import Path
 from extra.models.efficientnet import EfficientNet
 from tinygrad.tensor import Tensor
+from tinygrad.device import Device
 from tinygrad.nn.state import get_state_dict, safe_save, safe_load, load_state_dict
 from extra.export_model import export_model
-from tinygrad.helpers import getenv, fetch
+from tinygrad.helpers import fetch
 import ast

 if __name__ == "__main__":
@@ -12,13 +13,13 @@ if __name__ == "__main__":
  dirname = Path(__file__).parent
  # exporting a model that's loaded from safetensors doesn't work without loading in from safetensors first
  # loading the state dict from a safetensor file changes the generated kernels
-  if getenv("WEBGPU"):
+  if Device.DEFAULT == "WEBGPU":
    safe_save(get_state_dict(model), (dirname / "net.safetensors").as_posix())
    load_state_dict(model, safe_load(str(dirname / "net.safetensors")))
-  mode = "clang" if getenv("CPU", "") != "" else "webgpu" if getenv("WEBGPU", "") != "" else ""
+  mode = "clang" if Device.DEFAULT == "CPU" else "webgpu" if Device.DEFAULT == "WEBGPU" else ""
  prg, inp_sizes, out_sizes, state = export_model(model, mode, Tensor.randn(1,3,224,224))
-  if getenv("CPU", "") == "":
-    ext = "js" if getenv("WEBGPU", "") != "" else "json"
+  if Device.DEFAULT != "CPU":
+    ext = "js" if Device.DEFAULT == "WEBGPU" else "json"
    with open(dirname / f"net.{ext}", "w") as text_file:
      text_file.write(prg)
  else:
@@ -68,6 +69,6 @@ if __name__ == "__main__":
    else printf("%s\\n", lbls[best_idx]);
  }""")

-    # CPU=1 python3 examples/compile_efficientnet.py | clang -O2 -lm -x c - -o recognize && DEBUG=1 time ./recognize docs/showcase/stable_diffusion_by_tinygrad.jpg
+    # DEV=CPU python3 examples/compile_efficientnet.py | clang -O2 -lm -x c - -o recognize && DEBUG=1 time ./recognize docs/showcase/stable_diffusion_by_tinygrad.jpg
    # category : 281 (tabby, tabby cat) with 9.452788
    print('\n'.join(cprog))
@@ -35,12 +35,11 @@ def compile_onnx_model(onnx_model):
  tinyonnx = TinyOnnx(onnx_model)
  the_input = Tensor.randn(1,32)

-  run, special_names = jit_model(tinyonnx, the_input)
+  linear, output_bufs = jit_model(tinyonnx, the_input)
+  the_output = [tinyonnx.forward(the_input)]

-  functions, statements, bufs, bufs_to_save = compile_net(run, special_names)
+  functions, statements, bufs, bufs_to_save = compile_net(linear, output_bufs)
  prg = export_model_clang(functions, statements, bufs, {}, ["input0"], ["output0"])
-
-  the_output = run(the_input)
  cprog = ["#include <string.h>", "#include <stdio.h>", "#include <stdlib.h>"]
  cprog.append(prg)

@@ -5,8 +5,9 @@ with contextlib.suppress(ImportError): import tiktoken
 from tinygrad import Tensor, TinyJit, Device, GlobalCounters, Variable, dtypes
 from tinygrad.uop.ops import UOp
 from tinygrad.helpers import Timing, DEBUG, JIT, getenv, fetch, colored, trange
+from tinygrad.llm.gguf import gguf_load
 from tinygrad.nn import Embedding, Linear, LayerNorm
-from tinygrad.nn.state import gguf_load, torch_load, load_state_dict, get_state_dict
+from tinygrad.nn.state import torch_load, load_state_dict, get_state_dict
 from extra.bench_log import BenchEvent, WallTimeEvent

 MAX_CONTEXT = getenv("MAX_CONTEXT", 128)
@@ -19,8 +19,8 @@ cifar_std = [0.24703225141799082, 0.24348516474564, 0.26158783926049628]
 BS, STEPS = getenv("BS", 512), getenv("STEPS", 1000)
 EVAL_BS = getenv("EVAL_BS", BS)
 GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]
-assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"
-assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"
+assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}"
+assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}"

 class UnsyncedBatchNorm:
  def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1, num_devices=len(GPUS)):
@@ -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.uop.size * x.dtype.itemsize for x in get_parameters(llama.model))
+  param_bytes = sum(x.nbytes() 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)
@@ -2,7 +2,8 @@ from pathlib import Path
 from typing import List
 import json, argparse, random, time, os
 from extra.models.llama import Transformer, convert_from_huggingface, convert_from_gguf, fix_bf16
-from tinygrad.nn.state import safe_load, torch_load, load_state_dict, get_parameters, gguf_load
+from tinygrad.llm.gguf import gguf_load
+from tinygrad.nn.state import safe_load, torch_load, load_state_dict, get_parameters
 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
@@ -324,7 +325,7 @@ if __name__ == "__main__":

  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(args.shard)) if args.shard > 1 else Device.DEFAULT
  model = build_transformer(args.model, model_size=args.size, quantize=args.quantize, device=device)
-  param_bytes = sum(x.uop.size * x.dtype.itemsize for x in get_parameters(model))
+  param_bytes = sum(x.nbytes() 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
@@ -2,13 +2,14 @@
 import os
 if "NOOPT" not in os.environ: os.environ["NOOPT"] = "1"
 from tinygrad import Device, nn, Tensor, dtypes
-Device.DEFAULT = "CPU"
 from train_gpt2 import GPT, GPTConfig
-from tinygrad.helpers import dedup, flatten, getenv, GlobalCounters, to_function_name
+from tinygrad.helpers import DEV, dedup, flatten, getenv, GlobalCounters, to_function_name
 from tinygrad.engine.realize import get_kernel
-from tinygrad.engine.memory import memory_planner
+from tinygrad.schedule.memory import memory_planner
 from tinygrad.uop.ops import Ops

+DEV.value = "CPU"
+
 TIMING = getenv("TIMING")

 if __name__ == "__main__":
@@ -65,17 +65,7 @@ def loader_process(q_in, q_out, X:Tensor, seed):
      else:
        # pad data with training mean
        img = np.tile(np.array([[[123.68, 116.78, 103.94]]], dtype=np.uint8), (224, 224, 1))
-
-      # broken out
-      #img_tensor = Tensor(img.tobytes(), device='CPU')
-      #storage_tensor = X[idx].contiguous().realize().lazydata.base.realized
-      #storage_tensor._copyin(img_tensor.numpy())
-
-      # faster
-      X[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = img.tobytes()
-
-      # ideal
-      #X[idx].assign(img.tobytes())   # NOTE: this is slow!
+      X[idx].flatten().assign(img.tobytes())
      q_out.put(idx)
    q_out.put(None)

@@ -264,8 +254,8 @@ def load_unet3d_data(preprocessed_dataset_dir, seed, queue_in, queue_out, X:Tens
      x = random_brightness_augmentation(x)
      x = gaussian_noise(x)

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

    queue_out.put(idx)
  queue_out.put(None)
@@ -379,12 +369,12 @@ 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().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_boxes.tobytes()
-      labels[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_labels.tobytes()
-      matches[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = match_idxs.tobytes()
-      anchors[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = anchor.tobytes()
+      boxes[idx].flatten().assign(clipped_boxes.tobytes())
+      labels[idx].flatten().assign(clipped_labels.tobytes())
+      matches[idx].flatten().assign(match_idxs.tobytes())
+      anchors[idx].flatten().assign(anchor.tobytes())

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

    queue_out.put(idx)
  queue_out.put(None)
@@ -406,6 +396,7 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
      queue_in.put((idx, img, tgt))

  def _setup_shared_mem(shm_name:str, size:tuple[int, ...], dtype:dtypes) -> tuple[shared_memory.SharedMemory, Tensor]:
+    shm_name = f"{shm_name}_{os.getpid()}"
    if os.path.exists(f"/dev/shm/{shm_name}"): os.unlink(f"/dev/shm/{shm_name}")
    shm = shared_memory.SharedMemory(name=shm_name, create=True, size=prod(size))
    shm_tensor = Tensor.empty(*size, dtype=dtype, device=f"disk:/dev/shm/{shm_name}")
@@ -325,19 +325,18 @@ def eval_stable_diffusion():
  # NOTE: the clip weights are the same between model.cond_stage_model and clip_encoder
  eval_timesteps = list(reversed(range(1, 1000, 20)))

-  original_device, Device.DEFAULT = Device.DEFAULT, "CPU"
-  # The choice of alphas_prev[0] = alphas_cumprod[0] seems arbitrary, but it's how the mlperf ref does it:
-  #   alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
-  eval_alphas_prev = model.alphas_cumprod[0:1].cat(model.alphas_cumprod[list(range(1, 1000, 20))[:-1]]).to(GPUS).realize()
-  inception = FidInceptionV3().load_from_pretrained(CKPTDIR / "inception" / "pt_inception-2015-12-05-6726825d.pth")
-  vision_cfg = {'width': 1280, 'layers': 32, 'd_head': 80, 'image_size': 224, 'patch_size': 14}
-  text_cfg = {'width': 1024, 'n_heads': 16, 'layers': 24, 'vocab_size': 49408, 'ctx_length': 77}
-  clip.gelu = gelu_erf
-  clip_encoder = OpenClipEncoder(1024, text_cfg, vision_cfg)
-  loaded = torch_load(CKPTDIR / "clip" / "open_clip_pytorch_model.bin")
-  loaded.update({"attn_mask": clip_encoder.attn_mask, "mean": clip_encoder.mean, "std": clip_encoder.std})
-  load_state_dict(clip_encoder, loaded)
-  Device.DEFAULT=original_device
+  with Context(DEV="CPU"):
+    # The choice of alphas_prev[0] = alphas_cumprod[0] seems arbitrary, but it's how the mlperf ref does it:
+    #   alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+    eval_alphas_prev = model.alphas_cumprod[0:1].cat(model.alphas_cumprod[list(range(1, 1000, 20))[:-1]]).to(GPUS).realize()
+    inception = FidInceptionV3().load_from_pretrained(CKPTDIR / "inception" / "pt_inception-2015-12-05-6726825d.pth")
+    vision_cfg = {'width': 1280, 'layers': 32, 'd_head': 80, 'image_size': 224, 'patch_size': 14}
+    text_cfg = {'width': 1024, 'n_heads': 16, 'layers': 24, 'vocab_size': 49408, 'ctx_length': 77}
+    clip.gelu = gelu_erf
+    clip_encoder = OpenClipEncoder(1024, text_cfg, vision_cfg)
+    loaded = torch_load(CKPTDIR / "clip" / "open_clip_pytorch_model.bin")
+    loaded.update({"attn_mask": clip_encoder.attn_mask, "mean": clip_encoder.mean, "std": clip_encoder.std})
+    load_state_dict(clip_encoder, loaded)

  @TinyJit
  def denoise_step(x:Tensor, x_x:Tensor, t_t:Tensor, uc_c:Tensor, sqrt_alphas_cumprod_t:Tensor, sqrt_one_minus_alphas_cumprod_t:Tensor,
@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing

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

@@ -246,7 +246,7 @@ def train_resnet():

      if i == BENCHMARK:
        assert not math.isnan(loss)
-        median_step_time = sorted(step_times)[(BENCHMARK + 1) // 2]  # in seconds
+        median_step_time = sorted(step_times)[BENCHMARK // 2]  # in seconds
        estimated_total_minutes = int(median_step_time * steps_in_train_epoch * epochs / 60)
        print(f"Estimated training time: {estimated_total_minutes // 60}h{estimated_total_minutes % 60}m")
        print(f"epoch global_ops: {steps_in_train_epoch * GlobalCounters.global_ops:_}, "
@@ -593,7 +593,7 @@ def train_retinanet():

      if i == BENCHMARK:
        assert not math.isnan(loss)
-        median_step_time = sorted(step_times)[(BENCHMARK + 1) // 2]  # in seconds
+        median_step_time = sorted(step_times)[BENCHMARK // 2]  # in seconds
        estimated_total_minutes = int(median_step_time * steps_in_train_epoch * EPOCHS / 60)
        print(f"Estimated training time: {estimated_total_minutes // 60}h{estimated_total_minutes % 60}m")
        print(f"epoch global_ops: {steps_in_train_epoch * GlobalCounters.global_ops:_}, "
@@ -868,7 +868,7 @@ def train_unet3d():
        i += 1

        if i == BENCHMARK:
-          median_step_time = sorted(step_times)[(BENCHMARK + 1) // 2]  # in seconds
+          median_step_time = sorted(step_times)[BENCHMARK // 2]  # in seconds
          estimated_total_minutes = int(median_step_time * SAMPLES_PER_EPOCH * NUM_EPOCHS / 60)
          print(f"Estimated training time: {estimated_total_minutes // 60}h{estimated_total_minutes % 60}m")
          if (TRAIN_BEAM or EVAL_BEAM) and epoch == start_epoch: break
@@ -1167,7 +1167,7 @@ def train_bert():
      i += 1

      if i == BENCHMARK:
-        median_step_time = sorted(step_times)[(BENCHMARK + 1) // 2]  # in seconds
+        median_step_time = sorted(step_times)[BENCHMARK // 2]  # in seconds
        estimated_total_minutes = int(median_step_time * train_steps / 60)
        print(f"Estimated training time: {estimated_total_minutes // 60}h{estimated_total_minutes % 60}m")
        print(f"epoch global_ops: {train_steps * GlobalCounters.global_ops:_}, "
@@ -1282,11 +1282,14 @@ def train_bert():
        previous_step = i

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

+  INITMLPERF = getenv("INITMLPERF")
+  RUNMLPERF = getenv("RUNMLPERF")
+  LOGMLPERF = getenv("LOGMLPERF")
  BENCHMARK = getenv("BENCHMARK")

  config = {}
@@ -1309,15 +1312,61 @@ def train_llama3():
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)

-  # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 JITBEAM=2 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=512 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py
-  # trains to 7
+  if LOGMLPERF:
+    from mlperf_logging import mllog
+    import mlperf_logging.mllog.constants as mllog_constants
+
+    mllog.config(filename=f"result_llama31_{SEED}.log")
+    mllog.config(root_dir=Path(__file__).parents[3].as_posix())
+    MLLOGGER = mllog.get_mllogger()
+    MLLOGGER.logger.propagate = False
+
+    LLAMA_BENCHMARK = mllog_constants.LLAMA31_405B if getenv("LLAMA3_SIZE", "8B") == "405B" else mllog_constants.LLAMA31_8B
+
+    if INITMLPERF:
+      assert BENCHMARK, "BENCHMARK must be set for INITMLPERF"
+      MLLOGGER.event(key=mllog_constants.SUBMISSION_ORG, value="tinycorp")
+      MLLOGGER.event(key=mllog_constants.SUBMISSION_PLATFORM, value=getenv("SUBMISSION_PLATFORM", "tinybox"))
+      MLLOGGER.event(key=mllog_constants.SUBMISSION_DIVISION, value=mllog_constants.CLOSED)
+      MLLOGGER.event(key=mllog_constants.SUBMISSION_STATUS, value=mllog_constants.ONPREM)
+
+      MLLOGGER.event(key=mllog_constants.SUBMISSION_BENCHMARK, value=LLAMA_BENCHMARK)
+
+      diskcache_clear()
+      MLLOGGER.event(key=mllog_constants.CACHE_CLEAR, value=True)
+      MLLOGGER.start(key=mllog_constants.INIT_START, value=None)
+
+    if RUNMLPERF:
+      MLLOGGER.start(key=mllog_constants.RUN_START, value=None)
+      MLLOGGER.event(key=mllog_constants.SEED, value=SEED)
+
+      MLLOGGER.event(key=mllog_constants.GLOBAL_BATCH_SIZE, value=GBS)
+      MLLOGGER.event(key=mllog_constants.MAX_SEQUENCE_LENGTH, value=SEQLEN)
+      MLLOGGER.event(key=mllog_constants.MAX_STEPS, value=MAX_STEPS)
+      MLLOGGER.event(key=mllog_constants.GRADIENT_ACCUMULATION_STEPS, value=grad_acc)
+      MLLOGGER.event(key=mllog_constants.EVAL_SAMPLES, value=EVAL_SAMPLES)
+      MLLOGGER.event(key=mllog_constants.TRAIN_SAMPLES, value=SAMPLES)
+
+      MLLOGGER.event(key=mllog_constants.OPT_NAME, value=mllog_constants.ADAMW)
+      MLLOGGER.event(key=mllog_constants.OPT_BASE_LR, value=LR)
+      MLLOGGER.event(key=mllog_constants.OPT_END_LR, value=END_LR)
+      MLLOGGER.event(key=mllog_constants.OPT_ADAMW_BETA_1, value=0.9)
+      MLLOGGER.event(key=mllog_constants.OPT_ADAMW_BETA_2, value=0.95)
+      MLLOGGER.event(key=mllog_constants.OPT_ADAMW_EPSILON, value=1e-5)
+      MLLOGGER.event(key=mllog_constants.OPT_ADAMW_WEIGHT_DECAY, value=0.1)
+      MLLOGGER.event(key=mllog_constants.OPT_LR_WARMUP_STEPS, value=WARMUP_STEPS)
+      MLLOGGER.event(key=mllog_constants.NUM_WARMUP_STEPS, value=WARMUP_STEPS)
+      MLLOGGER.event(key=mllog_constants.OPT_LR_DECAY_STEPS, value=MAX_STEPS - WARMUP_STEPS)
+      MLLOGGER.event(key=mllog_constants.OPT_LR_DECAY_SCHEDULE, value="cosine with linear warmup")
+      MLLOGGER.event(key=mllog_constants.OPT_GRADIENT_CLIP_NORM, value=1.0)
+  else:
+    MLLOGGER = None

  opt_adamw_beta_1 = 0.9
  opt_adamw_beta_2 = 0.95
  opt_adamw_epsilon = 1e-5
  opt_adamw_weight_decay = 0.1

-  opt_gradient_clip_norm = 1.0
  opt_learning_rate_warmup_steps = WARMUP_STEPS
  opt_learning_rate_decay_steps = MAX_STEPS - opt_learning_rate_warmup_steps
  opt_base_learning_rate = LR
@@ -1335,49 +1384,43 @@ def train_llama3():
  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
  # vocab_size from the mixtral tokenizer
  if not SMALL: model_params |= {"vocab_size": 32000}
+  real_vocab_size = model_params['vocab_size']
  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
  print(f"model parameters: {model_params}")

-  model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
+  # pad vocab
+  if (MP := getenv("MP", 1)) > 1: model_params['vocab_size'] = round_up(model_params['vocab_size'], 256 * MP)
+  vocab_mask:Tensor = Tensor.arange(model_params['vocab_size']).reshape(1, 1, -1) >= real_vocab_size
+
+  model = FlatTransformer(**model_params, max_context=SEQLEN)
+
  params = get_parameters(model)
-  # weights are all bfloat16 for now
-  assert params and all(p.dtype == dtypes.bfloat16 for p in params)

-  if getenv("FAKEDATA"):
+  if getenv("EMPTYWEIGHT"):
    for v in get_parameters(model):
-      v = v.assign(Tensor.empty(v.shape))
+      v = v.assign(Tensor.empty(v.shape, dtype=v.dtype))

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

-  if (MP := getenv("MP", 1)) > 1:
-    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
-    for k,v in get_state_dict(model).items():
-      if 'scale' in k: v.shard_(device, axis=None)  # from quantized
-      elif '.attention.wq' in k: v.shard_(device, axis=0)
-      elif '.attention.wk' in k: v.shard_(device, axis=0)
-      elif '.attention.wv' in k: v.shard_(device, axis=0)
-      elif '.attention.wo' in k: v.shard_(device, axis=1)
-      elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
-      elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
-      elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
-      elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
-      elif 'output.weight' in k: v.shard_(device, axis=0)
-      else:
-        # attention_norm, ffn_norm, norm
-        v.shard_(device, axis=None)
-      # prevents memory spike on device 0
-      v.realize()
+  model.shard(device, is_mp)

-  optim = GradAccClipAdamW(get_parameters(model), lr=0.0,
-                b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc)
+  if is_dp: vocab_mask.shard_(device, axis=None).realize()
+  if is_mp: vocab_mask.shard_(device, axis=2).realize()
+
+  is_offload_optim = bool(getenv("OFFLOAD_OPTIM"))
+  is_fake_offload = Device.DEFAULT == "NULL"
+  optim_device = ("CPU" if not is_fake_offload else "NULL:99") if is_offload_optim else None
+  optim = GradAccClipAdamW(params, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2,
+                           eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device)

-  # init grads
  for p in optim.params:
-    p.grad = p.zeros_like().contiguous().realize()
-  grads: list[Tensor] = [p.grad for p in optim.params]
+    grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
+    p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device).contiguous()
+  grads = [p.grad for p in optim.params]

  scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)

@@ -1390,53 +1433,66 @@ def train_llama3():
    print(f"loading optim checkpoint from {fn}")
    load_state_dict(scheduler, safe_load(fn), realize=False)

+  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
+  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts] if hasattr(model, "_fp8_grad_amax") else []
+  fp8_inv_scales = list(model._fp8_inv_scale.values())
+
+  from tinygrad.nn.state import get_state_dict
+  model_state = get_state_dict(model)
+  for wname in ["wqkv", "wo", "w13", "w2"]:
+    w = model_state[wname]
+    w._inv_scale = model._fp8_inv_scale[wname]
+    if optim.master_params:
+      idx = next(j for j, p in enumerate(optim.params) if p is w)
+      optim.master_params[idx].assign((optim.master_params[idx] * w._inv_scale.reshape(-1, *([1]*(w.ndim-1)))).contiguous())
+
  @TinyJit
  def minibatch(tokens:Tensor):
-    tokens = tokens.to(None)
-    if (DP := getenv("DP", 1)) > 1:
-      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
-      tokens = tokens.shard(device, 0)
-    if (MP := getenv("MP", 1)) > 1:
-      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
-      tokens = tokens.shard(device)
-    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
-    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    loss.backward()
-    assert all(p.grad is g for p,g in zip(optim.params, grads))
-    Tensor.realize(loss, *grads)
-    return loss.flatten().float().to("CPU")
+    if is_dp: tokens = tokens.to(None).shard(device, 0)
+    if is_mp: tokens = tokens.shard(device)
+    if not is_sharding: tokens = tokens.to(None)
+    logits:Tensor = model(tokens[:, :-1])
+    if getenv("FAST_CE", 0):
+      from extra.llama_kernels.fused_ce import fused_ce_loss
+      loss = fused_ce_loss(logits.cast(dtypes.bfloat16), tokens[:, 1:], label_smoothing=0.0)
+    else:
+      loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
+
+    for g, new_g in zip(grads, loss.gradient(*optim.params)):
+      apply_grad(g, new_g.uop)
+
+    loss_cpu = loss.flatten().float().to("CPU")
+    return loss_cpu.realize(*grads, *fp8_amax, *fp8_grad_amax)

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

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

-    lr = optim.lr
-    Tensor.realize(lr, *grads)
+    lr_cpu = optim.lr.float().to("CPU")
+    grad_norm_cpu = grad_norm.float().to("CPU")
+    Tensor.realize(lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales)

-    return lr.float().to("CPU")
+    return lr_cpu, grad_norm_cpu

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

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

  def get_train_iter():
    if getenv("FAKEDATA", 0):
@@ -1461,51 +1517,60 @@ def train_llama3():
  train_iter = get_train_iter()
  i, sequences_seen = resume_ckpt, 0
  step_times = []
+
+  if MLLOGGER and RUNMLPERF:
+    MLLOGGER.start(key=mllog_constants.EPOCH_START, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
+    MLLOGGER.start(key=mllog_constants.BLOCK_START, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
+
  while i < MAX_STEPS:
    GlobalCounters.reset()
+    actual_gbs = GBS if i >= 2 else BS
    if getenv("TRAIN", 1):
      profile_marker(f"train @ {i}")
      st = time.perf_counter()

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

      gt = time.perf_counter()
-      lr = optim_step()
-      ot = time.perf_counter()
-
-      loss = loss.float().item()
-      lr = lr.item()
-
+      ret = optim_step()
+      lr, grad_norm = ret[0].item(), ret[1].item()
      et = time.perf_counter()
+
+      loss = sum(losses) / len(losses)
+      optim_time = et - gt
+      dev_time += optim_time
      step_time = et - st
      gbs_time = gt - st
-      optim_time = ot - gt
-      data_time = dt - ist
-      dev_time = step_time - data_time * grad_acc
      if BENCHMARK: step_times.append(step_time)

      i += 1
-      sequences_seen += GBS
+      sequences_seen += actual_gbs

      mem_gb = GlobalCounters.mem_used / 1e9
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * max(getenv("DP", 1), getenv("MP", 1)) * 2.3e15)) * 100
+      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 100
      tqdm.write(
          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
-          f"{lr:.12f} LR, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
+          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
+      if DEBUG >= 1: tqdm.write("  mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))

      if WANDB:
        wandb.log({
-          "lr": lr, "train/loss": loss,
+          "train/loss": loss,
+          "train/lr": lr,
+          "train/grad_norm": grad_norm,
          "train/step_time": step_time,
          "train/gbs_time": gbs_time,
          "train/optim_time": optim_time,
@@ -1528,42 +1593,58 @@ def train_llama3():
        safe_save(get_state_dict(scheduler), fn)

      if i == BENCHMARK:
-        median_step_time = sorted(step_times)[(BENCHMARK + 1) // 2]
-        estimated_total_minutes = int(median_step_time * (SAMPLES // GBS) / 60)
+        median_step_time = sorted(step_times)[BENCHMARK // 2]
+        estimated_steps = 200_000 // GBS if getenv("LLAMA3_SIZE", "8B") == "8B" else MAX_STEPS
+        estimated_total_minutes = int(median_step_time * estimated_steps / 60)
        print(f"Estimated training time: {estimated_total_minutes // 60}h{estimated_total_minutes % 60}m")
        print(f"epoch global_ops: {GlobalCounters.global_ops:_}, "
              f"epoch global_mem: {GlobalCounters.global_mem:_}")

-    if (sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
+    if (sequences_seen // EVAL_FREQ != (sequences_seen - actual_gbs) // EVAL_FREQ and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
      if EVAL_BS == 0: return
      tqdm.write(f"evaluating after {sequences_seen} sequences")
      profile_marker(f"eval @ {i}")

+      if MLLOGGER and RUNMLPERF:
+        MLLOGGER.end(key=mllog_constants.BLOCK_STOP, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
+        MLLOGGER.start(key=mllog_constants.EVAL_START, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
+
      # run eval
      eval_losses = []
      eval_iter = get_eval_iter()
-      tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")
+      tqdm.write(f"evaluating {EVAL_SAMPLES//EVAL_BS} batches of {EVAL_BS} sequences")

      for j,tokens in tqdm(enumerate(eval_iter), total=EVAL_SAMPLES//EVAL_BS):
        eval_losses += eval_step(tokens).tolist()

        if BENCHMARK and (j+1) == min(BENCHMARK, EVAL_SAMPLES//EVAL_BS):
+          if MLLOGGER and INITMLPERF:
+            MLLOGGER.end(key=mllog_constants.INIT_STOP, value=None)
          return

-      log_perplexity = Tensor(eval_losses).mean().float().item()
+      log_perplexity = sum(eval_losses) / len(eval_losses)

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

+      if MLLOGGER and RUNMLPERF:
+        MLLOGGER.event(key=mllog_constants.EVAL_ACCURACY, value=log_perplexity, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
+        MLLOGGER.end(key=mllog_constants.EVAL_STOP, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
+
      if WANDB:
        wandb.log({"eval/log_perplexity": log_perplexity, "eval/sequences_seen": sequences_seen})

      if log_perplexity < EVAL_TARGET:
        tqdm.write(f"target achieved after {sequences_seen} sequences")
+        if MLLOGGER and RUNMLPERF:
+          MLLOGGER.end(key=mllog_constants.EPOCH_STOP, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})
+          MLLOGGER.end(key=mllog_constants.RUN_STOP, metadata={mllog_constants.STATUS: mllog_constants.SUCCESS})
        if getenv("CKPT"):
          if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
          fn = f"{ckpt_dir}/llama3.safe"
          safe_save(get_state_dict(model), fn)
        break
+      if MLLOGGER and RUNMLPERF:
+        MLLOGGER.start(key=mllog_constants.BLOCK_START, metadata={mllog_constants.SAMPLES_COUNT: sequences_seen})

 def train_stable_diffusion():
  from extra.models.unet import UNetModel
@@ -0,0 +1,323 @@
+import math, os
+if __name__ == "__main__":
+  os.environ["DEFAULT_FLOAT"] = "bfloat16"
+  os.environ["OPTIM_DTYPE"] = "bfloat16"
+  if "DEV" not in os.environ: os.environ["DEV"] = "NULL"
+  # CDNA
+  os.environ["EMULATE"] = "AMD_CDNA4"
+  os.environ["DEVICE_IN_FUNCTION_BUG"] = "1"
+  os.environ["ALL2ALL"] = "1"
+  os.environ["USE_ATOMICS"] = "1"
+  if "HK_FLASH_ATTENTION" not in os.environ:
+    os.environ["HK_FLASH_ATTENTION"] = "1"
+    if "ASM_GEMM" not in os.environ:
+      os.environ["ASM_GEMM"] = "1"
+from tinygrad import Tensor, nn, function, getenv, dtypes, TinyJit
+from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
+from tinygrad.uop.ops import Ops, UOp
+from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
+from extra.llama_kernels.rmsnorm import rmsnorm
+from extra.llama_kernels import FP8_MAX, local_abs_max
+
+ASM_GEMM = getenv("ASM_GEMM", 0)
+FUSED_INPUT_QUANTIZE = getenv("FUSED_INPUT_QUANTIZE", 0)
+FUSED_ADD_NORM_MUL_QUANTIZE = getenv("FUSED_ADD_NORM_MUL_QUANTIZE", 0)
+FUSED_SILU_W13 = getenv("FUSED_SILU_W13", 0)
+
+FP8_DTYPE = dtypes.fp8e4m3
+FP8_GRAD_DTYPE = dtypes.fp8e5m2
+
+def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
+  new_amax = (local_abs_max(x) if isinstance(x.device, tuple) else x.abs().max()).detach().cast(dtypes.float32)
+  scale = FP8_MAX / ((amax_state if amax_state is not None else new_amax) + 1e-8)
+  x_scaled = x * scale
+  x_clamped = x_scaled + (x_scaled.detach().clamp(-FP8_MAX, FP8_MAX) - x_scaled.detach())  # STE
+  return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal(), new_amax
+
+def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
+           x_fp8:Tensor|None=None, x_scale:Tensor|None=None, x_new_amax:Tensor|None=None,
+           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
+  if not fp8:
+    if ASM_GEMM:
+      from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
+      if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T),)
+    return (x @ w.T,)
+  assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
+  if x_fp8 is None:
+    if FUSED_INPUT_QUANTIZE and amax_x is not None:
+      from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed
+      x_fp8, x_scale, x_new_amax, _ = quantize_fp8_delayed(x, amax_x, FP8_DTYPE)
+    else:
+      x_fp8, x_scale, x_new_amax = quantize_fp8(x, amax_state=amax_x)
+  if ASM_GEMM:
+    from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
+    if can_use_asm_gemm(x_fp8, w.T):
+      return asm_gemm(x_fp8, w.T, x_scale=x_scale, w_scale=w_inv_scale, grad_amax_state=grad_amax_state), x_new_amax, x_fp8, w
+  return x_fp8.dot(w.T, dtype=dtypes.float) * x_scale * w_inv_scale, x_new_amax, x_fp8, w
+
+def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor, grad_amax_state:Tensor):
+  if FUSED_ADD_NORM_MUL_QUANTIZE:
+    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_rmsnorm_mul_quantize_fp8
+    x_fp8, x_inv_scale, new_amax, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE)
+    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, x_scale=x_inv_scale, x_new_amax=new_amax, grad_amax_state=grad_amax_state)
+    return out, x_normed, rrms, ret
+  x_normed, rrms = rmsnorm(x, eps)
+  out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale, grad_amax_state=grad_amax_state)
+  return out, x_normed, rrms, ret
+
+def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor):
+  if FUSED_ADD_NORM_MUL_QUANTIZE:
+    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
+    x_fp8, x_inv_scale, new_amax, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE)
+    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, x_scale=x_inv_scale, x_new_amax=new_amax)
+    return out, h, x_normed, rrms, ret
+  h = x + residual
+  x_normed, rrms = rmsnorm(h, eps)
+  out, *ret = matmul(x_normed * norm, w, amax_x=amax_x, w_inv_scale=w_inv_scale)
+  return out, h, x_normed, rrms, ret
+
+def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
+                             amax_x2:Tensor,
+                             grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
+  if FUSED_SILU_W13:
+    from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
+    x2_fp8, x2_inv_scale, new_amax_x2 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13)
+    out, *ret = matmul(None, w2, w_inv_scale=s_2, x_fp8=x2_fp8, x_scale=x2_inv_scale, x_new_amax=new_amax_x2, grad_amax_state=grad_amax_xout)
+    return out, ret
+  hidden = x_w13.shape[-1] // 2
+  x_w1, x_w3 = x_w13[..., :hidden], x_w13[..., hidden:]
+  out, *ret = matmul(x_w1.silu() * x_w3, w2, amax_x=amax_x2, w_inv_scale=s_2, grad_amax_state=grad_amax_xout)
+  return out, ret
+
+class FlatTransformer:
+  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
+               rope_theta:int=10000, max_context:int=1024):
+    self.vocab_size = vocab_size
+    self.n_layers = n_layers
+    self.n_heads = n_heads
+    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
+    self.head_dim = dim // n_heads
+    self.n_rep = self.n_heads // self.n_kv_heads
+    self.hidden_dim = hidden_dim
+
+    scaled_std = 0.02 / math.sqrt(2 * n_layers)
+
+    # Attention
+    self._init_inv_scales = []  # populated by lin_per_layer
+    self.wqkv = self.lin_per_layer(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2)
+    self.wo = self.lin_per_layer(self.n_heads * self.head_dim, dim, std=scaled_std)
+
+    # FeedForward
+    self.w13 = self.lin_per_layer(dim, hidden_dim * 2)
+    self.w2 = self.lin_per_layer(hidden_dim, dim, std=scaled_std)
+
+    self.norm_eps = norm_eps
+    self.attention_norm = Tensor.ones(n_layers, dim).contiguous()
+    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
+
+    # output
+    self.norm = nn.RMSNorm(dim, norm_eps)
+    self.tok_embeddings = nn.Embedding(vocab_size, dim)
+    self.tok_embeddings.weight = Tensor.normal(vocab_size, dim, mean=0.0, std=0.02, dtype=dtypes.bfloat16)
+    self.output = Tensor.normal(1, vocab_size, dim, mean=0.0, std=0.02, dtype=dtypes.bfloat16)
+    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
+
+    def _amax(): return Tensor.full((), FP8_MAX, dtype=dtypes.float32).contiguous().requires_grad_(False)
+    names = ["xqkv", "xo", "x13", "x2"]
+    self._fp8_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
+    grad_names = ["xqkv", "xo", "xw13", "xout"]
+    self._fp8_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
+    w_names = ["wqkv", "wo", "w13", "w2"]
+    self._fp8_inv_scale = {wname: inv_scales.float().contiguous().requires_grad_(False)
+                           for wname, inv_scales in zip(w_names, self._init_inv_scales)}
+    del self._init_inv_scales
+
+  def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02):
+    if getenv("ZEROS"): w = Tensor.zeros(self.n_layers, out_features, in_features)
+    else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std)
+    amax = w.abs().flatten(1).max(1).detach()
+    scale = FP8_MAX / (amax + 1e-8)
+    self._init_inv_scales.append((amax + 1e-8) / FP8_MAX)
+    return (w * scale.reshape(-1, 1, 1)).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE)
+
+  def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
+                amax_xqkv:Tensor, amax_xo:Tensor, s_qkv:Tensor, s_o:Tensor,
+                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor):
+    bsz, seqlen, _ = x.shape
+    new_amaxs, saves = [], []
+
+    xqkv, x_normed, rrms, ret = norm_quantize_matmul(x, attention_norm, wqkv, s_qkv, self.norm_eps,
+                                                     amax_x=amax_xqkv, grad_amax_state=grad_amax_xqkv)
+    saves.extend([x_normed, rrms])
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [xqkv])
+    xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
+    xq = xqkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
+    xk = xqkv[:, :, :, self.n_rep].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
+    xv = xqkv[:, :, :, self.n_rep+1].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
+
+    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
+    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)
+    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
+    if getenv("HK_FLASH_ATTENTION"):
+      from extra.thunder.amd.fa import flash_attention
+      attn, *save = flash_attention(xq, xk, xv, is_causal=True)
+      saves.extend(save)
+    else:
+      attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True)
+    attn = attn.transpose(1, 2).reshape(bsz, seqlen, -1)
+
+    out, *ret = matmul(attn, wo, amax_x=amax_xo, w_inv_scale=s_o, grad_amax_state=grad_amax_xo)
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [out])
+    return (out, *new_amaxs, *saves)
+
+  def feed_forward(self, x:Tensor, residual:Tensor, ffn_norm:Tensor, w13:Tensor, w2:Tensor,
+                   amax_x13:Tensor, amax_x2:Tensor, s_13:Tensor, s_2:Tensor,
+                   grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
+    new_amaxs, saves = [], []
+
+    x_w13, h, x_normed, rrms, ret = add_norm_quantize_matmul(x, residual, ffn_norm, w13, s_13, self.norm_eps,
+                                                             amax_x=amax_x13)
+    saves.extend([x_normed, rrms])
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [x_w13])
+
+    out, ret = silu_w13_quantize_matmul(x_w13, w2, s_2, amax_x2=amax_x2, grad_amax_xw13=grad_amax_xw13, grad_amax_xout=grad_amax_xout)
+    new_amaxs.extend(ret[:1])
+    saves.extend(ret[1:] + [out])
+    return (out, h, *new_amaxs, *saves)
+
+  @function(precompile=True, precompile_backward=True)
+  def run_layer(self, x:Tensor, freqs_cis:Tensor,
+                attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
+                ffn_norm:Tensor, w13:Tensor, w2:Tensor,
+                amax_xqkv:Tensor, amax_xo:Tensor,
+                amax_x13:Tensor, amax_x2:Tensor,
+                s_qkv:Tensor, s_o:Tensor, s_13:Tensor, s_2:Tensor,
+                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor,
+                grad_amax_xw13:Tensor, grad_amax_xout:Tensor):
+    attn, *attn_ret = self.attention(x, freqs_cis, attention_norm, wqkv, wo,
+                                     amax_xqkv=amax_xqkv, amax_xo=amax_xo, s_qkv=s_qkv, s_o=s_o,
+                                     grad_amax_xqkv=grad_amax_xqkv, grad_amax_xo=grad_amax_xo)
+    attn_amaxs, attn_saves = attn_ret[:2], attn_ret[2:]
+    ffn, h, *ffn_ret = self.feed_forward(x, attn, ffn_norm, w13, w2,
+                                                   amax_x13=amax_x13, amax_x2=amax_x2, s_13=s_13, s_2=s_2,
+                                                   grad_amax_xw13=grad_amax_xw13, grad_amax_xout=grad_amax_xout)
+    ffn_amaxs, ffn_saves = ffn_ret[:2], ffn_ret[2:]
+    h = h + ffn
+    return (h, *attn_amaxs, *ffn_amaxs, *attn_saves, *ffn_saves)
+
+  def shard(self, device:tuple[str, ...], mp:bool=False):
+    from tinygrad.nn.state import get_parameters
+    if not mp:
+      for v in get_parameters(self): v.shard_(device, axis=None)
+    else:
+      # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
+      self.wqkv.shard_(device, axis=1).realize()          # (n_layers, out, dim) shard out
+      self.wo.shard_(device, axis=2).realize()            # (n_layers, dim, in) shard in
+      self.w13.shard_(device, axis=1).realize()           # (n_layers, hidden*2, dim) shard out
+      self.w2.shard_(device, axis=2).realize()            # (n_layers, dim, hidden) shard in
+      self.attention_norm.shard_(device, axis=None).realize()
+      self.ffn_norm.shard_(device, axis=None).realize()
+      self.norm.weight.shard_(device, axis=None).realize()
+      self.tok_embeddings.weight.shard_(device, axis=0).realize()
+      self.output.shard_(device, axis=1).realize()
+      self.freqs_cis.shard_(device, axis=None).realize()
+      for amax_dict in (self._fp8_amax, self._fp8_grad_amax):
+        for name in amax_dict:
+          for i in range(len(amax_dict[name])):
+            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().requires_grad_(False)
+      for name in self._fp8_inv_scale:
+        self._fp8_inv_scale[name] = self._fp8_inv_scale[name].to(device).contiguous().requires_grad_(False)
+
+  def __call__(self, tokens:Tensor):
+    h = self.tok_embeddings(tokens)
+    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
+    a, ga, s = self._fp8_amax, self._fp8_grad_amax, self._fp8_inv_scale
+    for i in range(self.n_layers):
+      h, *ret = self.run_layer(h, freqs_cis,
+                               self.attention_norm[i], self.wqkv[i], self.wo[i],
+                               self.ffn_norm[i], self.w13[i], self.w2[i],
+                               amax_xqkv=a["xqkv"][i], amax_xo=a["xo"][i],
+                               amax_x13=a["x13"][i], amax_x2=a["x2"][i],
+                               s_qkv=s["wqkv"][i], s_o=s["wo"][i],
+                               s_13=s["w13"][i], s_2=s["w2"][i],
+                               grad_amax_xqkv=ga["xqkv"][i], grad_amax_xo=ga["xo"][i],
+                               grad_amax_xw13=ga["xw13"][i], grad_amax_xout=ga["xout"][i])
+      for name, new_val in zip(["xqkv", "xo", "x13", "x2"], ret[:5]):
+        a[name][i].assign(new_val)
+
+    logits = matmul(self.norm(h), self.output[0], fp8=False)[0]
+    return logits
+
+def _get_pads(uop:UOp) -> list[UOp]:
+  if uop.op == Ops.ADD: return _get_pads(uop.src[0]) + _get_pads(uop.src[1])
+  return [uop]
+
+def apply_grad(grad_buf:Tensor, new_grad:UOp):
+  pads = _get_pads(new_grad)
+  new_grad = new_grad.cast(grad_buf.dtype)
+  if len(pads) <= 1:
+    store = grad_buf.uop.store(grad_buf.uop + new_grad)
+    grad_buf.uop = grad_buf.uop.after(store)
+    return
+  sorted_pads = sorted(pads, key=lambda p: p.marg[0][0] if p.op == Ops.PAD else 0)
+  inners = [Tensor(p.src[0] if p.op == Ops.PAD else p, device=grad_buf.device).cast(grad_buf.dtype) for p in sorted_pads]
+  if getenv("FUSED_PAD_GRAD_ACCUM", 0):
+    from extra.llama_kernels.fused_pad_grad_accum import fused_pad_grad_accum, can_fused_pad_grad_accum
+    if can_fused_pad_grad_accum(grad_buf, inners):
+      grad_buf.uop = fused_pad_grad_accum(grad_buf, inners).uop
+      return
+  grad_buf.assign(grad_buf + inners[0].cat(*inners[1:], dim=0))
+
+if __name__ == "__main__":
+  config = {}
+  BS                 = config["BS"]                     = getenv("BS", 16)
+  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
+
+  from examples.llama3 import MODEL_PARAMS
+  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
+  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
+  model = FlatTransformer(**model_params, max_context=SEQLEN)
+  state = nn.state.get_state_dict(model)
+  print("tensor count:", len(state))
+
+  # shard the model
+  from tinygrad import Device
+  if (DP := getenv("DP", 1)) > 1:
+    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)))
+  if (MP := getenv("MP", 1)) > 1:
+    model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)), mp=True)
+
+  # preallocate all the grad buffers and zero them out
+  grads = {x:Tensor.zeros(x.shape, dtype=x.dtype, device=x.device).contiguous()
+           for x in state.values() if x.requires_grad is None}
+
+  # print model size
+  sz = 0
+  for k,v in state.items():
+    print(f"{colored(k, 'green' if v in grads else 'white'):30s} {str(v.shape):30s} {str(v.dtype):20s} {v.device}  {v.nbytes()/1e9:.2f} GB")
+    sz += v.nbytes()
+  print(f"total sz: {sz/1e9:.2f} GB")
+
+  with Timing("fake data: "): tokens = Tensor.randint(BS, SEQLEN+1, low=0, high=model.vocab_size, dtype=dtypes.int)
+  with Timing("realize weights/grads/data: "): Tensor.realize(*state.values(), *grads.values(), tokens)
+  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
+  if DP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(DP)), axis=0)
+  if MP > 1: tokens = tokens.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)))
+
+  @TinyJit
+  def jit_step(tokens:Tensor):
+    with Timing("python forward: "): loss = model(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
+    with Timing("python backward: "):
+      for t,g in zip(grads, loss.gradient(*grads)):
+        apply_grad(grads[t], g.uop)
+    with Timing("run step: "): loss.realize(*grads.values())
+
+  for i in range(6):
+    GlobalCounters.reset()
+    profile_marker(f"step {i}")
+    with Timing(colored(f"*** step {i}: ", "red")):
+      jit_step(tokens)
+  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
@@ -0,0 +1,140 @@
+import os
+os.environ["WQKV"] = "1"
+import unittest
+import numpy as np
+from tinygrad import Tensor, nn, dtypes
+from tinygrad.nn.state import get_parameters
+from tinygrad.device import is_dtype_supported, Device
+from examples.mlperf.models.llama import Transformer
+from examples.mlperf.models.flat_llama import FlatTransformer
+
+def copy_weights(flat:FlatTransformer, ref:Transformer):
+  n_layers = flat.n_layers
+  Tensor.realize(*nn.state.get_state_dict(ref).values())
+  flat.wqkv.assign(Tensor(np.stack([ref.layers[i].attention.wqkv.weight.numpy() for i in range(n_layers)])))
+  flat.wo.assign(Tensor(np.stack([ref.layers[i].attention.wo.weight.numpy() for i in range(n_layers)])))
+  flat.w1.assign(Tensor(np.stack([ref.layers[i].feed_forward.w1.weight.numpy() for i in range(n_layers)])))
+  flat.w2.assign(Tensor(np.stack([ref.layers[i].feed_forward.w2.weight.numpy() for i in range(n_layers)])))
+  flat.w3.assign(Tensor(np.stack([ref.layers[i].feed_forward.w3.weight.numpy() for i in range(n_layers)])))
+  flat.attention_norm.assign(Tensor(np.stack([ref.layers[i].attention_norm.weight.numpy() for i in range(n_layers)])))
+  flat.ffn_norm.assign(Tensor(np.stack([ref.layers[i].ffn_norm.weight.numpy() for i in range(n_layers)])))
+  flat.norm.weight.assign(Tensor(ref.norm.weight.numpy()))
+  flat.tok_embeddings.weight.assign(Tensor(ref.tok_embeddings.weight.numpy()))
+  flat.output.weight.assign(Tensor(ref.output.weight.numpy()))
+
+class TestFlatLlama(unittest.TestCase):
+  def test_forward_match(self):
+    Tensor.manual_seed(42)
+    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
+    ref = Transformer(**params)
+    flat = FlatTransformer(**params)
+    copy_weights(flat, ref)
+    Tensor.realize(*nn.state.get_state_dict(flat).values())
+
+    tokens = Tensor([[1, 50, 100, 999, 2]])
+    ref_logits = ref(tokens).realize()
+    flat_logits = flat(tokens).realize()
+    self.assertEqual(ref_logits.shape, flat_logits.shape)
+    diff = (ref_logits - flat_logits).abs().max().item()
+    self.assertLess(diff, 1e-5, f"forward mismatch: max abs diff {diff}")
+
+  def test_backward_match(self):
+    Tensor.manual_seed(42)
+    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
+    ref = Transformer(**params)
+    flat = FlatTransformer(**params)
+    copy_weights(flat, ref)
+
+    for p in get_parameters(ref): p.requires_grad_(True)
+    for p in get_parameters(flat): p.requires_grad_(True)
+    Tensor.realize(*nn.state.get_state_dict(flat).values())
+
+    tokens = Tensor([[1, 50, 100, 999, 2, 10]])
+
+    ref_loss = ref(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
+    ref_loss.backward()
+    ref_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(ref).items() if v.grad is not None}
+
+    flat_loss = flat(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
+    flat_loss.backward()
+    flat_grads = {k: v.grad.numpy() for k, v in nn.state.get_state_dict(flat).items() if v.grad is not None}
+
+    # check loss matches
+    self.assertAlmostEqual(ref_loss.item(), flat_loss.item(), places=4)
+
+    # check output weight grad matches
+    diff = abs(ref_grads["output.weight"] - flat_grads["output.weight"]).max()
+    self.assertLess(diff, 1e-4, f"output.weight grad mismatch: max abs diff {diff}")
+
+    # check per-layer weight grads match
+    for i in range(params["n_layers"]):
+      for flat_key, ref_key in [
+        ("wqkv", f"layers.{i}.attention.wqkv.weight"),
+        ("wo", f"layers.{i}.attention.wo.weight"),
+        ("w1", f"layers.{i}.feed_forward.w1.weight"),
+        ("w2", f"layers.{i}.feed_forward.w2.weight"),
+        ("w3", f"layers.{i}.feed_forward.w3.weight"),
+      ]:
+        diff = abs(ref_grads[ref_key] - flat_grads[flat_key][i]).max()
+        self.assertLess(diff, 1e-4, f"layer {i} {flat_key} grad mismatch: max abs diff {diff}")
+
+  @unittest.skipUnless(Device.DEFAULT == "CPU", "multi-device CPU test")
+  def test_forward_match_mp(self):
+    Tensor.manual_seed(42)
+    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
+    from tinygrad import Device
+    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
+    ref = Transformer(**params)
+    flat = FlatTransformer(**params)
+    copy_weights(flat, ref)
+    Tensor.realize(*nn.state.get_state_dict(flat).values())
+    flat.shard(devices, mp=True)
+
+    tokens = Tensor([[1, 50, 100, 999, 2]], device=devices[0])
+    ref_logits = ref(tokens.to(devices[0])).numpy()
+    flat_logits = flat(tokens.shard(devices)).numpy()
+    self.assertEqual(ref_logits.shape, flat_logits.shape)
+    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
+
+  @unittest.skipUnless(Device.DEFAULT == "CPU", "multi-device CPU test")
+  def test_forward_match_dp(self):
+    Tensor.manual_seed(42)
+    params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
+    from tinygrad import Device
+    devices = (f"{Device.DEFAULT}:0", f"{Device.DEFAULT}:1")
+    ref = Transformer(**params)
+    flat = FlatTransformer(**params)
+    copy_weights(flat, ref)
+    Tensor.realize(*nn.state.get_state_dict(flat).values())
+    flat.shard(devices)
+
+    tokens = Tensor([[1, 50, 100, 999, 2], [2, 100, 50, 1, 999]], device=devices[0])
+    ref_logits = ref(tokens.to(devices[0])).numpy()
+    flat_logits = flat(tokens.shard(devices, axis=0)).numpy()
+    self.assertEqual(ref_logits.shape, flat_logits.shape)
+    np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
+
+  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3), "fp8 not supported on this device")
+  def test_forward_fp8(self):
+    import examples.mlperf.models.flat_llama as flat_llama_mod
+    old_fp8 = flat_llama_mod.FP8
+    try:
+      flat_llama_mod.FP8 = 1
+      Tensor.manual_seed(42)
+      params = dict(dim=128, hidden_dim=256, n_heads=4, n_kv_heads=2, n_layers=2, norm_eps=1e-5, vocab_size=1024, rope_theta=10000, max_context=64)
+      ref = Transformer(**params)
+      flat = FlatTransformer(**params)
+      copy_weights(flat, ref)
+      Tensor.realize(*nn.state.get_state_dict(flat).values())
+
+      tokens = Tensor([[1, 50, 100, 999, 2]])
+      ref_logits = ref(tokens).numpy()
+      flat_logits = flat(tokens).numpy()
+      self.assertEqual(ref_logits.shape, flat_logits.shape)
+      # FP8 has lower precision, allow larger tolerance
+      np.testing.assert_allclose(flat_logits, ref_logits, atol=1.0, rtol=0.1)
+    finally:
+      flat_llama_mod.FP8 = old_fp8
+
+if __name__ == "__main__":
+  unittest.main()
@@ -1,24 +1,90 @@
 from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes
-from tinygrad.nn.optim import LAMB
-from tinygrad.helpers import FUSE_OPTIM
+from tinygrad.nn.optim import Optimizer
+from tinygrad.helpers import FUSE_OPTIM, getenv
+from tinygrad.uop.ops import UOp, Ops

-class GradAccClipAdamW(LAMB):
-  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, fused=FUSE_OPTIM):
-    super().__init__(params, lr, b1, b2, eps, weight_decay, adam=True, fused=FUSE_OPTIM)
+STOCHASTIC_ROUND = getenv("STOCHASTIC_ROUND", 0)
+MASTER_WEIGHTS = getenv("MASTER_WEIGHTS", 0)
+
+def stochastic_round_bf16(x:Tensor) -> Tensor:
+  bits = x.bitcast(dtypes.uint32)
+  if isinstance(x.device, tuple):
+    shape = x.uop.shard_shape if x.uop.axis is not None else x.shape
+    noise = Tensor(UOp(Ops.MSTACK, dtypes.default_float, tuple(Tensor.rand(*shape, device=d).uop for d in x.device)))
+  else:
+    noise = x.rand_like()
+  noise = (noise * 0xFFFF).cast(dtypes.uint32)
+  return ((bits + noise) & 0xFFFF0000).bitcast(dtypes.float32).cast(dtypes.bfloat16)
+
+class GradAccClipAdamW(Optimizer):
+  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, device=None, fused=FUSE_OPTIM):
+    super().__init__(params, lr, device, fused)
+    self.b1, self.b2, self.eps, self.wd = b1, b2, eps, weight_decay
+    self.b1_t, self.b2_t = (Tensor.ones((1,), dtype=dtypes.float32, device=self.device, requires_grad=False) for _ in [b1, b2])
+    self.m = self._new_optim_param()
+    self.v = self._new_optim_param()
    self.grad_acc, self.clip_norm = grad_acc, clip_norm
+    self.master_params:list[Tensor]|None = [p.float().contiguous() for p in self.params] if MASTER_WEIGHTS and self.params[0].dtype != dtypes.float32 else None
+
+  def fstep(self, grads:list[Tensor]):
+    if self.fused:
+      out, extra = self._step([], grads)
+      updates = [out[0][self.pos_params[i]:self.pos_params[i+1]].reshape(tt.shape) for i, tt in enumerate(self.params)]
+    else:
+      updates, extra = self._step([], grads)
+    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i], self.master_params[i] if self.master_params else None))
+    # collect inv_scale tensors attached to fp8 params (set by _apply_update)
+    fp8_inv_scales = [tt._inv_scale for tt in self.params if hasattr(tt, '_inv_scale')]
+    to_realize = extra+self.params+self.buffers+(self.master_params or [])+fp8_inv_scales
+
+    Tensor.realize(*to_realize)
+    return extra[-1]

  def _step(self, params:list[Tensor], grads:list[Tensor]) -> tuple[list[Tensor], list[Tensor]]:
+    grads = list(grads)
+
+    for i in range(len(grads)):
+      if grads[i].device != self.m[i].device: grads[i] = grads[i].to(self.m[i].device)
+
    if self.fused:
-      grads[0] = grads[0] / self.grad_acc
+      grads[0].assign(grads[0] / self.grad_acc)
      total_norm = grads[0].float().square().sum().sqrt()
-      grads[0] = (grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype)
+      grads[0].assign((grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype))
    else:
-      total_norm = Tensor.zeros((), dtype=dtypes.float32, device=self.device)
-      for g in grads:
-        total_norm += g.float().square().sum()
-      total_norm = total_norm.sqrt()
      for i in range(len(grads)):
-        grads[i] = grads[i] / self.grad_acc
-        grads[i] = (grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype)
-    return super()._step(params, grads)
+        grads[i].assign(grads[i] / self.grad_acc)
+      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
+      for i in range(len(grads)):
+        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype))
+
+    ret = []
+    self.b1_t *= self.b1
+    self.b2_t *= self.b2
+    for i, g in enumerate(grads):
+      m_new = self.b1 * self.m[i].float() + (1.0 - self.b1) * g.float()
+      v_new = self.b2 * self.v[i].float() + (1.0 - self.b2) * (g.float() * g.float())
+      self.m[i].assign(m_new.cast(self.m[i].dtype))
+      self.v[i].assign(v_new.cast(self.v[i].dtype))
+      m_hat = m_new / (1.0 - self.b1_t)
+      v_hat = v_new / (1.0 - self.b2_t)
+      up = m_hat / (v_hat.sqrt() + self.eps)
+      ret.append(self.lr * up)
+    return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]
+
+  def _apply_update(self, t:Tensor, up:Tensor, master:Tensor|None=None) -> Tensor:
+    w = master if master is not None else t
+    wd = self.wd if t.ndim >= 3 else 0.0
+    up = up.float().shard_like(w) + self.lr.to(w.device) * wd * w.detach()
+    new_w = w.detach() - up
+    if master is not None: master.assign(new_w)
+    if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16: return stochastic_round_bf16(new_w)
+    if t.dtype in dtypes.fp8s:
+      from examples.mlperf.models.flat_llama import FP8_MAX
+      amax = new_w.float().abs().flatten(1).max(1).detach()  # per-layer amax for (n_layers, out, in)
+      scale = FP8_MAX / (amax + 1e-8)
+      fp8_w = (new_w * scale.reshape(-1, *([1]*(new_w.ndim-1)))).clamp(-FP8_MAX, FP8_MAX).cast(t.dtype)
+      if hasattr(t, '_inv_scale'):
+        t._inv_scale.assign(((amax + 1e-8) / FP8_MAX).cast(t._inv_scale.dtype))
+      return fp8_w
+    return new_w.cast(t.dtype)
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128

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

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024

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

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_8xMI300X"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

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

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

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

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

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

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="resnet"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

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

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="resnet"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -1,7 +1,7 @@
 #!/bin/bash
 set -e  # Exit on any error

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="retinanet"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128

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

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_8xMI300X"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="resnet"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="resnet"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="retinanet"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="retinanet"
 export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
 export BASEDIR="/raid/datasets/openimages"
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128

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

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_8xMI300X"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
@@ -1,6 +1,6 @@
 #!/bin/bash

-export PYTHONPATH="." AMD=1 DEBUG=0 JIT=1 FLASH_ATTENTION=1
+export PYTHONPATH="." DEV=AMD DEBUG=0 JIT=1 FLASH_ATTENTION=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024

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

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_8xMI350X"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." NV=1
+export PYTHONPATH="." DEV=NV
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
@@ -1,6 +1,6 @@
 #!/bin/bash

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

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

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

@@ -2,7 +2,7 @@
 set -e  # Exit on any error
 set -o pipefail  # Make pipeline fail if any command fails

-export PYTHONPATH="." AMD=1
+export PYTHONPATH="." DEV=AMD
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_red"
 export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export DEV=${DEV:-AMD}
+export EMULATE="AMD_CDNA4"
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-0}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-1} MP=${MP:-8}
+export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4/"
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export FAKEDATA=1 BENCHMARK=10
+if [ -z "$FULL_LAYERS" ]; then
+  export LLAMA_LAYERS=2
+fi
+
+python3 examples/mlperf/model_train.py
@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export DEV=${DEV:-AMD}
+export EMULATE="AMD_CDNA4"
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-0}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-1} MP=${MP:-8}
+export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1152}
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4/"
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+python3 examples/mlperf/model_train.py
@@ -2,18 +2,27 @@

 export PYTHONPATH="."
 export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export FAST_CE=${FASE_CE:-1}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}

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

 export MODEL="llama3"
@@ -21,7 +30,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -32,7 +41,7 @@ export DATA_SEED=${DATA_SEED:-5760}
 export JITBEAM=${JITBEAM:-3}
 export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1

-export FAKEDATA=1 BENCHMARK=10
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
  export LLAMA_LAYERS=2
 fi
@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export DEV=${DEV:-AMD}
+export EMULATE="AMD_CDNA4"
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-2}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-0}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-5760}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export FAKEDATA=1 BENCHMARK=10
+if [ -z "$FULL_LAYERS" ]; then
+  export LLAMA_LAYERS=2
+fi
+
+python3 examples/mlperf/model_train.py
@@ -2,18 +2,27 @@

 export PYTHONPATH="."
 export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
+export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
+export FAST_CE=${FASE_CE:-1}
+export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
+export FUSED_ADD_NORM_MUL_QUANTIZE=${FUSED_ADD_NORM_MUL_QUANTIZE:-1}
+export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
+export FUSED_PAD_GRAD_ACCUM=${FUSED_PAD_GRAD_ACCUM:-1}

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

 export MODEL="llama3"
@@ -21,7 +30,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+
+export PYTHONPATH="."
+export DEV=${DEV:-AMD}
+export EMULATE="AMD_CDNA4"
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export DEBUG=${DEBUG:-0}
+export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ALL2ALL=${ALL2ALL:-1}
+export USE_ATOMICS=${USE_ATOMICS:-0}
+export ASM_GEMM=${ASM_GEMM:-1}
+export WQKV=${WQKV:-1}
+export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-32}
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=${SEQLEN:-8192}
+
+export SEED=${SEED:-$RANDOM}
+export DATA_SEED=${DATA_SEED:-5760}
+
+export JITBEAM=${JITBEAM:-3}
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+python3 examples/mlperf/model_train.py
@@ -1,6 +1,6 @@
 #!/bin/bash
 export BENCHMARK=5
 export EVAL_BS=0
-export VIZ=${VIZ:--1}
-examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
-PYTHONPATH="." extra/viz/cli.py --profile --device "AMD" --top 20
+VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=0 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam.sh
+SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
+python -m tinygrad.viz.cli -s "$SRC" --top 20
@@ -0,0 +1,54 @@
+#!/usr/bin/env bash
+set -e  # Exit on any error
+set -o pipefail  # Make pipeline fail if any command fails
+
+export PYTHONPATH="."
+export DEV=AMD
+export CHECK_OOB=0
+export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
+export DEVICE_IN_FUNCTION_BUG=1
+
+export HK_FLASH_ATTENTION=1
+export ALL2ALL=1
+export USE_ATOMICS=1
+export ASM_GEMM=1
+export WQKV=1
+export MASTER_WEIGHTS=1
+export FP8=1
+export ALLREDUCE_CAST=1
+export FAST_CE=1
+export FUSED_INPUT_QUANTIZE=1
+export FUSED_ADD_NORM_MUL_QUANTIZE=1
+export FUSED_SILU_W13=1
+export FUSED_PAD_GRAD_ACCUM=1
+
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=8 MP=1 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=2
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=8B
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
+export SAMPLES=$((MAX_STEPS * GBS))
+export SEQLEN=8192
+
+export SEED=$RANDOM
+export DATA_SEED=$SEED
+
+export JITBEAM=3
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
+
+export LOGMLPERF=1
+
+DATETIME=$(date "+%m%d%H%M")
+LOGFILE="llama31_8b_8xMI350x_${DATETIME}_${SEED}.log"
+
+# beam
+FAKEDATA=1 BENCHMARK=10 INITMLPERF=1 LLAMA_LAYERS=2 python3 examples/mlperf/model_train.py | tee "$LOGFILE"
+
+# run
+RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a "$LOGFILE"
@@ -1,6 +1,6 @@
 #!/bin/bash

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

--- a/Show More
+++ b/Show More