Merge branch 'master' into rdna4_gemm

* delete that

* remove

* delete graph config

.github/actions/setup-tinygrad/action.yml

@@ -54,7 +54,7 @@ runs:
   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 }}
 
@@ -70,7 +70,7 @@ runs:
     - 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 }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
@@ -85,7 +85,7 @@ runs:
         key: downloads-${{ 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 }}
@@ -202,7 +202,7 @@ runs:
         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: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
@@ -269,7 +269,7 @@ 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:

.github/workflows/autogen.yml

@@ -28,7 +28,7 @@ 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:
@@ -57,6 +57,7 @@ jobs:
         python3 -c "from tinygrad.runtime.autogen import mesa"
         python3 -c "from tinygrad.runtime.autogen import avcodec"
         python3 -c "from tinygrad.runtime.autogen import llvm_qcom"
+        python3 -c "from tinygrad.runtime.autogen import mlx5"
         REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
     - name: Check for differences
       run: |
@@ -68,7 +69,7 @@ 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
@@ -79,7 +80,7 @@ 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:
@@ -99,7 +100,7 @@ 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
@@ -110,7 +111,7 @@ 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:
@@ -138,7 +139,7 @@ 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

.github/workflows/benchmark.yml

@@ -32,7 +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: |
@@ -49,7 +49,40 @@ jobs:
         source /tmp/tinygrad_pytest_ci/bin/activate
         pytest -nauto --durations=20
     - name: openpilot compile3 0.10.1 driving_vision
-      run: FLOAT16=1 CL=1 IMAGE=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+      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 other stuff
+      run: |
+        mkdir -p extra/remu/target/release/
+        ln -s ~/tinygrad/extra/remu/target/release/libremu.so extra/remu/target/release/libremu.so
+    - name: setup staging db
+      run: |
+        echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
+        rm -f /tmp/pytest-db-ci*
+    - name: Run pytest -nauto
+      run: |
+        source /tmp/tinygrad_pytest_ci/bin/activate
+        pytest -nauto --durations=20
 
   testmacbenchmark:
     name: Mac Benchmark
@@ -64,7 +97,7 @@ 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
@@ -94,17 +127,17 @@ 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)
@@ -112,7 +145,7 @@ 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
@@ -154,7 +187,7 @@ 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: |
@@ -174,7 +207,7 @@ 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: |
@@ -185,17 +218,19 @@ 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=. GMMU=0 DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
+      run: sudo -E PYTHONPATH=. 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=. GMMU=0 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=. GMMU=0 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=. GMMU=0 AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+    #  run: sudo -E PYTHONPATH=. 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
@@ -207,7 +242,7 @@ 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
@@ -227,67 +262,67 @@ 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: |
@@ -305,7 +340,7 @@ 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
@@ -325,35 +360,35 @@ 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 ASSERT_FPS=1400 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=120 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
 
@@ -367,7 +402,7 @@ 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
@@ -409,18 +444,18 @@ 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
@@ -429,43 +464,43 @@ 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
 
@@ -479,7 +514,7 @@ 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
@@ -504,20 +539,20 @@ jobs:
     - name: reset process replay
       run: test/external/process_replay/reset.py
     - name: Train MNIST
-      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+      run: time PYTHONPATH=. 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=230 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
+      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
@@ -534,7 +569,7 @@ 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
@@ -559,14 +594,14 @@ 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
 
@@ -580,7 +615,7 @@ 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: |
@@ -591,7 +626,7 @@ jobs:
     - name: openpilot compile3 0.11.0 driving_vision
       run: BENCHMARK_LOG=openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
     - name: IR3 openpilot compile3 0.11.0 driving_vision
-      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM QCOM_IR3=1 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
+      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=3 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
     - name: openpilot compile3 0.11.0 dmonitoring
@@ -609,9 +644,9 @@ jobs:
         # 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
 
@@ -625,16 +660,16 @@ 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: |
         echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
     - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." GMMU=0 DEV=AMD AMD_LLVM=1 AMD_IFACE=USB ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
+      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=AMD AMD_IFACE=USB ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
+      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
 
   testreddriverbenchmark:
     name: AM Benchmark
@@ -646,7 +681,7 @@ 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
@@ -671,39 +706,39 @@ 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
     # 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 AMD=1 AMD_IFACE=PCI python3 test/test_tiny.py
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD=1 AMD_AQL=1 AMD_IFACE=PCI python3 test/test_tiny.py
+        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
@@ -718,7 +753,7 @@ 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
@@ -743,30 +778,30 @@ 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 NV=1 python3 test/test_tiny.py
+        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

.github/workflows/benchmark_search.yml

@@ -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

.github/workflows/docs.yml

@@ -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

.github/workflows/mlperf.yml

@@ -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

.github/workflows/python-publish.yml

@@ -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

.github/workflows/szdiff.yml

@@ -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

.github/workflows/test.yml

@@ -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:
@@ -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 MAX_BUFFER_SIZE=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
+      run: DEV=PYTHON python3 -m pytest -rA test/backend/test_renderer_failures.py::TestRendererFailures
     - name: Test IMAGE support
       run: |
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_gemm
-        IMAGE=1 PYTHON=1 python3 test/backend/test_ops.py TestOps.test_simple_conv2d
+        IMAGE=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:
@@ -251,7 +251,7 @@ jobs:
 
     steps:
     - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
@@ -261,19 +261,19 @@ jobs:
         llvm: 'true'
         amd: 'true'
     - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
-    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+      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: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+    #   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: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
+      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: 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
+      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: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
+    #  run: DEV=NULL python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
 
   unittest:
     name: Unit Tests
@@ -282,7 +282,7 @@ jobs:
 
     steps:
     - name: Checkout Code
-      uses: actions/checkout@v4
+      uses: actions/checkout@v6
     - name: Setup Environment
       uses: ./.github/actions/setup-tinygrad
       with:
@@ -297,8 +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
+        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
@@ -325,7 +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:
@@ -341,7 +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:
@@ -362,7 +362,7 @@ 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:
@@ -371,9 +371,8 @@ jobs:
           opencl: 'true'
       - name: Test CL IMAGE=1 ops
         run: |
-          CL=1 IMAGE=1 python -m pytest -n=auto test/backend/test_ops.py --durations=20
-        # TODO: training is broken
-        # CL=1 IMAGE=1 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
 
@@ -383,7 +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:
@@ -391,13 +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
@@ -408,7 +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:
@@ -418,13 +417,13 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1486 ALLOWED_GATED_READ_IMAGE=17 FLOAT16=1 CL=1 IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=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=1 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: CL=1 IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
+        run: 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: IMAGE=1 FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+        run: 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
 
@@ -437,7 +436,7 @@ jobs:
 
     steps:
       - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Setup Environment
         uses: ./.github/actions/setup-tinygrad
         with:
@@ -446,15 +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
 
@@ -464,7 +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:
@@ -474,21 +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
 
@@ -500,7 +499,7 @@ 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:
@@ -521,7 +520,7 @@ 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:
@@ -530,11 +529,11 @@ 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
 
@@ -544,7 +543,7 @@ 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:
@@ -552,9 +551,9 @@ 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 ******
 
@@ -564,7 +563,7 @@ 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:
@@ -573,11 +572,11 @@ 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)
@@ -585,7 +584,7 @@ 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:
@@ -594,9 +593,9 @@ 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
@@ -607,11 +606,11 @@ 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)
@@ -619,7 +618,7 @@ 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:
@@ -629,11 +628,11 @@ 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
 
@@ -642,12 +641,12 @@ jobs:
     runs-on: ubuntu-24.04
     timeout-minutes: 20
     env:
-      AMD: 1
+      DEV: AMD
       PYTHON_REMU: 1
       MOCKGPU: 1
     steps:
       - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Setup Environment
         uses: ./.github/actions/setup-tinygrad
         with:
@@ -668,33 +667,34 @@ jobs:
       - name: Install rocprof-trace-decoder
         run: sudo PYTHONPATH="." ./extra/sqtt/install_rocprof_decoder.py
       - name: Run AMD renderer tests
-        run: AMD_LLVM=0 python -m pytest -n=auto test/amd/ --durations 20
-      - name: Run AMD renderer tests (AMD_LLVM=1)
-        run: AMD_LLVM=1 python -m pytest -n=auto test/amd/ --durations 20
+        run: python -m pytest -n=auto test/amd/ --durations 20
+      - name: Run AMD renderer tests (AMD:LLVM)
+        run: DEV=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::gfx1100 python extra/mmapeak/mmapeak.py
+          PYTHONPATH=. DEV=NULL::gfx1201 python3 -m pytest -n=auto test/testextra/test_tk.py test/backend/test_asm_gemm.py
+      - name: Run matmul on MOCKGPU
+        run: |
+          PYTHONPATH="." DEV=AMD MOCKGPU=1 N=256 python3 extra/gemm/amd_asm_matmul.py
+          PYTHONPATH="." DEV=AMD MOCKGPU=1 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=AMD:LLVM python test/device/test_amd_llvm.py
 
   testmockam:
     name: Linux (am)
     runs-on: ubuntu-24.04
     timeout-minutes: 15
     env:
-      AMD: 1
+      DEV: PCI+AMD
       MOCKGPU: 1
-      AMD_IFACE: PCI
     steps:
       - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Setup Environment
         uses: ./.github/actions/setup-tinygrad
         with:
@@ -704,29 +704,35 @@ jobs:
       - name: Run test_tiny on MOCKAM
         run: python test/test_tiny.py
       - name: Run test_tiny on MOCKAM USB
-        run: GMMU=0 AMD_IFACE=USB python test/test_tiny.py
+        run: GMMU=0 DEV=USB+AMD python test/test_tiny.py
       - name: Run test_hcq on MOCKAM
         run: python -m pytest test/device/test_hcq.py
+      - name: Run disk copy tests on MOCKAM
+        run: python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk
+      - name: Run test_tiny on MOCKAM 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, cdna4]
+        arch: [gfx1100, gfx1201, gfx950]
 
     name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
     runs-on: ubuntu-22.04
     timeout-minutes: 15
     env:
-      AMD: 1
+      DEV: AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
       MOCKGPU: 1
-      MOCKGPU_ARCH: ${{ matrix.arch }}
       SKIP_SLOW_TEST: 1
-      AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
     steps:
       - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Setup Environment
         uses: ./.github/actions/setup-tinygrad
         with:
@@ -740,6 +746,8 @@ jobs:
           DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
       - name: Run pytest (amd)
         run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
+      - name: Run disk copy tests
+        run: python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk
       - name: Run TRANSCENDENTAL math
         run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
       - name: Run process replay tests
@@ -759,7 +767,7 @@ jobs:
       FORWARD_ONLY: 1
     steps:
       - name: Checkout Code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Setup Environment
         uses: ./.github/actions/setup-tinygrad
         with:
@@ -768,7 +776,7 @@ jobs:
           cuda: 'true'
           ocelot: 'true'
       - name: Set env
-        run: printf "${{ 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=CUDA:PTX' || matrix.backend == 'nv' && 'DEV=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"
@@ -792,7 +800,7 @@ 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:
@@ -802,7 +810,7 @@ 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"
@@ -822,7 +830,7 @@ 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:
@@ -834,44 +842,41 @@ 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: METAL=1 python3 -m pytest test/device/test_metal.py
+      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: 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: "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: "NV:PTX"
         FORWARD_ONLY: 1
         # TODO: failing due to library loading error
         CAPTURE_PROCESS_REPLAY: 0
@@ -886,7 +891,7 @@ 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:
@@ -894,9 +899,9 @@ 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
@@ -915,7 +920,7 @@ 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:
@@ -927,7 +932,7 @@ 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:
@@ -936,7 +941,7 @@ 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"
@@ -962,7 +967,7 @@ 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:
@@ -971,7 +976,7 @@ 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'}}" >> $GITHUB_ENV
       - name: Run unit tests
         if: matrix.backend=='llvm'
         # test_newton_schulz hits RecursionError
@@ -979,7 +984,7 @@ jobs:
       - name: Run NULL backend tests
         if: matrix.backend=='llvm'
         shell: bash
-        run: CPU=0 CPU_LLVM=0 NULL=1 python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
+        run: DEV=NULL python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
       - name: Run pytest (${{ matrix.backend }})
         shell: bash
         run: |
@@ -998,7 +1003,7 @@ 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:
@@ -1008,7 +1013,7 @@ 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: |
@@ -1021,7 +1026,7 @@ 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:
@@ -1031,7 +1036,7 @@ jobs:
           python-version: '3.12'
       - name: Set env
         shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\nNULL_QCOMCL=1" >> $GITHUB_ENV
+        run: printf "DEV=NULL:QCOMCL:a630\nNULL_ALLOW_COPYOUT=1" >> $GITHUB_ENV
       - name: Run test_ops
         shell: bash
         run: |

docs/developer/am.md

@@ -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:
 

docs/env_vars.md

@@ -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.
 
@@ -31,12 +31,7 @@ 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
@@ -48,7 +43,23 @@ ALLOW_TF32          | [1]        | enable TensorFloat-32 tensor cores on Ampere
 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
+
+### Debug breakdown
 
 Variable | Value | Description
 ---|---|---

docs/runtime.md

@@ -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` |
 | [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) |
 
 
@@ -72,7 +72,7 @@ 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:

docs/tinygpu.md

@@ -0,0 +1,59 @@
+# 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 tinygrad/apps/llm.py
+```

examples/compile_efficientnet.py

@@ -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))

examples/llm.c/export.py

@@ -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.uop.ops import Ops
 
+DEV.value = "CPU"
+
 TIMING = getenv("TIMING")
 
 if __name__ == "__main__":

examples/mlperf/model_eval.py

@@ -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,

examples/mlperf/model_train.py

@@ -13,8 +13,6 @@ from extra.bench_log import BenchEvent, WallTimeEvent
 # TODO: fix benchmark logging and use tinygrad tqdm
 from tqdm import tqdm
 
-from tinygrad.uop.ops import UOp
-
 def train_resnet():
   from extra.models import resnet
   from examples.mlperf.dataloader import batch_load_resnet
@@ -1284,11 +1282,14 @@ def train_bert():
         previous_step = i
 
 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad
+  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8
   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 = {}
@@ -1311,15 +1312,60 @@ 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_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
@@ -1348,12 +1394,10 @@ def train_llama3():
   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"):
     for v in get_parameters(model):
-      v = v.assign(Tensor.empty(v.shape))
+      v = v.assign(Tensor.empty(v.shape, dtype=v.dtype))
 
   is_dp = (DP := getenv("DP", 1)) > 1
   is_mp = (MP := getenv("MP", 1)) > 1
@@ -1373,7 +1417,9 @@ def train_llama3():
                            eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device)
 
   # init grads
-  grads = [Tensor.zeros_like(p).contiguous() for p in optim.params]
+  for p in optim.params:
+    p.grad = Tensor.zeros(p.shape, dtype=p.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)
 
@@ -1386,6 +1432,8 @@ 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] if FP8 else []
+
   @TinyJit
   def minibatch(tokens:Tensor):
     if is_dp: tokens = tokens.to(None).shard(device, 0)
@@ -1394,11 +1442,11 @@ def train_llama3():
     logits:Tensor = model(tokens[:, :-1])
     loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
 
-    for i,(t,g) in enumerate(zip(optim.params, loss.gradient(*optim.params))):
-      grads[i].replace(Tensor(grads[i].uop.after(UOp.group(*apply_grad(grads[i].uop, g.uop))), device=t.device))
+    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)
+    return loss_cpu.realize(*grads, *fp8_amax)
 
   @TinyJit
   def optim_step():
@@ -1453,6 +1501,11 @@ 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
@@ -1491,7 +1544,7 @@ def train_llama3():
 
       mem_gb = GlobalCounters.mem_used / 1e9
       gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * max(getenv("DP", 1), getenv("MP", 1)) * 2.3e15)) * 100
+      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 2.3e15)) * 100
       tqdm.write(
           f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
           f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
@@ -1535,6 +1588,10 @@ def train_llama3():
       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()
@@ -1544,22 +1601,34 @@ def train_llama3():
         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.event(key=mllog_constants.TRAIN_SAMPLES, value=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

examples/mlperf/models/flat_llama.py

@@ -18,21 +18,32 @@ from tinygrad.uop.ops import Ops, UOp
 from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
 
 FP8 = getenv("FP8", 0)
+WQKV = getenv("WQKV", 0)
 
 FP8_DTYPE = dtypes.fp8e4m3
+FP8_GRAD_DTYPE = dtypes.fp8e5m2
 FP8_MAX = 448.0
 
-def quantize_fp8(x:Tensor):
-  scale = FP8_MAX / (x.abs().max().detach() + 1e-8)
+def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
+  if amax_state is not None:
+    scale = FP8_MAX / (amax_state + 1e-8)
+    amax_state.assign(x.abs().max().detach())
+  else:
+    scale = FP8_MAX / (x.abs().max().detach() + 1e-8)
   x_scaled = x * scale
   x_clamped = x_scaled + (x_scaled.detach().clamp(-FP8_MAX, FP8_MAX) - x_scaled.detach())  # STE
   return x_clamped.cast(FP8_DTYPE), scale.float().reciprocal()
 
-def matmul(x:Tensor, w:Tensor) -> Tensor:
-  if not FP8: return x @ w.T
-  # weights are already FP8, just quantize activations
-  x_fp8, x_scale = quantize_fp8(x)
-  return x_fp8.dot(w.T, dtype=dtypes.float) * x_scale
+def matmul(x:Tensor, w:Tensor, fp8=FP8, amax_x:Tensor|None=None, amax_w:Tensor|None=None) -> Tensor:
+  if not fp8: return x @ w.T
+  from tinygrad.helpers import ASM_GEMM
+  x_fp8, x_scale = quantize_fp8(x, amax_state=amax_x)
+  w_fp8, w_scale = quantize_fp8(w, amax_state=amax_w)
+  combined_scale = x_scale * w_scale
+  if ASM_GEMM:
+    from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
+    if can_use_asm_gemm(x_fp8, w_fp8.T): return asm_gemm(x_fp8, w_fp8.T, combined_scale=combined_scale)
+  return x_fp8.dot(w_fp8.T, dtype=dtypes.float) * combined_scale
 
 def rmsnorm(x_in:Tensor, eps:float):
   x = x_in.float()
@@ -49,13 +60,20 @@ class FlatTransformer:
     self.head_dim = dim // n_heads
     self.n_rep = self.n_heads // self.n_kv_heads
 
+    scaled_std = 0.02 / math.sqrt(2 * n_layers)
+
     # Attention
-    self.wqkv = self.lin_per_layer(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2)
-    self.wo = self.lin_per_layer(self.n_heads * self.head_dim, dim)
+    if WQKV:
+      self.wqkv = self.lin_per_layer(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2)
+    else:
+      self.wq = self.lin_per_layer(dim, self.n_heads * self.head_dim)
+      self.wk = self.lin_per_layer(dim, self.n_kv_heads * self.head_dim)
+      self.wv = self.lin_per_layer(dim, self.n_kv_heads * self.head_dim)
+    self.wo = self.lin_per_layer(self.n_heads * self.head_dim, dim, std=scaled_std)
 
     # FeedForward
     self.w1 = self.lin_per_layer(dim, hidden_dim)
-    self.w2 = self.lin_per_layer(hidden_dim, dim)
+    self.w2 = self.lin_per_layer(hidden_dim, dim, std=scaled_std)
     self.w3 = self.lin_per_layer(dim, hidden_dim)
 
     self.norm_eps = norm_eps
@@ -65,44 +83,71 @@ class FlatTransformer:
     # output
     self.norm = nn.RMSNorm(dim, norm_eps)
     self.tok_embeddings = nn.Embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False)
+    self.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 lin_per_layer(self, in_features:int, out_features:int):
-    bound = 1 / math.sqrt(in_features)
-    dt = FP8_DTYPE if FP8 else None
-    if getenv("ZEROS"): return Tensor.zeros(self.n_layers, out_features, in_features, dtype=dt)
-    return Tensor.uniform(self.n_layers, out_features, in_features, low=-bound, high=bound, dtype=dt)
+    if FP8:
+      def _amax(): return Tensor.full((), FP8_MAX).contiguous().requires_grad_(False)
+      names = (["xqkv", "wqkv"] if WQKV else ["xq", "wq", "xk", "wk", "xv", "wv"]) + \
+              ["xo", "wo", "x1", "w1", "x2", "w2", "x3", "w3"]
+      # _fp8_amax[name][layer_idx] = scalar amax tensor
+      self._fp8_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
+      self._fp8_amax["xout"] = [_amax()]
+      self._fp8_amax["wout"] = [_amax()]
 
-  def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wqkv:Tensor, wo:Tensor):
+  def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02):
+    if getenv("ZEROS"): return Tensor.zeros(self.n_layers, out_features, in_features)
+    return Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std)
+
+  def attention(self, x:Tensor, freqs_cis:Tensor, attention_norm:Tensor, wo:Tensor, wqkv:Tensor|None=None,
+                wq:Tensor|None=None, wk:Tensor|None=None, wv:Tensor|None=None,
+                amax_xqkv=None, amax_wqkv=None, amax_xq=None, amax_wq=None, amax_xk=None, amax_wk=None,
+                amax_xv=None, amax_wv=None, amax_xo=None, amax_wo=None):
     x = rmsnorm(x, self.norm_eps) * attention_norm
-    xqkv = matmul(x, wqkv)
+    bsz, seqlen, _ = x.shape
 
-    bsz, seqlen, _ = xqkv.shape
-    # interleaved layout: each kv group has [n_rep q heads, 1 k head, 1 v head] for clean MP sharding
-    xqkv = xqkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
-    xq = xqkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
-    xk = xqkv[:, :, :, self.n_rep].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
-    xv = xqkv[:, :, :, self.n_rep+1].reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
+    if wqkv is not None:
+      xqkv = matmul(x, wqkv, amax_x=amax_xqkv, amax_w=amax_wqkv)
+      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)
+    else:
+      assert wq is not None and wk is not None and wv is not None
+      xq = matmul(x, wq, amax_x=amax_xq, amax_w=amax_wq).reshape(bsz, seqlen, self.n_heads, self.head_dim)
+      xk = matmul(x, wk, amax_x=amax_xk, amax_w=amax_wk).reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
+      xv = matmul(x, wv, amax_x=amax_xv, amax_w=amax_wv).reshape(bsz, seqlen, self.n_kv_heads, self.head_dim)
 
     xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
+    if FP8: xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)
     xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
     attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
     attn = attn.reshape(bsz, seqlen, -1)
-    return matmul(attn, wo)
+    return matmul(attn, wo, amax_x=amax_xo, amax_w=amax_wo)
 
-  def feed_forward(self, x:Tensor, ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
+  def feed_forward(self, x:Tensor, ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor,
+                   amax_x1=None, amax_w1=None, amax_x2=None, amax_w2=None, amax_x3=None, amax_w3=None):
     x = rmsnorm(x, self.norm_eps) * ffn_norm
-    x_w1 = matmul(x, w1).silu()
-    x_w3 = matmul(x.contiguous_backward(), w3)
-    return matmul(x_w1 * x_w3, w2)
+    x_w1 = matmul(x, w1, amax_x=amax_x1, amax_w=amax_w1).silu()
+    x_w3 = matmul(x.contiguous_backward(), w3, amax_x=amax_x3, amax_w=amax_w3)
+    return matmul(x_w1 * x_w3, w2, amax_x=amax_x2, amax_w=amax_w2)
 
   @function(precompile=True, precompile_backward=True)
   def run_layer(self, x:Tensor, freqs_cis:Tensor,
-                attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor):
-    h = x + self.attention(x, freqs_cis, attention_norm, wqkv, wo)
-    return h + self.feed_forward(h, ffn_norm, w1, w2, w3)
+                attention_norm:Tensor, wo:Tensor,
+                ffn_norm:Tensor, w1:Tensor, w2:Tensor, w3:Tensor,
+                wqkv:Tensor|None=None, wq:Tensor|None=None, wk:Tensor|None=None, wv:Tensor|None=None,
+                amax_xqkv=None, amax_wqkv=None, amax_xq=None, amax_wq=None, amax_xk=None, amax_wk=None,
+                amax_xv=None, amax_wv=None, amax_xo=None, amax_wo=None,
+                amax_x1=None, amax_w1=None, amax_x2=None, amax_w2=None, amax_x3=None, amax_w3=None):
+    h = x + self.attention(x, freqs_cis, attention_norm, wo, wqkv=wqkv, wq=wq, wk=wk, wv=wv,
+                           amax_xqkv=amax_xqkv, amax_wqkv=amax_wqkv, amax_xq=amax_xq, amax_wq=amax_wq,
+                           amax_xk=amax_xk, amax_wk=amax_wk, amax_xv=amax_xv, amax_wv=amax_wv,
+                           amax_xo=amax_xo, amax_wo=amax_wo)
+    return h + self.feed_forward(h, ffn_norm, w1, w2, w3,
+                                 amax_x1=amax_x1, amax_w1=amax_w1, amax_x2=amax_x2, amax_w2=amax_w2,
+                                 amax_x3=amax_x3, amax_w3=amax_w3)
 
   def shard(self, device:tuple[str, ...], mp:bool=False):
     from tinygrad.nn.state import get_parameters
@@ -110,7 +155,12 @@ class FlatTransformer:
       for v in get_parameters(self): v.shard_(device, axis=None)
     else:
       # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
-      self.wqkv.shard_(device, axis=1).realize()          # (n_layers, out, dim) shard out
+      if WQKV:
+        self.wqkv.shard_(device, axis=1).realize()          # (n_layers, out, dim) shard out
+      else:
+        self.wq.shard_(device, axis=1).realize()            # (n_layers, n_heads*head_dim, dim) shard out
+        self.wk.shard_(device, axis=1).realize()            # (n_layers, n_kv_heads*head_dim, dim) shard out
+        self.wv.shard_(device, axis=1).realize()            # (n_layers, n_kv_heads*head_dim, dim) shard out
       self.wo.shard_(device, axis=2).realize()             # (n_layers, dim, in) shard in
       self.w1.shard_(device, axis=1).realize()             # (n_layers, hidden, dim) shard out
       self.w2.shard_(device, axis=2).realize()             # (n_layers, dim, hidden) shard in
@@ -119,28 +169,47 @@ class FlatTransformer:
       self.ffn_norm.shard_(device, axis=None).realize()
       self.norm.weight.shard_(device, axis=None).realize()
       self.tok_embeddings.weight.shard_(device, axis=0).realize()
-      self.output.weight.shard_(device, axis=0).realize()
+      self.output.shard_(device, axis=1).realize()
       self.freqs_cis.shard_(device, axis=None).realize()
 
   def __call__(self, tokens:Tensor):
     h = self.tok_embeddings(tokens)
     freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
+    a = self._fp8_amax if FP8 else None
     for i in range(self.n_layers):
+      if WQKV:
+        attn_kwargs = {"wqkv": self.wqkv[i]}
+        amax_attn = {"amax_xqkv": a["xqkv"][i], "amax_wqkv": a["wqkv"][i]} if a else {}
+      else:
+        attn_kwargs = {"wq": self.wq[i], "wk": self.wk[i], "wv": self.wv[i]}
+        amax_attn = {"amax_xq": a["xq"][i], "amax_wq": a["wq"][i],
+                     "amax_xk": a["xk"][i], "amax_wk": a["wk"][i],
+                     "amax_xv": a["xv"][i], "amax_wv": a["wv"][i]} if a else {}
+      amax_layer = {"amax_xo": a["xo"][i], "amax_wo": a["wo"][i],
+                    "amax_x1": a["x1"][i], "amax_w1": a["w1"][i],
+                    "amax_x2": a["x2"][i], "amax_w2": a["w2"][i],
+                    "amax_x3": a["x3"][i], "amax_w3": a["w3"][i]} if a else {}
       h = self.run_layer(h, freqs_cis,
-                         self.attention_norm[i], self.wqkv[i], self.wo[i],
-                         self.ffn_norm[i], self.w1[i], self.w2[i], self.w3[i])
-    logits = self.output(self.norm(h))
+                         self.attention_norm[i], self.wo[i],
+                         self.ffn_norm[i], self.w1[i], self.w2[i], self.w3[i],
+                         **attn_kwargs, **amax_attn, **amax_layer)
+    logits = (self.norm(h).contiguous().contiguous_backward() @ self.output[0].T).contiguous_backward()
     return logits
 
-# TODO: this shouldn't be needed, but it prevents a copy of the grads. CAT can help
-def apply_grad(old_grad:UOp, new_grad:UOp) -> list[UOp]:
-  if new_grad.op == Ops.ADD:
-    return apply_grad(old_grad, new_grad.src[0])+apply_grad(old_grad, new_grad.src[1])
-  elif new_grad.op == Ops.PAD:
-    grad_shrink = tuple([(p[0], s+p[0]) for s,p in zip(new_grad.src[0].shape, new_grad.marg)])
-    return apply_grad(old_grad.shrink(grad_shrink), new_grad.src[0])
-  else:
-    return [old_grad.store(old_grad + new_grad)]
+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]
+  grad_buf.assign(grad_buf + inners[0].cat(*inners[1:], dim=0))
 
 if __name__ == "__main__":
   config = {}
@@ -162,7 +231,8 @@ if __name__ == "__main__":
     model.shard(tuple(f"{Device.DEFAULT}:{i}" for i in range(MP)), mp=True)
 
   # preallocate all the grad buffers and zero them out
-  grads = {x:Tensor.zeros_like(x).contiguous() for x in state.values() if x.requires_grad is None}
+  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
@@ -182,7 +252,7 @@ if __name__ == "__main__":
     with Timing("python forward: "): loss = model(tokens[:, :-1]).sparse_categorical_crossentropy(tokens[:, 1:])
     with Timing("python backward: "):
       for t,g in zip(grads, loss.gradient(*grads)):
-        grads[t] = Tensor(grads[t].uop.after(UOp.group(*apply_grad(grads[t].uop, g.uop))), device=t.device)
+        apply_grad(grads[t], g.uop)
     with Timing("run step: "): loss.realize(*grads.values())
 
   for i in range(6):

examples/mlperf/models/test_flat_llama.py

@@ -4,7 +4,7 @@ import unittest
 import numpy as np
 from tinygrad import Tensor, nn, dtypes
 from tinygrad.nn.state import get_parameters
-from tinygrad.device import is_dtype_supported
+from tinygrad.device import is_dtype_supported, Device
 from examples.mlperf.models.llama import Transformer
 from examples.mlperf.models.flat_llama import FlatTransformer
 
@@ -78,7 +78,7 @@ class TestFlatLlama(unittest.TestCase):
         diff = abs(ref_grads[ref_key] - flat_grads[flat_key][i]).max()
         self.assertLess(diff, 1e-4, f"layer {i} {flat_key} grad mismatch: max abs diff {diff}")
 
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
+  @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)
@@ -96,7 +96,7 @@ class TestFlatLlama(unittest.TestCase):
     self.assertEqual(ref_logits.shape, flat_logits.shape)
     np.testing.assert_allclose(flat_logits, ref_logits, atol=1e-4, rtol=1e-4)
 
-  @unittest.skipUnless(os.getenv("CPU", "") == "1", "multi-device CPU test")
+  @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)

examples/mlperf/optim.py

@@ -1,7 +1,21 @@
 from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes
 from tinygrad.nn.optim import Optimizer
-from tinygrad.helpers import FUSE_OPTIM
+from tinygrad.helpers import FUSE_OPTIM, getenv
+from tinygrad.uop.ops import UOp, Ops
+
+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):
@@ -11,6 +25,7 @@ class GradAccClipAdamW(Optimizer):
     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:
@@ -18,8 +33,8 @@ class GradAccClipAdamW(Optimizer):
       updates = [out[0][self.pos_params[i]:self.pos_params[i+1]].reshape(tt.shape) for i, tt in enumerate(self.params)]
     else:
       updates, extra = self._step([], grads)
-    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i]))
-    to_realize = extra+self.params+self.buffers
+    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))
+    to_realize = extra+self.params+self.buffers+(self.master_params or [])
 
     Tensor.realize(*to_realize)
     return extra[-1]
@@ -45,15 +60,21 @@ class GradAccClipAdamW(Optimizer):
     self.b1_t *= self.b1
     self.b2_t *= self.b2
     for i, g in enumerate(grads):
-      self.m[i].assign((self.b1 * self.m[i] + (1.0 - self.b1) * g).cast(self.m[i].dtype))
-      self.v[i].assign((self.b2 * self.v[i] + (1.0 - self.b2) * (g * g)).cast(self.v[i].dtype))
-      m_hat = (self.m[i] / (1.0 - self.b1_t)).cast(self.m[i].dtype)
-      v_hat = (self.v[i] / (1.0 - self.b2_t)).cast(self.v[i].dtype)
+      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).cast(g.dtype))
+      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) -> Tensor:
+  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.shard_like(t) + self.lr.to(t.device) * wd * t.detach()
-    return t.detach() - up.cast(t.dtype)
+    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)
+    return new_w.cast(t.dtype)

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_1xMI300X/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_beam.sh

@@ -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

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/resnet/implementations/tinybox_green/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/resnet/implementations/tinybox_green/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/resnet/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/dev_beam.sh

@@ -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"

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/dev_run.sh

@@ -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"

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_red/dev_beam.sh

@@ -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"

examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_red/dev_run.sh

@@ -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"

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_1xMI300X/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_beam.sh

@@ -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

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_green/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_green/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/retinanet/implementations/tinybox_green/dev_beam.sh

@@ -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"

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/retinanet/implementations/tinybox_green/dev_run.sh

@@ -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"

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/retinanet/implementations/tinybox_red/dev_beam.sh

@@ -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"

examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/retinanet/implementations/tinybox_red/dev_run.sh

@@ -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"

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_1xMI300X/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_beam.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI350X/dev_run_fa.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI350X/dev_run_fp8.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI350X/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_green/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_red/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -12,10 +12,12 @@ export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
+export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-4}
 export GBS=$((BS * GRADIENT_ACC_STEPS))
 
 export MODEL="llama3"
@@ -34,7 +36,7 @@ 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=1 BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
   export LLAMA_LAYERS=2
 fi

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -12,7 +12,9 @@ export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}
+export WQKV=${WQKV:-1}
+export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
+export FP8=${FP8:-1}
 
 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-4}

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/profile.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
 export BENCHMARK=5
 export EVAL_BS=0
-VIZ=${VIZ:--1} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
-extra/viz/cli.py --profile --device "AMD" --limit 20
+VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=0 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_beam.sh
+extra/viz/cli.py --profile -s "${DEV:-AMD}"

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/run_and_time.sh

@@ -0,0 +1,49 @@
+#!/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 EMULATE="AMD_CDNA4"
+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 DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DP=8 MP=1 BS=8 EVAL_BS=8 GRADIENT_ACC_STEPS=4
+export GBS=$((BS * GRADIENT_ACC_STEPS))
+
+export MODEL="llama3"
+export BASEDIR="/raid/datasets/c4-8b/"
+export SMALL=1
+export LLAMA3_SIZE=8B
+export EVAL_TARGET=3.3 EVAL_FREQ=12288
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export WARMUP_STEPS=$((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"

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/resnet/implementations/tinybox_green/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/resnet/implementations/tinybox_green/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/resnet/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/dev_beam.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/dev_run.sh

@@ -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
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/dev_beam.sh

@@ -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"

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/dev_run.sh

@@ -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"

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/retinanet/implementations/tinybox_red/dev_beam.sh

@@ -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"

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/retinanet/implementations/tinybox_red/dev_run.sh

@@ -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"

examples/mlperf/training_submission_v6.0/tinycorp/systems/tinybox_8xMI350X.json

@@ -0,0 +1,38 @@
+{
+    "submitter": "tinycorp",
+    "division": "closed",
+    "status": "Available on-premise",
+    "system_name": "tinybox 8xMI350X",
+    "number_of_nodes": "1",
+    "host_processors_per_node": "2",
+    "host_processor_model_name": "AMD EPYC 9575F",
+    "host_processor_core_count": "32",
+    "host_processor_vcpu_count": "64",
+    "host_processor_frequency": "",
+    "host_processor_caches": "",
+    "host_processor_interconnect": "",
+    "host_memory_capacity": "3072 GiB",
+    "host_storage_type": "NVMe SSD",
+    "host_storage_capacity": "4TB",
+    "host_networking": "",
+    "host_networking_topology": "",
+    "host_memory_configuration": "24x 128GB DDR5",
+    "accelerators_per_node": "8",
+    "accelerator_model_name": "AMD Instinct MI350X 288GB HBM3e",
+    "accelerator_host_interconnect": "PCIe 5.0 x16",
+    "accelerator_frequency": "",
+    "accelerator_on-chip_memories": "",
+    "accelerator_memory_configuration": "HBM3",
+    "accelerator_memory_capacity": "288GB",
+    "accelerator_interconnect": "",
+    "accelerator_interconnect_topology": "",
+    "cooling": "air",
+    "hw_notes": "",
+    "framework": "tinygrad, branch mlperf_training_v6.0",
+    "other_software_stack": {
+        "python": "3.12.3",
+        "ROCm": "7.1.1"
+    },
+    "operating_system": "Ubuntu 24.04.3 LTS",
+    "sw_notes": ""
+  }

examples/tinychat/tinychat-browser/compile.py

@@ -3,7 +3,7 @@ from extra.export_model import export_model
 from examples.llama3 import build_transformer, Tokenizer
 from tinygrad.nn.state import get_state_dict, load_state_dict
 from tinygrad import Device, Variable, Tensor, dtypes, TinyJit
-from tinygrad.helpers import fetch, Context
+from tinygrad.helpers import DEV, fetch, Context
 from tiktoken.load import load_tiktoken_bpe, dump_tiktoken_bpe
 
 def prepare_browser_chunks(model):
@@ -115,7 +115,7 @@ if __name__=="__main__":
   start_pos = Variable("start_pos", 0, max_context).bind(0)
   model_input = lambda: [Tensor([[tok]]), start_pos, TEMPERATURE, TOP_K, TOP_P, ALPHA_F, ALPHA_P]
 
-  Device.DEFAULT="CPU"
+  DEV.value = "CPU"
   model = build_transformer(model_path, model_size="1B", quantize="int8", scale_dtype=dtypes.float32, device=Device.DEFAULT, max_context=max_context)
   state_dict = get_state_dict(model)
   validate_model(model, tokenizer)
@@ -129,7 +129,7 @@ if __name__=="__main__":
   with open(os.path.join(os.path.dirname(__file__), f"{model_name}.c"), "w") as f: f.write(cprog)
   with open(os.path.join(os.path.dirname(__file__), "net_clang.js"), "w") as f: f.write(js_wrapper)
 
-  Device.DEFAULT="WEBGPU"
+  DEV.value = "WEBGPU"
   # float16 is not yet supported for dawn/Vulkan/NVIDIA stack, see: https://issues.chromium.org/issues/42251215
   # therefore for now, we used CLANG to quantize the float16 llama to int8 with float32 scales, then load to WEBGPU
   model = build_transformer(model_path, model_size="1B", quantize="int8", max_context=max_context, load_weights=False)

examples/webgpu/stable_diffusion/compile.py

@@ -4,8 +4,8 @@ from extra.f16_decompress import u32_to_f16
 from examples.stable_diffusion import StableDiffusion
 from tinygrad.nn.state import get_state_dict, safe_save, safe_load_metadata, torch_load, load_state_dict
 from tinygrad.tensor import Tensor
-from tinygrad import Device, dtypes
-from tinygrad.helpers import fetch
+from tinygrad import dtypes
+from tinygrad.helpers import DEV, fetch
 from typing import NamedTuple, Any, List
 import requests
 import argparse
@@ -80,7 +80,7 @@ if __name__ == "__main__":
   parser = argparse.ArgumentParser(description='Run Stable Diffusion', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
   parser.add_argument('--remoteweights', action='store_true', help="Use safetensors from Huggingface, or from local")
   args = parser.parse_args()
-  Device.DEFAULT = "WEBGPU"
+  DEV.value = "WEBGPU"
 
   model = StableDiffusion()
 

examples/webgpu/yolov8/compile.py

@@ -4,10 +4,11 @@ from tinygrad.tensor import Tensor
 from tinygrad.nn.state import safe_save
 from extra.export_model import export_model
 from tinygrad.device import Device
+from tinygrad.helpers import DEV
 from tinygrad.nn.state import safe_load, load_state_dict
 
 if __name__ == "__main__":
-    Device.DEFAULT = "WEBGPU"
+    DEV.value = "WEBGPU"
     yolo_variant = 'n'
     yolo_infer = YOLOv8(w=0.25, r=2.0, d=0.33, num_classes=80)
     state_dict = safe_load(get_weights_location(yolo_variant))

extra/gemm/amd_asm_matmul.py

@@ -441,9 +441,9 @@ THREADS = 128
 
 def test_matmul():
   dev = Device[Device.DEFAULT]
-  print(f"Device arch: {dev.renderer.arch}")
+  print(f"Device arch: {dev.renderer.target.arch}")
 
-  insts = build_kernel(N, dev.renderer.arch)
+  insts = build_kernel(N, dev.renderer.target.arch)
 
   rng = np.random.default_rng(42)
   a = Tensor(rng.random((N, N), dtype=np.float32) - 0.5)

extra/gemm/amd_copy_matmul.py

@@ -61,7 +61,7 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
 
   # accumulator (unified: both paths use (TM, TN) with scalar dtypes.float)
   acc = UOp.placeholder((TM, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(UOp.const(dtypes.float, 0).reshape((1,)*len(acc.shape)).expand(acc.shape)))
+  acc = acc.after(acc.store(acc.zeros_like()))
 
   if use_wmma:
     k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)

extra/gemm/cdna_asm_gemm.py

@@ -1,10 +1,12 @@
-import atexit, functools
+import atexit, functools, pathlib
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.dtype import AddrSpace
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
 from tinygrad.renderer import Estimates
 from tinygrad.helpers import getenv, all_same, DEBUG
+from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from tinygrad.runtime.autogen.amd.cdna.ins import *
+from examples.mlperf.models.flat_llama import FP8_DTYPE, FP8_GRAD_DTYPE, matmul, quantize_fp8
 
 # ** CDNA4 assembly gemm
 
@@ -2623,6 +2625,26 @@ def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str) -> UOp:
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname),
                                 UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
 
+# ** FP8 GEMM custom kernel
+
+@functools.cache
+def custom_hk_fp8_gemm(C:UOp, A:UOp, B:UOp, S:UOp, dname:str) -> UOp:
+  # A is (batch, M, K), B is (N, K) transposed, S is combined scale (scalar float)
+  M, K = A.shape[0]*A.shape[1], A.shape[2]
+  N, K2 = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2, f"{A.shape} {B.shape}"
+  block_size = 256
+  threads = UOp.special(64 * 8, "lidx0")
+  workgroups = UOp.special((M // block_size) * (N // block_size), "gidx0")
+  sink = UOp.sink(C.base, A.base, B.base, S.base, threads, workgroups,
+                  arg=KernelInfo(f"hk_fp8_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K, mem=(M*K+N*K)*A.dtype.itemsize+M*N*C.dtype.itemsize)))
+  kittens_path = pathlib.Path(__file__).parent.parent/"thunder"/"amd"
+  src = (kittens_path/"gemm_fp8.cpp").read_text()
+  lib = HIPCCCompiler("gfx950", [f"-I{(kittens_path/'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-ffast-math",
+                                 "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DGEMM_M={M}", f"-DGEMM_N={N}", f"-DGEMM_K={K}"]).compile_cached(src)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
+                               UOp(Ops.BINARY, arg=lib)))
+
 counters = {"used":0, "todos":[]}
 def todo(msg:str) -> bool: counters["todos"].append(msg); return False
 def _asm_gemm_report():
@@ -2634,7 +2656,7 @@ atexit.register(_asm_gemm_report)
 
 def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
   if a.dtype != b.dtype: return todo(f"dtypes must match {a.dtype} != {b.dtype}")
-  if a.dtype not in {dtypes.bfloat16, dtypes.float16}: return todo(f"only bfloat16/float16, got {a.dtype}")
+  if a.dtype not in {dtypes.bfloat16, dtypes.float16, FP8_DTYPE}: return todo(f"only bfloat16/float16/fp8, got {a.dtype}")
   batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
   N = b.shape[1]
   if isinstance(a.device, tuple):
@@ -2647,7 +2669,7 @@ def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
     else: return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
     dname = a.device[0]
   else: dname = a.device
-  arch = getattr(Device[dname].renderer, "arch", "")
+  arch = Device[dname].renderer.target.arch
   if batch not in {1, 2}: return todo(f"GEMM batch size {batch}")
   # blacklist slow matmul
   # TODO: why is this slow?
@@ -2675,18 +2697,32 @@ def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
 # ** backward gemm, might use the asm gemm
 
 def custom_gemm_bw(gradient:UOp, kernel:UOp):
-  out, a, b = kernel.src[1:]
-  assert all_same([gradient.device, a.device, b.device, out.device])
-  a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
-  # TODO: this needs to be cleaned up and done properly, the batch dim of grad and a multi need to align
-  g_t = g_t[:a.shape[0]]
-  grad_a = (g_t @ b_t.T).uop
-  grad_b = (a_t.permute(2, 0, 1).reshape(a_t.shape[2], -1) @ g_t.reshape(-1, g_t.shape[-1])).uop
-  return (None, grad_a, grad_b)
+  inputs = kernel.src[1:]
+  # fp8 scaled gemm has 4 inputs (out, a, b, scale), others have 3 (out, a, b)
+  if len(inputs) == 4:
+    out, a, b, scale = inputs
+    a_t, b_t, g_t, s_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device), Tensor(scale, device=a.device)
+    g_t = g_t[:a.shape[0]]
+    # backward GEMMs in fp8 with scale applied inside kernel to prevent bf16 overflow
+    g_fp8, g_scale = quantize_fp8(g_t)
+    bw_scale = g_scale * s_t
+    # dgrad: g_fp8 @ weight (asm_gemm computes a@b)
+    grad_a = asm_gemm(g_fp8, b_t, combined_scale=bw_scale)
+    # wgrad: g_fp8.T @ activation = (N, batch*seq) @ (batch*seq, K) → use permute to preserve sharding
+    grad_b = asm_gemm(g_fp8.permute(2, 0, 1).reshape(g_t.shape[-1], -1), a_t.reshape(-1, a_t.shape[-1]), combined_scale=bw_scale)
+    return (None, grad_a.uop, grad_b.uop, None)
+  else:
+    out, a, b = inputs
+    assert all_same([gradient.device, a.device, b.device, out.device])
+    a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
+    g_t = g_t[:a.shape[0]]
+    grad_a = (g_t @ b_t.T).uop
+    grad_b = (a_t.permute(2, 0, 1).reshape(a_t.shape[2], -1) @ g_t.reshape(-1, g_t.shape[-1])).uop
+    return (None, grad_a, grad_b)
 
 # ** main gemm function
 
-def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
+def asm_gemm(a:Tensor, b:Tensor, combined_scale:Tensor|None=None) -> Tensor:
   assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
   counters["used"] += 1
   unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
@@ -2695,6 +2731,7 @@ def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
     a = a.reshape(a.shape[0]*a.shape[1], a.shape[2])
   squeeze = a.ndim == 2
   if squeeze: a = a.unsqueeze(0)
+  out_dtype = dtypes.bfloat16 if a.dtype == FP8_DTYPE else a.dtype
 
   batch, M, K = a.shape
   N = b.shape[1]
@@ -2705,19 +2742,24 @@ def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
 
   if is_multi:
     if n_sharded:
-      out = Tensor(Tensor.invalid(batch, M, N//len(a.device), dtype=a.dtype, device=a.device).uop.multi(2), device=a.device)
+      out = Tensor(Tensor.invalid(batch, M, N//len(a.device), dtype=out_dtype, device=a.device).uop.multi(2), device=a.device)
     elif m_sharded:
-      out = Tensor(Tensor.invalid(batch, M, N, dtype=a.dtype, device=a.device).uop.multi(1), device=a.device)
+      out = Tensor(Tensor.invalid(batch, M, N, dtype=out_dtype, device=a.device).uop.multi(1), device=a.device)
     else:
-      out = Tensor(Tensor.invalid(batch//len(a.device) if a.uop.axis==0 else batch, M, N, dtype=a.dtype, device=a.device).uop.multi(0),
+      out = Tensor(Tensor.invalid(batch//len(a.device) if a.uop.axis==0 else batch, M, N, dtype=out_dtype, device=a.device).uop.multi(0),
                    device=a.device)
   else:
-    out = Tensor.invalid(batch, M, N, dtype=a.dtype, device=a.device)
+    out = Tensor.invalid(batch, M, N, dtype=out_dtype, device=a.device)
 
-  renderer = Device[a.device[0] if is_multi else a.device].renderer
-  dname, arch = renderer.device, getattr(renderer, "arch", "")
+  renderer = Device[dname:=(a.device[0] if is_multi else a.device)].renderer
+  dname, arch = dname.split(":")[0], renderer.target.arch
   if arch.startswith("gfx950") and getenv("USE_ASM", 1):
-    out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname), grad_fxn=custom_gemm_bw)[0]
+    # fp8 gemm computes [email protected], with optional combined scale applied inside kernel before bf16 store
+    if a.dtype == FP8_DTYPE:
+      scale = combined_scale if combined_scale is not None else Tensor(1.0, dtype=dtypes.float, device=a.device)
+      out = Tensor.custom_kernel(out, a, b.T, scale, fxn=functools.partial(custom_hk_fp8_gemm, dname=dname), grad_fxn=custom_gemm_bw)[0]
+    else:
+      out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname), grad_fxn=custom_gemm_bw)[0]
   else:
     out = Tensor.custom_kernel(out, a, b, fxn=custom_uop_gemm, grad_fxn=custom_gemm_bw)[0]
   if k_sharded: out = out.sum(0)

extra/gemm/rdna4_asm_matmul.py

@@ -0,0 +1,245 @@
+# RDNA4 128x128 GEMM using WMMA — optimized DS scheduling
+import numpy as np
+from tinygrad import Tensor, Device, Context, GlobalCounters
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.helpers import getenv, colored
+from tinygrad.dtype import dtypes, AddrSpace
+from tinygrad.engine.realize import Estimates
+from tinygrad.renderer.amd.dsl import s, v, VCC_LO, NULL, src, ttmp
+from tinygrad.runtime.autogen.amd.rdna4.ins import *
+
+BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 16
+TILES_M, TILES_N = 4, 4
+THREADS, ELEM = 128, 2
+LDS_A_ROW = BLOCK_K*ELEM   # 32
+LDS_B_ROW = BLOCK_N*ELEM  # 256
+LDS_A_SIZE = BLOCK_M * LDS_A_ROW  # 4096
+LDS_B_SIZE = BLOCK_K * LDS_B_ROW  # 4096
+LDS_SIZE = LDS_A_SIZE + LDS_B_SIZE  # 8192
+LDS_B_OFF = LDS_A_SIZE
+ACC, DA, DB, FA, FB, ET = 60, 188, 196, 204, 44, 10
+
+def build_kernel(N, arch='gfx1200'):
+  assert N % BLOCK_M == 0 and N >= 256
+  NO_ALU, NO_DS, NO_GLOBAL = getenv("NO_ALU", 0), getenv("NO_DS", 0), getenv("NO_GLOBAL", 0)
+  I, L, B = [], {}, []
+  def e(i): I.append(i); return i
+  def label(n): L[n] = sum(i.size() for i in I)
+  def br(i, t): B.append((len(I)-1, t))
+
+  e(s_load_b128(sdata=s[4:7], sbase=s[0:1], ioffset=0, soffset=NULL))
+  e(s_load_b64(sdata=s[8:9], sbase=s[0:1], ioffset=0x10, soffset=NULL))
+  e(s_wait_kmcnt(simm16=0))
+  e(s_mov_b32(s[10], ttmp[9])); e(s_and_b32(s[11], ttmp[7], 0xFFFF))
+  e(s_lshl_b32(s[10], s[10], 7)); e(s_lshl_b32(s[11], s[11], 7))
+  e(s_mov_b32(s[12], N)); e(s_lshl_b32(s[13], s[12], 1))
+  e(s_mul_i32(s[14], s[12], BLOCK_K*ELEM))
+  e(s_add_co_i32(s[17], s[12], -2*BLOCK_K))  # loop bound
+
+  e(v_and_b32_e32(v[1], 31, v[0])); e(v_lshrrev_b32_e32(v[2], 5, v[0]))
+  e(v_and_b32_e32(v[3], 1, v[2])); e(v_lshrrev_b32_e32(v[2], 1, v[2]))
+
+  e(v_lshlrev_b32_e32(v[4], 5, v[0]))
+  # B store: transposed layout for stride-32 reads. addr = LDS_B_OFF + (tid%8)*512 + (tid/8)*32
+  e(v_and_b32_e32(v[48], 7, v[0])); e(v_lshlrev_b32_e32(v[5], 9, v[48]))   # (tid%8)*512
+  e(v_lshrrev_b32_e32(v[48], 3, v[0])); e(v_lshlrev_b32_e32(v[48], 5, v[48]))  # (tid/8)*32
+  e(v_add_nc_u32_e32(v[5], v[5], v[48])); e(v_add_nc_u32_e32(v[5], LDS_B_OFF, v[5]))
+
+  e(v_add_nc_u32_e32(v[48], s[11], v[0]))
+  e(v_mul_lo_u32(v[6], v[48], N*ELEM)); e(v_mov_b32_e32(v[7], 0))
+  e(v_lshrrev_b32_e32(v[48], 3, v[0])); e(v_mul_lo_u32(v[8], v[48], N*ELEM))
+  e(v_and_b32_e32(v[48], 7, v[0])); e(v_lshlrev_b32_e32(v[48], 5, v[48]))
+  e(v_add_nc_u32_e32(v[8], v[8], v[48]))
+  e(s_mul_i32(s[15], s[10], ELEM)); e(v_add_nc_u32_e32(v[8], s[15], v[8]))
+  e(v_mov_b32_e32(v[9], 0))
+
+  # LDS read addrs with padded strides (eliminates bank conflicts)
+  # A: (lane%16)*LDS_A_ROW + (lane/16)*16 + wave_m*64*LDS_A_ROW
+  # B: (lane%16)*LDS_B_ROW + (lane/16)*16 + wave_n*64*ELEM + LDS_B_OFF
+  LLA, LLB = 40, 43
+  e(v_and_b32_e32(v[50], 15, v[1])); e(v_lshrrev_b32_e32(v[51], 4, v[1]))
+  e(v_lshlrev_b32_e32(v[LLA], 5, v[50]))        # (lane%16) * 32
+  e(v_lshlrev_b32_e32(v[51], 4, v[51]))        # (lane/16) * 16
+  e(v_add_nc_u32_e32(v[LLA], v[LLA], v[51]))
+  e(v_lshlrev_b32_e32(v[52], 11, v[2]))         # wave_m * 2048
+  e(v_add_nc_u32_e32(v[LLA], v[LLA], v[52]))
+  # B read: transposed layout. addr = LDS_B_OFF + (lane%16)*32 + (lane/16)*16 + wave_n*2*512
+  # wave_n selects column panels: wave_n*2 panels (each panel=16 cols, wave_n covers 64 cols = 4 panels)
+  # But wave_n*2*512 = wave_n*1024. Hmm, wave_n covers cols [wave_n*64 : (wave_n+1)*64].
+  # Each panel = 16 cols = 512 bytes. wave_n*64/16 = wave_n*4 panels. Offset = wave_n*4*512 = wave_n*2048.
+  e(v_lshlrev_b32_e32(v[LLB], 5, v[50]))         # (lane%16) * 32 (stride 32!)
+  e(v_add_nc_u32_e32(v[LLB], v[LLB], v[51]))     # + (lane/16)*16
+  e(v_lshlrev_b32_e32(v[52], 11, v[3]))           # wave_n * 2048
+  e(v_add_nc_u32_e32(v[LLB], v[LLB], v[52]))
+  e(v_add_nc_u32_e32(v[LLB], LDS_B_OFF, v[LLB]))
+
+  for i in range(0, 128, 2):
+    e(VOPD(VOPDOp.V_DUAL_MOV_B32, VOPDOp.V_DUAL_MOV_B32, vdstx=v[ACC+i], vdsty=v[ACC+i+1], srcx0=0, srcy0=0))
+  e(s_mov_b32(s[16], 0))
+
+  if not NO_GLOBAL:
+    for i in range(2): e(global_load_b128(vdst=v[DA+i*4:DA+i*4+3], vaddr=v[6:7], saddr=s[4:5], ioffset=i*16))
+    for i in range(2): e(global_load_b128(vdst=v[DB+i*4:DB+i*4+3], vaddr=v[8:9], saddr=s[6:7], ioffset=i*16))
+    e(s_wait_loadcnt(simm16=0))
+  if not NO_DS:
+    for i in range(2): e(ds_store_b128(addr=v[4], data0=v[DA+i*4:DA+i*4+3], offset0=(i*16)&0xFF, offset1=(i*16)>>8))
+    for i in range(2): e(ds_store_b128(addr=v[5], data0=v[DB+i*4:DB+i*4+3], offset0=(i*16)&0xFF, offset1=(i*16)>>8))
+  if not NO_GLOBAL:
+    e(v_add_nc_u32_e32(v[6], BLOCK_K*ELEM, v[6]))
+    e(v_add_nc_u32_e32(v[8], s[14], v[8]))
+
+  # =============================================================================
+  def emit_iter_body(load_set='AB'):
+    if not NO_DS:
+      e(s_wait_dscnt(simm16=0))
+      e(s_barrier_signal(ssrc0=src[193])); e(s_barrier_wait(simm16=0xFFFF))
+    if not NO_GLOBAL:
+      if 'A' in load_set:
+        for i in range(2): e(global_load_b128(vdst=v[DA+i*4:DA+i*4+3], vaddr=v[6:7], saddr=s[4:5], ioffset=i*16))
+        e(v_add_nc_u32_e32(v[6], BLOCK_K*ELEM, v[6]))
+      if 'B' in load_set:
+        for i in range(2): e(global_load_b128(vdst=v[DB+i*4:DB+i*4+3], vaddr=v[8:9], saddr=s[6:7], ioffset=i*16))
+        e(v_add_nc_u32_e32(v[8], s[14], v[8]))
+    if not NO_DS:
+      # Issue 6 loads: A[0:3] + B[0] + B[1]. B[2:3] interleaved with WMMAs.
+      for tm in range(TILES_M):
+        aoff = tm * 16 * LDS_A_ROW
+        e(ds_load_b128(vdst=v[FA+tm*4:FA+tm*4+3], addr=v[LLA], offset0=aoff&0xFF, offset1=aoff>>8))
+      e(ds_load_b128(vdst=v[FB:FB+3], addr=v[LLB], offset0=0, offset1=0))
+      e(ds_load_b128(vdst=v[FB+4:FB+7], addr=v[LLB], offset0=0, offset1=2))
+      e(s_wait_dscnt(simm16=0))  # wait for 6 loads (no stall!)
+    if not NO_ALU:
+      # B[0] WMMAs — issue B[2] during compute
+      if not NO_DS: e(ds_load_b128(vdst=v[FB+8:FB+11], addr=v[LLB], offset0=0, offset1=4))
+      for tm in range(TILES_M):
+        ac = ACC + (tm*TILES_N+0)*8
+        e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB:FB+3], src2=v[ac:ac+7]))
+      # B[1] WMMAs — issue B[3] during compute
+      if not NO_DS:
+        e(ds_load_b128(vdst=v[FB+12:FB+15], addr=v[LLB], offset0=0, offset1=6))
+      for tm in range(TILES_M):
+        ac = ACC + (tm*TILES_N+1)*8
+        e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB+4:FB+7], src2=v[ac:ac+7]))
+      # B[2] WMMAs — B[2] loaded during B[0] WMMAs (~100 cycles ago)
+      if not NO_DS: e(s_wait_dscnt(simm16=1))  # B[2] done, B[3] may still be loading
+      for tm in range(TILES_M):
+        ac = ACC + (tm*TILES_N+2)*8
+        e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB+8:FB+11], src2=v[ac:ac+7]))
+      # B[3] WMMAs
+      if not NO_DS: e(s_wait_dscnt(simm16=0))
+      for tm in range(TILES_M):
+        ac = ACC + (tm*TILES_N+3)*8
+        e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB+12:FB+15], src2=v[ac:ac+7]))
+    if not NO_GLOBAL and not NO_DS: e(s_wait_loadcnt(simm16=0))
+    if not NO_DS:
+      for i in range(2): e(ds_store_b128(addr=v[4], data0=v[DA+i*4:DA+i*4+3], offset0=(i*16)&0xFF, offset1=(i*16)>>8))
+      for i in range(2): e(ds_store_b128(addr=v[5], data0=v[DB+i*4:DB+i*4+3], offset0=(i*16)&0xFF, offset1=(i*16)>>8))
+    e(s_add_co_i32(s[16], s[16], BLOCK_K))
+
+  label('LOOP')
+  emit_iter_body(load_set='A')
+  emit_iter_body(load_set='B')
+  e(s_cmp_lt_i32(s[16], s[17])); e(s_cbranch_scc1(simm16=0)); br(I[-1], 'LOOP')
+
+  emit_iter_body(load_set='AB')  # tail with prefetch
+
+  # Final iteration: no prefetch, no ds_store needed
+  if not NO_DS:
+    e(s_wait_dscnt(simm16=0))
+    e(s_barrier_signal(ssrc0=src[193])); e(s_barrier_wait(simm16=0xFFFF))
+  if not NO_DS:
+    for tm in range(TILES_M):
+      aoff = tm * 16 * LDS_A_ROW
+      e(ds_load_b128(vdst=v[FA+tm*4:FA+tm*4+3], addr=v[LLA], offset0=aoff&0xFF, offset1=aoff>>8))
+    e(ds_load_b128(vdst=v[FB:FB+3], addr=v[LLB], offset0=0, offset1=0))
+    e(ds_load_b128(vdst=v[FB+4:FB+7], addr=v[LLB], offset0=0, offset1=2))
+    e(s_wait_dscnt(simm16=0))
+  if not NO_ALU:
+    if not NO_DS: e(ds_load_b128(vdst=v[FB+8:FB+11], addr=v[LLB], offset0=0, offset1=4))
+    for tm in range(TILES_M):
+      ac = ACC + (tm*TILES_N+0)*8
+      e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB:FB+3], src2=v[ac:ac+7]))
+    if not NO_DS: e(ds_load_b128(vdst=v[FB+12:FB+15], addr=v[LLB], offset0=0, offset1=6))
+    for tm in range(TILES_M):
+      ac = ACC + (tm*TILES_N+1)*8
+      e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB+4:FB+7], src2=v[ac:ac+7]))
+    if not NO_DS: e(s_wait_dscnt(simm16=1))
+    for tm in range(TILES_M):
+      ac = ACC + (tm*TILES_N+2)*8
+      e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB+8:FB+11], src2=v[ac:ac+7]))
+    if not NO_DS: e(s_wait_dscnt(simm16=0))
+    for tm in range(TILES_M):
+      ac = ACC + (tm*TILES_N+3)*8
+      e(v_wmma_f32_16x16x16_f16(vdst=v[ac:ac+7], src0=v[FA+tm*4:FA+tm*4+3], src1=v[FB+12:FB+15], src2=v[ac:ac+7]))
+
+  label('EPILOGUE')
+  e(v_and_b32_e32(v[ET], 15, v[1]))
+  e(v_lshrrev_b32_e32(v[ET+1], 4, v[1])); e(v_lshlrev_b32_e32(v[ET+1], 3, v[ET+1]))
+  e(v_lshlrev_b32_e32(v[ET+2], 6, v[2])); e(v_add_nc_u32_e32(v[ET+2], s[11], v[ET+2]))
+  e(v_lshlrev_b32_e32(v[ET+3], 6, v[3])); e(v_add_nc_u32_e32(v[ET+3], s[10], v[ET+3]))
+  e(v_add_nc_u32_e32(v[ET+3], v[ET+3], v[ET])); e(v_mov_b32_e32(v[ET+5], 0))
+
+  for tm in range(TILES_M):
+    for tn in range(TILES_N):
+      ac = ACC + (tm*TILES_N+tn)*8; r_off, c_off = tm*16, tn*16
+      e(v_add_nc_u32_e32(v[ET+6], r_off, v[ET+2])); e(v_add_nc_u32_e32(v[ET+6], v[ET+1], v[ET+6]))
+      e(v_mul_lo_u32(v[ET+4], v[ET+6], s[12])); e(v_add_nc_u32_e32(v[ET+4], v[ET+4], v[ET+3]))
+      if c_off: e(v_add_nc_u32_e32(v[ET+4], c_off, v[ET+4]))
+      e(v_lshlrev_b32_e32(v[ET+4], 1, v[ET+4]))
+      for elem in range(8):
+        e(v_cvt_f16_f32_e32(v[ET+7], v[ac+elem]))
+        e(global_store_b16(vaddr=v[ET+4:ET+5], vsrc=v[ET+7], saddr=s[8:9]))
+        if elem < 7: e(v_add_nc_u32_e32(v[ET+4], s[13], v[ET+4]))
+
+  e(s_wait_storecnt(simm16=0)); e(s_sendmsg(simm16=3)); e(s_endpgm())
+
+  for idx, target in B:
+    off = (L[target] - sum(i.size() for i in I[:idx+1])) // 4
+    assert -32768 <= off <= 32767; I[idx].simm16 = off
+  return I
+
+N = getenv("N", 4096)
+
+def test_matmul():
+  dev = Device[Device.DEFAULT]
+  arch = getattr(dev.renderer, 'arch', 'gfx1200')
+  print(f"Device arch: {arch}")
+  insts = build_kernel(N, arch)
+
+  rng = np.random.default_rng(42)
+  a = Tensor(rng.random((N, N), dtype=np.float32).astype(np.float16))
+  b = Tensor(rng.random((N, N), dtype=np.float32).astype(np.float16))
+  c = Tensor.empty(N, N, dtype=dtypes.half)
+  Tensor.realize(a, b, c)
+
+  grid, local = (N//BLOCK_N, N//BLOCK_M, 1), (THREADS, 1, 1)
+  print(f"Grid: {grid}, Local: {local}")
+
+  dname = Device.DEFAULT
+  def asm_kernel(A, B, C):
+    gidxs = [UOp.special(n, f"gidx{i}") for i,n in enumerate(grid)]
+    lidxs = [UOp.special(THREADS, "lidx0")]
+    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=max(LDS_SIZE, 65536//getenv("LIMIT_OCC",2)), addrspace=AddrSpace.LOCAL), (), 'lds')
+    sink = UOp.sink(A.base, B.base, C.base, lds, *gidxs, *lidxs,
+                    arg=KernelInfo(name=colored("kernel","cyan"), estimates=Estimates(ops=N*N*N*2, mem=N*N*2*3)))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+
+  c = Tensor.custom_kernel(a, b, c, fxn=asm_kernel)[2]
+  ei = c.schedule()[0].lower()
+
+  ets = []
+  with Context(DEBUG=2):
+    for _ in range(getenv("CNT", 5)): ets.append(ei.run(wait=True))
+  print(f"REAL TFLOPS {N*N*N*2 / min(ets) * 1e-12:.2f}")
+
+  if getenv("VERIFY", 1):
+    GlobalCounters.reset()
+    c_np = c.float().numpy()
+    a_np, b_np = a.float().numpy(), b.float().numpy()
+    ref = a_np @ b_np
+    err = np.sqrt(np.mean((c_np - ref)**2)) / np.sqrt(np.mean(ref**2))
+    print(f"relative RMSE {err:.6f}")
+    if err != err or err > 0.05: raise RuntimeError(f"matmul is wrong! RMSE={err}")
+
+if __name__ == "__main__":
+  test_matmul()

extra/mlx_driver/connect.py

@@ -0,0 +1,68 @@
+#!/usr/bin/env python3
+import subprocess, json, sys, os
+
+REMOTE_HOST = os.getenv("REMOTE_HOST", "192.168.52.154")
+LOCAL_PCI   = os.getenv("MLX_PCI", "0000:41:00.0")
+REMOTE_PCI  = os.getenv("REMOTE_PCI", "0000:41:00.0")
+LOCAL_IP    = os.getenv("LOCAL_IP", "10.0.0.1")
+REMOTE_IP   = os.getenv("REMOTE_IP", "10.0.0.2")
+SSH         = ["ssh", "-o", "StrictHostKeyChecking=no", REMOTE_HOST]
+TINYGRAD    = os.path.dirname(os.path.abspath(__file__)) + "/../.."
+
+print("syncing code to remote")
+subprocess.run(["rsync", "-az", "--exclude=.git", "--exclude=__pycache__", "--exclude=*.pyc",
+                TINYGRAD + "/", f"{REMOTE_HOST}:~/tinygrad/"], check=True)
+
+print("booting remote")
+remote = subprocess.Popen(
+  SSH + [f"cd ~/tinygrad && sudo PYTHONPATH=. MLX_DEBUG=1 MLX_PCI={REMOTE_PCI} MLX_IP={REMOTE_IP} python3 extra/mlx_driver/mlxdev.py --server"],
+  stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=sys.stderr, text=True)
+
+remote_info = None
+for line in iter(remote.stdout.readline, ''):
+  print(f"  [remote] {line}", end='')
+  try: remote_info = json.loads(line.strip()); break
+  except json.JSONDecodeError: pass
+assert remote_info, "failed to get remote connection info"
+
+print("booting local")
+sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "../.."))
+from extra.mlx_driver.mlxdev import MLXDev, MLXQP
+from tinygrad.runtime.support.system import PCIDevice
+
+local_dev = MLXDev(PCIDevice("mlx5", LOCAL_PCI), ip=LOCAL_IP)
+local_qp = MLXQP(local_dev)
+local_info = {"qpn": local_qp.qpn, "mac": local_dev.mac.to_bytes(6,'big').hex(), "gid": local_dev.local_gid.hex()}
+
+remote.stdin.write(json.dumps(local_info) + "\n")
+remote.stdin.flush()
+for line in iter(remote.stdout.readline, ''):
+  print(f"  [remote] {line}", end='')
+  if "connected" in line: break
+
+local_qp.connect(remote_info["qpn"], int(remote_info["mac"], 16), int(remote_info["gid"], 16))
+print("both QPs in RTS")
+
+remote_target = None
+for line in iter(remote.stdout.readline, ''):
+  print(f"  [remote] {line}", end='')
+  try: remote_target = json.loads(line.strip()); break
+  except json.JSONDecodeError: pass
+assert remote_target
+
+test_msg = b"Test message, rdma works!"
+src_mem, src_paddrs = local_dev.pci_dev.alloc_sysmem(0x1000)
+for i, b in enumerate(test_msg): src_mem[i] = b
+
+print(f"RDMA WRITE {len(test_msg)}B to remote phys 0x{remote_target['target_addr']:x}")
+local_qp.rdma_write(remote_target["target_addr"], remote_target["rkey"], src_paddrs[0], local_dev.mkey, len(test_msg))
+
+remote.stdin.write("done\n")
+remote.stdin.flush()
+for line in iter(remote.stdout.readline, ''):
+  print(f"  [remote] {line}", end='')
+  if "AS TEXT" in line: break
+
+remote.stdin.close()
+remote.wait()
+print("RDMA WRITE test complete")

extra/mlx_driver/loopback.py

@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# GMMU=0 MLX_PCI=0000:41:00.0 PYTHONPATH=. python3 extra/mlx_driver/loopback.py
+import struct
+from tinygrad.helpers import getenv, round_up
+from tinygrad.device import Device, BufferSpec
+from tinygrad.runtime.support.system import PCIDevice
+from tinygrad.runtime.support.memory import AddrSpace
+from tinygrad.runtime.ops_amd import AMDComputeQueue
+from extra.mlx_driver.mlxdev import MLXDev, MLXQP, to_be
+
+BUF_SIZE = 0x1000
+MLX_PCI = getenv("MLX_PCI", "0000:41:00.0")
+MLX_IP = getenv("MLX_IP", "10.0.0.1")
+
+def map_phys_to_gpu(gpu, paddr, size):
+  size = round_up(size, 0x1000)
+  va = gpu.iface.dev_impl.mm.alloc_vaddr(size, align=0x1000)
+  gpu.iface.dev_impl.mm.map_range(va, size, [(paddr, size)], aspace=AddrSpace.SYS, snooped=True, uncached=True)
+  return va
+
+print("[init] AMD GPU...")
+gpu = Device["AMD"]
+
+print(f"[init] MLX5 at {MLX_PCI}")
+dev = MLXDev(PCIDevice("mlx5", MLX_PCI), ip=MLX_IP)
+qp = MLXQP(dev)
+
+print(f"[init] loopback connect QP 0x{qp.qp_info['qpn']:x}")
+qp.connect(qp.qp_info['qpn'], dev.mac, int.from_bytes(dev.local_gid, 'big'))
+
+# allocate src/dst via AMD GPU allocator
+buf_src = gpu.allocator.alloc(BUF_SIZE, BufferSpec(nolru=True))
+buf_dst = gpu.allocator.alloc(BUF_SIZE, BufferSpec(nolru=True))
+
+bar_base = gpu.iface.pci_dev.bar_info(gpu.iface.vram_bar)[0]
+src_paddr = buf_src.meta.mapping.paddrs[0][0] + bar_base
+dst_paddr = buf_dst.meta.mapping.paddrs[0][0] + bar_base
+print(f"src paddr=0x{src_paddr:x} dst paddr=0x{dst_paddr:x}")
+
+# fill src, zero dst
+test_msg = b"Hello from loopback send/recv!"
+gpu.allocator._copyin(buf_src, memoryview(bytearray(test_msg.ljust(BUF_SIZE, b'\x00'))))
+gpu.allocator._copyin(buf_dst, memoryview(bytearray(BUF_SIZE)))
+gpu.synchronize()
+
+# post recv WQE on RQ from CPU (scatter entry: byte_count, lkey, addr)
+rq_mask = (1 << 4) - 1  # log_rq_size=4
+rq_wqe = qp.qp_buf.view((qp.rq_head & rq_mask) * 16, 16)
+rq_wqe[:] = struct.pack('>IIQ', len(test_msg), dev.mkey, dst_paddr)
+qp.rq_head += 1
+# ring recv doorbell from CPU (DBR offset 0 = recv counter)
+dev.dbr[qp.qp_dbr // 4] = to_be('I', qp.rq_head)
+
+# build send WQE in SQ from CPU (opcode 0x0a = SEND, ds_count=2)
+sq_head = qp.sq_head
+sq_mask = (1 << qp.log_sq_size) - 1
+wqe = qp.qp_buf.view(qp.sq_offset + (sq_head & sq_mask) * 64, 64)
+wqe[:] = bytes(64)
+wqe[0:8] = struct.pack('>II', (sq_head << 8) | 0x0a, (qp.qp_info['qpn'] << 8) | 2)
+wqe[11] = 0x08  # CE: signal completion
+wqe[16:32] = struct.pack('>IIQ', len(test_msg), dev.mkey, src_paddr)
+qp.sq_head += 1
+doorbell_val = to_be('Q', int.from_bytes(bytes(wqe[0:8]), 'big'))
+
+# map MLX5 UAR and DBR into GPU VA
+uar_paddr = dev.pci_dev.bar_info(0)[0] + dev.uar * 0x1000
+uar_gpu_va = map_phys_to_gpu(gpu, uar_paddr, 0x1000)
+dbr_gpu_va = map_phys_to_gpu(gpu, dev.dbr_paddrs[0], 0x1000)
+print(f"UAR gpu_va=0x{uar_gpu_va:x} DBR gpu_va=0x{dbr_gpu_va:x}")
+
+# GPU rings send doorbell via compute queue release_mem
+q = AMDComputeQueue(gpu)
+q.wait(gpu.timeline_signal, gpu.timeline_value - 1)
+# write DBR (32-bit sq_head) - send doorbell at qp_dbr + 4
+q.release_mem(dbr_gpu_va + qp.qp_dbr + 4, to_be('I', qp.sq_head), q.pm4.data_sel__mec_release_mem__send_32_bit_low,
+              q.pm4.int_sel__mec_release_mem__none)
+# write UAR doorbell (64-bit)
+q.release_mem(uar_gpu_va + 0x800, doorbell_val, q.pm4.data_sel__mec_release_mem__send_64_bit_data,
+              q.pm4.int_sel__mec_release_mem__none)
+q.signal(gpu.timeline_signal, gpu.next_timeline())
+q.submit(gpu)
+
+print("GPU kicked doorbell, waiting...")
+gpu.synchronize()
+
+# poll CQ from CPU (send + recv completions)
+qp.poll_cq()
+qp.poll_cq()
+
+# read back
+result = bytearray(BUF_SIZE)
+gpu.allocator._copyout(memoryview(result), buf_dst)
+gpu.synchronize()
+
+got = bytes(result[:len(test_msg)])
+print(f"result: {got}")
+assert got == test_msg, f"MISMATCH: {got} != {test_msg}"
+print("RDMA loopback send/recv test passed (GPU-kicked)")