test_mismatch_reduce

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

@@ -121,7 +121,7 @@ runs:
         echo 'Acquire::GzipIndexes "true";' | sudo tee /etc/apt/apt.conf.d/gzip
         echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
         echo 'Binary::apt::APT::Keep-Downloaded-Packages "true";' | sudo tee -a /etc/apt/apt.conf.d/99keep-debs
-
+    
     - name: Add OpenCL Repo
       if: inputs.opencl == 'true' && runner.os == 'Linux'
       shell: bash
@@ -174,7 +174,7 @@ runs:
         if [[ "${{ inputs.llvm }}" == "true" ]]; then
           pkgs+=" libllvm20 clang-20 lld-20"
         fi
-
+        
         echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
         echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"
 
@@ -183,21 +183,21 @@ runs:
       uses: actions/cache@v4
       with:
         path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.APT_CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}
 
     - name: Run apt Update + Install
       if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
       shell: bash
       run: |
         sudo apt -qq update || true
-
+        
         # ******** do install ********
         if [[ -n "${{ steps.apt-pkgs.outputs.pkgs }}" ]]; then
           sudo apt-get -y --allow-unauthenticated --no-install-recommends install ${{ steps.apt-pkgs.outputs.pkgs }}
         fi
-
+        
         sudo chown -R $USER:$USER /var/cache/apt/archives/
-
+    
     # **** AMD ****
     - name: Setup AMD (Linux)
       if: inputs.amd == 'true' && runner.os == 'Linux'
@@ -234,7 +234,7 @@ runs:
         cache-name: cache-gpuocelot-build
       with:
         path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-0
     - name: Clone/compile gpuocelot
       if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
       shell: bash

.github/workflows/benchmark.yml

@@ -62,8 +62,10 @@ jobs:
       run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
     - name: Run model inference benchmark
       run: METAL=1 python3.11 test/external/external_model_benchmark.py
+    - name: Run huggingface_onnx test
+      run: METAL=1 python3.11 extra/huggingface_onnx/run_models.py test --debug FacebookAI/xlm-roberta-large
     - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: BIG=2 MPS=1 python3.11 test/test_speed_v_torch.py | tee torch_speed.txt
     - name: Test tensor cores
       run: METAL=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
     - name: Test AMX tensor cores
@@ -187,7 +189,7 @@ jobs:
     - name: Run model inference benchmark
       run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
     - name: Test speed vs torch
-      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: NV=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test benchmark allreduce
@@ -323,7 +325,7 @@ jobs:
       run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
     - name: Run 10 MLPerf Bert training steps (6 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (NVIDIA Training)
@@ -389,7 +391,7 @@ jobs:
     #- name: Test speed vs torch
     #  run: |
     #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/test_speed_v_torch.py | tee torch_speed.txt
     - name: Test speed vs theoretical
       run: AMD=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Test tensor cores
@@ -574,7 +576,7 @@ jobs:
       run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
     - name: Run 10 MLPerf Bert training steps (6 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (AMD MLPerf)
@@ -603,18 +605,18 @@ jobs:
         rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
     - name: reset process replay
       run: test/external/process_replay/reset.py
-    - name: benchmark openpilot 0.9.9 driving_vision
+    - name: validate openpilot 0.9.7
-      run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+      run: PYTHONPATH=. FLOAT16=0 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
-    - name: benchmark openpilot 0.9.9 driving_policy
+    - name: benchmark openpilot 0.9.7
-      run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+      run: BENCHMARK_LOG=openpilot_0_9_7 PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_7.txt
-    - name: benchmark openpilot 0.9.9 dmonitoring
+    - name: benchmark openpilot w IMAGE=2 0.9.7
-      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: BENCHMARK_LOG=openpilot_0_9_7_image PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
-    - name: openpilot compile3 0.9.9 driving_vision
+    - name: openpilot compile3 0.9.7
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx
-    - name: openpilot compile3 0.9.9 driving_policy
+    - name: openpilot compile3 0.9.7+ tomb raider
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/e8bea2c78ffa92685ece511e9b554122aaf1a79d/selfdrive/modeld/models/supercombo.onnx
-    - name: openpilot compile3 0.9.9 dmonitoring
+    - name: openpilot dmonitoring compile3 0.9.7
-      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/dmonitoring_model.onnx
     - name: openpilot compile3 Space Lab policy + vision
       run: |
         PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
@@ -690,12 +692,11 @@ jobs:
       run: AMD=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
     - name: Run full CIFAR training w 1 GPU
       run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
-    # TODO: enable
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-    # - 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 | tee am_train_resnet_one_gpu.txt
-    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
+      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (AM Driver)
@@ -753,7 +754,7 @@ jobs:
       run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
+      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
     - uses: actions/upload-artifact@v4
       with:
         name: Speed (NV Driver)

.github/workflows/test.yml

@@ -1,10 +1,8 @@
 name: Unit Tests
 env:
   # increment this when downloads substantially change to avoid the internet
-  DOWNLOAD_CACHE_VERSION: '12'
+  DOWNLOAD_CACHE_VERSION: '10'
-  PYTHON_CACHE_VERSION: '3'
+  PYTHON_CACHE_VERSION: '2'
-  APT_CACHE_VERSION: '1'
-  BUILD_CACHE_VERSION: '1'
   CAPTURE_PROCESS_REPLAY: 1
   GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
 
@@ -32,9 +30,9 @@ jobs:
     - name: External Benchmark Schedule
       run: PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
     - name: Speed Test
-      run: LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: LLVM=1 python3 test/test_speed_v_torch.py
     - name: Speed Test (BEAM=2)
-      run: BEAM=2 LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 LLVM=1 python3 test/test_speed_v_torch.py
 
   docs:
     name: Docs
@@ -48,11 +46,6 @@ jobs:
       with:
         deps: docs
         pydeps: "capstone"
-    - name: Build wheel and show size
-      run: |
-        pip install build
-        python -m build --wheel --outdir dist
-        ls -lh dist/*.whl
     - name: Use as an external package
       run: |
         mkdir $HOME/test_external_dir
@@ -336,6 +329,7 @@ jobs:
       run: |
         pip3 install --upgrade --force-reinstall ruff==0.11.0
         python3 -m ruff check .
+        python3 -m ruff check extra/onnx.py
         python3 -m ruff check examples/mlperf/ --ignore E501
     - name: Lint tinygrad with pylint
       run: python -m pylint tinygrad/
@@ -343,6 +337,7 @@ jobs:
       run: |
         python -m mypy --strict-equality --lineprecision-report .
         cat lineprecision.txt
+        python -m mypy --strict-equality extra/onnx.py
 
   unittest:
     name: Unit Tests
@@ -380,8 +375,8 @@ jobs:
         PYTHONPATH=. python extra/optimization/extract_dataset.py
         gzip -c /tmp/sops > extra/datasets/sops.gz
         DEBUG=1 MIN_ASTS=1 PYTHONPATH=. python extra/optimization/get_action_space.py
-    - name: Repo line count < 17000 lines
+    - name: Repo line count < 15500 lines
-      run: MAX_LINE_COUNT=17000 python sz.py
+      run: MAX_LINE_COUNT=15500 python sz.py
 
   fuzzing:
     name: Fuzzing
@@ -460,7 +455,7 @@ jobs:
   testopenpilot:
     name: 'openpilot Compile Tests'
     runs-on: ubuntu-22.04
-    timeout-minutes: 15
+    timeout-minutes: 10
     env:
       IGNORE_OOB: 0
     steps:
@@ -547,8 +542,8 @@ jobs:
         run: PYTHONPATH="." GPU=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
       - name: Test MLPerf stuff
         run: GPU=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: Test llama 3 training
+      - name: Run handcode_opt
-        run: MAX_BUFFER_SIZE=0 PYTHONPATH="." DEV=NULL SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=8 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: PYTHONPATH=. MODEL=resnet GPU=1 DEBUG=1 BS=4 HALF=0 python3 examples/handcode_opt.py
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
@@ -591,29 +586,6 @@ jobs:
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
-  testdevectorize:
-    name: Linux (devectorize)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    env:
-      IGNORE_OOB: 0
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: devectorize-minimal
-        deps: testing_minimal
-        pydeps: "pillow"
-        llvm: "true"
-    - name: Test LLVM=1 DEVECTORIZE=0
-      run: LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
-    - name: Test LLVM=1 DEVECTORIZE=0 for model
-      run: PYTHONPATH="." LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
-    - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 DEVECTORIZE=0 FUSE_ARANGE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
-
   testdsp:
     name: Linux (DSP)
     runs-on: ubuntu-24.04
@@ -649,6 +621,12 @@ jobs:
       run: CC=clang-20 PYTHONPATH="." DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
     - name: Test quantize onnx
       run: PYTHONPATH="." DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py
+    - name: Test LLVM=1 DEVECTORIZE=0
+      run: LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
+    - name: Test LLVM=1 DEVECTORIZE=0 for model
+      run: PYTHONPATH="." LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
+    - name: Test CPU=1 DEVECTORIZE=0
+      run: CPU=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
 
   testwebgpu:
     name: Linux (WebGPU)
@@ -708,9 +686,9 @@ jobs:
           DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
       - name: Run LLVM test
         if: matrix.backend=='amdllvm'
-        run: python test/device/test_amd_llvm.py
+        run: python test/test_amd_llvm.py
       - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py --durations=20
+        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/test_hcq.py --durations=20
       - name: Run pytest (amd)
         run: python -m pytest test/external/external_test_am.py --durations=20
       - name: Run TRANSCENDENTAL math
@@ -835,14 +813,14 @@ jobs:
         AMD: 1
         FORWARD_ONLY: 1
       run: |
-        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
+        python3 -m pytest -n=auto test/test_hcq.py test/test_tiny.py --durations=20
     - name: Run pytest (amd with llvm backend)
       env:
         MOCKGPU: 1
         AMD: 1
         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
+        python -m pytest -n=auto test/test_hcq.py test/test_tiny.py test/test_amd_llvm.py --durations=20
     - name: Run pytest (ptx)
       env:
         MOCKGPU: 1
@@ -850,7 +828,7 @@ jobs:
         NV: 1
         FORWARD_ONLY: 1
       run: |
-        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
+        python3 -m pytest -n=auto test/test_hcq.py test/test_tiny.py --durations=20
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
@@ -961,18 +939,18 @@ jobs:
         env:
           HOST: 127.0.0.1:6667*6,127.0.0.1:6668*6
         run: |
-          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_remote.py test/test_tensor_variable.py --durations 20
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_remote.py test/test_tensor_variable.py
       - name: Run REMOTE=1 Test (GPU)
         env:
           HOST: 127.0.0.1:7667*6
         run: |
-          python3 -m pytest test/test_tiny.py test/test_image_dtype.py test/test_jit.py --durations 20
+          python3 -m pytest test/test_tiny.py test/test_image_dtype.py test/test_jit.py
           IMAGE=2 python3 -m pytest test/test_tiny.py test/test_image_dtype.py
       - name: Run REMOTE=1 Test (CPU)
         env:
           HOST: 127.0.0.1:8667*6
         run: |
-          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py --durations 20
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py
       - name: Show remote server logs
         if: always()
         run: |
@@ -999,7 +977,6 @@ jobs:
         with:
           key: macos-${{ matrix.backend }}-minimal
           deps: testing_minimal
-          pydeps: "capstone"
           llvm: ${{ matrix.backend == 'llvm' && 'true' }}
       - name: Set env
         run: printf "${{ matrix.backend == 'llvm' && 'LLVM=1' || matrix.backend == 'cpu' && 'CPU=1' || matrix.backend == 'metal' && 'METAL=1'}}" >> $GITHUB_ENV

autogen_stubs.sh

@@ -198,7 +198,11 @@ generate_amd() {
   clang2py -k cdefstum \
     extra/hip_gpu_driver/sdma_registers.h \
     extra/hip_gpu_driver/nvd.h \
+    extra/hip_gpu_driver/kfd_pm4_headers_ai.h \
+    extra/hip_gpu_driver/soc21_enum.h \
+    extra/hip_gpu_driver/sdma_v6_0_0_pkt_open.h \
     extra/hip_gpu_driver/gc_11_0_0_offset.h \
+    extra/hip_gpu_driver/gc_10_3_0_offset.h \
     extra/hip_gpu_driver/sienna_cichlid_ip_offset.h \
     --clang-args="-I/opt/rocm/include -x c++" \
     -o $BASE/amd_gpu.py
@@ -372,6 +376,26 @@ generate_am() {
     -o $BASE/am/pm4_nv.py
   fixup $BASE/am/pm4_nv.py
 
+  clang2py -k cdefstum \
+    $AMKERN_INC/vega10_enum.h \
+    -o $BASE/am/vega10.py
+  fixup $BASE/am/vega10.py
+
+  clang2py -k cdefstum \
+    $AMKERN_INC/navi10_enum.h \
+    -o $BASE/am/navi10.py
+  fixup $BASE/am/navi10.py
+
+  clang2py -k cdefstum \
+    $AMKERN_INC/soc21_enum.h \
+    -o $BASE/am/soc21.py
+  fixup $BASE/am/soc21.py
+
+  clang2py -k cdefstum \
+    $AMKERN_INC/soc24_enum.h \
+    -o $BASE/am/soc24.py
+  fixup $BASE/am/soc24.py
+
   clang2py -k cdefstum \
     extra/hip_gpu_driver/sdma_registers.h \
     $AMKERN_AMD/amdgpu/vega10_sdma_pkt_open.h \

docs/developer/layout.md

@@ -18,11 +18,11 @@ Group UOps into kernels.
 
 ---
 
-## tinygrad/codegen/opt
+## tinygrad/opt
 
 Transforms the ast into an optimized ast. This is where BEAM search and heuristics live.
 
-::: tinygrad.codegen.opt.get_optimized_ast
+::: tinygrad.opt.get_optimized_ast
     options:
         members: false
         show_labels: false

docs/tinybox.md

@@ -47,8 +47,8 @@ Reboot after making these changes or restart the `displayservice.service` servic
 
 The [default tinybox image](https://github.com/tinygrad/tinyos) ships with tinygrad and PyTorch. While we develop tinygrad, the box is universal hardware. Use whatever framework you desire, run notebooks, download demos, install more things, train, inference, live, laugh, love, you aren't paying per hour for this box so the only limit is your imagination.
 
-## Building the OS image
+## tinychat
 
-The OS image is built using `ubuntu-image` from <https://github.com/tinygrad/tinyos>.
+Since LLMs are so popular, we ship with a built in tinygrad based chatbot using a LLaMA-3 finetune. Visit the IP (not the BMC IP) of your tinybox in a web browser on your computer or phone, and you'll find a friendly looking chat interface. This chatbot also provides an OpenAI compatible LLM API on that port, so you can script it.
 
-After cloning, run `make green` or `make red` to build a tinybox green or tinybox red image respectively.
+The conversations you have with this chatbot are between you and your tinybox. Also, the history in the web app is saved on the client, not the tinybox.

examples/beautiful_mnist.py

@@ -1,12 +1,12 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
-from typing import Callable
+from typing import List, Callable
 from tinygrad import Tensor, TinyJit, nn, GlobalCounters
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist
 
 class Model:
   def __init__(self):
-    self.layers: list[Callable[[Tensor], Tensor]] = [
+    self.layers: List[Callable[[Tensor], Tensor]] = [
       nn.Conv2d(1, 32, 5), Tensor.relu,
       nn.Conv2d(32, 32, 5), Tensor.relu,
       nn.BatchNorm(32), Tensor.max_pool2d,
@@ -21,15 +21,17 @@ if __name__ == "__main__":
   X_train, Y_train, X_test, Y_test = mnist(fashion=getenv("FASHION"))
 
   model = Model()
-  opt = (nn.optim.Adam if not getenv("MUON") else nn.optim.Muon)(nn.state.get_parameters(model))
+  opt = nn.optim.Adam(nn.state.get_parameters(model))
 
   @TinyJit
   @Tensor.train()
   def train_step() -> Tensor:
     opt.zero_grad()
     samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
+    # TODO: this "gather" of samples is very slow. will be under 5s when this is fixed
     loss = model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
-    return loss.realize(*opt.schedule_step())
+    opt.step()
+    return loss
 
   @TinyJit
   def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100

examples/handcode_opt.py

@@ -0,0 +1,134 @@
+from extra.models.resnet import ResNet50
+from extra.mcts_search import mcts_search
+from examples.mlperf.helpers import get_mlperf_bert_model
+from tinygrad import Tensor, Device, dtypes, nn
+from tinygrad.opt.kernel import Kernel
+from tinygrad.opt.heuristic import hand_coded_optimizations
+from tinygrad.uop.ops import Ops, sym_infer
+from tinygrad.device import Compiled
+from tinygrad.opt.search import beam_search, bufs_from_lin
+from tinygrad.helpers import DEBUG, ansilen, getenv, colored, TRACEMETA
+from extra.optimization.helpers import time_linearizer
+from tinygrad.engine.realize import get_program
+
+def get_sched_resnet():
+  mdl = ResNet50()
+  optim = (nn.optim.LARS if getenv("LARS") else nn.optim.SGD)(nn.state.get_parameters(mdl))
+  BS = getenv("BS", 64)
+
+  # run model twice to get only what changes, these are the kernels of the model
+  for _ in range(2):
+    out = mdl(Tensor.empty(BS, 3, 224, 224))
+    targets = [out]
+    if getenv("BACKWARD"):
+      optim.zero_grad()
+      out.sparse_categorical_crossentropy(Tensor.empty(BS, dtype=dtypes.int)).backward()
+      targets += [x for x in optim.schedule_step()]
+    sched = Tensor.schedule(*targets)
+    print(f"schedule length {len(sched)}")
+  return sched
+
+def get_sched_bert():
+  mdl = get_mlperf_bert_model()
+  optim = nn.optim.LAMB(nn.state.get_parameters(mdl))
+
+  # fake data
+  BS = getenv("BS", 9)
+  input_ids = Tensor.empty((BS, 512), dtype=dtypes.float32)
+  segment_ids = Tensor.empty((BS, 512), dtype=dtypes.float32)
+  attention_mask = Tensor.empty((BS, 512), dtype=dtypes.default_float)
+  masked_positions = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  masked_lm_ids = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  masked_lm_weights = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  next_sentence_labels = Tensor.empty((BS, 1), dtype=dtypes.float32)
+
+  # run model twice to get only what changes, these are the kernels of the model
+  for _ in range(2):
+    lm_logits, seq_relationship_logits = mdl(input_ids, attention_mask, masked_positions, segment_ids)
+    targets = [lm_logits, seq_relationship_logits]
+    if getenv("BACKWARD"):
+      optim.zero_grad()
+      loss = mdl.loss(lm_logits, seq_relationship_logits, masked_lm_ids, masked_lm_weights, next_sentence_labels)
+      # ignore grad norm and loss scaler for now
+      loss.backward()
+      targets += [x for x in optim.schedule_step()]
+    sched = Tensor.schedule(*targets)
+    print(f"schedule length {len(sched)}")
+  return sched
+
+if __name__ == "__main__":
+  if getenv("HALF", 1):
+    dtypes.default_float = dtypes.half
+
+  # the device we are optimizing for
+  device: Compiled = Device[Device.DEFAULT]
+  if getenv("BACKWARD"): Tensor.training = True
+  print(f"optimizing for {Device.DEFAULT}")
+
+  sched = globals()[f"get_sched_{getenv('MODEL', 'resnet')}"]()
+  sched = [x for x in sched if x.ast.op is Ops.SINK]
+
+  # focus on one kernel
+  if getenv("KERNEL", -1) >= 0: sched = sched[getenv("KERNEL", -1):getenv("KERNEL", -1)+1]
+
+  # work with the schedule
+  total_tm = 0
+  running_gflops = 0
+  usage = {}
+  for i,si in enumerate(sched):
+    if DEBUG >= 3: print(si.ast)
+
+    rawbufs = bufs_from_lin(Kernel(si.ast))
+
+    # "linearize" the op into uops in different ways
+    lins: list[tuple[Kernel, str]] = []
+
+    # always try hand coded opt
+    lin = Kernel(si.ast, opts=device.renderer)
+    lin.apply_opts(hand_coded_optimizations(lin))
+    lins.append((lin, "HC"))
+
+    # maybe try tensor cores
+    lin = Kernel(si.ast, opts=device.renderer)
+    if lin.apply_tensor_cores():
+      lins.append((lin, "TC"))
+
+    # try a beam search
+    if beam:=getenv("BEAM"):
+      lin = Kernel(si.ast, opts=device.renderer)
+      lin = beam_search(lin, rawbufs, beam, bool(getenv("BEAM_ESTIMATE", 1)))
+      lins.append((lin, "BEAM"))
+
+    # try MCTS
+    if mcts:=getenv("MCTS"):
+      lin = Kernel(si.ast, opts=device.renderer)
+      lin = mcts_search(lin, rawbufs, mcts)
+      lins.append((lin, "MCTS"))
+
+    # benchmark the programs
+    choices = []
+    for lin, nm in lins:
+      tm = time_linearizer(lin, rawbufs, allow_test_size=False, cnt=10, disable_cache=True)
+      ops = (prg:=get_program(lin.get_optimized_ast(), lin.opts)).estimates.ops
+      gflops = sym_infer(ops, {k:k.min for k in lin.ast.variables()})*1e-9/tm
+      choices.append((tm, gflops, lin, prg, nm))
+
+    sorted_choices = sorted(choices, key=lambda x: x[0])
+    if DEBUG >= 1: # print all kernels
+      for tm, gflops, lin, prg, nm in choices:
+        print(f"                 kernel {i:2d} {lin.name+' '*(37-ansilen(lin.name))} {str(prg.global_size):18s} {str(prg.local_size):12s} takes {tm*1000:7.2f} ms, {gflops:6.0f} GFLOPS -- {colored(nm, 'green') if lin is sorted_choices[0][2] else nm}")
+
+    tm, gflops, lin, prg, nm = sorted_choices[0]
+    if getenv("SRC"):
+      print(si.ast)
+      print(lin.applied_opts)
+      print(get_program(lin.get_optimized_ast(), lin.opts).src)
+    total_tm += tm
+    running_gflops += gflops * tm
+    if (key := str([str(m) for m in si.metadata])) not in usage: usage[key] = (0, 0)
+    usage[key] = (usage[key][0] + tm, usage[key][1] + 1)
+    print(f"*** {total_tm*1000:7.2f} ms : kernel {i:2d} {lin.name+' '*(37-ansilen(lin.name))} {str(prg.global_size):18s} {str(prg.local_size):12s} takes {tm*1000:7.2f} ms, {gflops:6.0f} GFLOPS {[repr(m) if TRACEMETA >= 2 else str(m) for m in si.metadata]}")
+  print(f"******* total {total_tm*1000:.2f} ms, {running_gflops/total_tm:6.0f} GFLOPS")
+  print("usage:")
+  for k in sorted(usage, key=lambda x: -usage[x][0])[:10]:
+    print(f"{usage[k][0]*1000:.2f} ms: {k} ({usage[k][1]} times)")

examples/mlperf/dataloader.py

@@ -1,4 +1,6 @@
-import os, random, pickle, queue, struct, math, functools, hashlib, time
+import functools
+import hashlib
+import os, random, pickle, queue, struct, math
 from typing import List
 from pathlib import Path
 from multiprocessing import Queue, Process, shared_memory, connection, Lock, cpu_count
@@ -530,21 +532,21 @@ class BinIdxDataset:
 
     start = self.idx.tell()
     end = start + self.count * dtypes.int32.itemsize
-    self.sizes = self.idx_t[start:end].bitcast(dtypes.int32).numpy()
+    self.sizes = self.idx_t[start:end].bitcast(dtypes.int32)
 
     start = end
     end = start + self.count * dtypes.int64.itemsize
-    self.pointers = self.idx_t[start:end].bitcast(dtypes.int64).numpy()
+    self.pointers = self.idx_t[start:end].bitcast(dtypes.int64)
 
     start = end
     end = start + doc_count * dtypes.int64.itemsize
-    self.doc_idx = self.idx_t[start:end].bitcast(dtypes.int64).numpy()
+    self.doc_idx = self.idx_t[start:end].bitcast(dtypes.int64)
 
     # bin file
     self.bin_t = Tensor(base_path.with_name(f"{base_path.name}.bin"))
 
   def _index(self, idx) -> tuple[int, int]:
-    return int(self.pointers[idx]), int(self.sizes[idx])
+    return self.pointers[idx].item(), self.sizes[idx].item()
 
   def get(self, idx, offset:int=0, length:int|None=None):
     ptr, size = self._index(idx)
@@ -564,13 +566,10 @@ class GPTDataset:
     # check for cache
     cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
     cache_path = base_path.with_name(f"{base_path.name}.{cache_hash}.index_cache")
-    print(f"try loading GPTDataset from {cache_path}...")
     if cache_path.exists():
-      print("cache found, loading...")
       with open(cache_path, "rb") as f:
         self.doc_idx, self.sample_idx, self.shuffle_idx = pickle.load(f)
     else:
-      print("cache not found, building index...")
       self.doc_idx = self._build_doc_idx()
       self.sample_idx = self._build_sample_idx()
       self.shuffle_idx = self._build_shuffle_idx()
@@ -629,20 +628,14 @@ class GPTDataset:
 
   # https://github.com/NVIDIA/Megatron-LM/blob/94bd476bd840c2fd4c3ebfc7448c2af220f4832b/megatron/core/datasets/gpt_dataset.py#L558
   def _build_doc_idx(self):
-    print(f"building doc_idx for {self.num_epochs=}, {self.indexed_dataset.count=}")
+    doc_idx = np.mgrid[:self.num_epochs, :self.indexed_dataset.count][1]
-    st = time.perf_counter()
+    doc_idx = doc_idx.reshape(-1)
-    # doc_idx = np.mgrid[:self.num_epochs, :self.indexed_dataset.count][1]
-    doc_idx = np.arange(self.indexed_dataset.count).reshape(1, -1).repeat(self.num_epochs, axis=0).flatten()
     doc_idx = doc_idx.astype(np.int32)
-    at = time.perf_counter()
     if self.shuffle: self.rng.shuffle(doc_idx)
-    print(f"doc_idx built in {at - st:.3f}s, shuffled in {time.perf_counter() - at:.3f}s")
     return doc_idx
 
   def _build_sample_idx(self):
-    print(f"building sample_idx for {self.samples=}, {self.seqlen=}, {self.doc_idx.shape[0]=}")
+    sample_idx = np.empty((self.samples + 1, 2), dtype=np.int32)
-    sample_idx_max = max(self.doc_idx.shape[0], self.indexed_dataset.sizes.max())
-    sample_idx = np.empty((self.samples + 1, 2), dtype=np.int64 if sample_idx_max > dtypes.int32.max else np.int32)
 
     sample_idx_idx, doc_idx_idx, doc_offset = 0, 0, 0
     sample_idx[sample_idx_idx, 0], sample_idx[sample_idx_idx, 1] = doc_idx_idx, doc_offset
@@ -652,7 +645,7 @@ class GPTDataset:
       remaining_seqlen = self.seqlen + 1
       while remaining_seqlen > 0:
         doc_idx = int(self.doc_idx[doc_idx_idx])
-        doc_len = int(self.indexed_dataset.sizes[doc_idx]) - doc_offset
+        doc_len = self.indexed_dataset.sizes[doc_idx].item() - doc_offset
         remaining_seqlen -= doc_len
         if remaining_seqlen <= 0:
           doc_offset += remaining_seqlen + doc_len - 1
@@ -661,7 +654,7 @@ class GPTDataset:
           if doc_idx_idx == len(self.doc_idx) - 1:
             assert sample_idx_idx == self.samples
             doc_idx = int(self.doc_idx[doc_idx_idx])
-            doc_offset = int(self.indexed_dataset.sizes[doc_idx]) - 1
+            doc_offset = self.indexed_dataset.sizes[doc_idx].item() - 1
             break
           doc_idx_idx += 1
           doc_offset = 0
@@ -672,18 +665,13 @@ class GPTDataset:
     return sample_idx
 
   def _build_shuffle_idx(self):
-    print(f"building shuffle_idx for {self.samples=}")
-    st = time.perf_counter()
     shuffle_idx = np.arange(self.samples, dtype=np.int32)
-    at = time.perf_counter()
     if self.shuffle: self.rng.shuffle(shuffle_idx)
-    print(f"shuffle_idx built in {at - st:.3f}s, shuffled in {time.perf_counter() - at:.3f}s")
     return shuffle_idx
 
 class BlendedGPTDataset:
   def __init__(self, paths:list[Path], weights:list[float], samples:int, seqlen:int, seed:int, shuffle:bool):
-    self.shuffle = shuffle
+    self.seed = seed
-    self.rng = np.random.RandomState(seed)
 
     # normalize weights
     total_weight = sum(weights)
@@ -695,47 +683,10 @@ class BlendedGPTDataset:
 
     self.datasets = [GPTDataset(path, samples_per_blend[i], seqlen, seed + i, shuffle) for i,path in enumerate(paths)]
 
-    # check for cache
-    cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
-    cache_path = paths[0].with_name(f"{paths[0].name}.{cache_hash}.blend_cache")
-    print(f"try loading BlendedGPTDataset from {cache_path}...")
-    if cache_path.exists():
-      print("cache found, loading...")
-      with open(cache_path, "rb") as f:
-        self.dataset_idx, self.dataset_sample_idx = pickle.load(f)
-    else:
-      print("cache not found, building index...")
-      self.dataset_idx, self.dataset_sample_idx = self._build_blend_idx()
-      # save cache
-      with open(cache_path, "wb") as f:
-        pickle.dump((self.dataset_idx, self.dataset_sample_idx), f)
-
   def get(self, idx:int):
-    tokens = self.datasets[self.dataset_idx[idx]][self.dataset_sample_idx[idx]]
+    tokens = self.datasets[0][idx]
     return tokens
 
-  def _build_blend_idx(self):
-    dataset_idx = np.zeros(self.samples, dtype=np.int16)
-    dataset_sample_idx = np.zeros(self.samples, dtype=np.int64)
-
-    unspent_datasets = set(range(len(self.datasets)))
-    dataset_sample_counts = [0] * len(self.datasets)
-
-    for i in tqdm(range(self.samples)):
-      error_argmax, error_max = 0, 0.0
-      for di in unspent_datasets:
-        error = self.weights[di] * max(i, 1) - dataset_sample_counts[di]
-        if error > error_max:
-          error_max = error
-          error_argmax = di
-
-      dataset_idx[i] = error_argmax
-      dataset_sample_idx[i] = dataset_sample_counts[error_argmax]
-
-      dataset_sample_counts[error_argmax] += 1
-
-    return dataset_idx, dataset_sample_idx
-
 def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
   if val:
     dataset = BlendedGPTDataset([
@@ -788,8 +739,8 @@ if __name__ == "__main__":
 
   def load_llama3(val):
     bs = 24
-    samples = 5760 if val else 1_200_000 * 1152
+    samples = 5760 if val else 1_200_000
-    seqlen = 8192
+    seqlen = 512
 
     max_, min_ = 0, math.inf
     for tokens in tqdm(batch_load_llama3(bs, samples, seqlen, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=5760, val=bool(val)), total=samples//bs):

examples/mlperf/model_train.py

@@ -1296,7 +1296,7 @@ def train_llama3():
   SEED               = config["SEED"]                   = getenv("SEED", 5760)
   SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
   TRAIN_ON_VAL       = config["TRAIN_ON_VAL"]           = getenv("TRAIN_ON_VAL", 0)
-  SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000 * 1152)
+  SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000)
 
   # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 FUSE_ARANGE=1 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
@@ -1318,33 +1318,29 @@ def train_llama3():
   if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: params['n_layers'] = llama_layers
   model = Transformer(**params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
 
-  if getenv("FAKEDATA"):
-    for v in get_parameters(model):
-      v = v.assign(Tensor.empty(v.shape))
-
   if (DP := getenv("DP", 1)) > 1:
     device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
     for v in get_parameters(model):
       v.shard_(device, axis=None)
 
-  if (MP := getenv("MP", 1)) > 1:
+  # TODO: MP
-    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
+  # if (GPUS := getenv("GPUS", 1)) > 1:
-    for k,v in get_state_dict(model).items():
+  #   device = tuple(f"{Device.DEFAULT}:{i}" for i in range(GPUS))
-      if 'scale' in k: v.shard_(device, axis=None)  # from quantized
+  #   for k,v in get_state_dict(model).items():
-      elif '.attention.wq' in k: v.shard_(device, axis=0)
+  #     if 'scale' in k: v.shard_(device, axis=None)  # from quantized
-      elif '.attention.wk' in k: v.shard_(device, axis=0)
+  #     # elif '.attention.wq' in k: v.shard_(device, axis=0)
-      elif '.attention.wv' in k: v.shard_(device, axis=0)
+  #     # elif '.attention.wk' in k: v.shard_(device, axis=0)
-      elif '.attention.wo' in k: v.shard_(device, axis=1)
+  #     # elif '.attention.wv' in k: v.shard_(device, axis=0)
-      elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
+  #     # elif '.attention.wo' in k: v.shard_(device, axis=1)
-      elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
+  #     # elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
-      elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
+  #     # elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
-      elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
+  #     # elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
-      elif 'output.weight' in k: v.shard_(device, axis=0)
+  #     # elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
-      else:
+  #     elif 'output.weight' in k: v.shard_(device, axis=0)  # 243.32
-        # attention_norm, ffn_norm, norm
+  #     else:
-        v.shard_(device, axis=None)
+  #       # print(k)
-      # prevents memory spike on device 0
+  #       # attention_norm, ffn_norm, norm
-      v.realize()
+  #       v.shard_(device, axis=None)
 
   optim = AdamW(get_parameters(model), lr=0.0,
                 b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay)
@@ -1359,9 +1355,6 @@ def train_llama3():
       if (DP := getenv("DP", 1)) > 1:
         device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
         batch = batch.shard(device, 0)
-      if (MP := getenv("MP", 1)) > 1:
-        device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
-        batch = batch.shard(device)
       logits:Tensor = model(batch[:, :-1], start_pos=0, temperature=math.nan)
       loss = logits.sparse_categorical_crossentropy(batch[:, 1:])
       loss.backward()
@@ -1398,12 +1391,11 @@ def train_llama3():
     t = time.perf_counter()
     GlobalCounters.reset()
     loss, lr = train_step(model, tokens, grad_acc)
-    loss = loss.float().item()
     # above as tqdm.write f-string
-    tqdm.write(f"{loss:.4f} loss, {lr.item():.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, {time.perf_counter()-t:.2f} s")
+    tqdm.write(f"{loss.item():.4f} loss, {lr.item():.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, {time.perf_counter()-t:.2f} s")
     if (fname:=getenv("LOSS_FILE", "")):
       with open(fname, "a") as f:
-        f.write(f"{i} {loss:.4f} {lr.item():.12f} {GlobalCounters.mem_used / 1e9:.2f}\n")
+        f.write(f"{i} {loss.item():.4f} {lr.item():.12f} {GlobalCounters.mem_used / 1e9:.2f}\n")
 
     if getenv("CKPT") and (i % 200 == 0 or i == 10):
       tqdm.write("saving checkpoint")

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

@@ -4,8 +4,6 @@ export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
 
-export IGNORE_OOB=1
-
 export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1
 # export BEAM_LOG_SURPASS_MAX=1

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

@@ -5,8 +5,6 @@ export MODEL="bert"
 export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 
-export IGNORE_OOB=1
-
 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
 export BASEDIR="/raid/datasets/wiki"

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

@@ -8,8 +8,6 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900
 
-export IGNORE_OOB=1
-
 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
 export BASEDIR="/raid/datasets/wiki"

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

@@ -11,8 +11,6 @@ export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
 export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
 export TRAIN_STEPS=3900
 
-export IGNORE_OOB=1
-
 export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
 export BASEDIR="/raid/datasets/wiki"

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

@@ -2,9 +2,9 @@
 
 export PYTHONPATH="." NV=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
 
 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1

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

@@ -2,9 +2,9 @@
 
 export PYTHONPATH="." NV=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
 
 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1

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

@@ -5,9 +5,9 @@ set -o pipefail  # Make pipeline fail if any command fails
 export PYTHONPATH="." NV=1
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
 
 export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1

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

@@ -2,9 +2,9 @@
 
 export PYTHONPATH="." AMD=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
 
 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1

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

@@ -2,9 +2,9 @@
 
 export PYTHONPATH="." AMD=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
 
 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1

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

@@ -5,9 +5,9 @@ set -o pipefail  # Make pipeline fail if any command fails
 export PYTHONPATH="." AMD=1
 export MODEL="bert"
 export SUBMISSION_PLATFORM="tinybox_red"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
 
-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0
 
 export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
 export IGNORE_JIT_FIRST_BEAM=1

examples/olmoe.py

@@ -1,7 +1,8 @@
 # https://arxiv.org/pdf/2409.02060
-import time, functools
+import time
 import numpy as np
 np.set_printoptions(suppress=True, linewidth=1000)
+import functools
 from tinygrad import Tensor, nn, Device, GlobalCounters
 from tinygrad.helpers import Timing, getenv
 from extra.models.llama import Transformer, convert_from_huggingface
@@ -16,7 +17,7 @@ class MixtureFeedForward:
   def __call__(self, x:Tensor) -> Tensor:
     assert x.shape[0] == 1, "only BS=1"
     assert x.shape[1] == 1, "only length=1"
-    g = self.gate(x).softmax(-1)
+    g = self.gate(x).float().softmax(-1)
 
     g = g.squeeze() # (BS, length, num_experts) -> (num_experts,)
     probs, sel = g.topk(self.activated_experts)
@@ -24,7 +25,7 @@ class MixtureFeedForward:
     # run MoE
     x_up_gate = x.dot(self.gate_proj[sel].permute(0,2,1)).silu() * x.dot(self.up_proj[sel].permute(0,2,1))
     x_down = x_up_gate.dot(self.down_proj[sel].permute(0,2,1))
-    return (x_down * probs.reshape(self.activated_experts, 1, 1)).sum(axis=0)
+    return (x_down.float() * probs.reshape(self.activated_experts, 1, 1)).sum(axis=0)
 
 # model is bf16, 1.3B active, 6.9B total
 # M3 Max is 400 GB/s, so 400/2.6 = ~154 tok/s

examples/yolov3.py

@@ -71,8 +71,8 @@ def bbox_iou(box1, box2):
   # get the coordinates of the intersection rectangle
   inter_rect_x1 = np.maximum(b1_x1, b2_x1)
   inter_rect_y1 = np.maximum(b1_y1, b2_y1)
-  inter_rect_x2 = np.minimum(b1_x2, b2_x2)
+  inter_rect_x2 = np.maximum(b1_x2, b2_x2)
-  inter_rect_y2 = np.minimum(b1_y2, b2_y2)
+  inter_rect_y2 = np.maximum(b1_y2, b2_y2)
   #Intersection area
   inter_area = np.clip(inter_rect_x2 - inter_rect_x1 + 1, 0, 99999) * np.clip(inter_rect_y2 - inter_rect_y1 + 1, 0, 99999)
   #Union Area
@@ -297,13 +297,13 @@ class Darknet:
           # Get the number of weights of batchnorm
           num_bn_biases = math.prod(bn.bias.shape)
           # Load weights
-          bn_biases = Tensor(weights[ptr:ptr + num_bn_biases].astype(np.float32))
+          bn_biases = Tensor(weights[ptr:ptr + num_bn_biases])
           ptr += num_bn_biases
-          bn_weights = Tensor(weights[ptr:ptr+num_bn_biases].astype(np.float32))
+          bn_weights = Tensor(weights[ptr:ptr+num_bn_biases])
           ptr += num_bn_biases
-          bn_running_mean = Tensor(weights[ptr:ptr+num_bn_biases].astype(np.float32))
+          bn_running_mean = Tensor(weights[ptr:ptr+num_bn_biases])
           ptr += num_bn_biases
-          bn_running_var = Tensor(weights[ptr:ptr+num_bn_biases].astype(np.float32))
+          bn_running_var = Tensor(weights[ptr:ptr+num_bn_biases])
           ptr += num_bn_biases
           # Cast the loaded weights into dims of model weights
           bn_biases = bn_biases.reshape(shape=tuple(bn.bias.shape))
@@ -319,7 +319,7 @@ class Darknet:
           # load biases of the conv layer
           num_biases = math.prod(conv.bias.shape)
           # Load weights
-          conv_biases = Tensor(weights[ptr: ptr+num_biases].astype(np.float32))
+          conv_biases = Tensor(weights[ptr: ptr+num_biases])
           ptr += num_biases
           # Reshape
           conv_biases = conv_biases.reshape(shape=tuple(conv.bias.shape))
@@ -327,7 +327,7 @@ class Darknet:
           conv.bias = conv_biases
         # Load weighys for conv layers
         num_weights = math.prod(conv.weight.shape)
-        conv_weights = Tensor(weights[ptr:ptr+num_weights].astype(np.float32))
+        conv_weights = Tensor(weights[ptr:ptr+num_weights])
         ptr += num_weights
         conv_weights = conv_weights.reshape(shape=tuple(conv.weight.shape))
         conv.weight = conv_weights
@@ -371,7 +371,7 @@ class Darknet:
 if __name__ == "__main__":
   model = Darknet(fetch('https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov3.cfg').read_bytes())
   print("Loading weights file (237MB). This might take a while…")
-  model.load_weights('https://github.com/shadiakiki1986/yolov3.weights/releases/download/3.0.1/yolov3.weights')
+  model.load_weights('https://pjreddie.com/media/files/yolov3.weights')
   if len(sys.argv) > 1:
     url = sys.argv[1]
   else:

extra/amdpci/proclogs.py

@@ -1,16 +1,6 @@
-import re, ctypes, sys, importlib
+import re, ctypes, sys
 
-from tinygrad.runtime.support.am.amdev import AMDev, AMRegister
+from tinygrad.runtime.autogen.am import am, mp_11_0, mp_13_0_0, nbio_4_3_0, mmhub_3_0_0, gc_11_0_0, osssys_6_0_0
-class AMDFake(AMDev):
-  def __init__(self, devfmt, vram, doorbell, mmio, dma_regions=None):
-    self.devfmt, self.vram, self.doorbell64, self.mmio, self.dma_regions = devfmt, vram, doorbell, mmio, dma_regions
-    self._run_discovery()
-    self._build_regs()
-
-amdev = importlib.import_module("tinygrad.runtime.support.am.amdev")
-amdev.AMDev = AMDFake
-
-from tinygrad.runtime.ops_amd import PCIIface
 
 def parse_amdgpu_logs(log_content, register_names=None):
   register_map = register_names
@@ -33,11 +23,16 @@ def parse_amdgpu_logs(log_content, register_names=None):
   return processed_log
 
 def main():
+  regs_offset = {13: {0: [3072, 37784576]}, 28: {0: [93184, 37754880], 1: [201327616, 201461760], 2: [209716224, 209850368], 3: [218104832, 218238976], 4: [226493440, 226627584], 5: [234882048, 235016192], 6: [243270656, 243404800]}, 21: {0: [28672, 12582912, 37795840, 130023424, 306184192], 1: [201326592, 201463808, 201465856, 204210176, 204472320], 2: [209715200, 209852416, 209854464, 212598784, 212860928], 3: [218103808, 218241024, 218243072, 220987392, 221249536], 4: [226492416, 226629632, 226631680, 229376000, 229638144], 5: [234881024, 235018240, 235020288, 237764608, 238026752], 6: [243269632, 243406848, 243408896, 246153216, 246415360]}, 22: {0: [18, 192, 13504, 36864, 37764096]}, 1: {0: [4704, 40960, 114688, 37760000]}, 2: {0: [3872, 37790720]}, 11: {0: [70656, 38103040]}, 12: {0: [106496, 37783552]}, 15: {0: [90112, 14417920, 14680064, 14942208, 38009856]}, 16: {0: [90112, 14417920, 14680064, 14942208, 38009856]}, 14: {0: [0, 20, 3360, 66560, 37859328, 67371008]}, 26: {0: [0, 20, 3360, 66560, 37859328, 67371008]}, 23: {0: [4256, 37789696]}, 33: {0: [0, 20, 3360, 66560, 37859328, 67371008]}, 25: {0: []}, 3: {0: [4704, 40960, 114688, 37760000]}, 4: {0: [4704, 40960, 114688, 37760000]}, 24: {0: [92160, 92672, 37752832, 54788096]}, 27: {0: [91648, 37751808], 1: [201339904, 201458176], 2: [209728512, 209846784], 3: [218117120, 218235392], 4: [226505728, 226624000], 5: [234894336, 235012608], 6: [243282944, 243401216]}, 29: {0: [201342976, 201344000, 205520896, 205537280], 1: [209731584, 209732608, 213909504, 213925888], 2: [218120192, 218121216, 222298112, 222314496], 3: [226508800, 226509824, 230686720, 230703104], 4: [234897408, 234898432, 239075328, 239091712], 5: [243286016, 243287040, 247463936, 247480320]}, 17: {0: [30720, 32256], 1: [31488, 73728]}}
+
   reg_names = {}
-  dev = PCIIface(None, 0)
+  def _prepare_registers(modules):
-  for x, y in dev.dev_impl.__dict__.items():
+    for base, m in modules:
-    if isinstance(y, AMRegister):
+      for k, regval in m.__dict__.items():
-      for inst, addr in y.addr.keys(): reg_names[addr] = f"{x}, xcc={inst}"
+        if k.startswith("reg") and not k.endswith("_BASE_IDX") and (base_idx:=getattr(m, f"{k}_BASE_IDX", None)) is not None:
+          reg_names[regs_offset[am.__dict__.get(f"{base}_HWIP")][0][base_idx] + regval] = k
+
+  _prepare_registers([("MP0", mp_13_0_0), ("NBIO", nbio_4_3_0), ("MMHUB", mmhub_3_0_0), ("GC", gc_11_0_0), ("OSSSYS", osssys_6_0_0)])
 
   with open(sys.argv[1], 'r') as f:
     log_content = log_content_them = f.read()

extra/assembly/assembly.py

@@ -1,5 +1,5 @@
 from typing import Tuple, List, NamedTuple, Any, Dict, Optional, Union, DefaultDict, cast
-from tinygrad.codegen.opt.kernel import Ops, MemOp, UOp
+from tinygrad.opt.kernel import Ops, MemOp, UOp
 from tinygrad.uop.ops import BinaryOps, UnaryOps
 from tinygrad.dtype import DType, dtypes
 from tinygrad.helpers import DEBUG

extra/assembly/assembly_arm64.py

@@ -3,7 +3,7 @@ from platform import system
 from typing import Tuple, Dict, List, Optional
 from tinygrad import dtypes
 from tinygrad.uop.ops import BinaryOps, UnaryOps, TernaryOps
-from tinygrad.codegen.opt.kernel import Ops, UOp
+from tinygrad.opt.kernel import Ops, UOp
 from tinygrad.helpers import CI
 from tinygrad.codegen.assembly import uops_to_asmstyle, AssemblyLanguage
 

extra/assembly/assembly_ptx.py

@@ -1,7 +1,7 @@
 from typing import List
 import struct
 from tinygrad.codegen.assembly import uops_to_asmstyle, AssemblyLanguage
-from tinygrad.codegen.opt.kernel import Ops, UOp
+from tinygrad.opt.kernel import Ops, UOp
 from tinygrad import dtypes
 from tinygrad.uop.ops import BinaryOps, UnaryOps, TernaryOps
 from tinygrad.runtime.ops_cuda import arch

extra/assembly/assembly_rdna.py

@@ -2,7 +2,7 @@ import yaml
 from typing import Tuple, Set, Dict
 from tinygrad import dtypes
 from tinygrad.codegen.assembly import AssemblyCodegen, Register
-from tinygrad.codegen.opt.kernel import Ops
+from tinygrad.opt.kernel import Ops
 from tinygrad.uop.ops import BinaryOps, UnaryOps, TernaryOps
 from tinygrad.runtime.ops_gpu import ROCM_LLVM_PATH
 

extra/backends/triton.py

@@ -2,7 +2,7 @@ from typing import Dict, List, Final, Callable, DefaultDict
 from collections import defaultdict
 from tinygrad.uop.ops import UnaryOps, BinaryOps, TernaryOps, Op
 from tinygrad.helpers import DType, PtrDType, dtypes, ImageDType, DEBUG, getenv
-from tinygrad.codegen.opt.kernel import  UOp, Ops
+from tinygrad.opt.kernel import  UOp, Ops
 from triton.compiler import compile as triton_compile
 import linecache
 import math

extra/gemm/amd_uop_matmul.py

@@ -2,11 +2,11 @@ from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, graph_rewrite, AxisType, PatternMatcher, UPat
 from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
 from tinygrad.dtype import AddrSpace
+from tinygrad.schedule.kernelize import merge_views, view_left
 from tinygrad.helpers import getenv, colored, prod, unwrap
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.shape.view import strides_for_shape
-from tinygrad.codegen.opt.kernel import axis_colors
+from tinygrad.opt.kernel import axis_colors
-from tinygrad.codegen.opt.swizzler import merge_views, view_left
 
 def to_colored(full_shape, axis_types): return '_'.join([colored(str(s), axis_colors[at]) for s,at in zip(full_shape, axis_types)])
 

extra/gemm/max_matmul.py

@@ -5,9 +5,9 @@ from typing import Optional, List, Tuple, cast, Dict, Final, DefaultDict, Self
 from tinygrad.engine.realize import get_program
 
 # for copied uops
-from tinygrad.codegen.opt.kernel import Kernel, KernelOptError
+from tinygrad.opt.kernel import Kernel, KernelOptError
 from tinygrad.uop.ops import UOp, Ops, BinaryOps, UnaryOps, TernaryOps, KernelInfo
-from tinygrad.codegen.opt.search import Opt, OptOps
+from tinygrad.opt.search import Opt, OptOps
 from tinygrad import Device, dtypes, Tensor
 from tinygrad.dtype import PtrDType, DType, DTYPES_DICT
 from tinygrad.shape.shapetracker import ShapeTracker

extra/gemm/simple_matmul.py

@@ -2,7 +2,7 @@ import numpy as np
 from tinygrad import dtypes, Tensor
 from tinygrad.helpers import getenv, get_single_element
 from tinygrad.dtype import _to_np_dtype
-from tinygrad.codegen.opt.kernel import OptOps
+from tinygrad.opt.kernel import OptOps
 from tinygrad.engine.realize import lower_schedule
 
 dtype_in = dtypes.half if getenv("HALF") else dtypes.bfloat16 if getenv("BFLOAT16") else dtypes.float

extra/gemm/tinygrad_nv_matmul.py

@@ -1,6 +1,6 @@
 from tinygrad import Tensor, dtypes, Device
 from tinygrad.helpers import getenv, DEBUG
-from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps
+from tinygrad.opt.kernel import Kernel, Opt, OptOps
 from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
 from dataclasses import replace
 

extra/gemm/tvm_gemm.py

@@ -37,7 +37,7 @@ B = Tensor.rand(K, N, device="CPU")
 C = (A.reshape(M, 1, K) * B.permute(1,0).reshape(1, N, K)).sum(axis=2)
 
 sched = C.schedule()
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.device import CompilerOptions
 lin = Kernel(sched[-1].ast, CompilerOptions(has_local=False, supports_float4=False))
 lin.to_program()

extra/hcqfuzz/tests/bert.py

@@ -8,6 +8,7 @@ bert_train_params = {
   "BS": 96,
   "EVAL_BS": 96,
   "FUSE_ARANGE": 1,
+  "FUSE_ARANGE_UINT": 0,
   "BASEDIR": "/raid/datasets/wiki",
 }
 

extra/huggingface_onnx/README.md

@@ -1,61 +0,0 @@
-# HuggingFace ONNX
-
-Tool for discovering, downloading, and validating ONNX models from HuggingFace.
-
-## Extra Dependencies
-
-```bash
-pip install huggingface_hub pyyaml requests onnx onnxruntime numpy
-```
-
-## Huggingface Manager (discovering and downloading)
-
-The `huggingface_manager.py` script discovers top ONNX models from HuggingFace, collects metadata, and optionally downloads them.
-
-```bash
-# Download top 50 models sorted by downloads
-python huggingface_manager.py --limit 50 --download
-
-# Just collect metadata (no download)
-python huggingface_manager.py --limit 100
-
-# Sort by likes instead of downloads
-python huggingface_manager.py --limit 20 --sort likes --download
-
-# Custom output file
-python huggingface_manager.py --limit 10 --output my_models.yaml
-```
-
-### Output Format
-
-The tool generates a YAML file with the following structure:
-
-```yaml
-repositories:
-  "model-name":
-    url: "https://huggingface.co/model-name"
-    download_path: "/path/to/models/..."  # when --download used
-    files:
-      - file: "model.onnx"
-        size: "90.91MB"
-total_size: "2.45GB"
-created_at: "2024-01-15T10:30:00Z"
-```
-
-## Run Models (validation)
-
-The `run_models.py` script validates ONNX models against ONNX Runtime for correctness.
-
-```bash
-# Validate models from a YAML configuration file
-python run_models.py --validate huggingface_repos.yaml
-
-# Debug specific repository (downloads and validates all ONNX models)
-python run_models.py --debug sentence-transformers/all-MiniLM-L6-v2
-
-# Debug specific model file
-python run_models.py --debug sentence-transformers/all-MiniLM-L6-v2/onnx/model.onnx
-
-# Debug with model truncation for debugging and validating intermediate results
-DEBUGONNX=1 python run_models.py --debug sentence-transformers/all-MiniLM-L6-v2/onnx/model.onnx --truncate 10
-```

extra/huggingface_onnx/collect_metadata.py

@@ -0,0 +1,85 @@
+import yaml, time, requests, argparse
+from pathlib import Path
+from huggingface_hub import list_models, HfApi
+from tinygrad.helpers import tqdm
+
+HUGGINGFACE_URL = "https://huggingface.co"
+SKIPPED_FILES = [
+  "fp16", "int8", "uint8", "quantized",      # numerical accuracy issues
+  "avx2", "arm64", "avx512", "avx512_vnni",  # numerical accuracy issues
+  "q4", "q4f16", "bnb4",                     # unimplemented quantization
+  "model_O4",                                # requires non cpu ort runner and MemcpyFromHost op
+  "merged",                                  # TODO implement attribute with graph type and Loop op
+]
+SKIPPED_REPO_PATHS = [
+  # Invalid model-index
+  "AdamCodd/vit-base-nsfw-detector",
+  # TODO: implement attribute with graph type and Loop op
+  "minishlab/potion-base-8M", "minishlab/M2V_base_output", "minishlab/potion-retrieval-32M",
+  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, GroupQueryAttention
+  "HuggingFaceTB/SmolLM2-360M-Instruct",
+  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, RotaryEmbedding, MultiHeadAttention
+  "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+  # TODO: implmement RandomNormalLike
+  "stabilityai/stable-diffusion-xl-base-1.0", "stabilityai/sdxl-turbo", 'SimianLuo/LCM_Dreamshaper_v7',
+  # TODO: implement NonZero
+  "mangoapps/fb_zeroshot_mnli_onnx",
+  # TODO huge Concat in here with 1024 (1, 3, 32, 32) Tensors, and maybe a MOD bug with const folding
+  "briaai/RMBG-2.0",
+]
+
+def get_top_repos(n: int, sort: str) -> list[str]: # list["FacebookAI/xlm-roberta-large", ...]
+  print(f"** Getting top {n} models sorted by {sort} **")
+  repos = []
+  i = 0
+  for model in list_models(filter="onnx", sort=sort):
+    if model.id in SKIPPED_REPO_PATHS: continue
+    print(f"{i+1}/{n}: {model.id} ({getattr(model, sort)})")
+    repos.append(model.id)
+    i += 1
+    if i == n: break
+  return repos
+
+def get_metadata(repos:list[str]) -> dict:
+  api = HfApi()
+  repos_metadata = {"repositories": {}}
+  total_size = 0
+
+  # TODO: speed head requests up with async?
+  for repo in tqdm(repos, desc="Getting metadata"):
+    files_metadata = []
+    model_info = api.model_info(repo)
+
+    for file in model_info.siblings:
+      filename = file.rfilename
+      if not (filename.endswith('.onnx') or filename.endswith('.onnx_data')): continue
+      if any(skip_str in filename for skip_str in SKIPPED_FILES): continue
+      head = requests.head(f"{HUGGINGFACE_URL}/{repo}/resolve/main/{filename}", allow_redirects=True)
+      file_size = file.size or int(head.headers.get('Content-Length', 0))
+      files_metadata.append({"file": filename, "size": f"{file_size/1e6:.2f}MB"})
+      total_size += file_size
+
+    repos_metadata["repositories"][repo] = {
+      "url": f"{HUGGINGFACE_URL}/{repo}",
+      "download_path": None,
+      "files": files_metadata,
+    }
+  repos_metadata['total_size'] = f"{total_size/1e9:.2f}GB"
+  repos_metadata['created_at'] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+  return repos_metadata
+
+if __name__ == "__main__":
+  sort = "downloads" # recent 30 days downloads
+  huggingface_onnx_dir = Path(__file__).parent
+
+  parser = argparse.ArgumentParser(description="Produces a YAML file with metadata of top huggingface onnx models")
+  parser.add_argument("--limit", type=int, required=True, help="Number of top repositories to process (e.g., 100)")
+  parser.add_argument("--output", type=str, default="huggingface_repos.yaml", help="Output YAML file name to save the report")
+  args = parser.parse_args()
+
+  top_repos = get_top_repos(args.limit, sort)
+  metadata = get_metadata(top_repos)
+  yaml_path = huggingface_onnx_dir / args.output
+  with open(yaml_path, 'w') as f:
+    yaml.dump(metadata, f, sort_keys=False)
+    print(f"YAML saved to: {str(yaml_path)}")

extra/huggingface_onnx/download_models.py

@@ -0,0 +1,29 @@
+import yaml, argparse
+from pathlib import Path
+from huggingface_hub import snapshot_download
+
+def download_models(yaml_file: str, download_dir: str) -> None:
+  with open(yaml_file, 'r') as f: metadata = yaml.safe_load(f)
+  n = len(metadata["repositories"])
+
+  for i, (model_id, model_data) in enumerate(metadata["repositories"].items()):
+    print(f"Downloading {i+1}/{n}: {model_id}...")
+    allow_patterns = [file_info["file"] for file_info in model_data["files"]]
+    root_path = Path(snapshot_download(repo_id=model_id, allow_patterns=allow_patterns, cache_dir=download_dir))
+    # download configs too (the sizes are small)
+    snapshot_download(repo_id=model_id, allow_patterns=["*config.json"], cache_dir=download_dir)
+    print(f"Downloaded model files to: {root_path}")
+    model_data["download_path"] = str(root_path)
+
+  # Save the updated metadata back to the YAML file
+  with open(yaml_file, 'w') as f: yaml.dump(metadata, f, sort_keys=False)
+  print("Download completed according to YAML file.")
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(description="Download models from Huggingface Hub based on a YAML configuration file.")
+  parser.add_argument("input", type=str, help="Path to the input YAML configuration file containing model information.")
+  args = parser.parse_args()
+
+  models_folder = Path(__file__).parent / "models"
+  models_folder.mkdir(parents=True, exist_ok=True)
+  download_models(args.input, str(models_folder))

extra/huggingface_onnx/huggingface_manager.py

@@ -1,230 +0,0 @@
-import yaml
-import time
-import requests
-import argparse
-from pathlib import Path
-from huggingface_hub import list_models, HfApi, snapshot_download
-from tinygrad.helpers import _ensure_downloads_dir
-DOWNLOADS_DIR = _ensure_downloads_dir() / "models"
-from tinygrad.helpers import tqdm
-
-def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, cache_dir: str|Path|None=None,
-                                 tries: int=2, **kwargs) -> Path:
-  for attempt in range(tries):
-    try:
-      return Path(snapshot_download(
-        repo_id=repo_id,
-        allow_patterns=allow_patterns,
-        cache_dir=str(cache_dir) if cache_dir is not None else None,
-        **kwargs
-      ))
-    except Exception as e:
-      if attempt == tries-1: raise
-      time.sleep(1)
-
-# Constants for filtering models
-HUGGINGFACE_URL = "https://huggingface.co"
-SKIPPED_FILES = [
-  "fp16", "int8", "uint8", "quantized",      # numerical accuracy issues
-  "avx2", "arm64", "avx512", "avx512_vnni",  # numerical accuracy issues
-  "q4", "q4f16", "bnb4",                     # unimplemented quantization
-  "model_O4",                                # requires non cpu ort runner and MemcpyFromHost op
-  "merged",                                  # TODO implement attribute with graph type and Loop op
-]
-
-SKIPPED_REPO_PATHS = [
-  # Invalid model-index
-  "AdamCodd/vit-base-nsfw-detector",
-  # TODO: implement attribute with graph type and Loop op
-  "minishlab/potion-base-8M", "minishlab/M2V_base_output", "minishlab/potion-retrieval-32M",
-  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, GroupQueryAttention
-  "HuggingFaceTB/SmolLM2-360M-Instruct",
-  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, RotaryEmbedding, MultiHeadAttention
-  "HuggingFaceTB/SmolLM2-1.7B-Instruct",
-  # TODO: implement RandomNormalLike
-  "stabilityai/stable-diffusion-xl-base-1.0", "stabilityai/sdxl-turbo", 'SimianLuo/LCM_Dreamshaper_v7',
-  # TODO: implement NonZero
-  "mangoapps/fb_zeroshot_mnli_onnx",
-  # TODO huge Concat in here with 1024 (1, 3, 32, 32) Tensors, and maybe a MOD bug with const folding
-  "briaai/RMBG-2.0",
-]
-
-
-class HuggingFaceONNXManager:
-  def __init__(self):
-    self.base_dir = Path(__file__).parent
-    self.models_dir = DOWNLOADS_DIR
-    self.api = HfApi()
-
-  def discover_models(self, limit: int, sort: str = "downloads") -> list[str]:
-    print(f"Discovering top {limit} ONNX models sorted by {sort}...")
-    repos = []
-    i = 0
-
-    for model in list_models(filter="onnx", sort=sort):
-      if model.id in SKIPPED_REPO_PATHS:
-        continue
-
-      print(f"  {i+1}/{limit}: {model.id} ({getattr(model, sort)})")
-      repos.append(model.id)
-      i += 1
-      if i == limit:
-        break
-
-    print(f"Found {len(repos)} suitable ONNX models")
-    return repos
-
-  def collect_metadata(self, repos: list[str]) -> dict:
-    print(f"Collecting metadata for {len(repos)} repositories...")
-    metadata = {"repositories": {}}
-    total_size = 0
-
-    for repo in tqdm(repos, desc="Collecting metadata"):
-      try:
-        files_metadata = []
-        model_info = self.api.model_info(repo)
-
-        for file in model_info.siblings:
-          filename = file.rfilename
-          if not (filename.endswith('.onnx') or filename.endswith('.onnx_data')):
-            continue
-          if any(skip_str in filename for skip_str in SKIPPED_FILES):
-            continue
-
-          # Get file size from API or HEAD request
-          try:
-            head = requests.head(
-              f"{HUGGINGFACE_URL}/{repo}/resolve/main/{filename}",
-              allow_redirects=True,
-              timeout=10
-            )
-            file_size = file.size or int(head.headers.get('Content-Length', 0))
-          except requests.RequestException:
-            file_size = file.size or 0
-
-          files_metadata.append({
-            "file": filename,
-            "size": f"{file_size/1e6:.2f}MB"
-          })
-          total_size += file_size
-
-        if files_metadata:  # Only add repos with valid ONNX files
-          metadata["repositories"][repo] = {
-            "url": f"{HUGGINGFACE_URL}/{repo}",
-            "download_path": None,
-            "files": files_metadata,
-          }
-
-      except Exception as e:
-        print(f"WARNING: Failed to collect metadata for {repo}: {e}")
-        continue
-
-    metadata['total_size'] = f"{total_size/1e9:.2f}GB"
-    metadata['created_at'] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
-
-    print(f"Collected metadata for {len(metadata['repositories'])} repositories")
-    print(f"Total estimated download size: {metadata['total_size']}")
-
-    return metadata
-
-  def download_models(self, metadata: dict) -> dict:
-    self.models_dir.mkdir(parents=True, exist_ok=True)
-
-    repos = metadata["repositories"]
-    n = len(repos)
-
-    print(f"Downloading {n} repositories to {self.models_dir}...")
-
-    for i, (model_id, model_data) in enumerate(repos.items()):
-      print(f"  Downloading {i+1}/{n}: {model_id}...")
-
-      try:
-        # Download ONNX model files
-        allow_patterns = [file_info["file"] for file_info in model_data["files"]]
-        root_path = snapshot_download_with_retry(
-          repo_id=model_id,
-          allow_patterns=allow_patterns,
-          cache_dir=str(self.models_dir)
-        )
-
-        # Download config files (usually small)
-        snapshot_download_with_retry(
-          repo_id=model_id,
-          allow_patterns=["*config.json"],
-          cache_dir=str(self.models_dir)
-        )
-
-        model_data["download_path"] = str(root_path)
-        print(f"    Downloaded to: {root_path}")
-
-      except Exception as e:
-        print(f"    ERROR: Failed to download {model_id}: {e}")
-        model_data["download_path"] = None
-        continue
-
-    successful_downloads = sum(1 for repo in repos.values() if repo["download_path"] is not None)
-    print(f"Successfully downloaded {successful_downloads}/{n} repositories")
-    print(f"All models saved to: {self.models_dir}")
-
-    return metadata
-
-  def save_metadata(self, metadata: dict, output_file: str):
-    yaml_path = self.base_dir / output_file
-    with open(yaml_path, 'w') as f:
-      yaml.dump(metadata, f, sort_keys=False)
-    print(f"Metadata saved to: {yaml_path}")
-
-  def discover_and_download(self, limit: int, output_file: str = "huggingface_repos.yaml",
-                          sort: str = "downloads", download: bool = True):
-    print(f"Starting HuggingFace ONNX workflow...")
-    print(f"   Limit: {limit} models")
-    print(f"   Sort by: {sort}")
-    print(f"   Download: {'Yes' if download else 'No'}")
-    print(f"   Output: {output_file}")
-    print("-" * 50)
-
-    repos = self.discover_models(limit, sort)
-
-    metadata = self.collect_metadata(repos)
-
-    if download:
-      metadata = self.download_models(metadata)
-
-    self.save_metadata(metadata, output_file)
-
-    print("-" * 50)
-    print("Workflow completed successfully!")
-    if download:
-      successful = sum(1 for repo in metadata["repositories"].values()
-                     if repo["download_path"] is not None)
-      print(f"{successful}/{len(metadata['repositories'])} models downloaded")
-
-    return metadata
-
-
-if __name__ == "__main__":
-  parser = argparse.ArgumentParser(
-    description="HuggingFace ONNX Model Manager - Discover, collect metadata, and download ONNX models",
-  )
-
-  parser.add_argument("--limit", type=int, help="Number of top repositories to process")
-  parser.add_argument("--output", type=str, default="huggingface_repos.yaml",
-                     help="Output YAML file name (default: huggingface_repos.yaml)")
-  parser.add_argument("--sort", type=str, default="downloads",
-                     choices=["downloads", "likes", "created", "modified"],
-                     help="Sort criteria for model discovery (default: downloads)")
-
-  parser.add_argument("--download", action="store_true", default=False,
-                     help="Download models after collecting metadata")
-
-  args = parser.parse_args()
-
-  if not args.limit: parser.error("--limit is required")
-
-  manager = HuggingFaceONNXManager()
-  manager.discover_and_download(
-    limit=args.limit,
-    output_file=args.output,
-    sort=args.sort,
-    download=args.download
-  )

extra/huggingface_onnx/run_models.py

@@ -1,11 +1,10 @@
-import onnx, yaml, tempfile, time, argparse, json
+import onnx, yaml, tempfile, time, collections, pprint, argparse, json
 from pathlib import Path
-from typing import Any
 from tinygrad.frontend.onnx import OnnxRunner
+from extra.onnx import get_onnx_ops
 from extra.onnx_helpers import validate, get_example_inputs
-from extra.huggingface_onnx.huggingface_manager import DOWNLOADS_DIR, snapshot_download_with_retry
 
-def get_config(root_path: Path) -> dict[str, Any]:
+def get_config(root_path: Path):
   ret = {}
   for path in root_path.rglob("*config.json"):
     config = json.load(path.open())
@@ -13,19 +12,19 @@ def get_config(root_path: Path) -> dict[str, Any]:
       ret.update(config)
   return ret
 
-def get_tolerances(file_name: str) -> tuple[float, float]:
+def run_huggingface_validate(onnx_model_path, config, rtol, atol):
+  onnx_runner = OnnxRunner(onnx_model_path)
+  inputs = get_example_inputs(onnx_runner.graph_inputs, config)
+  validate(onnx_model_path, inputs, rtol=rtol, atol=atol)
+
+def get_tolerances(file_name): # -> rtol, atol
   # TODO very high rtol atol
   if "fp16" in file_name: return 9e-2, 9e-2
   if any(q in file_name for q in ["int8", "uint8", "quantized"]): return 4, 4
   return 4e-3, 3e-2
 
-def run_huggingface_validate(onnx_model_path: str | Path, config: dict[str, Any], rtol: float, atol: float):
-  onnx_runner = OnnxRunner(onnx_model_path)
-  inputs = get_example_inputs(onnx_runner.graph_inputs, config)
-  validate(onnx_model_path, inputs, rtol=rtol, atol=atol)
-
 def validate_repos(models:dict[str, tuple[Path, Path]]):
-  print(f"** Validating {len(models)} models **")
+  print(f"** Validating {len(model_paths)} models **")
   for model_id, (root_path, relative_path) in models.items():
     print(f"validating model {model_id}")
     model_path = root_path / relative_path
@@ -37,6 +36,25 @@ def validate_repos(models:dict[str, tuple[Path, Path]]):
     et = time.time() - st
     print(f"passed, took {et:.2f}s")
 
+def retrieve_op_stats(models:dict[str, tuple[Path, Path]]) -> dict:
+  ret = {}
+  op_counter = collections.Counter()
+  unsupported_ops = collections.defaultdict(set)
+  supported_ops = get_onnx_ops()
+  print(f"** Retrieving stats from {len(model_paths)} models **")
+  for model_id, (root_path, relative_path) in models.items():
+    print(f"examining {model_id}")
+    model_path = root_path / relative_path
+    onnx_runner = OnnxRunner(model_path)
+    for node in onnx_runner.graph_nodes:
+      op_counter[node.op] += 1
+      if node.op not in supported_ops:
+        unsupported_ops[node.op].add(model_id)
+    del onnx_runner
+  ret["unsupported_ops"] = {k:list(v) for k, v in unsupported_ops.items()}
+  ret["op_counter"] = op_counter.most_common()
+  return ret
+
 def debug_run(model_path, truncate, config, rtol, atol):
   if truncate != -1:
     model = onnx.load(model_path)
@@ -53,9 +71,12 @@ def debug_run(model_path, truncate, config, rtol, atol):
     run_huggingface_validate(model_path, config, rtol, atol)
 
 if __name__ == "__main__":
-  parser = argparse.ArgumentParser(description="Huggingface ONNX Model Validator")
+  parser = argparse.ArgumentParser(description="Huggingface ONNX Model Validator and Ops Checker")
-  parser.add_argument("--validate", type=str, default="",
+  parser.add_argument("input", type=str, help="Path to the input YAML configuration file containing model information.")
-                      help="Validate correctness of models from the specified YAML configuration file")
+  parser.add_argument("--check_ops", action="store_true", default=False,
+                      help="Check support for ONNX operations in models from the YAML file")
+  parser.add_argument("--validate", action="store_true", default=False,
+                      help="Validate correctness of models from the YAML file")
   parser.add_argument("--debug", type=str, default="",
                       help="""Validates without explicitly needing a YAML or models pre-installed.
                       provide repo id (e.g. "minishlab/potion-base-8M") to validate all onnx models inside the repo
@@ -64,13 +85,13 @@ if __name__ == "__main__":
   parser.add_argument("--truncate", type=int, default=-1, help="Truncate the ONNX model so intermediate results can be validated")
   args = parser.parse_args()
 
-  if not (args.validate or args.debug):
+  if not (args.check_ops or args.validate or args.debug):
-    parser.error("Please provide either --validate <yaml_file> or --debug <repo_id>.")
+    parser.error("Please provide either --validate, --check_ops, or --debug.")
   if args.truncate != -1 and not args.debug:
     parser.error("--truncate and --debug should be used together for debugging")
 
-  if args.validate:
+  if args.check_ops or args.validate:
-    with open(args.validate, 'r') as f:
+    with open(args.input, 'r') as f:
       data = yaml.safe_load(f)
       assert all(repo["download_path"] is not None for repo in data["repositories"].values()), "please run `download_models.py` for this yaml"
       model_paths = {
@@ -80,16 +101,22 @@ if __name__ == "__main__":
         if model["file"].endswith(".onnx")
       }
 
-    validate_repos(model_paths)
+    if args.check_ops:
+      pprint.pprint(retrieve_op_stats(model_paths))
+
+    if args.validate:
+      validate_repos(model_paths)
 
   if args.debug:
+    from huggingface_hub import snapshot_download
+    download_dir = Path(__file__).parent / "models"
     path:list[str] = args.debug.split("/")
     if len(path) == 2:
       # repo id
       # validates all onnx models inside repo
       repo_id = "/".join(path)
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=DOWNLOADS_DIR)
+      root_path = Path(snapshot_download(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=download_dir))
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      snapshot_download(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=download_dir)
       config = get_config(root_path)
       for onnx_model in root_path.rglob("*.onnx"):
         rtol, atol = get_tolerances(onnx_model.name)
@@ -101,8 +128,8 @@ if __name__ == "__main__":
       onnx_model = path[-1]
       assert path[-1].endswith(".onnx")
       repo_id, relative_path = "/".join(path[:2]), "/".join(path[2:])
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=DOWNLOADS_DIR)
+      root_path = Path(snapshot_download(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=download_dir))
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      snapshot_download(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=download_dir)
       config = get_config(root_path)
       rtol, atol = get_tolerances(onnx_model)
       print(f"validating {relative_path} with truncate={args.truncate}, {rtol=}, {atol=}")

extra/mcts_search.py

@@ -4,9 +4,9 @@ import numpy as np
 np.set_printoptions(suppress=True)
 import math, functools, time, random, statistics
 from tinygrad.helpers import DEBUG, getenv, CACHELEVEL, diskcache_get, diskcache_put, colored, Profiling
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.device import Buffer, Device, CompileError
-from tinygrad.codegen.opt.search import _ensure_buffer_alloc, get_kernel_actions, _time_program
+from tinygrad.opt.search import _ensure_buffer_alloc, get_kernel_actions, _time_program
 from tinygrad.engine.realize import get_program
 
 class MCTSNode:

extra/models/llama.py

@@ -181,11 +181,13 @@ class Transformer:
   def forward(self, tokens:Tensor, start_pos:Union[Variable,int], temperature:float, top_k:int, top_p:float, alpha_f:float, alpha_p:float):
     _bsz, seqlen = tokens.shape
     h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, start_pos:start_pos+seqlen, :, :, :]
+
+    self.freqs_cis = self.freqs_cis.cast(h.dtype).contiguous()
+    freqs_cis = self.freqs_cis[:, start_pos:start_pos+seqlen, :, :, :]
 
     mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1) if seqlen > 1 else None
     for layer in self.layers: h = layer(h, start_pos, freqs_cis, mask)
-    logits = self.output(self.norm(h))
+    logits = self.output(self.norm(h)).float()
     if math.isnan(temperature): return logits
 
     return sample(logits[:, -1, :].flatten(), temperature, top_k, top_p, alpha_f, alpha_p)

extra/onnx_helpers.py

@@ -1,6 +1,7 @@
 from tinygrad import Tensor
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.frontend.onnx import OnnxRunner, OnnxValue
+from tinygrad.frontend.onnx import OnnxRunner
+from extra.onnx import OnnxValue
 import numpy as np
 import onnxruntime as ort
 

extra/optimization/extract_policynet.py

@@ -5,9 +5,9 @@ from tinygrad.nn import Linear
 from tinygrad.tensor import Tensor
 from tinygrad.nn.optim import Adam
 from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import actions
+from tinygrad.opt.search import actions
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, assert_same_lin
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.helpers import getenv
 
 # stuff needed to unpack a kernel
@@ -17,7 +17,7 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.uop.ops import Variable
 inf, nan = float('inf'), float('nan')
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps
 
 INNER = 256
 class PolicyNet:

extra/optimization/extract_sa_pairs.py

@@ -10,11 +10,11 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.uop.ops import Variable
 inf, nan = float('inf'), float('nan')
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps
 
 # more stuff
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import actions
+from tinygrad.opt.search import actions
 from extra.optimization.helpers import lin_to_feats
 from extra.optimization.pretrain_valuenet import ValueNet
 from tinygrad.nn.optim import Adam

extra/optimization/get_action_space.py

@@ -1,8 +1,8 @@
 import random
 from extra.optimization.helpers import load_worlds, ast_str_to_lin
-from tinygrad.codegen.opt.search import actions
+from tinygrad.opt.search import actions
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from tinygrad.helpers import tqdm
 
 tactions = set()

extra/optimization/helpers.py

@@ -1,6 +1,6 @@
 # stuff needed to unpack a kernel
 from tinygrad import Variable
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.dtype import dtypes, PtrDType
 from tinygrad.shape.shapetracker import ShapeTracker
@@ -11,7 +11,7 @@ inf, nan = float('inf'), float('nan')
 UOps = Ops
 
 # kernel unpacker
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 def ast_str_to_ast(ast_str:str) -> UOp: return eval(ast_str)
 def ast_str_to_lin(ast_str:str, opts=None): return Kernel(ast_str_to_ast(ast_str), opts=opts)
 def kern_str_to_lin(kern_str:str, opts=None):
@@ -103,7 +103,7 @@ def lin_to_feats(lin:Kernel, use_sts=True):
   return ret
 
 from tinygrad.device import Device, Buffer
-from tinygrad.codegen.opt.search import _ensure_buffer_alloc, _time_program
+from tinygrad.opt.search import _ensure_buffer_alloc, _time_program
 from tinygrad.helpers import to_function_name, CACHELEVEL, diskcache_get, diskcache_put
 
 def time_linearizer(lin:Kernel, rawbufs:list[Buffer], allow_test_size=True, max_global_size=65536, cnt=3, disable_cache=False, clear_l2=False) -> float:  # noqa: E501

extra/optimization/pretrain_valuenet.py

@@ -1,4 +1,4 @@
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tqdm import tqdm, trange
 import math
 import random
@@ -14,7 +14,7 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.uop.ops import Variable
 inf, nan = float('inf'), float('nan')
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps
 
 from extra.optimization.helpers import lin_to_feats, MAX_DIMS
 

extra/optimization/rl.py

@@ -3,7 +3,7 @@ import numpy as np
 import math, random
 from tinygrad.tensor import Tensor
 from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import actions, bufs_from_lin, get_kernel_actions
+from tinygrad.opt.search import actions, bufs_from_lin, get_kernel_actions
 from tinygrad.nn.optim import Adam
 from extra.optimization.extract_policynet import PolicyNet
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, time_linearizer

extra/optimization/run_qnet.py

@@ -1,6 +1,6 @@
 from typing import List, Tuple
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import get_kernel_actions, actions
+from tinygrad.opt.search import get_kernel_actions, actions
 
 _net = None
 def beam_q_estimate(beam:List[Tuple[Kernel, float]]) -> List[Tuple[Kernel, float]]:

extra/optimization/search.py

@@ -4,8 +4,8 @@ from extra.optimization.helpers import ast_str_to_lin, time_linearizer
 from tinygrad import dtypes
 from tinygrad.helpers import BEAM, getenv
 from tinygrad.device import Device, Compiled
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
+from tinygrad.opt.search import beam_search, bufs_from_lin
 
 
 if __name__ == '__main__':

extra/optimization/test_net.py

@@ -6,8 +6,8 @@ from copy import deepcopy
 from tinygrad.helpers import getenv, colored
 from tinygrad.tensor import Tensor
 from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import bufs_from_lin, actions, get_kernel_actions
+from tinygrad.opt.search import bufs_from_lin, actions, get_kernel_actions
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, time_linearizer
 from extra.optimization.extract_policynet import PolicyNet
 from extra.optimization.pretrain_valuenet import ValueNet

extra/optimization/test_time_linearizer.py

@@ -1,5 +1,5 @@
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, time_linearizer
-from tinygrad.codegen.opt.search import bufs_from_lin, get_kernel_actions
+from tinygrad.opt.search import bufs_from_lin, get_kernel_actions
 
 if __name__ == "__main__":
   ast_strs = load_worlds()

extra/replay_pkl.py

@@ -6,8 +6,8 @@ from tinygrad.helpers import getenv, BEAM
 from tinygrad.engine.jit import TinyJit
 from tinygrad.engine.realize import CompiledRunner, ExecItem, ScheduleItem, lower_schedule_item, get_program
 from tinygrad.renderer import ProgramSpec
-from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps
+from tinygrad.opt.kernel import Kernel, Opt, OptOps
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
 import numpy as np
 
 def move_jit_captured_to_dev(captured, device="DSP"):

extra/torch_backend/backend.py

@@ -223,18 +223,15 @@ def max_unpool2d(self:torch.Tensor, indices:torch.Tensor, output_size):
 
 @torch.library.impl("aten::arange", "privateuseone")
 def arange(end, dtype=None, device=None, pin_memory=None):
-  has_float = isinstance(end, float)
+  return wrap(Tensor.arange(0, end, dtype=_from_torch_dtype(dtype or torch.get_default_dtype())))
-  return wrap(Tensor.arange(0, end, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
 
 @torch.library.impl("aten::arange.start", "privateuseone")
 def arange_start(start, end, dtype=None, device=None, pin_memory=None):
-  has_float = any(isinstance(x, float) for x in (start, end))
+  return wrap(Tensor.arange(start, end, dtype=_from_torch_dtype(dtype or torch.get_default_dtype())))
-  return wrap(Tensor.arange(start, end, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
 
 @torch.library.impl("aten::arange.start_step", "privateuseone")
 def arange_start_step(start, end, step, dtype=None, device=None, pin_memory=None):
-  has_float = any(isinstance(x, float) for x in (start, end, step))
+  return wrap(Tensor.arange(start, end, step, dtype=_from_torch_dtype(dtype or torch.get_default_dtype())))
-  return wrap(Tensor.arange(start, end, step, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
 
 @torch.library.impl("aten::convolution_overrideable", "privateuseone")
 def convolution_overrideable(input, weight, bias, stride, padding, dilation, transposed, output_padding, groups):
@@ -371,7 +368,6 @@ from torch._decomp import get_decompositions
 decomps = [
   aten.native_batch_norm, aten.native_batch_norm_backward,
   aten.native_layer_norm_backward,
-  aten.linalg_cross,
   aten.addmm,
   aten.addcmul,
   aten.addcdiv,

extra/torch_backend/test.py

@@ -135,7 +135,7 @@ class TestTorchBackend(unittest.TestCase):
     print(c.cpu())
 
   def test_maxpool2d_backward(self):
-    x = torch.arange(3*3, dtype=torch.float32, device=device).reshape(1, 1, 3, 3).requires_grad_(True)
+    x = torch.arange(3*3, device=device).reshape(1, 1, 3, 3).requires_grad_(True)
     torch.nn.functional.max_pool2d(x, kernel_size=2, stride=1).sum().backward()
     np.testing.assert_equal(x.grad.squeeze().cpu().numpy(), [[0, 0, 0], [0, 1, 1], [0, 1, 1]])
 
@@ -203,12 +203,6 @@ class TestTorchBackend(unittest.TestCase):
     b = torch.linalg.det(a)
     np.testing.assert_equal(b.cpu().numpy(), 120.0)
 
-  def test_linalg_cross(self):
-    a = torch.tensor([[1, 0, 0], [0, 1, 0]], dtype=torch.float32, device=device)
-    b = torch.tensor([[0, 0, 1]], dtype=torch.float32, device=device)
-    cross = torch.linalg.cross(a, b)
-    np.testing.assert_equal(cross.cpu().numpy(), np.array([[0, -1, 0], [1, 0, 0]], dtype=np.float32))
-
   def test_scalar_assign(self):
     a = torch.tensor([1, 2, 3], device=device)
     a[1] = 4

extra/torch_muon.py

@@ -1,75 +0,0 @@
-import torch
-
-#credit to KellerJordan at https://github.com/KellerJordan/Muon/tree/master
-#some changes: classic momentum instead of weighting gradient
-#added ns_steps, ns_params, nesterov as hyperparams
-def zeropower_via_newtonschulz5(G:torch.tensor, steps:int, params:tuple[int, ...]):
-  """
-  Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
-  quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
-  of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
-  zero even beyond the point where the iteration no longer converges all the way to one everywhere
-  on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
-  where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
-  performance at all relative to UV^T, where USV^T = G is the SVD.
-  """
-  assert G.ndim >= 2 # batched Muon implementation by @scottjmaddox, and put into practice in the record by @YouJiacheng
-
-  a, b, c = params
-  X = G
-  if G.size(-2) > G.size(-1):
-    X = X.mT
-
-  # Ensure spectral norm is at most 1
-  X = X / (X.norm(dim=(-2, -1), keepdim=True) + 1e-7)
-  # Perform the NS iterations
-  for _ in range(steps):
-    A = X @ X.mT
-    B = b * A + c * A @ A # quintic computation strategy adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
-    X = a * X + B @ X
-
-  if G.size(-2) > G.size(-1):
-    X = X.mT
-
-  return X
-
-def muon_update(grad, momentum, beta=0.95, ns_steps=5, ns_params=(3.4445, -4.7750,  2.0315), nesterov=True):
-  if beta:
-    momentum.mul_(beta).add_(grad)
-    update = grad.add(momentum,alpha=beta) if nesterov else momentum
-  else: update = grad
-  if update.ndim == 4: # for the case of conv filters
-    update = update.view(len(update), -1)
-  update = zeropower_via_newtonschulz5(update, steps=ns_steps, params=ns_params)
-  return update
-
-class SingleDeviceMuon(torch.optim.Optimizer):
-  """
-  Muon variant for usage in non-distributed settings.
-  """
-  def __init__(self, params, lr=0.02, weight_decay=0.0, momentum=0.95, ns_steps=5, ns_params=(3.4445, -4.7750,  2.0315), nesterov=True):
-    defaults = dict(lr=lr, weight_decay=weight_decay, momentum=momentum, ns_steps=ns_steps, ns_params=ns_params, nesterov=nesterov)
-    super().__init__(params, defaults)
-
-  @torch.no_grad()
-  def step(self, closure=None):
-
-    loss = None
-    if closure is not None:
-      with torch.enable_grad():
-        loss = closure()
-
-    for group in self.param_groups:
-      for p in group["params"]:
-        if p.grad is None:
-          p.grad = torch.zeros_like(p)  # Force synchronization
-        state = self.state[p]
-        if len(state) == 0:
-          state["momentum_buffer"] = torch.zeros_like(p)
-        update = muon_update(p.grad, state["momentum_buffer"], beta=group["momentum"], ns_steps=group["ns_steps"],
-                             ns_params=group["ns_params"], nesterov=group["nesterov"])
-        p.mul_(1.0 - group["lr"] * group["weight_decay"])
-
-        p.add_(update.reshape(p.shape), alpha=-group["lr"])
-
-    return loss

pytest.ini

@@ -1,2 +0,0 @@
-[pytest]
-norecursedirs = extra

setup.py

@@ -18,34 +18,16 @@ testing_minimal = [
 ]
 
 setup(name='tinygrad',
-      version='0.11.0',
+      version='0.10.3',
       description='You like pytorch? You like micrograd? You love tinygrad! <3',
       author='George Hotz',
       license='MIT',
       long_description=long_description,
       long_description_content_type='text/markdown',
-      packages = [
+      packages = ['tinygrad', 'tinygrad.runtime.autogen', 'tinygrad.runtime.autogen.am', 'tinygrad.codegen', 'tinygrad.nn',
-        'tinygrad',
+                  'tinygrad.renderer', 'tinygrad.engine', 'tinygrad.viz', 'tinygrad.runtime', 'tinygrad.runtime.support', 'tinygrad.schedule',
-        'tinygrad.apps',
+                  'tinygrad.runtime.support.am', 'tinygrad.runtime.graph', 'tinygrad.shape', 'tinygrad.uop', 'tinygrad.opt',
-        'tinygrad.codegen',
+                  'tinygrad.runtime.support.nv', 'tinygrad.apps'],
-        'tinygrad.codegen.opt',
-        'tinygrad.engine',
-        'tinygrad.frontend',
-        'tinygrad.nn',
-        'tinygrad.renderer',
-        'tinygrad.runtime',
-        'tinygrad.runtime.autogen',
-        'tinygrad.runtime.autogen.am',
-        'tinygrad.runtime.autogen.nv',
-        'tinygrad.runtime.graph',
-        'tinygrad.runtime.support',
-        'tinygrad.runtime.support.am',
-        'tinygrad.runtime.support.nv',
-        'tinygrad.schedule',
-        'tinygrad.shape',
-        'tinygrad.uop',
-        'tinygrad.viz',
-      ],
       package_data = {'tinygrad': ['py.typed'], 'tinygrad.viz': ['index.html', 'assets/**/*', 'js/*']},
       classifiers=[
         "Programming Language :: Python :: 3",

test/external/external_benchmark_hcopt.py

@@ -1,7 +1,7 @@
 import random
 from tinygrad.helpers import getenv
-from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
+from tinygrad.opt.search import beam_search, bufs_from_lin
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, time_linearizer
 
 def optimize_kernel(k):

test/external/external_benchmark_kernel_launch.py

@@ -2,7 +2,7 @@ import time
 from tinygrad import Tensor, TinyJit, Device, Context
 from tinygrad.helpers import Profiling, Timing, GlobalCounters
 
-# python3 test/speed/external_test_speed_v_torch.py TestSpeed.test_add_a
+# python3 test/test_speed_v_torch.py TestSpeed.test_add_a
 
 @TinyJit
 def plus(a:Tensor, b:Tensor): return a+b

test/external/external_benchmark_schedule.py

@@ -1,8 +1,12 @@
+from typing import List
 from extra.models.resnet import ResNet50
-from tinygrad import Tensor, nn, Device
+from tinygrad import Tensor, nn
-from tinygrad.helpers import Profiling, Timing, getenv
+from tinygrad.helpers import Profiling, Timing, getenv, BEAM, NOOPT, DEBUG, Context, ansilen
 from tinygrad.uop.ops import Ops
+from tinygrad.opt.kernel import Kernel
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from tinygrad.codegen import get_rewrites_for_renderer, apply_rewrites, rewrites_for_linearizer
+from tinygrad.opt.search import beam_search, bufs_from_lin
 from tinygrad.uop.spec import type_verify
 
 if __name__ == "__main__":
@@ -27,13 +31,26 @@ if __name__ == "__main__":
       if not SCHEDULE_ONLY:
         asts = list({x.ast.key:x.ast for x in sched if x.ast.op is Ops.SINK}.values())
         if (restrict_kernel := getenv("RESTRICT_KERNEL", -1)) != -1: asts = asts[restrict_kernel:restrict_kernel+1]
+        kernels: List[Kernel] = []
+        with Timing(f"***** model opts({len(asts):2d}) in  "):
+          with Profiling(PROFILE >= 3):
+            for ast in asts:
+              k = Kernel(ast)
+              if BEAM:
+                with Context(DEBUG=max(2, DEBUG.value)): k = beam_search(k, bufs_from_lin(k), BEAM.value)
+              elif NOOPT: pass
+              else: k.apply_opts(hand_coded_optimizations(k))
+              kernels.append(k)
+
+        with Timing("***** model prep in      "):
+          kernels = [(k, k.get_optimized_ast(), get_rewrites_for_renderer(k.opts, linearizer=False)) for k in kernels]
 
-        rewrites = get_rewrites_for_renderer(Device.default.renderer, linearizer=False)
         with Profiling(PROFILE, fn="/tmp/rewrite.prof"):
           with Timing("***** model rewrite in   "):
             rewritten_uops = []
-            for u in asts:
+            for i,(k,u,rewrites) in enumerate(kernels):
-              rewritten_uops.append(apply_rewrites(u, rewrites))
+              with Timing(f"rewrite {i:2d} {k.name}{' '*(50-ansilen(k.name))}", enabled=getenv("VERBOSE", 0)):
+                rewritten_uops.append(apply_rewrites(u, rewrites))
 
         if LINEARIZE:
           with Timing("***** model linearize in "):

test/external/external_benchmark_sdxl_softmax.py

@@ -8,7 +8,7 @@ if __name__ == "__main__":
   GlobalCounters.reset()
   t.softmax(-1, dtype="half", _single_kernel=True).realize()
 
-  from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps
+  from tinygrad.opt.kernel import Kernel, Opt, OptOps
   from tinygrad.helpers import get_single_element
   GlobalCounters.reset()
   si = get_single_element(t.softmax(-1, dtype="half", _single_kernel=True).schedule())

test/external/external_debug_metal_sd_conv.py

@@ -1,8 +1,8 @@
 # ruff: noqa: E501
-from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps
+from tinygrad.opt.kernel import Kernel, Opt, OptOps
 from tinygrad.dtype import dtypes
 from tinygrad.engine.realize import CompiledRunner, get_program
-from tinygrad.codegen.opt.search import bufs_from_lin
+from tinygrad.opt.search import bufs_from_lin
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View

test/external/external_test_hcq_fuzz_failures.py

@@ -4,10 +4,10 @@ os.environ["VALIDATE_HCQ"]="1"
 
 import unittest, random
 import numpy as np
-from tinygrad.codegen.opt.kernel import Kernel, KernelOptError
+from tinygrad.opt.kernel import Kernel, KernelOptError
 from tinygrad.device import is_dtype_supported
 from tinygrad.uop.ops import UOp, Ops
-from tinygrad.codegen.opt.search import Opt, OptOps
+from tinygrad.opt.search import Opt, OptOps
 from tinygrad import Device, dtypes, Tensor
 from test.external.fuzz_linearizer import compare_linearizer, compare_states, get_fuzz_rawbuf_like
 

test/external/external_test_hip_compile.py

@@ -3,7 +3,7 @@ from tinygrad.runtime.support.hip_comgr import compile_hip
 from tinygrad import Tensor
 from tinygrad.device import Device
 from tinygrad.engine.schedule import create_schedule
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 
 class TestHIPCompileSpeed(unittest.TestCase):
   @unittest.skipIf(Device.DEFAULT != "HIP", "only run on HIP")

test/external/external_test_nv.py

@@ -2,11 +2,11 @@ import unittest, struct, array, ctypes
 from tinygrad import Device, dtypes, Tensor
 from tinygrad.helpers import to_mv
 from tinygrad.runtime.ops_nv import NVDevice, HWQueue
-from tinygrad.codegen.opt.search import Opt, OptOps
+from tinygrad.opt.search import Opt, OptOps
 from tinygrad.engine.realize import get_runner, CompiledRunner, get_program
 from test.external.fuzz_linearizer import get_fuzz_rawbufs
 
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.uop.ops import LazyOp, Ops, ReduceOps, BufferOps, MemBuffer
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View

test/external/external_test_onnx_backend.py

@@ -53,7 +53,6 @@ backend_test.exclude('test_dynamicquantizelinear_cpu')
 backend_test.exclude('test_dynamicquantizelinear_expanded_cpu')
 
 # BUG: ORT fails these with numerical error but we match ORT numerically
-# see: https://onnx.ai/backend-scoreboard/onnxruntime_details_stable.html
 # tested in external_test_onnx_ops.py::TestMainOnnxOps.test_qlinearmatmul_2D_int8_float16
 backend_test.exclude('test_qlinearmatmul_2D_int8_float16_cpu')
 # tested in external_test_onnx_ops.py::TestMainOnnxOps.test_qlinearmatmul_3D_int8_float16
@@ -66,10 +65,6 @@ backend_test.exclude('test_qlinearmatmul_3D_int8_float32_cpu')
 backend_test.exclude('test_maxunpool_export_with_output_shape_cpu')
 # tested in external_test_onnx_ops.py::TestMainOnnxOps.test_averagepool_3d_dilations_large_count_include_pad_is_1_ceil_mode_is_True
 backend_test.exclude('test_averagepool_3d_dilations_large_count_include_pad_is_1_ceil_mode_is_True_cpu')
-# tested in external_test_onnx_ops.py::TestMainOnnxOps.test_resize_downsample_scales_linear_align_corners
-backend_test.exclude('test_resize_downsample_scales_linear_align_corners_cpu')
-# tested in external_test_onnx_ops.py::TestMainOnnxOps.test_resize_downsample_scales_cubic_align_corners
-backend_test.exclude('test_resize_downsample_scales_cubic_align_corners_cpu')
 
 # about different dtypes
 if not is_dtype_supported(dtypes.float64):
@@ -170,6 +165,10 @@ backend_test.exclude('test_deform_conv_*')
 backend_test.exclude('test_lppool_*')
 backend_test.exclude('test_scan_*')
 backend_test.exclude('test_split_to_sequence_*')
+backend_test.exclude('test_resize_downsample_scales_cubic_*') # unsure how to implement cubic
+backend_test.exclude('test_resize_downsample_sizes_cubic_*') # unsure how to implement cubic
+backend_test.exclude('test_resize_upsample_scales_cubic_*') # unsure how to implement cubic
+backend_test.exclude('test_resize_upsample_sizes_cubic_*') # unsure how to implement cubic
 backend_test.exclude('test_ai_onnx_ml_tree_ensemble_*') # https://github.com/onnx/onnx/blob/main/onnx/reference/ops/aionnxml/op_tree_ensemble.py#L121
 
 # rest of the failing tests
@@ -179,8 +178,6 @@ backend_test.exclude('test_resize_tf_crop_and_resize_axes_3_2_cpu') # tf_crop_an
 backend_test.exclude('test_resize_tf_crop_and_resize_extrapolation_value_cpu') # tf_crop_and_resize value not implemented
 backend_test.exclude('test_resize_downsample_scales_linear_antialias_cpu') # antialias not implemented
 backend_test.exclude('test_resize_downsample_sizes_linear_antialias_cpu') # antialias not implemented
-backend_test.exclude('test_resize_downsample_scales_cubic_antialias_cpu') # antialias not implemented
-backend_test.exclude('test_resize_downsample_sizes_cubic_antialias_cpu') # antialias not implemented
 backend_test.exclude('test_ai_onnx_ml_label_encoder_tensor_value_only_mapping_cpu') # bad data type string
 backend_test.exclude('test_ai_onnx_ml_label_encoder_tensor_mapping_cpu') # bad data type string
 

test/external/external_test_onnx_ops.py

@@ -75,49 +75,6 @@ class TestMainOnnxOps(TestOnnxOps):
     outputs = ["y"]
     self.helper_test_single_op("Gather", inputs, attributes, outputs)
 
-  # NOTE: resize OP is sensitive to numerical errors
-  def _test_resize_scales(self, scale_values, **kwargs):
-    for sc in scale_values:
-      for ct_mode in ["half_pixel", "align_corners", "asymmetric", "pytorch_half_pixel", "half_pixel_symmetric"]:
-        with self.subTest(coordinate_transformation_mode=ct_mode, scale=sc, **kwargs):
-          X = np.array([[[[1, 2, 3, 4],
-                          [5, 6, 7, 8],
-                          [9,10,11,12]]]], dtype=np.float32)
-          scales = np.array([1.0, 1.0, sc, sc], dtype=np.float32)
-          inputs = {"X": X, "roi": np.array([], dtype=np.float32), "scales": scales}
-          attributes = {"coordinate_transformation_mode": ct_mode, **kwargs}
-          outputs = ["out"]
-          self.helper_test_single_op("Resize", inputs, attributes, outputs)
-
-  def test_resize_linear_mode(self):
-    self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 3.5, 20.0], mode="linear")
-
-  def test_resize_nearest_mode(self):
-    # excluded 3.5 because some values divide into slight numerical differences, which when rounded gives wrong results
-    self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 20.0], mode="nearest")
-
-  def test_resize_cubic_mode(self):
-    self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 3.5, 20.0], mode="cubic", exclude_outside=1)
-    self._test_resize_scales([0.01, 0.25, 0.5, 0.51, 0.6, 1.0, 1.5, 2.0, 3.5, 20.0], mode="cubic", exclude_outside=0)
-
-  def test_resize_downsample_scales_linear_align_corners(self):
-    # https://github.com/onnx/onnx/blob/main/docs/Operators.md#examples-131
-    X = np.array([[[[1, 2, 3, 4], [5, 6, 7, 8]]]], dtype=np.float32)
-    scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32)
-    inputs = {"X": X, "roi": np.array([], dtype=np.float32), "scales": scales}
-    attributes = {"mode": "linear", "coordinate_transformation_mode": "align_corners"}
-    outputs = ["out"]
-    self.helper_test_single_op("Resize", inputs, attributes, outputs)
-
-  def test_resize_downsample_scales_cubic_align_corners(self):
-    # https://github.com/onnx/onnx/blob/main/docs/Operators.md#examples-131
-    X = np.array([[[[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]]], dtype=np.float32)
-    scales = np.array([1.0, 1.0, 0.8, 0.8], dtype=np.float32)
-    inputs = {"X": X, "roi": np.array([], dtype=np.float32), "scales": scales}
-    attributes = {"mode": "cubic", "coordinate_transformation_mode": "align_corners"}
-    outputs = ["out"]
-    self.helper_test_single_op("Resize", inputs, attributes, outputs)
-
   def test_maxunpool_export_with_output_shape(self):
     # https://github.com/onnx/onnx/blob/main/docs/Operators.md#examples-91
     xT = np.array([[[[5, 6], [7, 8]]]], dtype=np.float32)
@@ -294,6 +251,7 @@ class TestTrainingOnnxOps(TestOnnxOps):
         outputs = ["X_out", "V_out"]
         self._validate_training("Momentum", onnx_fxn, inputs, attributes, outputs)
 
+  @unittest.expectedFailure  # TODO: regression from removing StrEnum in Domain
   def test_adam_t_greater_than_zero(self):
     from onnx.backend.test.case.node.adam import apply_adam
     for t in [1, 3, 100]:

test/external/external_test_onnx_runner.py

@@ -3,7 +3,8 @@ import numpy as np
 from tinygrad import dtypes, Tensor
 from tinygrad.uop.ops import Ops
 from tinygrad.device import is_dtype_supported
-from tinygrad.frontend.onnx import OnnxRunner, OnnxDataType
+from extra.onnx import data_types
+from tinygrad.frontend.onnx import OnnxRunner
 from hypothesis import given, strategies as st
 
 # copied from test_const_folding.py
@@ -85,8 +86,8 @@ class TestOnnxRunner(unittest.TestCase):
       output = runner({'inp': Tensor([1])})['output']
       np.testing.assert_equal(output.numpy(), weights + 1)
 
-all_dtypes = list(OnnxDataType)
+all_dtypes = list(data_types.keys())
-device_supported_dtypes = {odt for odt in OnnxDataType if is_dtype_supported(odt.to_dtype())}
+device_supported_dtypes = {odt for odt, dtype in data_types.items() if is_dtype_supported(dtype)}
 
 class TestOnnxRunnerDtypes(unittest.TestCase):
   """
@@ -94,7 +95,7 @@ class TestOnnxRunnerDtypes(unittest.TestCase):
   External tensors (inputs) preserve their original dtype - user must ensure compatibility with device.
   """
   def _get_expected_dtype(self, onnx_dtype: int, is_input: bool):
-    true_dtype = OnnxDataType(onnx_dtype).to_dtype()
+    true_dtype = data_types[onnx_dtype]
     # inputs always preserve their true dtype.
     if is_input:
       return true_dtype

test/external/external_test_train_gpt2.py

@@ -2,11 +2,11 @@
 import unittest
 
 from tinygrad.uop.ops import UOp, Ops
-from tinygrad.codegen.opt.search import Opt, OptOps
+from tinygrad.opt.search import Opt, OptOps
 from tinygrad.dtype import dtypes
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 
 from test.external.fuzz_linearizer import run_linearizer
 

test/external/fuzz_fast_idiv.py

@@ -1,26 +1,33 @@
 import random
-import z3
+from z3 import Int, Solver, sat
-from tinygrad import dtypes
+from tinygrad import dtypes, Device
-from tinygrad.uop.spec import z3_renderer, z3_cdiv
+from tinygrad.uop.ops import UOp, Ops, UPat, graph_rewrite, PatternMatcher
-from tinygrad.uop.ops import UOp, graph_rewrite
+from tinygrad.codegen.optional import fast_idiv
-from tinygrad.uop.decompositions import fast_idiv
 random.seed(42)
 
-powers_of_two = [2**i for i in range(64)]
+z3_renderer = PatternMatcher([
+  (UPat((Ops.DEFINE_VAR, Ops.SPECIAL), name="x"), lambda x: UOp(Ops.NOOP, arg=x.arg[0])),
+  # Because fast_idiv only works for non-negative integers we can emulate machine arithmetic with modulo operations.
+  (UPat(Ops.SHR, src=UPat(Ops.NOOP), name="x"), lambda x: UOp(Ops.NOOP, arg=f"(({x.src[0].arg}/(2**{x.src[1].arg}))%{dtypes.max(x.dtype)+1})")),
+  (UPat(Ops.MUL, src=UPat(Ops.NOOP), name="x"), lambda x: UOp(Ops.NOOP, arg=f"(({x.src[0].arg}*{x.src[1].arg})%{dtypes.max(x.dtype)+1})")),
+  (UPat((Ops.CONST, Ops.VCONST), name="x"), lambda x: UOp(Ops.NOOP, arg=str(x.arg))),
+  (UPat(Ops.CAST, src=UPat(Ops.NOOP), name="x"), lambda x:  UOp(Ops.NOOP, arg=f"{x.src[0].arg}")),
+])
+
+def render(self) -> str:
+  ret = graph_rewrite(self.simplify(), z3_renderer)
+  return ret.arg if ret.op is Ops.NOOP else str(ret)
+
 if __name__ == "__main__":
-  for i in range(10_000):
+  x = Int('x')
-    if i % 1000 == 0:
+  for _ in range(10_000):
-      print(f"Progress: {i}")
     dt = random.choice(dtypes.ints)
-    u = UOp.variable('x', random.randint(dt.min, 0), random.randint(1, dt.max), dtype=dt)
+    u = UOp(Ops.DEFINE_VAR, dt, arg=('x', 0, random.randint(1, dtypes.max(dt))), src=())
     d = random.randint(1, max(1, u.arg[2]))
-    if d in powers_of_two: continue
+
-    expr = fast_idiv(None, u, d)
+    expr = fast_idiv(Device[Device.DEFAULT].renderer, u, d)
     if expr is None: continue
-
+    solver = Solver()
-    solver = z3.Solver()
+    solver.add(x>=u.arg[1], x<=u.arg[2])
-    z3_sink = graph_rewrite(expr.sink(u), z3_renderer, ctx=(solver, {}))
+    if solver.check(eval(render(expr)) != x/d) == sat:
-    z3_expr, x = z3_sink.src[0].arg, z3_sink.src[1].arg
+      assert False, f"Failed: {render(expr)} != x//{d} at x={solver.model()[x]}\nx={u}\nd={d}"
-
-    if solver.check(z3_expr != z3_cdiv(x, d)) == z3.sat:
-      assert False, f"Failed: {expr.render()} != x//{d} at x={solver.model()}\nx={u}\nd={d}\n{z3_expr=}\n{x/d=}"

test/external/fuzz_linearizer.py

@@ -21,9 +21,9 @@ if os.getenv("VALIDATE_HCQ", 0) != 0:
 
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps
-from tinygrad.codegen.opt.search import get_kernel_actions, bufs_from_lin
+from tinygrad.opt.search import get_kernel_actions, bufs_from_lin
 from tinygrad.engine.realize import CompiledRunner
 from tinygrad.helpers import getenv, from_mv, prod, colored, Context, DEBUG, Timing
 from tinygrad.uop.ops import UOp, Ops

test/external/process_replay/process_replay.py

@@ -1,7 +1,6 @@
 #!/usr/bin/env python3
 # compare kernels created by HEAD against master
 import os, multiprocessing, logging, pickle, sqlite3, difflib, warnings, itertools, functools, base64, codecs
-from dataclasses import replace
 from typing import Callable, Any
 
 ASSERT_DIFF = int((flag:="[pr]") in os.getenv("COMMIT_MESSAGE", flag) or flag in os.getenv("PR_TITLE", flag))
@@ -12,9 +11,7 @@ try:
   from tinygrad.renderer import Renderer, ProgramSpec
   from tinygrad.engine.realize import get_program
   from tinygrad.uop.ops import UOp, Ops, KernelInfo
-  from tinygrad.codegen.opt.kernel import Opt
   from tinygrad.helpers import VERSION, Context, ContextVar, colored, db_connection, getenv, tqdm
-  from tinygrad.device import Device
 except ImportError as e:
   print(repr(e))
   exit(int(ASSERT_DIFF))
@@ -50,13 +47,9 @@ def replay_kernelize(ret:dict[UOp, UOp], big_sink:UOp) -> tuple[str, str, tuple[
     return "\n".join([f"{len(asts)} kernels", *asts])
   return to_str(new_sink), to_str(ret[big_sink]), (big_sink,)
 
-def replay_get_program(p:ProgramSpec, ast:UOp, renderer:Renderer|None=None, opts:list[Opt]|None=None) -> tuple[str, str, tuple[Any, ...]]:
+def replay_get_program(p:ProgramSpec, ast:UOp, renderer:Renderer) -> tuple[str, str, tuple[Any, ...]]:
-  # NOTE: this always uses the opts_to_apply path
+  input_ast = ast.replace(arg=KernelInfo(opts_to_apply=p.applied_opts, name=p.name)) if ast.arg is None else ast
-  sink_arg = ast.arg or KernelInfo(opts_to_apply=p.applied_opts)
+  p2 = get_program(input_ast, renderer)
-  input_ast = ast.replace(arg=replace(sink_arg, name=p.name))
-  # if no renderer was provided, open the device to get it
-  if renderer is None: renderer = Device[p.device].renderer
-  p2 = get_program(input_ast, renderer=renderer)
   def to_str(ret:ProgramSpec) -> str:
     # PYTHON renderer pickles UOps, first unpickle and decode here
     if p.device.startswith("PYTHON"): return "\n".join([str(x) for x in pickle.loads(base64.b64decode(ret.src))])

test/external/speed_beam_v_hcopt.py

@@ -1,7 +1,7 @@
 from tinygrad import Device
 from tinygrad.helpers import getenv, DEBUG, BEAM
-from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
+from tinygrad.opt.search import beam_search, bufs_from_lin
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, time_linearizer
 
 if __name__ == "__main__":

test/external/speed_compare_amd_am.py

@@ -2,8 +2,8 @@ from tinygrad import Device, dtypes
 from tinygrad.helpers import getenv, colorize_float, DEBUG
 from extra.optimization.helpers import load_worlds, ast_str_to_lin
 from test.external.fuzz_linearizer import get_fuzz_rawbufs
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
-from tinygrad.codegen.opt.search import bufs_from_lin
+from tinygrad.opt.search import bufs_from_lin
 from tinygrad.engine.realize import CompiledRunner, get_program
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.runtime.ops_amd import AMDDevice

test/external/speed_compare_cuda_nv.py

@@ -2,8 +2,8 @@ from tinygrad import Device, dtypes
 from tinygrad.helpers import getenv, colorize_float
 from extra.optimization.helpers import load_worlds, ast_str_to_lin
 from test.external.fuzz_linearizer import get_fuzz_rawbufs
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
-from tinygrad.codegen.opt.search import bufs_from_lin
+from tinygrad.opt.search import bufs_from_lin
 from tinygrad.engine.realize import CompiledRunner, get_program
 from tinygrad.tensor import _to_np_dtype
 import numpy as np

test/external/speed_compare_cuda_ptx.py

@@ -1,10 +1,10 @@
 import itertools
 from tinygrad import Device
 from tinygrad.engine.realize import CompiledRunner, get_program
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from tinygrad.helpers import getenv, colorize_float
 from extra.optimization.helpers import load_worlds, ast_str_to_lin
-from tinygrad.codegen.opt.search import bufs_from_lin
+from tinygrad.opt.search import bufs_from_lin
 from tinygrad.runtime.ops_cuda import PTXCompiler, PTXRenderer, CUDACompiler
 
 if __name__ == "__main__":

test/external/verify_kernel.py

@@ -3,7 +3,7 @@ from collections import defaultdict
 from extra.optimization.helpers import kern_str_to_lin, time_linearizer
 from test.external.fuzz_linearizer import compare_linearizer
 from tinygrad.helpers import colored
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 
 # Use this with the LOGKERNS options to verify that all executed kernels are valid and evaluate to the same ground truth results
 

test/mockgpu/amd/amddriver.py

@@ -87,19 +87,16 @@ class AMDDriver(VirtDriver):
         functools.partial(TextFileDesc, text=gpu_props.format(drm_render_minor=gpu_id))),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0',
                functools.partial(DirFileDesc, child_names=[str(am.GC_HWID), str(am.SDMA0_HWID), str(am.NBIF_HWID)])),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/major', functools.partial(TextFileDesc, text='11')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/minor', functools.partial(TextFileDesc, text='0')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.GC_HWID}/0/base_addr',
                functools.partial(TextFileDesc, text='0x00001260\n0x0000A000\n0x0001C000\n0x02402C00')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/major', functools.partial(TextFileDesc, text='6')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/minor', functools.partial(TextFileDesc, text='0')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.SDMA0_HWID}/0/base_addr',
                functools.partial(TextFileDesc, text='0x00001260\n0x0000A000\n0x0001C000\n0x02402C00')),
-      VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}', functools.partial(DirFileDesc, child_names=['0'])),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/major', functools.partial(TextFileDesc, text='4')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/minor', functools.partial(TextFileDesc, text='3')),
       VirtFile(f'/sys/class/drm/renderD{gpu_id}/device/ip_discovery/die/0/{am.NBIF_HWID}/0/revision', functools.partial(TextFileDesc, text='0')),

test/mockgpu/amd/amdgpu.py

@@ -1,7 +1,7 @@
 import ctypes, time
 from test.mockgpu.gpu import VirtGPU
 from tinygrad.helpers import getbits, to_mv, init_c_struct_t
-import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4
+import tinygrad.runtime.autogen.amd_gpu as amd_gpu, tinygrad.runtime.autogen.am.pm4_nv as pm4, tinygrad.runtime.autogen.am.soc21 as soc21
 
 SDMA_MAX_COPY_SIZE = 0x400000
 
@@ -14,9 +14,6 @@ regSQ_THREAD_TRACE_BUF0_SIZE = 0x39e9 + amd_gpu.GC_BASE__INST0_SEG1
 regSQ_THREAD_TRACE_WPTR = 0x39ef + amd_gpu.GC_BASE__INST0_SEG1
 regSQ_THREAD_TRACE_STATUS = 0x39f4 + amd_gpu.GC_BASE__INST0_SEG1
 
-class SQTT_EVENTS:
-  THREAD_TRACE_FINISH = 0x00000037
-
 CACHE_FLUSH_AND_INV_TS_EVENT = 0x14
 
 WAIT_REG_MEM_FUNCTION_ALWAYS = 0
@@ -211,7 +208,7 @@ class PM4Executor(AMDQueue):
     assert n == 0
     event_dw = self._next_dword()
     match (event_dw & 0xFF): # event type
-      case SQTT_EVENTS.THREAD_TRACE_FINISH:
+      case soc21.THREAD_TRACE_FINISH:
         old_idx = self.gpu.regs.grbm_index
         for se in range(self.gpu.regs.n_se):
           self.gpu.regs.grbm_index = 0b011 << 29 | se << 16 # select se, broadcast sa and instance

test/models/test_onnx.py

@@ -9,15 +9,7 @@ except ModuleNotFoundError:
   raise unittest.SkipTest("onnx not installed, skipping onnx test")
 from tinygrad.frontend.onnx import OnnxRunner
 from tinygrad.tensor import Tensor
-from tinygrad.device import Device
+from tinygrad.helpers import CI, fetch, temp
-from tinygrad.helpers import CI, fetch, temp, Context
-
-try:
-  from extra.onnx_helpers import validate
-  from extra.huggingface_onnx.huggingface_manager import DOWNLOADS_DIR, snapshot_download_with_retry
-  HUGGINGFACE_AVAILABLE = True
-except ModuleNotFoundError:
-  HUGGINGFACE_AVAILABLE = False
 
 def run_onnx_torch(onnx_model, inputs):
   import torch
@@ -32,7 +24,6 @@ OPENPILOT_MODEL = "https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/mod
 np.random.seed(1337)
 
 class TestOnnxModel(unittest.TestCase):
-  @unittest.skip("this isn't a test, it can't fail")
   def test_benchmark_openpilot_model(self):
     onnx_model = fetch(OPENPILOT_MODEL)
     run_onnx = OnnxRunner(onnx_model)
@@ -146,36 +137,5 @@ class TestOnnxModel(unittest.TestCase):
     print(cls, _LABELS[cls])
     assert "car" in _LABELS[cls] or _LABELS[cls] == "convertible"
 
-@unittest.skipUnless(HUGGINGFACE_AVAILABLE and Device.DEFAULT == "METAL", "only run on METAL")
-class TestHuggingFaceOnnxModels(unittest.TestCase):
-  @classmethod
-  def setUpClass(cls):
-    cls._ctx = Context(MAX_BUFFER_SIZE=0)
-    cls._ctx.__enter__()
-
-  @classmethod
-  def tearDownClass(cls):
-    cls._ctx.__exit__()
-
-  def _validate(self, repo_id, model_file, custom_inputs, rtol=1e-4, atol=1e-4):
-    onnx_model_path = snapshot_download_with_retry(
-      repo_id=repo_id,
-      allow_patterns=["*.onnx", "*.onnx_data"],
-      cache_dir=str(DOWNLOADS_DIR)
-    )
-    onnx_model_path = onnx_model_path / model_file
-    file_size = onnx_model_path.stat().st_size
-    print(f"Validating model: {repo_id}/{model_file} ({file_size/1e6:.2f}M)")
-    validate(onnx_model_path, custom_inputs, rtol=rtol, atol=atol)
-
-  def test_xlm_roberta_large(self):
-    repo_id = "FacebookAI/xlm-roberta-large"
-    model_file = "onnx/model.onnx"
-    custom_inputs = {
-      "input_ids": np.random.randint(0, 250002, (1, 11), dtype=np.int64),
-      "attention_mask": np.ones((1, 11), dtype=np.int64),
-    }
-    self._validate(repo_id, model_file, custom_inputs)
-
 if __name__ == "__main__":
   unittest.main()

test/models/test_whisper.py

@@ -16,7 +16,7 @@ TRANSCRIPTION_2 = "a slightly longer audio file so that we can test batch transc
 TEST_FILE_3_URL = 'https://homepage.ntu.edu.tw/~karchung/miniconversations/mc45.mp3'
 TRANSCRIPTION_3 = "Just lie back and relax. Is the level of pressure about right? Yes, it's fine, and I'd like conditioner please. Sure. I'm going to start the second lathering now. Would you like some Q-tips? How'd you like it cut? I'd like my bangs and the back trimmed, and I'd like the rest thinned out a bit and layered. Where would you like the part? On the left, right about here. Here, have a look. What do you think? It's fine. Here's a thousand anti-dollars. It's 30-ant extra for the rants. Here's your change and receipt. Thank you, and please come again. So how do you like it? It could have been worse, but you'll notice that I didn't ask her for her card. Hmm, yeah. Maybe you can try that place over there next time."   # noqa: E501
 
-@unittest.skipIf(Device.DEFAULT in ["CPU", "LLVM"], "slow")
+@unittest.skipIf(CI and Device.DEFAULT in ["CPU"], "slow")
 @unittest.skipUnless(is_dtype_supported(dtypes.float16), "need float16 support")
 class TestWhisper(unittest.TestCase):
   @classmethod

test/test_arange.py

@@ -3,11 +3,10 @@ import numpy as np
 from tinygrad import Tensor, GlobalCounters, dtypes, nn, Device, Variable
 from tinygrad.helpers import CI, Context, getenv
 from tinygrad.engine.realize import run_schedule
-from tinygrad.codegen.opt.kernel import Opt, OptOps, Kernel, KernelOptError
+from tinygrad.opt.kernel import Opt, OptOps, Kernel, KernelOptError
 from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
-from tinygrad.codegen.opt.search import get_kernel_actions
+from tinygrad.opt.search import get_kernel_actions
 from tinygrad.uop.ops import Ops
-from tinygrad.codegen import apply_rewrites, rewrites_for_views
 
 class TestArange(unittest.TestCase):
   def _get_flops(self, N, opts=None):
@@ -15,7 +14,10 @@ class TestArange(unittest.TestCase):
     tt = Tensor.arange(N)
     sched = tt.schedule()
     self.assertEqual(len(sched), 1)
-    p = get_program(sched[-1].ast, opts=opts)
+    k = Kernel(sched[-1].ast)
+    if opts is not None:
+      for o in opts: k.apply_opt(o)
+    p = get_program(k.get_optimized_ast(), k.opts)
     print(p.name)
     #print(p.src)
     ExecItem(CompiledRunner(p), [tt.uop.buffer]).run()
@@ -50,11 +52,11 @@ class TestArange(unittest.TestCase):
     def test_complexity_w_local_and_padto(self): return self.test_complexity([Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.PADTO, axis=1, arg=32)])
 
   def test_all_opts(self, opts=None, exclude=None):
-    k = Kernel(apply_rewrites(Tensor.arange(256).schedule()[-1].ast, rewrites_for_views))
+    k = Kernel(Tensor.arange(256).schedule()[-1].ast)
     if opts is not None:
       for o in opts: k.apply_opt(o)
     all_opts_256 = [kk.applied_opts for kk in get_kernel_actions(k, include_0=False).values()]
-    k = Kernel(apply_rewrites(Tensor.arange(2560).schedule()[-1].ast, rewrites_for_views))
+    k = Kernel(Tensor.arange(2560).schedule()[-1].ast)
     if opts is not None:
       for o in opts: k.apply_opt(o)
     all_opts_2560 = [kk.applied_opts for kk in get_kernel_actions(k, include_0=False).values()]

test/test_compile_failures.py

@@ -1,10 +1,7 @@
-import unittest, io
+import unittest
-from contextlib import redirect_stdout
 from tinygrad import Tensor, dtypes, Device
-from tinygrad.helpers import OSX
 from tinygrad.engine.realize import lower_schedule
 from tinygrad.device import is_dtype_supported
-from tinygrad.engine.realize import get_program
 
 class TestCompileFailures(unittest.TestCase):
   def compile(self, out:Tensor):
@@ -17,17 +14,5 @@ class TestCompileFailures(unittest.TestCase):
   def test_add_max_uchar(self):
     self.compile((Tensor.empty(1024, dtype='uint8') + Tensor.empty(1024, dtype='uint8')).max())
 
-class TestDisassembly(unittest.TestCase):
-  # TODO: fails on llvm. llvm.LLVMGetHostCPUName() returns "generic"
-  @unittest.skipUnless(Device.DEFAULT in ("CPU",) and OSX, "m series cpus support fp16 arithmetic")
-  def test_float16_alu(self):
-    c = Tensor([1], dtype=dtypes.float16) + Tensor([1], dtype=dtypes.float16)
-    s = c.schedule()[-1]
-    p = get_program(s.ast, Device[Device.DEFAULT].renderer)
-    lib = Device[Device.DEFAULT].compiler.compile(p.src)
-    out = io.StringIO()
-    with redirect_stdout(out): Device[Device.DEFAULT].compiler.disassemble(lib)
-    assert "fcvt" not in out.getvalue()
-
 if __name__ == '__main__':
   unittest.main()

test/test_const_folding.py

@@ -139,16 +139,18 @@ class TestBitcastConstFolding(unittest.TestCase):
 class TestIndexingConstFolding(unittest.TestCase):
   def test_scalar_index(self):
     t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
-    _check_ast_count(1, t[:,:,Tensor(1),:])
+    # TODO: fold these
-    _check_ast_count(1, t[:,:,Tensor(1)+2,:])
+    _check_ast_count(2, t[:,:,Tensor(1),:])
-    _check_ast_count(1, t[:,:,Tensor(1),Tensor(0)])
+    _check_ast_count(2, t[:,:,Tensor(1)+2,:])
+    _check_ast_count(2, t[:,:,Tensor(1),Tensor(0)])
 
+  @unittest.expectedFailure
   def test_const_tensor_index(self):
-    # TODO: these can be 0, implement const tensor folded indexing
+    # TODO: implement const tensor folded indexing
     t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
-    _check_ast_count(1, t[:,:,Tensor.ones(2,1,dtype=dtypes.int),:])
+    _check_ast_count(0, t[:,:,Tensor.ones(2,1),:])
-    _check_ast_count(1, t[:,:,Tensor.ones(1,2,dtype=dtypes.int)+2,:])
+    _check_ast_count(0, t[:,:,Tensor.ones(1,2)+2,:])
-    _check_ast_count(1, t[:,:,Tensor.ones(1,1,dtype=dtypes.int),Tensor.zeros(2,1,2,dtype=dtypes.int)])
+    _check_ast_count(0, t[:,:,Tensor.ones(1,1),Tensor.zeros(2,1,2)])
 
 class TestMovedConstFolding(unittest.TestCase):
   def test_add_shrunk_zero(self):
@@ -289,12 +291,17 @@ class TestMultiConstFolding(unittest.TestCase):
     np.testing.assert_equal((t + zero).numpy(), np.arange(16))
     np.testing.assert_equal((t * zero).numpy(), [0] * 16)
     np.testing.assert_equal((t * one).numpy(), np.arange(16))
+
+  def test_multi_todo_pow(self):
+    ds = tuple(f"{Device.DEFAULT}:{i}" for i in range(4))
+    t = Tensor.arange(16).float().to(ds).realize()
+    zero = Tensor.zeros(16).to(ds).realize()
+    one = Tensor.ones(16).to(ds).realize()
+
+    # TODO: fix pow folding
     _check_ast_count(0, t ** zero)
     _check_ast_count(0, t ** one)
     _check_ast_count(0, one ** t)
-    np.testing.assert_equal((t ** zero).numpy(), [1] * 16)
-    np.testing.assert_equal((t ** one).numpy(), np.arange(16))
-    np.testing.assert_equal((one ** t).numpy(), [1] * 16)
 
 class TestTautologicalCompare(unittest.TestCase):
   # without const folding, these would have triggered -Wtautological-compare in clang

test/test_copy_speed.py

@@ -3,7 +3,7 @@ from tinygrad import Tensor, Device, TinyJit
 from tinygrad.helpers import Timing, CI, OSX
 import multiprocessing.shared_memory as shared_memory
 
-N = 256
+N = 256 if CI else 4096
 class TestCopySpeed(unittest.TestCase):
   @classmethod
   def setUpClass(cls): Device[Device.DEFAULT].synchronize()

test/test_edgecases.py

@@ -62,6 +62,7 @@ class TestNaNEdgeCases(unittest.TestCase):
 class TestEmptyTensorEdgeCases(unittest.TestCase):
   # we don't need more of these
 
+  @unittest.expectedFailure
   def test_sort_empty(self):
     # Sorting an empty tensor works in PyTorch and should return empty
     # values and indices. tinygrad raises an error instead.
@@ -218,6 +219,7 @@ class TestAssignIssues(unittest.TestCase):
     t.shrink(((1, 3), (1, 3))).assign(Tensor.ones(2, 2))
     np.testing.assert_allclose(t.numpy(), torch_tensor.numpy())
 
+  @unittest.expectedFailure
   def test_assign_broadcast(self):
     # broadcasting during assign should behave like PyTorch
     torch_tensor = torch.zeros(3, 5)
@@ -256,11 +258,12 @@ class TestEdgeCases(unittest.TestCase):
     out = Tensor(arr).pad((1, -1, 1, -1), mode='circular')
     np.testing.assert_equal(out.numpy(), torch_out.numpy())
 
+  @unittest.expectedFailure
   def test_arange_float_step(self):
     # float steps should match PyTorch exactly
     torch_out = torch.arange(0, 2, 0.3).numpy()
     out = Tensor.arange(0, 2, 0.3).numpy()
-    np.testing.assert_allclose(out, torch_out, atol=1e-7)
+    np.testing.assert_allclose(out, torch_out)
 
   @unittest.skip("this is flaky")
   @unittest.expectedFailure

test/test_hcq.py

@@ -6,7 +6,7 @@ from tinygrad.runtime.support.hcq import HCQCompiled, HCQBuffer
 from tinygrad.runtime.autogen import libc
 from tinygrad.runtime.support.system import PCIIfaceBase
 from tinygrad.engine.realize import get_runner, CompiledRunner, get_program
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Kernel, Opt, OptOps
 from tinygrad import Variable
 
 MOCKGPU = getenv("MOCKGPU")
@@ -163,8 +163,10 @@ class TestHCQ(unittest.TestCase):
     a = Tensor.randint((3, 3, 3), dtype=dtypes.int, device=Device.DEFAULT).realize()
     b = a + 1
     si = b.schedule()[-1]
+    k = Kernel(si.ast, opts=TestHCQ.d0.renderer)
+    for i in range(3): k.apply_opt(Opt(op=OptOps.LOCAL, axis=0, arg=3))
 
-    runner = CompiledRunner(get_program(si.ast, TestHCQ.d0.renderer, opts=[Opt(op=OptOps.LOCAL, axis=0, arg=3) for _ in range(3)]))
+    runner = CompiledRunner(get_program(k.get_optimized_ast(), k.opts))
 
     zb = Buffer(Device.DEFAULT, 3 * 3 * 3, dtypes.int, options=BufferSpec(cpu_access=True, nolru=True)).ensure_allocated()
     zt = Buffer(Device.DEFAULT, 3 * 3 * 3, dtypes.int, options=BufferSpec(cpu_access=True, nolru=True)).ensure_allocated()
@@ -336,7 +338,7 @@ class TestHCQ(unittest.TestCase):
     et = float(sig_en.timestamp - sig_st.timestamp)
 
     print(f"exec kernel time: {et:.2f} us")
-    assert 0.1 <= et <= (100000 if MOCKGPU or Device.DEFAULT in {"CPU", "LLVM"} else 100)
+    assert 0.1 <= et <= (15000 if MOCKGPU or Device.DEFAULT in {"CPU", "LLVM"} else 100)
 
   def test_speed_copy_bandwidth(self):
     if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")

test/test_jit.py

@@ -718,6 +718,8 @@ class TestJitGraphSplit(unittest.TestCase):
   def ji_xfer(self): return {"type": "xfer"}
 
   def test_jit_split_simple(self):
+    if Device.DEFAULT == "REMOTE": raise unittest.SkipTest("REMOTE gpu is broken")
+
     @TinyJit
     def f(inp):
       op0 = self.compute(Device.DEFAULT, inp)
@@ -790,9 +792,9 @@ class TestJitGraphSplit(unittest.TestCase):
       multigraph=[self.ji_graph(5)],
       hcqgraph=[self.ji_graph(5)])
 
+  @unittest.skip("flaky")
   def test_jit_multidev_xfer(self):
     if Device.DEFAULT in {"CPU", "LLVM"}: raise unittest.SkipTest("CPU/LLVM is not a valid default device for this test (zero-copies)")
-    if Device.DEFAULT == "METAL" or REAL_DEV == "METAL": raise unittest.SkipTest("Metal is flaky, with multidevice (same as metal llama 4gpu?)")
 
     try: Device[f"{Device.DEFAULT}:1"]
     except Exception: raise unittest.SkipTest("no multidevice")
@@ -817,6 +819,7 @@ class TestJitGraphSplit(unittest.TestCase):
   @unittest.skipIf(getenv("MOCKGPU"), "MockGPU does not support parallel copies")
   def test_jit_multidev_copy(self):
     if Device.DEFAULT in {"CPU", "LLVM"}: raise unittest.SkipTest("CPU/LLVM is not a valid default device for this test (zero-copies)")
+    if Device.DEFAULT == "REMOTE": raise unittest.SkipTest("REMOTE gpu is broken")
 
     @TinyJit
     def f(inp):

test/test_linalg.py

@@ -62,15 +62,5 @@ class TestLinAlg(unittest.TestCase):
       orthogonality_helper(Q)
       reconstruction_helper([Q,R],a)
 
-  def test_newton_schulz(self):
-    coefficients = [(2, -1.5, 0.5), (2.0, -1.4, 0.2, 0.2)]#these params map to the sign function
-    sizes = [(2,2), (3,2), (2,3), (2,2,2)]
-    for coefs in coefficients:
-      for size in sizes:
-        a = Tensor.randn(size)
-        b = Tensor.newton_schulz(a, steps=20, params=coefs, eps=0.0)
-        # ns(A) = U @ Vt -> (U @ Vt) @ (U @ Vt)t = I
-        orthogonality_helper(b if size[-1] > size[-2] else b.transpose(-2, -1), tolerance=1e-1)
-
 if __name__ == "__main__":
   unittest.main()

test/test_linearizer.py

@@ -2,7 +2,7 @@ import numpy as np
 import unittest
 from dataclasses import replace
 
-from tinygrad.codegen.opt.kernel import Opt, OptOps, KernelOptError, Kernel, AxisType
+from tinygrad.opt.kernel import Opt, OptOps, KernelOptError, Kernel, AxisType
 from tinygrad.codegen.gpudims import get_grouped_dims
 from tinygrad.uop.ops import UOp, Ops, GroupOp, KernelInfo
 from tinygrad.device import Device, Buffer, is_dtype_supported
@@ -10,12 +10,9 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import run_schedule, lower_schedule, CompiledRunner, get_program
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from tinygrad.helpers import prod, Context, getenv, CI, flatten, dedup, AMX, AMD_LLVM
 from tinygrad.dtype import DType, dtypes, AddrSpace
-from tinygrad.codegen import apply_rewrites, rewrites_for_views
-
-def push_views(ast): return apply_rewrites(ast, rewrites_for_views)
 
 def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
   if isinstance(r, Tensor): r = [r]
@@ -25,7 +22,7 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
   # now all input buffers in s[-1] should be realized
   # create fresh buffers for the outputs
   bufs = [Buffer((x).device, x.size, x.dtype).allocate() if i < len(s[-1].ast.src) else x for i,x in enumerate(s[-1].bufs)]
-  return push_views(s[-1].ast), bufs
+  return s[-1].ast, bufs
 
 def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0, use_tensor_cores:int=1):
   a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
@@ -124,7 +121,7 @@ class TestLinearizer(unittest.TestCase):
     with Context(FUSE_ARANGE=1):
       sink = dataset[idxs].contiguous().kernelize().uop.base.src[1].arg.ast
     real_index = dataset.numpy()[idxs.numpy()].reshape(4, 256, 1, 1)
-    helper_linearizer_ast(push_views(sink), [dataset, idxs], wanna_output=[real_index])
+    helper_linearizer_ast(sink, [dataset, idxs], wanna_output=[real_index])
 
   def test_two_nested_range(self):
     a = Tensor.randn(2, ).realize()
@@ -217,7 +214,9 @@ class TestLinearizer(unittest.TestCase):
     # these are of size 3 to avoid float4 coalesce
     r = a[:-1] + a[1:]
 
-    uops = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=0)]).uops
+    k = Kernel(r.schedule()[-1].ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=0))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
     num_loads = len([uop for uop in uops if uop.op is Ops.LOAD])
     assert num_loads <= 4, "more load uops than needed"
     assert num_loads >= 4, "unexpected number of uops, maybe this test needs updating?"
@@ -228,7 +227,9 @@ class TestLinearizer(unittest.TestCase):
     a, b = Tensor.randn(1).realize(), Tensor.randn(1).realize()
     r = a.expand([2]) + b.expand([2])
 
-    uops = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=0)]).uops
+    k = Kernel(r.schedule()[-1].ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=0))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
     num_ops = len([uop for uop in uops if uop.op in GroupOp.ALU])
     assert num_ops <= 1, "more alu uops than needed"
 
@@ -237,7 +238,10 @@ class TestLinearizer(unittest.TestCase):
     x, w = Tensor.randn((1,1,3)).realize(), Tensor.randn((1,1,2)).realize()
     r = Tensor.conv2d(x,w,padding=1).relu()
 
-    uops = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=0)]).uops
+    k = Kernel(r.schedule()[-1].ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=0))
+    k.apply_opt(Opt(op=OptOps.UNROLL, axis=0, arg=0))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
     accs = [u for u in uops if u.op is Ops.DEFINE_REG]
     stores = [u for u in uops if u.op is Ops.STORE]
     assert len(accs) == 0  # it's removed now
@@ -249,7 +253,9 @@ class TestLinearizer(unittest.TestCase):
   @unittest.skipUnless(Device.DEFAULT == "CPU", "test only for CPU")
   def test_upcast_with_locals_cpu(self):
     out = Tensor.ones(64,64).contiguous() @ Tensor.ones(64,64).contiguous()
-    prg = get_program(out.schedule()[-1].ast, opts=[Opt(OptOps.LOCAL, axis=0, arg=4)]).uops
+    k = Kernel(out.schedule()[-1].ast)
+    k.apply_opt(Opt(OptOps.LOCAL, axis=0, arg=4))
+    prg = get_program(k.get_optimized_ast(), k.opts)
     self.assertEqual(len(prg.src.split("for")), 5)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
@@ -259,8 +265,10 @@ class TestLinearizer(unittest.TestCase):
   def test_upcast_with_locals(self):
     x, y = Tensor.rand(1,128), Tensor.rand(128, 128)
     r = (x@y).relu()
+    realized_ast = r.schedule()[-1].ast
     opts_to_apply = [Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.LOCAL, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=4)]
-    program = get_program(r.schedule()[-1].ast, opts=opts_to_apply)
+    realized_ast = realized_ast.replace(arg=KernelInfo(opts_to_apply=tuple(opts_to_apply)))
+    program = get_program(realized_ast, Device[Device.DEFAULT].renderer)
 
     stores = [u for u in program.uops if u.op is Ops.STORE and u.src[0].dtype.addrspace != AddrSpace.REG]
 
@@ -274,7 +282,10 @@ class TestLinearizer(unittest.TestCase):
   def test_zero_fold(self):
     a, b = Tensor.randn(1).realize(), Tensor.randn(1).realize()
     r = Tensor.stack(a, b)
-    uops = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=0)]).uops
+
+    k = Kernel(r.schedule()[-1].ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=0))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
     num_ops = len([uop for uop in uops if uop.op in GroupOp.ALU])
     assert num_ops == 0, "more alu uops than needed"
 
@@ -284,14 +295,16 @@ class TestLinearizer(unittest.TestCase):
       if is_dtype_supported(tensor_dtype) and is_dtype_supported(acc_dtype):
         a = Tensor([1, 2, 3], dtype=tensor_dtype).sum()
         realized_ast = a.schedule()[-1].ast
-        program = get_program(realized_ast, opts=[])
+        realized_ast = realized_ast.replace(arg=KernelInfo(opts_to_apply=tuple()))
+        program = get_program(realized_ast, Device[Device.DEFAULT].renderer)
         local = [uop for uop in program.uops if uop.op is Ops.DEFINE_REG]
         assert local[0].dtype.base == acc_dtype
 
   def test_arg_acc_dtype(self):
     def helper_arg_acc_dtype(c: Tensor, expected_dtype:DType):
       realized_ast = c.schedule()[-1].ast
-      program = get_program(realized_ast, opts=[])
+      realized_ast = realized_ast.replace(arg=KernelInfo(opts_to_apply=tuple()))
+      program = get_program(realized_ast, Device[Device.DEFAULT].renderer)
       local = [uop for uop in program.uops if uop.op is Ops.DEFINE_REG]
       self.assertEqual(local[0].dtype.base, expected_dtype)
 
@@ -328,7 +341,7 @@ class TestLinearizer(unittest.TestCase):
       a, b = Tensor.rand(m, k, dtype=tc.dtype_in), Tensor.rand(k, n, dtype=tc.dtype_in)
       r = a.matmul(b, dtype=tc.dtype_out)
       sched = r.schedule()
-      realized_ast = push_views(sched[-1].ast)
+      realized_ast = sched[-1].ast
       kernel = Kernel(realized_ast)
       kernel.apply_tensor_cores(1, axis=0, tc_select=-1, tc_opt=2)
       prg = get_program(kernel.get_optimized_ast(), kernel.opts)
@@ -410,7 +423,7 @@ class TestLinearizer(unittest.TestCase):
         np.testing.assert_allclose(result, golden_result, atol=0.1, rtol=0.2)
 
       # check that get_kernel_actions produces all 9 options
-      from tinygrad.codegen.opt.search import get_kernel_actions
+      from tinygrad.opt.search import get_kernel_actions
       tc_actions = [k for i, k in get_kernel_actions(Kernel(realized_ast), False).items() if k.applied_opts[0].op == OptOps.TC]
 
       available_tc = len([x for x in Device[Device.DEFAULT].renderer.tensor_cores if x.dtype_in == tc.dtype_in and x.dtype_out == tc.dtype_out])
@@ -749,7 +762,11 @@ class TestFloat4(unittest.TestCase):
     c = a + b
 
     s = c.schedule()[0]
-    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=2)]).uops
+    k = Kernel(s.ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=4))
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=2))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
+
     assert TestFloat4.count_float4(uops) == (4, 2)
 
   @unittest.skipUnless(Device.DEFAULT in {"CPU", "LLVM"} and AMX, "Only CPU with AMX upcasts float up to size 16")
@@ -760,7 +777,10 @@ class TestFloat4(unittest.TestCase):
       c = a + b
 
       s = c.schedule()[0]
-      return get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=shift)]).uops
+      k = Kernel(s.ast)
+      k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=4))
+      k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=shift))
+      return get_program(k.get_optimized_ast(), k.opts).uops
 
     sizes = [12, 8, 16]
     shifts = [3, 2, 4]
@@ -790,7 +810,10 @@ class TestFloat4(unittest.TestCase):
     c = a + b
 
     s = c.schedule()[0]
-    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=2)]).uops
+    k = Kernel(s.ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=1, arg=4))
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=1, arg=2))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
 
     assert TestFloat4.count_float4(uops) == (0, 2)
 
@@ -802,7 +825,10 @@ class TestFloat4(unittest.TestCase):
       c = a + b
 
       s = c.schedule()[0]
-      return get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=shift)]).uops
+      k = Kernel(s.ast)
+      k.shift_to(1, 4, AxisType.UPCAST)  # manual trigger float4 dim
+      k.shift_to(1, shift, AxisType.UPCAST, insert_at=k.shape_len-1)
+      return get_program(k.get_optimized_ast(), k.opts).uops
 
     sizes = [13, 9, 17]
     shifts = [3, 2, 4]
@@ -820,7 +846,9 @@ class TestFloat4(unittest.TestCase):
     # float4 should be emitted (the reduce axis of size 4 is the float4 axis here)
 
     s = c.schedule()[0]
-    uops = get_program(s.ast, opts=[Opt(op=OptOps.UNROLL, axis=0, arg=4)]).uops
+    k = Kernel(s.ast)
+    k.apply_opt(Opt(op=OptOps.UNROLL, axis=0, arg=4))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
 
     assert TestFloat4.count_float4(uops) == (0, 0)
 
@@ -834,7 +862,10 @@ class TestFloat4(unittest.TestCase):
     # UPDATE: now we do this fusion
 
     s = c.schedule()[0]
-    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=0)]).uops
+    k = Kernel(s.ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=0))
+    k.apply_opt(Opt(op=OptOps.UNROLL, axis=0, arg=0))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
 
     assert TestFloat4.count_float4(uops) in {(0,1), (1,1)}
 
@@ -847,7 +878,9 @@ class TestFloat4(unittest.TestCase):
     # since the top axis is not contiguous.
 
     s = c.schedule()[0]
-    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4)]).uops
+    k = Kernel(s.ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=4))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
 
     assert TestFloat4.count_float4(uops) == (0, 1)
 
@@ -859,7 +892,9 @@ class TestFloat4(unittest.TestCase):
     # should float4 b but not a
 
     s = c.schedule()[0]
-    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4)]).uops
+    k = Kernel(s.ast)
+    k.apply_opt(Opt(op=OptOps.UPCAST, axis=0, arg=4))
+    uops = get_program(k.get_optimized_ast(), k.opts).uops
 
     assert TestFloat4.count_float4(uops) == (1, 1)
 
@@ -948,7 +983,7 @@ class TestHandCodedOpts(unittest.TestCase):
     layer_2 = Tensor.cat(layer_1.unsqueeze(0), Tensor.empty(6, 20))
 
     s = layer_2.schedule()[-1]
-    k = Kernel(push_views(s.ast))
+    k = Kernel(s.ast)
     k.apply_opts(hand_coded_optimizations(k))
     assert len(k.bufs) == 6  # make sure all ops are done in one kernel
     # masked upcast should upcast masked axis of size 7
@@ -961,7 +996,7 @@ class TestHandCodedOpts(unittest.TestCase):
     monster = Tensor.stack(*[Tensor.stack(*[Tensor.empty(16) for _ in range(6)]) for _ in range(6)])
 
     s = monster.schedule()[-1]
-    k = Kernel(push_views(s.ast))
+    k = Kernel(s.ast)
     k.apply_opts(hand_coded_optimizations(k))
     assert len(k.bufs) == 37  # make sure all ops are done in one kernel
     # should upcast the two Tensor.stacks
@@ -977,7 +1012,7 @@ class TestHandCodedOpts(unittest.TestCase):
       wino_schedule = out.schedule()
       # collect upcasts of tile transform kernels
       for i, si in enumerate(wino_schedule):
-        k = Kernel(push_views(si.ast))
+        k = Kernel(si.ast)
         k.apply_opts(hand_coded_optimizations(k))
         if k.reduceop is not None: continue  # not a tile transform kernel (there is a gemm reduce kernel)
         if len(k.bufs) < 22: continue  # not a tile transform kernel (there's a permute kernel at the end)
@@ -989,7 +1024,7 @@ class TestHandCodedOpts(unittest.TestCase):
 
       backward_schedule = Tensor.schedule(x.grad, w.grad)
       for si in backward_schedule:
-        k = Kernel(push_views(si.ast))
+        k = Kernel(si.ast)
         k.apply_opts(hand_coded_optimizations(k))
         if len(k.bufs) < 20: continue  # not a tile transform kernel
         # heuristic number to make sure that at least some upcasts but not too many upcasts are being done

test/test_linearizer_dumb.py

@@ -8,8 +8,8 @@ from tinygrad.device import is_dtype_supported
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.helpers import getenv
 from tinygrad.shape.shapetracker import ShapeTracker, View
-from tinygrad.codegen.opt.search import Opt, OptOps
+from tinygrad.opt.search import Opt, OptOps
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.engine.realize import get_program
 
 class TestLinearizerDumb(unittest.TestCase):

test/test_linearizer_overflows.py

@@ -1,8 +1,8 @@
 # ruff: noqa: E501
 import unittest
 from tinygrad import dtypes
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import Opt, OptOps, bufs_from_lin
+from tinygrad.opt.search import Opt, OptOps, bufs_from_lin
 from extra.optimization.helpers import time_linearizer
 
 # stuff needed to unpack a kernel

test/test_multitensor.py

@@ -373,11 +373,11 @@ class TestMultiTensor(unittest.TestCase):
     np.testing.assert_allclose(y.numpy(), y_shard.numpy(), atol=1e-6, rtol=1e-6)
 
   # NOTE: this is failing on LLVM CI, no idea why. Works locally.
-  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU", "AMD"), "slow, and flaky on LLVM/CPU")
+  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU"), "slow, and flaky on LLVM/CPU")
   def test_data_parallel_resnet(self):
     from extra.models.resnet import ResNet18
 
-    fake_image = Tensor.rand((2, 3, 224//16, 224//16))
+    fake_image = Tensor.rand((2, 3, 224//8, 224//8))
     fake_image_sharded = fake_image.shard(devices_2, axis=0)
     m = ResNet18()
     m.load_from_pretrained()
@@ -409,10 +409,10 @@ class TestMultiTensor(unittest.TestCase):
     # sometimes there is zeros in these grads... why?
     np.testing.assert_allclose(grad, shard_grad, atol=1e-5, rtol=1e-5)
 
-  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU", "AMD"), "slow, and flaky on LLVM/CPU")
+  @unittest.skipIf(CI and REAL_DEV in ("CUDA", "NV", "LLVM", "CPU"), "slow, and flaky on LLVM/CPU")
   def test_data_parallel_resnet_train_step(self):
     from extra.models.resnet import ResNet18
-    fake_image = Tensor.rand((2, 3, 224//16, 224//16))
+    fake_image = Tensor.rand((2, 3, 224//8, 224//8))
     labels = Tensor.randint(2, low=0, high=1000)
     m = ResNet18()
     self._test_model_train_step(m, fake_image, labels)

test/test_net_speed.py

@@ -0,0 +1,87 @@
+#!/usr/bin/env python
+import time
+import unittest
+import torch
+from tinygrad import Tensor, Device
+from tinygrad.helpers import Profiling, CI
+
+@unittest.skipIf(CI and Device.DEFAULT in {"CUDA", "NV"}, "slow")
+class TestConvSpeed(unittest.TestCase):
+
+  def test_mnist(self):
+    # https://keras.io/examples/vision/mnist_convnet/
+    conv = 3
+    inter_chan, out_chan = 32, 64
+
+    # ****** torch baseline *******
+
+    torch.backends.mkldnn.enabled = False
+
+    conv = 3
+    inter_chan, out_chan = 32, 64
+    c1 = torch.randn(inter_chan,1,conv,conv, requires_grad=True)
+    c2 = torch.randn(out_chan,inter_chan,conv,conv, requires_grad=True)
+    l1 = torch.randn(out_chan*5*5, 10, requires_grad=True)
+
+    c2d = torch.nn.functional.conv2d
+    mp = torch.nn.MaxPool2d((2,2))
+    lsm = torch.nn.LogSoftmax(dim=1)
+
+    cnt = 5
+    fpt, bpt = 0.0, 0.0
+    for i in range(cnt):
+      et0 = time.time()
+      x = torch.randn(128, 1, 28, 28, requires_grad=True)
+      x = mp(c2d(x,c1).relu())
+      x = mp(c2d(x,c2).relu())
+      x = x.reshape(x.shape[0], -1)
+      out = lsm(x.matmul(l1))
+      out = out.mean()
+      et1 = time.time()
+      out.backward()
+      et2 = time.time()
+      fpt += (et1-et0)
+      bpt += (et2-et1)
+
+    fpt_baseline = (fpt*1000/cnt)
+    bpt_baseline = (bpt*1000/cnt)
+    print("torch forward pass:  %.3f ms" % fpt_baseline)
+    print("torch backward pass: %.3f ms" % bpt_baseline)
+
+    # ****** tinygrad compare *******
+
+    c1 = Tensor(c1.detach().numpy(), requires_grad=True)
+    c2 = Tensor(c2.detach().numpy(), requires_grad=True)
+    l1 = Tensor(l1.detach().numpy(), requires_grad=True)
+
+    cnt = 5
+    fpt, bpt = 0.0, 0.0
+    for i in range(1+cnt):
+      et0 = time.time()
+      x = Tensor.randn(128, 1, 28, 28)
+      x = x.conv2d(c1).relu().avg_pool2d()
+      x = x.conv2d(c2).relu().max_pool2d()
+      x = x.reshape(shape=(x.shape[0], -1))
+      out = x.dot(l1).log_softmax()
+      out = out.mean()
+      out.backward()  # NOTE: we have to now compute this here, but it doesn't realize
+      out.realize()
+      et1 = time.time()
+      [x.grad.realize() for x in [c1, c2, l1]]
+      et2 = time.time()
+      if i == 0:
+        pr = Profiling(sort='time', frac=0.2)
+        pr.__enter__()
+      else:
+        fpt += (et1-et0)
+        bpt += (et2-et1)
+
+    pr.__exit__()
+    fpt = (fpt*1000/cnt)
+    bpt = (bpt*1000/cnt)
+    print("forward pass:  %.3f ms, %.2fx off baseline %.3f ms" % (fpt, fpt/fpt_baseline, fpt_baseline))
+    print("backward pass: %.3f ms, %.2fx off baseline %.3f ms" % (bpt, bpt/bpt_baseline, bpt_baseline))
+
+
+if __name__ == '__main__':
+  unittest.main()