tinygrad/.github/workflows/test.yml

name: Unit Tests
env:
  # increment this when downloads substantially change to avoid the internet
  CACHE_VERSION: '15'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
  IGNORE_OOB: 0

on:
  push:
    branches:
      - master
  pull_request:
  workflow_dispatch:

jobs:
  llvmspeed:
    name: LLVM Speed
    runs-on: ubuntu-24.04
    timeout-minutes: 20
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: llvm-speed
        deps: testing_minimal
        llvm: 'true'
    - name: Speed Test
      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py

  docs:
    name: Docs
    runs-on: ubuntu-22.04
    timeout-minutes: 10
    env:
      IGNORE_OOB: 1
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        deps: docs
        pydeps: "capstone torch"
    - 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
        cd $HOME/test_external_dir
        python -m venv venv
        source venv/bin/activate
        pip install $GITHUB_WORKSPACE
        python -c "from tinygrad.tensor import Tensor; print(Tensor([1,2,3,4,5]))"
        pip install mypy
        mypy -c "from tinygrad.tensor import Tensor; print(Tensor([1,2,3,4,5]))"
    - name: Run beautiful_mnist with tinygrad only
      run: |
        mkdir $GITHUB_WORKSPACE/test_dir
        cd $GITHUB_WORKSPACE/test_dir
        python -m venv venv
        source venv/bin/activate
        pip install $GITHUB_WORKSPACE
        cp $GITHUB_WORKSPACE/examples/beautiful_mnist.py .
        BS=2 STEPS=10 python beautiful_mnist.py
    - name: Test Docs Build
      run: python -m mkdocs build --strict
    - name: Test Docs
      run: python docs/abstractions3.py
    - name: Test README
      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
    - name: Test Quickstart
      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' docs/quickstart.md > quickstart.py &&  python quickstart.py
    - name: Test DEBUG
      run: DEBUG=100 python3 -c "from tinygrad import Tensor; N = 1024; a, b = Tensor.rand(N, N), Tensor.rand(N, N); c = (a.reshape(N, 1, N) * b.T.reshape(1, N, N)).sum(axis=2); print((c.numpy() - (a.numpy() @ b.numpy())).mean())"
    - name: Compile EfficientNet to C and test it
      run: |
        CPU=1 CPU_LLVM=0 python examples/compile_efficientnet.py > recognize.c
        clang -O2 recognize.c -lm -o recognize
        cat test/models/efficientnet/Chicken.jpg | ./recognize | grep cock

  torchbackend:
    name: Torch Backend Tests
    runs-on: ubuntu-latest
    timeout-minutes: 15
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: torch-backend-pillow-torchvision-et-pt
        deps: testing_minimal
        pydeps: "pillow torchvision expecttest"
        llvm: 'true'
    - name: Install ninja
      run: |
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
    - name: Test one op
      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_ops.py TestOps.test_add
    - name: Test ResNet-18
      run: DEBUG=2 python3 extra/torch_backend/example.py
    - name: custom tests
      run: python3 extra/torch_backend/test.py
    - name: Test one op in torch tests
      run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
    - name: Test Ops with TINY_BACKEND
      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
    - name: Test in-place operations on views
      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
    - name: Test multi-gpu
      run: CPU=1 CPU_LLVM=1 GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
    - name: Test kernel fusion
      run: python3 extra/torch_backend/test_kernel_fusion.py


  torchbackendmore:
    name: Torch Backend Tests More
    runs-on: ubuntu-latest
    timeout-minutes: 15
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: torch-backend-pillow-torchvision-et-pt
        deps: testing_minimal
        llvm: 'true'
    - name: Install ninja
      run: |
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
    - name: Test beautiful_mnist in torch with TINY_BACKEND
      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
    - name: Test some torch tests (expect failure)
      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true

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

  linter:
    name: Linters
    runs-on: ubuntu-latest
    timeout-minutes: 10

    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: linting-only
        python-version: '3.11'
        deps: linting
    - name: Lint bad-indentation and trailing-whitespace with pylint
      run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
    - name: Lint with ruff
      run: |
        pip3 install --upgrade --force-reinstall ruff==0.14.10
        pre-commit run ruff --all-files
        python3 -m ruff check examples/mlperf/ --ignore E501
        python3 -m ruff check extra/thunder/tiny/ --ignore E501 --ignore F841 --ignore E722
        python3 -m ruff check extra/torch_backend/backend.py
    - name: Run mypy
      run: |
        python -m mypy --lineprecision-report .
        cat lineprecision.txt
    - name: Run TYPED=1
      run: TYPED=1 python -c "import tinygrad"

  unittest:
    name: Unit Tests
    runs-on: ubuntu-latest
    timeout-minutes: 15

    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-12
        pydeps: "pillow numpy ftfy regex"
        deps: testing_unit
        llvm: 'true'
    - name: Check Device.DEFAULT
      run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
    - name: Run unit tests
      run: |
        CPU=1 python test/unit/test_device.py TestRunAsModule.test_module_runs
        CPU=1 python -m pytest -n=auto test/unit/ --durations=20 --deselect=test/unit/test_device.py::TestRunAsModule::test_module_runs
    - name: Run targetted tests on NULL backend
      run: NULL=1 python3 -m unittest test.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
    # TODO: too slow
    # - name: Run SDXL on NULL backend
    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
    - name: Run Clip tests for SD MLPerf on NULL backend
      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
    - name: Run AMD emulated BERT training on NULL backend
      run: EMULATE=AMD_RDNA4 NULL=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    # TODO: support fake weights
    #- name: Run LLaMA 7B on 4 fake devices
    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
    - name: Run GC tests
      run: python test/external/external_uop_gc.py
    - name: External Benchmark Schedule
      run: python3 test/external/external_benchmark_schedule.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
    - name: Regen dataset on test_tiny
      run: |
        test/external/process_replay/reset.py
        CAPTURE_PROCESS_REPLAY=1 python test/test_tiny.py TestTiny.test_plus
        python extra/optimization/extract_dataset.py
        gzip -c /tmp/sops > extra/datasets/sops.gz
        #DEBUG=1 MIN_ASTS=1 python extra/optimization/get_action_space.py
    - name: Repo line count < 20000 lines
      run: MAX_LINE_COUNT=20000 python sz.py

  spec:
    strategy:
      fail-fast: false
      matrix:
        group: [1, 2]
    name: SPEC=2 (${{ matrix.group }})
    runs-on: ubuntu-latest
    timeout-minutes: 15
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: spec-unit
        deps: testing_unit
        python-version: '3.14'
    - name: Test SPEC=2
      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/test_custom_kernel.py --ignore test/unit/test_hashing.py --ignore test/unit/test_autogen.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}

  fuzzing:
    name: Fuzzing
    runs-on: ubuntu-latest
    timeout-minutes: 10
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: fuzzing-unit
        deps: testing_unit
    - name: Fuzz Test symbolic
      run: python test/external/fuzz_symbolic.py
    - name: Fuzz Test symbolic (symbolic divisors)
      run: python test/external/fuzz_symbolic_symbolic_div.py
    - name: Fuzz Test fast idiv
      run: python test/external/fuzz_fast_idiv.py
    - name: Fuzz Test shape ops
      run: python test/external/fuzz_shape_ops.py

  testopenclimage:
    name: CL IMAGE Tests
    runs-on: ubuntu-22.04
    timeout-minutes: 15
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: gpu-image
          deps: testing_minimal
          opencl: 'true'
      - name: Test CL IMAGE=2 ops
        run: |
          CL=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
        # TODO: training is broken
        # CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

  testgpumisc:
    name: CL Misc tests
    runs-on: ubuntu-22.04
    timeout-minutes: 10
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: gen-dataset
          deps: testing_minimal
          opencl: 'true'
      - name: Generate Dataset
        run: CL=1 extra/optimization/generate_dataset.sh
      - name: Run Kernel Count Test
        run: CL=1 python -m pytest -n=auto test/external/external_test_opt.py
      - name: Run fused optimizer tests
        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/test_optim.py -k "not muon"
      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
          name: sops.gz
          path: /tmp/sops.gz

  testopenpilot:
    name: openpilot Compile Tests
    runs-on: ubuntu-22.04
    timeout-minutes: 15
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: openpilot-compile
          deps: testing
          opencl: 'true'
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1397 ALLOWED_GATED_READ_IMAGE=94 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp16
        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp32 (test correctness)
        run: DEBUGCL=1 CL=1 IMAGE=2 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
      - name: Test openpilot LLVM compile fp16
        run: FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

# ****** ONNX Tests ******

  testonnxcpu:
    name: ONNX (CPU) Tests
    runs-on: ubuntu-22.04
    timeout-minutes: 20

    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: onnxoptc
          deps: testing
          python-version: '3.11'
          llvm: 'true'
      - name: Test ONNX (CPU)
        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      - name: Test ONNX (LLVM)
        run: CPU=1 CPU_LLVM=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      - name: Test ONNX Runner (CPU)
        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_runner.py
      - name: Test Additional ONNX Ops (CPU)
        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_ops.py
      - name: Test Quantize ONNX
        run: CPU=1 CPU_LLVM=0 python3 test/test_quantize_onnx.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

  testopencl:
    name: ONNX (CL)+Optimization Tests
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: onnxoptl
          deps: testing
          pydeps: "tensorflow==2.19"
          python-version: '3.11'
          opencl: 'true'
      - name: Test ONNX (CL)
        run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      #- name: Test Optimization Helpers
      #  run: DEBUG=1 python3 extra/optimization/test_helpers.py
      #- name: Test Action Space
      #  run: DEBUG=1 CL=1 python3 extra/optimization/get_action_space.py
      - name: Test Beam Search
        run: CL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
      - name: NULL=1 beautiful_mnist_multigpu
        run: NULL=1 python examples/beautiful_mnist_multigpu.py
      - name: Test Bert training
        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

  testllm:
    name: Test LLM
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    env:
      IGNORE_OOB: 1
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: apps_llm
      - name: Test 1B LLM
        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm | grep -i rooster

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

  testmodels:
    name: Models (llvm+cpu+gpu)
    runs-on: ubuntu-22.04
    timeout-minutes: 15
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: models
          deps: testing
          opencl: 'true'
          llvm: 'true'
      - name: Test models (llvm)
        run: CPU=1 CPU_LLVM=1 python -m pytest -n=auto test/models --durations=20
      - name: Test models (opencl)
        run: CL=1 python -m pytest -n=auto test/models --durations=20
      - name: Test models (cpu)
        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/models --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

  testmetalmodels:
    name: Models (metal)
    runs-on: macos-14
    timeout-minutes: 20
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: metal
        deps: testing
        python-version: '3.11'
    - name: Test models (Metal)
      run: METAL=1 python -m pytest -n=auto test/models --durations=20
    - name: Test LLaMA compile speed
      run: METAL=1 python test/external/external_test_speed_llama.py

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

  testdevectorize:
    name: Linux (devectorize)
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    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: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
    - name: Test LLVM=1 DEVECTORIZE=0 for model
      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
    - name: Test CPU=1 DEVECTORIZE=0
      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py

  testdsp:
    name: Linux (DSP)
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: dsp-minimal
        deps: testing_minimal
        pydeps: "onnx==1.18.0 onnxruntime pillow"
        llvm: "true"
    - name: Set up Docker Buildx
      uses: docker/setup-buildx-action@v3
    - name: Build QEMU Docker with cache
      uses: docker/build-push-action@v4
      with:
        file: extra/dsp/Dockerfile
        push: false
        load: true
        tags: qemu-hexagon:latest
        cache-from: type=gha
        cache-to: type=gha,mode=min
    - name: Set MOCKDSP env
      run: printf "MOCKDSP=1" >> $GITHUB_ENV
    - name: Run test_tiny on DSP
      run: DEBUG=2 DSP=1 python test/test_tiny.py
    - name: Test transcendentals
      run: CC=clang-20 DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
    - name: Test quantize onnx
      run: DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py

  testwebgpu:
    name: Linux (WebGPU)
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: webgpu-minimal
        deps: testing_minimal
        python-version: '3.11'
        webgpu: 'true'
    - name: Check Device.DEFAULT (WEBGPU) and print some source
      run: |
        WEBGPU=1 python -c "from tinygrad import Device; assert Device.DEFAULT == 'WEBGPU', Device.DEFAULT"
        WEBGPU=1 DEBUG=4 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
    - name: Run selected webgpu tests
      run: |
          WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit \
          --ignore=test/test_copy_speed.py --ignore=test/test_rearrange_einops.py \
          --ignore=test/test_fuzz_shape_ops.py --durations=20
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

  testamd:
    strategy:
      fail-fast: false
      matrix:
        backend: [amd, amdllvm]

    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    env:
      AMD: 1
      MOCKGPU: 1
      FORWARD_ONLY: 1
      AMD_LLVM: ${{ matrix.backend == 'amdllvm' && '1' || matrix.backend != 'amdllvm' && '0' }}
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
          deps: testing_minimal
          amd: 'true'
          llvm: ${{ matrix.backend == 'amdllvm' && 'true' }}
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run LLVM test
        if: matrix.backend=='amdllvm'
        run: python test/device/test_amd_llvm.py
      - name: Run pytest (amd)
        run: python -m pytest -n=auto test/test_ops.py test/test_dtype.py test/test_dtype_alu.py test/test_linearizer.py test/test_randomness.py test/test_jit.py test/test_graph.py test/test_multitensor.py test/device/test_hcq.py test/testextra/test_cfg_viz.py --durations=20
      - name: Run pytest (amd)
        run: python -m pytest test/external/external_test_am.py --durations=20
      - name: Run TRANSCENDENTAL math
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
      - name: Run TestOps.test_add with SQTT
        run: |
          VIZ=1 PMC=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

  testamdasm:
    name: AMD ASM IDE
    runs-on: ubuntu-24.04
    timeout-minutes: 10
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: rdna3-emu
          deps: testing_minimal
          amd: 'true'
          python-version: '3.13'
      - name: Install LLVM 21
        run: |
          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
          echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
          sudo apt-get update
          sudo apt-get install llvm-21 llvm-21-tools cloc
      - name: RDNA3 Line Count
        run: cloc --by-file extra/assembly/amd/*.py
      - name: Run RDNA3 emulator tests
        run: python -m pytest -n=auto extra/assembly/amd/ --durations 20
      - name: Run RDNA3 emulator tests (AMD_LLVM=1)
        run: AMD_LLVM=1 python -m pytest -n=auto extra/assembly/amd/ --durations 20
      - name: Run RDNA3 dtype tests
        run: AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=0 pytest -n=auto test/test_dtype_alu.py test/test_dtype.py
      - name: Run RDNA3 dtype tests (AMD_LLVM=1)
        run: AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=1 pytest -n=auto test/test_dtype_alu.py test/test_dtype.py
      # TODO: run all once emulator is faster
      - name: Run RDNA3 ops tests
        run: SKIP_SLOW_TEST=1 AMD=1 PYTHON_REMU=1 MOCKGPU=1 AMD_LLVM=0 pytest -n=auto test/test_ops.py -k "test_sparse_categorical_crossentropy or test_tril"

  testamdautogen:
    name: AMD autogen
    runs-on: ubuntu-24.04
    timeout-minutes: 10
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: rdna3-autogen
          pydeps: "pdfplumber"
      - name: Verify AMD autogen is up to date
        run: |
          python -m extra.assembly.amd.pdf --arch all
          git diff --exit-code extra/assembly/amd/autogen/

  testnvidia:
    strategy:
      fail-fast: false
      matrix:
        backend: [ptx, nv]

    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    env:
      MOCKGPU: 1
      FORWARD_ONLY: 1
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
          deps: testing_minimal
          cuda: 'true'
          ocelot: 'true'
      - name: Set env
        run: printf "${{ matrix.backend == 'PTX' && 'CUDA=1\nCUDA_PTX=1' || matrix.backend == 'nv' && 'NV=1\nSKIP_SLOW_TEST=1' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CUDA','NV'], Device.DEFAULT"
          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (cuda)
        # skip multitensor because it's slow
        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --ignore test/test_gc.py --ignore test/test_multitensor.py --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

  testcpuopencl:
    strategy:
      fail-fast: false
      matrix:
        backend: [llvm, cpu, opencl, lvp]

    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: ${{ matrix.backend }}-minimal
          deps: testing_minimal
          opencl: ${{ matrix.backend == 'opencl' && 'true' }}
          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
      - name: Set env
        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
      - name: Run pytest (${{ matrix.backend }})
        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
      - name: Run TRANSCENDENTAL math
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

# ****** OSX Tests ******

  testmetal:
    name: MacOS (unit)
    runs-on: macos-14
    timeout-minutes: 20
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: metal
        deps: testing
        python-version: '3.11'
        amd: 'true'
        cuda: 'true'
        ocelot: 'true'
        llvm: 'true'
    - name: Run unit tests
      env:
        LIBCLANG_PATH: '/opt/homebrew/opt/llvm@20/lib/libclang.dylib'
      run: METAL=1 python -m pytest -n=auto test/unit/ --durations=20
    - name: Run ONNX
      run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
    - name: Test tensor core ops (fake)
      run: METAL=1 DEBUG=3 TC=2 python test/test_ops.py TestOps.test_gemm
    - name: Test tensor core ops (real)
      run: METAL=1 DEBUG=3 python test/test_ops.py TestOps.test_big_gemm
    - name: Test Beam Search
      run: METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
    #- name: Fuzz Test linearizer
    #  run: METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
    - name: Run TRANSCENDENTAL math
      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
    - name: Run pytest (amd)
      env:
        MOCKGPU: 1
        AMD: 1
        AMD_LLVM: 0
        FORWARD_ONLY: 1
      run: |
        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
    - name: Run pytest (amd with llvm backend)
      env:
        MOCKGPU: 1
        AMD: 1
        AMD_LLVM: 1
        FORWARD_ONLY: 1
      run: |
        python -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py test/device/test_amd_llvm.py --durations=20
    - name: Run pytest (ptx)
      env:
        MOCKGPU: 1
        NV_PTX: 1
        NV: 1
        FORWARD_ONLY: 1
      run: |
        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

  osxwebgpu:
    name: MacOS (WebGPU)
    runs-on: macos-14
    timeout-minutes: 10
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
        key: osx-webgpu
        deps: testing
        webgpu: 'true'
    - name: Test infinity math in WGSL
      run: WEBGPU=1 python -m pytest -n=auto test/test_renderer_failures.py::TestWGSLFailures::test_multiply_infinity --durations=20
    - name: Build WEBGPU Efficientnet
      run: WEBGPU=1 WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m examples.compile_efficientnet
    - name: Clean npm cache
      run: npm cache clean --force
    - name: Install Puppeteer
      run: npm install puppeteer
    # this is also flaky
    #- name: Run WEBGPU Efficientnet
    #  run: node test/web/test_webgpu.js
    # this is flaky
    #- name: Run VIZ tests as external package
    #  run: |
    #    mkdir $GITHUB_WORKSPACE/test_dir
    #    cd $GITHUB_WORKSPACE/test_dir
    #    python -m venv venv
    #    source venv/bin/activate
    #    pip install $GITHUB_WORKSPACE
    #    cp $GITHUB_WORKSPACE/test/web/test_viz.js .
    #    node test_viz.js
    - name: Test ONNX Runner (WEBGPU)
      run: WEBGPU=1 python3 test/external/external_test_onnx_runner.py

  osxtests:
    strategy:
      fail-fast: false
      matrix:
        backend: [metal, llvm, cpu, lvp]
    name: MacOS (${{ matrix.backend }})
    runs-on: macos-15
    timeout-minutes: 20
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: macos-${{ matrix.backend }}-minimal
          deps: testing_minimal
          pydeps: "capstone"
          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
      - name: Set env
        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
      - name: Run pytest (${{ matrix.backend }})
        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
      - name: Run macOS-specific unit test
        if: matrix.backend == 'cpu'
        run: python3 -m pytest test/unit/test_disk_tensor.py::TestDiskTensor::test_copy_to_cpu_not_truncated

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

  wintests:
    strategy:
      fail-fast: false
      matrix:
        backend: [llvm, cpu, webgpu]

    name: Windows (${{ matrix.backend }})
    runs-on: windows-latest
    timeout-minutes: 15
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: windows-${{ matrix.backend }}-minimal
          deps: testing_unit
          pydeps: ${{ matrix.backend == 'webgpu' && 'dawn-python' || '' }}
      - name: Set env
        shell: bash
        run:  printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'webgpu' && 'WEBGPU=1'}}" >> $GITHUB_ENV
      - name: Run unit tests
        if: matrix.backend=='llvm'
        # test_newton_schulz hits RecursionError
        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_elf.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
      - name: Run pytest (${{ matrix.backend }})
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
          python -m pytest -n=auto test/test_tiny.py test/test_ops.py --durations=20

# ****** Compile-only Tests ******

  compiletests:
    strategy:
      fail-fast: false
      matrix:
        backend: [ir3, nak]
    name: Compile-only (${{ matrix.backend }})
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
          key: compile-${{ matrix.backend }}
          deps: testing_minimal
          mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
          python-version: '3.14'
      - name: Set env
        shell: bash
        run: printf "NULL=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
      - name: Run test_ops
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
          DEBUG=4 python3 test/test_ops.py TestOps.test_add
          python -m pytest -n=auto test/test_ops.py --durations=20