precompile_backward tests for sched_cache (#17544)

* work

* back

* work

* keep +

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

@@ -41,10 +41,6 @@ inputs:
     description: "Install LLVM?"
     required: false
     default: 'false'
-  tinydreno:
-    description: "Install tinydreno"
-    required: false
-    default: 'false'
   qemu:
     description: "Install qemu"
     required: false
@@ -277,12 +273,6 @@ runs:
       shell: bash
       run: brew install llvm@20
 
-    # *** tinydreno ***
-    - name: Install tinydreno (linux)
-      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
-
     # *** OpenCL ***
     - name: Install rusticl
       if: inputs.opencl == 'true'

.github/workflows/platform.yml

@@ -0,0 +1,181 @@
+name: Platform Tests
+env:
+  # increment this when downloads substantially change to avoid the internet
+  CACHE_VERSION: '19'
+  CAPTURE_PROCESS_REPLAY: ${{ github.event_name == 'pull_request' && contains(github.event.pull_request.title, '[pr]') && '1' || '0' }}
+  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+  PYTHONPATH: ${{ github.workspace }}
+  CHECK_OOB: 1
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+  workflow_dispatch:
+
+concurrency:
+  group: platform-${{ github.event_name }}-${{ github.event_name == 'pull_request' && github.event.pull_request.number || github.run_id }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+
+# ****** OSX Tests ******
+
+  unittestmacos:
+    name: MacOS (unit)
+    runs-on: macos-26
+    timeout-minutes: 20
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: unittest-macos
+        deps: testing_unit
+    - name: Run unit tests
+      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
+    - name: Test tensor core ops (fake)
+      run: DEV=METAL DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
+    - name: Test tensor core ops (real)
+      run: DEV=METAL DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
+    - name: Test Beam Search
+      run: DEV=METAL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
+    - name: Test Device Specific
+      run: DEV=METAL python3 -m pytest test/device/test_metal.py
+    #- name: Fuzz Test linearizer
+    #  run: DEV=METAL DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  unittestmacosmock:
+    name: MacOS (unit, mock)
+    runs-on: macos-26
+    timeout-minutes: 20
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: unittest-macos-mock
+        deps: testing_unit
+        amd: 'true'
+        ocelot: 'true'
+    - name: Run NULL backend tests
+      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
+    - name: Run pytest (amd)
+      env:
+        DEV: MOCKKFD+AMD
+        FORWARD_ONLY: 1
+      run: |
+        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
+    - name: Run pytest (ptx)
+      env:
+        DEV: "MOCK+NV:PTX"
+        FORWARD_ONLY: 1
+        # TODO: failing due to library loading error
+        CAPTURE_PROCESS_REPLAY: 0
+      run: |
+        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py \
+          test/testextra/test_hevc.py::TestHevc::test_hevc_decode_compile --durations=20
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testmetal:
+    strategy:
+      fail-fast: false
+      matrix:
+        group: [1, 2]
+    name: MacOS (DEV=METAL) (${{ matrix.group }})
+    runs-on: macos-26
+    timeout-minutes: 20
+    env:
+      DEV: METAL
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: macos-metal
+        deps: testing_unit
+    - name: Check Device.DEFAULT and print some source
+      run: |
+        python -c "from tinygrad import Device; assert Device.DEFAULT == 'METAL'"
+        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+    - name: Run backend tests
+      run: python -m pytest -n=auto test/backend --durations=20 --splits 2 --group ${{ matrix.group }}
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testmacos:
+    strategy:
+      fail-fast: false
+      matrix:
+        dev:
+          - 'CPU:CLANG'
+          - 'CPU:LLVM'
+          - 'CPU:LVP'
+          - 'WEBGPU'
+
+    name: MacOS (DEV=${{ matrix.dev }})
+    runs-on: macos-26
+    timeout-minutes: 20
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: macos-${{ matrix.dev }}
+        deps: "testing_unit${{ contains(matrix.dev, 'LVP') && ' mesa' || '' }}"
+        llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') }}
+        webgpu: ${{ matrix.dev == 'WEBGPU' }}
+    - name: Set env
+      run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
+    - name: Check Device.DEFAULT and print some source
+      run: |
+        python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
+        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+    - name: Run test_tiny
+      run: python -m pytest -n=auto test/test_tiny.py --durations=20
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+# ****** Windows Tests ******
+
+  testwindows:
+    strategy:
+      fail-fast: false
+      matrix:
+        dev:
+          - 'CPU:CLANG'
+          - 'CPU:LLVM'
+          - 'CPU:X86'
+          - 'WEBGPU'
+
+    name: Windows (DEV=${{ matrix.dev }})
+    runs-on: windows-2025
+    timeout-minutes: 15
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v6
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: windows-${{ matrix.dev }}-minimal
+          deps: testing_unit
+          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
+      - name: Set env
+        shell: bash
+        run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
+      - name: Check Device.DEFAULT and print some source
+        shell: bash
+        run: |
+          python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
+          DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+      - name: Run test_tiny
+        shell: bash
+        run: python -m pytest -n=auto test/test_tiny.py --durations=20

.github/workflows/test.yml

@@ -21,7 +21,7 @@ concurrency:
 jobs:
   docs:
     name: Docs
-    runs-on: &linux ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 10
     env:
       CHECK_OOB: 0
@@ -61,7 +61,7 @@ jobs:
 
   torchbackend:
     name: Torch Backend Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -88,7 +88,7 @@ jobs:
 
   bepython:
     name: Python Backend
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -126,7 +126,7 @@ jobs:
 
   linter:
     name: Linters
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 10
 
     steps:
@@ -157,7 +157,7 @@ jobs:
 
   nulltest:
     name: Null Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
 
     steps:
@@ -191,7 +191,7 @@ jobs:
 
   unittest:
     name: Unit Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
 
     steps:
@@ -228,7 +228,7 @@ jobs:
       matrix:
         group: [1, 2]
     name: SPEC=2 (${{ matrix.group }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -244,7 +244,7 @@ jobs:
 
   fuzzing:
     name: Fuzzing
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 10
     steps:
     - name: Checkout Code
@@ -260,7 +260,7 @@ jobs:
 
   testopenclimage:
     name: CL IMAGE Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
       - name: Checkout Code
@@ -280,7 +280,7 @@ jobs:
 
   testopenpilot:
     name: openpilot Compile Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
       - name: Checkout Code
@@ -294,7 +294,7 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=54 FLOAT16=1 DEV="CL::IMAGE_PITCH_ALIGNMENT=64" IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=38 FLOAT16=1 DEV="CL::IMAGE_PITCH_ALIGNMENT=64" IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
           # IMAGE_PITCH_ALIGNMENT=64 matches adreno 630
       - name: Test openpilot CL compile fp32 (test correctness)
         run: |
@@ -309,7 +309,7 @@ jobs:
 
   testonnxcpu:
     name: ONNX (CPU) Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
 
     steps:
@@ -328,7 +328,7 @@ jobs:
 
   testoptim:
     name: Optimization Tests
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     steps:
       - name: Checkout Code
@@ -360,7 +360,7 @@ jobs:
 
   testllm:
     name: Test LLM
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     env:
       CHECK_OOB: 0
@@ -387,7 +387,7 @@ jobs:
 
   testmodels:
     name: Models
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
       - name: Checkout Code
@@ -407,7 +407,7 @@ jobs:
 
   testdsp:
     name: Linux (DSP)
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     steps:
     - name: Checkout Code
@@ -435,7 +435,7 @@ jobs:
           - 'WEBGPU'
 
     name: Linux (DEV=${{ matrix.dev }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     steps:
     - name: Checkout Code
@@ -461,7 +461,7 @@ jobs:
 
   testamdasm:
     name: AMD ASM IDE
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     env:
       DEV: MOCKKFD+AMD
@@ -507,7 +507,7 @@ jobs:
 
   hcq2:
     name: hcq2
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 5
     steps:
       - name: Checkout Code
@@ -521,16 +521,15 @@ jobs:
       - name: Run HCQ2 tests
         run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/test_tiny.py
       - name: Run HCQ2 multi-device tests
-        run: |
-          HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_multitensor.py \
-            TestMultiTensor.test_simple_add TestMultiTensor.test_shard_reduce \
-            TestMultiTensor.test_backward_sum TestMultiTensor.test_matmul_shard_0_0
+        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest -n=auto test/backend/test_multitensor.py
       - name: Run HCQ2 JIT tests
         run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_jit.py
+      - name: Run HCQ2 unit tests
+        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest test/device/test_hcq2.py
 
   testmockam:
     name: Linux (am)
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     env:
       DEV: MOCKPCI+AMD
@@ -566,7 +565,7 @@ jobs:
         arch: [gfx1100, gfx1201, gfx950]
 
     name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
     env:
       DEV: MOCKKFD+AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
@@ -585,8 +584,11 @@ jobs:
         run: |
           python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
           DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
+      - name: Run MXFP4 Llama training on NULL backend
+        if: ${{ matrix.backend == 'amd' && matrix.arch == 'gfx950' }}
+        run: PYTHONPATH=. DEV=NULL:HIP:gfx950 MXFP4=1 LLAMA_LAYERS=2 BENCHMARK=3 NULL_ALLOW_COPYOUT=1 NO_HIPCC=1 ROCM_PATH=/opt/rocm JITBEAM=0 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/profile.sh
       - name: Run pytest (amd)
-        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM test/opt/test_tensor_cores.py --durations=20
+        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
       - name: Run disk copy tests
         run: python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk
       - name: Run TRANSCENDENTAL math
@@ -601,7 +603,7 @@ jobs:
         backend: [ptx, nv]
 
     name: Linux (${{ matrix.backend }})
-    runs-on: *linux
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 20
     env:
       FORWARD_ONLY: 1
@@ -629,175 +631,23 @@ jobs:
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
-# ****** OSX Tests ******
-
-  unittestmacos:
-    name: MacOS (unit)
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-macos
-        deps: testing_unit
-    - name: Run unit tests
-      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
-    - name: Test tensor core ops (fake)
-      run: DEV=METAL DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
-    - name: Test tensor core ops (real)
-      run: DEV=METAL DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
-    - name: Test Beam Search
-      run: DEV=METAL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
-    - name: Test Device Specific
-      run: DEV=METAL python3 -m pytest test/device/test_metal.py
-    #- name: Fuzz Test linearizer
-    #  run: DEV=METAL DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  unittestmacosmock:
-    name: MacOS (unit, mock)
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-macos-mock
-        deps: testing_unit
-        amd: 'true'
-        ocelot: 'true'
-    - name: Run NULL backend tests
-      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
-    - name: Run pytest (amd)
-      env:
-        DEV: MOCKKFD+AMD
-        FORWARD_ONLY: 1
-      run: |
-        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
-    - name: Run pytest (ptx)
-      env:
-        DEV: "MOCK+NV:PTX"
-        FORWARD_ONLY: 1
-        # TODO: failing due to library loading error
-        CAPTURE_PROCESS_REPLAY: 0
-      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
-
-  testmetal:
-    strategy:
-      fail-fast: false
-      matrix:
-        group: [1, 2]
-    name: MacOS (DEV=METAL) (${{ matrix.group }})
-    runs-on: macos-26
-    timeout-minutes: 20
-    env:
-      DEV: METAL
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: macos-metal
-        deps: testing_unit
-    - name: Check Device.DEFAULT and print some source
-      run: |
-        python -c "from tinygrad import Device; assert Device.DEFAULT == 'METAL'"
-        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run backend tests
-      run: python -m pytest -n=auto test/backend --durations=20 --splits 2 --group ${{ matrix.group }}
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmacos:
-    strategy:
-      fail-fast: false
-      matrix:
-        dev:
-          - 'CPU:CLANG'
-          - 'CPU:LLVM'
-          - 'CPU:LVP'
-          - 'WEBGPU'
-
-    name: MacOS (DEV=${{ matrix.dev }})
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: macos-${{ matrix.dev }}
-        deps: "testing_unit${{ contains(matrix.dev, 'LVP') && ' mesa' || '' }}"
-        llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') }}
-        webgpu: ${{ matrix.dev == 'WEBGPU' }}
-    - name: Set env
-      run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
-    - name: Check Device.DEFAULT and print some source
-      run: |
-        python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
-        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run test_tiny
-      run: python -m pytest -n=auto test/test_tiny.py --durations=20
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-# ****** Windows Tests ******
-
-  testwindows:
-    strategy:
-      fail-fast: false
-      matrix:
-        dev:
-          - 'CPU:CLANG'
-          - 'CPU:LLVM'
-          - 'CPU:X86'
-          - 'WEBGPU'
-
-    name: Windows (DEV=${{ matrix.dev }})
-    runs-on: windows-2025
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: windows-${{ matrix.dev }}-minimal
-          deps: testing_unit
-          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
-      - name: Set env
-        shell: bash
-        run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
-      - name: Check Device.DEFAULT and print some source
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
-          DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-      - name: Run test_tiny
-        shell: bash
-        run: python -m pytest -n=auto test/test_tiny.py --durations=20
-
 # ****** Compile-only Tests ******
 
   compiletests:
     strategy:
       fail-fast: false
       matrix:
-        backend: [ir3, nak]
-    name: Compile-only (${{ matrix.backend }})
-    runs-on: *linux
+        dev:
+          - 'NULL:IR3:a630'
+          - 'NULL:QCOMCL:a630'
+          - 'NULL:NAK:sm_120'
+    name: Compile-only (DEV=${{ matrix.dev }})
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
     timeout-minutes: 15
+    env:
+      NULL_ALLOW_COPYOUT: 1
+      DEV: ${{ matrix.dev }}${{ contains(matrix.dev, 'a630') && ',IMAGE_PITCH_ALIGNMENT=64' || '' }}
+      IMAGE: ${{ contains(matrix.dev, 'a630') && '1' || '0' }}
     steps:
       - name: Checkout Code
         uses: actions/checkout@v6
@@ -806,51 +656,14 @@ jobs:
         with:
           key: compile-${{ matrix.backend }}
           deps: "testing_unit mesa"
-      - name: Set env
+          qemu: ${{ contains(matrix.dev, 'QCOMCL') }}
+      - name: Test IMAGE
         shell: bash
-        run: printf "NULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'DEV=NULL:IR3:a630' || matrix.backend == 'nak' && 'DEV=NULL:NAK:sm_120' }}" >> $GITHUB_ENV
+        if: contains(matrix.dev, 'a630')
+        run: DEBUG=7 python3 test/backend/test_ops.py TestOps.test_gemm | grep isam
       - name: Run test_ops
         shell: bash
         run: |
           python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
           DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
           python -m pytest -n=auto test/backend/test_ops.py --durations=20
-      - name: Run test_ops (IMAGE)
-        if: matrix.backend == 'ir3'
-        shell: bash
-        env:
-          IMAGE: 1
-          DEV: "NULL:IR3:a630,IMAGE_PITCH_ALIGNMENT=64"
-        run: |
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_gemm | grep image_load
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-  qcomclcompiletests:
-    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-qcomcl
-          deps: testing_unit
-          tinydreno: 'true'
-      - name: Set env
-        shell: bash
-        run: printf "DEV=NULL:QCOMCL:a630\nNULL_ALLOW_COPYOUT=1" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-      - name: Run test_ops (IMAGE)
-        shell: bash
-        env:
-          IMAGE: 1
-          DEV: "NULL:QCOMCL:a630,IMAGE_PITCH_ALIGNMENT=64"
-        run: |
-          DEBUG=4 python test/backend/test_ops.py TestOps.test_gemm | grep read_imagef
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20

examples/mlperf/model_train.py

@@ -1458,7 +1458,8 @@ def train_llama3():
 
   # realize everything here
   if optim.master_params: Tensor.realize(*optim.master_params)
-  Tensor.realize(*optim.params, *fp8_inv_scales, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
+  loss_acc = Tensor.zeros(1, dtype=dtypes.float32, device=device)
+  Tensor.realize(loss_acc, *optim.params, *fp8_inv_scales, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
 
   @TinyJit
   def minibatch(tokens:Tensor):
@@ -1476,8 +1477,8 @@ def train_llama3():
     for g, new_g in zip(grads, loss.gradient(*optim.params)):
       apply_grad(g, new_g.uop)
 
-    loss_cpu = loss.flatten().float().to("CPU")
-    return loss_cpu.realize(*grads, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
+    loss_acc.assign(loss_acc + loss.flatten().float())
+    return loss_acc.realize(*grads, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
 
   @TinyJit
   def optim_step():
@@ -1490,9 +1491,10 @@ def train_llama3():
 
     lr_cpu = optim.lr.float().to("CPU")
     grad_norm_cpu = grad_norm.float().to("CPU")
-    Tensor.realize(lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales, *fp8_amax, *fp8_grad_amax)
+    loss_cpu = loss_acc.to("CPU")
+    Tensor.realize(lr_cpu, grad_norm_cpu, loss_cpu, loss_acc.assign(0), *grads, *fp8_inv_scales, *fp8_amax, *fp8_grad_amax)
 
-    return lr_cpu, grad_norm_cpu
+    return lr_cpu, grad_norm_cpu, loss_cpu
 
   @TinyJit
   @Context(TRAINING=0)
@@ -1547,8 +1549,8 @@ def train_llama3():
       st = time.perf_counter()
 
       stopped = False
-      losses, data_time, dev_time = [], 0, 0
-      for _ in range(grad_acc if i >= 2 else 1):
+      data_time, dev_time = 0, 0
+      for _ in range(accum_steps:=grad_acc if i >= 2 else 1):
         ist = time.perf_counter()
         try: tokens = next(train_iter)
         except StopIteration:
@@ -1556,16 +1558,15 @@ def train_llama3():
           break
         mst = time.perf_counter()
         data_time += mst - ist
-        losses.append(minibatch(tokens).item())
+        minibatch(tokens)
         dev_time += time.perf_counter() - mst
       if stopped: break
 
       gt = time.perf_counter()
       ret = optim_step()
-      lr, grad_norm = ret[0].item(), ret[1].item()
+      lr, grad_norm, loss = ret[0].item(), ret[1].item(), ret[2].item() / accum_steps
       et = time.perf_counter()
 
-      loss = sum(losses) / len(losses)
       optim_time = et - gt
       dev_time += optim_time
       step_time = et - st

examples/mlperf/models/flat_llama.py

@@ -114,6 +114,11 @@ def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
                              amax_x2:Tensor|None, next_amax_x2:Tensor|None,
                              grad_amax_xw13:Tensor|None, next_grad_amax_xw13:Tensor|None,
                              grad_amax_xout:Tensor|None, next_grad_amax_xout:Tensor|None):
+  if FUSED_SILU_W13 and MXFP4:
+    from extra.llama_kernels.swiglu import swiglu
+    out, *ret = matmul(swiglu(x_w13), w2, amax_x=amax_x2, w_inv_scale=s_2, grad_amax_state=grad_amax_xout,
+                       next_grad_amax_state=next_grad_amax_xout, next_amax_x=next_amax_x2)
+    return out, ret
   if FUSED_SILU_W13 and not MXFP4:
     from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
     x2_fp8 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13,

examples/mlperf/models/gpt_oss.py

@@ -12,7 +12,7 @@ from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
 from tinygrad.uop.ops import Ops, UOp
 from extra.models.llama import apply_rotary_emb
 from extra.llama_kernels.rmsnorm import rmsnorm
-from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8
+from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8, asm_gemm, can_use_asm_gemm
 from extra.gemm.moe_gemm import grouped_mx_gemm
 from extra.gemm.moe_routing import route, dispatch, combine
 
@@ -182,12 +182,12 @@ class GPTOSS:
     xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
     xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)  # (B,N,H,D)/(B,N,KV,D)
 
-    if sliding:
-      attn = self._sliding_attention(xq, xk, xv, sinks)
-    elif getenv("HK_FLASH_ATTENTION"):
+    if getenv("HK_FLASH_ATTENTION"):
       from extra.thunder.amd.fa import flash_attention
-      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks)
+      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks, window=self.sliding_window if sliding else 0)
       attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
+    elif sliding:
+      attn = self._sliding_attention(xq, xk, xv, sinks)
     else:
       xqm = xq.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
       xkm, xvm = xk.permute(0, 2, 1, 3).unsqueeze(2), xv.permute(0, 2, 1, 3).unsqueeze(2)
@@ -263,7 +263,11 @@ class GPTOSS:
                         w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
       h, *_ = self.run_layer(h, freqs_cis, mask_full, i % 2 == 0, attn_kwargs, ffn_kwargs, save=save)
 
-    logits = self.norm(h) @ self.output.T
+    h_normed = self.norm(h)
+    pad = (-self.dim) % 256
+    h_padded, w_padded = h_normed.pad((None, None, (0, pad))), self.output.pad(((0, 0), (0, pad)))
+    if ASM_GEMM and can_use_asm_gemm(h_padded, w_padded.T): logits = asm_gemm(h_padded, w_padded.T)
+    else: logits = h_normed @ self.output.T
     return logits
 
 def _get_pads(uop:UOp) -> list[UOp]:

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -1,8 +1,8 @@
 #!/usr/bin/env bash
 
 export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
+export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
+export PATH="$ROCM_PATH/bin:$PATH"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -16,7 +16,7 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export FP8=${FP8:-1}
+export MXFP4=${MXFP4:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
 export FAST_CE=${FAST_CE:-1}
 export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
 
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))
 
@@ -46,7 +46,7 @@ export DATA_SEED=${DATA_SEED:-5760}
 export JITBEAM=${JITBEAM:-3}
 export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
 
-export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
+export FAKEDATA=${FAKEDATA:-$([[ "$DEV" == NULL:* ]] && echo 1 || echo 0)} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
   export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
 fi

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -1,8 +1,8 @@
 #!/usr/bin/env bash
 
 export PYTHONPATH="."
-export PATH="/opt/rocm-7.1.1/bin:$PATH"
-export ROCM_PATH="/opt/rocm-7.1.1"
+export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
+export PATH="$ROCM_PATH/bin:$PATH"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -16,7 +16,7 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export FP8=${FP8:-1}
+export MXFP4=${MXFP4:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
 export FAST_CE=${FAST_CE:-1}
 export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}
 
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))
 

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/profile.sh

@@ -1,4 +1,6 @@
 #!/bin/bash
+set -e
+
 export BENCHMARK=${BENCHMARK:-5}
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -11,6 +11,7 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
+export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh

@@ -11,6 +11,7 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
+export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}

extra/export_model.py

@@ -35,7 +35,7 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li
     return name
 
   for call in iter_kernel_calls(linear):
-    arg_uops = [b for b in call.src[1:] if b.op is not Ops.BIND]
+    arg_uops = [b for b in call.src[1:] if not b.is_bound_var]
     prg = to_program(call.src[0], Device[arg_uops[0].device].renderer)
     info = prg.arg
     functions[info.function_name] = prg.src[2].arg

extra/gemm/cdna_asm_gemm.py

@@ -6,6 +6,7 @@ from tinygrad.renderer import Estimates
 from tinygrad.helpers import getenv, all_same, DEBUG, ceildiv
 from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8
+from extra.llama_kernels.quantize_mxfp4 import quantize_mxfp4
 
 TILE_M, TILE_N, TILE_K = 256, 256, 64
 
@@ -121,10 +122,30 @@ def custom_mxfp4_gemm(C:UOp, A:UOp, B:UOp, scale_a:UOp, scale_b:UOp, *extra:UOp,
   groups_x, groups_y = UOp.special(ceildiv(N, tile_n), "gidx0"), UOp.special(ceildiv(M, tile_m), "gidx1")
   lds = UOp.placeholder((163840,), dtypes.uint8, 0, AddrSpace.LOCAL)
   sink = UOp.sink(C.base, A.base, B.base, scale_a.base, scale_b.base, *(x.base for x in extra), lds, threads, groups_x, groups_y,
-                  arg=KernelInfo(f"custom_mxfp4_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K)))
+                  arg=KernelInfo(f"mxfp4_gemm_{M}_{N}_{K}",
+                                 estimates=Estimates(ops=2*M*N*K, mem=(M*half_k+N*half_k)*A.dtype.itemsize+M*N*C.dtype.itemsize)))
   insts = build_kernel(M, N, K, tile_m, tile_n)
   return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=x) for x in insts))))
 
+def _mxfp4_gemm_quantized(a_q:Tensor, b_q:Tensor, scale_a:Tensor, scale_b:Tensor) -> Tensor:
+  M, half_k = a_q.shape
+  N, half_k_b = b_q.shape
+  assert half_k == half_k_b
+  is_multi = isinstance(a_q.device, tuple)
+  reduce_out = is_multi and (a_q.uop.axis == 1 or b_q.uop.axis == 1)
+  if not is_multi: out = Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a_q.device)
+  elif reduce_out: out = Tensor(Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a_q.device).uop.unshard(0), device=a_q.device)
+  elif a_q.uop.axis == 0:
+    out = Tensor(Tensor.invalids(1, M//len(a_q.device), N, dtype=dtypes.bfloat16, device=a_q.device).uop.unshard(1), device=a_q.device)
+  elif b_q.uop.axis == 0:
+    out = Tensor(Tensor.invalids(1, M, N//len(a_q.device), dtype=dtypes.bfloat16, device=a_q.device).uop.unshard(2), device=a_q.device)
+  else: out = Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a_q.device)
+  tile_m, tile_n = next((tm, tn) for tm, tn in ((256, 256), (192, 256), (128, 512)) if M % tm == N % tn == 0)
+  out = Tensor.custom_kernel(out, a_q, b_q, scale_a, scale_b,
+                             fxn=functools.partial(custom_mxfp4_gemm, tile_m=tile_m, tile_n=tile_n))[0]
+  if reduce_out: out = out.sum(0)
+  return out.squeeze(0)
+
 def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
   # 1x32 block scaling along the last axis
   *batch, K = x.shape
@@ -137,50 +158,6 @@ def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
   packed = mx_pack(e8) if len(batch) == 1 and scale_K % 4 == 0 else None
   return x_clamped.cast(FP8_DTYPE), e8, packed
 
-def _mxfp4_shuffle_weight(x:Tensor) -> Tensor:
-  # shuffle_weight(x, layout=(16, 16)) on the packed uint8 buffer.
-  if x.ndim == 3:
-    ndev, rows, half_k = x.shape
-    return x.reshape(ndev, rows//16, 16, half_k//32, 2, 16).permute(0, 1, 3, 4, 2, 5).reshape(ndev, rows, half_k).contiguous()
-  rows, half_k = x.shape
-  return x.reshape(rows//16, 16, half_k//32, 2, 16).permute(0, 2, 3, 1, 4).reshape(rows, half_k).contiguous()
-
-def _mxfp4_shuffle_scales(x:Tensor) -> Tensor:
-  # e8m0_shuffle: each 256x8 scale tile is arranged for the raw MFMA scale loads.
-  if x.ndim == 3:
-    ndev, rows, scale_k = x.shape
-    return x.reshape(ndev, rows//32, 2, 16, scale_k//8, 2, 4).permute(0, 1, 4, 6, 3, 5, 2).reshape(ndev, rows, scale_k).contiguous()
-  rows, scale_k = x.shape
-  return x.reshape(rows//32, 2, 16, scale_k//8, 2, 4).permute(0, 3, 5, 2, 4, 1).reshape(rows, scale_k).contiguous()
-
-def quantize_mxfp4(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
-  # OCP MXFP4: 1x32 blocks, E2M1 values packed low-nibble first, and E8M0 scales.
-  *batch, K = x.shape
-  rows = math.prod(batch)
-  assert x.ndim >= 2 and K % 256 == 0 and rows % 32 == 0, \
-    f"mxfp4 quantization needs rows%32 and K%256, got {x.shape}"
-  xb = x.float().reshape(*batch, K//32, 32)
-  amax = xb.abs().max(axis=-1)
-
-  # even scale rounding: round the fp32 significand before choosing 2^(floor(log2)-2).
-  amax_rounded = ((amax.bitcast(dtypes.uint32) + 0x200000) & 0xFF800000).bitcast(dtypes.float32)
-  scale_exp = (amax_rounded.maximum(2**-126).log2().floor() - 2).clamp(-127, 127)
-  e8 = (scale_exp + 127).cast(dtypes.uint8)
-  scaled = xb * (-scale_exp).exp2().reshape(*batch, K//32, 1)
-
-  mag = scaled.abs()
-  code = sum(x.cast(dtypes.uint8) for x in
-             (mag > .25, mag >= .75, mag > 1.25, mag >= 1.75, mag > 2.5, mag >= 3.5, mag > 5.0))
-  code = code | ((scaled < 0).cast(dtypes.uint8) << 3)
-  code = code.reshape(*batch, K)
-  packed = code[..., 0::2] | (code[..., 1::2] << 4)
-  if isinstance(x.device, tuple) and x.uop.axis == x.ndim-2 and x.shape[x.uop.axis] == len(x.device):
-    axis = x.uop.axis
-    order = (axis, *range(axis), *range(axis+1, e8.ndim))
-    e8_local = e8.permute(order)
-    return packed, e8, _mxfp4_shuffle_scales(e8_local.reshape(e8_local.shape[0], -1, K//32))
-  return packed, e8, _mxfp4_shuffle_scales(e8.reshape(rows, K//32))
-
 def mx_pack(e8:Tensor) -> Tensor:
   rows, scale_K = e8.shape
   return e8.reshape(rows, scale_K // 4, 4).bitcast(dtypes.uint32).reshape(rows, scale_K // 4).permute(1, 0).contiguous()
@@ -405,15 +382,16 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F
 # ** mxfp4 gemm backward
 
 def custom_mxfp4_gemm_bw(gradient:UOp, kernel:UOp):
-  # The raw kernel consumes quantized buffers, while the final two inputs retain the BF16 operands for STE gradients.
-  inputs = kernel.src[1:]  # (out, a_q, b_q, scale_a, scale_b, a, w)
-  assert len(inputs) == 7
+  inputs = kernel.src[1:]  # out, row operands/scales, BF16 operands, column operands/scales
+  assert len(inputs) == 11
   a, w = Tensor(inputs[5], device=inputs[5].device), Tensor(inputs[6], device=inputs[6].device)
+  a_col, scale_a_col = Tensor(inputs[7], device=a.device), Tensor(inputs[8], device=a.device)
+  w_col, scale_w_col = Tensor(inputs[9], device=a.device), Tensor(inputs[10], device=a.device)
   g = Tensor(gradient, device=a.device)[:a.shape[0]].cast(dtypes.bfloat16)
-  grad_a = asm_gemm(g, w, mxfp4=True)
-  a_flat, g_flat = a.reshape(-1, a.shape[-1]), g.reshape(-1, g.shape[-1])
-  grad_w = asm_gemm(g_flat.T, a_flat, mxfp4=True)
-  return (None, None, None, None, None, grad_a.uop, grad_w.uop)
+  g_row, scale_g_row, g_col, scale_g_col = quantize_mxfp4(g, flatten_row=True)
+  grad_a = _mxfp4_gemm_quantized(g_row, w_col, scale_g_row, scale_w_col).reshape(*a.shape[:-1], w.shape[-1])
+  grad_w = _mxfp4_gemm_quantized(g_col, a_col, scale_g_col, scale_a_col).reshape(w.shape)
+  return (None, None, None, None, None, grad_a.uop, grad_w.uop, None, None, None, None)
 
 # ** main gemm function
 
@@ -459,16 +437,10 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
       tile_m, tile_n = next((tm, tn) for tm, tn in ((256, 256), (192, 256), (128, 512)) if (batch*M) % tm == N % tn == 0)
       fxn = functools.partial(custom_mxfp4_gemm, tile_m=tile_m, tile_n=tile_n)
       w = b.T
-      if k_sharded:
-        ndev = len(a.device)
-        a_q, _, scale_a = quantize_mxfp4(a.reshape(batch, M, ndev, K))
-        b_q, _, scale_b = quantize_mxfp4(w.reshape(w.shape[0], ndev, K))
-        b_q = _mxfp4_shuffle_weight(b_q.permute(1, 0, 2))
-      else:
-        a_q, _, scale_a = quantize_mxfp4(a.reshape(batch*M, K))
-        b_q, _, scale_b = quantize_mxfp4(w)
-        a_q, b_q = a_q.reshape(batch, M, K//2).contiguous(), _mxfp4_shuffle_weight(b_q)
-      out = Tensor.custom_kernel(out, a_q, b_q, scale_a, scale_b, a, w, fxn=fxn, grad_fxn=custom_mxfp4_gemm_bw)[0]
+      a_q, scale_a, a_col, scale_a_col = quantize_mxfp4(a, shuffle_col=True)
+      b_q, scale_b, b_col, scale_b_col = quantize_mxfp4(w, shuffle_row=True, shuffle_col=True)
+      out = Tensor.custom_kernel(out, a_q, b_q, scale_a, scale_b, a, w,
+                                 a_col, scale_a_col, b_col, scale_b_col, fxn=fxn, grad_fxn=custom_mxfp4_gemm_bw)[0]
     elif mx:
       # mxfp8 1x32 block scaling
       if mx_scales is not None:

extra/gemm/moe_routing.py

@@ -5,7 +5,8 @@ BLOCK_ROW = 256
 
 def _sharded_invalids(shape:tuple[int, ...], dtype, device) -> Tensor:
   if isinstance(device, tuple):
-    return Tensor.invalids(*shape, dtype=dtype, device=device[0]).shard(device, axis=0)
+    per = Tensor.invalids(shape[0]//len(device), *shape[1:], dtype=dtype, device=device)
+    return Tensor(per.uop.unshard(0), device=device)
   return Tensor.invalids(*shape, dtype=dtype, device=device)
 
 def _atomic_add(device:str) -> str:

extra/gemm/triton_nv_matmul.py

@@ -79,7 +79,7 @@ if __name__ == "__main__":
   linear, var_vals = C.linear_with_vars()
   last_call = linear.src[-1]
   ast = last_call.src[0]
-  bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]
+  bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]
 
   src = compiled.asm["ptx"]
   # specify the shared memory here so we don't need to do it dynamically

extra/hcq2/ops_amd2.py

@@ -3,7 +3,7 @@ from typing import cast, Any, Callable
 import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit
 assert sys.platform != 'win32'
 from dataclasses import dataclass
-from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_cmdbuf
+from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, encode_kernargs_clike, make_cmdbuf
 from tinygrad.runtime.support.hcq2 import make_binary_patch
 from tinygrad.uop.ops import sint, UOp
 from tinygrad.device import Compiled, BufferSpec, Buffer, Device
@@ -288,14 +288,16 @@ def amd_build_program(prg:UOp) -> UOp:
 
 class AMDAllocator(HCQAllocator['AMDDevice']):
   def __init__(self, dev:AMDDevice):
-    super().__init__(dev, supports_copy_from_disk=dev.has_sdma_queue, supports_transfer=dev.has_sdma_queue and not dev.is_usb())
+    super().__init__(dev, supports_copy_from_disk=dev.has_copy_queue, supports_transfer=dev.has_copy_queue and not dev.is_usb())
 
-  def _alloc(self, size:int, options:BufferSpec) -> HCQ2Buffer:
-    return self.dev.iface.alloc(size, host=options.host, uncached=options.uncached, cpu_access=options.cpu_access or not self.dev.has_sdma_queue)
+  def _alloc(self, size:int, options:BufferSpec) -> HCQBuffer:
+    return self.dev.iface.alloc(size, host=options.host, uncached=options.uncached, cpu_access=options.cpu_access or not self.dev.has_copy_queue)
 
   def _do_free(self, opaque, options:BufferSpec): self.dev.iface.free(opaque)
 
-  def _do_map(self, buf:HCQ2Buffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)
+  def _do_map(self, buf:HCQBuffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)
+
+  def _do_unmap(self, buf:HCQBuffer): self.dev.iface.unmap(buf)
 
 @dataclass
 class AMDQueueDesc:
@@ -388,15 +390,24 @@ class KFDIface:
     return hcqbuf
 
   def free(self, mem):
+    self._unmap(mem)
+    if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
+    kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
+
+  def unmap(self, mem):
+    self._unmap(mem)
+    if getattr(mem, '_owns_kfd_handle', False): kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
+
+  def _unmap(self, mem):
     gpus = (ctypes.c_int32 * 1)(self.gpu_id)
     stm = kfd.AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(gpus), n_devices=1)
     assert stm.n_success == 1
-    if mem.owner == self.dev:
-      if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
-      kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
 
   def map(self, mem):
-    if mem.owner is not None and mem.owner._is_cpu(): return self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)
+    if mem.owner is not None and mem.owner._is_cpu():
+      mapped = self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)
+      mapped._owns_kfd_handle = True
+      return mapped
 
     c_gpus = (ctypes.c_int32 * 1)(self.gpu_id)
     stm = kfd.AMDKFD_IOC_MAP_MEMORY_TO_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(c_gpus), n_devices=1)
@@ -468,6 +479,7 @@ class PCIIface(PCIIfaceBase):
 
   def require_profile_mode(self): return True
   def is_wgp_active(self, xcc, se, sa, wgp) -> bool: return True # TODO: account for WGP disablement on some asics.
+  def unmap(self, mem): self.free(mem)
 
   def _compute_props(self):
     self.ip_versions = self.dev_impl.ip_ver
@@ -539,9 +551,12 @@ class AMDDevice(HCQ2Compiled):
   ])
 
   timestamp_divider = 100.0  # AMD GPU clock: ticks/us
+  max_scratch_psize = 0
 
   ifaces = [KFDIface, PCIIface, _mock(KFDIface, "MOCKIface"), _mock(KFDIface), _mock(PCIIface)]
 
+  def device_props(self): return self.iface.props
+
   def is_am(self) -> bool: return isinstance(self.iface, (PCIIface,))
   def is_usb(self) -> bool: return False
 
@@ -578,7 +593,7 @@ class AMDDevice(HCQ2Compiled):
 
     self.max_copy_size = 0x40000000 if self.iface.ip_versions[am.SDMA0_HWIP][0] >= 5 else 0x400000
     self.sdma_queues:dict = {}
-    self.has_sdma_queue = True # self.sdma_queue(0) is not None, TODO: think of this
+    self.has_copy_queue = not getenv("AMD_DISABLE_SDMA")
 
     super().__init__(device, AMDAllocator(self), [HIPRenderer, AMDLLVMRenderer, HIPCCRenderer], None, can_recover=self.is_am(), arch=self.arch)
 
@@ -691,7 +706,7 @@ class AMDDevice(HCQ2Compiled):
     return tmpring
 
   def scratch_buffer(self, private_segment_size):
-    private_segment_size = max(private_segment_size, 128)
+    AMDDevice.max_scratch_psize = private_segment_size = max(private_segment_size, 128, AMDDevice.max_scratch_psize)
     if self.max_private_segment_size < private_segment_size:
       lanes_per_wave = 64 # wave64
       mem_alignment_size = 256 if self.target[0] != 9 else 1024

extra/llama_kernels/quantize_mxfp4/__init__.py

@@ -0,0 +1,36 @@
+import functools, math, pathlib
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import alloc_like, compile_hip
+
+@functools.cache
+def _custom_quantize_mxfp4(row_fp4:UOp, row_scale:UOp, col_fp4:UOp, col_scale:UOp, x:UOp, *, shuffle_row:bool, shuffle_col:bool) -> UOp:
+  M, N = math.prod(x.shape[:-1]), x.shape[-1]
+  assert M % 256 == 0 and N % 256 == 0, f"MXFP4 quantization requires multiples of 256, got {x.shape}"
+  name = f"quantize_mxfp4_{int(shuffle_row)}_{int(shuffle_col)}_{M}_{N}"
+  mem = M*N*2 + M*N + M*N//16  # read bf16, write row+col fp4 + e8m0
+  outputs = (row_fp4, row_scale, col_fp4, col_scale)
+  sink = UOp.sink(*(o.base for o in outputs), x.base,
+                  *(UOp(Ops.CUSTOM, dtypes.void, (o.base.index(0),), arg="") for o in outputs),
+                  UOp.special(256, "lidx0"), UOp.special(M//128, "gidx0"), UOp.special(N//64, "gidx1"),
+                  arg=KernelInfo(name, estimates=Estimates(ops=12*M*N, mem=mem)))
+  src = (pathlib.Path(__file__).parent/"quantize_mxfp4.cpp").read_text()
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
+    UOp(Ops.BINARY, arg=compile_hip(src, [f"-DKERNEL_NAME={name}", f"-DM_DIM={M}", f"-DN_DIM={N}",
+                                             f"-DSHUFFLE_ROWWISE_FP4_VALUE={int(shuffle_row)}",
+                                             f"-DSHUFFLE_COLWISE_FP4_VALUE={int(shuffle_col)}"]))))
+
+def quantize_mxfp4(x:Tensor, *, shuffle_row:bool=False, shuffle_col:bool=False, flatten_row:bool=False) -> tuple[Tensor, Tensor, Tensor, Tensor]:
+  assert x.dtype == dtypes.bfloat16 and x.ndim >= 2, f"expected BF16 matrix, got {x.dtype} {x.shape}"
+  M, N = math.prod(x.shape[:-1]), x.shape[-1]
+  assert M % 256 == 0 and N % 256 == 0, f"MXFP4 quantization requires multiples of 256, got {x.shape}"
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  row_axis = 0 if flatten_row and axis is not None else axis
+  col_axis = None if axis is None else (0 if axis == x.ndim-1 else 1)
+  outputs = (alloc_like((M, N//2) if flatten_row else (*x.shape[:-1], N//2), dtypes.uint8, x.device, row_axis),
+             alloc_like((M, N//32) if flatten_row else (*x.shape[:-1], N//32), dtypes.uint8, x.device, row_axis),
+             alloc_like((N, M//2), dtypes.uint8, x.device, col_axis),
+             alloc_like((N, M//32), dtypes.uint8, x.device, col_axis))
+  fxn = functools.partial(_custom_quantize_mxfp4, shuffle_row=shuffle_row, shuffle_col=shuffle_col)
+  return tuple(Tensor.custom_kernel(*outputs, x, fxn=fxn)[:4])

extra/llama_kernels/quantize_mxfp4/quantize_mxfp4.cpp

@@ -0,0 +1,226 @@
+// Copyright (c) 2025-2026, Advanced Micro Devices, Inc. All rights reserved.
+// SPDX-License-Identifier: MIT
+
+#include <hip/hip_runtime.h>
+#include <cstdint>
+
+#if !defined(KERNEL_NAME) || !defined(M_DIM) || !defined(N_DIM) || !defined(SHUFFLE_ROWWISE_FP4_VALUE) || \
+    !defined(SHUFFLE_COLWISE_FP4_VALUE)
+#error kernel dimensions and layouts must be defined
+#endif
+
+namespace {
+
+constexpr int BLOCK = 32;
+constexpr int TILE_M = 128;
+constexpr int TILE_N = 64;
+constexpr int THREADS = 256;
+constexpr int THREADS_PER_ROW = 8;
+constexpr int VALUES_PER_THREAD = 4;
+constexpr int SMEM_STRIDE = BLOCK + 2;
+constexpr int M = M_DIM;
+constexpr int N = N_DIM;
+constexpr int M_PACKED = M / 2;
+constexpr int N_PACKED = N / 2;
+constexpr int M_SCALES = M / BLOCK;
+constexpr int N_SCALES = N / BLOCK;
+constexpr bool SHUFFLE_ROWWISE_FP4 = SHUFFLE_ROWWISE_FP4_VALUE;
+constexpr bool SHUFFLE_COLWISE_FP4 = SHUFFLE_COLWISE_FP4_VALUE;
+
+static_assert(M % 256 == 0 && N % 256 == 0);
+
+struct Quantized4 {
+  uint16_t fp4;
+  uint8_t scale;
+};
+
+__device__ __forceinline__ float swizzle_xor1(float value) {
+  float result;
+  asm volatile("ds_swizzle_b32 %0, %1 offset:0x041f\n\ts_waitcnt lgkmcnt(0)" : "=v"(result) : "v"(value));
+  return result;
+}
+
+__device__ __forceinline__ float swizzle_xor2(float value) {
+  float result;
+  asm volatile("ds_swizzle_b32 %0, %1 offset:0x081f\n\ts_waitcnt lgkmcnt(0)" : "=v"(result) : "v"(value));
+  return result;
+}
+
+__device__ __forceinline__ float swizzle_xor4(float value) {
+  float result;
+  asm volatile("ds_swizzle_b32 %0, %1 offset:0x101f\n\ts_waitcnt lgkmcnt(0)" : "=v"(result) : "v"(value));
+  return result;
+}
+
+__device__ __forceinline__ float max8(float value) {
+  value = fmaxf(value, swizzle_xor4(value));
+  value = fmaxf(value, swizzle_xor2(value));
+  return fmaxf(value, swizzle_xor1(value));
+}
+
+__device__ __forceinline__ float4 load_bf16x4(const uint16_t* values) {
+  const uint32_t lo = *reinterpret_cast<const uint32_t*>(values);
+  const uint32_t hi = *reinterpret_cast<const uint32_t*>(values + 2);
+  return make_float4(__uint_as_float(lo << 16), __uint_as_float(lo & 0xffff0000u),
+                     __uint_as_float(hi << 16), __uint_as_float(hi & 0xffff0000u));
+}
+
+__device__ __forceinline__ void hadamard16(float4& value, int lane) {
+  const float a0 = value.x + value.y, a1 = value.x - value.y;
+  const float a2 = value.z + value.w, a3 = value.z - value.w;
+  value = make_float4(a0 + a2, a1 + a3, a0 - a2, a1 - a3);
+
+  const float4 xor1 = make_float4(swizzle_xor1(value.x), swizzle_xor1(value.y), swizzle_xor1(value.z), swizzle_xor1(value.w));
+  value = lane & 1 ? make_float4(xor1.x - value.x, xor1.y - value.y, xor1.z - value.z, xor1.w - value.w)
+                   : make_float4(xor1.x + value.x, xor1.y + value.y, xor1.z + value.z, xor1.w + value.w);
+
+  const float4 xor2 = make_float4(swizzle_xor2(value.x), swizzle_xor2(value.y), swizzle_xor2(value.z), swizzle_xor2(value.w));
+  value = lane & 2 ? make_float4(xor2.x - value.x, xor2.y - value.y, xor2.z - value.z, xor2.w - value.w)
+                   : make_float4(xor2.x + value.x, xor2.y + value.y, xor2.z + value.z, xor2.w + value.w);
+  value.x *= 0.25f;
+  value.y *= 0.25f;
+  value.z *= 0.25f;
+  value.w *= 0.25f;
+}
+
+__device__ __forceinline__ uint8_t e8m0_scale(float amax, float& scale) {
+  if (amax == 0.0f) {
+    scale = 1.0f;
+    return 127;
+  }
+
+  const uint32_t rounded = (__float_as_uint(amax) + 0x200000u) & 0xff800000u;
+  int exponent = static_cast<int>((rounded >> 23) & 0xff) - 129;
+  exponent = exponent < -127 ? -127 : exponent > 127 ? 127 : exponent;
+  scale = exponent == -127 ? __uint_as_float(0x00400000u) : __uint_as_float(static_cast<uint32_t>(exponent + 127) << 23);
+  return static_cast<uint8_t>(exponent + 127);
+}
+
+__device__ __forceinline__ uint16_t pack_fp4(float4 value, float scale) {
+  uint32_t lo = 0, hi = 0;
+  asm volatile("v_cvt_scalef32_pk_fp4_f32 %0, %1, %2, %3" : "+v"(lo) : "v"(value.x), "v"(value.y), "v"(scale));
+  asm volatile("v_cvt_scalef32_pk_fp4_f32 %0, %1, %2, %3" : "+v"(hi) : "v"(value.z), "v"(value.w), "v"(scale));
+  return static_cast<uint16_t>(lo | (hi << 8));
+}
+
+__device__ __forceinline__ Quantized4 quantize(float4 value, int lane) {
+  hadamard16(value, lane);
+  const float local_max = fmaxf(fmaxf(fabsf(value.x), fabsf(value.y)), fmaxf(fabsf(value.z), fabsf(value.w)));
+  float scale;
+  const uint8_t e8m0 = e8m0_scale(max8(local_max), scale);
+  return {pack_fp4(value, scale), e8m0};
+}
+
+__device__ __forceinline__ void store_scale(uint8_t* output, int row, int col, int cols, uint8_t value) {
+  const int tile = ((row >> 5) * (cols >> 3) + (col >> 3)) << 8;
+  const int offset = ((col & 3) << 6) + ((row & 15) << 2) + (((col >> 2) & 1) << 1) + ((row >> 4) & 1);
+  output[tile + offset] = value;
+}
+
+template<bool Shuffled>
+__device__ __forceinline__ void store_fp4(uint8_t* output, int row, int col, int packed_cols, uint16_t value) {
+  int index = row * packed_cols + col;
+  if constexpr (Shuffled) {
+    const int tile = (row >> 4) * (packed_cols << 4) + (col >> 5) * 512;
+    const int offset = ((col >> 4) & 1) * 256 + (row & 15) * 16 + (col & 15);
+    index = tile + offset;
+  }
+  *reinterpret_cast<uint16_t*>(output + index) = value;
+}
+
+__device__ __forceinline__ void load_tile(uint16_t* tile, const uint16_t* input, int tile_m, int tile_n) {
+  const int row = threadIdx.x / THREADS_PER_ROW;
+  const int col = threadIdx.x % THREADS_PER_ROW * VALUES_PER_THREAD;
+  const uint64_t packed = *reinterpret_cast<const uint64_t*>(input + (tile_m + row) * N + tile_n + col);
+  *reinterpret_cast<uint32_t*>(tile + row * SMEM_STRIDE + col) = static_cast<uint32_t>(packed);
+  *reinterpret_cast<uint32_t*>(tile + row * SMEM_STRIDE + col + 2) = static_cast<uint32_t>(packed >> 32);
+}
+
+__device__ __forceinline__ void quantize_row(uint16_t* tile, uint8_t* fp4_output, uint8_t* scale_output,
+                                             int tile_m, int tile_n, int local_row, int lane) {
+  const int row = tile_m + local_row;
+  const int col = lane * VALUES_PER_THREAD;
+  const Quantized4 result = quantize(load_bf16x4(tile + local_row * SMEM_STRIDE + col), lane);
+  store_fp4<SHUFFLE_ROWWISE_FP4>(fp4_output, row, (tile_n + col) / 2, N_PACKED, result.fp4);
+  if (lane == 0) store_scale(scale_output, row, tile_n / BLOCK, N_SCALES, result.scale);
+}
+
+__device__ __forceinline__ Quantized4 quantize_col(uint16_t* tile, int col, int lane) {
+  const int row = lane * VALUES_PER_THREAD;
+  return quantize(make_float4(
+    __uint_as_float(static_cast<uint32_t>(tile[(row + 0) * SMEM_STRIDE + col]) << 16),
+    __uint_as_float(static_cast<uint32_t>(tile[(row + 1) * SMEM_STRIDE + col]) << 16),
+    __uint_as_float(static_cast<uint32_t>(tile[(row + 2) * SMEM_STRIDE + col]) << 16),
+    __uint_as_float(static_cast<uint32_t>(tile[(row + 3) * SMEM_STRIDE + col]) << 16)), lane);
+}
+
+} // namespace
+
+extern "C" __global__ __launch_bounds__(THREADS, 8)
+void KERNEL_NAME(uint8_t* __restrict__ rowwise_fp4, uint8_t* __restrict__ rowwise_scale,
+                 uint8_t* __restrict__ colwise_fp4, uint8_t* __restrict__ colwise_scale,
+                 const uint16_t* __restrict__ input) {
+  __shared__ uint16_t tile[BLOCK * SMEM_STRIDE];
+  const int tid = threadIdx.x;
+  const int line = tid / THREADS_PER_ROW;
+  const int lane = tid % THREADS_PER_ROW;
+  const int block_m = blockIdx.x * TILE_M;
+  const int block_n = blockIdx.y * TILE_N;
+
+  if constexpr (!SHUFFLE_COLWISE_FP4) {
+    uint16_t col_fp4[TILE_N / BLOCK][TILE_M / BLOCK];
+    uint8_t col_scale[TILE_N / BLOCK][TILE_M / BLOCK];
+
+    for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++) {
+      for (int chunk_n = 0; chunk_n < TILE_N / BLOCK; chunk_n++) {
+        const int tile_m = block_m + chunk_m * BLOCK;
+        const int tile_n = block_n + chunk_n * BLOCK;
+        load_tile(tile, input, tile_m, tile_n);
+        __syncthreads();
+
+        quantize_row(tile, rowwise_fp4, rowwise_scale, tile_m, tile_n, line, lane);
+        const Quantized4 result = quantize_col(tile, line, lane);
+        col_fp4[chunk_n][chunk_m] = result.fp4;
+        col_scale[chunk_n][chunk_m] = result.scale;
+        __syncthreads();
+      }
+    }
+
+    for (int chunk_n = 0; chunk_n < TILE_N / BLOCK; chunk_n++) {
+      for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++)
+        tile[line * BLOCK + chunk_m * THREADS_PER_ROW + lane] = col_fp4[chunk_n][chunk_m];
+      __syncthreads();
+
+      for (int round = 0; round < BLOCK / THREADS_PER_ROW; round++) {
+        const int col = round * THREADS_PER_ROW + tid / BLOCK;
+        const int row_pair = tid % BLOCK;
+        *reinterpret_cast<uint16_t*>(colwise_fp4 + (block_n + chunk_n * BLOCK + col) * M_PACKED + block_m / 2 + row_pair * 2) =
+          tile[col * BLOCK + row_pair];
+      }
+
+      if (lane == 0) {
+        const int col = block_n + chunk_n * BLOCK + line;
+        for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++)
+          store_scale(colwise_scale, col, block_m / BLOCK + chunk_m, M_SCALES, col_scale[chunk_n][chunk_m]);
+      }
+      __syncthreads();
+    }
+  } else {
+    for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++) {
+      for (int chunk_n = 0; chunk_n < TILE_N / BLOCK; chunk_n++) {
+        const int tile_m = block_m + chunk_m * BLOCK;
+        const int tile_n = block_n + chunk_n * BLOCK;
+        load_tile(tile, input, tile_m, tile_n);
+        __syncthreads();
+
+        quantize_row(tile, rowwise_fp4, rowwise_scale, tile_m, tile_n, line, lane);
+        const int row = lane * VALUES_PER_THREAD;
+        const int col = tile_n + line;
+        const Quantized4 result = quantize_col(tile, line, lane);
+        store_fp4<true>(colwise_fp4, col, (tile_m + row) / 2, M_PACKED, result.fp4);
+        if (lane == 0) store_scale(colwise_scale, col, tile_m / BLOCK, M_SCALES, result.scale);
+        __syncthreads();
+      }
+    }
+  }
+}

extra/llama_kernels/swiglu/__init__.py

@@ -0,0 +1,49 @@
+import functools, math
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import alloc_like
+
+LOG2E = 1.4426950408889634
+
+@functools.cache
+def _custom_swiglu(out:UOp, x_w13:UOp) -> UOp:
+  rows, hidden = math.prod(x_w13.shape[:-1]), x_w13.shape[-1]//2
+  n_elems = rows * hidden
+  out, x_w13 = out.reshape(n_elems), x_w13.reshape(rows, 2*hidden)
+  i = UOp.range(n_elems, 0)
+  row, col = i // hidden, i % hidden
+  act, gate = x_w13[row, col].cast(dtypes.float), x_w13[row, hidden+col].cast(dtypes.float)
+  sigmoid = (1.0 + (-LOG2E * act).exp2()).reciprocal()
+  store = out[i].store((act * sigmoid * gate).cast(out.dtype))
+  return store.end(i).sink(arg=KernelInfo(f"swiglu_fwd_{n_elems}", estimates=Estimates(ops=5*n_elems, mem=6*n_elems)))
+
+@functools.cache
+def _custom_swiglu_bwd(grad_out:UOp, x_w13:UOp, grad_act:UOp) -> UOp:
+  rows, hidden = math.prod(x_w13.shape[:-1]), x_w13.shape[-1]//2
+  n_elems = rows * hidden
+  grad_out, x_w13, grad_act = grad_out.reshape(rows, 2*hidden), x_w13.reshape(rows, 2*hidden), grad_act.reshape(n_elems)
+  i = UOp.range(n_elems, 0)
+  row, col = i // hidden, i % hidden
+  act, gate = x_w13[row, col].cast(dtypes.float), x_w13[row, hidden+col].cast(dtypes.float)
+  grad = grad_act[i].cast(dtypes.float)
+  sigmoid = (1.0 + (-LOG2E * act).exp2()).reciprocal()
+  silu = act * sigmoid
+  dact = grad_out[row, col].store((grad * (sigmoid + silu * (1.0 - sigmoid)) * gate).cast(grad_out.dtype))
+  dgate = grad_out.after(dact)[row, hidden+col].store((grad * silu).cast(grad_out.dtype))
+  return dgate.end(i).sink(arg=KernelInfo(f"swiglu_bwd_{n_elems}", estimates=Estimates(ops=10*n_elems, mem=10*n_elems)))
+
+def _swiglu_bwd(gradient:UOp, kernel:UOp):
+  _, x_w13 = kernel.src[1:]
+  axis = x_w13.axis if isinstance(x_w13.device, tuple) else None
+  grad_out = alloc_like(x_w13.shape, dtypes.bfloat16, x_w13.device, axis)
+  grad_out, *_ = Tensor.custom_kernel(grad_out, Tensor(x_w13, device=x_w13.device), Tensor(gradient, device=x_w13.device),
+                                      fxn=_custom_swiglu_bwd)
+  return (None, grad_out.uop)
+
+def swiglu(x_w13:Tensor) -> Tensor:
+  assert x_w13.dtype == dtypes.bfloat16 and x_w13.ndim >= 2 and x_w13.shape[-1] % 32 == 0
+  *prefix, two_k = x_w13.shape
+  axis = x_w13.uop.axis if isinstance(x_w13.device, tuple) else None
+  out = alloc_like((*prefix, two_k//2), dtypes.bfloat16, x_w13.device, axis)
+  return Tensor.custom_kernel(out, x_w13, fxn=_custom_swiglu, grad_fxn=_swiglu_bwd)[0]

extra/runbook_digitalocean_mi350x.md

@@ -0,0 +1,217 @@
+# Runbook: Llama 3 8B Training on DigitalOcean MI350X
+
+## Machine Specs
+- 8x MI350X GPUs (gfx950, device ID 75b0), 288GB VRAM each
+- 2TB RAM, 192 CPUs, 2TB disk
+- ROCm 7.14 at `/opt/rocm` (NOT `/opt/rocm-7.1.1` like the submission scripts assume)
+- Python 3.12
+
+## Phase 1: System Setup
+
+### 1.1 Install packages
+```bash
+apt-get update
+apt-get install -y python3-pip python3-venv git tmux rclone clang
+```
+
+### 1.2 Install Python deps
+```bash
+python3 -m pip install --break-system-packages --ignore-installed typing-extensions numpy tqdm wandb tiktoken sentencepiece
+```
+Note: `--ignore-installed typing-extensions` is needed because the base image ships typing-extensions 4.10.0 without a RECORD file, so pip cannot uninstall it.
+
+### 1.3 Install ROCm dev headers
+The base image has ROCm runtime but NOT the HIP dev headers. Need:
+```bash
+apt-get install -y amdrocm-core-dev
+```
+This installs `hip/hip_runtime.h` at `/opt/rocm/core-7.14/include/hip/hip_runtime.h`.
+The symlink `/opt/rocm/include` → `/opt/rocm/core-7.14/include` makes it available at `/opt/rocm/include/hip/hip_runtime.h`.
+
+### 1.4 Configure ROCm comgr
+ROCm 7.14 ships comgr 3.3 at `/opt/rocm/lib/libamd_comgr.so`. tinygrad's DLL loader needs explicit env vars to find it (it searches for `libcomgr.so*` by default, not `libamd_comgr.so*`). Set these in the run command:
+```bash
+export COMGR_PATH=/opt/rocm/lib/libamd_comgr.so
+export COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so
+```
+Also add ROCm libs to ldconfig so comgr's shared library dependencies resolve:
+```bash
+cat > /etc/ld.so.conf.d/rocm.conf << 'EOF'
+/opt/rocm/lib
+/opt/rocm/lib/llvm/lib
+/opt/rocm/lib/rocm_sysdeps/lib
+EOF
+ldconfig
+```
+
+### 1.5 Install geohot tmux config
+```bash
+curl -sL https://raw.githubusercontent.com/geohot/configuration/master/.tmux.conf -o ~/.tmux.conf
+```
+
+### 1.6 Reload amdgpu driver
+tinygrad's HCQ backend needs `/dev/kfd` which is created by the amdgpu kernel driver.
+If the driver was unloaded, reload it:
+```bash
+modprobe amdgpu
+ls /dev/kfd  # should exist
+```
+
+## Phase 2: Clone tinygrad
+```bash
+cd /root
+git clone https://github.com/tinygrad/tinygrad.git
+cd tinygrad
+python3 -m pip install --break-system-packages -e .
+```
+
+## Phase 3: Download C4 Dataset
+
+The C4 data is on the MLCommons Cloudflare R2 bucket in Megatron-LM indexed format.
+
+```bash
+rclone config create mlc-training s3 provider=Cloudflare \
+  access_key_id=76ea42eadb867e854061a1806220ee1e \
+  secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 \
+  endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
+
+mkdir -p /raid/datasets/c4-8b
+rclone copy mlc-training:mlcommons-training-wg-public/llama3_1/datasets/c4/llama3_1_8b/ /raid/datasets/c4-8b/ -P
+```
+
+Files downloaded (~85GB total, ~6 minutes):
+- `c4-train.en_6_text_document.bin` (79 GB)
+- `c4-train.en_6_text_document.idx` (870 MB)
+- `c4-validation-91205-samples.en_text_document.bin` (159 MB)
+- `c4-validation-91205-samples.en_text_document.idx` (1.8 MB)
+- `LICENSE.txt`, `NOTICE.txt`
+
+**Wait for rclone to fully complete before starting training.** Starting training while the dataset is still downloading will read a truncated .bin file, causing `ValueError: all input arrays must have the same shape` in the dataloader. The stale `.index_cache` and `.blend_cache` files must also be deleted if this happens:
+```bash
+rm -f /raid/datasets/c4-8b/*.index_cache /raid/datasets/c4-8b/*.blend_cache
+```
+
+## Phase 4: wandb Login
+```bash
+wandb login
+```
+Enter API key from https://wandb.ai/authorize
+
+Alternatively, pass the key directly:
+```bash
+wandb login <API_KEY>
+```
+
+## Phase 5: Run Training
+
+Run training in tmux so it survives SSH disconnects:
+```bash
+tmux new-session -d -s train 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm WANDB=1 bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh 2>&1 | tee /root/train.log'
+```
+Attach with `tmux attach -t train`.
+
+### 5.1 Smoke test (beam search, 2 layers, real data)
+Always run beam first to validate the pipeline:
+```bash
+tmux new-session -d -s beam 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh 2>&1 | tee /root/beam.log'
+```
+
+The beam test runs 10 training steps with 2 layers. Expected results:
+- ~0.29s per step after warmup
+- ~700K GFLOPS, ~7% MFU (low because only 2 layers)
+- ~380 GB VRAM used
+- Loss stable at ~12.55 with random init
+
+### 5.2 Full training run
+```bash
+tmux new-session -d -s train 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm WANDB=1 bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh 2>&1 | tee /root/train.log'
+```
+
+## Environment Variable Reference
+
+| Variable | Value | Why |
+|---|---|---|
+| `COMGR_PATH` | `/opt/rocm/lib/libamd_comgr.so` | tinygrad's DLL loader needs explicit path to find comgr 3.3 |
+| `COMGR_3_PATH` | `/opt/rocm/lib/libamd_comgr.so` | comgr 3.x uses a separate `comgr_3` module with its own path var |
+| `CC` | `/opt/rocm/core-7.14/lib/llvm/bin/clang` | System clang doesn't know gfx950; must use ROCm's bundled clang |
+| `DEV` | `AMD:HIP` | Force HIPRenderer (comgr-based) over HIPCCRenderer (hipcc subprocess) |
+| `ROCM_PATH` | `/opt/rocm` | Script defaults to `/opt/rocm-7.1.1` which doesn't exist |
+| `WANDB` | `1` | Enable wandb logging (off by default) |
+
+## Architecture
+
+| Component | Source file |
+|---|---|
+| Model | `examples/mlperf/models/flat_llama.py` — FlatTransformer, FP8 MXFP4 weights, fused QKV, flash attention |
+| Trainer | `examples/mlperf/model_train.py` → `train_llama3()` |
+| Optimizer | `examples/mlperf/optim.py` — GradAccClipAdamW, master weights, FP8 re-quant |
+| LR schedule | `examples/mlperf/lr_schedulers.py` — CosineAnnealingLRWithWarmup |
+| Dataloader | `examples/mlperf/dataloader.py` — Megatron-LM indexed bin format |
+| ASM GEMM | `extra/gemm/cdna_asm_gemm.py` — gfx950 MFMA assembly, MXFP4 |
+| Flash attention | `extra/thunder/amd/fa.py` |
+| Fused kernels | `extra/llama_kernels/` — rmsnorm, silu, quantize, fused_ce |
+| GPU driver | `tinygrad/runtime/ops_amd.py` — HCQ, direct KFD ioctl |
+| Renderer | `tinygrad/renderer/cstyle.py` — HIPRenderer for gfx950 |
+| comgr compiler | `tinygrad/runtime/support/compiler_amd.py` — HIPCompiler using comgr 3.3 |
+
+## Troubleshooting
+
+### `'hip/hip_runtime.h' file not found`
+Install `amdrocm-core-dev`:
+```bash
+apt-get install -y amdrocm-core-dev
+```
+
+### `'gfx950' is not a recognized processor` + LLVM crash
+System clang doesn't know gfx950. Set `CC=/opt/rocm/core-7.14/lib/llvm/bin/clang`.
+
+### `comgr not available: try setting COMGR_PATH?`
+Add ROCm libs to ldconfig and set `COMGR_PATH` and `COMGR_3_PATH`:
+```bash
+# /etc/ld.so.conf.d/rocm.conf should contain /opt/rocm/lib paths
+ldconfig
+```
+
+### `comgr not available: try setting COMGR_3_PATH?`
+comgr 3.x uses a separate module. Set `COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so` too.
+
+### `No such file or directory: 'clang'`
+Install clang: `apt-get install -y clang` (for CPU compilation).
+For gfx950 HIP compilation, comgr (not clang) is used — ensure the ROCm 7.14 comgr 3.3 is properly loaded via `COMGR_PATH` and `COMGR_3_PATH`.
+
+## Appendix: KVM Virtualization Observations
+
+### Virtualization detection
+```
+$ systemd-detect-virt
+kvm
+$ lspci -nn | grep AMD
+83:00.0 ... Device [1002:75b0]
+```
+CPU flags include `hypervisor`. `dmesg` shows `Hypervisor detected: KVM`.
+
+### Working path: amdgpu driver (KFDIface)
+The amdgpu driver loads on boot and binds to all 8 GPUs, creating `/dev/kfd` and 64 renderD nodes (`/dev/dri/renderD128` through `/dev/dri/renderD191`). tinygrad's `KFDIface` enumerates GPUs through `/sys/devices/virtual/kfd/kfd/topology/nodes` and uses `/dev/kfd` for ioctl. No PCI device ID patching is needed — the KFD path does not use `PCIIface` or `AMDev._run_discovery()`.
+
+This is the working configuration. No code changes to tinygrad are required.
+
+### PCIIface path (does not work on this VM)
+For reference, the `PCIIface` path was also explored but does not work in this KVM guest:
+
+- `PCIIface` in `ops_amd.py` does not list device ID `0x75b0`. Adding it allows PCI detection but `AMDev._run_discovery()` fails because the VRAM BAR reads all `0xFF`.
+- This was observed with the GPU unbound from any driver, after PCI reset, and with VFIO bound.
+- VFIO binding (`vfio-pci` with `enable_unsafe_noiommu_mode=1`) succeeded but VRAM BAR still reads all `0xFF`.
+- No IOMMU in guest — `dmesg` has no `AMD-Vi` entries, PCI devices have no `iommu_group` symlink.
+
+### amdgpu driver behavior
+On first boot, amdgpu loaded and bound to all 8 GPUs. On one boot it failed to initialize:
+```
+[  799.780369] amdgpu 0000:83:00.0: Failed to alloc msi vectors
+[  799.781476] amdgpu 0000:83:00.0: sw_init of IP block <vega20_ih> failed -22
+[  799.782724] amdgpu 0000:83:00.0: amdgpu_device_ip_init failed
+[  799.793885] amdgpu 0000:83:00.0: Fatal error during GPU init
+```
+On a subsequent boot, amdgpu initialized successfully (SMU initialized, VRAM ready). After unbinding all 8 GPUs from amdgpu, `rmmod amdgpu` wedged the module (stuck in "Unloading" state in `/proc/modules`), requiring a full VM reboot.
+
+### No fan control
+No `fan*` or `pwm*` hwmon entries exist. Only `temp*`, `power*`, `freq*` are exposed. GPU temps read 56-63°C, power ~265W per GPU.

extra/thunder/amd/fa.py

@@ -2,7 +2,7 @@ import math, pathlib, functools, struct
 
 from tinygrad import Device, Tensor
 from tinygrad.dtype import DTypeLike, dtypes
-from tinygrad.helpers import DEBUG
+from tinygrad.helpers import DEBUG, getenv
 from tinygrad.renderer import Estimates
 from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from tinygrad.runtime.support.elf import elf_loader
@@ -110,7 +110,49 @@ def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
   return _sharded_empty(ref.shape, ref, axis)
 
 @functools.cache
-def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink):
+def _windowed_lse(xq:Tensor, xk:Tensor, sinks, W:int) -> Tensor:
+  B, N, H, hd = xq.shape
+  H_KV = xk.shape[2]; R = H // H_KV; nb = N // W; sm = hd ** -0.5
+  q = xq.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
+  k = xk.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
+  k_prev = k.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
+  sc_d = (q @ k.transpose(-1, -2)) * sm
+  sc_p = (q @ k_prev.transpose(-1, -2)) * sm
+  li, lj = Tensor.arange(W).reshape(W, 1), Tensor.arange(W).reshape(1, W)
+  pv = (Tensor.arange(nb).reshape(nb, 1, 1) >= 1)
+  sc_d = (lj <= li).where(sc_d, -float("inf"))
+  sc_p = ((li < lj) & pv).where(sc_p, -float("inf"))
+  m = sc_d.max(-1, keepdim=True).maximum(sc_p.max(-1, keepdim=True))
+  if sinks is not None: m = m.maximum(sinks.reshape(1, H_KV, R, 1, 1, 1).float())
+  denom = (sc_d - m).exp().sum(-1, keepdim=True) + (sc_p - m).exp().sum(-1, keepdim=True)
+  if sinks is not None: denom = denom + (sinks.reshape(1, H_KV, R, 1, 1, 1).float() - m).exp()
+  return (m + denom.log()).reshape(B, H, N).unsqueeze(2)  # (B, H, 1, N), matches saved l_vec
+
+def _windowed_delta(xq:Tensor, xk:Tensor, xv:Tensor, do:Tensor, sinks, W:int) -> Tensor:
+  B, N, H, hd = xq.shape
+  H_KV = xk.shape[2]; R = H // H_KV; nb = N // W; sm = hd ** -0.5
+  q = xq.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
+  k = xk.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
+  v = xv.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
+  dob = do.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
+  k_prev = k.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
+  v_prev = v.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
+  sc_d = (q @ k.transpose(-1, -2)) * sm
+  sc_p = (q @ k_prev.transpose(-1, -2)) * sm
+  li, lj = Tensor.arange(W).reshape(W, 1), Tensor.arange(W).reshape(1, W)
+  pv = (Tensor.arange(nb).reshape(nb, 1, 1) >= 1)
+  sc_d = (lj <= li).where(sc_d, -float("inf"))
+  sc_p = ((li < lj) & pv).where(sc_p, -float("inf"))
+  m = sc_d.max(-1, keepdim=True).maximum(sc_p.max(-1, keepdim=True))
+  if sinks is not None: m = m.maximum(sinks.reshape(1, H_KV, R, 1, 1, 1).float())
+  e_d, e_p = (sc_d - m).exp(), (sc_p - m).exp()
+  denom = e_d.sum(-1, keepdim=True) + e_p.sum(-1, keepdim=True)
+  if sinks is not None: denom = denom + (sinks.reshape(1, H_KV, R, 1, 1, 1).float() - m).exp()
+  o = ((e_d / denom) @ v) + ((e_p / denom) @ v_prev)
+  delta = (dob * o).sum(-1)
+  return delta.reshape(B, H, N).unsqueeze(2)
+
+def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=0):
   def grad(dou:UOp, ker:UOp) -> tuple:
     do = Tensor(dou, device=dou.device)
     attn = Tensor(ker.src[1].after(ker), device=ker.src[1].device)
@@ -118,6 +160,8 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
     xq = Tensor(ker.src[3], device=ker.src[3].device)
     xk = Tensor(ker.src[4], device=ker.src[4].device)
     xv = Tensor(ker.src[5], device=ker.src[5].device)
+    if window:
+      l_vec = _windowed_lse(xq, xk, Tensor(ker.src[6], device=ker.src[6].device) if has_sink else None, window)
 
     dq = _sharded_empty((B, H, N, D), xq, axis=shard_axis_t)
     GROUP_SIZE = H_local // H_KV_local
@@ -128,8 +172,10 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
     # delta_vec = (do * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()
     delta_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
     delta_vec, dq = Tensor.custom_kernel(delta_vec, dq, attn, do, fxn=functools.partial(custom_fa_backward_pre, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:2]
+    if window:
+      delta_vec = _windowed_delta(xq, xk, xv, do, Tensor(ker.src[6], device=ker.src[6].device) if has_sink else None, window)
 
-    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:3]
+    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, window=window))[:3]
 
     if D == 64:
       dq = dq.reshape(B, H, N//16, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2).permute(0, 1, 2, 8, 9, 10, 11, 3, 4, 6, 7, 5, 12).reshape(B, H, N, D).transpose(1, 2)
@@ -149,7 +195,7 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
   return grad
 
 # TODO: remove write_flat once scheduler can remove reshapes between custom_kernel. TestCustomKernel.test_simple_reshape
-def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None):
+def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None, window:int=0):
   assert attn_mask is None, "attn_mask not supported"
   assert is_causal, "only causal attention supported"
 
@@ -176,18 +222,18 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
   attn = _sharded_empty((B, N, H * D), xq, axis=shard_axis) if write_flat else _sharded_empty_like(xq, axis=shard_axis)
   l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
 
-  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink)
+  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=window)
 
   fwd_inputs = (attn, l_vec, xq, xk, xv) + ((sinks,) if has_sink else ())
-  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink), grad_fxn=grad)[:2]
+  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink, window=window), grad_fxn=grad)[:2]
 
   return attn, attn, l_vec
 
 @functools.cache
-def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True):
+def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True, window:int=0):
   code = (pathlib.Path(__file__).parent / "fa_fwd_causal.cpp").read_text()
   compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}", f"-DWINDOW={window}"]
 
   Q_BLOCK_SIZE = 32
   NUM_WARPS = 8
@@ -206,10 +252,11 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None
                   arg=KernelInfo(name="custom_fa_forward", estimates=estimates))
 
   lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  lib = bytearray(lib)
-  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-  struct.pack_into('<I', lib, rodata_off, 160000)
-  lib = bytes(lib)
+  if not getenv("NO_HIPCC"):
+    lib = bytearray(lib)
+    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+    struct.pack_into('<I', lib, rodata_off, 160000)
+    lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
              src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
@@ -236,19 +283,20 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
                   arg=KernelInfo(name="custom_fa_backward_pre", estimates=estimates))
 
   lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  lib = bytearray(lib)
-  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-  struct.pack_into('<I', lib, rodata_off, 160000)
-  lib = bytes(lib)
+  if not getenv("NO_HIPCC"):
+    lib = bytearray(lib)
+    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+    struct.pack_into('<I', lib, rodata_off, 160000)
+    lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
              src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
 
 @functools.cache
-def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, window:int=0):
   code = (pathlib.Path(__file__).parent / "fa_bwd_causal.cpp").read_text()
   compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DWINDOW={window}"]
 
   BLOCK_SIZE_KV = 256
   GROUP_SIZE = H // H_KV
@@ -268,10 +316,11 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
                   arg=KernelInfo(name="custom_fa_backward", estimates=estimates))
 
   lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  lib = bytearray(lib)
-  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-  struct.pack_into('<I', lib, rodata_off, 160000)
-  lib = bytes(lib)
+  if not getenv("NO_HIPCC"):
+    lib = bytearray(lib)
+    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+    struct.pack_into('<I', lib, rodata_off, 160000)
+    lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
              src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
@@ -298,10 +347,11 @@ def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int,
                   arg=KernelInfo(name="custom_fa_backward_post", estimates=estimates))
 
   lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  lib = bytearray(lib)
-  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-  struct.pack_into('<I', lib, rodata_off, 160000)
-  lib = bytes(lib)
+  if not getenv("NO_HIPCC"):
+    lib = bytearray(lib)
+    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+    struct.pack_into('<I', lib, rodata_off, 160000)
+    lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
              src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

extra/thunder/amd/fa_bwd_causal.cpp

@@ -43,6 +43,10 @@ constexpr int SLICE_QO = 32;
 constexpr int DOT_SLICE_QO = 16;
 constexpr int WARP_SIZE_KV = 64; // warp size for KV
 constexpr bool causal = true;
+// WINDOW>0: sliding-window backward (query i sees keys in [i-WINDOW+1, i])
+#ifndef WINDOW
+#define WINDOW 0
+#endif
 
 #define NUM_WARPS 4
 #define NUM_THREADS (kittens::WARP_THREADS * NUM_WARPS)
@@ -88,7 +92,12 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
   const int k_start_min = j_min * WARP_SIZE_KV;
   // first Q step that can overlap this K_span:
   const int first_step = max(0, k_start_min / STEP_QO);
+#if WINDOW
+  // cap the Q loop, padded by 2 masked steps: the epilogue's deferred dq path miscomputes in-window tail queries
+  const int num_steps_per_head = min(total_steps_per_head - first_step, (BLOCK_SIZE_KV + WINDOW) / STEP_QO + 2);
+#else
   const int num_steps_per_head = total_steps_per_head - first_step;
+#endif
   const int num_steps = num_steps_per_head * HEADS_PER_WG;
   const int k_pos = j * WARP_SIZE_KV;
 
@@ -380,6 +389,13 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             mov<0, 1, neg_inf_v>(P_ij);
             mov<0, 2, neg_inf_v>(P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+          // window lower boundary, mirror of the causal edge
+          } else if (q_pos - k_pos == WINDOW) {
+            make_window<0, 0, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -640,6 +656,13 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             make_causal<0, 1, neg_inf_v>(P_ij, P_ij);
             mov<0, 2, neg_inf_v>(P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+          } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            make_window<0, 1, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -899,6 +922,14 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             // Apply the causal mask to [0, 2] and set [0, 3:4] to -inf
             make_causal<0, 2, neg_inf_v>(P_ij, P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+          } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            mov<0, 1, neg_inf_v>(P_ij);
+            make_window<0, 2, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1157,6 +1188,15 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
           } else if (q_pos == k_pos) {
             // Apply the causal mask to [0, 3]
             make_causal<0, 3, neg_inf_v>(P_ij, P_ij);
+#if WINDOW
+          } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            mov<0, 1, neg_inf_v>(P_ij);
+            mov<0, 2, neg_inf_v>(P_ij);
+            make_window<0, 3, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1436,6 +1476,13 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             mov<0, 1, neg_inf_v>(P_ij);
             mov<0, 2, neg_inf_v>(P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+          // window lower boundary, mirror of the causal edge
+          } else if (q_pos - k_pos == WINDOW) {
+            make_window<0, 0, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1699,6 +1746,13 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             make_causal<0, 1, neg_inf_v>(P_ij, P_ij);
             mov<0, 2, neg_inf_v>(P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+          } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            make_window<0, 1, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1958,6 +2012,14 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             // Apply the causal mask to [0, 2] and set [0, 3:4] to -inf
             make_causal<0, 2, neg_inf_v>(P_ij, P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+          } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            mov<0, 1, neg_inf_v>(P_ij);
+            make_window<0, 2, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -2216,6 +2278,15 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
           } else if (q_pos == k_pos) {
             // Apply the causal mask to [0, 3]
             make_causal<0, 3, neg_inf_v>(P_ij, P_ij);
+#if WINDOW
+          } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            mov<0, 1, neg_inf_v>(P_ij);
+            mov<0, 2, neg_inf_v>(P_ij);
+            make_window<0, 3, neg_inf_v>(P_ij, P_ij);
+          } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
           }
         }
         mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -2487,6 +2558,12 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             mov<0, 1, neg_inf_v>(P_ij);
             mov<0, 2, neg_inf_v>(P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+        } else if (q_pos - k_pos == WINDOW) {
+            make_window<0, 0, neg_inf_v>(P_ij, P_ij);
+        } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
         }
       }
       mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -2748,6 +2825,13 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             make_causal<0, 1, neg_inf_v>(P_ij, P_ij);
             mov<0, 2, neg_inf_v>(P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+        } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            make_window<0, 1, neg_inf_v>(P_ij, P_ij);
+        } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
         }
       }
       mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -3004,6 +3088,14 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
             // Apply the causal mask to [0, 2] and set [0, 3:4] to -inf
             make_causal<0, 2, neg_inf_v>(P_ij, P_ij);
             mov<0, 3, neg_inf_v>(P_ij);
+#if WINDOW
+        } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            mov<0, 1, neg_inf_v>(P_ij);
+            make_window<0, 2, neg_inf_v>(P_ij, P_ij);
+        } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
         }
       }
       mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -3260,6 +3352,15 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
         } else if (q_pos == k_pos) {
             // Apply the causal mask to [0, 3]
             make_causal<0, 3, neg_inf_v>(P_ij, P_ij);
+#if WINDOW
+        } else if (q_pos - k_pos == WINDOW) {
+            mov<0, 0, neg_inf_v>(P_ij);
+            mov<0, 1, neg_inf_v>(P_ij);
+            mov<0, 2, neg_inf_v>(P_ij);
+            make_window<0, 3, neg_inf_v>(P_ij, P_ij);
+        } else if (q_pos - k_pos > WINDOW) {
+            mov<neg_inf_v>(P_ij);
+#endif
         }
       }
       mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);

extra/thunder/amd/fa_fwd_causal.cpp

@@ -34,6 +34,10 @@ constexpr int ATTN_D = 128; // dimension
 constexpr int Q_BLOCK_SIZE = 32; // q block size
 constexpr int KV_BLOCK_SIZE = 64; // kv block size
 constexpr bool causal = true;
+// WINDOW>0: sliding-window attention, query i attends keys in [i-WINDOW+1, i]
+#ifndef WINDOW
+#define WINDOW 0
+#endif
 
 #define NUM_WARPS 8
 #define NUM_THREADS (kittens::WARP_THREADS * NUM_WARPS)
@@ -82,11 +86,26 @@ template<typename T=float, typename L=col_l, typename S=rt_16x32_4_s> using attn
 
 /**********************************************************/
 template<int THR_X, int THR_Y>
-__device__ inline void mask_vec2_imm(uint32_t rel_vgpr, uint32_t neg_inf_vgpr,
+__device__ inline void mask_vec2_imm(uint32_t rel_vgpr, uint32_t rel_hi_vgpr, uint32_t neg_inf_vgpr,
                                      uint32_t& x_ref, uint32_t& y_ref) {
 
     uint64_t x_mask, y_mask;
-    // uint32_t ox, oy;
+#if WINDOW
+    // causal+window in one asm block to not disturb register allocation
+    asm volatile(
+        "v_cmp_lt_i32_e64 %0, %4, %5\n\t"
+        "v_cmp_lt_i32_e64 %1, %4, %7\n\t"
+        "v_cndmask_b32_e64 %2, %2, %6, %0\n\t"
+        "v_cndmask_b32_e64 %3, %3, %6, %1\n\t"
+        "v_cmp_ge_i32_e64 %0, %8, %5\n\t"
+        "v_cmp_ge_i32_e64 %1, %8, %7\n\t"
+        "v_cndmask_b32_e64 %2, %2, %6, %0\n\t"
+        "v_cndmask_b32_e64 %3, %3, %6, %1\n\t"
+        : "=s"(x_mask), "=s"(y_mask), "+v"(x_ref), "+v"(y_ref)
+        : "v"(rel_vgpr), "n"(THR_X), "v"(neg_inf_vgpr), "n"(THR_Y), "v"(rel_hi_vgpr)
+        : "vcc"
+    );
+#else
     asm volatile(
         // x: rel < THR_X ?
         "v_cmp_lt_i32_e64 %0, %6, %7\n\t"
@@ -99,7 +118,7 @@ __device__ inline void mask_vec2_imm(uint32_t rel_vgpr, uint32_t neg_inf_vgpr,
           "n"(THR_X), "v"(neg_inf_vgpr), "n"(THR_Y)
         : "vcc"
     );
-    // x_ref = ox; y_ref = oy;
+#endif
 }
 
 template<ducks::rt::col_layout RT>
@@ -122,6 +141,8 @@ __device__ inline void mask_kv_tile(RT &dst, int q_abs, int k_abs, uint32_t neg_
         // (smaller rel ⇒ more "future" keys that must be -inf)
         const int rel0 = q_pos - (k_base + row_base);
         const uint32_t rel = static_cast<uint32_t>(rel0);
+        // rel-WINDOW keeps THR within the inline-constant range
+        const uint32_t rel_hi = static_cast<uint32_t>(rel0 - WINDOW);
 
         #pragma unroll
         for (int j = 0; j < dst.width; ++j) {
@@ -145,14 +166,14 @@ __device__ inline void mask_kv_tile(RT &dst, int q_abs, int k_abs, uint32_t neg_
             //  - reuse a single neg_inf register
             //  - keep VCC live across the pair
             //  - avoid reloading -inf or recomputing rel
-            mask_vec2_imm< 0, 1 >(rel, neg_inf_v, d0x, d0y);
-            mask_vec2_imm< 2, 3 >(rel, neg_inf_v, d1x, d1y);
-            mask_vec2_imm< 8, 9 >(rel, neg_inf_v, d2x, d2y);
-            mask_vec2_imm<10,11 >(rel, neg_inf_v, d3x, d3y);
-            mask_vec2_imm<16,17 >(rel, neg_inf_v, d4x, d4y);
-            mask_vec2_imm<18,19 >(rel, neg_inf_v, d5x, d5y);
-            mask_vec2_imm<24,25 >(rel, neg_inf_v, d6x, d6y);
-            mask_vec2_imm<26,27 >(rel, neg_inf_v, d7x, d7y);
+            mask_vec2_imm< 0, 1 >(rel, rel_hi, neg_inf_v, d0x, d0y);
+            mask_vec2_imm< 2, 3 >(rel, rel_hi, neg_inf_v, d1x, d1y);
+            mask_vec2_imm< 8, 9 >(rel, rel_hi, neg_inf_v, d2x, d2y);
+            mask_vec2_imm<10,11 >(rel, rel_hi, neg_inf_v, d3x, d3y);
+            mask_vec2_imm<16,17 >(rel, rel_hi, neg_inf_v, d4x, d4y);
+            mask_vec2_imm<18,19 >(rel, rel_hi, neg_inf_v, d5x, d5y);
+            mask_vec2_imm<24,25 >(rel, rel_hi, neg_inf_v, d6x, d6y);
+            mask_vec2_imm<26,27 >(rel, rel_hi, neg_inf_v, d7x, d7y);
         }
     }
 }
@@ -201,6 +222,16 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     else max_num_tiles = num_tiles;
     const int q_start_pos = tile_idx * Q_BLOCK_SIZE;
 
+#if WINDOW
+    // start at the first in-window tile; clamp keeps >=4 tiles for the pipeline unroll
+    const int block_min_q = block_tile_idx * NUM_WARPS * Q_BLOCK_SIZE;
+    int min_tile = (block_min_q - WINDOW + 1) / KV_BLOCK_SIZE;
+    if (min_tile < 0) min_tile = 0;
+    if (min_tile > max_num_tiles - 4) min_tile = max(0, max_num_tiles - 4);
+#else
+    constexpr int min_tile = 0;
+#endif
+
     constexpr float TEMPERATURE_SCALE = (D == 128) ? 0.08838834764f*1.44269504089f : 0.125f*1.44269504089f;
     uint32_t neg_inf_v = 0xff800000;
 
@@ -231,7 +262,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     G::prefill_swizzled_offsets<1, false>(k_smem[0], g.Kg, swizzled_offsets_K);
     G::prefill_swizzled_offsets<1, false>(v_smem[0], g.Vg, swizzled_offsets_V);
 
-    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, 0, head_idx_kv, 0}, swizzled_offsets_K);
+    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, min_tile, head_idx_kv, 0}, swizzled_offsets_K);
     __builtin_amdgcn_s_waitcnt(0);
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
@@ -243,9 +274,9 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     transpose(q_reg_transposed, q_reg);
 
     // All warps then collaboratively load in the first slice of V (V0) and the second slice of K (K1) into shared memory
-    G::load<1, false>(k_smem[1], g.Kg, {batch_idx, 1, head_idx_kv, 0}, swizzled_offsets_K);
+    G::load<1, false>(k_smem[1], g.Kg, {batch_idx, min_tile + 1, head_idx_kv, 0}, swizzled_offsets_K);
     // All warps then load in the first slice of K (K0)
-    G::load<1, false>(v_smem[0], g.Vg, {batch_idx, 0, head_idx_kv, 0}, swizzled_offsets_V);
+    G::load<1, false>(v_smem[0], g.Vg, {batch_idx, min_tile, head_idx_kv, 0}, swizzled_offsets_V);
     load(k_reg, k_smem[0]);
     __builtin_amdgcn_sched_barrier(0);
     asm volatile("s_waitcnt lgkmcnt(0)");
@@ -259,13 +290,20 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     mma_AtB(att_block[0], k_reg_transposed, q_reg_transposed, att_block[0]);
     __builtin_amdgcn_sched_barrier(0);
     if constexpr (causal) {
-        const int kv_end_pos = (1) * KV_BLOCK_SIZE;
-        if (__builtin_expect(q_start_pos < kv_end_pos, 0)) {  // Only mask if needed
-            mask_kv_tile(att_block[0], tile_idx, 0, neg_inf_v, lane);
+        const int kv_end_pos = (min_tile + 1) * KV_BLOCK_SIZE;
+        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, WINDOW ? 1 : 0)) {
+            mask_kv_tile(att_block[0], tile_idx, min_tile, neg_inf_v, lane);
         }
     }
     // Each warp performs a partial softmax of QK0 (i.e. some of the online softmax up until but not including the second exponential scaling of the attention block likely)
+#if WINDOW
+    // floor the max: min_tile can be fully masked, which would NaN via exp2(-inf - -inf)
+    zero(max_vec_prev);
+    add(max_vec_prev, max_vec_prev, -1e4f);
+    col_max(max_vec, att_block[0], max_vec_prev);
+#else
     col_max(max_vec, att_block[0]);
+#endif
 
     copy(max_vec_prev, max_vec);
     exp2(scale_vec, scale_vec);
@@ -284,21 +322,25 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     // All warps then load in the second slice of K (K1)
     load(k_reg, k_smem[1]);
     // All warps then collaboratively load in the third slice of K (K2) into shared memory
-    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, 2, head_idx_kv, 0}, swizzled_offsets_K);
+    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, min_tile + 2, head_idx_kv, 0}, swizzled_offsets_K);
     // All warps then collaboratively load in the second slice of V (V1) into shared memory
-    G::load<1, false>(v_smem[1], g.Vg, {batch_idx, 1, head_idx_kv, 0}, swizzled_offsets_V);
+    G::load<1, false>(v_smem[1], g.Vg, {batch_idx, min_tile + 1, head_idx_kv, 0}, swizzled_offsets_V);
     asm volatile("s_waitcnt lgkmcnt(0)");
     asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
     __builtin_amdgcn_sched_barrier(0);
     __builtin_amdgcn_s_barrier();
 
     // hot loop
-    for (int j = 3; j < max_num_tiles - 1; j += 2) {
+    for (int j = min_tile + 3; j < max_num_tiles - 1; j += 2) {
         // Cluster 0:
         //      QK1
         zero(att_block[1]);
         transpose(k_reg_transposed, k_reg);
         mma_AtB(att_block[1], k_reg_transposed, q_reg_transposed, att_block[1]);
+#if WINDOW
+        // window masks interior tiles that causal skips
+        mask_kv_tile(att_block[1], tile_idx, j - 2, neg_inf_v, lane);
+#endif
         //      Finish softmax for QK0
         exp2(att_block[0].tiles[1][0], att_block[0].tiles[1][0]);
         mul(norm_vec, norm_vec, scale_vec);
@@ -379,7 +421,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
         load(v_reg, v_smem[1]);
         if constexpr (causal) {
             const int kv_end_pos = (j) * KV_BLOCK_SIZE;
-            if (q_start_pos < kv_end_pos) {  // Only mask if needed
+            if (WINDOW || q_start_pos < kv_end_pos) {
                 mask_kv_tile(att_block[0], tile_idx, j - 1, neg_inf_v, lane);
             }
         }
@@ -447,7 +489,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     load(v_reg, v_smem[0]);
     if constexpr (causal) {
         const int kv_end_pos = (max_num_tiles - 2) * KV_BLOCK_SIZE;
-        if (__builtin_expect(q_start_pos < kv_end_pos, 0)) {  // Only mask if needed
+        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, WINDOW ? 1 : 0)) {
             mask_kv_tile(att_block[1], tile_idx, max_num_tiles - 3, neg_inf_v, lane);
         }
     }
@@ -510,7 +552,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     load(v_reg, v_smem[1]);
     if constexpr (causal) {
         const int kv_end_pos = (max_num_tiles - 1) * KV_BLOCK_SIZE;
-        if (__builtin_expect(q_start_pos < kv_end_pos, 1)) {  // Only mask if needed
+        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, 1)) {
             mask_kv_tile(att_block[0], tile_idx, max_num_tiles - 2, neg_inf_v, lane);
         }
     }
@@ -572,7 +614,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
     load(v_reg, v_smem[0]);
     if constexpr (causal) {
         const int kv_end_pos = (max_num_tiles) * KV_BLOCK_SIZE;
-        if (__builtin_expect(q_start_pos < kv_end_pos, 1)) {  // Only mask if needed
+        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, 1)) {
             mask_kv_tile(att_block[1], tile_idx, max_num_tiles - 1, neg_inf_v, lane);
         }
     }

extra/thunder/amd/include/utils.cpp

@@ -97,4 +97,33 @@ __device__ inline static void atomic_pk_add_bf16_with_warpid(const GL &dst, cons
             }(std::make_index_sequence<RT::width>{});
         }.template operator()<Ns>(), ...);
     }(std::make_index_sequence<RT::height>{});
-}
+}
+// make_window: complement of make_causal for the window lower boundary (q_pos-k_pos == WINDOW). masks = ~(causal masks)
+template<int N, int M, int GPR, ducks::art::all T0, ducks::art::all T1>
+__device__ static inline void make_window(T0 &dst, const T1 &src) {
+    static_assert(std::is_same_v<typename T0::T, float> && std::is_same_v<typename T1::T, float>, "Only float to float window mask is supported");
+    static_assert(std::is_same_v<typename T0::layout, typename T1::layout>, "Only same layout is supported");
+    static_assert(std::is_same_v<typename T0::shape, typename T1::shape>, "Only same shape is supported");
+
+    if constexpr (std::is_same_v<typename T0::layout, typename ducks::rt_layout::col> && std::is_same_v<typename T0::shape, typename ducks::rt_shape::rt_16x16>) {
+        using range_type_T0 = ducks::art::get_nth_range_t<typename T0::register_ranges, N * T0::width + M>;
+        using registers_T0 = ducks::art::split_many_t<ducks::art::type_list<range_type_T0>, 1>;
+        using range_type_T1 = ducks::art::get_nth_range_t<typename T1::register_ranges, N * T1::width + M>;
+        using registers_T1 = ducks::art::split_many_t<ducks::art::type_list<range_type_T1>, 1>;
+        static_assert(registers_T0::size == registers_T1::size);
+
+        uint64_t window_mask = 0x1FFF01FF001F0001;
+        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 0>::lo, ducks::art::get_nth_range_t<registers_T1, 0>::lo, GPR>(window_mask);
+
+        window_mask = 0x3FFF03FF003F0003;
+        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 1>::lo, ducks::art::get_nth_range_t<registers_T1, 1>::lo, GPR>(window_mask);
+
+        window_mask = 0x7FFF07FF007F0007;
+        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 2>::lo, ducks::art::get_nth_range_t<registers_T1, 2>::lo, GPR>(window_mask);
+
+        window_mask = 0xFFFF0FFF00FF000F;
+        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 3>::lo, ducks::art::get_nth_range_t<registers_T1, 3>::lo, GPR>(window_mask);
+    } else {
+        static_assert(false, "Unsupported window mask");
+    }
+}

test/amd/hw/test_vopc.py

@@ -471,6 +471,20 @@ class TestCmpFloat(unittest.TestCase):
     st = run_program(instructions, n_lanes=1)
     self.assertEqual(st.vcc & 1, 1, "Expected vcc=1 (1.0 != 2.0)")
 
+  def test_v_cmp_eq_f16_src0_hi(self):
+    """v_cmp_eq_f16 with src0 from high half (true16 384+n encoding)."""
+    cmp = v_cmp_eq_f16_e32(v[0], v[1])
+    cmp._raw += 128  # src0 v[0] -> v[0].h, the dsl can't encode hi-half src0 yet
+    instructions = [
+      s_mov_b32(s[0], 0x42003c00),  # hi=3.0, lo=1.0
+      v_mov_b32_e32(v[0], s[0]),
+      s_mov_b32(s[0], 0x47004200),  # hi=7.0, lo=3.0
+      v_mov_b32_e32(v[1], s[0]),
+      cmp,
+    ]
+    st = run_program(instructions, n_lanes=1)
+    self.assertEqual(st.vcc & 1, 1, "Expected vcc=1 (v0.hi 3.0 == v1.lo 3.0)")
+
   def test_v_cmp_nge_f16_inf_self(self):
     """v_cmp_nge_f16 comparing -inf with itself (unordered less than).
 

test/backend/test_asm_gemm.py

@@ -1,7 +1,7 @@
 import unittest
 from tinygrad import Tensor, Device, dtypes, Context
 from tinygrad.helpers import getenv, system, DEV
-from extra.gemm.cdna_asm_gemm import asm_gemm, hk_bf16_atb_gemm, quantize_mxfp4
+from extra.gemm.cdna_asm_gemm import asm_gemm, hk_bf16_atb_gemm
 from test.helpers import needs_second_gpu
 from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8, FP8_MAX
 
@@ -157,13 +157,20 @@ class TestMXFP4(unittest.TestCase):
 
   def test_quantize(self):
     import numpy as np
-    block = np.array([0, .26, .74, .75, 1.26, 1.75, 2.51, 3.5, 5.1, 6, -6] + [0] * 21, dtype=np.float32)
-    x = Tensor(np.tile(block, (32, 8)), dtype=dtypes.bfloat16)
-    packed, scale, _ = quantize_mxfp4(x)
-    p = packed.numpy()
-    codes = np.stack((p & 0xF, p >> 4), axis=-1).reshape(32, 256)
-    np.testing.assert_array_equal(codes[0, :11], [0, 1, 1, 2, 3, 4, 5, 6, 7, 7, 15])
-    np.testing.assert_array_equal(scale.numpy(), np.full((32, 8), 127, dtype=np.uint8))
+    from extra.llama_kernels.quantize_mxfp4 import quantize_mxfp4
+    rng = np.random.default_rng(0)
+    x = np.triu(rng.standard_normal((256, 256), dtype=np.float32))
+    x += np.triu(x, 1).T
+    x[:32, :32] = 0
+    row, row_scale, col, col_scale = quantize_mxfp4(Tensor(x, dtype=dtypes.bfloat16))
+    Tensor.realize(row, row_scale, col, col_scale)
+    row, row_scale = row.numpy(), row_scale.numpy()
+    col, col_scale = col.numpy(), col_scale.numpy()
+    np.testing.assert_array_equal(row, col)
+    np.testing.assert_array_equal(row_scale, col_scale)
+    self.assertTrue(row.any())
+    self.assertTrue((row_scale == 127).any())
+    self.assertTrue((row_scale != 127).any())
 
   def test_correctness(self):
     import numpy as np
@@ -171,17 +178,9 @@ class TestMXFP4(unittest.TestCase):
     rng = np.random.default_rng(1)
     a = Tensor(rng.standard_normal((M, K), dtype=np.float32), dtype=dtypes.bfloat16)
     b = Tensor(rng.standard_normal((N, K), dtype=np.float32), dtype=dtypes.bfloat16)
-    out = asm_gemm(a, b.T, mxfp4=True).realize()
-    # reference gemm
-    a_packed, scale_a, _ = quantize_mxfp4(a)
-    b_packed, scale_b, _ = quantize_mxfp4(b)
-    def unpack(x): return np.stack((x & 0xF, x >> 4), axis=-1).reshape(x.shape[0], -1)
-    code_a, code_b = unpack(a_packed.numpy()), unpack(b_packed.numpy())
-    lut = np.array([0, .5, 1, 1.5, 2, 3, 4, 6, -0., -.5, -1, -1.5, -2, -3, -4, -6], dtype=np.float32)
-    a_dequant = lut[code_a] * np.repeat(np.exp2(scale_a.numpy().astype(np.int16)-127), 32, axis=1)
-    b_dequant = lut[code_b] * np.repeat(np.exp2(scale_b.numpy().astype(np.int16)-127), 32, axis=1)
-    ref = Tensor(a_dequant @ b_dequant.T, dtype=dtypes.bfloat16).realize().numpy()
-    np.testing.assert_array_equal(out.numpy(), ref)
+    out = asm_gemm(a, b.T, mxfp4=True).realize().numpy().astype(np.float32)
+    ref = a.numpy().astype(np.float32) @ b.numpy().astype(np.float32).T
+    self.assertLess(np.linalg.norm(out-ref) / np.linalg.norm(ref), 0.2)
 
   def test_empty(self):
     M, N, K = getenv("M", 16384), getenv("N", 4096), getenv("K", 14336)

test/backend/test_custom_kernel.py

@@ -190,6 +190,12 @@ class TestCustomKernel(unittest.TestCase):
     b = Tensor.custom_kernel(tst, a, fxn=custom_sum)[0]
     self.assertEqual(b.item(), 15)
 
+  def test_sum_outside(self):
+    a = Tensor([1.0, 2, 3, 4, 5])+1
+    tst = Tensor.empty(1)
+    b = Tensor.custom_kernel(tst, a, fxn=custom_sum)[0]
+    self.assertEqual(b.item(), 20)
+
   def test_sum_int(self):
     a = Tensor([1, 2, 3, 4, 5])
     tst = Tensor.empty(1, dtype=a.dtype)
@@ -287,7 +293,7 @@ class TestCustomKernel(unittest.TestCase):
     GlobalCounters.reset()
     c.realize()
     assert all(i == 3. for i in c.flatten().tolist()), f"all 3 {c.tolist()}"
-    assert_kernel_count(3)
+    assert_kernel_count(2)
 
   def test_multi_after_schedule_order(self):
     """Test correct scheduling order when custom_kernel has multiple outputs.
@@ -405,10 +411,8 @@ class TestCustomKernel(unittest.TestCase):
     assert_kernel_count(2)
     self.assertEqual(z.tolist(), x.add(2).tolist())
 
-  @unittest.expectedFailure
   def test_custom_kernel_sched_copy(self): self.test_custom_kernel_sched(use_custom=True)
 
-  @unittest.expectedFailure
   def test_sliced_buffer_function(self):
     x = Tensor.arange(32).reshape(8, 4).clone().realize()
     from tinygrad import function
@@ -418,8 +422,8 @@ class TestCustomKernel(unittest.TestCase):
       return Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
     GlobalCounters.reset()
     y = run(x[0]).realize()
-    # it's copying the input and the output
-    assert_kernel_count(1)
+    # backends that support contiguous views don't launch extra kernels
+    assert_kernel_count(2 if x[0].uop.contiguous_view() is None else 1)
     self.assertEqual(y.tolist(), [1, 2, 3, 4])
 
   @Context(DEV="CPU")
@@ -429,12 +433,28 @@ class TestCustomKernel(unittest.TestCase):
     # TODO: it currently requires a compiler for Ops.BINARY
     from tinygrad.device import Device
     binary = Device[a.device].renderer.compiler.compile(src)
-    def custom_src_kernel(A:UOp) -> UOp:
+    def custom_src_kernel(A:UOp, B:UOp) -> UOp:
       sink = UOp.sink(A, arg=KernelInfo(name="test_src"))
       return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(sink.toposort())), UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=binary)))
-    a = Tensor.custom_kernel(a.reshape(2, 2).T, fxn=custom_src_kernel)[0]
-    self.assertEqual(a.tolist(), [[1, 2], [1, 3]])
+    a = Tensor.custom_kernel(a.reshape(2, 2).clone(), a.reshape(2, 2).T, fxn=custom_src_kernel)[0]
+    self.assertEqual(a.tolist(), [[1, 1], [2, 3]])
 
+  @Context(DEV="CPU")
+  def test_simple_from_source_alt(self):
+    a = Tensor.arange(4).clone().realize()
+    src = "void copy(int* restrict out, int* restrict in) { for (int i = 0; i < 4; i++) out[i] = in[i]; }"
+    from tinygrad.device import Device
+    binary = Device[a.device].renderer.compiler.compile(src)
+    def custom_src_kernel(out:UOp, inp:UOp) -> UOp:
+      sink = UOp.sink(out, inp, arg=KernelInfo(name="copy"))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(sink.toposort())), UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=binary)))
+    out = Tensor.custom_kernel(Tensor.empty_like(a), a+1, fxn=custom_src_kernel)[0]
+    GlobalCounters.reset()
+    out.realize()
+    assert_kernel_count(2)
+    self.assertEqual(out.tolist(), [1, 2, 3, 4])
+
+  @unittest.skip("this shouldn't be expected to work")
   def test_inplace_transpose(self):
     def custom_assign_row_max_kernel(A:UOp) -> UOp:
       row = UOp.range(A.shape[0], 0)
@@ -471,8 +491,8 @@ class TestCustomKernelInput(unittest.TestCase):
 
   def test_reshape(self): self._test_mop(lambda x: x.reshape(16, 2), max_kernels=2)
   def test_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T, max_kernels=3)
-  def test_double_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T.T, max_kernels=3)
-  def test_shrink(self): self._test_mop(lambda x: x[:4], max_kernels=2)
+  def test_double_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T.T, max_kernels=2)
+  def test_shrink(self): self._test_mop(lambda x: x[:4], max_kernels=1)
   def test_pad(self): self._test_mop(lambda x: x[:4].pad(((0, 4),)), max_kernels=2)
   def test_flip(self): self._test_mop(lambda x: x.flip(0), max_kernels=2)
   def test_offset_shrink(self): self._test_mop(lambda x: x[4:8], max_kernels=2)

test/backend/test_dtype.py

@@ -169,6 +169,13 @@ class TestFp8sConversions(unittest.TestCase):
   def test_fp8e5m2fnuz_to_float(self, x):
     np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e5m2fnuz).float().item())
 
+  def test_fp8e5m2fnuz_to_float_smallest_normals(self):
+    # fnuz bias exceeds half's, so exp-1 normals land below half's normal range: they flush to zero like denormals
+    if dtypes.half not in supported_dtypes or dtypes.half in EMULATED_DTYPES.tolist(dtypes) or dtypes.fp8e5m2fnuz in supported_dtypes:
+      self.skipTest("needs the emulated fp8 with a native half intermediate")
+    vals = Tensor([0x04, 0x05, 0x06, 0x07], dtype=dtypes.uint8).bitcast(dtypes.fp8e5m2fnuz).float().numpy()
+    np.testing.assert_equal(vals, [0., 0., 0., 0.])
+
 class TestBFloat16DType(unittest.TestCase):
   def test_bf16_to_float(self):
     _test_cast(Tensor([100000], dtype=dtypes.bfloat16), dtypes.float32)

test/backend/test_dtype_alu.py

@@ -399,9 +399,10 @@ class TestDTypeALU(unittest.TestCase):
     if float_dtype not in supported_dtypes: float_dtype = dtypes.float32
     universal_test_cast(a, float_dtype, unsigned_dtype)
 
-  @unittest.expectedFailure
-  def test_unsafe_cast_float_to_int_failure(self):
-    val = float(dtypes.int32.max - 1)
+  def test_unsafe_cast_float_to_int(self):
+    # the value is off the float32 grid but rounds in-range: the buffer and const-fold paths must agree
+    # (out-of-range float->int cast stays undefined: hardware may saturate where the fold wraps)
+    val = 2147483000.0
     t1 = Tensor([val], dtype=dtypes.float32).cast(dtypes.int32)
     t2 = Tensor(val, dtype=dtypes.float32).cast(dtypes.int32)
     np.testing.assert_equal(t1.item(), t2.item())

test/backend/test_jit.py

@@ -360,7 +360,7 @@ class TestJitGraphSplit(unittest.TestCase):
     self.expect(f, inp, inp_cpu,
       graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
       multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
-      hcqgraph=[self.ji_graph(4)])
+      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()]) # cpu is hcq2 now, it does not join hcq graphs
 
   def test_jit_cpu_several(self):
     if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")
@@ -377,9 +377,9 @@ class TestJitGraphSplit(unittest.TestCase):
     inp = Tensor.randn(10, 10, device=Device.DEFAULT).realize()
     inp_cpu = Tensor.randn(10, 10, device="CPU").realize()
     self.expect(f, inp, inp_cpu,
-      graph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
-      multigraph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
-      hcqgraph=[self.ji_graph(5)])
+      graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
+      multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
+      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()])
 
   def test_jit_multidev(self):
     if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")

test/backend/test_linearizer.py

@@ -30,7 +30,7 @@ class TestLinearizer(unittest.TestCase):
     c = ((a.shrink(((0, 2),)) - a.shrink(((2, 4),))) - (b.shrink(((0, 2),)) - b.shrink(((2, 4),))))
     linear = c.schedule_linear()
     run_linear(linear)
-    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if s.op is not Ops.BIND]
+    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if not s.is_bound_var]
     assert len(rawbufs) == 3 and set(rawbufs[1:]) == {a.uop.base.realized, b.uop.base.realized}
     np_c = (np_a[:2] - np_a[2:]) - (np_b[:2] - np_b[2:])
     np.testing.assert_allclose(np_c, c.numpy(), atol=1e-4, rtol=1e-4)
@@ -411,7 +411,7 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
   last_call = linear.src[-1]
   ast = last_call.src[0]
   assert ast.op is Ops.SINK, f"helper_realized_ast expects a SINK {last_call}"
-  last_bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]
+  last_bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]
   # now all input buffers in last_call should be realized
   # create fresh buffers for the outputs
   bufs = [Buffer(x.device, x.size, x.dtype).allocate() if i < len(ast.src) else x for i,x in enumerate(last_bufs)]
@@ -437,7 +437,7 @@ def reset_bufs(bufs:list[Buffer]):
   for buf in bufs: buf.copy_from(Buffer("PYTHON", buf.size, buf.dtype, opaque=memoryview(bytearray(buf.nbytes))))
 
 def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[],
-                               apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[]):
+                               apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[], check_default_opt=True):
   outbufs = real_bufs[:len(realized_ast.src)]
   wanna_output = [np.array(x).flatten() for x in wanna_output]
   buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in real_bufs]
@@ -459,9 +459,7 @@ def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[]
     for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)
 
   # Check correctness of handcoded optimiztions.
-  reset_bufs(outbufs)
-  run_prg(opts=None)
-  for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)
+  if check_default_opt: check_opt(None)
   for x in opts: # Check custom transformations if any.
     check_opt(([Opt(OptOps.TC, 0, (TC_SELECT.value, TC_OPT.value, 1))] if apply_tc else [])+x)
 

test/backend/test_llama_kernels.py

@@ -5,6 +5,7 @@ from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8
 from extra.llama_kernels.fused_ce import fused_ce_loss
 from extra.llama_kernels import local_abs_max
 from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed, quantize_fp8_scalar
+from extra.llama_kernels.swiglu import swiglu
 from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
 from extra.thunder.amd.fa import custom_fused_qkv_rope_backward, fused_qkv_rope
 from test.helpers import needs_second_gpu, assert_kernel_count
@@ -161,5 +162,31 @@ class TestFusedQKVRoPE(unittest.TestCase):
     ref = Tensor.cat(dq_ref, dk_ref, dv_ref, dim=3).reshape(*dx.shape).realize()
     with Context(DEBUG=0): self.assertTrue(dx.allclose(ref, atol=2e-2, rtol=2e-2).item(), "backward mismatch")
 
+def run_swiglu(test:unittest.TestCase, shape:tuple[int, ...]) -> None:
+  Tensor.manual_seed(0)
+  x = (Tensor.randn(*shape) * 2).cast(dtypes.bfloat16).realize()
+  hidden = x.shape[-1] // 2
+  out, ref = swiglu(x), x[..., :hidden].silu() * x[..., hidden:]
+  Tensor.realize(out, ref)
+  with Context(DEBUG=0): test.assertTrue(out.allclose(ref, atol=2.5e-1, rtol=3e-2).item(), "SwiGLU forward mismatch")
+
+  grad = (Tensor.randn(*out.shape) * 2).cast(dtypes.bfloat16).realize()
+  grad_x, grad_ref = out.gradient(x, gradient=grad)[0], ref.gradient(x, gradient=grad)[0]
+  Tensor.realize(grad_x, grad_ref)
+  test.assertEqual(grad_x.shape, shape)
+  test.assertEqual(grad_x.dtype, dtypes.bfloat16)
+  with Context(DEBUG=0): test.assertTrue(grad_x.allclose(grad_ref, atol=2.5e-1, rtol=3e-2).item(), "SwiGLU backward mismatch")
+
+class TestSwiGLU(unittest.TestCase):
+  def setUp(self):
+    if dtypes.bfloat16 not in Device[Device.DEFAULT].renderer.supported_dtypes(): self.skipTest("need bfloat16")
+
+  def test_simple(self): run_swiglu(self, (2, 32, 64))
+
+  def test_llama_shape(self):
+    if Device.DEFAULT != "AMD" or not Device[Device.DEFAULT].renderer.target.arch.startswith("gfx950"):
+      self.skipTest("only run on real machine for speed")
+    run_swiglu(self, (2, 8192, 28672))
+
 if __name__ == '__main__':
   unittest.main()

test/backend/test_multitensor.py

@@ -1,9 +1,9 @@
 import unittest, random
 from tinygrad import Tensor, Device, nn, GlobalCounters, TinyJit, dtypes, Variable
-from tinygrad.uop.ops import Ops, UOp, AxisType
+from tinygrad.uop.ops import Ops, UOp, AxisType, graph_rewrite
 from tinygrad.helpers import getenv, prod, Context
 from tinygrad.nn.state import get_parameters
-from tinygrad.engine.realize import run_linear, compile_linear
+from tinygrad.engine.realize import run_linear, compile_linear, pm_beam, pm_compile
 import numpy as np
 from hypothesis import given, strategies as strat, settings
 from test.helpers import not_support_multi_device, needs_second_gpu, slow, call_is_graph, check_schedule, assert_kernel_count
@@ -79,9 +79,9 @@ class TestMultiTensor(unittest.TestCase):
   def test_shard_beam(self):
     cpu_2 = ("CPU:1", "CPU:2")
     src = Tensor.ones(16).shard(cpu_2, 0).realize()
-    pad = src.to(cpu_2[::-1]).schedule_linear().src[0]
-    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): prg = compile_linear(UOp(Ops.LINEAR, src=(pad,))).src[0].src[0]
-    self.assertNotEqual(prg.src[0].arg.applied_opts, ())
+    lin = UOp(Ops.LINEAR, src=(src.to(cpu_2[::-1]).schedule_linear().src[0],))
+    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): call = graph_rewrite(graph_rewrite(lin, pm_beam, ctx=1, walk=True), pm_compile, walk=True).src[0]
+    self.assertNotEqual(call.src[0].src[0].arg.applied_opts, ())
 
   def test_shard_same_device(self):
     X = Tensor.ones(256).contiguous().realize()

test/backend/test_ops.py

@@ -1535,6 +1535,8 @@ class TestOps(unittest.TestCase):
 
   def test_prod(self):
     helper_test_op(None, lambda x: x.prod(), vals=[[1.0, 2.0, 3.0]])
+    helper_test_op(None, lambda x: x.prod(), vals=[[0.0, 2.0, 3.0]])
+    helper_test_op(None, lambda x: x.prod(), vals=[[0.0, 0.0, 3.0]])
     with Context(NOOPT=1): helper_test_op(None, lambda x: x.prod(), vals=[[1.0, 2.0, 3.0]])
     helper_test_op([(3,4,5,6)], lambda x: x.prod(dim=3), lambda x: x.prod(axis=3))
     helper_test_op([(3,4,5,6)], lambda x: x.prod(dim=1), lambda x: x.prod(axis=1))

test/backend/test_pickle.py

@@ -2,7 +2,7 @@ import unittest, pickle, types, tracemalloc
 import numpy as np
 from tinygrad import Tensor, Device, TinyJit, Variable, dtypes
 from tinygrad.helpers import GlobalCounters, ContextVar, Context, DEV
-from tinygrad.uop.ops import PatternMatcher, UPat, UOp
+from tinygrad.uop.ops import PatternMatcher, UPat, UOp, deconstruct_function
 
 class TestPickle(unittest.TestCase):
   def test_pickle_code_object(self):
@@ -11,6 +11,11 @@ class TestPickle(unittest.TestCase):
     fxn = types.FunctionType(pickle.loads(code_str), globals())
     self.assertEqual(fxn(2), 4)
 
+  def test_deconstruct_function_nested_comprehension(self):
+    # pre PEP 709, each comprehension is its own code object, so dtypes here is referenced two code objects deep
+    def fxn(): return [[dtypes.int for _ in range(2)] for _ in range(2)]
+    self.assertEqual(types.FunctionType(*deconstruct_function(fxn))(), fxn())
+
   def test_pickle_pattern_matcher(self):
     pm = PatternMatcher([(UPat.cvar('x'), lambda x: x*2)])
     sink = UOp.const(2)

test/device/test_hcq2.py

@@ -0,0 +1,14 @@
+import unittest, numpy as np
+from unittest.mock import patch
+from tinygrad import Device, Tensor
+from tinygrad.helpers import getenv
+from tinygrad.runtime.support.hcq2 import HCQ_DEVS, all_devices_in
+
+@unittest.skipUnless(getenv("HCQ2") and all_devices_in(Device.DEFAULT, HCQ_DEVS), "hcq2 device required")
+class TestHCQ2(unittest.TestCase):
+  def test_copy_without_copy_queue(self):
+    with patch.object(Device[Device.DEFAULT], "has_copy_queue", False):
+      np.testing.assert_equal(Tensor(np.arange(61, dtype=np.float32)).to(Device.DEFAULT).contiguous().realize().numpy(), np.arange(61))
+
+if __name__ == "__main__":
+  unittest.main()

test/external/external_benchmark_all2all.py

@@ -0,0 +1,17 @@
+from tinygrad import Device, Tensor, TinyJit, dtypes
+from tinygrad.helpers import Timing, Context
+
+GPUS, DEPTH, SZ = 8, 4, 128 * 2**20
+WARMUP, ITERS = 3, 5
+devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(GPUS))
+bufs = tuple(Tensor.empty(SZ, dtype=dtypes.uint8, device=dev).contiguous().realize() for _ in range(DEPTH) for dev in devs)
+
+@TinyJit
+def all_to_all(*srcs:Tensor): return Tensor.realize(*(src.to(dst) for i,src in enumerate(srcs) for j,dst in enumerate(devs) if i % GPUS != j))
+
+if __name__ == "__main__":
+  with Context(ALL2ALL=1, JIT_BATCH_SIZE=0):
+    for i in range(-WARMUP, ITERS):
+      with Timing("ALL2ALL ", lambda ns: f" {SZ*GPUS*(GPUS-1)*DEPTH/ns:.2f} GB/s", enabled=i>=0):
+        all_to_all(*bufs)
+        for dev in devs: Device[dev].synchronize()

test/external/external_test_schedule_scaling.py

@@ -1,5 +1,5 @@
-import unittest, time
-from tinygrad import Tensor
+import unittest, time, itertools
+from tinygrad import Tensor, Context
 
 class TestScheduleScaling(unittest.TestCase):
   """Test that .schedule() scales linearly with graph size (no O(n^2) behavior)."""
@@ -130,5 +130,18 @@ class TestScheduleScaling(unittest.TestCase):
       return parts[0].cat(*parts[1:])
     self._assert_linear(concat_chain)
 
+  @Context(DEV="NULL:HIP:gfx1100")
+  def test_custom_kernel_assign_scaling(self):
+    from tinygrad.uop.ops import UOp, Ops, KernelInfo
+    from tinygrad.runtime.autogen.amd.rdna3.ins import s_nop
+    count = itertools.count(0)
+    def custom_kernel_assign(n):
+      def custom_asm(out):
+        return UOp(Ops.PROGRAM, src=(UOp.sink(out, arg=KernelInfo(f"fxn_{next(count)}")),
+                                     UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=s_nop(i)) for i in range(n*8)))))
+      call = Tensor.custom_kernel(Tensor.empty(1), fxn=custom_asm)[0]
+      return Tensor.cat(*[Tensor.empty(1).assign(call+i) for i in range(n)])
+    self._assert_linear(custom_kernel_assign, n_small=50, n_large=500)
+
 if __name__ == '__main__':
   unittest.main(verbosity=2)

test/helpers.py

@@ -48,6 +48,8 @@ def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Te
   else:
     assert isinstance(t, UOp), f"can't schedule {t}"
     linear, var_vals = Tensor(t).linear_with_vars()
+  # test compiling the linear
+  compile_linear(linear)
   kernel_cnt = sum((len(call.device) if isinstance(call.device, tuple) else 1)
                    for call in linear.src if call.src[0].op is Ops.SINK or not filter_sink)
   if kernel_cnt != allowed:
@@ -57,8 +59,6 @@ def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Te
         print("kernel", i+1)
         print(call.src[0])
     raise KernelCountException(allowed, kernel_cnt)
-  # test compiling the linear
-  compile_linear(linear)
   return linear, var_vals
 
 def assert_kernel_count(expected:int):
@@ -86,7 +86,7 @@ def assert_jit_cache_len(fxn, expected_len):
   if linear is None or not linear.src:
     if expected_len != 0: raise KernelCountException(expected_len, 0)
     return
-  if expected_len and all(call_is_hcq(call) for call in linear.src): expected_len = 3 # HCQ2: merged same-queue calls + finalizer + bumps
+  if expected_len and all(call_is_hcq(call) for call in linear.src): expected_len = 4 # HCQ2: fence + reset + merged same-queue calls + finalizer
   if call_is_graph(linear.src[0]):
     if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
     inner = linear.src[0].src[0].src[0]  # LINEAR UOp inside CUSTOM_FUNCTION

test/mockgpu/amd/emu.py

@@ -260,19 +260,6 @@ def _cond(cond, if_true, if_false):
 def _cond_hi16(cond, val: UOp) -> UOp: return _cond(cond, _hi16(val), val)
 def _apply_opsel(val: UOp, sel_bit: int, opsel: int) -> UOp: return _hi16(val) if opsel & (1 << sel_bit) else val
 
-def _set_lane_bit(old: UOp, lane: UOp, val: UOp, exec_mask: UOp) -> UOp:
-  """Set/clear a single bit in a mask based on lane index, respecting exec mask."""
-  if old.dtype in (dtypes.uint64, dtypes.int64):
-    dt = dtypes.uint64
-    mask = UOp.const(1, dt) << lane.cast(dt)
-    new_bit = _to_u32(val).cast(dt) << lane.cast(dt)
-    cleared = old.cast(dt) & (mask ^ UOp.const(0xFFFFFFFFFFFFFFFF, dt))
-    return _lane_active(exec_mask, lane).where(cleared | new_bit, old.cast(dt))
-  mask = _c(1) << lane.cast(dtypes.uint32)
-  new_bit = _to_u32(val) << lane.cast(dtypes.uint32)
-  cleared = old & (mask ^ _c(MASK32))
-  return _lane_active(exec_mask, lane).where(cleared | new_bit, old)
-
 def _val_to_u32(val: UOp) -> UOp:
   """Convert any value to uint32 for storage (bitcast floats, cast ints)."""
   if val.dtype == dtypes.uint32: return val
@@ -532,6 +519,19 @@ class _Ctx:
       return [self.wsgpr_dyn(reg, lo), self.wsgpr_dyn(reg + _c(1), hi)]
     return [self.wsgpr_dyn(reg, val)]
 
+  def wmask_lane_bit(self, reg: UOp, lane: UOp, val: UOp, exec_mask: UOp) -> list[UOp]:
+    """Set/clear bit `lane` of the mask at `reg` from val for exec-active lanes, preserving memory for inactive lanes"""
+    active, bit = _lane_active(exec_mask, lane), _to_u32(val)
+    if self.wave_size <= 32:
+      old = self.rsgpr_dyn(reg)
+      mask = _c(1) << lane.cast(dtypes.uint32)
+      return [self.wsgpr_dyn(reg, active.where((old & (mask ^ _c(MASK32))) | (bit << lane.cast(dtypes.uint32)), old))]
+    off = (lane & _c(31, dtypes.int)).cast(dtypes.uint32)
+    mask = _c(1) << off
+    def half(old: UOp, sel: UOp) -> UOp: return sel.where(active.where((old & (mask ^ _c(MASK32))) | (bit << off), old), old)
+    return [self.wsgpr_dyn(reg, half(self.rsgpr_dyn(reg), lane < _c(32, dtypes.int))),
+            self.wsgpr_dyn(reg + _c(1), half(self.rsgpr_dyn(reg + _c(1)), _c(32, dtypes.int) <= lane))]
+
   def rmask(self, reg: UOp) -> UOp:
     """Read a lane mask (VCC/EXEC). Combines lo/hi for wave64."""
     if self.wave_size > 32: return _u64(self.rsgpr_dyn(reg), self.rsgpr_dyn(reg + _c(1)))
@@ -718,9 +718,7 @@ class _Ctx:
         raw_stores.append(('vgpr_direct', self.vgpr.index(val[0].valid(active)).store(new_val)))
         continue
       if 'D0' in dest and '[laneId]' in dest:
-        old_vcc = self.rmask(_c(VCC_LO.offset))
-        new_vcc = _set_lane_bit(old_vcc, lane, val, exec_mask)
-        raw_stores.extend([('vcc', s) for s in self.wmask(_c(VCC_LO.offset), new_vcc)])
+        raw_stores.extend([('vcc', s) for s in self.wmask_lane_bit(_c(VCC_LO.offset), lane, val, exec_mask)])
       elif dest.startswith('D0'):
         dest_suffix = re.match(r'D0\.(\w+)', dest)
         if dest_suffix is not None:
@@ -1039,13 +1037,11 @@ def _compile_sdwa(inst: irc.VOP1_SDWA | irc.VOP2_SDWA | irc.VOP2_SDWA_SDST | irc
         result = _sdwa_write(old, result, dst_sel, dst_unused)
       stores.append(ctx.wvgpr_dyn(vdst_reg, lane, result, exec_mask))
     elif dest.startswith('VCC'):
-      old_vcc = ctx.rmask(_c(VCC_LO.offset))
-      stores.extend(ctx.wmask(_c(VCC_LO.offset), _set_lane_bit(old_vcc, lane, val, exec_mask)))
+      stores.extend(ctx.wmask_lane_bit(_c(VCC_LO.offset), lane, val, exec_mask))
   if vcc_val is not None:
     # Initialize sdst to 0 before lane loop (old value may be unrelated data), then set lane bits in loop
     init_stores = [ctx.wsgpr_dyn(sdst_off, _c(0)), ctx.wsgpr_dyn(sdst_off + _c(1), _c(0))]
-    old_sdst = ctx.rmask(sdst_off)
-    stores.extend(ctx.wmask(sdst_off, _set_lane_bit(old_sdst, lane, vcc_val, exec_mask)))
+    stores.extend(ctx.wmask_lane_bit(sdst_off, lane, vcc_val, exec_mask))
     if stores:
       return UOp.sink(*init_stores, UOp.sink(*stores).end(lane), *ctx.inc_pc())
     return UOp.sink(*init_stores, *ctx.inc_pc())
@@ -1150,6 +1146,9 @@ def _compile_vopc(inst: ir3.VOPC|ir3.VOPC_DPP16|ir3.VOP3|ir4.VOPC|ir4.VOPC_DPP16
   def get_cmp_bit(lane) -> UOp:
     lc = lane.cast(dtypes.int) if isinstance(lane, UOp) else _c(lane, dtypes.int)
     s0 = _load_dpp16_src0(ctx, inst, lc, _c(0)) if is_dpp16 else ctx.rsrc_dyn(src0_off, lc, bits['s0'], literal, is_f64)
+    if is_vopc and not isinstance(inst, irc.VOPC) and bits['s0'] == 16 and not is_dpp16:
+      src0_hi = src0_off >= _c(384)
+      s0 = src0_hi.where(_hi16(ctx.rvgpr_dyn(src0_hi.where(src0_off - _c(384), _c(0)), lc)), s0)
     s1 = _cond_hi16(vsrc1_hi, ctx.rsrc_dyn(src1_off, lc, bits['s1'], literal, is_f64)) if bits['s0'] == 16 \
       else ctx.rsrc_dyn(src1_off, lc, bits['s1'], literal, is_f64)
     if bits['s0'] == 16 and opsel: s0, s1 = _apply_opsel(s0, 0, opsel), _apply_opsel(s1, 1, opsel)

test/models/test_whisper.py

@@ -74,6 +74,7 @@ class TestWhisper(unittest.TestCase):
         err
       )
 
+  @slow
   def test_transcribe_file1(self):
     self.assertEqual(transcribe_file(self.model, self.enc, TEST_FILE_1),  TRANSCRIPTION_1)
 
@@ -89,6 +90,7 @@ class TestWhisper(unittest.TestCase):
     self.assertEqual(TRANSCRIPTION_1,  transcriptions[0])
     self.assertEqual(TRANSCRIPTION_2,  transcriptions[1])
 
+  @slow
   def test_transcribe_batch21(self):
     waveforms = [load_file_waveform(TEST_FILE_2), load_file_waveform(TEST_FILE_1)]
     transcriptions = transcribe_waveform(self.model, self.enc, waveforms)

test/null/test_const_folding.py

@@ -51,6 +51,14 @@ class TestWeakConstFolding(unittest.TestCase):
   def test_invalid_poison(self):
     self.assertTrue(UOp.invalid().alu(Ops.CDIV, UOp.const(0)).simplify().is_invalid)
 
+  def test_single_rounding_log10_backward(self):
+    # log10 backward folds log10(2)/log(2) = 1/log(10) in one rounding, not the double-rounded 1/float32(log(10))
+    x = Tensor([1.0, 2.0, 3.0])
+    ast = next(s.src[0] for s in x.log10().sum().gradient(x)[0].schedule_linear().src if s.src[0].op is Ops.SINK)
+    const = next(u.arg for u in full_rewrite(ast).toposort() if u.op is Ops.CONST and u.dtype is dtypes.float32)
+    # correctly rounded: within half a float32 ulp of the exact value (folding at float32 lands 0.66 ulp off)
+    self.assertLess(abs(const - 1/math.log(10)), 2**-26)
+
 class TestBinaryOpsConstFolding(unittest.TestCase):
   def test_add_literal_zero(self):
     _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + 0)

test/null/test_graph_rewrite.py

@@ -143,13 +143,13 @@ class TestModuloAndDivisionFolding(unittest.TestCase):
     self.assertIs(apply_rewrite(x_var_uop.cast(dtypes.weakint) % 10).render(simplify=False), x_var_uop.render(simplify=False))
 
   def test_full_graph_rewrite_division_with_remainder(self):
-    x_var_uop = UOp.variable('x', 7, 9)
+    x_var_uop = UOp.variable('x', 7, 9, param=True)
     optimized_sink = apply_rewrite(x_var_uop // 2)
     for x_value in range(7, 10):
       self.assertEqual(x_value // 2, evaluate_uop(optimized_sink, {'x': x_value}))
 
   def test_full_graph_rewrite_complex_mod_div_expression(self):
-    x_var_uop = UOp.variable('x', 1, 10)
+    x_var_uop = UOp.variable('x', 1, 10, param=True)
     optimized_sink = apply_rewrite(((x_var_uop * 5) % 3) // 2)
     for x_value in range(1, 11):
       original_result = ((x_value * 5) % 3) // 2

test/null/test_memory_planner.py

@@ -27,10 +27,15 @@ def _make_linear(buffer_lists, copies=None):
     calls.append(UOp(Ops.CALL, src=(src0, *bufs)))
   return UOp(Ops.LINEAR, src=tuple(calls))
 
+def _get_planned_view(buf:UOp) -> tuple[UOp, int, int]|None:
+  view = buf.src[0] if buf.op is Ops.BITCAST else buf
+  if view.op is not Ops.SHRINK or view.src[0].op is not Ops.BUFFER: return None
+  return (arena:=view.src[0]), view.src[1].val * arena.dtype.itemsize, view.src[2].val * arena.dtype.itemsize
+
 def _get_arena(buf, linear, result):
   for orig_si, new_si in zip(linear.src, result.src):
     for orig, new in zip(orig_si.src[1:], new_si.src[1:]):
-      if orig is buf and new.op is Ops.SLICE: return new.src[0]
+      if orig is buf and (planned:=_get_planned_view(new)) is not None: return planned[0]
   return None
 
 def check_assign(buffer_lists, copies=None):
@@ -41,8 +46,8 @@ def check_assign(buffer_lists, copies=None):
   replace_map: dict[int, tuple[UOp, int, int]] = {}
   for orig_si, new_si in zip(linear.src, result.src):
     for orig, new in zip(orig_si.src[1:], new_si.src[1:]):
-      if new.op is Ops.SLICE and id(orig) not in replace_map:
-        replace_map[id(orig)] = (new.src[0], new.src[1].val * new.src[0].dtype.itemsize, new.arg * new.dtype.itemsize)
+      if (planned:=_get_planned_view(new)) is not None and id(orig) not in replace_map:
+        replace_map[id(orig)] = planned
 
   # verify pinned buffers are not planned
   for buf in held_bufs:

test/null/test_schedule.py

@@ -858,6 +858,65 @@ class TestSchedule(unittest.TestCase):
     x = Tensor.rand(32)
     check_schedule(x, 1, [Tensor._device_rng_counters[x.device]])
 
+  # **** custom kernel realize tests
+
+  @staticmethod
+  def _copy_fxn(name:str="copy"):
+    def copy_kernel(out:UOp, inp:UOp) -> UOp:
+      i = UOp.range(inp.numel(), 0)
+      return UOp.group(out[i].store(inp[i])).end(i).sink(arg=KernelInfo(name=name))
+    return copy_kernel
+
+  def _copy_call(self, out:Tensor, expr:Tensor, name:str="copy") -> Tensor:
+    # forge a custom kernel call with params and call args, like llm/kernels does (no Tensor.custom_kernel contiguous)
+    params = tuple(UOp.placeholder_like(u, slot=i) for i,u in enumerate((out.uop, expr.uop)))
+    return Tensor(out.uop.after(self._copy_fxn(name)(*params).call(out.uop, expr.uop)))
+
+  def test_custom_kernel_buffer_src(self):
+    # custom kernels need buffers: a buffer input must never add a realize kernel
+    y = Tensor.ones(64).contiguous().realize()
+    out = Tensor.empty_like(y)
+    check_schedule(self._copy_call(out, y), 1)
+
+  def test_custom_kernel_view_src(self):
+    # a RESHAPE over a buffer resolves to the buffer state (RESHAPEs on call args are stripped), no realize kernel
+    y = Tensor.ones(64).contiguous().realize()
+    out = Tensor.empty_like(y)
+    check_schedule(self._copy_call(out, y.reshape(8, 8).reshape(64)), 1)
+
+  def test_custom_kernel_elementwise_src(self):
+    # a computed input is not a buffer state: the call args are unwrapped to their base buffer,
+    # so the compute would be silently dropped. this must raise instead of producing wrong results
+    y = Tensor.ones(64).contiguous().realize()
+    out = Tensor.empty_like(y)
+    check_schedule(self._copy_call(out, y + y), 2)
+
+  def test_custom_kernel_lazy_const_src(self):
+    # a lazy const expression above the call has no buffer at all. this used to crash rangeify with a KeyError
+    x = Tensor.linspace(-1.0, 1.0, 64)
+    out = Tensor.empty_like(x)
+    check_schedule(self._copy_call(out, x), 2)
+
+  def test_custom_kernel_offset_view_src(self):
+    # a SHRINK with an offset over a buffer is not a buffer state either, the offset would be silently dropped
+    y = Tensor.ones(128).contiguous().realize()
+    out = Tensor.empty(64)
+    check_schedule(self._copy_call(out, y[16:80]), 2)
+
+  def test_custom_kernel_computed_src_api(self):
+    # the supported way to pass computed inputs: Tensor.custom_kernel makes inputs contiguous (one realize kernel)
+    y = Tensor.ones(64).contiguous().realize()
+    out = Tensor.empty_like(y)
+    check_schedule(Tensor.custom_kernel(out, y + y, fxn=self._copy_fxn())[0], 2)
+
+  def test_custom_kernel_on_custom_kernel(self):
+    # the output of a custom kernel is a buffer state, chaining custom kernels must not add kernels
+    y = Tensor.ones(64).contiguous().realize()
+    k1 = self._copy_call(Tensor.empty_like(y), y, name="k1")
+    k2 = self._copy_call(Tensor.empty_like(y), k1, name="k2")
+    sched, _ = check_schedule(k2, 2)
+    self.assertEqual([call.src[0].arg.name for call in sched.src], ["k1", "k2"])
+
   def test_empty_is_not_realized(self):
     a = Tensor.empty(10)
     child = a+2

test/null/test_simplify_valid_idx.py

@@ -2,7 +2,8 @@ import unittest, itertools
 
 from tinygrad.codegen.late.coalesce import indexing_simplify
 from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import UOp, Ops, graph_rewrite, pm_lower_index_dtype
+from tinygrad.uop.ops import UOp, Ops, graph_rewrite
+from tinygrad.uop.weak import pm_lower_index_dtype
 from tinygrad.uop.symbolic import simplify_valid, sym, pm_move_where_on_load
 from tinygrad.helpers import Context
 from test.helpers import full_rewrite
@@ -24,7 +25,7 @@ def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UO
   ))
 
 def Special(expr, nmax): return UOp(Ops.SPECIAL, src=(UOp.const(nmax),), arg=expr)
-def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
+def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax, param=True)
 def Range(n, nmax): return UOp.range(nmax, n)
 
 class TestValidIdxSimplification(unittest.TestCase):
@@ -332,7 +333,7 @@ class TestImageSimplification(unittest.TestCase):
     load = get_load_image_uop(shape, valid, idx)
 
     self.check(load,
-               "((((idx2*2)+r0)<11)&((((idx1*8)+r1)<3)!=True))",
+               "(((idx2*2)+r0)<11)",
                "(idx0+(idx1*512+r1*64)+-192)",
                "((((idx2*2)+r0)+(((idx1+((r1+5)//8))+1)//2))+-4)")
 
@@ -460,7 +461,7 @@ class TestImageSimplification(unittest.TestCase):
       self.check(load, None, "(gidx0+lidx0*1024+r0*1024+lidx1*128+-3168)", "0")
     except AssertionError:
       # TODO: fold valid
-      self.check(load, "(((lidx1<1)!=True)&(((lidx0+r0)<3)!=True)&((lidx0+r0)<19))",
+      self.check(load, "(((lidx1<1)!=True)&((lidx0+r0)<19))",
                        "(gidx0+lidx1*128+(lidx0*1024+r0*1024)+-3168)", "0")
 
   def test_simplify10(self):
@@ -479,7 +480,7 @@ class TestImageSimplification(unittest.TestCase):
       self.check(load, None, "(lidx2+gidx0*4+lidx0*1024+r0*1024+lidx1*256+-3264)", "0")
     except AssertionError:
       # TODO: fold valid
-      self.check(load, "(((lidx1<1)!=True)&(((lidx0+r0)<3)!=True)&((lidx0+r0)<11))",
+      self.check(load, "(((lidx1<1)!=True)&((lidx0+r0)<11))",
                        "(lidx2+gidx0*4+lidx1*256+(lidx0*1024+r0*1024)+-3264)", "0")
 
   def test_drop_non_monotonic_window(self):

test/null/test_uop_graph.py

@@ -1,10 +1,11 @@
 import unittest, pytest
-from tinygrad import dtypes, Variable
+from tinygrad import dtypes, Variable, Device
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import DEBUG, Context
-from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, graph_rewrite, GroupOp, AxisType, broadcast_axes
+from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, graph_rewrite, GroupOp, AxisType, broadcast_axes, KernelInfo
 from tinygrad.uop.symbolic import sym
 from test.helpers import to_uops_list
+from tinygrad.codegen import full_rewrite_to_sink
 
 simple_pm = PatternMatcher([
   (UPat.cvar('x', dtypes.weakint), lambda x: UOp.const(1.0) + UOp.const(2.0)),
@@ -156,7 +157,7 @@ class TestGraphRewrite(unittest.TestCase):
     self.assertEqual(nout.val, 3.0)
 
   def test_depth_2_fold(self):
-    v = UOp.variable("v", 0, 1, dtypes.float)
+    v = UOp.variable("v", 0, 1, dtypes.float, param=True)
     c1 = UOp.const(1.0)
     c2 = UOp.const(2.0)
     nout = graph_rewrite(v+c1+c2, simple_pm)
@@ -338,7 +339,7 @@ class TestUOpGraph(unittest.TestCase):
     self.assertEqual(len([x for x in uops if x.op is Ops.CAST]), 1)
 
   def test_depth_2_const_fold(self):
-    v = UOp.variable("tmp", 0, 1, dtypes.int)
+    v = UOp.variable("tmp", 0, 1, dtypes.int, param=True)
     c2 = UOp.const(2, dtypes.int)
     c4 = UOp.const(4, dtypes.int)
     vc = v+c2
@@ -536,6 +537,15 @@ class TestReduceCollapse(unittest.TestCase):
     # Should become add of two separate reduces
     self.assertEqual(result.op, Ops.ADD)
 
+  def test_reduce_shapeless_const_unroll(self):
+    """a REDUCE over a shapeless CONST (e.g. x*0 folded late in codegen) must collapse before the expander"""
+    out = UOp.param(0, dtypes.float, (1,))
+    red = UOp.const(3.0).cast(dtypes.float).reduce(UOp.range(4, 0, AxisType.UNROLL), arg=(Ops.ADD, 0))
+    ast = UOp.sink(out.index(UOp.const(0)).store(red)).replace(arg=KernelInfo())
+    uops = full_rewrite_to_sink(ast, Device["CPU"].renderer, optimize=False).toposort()
+    self.assertNotIn(Ops.REDUCE, [u.op for u in uops])
+    self.assertIn(12.0, [u.val for u in uops if u.op is Ops.CONST])
+
 class TestMovementOps(unittest.TestCase):
   def test_pm_mops_partial_reshape_index_removes_reshape(self):
     from tinygrad.schedule.rangeify import pm_mops

test/null/test_uop_symbolic.py

@@ -3,10 +3,9 @@ import unittest, pickle, functools, math
 import z3
 
 from tinygrad.dtype import dtypes, ConstType, DType, Invalid
-from test.helpers import get_uops
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad.uop.spec import spec_shared, type_verify
-from tinygrad.uop.symbolic import sym, commutative, pm_simplify_valid, pm_move_where_on_load
+from tinygrad.uop.symbolic import sym, pm_fold_cast_const, commutative, pm_simplify_valid, pm_move_where_on_load
 from tinygrad.uop.validate import uops_to_z3
 
 def check_uop_against_string(self, v:UOp, s:str):
@@ -16,7 +15,8 @@ def check_uop_against_string(self, v:UOp, s:str):
   s_eval = graph_rewrite(s_eval, commutative, name="cannonicalize eval")
   self.assertIs(s_eval, v, f"eval did not match simplified: {s_eval} != {v.render()} for {s}")
 
-def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint): return UOp.variable(name,min_val,max_val,dtype)
+def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint):
+  return UOp.variable(name, min_val, max_val, dtype, param=True)
 def uconst(val): return UOp.const(val)
 def usum(ops): return functools.reduce(lambda x,y: x+y, ops)
 def uand(ops): return functools.reduce(lambda x,y: x*y, ops)
@@ -35,7 +35,7 @@ class TestSymbolic(unittest.TestCase):
     self.assertEqual(solver.check(expr1 != expr2), z3.unsat, "simplified expression not equal to original")
 
   def helper_test_variable(self, v, n, m, s, test_z3:bool=True):
-    v_simplified = graph_rewrite(v, sym, name="simplify symbolic uop")
+    v_simplified = graph_rewrite(v, sym+pm_fold_cast_const, name="simplify symbolic uop")
     if test_z3: self.check_equal_z3(v, v_simplified)
     nmin, nmax = v_simplified.vmin, v_simplified.vmax
     check_uop_against_string(self, v_simplified, s)
@@ -442,7 +442,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(uand([uconst(1), Variable("a", 0, 1)]), 0, 1, "a")
 
   def test_masked_shr_fold(self):
-    x = UOp.variable('x', 0, 255, dtype=dtypes.uint32)
+    x = UOp.variable('x', 0, 255, dtype=dtypes.uint32, param=True)
     self.helper_test_variable((x & -4) >> 2, 0, 63, "(x>>2)")
 
   def test_bool_or_not_tautology(self):
@@ -483,12 +483,12 @@ class TestSymbolic(unittest.TestCase):
 
   def test_div_drop_small_terms(self):
     # from openpilot, shouldnt simplify
-    gidx0 = UOp.variable("gidx0", 0, 10)
-    gidx1 = UOp.variable("gidx1", 0, 10)
-    lidx0 = UOp.variable("lidx0", 0, 1)
-    lidx1 = UOp.variable("lidx1", 0, 1)
-    ridx1005 = UOp.variable("ridx1005", 0, 2)
-    ridx1006 = UOp.variable("ridx1006", 0, 2)
+    gidx0 = UOp.variable("gidx0", 0, 10, param=True)
+    gidx1 = UOp.variable("gidx1", 0, 10, param=True)
+    lidx0 = UOp.variable("lidx0", 0, 1, param=True)
+    lidx1 = UOp.variable("lidx1", 0, 1, param=True)
+    ridx1005 = UOp.variable("ridx1005", 0, 2, param=True)
+    ridx1006 = UOp.variable("ridx1006", 0, 2, param=True)
     self.helper_test_variable((lidx1+((gidx1*18)+(ridx1005*18)+(lidx0*162))+(gidx0*2)+(ridx1006*2)+-40)//18, -3, 20,
       "(gidx1+ridx1005+lidx0*9+(gidx0+ridx1006+7)//9+-3)")
 
@@ -948,6 +948,11 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(cond.where(u0, u1), 0, 1, "((a<2)!=True)")
     self.helper_test_variable(cond.where(u0, u1).where(u0, u1), 0, 1, "(a<2)")
 
+  def test_equivalent_const_max(self):
+    x = Variable("x", -10, 10)
+    self.helper_test_variable((x < 0).where(0, x), 0, 10, "x.maximum(0)")
+    self.helper_test_variable((0 < x).where(x, 0), 0, 10, "x.maximum(0)")
+
   def test_where_combine(self):
     cond = Variable("x", 0, 3) < 2
     a = Variable("a", 0, 3)
@@ -992,7 +997,7 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(cond.ne(False), 0, 1, "(x<2)")
 
   def test_bitcast_chain(self):
-    a = UOp.variable("a", 0, 3, dtype=dtypes.int32)
+    a = UOp.variable("a", 0, 3, dtype=dtypes.int32, param=True)
     self.assertIs(graph_rewrite(a.bitcast(dtypes.float32).bitcast(a.dtype), sym), a)
 
   def test_negation_in_where(self):
@@ -1007,22 +1012,6 @@ class TestSymbolic(unittest.TestCase):
     b = Variable("b", 0, 3)
     self.helper_test_variable(-a<-b, False, True, "(b<a)")
 
-  def test_where_cast(self):
-    s = Variable("s", 0, 3, dtypes.int)
-    cond = s < 2
-    a = Variable("a", 0, 3, dtypes.int)
-    b = Variable("b", 0, 3, dtypes.int)
-    expr = cond.where(a, b).cast(dtypes.half)
-
-    # TODO: copied from render, render does not support cast
-    glbl = UOp.param(0, dtypes.int, (1,))
-    uops = get_uops(UOp(Ops.STORE, src=(glbl.index(UOp.const(0, dtypes.int)), expr)).sink())
-    rewritten_uop = [uop for uop in uops if uop.op is Ops.STORE][0].src[1]
-
-    # the vars are now scalar PARAMs
-    pvar = {u.expr: u for u in rewritten_uop.toposort() if u.op is Ops.PARAM}
-    self.assertEqual(rewritten_uop, (pvar['s']<UOp.const(2, dtypes.int)).where(pvar['a'].cast(dtypes.half), pvar['b'].cast(dtypes.half)))
-
   def test_where_merge_branches(self):
     cond1 = Variable("s", 0, 10) < 6
     cond2 = Variable("s", 0, 10) > 2
@@ -1175,7 +1164,7 @@ class TestSymbolicVariables(unittest.TestCase):
     assert (a//4 + a//6).variables() == [a]
 
   def test_variable_min_eq_max_bind_folds(self):
-    b = Variable("x", 1, 1).bind(1)
+    b = UOp.variable("x", 1, 1).bind(1)
     s = b.simplify()
     self.assertEqual(s.op, Ops.CONST)
     self.assertEqual(s.val, 1)
@@ -1369,6 +1358,16 @@ class TestInvalidIndex(unittest.TestCase):
     c2 = UOp.const((1, Invalid, 1, 1))
     self.assertIs((c1+c2).simplify(), UOp.const((2, Invalid, Invalid, Invalid)))
 
+  def test_gated_load_keeps_index_valid(self):
+    # the load executes even on gated-off iterations: gated_given_valid must not erase its mask (PADTO OOB shape)
+    buf = UOp.param(0, dtypes.bool, (17,))
+    ridx = Variable("ridx", 0, 31)
+    cond = ridx < 17
+    load = buf.index(ridx.valid(cond))
+    out = graph_rewrite(cond.where(load.where(uconst(2), uconst(0)), UOp.invalid()), sym)
+    idx = next(u for u in out.toposort() if u.op is Ops.INDEX)
+    self.assertIs(idx.src[1].get_valid(), cond.simplify())
+
 class TestStoreLoadFolding(unittest.TestCase):
   """Tests for store(index, load(index)) -> NOOP rule. This rule matches patterns that EMERGE during simplification."""
   def test_store_load_folding(self):

test/null/test_uop_vmin_vmax.py

@@ -1,6 +1,6 @@
 import unittest, math
 from tinygrad.uop.ops import UOp, Ops
-from tinygrad.dtype import dtypes, Invalid
+from tinygrad.dtype import dtypes, Invalid, truncate
 
 class TestVminVmaxProperties(unittest.TestCase):
   def test_vmin_vmax_constant(self):
@@ -168,6 +168,10 @@ class TestVminVmaxProperties(unittest.TestCase):
     x = UOp.const(4.5).cast(dtypes.float)
     self.assertIs(x.ne(x.cast(dtypes.int).cast(dtypes.float)).simplify().arg, True)
 
+  def test_vmin_vmax_cast_int_to_float_grid(self):
+    # a cast to float only takes values on the float grid, so its bounds are the source bounds rounded at the destination
+    self.assertEqual(UOp.variable('x', 0, 16777219, dtypes.int).cast(dtypes.float)._min_max, (0.0, 16777220.0))
+
   def test_vmin_vmax_invalid(self):
     i = UOp.invalid()
     self.assertNotEqual(i.vmin, i.vmax)
@@ -317,8 +321,8 @@ class TestVminVmaxVConst(unittest.TestCase):
   def test_vmin_vmax_vconst_with_floats(self):
     # vmin and vmax for a vector constant of float values
     uop = UOp.const((1.5, -3.2, 0.0))
-    self.assertEqual(uop.vmin, -3.2)
-    self.assertEqual(uop.vmax, 1.5)
+    self.assertEqual(uop.vmin, truncate[dtypes.default_float](-3.2))
+    self.assertEqual(uop.vmax, truncate[dtypes.default_float](1.5))
 
   def test_vmin_vmax_vconst_with_bools(self):
     # vmin and vmax for a vector constant of bool values

test/null/test_uops.py

@@ -5,7 +5,8 @@ from tinygrad.tensor import Tensor
 from tinygrad.helpers import Timing, Context, cdiv
 from tinygrad.dtype import dtypes, AddrSpace, ConstFloat, Invalid  # noqa: F401
 from tinygrad.device import Device
-from tinygrad.uop.ops import Ops, ParamArg, PatternMatcher, UOp, UPat, dtype_from_uop, exec_alu, graph_rewrite, pm_lower_index_dtype  # noqa: F401  # ParamArg used by eval(str(uop)) roundtrip tests
+from tinygrad.uop.ops import Ops, ParamArg, PatternMatcher, UOp, UPat, dtype_from_uop, exec_alu, graph_rewrite  # noqa: F401  # ParamArg used by eval(str(uop)) roundtrip tests
+from tinygrad.uop.weak import pm_lower_index_dtype
 from tinygrad.uop.spec import spec_program, spec_shared, type_verify
 from tinygrad.uop.symbolic import sym, pm_remove_invalid
 from test.helpers import eval_uop, to_uops_list

test/null/test_validate_oob.py

@@ -1,10 +1,12 @@
 import unittest
-from tinygrad import dtypes, Variable
+from tinygrad import dtypes
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import Ops, UOp, AxisType
 from test.helpers import to_uops_list
 
+def Variable(name, nmin, nmax): return UOp.variable(name, nmin, nmax, param=True)
+
 class TestValidateOOB(unittest.TestCase):
   """Test z3 validation of index bounds for different ALU ops and patterns."""
 

test/null/test_viz.py

@@ -305,10 +305,10 @@ class TestVizTree(unittest.TestCase):
 
   def test_tree_view(self):
     with save_viz() as viz:
-      a = UOp.variable("a",0,10)
-      b = UOp.variable("b",0,10)
-      c = UOp.variable("c",0,10)
-      d = UOp.variable("d",0,10)
+      a = UOp.variable("a",0,10,param=True)
+      b = UOp.variable("b",0,10,param=True)
+      c = UOp.variable("c",0,10,param=True)
+      d = UOp.variable("d",0,10,param=True)
       sink = UOp.sink(a+b, c+d)
       def tree_rewrite(): return graph_rewrite(sink, root, name="root")
       tree_rewrite()

test/opt/test_kernel_opts.py

@@ -239,6 +239,15 @@ class TestKernelOpts(unittest.TestCase):
     helper_linearizer_opt(a.sum().exp(), [[Opt(OptOps.PADTO, 0, 32)],])
     helper_linearizer_opt(a.sum(0).exp(), [[Opt(OptOps.PADTO, 1, 32)],])
 
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  @unittest.expectedFailure
+  def test_padto_group_full_unroll_sum(self):
+    a = Tensor.ones(2, 28, 4096, dtype=dtypes.bfloat16).realize()
+    out = ((a * 0.5).float().square()).sum(axis=(0, 2))
+    opts_to_apply = [Opt(OptOps.GROUPTOP, 1, 256), Opt(OptOps.PADTO, 3, 32), Opt(OptOps.UNROLL, 2, 0), Opt(OptOps.UPCAST, 0, 7)]
+    helper_linearizer_opt(out, [opts_to_apply], check_default_opt=False)
+
   def test_padto_sum(self):
     N = 18
     # NOTE: this setup prevents 17 * 17 contiguous merged into one dimension

test/opt/test_tensor_cores.py

@@ -79,7 +79,8 @@ class TestTensorCores(unittest.TestCase):
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   def test_tensor_cores(self):
     for tc in Device[Device.DEFAULT].renderer.tensor_cores:
-      helper_tc_allclose(tc.dims[0], tc.dims[1], tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)
+      with self.subTest(tc=tc):
+        helper_tc_allclose(tc.dims[0], tc.dims[1], tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   def test_tensor_cores_nested_reduce(self):
@@ -185,10 +186,10 @@ class TestTensorCores(unittest.TestCase):
     # skip fp8 tcs: the unoptimized ALU baseline quantizes products to fp8 (JAX promotion), which legitimately
     # differs from the MFMA path (f32 accumulation), so the baseline-vs-TC numerical gate can't hold for fp8.
     tc = next(tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in not in dtypes.fp8s)
-    x, y = Tensor.rand(64, 64, dtype=tc.dtype_in), Tensor.rand(64, 64, dtype=tc.dtype_in)
+    x, y = Tensor.rand(16, 64, dtype=tc.dtype_in), Tensor.rand(64, 16, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out)
     opts = [Opt(OptOps.UNROLL, 0, 2)]
-    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3, check_default_opt=False)
     for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[1].src):
       if u.op is Ops.WMMA:
         assert u.src[-1].src[0].op != Ops.STORE
@@ -199,10 +200,10 @@ class TestTensorCores(unittest.TestCase):
   @unittest.skipIf(Device.DEFAULT in {"CPU"}, "CPU does not support using a different type for accumulation")
   def test_tensor_cores_unroll_casted_phi(self):
     tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out and tc.dtype_in not in dtypes.fp8s][0]
-    x, y = Tensor.rand(64, 64, dtype=tc.dtype_in), Tensor.rand(64, 64, dtype=tc.dtype_in)
+    x, y = Tensor.rand(16, 64, dtype=tc.dtype_in), Tensor.rand(64, 16, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out)
     opts = [Opt(OptOps.UNROLL, 0, 2)]
-    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3, check_default_opt=False)
     for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[1].src):
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
@@ -215,10 +216,10 @@ class TestTensorCores(unittest.TestCase):
   def test_tensor_cores_unroll_casted_phi_with_children(self):
     # all STORE children are outside the loop
     tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out and tc.dtype_in not in dtypes.fp8s][0]
-    x, y = Tensor.rand(64, 64, dtype=tc.dtype_in), Tensor.rand(64, 64, dtype=tc.dtype_in)
+    x, y = Tensor.rand(16, 64, dtype=tc.dtype_in), Tensor.rand(64, 16, dtype=tc.dtype_in)
     r = x.matmul(y, dtype=tc.dtype_out).relu()
     opts = [Opt(OptOps.UNROLL, 0, 2)]
-    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3, check_default_opt=False)
     for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[1].src):
       if u.op is Ops.WMMA:
         #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))

test/testextra/test_hevc.py

@@ -1,7 +1,9 @@
 import unittest
 
-from tinygrad import Tensor, Device, dtypes
-from tinygrad.helpers import fetch, round_up
+from tinygrad import Tensor, Device, Variable, dtypes
+from tinygrad.helpers import DEV, fetch, round_up
+from tinygrad.engine.realize import compile_linear
+from tinygrad.uop.ops import Ops
 from extra.hevc.hevc import parse_hevc_file_headers, nv_gpu
 from extra.hevc.decode import hevc_decode
 
@@ -63,7 +65,7 @@ class TestHevc(unittest.TestCase):
     self.assertEqual(list(frame3.initreflistidxl1), [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
     self.assertEqual(list(frame3.RefDiffPicOrderCnts), [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
 
-  @unittest.skipUnless(Device.DEFAULT == "NV", "NV only")
+  @unittest.skipUnless(Device.DEFAULT == "NV" and not DEV.interface.startswith("MOCK"), "real NV only")
   def test_hevc_decode(self):
     url = "https://github.com/haraschax/filedump/raw/09a497959f7fa6fd8dba501a25f2cdb3a41ecb12/comma_video.hevc"
     dat = fetch(url, headers={"Range": f"bytes=0-{512<<10}"}).read_bytes()
@@ -83,5 +85,22 @@ class TestHevc(unittest.TestCase):
       self.assertEqual(f.dtype, dtypes.uint8)
       self.assertEqual(f.device, "NV")
 
+  @unittest.skipUnless(Device.DEFAULT == "NV", "NV only")
+  def test_hevc_decode_compile(self):
+    url = "https://github.com/haraschax/filedump/raw/09a497959f7fa6fd8dba501a25f2cdb3a41ecb12/comma_video.hevc"
+    dat = fetch(url, headers={"Range": f"bytes=0-{512<<10}"}).read_bytes()
+
+    opaque, frame_info, _, _, luma_w, luma_h, _ = parse_hevc_file_headers(dat)
+    offset, sz, frame_pos, max_hist, _ = frame_info[1]
+    out_image_size = luma_h + (luma_h + 1) // 2, round_up(luma_w, 64)
+    history = [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV") for _ in range(max_hist)]
+    decoded = Tensor(dat, device="NV")[offset:offset+sz].decode_hevc_frame(
+      Variable("pos", 0, max_hist + 1).bind(frame_pos), out_image_size, opaque[1], history)
+
+    compiled = compile_linear(decoded.linear_with_vars()[0])
+    self.assertTrue(any(call.src[0].op is Ops.PROGRAM for call in compiled.src))
+    encdec_calls = [call for call in compiled.src if call.src[0].op is Ops.CUSTOM_FUNCTION and call.src[0].arg == "encdec"]
+    self.assertEqual(len(encdec_calls), 1)
+
 if __name__ == "__main__":
   unittest.main()

test/unit/test_assign.py

@@ -540,6 +540,10 @@ class TestAssign(unittest.TestCase):
     c = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
     c[0:2].bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000], dtype=dtypes.uint32)).realize()
     np.testing.assert_allclose(c.numpy(), [4.0, 3.0, 3.0, 4.0])
+    # without .realize()
+    a = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
+    a.bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000, 0x40000000, 0x3f800000], dtype=dtypes.uint32))
+    np.testing.assert_allclose(a.numpy(), [4.0, 3.0, 2.0, 1.0])
 
   def test_assign_bitcast_different_size(self):
     # assign to a shape-changing bitcast view (only works on DISK currently)

test/unit/test_call.py

@@ -212,6 +212,18 @@ class TestCallSchedule(unittest.TestCase):
     out = f(a, v.bind(5))
     np.testing.assert_allclose(out.numpy(), [5., 10., 15.])
 
+  def test_precompile_scoped_bind_arg(self):
+    @function(precompile=True)
+    def f(x:Tensor, scale:UOp) -> Tensor: return x * scale
+    a = Tensor.ones(3)
+    x = f(a, UOp.variable("scale_a", 1, 100).bind(2))
+    y = f(a, UOp.variable("scale_b", 1, 100).bind(3))
+    fx = next(u for u in x.uop.toposort() if u.op is Ops.FUNCTION)
+    fy = next(u for u in y.uop.toposort() if u.op is Ops.FUNCTION)
+    self.assertEqual(fx.src[0].key, fy.src[0].key)
+    np.testing.assert_equal(x.numpy(), [2, 2, 2])
+    np.testing.assert_equal(y.numpy(), [3, 3, 3])
+
   def test_precompile_schedule_cache_hit(self):
     """two instances of the same @function should produce identical function body keys (schedule cache hit)"""
     @function(precompile=True)
@@ -347,5 +359,15 @@ class TestCallMultiSharded(unittest.TestCase):
     np.testing.assert_allclose(a.grad.numpy(), b.numpy(), rtol=1e-5)
     np.testing.assert_allclose(b.grad.numpy(), a.numpy(), rtol=1e-5)
 
+  def test_symbolic_reshape_shard_axis(self):
+    toks = UOp.variable("toks", 1, 2).bind(2)
+    devs = ("CPU:0", "CPU:1")
+    x = Tensor(np.arange(16, dtype=np.float32).reshape(1, 2, 8)).shard(devs, axis=2).realize()
+    @function
+    def f(x:Tensor) -> Tensor: return x.reshape(1, x.shape[1], 2, 4)
+    out = f(x[:, :toks]).realize()
+    self.assertEqual(out.uop.axis, 2)
+    np.testing.assert_equal(out[:1, :2].to(devs[0]).numpy(), np.arange(16, dtype=np.float32).reshape(1, 2, 2, 4))
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_dtype_spec.py

@@ -222,6 +222,12 @@ class TestAutoCastType(unittest.TestCase):
     t.square().mean().backward()
     np.testing.assert_allclose(t.grad.numpy().flatten(), [60000 * 2 / (N*N)] * N*N)
 
+  @unittest.skipUnless(dtypes.half in supported_dtypes, "need half")
+  def test_var_half_precision_large_n(self):
+    # the element count (70000) exceeds half max (65504): the denominator must not be materialized in half
+    t = Tensor([[0.0, 1.0]], dtype=dtypes.half).expand(35000, 2).contiguous()
+    np.testing.assert_allclose(t.var().numpy(), 0.25, rtol=1e-3)
+
   @unittest.skipIf(Device.DEFAULT == "WEBGPU", "Precision error")
   @unittest.skipUnless(dtypes.half in supported_dtypes, "need half")
   def test_softmax_dtype(self):

test/unit/test_dtype_weak.py

@@ -3,7 +3,8 @@ import tempfile, unittest, math
 from tinygrad import Tensor, dtypes, TinyJit
 from tinygrad.helpers import Context
 from tinygrad.dtype import least_upper_float
-from tinygrad.uop.ops import UOp, Ops, dtype_from_uop, graph_rewrite, pm_lower_index_dtype, pm_commit_weak
+from tinygrad.uop.ops import UOp, Ops, dtype_from_uop, graph_rewrite
+from tinygrad.uop.weak import pm_lower_index_dtype, pm_commit_weak
 from tinygrad.uop.symbolic import symbolic_simple
 from tinygrad.uop.spec import spec_shared, type_verify
 from tinygrad.engine.jit import JitError

test/unit/test_hcq_graph.py

@@ -25,10 +25,10 @@ class TestHCQUnit(unittest.TestCase):
     cpu_call = UOp(Ops.PROGRAM, src=(UOp.sink(),)).call(UOp.new_buffer("CPU", 1, dtypes.float))
     gpu_devs = [d0]
 
-    # local MMIO: GPU works alone and with CPU in batch (cpu_support=True)
+    # CPU uses HCQ2 and is no longer batched into legacy HCQ graphs.
     assert HCQGraph.supports_uop(gpu_devs, gpu_call) is True
-    assert HCQGraph.supports_uop(gpu_devs, cpu_call) is True
-    assert HCQGraph.supports_uop(gpu_devs + [cpu_dev], gpu_call) is True
+    assert HCQGraph.supports_uop(gpu_devs, cpu_call) is False
+    assert HCQGraph.supports_uop(gpu_devs + [cpu_dev], gpu_call) is False
 
     # USB MMIO: GPU-only still works, but CPU batching must be rejected (cpu_support=False)
     orig_view = d0.timeline_signal.base_buf.view

test/unit/test_llm_server.py

@@ -13,12 +13,18 @@ V_TOKS = UOp.variable("toks", 1, 32)  # 32 is the default chunk_size in generate
 class TestTransformerGenerate(unittest.TestCase):
   def test_warmup(self):
     model, calls = Transformer(TEST_CONFIG), []
-    def generate(tokens):
+    def generate(tokens, **kwargs):
       calls.append(tokens)
       yield from (1, 2)
     with patch.object(model, "generate", generate): model.warmup()
     self.assertEqual(calls, [[0], [0]])
 
+  def test_warmup_then_generate_with_default_chunk(self):
+    # warmup must not capture JIT graphs that generate()'s default chunk_size then rejects
+    model = Transformer(TEST_CONFIG)
+    model.warmup()
+    self.assertIsInstance(next(model.generate([5, 6, 7, 8])), int)
+
   def test_first_recurrent_generate_before_state_init(self):
     model = Transformer(TEST_CONFIG)
     model.has_recurrent_block = True
@@ -38,6 +44,15 @@ class TestTransformerGenerate(unittest.TestCase):
       next(model.generate([1, 2, 3, 4, 5, 42, 10]))
     self.assertEqual(calls, [((1, 1), V_START_POS.bind(5)), ((1, 1), V_START_POS.bind(6))])
 
+  def test_recurrent_divergent_prompt_restarts(self):
+    model, calls = Transformer(TEST_CONFIG), []
+    model.has_recurrent_block, model._cached_tokens = True, [1, 2, 9]
+    def mock_call(self, tokens, start_pos, temperature):
+      calls.append(start_pos)
+      return Tensor([[42]])
+    with patch.object(Transformer, '__call__', mock_call): next(model.generate([1, 2, 10, 11]))
+    self.assertEqual(calls[0], V_START_POS.bind(0))
+
   def test_template_starts_reasoning(self):
     router = StreamRouter(reasoning=True)
     self.assertEqual(list(router.route("reasoning</think>answer")),
@@ -178,6 +193,12 @@ class TestTransformerGenerate(unittest.TestCase):
     # with temperature=2.0, we should see at least 2 distinct outputs across 5 runs
     self.assertGreater(len(runs), 1, "high temperature should produce varied outputs")
 
+  def test_recurrent_temperature_high_produces_variety(self):
+    model = Transformer(TEST_CONFIG)
+    model.has_recurrent_block = True
+    outputs = {model.forward(Tensor([[1]]), 0, Tensor([2.0])).item() for _ in range(5)}
+    self.assertGreater(len(outputs), 1)
+
   def test_temperature_passed_to_forward(self):
     """Temperature from generate should be passed through to __call__."""
     model = Transformer(TEST_CONFIG)

test/unit/test_metal_graph.py

@@ -1,5 +1,4 @@
 import unittest
-from unittest.mock import MagicMock
 from tinygrad import Device
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.dtype import dtypes
@@ -11,36 +10,27 @@ class TestMetalGraph(unittest.TestCase):
     self.MetalGraph = MetalGraph
     self.dev = Device[Device.DEFAULT]
 
-  def metal_buf(self, offset):
-    buf = MagicMock()
-    if offset > 0:
-      buf.op = Ops.SLICE
-      src = MagicMock()
-      src.dtype = dtypes.uint8
-      buf.src = (src, UOp.const(offset))
-      buf.dtype = dtypes.uint8
-    else:
-      buf.op = Ops.BUFFER
-    buf.device = Device.DEFAULT
-    return buf
+  def metal_buf(self, offset, bitcast=False):
+    size = 4 if bitcast else 1
+    buf = UOp.new_buffer(Device.DEFAULT, offset+size, dtypes.uint8)
+    if offset: buf = buf[offset:offset+size]
+    return buf.bitcast(dtypes.float32) if bitcast else buf
 
-  def call(self, *bufs):
-    c = MagicMock()
-    c.src = (MagicMock(op=Ops.PROGRAM),) + tuple(bufs)
-    return c
+  def supports_uop(self, *bufs):
+    return self.MetalGraph.supports_uop([self.dev], UOp(Ops.PROGRAM, src=(UOp.sink(),)).call(*bufs))
 
   def test_supports_uop_normal_offset(self):
-    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF))) is True
+    assert self.supports_uop(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF)) is True
 
   def test_supports_uop_overflow_offset(self):
-    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(0x100000000))) is False
+    assert self.supports_uop(self.metal_buf(0), self.metal_buf(0x100000000)) is False
 
-  def test_supports_uop_nonmetal_buf(self):
-    # non-SLICE ops should not be checked for offset
-    buf = MagicMock()
-    buf.op = Ops.BUFFER
-    buf.device = Device.DEFAULT
-    self.MetalGraph.supports_uop([self.dev], self.call(buf))
+  def test_supports_uop_non_view_buf(self):
+    assert self.supports_uop(self.metal_buf(0)) is True
+
+  def test_supports_uop_bitcast(self):
+    assert self.supports_uop(self.metal_buf(0xFFFFFFFF, bitcast=True)) is True
+    assert self.supports_uop(self.metal_buf(0x100000000, bitcast=True)) is False
 
 if __name__ == "__main__":
   unittest.main()

test/unit/test_multitensor.py

@@ -384,6 +384,12 @@ class TestMultiTensor(unittest.TestCase):
       np.testing.assert_allclose(r.numpy(), np.ones(256)+np.ones(256), atol=1e-4, rtol=1e-5)
     assert jf.captured is not None
 
+  def test_symbolic_broadcast_copy(self):
+    rows = Variable("rows", 1, 4).bind(3)
+    out = Tensor.ones(rows, 8).to(devices_2).realize()
+    self.assertEqual(out.shape, (rows, 8))
+    np.testing.assert_equal(out[:3].to(Device.DEFAULT).numpy(), np.ones((3, 8)))
+
   def test_multitensor_jit_in_list(self):
     # test MULTI tensor inside a list container - exercises the container unpacking + MULTI unpacking
     @TinyJit

test/unit/test_schedule_cache.py

@@ -1,11 +1,16 @@
 import unittest
 import functools
-from tinygrad import Tensor, Variable, UOp
+from tinygrad import Tensor, Variable, UOp, function
 from tinygrad.uop.ops import KernelInfo
 from tinygrad.schedule import schedule_cache
 
-def custom_set0_kernel(A:UOp, num:int) -> UOp:
-  return A[0].set(num).sink(arg=KernelInfo(f"custom_set0_{num}"))
+def custom_add_kernel(A:UOp, B:UOp, num:int=0) -> UOp:
+  return A[0].set(B[0] + num).sink(arg=KernelInfo(f"custom_add_{num}"))
+
+def custom_add_backward(grad_output:UOp, _) -> tuple[None, UOp]:
+  grad = Tensor.invalids(*grad_output.shape, dtype=grad_output.dtype, device=grad_output.device)
+  grad = Tensor.custom_kernel(grad, Tensor(grad_output, device=grad_output.device), fxn=functools.partial(custom_add_kernel, num=0))[0]
+  return None, grad.uop
 
 class TestScheduleCache(unittest.TestCase):
   def test_bound_variable_reuses_cache(self):
@@ -25,27 +30,27 @@ class TestScheduleCache(unittest.TestCase):
 
   def test_custom_kernel(self):
     for i in range(4):
-      a = Tensor.empty(1)
-      a = Tensor.custom_kernel(a, fxn=functools.partial(custom_set0_kernel, num=i))[0]
+      a, b = Tensor.empty(1), Tensor.ones(1)
+      a = Tensor.custom_kernel(a, b, fxn=functools.partial(custom_add_kernel, num=i))[0]
       a.realize()
-      self.assertEqual(a.item(), i)
+      self.assertEqual(a.item(), i+1)
 
   def test_same_custom_function_reuses_cache(self):
     schedule_cache.clear()
-    fxn = functools.partial(custom_set0_kernel, num=10)
+    fxn = functools.partial(custom_add_kernel, num=10)
 
     # first run
-    a = Tensor.empty(1)
-    a = Tensor.custom_kernel(a, fxn=fxn)[0]
+    a, x = Tensor.empty(1), Tensor.ones(1)
+    a = Tensor.custom_kernel(a, x, fxn=fxn)[0]
     a.realize()
-    self.assertEqual(a.item(), 10)
+    self.assertEqual(a.item(), 11)
     cache_size_after_first = len(schedule_cache)
 
     # second run with same function should reuse cache
-    b = Tensor.empty(1)
-    b = Tensor.custom_kernel(b, fxn=fxn)[0]
+    b, x = Tensor.empty(1), Tensor.ones(1)
+    b = Tensor.custom_kernel(b, x, fxn=fxn)[0]
     b.realize()
-    self.assertEqual(b.item(), 10)
+    self.assertEqual(b.item(), 11)
     self.assertEqual(len(schedule_cache), cache_size_after_first)
 
   def test_simple(self):
@@ -65,5 +70,30 @@ class TestScheduleCache(unittest.TestCase):
       print(num)
     self.assertEqual(len(schedule_cache), start_len_schedule_cache)
 
+  @unittest.expectedFailure
+  def test_simple_precompile(self):
+    @function(precompile=True, precompile_backward=True)
+    def f(x:Tensor) -> Tensor:
+      out = Tensor.invalids(*x.shape, dtype=x.dtype, device=x.device)
+      out = Tensor.custom_kernel(out, x, fxn=functools.partial(custom_add_kernel, num=10), grad_fxn=custom_add_backward)[0]
+      return out + x
+
+    # warmup
+    x = Tensor.ones(1).realize()
+    out = f(x)
+    out.backward(x)
+    self.assertEqual(out.item(), 12)
+    self.assertEqual(x.grad.item(), 2)
+
+    # use the cache next time function is called
+    start_len_schedule_cache = len(schedule_cache)
+    for _ in range(3):
+      x = Tensor.ones(1).realize()
+      out = f(x)
+      out.backward(x)
+      self.assertEqual(out.item(), 12)
+      self.assertEqual(x.grad.item(), 2)
+    self.assertEqual(len(schedule_cache), start_len_schedule_cache)
+
 if __name__ == "__main__":
   unittest.main()

tinygrad/codegen/__init__.py

@@ -2,8 +2,8 @@ from dataclasses import replace, dataclass
 import itertools, functools
 from tinygrad.helpers import DISABLE_FAST_IDIV, TRANSCENDENTAL, SPEC, DEBUG, VIZ, IMAGE, NOOPT, EMULATED_DTYPES, NOLOCALS, USE_TC
 from tinygrad.helpers import ALLOW_TF32, DEFAULT_FLOAT, DEFAULT_INT, TracingKey, Context, panic
-from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype, Ops, UPat, rewrite_group, KernelInfo, ProgramInfo, GroupOp
-from tinygrad.uop.ops import AxisType, pm_commit_weak, pm_cast_weak
+from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, Ops, UPat, rewrite_group, KernelInfo, ProgramInfo, GroupOp, AxisType
+from tinygrad.uop.weak import pm_lower_index_dtype, pm_commit_weak, pm_cast_weak
 from tinygrad.uop.render import pyrender
 from tinygrad.uop.spec import type_verify, spec_tensor, spec_program
 from tinygrad.renderer import Renderer, Estimates
@@ -12,7 +12,7 @@ from tinygrad.dtype import dtypes, AddrSpace
 
 # import all pattern matchers here
 from tinygrad.codegen.gpudims import pm_add_gpudims
-from tinygrad.uop.symbolic import sym, symbolic_simple, symbolic, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
+from tinygrad.uop.symbolic import sym, symbolic_simple, symbolic, pm_fold_cast_const, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
 from tinygrad.uop.movement import mop_cleanup
 from tinygrad.codegen.decomp.dtype import pm_dtype_decomps
 from tinygrad.codegen.decomp.op import get_late_rewrite_patterns, get_simplifying_rewrite_patterns
@@ -20,7 +20,7 @@ from tinygrad.codegen.decomp.transcendental import get_transcendental_patterns
 from tinygrad.codegen.late.coalesce import indexing_simplify
 from tinygrad.codegen.opt.postrange import apply_opts
 from tinygrad.codegen.late.gater import pm_move_gates_from_index
-from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse
+from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse, pm_reduce_unparented
 from tinygrad.schedule.multi import multi_pm
 from tinygrad.schedule.rangeify import pm_mops
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
@@ -310,7 +310,8 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
     sink = apply_opts(sink, ren, beam=ast.arg.beam)
 
   # ** expander (expand_rewrite) **
-  sink = graph_rewrite(sink, sym+pm_move_where_on_load+pm_flatten_range, name="postopt symbolic")
+  # reduce_unparented: a REDUCE whose src folded to a CONST (e.g. x*0) has no parented ranges, collapse it before the expander
+  sink = graph_rewrite(sink, sym+pm_move_where_on_load+pm_flatten_range+pm_reduce_unparented, name="postopt symbolic")
 
   # expand
   sink = graph_rewrite(sink, expander2, ctx=build_range_map(sink), name="expander")
@@ -336,14 +337,16 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
 
   # do memory coalescing (late)
   sink = memory_coalescing(sink, ren)
-  sink = graph_rewrite(sink, symbolic_simple+ew_devectorizer+pm_simplify_add_image, name="add images", ctx=({}, ren), bottom_up=True)
+  sink = graph_rewrite(sink, symbolic_simple+ew_devectorizer+pm_simplify_add_image,
+                       name="add images", ctx=({}, ren), bottom_up=True)
 
   # extra symbolic before decomp. crashes without this?
-  sink = graph_rewrite(sink, sym, name="extra symbolic")
+  # NOTE: also run indexing_simplify here, while the index is still weakint and (x+y)*c -> x*c+y*c applies
+  sink = graph_rewrite(sink, sym+indexing_simplify, name="extra symbolic")
 
   # lower index dtype
   # NOTE: we need indexing_simplify to remove the cast to long using the Invalid
-  sink = graph_rewrite(sink, pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")
+  sink = graph_rewrite(sink, symbolic_simple+pm_fold_cast_const+pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")
 
   # final symbolic before decomp
   sink = graph_rewrite(sink, symbolic, name="final symbolic")

tinygrad/codegen/decomp/dtype.py

@@ -99,7 +99,8 @@ def f2f(v, fr:DType, to:DType, sat=True):
     if fr in dtypes.fp8_fnuz:
       fnuz_nan = sign.ne(0) & nosign.eq(0)
       qnan = shl(shl(1, te) - 1, tm) | shl(1, tm - 1)
-      return fnuz_nan.where(qnan, sign | exp.eq(0).where(0, norm)).bitcast(to)
+      # the fnuz bias can exceed the target's: exp in [1, fb-tb] is normal in fr but lands below to's normal range, so it flushes like a denormal
+      return fnuz_nan.where(qnan, sign | (exp < max(fb - tb, 0) + 1).where(0, norm)).bitcast(to)
     # fp8e4m3 has only one nan
     is_nan = (nosign.eq(shl(1, fm + fe) - 1) if fr == dtypes.fp8e4m3 else exp.eq(shl(1, fe) - 1))
     return (sign | exp.eq(0).where(0, is_nan.where(nan, norm))).bitcast(to)

tinygrad/codegen/gpudims.py

@@ -57,7 +57,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
 
   # get the idxs
   ki: KernelInfo = s.arg
-  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int).cast(dtypes.weakint)]
+  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int, param=True).cast(dtypes.weakint)]
   elif ki.dont_use_locals:
     assert not local_dims, "can't use locals if there's no local dims"
     idxs = get_grouped_dims("idx", global_shape, ctx.global_max, reverse=True)
@@ -89,7 +89,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
 
 pm_device_to_var = PatternMatcher([
   # the DEVICE axis is not a program axis, it's bound per device at launch. lower it to the _device_num variable (like SPECIAL for devices)
-  (UPat(Ops.RANGE, name="r"), lambda r: UOp.variable("_device_num", 0, r.vmax, dtype=r.dtype) if r.arg[-1] is AxisType.DEVICE else None),
+  (UPat(Ops.RANGE, name="r"), lambda r: UOp.variable("_device_num", 0, r.vmax, dtype=r.dtype, param=True) if r.arg[-1] is AxisType.DEVICE else None),
   # ENDs that closed a DEVICE range no longer close it
   (UPat(Ops.END, name="e"), lambda e: e.replace(src=(e.src[0],)+tuple(s for s in e.src[1:] if s.op is not Ops.PARAM))
    if any(s.op is Ops.PARAM and s.arg.name == '_device_num' for s in e.src[1:]) else None),

tinygrad/codegen/late/coalesce.py

@@ -1,8 +1,8 @@
 import itertools, functools
 from collections import defaultdict
 from tinygrad.dtype import dtypes, AddrSpace, Invalid, DType
-from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp, shape_to_shape_arg
-from tinygrad.uop.symbolic import uop_given_valid, parse_valid, invalid_gate
+from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp, shape_to_shape_arg, graph_rewrite
+from tinygrad.uop.symbolic import uop_given_valid, parse_valid, invalid_gate, sym
 from tinygrad.helpers import getenv, IMAGE, OSX, ceildiv, is_image_shape
 from tinygrad.renderer import Renderer
 
@@ -26,12 +26,15 @@ def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
     # check if idx is out of bound when X is on the wrong side of the bound: X in [c+1, vmax] or [vmin, c-1]
     lo, hi = (c + 1, X.vmax) if is_upper_bound else (X.vmin, c - 1)
     if lo <= hi:
-      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype)
-      for coord,b in zip(idx.src, (width, height)):
-        rw = coord.substitute({X:fake}).simplify()
-        if rw.vmin >= b or rw.vmax < 0:
-          drop_stmt.append(stmt)
-          break
+      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype, param=True)
+      subs = [{X: fake}]
+      # idx may not have X itself, so also substitute a term of X: v -> fake - (X - v)
+      terms = list(X.split_uop(Ops.ADD))
+      v = next((u for u in terms if u.op in GroupOp.Irreducible and u.op is not Ops.CONST), None)
+      if v is not None and (rest:=[u for u in terms if u is not v]): subs.append({v: fake - UOp.usum(*rest)})
+      if any((testidx:=graph_rewrite(coord.substitute(sub), sym)).vmin >= b or testidx.vmax < 0
+             for sub in subs for coord,b in zip(idx.src, (width, height))):
+        drop_stmt.append(stmt)
   return drop_stmt
 
 def simplify_valid_load(buf:UOp, start_idx:UOp, valid:UOp) -> UOp|None:

tinygrad/codegen/late/gater.py

@@ -3,7 +3,7 @@ from tinygrad.uop.ops import PatternMatcher, UPat, Ops
 from tinygrad.dtype import Invalid, dtypes
 
 def move_where_load(gate, l, a, w):
-  return l.replace(src=(l.src[0], l.vconst_like(0) if a.is_invalid else
+  return l.replace(src=(l.src[0], l.vconst_like(0) if a.is_invalid else l.const_like(a.val) if a.op is Ops.CONST else
     a.src[0] if a.op is Ops.CAST and a.src[0].dtype == l.dtype else a.cast(l.dtype), l.src[2])).cast(w.dtype)
 
 pm_move_gates_from_index = PatternMatcher([

tinygrad/codegen/opt/postrange.py

@@ -332,9 +332,9 @@ class Scheduler:
   @property
   def group_for_reduces(self) -> int: return len(self.axes_of(AxisType.GROUP_REDUCE))
 
-def bufs_from_ast(ast:UOp, dname:str) -> list[Buffer]:
+def args_from_ast(ast:UOp, dname:str) -> tuple[list[Buffer], dict[str, int]]:
   glbls = sorted([x for x in ast.backward_slice if x.op is Ops.PARAM and x.arg.slot >= 0], key=lambda x: x.arg.slot)
-  return [Buffer(dname, x.max_numel(), x.dtype) for x in glbls]
+  return [Buffer(dname, x.max_numel(), x.dtype) for x in glbls], {k.expr:int(k.vmax+k.vmin)//2 for k in ast.variables()}
 
 def apply_opts(ast:UOp, ren:Renderer, beam:int=0) -> UOp:
   if ast.tag is not None: return ast
@@ -344,10 +344,10 @@ def apply_opts(ast:UOp, ren:Renderer, beam:int=0) -> UOp:
     for opt in ast.arg.opts_to_apply: k.apply_opt(opt)
   elif beam >= 1:
     from tinygrad.codegen.opt.search import beam_search
-    rawbufs = bufs_from_ast(ast, ren.target.device)
+    rawbufs, var_vals = args_from_ast(ast, ren.target.device)
     # beam search may open devices
     with Context(ALLOW_DEVICE_USAGE=1):
-      k = beam_search(k, rawbufs, beam, bool(getenv("BEAM_ESTIMATE", 1)))
+      k = beam_search(k, rawbufs, var_vals, beam, bool(getenv("BEAM_ESTIMATE", 1)))
   elif not NOOPT and (ast.arg is None or ast.arg.applied_opts == ()):
     from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
     # NOTE: hand_coded_optimizations doesn't support multiblock opts yet

tinygrad/codegen/opt/search.py

@@ -1,6 +1,6 @@
 import math, time, multiprocessing, traceback, signal, atexit
 from dataclasses import replace
-from tinygrad.uop.ops import sym_infer, AxisType, UOp
+from tinygrad.uop.ops import sym_infer, AxisType, UOp, Ops
 from tinygrad.uop.render import pyrender
 from tinygrad.device import Device, Buffer
 from tinygrad.helpers import prod, flatten, DEBUG, CACHELEVEL, diskcache_get, diskcache_put, getenv, Context, colored, time_to_str
@@ -62,7 +62,8 @@ def _try_compile(x:tuple[int,Scheduler]) -> tuple[int, tuple[UOp, float]|None]:
   ret = None
   try:
     st = time.perf_counter()
-    prg = to_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].ren)
+    ast, dev = x[1].copy().get_optimized_ast(name_override="test"), x[1].ren.target.device
+    prg = to_program(ast.substitute({p: p.replace(arg=replace(p.arg, device=dev)) for p in ast.toposort() if p.op is Ops.PARAM}), x[1].ren)
     et = time.perf_counter() - st
     uops = prg.src[1].src
     if len(uops) >= (uops_max:=getenv("BEAM_UOPS_MAX", 3000)) > 0:
@@ -111,7 +112,7 @@ def get_kernel_actions(s:Scheduler, include_0=True, max_up:int|None=None) -> dic
   return acted
 
 beam_pool, BEAM_DEBUG = None, getenv("BEAM_DEBUG")
-def beam_search(s:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
+def beam_search(s:Scheduler, rawbufs:list[Buffer], var_vals:dict[str,int], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
   global beam_pool
   key = {"ast": s.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": s.ren.target.device, "suffix": s.ren.suffix}
   if not disable_cache and CACHELEVEL >= 1 and (val:=diskcache_get("beam_search", key)) is not None:
@@ -136,7 +137,6 @@ def beam_search(s:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True
 
   try:
     rawbufs = _ensure_buffer_alloc(rawbufs)
-    var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in s.ast.variables()}
     exiting, st = False, time.perf_counter()
     dev = Device[s.ren.target.device]
     while not exiting:

tinygrad/codegen/opt/tc.py

@@ -1,6 +1,7 @@
 import math, functools
 from dataclasses import dataclass
 from tinygrad.dtype import DType, dtypes
+from tinygrad.uop.ops import PatternMatcher, UOp, UPat, Ops
 
 @dataclass(frozen=True)
 class TensorCore: # D = A * B + C, A is (M x K), B is (K x N), C and D are (M x N)
@@ -135,6 +136,42 @@ amd_cdna4 = amd_cdna_1616128 + amd_cdna_161632 + amd_cdna_161616
 
 def get_amd(arch): return {"gfx942": amd_cdna3, "gfx950": amd_cdna4, "gfx1200": amd_rdna4, "gfx1201": amd_rdna4}.get(arch, amd_rdna3)
 
+pm_validate_wmma_rdna3 = PatternMatcher([
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.int32), lambda x: x.replace(
+    src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
+    if x.src[0].dtype == dtypes.int8 and x.src[0].max_numel() == 16 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.half), lambda x: UOp(Ops.STACK, src=tuple(x.replace(
+      src=(x.src[0], x.src[1], UOp(Ops.STACK, src=tuple(x.src[2].index(UOp.const(j//2, dtypes.int16))
+      if j%2 == 0 else UOp.const(0.0, x.src[2].dtype)
+      for j in range(x.max_numel()*2)))),
+      arg=(*x.arg[:4], None)).index(UOp.const(i*2, dtypes.int16))
+      for i in range(x.max_numel()))) if x.max_numel() == 8 else None),
+  (UPat(Ops.WMMA, name="x"), lambda x: x.replace(
+    src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
+    if x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 16 else None),
+])
+
+pm_validate_wmma_rdna4 = PatternMatcher([
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.bfloat16), lambda x: x.replace(
+    dtype=dtypes.uint16,
+    src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2].bitcast(dtypes.uint16)))
+      .bitcast(dtypes.bfloat16) if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
+    if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None)
+])
+
+pm_validate_wmma_cdna = PatternMatcher([
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
+    if x.arg[0][2] == 128 and x.src[0].dtype.itemsize <= 8 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
+    if x.max_numel() == 4 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 4 else None),
+  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
+    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64), x.src[2]))
+    if x.max_numel() == 4 and x.src[0].dtype in dtypes.fp8_ocp and x.src[0].max_numel() == 8 else None),
+])
 # ***** Apple Metal *****
 
 metal = [TensorCore(dims=(8,8,8), threads=32, elements_per_thread=(2,2,2), dtype_in=di, dtype_out=do,

tinygrad/codegen/simplify.py

@@ -137,7 +137,7 @@ def reduce_collapse(red:UOp, u:UOp, pm:PatternMatcher=pm_reduce_collapse) -> UOp
     for u in included:
       for s in u.src:
         if s in included or s in replaces or s.op in {Ops.CONST, Ops.PARAM, Ops.BUFFER}: continue
-        replaces[s] = UOp.variable(f'in{len(replaces)}', s.vmin, s.vmax, s.dtype)
+        replaces[s] = UOp.variable(f'in{len(replaces)}', s.vmin, s.vmax, s.dtype, param=True)
     collapse_fxn = u.substitute(replaces).reduce(r, arg=Ops.ADD)
     sink = graph_rewrite(collapse_fxn, pm, name="reduce_collapse")
     if not no_range(sink): return None

tinygrad/device.py

@@ -336,6 +336,8 @@ class Compiled:
   pm_lower:Any = None
   pm_bufferize:Any = None
 
+  has_copy_queue:bool = True
+
   def __init__(self, device:str, allocator:Allocator, renderers:list[type[Renderer]], runtime:type[Program[Self]]|None, graph=None, arch=None):
     from tinygrad.renderer import Renderer
     self.device, self.allocator, self.runtime_t, self.graph, self.renderers = device, allocator, runtime, graph, renderers or [Renderer]

tinygrad/dtype.py

@@ -80,7 +80,7 @@ class DType(metaclass=DTypeMetaClass):
     # NOTE: float('nan') != float('nan'), so we canonicalize here
     if isinstance(val, float) and math.isnan(val): val = math.nan
     # int is the default. wrap floats in ConstFloat to distinguish -0.0 from 0.0 in cache
-    return ConstFloat(float(val)) if dtypes.is_float(self) else bool(val) if dtypes.is_bool(self) else int(val)
+    return ConstFloat(truncate.get(self, float)(float(val))) if dtypes.is_float(self) else bool(val) if dtypes.is_bool(self) else int(val)
 
 
 class DTypes:

tinygrad/engine/jit.py

@@ -44,9 +44,7 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
     current_batch, current_batch_devs = [], []
 
   for si in linear.src:
-    if si.src[0].op is Ops.SLICE: continue
-
-    devs = dedup([Device[x] for b in si.src[1:] if b.op is not Ops.BIND for x in (b.device if isinstance(b.device, tuple) else (b.device,))])
+    devs = dedup([Device[x] for b in si.src[1:] if not b.is_bound_var for x in (b.device if isinstance(b.device, tuple) else (b.device,))])
     graph_t = graph_class(devs[0]) if devs[0].graph is not None else None
 
     can_graph = graph_t is not None and graph_t.supports_uop(devs, si)
@@ -180,7 +178,7 @@ class CapturedJit(Generic[ReturnType]):
       if call.op is not Ops.CALL: continue
       arg_uops = get_call_arg_uops(call)
       outs, ins = get_call_outs_ins(call)
-      out |= {arg_uops[k] for k in set(outs) - set(ins) if arg_uops[k].op in (Ops.BUFFER, Ops.SLICE)}
+      out |= {b for k in set(outs) - set(ins) if (b:=u if (cv:=(u:=arg_uops[k]).contiguous_view()) is None else cv[0]).op is Ops.BUFFER}
     return out
 
   def __call__(self, input_uops:list[UOp], var_vals:dict[str, int]) -> ReturnType:

tinygrad/engine/realize.py

@@ -3,21 +3,21 @@ from typing import cast, Iterator, Any, Sequence
 import time, random, itertools, math, contextlib, weakref, array
 from dataclasses import dataclass, replace, field
 from tinygrad.helpers import colored, DEBUG, GlobalCounters, ansilen, all_int, prod, flatten, Context, getenv, to_tuple
-from tinygrad.helpers import BEAM, size_to_str, time_to_str, VALIDATE_WITH_CPU, PROFILE, ProfilePointEvent, cpu_events
-from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, AxisType, sym_infer, buffers, graph_rewrite
-from tinygrad.device import Device, Buffer, MultiBuffer
+from tinygrad.helpers import BEAM, size_to_str, time_to_str, VALIDATE_WITH_CPU, PROFILE, ProfilePointEvent, cpu_events, wait_cond
+from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, AxisType, sym_infer, graph_rewrite
+from tinygrad.device import Device, Buffer, MultiBuffer, ProfileGraphEntry
 from tinygrad.renderer import Estimates
 from tinygrad.codegen import to_program
-from tinygrad.codegen.opt.postrange import bufs_from_ast
+from tinygrad.codegen.opt.postrange import args_from_ast
 
 # **************** Helpers ****************
 
-def get_call_arg_uops(call:UOp) -> tuple[UOp, ...]: return tuple(s for s in call.src[1:] if s.op is not Ops.BIND)
+def get_call_arg_uops(call:UOp) -> tuple[UOp, ...]: return tuple(s for s in call.src[1:] if not s.is_bound_var)
 
 def get_call_outs_ins(call:UOp) -> tuple[tuple[int, ...], tuple[int, ...]]:
   ast = call.src[0]
   if ast.op is Ops.PROGRAM: return tuple(ast.arg.outs), tuple(ast.arg.ins)
-  if ast.op in (Ops.COPY, Ops.SLICE): return (0,), (1,)
+  if ast.op is Ops.COPY: return (0,), (1,)
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return (0,), tuple(range(1, len(get_call_arg_uops(call))))
   return (), ()
 
@@ -27,9 +27,6 @@ def get_call_name(call:UOp, bufs:Sequence[Buffer|UOp], var_vals:dict[str, int]|N
 
   ast, arg_uops = call.src[0], get_call_arg_uops(call)
   if ast.op is Ops.PROGRAM: return ast.arg.name
-  if ast.op is Ops.SLICE:
-    offset = ast.src[1].val * arg_uops[1].dtype.itemsize
-    return colored(f"view {_uop_sz_to_str(arg_uops[0]):>10} @ {offset:<10d}", "yellow")
   if ast.op is Ops.COPY: return colored(f"copy {_uop_sz_to_str(arg_uops[0]):>10}, {_dev_str(bufs[0]):>7s} <- {_dev_str(bufs[1]):7s}", "yellow")
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return colored(f"enc/dec {_uop_sz_to_str(arg_uops[0])}", "yellow")
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph": return colored(f"batched {len(ast.src[0].src)}", "cyan")
@@ -90,12 +87,13 @@ def optimize_local_size(call:UOp, prg:UOp) -> UOp|None:
 
   if (local_size:=local_size_cache.get(prg.key)) is None:
     # reuse one loaded runtime across candidates, only launch dims vary
-    bufs, runtime = [b.allocate() for b in bufs_from_ast(prg.src[0], device)], get_runtime(device, prg, cache=False)
+    (bufs, var_vals), runtime = args_from_ast(prg.src[0], device), get_runtime(device, prg, cache=False)
+    bufs  = [b.allocate() for b in bufs]
     def try_exec(local_size):
       try:
         new_gs = tuple(g//l if g%l == 0 else g/l for g,l in zip(prg.arg.global_size, local_size))
         return runtime(*[bufs[i].get_buf(device) for i in prg.arg.globals], global_size=new_gs, local_size=(*local_size,),
-                       vals=prg.arg.vals({}), wait=True)
+                       vals=prg.arg.vals(var_vals), wait=True)
       except Exception: return float('inf')
 
     MAX_WORKGROUP = 1024
@@ -139,7 +137,7 @@ class ExecContext:
   cache: bool = True
 
 def _resolve(b:UOp, inputs:tuple[UOp, ...]) -> UOp:
-  if b.op in (Ops.SLICE, Ops.MSELECT) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg.slot], *b.src[1:]))
+  if b.op in (Ops.MSELECT, Ops.SHRINK) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg.slot], *b.src[1:]))
   if b.op is Ops.MSTACK: return b.replace(src=tuple(_resolve(x, inputs) for x in b.src))
   return inputs[b.arg.slot] if b.op is Ops.PARAM else b
 def resolve_params(call:UOp, inputs:tuple[UOp, ...]) -> list[UOp]: return [_resolve(b, inputs) for b in get_call_arg_uops(call)]
@@ -153,13 +151,6 @@ def unwrap_multi(call:UOp, resolved:list[UOp]) -> Iterator[tuple[list[Buffer], d
                    for x in call.src[0].toposort())
     for j, per_dev in enumerate(zip(*[cast(MultiBuffer, b).bufs for b in bufs])): yield list(per_dev), {"_device_num": j} if has_dnum else {}
 
-def exec_view(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
-  resolved = resolve_params(call, ctx.input_uops)
-  bufs = [cast(Buffer, b.buffer) for b in resolved]
-  bv = bufs[1].view(resolved[0].max_numel(), ast.dtype, ast.src[1].val*bufs[1].dtype.itemsize)
-  with track_stats(ctx, call, bv.device, [bv, bufs[1]], ctx.var_vals): buffers[resolved[0]] = bv
-  return None
-
 def exec_copy(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
   for bufs, device_vars in unwrap_multi(call, resolve_params(call, ctx.input_uops)):
     dest, src = bufs[0].ensure_allocated(), bufs[1].ensure_allocated()
@@ -209,21 +200,30 @@ def exec_graph(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
   return t[0]
 
 def exec_hcq(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
-  if (inputs:=call.arg.aux.inputs) is not None:
+  if (info:=call.arg.aux).inputs is not None:
     bufs = [_resolve(ctx.input_uops[i], ctx.input_uops).buffer for i in call.arg.aux.input_idxs]
-    table = call.src[1+inputs].buffer
+    table = call.src[1+info.inputs].buffer
     for j,dev in enumerate(call.arg.aux.device):
       addrs = array.array('Q', [(b.bufs[j] if isinstance(b, MultiBuffer) else b).get_buf(dev).va_addr for b in bufs])
-      buf = table.bufs[j] if isinstance(table, MultiBuffer) else table
-      buf.ensure_allocated()._buf.cpu_view().view(fmt='Q')[:len(addrs)] = addrs
+      mv = (table.bufs[j] if isinstance(table, MultiBuffer) else table).ensure_allocated()._buf.cpu_view().view(fmt='Q')
+      wait_cond(lambda: mv[0], value=0, timeout_ms=ctx.timeout or getenv("HCQDEV_WAIT_TIMEOUT_MS", 30000), msg=f"{dev} hang detected")
+      mv[:len(addrs)] = addrs
 
-  exec_kernel(replace(ctx, update_stats=False), call, ast)
+  exec_kernel(replace(ctx, update_stats=DEBUG>=3), call, ast)
 
-  st = time.perf_counter()
-  for d in call.arg.aux.device:
-    with track_stats(ctx, call, d, [], ctx.var_vals):
-      if ctx.wait: cast(Any, Device[d]).synchronize(timeout=ctx.timeout)
-  return time.perf_counter() - st
+  tms = []
+  for devices,name,estimates,prof in info.kernels:
+    for device in devices:
+      tm = None
+      if prof:
+        (d:=cast(Any, Device[device])).prof_ents[prof[0]] = ProfileGraphEntry(device, name, *prof)
+        if ctx.wait:
+          d.synchronize(timeout=ctx.timeout)
+          st, en = (d.signal(x)._buf.cpu_view().view(fmt='Q')[0] for x in prof)
+          tms.append(tm:=float(en-st)/d.timestamp_divider/1e6)
+      stat_call = call.replace(arg=replace(call.arg, name=name, aux=replace(info, estimates=estimates, kernels=())))
+      with track_stats(ctx, stat_call, device, [], ctx.var_vals) as et: et[0] = tm
+  return max(tms) if tms else None
 
 # flatten LINEAR-in-LINEAR: any nested LINEAR child gets inlined into its parent's src
 pm_flatten_linear = PatternMatcher([
@@ -254,7 +254,6 @@ pm_optimize_local_size = PatternMatcher([
 ])
 
 pm_exec = PatternMatcher([
-  (UPat(Ops.CALL, src=(UPat(Ops.SLICE, name="ast"),), name="call", allow_any_len=True), exec_view),
   (UPat(Ops.CALL, src=(UPat(Ops.COPY, name="ast"),), name="call", allow_any_len=True), exec_copy),
   (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="ast"),), name="call", allow_any_len=True), exec_kernel),
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="encdec", name="ast"),), name="call", allow_any_len=True), exec_encdec),
@@ -265,11 +264,11 @@ pm_exec = PatternMatcher([
 
 if getenv("HCQ2"): from tinygrad.runtime.support.hcq2 import hcq_compile, hcq_link # noqa: E402 # down here, hcq2 imports the helpers above
 
-def compile_linear(linear:UOp, beam:int|None=None, validate=False, input_uops:list[UOp]|None=None) -> UOp:
+def compile_linear(linear:UOp, beam:int|None=None, validate=False, input_uops:list[UOp]|None=None, profile:bool|None=None) -> UOp:
   if validate: linear = graph_rewrite(linear, pm_validate, name="validate", walk=True)
   if (beam_val:=BEAM.value if beam is None else beam) >= 1: linear = graph_rewrite(linear, pm_beam, ctx=beam_val, walk=True)
   linear = graph_rewrite(linear, pm_compile, name="precompile kernels", walk=True)
-  if getenv("HCQ2"): linear = hcq_compile(linear, input_uops)
+  if getenv("HCQ2"): linear = hcq_compile(linear, input_uops, bool(PROFILE or DEBUG >= 2) if profile is None else profile)
   return graph_rewrite(linear, pm_optimize_local_size, name="optimize local size", walk=True)
 
 def link_linear(linear:UOp, cache=True) -> UOp: return hcq_link(linear, cache=cache) if getenv("HCQ2") else linear
@@ -287,5 +286,5 @@ def time_call(call:UOp, var_vals:dict[str, int]|None=None, timeout:int|None=None
       from tinygrad.tensor import Tensor
       with Context(DEBUG=0, BEAM=0, CAPTURING=0, TRACK_MATCH_STATS=0): Tensor.ones(1024, 1024).contiguous().realize(do_update_stats=False)
   ctx = ExecContext(var_vals or {}, update_stats=False, wait=True, timeout=timeout, cache=False)
-  linear = link_linear(compile_linear(UOp(Ops.LINEAR, src=(call,)), beam=0), cache=ctx.cache)
+  linear = link_linear(compile_linear(UOp(Ops.LINEAR, src=(call,)), beam=0, profile=True), cache=ctx.cache)
   return max(pm_exec.rewrite(c, ctx) or 0.0 for c in linear.src)

tinygrad/function.py

@@ -12,7 +12,7 @@ def add_to_ctx(ctx, x:UOp):
   return ret
 
 pm_ctx = PatternMatcher([
-  (UPat((Ops.BUFFER, Ops.BIND), name="x"), add_to_ctx),
+  (UPat(Ops.BUFFER, name="x"), add_to_ctx),
   (UPat((Ops.AFTER, Ops.CONTIGUOUS), name="x"),
    lambda ctx,x: add_to_ctx(ctx,x) if not x.op_in_backward_slice_with_self(Ops.PARAM) and x.op_in_backward_slice_with_self(Ops.BUFFER) else None),
 ])

tinygrad/helpers.py

@@ -1,9 +1,9 @@
 from __future__ import annotations
 import time
 START_TIME = time.perf_counter()
-import os, functools, platform, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass, gc
+import os, functools, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass, gc
 from collections import defaultdict
-import subprocess, shutil, math, types, copyreg, inspect, importlib, decimal, itertools, difflib
+import shutil, math, types, copyreg, inspect, importlib, decimal, itertools, difflib
 from dataclasses import dataclass, field, replace
 from typing import ClassVar, Iterable, Any, TypeVar, Callable, Sequence, TypeGuard, Iterator, Generic, Generator, cast, overload
 
@@ -13,8 +13,7 @@ U = TypeVar("U")
 def prod(x:Iterable[T]) -> T|int: return functools.reduce(operator.mul, x, 1)
 
 # NOTE: helpers is not allowed to import from anything else in tinygrad
-OSX, WIN = platform.system() == "Darwin", sys.platform == "win32"
-ARCH_X86 = any(x in platform.processor() for x in ("Intel", "i386", "x86_64"))
+OSX, WIN = sys.platform == "darwin", sys.platform == "win32"
 BASEDIR = pathlib.Path(__file__).parent
 
 # fix colors on Windows, https://stackoverflow.com/questions/12492810/python-how-can-i-make-the-ansi-escape-codes-to-work-also-in-windows
@@ -231,7 +230,7 @@ class _DEV(ContextVar):
 
 DEV, DEBUG, BEAM, NOOPT = _DEV("DEV", ""), ContextVar("DEBUG", 0), ContextVar("BEAM", 0), ContextVar("NOOPT", 0)
 IMAGE, FLOAT16, OPENPILOT_HACKS = ContextVar("IMAGE", 0), ContextVar("FLOAT16", 0), ContextVar("OPENPILOT_HACKS", 0)
-JIT, JIT_BATCH_SIZE = ContextVar("JIT", 2 if OSX and ARCH_X86 else 1), ContextVar("JIT_BATCH_SIZE", 32)
+JIT, JIT_BATCH_SIZE = ContextVar("JIT", 1), ContextVar("JIT_BATCH_SIZE", 32)
 CHUNK_SIZE = 2**20  # TinyFS content-addressed store: blob chunk + hash-tree node granularity
 WINO, CAPTURING, TRACEMETA, NO_COLOR = ContextVar("WINO", 0), ContextVar("CAPTURING", 1), ContextVar("TRACEMETA", 1), ContextVar("NO_COLOR", 0)
 TRAINING = ContextVar("TRAINING", 0)
@@ -454,9 +453,9 @@ def _ensure_downloads_dir() -> pathlib.Path:
   if pathlib.Path("/etc/tinybox-release").is_file():
     # try creating dir with sudo
     if not (downloads_dir := pathlib.Path("/raid/downloads")).exists():
-      subprocess.run(["sudo", "mkdir", "-p", downloads_dir], check=True)
-      subprocess.run(["sudo", "chown", "tiny:root", downloads_dir], check=True)
-      subprocess.run(["sudo", "chmod", "775", downloads_dir], check=True)
+      system(f"sudo mkdir -p {downloads_dir}")
+      system(f"sudo chown tiny:root {downloads_dir}")
+      system(f"sudo chmod 775 {downloads_dir}")
     return downloads_dir
   return pathlib.Path(cache_dir) / "downloads"
 
@@ -497,6 +496,7 @@ def fetch_fw(path:str, name:str, sha256:str) -> bytes:
 # *** Exec helpers
 
 def system(cmd:str, **kwargs) -> str:
+  import subprocess
   st = time.perf_counter()
   try: ret = subprocess.check_output(cmd.split(), stderr=subprocess.STDOUT, **kwargs).decode().strip()
   except subprocess.CalledProcessError as e:

tinygrad/llm/model.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 import functools, itertools, pathlib
 from dataclasses import dataclass, replace
-from tinygrad import Tensor, nn, UOp, TinyJit, getenv, function
+from tinygrad import Tensor, nn, UOp, TinyJit, getenv, function, dtypes
 from tinygrad.nn import Linear
 from tinygrad.llm.gguf import gguf_load
 from tinygrad.uop.ops import resolve
@@ -187,8 +187,8 @@ class TransformerBlock(FFNBlock):
 
   def _init_state(self, x:Tensor):
     if not hasattr(self, "cache_kv"):
-      # TODO: how is the dtype of this determined?
-      self.cache_kv = Tensor.empty(2, x.shape[0], self.config.n_kv_heads, self.config.max_context, self.config.head_dim, device=x.device)
+      self.cache_kv = Tensor.empty(2, x.shape[0], self.config.n_kv_heads, self.config.max_context, self.config.head_dim,
+                                   dtype=dtypes.default_float, device=x.device)
       self.freqs_cis = precompute_freqs_cis(self.config.rope_dim, self.config.max_context, self.config.rope_theta, device=x.device)
 
 class MLATransformerBlock(FFNBlock):
@@ -261,13 +261,14 @@ class GatedDeltaNetBlock(FFNBlock):
   def _attention(self, x:Tensor, start_pos:int|UOp) -> Tensor:
     B, T, _ = x.shape
     assert T == 1, "GatedDeltaNetBlock currently only supports T=1"
+    is_kda = hasattr(self, "ssm_g_a")
 
     # input processing
     x = x.half()
-    out_gate = self.ssm_g_b(self.ssm_g_a(x)) if hasattr(self, "ssm_g_a") else self.attn_gate(x)
+    out_gate = self.ssm_g_b(self.ssm_g_a(x)) if is_kda else self.attn_gate(x)
     out_gate = out_gate.reshape(B, 1, self.num_v_heads, self.head_v_dim)
     beta = self.ssm_beta(x).sigmoid().reshape(B, self.num_v_heads, 1, 1)
-    alpha = self.ssm_f_b(self.ssm_f_a(x)) if hasattr(self, "ssm_f_a") else self.ssm_alpha(x)
+    alpha = self.ssm_f_b(self.ssm_f_a(x)) if is_kda else self.ssm_alpha(x)
     alpha = ((alpha.float() + self.ssm_dt["bias"]).softplus().reshape(B, self.num_v_heads, -1) *
              self.ssm_a.reshape(1, self.num_v_heads, -1)).exp().unsqueeze(-2)
 
@@ -291,14 +292,13 @@ class GatedDeltaNetBlock(FFNBlock):
 
     # output
     core_attn_out = self.ssm_norm((recurrent_state@q).squeeze(-1).reshape(B, 1, self.num_v_heads, self.head_v_dim))
-    out_gate = out_gate.sigmoid() if hasattr(self, "ssm_g_a") else out_gate.silu()
+    out_gate = out_gate.sigmoid() if is_kda else out_gate.silu()
     return self.ssm_out((core_attn_out * out_gate).reshape(B, 1, -1).cast(x.dtype))
 
   # recurrent state can't be partially reused after divergence, force a full rebuild
   def _state_reset_ops(self):
     return [self.conv_state.assign(self.conv_state.const_like(0)),
             self.recurrent_state.assign(self.recurrent_state.const_like(0))] if hasattr(self, "conv_state") else []
-  def _reusable_prefix_len(self, prefix_len:int, cached_len:int) -> int: return 0 if prefix_len != cached_len else prefix_len
 
   def _init_state(self, x):
     if not hasattr(self, "conv_state"):
@@ -326,7 +326,8 @@ class Transformer:
   def forward(self, tokens:Tensor, start_pos:int|UOp, temperature:Tensor) -> Tensor:
     x = self.token_embd(tokens).float()                   # (B, T, D)
     for block in self.blk: x = block(x, start_pos)
-    logits = self.output(self.output_norm(x))[:, -1, :]
+    # only run the output projection on the last token
+    logits = self.output(self.output_norm(x[:, -1:]))[:, -1, :]
     # Gumbel-max trick: argmax(logits/temp - log(-log(uniform))) is equivalent to sampling from softmax(logits/temp)
     return (logits / temperature.maximum(1e-12) - (Tensor.rand_like(logits).maximum(1e-12).log().neg()).log()).argmax(-1, keepdim=True)
 
@@ -420,6 +421,10 @@ class Transformer:
     for _ in range(2): list(zip(range(2), self.generate([0])))
 
   def get_start_pos(self, tokens:list[int]) -> int:
+    # recurrent state can't be partially reused after divergence: reuse it only when tokens extend the cached prefix
+    if self.has_recurrent_block:
+      return len(self._cached_tokens) if self._cached_tokens and len(self._cached_tokens) < len(tokens) \
+        and tokens[:len(self._cached_tokens)] == self._cached_tokens else 0
     prefix_len = sum(1 for _ in itertools.takewhile(lambda ab: ab[0] == ab[1], zip(tokens[:-1], self._cached_tokens)))
     return min(block._reusable_prefix_len(prefix_len, len(self._cached_tokens)) for block in self.blk)
 

tinygrad/mixin/elementwise.py

@@ -221,7 +221,7 @@ class ElementwiseMixin(CreationMixin):
     if dtypes.is_int(a.dtype) and dtypes.is_int(b.dtype): return a.alu(Ops.CMOD, b)
     return a - a.div(b, rounding_mode="trunc") * b
 
-  def div(self, x: Self | ConstType, reverse: bool = False, rounding_mode: Literal["trunc", "floor"] | None = None) -> Self:
+  def div(self, x: 'Self|ConstType|UOp', reverse: bool = False, rounding_mode: Literal["trunc", "floor"] | None = None) -> Self:
     """
     Divides `self` by `x`.
     Equivalent to `self / x`.

tinygrad/mixin/gradient.py

@@ -7,7 +7,11 @@ from tinygrad.dtype import sum_acc_dtype
 def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
   if op == Ops.ADD: return (ctx._broadcast_to(ret.src[0].shape),)
   if op == Ops.MAX: return (((mask:=ret.src[0].eq(ret).cast(ctx.dtype))/mask._rop(Ops.ADD, tuple(range(ret.arg[1])))) * ctx,)
-  if op == Ops.MUL: return (ctx * ret / ret.src[0],)
+  if op == Ops.MUL:
+    # d(prod x)/dx_j = prod_{i!=j} x_i: ret/x_j whenever x_j != 0 (any zero makes ret 0), else the product of the others
+    safe_x, axes = (is_zero:=(x:=ret.src[0]).eq(0)).where(1, x), tuple(range(ret.arg[1]))
+    zero_count = is_zero.cast(sum_acc_dtype(is_zero.dtype))._rop(Ops.ADD, axes)
+    return (ctx * is_zero.where(zero_count.eq(1).where(safe_x._rop(Ops.MUL, axes), 0), ret/safe_x),)
 
 def _compact_params(body:UOp, all_args:tuple[UOp, ...]) -> tuple[UOp, tuple[UOp, ...]]:
   """Remove unused PARAMs from body and return compacted (body, args)."""

tinygrad/mixin/movement.py

@@ -268,7 +268,8 @@ class MovementMixin:
     return self.shrink(tuple([None if ns is None else (0, ns) for ns in argfix(shape, *args)]))
 
   def pad_to(self, shape, *args) -> Self:
-    return self._mop(Ops.PAD, tuple((0, s if ns is None else ns) for s,ns in zip(self.shape, argfix(shape, *args), strict=True)))
+    ret = self._mop(Ops.PAD, tuple((0, s if ns is None else ns) for s,ns in zip(self.shape, argfix(shape, *args), strict=True)))
+    return self if ret.shape == self.shape else ret
 
   def view(self, shape, *args) -> Self:
     """`.view` is an alias for `.reshape`."""

tinygrad/mixin/op.py

@@ -514,7 +514,7 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
     output_dtype = self.dtype if dtypes.is_float(self.dtype) else dtypes.float32
     numerator = self.cast(sum_acc_dtype(self.dtype)).sum(axis=axis, keepdim=keepdim)
     denominator = prod([si for si, so in zip(self.shape, self.sum(axis=axis, keepdim=True).shape) if resolve(si != so)])
-    return numerator.div(denominator).cast(output_dtype)  # type: ignore[arg-type]
+    return numerator.div(denominator).cast(output_dtype)
 
   def var(self, axis:int|Sequence[int]|None=None, keepdim=False, correction=1) -> Self:
     """
@@ -538,12 +538,11 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
     print(t.var(axis=1).numpy())
     ```
     """
+    output_dtype = self.dtype if dtypes.is_float(self.dtype) else dtypes.float32
     squares = (self - self.mean(axis=axis, keepdim=True)).square()
     n = prod([si for si, so in zip(self.shape, squares.sum(axis=axis, keepdim=True).shape) if resolve(si != so)])
-    reduced = squares.sum(axis=axis, keepdim=keepdim)
-    denominator = reduced.const_like(n) - correction  # type: ignore[arg-type]
-    # TODO: remove relu?
-    return reduced.div(denominator.relu())
+    numerator = squares.cast(sum_acc_dtype(self.dtype)).sum(axis=axis, keepdim=keepdim)
+    return numerator.div(smax(n - correction, 0)).cast(output_dtype)
 
   def var_mean(self, axis:int|Sequence[int]|None=None, keepdim=False, correction=1) -> tuple[Self, Self]:
     """
@@ -1057,14 +1056,16 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
     assert not (align_corners and mode != "linear"), "align_corners option can only be set with the interpolating mode linear"
     x, expand = self, list(self.shape)
     for i in range(-1,-len(size)-1,-1):
-      scale = (int(self.shape[i]) - int(align_corners)) / (size[i] - int(align_corners))
-      arr, reshape = type(self).arange(size[i], dtype=dtypes.float32), [1] * self.ndim
+      in_sz, reshape = int(self.shape[i]), [1] * self.ndim
       reshape[i] = expand[i] = size[i]
       if mode == "linear":
-        index = (scale*arr if align_corners else (scale*(arr+0.5))-0.5).clip(0, self.shape[i]-1)
-        low, high, perc = [y.reshape(reshape).expand(expand) for y in (index.floor().int(), index.ceil().int(), index - index.floor())]
+        arr = type(self).arange(size[i])
+        num, den = (arr*(in_sz-1), size[i]-1) if align_corners else ((arr*2+1)*in_sz - size[i], size[i]*2)
+        num = num.clip(0, (in_sz-1)*den)
+        low, high, perc = [y.reshape(reshape).expand(expand) for y in (num//den, (num+den-1)//den, (num % den).cast(dtypes.float32)/den)]
         x = x.gather(i, low).lerp(x.gather(i, high), perc)
       else:
+        scale, arr = in_sz / size[i], type(self).arange(size[i], dtype=dtypes.float32)
         index = (scale*(arr+0.5) if mode=="nearest-exact" else scale*arr).cast(dtypes.int32).reshape(reshape).expand(expand)
         x = x.gather(i, index)
     return x.cast(self.dtype)

tinygrad/nn/state.py

@@ -1,4 +1,4 @@
-import json, math, pathlib, zipfile, pickle, tarfile, struct, functools, io, zlib
+import json, math, pathlib, struct, functools, io, zlib
 from collections import OrderedDict
 from typing import Any, Callable, BinaryIO, Iterable, cast
 from tinygrad.tensor import Tensor
@@ -219,6 +219,7 @@ def load_state_dict(model, state_dict:dict[str, Tensor], strict=True, verbose=Tr
 
 @accept_filename
 def zip_extract(t: Tensor) -> dict[str, Tensor]:
+  import zipfile
   files: dict[str, Tensor] = {}
   with zipfile.ZipFile(TensorIO(t), "r") as myzip:
     # sadly, the extra length needs to be read from the local header of each file.
@@ -249,6 +250,7 @@ def tar_extract(t: Tensor) -> dict[str, Tensor]:
   tensors = nn.state.tar_extract(Tensor(pathlib.Path("archive.tar")))
   ```
   """
+  import tarfile
   with tarfile.open(fileobj=TensorIO(t), mode="r") as tar:
     return {member.name:t[member.offset_data:member.offset_data+member.size] for member in tar if member.type == tarfile.REGTYPE}
 
@@ -303,6 +305,7 @@ def torch_load(t:Tensor) -> dict[str, Tensor]:
                "FloatTensor": None, "Parameter": Parameter}
   whitelist = {"torch", "collections", "numpy", "_codecs"}  # NOTE: this is not for security, only speed
   class Dummy: pass
+  import pickle, zipfile, tarfile
   class TorchPickle(pickle.Unpickler):
     def find_class(self, module, name):
       module_root = module.split(".")[0]

tinygrad/renderer/cstyle.py

@@ -264,6 +264,7 @@ class ClangRenderer(CStyleLanguage):
   nan = '__builtin_nanf("")'
 
   # language options
+  barrier = "__atomic_thread_fence(__ATOMIC_SEQ_CST);"
   buffer_suffix = " restrict"
   type_map = {dtypes.bool:"_Bool", dtypes.half:"__fp16"}
   code_for_op = {**({k:v for k,v in CStyleLanguage.code_for_op.items() if k not in [Ops.EXP2, Ops.SIN, Ops.LOG2, Ops.TRUNC, Ops.RECIPROCAL]}),

tinygrad/renderer/llvmir.py

@@ -142,7 +142,7 @@ base_rewrite = PatternMatcher([
   (UPat(Ops.IF, name="x"), lambda ctx,x: f"  br i1 {ctx[x.src[0]]}, label %ifbody_{ctx[x][1:]}, label %ifskip_{ctx[x][1:]}\nifbody_{ctx[x][1:]}:"),
   (UPat(Ops.ENDIF, name="x"), lambda ctx,x: f"  br label %ifskip_{ctx[x.src[0]][1:]}\nifskip_{ctx[x.src[0]][1:]}:"),
 
-  (UPat(Ops.BARRIER), lambda ctx: "")
+  (UPat(Ops.BARRIER), lambda ctx: "  fence seq_cst")
 ])
 
 class LLVMRenderer(Renderer):
@@ -238,8 +238,8 @@ class AMDLLVMRenderer(LLVMRenderer):
     (UPat(Ops.CAST, dtypes.fp8s, (UPat(dtype=dtypes.float),), name="x",), lambda ctx,x:
       f"  {ctx[x]} = call i8 @f32_to_fp8({ldt(x.src[0].dtype)}  {ctx[x.src[0]]}, i1 {'1' if x.dtype == dtypes.fp8e5m2 else '0'})"),
     (UPat(Ops.CAST, dtypes.float, (UPat.var("y", dtypes.fp8s),), name="x",), lambda ctx,x,y:
-      f"  {ctx[x.src[0]]}_i32 = zext i8 {ctx[x.src[0]]} to i32\n"
-      f"  {ctx[x]} = call float @llvm.amdgcn.cvt.f32.{'bf8' if y.dtype == dtypes.fp8e5m2 else 'fp8'}(i32 {ctx[x.src[0]]}_i32, i32 0)"),
+      f"  {ctx[x]}_i32 = zext i8 {ctx[x.src[0]]} to i32\n"
+      f"  {ctx[x]} = call float @llvm.amdgcn.cvt.f32.{'bf8' if y.dtype == dtypes.fp8e5m2 else 'fp8'}(i32 {ctx[x]}_i32, i32 0)"),
   ]) + base_rewrite
   extra_matcher = LLVMRenderer.extra_matcher + create_non_native_float_pats(dtypes.fp8s) + PatternMatcher([
     # amd llvm intrinsics llvm.log2/llvm.exp2 don't support double
@@ -279,43 +279,9 @@ exit: %packed = phi i32 [%packed_bf8, %do_bf8], [%packed_fp8, %do_fp8]\n  %trunc
       (UPat(Ops.WMMA, name="wmma"), lambda ctx, wmma, rdna4=AMDLLVMRenderer.is_rdna4(target.arch), cdna=self.is_cdna:
         render_wmma_amd(ctx, wmma, cdna, rdna4))
     ])
-    if self.is_cdna:
-      self.extra_matcher += PatternMatcher([
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
-          if x.arg[0][2] == 128 and x.src[0].dtype.itemsize <= 8 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-          if x.max_numel() == 4 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 4 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64), x.src[2]))
-          if x.max_numel() == 4 and x.src[0].dtype in dtypes.fp8_ocp and x.src[0].max_numel() == 8 else None),
-      ])
-    if target.arch in {"gfx1100", "gfx1151"}:
-      self.extra_matcher += PatternMatcher([
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.int32), lambda x: x.replace(
-          src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
-          if x.src[0].dtype == dtypes.int8 and x.src[0].max_numel() == 16 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.half), lambda x: UOp(Ops.STACK, src=tuple(x.replace(
-          src=(x.src[0], x.src[1], UOp(Ops.STACK, src=tuple(x.src[2].index(UOp.const(j//2, dtypes.int16))
-            if j%2 == 0 else UOp.const(0.0, x.src[2].dtype)
-            for j in range(x.max_numel()*2)))),
-          arg=(*x.arg[:4], None)).index(UOp.const(i*2, dtypes.int16))
-          for i in range(x.max_numel()))) if x.max_numel() == 8 else None),
-        (UPat(Ops.WMMA, name="x"), lambda x: x.replace(
-          src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-          if x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 16 else None),
-      ])
-    if target.arch in {"gfx1200", "gfx1201"}:
-      self.extra_matcher += PatternMatcher([
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.bfloat16), lambda x: x.replace(
-          dtype=dtypes.uint16,
-          src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2].bitcast(dtypes.uint16)))
-            .bitcast(dtypes.bfloat16) if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None),
-        (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-          lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-          if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None)
-      ])
+    if self.is_cdna: self.extra_matcher += tc.pm_validate_wmma_cdna
+    if target.arch in {"gfx1100", "gfx1151"}: self.extra_matcher += tc.pm_validate_wmma_rdna3
+    if target.arch in {"gfx1200", "gfx1201"}: self.extra_matcher += tc.pm_validate_wmma_rdna4
 
   def supported_dtypes(self): return {d for d in super().supported_dtypes()
                                       if (d not in dtypes.fp8_ocp or self.target.arch == "gfx950") and d not in dtypes.fp8_fnuz}

tinygrad/renderer/nir.py

@@ -137,7 +137,8 @@ class NIRRenderer(Renderer):
     (UPat(Ops.CAST, (dtypes.uchar, dtypes.ushort), src=(UPat.var("x", dtypes.floats),), name="c"), lambda x,c: x.cast(dtypes.int32).cast(c.dtype)),
     # load/store use pointer arithmetic, and the cast does nothing. NOTE: this doesn't apply to image indexing cause it's 1-D
     (UPat((Ops.INDEX, Ops.SHRINK), src=(UPat.var("buf"), UPat.var("off")), allow_any_len=True, name="x"), lambda x,buf,off: x.replace(
-      src=(buf,off.cast(dtypes.long))+x.src[2:]) if buf.addrspace != AddrSpace.REG and not is_image_shape(buf._shape) else None),
+      src=(buf,UOp.const(off.val, dtypes.long) if off.op is Ops.CONST else off.cast(dtypes.long))+x.src[2:])
+      if buf.addrspace != AddrSpace.REG and not is_image_shape(buf._shape) else None),
     # images need index to be int for nir (coordinates only: the INDEX keeps its access dtype)
     (UPat.var("buf").index(UPat.var("idx_y"), UPat.var("idx_x"), name="x"),
      lambda x,buf,idx_y,idx_x: x.replace(src=(buf, idx_y.cast(dtypes.int), idx_x.cast(dtypes.int)))),

tinygrad/runtime/autogen/__init__.py

@@ -1,4 +1,4 @@
-import glob, importlib, os, pathlib, shutil, subprocess, tarfile, tempfile
+import glob, importlib, os, pathlib, subprocess
 from tinygrad.helpers import fetch, flatten, system, getenv
 
 root = (here:=pathlib.Path(__file__).parent).parents[2]
@@ -31,6 +31,7 @@ def load(name, files, **kwargs):
   if not (f:=(root/(path:=kwargs.pop("path", __name__)).replace('.','/')/f"{name}.py")).exists() or getenv('REGEN'):
     files, kwargs['args'] = files() if callable(files) else files, args() if callable(args:=kwargs.get('args', [])) else args
     if (srcs:=kwargs.pop('srcs', None)):
+      import tempfile, tarfile
       srcpath = (td:=tempfile.TemporaryDirectory(f"autogen-src-{name.replace('/','-')}")).name + "/"
       for src in (srcs if isinstance(srcs, list) else [srcs]):
         if 'tar' in src:
@@ -157,7 +158,7 @@ def __getattr__(nm):
           *[f"python3 src/compiler/nir/nir_{s}_h.py --outdir gen" for s in ["intrinsics", "intrinsics_indices"]]]), cwd=path, shell=True, check=True),
   srcs="https://gitlab.freedesktop.org/mesa/mesa/-/archive/mesa-25.2.7/mesa-25.2.7.tar.gz",
   dll=f"'tinymesa_cpu' if DEV.renderer == 'LVP' else 'tinymesa', {tinymesa_path}, emsg='pip install tinymesa==25.2.7.2'",
-  prolog=["from tinygrad.helpers import DEV", "import gzip, base64, platform, sysconfig, os"],
+  prolog=["from tinygrad.helpers import DEV", "import gzip, base64, sysconfig, os"],
   epilog=lambda path: [system(f"{root}/extra/mesa/lvp_nir_options.sh {path}")])
     case "libclang":
       return load("libclang",

tinygrad/runtime/autogen/mesa.py

@@ -5,7 +5,7 @@ from typing import Literal, TypeAlias
 from tinygrad.runtime.support.c import _IO, _IOW, _IOR, _IOWR
 from tinygrad.runtime.support import c
 from tinygrad.helpers import DEV
-import gzip, base64, platform, sysconfig, os
+import gzip, base64, sysconfig, os
 dll = c.DLL('mesa', 'tinymesa_cpu' if DEV.renderer == 'LVP' else 'tinymesa', os.path.join(sysconfig.get_paths()['platlib'], 'tinymesa'), emsg='pip install tinymesa==25.2.7.2')
 class struct_u_printf_info(c.Struct): pass
 u_printf_info: TypeAlias = struct_u_printf_info

tinygrad/runtime/graph/hcq.py

@@ -6,7 +6,6 @@ from tinygrad.device import Buffer, BufferSpec, Compiled, Device, MultiBuffer, P
 from tinygrad.dtype import dtypes
 from tinygrad.uop.ops import UOp, Ops, Variable
 from tinygrad.engine.jit import GraphRunner, MultiGraphRunner
-from tinygrad.runtime.ops_rdma import RDMACopyQueue
 
 class HCQGraph(MultiGraphRunner):
   def __init__(self, *args, **kwargs):
@@ -50,7 +49,7 @@ class HCQGraph(MultiGraphRunner):
 
     self.comp_queues: dict[HCQCompiled, HWQueue] = {dev: unwrap(dev.hw_compute_queue_t)() for dev in self.devices}
     self.copy_queues: dict[tuple[HCQCompiled, int], HWQueue] = {} # lazy allocation, keyed by (device, queue_idx)
-    self.rdma_queues: dict[tuple[HCQCompiled, HCQCompiled], RDMACopyQueue] = {} # lazy allocation, keyed by device pair
+    self.rdma_queues: dict[tuple[HCQCompiled, HCQCompiled], "RDMACopyQueue"] = {} # lazy allocation, keyed by device pair
     self.num_copy_queues: int = getenv("HCQ_NUM_SDMA", min(len(self.devices), 8) if ALL2ALL >= 1 else 1)
     self.num_rdma_ops: dict[tuple[HCQCompiled, HCQCompiled], int] = collections.defaultdict(int)
 
@@ -104,6 +103,7 @@ class HCQGraph(MultiGraphRunner):
       elif is_rdma:
         enqueue_queue = self.comp_queues[enqueue_dev]
         rdma_key = (cast(HCQCompiled, Device[bufs[0].device]).rdma_dev(), enqueue_dev.rdma_dev())
+        from tinygrad.runtime.ops_rdma import RDMACopyQueue
         self.rdma_queues.setdefault(rdma_key, RDMACopyQueue(enqueue_dev.rdma_dev()))
       else:
         assert (enqueue_dev.hw_copy_queue_t is not None), "device must implement a copy queue"

tinygrad/runtime/graph/metal.py

@@ -113,5 +113,6 @@ class MetalGraph(GraphRunner):
   @staticmethod
   def supports_uop(batch_devs, new_call:UOp) -> bool:
     # Metal ICB replay encodes offsets as uint32; reject if any Metal buffer offset exceeds 32-bit range.
-    if any(b.op is Ops.SLICE and b.src[1].val * b.src[0].dtype.itemsize > 0xFFFFFFFF for b in new_call.src[1:]): return False
+    for shrink in [s for src in new_call.src[1:] if (s:=src.src[0] if src.op is Ops.BITCAST else src).op is Ops.SHRINK]:
+      if shrink.src[1].val * shrink.src[0].dtype.itemsize > 0xFFFFFFFF: return False
     return GraphRunner.supports_uop(batch_devs, new_call)

tinygrad/runtime/ops_cpu.py

@@ -1,11 +1,11 @@
 from __future__ import annotations
-import platform, sys, os, ctypes, functools, mmap, threading, array, itertools
-from dataclasses import replace
-from typing import cast
-from tinygrad.helpers import to_mv, OSX, WIN, Context, mv_address, suppress_finalizing, unwrap, data64_le, partition
-from tinygrad.device import Buffer, BufferSpec, TinyELF
-from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, HCQArgsState, HCQSignal, HCQProgram, MMIOInterface
-from tinygrad.runtime.support.hcq import CLikeArgsState
+import platform, sys, os, ctypes, functools, mmap, threading, array, struct, time
+from dataclasses import dataclass, replace
+from typing import cast, Callable
+from tinygrad.helpers import to_mv, from_mv, OSX, WIN, Context, mv_address, suppress_finalizing, unwrap, data64_le, to_tuple
+from tinygrad.device import Buffer, BufferSpec, TinyELF, Program, Device
+from tinygrad.runtime.support.hcq import HCQBuffer, MMIOInterface
+from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, make_cmdbuf, make_signal
 from tinygrad.renderer.cstyle import ClangRenderer
 from tinygrad.renderer.llvmir import CPULLVMRenderer
 from tinygrad.renderer.nir import LVPRenderer
@@ -13,11 +13,15 @@ from tinygrad.renderer.isa.x86 import X86Renderer
 from tinygrad.runtime.support.elf import jit_loader
 from tinygrad.runtime.autogen import libc
 from tinygrad.codegen import do_to_program
+from tinygrad.engine.realize import pm_flatten_linear, get_call_arg_uops, get_runtime
 from tinygrad import UOp, dtypes
 from tinygrad.dtype import AddrSpace
-from tinygrad.uop.ops import sint, KernelInfo, Ops, UPat, PatternMatcher, graph_rewrite
+from tinygrad.uop.ops import KernelInfo, Ops, UPat, PatternMatcher, graph_rewrite
 
-MAX_ARGS, CMD_SIZE, RING_SLOTS = 63, 64, (16 << 10)
+MAX_ARGS, CMD_SIZE, RING_SLOTS, FUNCS = 63, 64, (16 << 10), (() if WIN else ('clock_gettime', 'sem_wait', 'sem_post'))
+
+# *****************
+# 1. workers
 
 def signal_prog():
   val = UOp.param(1, dtypes.int, (), vmin_vmax=(0, dtypes.int.max), name="value", addrspace=AddrSpace.ALU)
@@ -35,79 +39,86 @@ def timestamp_prog():
     val = ts.after(call)[0].load() * 1_000_000_000 + ts.after(call)[1].load()
   return UOp.param(0, dtypes.uint64, (1,))[0].store(val)
 
-def quit_prog():
-  fn = UOp.param(0, dtypes.uint64, (1 if WIN else 3,))
-  if WIN: return fn[0].load().call(UOp.const(0, dtypes.uint64), ret_dtype=dtypes.void) # ExitThread(0)
-  sem = UOp.param(1, dtypes.uint64, (1,))
-
-  close = fn[2].load().call(sem[0], ret_dtype=dtypes.void) # sem_close(sem)
-  return fn.after(close)[0].load().call(UOp.const(0, dtypes.uint64), ret_dtype=dtypes.void) # pthread_exit(0)
-
 def worker_prog():
   ring = UOp.param(0, dtypes.uint64, (RING_SLOTS * CMD_SIZE,), volatile=True)
-  wait, sem = UOp.param(1, dtypes.uint64, (1,), volatile=True), UOp.param(2, dtypes.uint64, (1,))
-  cur = UOp.range(2**64-1, 0, dtype=dtypes.uint64)
+  wait, done = UOp.param(1, dtypes.uint64, (1,), volatile=True), UOp.param(2, dtypes.uint64, (1,), volatile=True)
+  sem, cur = UOp.param(3, dtypes.uint64, (1,)), UOp.range(2**64-1, 0, dtype=dtypes.uint64) # sem is unused on windows, it has to come last
 
   # spin on windows, sem_wait to sleep on posix
   if WIN: ready = (v:=wait.after(lw:=UOp.loop(1), cur)[0].load()).end(lw, v <= cur)
-  else: ready = wait.after(cur)[0].load().call(sem.after(cur)[0], ret_dtype=dtypes.void)
+  else: ready = (rv:=wait.after(lw:=UOp.loop(1), cur)[0].load().call(sem.after(cur)[0], ret_dtype=dtypes.int)).end(lw, rv != 0)
 
   entry = [ring.after(ready).index((cur % RING_SLOTS) * CMD_SIZE + i).load() for i in range(CMD_SIZE)]
-  return entry[0].call(*entry[1:], ret_dtype=dtypes.void).end(cur)
+  return done.after(entry[0].call(*entry[1:], ret_dtype=dtypes.void)).index(0).store(cur + 1).end(cur)
 
-def host_wait(ctx, dst:UOp, val:UOp) -> UOp:
-  return (cur:=dst.after(loop:=UOp.loop(next(ctx))).index(UOp.const(0, dtypes.int)).load()).end(loop, cur < val)
+@dataclass
+class CPUWorker: ring:Buffer; put:Buffer; sem:Buffer; sys:Buffer; done:Buffer; thread:threading.Thread  # noqa: E702
 
-pm_host_opsel = PatternMatcher([(UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))), host_wait)])
+# *****************
+# 2. queue encoders
 
-def encode_host_queue(q:UOp) -> UOp:
-  # TODO: subset of hcq2 for now
-  spins, (store,) = partition(graph_rewrite(q, pm_host_opsel, ctx=itertools.count(), walk=True, name="host opsel").src, lambda u: u.op is Ops.END)
-  assert store.op is Ops.INS and store.arg == "store", f"host queue cannot encode {store.op} {store.arg}"
-  return store.src[0].after(*spins).index(UOp.const(0, dtypes.int)).store(store.src[1])
+def cpu_cmd(devs:tuple[str, ...], prog, *args:UOp) -> UOp:
+  progs = [get_runtime(d, prog) if isinstance(prog, UOp) else cast(CPUDevice, Device[d]).prgs[prog] for d in devs]
+  addrs = tuple(UOp.const(p.addr, dtypes.uint64) for p in progs)
+  words = ((addrs[0] if len(addrs) == 1 else UOp(Ops.STACK, dtypes.uint64, addrs)),) + args
+  return UOp(Ops.INS, dtypes.void, words + (UOp.const(0, dtypes.uint64),) * (CMD_SIZE - len(words)), arg="cmd")
 
-class CPUComputeQueue(HWQueue):
-  def __init__(self, dev):
-    super().__init__()
-    self.dev = dev
-  def _cmd(self, prog, args=(), vals=()): return self.exec(prg:=self.dev.prgs[prog], prg.fill_kernargs(args, vals), None, None)
-  def memory_barrier(self): return self
-  def exec(self, prg:CPUProgram, args_state:HCQArgsState, global_size, local_size):
-    if (lvp:=isinstance(args_state, LVPArgsState)): self.bind_args_state(args_state)
-    args:list[sint|None] = [args_state.buf.va_addr] if lvp else [*[x.va_addr for x in args_state.bufs], *args_state.vals]
-    assert len(args) <= MAX_ARGS, f"CPU programs support at most {MAX_ARGS} arguments, got {len(args)}"
-    for tid in range(1 if lvp else (global_size or (1,))[0]):
-      if not lvp and 'core_id' in prg.runtimevars: args[prg.runtimevars['core_id']] = tid
-      self.q(prg, *[unwrap(x) for x in args], *([0] * (MAX_ARGS - len(args))))
-    return self
-  def wait(self, signal, value=0): return self._cmd(wait_prog, (signal.base_buf,), (value,))
-  def timestamp(self, signal): return self._cmd(timestamp_prog, (signal.base_buf.offset(8, 8), self.dev.func_table._buf.offset(0, 8)))
-  def signal(self, signal, value:sint=0): return self._cmd(signal_prog, (signal.base_buf,), (value,))
-  def _submit(self, dev):
-    dev.ensure_worker()
-    ring_view = dev.ring.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
-    for off in range(0, len(self._q), CMD_SIZE):
-      entry = [self._q[off].addr, *self._q[off+1:off+CMD_SIZE]]
-      ring_view[(base:=(dev.ring_pos % RING_SLOTS) * CMD_SIZE):base+CMD_SIZE] = array.array('Q', (int(x) & ((1<<64)-1) for x in entry))
-      dev.ring_pos += 1
-      if WIN: dev.sys.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = dev.ring_pos
-      else: assert libc.sem_post(dev.sem) == 0
+def cpu_exec(ctx:tuple[str, ...], call:UOp, prg:UOp) -> UOp:
+  args = [get_call_arg_uops(call)[i].getaddr(ctx) for i in prg.arg.globals] + [v.cast(dtypes.uint64) for v in prg.arg.vars]
+  if (core:=prg.arg.runtimevars.get('core_id')) is None: return cpu_cmd(ctx, prg, *args)
 
-class LVPArgsState(CLikeArgsState):
-  def __init__(self, buf, prg, bufs, vals=()): super().__init__(buf, prg, bufs, vals, [*data64_le(buf.va_addr + 12), (len(bufs) + len(vals)) * 2])
+  la = [cpu_cmd(ctx,prg,*args[:(cid:=(len(prg.arg.globals)+core))],UOp.const(t, dtypes.uint64),*args[cid+1:]) for t in range(prg.arg.global_size[0])]
+  return UOp(Ops.LINEAR, dtypes.void, tuple(la))
+
+pm_cpu_opsel = PatternMatcher([
+  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="prg"),), name="call", allow_any_len=True), cpu_exec),
+
+  (UPat(Ops.INS, arg="barrier"), lambda: UOp(Ops.NOOP, dtypes.void, ())),
+  (UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val"))),
+   lambda ctx, dst, val: cpu_cmd(ctx, wait_prog, dst.getaddr(ctx), val.cast(dtypes.uint64))),
+  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))),
+   lambda ctx, dst, val: cpu_cmd(ctx, signal_prog, dst.getaddr(ctx), val.cast(dtypes.uint64))),
+  (UPat(Ops.INS, arg="timestamp", src=(UPat(name="dst"),)),
+   lambda ctx, dst: cpu_cmd(ctx, timestamp_prog, dst.getaddr(ctx), *(() if WIN else (make_signal(ctx, tag="func:clock_gettime").getaddr(ctx),)))),
+])
+
+def encode_queue(q:UOp) -> UOp:
+  devs, queue = to_tuple(q.arg[0]), q.arg[1]
+  lin = graph_rewrite(q, pm_cpu_opsel+pm_flatten_linear, ctx=devs, walk=True, name=f"{queue} opsel")
+
+  cnt = sum(len(ins.src) for ins in lin.src) // CMD_SIZE
+  assert cnt < RING_SLOTS, f"submit of {cnt} entries doesn't fit the ring"
+  cmdbuf = make_cmdbuf(lin, devs, buf=UOp.placeholder((cnt*CMD_SIZE,), dtypes.uint64, next(UOp.unique_num), device=devs).rtag("cmdbuf"))
+  ring = UOp.placeholder((ring_words:=RING_SLOTS*CMD_SIZE,), dtypes.uint64, 0, device=devs, volatile=True).rtag(f"{queue}_ring")
+  put, done, sem, sysbuf = (make_signal(devs, tag=f"{queue}_{name}") for name in ("put", "done", "sem", "sys"))
+
+  # submits are serialized on the submitter, so they can bump put without atomics
+  ran = done.after(l:=UOp.loop(next(UOp.unique_num))).index(0).load()
+  room = ran.end(l, put.index(0).load() - ran > RING_SLOTS - cnt) # wait until cnt entries fit in the ring
+  base = ((put.after(room).index(0).load() % RING_SLOTS) * CMD_SIZE).cast(dtypes.int)
+  e = UOp.range(cnt, next(UOp.unique_num), dtype=dtypes.int, src=(cmdbuf, ring))
+  copy = UOp.group(*[ring.index((base + e*CMD_SIZE + w) % ring_words).store(cmdbuf.index(e*CMD_SIZE + w).load()) for w in range(CMD_SIZE)])
+
+  # wake the worker after each entry, keeping the post with the stores stops it from hoisting out of the loop
+  wake = copy.end(e) if WIN else make_signal(devs, tag="func:sem_post").after(copy).index(0).load().call(sem.index(0), ret_dtype=dtypes.void).end(e)
+  bumped = put.after(wake).index(0).store(put.index(0).load() + cnt)
+  return sysbuf.after(bumped).index(0).store(put.index(0).load() + cnt) if WIN else bumped
+
+# *****************
 
 # NOTE: MAP_JIT is added to mmap module in python 3.13
 MAP_JIT = 0x0800
 
-class CPUProgram(HCQProgram['CPUDevice']):
+class CPUProgram(Program['CPUDevice']):
   rt_lib = None
   try: rt_lib = ctypes.CDLL(ctypes.util.find_library('System' if OSX else 'kernel32') if OSX or WIN else 'libgcc_s.so.1')
   except OSError: pass
 
   def __init__(self, dev:CPUDevice, obj:TinyELF):
-    self.signature, self.runtimevars = obj.signature, {name:slot for name,slot,*_ in obj.signature if name == 'core_id'}
+    self.dev, self.name, self.signature = dev, obj.name, obj.signature
+    self.runtimevars = {name:slot for name,slot,*_ in obj.signature if name == 'core_id'}
+    self.lvp = obj.target.renderer == "LVP"
 
-    LVP = obj.target.renderer == "LVP"
     if sys.platform == "win32": # mypy doesn't understand when WIN is used here
       PAGE_EXECUTE_READWRITE, MEM_COMMIT, MEM_RESERVE = 0x40, 0x1000, 0x2000
       ctypes.windll.kernel32.VirtualAlloc.restype = ctypes.c_void_p
@@ -117,7 +128,7 @@ class CPUProgram(HCQProgram['CPUDevice']):
       ctypes.windll.kernel32.GetCurrentProcess.restype = ctypes.c_void_p
       proc = ctypes.windll.kernel32.GetCurrentProcess()
       ctypes.windll.kernel32.FlushInstructionCache(ctypes.c_void_p(proc), ctypes.c_void_p(self.addr), ctypes.c_size_t(len(obj.lib)))
-      self.fxn = ctypes.CFUNCTYPE(None)(self.addr)
+      self.fxn = ctypes.CFUNCTYPE(None, ctypes.c_void_p)(self.addr) if self.lvp else ctypes.CFUNCTYPE(None)(self.addr)
     else:
       # On apple silicon with SPRR enabled (it always is in macos) RWX pages are unrepresentable: https://blog.svenpeter.dev/posts/m1_sprr_gxf/
       # MAP_JIT allows us to easily flip pages from RW- to R-X and vice versa. It is a noop on intel cpus. (man pthread_jit_write_protect_np)
@@ -125,7 +136,7 @@ class CPUProgram(HCQProgram['CPUDevice']):
       self.addr = mv_address(self.mem)
 
       if OSX: unwrap(CPUProgram.rt_lib).pthread_jit_write_protect_np(False)
-      lib = jit_loader(obj.lib, base=ctypes.addressof(ctypes.c_void_p.from_buffer(self.mem)), link_libs=['m']) if LVP else obj.lib
+      lib = jit_loader(obj.lib, base=ctypes.addressof(ctypes.c_void_p.from_buffer(self.mem)), link_libs=['m']) if self.lvp else obj.lib
       self.mem.write(lib)
       if OSX: unwrap(CPUProgram.rt_lib).pthread_jit_write_protect_np(True)
 
@@ -138,15 +149,30 @@ class CPUProgram(HCQProgram['CPUDevice']):
         # msync should be a universal POSIX way to do this
         libc.msync(ctypes.c_void_p(self.addr), len(lib), libc.MS_SYNC | libc.MS_INVALIDATE)
 
-      self.fxn = ctypes.CFUNCTYPE(None)(self.addr)
+      self.fxn = ctypes.CFUNCTYPE(None, ctypes.c_void_p)(self.addr) if self.lvp else ctypes.CFUNCTYPE(None)(self.addr)
 
-    super().__init__(LVPArgsState if LVP else HCQArgsState, dev, obj, kernargs_alloc_size=12+256 if LVP else 0)
+  def __call__(self, *bufs:HCQBuffer, global_size:tuple[int,int,int]=(1,1,1), local_size:tuple[int,int,int]=(1,1,1),
+               vals:tuple[int|None, ...]=(), wait:bool=False, timeout:int|None=None) -> float|None:
+    st = time.perf_counter()
+    if self.lvp:
+      lvp_args = bytearray(12 + (len(bufs) + len(vals)) * 8)
+      addr = mv_address(lvp_args)
+      struct.pack_into(f'<3I{len(bufs)}Q', lvp_args, 0, *data64_le(addr+12), (len(bufs)+len(vals))*2, *[b.va_addr for b in bufs])
+      for v,(off,dt) in zip(vals, TinyELF.iter_sig(self.signature[-len(vals):], len(bufs)*8)): struct.pack_into(f'<{dt.fmt}', lvp_args, 12+off, v)
+      self.fxn(addr)
+    else:
+      args = [*[cast(int, b.va_addr) for b in bufs], *cast(tuple[int, ...], vals)]
+      assert len(args) <= MAX_ARGS, f"CPU programs support at most {MAX_ARGS} arguments, got {len(args)}"
+      for tid in range(global_size[0]):
+        if 'core_id' in self.runtimevars: args[self.runtimevars['core_id']] = tid
+        self.fxn(*[ctypes.c_uint64(x) for x in args])
+    return time.perf_counter() - st if wait else None
 
   @suppress_finalizing
   def __del__(self):
     if sys.platform == 'win32': ctypes.windll.kernel32.VirtualFree(ctypes.c_void_p(self.addr), ctypes.c_size_t(0), 0x8000) #0x8000 - MEM_RELEASE
 
-class CPUAllocator(HCQAllocator):
+class CPUAllocator(HCQAllocator['CPUDevice']):
   def __init__(self, dev:CPUDevice): super().__init__(dev, supports_copy_from_disk=False, supports_transfer=False)
   def _alloc(self, size:int, options:BufferSpec) -> HCQBuffer:
     if options.external_ptr is not None: addr, buf = options.external_ptr, None
@@ -154,68 +180,55 @@ class CPUAllocator(HCQAllocator):
     else: addr = mv_address(buf:=mmap.mmap(-1, size, mmap.MAP_ANON | mmap.MAP_SHARED, mmap.PROT_READ | mmap.PROT_WRITE))
     return HCQBuffer(va:=addr, sz:=size, meta=buf, view=MMIOInterface(va, sz, fmt='B'), owner=self.dev)
   def _as_buffer(self, src) -> memoryview: return to_mv(src.va_addr, src.size)
+  def _copyin(self, dest:HCQBuffer, src:memoryview):
+    self.dev.synchronize()
+    ctypes.memmove(int(dest.va_addr), from_mv(src), len(src))
+  def _copyout(self, dest:memoryview, src:HCQBuffer):
+    self.dev.synchronize()
+    ctypes.memmove(from_mv(dest), int(src.va_addr), len(dest))
   def _do_map(self, buf:HCQBuffer):
     if buf.view is None or not isinstance(buf.view, MMIOInterface): raise RuntimeError("Cannot map buffer without view to cpu")
     return HCQBuffer(buf.view.addr, buf.size, view=buf.view, owner=buf.owner)
   def _unmap(self, mb): pass  # CPU _do_map returns a view wrapper, nothing to release
 
-class CPUDevice(HCQCompiled):
-  pm_lower = PatternMatcher([
-    (UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),)), encode_host_queue)])
-
-  pm_bufferize = PatternMatcher([
-    (UPat(Ops.PARAM, tag="sentinel_signal"), lambda ctx: ctx[0].signal("sentinel", (1 << 64) - 1)),
-    (UPat(Ops.PARAM, tag="timeline_signal"), lambda ctx: ctx[0].signal("timeline")),
-    (UPat(Ops.PARAM, tag="timeline_value"), lambda ctx: ctx[0].signal("value", 1)),
-    (UPat(Ops.PARAM, tag="signal", name="b"), lambda ctx, b: ctx[0].signal(b.arg.slot)),
-  ])
-
-  @functools.cache
-  def signal(self, name:str|int, init_value:int=0) -> Buffer:
-    (buf:=Buffer(self.device, 1, dtypes.uint64, preallocate=True)).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = init_value
-    return buf
+class CPUDevice(HCQ2Compiled):
+  wait_timeout_ms, has_copy_queue = 30000, False
+  pm_lower = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),)), encode_queue)])
 
   def __init__(self, device:str=""):
-    super().__init__(device, CPUAllocator(self), [ClangRenderer, CPULLVMRenderer, LVPRenderer, X86Renderer], CPUProgram, HCQSignal,
-      functools.partial(CPUComputeQueue, self), arch={'amd64':'x86_64', 'aarch64':'arm64'}.get(m:=platform.machine().lower(), m)+",native")
+    super().__init__(device, CPUAllocator(self), [ClangRenderer, CPULLVMRenderer, LVPRenderer, X86Renderer], CPUProgram,
+      arch={'amd64':'x86_64', 'aarch64':'arm64'}.get(m:=platform.machine().lower(), m)+",native")
 
-    self.ring_pos = 0
+    self.pm_bufferize = PatternMatcher(
+      [(UPat(Ops.PARAM, tag=f"COMPUTE:0_{n}"), lambda ctx, n=n: getattr(ctx[0].worker, n)) for n in ("ring", "put", "sem", "sys", "done")] +
+      [(UPat(Ops.PARAM, tag=f"func:{f}"), lambda ctx, f=f: ctx[0].func_ptr(f)) for f in FUNCS]) + self.pm_bufferize
 
-    # posix uses sem to put cpus into sleep
-    self.sem_addr = 0
-    if not WIN:
-      self.sem = libc.sem_open(sem_name:=f"/tinygrad-{os.getpid()}-{id(self):x}".encode(), os.O_CREAT|os.O_EXCL, 0o600, 0) # type: ignore[call-arg]
-      self.sem_addr = unwrap(ctypes.cast(self.sem, ctypes.c_void_p).value)
-      if self.sem_addr == ctypes.c_void_p(-1).value or libc.sem_unlink(sem_name): raise OSError(ctypes.get_errno(), "semaphore")
-
-    # TODO: move to hcq2
     with Context(EMULATED_DTYPES="", TRACK_MATCH_STATS=0):
-      prgs = {f: f().sink(arg=KernelInfo(f.__name__), tag=1) for f in (signal_prog, wait_prog, timestamp_prog, quit_prog, worker_prog)}
-      self.prgs = {f: self.runtime(do_to_program(v, ClangRenderer(replace(self.renderer.target, renderer="CLANG"))).to_elf()) for f,v in prgs.items()}
+      clang = ClangRenderer(replace(self.renderer.target, renderer="CLANG"))
+      self.prgs:dict[Callable, CPUProgram] = {f: CPUProgram(self, do_to_program(f().sink(arg=KernelInfo(f.__name__), tag=1), clang).to_elf())
+                                              for f in (signal_prog, wait_prog, timestamp_prog, worker_prog)}
 
-  @functools.cached_property
-  def ring(self) -> Buffer: return Buffer(self.device, RING_SLOTS * CMD_SIZE, dtypes.uint64, preallocate=True)
-  @functools.cached_property
-  def sys(self) -> Buffer: return Buffer(self.device, 1, dtypes.uint64, preallocate=True)
-  @functools.cached_property
-  def sem_buf(self) -> Buffer: return Buffer(self.device, 1, dtypes.uint8, options=BufferSpec(external_ptr=self.sem_addr), preallocate=True)
+  def func_ptr(self, name:str) -> Buffer: return self.func_table.view(1, dtypes.uint64, FUNCS.index(name)*8).ensure_allocated()
 
-  # TODO: move to hcq2 infra
   @functools.cached_property
   def func_table(self) -> Buffer:
-    fns = ([0, ctypes.windll.kernel32.ExitThread, 0, 0] if WIN else  # type: ignore[attr-defined]
-           [libc.dll.clock_gettime, libc.dll.pthread_exit, libc.dll.sem_wait, libc.dll.sem_close])
-    addrs = array.array('Q', [unwrap(ctypes.cast(f, ctypes.c_void_p).value) if f else 0 for f in fns])
-    (ft:=Buffer(self.device, len(fns), dtypes.uint64, preallocate=True)).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[:] = addrs
+    lib = ctypes.windll.kernel32 if sys.platform == "win32" else libc.dll # type: ignore[attr-defined]
+    (ft:=Buffer(self.device, len(FUNCS), dtypes.uint64, preallocate=True))._buf.cpu_view().view(fmt='Q')[:] = \
+      array.array('Q', [unwrap(ctypes.cast(getattr(lib, f), ctypes.c_void_p).value) for f in FUNCS])
     return ft
 
-  @functools.cache
-  def ensure_worker(self):
-    threading.Thread(target=cast(CPUProgram, self.prgs[worker_prog]).fxn, daemon=True, args=[ctypes.c_uint64(x) for x in
-      [self.ring._buf.va_addr, self.sys._buf.va_addr if WIN else self.func_table._buf.va_addr+16, self.sem_addr]]).start()
+  @functools.cached_property
+  def worker(self) -> CPUWorker:
+    ring, put, sysbuf, done = (Buffer(self.device, sz, dtypes.uint64, preallocate=True) for sz in (RING_SLOTS*CMD_SIZE, 1, 1, 1))
+    addr, hsem = 0, None
 
-  def finalize(self):
-    if self.ring_pos == 0: return # the worker starts with the first submit
-    ft = self.func_table._buf
-    CPUComputeQueue(self)._cmd(quit_prog, (ft.offset(8, 8),) if WIN else (ft.offset(8, 24), self.sem_buf._buf)).submit(self)
-    self.ring_pos = 0
+    # sem are posix-only
+    if not WIN:
+      hsem = libc.sem_open(nm:=f"/tinygrad-{os.getpid()}-{id(ring):x}".encode(), os.O_CREAT|os.O_EXCL, 0o600, 0) # type: ignore[call-arg]
+      if (addr:=unwrap(ctypes.cast(hsem, ctypes.c_void_p).value)) == ctypes.c_void_p(-1).value or libc.sem_unlink(nm):
+        raise OSError(ctypes.get_errno(), "semaphore")
+    sem = Buffer(self.device, 1, dtypes.uint64, options=BufferSpec(external_ptr=addr), preallocate=True)
+
+    worker_args = [ring._buf.va_addr, sysbuf._buf.va_addr if WIN else self.func_ptr('sem_wait')._buf.va_addr, done._buf.va_addr, addr]
+    (worker:=threading.Thread(target=self.prgs[worker_prog].fxn, daemon=True, args=[ctypes.c_uint64(x) for x in worker_args])).start()
+    return CPUWorker(ring, put, sem, sysbuf, done, worker)

tinygrad/runtime/ops_python.py

@@ -3,10 +3,10 @@
 # works to test the tensor cores, and all the uops in general
 # this is the (living) definition of uops
 from typing import Any, TYPE_CHECKING
-import pickle, base64, itertools, time, sys, functools
+import pickle, base64, itertools, time, sys, functools, ctypes
 from dataclasses import replace
 from tinygrad.dtype import bitcast, DType, dtypes, AddrSpace, truncate, storage_fmt_for_dtype, to_storage_scalar, from_storage_scalar
-from tinygrad.helpers import all_same, getenv, flatten, Target, IMAGE, is_image_shape, cpu_profile
+from tinygrad.helpers import all_same, getenv, flatten, Target, IMAGE, is_image_shape, cpu_profile, mv_address
 from tinygrad.device import Buffer, Compiled, Compiler, Allocator, Program, TinyELF
 from tinygrad.codegen.opt import tc
 from tinygrad.uop.ops import exec_alu, python_alu, Ops, UOp, GroupOp
@@ -134,6 +134,13 @@ class PythonProgram(Program['PythonDevice']):
                                for k in range(len(src_values))], j, u.dtype) for j in range(load_sz)]
           else:
             values[u] = load(src_values, 0, u.dtype)
+        elif u.op is Ops.CALL:
+          assert u.dtype is dtypes.void
+          cfunc = ctypes.CFUNCTYPE(None, *[ctypes.c_uint64] * (len(src_values)-1))
+          values[u] = []
+          for args,gate in zip(zip(*src_values), exec_masks[-1]):
+            call_args = [(mv_address(x[0]) + x[1]*dt.itemsize) if isinstance(x, tuple) else x for x,dt in zip(args, src_dtypes)]
+            values[u].append(cfunc(call_args[0])(*call_args[1:]) if gate else None)
         elif u.op is Ops.WMMA:
           first_src_dtype = u.src[0].dtype
           assert isinstance(first_src_dtype, DType) # mypy

tinygrad/runtime/support/compiler_amd.py

@@ -89,9 +89,11 @@ class HIPCompiler(Compiler):
 
 class HIPCCCompiler(Compiler):
   def __init__(self, arch:str, extra_options:list[str]=[]):
-    self.arch, self.extra_options = arch, extra_options
-    super().__init__(f"compile_hipcc_{self.arch}_{hashlib.sha256(' '.join(extra_options).encode()).hexdigest()[:8]}")
+    self.arch, self.extra_options, self.no_hipcc = arch, extra_options, getenv("NO_HIPCC")
+    super().__init__(f"compile_hipcc_{self.arch}_{hashlib.sha256(' '.join(extra_options).encode()).hexdigest()[:8]}"+
+                     ("_nohipcc" if self.no_hipcc else ""))
   def compile(self, src:str) -> bytes:
+    if self.no_hipcc: return b""
     with tempfile.NamedTemporaryFile(suffix=".cpp") as srcf, tempfile.NamedTemporaryFile(suffix=".bc") as bcf:
       with tempfile.NamedTemporaryFile(suffix=".hsaco") as libf:
         srcf.write(src.encode())

tinygrad/runtime/support/compiler_qcom.py

@@ -1,6 +1,6 @@
-import ctypes, struct
+import ctypes, struct, platform, pathlib, shutil, subprocess, sys, tarfile, tempfile
 from tinygrad.device import Compiler
-from tinygrad.helpers import DEBUG, system
+from tinygrad.helpers import DEBUG, system, fetch, unwrap
 from tinygrad.runtime.support.compiler_mesa import disas_adreno
 # see https://github.com/sirhcm/tinydreno
 from tinygrad.runtime.autogen import llvm_qcom
@@ -10,10 +10,17 @@ def _read_lib(lib, off) -> int: return struct.unpack("I", lib[off:off+4])[0]
 class QCOMCompiler(Compiler):
   def __init__(self, arch:str):
     assert arch.split(',')[0] == "a630", "only a630 supported"
-    self.arch, self.chip_id, self.llvm_inst = arch, 0x6030001, llvm_qcom.cl_compiler_create_llvm_instance()
+    if platform.machine() == "aarch64": self.arch, self.chip_id, self.llvm_inst = arch, 0x6030001, llvm_qcom.cl_compiler_create_llvm_instance()
+    else:
+      self.arch, self.chip_id, self.fs = arch, 0x6030001, tempfile.TemporaryDirectory()
+      with tarfile.open(fetch('https://git.tinygrad.win/sirhcm/images/releases/download/v2/qcomcl.tar.gz')) as t: t.extractall(fs:=self.fs.name)
+      if (qemu:=shutil.which("qemu-aarch64-static")): argv = f"{qemu} -cpu max,pauth=off -L {fs} {fs}/usr/bin/python3 {__file__} {arch}"
+      else: argv = (f"docker run --rm -i --platform linux/aarch64 -v {fs}/usr:/usr -v {pathlib.Path(__file__).parents[2]}:/tinygrad "
+                    f"-e PYTHONPATH=/ -e QEMU_CPU=max,pauth=off gcr.io/distroless/static python3 /tinygrad/runtime/support/compiler_qcom.py {arch}")
+      self.compiler_process = subprocess.Popen(argv.split(), stdout=subprocess.PIPE, stdin=subprocess.PIPE, bufsize=0)
     super().__init__(f"compile_qcomcl_{arch}")
 
-  def __del__(self): llvm_qcom.cl_compiler_destroy_llvm_instance(self.llvm_inst)
+  def __del__(self): llvm_qcom.cl_compiler_destroy_llvm_instance(self.llvm_inst) if platform.machine() == "aarch64" else self.compiler_process.kill()
 
   def __reduce__(self): return QCOMCompiler, (self.arch,)
 
@@ -25,6 +32,10 @@ class QCOMCompiler(Compiler):
     return handle
 
   def compile(self, src) -> bytes:
+    if platform.machine() != "aarch64":
+      unwrap(self.compiler_process.stdin).write(struct.pack("I", len(src.encode())) + src.encode())
+      if (lib:=unwrap(self.compiler_process.stdout).read(struct.unpack("I", unwrap(self.compiler_process.stdout).read(4))[0])): return lib
+      raise RuntimeError("QCOM Compilation Error")
     ch = self.checked(llvm_qcom.cl_compiler_compile_source(self.llvm_inst, self.chip_id, llvm_qcom.CL_MODE_64BIT, b"", 0, 0, 0, src.encode(), 0,
                                                            llvm_qcom.CL_SRC_STR, None))
     if DEBUG >= 8: print(system("llvm-dis", input=ctypes.string_at((comp:=ch.contents.compiled.contents).llvm_bitcode, comp.llvm_bitcode_size)))
@@ -36,3 +47,13 @@ class QCOMCompiler(Compiler):
     return ret
 
   def disassemble(self, lib: bytes): disas_adreno(lib[(ofs:=_read_lib(lib, 0xc0)):ofs+_read_lib(lib, 0x100)], self.chip_id)
+
+if __name__ == "__main__":
+  compiler = QCOMCompiler(sys.argv[1])
+  while (amt:=sys.stdin.buffer.read(4)):
+    try: lib = compiler.compile(sys.stdin.buffer.read(struct.unpack("I", amt)[0]).decode())
+    except Exception as e:
+      lib = b""
+      print(e, file=sys.stderr, flush=True)
+    sys.stdout.buffer.write(struct.pack("I", len(lib)) + lib)
+    sys.stdout.buffer.flush()

tinygrad/runtime/support/hcq.py

@@ -517,8 +517,7 @@ class HCQCompiled(Compiled, Generic[SignalType]):
     if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()
 
 class HCQBuffer:
-  def __init__(self, va_addr:sint, size:int, meta:Any=None, _base:HCQBuffer|None=None, view:MMIOInterface|None=None,
-               owner:HCQCompiled|None=None):
+  def __init__(self, va_addr:sint, size:int, meta:Any=None, _base:HCQBuffer|None=None, view:MMIOInterface|None=None, owner:Any=None):
     self.va_addr, self.size, self.meta, self._base, self.view = va_addr, size, meta, _base, view
     self._devs, self.owner = ([owner] if owner is not None else []), owner
     self._mappings:dict[HCQCompiled, HCQBuffer] = {} # mapping to the other devices

tinygrad/runtime/support/hcq2.py

@@ -1,14 +1,15 @@
 from __future__ import annotations
-from typing import cast, Callable, TypeVar, Generic, Any, Sequence
-import struct, functools, time, collections, itertools
+from typing import cast, Callable, TypeVar, Generic, Any, Sequence, Iterable
+import struct, functools, time, collections, itertools, decimal, statistics
 from dataclasses import replace, dataclass
-from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, dedup, pluralize, JIT_BATCH_SIZE, unwrap
-from tinygrad.helpers import to_tuple, round_up, partition, data64_le, panic, ContextVar
+from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, dedup, pluralize, JIT_BATCH_SIZE, unwrap, PROFILE
+from tinygrad.helpers import to_tuple, round_up, partition, data64_le, panic, ContextVar, perf_counter_us, Context
 from tinygrad.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer, DepsTracker
+from tinygrad.device import ProfileDeviceEvent, ProfileGraphEntry, ProfileGraphEvent
 from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, rewrite_group, GroupOp
 from tinygrad.uop.symbolic import symbolic
 from tinygrad.dtype import dtypes, truncate
-from tinygrad.runtime.support.hcq import MMIOInterface
+from tinygrad.runtime.support.hcq import MMIOInterface, HCQBuffer
 from tinygrad.runtime.support.memory import BumpAllocator
 from tinygrad.renderer import Renderer, Estimates
 from tinygrad.engine.realize import to_program, get_call_arg_uops, get_call_name, get_call_outs_ins, estimate_uop
@@ -21,7 +22,7 @@ HCQDeviceType = TypeVar('HCQDeviceType', bound='HCQ2Compiled')
 
 HCQ_RUNTIME_DEV = ContextVar("HCQ_RUNTIME_DEV", "CPU")
 
-HCQ_DEVS = frozenset(("AMD",))
+HCQ_DEVS = frozenset(("AMD", "CPU"))
 HCQ_P2P_DEVS = HCQ_DEVS | frozenset(("CPU",))
 HCQ_CACHE_TAGS = frozenset(("program", "systems", "template"))
 
@@ -32,12 +33,13 @@ class HCQInfo:
 
   input_idxs:tuple[int, ...] = () # indexes into input_uops used by this call
   inputs:int|None = None
+  kernels:tuple[tuple[tuple[str, ...], str, Estimates, tuple[int, ...]], ...] = ()
 
 def all_devices_in(d:Any, c:frozenset[str]) -> bool: return {x.split(":")[0] for x in to_tuple(d)} <= c
 
 def unwrap_mstack(u):
   if u.op is Ops.MSTACK: return tuple(x for s in u.src for x in unwrap_mstack(s))
-  return unwrap_mstack(u.src[0]) if u.op in {Ops.MSELECT, Ops.SLICE} else (u,)
+  return unwrap_mstack(u.src[0]) if u.op is Ops.MSELECT else (u,)
 
 def is_value_known_at_link(val:UOp) -> bool:
   runtime_reads = [u for u in val.toposort() if u.op in (Ops.LOAD, Ops.INDEX)]
@@ -85,21 +87,36 @@ def encode_kernargs_clike(call:UOp, prg:UOp, devs:str|tuple[str, ...]) -> UOp:
 def replace_call_buffers(ctx:tuple[list[UOp], dict[UOp, int]], call:UOp) -> UOp|None:
   bufs, slots = ctx
   for s in call.src[1:]:
-    if s.op not in (Ops.PARAM, Ops.BIND) and slots.setdefault(s, len(bufs)) == len(bufs): bufs.append(s)
-  return call.replace(src=call.src[:1] + tuple(s if s.op in (Ops.PARAM, Ops.BIND) else s.param_like(slots[s]) for s in call.src[1:]))
+    if s.op is not Ops.PARAM and not s.is_bound_var and slots.setdefault(s, len(bufs)) == len(bufs): bufs.append(s)
+  return call.replace(src=call.src[:1] + tuple(s if s.op is Ops.PARAM or s.is_bound_var else s.param_like(slots[s]) for s in call.src[1:]))
 pm_replace_buffers = PatternMatcher([(UPat(Ops.CALL, name="call"), replace_call_buffers)])
 
 # *****************
 # 1.1. prep: staging copies
 
-def _need_staging(a, b): return all_devices_in(a.device, HCQ_DEVS) and not all_devices_in(b.device, HCQ_P2P_DEVS)
+def _need_staging(a, b): return all_devices_in(a.device, HCQ_DEVS - {"CPU"}) and not all_devices_in(b.device, HCQ_P2P_DEVS)
+
+def _get_enqueue_devs(call:UOp) -> Any|None:
+  if not (bufs:=call.src[1:]) or not all(all_devices_in(b.device, HCQ_P2P_DEVS) for b in bufs): return None
+  devs = min(bufs, key=lambda b: to_tuple(b.device)[0].startswith("CPU")).device # prio to enqueue on not CPU device
+  return devs if all_devices_in(devs, HCQ_DEVS) else None
 
 def stage_copy(dst:UOp, src:UOp) -> UOp|None:
   if not (_need_staging(src, dst) or _need_staging(dst, src)): return None
 
   stage = UOp.new_buffer("CPU", src.max_numel() * src.dtype.itemsize, dtypes.uint8)
   return UOp(Ops.LINEAR, src=(src.copy_to_device("CPU").call(stage, src), stage.copy_to_device(dst.device).call(dst, stage)))
-pm_insert_copy_staging = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src"))), stage_copy)])
+
+def kernel_copy(call:UOp, dst:UOp, src:UOp) -> UOp|None:
+  if (devs:=_get_enqueue_devs(call)) is None or Device[(dev:=to_tuple(devs)[0])].has_copy_queue: return None
+  d, s = (UOp.param(i, dst.dtype, (n:=dst.max_numel(),), device=devs) for i in range(2))
+  ast = d.index(r:=UOp.range(n, 0)).store(s.index(r).load()).end(r).sink(arg=KernelInfo(name="copy"), tag=1)
+  return call.replace(src=(to_program(ast, Device[dev].renderer), dst, src))
+
+pm_insert_copy_staging = PatternMatcher([
+  (UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src"))), stage_copy),
+  (UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src")), name="call"), kernel_copy)
+])
 
 # *****************
 # 2. deps
@@ -122,19 +139,19 @@ def _get_deps(ctx:DepsTracker, bufs_by_lane:list[list[Any]], write, key:tuple[tu
 
 def _build_wait_cmds(slots:dict[str, int], dep_lanes:list[tuple[tuple, int, int]], devices:tuple[str, ...], queue:str) -> tuple[list[UOp], set[int]]:
   # opt1: same-queue ops are fifo-ordered
-  if devices[0].split(":")[0] in {"AMD", "QCOM"} or queue.startswith("COPY"):
+  if devices[0].split(":")[0] in {"AMD", "QCOM", "CPU"} or queue.startswith("COPY"):
     dep_lanes = [(dep, dlane, lane) for dep, dlane, lane in dep_lanes if (dep[0][dlane], dep[1]) != (devices[lane], queue)]
 
   # opt2: keep latest dep per (dep device, queue, cur lane)
   latest = {((dep[0][dlane], dep[1]), lane): (dep, dlane) for dep, dlane, lane in sorted(dep_lanes, key=lambda x: x[0][2])}
-  deps:dict[tuple, list[int|None]] = collections.defaultdict(lambda: [None]*len(devices))
-  for (_, lane), (dep, dlane) in latest.items(): deps[dep][lane] = dlane
+  deps:dict[tuple, dict[int, list[int]]] = collections.defaultdict(lambda: collections.defaultdict(list))
+  for (_, lane), (dep, dlane) in latest.items(): deps[dep][lane].append(dlane)
 
   waits = []
-  for (ddevs, dqueue, dtag), lanes in deps.items():
-    sig = UOp.mstack(*[make_signal(d, tag="sentinel_signal") if dl is None else make_signal(ddevs[dl], slots[dqueue])
-                       for dl, d in zip(lanes, devices)])
-    waits.append(UOp(Ops.INS, arg="wait", src=(sig, UOp.const(dtag + 1, dtypes.uint64))))
+  for (ddevs, dqueue, dtag), by_lane in deps.items():
+    for ls in itertools.zip_longest(*(by_lane[lane] for lane in range(len(devices)))):
+      s = UOp.mstack(*[make_signal(d, tag="sentinel_signal") if dl is None else make_signal(ddevs[dl], slots[dqueue]) for dl, d in zip(ls, devices)])
+      waits.append(UOp(Ops.INS, arg="wait", src=(s, UOp.const(dtag + 1, dtypes.uint64))))
   return waits, {dtag for _, _, dtag in deps}
 
 def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[tuple[tuple[str, ...], str]],
@@ -145,7 +162,7 @@ def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[t
     for b in itertools.chain.from_iterable(_get_call_bufs_by_lane(call, devices)):
       for bd in to_tuple(b.device): dev_bufs[bd][id(b)] = b
 
-  n, fences, fins, signal_tags = len(batch_info), [], [], set()
+  n, fences, resets, fins, signal_tags = len(batch_info), [], [], [], set()
   for _, devgroup in itertools.groupby(sorted(dev_bufs), key=lambda d: d.split(":")[0]):
     devs = tuple(devgroup)
 
@@ -159,18 +176,19 @@ def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[t
     fin_submit = make_submit(*waits, UOp(Ops.INS, arg="store", src=(tl_signal, tl_value.index(0))), devs=devs, queue="COMPUTE:0")
     epoch = (epoch_slot:=tl_value.after(fin_submit).index(0)).load()
 
-    # fence once per device group on this schedule's previous epoch, then reset any queue signals used by the group
+    # fence once per device group on this schedule's previous epoch
     qs = dedup([qn for bdevs, qn in batch_info if set(bdevs) & set(devs)])
     sched_epoch = make_signal(devs, next(UOp.unique_num))
 
     wait_device_epoch = (done:=tl_signal.after(loop:=UOp.loop(0)).index(0).load()).end(loop, done < sched_epoch.index(0).load())
-    resets = [make_signal(devs, slots[q]).after(wait_device_epoch).index(0).store(0) for q in qs]
+    fences.append(make_call("hcq_fence", UOp.sink(wait_device_epoch), HCQInfo(devs)))
 
-    fences.append(make_call("hcq_fence", UOp.sink(*(resets or [wait_device_epoch])), HCQInfo(devs)))
+    # queues of other groups wait on these signals, so reset them only after every group reached its epoch
+    if qs: resets.append(make_call("hcq_reset", UOp.sink(*[make_signal(devs, slots[q]).index(0).store(0) for q in qs]), HCQInfo(devs)))
     fins.append(make_call("hcq_finalizer", UOp.sink(epoch_slot.store(epoch + 1), sched_epoch.after(fin_submit).index(0).store(epoch)), HCQInfo(devs)))
-  return fences, fins, signal_tags
+  return fences + resets, fins, signal_tags
 
-def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]]) -> list[UOp]:
+def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]], profile:bool) -> list[UOp]:
   batch_info = [(devices, "COMPUTE:0" if call.src[0].op is Ops.PROGRAM else "COPY:0") for call, devices in batch]
 
   # schedule deps
@@ -185,10 +203,10 @@ def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]]) -> list[UOp]:
     signal_tags |= cur_signal_tags
 
   # build fences and finalizers
-  fences, finalizers, finalizer_signal_tags = _build_finalizers(batch, batch_info, deps_tracker, slots)
+  fences, fins, finalizer_signal_tags = _build_finalizers(batch, batch_info, deps_tracker, slots)
   signal_tags |= finalizer_signal_tags
 
-  src = []
+  src, kerns = [], []
   for tag, ((call, _), (devices, queue), q) in enumerate(zip(batch, batch_info, call_waits)):
     # first queue use, sync prior device work with the device timeline
     if batch_info.index((devices, queue)) == tag:
@@ -197,20 +215,27 @@ def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]]) -> list[UOp]:
 
     # and make hcq call
     name, info = get_call_name(call, get_call_arg_uops(call)), HCQInfo(devices, estimate_uop(call))
-    q += [call.replace(arg=replace(call.arg, aux=info))]
+    ts_ids = [next(UOp.unique_num) for _ in range(2)] if profile else []
+    kerns.append((devices, name, info.estimates, tuple(ts_ids)))
+
+    ts_ins = [UOp(Ops.INS, arg="timestamp", src=(make_signal(devices, s),)) for s in ts_ids]
+    q += ts_ins[:1] + [call.replace(arg=replace(call.arg, aux=info))] + ts_ins[1:]
 
     # signal the queue if someone waits for us
     if tag in signal_tags: q += [UOp(Ops.INS, arg="store", src=(make_signal(devices, slots[queue]), UOp.const(tag + 1, dtypes.uint64)))]
-    src.append(make_call(name, make_submit(*q, devs=devices, queue=queue).sink(), info))
-  return fences + src + finalizers
+    src.append(make_call(f"submit {name}", make_submit(*q, devs=devices, queue=queue).sink(), info))
 
-def sched_hcq_batches(l:UOp) -> UOp:
+  # append batch timestamps to finalizers
+  fins = [f.replace(arg=replace(f.arg, aux=replace(a:=f.arg.aux, kernels=tuple(x for x in kerns if set(x[0]) & set(a.device))))) for f in fins]
+  return fences + src + fins
+
+def sched_hcq_batches(l:UOp, profile:bool) -> UOp:
   srcs:list[UOp] = []
   batch:list[tuple[UOp, tuple[str, ...]]] = []
   for call in l.src:
-    if (devs:=next((b.device for b in call.src[1:] if all_devices_in(b.device, HCQ_DEVS)), None)) is not None: batch.append((call, to_tuple(devs)))
-    else: srcs, batch = srcs + _finalize_batch(batch) + [call], []
-  return l.replace(src=tuple(srcs + _finalize_batch(batch)))
+    if (devs:=_get_enqueue_devs(call)) is not None: batch.append((call, to_tuple(devs)))
+    else: srcs, batch = srcs + _finalize_batch(batch, profile) + [call], []
+  return l.replace(src=tuple(srcs + _finalize_batch(batch, profile)))
 
 # *****************
 # 3. merge into queues
@@ -246,7 +271,7 @@ def merge_queues(linear:UOp) -> UOp:
   return linear.replace(src=tuple(new_src + [_merged_hcq_call(c) for c in opened_qs.values()]))
 
 pm_schedule_and_merge = PatternMatcher([(UPat(Ops.LINEAR, name="l"),
-  lambda ctx, l: merge_queues(sched_hcq_batches(l).substitute(ctx, walk=True, enter_calls=True)))])
+  lambda ctx, l: merge_queues(sched_hcq_batches(l, ctx[1]).substitute(ctx[0], walk=True, enter_calls=True)))])
 
 # *****************
 # 4.2. hcq lowering: ops to ir
@@ -321,6 +346,10 @@ def split_patches(call:UOp) -> UOp|None:
   scatter = make_scatter_loops(input_patches, tables[0], lt_patches)
   body = body.substitute({p:p.substitute(scatter | reads) for p in rt_patches})
 
+  if inputs: # fence inputs
+    fills.append((t:=tables[0][0]).after(make_binary_patch(t, bytes(t.max_numel() * 8)))) # zeroed at link, slot 0 is the host fence
+    body = body.replace(src=(UOp.sink(*body.src[0].src, t.after(*body.src[0].src).index(0).store(0)),)) # open it once consumed
+
   lt_srcs = collections.defaultdict(list)
   for p in lt_patches: lt_srcs[p.buf_uop].append(p)
   return call.replace(src=(body, *call.src[1:], *[b.after(*ps) for b,ps in lt_srcs.items()], *fills),
@@ -329,9 +358,13 @@ pm_split_patches = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION
 
 # *****************
 
+def _rank_ranges(uops:Iterable[UOp]) -> dict[UOp, UOp]:
+  return {r: r.replace(arg=(i,)+r.arg[1:]) for i,r in enumerate(sorted([u for u in uops if u.op is Ops.RANGE], key=lambda r: r.arg))}
+
 def replace_params(call:UOp) -> UOp|None:
   body, variables, param_ops = call.src[0], call.src[0].variables(), {Ops.PARAM, Ops.MSTACK}
-  args = dedup([s for u in body.toposort(gate=lambda u: u.op not in param_ops) for s in u.src if s.op in param_ops and s not in variables])
+  tops = body.toposort(gate=lambda u: u.op not in param_ops)
+  args = dedup([s for u in tops for s in u.src if s.op in param_ops and s not in variables])
 
   patched, refhold = partition(call.src[1:], lambda x: x.src[0] in args)
   by_root = {p.src[0]: p for p in patched}
@@ -343,8 +376,8 @@ def replace_params(call:UOp) -> UOp|None:
   refhold += [a for a in addrs if a not in held and all(b.op is not Ops.PARAM or b.tag is not None for b in unwrap_mstack(a))]
 
   sub = {(b:=u.without_after): UOp.param(i, u.dtype, shape=b.shape, device=HCQ_RUNTIME_DEV.value, volatile=b.op is Ops.PARAM and b.arg.volatile)
-         for i,u in enumerate(c_args)} | {v: v.replace(arg=replace(v.arg, slot=-1)) for v in variables if v.op is Ops.PARAM}
-  info = replace(call.arg.aux, inputs=next((i for i,u in enumerate(c_args) if u.tag == "inputs"), None))
+         for i,u in enumerate(c_args)} | {v: v.replace(arg=replace(v.arg, slot=-1)) for v in variables if v.op is Ops.PARAM} | _rank_ranges(tops)
+  info = replace(call.arg.aux, inputs=next((i for i,u in enumerate(c_args) if u.without_after.tag == "inputs"), None))
   return call.replace(src=(body.substitute(sub).replace(arg="hcq_args"), *c_args, *refhold),
                       arg=replace(call.arg, aux=info)) # TODO: call.after(*refhold)?
 pm_replace_params = PatternMatcher([
@@ -352,14 +385,15 @@ pm_replace_params = PatternMatcher([
 
 # *****************
 
-def resolve_getaddr_slice(bv:UOp, g:UOp) -> UOp:
+def resolve_getaddr_view(bv:UOp, g:UOp) -> UOp:
   base = bv.src[0].after(*g.src[0].src[1:] if g.src[0].op is Ops.AFTER else ())
-  itemsize = bv.src[0].dtype.itemsize if bv.src[0].without_after.op in (Ops.BUFFER, Ops.SLICE, Ops.MSTACK, Ops.MSELECT) else bv.dtype.itemsize
+  if bv.op is Ops.BITCAST: return UOp(Ops.GETADDR, src=(base,), arg=g.arg)
+  itemsize = bv.src[0].dtype.itemsize if bv.src[0].without_after.op in (Ops.BUFFER, Ops.MSTACK, Ops.MSELECT) else bv.dtype.itemsize
   return UOp(Ops.GETADDR, src=(base,), arg=g.arg) + UOp.const(bv.src[1].val * itemsize, dtypes.uint64)
 
 pm_early_simplify = PatternMatcher([
-  (UPat(Ops.GETADDR, src=(UPat.any(sl:=UPat(Ops.SLICE, name="bv"), sl.after(allow_any_len=True)),), name="g"), resolve_getaddr_slice),
-  (UPat(Ops.INDEX, src=(UPat(Ops.SLICE, name="bv"),), allow_any_len=True, name="x"),
+  (UPat(Ops.GETADDR, src=(UPat((Ops.SHRINK, Ops.BITCAST), name="bv").or_after(),), name="g"), resolve_getaddr_view),
+  (UPat(Ops.INDEX, src=(UPat(Ops.SHRINK, name="bv"),), allow_any_len=True, name="x"),
    lambda bv,x: x.replace(src=(bv.src[0], x.src[1] + bv.src[1].cast(x.src[1].dtype), *x.src[2:]))),
 ])
 
@@ -377,7 +411,7 @@ def pack_hcq_placeholders(call:UOp) -> UOp|None:
       sizes[b.tag] = offs[b] + b.max_numel()
   counts = collections.Counter(b.tag for b in bufs)
   bases = {b.tag:UOp.placeholder((sizes[b.tag],), b.dtype, next(UOp.unique_num), device=b.device).rtag(b.tag) for b in bufs if counts[b.tag] > 1}
-  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(offs.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
+  subs = {b:bases[b.tag][(off:=offs.get(b, 0)):off+b.max_numel()] for b in bufs if b.tag in bases}
   return call.replace(src=(call.src[0].substitute(subs, walk=True), *call.src[1:])) if subs else None
 pm_pack_placeholders = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), pack_hcq_placeholders)])
@@ -391,21 +425,21 @@ def callify_hcq(call:UOp, cf:UOp) -> UOp:
 pm_callify_hcq = PatternMatcher([(UPat(Ops.CALL, src=(
   UPat(Ops.CUSTOM_FUNCTION, arg="hcq_args", src=(UPat(Ops.SINK),), name="cf"),), name="call", allow_any_len=True), callify_hcq)])
 
-hcq_compile_cache:dict[bytes, UOp] = {}
+hcq_compile_cache:dict[tuple[bytes, bool], UOp] = {}
 
-@rewrite_group(lambda linear,input_uops,ret: f"HCQ Compile {pluralize('Kernel', len(ret.src))}")
-def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None) -> UOp:
+@rewrite_group(lambda linear,input_uops,profile,ret: f"HCQ Compile {pluralize('Kernel', len(ret.src))}")
+def hcq_compile(linear:UOp, input_uops:list[UOp]|None, profile:bool) -> UOp:
   if input_uops is not None:
     slots = {u:i for i,u in reversed(tuple(enumerate(input_uops)))}
     linear = graph_rewrite(linear, pm_replace_buffers, ctx=(input_uops, slots), walk=True, name="replace buffer")
 
-  if (final_linear:=(hcq_compile_cache.get(cache_key:=linear.key))) is None:
+  if (final_linear:=(hcq_compile_cache.get(cache_key:=(linear.key, profile)))) is None:
     # prep
     linear = linear.substitute(back_map:={s.param_like(i): s for i,s in enumerate(input_uops)} if input_uops is not None else {}, walk=True)
     linear = graph_rewrite(linear, pm_insert_copy_staging+pm_flatten_linear, name="insert copy staging")
 
     # schedule
-    linear = graph_rewrite(linear, pm_schedule_and_merge, ctx={s:p for p,s in back_map.items()}, walk=True, name="schedule and merge hcq")
+    linear = graph_rewrite(linear, pm_schedule_and_merge, ctx=({s:p for p,s in back_map.items()}, profile), walk=True, name="schedule and merge hcq")
 
     # lowering to hcq ir
     linear = graph_rewrite(linear, pm_encode_cmdbufs+pm_pack_placeholders, walk=True, name="encode and pack", enter_calls=True)
@@ -443,7 +477,7 @@ def fold_binary(buf:UOp, blob:UOp) -> UOp:
 def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
   for off,val in zip(off.src, val.src):
     for b,v in zip((bs:=mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)), val.src if val.op is Ops.STACK else (val,)*len(bs)):
-      data = struct.pack(f'<{v.dtype.fmt}', truncate[v.dtype](v.val))
+      data = struct.pack(f'<{v.dtype.fmt}', truncate[v.dtype]((v.src[0] if v.op is Ops.CAST else v).val))
       b.ensure_allocated().as_memoryview(force_zero_copy=True, no_sync=True).cast('B')[(bo:=off.val*buf.dtype.itemsize):bo+len(data)] = data
   return UOp(Ops.NOOP)
 
@@ -458,7 +492,7 @@ def resolve_getaddr(buf:UOp, g:UOp) -> UOp:
 
 pm_resolve_patches = PatternMatcher([
   # multi
-  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
+  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat.any(UPat(Ops.CONST), UPat(Ops.CAST, src=(UPat(Ops.CONST),)))], name="op"), push_stack),
   (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),
 
   # getaddr
@@ -468,7 +502,7 @@ pm_resolve_patches = PatternMatcher([
   (UPat(name="buf").index(UPat(Ops.RANGE), allow_any_len=True)
     .store(UPat.any(UPat(Ops.BINARY, name="blob"), UPat(Ops.BINARY, name="blob").bitcast()).index(UPat(Ops.RANGE), allow_any_len=True).load())
     .end(UPat(Ops.RANGE)), fold_binary),
-  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").index(UPat(Ops.STACK, name="off")).store(UPat(Ops.STACK, name="val")), fold_const_store),
+  (UPat({Ops.BUFFER, Ops.MSTACK}, name="buf").index(UPat(Ops.STACK, name="off")).store(UPat(Ops.STACK, name="val")), fold_const_store),
 ])
 
 pm_assert_no_afters = PatternMatcher([(UPat(Ops.AFTER, name="a"), lambda a: panic(RuntimeError, f"AFTER left at hcq_link: {a.src[0].op}"))])
@@ -495,6 +529,7 @@ def hcq_link(linear:UOp, cache=True) -> UOp:
 
 class HCQ2Compiled(Compiled):
   timestamp_divider: float = 1000.0
+  wait_timeout_ms: float = 30000.0
 
   def __init__(self, device:str, allocator:HCQAllocator, compilers:list[type[Renderer]], runtime, can_recover:bool=False, arch=None):
     self.device_id:int = int(device.split(":")[1]) if ":" in device else 0
@@ -512,6 +547,27 @@ class HCQ2Compiled(Compiled):
 
     self.rt_buffer = Buffer(self.device, 64 << 20, dtypes.uint8, options=BufferSpec(uncached=True, cpu_access=True))
     self.rt_allocator = BumpAllocator(64 << 20)
+    self.prof_ents:dict[int, ProfileGraphEntry] = {}
+
+  def collect_prof(self):
+    if PROFILE:
+      es = list(self.prof_ents.values())
+      sigs = [self.signal(i)._buf.cpu_view().view(fmt='Q')[0]/decimal.Decimal(self.timestamp_divider) for e in es for i in (e.st_id, e.en_id)]
+      Compiled.profile_events.append(ProfileGraphEvent([replace(e, st_id=2*i, en_id=2*i+1) for i,e in enumerate(es)], [], sigs))
+    self.prof_ents.clear()
+
+  def _at_profile_finalize(self):
+    from tinygrad.tensor import Tensor
+    tdiffs = []
+    for _ in range(5):
+      with Context(DEBUG=0, BEAM=0, TRACK_MATCH_STATS=0): Tensor.ones(1, device=self.device).contiguous().realize()
+      if not (ents:=list(self.prof_ents.values())): return
+      self.prof_ents.clear()
+      st = perf_counter_us()
+      self.synchronize()
+      gpu = max(self.signal(e.en_id)._buf.cpu_view().view(fmt='Q')[0] for e in ents)/decimal.Decimal(self.timestamp_divider)
+      tdiffs.append((st+perf_counter_us())/2 - gpu)
+    Compiled.profile_events.append(ProfileDeviceEvent(self.device, statistics.median(tdiffs), self.device_props()))
 
   def new_buffer(self, b:UOp, cache:bool) -> Buffer:
     if cache or b.tag in HCQ_CACHE_TAGS:
@@ -525,13 +581,16 @@ class HCQ2Compiled(Compiled):
     return buf
 
   def synchronize(self, timeout:int|None=None):
-    if not hasattr(self, 'iface'): return
+    if HCQ_RUNTIME_DEV.value != self.device: Device[HCQ_RUNTIME_DEV.value].synchronize()
+
     sig = self.signal("timeline").as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
     tl = self.signal("value", 1).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
     timeout = timeout if timeout is not None and self.can_recover else None
-    st = time.perf_counter()
-    while sig[0] < tl[0] - 1:
-      if time.perf_counter() - st > (timeout or 3000) / 1000: self.on_device_hang()
+    st, done = time.perf_counter(), sig[0]
+    while done < tl[0] - 1:
+      if done != (done:=sig[0]): st = time.perf_counter()
+      elif time.perf_counter() - st > (timeout or self.wait_timeout_ms) / 1000: self.on_device_hang()
+    if self.prof_ents: self.collect_prof()
 
   def on_device_hang(self): raise RuntimeError(f"{self.device} hang detected")
 
@@ -568,21 +627,23 @@ class HCQ2Buffer:
     return HCQ2Buffer(self.va_addr+offset, meta=self.meta, view=(self.view.view(offset=offset, size=size) if self.view is not None else None))
 
 class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
-  def _as_buffer(self, buf:HCQ2Buffer) -> memoryview:
+  def _as_buffer(self, buf:HCQBuffer) -> memoryview:
     return unwrap(buf.view).mv
 
-  def _map(self, buf:HCQ2Buffer) -> HCQ2Buffer:
+  def _map(self, buf:HCQBuffer) -> HCQBuffer:
     if not hasattr(self, '_do_map'): raise NotImplementedError("map failed: no method implemented")
     return self._do_map(buf)
 
+  def _do_unmap(self, mb): self.dev.iface.free(mb)
+
   @suppress_finalizing
-  def _free(self, buf:HCQ2Buffer, options:BufferSpec|None=None):
+  def _free(self, buf:HCQBuffer, options:BufferSpec|None=None):
     if options is not None and options.external_ptr is not None: return
     self.dev.synchronize()
     if hasattr(self, '_do_free'): self._do_free(buf, options)
 
   def _unmap(self, mb):
     self.dev.synchronize()
-    self.dev.iface.free(mb)
+    self._do_unmap(mb)
 
-  def _offset(self, buf, size:int, offset:int) -> HCQ2Buffer: return buf.offset(offset=offset, size=size)
+  def _offset(self, buf, size:int, offset:int) -> HCQBuffer: return buf.offset(offset=offset, size=size)

tinygrad/runtime/support/system.py

@@ -1,7 +1,8 @@
 from __future__ import annotations
-import os, mmap, array, functools, ctypes, select, contextlib, dataclasses, sys, itertools, struct, socket, subprocess, time, enum, atexit
+import os, mmap, array, functools, ctypes, ctypes.util, select, contextlib, dataclasses, sys, itertools, struct, socket
+import subprocess, time, enum, atexit
 from tinygrad.helpers import round_up, getenv, OSX, temp, ceildiv, unwrap, fetch, system, _ensure_downloads_dir, DEBUG, flatten, pluralize
-from tinygrad.runtime.autogen import libc, pci, vfio, iokit, corefoundation
+from tinygrad.runtime.autogen import libc, pci, vfio
 from tinygrad.runtime.support.hcq import FileIOInterface, MMIOInterface, HCQBuffer, hcq_filter_visible_devices
 from tinygrad.runtime.support.memory import VirtMapping, AddrSpace, BumpAllocator
 from tinygrad.runtime.support.usb import USB3, CustomASM24Controller, USBMMIOInterface
@@ -55,6 +56,7 @@ class _System:
   def pci_scan_bus(self, vendor:int, devices:tuple[tuple[int, tuple[int, ...]], ...], base_class:int|None=None) -> list[str]:
     all_devs = []
     if OSX:
+      from tinygrad.runtime.autogen import iokit, corefoundation
       def read_prop(svc, key) -> int:
         cfkey = corefoundation.CFStringCreateWithCString(None, key.encode(), corefoundation.kCFStringEncodingUTF8)
         cfdata = ctypes.cast(iokit.IORegistryEntryCreateCFProperty(svc, ctypes.cast(cfkey, iokit.CFStringRef), None, 0), corefoundation.CFDataRef)

tinygrad/schedule/__init__.py

@@ -8,14 +8,13 @@ from tinygrad.helpers import DEBUG, cpu_profile, TracingKey, SPEC, pluralize, SC
 
 # unwrap VIEW/CAST/etc to find the actual data source (kernel output, buffer, or multi-device op)
 def _unwrap_src(s: UOp) -> UOp:
-  while len(s.src) and s.op not in {Ops.AFTER, Ops.BUFFER, Ops.PARAM, Ops.MSELECT, Ops.MSTACK, Ops.BIND}: s = s.src[0]
+  while len(s.src) and s.op not in {Ops.AFTER, Ops.BUFFER, Ops.PARAM, Ops.MSELECT, Ops.MSTACK}: s = s.src[0]
   return s
 
-# a buffer state is AFTER | BUFFER | PARAM. MSELECT/MSTACK join per-device states, BIND is not a buffer dependency
+# a buffer state is AFTER | BUFFER | PARAM. MSELECT/MSTACK join per-device states
 def _states(s: UOp) -> list[UOp]:
   s = _unwrap_src(s)
   if s.op in {Ops.MSELECT, Ops.MSTACK}: return [st for ss in s.src for st in _states(ss)]
-  if s.op is Ops.BIND: return []
   assert s.op in {Ops.AFTER, Ops.BUFFER, Ops.PARAM}, f"input to kernel must resolve to a buffer state, not {s.op}"
   return [s]
 
@@ -71,7 +70,7 @@ def create_schedule(sched_sink:UOp) -> UOp:
       else:
         k = rk.src[0] if rk.op is Ops.END else rk
         assert k.op is Ops.CALL, f"unexpected op in queue: {k.op}"
-        buf_uops = tuple(_unwrap_src(s).buf_uop for s in k.src[1:] if s.op is not Ops.BIND)
+        buf_uops = tuple(_unwrap_src(s).buf_uop for s in k.src[1:] if not s.is_bound_var)
         linearized.append(k.src[0].call(*buf_uops))
       for x in children.get(rk, []):
         in_degree[x] -= 1
@@ -98,10 +97,15 @@ pm_post_sched_cache = PatternMatcher([
    create_new_buffer(ctx, b) if isinstance(b.arg, ParamArg) and b.addrspace is AddrSpace.GLOBAL else None),
 ])
 
+def resolve_linear_call(linear_call:UOp):
+  linear = graph_rewrite(linear_call.src[0], pm_post_sched_cache, ctx=({}, linear_call.src[1:]), walk=True, name="params to buffers")
+  # map the call body params back to the original Variables stored in the call args
+  binds = {f"p{i}":x.src[0].replace(op=Ops.PARAM) for i,x in enumerate(linear_call.src[1:]) if x.is_bound_var}
+  return linear.substitute({v:binds[v.expr] for v in linear.variables() if v.expr in binds}, enter_calls=True, name="resolve scalar params")
+
 pm_resolve_linear_call = PatternMatcher([
   # call LINEAR is resolved here
-  (UPat(Ops.CALL, src=(UPat(Ops.LINEAR),), name="linear_call", allow_any_len=True), lambda linear_call:
-   graph_rewrite(linear_call.src[0], pm_post_sched_cache, ctx=({}, linear_call.src[1:]), walk=True, name="params to buffers")),
+  (UPat(Ops.CALL, src=(UPat(Ops.LINEAR),), name="linear_call", allow_any_len=True), resolve_linear_call),
 ])+pm_flatten_linear
 
 schedule_cache: dict[bytes, UOp] = {}
@@ -180,13 +184,13 @@ def create_linear_with_vars(big_sink:UOp) -> tuple[UOp, dict[str, int]]:
 
   # vars used in the schedule
   used_vars = set().union(*[{v.expr for v in si.src[0].variables()} for si in linear.src])
-  # get var_vals
+  # get var_vals from the bound Variables in the call args
   var_vals: dict[str, int] = {}
   for b in big_sink.src[1:]:
-    if b.op is Ops.BIND:
-      nm = b.src[0].expr
+    if b.is_bound_var:
+      v, val = b.unbind()
+      nm = v.expr
       if nm not in used_vars: continue
-      val = b.src[1].val
       if var_vals.get(nm, val) != val: raise RuntimeError(f"bind mismatch on {nm}, {var_vals[nm]} != {val}")
       var_vals[nm] = val