kimi linear working

llm: reuse exact recurrent prefixes
llm: fully warm recurrent serving at startup
2026-08-10 15:37:20 +00:00 · 2026-08-10 14:06:15 +00:00 · 2026-08-10 14:06:15 +00:00 · 2026-08-10 14:06:15 +00:00 · 2026-08-10 14:06:15 +00:00 · 2026-08-10 14:06:15 +00:00
258 changed files with 13613 additions and 4285 deletions
@@ -137,6 +137,7 @@ runs:
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true')
      shell: bash
      run: |
+        sudo mkdir -p /var/cache/apt/archives
        sudo chown -R $USER:$USER /var/cache/apt/archives

        echo 'Acquire::GzipIndexes "true";' | sudo tee /etc/apt/apt.conf.d/gzip
@@ -214,6 +215,7 @@ runs:
          sudo apt-get -y --allow-unauthenticated --no-install-recommends install ${{ steps.apt-pkgs.outputs.pkgs }}
        fi

+        sudo mkdir -p /var/cache/apt/archives
        sudo chown -R $USER:$USER /var/cache/apt/archives/

    - name: Add clang to PATH (Linux)
@@ -94,6 +94,7 @@ jobs:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
+      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -104,6 +105,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -117,10 +121,10 @@ jobs:
      run: python3 test/external/process_replay/reset.py
    - name: Run llama3.2
      run: BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
-    - name: Run qwen3.5
-      # qwen3.5:35b-a3b doesn't fit on mac
+    - name: Run qwen3.6
+      # qwen3.6:35b-a3b doesn't fit on mac
      if: ${{ matrix.dev != 'METAL' }}
-      run: BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
+      run: BENCHMARK_LOG=qwen36_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.6:35b-a3b --benchmark --warmup
    - name: Run olmoe
      # just metal for now
      if: ${{ matrix.dev == 'METAL' }}
@@ -145,6 +149,7 @@ jobs:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
+      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -155,6 +160,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -194,6 +202,7 @@ jobs:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
+      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -204,6 +213,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p extra/datasets
@@ -240,6 +252,7 @@ jobs:
        shell: bash -e -o pipefail {0}
    env:
      DEV: ${{ matrix.dev }}
+      HCQ2: ${{ matrix.dev == 'AMD' && '1' || '0' }}
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -250,6 +263,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -292,6 +308,9 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Setup (NV)
+      if: ${{ matrix.dev == 'NV' }}
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -402,6 +421,35 @@ jobs:
      run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py

  testcommalatest:
+    name: comma Benchmark (0.11.2)
+    runs-on: [self-hosted, Linux, comma]
+    timeout-minutes: 12
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: openpilot compile3 0.11.2 supercombo
+      run: BENCHMARK_LOG=openpilot_0_11_2_supercombo PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
+    - name: openpilot compile3 0.11.2 supercombo (from pickle)
+      run: BENCHMARK_LOG=openpilot_0_11_2_supercombo_run_pickle RUN_PICKLE=1 PYTHONPATH="." ASSERT_MIN_STEP_TIME=26 DEV=QCOM taskset -c 4-7 python3 examples/openpilot/compile3.py
+    - name: IR3 openpilot compile3 0.11.2 supercombo
+      run: BENCHMARK_LOG=ir3_openpilot_0_11_2_supercombo PYTHONPATH="." ASSERT_MIN_STEP_TIME=41 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
+    - name: openpilot compile3 0.11.2 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_11_2_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/3e7b31dfbc0a5234f1baf196513b77fc6af12204b8a8ffe8ee0417e48352f316
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testcommaold:
    name: comma Benchmark (0.11.0)
    runs-on: [self-hosted, Linux, comma]
    timeout-minutes: 12
@@ -432,35 +480,6 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

-  testcommaold:
-    name: comma Benchmark (0.10.1)
-    runs-on: [self-hosted, Linux, comma]
-    timeout-minutes: 12
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: reset process replay
-      run: test/external/process_replay/reset.py
-    - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
-      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3.2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
  testqualcommdsp:
    name: DSP Benchmark
    runs-on: [self-hosted, Linux, comma4]
@@ -515,12 +534,12 @@ jobs:
      run: |
        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." GMMU=0 DEV=USB+AMD:LLVM ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot load_pickle 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_load_pickle PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
-    - name: openpilot run_pickle 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py
+    - name: openpilot compile3 big_driving_supercombo
+      run: BENCHMARK_LOG=usbgpu_openpilot_big_driving_supercombo PICKLE_OOB=1 PYTHONPATH="." TC_OPT=2 GMMU=0 DEV=USB+AMD:LLVM ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/10926f2c0911821ca0e72439c1c3bf3ec11f0a08789aa14b7ee8f25379b2afa4 openpilot.pkl
+    - name: openpilot load_pickle big_driving_supercombo
+      run: BENCHMARK_LOG=usbgpu_openpilot_big_driving_supercombo_load_pickle PICKLE_OOB=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_LOAD_TIME=25 python3 examples/openpilot/load_pickle.py openpilot.pkl
+    - name: openpilot run_pickle big_driving_supercombo
+      run: BENCHMARK_LOG=usbgpu_openpilot_big_driving_supercombo_run_pickle RUN_PICKLE=1 PICKLE_OOB=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py - openpilot.pkl
    - name: Test copy speeds
      run: SIZE=64e6 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3 test/external/external_test_usb_asm24.py TestDevCopySpeeds

@@ -0,0 +1,213 @@
+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
+
+
+  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
@@ -167,6 +167,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-13
+        python-version: '3.11'
        deps: testing_unit
        llvm: 'true'
        amd: 'true'
@@ -176,13 +177,14 @@ jobs:
      run: |
        DEV=NULL python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
        DEV=NULL VIZ=1 python3 -m pytest -n=auto test/null/test_viz.py
+        DEBUG=7 python -m tinygrad.viz.cli --json | jq empty
    # TODO: too slow
    # - name: Run SDXL on NULL backend
    #   run: DEV=NULL DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
    - name: Run Clip tests for SD MLPerf on NULL backend
      run: DEV=NULL python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
    - name: Run AMD emulated BERT training on NULL backend
-      run: DEV=NULL::gfx1201 NULL_ALLOW_COPYOUT=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: DEV=NULL::gfx1201 NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    # TODO: support fake weights
    #- name: Run LLaMA 7B on 4 fake devices
    #  run: DEV=NULL python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
@@ -199,6 +201,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-13
+        python-version: '3.11'
        pydeps: "pre-commit"
        deps: testing_unit
        llvm: 'true'
@@ -216,8 +219,8 @@ jobs:
      run: python3 test/external/external_benchmark_schedule.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
-    - name: Repo line count < 25000 lines
-      run: MAX_LINE_COUNT=25000 python sz.py
+    - name: Repo line count <= 26000 lines
+      run: MAX_LINE_COUNT=26000 python sz.py

  spec:
    strategy:
@@ -291,7 +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: |
@@ -350,6 +353,8 @@ jobs:
        run: DEV=NULL NULL_ALLOW_COPYOUT=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
      - name: Test llama 3 training
        run: DEV=NULL NULL_ALLOW_COPYOUT=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+      - name: Test gpt-oss training
+        run: DEV=NULL NULL_ALLOW_COPYOUT=1 SAMPLES=32 BS=2 SEQLEN=128 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 MXFP8=1 VOCAB_SIZE=32000 LAYERS=2 EXPERTS=4 MODEL=gptoss PYTHONPATH=. python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -478,7 +483,7 @@ jobs:
          wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
          echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-21 main" | sudo tee /etc/apt/sources.list.d/llvm.list
          sudo apt-get update
-          sudo apt-get install llvm-21 llvm-21-tools cloc
+          sudo apt-get install -y llvm-21 llvm-21-tools cloc
      - name: Install rocprof-trace-decoder
        run: sudo PYTHONPATH="." ./extra/sqtt/install_rocprof_decoder.py
      - name: Run AMD renderer tests
@@ -580,8 +585,11 @@ 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 --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 test/opt/test_tensor_cores.py --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
@@ -624,150 +632,6 @@ jobs:
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

-# ****** OSX Tests ******
-
-  unittestmacos:
-    name: MacOS (unit)
-    runs-on: &macos 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
-        amd: 'true'
-        ocelot: 'true'
-    - name: Run unit tests
-      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
-    - name: Run NULL backend tests
-      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --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 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
-    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
-    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:
@@ -804,33 +668,3 @@ jobs:
        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
@@ -3,3 +3,4 @@
 - Run tests with `-n12` for speed (e.g. `python -m pytest test/null/test_dtype.py -x -q -n12`)
 - Run `python -m mypy tinygrad/` to typecheck
 - Run `python -m ruff check .` to lint
+- Read `./tinygrad/viz/README.md` for profiling and debugging rewrite rules
@@ -88,7 +88,7 @@ def example_3_custom_uop(a:Tensor, correct):

    # store all the per lane accumulators to LOCAL
    local_accs = UOp.placeholder((LCLS,), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL)
-    local_accs = local_accs.after(local_accs[lane].store(acc[0]).barrier())
+    local_accs = local_accs.after(local_accs[lane].store(acc[0]))

    # accumulate LOCALs into a single per CU accumulator
    late_reduce_loop = UOp.range(LCLS, 3, AxisType.REDUCE)
@@ -1,7 +1,8 @@
 ::: tinygrad.dtype.DType

-::: tinygrad.dtype.dtypes
+::: tinygrad.dtype.DTypes
    options:
+        heading: dtypes
        members: true
        members_order: source
        show_labels: false
@@ -0,0 +1,224 @@
+# Kimi K3 on 8× MI350X
+
+This branch targets text generation directly from the official `moonshotai/Kimi-K3` checkpoint at `/raid/weights/kimi-k3`. It intentionally ignores the vision tower and multimodal projector. The checkpoint remains in its official 96-shard format; the loader never converts, rewrites, or creates a second 1.56 TB copy.
+
+The checked TP8 layout consumes 196.78 GB (183.27 GiB) of text weights per GPU. The compressed MLA cache adds 28.99 GB (27 GiB) per GPU at the full 1,048,576-token context, leaving approximately 62.23 GB of each nominal 288 GB MI350X for execution buffers and allocator overhead. Start much smaller.
+
+## Resume the current optimization session
+
+Work on branch `kimi_slop`. It was cleanly rebased onto `origin/kimi_slop` commit `553bdf68e` on 2026-08-10. The retained K3 commits after that base are `1b3732a6e`, `c6ac4961d`, `1d8620471`, `224bac031`, `f53f0e7e7`, and `2b1b8c22a`; verify the current hashes with `git log` because a later rebase may rewrite them. Before starting any benchmark, check that the worktree is clean and that no model process remains:
+
+```sh
+git status --short --branch
+git log --oneline --decorate -10
+pgrep -af 'tinygrad.llm.cli|benchmark_kimi_k3' || true
+```
+
+The active acceptance target is **more than 100 tok/s decode, more than 200 tok/s prefill, and less than 180 seconds cold startup** on TP8/gfx950. None is currently met. The authoritative official-checkpoint baseline is 389.84 seconds startup, 38.65 tok/s prefill, and 6.25 tok/s decode. The 1.56 TB checkpoint has a measured 6.9 GB/s single-XFS-NVMe read ceiling, giving a roughly 227-second physical cold-read floor; meeting the startup target therefore also requires a faster storage path, not only loader code.
+
+Use the fake-weight, one-layer loop for development. Do not repeatedly load the official checkpoint while optimizing:
+
+```sh
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode attention --iterations 30
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode block --iterations 30
+PROFILE=1 DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode block --iterations 5
+```
+
+The clean retained baseline is about 0.630 ms per attention layer and 1.37 ms per complete block, with fake initialization taking about 0.9/2 seconds respectively after the rebase. Since K3 has 93 sequential blocks, a 100 tok/s projection requires at most approximately 0.108 ms per complete block. Only run another 96-shard official validation after a candidate produces a large whole-block gain, remains finite and deterministic, and passes a direct numerical comparison. Test one candidate at a time and remove failed experiments before moving on.
+
+The immediate bottleneck is launch and synchronization granularity: an official four-token decode profile contained 6,304 kernel events, while packed expert work was only a small fraction of total GPU time. Continue with whole-component or whole-block fusion/replay work, not isolated expert microkernels. The latest fake-loop A/B retested the previously rejected dual gate/up and weighted-down MFMA prototypes: 1.374 ms baseline versus 1.375 ms fused, so they were removed again. A fused whole-core KDA recurrence was also slower in the exact fake attention gate (0.665 versus 0.633 ms) and must not be restored unchanged.
+
+Preserve these invariants when official validation resumes: use `/raid/weights/kimi-k3` directly, keep all 96 shards byte-for-byte untouched, run only one model process, begin at context 128, verify all eight devices are `gfx950`, and preserve the first failure instead of retrying over it. The most recent preserved official failure from a rejected KDA experiment was the invalid sequence `[198, 163840, 163840, 163840]`; token 163840 is outside the valid vocabulary. The retained path before that experiment produced deterministic in-range replay.
+
+After a synthetic candidate passes, run correctness and performance in this order: NULL gfx950 compile coverage, focused tests with `-n12` where supported, TP8 fake numerical comparison, official context-128 deterministic tokens, load/prefill/decode timing, and then context admission at 4K, 32K, 131K, and 262K. Run `python -m mypy tinygrad/` and `python -m ruff check .` when those tools are installed. Read `tinygrad/viz/README.md` before inspecting rewrite or device profiles.
+
+## Before renting the machine
+
+- Keep the existing 96 shards in `/raid/weights/kimi-k3`; no additional model-sized free space is required. Leave ordinary headroom for logs and temporary files.
+- The host should have roughly 3 TB RAM, in line with AMD's MI350X platform guidance. The loader itself is streaming and must not need checkpoint-sized RAM.
+- Use a recent kernel/ROCm stack supported by the host vendor, although tinygrad uses its own AMD userspace driver when `DEV=AMD`.
+- Clone this exact commit/branch and keep the official checkpoint directory separate from the repository.
+
+Validate the existing directory without modifying it:
+
+```sh
+python examples/kimi_k3_prepare.py /raid/weights/kimi-k3 --context 4096
+```
+
+For a metadata-only preflight, place the official `config.json` and `model.safetensors.index.json` in a directory and run:
+
+```sh
+python examples/kimi_k3_prepare.py /raid/weights/kimi-k3 --metadata-only
+```
+
+## Hardware admission checks
+
+Do these before loading weights. Stop if any device is missing or reports a different architecture.
+
+```sh
+lspci -d 1002:75a0
+amd-smi list
+DEV=AMD DEBUG=2 python - <<'PY'
+from tinygrad import Device
+for i in range(8):
+  dev = Device[f"AMD:{i}"]
+  print(i, dev.arch)
+PY
+```
+
+Expected architecture: `gfx950` on all eight devices. Then run the small TP8 graph tests:
+
+```sh
+python -m pytest test/unit/test_llm_k3.py test/null/test_kimi_k3.py -q -n12
+DEV=NULL:HIP:gfx950 NULL_ALLOW_COPYOUT=1 python -m pytest \
+  test/unit/test_llm_k3.py::TestKimiK3::test_chunked_recurrent_generate -q -n1
+DEV=AMD python examples/kimi_k3_smoke.py --devices 8
+```
+
+The last two commands are deliberately small. They compile CDNA4 kernels and then exercise the complete TP8 topology without loading the checkpoint.
+
+For performance iteration, use the exact-width fake-weight harness before another official load:
+
+```sh
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode attention --iterations 20
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode block --iterations 20
+```
+
+It retains K3's 7,168-wide residual stream, 12,288-wide KDA state, 96 heads, 128×128 recurrent matrices, TP8 layouts, top-k 16 routing, packed MXFP4 expert shapes, collectives, and decode JIT, but uses one layer and 16 fake experts. Fake attention weights initialize in about 0.9 seconds and the full block in about 3 seconds. The retained path measured 0.630 ms per fake attention layer and 1.367 ms per complete fake block, projecting about 7.87 tok/s across 93 identical blocks versus 6.25 tok/s for the official heterogeneous model. Treat this as a candidate admission benchmark, not a correctness substitute for official weights.
+
+## First official load
+
+Start at a short context so cache allocation and compilation are bounded. The loader reads disk-backed safetensors, TP-shards every destination before realizing it, and drops each source shard/projection immediately afterward.
+
+```sh
+/usr/bin/time -v env DEV=AMD DEBUG=1 python -m tinygrad.llm.cli \
+  --model /raid/weights/kimi-k3 --devices 8 --max_context 128 </dev/null 2>&1 | tee kimi-k3-load.log
+```
+
+Watch host RAM, swap, HBM, temperatures, and XGMI traffic from a second terminal. Do not start with a one-million-token cache. If loading fails, preserve the first exception and the last loader progress line; do not retry with a larger host-side cache.
+
+## Correctness and performance sequence
+
+1. Load with context 128 and generate one token.
+2. Repeat a fixed prompt twice and confirm token-for-token deterministic greedy output.
+3. Compare the first several greedy tokens against the official Transformers implementation at temperature zero.
+4. Benchmark decode only after two warm-up tokens.
+5. Benchmark prefill at 128, 512, 2K, and 8K tokens. Increase context only while HBM and compile time remain healthy.
+6. Use `VIZ=1` plus `python -m tinygrad.viz.cli` to inspect kernels; use `VIZ=2` only for short SQTT captures because it adds overhead.
+
+Example decode benchmark:
+
+```sh
+DEV=AMD DEBUG=1 python -m tinygrad.llm.cli --model /raid/weights/kimi-k3 \
+  --devices 8 --max_context 4096 --warmup --benchmark 20
+```
+
+## MI350X validation results (2026-08-10)
+
+The official directory was audited in place: 96 shards, 497,220 indexed tensors, 497,052 language tensors, and 1,560,860,324,864 total bytes. All eight devices reported `gfx950`. No checkpoint file was converted, copied, or modified, and every model run used a single process. The actual text tower is 1,559,965,606,912 bytes; its checked TP8 layout is 196,784,397,312 bytes per GPU.
+
+The preserved first full-checkpoint error was an `A_log` shape mismatch, `(128,) -> (96, 1)`. K3 stores one decay value per 128-wide KDA channel, not one per head. The loader now keeps this field replicated and applies the official channel-wise broadcast. A numerical unit test covers the distinction from the older head-wise Kimi Linear behavior.
+
+Load speed was fixed before generation. The original loader opened thousands of individual expert tensors and independently realized eight strided TP slices. The MI350 path now does the following without changing the checkpoint:
+
+- parses safetensor headers selectively, constructing disk-backed tensors only for the 2,460 non-expert entries consumed by that pass instead of materializing metadata objects for every expert entry twice;
+- copies contiguous axis-zero shards and replicas directly into their final device buffers;
+- reads a replicated tensor once and fans it out over XGMI instead of issuing eight identical direct reads (14.31 GB less RAID traffic);
+- stages an inner-axis tensor once and schedules all eight TP slices together;
+- reads each layer's contiguous 15.72 GB expert region once, reorders its lexicographically stored expert records on GPU 0, and realizes all six packed/scale destinations together;
+- retains only final MultiBuffer identities, drops the reorder graph, and flushes the 15.72 GB staging allocation before the next layer.
+
+One real expert layer leaves exactly 1,965,293,568 bytes resident on each GPU and zero bytes in the GPU-0 allocator cache. Complete context-128 loads measured 527.20 seconds before the final staging cleanup and 490.05/489.59 seconds afterward. Peak host RSS for the unprofiled correctness run was 2.11 GiB with zero swap. RAID variability produced later loads from 489.06 to 532.85 seconds.
+
+The selective-metadata and bounded-GC pass reduced non-expert loading from 125.77 to 57.77 seconds. A subsequent full official context-128 load completed in 411.49 seconds, 78.10 seconds (16.0%) faster than the 489.59-second baseline. It read the 96 shards in place with 1,049,688 KiB peak host RSS and zero swap; no weight payload was converted, copied, or modified. Direct-I/O probes measured approximately 6.9 GB/s aggregate for both one and eight concurrent 1 GiB reads. At that rate the 1.56 TB checkpoint has a roughly 227-second cold-read lower bound, so this RAID cannot meet a true cold sub-three-minute startup regardless of loader overhead.
+
+Expert staging graphs are acyclic and are released by reference counting after each layer, so the loader now suppresses unnecessary cyclic-collector scans only around that loop and restores its prior state on every exit. A quiet context-128 load then completed in 391.54 seconds, 30.14 seconds (7.1%) faster than the immediately preceding 421.68-second run, with 1.04 GiB peak RSS and zero swap, although storage variability contributes to run-to-run timing. The host used for these measurements actually mounts `/raid` from one 3.5 TB XFS NVMe, not a multi-drive RAID; shard 28 has 218 extents and live reads fell to roughly 160 MB/s there. This storage layout, plus the physical checkpoint size, remains the limiting cold-start constraint. The weights were not defragmented, copied, or modified.
+
+The fixed XTML prompt `Reply with exactly: OK` encodes to 93 tokens. After excluding the cold JIT capture from replay comparison, two greedy runs produced the identical eight-token sequence:
+
+```text
+[9545, 59991, 10580, 14404, 9545, 59991, 9545, 59991]
+```
+
+At context 128, steady prefill was 14.32 seconds (6.49 tok/s) and eight-token decode was 2.27 seconds (3.53 tok/s, 283.3 ms/token). The same first tokens remained stable at every admitted context. These rates are much lower than the planning estimates below and should be treated as the current measured baseline.
+
+The retained gfx950 serving pass enables the validated wave64 recurrent prefill kernel with 128-token chunks, uses exact BF16 decode projections, combines the routed/shared final TP partials into one collective, and tiles four adjacent packed-expert outputs during multi-token execution. On the same 93-token prompt, two replay trials produced the identical sequence `[198, 92652, 220, 80225]`. Prefill replay measured 2.418--2.482 seconds (37.47--38.46 tok/s), and eight-token decode measured 1.294 seconds (6.18 tok/s, 161.81 ms/token). Peak RSS was 2.77 GiB with zero swap. The packed prefill tile changes floating-point reduction order: direct official-layer comparison against the original kernel had maximum differences of 0.015625 for gate and 0.0078125 for down, and the end-to-end greedy sequence was stable across replay.
+
+A subsequent gfx950 decode pass split the 7,168-wide replicated BF16 projections across eight waves per 16 output channels and used CDNA4 BF16 MFMA, with one FP32 LDS reduction at the end. It is enabled only for batch-one/token-one replicated projections whose dimensions satisfy the hardware tile; prefill, the FP32 router, and the output-sharded 12,288-wide KDA gate remain unchanged. The official retained path uses it for MLA q-a/kv-a and KDA f-a. Isolated TP8 measurements improved replicated 128/576-output projections by about 16--18%; applying it to the already output-sharded KDA gate was slower and was rejected. Random-shape comparison against the generic graph had maximum/mean absolute BF16 differences of 2.0/0.1114 because the split changes reduction order. Against a serial FP32 accumulation rounded once to BF16, the 7,168-to-1,536 kernel was bit-exact in the tested sample.
+
+The final official context-128 validation loaded in 389.84 seconds with 2.71 GiB peak RSS and zero swap. Two replay trials produced the identical four-token sequence `[198, 59675, 9817, 12519]`; prefill remained 2.406 seconds (38.65 tok/s), while eight-token decode improved to 1.280 seconds (6.25 tok/s, 160.00 ms/token). A one-wave MFMA variant and a full-wave fused decode recurrence were both rejected: the former delivered 6.02 tok/s, and the latter 6.179 tok/s, while both changed the greedy sequence without a useful speed gain.
+
+A final load-first experiment increased the disk-to-HBM io_uring queue depth from one to the 32 existing bounded 2 MiB staging buffers. On a direct 1 GiB read from fragmented shard 28 it measured 6.834 GB/s versus 6.832 GB/s for the original path, so the change was rejected. The subsequent unmodified official 96-shard load completed in 389.48 seconds, confirming both the prior result and the single-NVMe lower bound. Peak RSS was 2.75 GiB with zero swap.
+
+Two direct packed-expert MFMA prototypes were also rejected after that load. A fused gate/up kernel was about 29% faster in isolation at the TP8-local shape, and a routed-down kernel which combined projection, probability weighting, and route reduction measured 1.45 ms versus 2.42 ms in isolation. End-to-end, however, stable replay produced `[198, 2338, 2127, 148297]`, prefill measured 38.87 tok/s, and decode measured 6.263 tok/s. That is indistinguishable from the retained 38.65/6.25 tok/s path while changing floating-point reduction order, so neither kernel was retained.
+
+A whole-core KDA decode experiment fused convolution, Q/K normalization, channel decay, recurrence, RMS normalization, output gating, and four persistent state updates. Its raw kernel replayed in about 109 microseconds per local KDA layer and matched a one-step synthetic reference within `9.77e-4` output and `8.13e-4` state maximum error. The exact-width fake-layer gate caught that it was slower than the retained attention path (0.665 versus 0.633 ms/layer). The already-running official validation was stopped after its first invalid greedy sequence, `[198, 163840, 163840, 163840]`, where 163840 is outside the checkpoint's vocabulary. The kernel was rejected and removed.
+
+| Maximum context | Load | Short-prompt replay | Result |
+|---:|---:|---:|---|
+| 128 | 489.59s | 14.32s | stable 8-token replay |
+| 4,096 | 489.06s | 14.32s | stable replay, zero swap |
+| 32,768 | 532.85s | 14.33s | stable replay, zero swap |
+| 131,072 | 520.91s | 14.37s | stable first token, zero swap |
+| 262,144 | 497.34s | 14.41s | stable first token, zero swap |
+
+These are maximum-context/cache admission tests with the same 93-token prompt, not full-length 32K/131K/262K prefills. The full cache allocation path was exercised, but filling those contexts remains a separate long-running throughput test.
+
+Runtime profiling bracketed four steady decode tokens. It recorded 6,304 kernel events and about 474--478 ms of summed GPU work across the eight devices inside a roughly 1.5-second profiled wall interval. The packed `mxfp4_expert_linear_wave64` kernels accounted for only about 22.5 ms summed; the largest families were small 1,792-wide reductions. This identifies launch/synchronization granularity as the immediate MI350 bottleneck rather than packed-weight bandwidth. `JIT_BATCH_SIZE=64` produced the same original 3.53 tok/s as 32. A gfx950 fused MXFP8 QDQ experiment was bit-exact but slower on the real device (about 95 microseconds versus 57--64 microseconds), so it was rejected. Combining the routed and shared final TP partials removed one collective per routed decode layer and helped raise unprofiled decode to 6.18 tok/s, but the remaining sequential launch boundaries still dominate.
+
+The checkpoint's bundled Transformers code was used as the architectural reference for channel decay and tensor mapping. A full independent Transformers/vLLM token comparison was not run on this host because the required `compressed_tensors`/serving backend is not installed; deterministic tinygrad replay and the numerical KDA, loader-layout, NULL gfx950 compile, and real TP8 smoke tests are the completed correctness gates.
+
+## Known hardware-only gate
+
+The correctness path now consumes packed MXFP4 expert weights directly on gfx950 with a wave64 software-decode kernel, so it does not create selected-expert BF16 weight expansions. MXFP8 activation quantization is still emulated. tinygrad has gfx950/CDNA4 BF16 and FP8 matrix-core support, but this branch does not yet have a hardware-validated native MXFP4×MXFP8 expert GEMM. Expect the first run to be a correctness bring-up, not production throughput. Capture profiles on MI350X before changing the representation: native FP4 work cannot be validated faithfully on the available gfx1100 cards.
+
+Recurrent prefill is fused. The gfx950 wave-parallel kernel was compared directly with the portable graph at the official per-GPU shape through 128 tokens: maximum core/state differences remained below `8e-6`/`1e-6`, outputs were finite, and replay was about 2.7 ms versus about 8 ms for the portable kernel in the isolated test. Full K3 therefore uses 128-token recurrent chunks on gfx950. Chunk size remains part of the numerical configuration because different reduction orders can select different final greedy tokens.
+
+The following serving changes apply to the official K3 path: recurrent-state reset graph capture, direct AMD scalar readback without rebuilding a scheduler graph, materialized gate/up boundaries, separate greedy decode JITs, K3's uncorrected routed probability semantics, gfx950 KDA Q/K/V and exact BF16 partial projections, one combined routed/shared final collective, a gfx950 greedy output-head kernel, the wave64 packed-expert path, and the multi-token four-output packed tile. Software MXFP8 remains in use.
+
+After hardware admission on MI350X, profile before porting those kernels. The likely implementation order is:
+
+1. A native packed MXFP4×MXFP8 grouped expert GEMM using CDNA4 matrix instructions.
+2. A wave64/MFMA KDA Q/K/V decode projection.
+3. Combined routed/shared down-projection TP partials so each layer performs one XGMI all-reduce.
+4. A CDNA4 output-head matvec and router matvec if they remain visible in the profile.
+
+Every port needs a direct numerical comparison with the generic graph and an end-to-end greedy-token comparison before performance measurements. The wave64 packed-expert kernel has compile coverage through `NULL:HIP:gfx950`; numerical and performance validation still require real MI350X hardware. None of the remaining gfx11-only kernels should be enabled on gfx950 by changing only the architecture guard.
+
+## MI350X performance expectation
+
+Treat the first rental as bring-up, not a guaranteed throughput run. The loader reads every official expert tensor once into a transient GPU-0 staging buffer (at most one packed projection), then redistributes TP8 slices over the GPU fabric; it does not generate files or require checkpoint-sized host RAM. A reasonable planning range for the full text model on eight MI350X cards is 3–8 minutes to stream and TP-shard the 1.56 TB checkpoint, 150–400 tok/s for initial short/medium prefill, and 25–60 tok/s decode with the software packed-expert path. After a native CDNA4 MXFP4×MXFP8 grouped expert kernel, wave64/MFMA recurrent projections, and XGMI collective tuning, 500+ tok/s prefill and roughly 80–150 tok/s decode are plausible targets. These ranges are engineering estimates, not measurements.
+
+The nominal HBM bandwidth is not the main uncertainty: eight MI350X devices have enough aggregate bandwidth for K3's active weights. Utilization is limited by 93 sequential layers, small routed projections, and synchronization after TP input-sharded projections. Record actual HBM and XGMI counters before deciding whether the next port should target matrix instructions or collective count.
+
+The official checkpoint also contains MoonViT-V2 and multimodal projector weights. They are skipped by the text loader. Image input remains a separate implementation and validation task.
+
+## Local TP4 performance baseline
+
+The pre-rental benchmark uses the converted `Kimi-Linear-48B-A3B-Instruct-MXFP4-v2` checkpoint on four gfx1100 GPUs. It is a useful regression test for the KDA/MLA/MoE text path, not a projection of K3 throughput on MI350X.
+
+```sh
+DEV=AMD JIT_BATCH_SIZE=64 python extra/benchmark_kimi.py \
+  /raid/models/Kimi-Linear-48B-A3B-Instruct-MXFP4-v2 \
+  --devices 4 --max-context 128 --prompt-tokens 32 --decode-tokens 32 --chunk-size 32
+```
+
+Results from 2026-08-10:
+
+- load from RAID: 44.28s for the 29.27 GB checkpoint
+- first 32-token prefill includes roughly 10s of compilation/capture
+- steady fresh-prompt prefill replay: 0.118s, 270.20 tok/s
+- steady context-32 decode replay: 101.82 tok/s, 9.82 ms/token
+- peak host RSS: 729.9 MiB; swap was not used
+
+The load, prefill, and decode targets are all met in the bounded prompt-32 run. Decode improved from 23.03 tok/s to 101.82 tok/s. The retained greedy output was checked across 32 decode steps; rejected half-wave and unrounded recurrent reductions were faster but diverged and eventually collapsed to a repeated token.
+
+Fully warmed HTTP serving was also measured with `--max_context 4096`. Startup, including weight load, capture, and replay of both serving shapes, took 113.73s. After a two-turn cache test, the first aligned 64-token request reported 271 tok/s prefill and 101 tok/s decode over 64 generated tokens. A 99-token prompt reported 254 tok/s prefill and 99 tok/s decode; decode falls slightly as MLA context grows.
+
+Recurrent serving uses only the captured 32-token prefill graph and captured single-token graph. Warmup uses two consecutive chunks so both initial and nonzero-position prefill execution are ready before the socket opens. A prompt tail shorter than 32 tokens runs through the single-token graph instead of compiling a new static shape, so no request-time JIT capture is required. Exact extensions reuse recurrent and KV state—the live second turn logged `in: 18 + 15`—while divergent prompts reset both safely. Very short prompts can report less than 200 aggregate prefill tok/s because fixed reset and single-token costs dominate; aligned and medium/long prompts exercise the 200+ tok/s prefill path.
+
+Four 7900 XTX cards provide 96 GB aggregate VRAM and about 3.84 TB/s aggregate physical memory bandwidth. Their nominal aggregate vector FP16 rate is about 245.6 TFLOP/s, or about 492 TFLOP/s through matrix instructions. Kimi Linear activates roughly 3.107B parameters per token; a simple active-weight accounting gives approximately 4.05 GB/token and an optimistic bandwidth-only ceiling near 948 tok/s. The measured decode rate is much lower because this MoE decode workload is a collection of small matrix-vector operations plus PCIe collectives, not one ideal streaming kernel.
+
+The generic loader currently rereads logical TP shards and accounts for roughly 227 GB of disk traffic for a TP4 load. RAID bandwidth hides that inefficiency locally, but a direct one-pass shard loader remains worthwhile before slow remote storage is used. It was not retained here because the attempted direct-shard graph exposed an unresolved scheduler/renderer edge; correctness and bounded memory take priority over avoiding the redundant reads.
+
+Different chunk sizes can choose a different final token because their matrix kernels use different floating-point reduction orders. Each measured shape was repeatable between cold and captured execution. For official K3 validation, compare logits/tokens against the reference at one fixed chunk size and greedy settings rather than requiring bitwise agreement between performance shapes.
@@ -9,8 +9,7 @@ from extra.lr_scheduler import OneCycleLR
 GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]

 # override tinygrad defaults
-dtypes.default_float = dtypes.half
-Context(FUSE_OPTIM=1).__enter__()
+Context(DEFAULT_FLOAT=dtypes.half, FUSE_OPTIM=1).__enter__()

 # from https://github.com/tysam-code/hlb-CIFAR10/blob/main/main.py
 batchsize = getenv("BS", 1024)
@@ -0,0 +1,9 @@
+import argparse
+from tinygrad.llm.kimi import convert_kimi
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(description="Convert official Kimi-Linear-48B-A3B BF16 weights to tinygrad MXFP4/BF16")
+  parser.add_argument("source", help="downloaded moonshotai/Kimi-Linear-48B-A3B-Instruct directory")
+  parser.add_argument("output", help="output directory")
+  args = parser.parse_args()
+  convert_kimi(args.source, args.output)
@@ -22,10 +22,6 @@ class Attention:
    self.head_dim = dim // n_heads

  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]) -> Tensor:
-    if mask is not None or start_pos.val == 0:
-      # no symbolic shape qkv when consuming prompts
-      start_pos = start_pos.val
-
    if HALF: x = x.half()
    xqkv = self.c_attn(x).reshape(None, None, 3, self.n_heads, self.head_dim)
    xq, xk, xv = [xqkv[:, :, i, :, :] for i in range(3)]
@@ -38,12 +34,8 @@ class Attention:
    # update the cache
    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()

-    if start_pos > 0:
-      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
-      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
-    else:
-      keys = xk
-      values = xv
+    keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
+    values = self.cache_kv[1][:, :start_pos+seqlen, :, :]

    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)
    return self.c_proj(xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2).reshape(bsz, seqlen, self.dim))
@@ -86,15 +78,14 @@ class Transformer:
      seqlen = tokens.shape[1]
      tok_emb = self.wte(tokens)

-    # not symbolic when consuming the prompt
-    selected_pos = (0, seqlen) if start_pos.val == 0 else (start_pos, start_pos+1)
-    pos_emb = self.wpe(self.allpos.shrink((None, selected_pos)))
+    # start_pos is a bound Variable, so everything below it stays symbolic
+    pos_emb = self.wpe(self.allpos.shrink((None, (start_pos, start_pos+seqlen))))

    h = tok_emb + pos_emb

    if HALF: h = h.half()

-    mask = Tensor.full((1, 1, seqlen, start_pos.val+seqlen), float("-inf"), dtype=h.dtype).triu(start_pos.val+1) if seqlen > 1 else None
+    mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype).triu(start_pos+1) if seqlen > 1 else None

    for hi in self.h: h = hi(h, start_pos, mask)

@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""Cheap preflight for an official moonshotai/Kimi-K3 checkout. Does not load model weights."""
+import argparse, json, pathlib, shutil
+from tinygrad.llm.kimi_k3 import KIMI_K3_TP8_BYTES_PER_GPU, audit_kimi_k3_checkpoint
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("model_dir", type=pathlib.Path)
+  parser.add_argument("--metadata-only", action="store_true", help="permit absent weight shards")
+  parser.add_argument("--context", type=int, default=4096, help="context length used for the memory estimate")
+  args = parser.parse_args()
+  stats = audit_kimi_k3_checkpoint(args.model_dir, require_shards=not args.metadata_only)
+  if not 1 <= args.context <= 1_048_576: raise ValueError("--context must be between 1 and 1048576")
+
+  # K3 has 24 MLA layers. Each token stores the 512-value compressed latent plus 64 RoPE values in BF16.
+  per_gpu_weights = KIMI_K3_TP8_BYTES_PER_GPU
+  mla_cache = 24 * args.context * (512 + 64) * 2
+  hbm = 288_000_000_000
+  print(json.dumps(stats, indent=2))
+  print(f"exact text weights/GPU under this TP8 layout: {per_gpu_weights/1e9:.2f} GB ({per_gpu_weights/2**30:.2f} GiB)")
+  print(f"replicated MLA cache/GPU at {args.context:,} tokens: {mla_cache/1e9:.2f} GB ({mla_cache/2**30:.2f} GiB)")
+  print(f"nominal MI350X headroom before runtime buffers: {(hbm-per_gpu_weights-mla_cache)/1e9:.2f} GB")
+  if not args.metadata_only:
+    usage = shutil.disk_usage(args.model_dir)
+    print(f"filesystem free space: {usage.free/1e9:.2f} GB")
+
+if __name__ == "__main__": main()
@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""Run a reduced, architecture-complete K3 prefill/decode on tensor-parallel devices."""
+import argparse, time
+from tinygrad import Tensor, Device, dtypes, nn
+from tinygrad.llm.kimi_k3 import _shard_kimi_k3, kimi_k3_smoke_config
+from tinygrad.llm.model import Transformer
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--devices", type=int, default=8)
+  args = parser.parse_args()
+  if args.devices not in (1, 2, 4, 8): raise ValueError("the K3 admission smoke test supports 1, 2, 4, or 8 devices")
+  devices = tuple(f"AMD:{i}" for i in range(args.devices))
+  model = Transformer(kimi_k3_smoke_config())
+  for name,value in nn.state.get_state_dict(model).items():
+    fill = 127 if name.endswith("weight_scale") else 0
+    dtype = value.dtype if value.dtype is dtypes.uint8 else dtypes.bfloat16
+    value.replace(Tensor.full(value.shape, fill, dtype=dtype, device="CPU"))
+  _shard_kimi_k3(model, devices)
+  temperature = Tensor([0.0], device=devices)
+  for label,tokens,start in (("prefill", [[1, 2]], 0), ("decode", [[3]], 2), ("decode replay", [[4]], 3)):
+    begin = time.perf_counter()
+    out = model(Tensor(tokens, dtype=dtypes.int32, device=devices), start, temperature).realize()
+    for device in devices: Device[device].synchronize()
+    print(f"{label}: shape={out.shape}, {time.perf_counter()-begin:.3f}s")
+
+if __name__ == "__main__": main()
@@ -1,11 +1,11 @@
 import os, random, pickle, queue, struct, math, functools, hashlib, time
 from typing import List
 from pathlib import Path
-from multiprocessing import Queue, Process, shared_memory, connection, Lock, cpu_count
+from multiprocessing import Queue, Process, shared_memory, connection, Lock

 import numpy as np
 from tinygrad import dtypes, Tensor
-from tinygrad.helpers import getenv, prod, Context, round_up, tqdm, OSX
+from tinygrad.helpers import getenv, prod, Context, round_up, tqdm, OSX, NUM_CPU_THREADS
 from tinygrad.nn.state import TensorIO

 ### ResNet
@@ -131,7 +131,7 @@ def batch_load_resnet(batch_size=64, val=False, shuffle=True, seed=None, pad_fir
    else: X = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/{shm_name}")
    Y = [None] * (batch_size*BATCH_COUNT)

-    for _ in range(cpu_count()):
+    for _ in range(NUM_CPU_THREADS.value):
      p = Process(target=loader_process, args=(q_in, q_out, X, seed))
      p.daemon = True
      p.start()
@@ -212,7 +212,7 @@ def batch_load_train_bert(BS:int, seed:int|None=None):
    rng.shuffle(fs)
    train_files.append(fs.pop(0))

-  cycle_length = min(getenv("NUM_CPU_THREADS", min(os.cpu_count(), 8)), len(train_files))
+  cycle_length = min(NUM_CPU_THREADS.value, len(train_files))
  assert cycle_length > 0, "cycle_length must be greater than 0"

  dataset = InterleavedDataset(train_files, cycle_length)
@@ -301,7 +301,7 @@ def batch_load_unet3d(preprocessed_dataset_dir:Path, batch_size:int=6, val:bool=
    X = Tensor.empty(*sz, dtype=dtypes.float32, device=f"disk:/dev/shm/{shm_name_x}")
    Y = Tensor.empty(*sz, dtype=dtypes.uint8, device=f"disk:/dev/shm/{shm_name_y}")

-    for _ in range(cpu_count()):
+    for _ in range(NUM_CPU_THREADS.value):
      proc = Process(target=load_unet3d_data, args=(preprocessed_dataset_dir, seed, queue_in, queue_out, X, Y))
      proc.daemon = True
      proc.start()
@@ -437,7 +437,7 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
  dataset_iter = iter(image_ids)

  try:
-    for _ in range(cpu_count()):
+    for _ in range(NUM_CPU_THREADS.value):
      proc = Process(
        target=load_retinanet_data,
        args=(base_dir, val, queue_in, queue_out, imgs, boxes, labels),
@@ -1282,10 +1282,10 @@ def train_bert():
        previous_step = i

 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8
+  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8, MXFP4
  from examples.llama3 import MODEL_PARAMS
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
-  from examples.mlperf.optim import GradAccClipAdamW
+  from examples.mlperf.optim import GradAccClipAdamW, clip_grads

  INITMLPERF = getenv("INITMLPERF")
  RUNMLPERF = getenv("RUNMLPERF")
@@ -1434,9 +1434,9 @@ def train_llama3():
    load_state_dict(scheduler, safe_load(fn), realize=False)

  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
-  fp8_next_amax = [t for ts in model._fp8_next_amax.values() for t in ts] if hasattr(model, "_fp8_next_amax") else []
-  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts] if hasattr(model, "_fp8_grad_amax") else []
-  fp8_next_grad_amax = [t for ts in model._fp8_next_grad_amax.values() for t in ts] if hasattr(model, "_fp8_next_grad_amax") else []
+  fp8_next_amax = [t for ts in model._fp8_next_amax.values() for t in ts]
+  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts]
+  fp8_next_grad_amax = [t for ts in model._fp8_next_grad_amax.values() for t in ts]
  fp8_inv_scales = list(model._fp8_inv_scale.values()) + list(model._fp8_next_inv_scale.values())

  from tinygrad.nn.state import get_state_dict
@@ -1458,12 +1458,12 @@ 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):
-    for nxt in fp8_next_amax: nxt.assign(0)
-    for nxt in fp8_next_grad_amax: nxt.assign(0)
+    model.reset_amax()
    if is_dp: tokens = tokens.to(None).shard(device, 0)
    if is_mp: tokens = tokens.shard(device)
    if not is_sharding: tokens = tokens.to(None)
@@ -1477,23 +1477,24 @@ 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():
-    grad_norm = optim.fstep(grads)
+    grad_norm = clip_grads(grads, grad_acc, 1.0)
+    optim.fstep(grads, grad_norm)
    scheduler.step()

    for g in grads: g.assign(0)
-    for cur, nxt in zip(fp8_amax, fp8_next_amax): cur.assign(nxt)
-    for cur, nxt in zip(fp8_grad_amax, fp8_next_grad_amax): cur.assign(nxt)
+    model.update_amax()

    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)
@@ -1548,8 +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:
@@ -1557,16 +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
@@ -1578,7 +1578,7 @@ def train_llama3():

      mem_gb = GlobalCounters.mem_used / 1e9
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 100
+      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * (9.2e15 if MXFP4 else 4.6e15))) * 100
      tqdm.write(
          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
@@ -1667,7 +1667,7 @@ def train_llama3():
 def train_gptoss():
  from examples.mlperf.models.gpt_oss import GPTOSS, GPT_OSS_20B, apply_grad, FP8_DTYPE
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
-  from examples.mlperf.optim import GradAccClipAdamW
+  from examples.mlperf.optim import GradAccClipAdamW, GradAccClipAdamWGroup, clip_grads

  BENCHMARK = getenv("BENCHMARK")

@@ -1710,9 +1710,10 @@ def train_gptoss():
    wandb.init(config=config, **wandb_args, project="MLPerf-gpt-oss")

  model_params = GPT_OSS_20B
-  model_params['vocab_size'] = 128256
+  model_params['vocab_size'] = getenv("VOCAB_SIZE", 128256)
  real_vocab_size = model_params['vocab_size']
  if (layers:=getenv("LAYERS")) != 0: model_params['n_layers'] = layers
+  if (experts:=getenv("EXPERTS")) != 0: model_params['n_experts'] = experts
  print(f"model parameters: {model_params}")

  model = GPTOSS(**model_params, max_context=SEQLEN)
@@ -1733,7 +1734,12 @@ def train_gptoss():
  is_offload_optim = bool(getenv("OFFLOAD_OPTIM"))
  is_fake_offload = Device.DEFAULT == "NULL"
  optim_device = ("CPU" if not is_fake_offload else "NULL:99") if is_offload_optim else None
-  optim = GradAccClipAdamW(params, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device)
+  params_wd = [p for p in params if p.ndim >= 3]
+  params_no_wd = [p for p in params if p.ndim < 3]
+  optim = GradAccClipAdamWGroup(
+    GradAccClipAdamW(params_wd, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device),
+    GradAccClipAdamW(params_no_wd, lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=0.0, grad_acc=grad_acc, device=optim_device),
+  )

  for p in optim.params:
    grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
@@ -1742,7 +1748,10 @@ def train_gptoss():

  from extra.gemm.cdna_asm_gemm import _mx_block_scale
  model_state = get_state_dict(model)
-  fp8_scale_names = {n: f"{n}_scale" for n, t in model_state.items() if t.dtype == FP8_DTYPE}
+  def _scale_key(n):
+    if "." in n and (c:=f"{(b:=n.rsplit('.',1))[0]}_scale.{b[1]}") in model_state: return c
+    return f"{n}_scale"
+  fp8_scale_names = {n: _scale_key(n) for n, t in model_state.items() if t.dtype == FP8_DTYPE}
  fp8_inv_scales = [model_state[sname] for sname in fp8_scale_names.values()]
  for wname, sname in fp8_scale_names.items():
    w, scale = model_state[wname], model_state[sname]
@@ -1775,7 +1784,8 @@ def train_gptoss():

  @TinyJit
  def optim_step():
-    grad_norm = optim.fstep(grads)
+    grad_norm = clip_grads(grads, grad_acc, 1.0)
+    optim.fstep(grads, grad_norm)
    scheduler.step()

    for g in grads: g.assign(0)
@@ -25,6 +25,7 @@ FUSED_SILU_W13 = getenv("FUSED_SILU_W13", 0)
 SPLIT_W13 = getenv("SPLIT_W13", 0)
 COLUMNWISE_WEIGHT_SCALE = getenv("COLUMNWISE_WEIGHT_SCALE", 0)
 MXFP8 = getenv("MXFP8", 0)
+MXFP4 = getenv("MXFP4", 0)

 FP8_DTYPE = dtypes.fp8e4m3
 FP8_GRAD_DTYPE = dtypes.fp8e5m2
@@ -44,6 +45,11 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
      from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
      if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T),)
    return (x @ w.T,)
+  if MXFP4:
+    assert x is not None, "MXFP4 matmul requires an unquantized input"
+    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm
+    if can_use_asm_gemm(x, w.T): return (asm_gemm(x, w.T, mxfp4=True),)
+    return (x @ w.T,)
  assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
  if MXFP8:
    from extra.gemm.cdna_asm_gemm import asm_gemm, quantize_mxfp8, mx_pack, can_use_asm_gemm, _mx_block_scale
@@ -77,9 +83,9 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
      return out, x_fp8
  return (x_fp8.dot(w.T, dtype=dtypes.float) * ((amax_x.float() + 1e-8) / FP8_MAX) * w_inv_scale).cast(dtypes.bfloat16), x_fp8

-def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor,
-                         next_amax_x:Tensor, grad_amax_state:Tensor, next_grad_amax_state:Tensor):
-  if FUSED_ADD_NORM_MUL_QUANTIZE:
+def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
+                         next_amax_x:Tensor|None, grad_amax_state:Tensor|None, next_grad_amax_state:Tensor|None):
+  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_rmsnorm_mul_quantize_fp8
    x_fp8, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x,
@@ -90,9 +96,9 @@ def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, ep
                     next_grad_amax_state=next_grad_amax_state, next_amax_x=next_amax_x)
  return out, x_normed, rrms, ret

-def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor,
-                             next_amax_x:Tensor, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
-  if FUSED_ADD_NORM_MUL_QUANTIZE:
+def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
+                             next_amax_x:Tensor|None, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
+  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
    x_fp8, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
    out, *ret = matmul(None, w, w_inv_scale=w_inv_scale, x_fp8=x_fp8, amax_x=amax_x,
@@ -105,10 +111,15 @@ def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w
  return out, h, x_normed, rrms, ret

 def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
-                             amax_x2:Tensor, next_amax_x2:Tensor,
-                             grad_amax_xw13:Tensor, next_grad_amax_xw13:Tensor,
-                             grad_amax_xout:Tensor, next_grad_amax_xout:Tensor):
-  if FUSED_SILU_W13:
+                             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,
                                                 next_grad_amax_state=next_grad_amax_xw13, amax_out=next_amax_x2)
@@ -158,14 +169,15 @@ class FlatTransformer:
    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).clone().is_param_(False)

    def _amax(): return Tensor.full((), FP8_MAX, dtype=dtypes.float32).contiguous().is_param_(False)
+    n_amax = 0 if MXFP4 else n_layers
    names = ["xqkv", "xo", "x2"]
    names += ["x1", "x3"] if SPLIT_W13 else ["x13"]
-    self._fp8_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
-    self._fp8_next_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
+    self._fp8_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
+    self._fp8_next_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
    grad_names = ["xqkv", "xo", "xout"]
    grad_names += ["xw1", "xw3"] if SPLIT_W13 else ["xw13"]
-    self._fp8_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
-    self._fp8_next_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
+    self._fp8_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
+    self._fp8_next_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
    w_scales = [("wqkv", s_qkv), ("wo", s_o), ("w2", s_2)]
    w_scales += [("w1", s_1), ("w3", s_3)] if SPLIT_W13 else [("w13", s_13)]
    self._fp8_inv_scale = {name: (s if MXFP8 else s.float()).contiguous().is_param_(False) for name, s in w_scales}
@@ -179,6 +191,9 @@ class FlatTransformer:
      from extra.gemm.cdna_asm_gemm import quantize_mxfp8
      w_q, w_e8, _ = quantize_mxfp8(w.reshape(self.n_layers * out_features, in_features))
      return w_q.reshape(self.n_layers, out_features, in_features), w_e8.reshape(self.n_layers, out_features, in_features // 32)
+    if MXFP4:
+      # FP4 is produced dynamically so optimizer updates always start from the current BF16 weight.
+      return w.cast(dtypes.bfloat16), Tensor.ones(self.n_layers)
    amax = (w.abs().max(axis=2) if COLUMNWISE_WEIGHT_SCALE else w.abs().flatten(1).max(1)).detach()
    scale = FP8_MAX / (amax + 1e-8)
    inv_scale = (amax + 1e-8) / FP8_MAX
@@ -186,9 +201,10 @@ class FlatTransformer:
    return (w * scale_b).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE), inv_scale

  def attention(self, x:Tensor, freqs_cis:Tensor, *, attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                amax_xqkv:Tensor, amax_xo:Tensor, s_qkv:Tensor, s_o:Tensor,
-                next_amax_xqkv:Tensor, next_amax_xo:Tensor,
-                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor, next_grad_amax_xqkv:Tensor, next_grad_amax_xo:Tensor):
+                amax_xqkv:Tensor|None, amax_xo:Tensor|None, s_qkv:Tensor, s_o:Tensor,
+                next_amax_xqkv:Tensor|None, next_amax_xo:Tensor|None,
+                grad_amax_xqkv:Tensor|None, grad_amax_xo:Tensor|None,
+                next_grad_amax_xqkv:Tensor|None, next_grad_amax_xo:Tensor|None):
    bsz, seqlen, _ = x.shape
    saves = []

@@ -310,28 +326,33 @@ class FlatTransformer:
          for i in range(len(amax_dict[name])):
            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().is_param_(False)

+  def reset_amax(self):
+    for st in (self._fp8_next_amax, self._fp8_next_grad_amax):
+      for ts in st.values():
+        for t in ts: t.assign(0)
+
+  def update_amax(self):
+    for cur, nxt in ((self._fp8_amax, self._fp8_next_amax), (self._fp8_grad_amax, self._fp8_next_grad_amax)):
+      for name in cur:
+        for c, n in zip(cur[name], nxt[name]): c.assign(n)
+
  def __call__(self, tokens:Tensor, save:bool=True):
    h = self.tok_embeddings(tokens)
    freqs_cis = self.freqs_cis.cast(h.dtype)
    if not getenv("HK_FLASH_ATTENTION"): freqs_cis = freqs_cis[:, :tokens.shape[1], :, :, :]
    a, na, ga, nga, s = self._fp8_amax, self._fp8_next_amax, self._fp8_grad_amax, self._fp8_next_grad_amax, self._fp8_inv_scale
+    def amax_kwargs(i:int, act_names:tuple[str, ...], grad_names:tuple[str, ...]) -> dict[str, Tensor|None]:
+      specs = (("amax_", a, act_names), ("next_amax_", na, act_names), ("grad_amax_", ga, grad_names), ("next_grad_amax_", nga, grad_names))
+      if MXFP4: return dict.fromkeys(f"{prefix}{name}" for prefix, _, names in specs for name in names)
+      return {f"{prefix}{name}":val[name][i] for prefix, val, names in specs for name in names}
    for i in range(self.n_layers):
-      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wo=self.wo[i],
-                         amax_xqkv=a["xqkv"][i], amax_xo=a["xo"][i], s_qkv=s["wqkv"][i], s_o=s["wo"][i],
-                         next_amax_xqkv=na["xqkv"][i], next_amax_xo=na["xo"][i],
-                         grad_amax_xqkv=ga["xqkv"][i], grad_amax_xo=ga["xo"][i],
-                         next_grad_amax_xqkv=nga["xqkv"][i], next_grad_amax_xo=nga["xo"][i])
-      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], w2=self.w2[i],
-                        amax_x2=a["x2"][i], s_2=s["w2"][i], grad_amax_xout=ga["xout"][i], next_grad_amax_xout=nga["xout"][i],
-                        next_amax_x2=na["x2"][i])
+      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wo=self.wo[i], s_qkv=s["wqkv"][i], s_o=s["wo"][i],
+                         **amax_kwargs(i, ("xqkv", "xo"), ("xqkv", "xo")))
+      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], w2=self.w2[i], s_2=s["w2"][i], **amax_kwargs(i, ("x2",), ("xout",)))
      if SPLIT_W13:
-        ffn_kwargs.update(w1=self.w1[i], w3=self.w3[i], amax_x1=a["x1"][i], amax_x3=a["x3"][i],
-                          next_amax_x1=na["x1"][i], next_amax_x3=na["x3"][i],
-                          s_1=s["w1"][i], s_3=s["w3"][i], grad_amax_xw1=ga["xw1"][i], grad_amax_xw3=ga["xw3"][i],
-                          next_grad_amax_xw1=nga["xw1"][i], next_grad_amax_xw3=nga["xw3"][i])
+        ffn_kwargs.update(w1=self.w1[i], w3=self.w3[i], s_1=s["w1"][i], s_3=s["w3"][i], **amax_kwargs(i, ("x1", "x3"), ("xw1", "xw3")))
      else:
-        ffn_kwargs.update(w13=self.w13[i], amax_x13=a["x13"][i], s_13=s["w13"][i], grad_amax_xw13=ga["xw13"][i],
-                          next_grad_amax_xw13=nga["xw13"][i], next_amax_x13=na["x13"][i])
+        ffn_kwargs.update(w13=self.w13[i], s_13=s["w13"][i], **amax_kwargs(i, ("x13",), ("xw13",)))
      h, *_ = self.run_layer(h, freqs_cis, attn_kwargs, ffn_kwargs, save=save)

    logits = matmul(self.norm(h), self.output[0], fp8=False)[0]
@@ -415,9 +436,7 @@ if __name__ == "__main__":
  @TinyJit
  def fwd_bwd(tokens:Tensor):
    with Timing("python forward: "):
-      for amax_dict in (model._fp8_next_amax, model._fp8_next_grad_amax):
-        for ts in amax_dict.values():
-          for nxt in ts: nxt.assign(0)
+      model.reset_amax()
      logits = model(tokens[:, :-1], save=llama_size=="8B")
      loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
    with Timing("python backward: "):
@@ -13,10 +13,14 @@ 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.moe_gemm import grouped_mx_gemm
+from extra.gemm.moe_routing import route, dispatch, combine

 FP8_DTYPE = dtypes.fp8e4m3
 FP8_MAX = 448.0
-INIT_STD = 0.008
+INIT_STD = 0.02
+ASM_GEMM = getenv("ASM_GEMM", 0)
+

 def _quant_dequant_fwd(x:Tensor) -> Tensor:
  # x (2d bf16) -> bf16 value after an mxfp8 round-trip (1x32 block scaling on the last axis)
@@ -50,8 +54,7 @@ def _dequant_fwd_fxn(wq_p, ws_p, device):
  return _dequant_fwd(Tensor(wq_p, device=device), Tensor(ws_p, device=device))

 def _dequant_bwd(grad:UOp, call:UOp) -> tuple:
-  w_scale = Tensor(call.src[2])
-  return ((Tensor(grad).cast(dtypes.bfloat16) * _mx_scale(w_scale).cast(dtypes.bfloat16)).uop, None)
+  return (Tensor(grad).cast(dtypes.bfloat16).uop, None)

 def dequant_weight(w_q:Tensor, w_scale:Tensor) -> Tensor:
  fxn = _dequant_fwd_fxn(w_q.as_param(0).uop, w_scale.as_param(1).uop, w_q.device)
@@ -60,10 +63,34 @@ def dequant_weight(w_q:Tensor, w_scale:Tensor) -> Tensor:

 def matmul_mx(x:Tensor, w_q:Tensor, w_scale:Tensor) -> Tensor:
  l_shape = x.shape[:-1]
+  if ASM_GEMM:
+    from extra.gemm.cdna_asm_gemm import asm_gemm, can_use_asm_gemm, mx_pack
+    x2, K, N = x.reshape(-1, x.shape[-1]), x.shape[-1], w_q.shape[0]
+    wq, ws = w_q, w_scale
+    if (pad := (-K) % 256):
+      x2 = x2.pad(((0, 0), (0, pad)))
+      wq = wq.pad(((0, 0), (0, pad)))
+      ws = ws.pad(((0, 0), (0, pad // 32)), value=127).cast(dtypes.uint8)
+    if (npad := (-N) % 256):
+      wq = wq.pad(((0, npad), (0, 0)))
+      ws = ws.pad(((0, npad), (0, 0)), value=127).cast(dtypes.uint8)
+    x_q, x_e8, x_si = quantize_mxfp8(x2)
+    if x_si is not None and can_use_asm_gemm(x_q, wq.T):
+      out = asm_gemm(x_q, wq.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(ws), ws), mx_w_stored=True)
+      return (out[:, :N] if npad else out).reshape(*l_shape, N).cast(dtypes.bfloat16)
  x_phys = quant_dequant_mx(x.reshape(-1, x.shape[-1])).reshape(*l_shape, x.shape[-1])
  w_phys = dequant_weight(w_q, w_scale)
  return (x_phys @ w_phys.T).cast(dtypes.bfloat16)

+def _pad_to_mult(t:Tensor, axis:int, mult:int=256) -> Tensor:
+  if (r := (-t.shape[axis]) % mult) == 0: return t
+  pads = [(0, 0)] * t.ndim
+  pads[axis] = (0, r)
+  return t.pad(tuple(pads))
+
+def _pad_cols(t:Tensor) -> Tensor: return _pad_to_mult(t, -1)
+def _pad_rows(t:Tensor) -> Tensor: return _pad_to_mult(t, -2)
+
 def swiglu(x:Tensor, limit:float=7.0, alpha:float=1.702) -> Tensor:
  x_glu, x_linear = x[..., ::2], x[..., 1::2]
  x_glu = x_glu.clamp(max_=limit)
@@ -100,9 +127,9 @@ class GPTOSS:
    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
    self.gate = Tensor.normal(n_layers, n_experts, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
    self.gate_bias = Tensor.zeros(n_layers, n_experts, dtype=dtypes.bfloat16).contiguous()
-    self.w_gate_up, self.w_gate_up_scale = self._quant_weight(n_layers, n_experts, intermediate_size * 2, dim)
+    self.w_gate_up, self.w_gate_up_scale = self._quant_weight(n_layers, n_experts, intermediate_size * 2, dim, moe=True)
    self.w_gate_up_bias = Tensor.zeros(n_layers, n_experts, intermediate_size * 2, dtype=dtypes.bfloat16).contiguous()
-    self.w_down, self.w_down_scale = self._quant_weight(n_layers, n_experts, dim, intermediate_size, std=scaled_std)
+    self.w_down, self.w_down_scale = self._quant_weight(n_layers, n_experts, dim, intermediate_size, std=scaled_std, moe=True)
    self.w_down_bias = Tensor.zeros(n_layers, n_experts, dim, dtype=dtypes.bfloat16).contiguous()

    # output
@@ -112,10 +139,15 @@ class GPTOSS:
    self.output = Tensor.normal(vocab_size, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
    self.freqs_cis = precompute_freqs_cis(head_dim, max_context * 2, rope_theta).contiguous().is_param_(False)

-  def _quant_weight(self, *shape:int, std:float=INIT_STD):
-    w = Tensor.zeros(*shape) if getenv("ZEROS") else Tensor.normal(*shape, mean=0.0, std=std)
-    w_q, w_e8, _ = quantize_mxfp8(w)
-    return w_q, w_e8.is_param_(False)
+  def _quant_weight(self, *shape:int, std:float=INIT_STD, moe:bool=False):
+    def _one(*s:int):
+      w = Tensor.zeros(*s) if getenv("ZEROS") else Tensor.normal(*s, mean=0.0, std=std)
+      w_q, w_e8, _ = quantize_mxfp8(_pad_cols(_pad_rows(w)) if moe else w)
+      return w_q, w_e8.is_param_(False)
+    if moe:
+      qs = [_one(*shape[1:]) for _ in range(shape[0])]
+      return [q[0] for q in qs], [q[1] for q in qs]
+    return _one(*shape)

  def _attn_mask(self, seqlen:int, dtype) -> Tensor:
    i, j = Tensor.arange(seqlen).reshape(seqlen, 1), Tensor.arange(seqlen).reshape(1, seqlen)
@@ -174,17 +206,32 @@ class GPTOSS:
                   w_down:Tensor, w_down_scale:Tensor, w_down_bias:Tensor):
    x_normed, rrms = rmsnorm(x, self.norm_eps)
    inp = x_normed * ffn_norm
-
    logits = inp.float() @ gate.float().T + gate_bias.float()
-    thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
-    weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
+    dim, inter = self.dim, self.intermediate_size

-    out = None
-    for e in range(self.n_experts):
-      gate_up = matmul_mx(inp, w_gate_up[e], w_gate_up_scale[e]) + w_gate_up_bias[e]
-      y = (matmul_mx(swiglu(gate_up, self.swiglu_limit), w_down[e], w_down_scale[e]) + w_down_bias[e]).contiguous()
-      contrib = weights[..., e:e+1].cast(y.dtype) * y
-      out = contrib if out is None else out + contrib
+    if getenv("GROUPED_MOE", 0):
+      bsz, seqlen = x.shape[:2]
+      inp, logits = inp.reshape(-1, dim), logits.reshape(-1, self.n_experts)
+      r = route(logits, self.experts_per_tok, self.n_experts)
+      onehot = r.rows_e.one_hot(self.n_experts).float()
+      xg = dispatch(_pad_cols(inp.cast(dtypes.bfloat16)), r)
+      h = grouped_mx_gemm(xg, (w_gate_up, w_gate_up_scale), r.off)[:, :2*inter] + (onehot @ w_gate_up_bias.float()).cast(dtypes.bfloat16)
+      y = swiglu(h, self.swiglu_limit)
+      z = grouped_mx_gemm(_pad_cols(y.cast(dtypes.bfloat16)), (w_down, w_down_scale), r.off)[:, :dim] \
+          + (onehot @ w_down_bias.float()).cast(dtypes.bfloat16)
+      out = combine(z, r, inp.shape[0], self.experts_per_tok).reshape(bsz, seqlen, dim)
+    else:
+      thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
+      weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
+
+      out = None
+      for e in range(self.n_experts):
+        gu_q, gu_s = w_gate_up[e][:2*inter, :dim].contiguous(), w_gate_up_scale[e][:2*inter, :dim//32].contiguous()
+        dn_q, dn_s = w_down[e][:dim, :inter].contiguous(), w_down_scale[e][:dim, :inter//32].contiguous()
+        gate_up = matmul_mx(inp, gu_q, gu_s) + w_gate_up_bias[e]
+        y = (matmul_mx(swiglu(gate_up, self.swiglu_limit), dn_q, dn_s) + w_down_bias[e]).contiguous()
+        contrib = weights[..., e:e+1].cast(y.dtype) * y
+        out = contrib if out is None else out + contrib
    return out, [x_normed, rrms]

  @function(precompile=True, precompile_backward=True)
@@ -1,8 +1,8 @@
 from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes
-from tinygrad.nn.optim import Optimizer
+from tinygrad.nn.optim import Optimizer, OptimizerGroup
 from tinygrad.helpers import FUSE_OPTIM, getenv
-from tinygrad.uop.ops import UOp, Ops
+from tinygrad.uop.ops import UOp, Ops, AxisType

 STOCHASTIC_ROUND = getenv("STOCHASTIC_ROUND", 0)
 MASTER_WEIGHTS = getenv("MASTER_WEIGHTS", 0)
@@ -21,6 +21,12 @@ def stochastic_round_bf16(x:Tensor) -> Tensor:
  noise = (noise * 0xFFFF).cast(dtypes.uint32)
  return ((bits + noise) & 0xFFFF0000).bitcast(dtypes.float32).cast(dtypes.bfloat16)

+def clip_grads(grads:list[Tensor], grad_acc, clip_norm) -> Tensor:
+  for g in grads: g.assign(g / grad_acc)
+  total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
+  for g in grads: g.assign((g * (clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(g.dtype))
+  return total_norm
+
 class GradAccClipAdamW(Optimizer):
  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, device=None, fused=FUSE_OPTIM):
    super().__init__(params, lr, device, fused)
@@ -36,46 +42,29 @@ class GradAccClipAdamW(Optimizer):
      self.master_params = None

  def _zero_shard(self, t:Tensor) -> Tensor:
-    if not self.zero or (t.shape[0] % len(self.device)) != 0: return t
-    return Tensor(t.uop._shard(0, len(self.device)).multi(0)).clone()
+    if not self.zero or t.ndim < 2 or (t.shape[0] % len(self.device)) != 0: return t
+    return Tensor(t.uop._shard(0, UOp.range(len(self.device), -1, AxisType.DEVICE)).unshard(0)).clone()

  def _zero_gather(self, t:Tensor) -> Tensor:
    if not isinstance(t.device, tuple) or t.uop.axis != 0: return t
    n, sz = len(t.device), t.shape[0] // len(t.device)
    return Tensor.cat(*[t[p*sz:(p+1)*sz] for p in range(n)], dim=0)

-  def fstep(self, grads:list[Tensor]):
-    if self.fused:
-      out, extra = self._step([], grads)
-      updates = [out[0][self.pos_params[i]:self.pos_params[i+1]].reshape(tt.shape) for i, tt in enumerate(self.params)]
-    else:
-      updates, extra = self._step([], grads)
+  def fschedule_step(self, grads:list[Tensor]) -> list[Tensor]:
+    updates, extra = self._step([], grads)
    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i], self.master_params[i] if self.master_params else None))
-    # collect inv_scale tensors attached to fp8 params (set by _apply_update)
    fp8_inv_scales = [tt._inv_scale for tt in self.params if hasattr(tt, '_inv_scale')]
    fp8_next_inv_scales = [tt._next_inv_scale for tt in self.params if hasattr(tt, '_next_inv_scale')]
-    to_realize = extra+self.params+self.buffers+(self.master_params or [])+fp8_inv_scales+fp8_next_inv_scales
+    return extra + self.params + self.buffers + (self.master_params or []) + fp8_inv_scales + fp8_next_inv_scales

-    Tensor.realize(*to_realize)
-    return extra[-1]
+  def fstep(self, grads:list[Tensor], grad_norm:Tensor|None=None):
+    Tensor.realize(*([grad_norm] if grad_norm is not None else []), *self.fschedule_step(grads))

  def _step(self, params:list[Tensor], grads:list[Tensor]) -> tuple[list[Tensor], list[Tensor]]:
    grads = list(grads)

    for i in range(len(grads)):
      if grads[i].device != self.m[i].device: grads[i] = grads[i].to(self.m[i].device)
-
-    if self.fused:
-      grads[0].assign(grads[0] / self.grad_acc)
-      total_norm = grads[0].float().square().sum().sqrt()
-      grads[0].assign((grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype))
-    else:
-      for i in range(len(grads)):
-        grads[i].assign(grads[i] / self.grad_acc)
-      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
-      for i in range(len(grads)):
-        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype))
-
    ret = []
    self.b1_t *= self.b1
    self.b2_t *= self.b2
@@ -88,7 +77,7 @@ class GradAccClipAdamW(Optimizer):
      v_hat = v_new / (1.0 - self.b2_t)
      up = m_hat / (v_hat.sqrt() + self.eps)
      ret.append(self.lr * up)
-    return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]
+    return ret, [self.b1_t, self.b2_t] + self.m + self.v

  def _apply_update(self, t:Tensor, up:Tensor, master:Tensor|None=None) -> Tensor:
    w = master if master is not None else t
@@ -109,7 +98,7 @@ class GradAccClipAdamW(Optimizer):
        if self.zero: w_q, w_e8 = self._zero_gather(w_q), self._zero_gather(w_e8)
        new_e8 = w_e8.reshape(t._inv_scale.shape)
        t._inv_scale.assign(new_e8.shard_like(t._inv_scale) if offloaded else new_e8)
-        ret = w_q.reshape(new_w.shape)
+        ret = w_q.reshape(t.shape)
        return ret.shard_like(t) if offloaded else ret
      from examples.mlperf.models.flat_llama import FP8_MAX
      if IMMEDIATE_SCALE:
@@ -132,3 +121,24 @@ class GradAccClipAdamW(Optimizer):
      return ret.shard_like(t) if offloaded else ret
    out = new_w.cast(t.dtype)
    return out.shard_like(t) if offloaded else out
+
+class GradAccClipAdamWGroup(OptimizerGroup):
+  def __init__(self, *optimizers:GradAccClipAdamW):
+    super().__init__(*optimizers)
+    for o in self.optimizers[1:]: o.lr = self.optimizers[0].lr
+  def fstep(self, grads:list[Tensor], grad_norm:Tensor|None=None):
+    offset = 0
+    to_realize = []
+    for o in self.optimizers:
+      n = len(o.params)
+      to_realize += o.fschedule_step(grads[offset:offset+n])
+      offset += n
+    Tensor.realize(*to_realize, *([grad_norm] if grad_norm is not None else []))
+  @property
+  def lr(self): return self.optimizers[0].lr
+  @property
+  def device(self): return self.optimizers[0].device
+  @property
+  def master_params(self):
+    mp = [mp for o in self.optimizers for mp in (o.master_params or [])]
+    return mp if mp else None
@@ -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}
@@ -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=${FUSED_INPUT_QUANTIZE:-1}
@@ -1,4 +1,6 @@
 #!/bin/bash
+set -e
+
 export BENCHMARK=${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
@@ -10,6 +10,7 @@ export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -10,6 +10,7 @@ export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
+export ASM_GEMM=${ASM_GEMM:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -1,4 +1,4 @@
-import os, sys, pickle, time, re
+import os, sys, pickle, time, re, tempfile, struct, shutil, io
 import numpy as np
 if "JIT_BATCH_SIZE" not in os.environ: os.environ["JIT_BATCH_SIZE"] = "0"

@@ -9,6 +9,39 @@ from tinygrad.nn.onnx import OnnxRunner

 OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
 OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
+PICKLE_OOB = getenv("PICKLE_OOB")
+
+def dump_pickle(obj, f):
+  if PICKLE_OOB:
+    # allows pickling when buffers don't fit in (CPU) RAM
+    # from openpilot/selfdrive/modeld/helpers.py
+    with tempfile.TemporaryFile(dir=".") as tmp:
+      def buffer_callback(pb: pickle.PickleBuffer):
+        m = pb.raw()
+        tmp.write(struct.pack('<q', m.nbytes))
+        tmp.write(m)
+        pb.release() # keep peak ram at ~1 buffer
+      stream = io.BytesIO()
+      pickle.Pickler(stream, protocol=5, buffer_callback=buffer_callback).dump(obj)
+      opcodes = stream.getvalue()
+      f.write(struct.pack('<q', len(opcodes)))
+      f.write(opcodes)
+      tmp.seek(0)
+      shutil.copyfileobj(tmp, f)
+  else: pickle.dump(obj, f)
+
+def load_pickle(f):
+  if PICKLE_OOB:
+    # allows unpickling when buffers don't fit in (CPU) RAM
+    # from openpilot/selfdrive/modeld/helpers.py
+    opcodes = f.read(struct.unpack('<q', f.read(8))[0])
+    def buffers():
+      while (h := f.read(8)):
+        pb = pickle.PickleBuffer(bytearray(struct.unpack('<q', h)[0]))
+        f.readinto(pb)
+        yield pb
+    return pickle.load(io.BytesIO(opcodes), buffers=buffers())
+  else: return pickle.load(f)

 def compile(onnx_file):
  run_onnx = OnnxRunner(onnx_file)
@@ -28,8 +61,8 @@ def compile(onnx_file):
    inputs = {k:Tensor(v.numpy(), device=Device.DEFAULT).realize() if 'img' in k else v for k,v in inputs.items()}
  print("created tensors")

-  run_onnx_jit = TinyJit(lambda **kwargs:
-                         next(iter(run_onnx({k:v.to(Device.DEFAULT) for k,v in kwargs.items()}).values())).cast('float32'), prune=True)
+  @TinyJit(prune=True)
+  def run_onnx_jit(**kwargs): return next(iter(run_onnx({k:v.to(Device.DEFAULT) for k,v in kwargs.items()}).values())).cast('float32')
  for i in range(3):
    GlobalCounters.reset()
    print(f"run {i}")
@@ -65,8 +98,7 @@ def compile(onnx_file):
  if (allowed_gated_read_image:=getenv("ALLOWED_GATED_READ_IMAGE", -1)) != -1:
    assert gated_read_image_count == allowed_gated_read_image, f"different gated read_image! {gated_read_image_count=}, {allowed_gated_read_image=}"

-  with open(OUTPUT, "wb") as f:
-    pickle.dump(run_onnx_jit, f)
+  with open(OUTPUT, "wb") as f: dump_pickle(run_onnx_jit, f)
  mdl_sz = os.path.getsize(onnx_file)
  pkl_sz = os.path.getsize(OUTPUT)
  print(f"mdl size is {mdl_sz/1e6:.2f}M")
@@ -136,7 +168,7 @@ def bench(run, inputs):

 if __name__ == "__main__":
  if getenv("RUN_PICKLE"):
-    with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)
+    with open(OUTPUT, "rb") as f: pickle_loaded = load_pickle(f)
    inputs = {name: Tensor(Tensor.randn(*view.shape, dtype=dtype).numpy(), device=device)
              for name, (view, _vars, dtype, device) in zip(pickle_loaded.captured.expected_names, pickle_loaded.captured.expected_input_info)}
    test_vs_compile(pickle_loaded, inputs)
@@ -144,7 +176,7 @@ if __name__ == "__main__":
    onnx_file = fetch(OPENPILOT_MODEL)
    inputs, outputs = compile(onnx_file)

-    with open(OUTPUT, "rb") as f: pickle_loaded = pickle.load(f)
+    with open(OUTPUT, "rb") as f: pickle_loaded = load_pickle(f)

    test_vs_compile(pickle_loaded, inputs, outputs)
    if getenv("SELFTEST"):
@@ -1,5 +1,6 @@
-import sys, pickle
+import sys
 from extra.bench_log import WallTimeEvent, BenchEvent
+from examples.openpilot.compile3 import load_pickle
 from tinygrad.helpers import getenv

 PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
@@ -7,7 +8,7 @@ PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
 load_times = []

 for _ in range(10):
-  with WallTimeEvent(BenchEvent.STEP) as wte: pickle.load(open(PKL, 'rb'))
+  with WallTimeEvent(BenchEvent.STEP) as wte: load_pickle(open(PKL, 'rb'))
  load_times.append(wte.time)
  print(f"pickle load: {wte.time:6.2f} s")

@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+"""Benchmark Kimi-Linear load, prefill, and decode on its TP4 checkpoint."""
+import argparse, resource, time
+from tinygrad import Device, TinyJit
+from tinygrad.helpers import profile_marker
+from tinygrad.llm.kimi import load_kimi
+
+def sync(devices:int) -> None:
+  for i in range(devices): Device[f"AMD:{i}"].synchronize()
+
+def timed_next(gen, devices:int) -> tuple[int, float]:
+  begin = time.perf_counter()
+  token = next(gen)
+  sync(devices)
+  return token, time.perf_counter()-begin
+
+def fresh_generate(model, prompt:list[int], chunk_size:int):
+  # Force recurrent/KV state reset so repeated runs and chunk sweeps measure the entire prompt,
+  # rather than silently reusing the prefix cached by the previous measurement.
+  model._cached_tokens = [-1] * len(prompt)
+  return model.generate(prompt.copy(), chunk_size=chunk_size)
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("model", help="converted Kimi-Linear-48B-A3B MXFP4-v2 directory")
+  parser.add_argument("--devices", type=int, default=4)
+  parser.add_argument("--max-context", type=int, default=128)
+  parser.add_argument("--prompt-tokens", type=int, default=32)
+  parser.add_argument("--decode-tokens", type=int, default=8)
+  parser.add_argument("--chunk-size", type=int, default=32)
+  parser.add_argument("--sweep-chunks", help="comma-separated prefill chunk sizes; uses the fastest for decode")
+  args = parser.parse_args()
+  if args.prompt_tokens < 1 or args.prompt_tokens + args.decode_tokens + 1 > args.max_context:
+    raise ValueError("prompt and decode tokens must fit within --max-context")
+
+  begin = time.perf_counter()
+  model = load_kimi(args.model, max_context=args.max_context, devices=args.devices)
+  sync(args.devices)
+  print(f"load: {time.perf_counter()-begin:.3f}s", flush=True)
+
+  prompt = [1] + [1000+i%1000 for i in range(args.prompt_tokens-1)]
+  chunks = [int(x) for x in args.sweep_chunks.split(",")] if args.sweep_chunks else [args.chunk_size]
+  if any(x < 1 or x > args.prompt_tokens for x in chunks): raise ValueError("prefill chunks must be between 1 and --prompt-tokens")
+  timings:list[tuple[float, int]] = []
+  prefill_jits:dict[int, TinyJit] = {}
+  for chunk in chunks:
+    # Recurrent prefill has a static token dimension. Give each swept shape its own capture;
+    # the rollout JIT remains shared and independently benchmarks chunk 1/decode.
+    if chunk != 1: model.prefill_jit = TinyJit(model.forward)
+    cold = fresh_generate(model, prompt, chunk)
+    first, cold_prefill = timed_next(cold, args.devices)
+    print(f"chunk {chunk}: cold prefill {cold_prefill:.3f}s, token={first}", flush=True)
+    warm = fresh_generate(model, prompt, chunk)
+    warm_first, prefill = timed_next(warm, args.devices)
+    if first != warm_first: raise RuntimeError(f"chunk {chunk} is not repeatable: cold={first}, warm={warm_first}")
+    timings.append((prefill, chunk))
+    if chunk != 1: prefill_jits[chunk] = model.prefill_jit
+    print(f"chunk {chunk}: prefill {prefill:.3f}s ({args.prompt_tokens/prefill:.3f} tok/s), token={first}", flush=True)
+
+  prefill, best_chunk = min(timings)
+  if best_chunk != 1: model.prefill_jit = prefill_jits[best_chunk]
+  warm = fresh_generate(model, prompt, best_chunk)
+  first, replay_prefill = timed_next(warm, args.devices)
+  _, cold_decode = timed_next(warm, args.devices)
+  _, capture_decode = timed_next(warm, args.devices)
+  print(f"selected chunk: {best_chunk}; prefill replay {replay_prefill:.3f}s "
+        f"({args.prompt_tokens/replay_prefill:.3f} tok/s), token={first}", flush=True)
+  print(f"cold decode: {cold_decode:.3f}s", flush=True)
+  print(f"capture decode: {capture_decode:.3f}s", flush=True)
+  profile_marker("kimi decode steady start")
+  begin = time.perf_counter()
+  output = [next(warm) for _ in range(args.decode_tokens)]
+  sync(args.devices)
+  decode = time.perf_counter()-begin
+  profile_marker("kimi decode steady end")
+  print(f"decode: {decode:.3f}s ({args.decode_tokens/decode:.3f} tok/s, {decode/args.decode_tokens*1e3:.3f} ms/tok), output={output}", flush=True)
+  print(f"peak RSS: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss/1024:.1f} MiB", flush=True)
+
+if __name__ == "__main__": main()
@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""Bounded correctness and load/prefill/decode benchmark for the official TP8 Kimi K3 checkpoint."""
+import argparse, resource, time
+
+from tinygrad import Device
+from tinygrad.helpers import profile_marker
+from tinygrad.llm.cli import KimiK3Template, SimpleTokenizer
+from tinygrad.llm.kimi_k3 import load_kimi_k3, load_kimi_tokenizer_data
+
+def sync(devices:int) -> None:
+  for i in range(devices): Device[f"AMD:{i}"].synchronize()
+
+def fresh_generate(model, prompt:list[int], chunk_size:int):
+  # Never reuse a prefix or recurrent state across correctness/benchmark trials.
+  model._cached_tokens = [-1] * len(prompt)
+  return model.generate(prompt.copy(), chunk_size=chunk_size, temperature=0.0)
+
+def timed_next(gen, devices:int) -> tuple[int, float]:
+  begin = time.perf_counter()
+  token = next(gen)
+  sync(devices)
+  return token, time.perf_counter()-begin
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("model", help="official unmodified Kimi K3 checkpoint directory")
+  parser.add_argument("--devices", type=int, default=8)
+  parser.add_argument("--max-context", type=int, default=128)
+  parser.add_argument("--prompt", default="Reply with exactly: OK")
+  parser.add_argument("--stable-tokens", type=int, default=8)
+  parser.add_argument("--decode-tokens", type=int, default=8)
+  parser.add_argument("--chunk-size", type=int, default=128)
+  args = parser.parse_args()
+
+  begin = time.perf_counter()
+  model = load_kimi_k3(args.model, max_context=args.max_context, devices=args.devices)
+  sync(args.devices)
+  load_time = time.perf_counter()-begin
+  print(f"load: {load_time:.3f}s", flush=True)
+
+  normal, special, bos, eos = load_kimi_tokenizer_data(args.model)
+  tok = SimpleTokenizer(normal, special, "kimi-k2", bos_id=bos, eos_id=eos, eot_id=eos)
+  rendered = KimiK3Template().render(messages=[{"role":"user", "content":args.prompt}], add_generation_prompt=True)
+  prompt = tok.encode(rendered)
+  needed = len(prompt) + max(args.stable_tokens, args.decode_tokens+3)
+  if needed > args.max_context: raise ValueError(f"prompt and output need {needed} tokens but max context is {args.max_context}")
+  print(f"prompt: {len(prompt)} tokens, chunk={args.chunk_size}", flush=True)
+
+  sequences:list[list[int]] = []
+  # TinyJit executes uncaptured once, captures the second call, and replays from the third call.
+  # Compare two replay paths rather than capture numerics/timing against replay.
+  for trial in range(4):
+    gen = fresh_generate(model, prompt, args.chunk_size)
+    sequence:list[int] = []
+    prefill = 0.0
+    for step in range(args.stable_tokens):
+      token, elapsed = timed_next(gen, args.devices)
+      sequence.append(token)
+      if step == 0: prefill = elapsed
+    if trial >= 2: sequences.append(sequence)
+    label = ("uncaptured warmup", "capture warmup", "stable trial 1", "stable trial 2")[trial]
+    print(f"{label}: prefill={prefill:.3f}s "
+          f"({len(prompt)/prefill:.3f} tok/s), tokens={sequence}", flush=True)
+  if sequences[0] != sequences[1]: raise RuntimeError(f"greedy output is not repeatable: {sequences}")
+  print(f"stable text: {tok.decode(sequences[0])!r}", flush=True)
+
+  gen = fresh_generate(model, prompt, args.chunk_size)
+  profile_marker("kimi k3 steady prefill start")
+  first, prefill = timed_next(gen, args.devices)
+  profile_marker("kimi k3 steady prefill end")
+  warmup = [timed_next(gen, args.devices)[0] for _ in range(2)]
+  profile_marker("kimi k3 steady decode start")
+  begin = time.perf_counter()
+  output = [next(gen) for _ in range(args.decode_tokens)]
+  sync(args.devices)
+  decode = time.perf_counter()-begin
+  profile_marker("kimi k3 steady decode end")
+  print(f"prefill replay: {prefill:.3f}s ({len(prompt)/prefill:.3f} tok/s), token={first}", flush=True)
+  print(f"decode after warmup {warmup}: {decode:.3f}s ({args.decode_tokens/decode:.3f} tok/s, "
+        f"{decode/args.decode_tokens*1e3:.3f} ms/tok), output={output}", flush=True)
+  print(f"peak RSS: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss/1024:.1f} MiB", flush=True)
+
+if __name__ == "__main__": main()
@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+"""Fast exact-shape K3 KDA/layer benchmark using bounded fake weights instead of the 1.56 TB checkpoint."""
+from __future__ import annotations
+import argparse, statistics, time
+from dataclasses import replace
+
+from tinygrad import Device, Tensor, TinyJit, dtypes, nn
+from tinygrad.helpers import profile_marker
+from tinygrad.llm.kimi_k3 import kimi_k3_config
+from tinygrad.llm.model import GatedDeltaNetBlock
+
+def tp_axis(name:str) -> int|None:
+  if "ffn_gate_exps.weight" in name or "ffn_up_exps.weight" in name: return 1
+  if "ffn_gate_exps.weight_scale" in name or "ffn_up_exps.weight_scale" in name: return 1
+  if "ffn_down_exps.weight" in name or "ffn_down_exps.weight_scale" in name: return 2
+  if name.endswith(("ffn_gate_shexp.weight", "ffn_up_shexp.weight")): return 0
+  if name.endswith(("ffn_down_shexp.weight", "ffn_routed_down.weight", "ffn_routed_up.weight", "ssm_out.weight")): return 1
+  if name.endswith(("attn_q.weight", "attn_k.weight", "attn_v.weight", "ssm_g_full.weight", "ssm_f_b.weight", "ssm_beta.weight")): return 0
+  if name.endswith(("ssm_q_conv1d.weight", "ssm_k_conv1d.weight", "ssm_v_conv1d.weight", "ssm_dt.bias")): return 0
+  return None
+
+def fake_value(name:str) -> tuple[int|float, object]:
+  if name.endswith("weight_scale"): return 120, dtypes.uint8
+  if name.endswith("_exps.weight"): return 0x11, dtypes.uint8
+  if name.endswith("ssm_a"): return -0.1, dtypes.float32
+  if name.endswith("ssm_dt.bias"): return 0.1, dtypes.float32
+  if "conv1d.weight" in name: return 0.1, dtypes.float32
+  if name.endswith("exp_probs_b.bias"): return 0.0, dtypes.float32
+  if name.endswith("norm.weight"): return 1.0, dtypes.bfloat16
+  return 0.001, dtypes.bfloat16
+
+def fake_tp_tensor(shape:tuple[int, ...], value:int|float, dtype, devices:tuple[str, ...], axis:int|None) -> Tensor:
+  if axis is not None and shape[axis] % len(devices): raise ValueError(f"shape {shape} is not TP{len(devices)} divisible on axis {axis}")
+  source = Tensor.full(shape, value, dtype=dtype, device=devices[0]).clone().realize()
+  return source.shard(devices, axis=axis).realize()
+
+def sync(devices:tuple[str, ...]) -> None:
+  for device in devices: Device[device].synchronize()
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--devices", type=int, default=8)
+  parser.add_argument("--mode", choices=("attention", "block"), default="attention")
+  parser.add_argument("--iterations", type=int, default=20)
+  args = parser.parse_args()
+  devices = tuple(f"AMD:{i}" for i in range(args.devices))
+  # One exact-width KDA layer, but only 16 fake routed experts. This retains top-k 16 and every
+  # official per-GPU matrix/state shape while keeping fake expert storage below 300 MB per layer.
+  config = replace(kimi_k3_config(4), num_blocks=1, num_experts=16, num_experts_per_tok=16, ssm_layers=(True,),
+                   attn_res_block_size=0)
+  block = GatedDeltaNetBlock(config, config.ssm)
+  begin = time.perf_counter()
+  for name,tensor in nn.state.get_state_dict(block).items():
+    if args.mode == "attention" and name.startswith(("ffn_", "exp_probs_")): continue
+    value, dtype = fake_value(name)
+    tensor.replace(fake_tp_tensor(tuple(int(x) for x in tensor.shape), value, dtype, devices, tp_axis(name)))
+  sync(devices)
+  print(f"fake weights: {time.perf_counter()-begin:.3f}s", flush=True)
+  x_source = (((Tensor.arange(config.dim, dtype=dtypes.float32).reshape(1, 1, config.dim) % 31) / 31) \
+    .cast(dtypes.bfloat16).to(devices[0])).clone().realize()
+  x = x_source.shard(devices, axis=None).realize()
+  block._init_state(x)
+  # Use direct buffer-backed state shards. The production path reaches this form after prefill;
+  # the fake harness begins immediately at decode and must not feed lazy clone graphs to TinyJit.
+  for state,axis in ((block.conv_state_q, 2), (block.conv_state_k, 2), (block.conv_state_v, 2), (block.recurrent_state, 1)):
+    state.replace(Tensor.zeros(*state.shape, dtype=state.dtype, device=devices[0]).shard(devices, axis=axis).realize())
+
+  @TinyJit
+  def run(inp:Tensor) -> Tensor:
+    if args.mode == "attention": return block._attention(block.attn_norm(inp), 0).realize()
+    return block(inp, 0).realize()
+
+  # uncaptured, capture, then replay only
+  run(x); sync(devices)
+  run(x); sync(devices)
+  samples:list[float] = []
+  profile_marker(f"fake K3 {args.mode} start")
+  for _ in range(args.iterations):
+    begin = time.perf_counter(); out = run(x); sync(devices); samples.append((time.perf_counter()-begin)*1e3)
+  profile_marker(f"fake K3 {args.mode} end")
+  print(f"{args.mode}: median={statistics.median(samples):.3f} ms/layer, min={min(samples):.3f} ms/layer, "
+        f"projected_93_layer_rate={1000/(statistics.median(samples)*93):.3f} tok/s, finite={out.float().isfinite().all().item()}")
+
+if __name__ == "__main__": main()
@@ -0,0 +1,31 @@
+import argparse, time
+from tinygrad.llm.model import Transformer
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--model", required=True, help="path to gguf model")
+  parser.add_argument("--max-context", type=int, default=8192, help="max context length (default: %(default)s)")
+  parser.add_argument("--prompt-tokens", type=int, default=1024, help="number of prompt tokens (default: %(default)s)")
+  parser.add_argument("--decode-tokens", type=int, default=16, help="number of tokens to decode (default: %(default)s)")
+  parser.add_argument("--chunk-size", type=int, default=32, help="chunk size for prefill (default: %(default)s)")
+  args = parser.parse_args()
+
+  st = time.perf_counter()
+  model, _ = Transformer.from_gguf(args.model, args.max_context)
+  print(f"load {time.perf_counter()-st:.3f}s", flush=True)
+
+  st = time.perf_counter()
+  model.warmup()
+  print(f"warm {time.perf_counter()-st:.3f}s", flush=True)
+
+  prompt = [257] + [1000+i%1000 for i in range(args.prompt_tokens-1)]
+  gen = model.generate(prompt, chunk_size=args.chunk_size)
+  st = time.perf_counter()
+  # first token is time-to-first-token; counted as part of prefill
+  output = [next(gen)]
+  pt = time.perf_counter()
+  print(f"prefill {args.prompt_tokens/(pt-st):.3f} tok/s", flush=True)
+
+  for _ in range(args.decode_tokens): output.append(next(gen))
+  et = time.perf_counter()
+  print(f"decode {args.decode_tokens/(et-pt):.3f} tok/s output {output}", flush=True)
@@ -241,8 +241,8 @@ export default {model_name};
 def export_model(model, target:str, *inputs, model_name: Optional[str] = "model", stream_weights=False):
  assert Device.DEFAULT in EXPORT_SUPPORTED_DEVICE, f"only {', '.join(EXPORT_SUPPORTED_DEVICE)} are supported"

-  # NOTE: CPU_COUNT=1, since export does not support threading
-  with Context(JIT=2, CPU_COUNT=1): linear, output_bufs = jit_model(model, *inputs)
+  # NOTE: NUM_CPU_THREADS=1, since export does not support threading
+  with Context(JIT=2, NUM_CPU_THREADS=1): linear, output_bufs = jit_model(model, *inputs)
  functions, statements, bufs, bufs_to_save = compile_net(linear, output_bufs)
  state = get_state_dict(model)
  weight_names = {(id(b), b.offset, b.size, b.dtype): name for name, x in state.items() if (b:=x.uop.base.realized) is not None}
@@ -264,7 +264,7 @@ def export_model(model, target:str, *inputs, model_name: Optional[str] = "model"
        if getattr(dim, "op", None) is Ops.ADD and len(dim.src) == 2 and \
           any(s.op is Ops.PARAM and s.addrspace is AddrSpace.ALU for s in dim.src) and any(s.op is Ops.CONST for s in dim.src):
          name, val = dim.src if dim.src[1].op is Ops.CONST else reversed(dim.src)
-          global_size[j] = f"_{name.expr}[0] + {val.arg}"
+          global_size[j] = f"_{name.expr}[0] + {val.val}"

  prg = ""
  if target == "clang":
@@ -18,9 +18,9 @@ def custom_matmul(output: UOp, inp: UOp, weight: UOp) -> UOp:
  SEQ = inp.shape[1]
  OUT = weight.shape[0]
  IN = weight.shape[-1]
-  seq_idx = UOp.range(SEQ, 2, AxisType.LOOP)
-  out_idx = UOp.range(OUT, 3, AxisType.LOOP)
-  batch_idx = UOp.range(output.size//SEQ//OUT, 1, AxisType.LOOP)
+  seq_idx = UOp.range(SEQ, 2)
+  out_idx = UOp.range(OUT, 3)
+  batch_idx = UOp.range(output.size//SEQ//OUT, 1)
  reduce_idx = UOp.range(IN, 0, AxisType.REDUCE)
  product = (inp.index((seq_idx*IN+reduce_idx+batch_idx*IN*SEQ)) * weight.index((out_idx*IN+reduce_idx))).cast(dtypes.float)
  reduced = product.reduce(reduce_idx, arg=Ops.ADD)
@@ -53,7 +53,7 @@ class FP8Linear:
    x_fp8, x_scale = quantize_to_fp8(x)
    GPUS = self.weight.device
    if isinstance(GPUS, tuple) and len(GPUS) > 1:
-      y = Tensor(Tensor.empty((batch//len(GPUS), seq, self.weight.shape[0]), dtype=dtypes.float, device=GPUS).uop.multi(0), device=GPUS)
+      y = Tensor(Tensor.empty((batch//len(GPUS), seq, self.weight.shape[0]), dtype=dtypes.float, device=GPUS).uop.unshard(0), device=GPUS)
    else:
      y = Tensor.empty((batch, seq, self.weight.shape[0]), dtype=dtypes.float)
    y = Tensor.custom_kernel(y, x_fp8, w_fp8, fxn=custom_matmul, grad_fxn=custom_matmul_backward)[0]
@@ -58,8 +58,8 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
  B_copy = B_local.permute((1,0)) if use_wmma else B_local
  A_store = A_copy.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(a[k_tile].reshape(-1, THREADS_PER_BLOCK)[:, tid])
  B_store = B_copy.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(b[k_tile].reshape(-1, THREADS_PER_BLOCK)[:, tid])
-  barrier = UOp.barrier(A_store, B_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
+  # NOTE: no explicit barrier needed, the AFTER on the LOCAL buffers implies it in late codegen
+  A_local, B_local = A_local.after(A_store, B_store), B_local.after(A_store, B_store)

  # -- COMPUTE --
  lane_m, lane_n = lane // LANES_PER_WAVE_N, lane % LANES_PER_WAVE_N
@@ -70,8 +70,8 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:

  if use_wmma:
    k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)
-    tile_m = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-    tile_n = UOp.range(TN, 201, AxisType.LOOP)
+    tile_m = UOp.range(TM // WMMA_ACC, 200)
+    tile_n = UOp.range(TN, 201)

    acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0,2,1)[tile_m, tile_n]
    a_frag = A_local.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_K // WMMA_K, WMMA_K)[wave_m, tile_m, lane_n, k]
@@ -96,8 +96,8 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
    b_frag = b_frag.reshape(1, TN).expand(TM, TN)
    acc_store = acc.store(acc.after(k) + (a_frag * b_frag))

-  # store accumulator and loop
-  acc = acc.after(acc_store.end(k).barrier().end(k_tile))
+  # store accumulator and loop (the barrier at the end of the loop is implied by the LOCAL buffers stored and loaded in the loop)
+  acc = acc.after(acc_store.end(k).end(k_tile))

  # store accumulator to output (unified)
  c = c.reshape(WAVES_M, TM//UNROLL_M, LANES_PER_WAVE_M, UNROLL_M,
@@ -1,203 +0,0 @@
-from tinygrad import Tensor, UOp, getenv
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-from tinygrad.dtype import AddrSpace, dtypes
-from tinygrad.helpers import DEBUG, GlobalCounters, Context
-import math
-
-BLOCK_M, BLOCK_N = 64, 64
-WARP_SIZE = 32
-WMMA_M, WMMA_N, WMMA_K = 16, 16, 16
-WAVES_M, WAVES_N = 4, 1
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 2, 16
-WMMA_ACC = WMMA_M // LANES_PER_WAVE_M
-THREADS_PER_BLOCK = WARP_SIZE * WAVES_M * WAVES_N
-LDS_PAD = 4  # pad LDS rows to reduce bank conflicts
-
-WMMA_ARG = (WMMA_M, WMMA_N, WMMA_K), 'AMD', 32
-LOG2E = math.log2(math.e)
-
-def warp_shfl_xor(val, offset, lane):
-  """Read val from lane ^ offset using ds_bpermute."""
-  idx = ((lane ^ offset) * 4).cast(dtypes.int)
-  if val.op is Ops.INDEX and val.addrspace == AddrSpace.REG: val = val.load()
-  return UOp(Ops.CUSTOM, dtypes.float, (idx, val),
-             arg="__builtin_bit_cast(float, __builtin_amdgcn_ds_bpermute({0}, __builtin_bit_cast(int, {1})))")
-
-def warp_reduce_max(val, lane):
-  """Tree reduce MAX across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = UOp(Ops.MAX, dtypes.float, (val, warp_shfl_xor(val, offset, lane)))
-  return val
-
-def warp_reduce_sum(val, lane):
-  """Tree reduce SUM across LANES_PER_WAVE_N=16 lanes."""
-  for offset in [8, 4, 2, 1]:
-    val = val + warp_shfl_xor(val, offset, lane)
-  return val
-
-def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
-  # inputs are (B*H, N, D)
-  BH, N, D = q.shape
-  assert N % BLOCK_M == 0 and N % BLOCK_N == 0, f"N={N} must be divisible by BLOCK_M={BLOCK_M} and BLOCK_N={BLOCK_N}"
-  assert D % WMMA_K == 0 and D % LANES_PER_WAVE_N == 0, f"D={D} must be divisible by WMMA_K={WMMA_K} and LANES_PER_WAVE_N={LANES_PER_WAVE_N}"
-  assert BLOCK_M % (WAVES_M * WMMA_M) == 0 and BLOCK_N % LANES_PER_WAVE_N == 0
-  TM = BLOCK_M // (WAVES_M * LANES_PER_WAVE_M)
-  TN = BLOCK_N // (WAVES_N * LANES_PER_WAVE_N)
-  TD = D // (WAVES_N * LANES_PER_WAVE_N)
-  SCALE = 1.0 / math.sqrt(D)
-
-  block_bh = UOp.range(BH, 0, AxisType.GLOBAL)
-  block_m = UOp.range(N // BLOCK_M, 1, AxisType.GLOBAL)
-
-  q = q.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-  k = k.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  v = v.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
-  o = o.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
-
-  wave_m = UOp.range(WAVES_M, 2, AxisType.LOCAL)
-  wave_n = UOp.range(WAVES_N, 3, AxisType.LOCAL)
-  lane = UOp.range(WARP_SIZE, -1, AxisType.WARP)
-  tid = (wave_m * WAVES_N + wave_n) * WARP_SIZE + lane
-  lane_m = lane // LANES_PER_WAVE_N
-  lane_n = lane % LANES_PER_WAVE_N
-
-  # LDS allocation: slot 0 = Q then P (shared), slot 1 = K then V
-  # TODO: the memory planner should be able to find this reuse
-  ELEMS_PER_THREAD = BLOCK_M * D // THREADS_PER_BLOCK
-  QP_lds = UOp.placeholder((BLOCK_M, D + LDS_PAD), dtypes.half, slot=0, addrspace=AddrSpace.LOCAL)
-  KV_lds = UOp.placeholder((BLOCK_N, D + LDS_PAD), dtypes.half, slot=1, addrspace=AddrSpace.LOCAL)[:, :D]
-
-  # register state
-  acc = UOp.placeholder((TM, TD), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  m_i = UOp.placeholder((TM,), dtypes.float, slot=3, addrspace=AddrSpace.REG)
-  l_i = UOp.placeholder((TM,), dtypes.float, slot=4, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(acc.const_like(0)))
-  m_i = m_i.after(m_i.store(m_i.const_like(-math.inf)))
-  l_i = l_i.after(l_i.store(l_i.const_like(0)))
-
-  # ====== KV tile loop ======
-  n_tile = UOp.range(N // BLOCK_N, 100, AxisType.REDUCE)
-
-  # load Q + K into LDS (Q reloaded each iteration since P overwrites slot 0)
-  Q_lds = QP_lds[:, :D]
-  Q_store = Q_lds.after(n_tile).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    q.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  K_store = KV_lds.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    k[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  qk_load_barrier = UOp.barrier(UOp.group(Q_store, K_store))
-  Q_lds = Q_lds.after(qk_load_barrier)
-  KV_lds_k = KV_lds.after(qk_load_barrier)
-
-  # -- S = Q @ K^T via WMMA (re-init each n_tile) --
-  S_reg = UOp.placeholder((TM, TN), dtypes.float, slot=6, addrspace=AddrSpace.REG)
-  S_reg = S_reg.after(S_reg.after(n_tile).store(S_reg.const_like(0)))
-  k_qk = UOp.range(D // WMMA_K, 101, AxisType.REDUCE)
-  tm1 = UOp.range(TM // WMMA_ACC, 200, AxisType.LOOP)
-  tn1 = UOp.range(TN, 201, AxisType.LOOP)
-  S_frag = S_reg.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0, 2, 1)[tm1, tn1]
-  q_frag = Q_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, D // WMMA_K, WMMA_K)[wave_m, tm1, lane_n, k_qk]
-  k_frag = KV_lds_k.reshape(WAVES_N, TN, WMMA_N, D // WMMA_K, WMMA_K)[wave_n, tn1, lane_n, k_qk]
-  qk = UOp.wmma(q_frag, k_frag, S_frag.after(k_qk), *WMMA_ARG)
-  qk_done = S_frag.store(qk).end(tm1, tn1).end(k_qk)
-  S_reg = S_reg.after(qk_done)
-
-  # -- softmax in registers with warp shuffles --
-  S_reg = S_reg.after(S_reg.store(S_reg * SCALE))
-
-  # per-thread local row max over TN=4 elements, then warp reduce across 16 lanes
-  m_ij = UOp.placeholder((TM,), dtypes.float, slot=7, addrspace=AddrSpace.REG)
-  m_ij = m_ij.after(m_ij.after(n_tile).store(m_ij.const_like(-math.inf)))
-  rm2 = UOp.range(TN, 261, AxisType.REDUCE)
-  m_ij = m_ij.after(m_ij.store(m_ij.after(rm2).maximum(S_reg[:, rm2])).end(rm2))
-  # warp reduce max (in-place)
-  ri_w = UOp.range(TM, 270, AxisType.LOOP)
-  m_ij = m_ij.after(m_ij[ri_w].store(warp_reduce_max(m_ij[ri_w], lane)).end(ri_w))
-
-  # compute P = exp(S - m_ij) in S_reg
-  S_reg = S_reg.after(S_reg.store(((S_reg - m_ij.reshape(TM, 1).expand(TM, TN)) * LOG2E).exp2()))
-
-  p_local = UOp.placeholder((TM,), dtypes.float, slot=8, addrspace=AddrSpace.REG)
-  p_local = p_local.after(p_local.after(n_tile).store(p_local.const_like(0)))
-  rp2 = UOp.range(TN, 291, AxisType.REDUCE)
-  p_local = p_local.after(p_local.store(p_local.after(rp2) + S_reg[:, rp2]).end(rp2))
-  ri_ws = UOp.range(TM, 295, AxisType.LOOP)
-  p_sum = p_local.after(p_local[ri_ws].store(warp_reduce_sum(p_local[ri_ws], lane)).end(ri_ws))
-
-  # write P = exp(S - m_ij) to P_lds (reuses slot 0, Q no longer needed)
-  P_lds = QP_lds[:, :BLOCK_N]
-  P_write = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TN, LANES_PER_WAVE_N)
-  P_write = P_write.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TN)
-  P_store = P_write[tid].store(S_reg.cast(dtypes.half))
-
-  # -- online softmax correction --
-  ri4 = UOp.range(TM, 330, AxisType.LOOP)
-  m_new_val = m_i[ri4].maximum(m_ij[ri4])
-  alpha_val = ((m_i[ri4] - m_new_val) * LOG2E).exp2()
-  beta_val = ((m_ij[ri4] - m_new_val) * LOG2E).exp2()
-  rj4 = UOp.range(TD, 331, AxisType.LOOP)
-  correction = UOp.group(
-    acc[ri4, rj4].store(alpha_val * acc[ri4, rj4]).end(rj4),
-    l_i[ri4].store(alpha_val * l_i[ri4] + beta_val * p_sum[ri4]),
-    m_i[ri4].store(m_new_val),
-  ).end(ri4)
-  acc = acc.after(correction)
-  l_i = l_i.after(correction)
-  m_i = m_i.after(correction)
-
-  # load V into KV_lds (must wait for QK WMMA to finish reading K from KV_lds)
-  V_store = KV_lds.after(qk_done).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
-    v[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
-  pv_barrier = UOp.barrier(UOp.group(P_store, V_store))
-  P_lds = P_lds.after(pv_barrier)
-  KV_lds_v = KV_lds.after(pv_barrier)
-
-  # -- acc += P @ V via WMMA --
-  k_pv = UOp.range(BLOCK_N // WMMA_K, 400, AxisType.REDUCE)
-  tm2 = UOp.range(TM // WMMA_ACC, 401, AxisType.LOOP)
-  tn2 = UOp.range(TD, 402, AxisType.LOOP)
-  acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TD).permute(0, 2, 1)[tm2, tn2]
-  p_frag = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_N // WMMA_K, WMMA_K)[wave_m, tm2, lane_n, k_pv]
-  v_frag = KV_lds_v.reshape(WAVES_N, TD, WMMA_N, BLOCK_N // WMMA_K, WMMA_K)[wave_n, tn2, lane_n, k_pv]
-  pv = UOp.wmma(p_frag, v_frag, acc_frag.after(k_pv), *WMMA_ARG)
-
-  # end KV tile loop
-  n_tile_end = acc_frag.store(pv).end(tm2, tn2).end(k_pv).barrier().end(n_tile)
-  acc = acc.after(n_tile_end)
-  l_i = l_i.after(n_tile_end)
-  m_i = m_i.after(n_tile_end)
-
-  # normalize: acc /= l_i
-  acc = acc.after(acc.store(acc * (1 / l_i).reshape(TM, 1).expand(TM, TD)))
-
-  # store output
-  o = o.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TD, LANES_PER_WAVE_N)
-  o = o.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TD)
-  return o[tid].store(acc).end(wave_m, wave_n, lane).end(block_m, block_bh).sink(arg=KernelInfo(opts_to_apply=()))
-
-if __name__ == "__main__":
-  B, H, N, D = getenv("B", 1), getenv("H", 32), getenv("N", 1024), getenv("D", 64)
-  q = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  k = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  v = Tensor.rand(B, H, N, D).cast(dtypes.half)
-  o = Tensor.empty(B, H, N, D, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(q, k, v)
-
-  q_flat, k_flat, v_flat, o_flat = q.reshape(B*H, N, D), k.reshape(B*H, N, D), v.reshape(B*H, N, D), o.reshape(B*H, N, D)
-  NUM_RUNS = getenv("CNT", 5)
-  ets = []
-  with Context(DEBUG=2):
-    for _ in range(NUM_RUNS):
-      GlobalCounters.reset()
-      tst = Tensor.custom_kernel(o_flat, q_flat, k_flat, v_flat, fxn=amd_flash_attention)[0].realize()
-      ets.append(GlobalCounters.time_sum_s)
-  print(f"best time: {min(ets)*1e3:.2f}ms")
-
-  if getenv("VERIFY", 1):
-    with Context(DEBUG=0):
-      ref = q.float().scaled_dot_product_attention(k.float(), v.float()).reshape(B*H, N, D).realize()
-      err = (ref - tst).square().mean().item()
-    print(f"mean squared error {err}")
-    if err > 1e-2:
-      raise RuntimeError("flash attention is wrong!")
-    else:
-      print("flash attention is correct!")
@@ -28,10 +28,10 @@ REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
 assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
 assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"

-def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
+def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.WEAK): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
 def copy(dest:UOp, src:UOp, rng:int, upcast=False):
  assert dest.shape == src.shape
-  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
+  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.WEAK)
  return dest[*rngs].store(src[*rngs]).end(*rngs)

 def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
@@ -66,9 +66,8 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
  B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
  B_local_store = copy(B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)

-  # TODO: can we automate barrier?
-  barrier = UOp.barrier(A_local_store, B_local_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
+  # NOTE: no explicit barrier needed, the AFTER on the LOCAL buffers implies it in late codegen
+  A_local, B_local = A_local.after(A_local_store, B_local_store), B_local.after(A_local_store, B_local_store)

  # open inner k range
  k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
@@ -102,7 +101,7 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iter_m, t_m] * B_row[iter_n, t_n]).end(iter_m, iter_n, t_m, t_n)

  # Close k, sync, and close K tiles
-  sink = sink.end(k).barrier().end(k_tile_range)
+  sink = sink.end(k).end(k_tile_range)

  # ---------------------------
  # REG -> GLOBAL (epilogue)
@@ -1,10 +1,12 @@
-import atexit, functools, pathlib
+import atexit, functools, math, pathlib
 from tinygrad import Tensor, Device, dtypes
+from tinygrad.dtype import AddrSpace
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
 from tinygrad.renderer import Estimates
-from tinygrad.helpers import getenv, all_same, DEBUG
+from tinygrad.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

@@ -72,7 +74,7 @@ def hk_fp8_atb_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, g_amax:Tensor|
    elif b.uop.axis == 2: inv, out_axis = Tensor.invalids(1, M, N // ndev, dtype=dtypes.bfloat16, device=a.device), 2
    elif a.uop.axis == 2: inv, out_axis = Tensor.invalids(1, M // ndev, N, dtype=dtypes.bfloat16, device=a.device), 1
    else: inv, out_axis, reduce_out = Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a.device), 0, True
-    out = Tensor(inv.uop.multi(out_axis), device=a.device)
+    out = Tensor(inv.uop.unshard(out_axis), device=a.device)
    dname = a.device[0]
  else:
    out = Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a.device)
@@ -107,6 +109,42 @@ def custom_hk_mxfp8_gemm(C:UOp, A:UOp, B:UOp, scale_A:UOp, scale_B:UOp, *extra:U
  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
                               UOp(Ops.BINARY, arg=lib)))

+# ** MXFP4 GEMM custom kernel
+
+@functools.cache
+def custom_mxfp4_gemm(C:UOp, A:UOp, B:UOp, scale_a:UOp, scale_b:UOp, *extra:UOp, tile_m:int, tile_n:int) -> UOp:
+  from extra.gemm.gemm_mxfp4 import build_kernel
+  M, half_k = math.prod(A.shape[:-1]), A.shape[-1]
+  N, half_k_b = math.prod(B.shape[:-1]), B.shape[-1]
+  K = half_k * 2
+  assert half_k == half_k_b and math.prod(C.shape[:-1]) == M and C.shape[-1] == N
+  threads = UOp.special(256, "lidx0")
+  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)))
+  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
@@ -171,13 +209,13 @@ def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
  M, K = A.shape[0]*A.shape[1], A.shape[2]
  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
  assert K == K2
-  m = UOp.range(M, 1, AxisType.LOOP)
-  n = UOp.range(N, 2, AxisType.LOOP)
+  m = UOp.range(M, 1)
+  n = UOp.range(N, 2)
  k = UOp.range(K, 0, AxisType.REDUCE)
-  mul = (A.flatten().index((m*UOp.const(dtypes.weakint, K)+k))*
-         B.flatten().index((k*UOp.const(dtypes.weakint, N)+n))).cast(dtypes.float32)
+  mul = (A.flatten().index((m*UOp.const(K)+k))*
+         B.flatten().index((k*UOp.const(N)+n))).cast(dtypes.float32)
  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype)
-  store = C.flatten().index((m*UOp.const(dtypes.weakint, N)+n)).store(red).end(m, n)
+  store = C.flatten().index((m*UOp.const(N)+n)).store(red).end(m, n)
  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))

 # ** bf16 A @ B.T kernel in C
@@ -234,7 +272,7 @@ def hk_bf16_atb_gemm(a:Tensor, b:Tensor) -> Tensor:
    elif b.uop.axis == 2: inv, out_axis = Tensor.invalids(1, M, N // ndev, dtype=a.dtype, device=a.device), 2
    elif a.uop.axis == 2: inv, out_axis = Tensor.invalids(1, M // ndev, N, dtype=a.dtype, device=a.device), 1
    else: inv, out_axis, reduce_out = Tensor.invalids(1, M, N, dtype=a.dtype, device=a.device), 0, True
-    out = Tensor(inv.uop.multi(out_axis), device=a.device)
+    out = Tensor(inv.uop.unshard(out_axis), device=a.device)
    dname = a.device[0]
  else:
    out = Tensor.invalids(1, M, N, dtype=a.dtype, device=a.device)
@@ -244,7 +282,6 @@ def hk_bf16_atb_gemm(a:Tensor, b:Tensor) -> Tensor:
  if reduce_out: out = out.sum(0)
  return out.squeeze(0) if out.ndim == 3 else out

-
 # ** backward gemm, might use the asm gemm

 def custom_gemm_bw(gradient:UOp, kernel:UOp, n_scales:int=2, has_grad_amax:bool=False, has_w_post:bool=False):
@@ -341,13 +378,30 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F
  if wp is not None: grad_b = grad_b / wp.reshape(-1, 1)
  return (None, grad_a.uop, grad_b.uop) + tuple(None for _ in inputs[3:])

+# ** mxfp4 gemm backward
+
+def custom_mxfp4_gemm_bw(gradient:UOp, kernel:UOp):
+  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)
+  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

 def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=None, grad_amax_state:Tensor|None=None,
             next_grad_amax_state:Tensor|None=None,
             w_post_scale:Tensor|None=None, mx:bool=False, mx_scales:tuple|None=None, mx_w_stored:bool=False, g_amax:Tensor|None=None,
-             a_pretranspose:Tensor|None=None) -> Tensor:
+             a_pretranspose:Tensor|None=None, mxfp4:bool=False) -> Tensor:
  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
+  if mxfp4:
+    assert not mx and mx_scales is None, "mxfp4 owns quantization; mx/mx_scales are for mxfp8"
+    assert a.dtype == dtypes.bfloat16, f"cannot quantize {a.dtype} to mxfp4"
  counters["used"] += 1
  unfold_batch = a.ndim == 3 and isinstance(a.device, tuple) and a.uop.axis == 2 and b.uop.axis == 0
  if unfold_batch:
@@ -355,7 +409,7 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
    a = a.reshape(a.shape[0]*a.shape[1], a.shape[2])
  squeeze = a.ndim == 2
  if squeeze: a = a.unsqueeze(0)
-  out_dtype = dtypes.bfloat16 if a.dtype == FP8_DTYPE else a.dtype
+  out_dtype = dtypes.bfloat16 if a.dtype == FP8_DTYPE or mxfp4 else a.dtype

  batch, M, K = a.shape
  N = b.shape[1]
@@ -366,11 +420,11 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N

  if is_multi:
    if n_sharded:
-      out = Tensor(Tensor.invalids(batch, M, N//len(a.device), dtype=out_dtype, device=a.device).uop.multi(2), device=a.device)
+      out = Tensor(Tensor.invalids(batch, M, N//len(a.device), dtype=out_dtype, device=a.device).uop.unshard(2), device=a.device)
    elif m_sharded:
-      out = Tensor(Tensor.invalids(batch, M, N, dtype=out_dtype, device=a.device).uop.multi(1), device=a.device)
+      out = Tensor(Tensor.invalids(batch, M, N, dtype=out_dtype, device=a.device).uop.unshard(1), device=a.device)
    else:
-      out = Tensor(Tensor.invalids(batch//len(a.device) if a.uop.axis==0 else batch, M, N, dtype=out_dtype, device=a.device).uop.multi(0),
+      out = Tensor(Tensor.invalids(batch//len(a.device) if a.uop.axis==0 else batch, M, N, dtype=out_dtype, device=a.device).uop.unshard(0),
                   device=a.device)
  else:
    out = Tensor.invalids(batch, M, N, dtype=out_dtype, device=a.device)
@@ -378,7 +432,15 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
  renderer = Device[dname:=(a.device[0] if is_multi else a.device)].renderer
  dname, arch = dname.split(":")[0], renderer.target.arch
  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
-    if mx:
+    if mxfp4:
+      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
+      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:
        a_si, a_e8, b_si, b_e8 = mx_scales
@@ -79,7 +79,7 @@ def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
  # this is the big accumulator
  acc = UOp.placeholder((BLOCK_N//TC_N, BLOCK_M//TC_M//WARPGROUP_SIZE), dtypes.float, 0, AddrSpace.REG)
  assert acc.size*WARP_SIZE*WARPGROUP_SIZE*4 == BLOCK_M*BLOCK_N
-  acc = acc[init_l:=UOp.range(acc.size, 500)].set(UOp.const(dtypes.float, (0.0,)*4), end=init_l)
+  acc = acc[init_l:=UOp.range(acc.size, 500)].set(UOp.const((0.0,)*4, dtypes.float), end=init_l)

  # create locals (note A is permuted, and the stride is changed to avoid bank conflicts)
  def make_locals(slot) -> tuple[UOp, UOp]:
@@ -29,7 +29,7 @@ TID_SIZE = WARPGROUP_SIZE*WARP_SIZE

 def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=()):
  assert dest.shape == src.shape
-  rngs = [UOp.range(s, rng+i, AxisType.UPCAST if i in upcast else AxisType.LOOP) for i,s in enumerate(src.shape)]
+  rngs = [UOp.range(s, rng+i, AxisType.UPCAST if i in upcast else AxisType.WEAK) for i,s in enumerate(src.shape)]
  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
  return dest.after(copy) if set else copy

@@ -55,7 +55,7 @@ def grouped_mx_wgrad(g:Tensor, xg:Tensor, expert_off:Tensor, n_experts:int) -> T
  dname = (g.device[0] if isinstance(g.device, tuple) else g.device).split(":")[0]
  is_multi = isinstance(g.device, tuple)
  inv = Tensor.invalids(1, n_experts * N, K, dtype=dtypes.bfloat16, device=g.device)
-  out = Tensor(inv.uop.multi(0), device=g.device) if is_multi else inv
+  out = Tensor(inv.uop.unshard(0), device=g.device) if is_multi else inv
  out = Tensor.custom_kernel(out, gT, xT, g_si, x_si, expert_off,
                             fxn=functools.partial(custom_hk_grouped_mxfp8_wgrad, dname=dname, n_experts=n_experts))[0]
  out = out.sum(0) if is_multi else out.squeeze(0)
@@ -103,7 +103,7 @@ def grouped_mx_gemm(x:Tensor, w:Tensor|tuple[Tensor, Tensor], expert_off:Tensor)
  if isinstance(x.device, tuple) and (row_axis := x.uop.axis) is not None:
    ndev = len(x.device)
    out = Tensor(Tensor.invalids(*(s // ndev if i == row_axis else s for i, s in enumerate(out_shape)),
-                                 dtype=dtypes.bfloat16, device=x.device).uop.multi(row_axis), device=x.device)
+                                 dtype=dtypes.bfloat16, device=x.device).uop.unshard(row_axis), device=x.device)
  else:
    out = Tensor.invalids(*out_shape, dtype=dtypes.bfloat16, device=x.device)
  return Tensor.custom_kernel(out, x_q, w_q, x_si, w_si, xe_in, w_e8, expert_off,
@@ -0,0 +1,131 @@
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+
+BLOCK_ROW = 256
+
+def _sharded_invalids(shape:tuple[int, ...], dtype, device) -> Tensor:
+  if isinstance(device, tuple):
+    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:
+  return "__hip_atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);" if device == "AMD" \
+    else "__atomic_fetch_add({0}, {1}, __ATOMIC_RELAXED);"
+
+def _blk_for(D:int) -> int:
+  blk = 64
+  while D % blk: blk //= 2
+  return blk
+
+def _kv_ranges(G, N, D, BLK):
+  g = UOp.range(G, 0)
+  m = UOp.range(N, 1)
+  jo = UOp.range(D // BLK, 2)
+  ji = UOp.range(BLK, 3, AxisType.LOCAL)
+  return g, m, jo * BLK + ji, jo, ji
+
+def _ggather_fwd_kernel(out:UOp, table:UOp, idx:UOp) -> UOp:
+  G, M, D = out.shape
+  g, m, j, jo, ji = _kv_ranges(G, M, D, _blk_for(D))
+  row = idx.index(g, m).cast(dtypes.weakint)
+  val = table.index(g, row, j).load()
+  return out.index(g, m, j).store(val).end(g, m, jo, ji).sink(
+    arg=KernelInfo(name=f"ggather_fwd_{M}_{D}", opts_to_apply=()))
+
+def _ggather_zero_kernel(out:UOp) -> UOp:
+  i = UOp.range(out.numel(), 0)
+  return out.flatten().index(i).store(UOp.const(0.0, out.dtype)).end(i).sink(arg=KernelInfo(name="ggather_zero"))
+
+def _sharded_zeros(shape:tuple[int, ...], dtype, device) -> Tensor:
+  return Tensor.custom_kernel(_sharded_invalids(shape, dtype, device), fxn=_ggather_zero_kernel)[0]
+
+def _ggather_bwd(gradient:UOp, kernel:UOp) -> tuple:
+  _, table_u, idx_u = kernel.src[1:4]
+  dev = table_u.device
+  device = (dev[0] if isinstance(dev, tuple) else dev).split(":")[0]
+  G, R, D = table_u.shape
+  gt = _sharded_zeros((G, R, D), dtypes.float32, dev)
+  go = Tensor(gradient, device=dev)
+  atomic_str = _atomic_add(device)
+  def _bwd_kernel(gtab:UOp, gout:UOp, idx:UOp) -> UOp:
+    Gk, M, Dk = gout.shape
+    g, m, j, jo, ji = _kv_ranges(Gk, M, Dk, _blk_for(Dk))
+    row = idx.index(g, m).cast(dtypes.weakint)
+    val = gout.index(g, m, j).load().cast(dtypes.float32)
+    atomic = UOp(Ops.CUSTOM, dtypes.void, (gtab.index(g, row, j), val), arg=atomic_str)
+    return atomic.end(g, m, jo, ji).sink(arg=KernelInfo(name=f"ggather_bwd_{M}_{Dk}", opts_to_apply=()))
+  grad_table = Tensor.custom_kernel(gt, go, Tensor(idx_u, device=dev), fxn=_bwd_kernel)[0]
+  return (None, grad_table.cast(table_u.dtype).uop, None)
+
+def grouped_gather_rows(table:Tensor, idx:Tensor, n_groups:int) -> Tensor:
+  G, R, D = table.shape
+  M = idx.shape[1]
+  out = _sharded_invalids((G, M, D), table.dtype, table.device)
+  return Tensor.custom_kernel(out, table, idx, fxn=_ggather_fwd_kernel, grad_fxn=_ggather_bwd)[0]
+
+def _gscatter_fwd_kernel(out:UOp, src:UOp, idx:UOp) -> UOp:
+  G, M, D = out.shape
+  k = idx.shape[1] // src.shape[1]
+  g, m, j, jo, ji = _kv_ranges(G, idx.shape[1], D, _blk_for(D))
+  row = idx.index(g, m).cast(dtypes.weakint)
+  val = src.index(g, (m // k).cast(dtypes.weakint), j).load()
+  return out.index(g, row, j).store(val).end(g, m, jo, ji).sink(
+    arg=KernelInfo(name=f"gscatter_fwd_{idx.shape[1]}_{D}", opts_to_apply=()))
+
+def _gscatter_bwd(gradient:UOp, kernel:UOp) -> tuple:
+  _, src_u, idx_u = kernel.src[1:4]
+  dev = src_u.device
+  G, T_l, D = src_u.shape
+  k = idx_u.shape[1] // T_l
+  sel = grouped_gather_rows(Tensor(gradient, device=dev), Tensor(idx_u, device=dev), G)
+  return (None, sel.reshape(G, T_l, k, D).sum(2).cast(src_u.dtype).uop, None)
+
+def grouped_scatter_rows(src:Tensor, idx:Tensor, m_l:int) -> Tensor:
+  G, T_l, D = src.shape
+  zero = _sharded_zeros((G, m_l, D), src.dtype, src.device)
+  return Tensor.custom_kernel(zero, src, idx, fxn=_gscatter_fwd_kernel, grad_fxn=_gscatter_bwd)[0]
+
+def m_max_for(t_local:int, experts_per_tok:int, n_experts:int) -> int:
+  return (-(-t_local * experts_per_tok // BLOCK_ROW) + n_experts) * BLOCK_ROW
+
+class Routing:
+  def __init__(self, weights:Tensor, dest_row:Tensor, off:Tensor, m_l:int, n_groups:int, t_local:int):
+    self.weights, self.dest_row = weights, dest_row
+    self.off = off
+    self.m_l, self.n_groups, self.t_local = m_l, n_groups, t_local
+
+  @property
+  def rows_e(self) -> Tensor:
+    G, E = self.off.shape[0], self.off.shape[1] - 1
+    tr = Tensor.arange(self.m_l // BLOCK_ROW, dtype=dtypes.int32).reshape(1, -1, 1) * BLOCK_ROW
+    tr = tr.shard(self.off.device) if isinstance(self.off.device, tuple) else tr.to(self.off.device)
+    tile_e = ((tr >= self.off[:, :E].reshape(G, 1, E)).sum(-1) - 1).cast(dtypes.int32)
+    return tile_e.reshape(-1, 1).expand(-1, BLOCK_ROW).reshape(-1)
+
+def n_groups_of(t:Tensor) -> int:
+  return len(t.device) if isinstance(t.device, tuple) else 1
+
+def route(logits:Tensor, experts_per_tok:int, n_experts:int) -> Routing:
+  T, E = logits.shape
+  k, G = experts_per_tok, n_groups_of(logits)
+  assert T % G == 0, f"tokens {T} must split across {G} devices"
+  T_l, m_l = T // G, m_max_for(T // G, k, n_experts)
+
+  topv, topi = logits.reshape(G, T_l, E).topk(k)
+  weights = topv.softmax(-1)
+  m = topi.reshape(G, T_l * k).cast(dtypes.int32).one_hot(E).cast(dtypes.int32)
+
+  pad = ((m.sum(1) + (BLOCK_ROW - 1)) // BLOCK_ROW) * BLOCK_ROW
+  off = pad.cumsum(1).pad(((0, 0), (1, 0)))
+  dest_row = ((m.cumsum(1) + off[:, :E].reshape(G, 1, E)) * m).sum(-1).sub(1).cast(dtypes.int32)
+  return Routing(weights, dest_row, off, m_l, G, T_l)
+
+def dispatch(x:Tensor, r:Routing) -> Tensor:
+  G, D = r.n_groups, x.shape[-1]
+  return grouped_scatter_rows(x.reshape(G, r.t_local, D), r.dest_row, r.m_l).reshape(G * r.m_l, D)
+
+def combine(y:Tensor, r:Routing, n_tokens:int, experts_per_tok:int) -> Tensor:
+  G, D, k = r.n_groups, y.shape[-1], experts_per_tok
+  sel = grouped_gather_rows(y.reshape(G, r.m_l, D), r.dest_row, G).reshape(G, r.t_local, k, D)
+  return (sel * r.weights.reshape(G, r.t_local, k, 1).cast(sel.dtype)).sum(2).reshape(n_tokens, D).cast(y.dtype)
@@ -0,0 +1,139 @@
+"""
+tilelang-style matmul_relu written with tinygrad UOp APIs.
+
+Reference tilelang kernel:
+  @tilelang.jit
+  def matmul_relu(A, B, block_M=64, block_N=64, block_K=64,
+                  dtype=T.float16, accum_dtype=T.float32):
+    M, N, K = T.const('M, N, K')
+    C = T.empty([M, N], dtype)
+    with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=128) as (bx, by):
+      A_shared = T.alloc_shared((block_M, block_K), dtype)
+      B_shared = T.alloc_shared((block_K, block_N), dtype)
+      C_local  = T.alloc_fragment((block_M, block_N), accum_dtype)
+      T.clear(C_local)
+      for ko in T.Pipelined(T.ceildiv(K, block_K), num_stages=3):
+        T.copy(A[by * block_M, ko * block_K], A_shared)
+        T.copy(B[ko * block_K, bx * block_N], B_shared)
+        T.gemm(A_shared, B_shared, C_local)
+      for i, j in T.Parallel(block_M, block_N):
+        C_local[i, j] = T.max(C_local[i, j], 0)
+      T.copy(C_local, C[by * block_M, bx * block_N])
+    return C
+"""
+
+from tinygrad.dtype import dtypes, AddrSpace, DType
+from tinygrad.uop.ops import UOp, Ops, AxisType, KernelInfo
+from tinygrad.helpers import cdiv, getenv
+from tinygrad.tensor import Tensor
+
+# ---------------------------------------------------------------------------
+# tilelang builtins, expressed with tinygrad UOp APIs
+# ---------------------------------------------------------------------------
+
+def alloc_shared(shape:tuple[int, ...], dtype:DType, slot:int) -> UOp:
+  """T.alloc_shared: one LOCAL buffer shared by all threads in the block."""
+  return UOp.placeholder(tuple(shape), dtype, slot, AddrSpace.LOCAL)
+
+def alloc_fragment(shape:tuple[int, ...], dtype:DType, slot:int, axes:tuple[int, ...], rngs:tuple[UOp, ...]) -> UOp:
+  """T.alloc_fragment: per-thread REG fragment + UNSHARD over the LOCAL thread grid."""
+  assert len(axes) == len(rngs)
+  assert all(tnum.op is Ops.RANGE and tnum.arg[-1] is AxisType.LOCAL for tnum in rngs), "fragments shard over LOCAL ranges"
+  by_axis = dict(zip(axes, rngs))
+  shard_shape = tuple(s // (int(by_axis[i].vmax)+1) if i in by_axis else s for i, s in enumerate(shape))
+  fragment = UOp.placeholder(shard_shape, dtype, slot, AddrSpace.REG)
+  return fragment.unshard(axes, rngs)
+
+# ---------------------------------------------------------------------------
+# GEMM kernel: C = relu(A @ B), float inputs (fp16 or fp32), fp32 fragment accumulator, no WMMA
+# ---------------------------------------------------------------------------
+
+# 64x64 output tile per block, 128 threads as an 8x16 grid; each thread owns an 8x4 fragment sub-tile
+# (the 2-D per-thread layout tilelang infers for this GEMM). The 4 contiguous columns (TN=4) are what
+# let codegen vectorize loads/stores to float4, matching tilelang's lowering exactly.
+BLOCK_M = BLOCK_N = BLOCK_K = 64
+TY = 8
+TX = 16
+THREADS = TY * TX
+TM = BLOCK_M // TY   # fragment rows per thread  (8)
+TN = BLOCK_N // TX   # fragment columns per thread (4)
+
+def matmul_relu_kernel(c:UOp, a:UOp, b:UOp) -> UOp:
+  """C[M, N] = relu(A[M, K] @ B[K, N]) -- one 64x64 tile per block, locals + a 2-D fragment."""
+  M, K = a.shape
+  K2, N = b.shape
+  assert K == K2 and a.dtype == b.dtype == c.dtype and not dtypes.is_int(a.dtype)
+  assert not (K % BLOCK_K or M % BLOCK_M or N % BLOCK_N), "test sizes must be multiples of the block sizes"
+
+  # with T.Kernel(T.ceildiv(N, BLOCK_N), T.ceildiv(M, BLOCK_M), threads=128) as (bx, by):
+  bx = UOp.range(cdiv(N, BLOCK_N), 0, AxisType.GLOBAL)
+  by = UOp.range(cdiv(M, BLOCK_M), 1, AxisType.GLOBAL)
+
+  # 16*8 threads = 128 threads
+  tx = UOp.range(TX, 2, AxisType.LOCAL)
+  ty = UOp.range(TY, 3, AxisType.LOCAL)
+
+  # shared + fragment (regs)
+  A_shared = alloc_shared((BLOCK_M, BLOCK_K), a.dtype, 0)
+  B_shared = alloc_shared((BLOCK_K, BLOCK_N), b.dtype, 1)
+  C_local = alloc_fragment((TM, TY, TX, TN), dtypes.float32, 0, (1, 2), (ty, tx))
+
+  # zero out the regs to start. this is expanded by the devectorizer
+  C_local = C_local.after(C_local.store(0.0))
+
+  # for ko in T.Pipelined(T.ceildiv(K, BLOCK_K), num_stages=3):
+  ko = UOp.range(cdiv(K, BLOCK_K), 6, AxisType.LOOP)
+
+  # index the outer matrices
+  a = a.rearrange("(m bm) (k bk) -> m k bm bk", bm=BLOCK_M, bk=BLOCK_K)[by, ko]
+  b = b.rearrange("(k bk) (n bn) -> k n bk bn", bk=BLOCK_K, bn=BLOCK_N)[ko, bx]
+  c = c.rearrange("(m bm) (n bn) -> m n bm bn", bm=BLOCK_M, bn=BLOCK_N)[by, bx]
+
+  # T.copy: A_shared <- a, B_shared <- b
+  def with_threads(x:UOp): return x.rearrange("(tm ty) (tx tn) -> ty tx tm tn", tm=TM, tn=TN)[ty, tx]
+  A_shared = A_shared.after(with_threads(A_shared).store(with_threads(a)))
+  B_shared = B_shared.after(with_threads(B_shared).store(with_threads(b)))
+
+  # T.gemm(A_shared, B_shared, C_local), no WMMA
+  kk = UOp.range(BLOCK_K, 11, AxisType.LOOP)
+  ir = UOp.range(TM, 12, AxisType.LOOP)
+  jj = UOp.range(TN, 13, AxisType.UPCAST)
+  acc = C_local.after(kk)[ir, ty, tx, jj] + A_shared[ir*TM + ty, kk].cast(dtypes.float32) * B_shared[kk, tx*TN + jj].cast(dtypes.float32)
+  # closing the ko loop here too; codegen adds the barrier so no thread overwrites the tiles while others still read them
+  C_local = C_local[ir, ty, tx, jj].set(acc, end=(kk, ir, jj, ko))
+
+  # c <- C_local (with relu and cast): every thread stores its shard's sub-view of the output tile
+  c_st = c.reshape(C_local.shape).store(C_local.relu().cast(c.dtype))
+
+  # close the locals and globals
+  return c_st.end(tx, ty, bx, by).sink(arg=KernelInfo(name="matmul_relu", opts_to_apply=()))
+
+# ---------------------------------------------------------------------------
+# python wrapper: same signature as the tilelang function
+# ---------------------------------------------------------------------------
+
+def matmul_relu(a:Tensor, b:Tensor) -> Tensor:
+  """C = relu(A @ B), fp16 in/out with an fp32 fragment accumulator."""
+  c = Tensor.empty(a.shape[0], b.shape[1], dtype=a.dtype, device=a.device)
+  return c.custom_kernel(a, b, fxn=matmul_relu_kernel)[0]
+
+# ---------------------------------------------------------------------------
+# test
+# ---------------------------------------------------------------------------
+
+if __name__ == "__main__":
+  from tinygrad import Device
+  assert Device[Device.DEFAULT].renderer.has_local, "this GPU-style kernel needs a backend with local memory (LOCAL ranges + barriers)"
+  M = K = N = getenv("N", 256)  # 4x4 grid of 64x64 tiles, 4 K chunks
+  dtype_in = dtypes.half if getenv("HALF") else dtypes.float
+
+  a = Tensor.randn(M, K, dtype=dtype_in).contiguous()
+  b = Tensor.randn(K, N, dtype=dtype_in).contiguous()
+  ref = (a @ b).relu().realize()
+
+  for _ in range(10):
+    out = matmul_relu(a, b).realize()
+
+  import numpy as np
+  np.testing.assert_allclose(out.numpy(), ref.numpy(), atol=1e-1, rtol=1e-2)
+  print("matmul_relu passed!")
@@ -1,544 +0,0 @@
-from __future__ import annotations
-from typing import cast, Callable, TypeVar, Generic, Any
-import struct, functools, time, collections, itertools
-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.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer
-from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, track_rewrites, GroupOp
-from tinygrad.uop.symbolic import symbolic
-from tinygrad.dtype import dtypes, truncate
-from tinygrad.runtime.support.hcq import MMIOInterface
-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, pm_flatten_linear
-from tinygrad.engine.jit import DepsTracker
-
-# *****************
-# 0. helpers
-
-HCQDeviceType = TypeVar('HCQDeviceType', bound='HCQ2Compiled')
-
-HCQ_RUNTIME_DEV = ContextVar("HCQ_RUNTIME_DEV", "CPU")
-
-HCQ_DEVS = frozenset(("AMD",))
-HCQ_P2P_DEVS = HCQ_DEVS | frozenset(("CPU",))
-HCQ_CACHE_TAGS = frozenset(("program", "systems", "template"))
-
-@dataclass(frozen=True)
-class HCQInfo:
-  name:str
-  estimates:Estimates
-  device:tuple[str, ...]
-  queue:str
-
-  input_idxs:tuple[int, ...] = () # indexes into input_uops used by this call
-  inputs:int|None = None
-
-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):
-  return tuple(x for s in u.src for x in unwrap_mstack(s)) if u.op is Ops.MSTACK else (unwrap_mstack(u.src[0]) if u.op in {Ops.MSELECT, Ops.SLICE} else (u,))
-
-def make_patch(buf:UOp, off:sint, val:UOp) -> UOp:
-  return buf.index(UOp.const(dtypes.int, off // buf.dtype.itemsize)).store(val.simplify().cast(buf.dtype))
-
-def make_binary_patch(buf:UOp, blob:bytes) -> UOp:
-  data = UOp(Ops.BINARY, src=(), arg=blob).bitcast(buf.dtype)
-  r = UOp.range(len(blob) // buf.dtype.itemsize, 0, dtype=dtypes.int, src=(buf, data))
-  return buf.index(r).store(data.index(r).load()).end(r)
-
-def make_cmdbuf(lin, devs):
-  blob, patches = b'', []
-  for s in (s for ins in lin.src for s in ins.src):
-    if (ssimp:=s.simplify()).op is not Ops.CONST: patches.append((len(blob), ssimp))
-    blob += struct.pack(f'<{ssimp.dtype.fmt}', ssimp.arg if ssimp.op is Ops.CONST else 0x0)
-  cmdbuf = UOp.placeholder((len(blob) // 4,), dtypes.uint32, next(UOp.unique_num), device=devs).rtag("cmdbuf")
-  return cmdbuf.after(make_binary_patch(cmdbuf, blob), *[make_patch(cmdbuf, off, s) for off, s in patches])
-
-def make_signal(devs, queue="COMPUTE:0", sentinel=False):
-  return UOp.placeholder((1,), dtypes.uint64, 0, device=devs).rtag("sentinel_signal" if sentinel else f"{queue}_timeline_signal")
-def make_signal_value(devs, queue="COMPUTE:0"):
-  return UOp.placeholder((1,), dtypes.uint64, 0, device=devs).rtag(f"{queue}_timeline_value")
-
-def make_submit(*cmds, devs:str|tuple[str, ...], queue:str) -> UOp:
-  return UOp.custom_function("submit_cmdbuf", UOp(Ops.LINEAR, src=tuple(cmds), arg=(to_tuple(devs), queue)))
-def get_submit(ast:UOp) -> UOp: return next(u for u in ast.toposort() if u.op is Ops.CUSTOM_FUNCTION and u.arg == "submit_cmdbuf")
-
-def encode_kernargs_clike(call:UOp, prg:UOp, devs:str|tuple[str, ...]) -> UOp:
-  data, info = prg.arg
-  buf = UOp.placeholder((data.kernargs_alloc_size // 4,), dtypes.uint32, next(UOp.unique_num), device=devs).rtag("kernargs")
-  words = [w for gi in info.globals for w in data64_le(get_call_arg_uops(call)[gi].getaddr(devs))] + list(info.vars)
-  return buf.after(*[make_patch(buf, i * 4, w) for i, w in enumerate(words)])
-
-# *****************
-# 0.1. prep: replace buffers with params
-
-def replace_call_buffers(ctx:list[UOp], call:UOp) -> UOp|None:
-  ctx += [s for s in call.src[1:] if s not in ctx and s.op not in (Ops.PARAM, Ops.BIND)]
-  return call.replace(src=call.src[:1] + tuple(s if s.op in (Ops.PARAM, Ops.BIND) else s.param_like(ctx.index(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 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)])
-
-# *****************
-# 2. deps
-
-class HCQDepsTracker(DepsTracker):
-  @staticmethod
-  def _key(buf:Any) -> tuple[Any, int, int]:
-    return (buf.arg.slot, 0, buf.max_numel() * buf.dtype.itemsize) if isinstance(buf, UOp) else DepsTracker._key(buf)
-
-def _get_call_bufs_by_lane(call:UOp, devices:tuple[str, ...]) -> list[list[Any]]:
-  refs = get_call_arg_uops(call)
-  return [[b if b.op is Ops.PARAM else mb.bufs[lane] if isinstance(mb:=b.buffer, MultiBuffer) else mb for b in refs] for lane in range(len(devices))]
-
-def _get_deps(ctx:DepsTracker, bufs_by_lane:list[list[Any]], write, key:tuple[tuple[str, ...], str, int]) -> list[tuple[tuple, int, int]]:
-  dep_lanes:list[tuple[tuple, int, int]] = []
-  for lane, bufs in enumerate(bufs_by_lane):
-    dep_lanes += [(dep, dlane, lane) for dep, dlane in ctx.access_resources(bufs, write if write is not None else range(len(bufs)), (key, lane))]
-  return dep_lanes
-
-def _build_wait_cmds(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"):
-    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
-
-  waits = []
-  for (ddevs, dqueue, dtag), lanes in deps.items():
-    sig = UOp.mstack(*[make_signal(d if dl is None else ddevs[dl], queue=dqueue, sentinel=dl is None) for dl, d in zip(lanes, devices)])
-    val = UOp.mstack(*[make_signal_value(d if dl is None else ddevs[dl], queue=dqueue) for dl, d in zip(lanes, devices)])
-    waits.append(UOp(Ops.INS, arg="wait", src=(sig, val.index(UOp.const(dtypes.int, 0)) + dtag)))
-  return waits, {dtag for _, _, dtag in deps}
-
-def _build_finalizers(batch:list[tuple[UOp, tuple[str, ...]]], batch_info:list[tuple[tuple[str, ...], str]],
-                      tracker:HCQDepsTracker) -> tuple[list[UOp], set[int]]:
-  # collect all buffers which belong to devices
-  dev_bufs:dict[str, dict[int, Any]] = collections.defaultdict(dict)
-  for call, devices in batch:
-    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
-
-  zero, n, finalizers, waited = UOp.const(dtypes.int, 0), len(batch_info), [], set()
-  for _, devgroup in itertools.groupby(sorted(dedup([d for devs, _ in batch_info for d in devs])), key=lambda d: d.split(":")[0]):
-    devs = tuple(devgroup)
-
-    # to finalize the batch, sync all accesses from other devices to buffers that belong to this device
-    fin_deps = [dl for dl in _get_deps(tracker, [list(dev_bufs[d].values()) for d in devs], None, key=(devs, "COMPUTE:0", n)) if dl[0][2] < n]
-    waits, cur_waited = _build_wait_cmds(fin_deps, devs, "COMPUTE:0")
-    waited |= cur_waited
-
-    # wait the syncs, store the device epoch; value bumps are a separate call: no lane may bump until every lane has patched its waits
-    store = UOp(Ops.INS, arg="store", src=(make_signal(devs), (tl:=make_signal_value(devs)).index(zero) + n))
-    submit = make_submit(*waits, store, devs=devs, queue="COMPUTE:0")
-    upd = [(tl, n + 1)] + [(make_signal_value(devs, queue=qn), n)
-                           for qn in dedup([qn for bdevs, qn in batch_info if set(bdevs) & set(devs)]) if qn != "COMPUTE:0"]
-    bump = UOp.barrier(*[s.index(zero, dtype=s.dtype).store(s.index(zero) + inc) for s, inc in upd])
-    finalizers += [UOp.custom_function("hcq", b.sink()).call(aux=HCQInfo("hcq_finalizer", Estimates(), devs, "COMPUTE:0")) for b in (submit, bump)]
-  return finalizers, waited
-
-def _finalize_batch(batch:list[tuple[UOp, tuple[str, ...]]]) -> 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
-  waited:set[int] = set()
-  deps_tracker = HCQDepsTracker()
-  call_waits:list[list[UOp]] = []
-  for tag, ((call, _), (devices, queue)) in enumerate(zip(batch, batch_info)):
-    deps = _get_deps(deps_tracker, _get_call_bufs_by_lane(call, devices), get_call_outs_ins(call)[0], key=(devices, queue, tag))
-    cmds, cur_waited = _build_wait_cmds(deps, devices, queue)
-    call_waits.append(cmds)
-    waited |= cur_waited
-
-  # build finalizers
-  finalizers, finalizer_waited = _build_finalizers(batch, batch_info, deps_tracker)
-  waited |= finalizer_waited
-
-  src = []
-  for tag, ((call, _), (devices, queue), q) in enumerate(zip(batch, batch_info, call_waits)):
-    # first queue use, sync prior device work with main signal
-    if batch_info.index((devices, queue)) == tag:
-      q = [UOp(Ops.INS, arg="barrier", src=()), UOp(Ops.INS, arg="wait", src=(make_signal(devices), make_signal_value(devices).index(0) - 1))] + q
-
-    # and make hcq call
-    info = HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), devices, queue)
-    q += [call.replace(arg=replace(call.arg, aux=info))]
-
-    # signal queue timeline if someone waits for us
-    if tag in waited: q += [UOp(Ops.INS, arg="store", src=(make_signal(devices, queue), make_signal_value(devices, queue).index(0) + tag))]
-    src.append(UOp.custom_function("hcq", make_submit(*q, devs=devices, queue=queue).sink()).call(name="hcq", aux=info))
-  return src + finalizers
-
-def sched_hcq_batches(l:UOp) -> 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)))
-pm_sched_hcq_batches = PatternMatcher([(UPat(Ops.LINEAR, name="l"), sched_hcq_batches)])
-
-# *****************
-# 3. merge into queues
-
-def _merged_hcq_call(calls:list[UOp]) -> UOp: # TODO: simplify?
-  if len(calls) == 1: return calls[0]
-  info = replace(calls[0].arg.aux, name=f"submit {calls[0].arg.aux.queue} ({len(calls)})",
-                 estimates=sum((c.arg.aux.estimates for c in calls), start=Estimates()))
-  cmds = [cmd for c in calls for cmd in get_submit(c).src[0].src]
-  return UOp.custom_function("hcq", make_submit(*cmds, devs=info.device, queue=info.queue).sink()).call(name="hcq", aux=info)
-
-def merge_queues(linear:UOp) -> UOp:
-  new_src:list[UOp] = []
-  opened_qs:dict[tuple[tuple[str, ...], str], list[UOp]] = {} # (devs, queue) -> list of hcq calls, kept in submit order
-  limits = collections.defaultdict(lambda: JIT_BATCH_SIZE.value)
-
-  for call in linear.src:
-    if not isinstance(info:=call.arg.aux, HCQInfo) or info.name == "hcq_finalizer": # non-hcq call or finalizer: close all open queues
-      new_src += [_merged_hcq_call(opened_qs.pop(k)) for k in list(opened_qs)] + [call]
-      continue
-
-    if (old:=opened_qs.pop(key:=(info.device, info.queue), None)) is not None:
-      if limits[key] and len(old) >= limits[key]: new_src, old, limits[key] = new_src + [_merged_hcq_call(old)], [], limits[key] * 2
-      new_rec = old + [call]
-    else:
-      # no such queue opened: close every open submit on this queue that shares a device, so submit order is kept
-      closing = [k for k in opened_qs if k[1] == info.queue and set(k[0]) & set(info.device)]
-      new_src += [_merged_hcq_call(opened_qs.pop(k)) for k in closing]
-      new_rec = [call]
-    opened_qs[(info.device, info.queue)] = new_rec
-  return linear.replace(src=tuple(new_src + [_merged_hcq_call(c) for c in opened_qs.values()]))
-pm_merge_queues = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), merge_queues)])
-
-# *****************
-# 4.2. hcq lowering: ops to ir
-
-def encode_cmdbuf(submit:UOp, lin:UOp) -> UOp|None:
-  if (pm:=Device.get_class(lin.arg[0][0]).pm_lower) is None: return None
-  return graph_rewrite(submit, pm, name=f"encode {lin.arg[0]}", enter_calls=True)
-pm_encode_cmdbufs = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="lin"),), name="submit"), encode_cmdbuf)])
-
-# *****************
-
-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)]
-  addressed_bufs = [b for g in val.toposort() if g.op is Ops.GETADDR for b in unwrap_mstack(g.buf_uop)]
-
-  # addr of input params is not known at link time
-  return not val.variables() and not runtime_reads and all(b.op is not Ops.PARAM or b.tag is not None for b in addressed_bufs)
-
-def is_link_patch(p:UOp, jit:bool) -> bool:
-  store = p.src[0] if (is_binary_patch:=p.op is Ops.END) else p
-  if not jit: return store.buf_uop.tag == "program"
-  return is_binary_patch or (store.op is Ops.STORE and is_value_known_at_link(store.src[1]))
-
-def trim_link_patches(ctx:tuple[bool, list[UOp]], a:UOp) -> UOp|None:
-  links, kept = partition(a.src[1:], lambda p: is_link_patch(p, ctx[0]))
-
-  # keep all patches from the link-time patches' subtrees in the C code
-  afters = [u for u in UOp.sink(*links).toposort() if u.op is Ops.AFTER]
-  ctx[1].extend(UOp.sink(*links).substitute({p: p.src[0] for p in afters}).src)
-  return a.src[0].after(*kept, *[d for p in afters for d in p.src[1:]]) if links else None
-pm_trim_link_patches = PatternMatcher([(UPat(Ops.AFTER, src=(UPat((Ops.PARAM, Ops.MSTACK)),), allow_any_len=True, name="a"), trim_link_patches)])
-
-def split_patches(ctx:bool, call:UOp) -> UOp|None:
-  body = graph_rewrite(call.src[0], pm_trim_link_patches, ctx=(ctx, lt_patches:=[]), name=f"trim link-time patches ({call.arg.aux.name})")
-
-  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()]))
-pm_split_patches = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), split_patches)])
-
-# *****************
-
-def make_addr_table(call:UOp, gaddrs:list[UOp], name:str) -> tuple[dict[UOp, UOp], tuple[UOp, ...]]:
-  bare = {g: g.replace(src=(g.src[0].without_after,)) for g in gaddrs}
-
-  order = sorted(dedup(bare.values()), key=lambda g: ((b:=unwrap_mstack(g.buf_uop)[0]).arg.slot, repr(b.tag)))
-  slots = {g:i for i,g in enumerate(order)}
-  table = UOp.placeholder((len(order),), dtypes.uint64, next(UOp.unique_num), device=call.arg.aux.device).rtag(name)
-
-  reads = {g: table.after(*g.src[0].src[1:] if g.src[0].op is Ops.AFTER else ()).index(UOp.const(dtypes.int, slots[bare[g]])).load() for g in gaddrs}
-  return reads, (table.after(*[make_patch(table, i * table.dtype.itemsize, addr) for addr, i in slots.items()]),) if slots else ()
-
-def make_blob_bufs(call:UOp, blobs:list[UOp]) -> tuple[dict[UOp, UOp], tuple[UOp, ...]]:
-  bufs = {b: UOp.placeholder((b.max_numel(),), b.dtype, next(UOp.unique_num), device=call.arg.aux.device).rtag("template") for b in blobs}
-  return bufs, tuple(buf.after(make_binary_patch(buf, b.src[0].arg)) for b,buf in bufs.items())
-
-def rm_rt_uops(call:UOp) -> UOp|None:
-  if not (rt_uops:=[u for u in call.src[0].toposort() if u.op is Ops.GETADDR or (u.op is Ops.BITCAST and u.src[0].op is Ops.BINARY)]): return None
-  gaddrs, blobs = partition(rt_uops, lambda u: u.op is Ops.GETADDR)
-  inputs, internals = partition(gaddrs, lambda g: all(x.op is Ops.PARAM and x.tag is None for x in unwrap_mstack(g.buf_uop)))
-  runtimes, systems = partition(internals, lambda g: any(x.tag in {"program", "kernargs", "cmdbuf"} for x in unwrap_mstack(g.buf_uop)))
-
-  # exec fills the inputs table with the input addresses every run, so it has no fill patches
-  (reads, _), *tables = [make_addr_table(call, gs, n) for gs,n in ((inputs, "inputs"), (runtimes, "runtime"), (systems, "systems"))] + \
-                        [make_blob_bufs(call, blobs)]
-  reads, fills = reads | {k:v for r,_ in tables for k,v in r.items()}, [f for _,fs in tables for f in fs]
-  return call.replace(src=(call.src[0].substitute(reads), *call.src[1:], *fills),
-                      arg=replace(call.arg, aux=replace(call.arg.aux, input_idxs=tuple(sorted(dedup(g.buf_uop.arg.slot for g in inputs))))))
-pm_rm_rt_uops = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), rm_rt_uops)])
-
-# *****************
-
-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])
-
-  patched, refhold = partition(call.src[1:], lambda x: x.src[0] in args)
-  by_root = {p.src[0]: p for p in patched}
-  c_args = [by_root.get(a, a) for a in args]
-
-  sub = {(b:=u.without_after): UOp.param(i, u.dtype, shape=b.shape, device=u.device) 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))
-  return call.replace(src=(body.substitute(sub), *c_args, *refhold), arg=replace(call.arg, aux=info)) # TODO: call.after(*refhold)?
-pm_replace_params = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), replace_params)])
-
-# *****************
-
-def resolve_getaddr_slice(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
-  return UOp(Ops.GETADDR, dtypes.uint64, src=(base,), arg=g.arg) + UOp.const(dtypes.uint64, bv.src[1].arg * itemsize)
-
-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"),
-   lambda bv,x: x.replace(src=(bv.src[0], x.src[1] + bv.src[1].cast(x.src[1].dtype), *x.src[2:]))),
-])
-
-# *****************
-# 5.3. pack placeholders buffers
-
-def pack_hcq_placeholders(call:UOp) -> UOp|None:
-  bufs = [b for b in call.src[0].toposort() if b.op is Ops.PARAM and b.tag in {"scratch", "kernargs"}]
-  offs, sizes = {}, {}
-  for b in bufs:
-    if b.tag == "scratch": sizes[b.tag] = max(sizes.get(b.tag, 0), b.max_numel())
-    else:
-      offs[b] = round_up(sizes.get(b.tag, 0), 128 // b.dtype.itemsize)
-      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(dtypes.weakint, offs.get(b, 0))), 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)])
-
-# *****************
-# 8. callify hcq programs
-
-pm_callify_hcq = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="hcq", src=(UPat(Ops.SINK),), name="cf"),
-  lambda cf: cf.replace(src=(to_program(cf.src[0].replace(arg=KernelInfo("hcq_submit"), tag=1), Device[HCQ_RUNTIME_DEV.value].renderer),)))])
-
-hcq_compile_cache:dict[tuple[bytes, bool], UOp] = {}
-
-@track_rewrites(lambda linear,input_uops,jit,ret: f"HCQ Compile {pluralize('Kernel', len(ret.src))}")
-def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None, jit=False) -> UOp:
-  if input_uops is not None: linear = graph_rewrite(linear, pm_replace_buffers, ctx=input_uops, walk=True, enter_calls=True, name="replace buffer")
-
-  if (final_linear:=(hcq_compile_cache.get(cache_key:=(linear.key, jit)))) 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_sched_hcq_batches, walk=True, name="schedule hcq batches")
-    linear = linear.substitute({s: p for p, s in back_map.items()}, walk=True, enter_calls=True)
-    linear = graph_rewrite(linear, pm_merge_queues, walk=True, name="merge queues")
-
-    # lowering to hcq ir
-    linear = graph_rewrite(linear, pm_encode_cmdbufs, walk=True, name="encode cmdbufs", enter_calls=True)
-    linear = graph_rewrite(linear, pm_pack_placeholders, walk=True, name="pack placeholders")
-
-    # pie
-    linear = graph_rewrite(linear, pm_split_patches, ctx=jit, walk=True, name="split rt/lt patches")
-    linear = graph_rewrite(linear, pm_early_simplify + symbolic, bottom_up=False, name="simplify packed placeholders", enter_calls=True)
-    linear = graph_rewrite(linear, pm_rm_rt_uops, walk=True, name="replace rt uops")
-    linear = graph_rewrite(linear, pm_replace_params, walk=True, name="replace with args")
-
-    # and compile it
-    final_linear = hcq_compile_cache[cache_key] = graph_rewrite(linear, pm_callify_hcq, name="callify hcq", enter_calls=True)
-
-  return final_linear
-
-# *****************
-# 6. bufferize placeholders: replace placeholders with real buffers.
-
-def bufferize_buf(ctx:bool, buf:UOp) -> UOp|None:
-  if buf.tag is None: return None
-  return UOp.mstack(*(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=(dv, ctx)), HCQ_RUNTIME_DEV.value)
-                      for dev in to_tuple(buf.device)))
-pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, name="buf"), bufferize_buf)])
-
-# *****************
-# 7. resolve patches
-
-def push_stack(op, s): return UOp(Ops.STACK, op.dtype.scalar(),
-  tuple(op.replace(dtype=op.dtype.scalar(), src=tuple(x if y is s else y for y in op.src)) for x in s.src))
-
-def fold_binary(buf:UOp, blob:UOp) -> UOp:
-  for b in (m.bufs if isinstance(m:=buf.buffer, MultiBuffer) else (m,)):
-    b.ensure_allocated().as_memoryview(force_zero_copy=True, no_sync=True).cast('B')[:len(blob.arg)] = blob.arg
-  return UOp(Ops.NOOP)
-
-def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
-  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.arg))
-    b.ensure_allocated().as_memoryview(force_zero_copy=True, no_sync=True).cast('B')[(byte_off:=off.arg*buf.dtype.itemsize):byte_off+len(data)] = data
-  return UOp(Ops.NOOP)
-
-def resolve_getaddr(buf:UOp, g:UOp) -> UOp:
-  assert buf.op in (Ops.BUFFER, Ops.MSTACK, Ops.MSELECT), f"{buf.op}"
-
-  devs, b = g.arg, buf.buffer
-  bufs = tuple(cast(Buffer, x.buffer) for x in buf.src) if buf.op is Ops.MSTACK else tuple(b.bufs if isinstance(b, MultiBuffer) else (b,)*len(devs))
-  assert len(bufs) == len(devs), f"can't resolve {len(bufs)} buffers on {len(devs)} devices"
-  addrs = tuple(UOp.const(dtypes.uint64, x.get_buf(d).va_addr) for x, d in zip(bufs, devs))
-  return addrs[0] if len(addrs) == 1 else UOp(Ops.STACK, dtypes.uint64, addrs)
-
-pm_resolve_patches = PatternMatcher([
-  # multi
-  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
-  (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),
-
-  # getaddr
-  (UPat(Ops.GETADDR, src=(UPat(name="buf"),), name="g"), resolve_getaddr),
-
-  # folders
-  (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.cvar("off"))
-    .store(UPat.any(UPat.cvar("val"), 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}"))])
-
-hcq_link_cache:dict[tuple[bytes, tuple[str, ...]], UOp] = {}
-
-def link_cache_key(a:UOp): return a.key, to_tuple(a.device)
-pm_link_cache = PatternMatcher([(UPat(Ops.AFTER, name="a"), lambda a: hcq_link_cache.get(link_cache_key(a)))])
-
-@track_rewrites(lambda _,jit,ret: f"HCQ Link {pluralize('Kernel', len(ret.src))}")
-def hcq_link(linear:UOp, jit=False) -> UOp:
-  cacheable = {(j,i):a for j,c in enumerate(linear.src) for i,a in enumerate(c.src[1:], 1)
-               if a.op is Ops.AFTER and unwrap_mstack(a.src[0])[0].tag in HCQ_CACHE_TAGS}
-  hits = {a.src[0]:hcq_link_cache[key] for a in cacheable.values() if (key:=link_cache_key(a)) in hcq_link_cache}
-  linear = graph_rewrite(linear, pm_link_cache, name="apply link cache").substitute(hits, walk=True)
-  linear = graph_rewrite(linear, pm_bufferize, ctx=jit, bottom_up=True, walk=True, name="bufferize placeholders")
-  linear = graph_rewrite(linear, pm_resolve_patches + symbolic, bottom_up=False, name="simplify patches")
-  linear = graph_rewrite(linear, pm_assert_no_afters, name="assert no afters")
-  for (j,i),a in cacheable.items(): hcq_link_cache.setdefault(link_cache_key(a), linear.src[j].src[i])
-  return linear
-
-# *****************
-# Device classes
-
-class HCQ2Compiled(Compiled):
-  timestamp_divider: float = 1000.0
-
-  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
-
-    self.pm_bufferize = PatternMatcher([
-      (UPat(Ops.PARAM, tag="sentinel_signal"), lambda ctx: ctx[0].timeline_signal("sentinel", (1 << 64) - 1)),
-      (UPat(Ops.PARAM, name="b"), lambda ctx, b: None if b.tag is None else ctx[0].new_buffer(b, jit=ctx[1]))
-    ])
-
-    super().__init__(device, allocator, compilers, lambda *a, **kw: None, None, arch=arch)
-
-    self.rt_buffer = Buffer(self.device, 64 << 20, dtypes.uint8, options=BufferSpec(uncached=True, cpu_access=True))
-    self.rt_allocator = BumpAllocator(64 << 20, wrap=False)
-
-  def new_buffer(self, b:UOp, jit:bool) -> Buffer:
-    if jit or b.tag in HCQ_CACHE_TAGS:
-      return Buffer(self.device, b.max_numel(), b.dtype, options=BufferSpec(uncached=True, cpu_access=True, nolru=True))
-    return self.rt_buffer.view(b.max_numel(), b.dtype, self.rt_allocator.alloc(b.max_numel() * b.dtype.itemsize, alignment=128))
-
-  @functools.cache
-  def timeline_signal(self, queue:str="COMPUTE:0", init_value:int=0) -> Buffer:
-    buf = Buffer(self.device, 1, dtypes.uint64, options=BufferSpec(host=True, uncached=True, cpu_access=True), preallocate=True)
-    buf.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = init_value
-    return buf
-
-  @functools.cache
-  def timeline_value(self, queue:str="COMPUTE:0", init_value:int=1) -> Buffer:
-    buf = Buffer("CPU", 1, dtypes.uint64, preallocate=True)
-    buf.as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] = init_value
-    return buf
-
-  def synchronize(self, timeout:int|None=None):
-    if not hasattr(self, 'iface'): return
-    sig = self.timeline_signal().as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
-    tl = self.timeline_value().as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')
-    st = time.perf_counter()
-    while sig[0] < tl[0] - 1:
-      if time.perf_counter() - st > (timeout or 3000) / 1000: self.on_device_hang()
-
-  def device_props(self) -> dict[str,Any]: return {} # to be overridden if needed. dict keys are backend dependent.
-
-  def count(self) -> int: return self.iface.count if hasattr(self, 'iface') else 1
-
-  def _select_iface(self):
-    assert (v:=getenv(k:=f'{type(self).__name__[:-6].upper()}_IFACE', "")) == "",  \
-      f"{k}={v} is deprecated, use DEV={replace(DEV.target(type(self).__name__[:-6]), interface=v)} instead"
-    assert hasattr(self, "ifaces"), "must have ifaces to select an iface"
-    t = DEV.target(dev:=type(self).__name__[:-6])
-    filtered = select_by_name(self.ifaces, lambda i: i.__name__[:-5], t.interface, f"{dev} has no interface {t.interface!r}")
-    filtered = [i for i in filtered if t.interface.startswith("MOCK") or not i.__name__[:-5].startswith("MOCK")] # never fall back to mock ifaces
-    return select_first_inited([functools.partial(cast(Callable, iface), self, self.device_id) for iface in filtered],
-                               f"No interface for {dev}:{self.device_id} is available")
-
-  def _is_cpu(self) -> bool: return hasattr(self, 'device') and self.device.split(":")[0] == "CPU"
-
-  def finalize(self):
-    try: self.synchronize() # try to finalize the device in any case
-    except RuntimeError as e: print(f"{self.device} synchronization failed before finalizing: {e}")
-
-    # if the device has an interface, call device_fini to clean up resources
-    if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()
-
-@dataclass
-class HCQ2Buffer:
-  va_addr:sint
-  meta:Any=None
-  view:MMIOInterface|None=None
-
-  def offset(self, offset:int, size:int) -> 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:
-    return unwrap(buf.view).mv
-
-  def _map(self, buf:HCQ2Buffer) -> HCQ2Buffer:
-    if not hasattr(self, '_do_map'): raise NotImplementedError("map failed: no method implemented")
-    return self._do_map(buf)
-
-  @suppress_finalizing
-  def _free(self, buf:HCQ2Buffer, 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)
-
-  def _offset(self, buf, size:int, offset:int) -> HCQ2Buffer: return buf.offset(offset=offset, size=size)
@@ -3,8 +3,8 @@ 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 extra.hcq2.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_cmdbuf
-from extra.hcq2.hcq2 import make_binary_patch
+from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, 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
 from tinygrad.dtype import dtypes
@@ -37,7 +37,7 @@ class PM4Ops(FastEnum):
  RELEASE_MEM = auto(); DISPATCH_DIRECT = auto(); EVENT_WRITE = auto()  # noqa: E702

 def pkt3(ctx, op:PM4Ops, *vals):
-  return UOp(Ops.INS, arg=op, src=tuple(UOp.const(dtypes.uint32, x)
+  return UOp(Ops.INS, arg=op, src=tuple(UOp.const(x, dtypes.uint32)
    for x in (ctx.pm4.PACKET3(getattr(ctx.pm4, f"PACKET3_{op.name}"), len(vals) - 1), *vals)))

 def wreg(ctx, reg:AMDReg, *args:sint, **kwargs:int):
@@ -146,31 +146,31 @@ pm_pm4_opsel = PatternMatcher([
  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), pm4_store),
 ])

-def pm4_submit(cmdbuf, devs):
-  size, zero = UOp.const(dtypes.uint32, cmdbuf.nbytes() // dtypes.uint32.itemsize), UOp.const(dtypes.int, 0)
-
-  # the compute queue's ring and its host-side ring/write/put pointers (placeholders, resolved in pm_bufferize)
-  for d in devs: q = Device[d].compute_queue
+def pm4_submit(ctx, lin):
+  # ensure compute queues are allocated
+  for d in (devs:=ctx.devs): q = Device[d].compute_queue
  ring, wptr, doorbell, put_ptr = (UOp.placeholder((b.size,), b.dtype, 0, device=devs).rtag(f"COMPUTE:0_{name}")
    for name, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))

-  # place the cmdbuf at the ring's write offset, wrapping the ring
-  put = put_ptr.index(zero)
-  next_put = put + size.cast(put.dtype)
-  i = UOp.range(size, 0, dtype=dtypes.int, src=(cmdbuf,))
-  ring_idx = ((put + i.cast(put.dtype)) % q.ring.size).cast(dtypes.int)
+  # the host fence at the start of the batch guarantees the ib is free to reuse
+  size_dw = sum(len(ins.src) for ins in lin.src)
+  assert size_dw < (1 << 20), f"indirect buffer of {size_dw} dwords doesn't fit one packet"

-  # copy the cmdbuf into the ring and advance the put/write pointers
-  copy_to_ring = ring.index(ring_idx).store(cmdbuf.index(i).load()).end(i)
-  bump_put_ptr = put_ptr.index(zero).store(next_put)
-  bump_wptr = wptr.index(zero).store(next_put)
+  ib = UOp.placeholder((size_dw,), dtypes.uint32, next(UOp.unique_num), device=devs, volatile=True).rtag("cmdbuf")
+  cmdbuf = make_cmdbuf(lin, devs, buf=ib)

-  # ring the doorbell once the copy and pointer bumps have landed
-  flush = UOp.barrier(copy_to_ring, bump_put_ptr, bump_wptr)
-  return doorbell.after(flush).index(zero).store(next_put)
+  # the ring itself only carries a packet pointing at the ib, wrapping the ring
+  put = put_ptr.index(zero:=UOp.const(0, dtypes.int))
+  pkt = (ctx.pm4.PACKET3(ctx.pm4.PACKET3_INDIRECT_BUFFER, 2), *data64_le(cmdbuf.getaddr(devs)), size_dw | ctx.pm4.INDIRECT_BUFFER_VALID)
+  write_pkt = UOp.barrier(*[ring.index(((put + off) % q.ring.size).cast(dtypes.int)).store(UOp.const(x, dtypes.uint32)) for off,x in enumerate(pkt)])

-pm_pm4_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"),
-  lambda lin: pm4_submit(make_cmdbuf(lin, to_tuple(lin.arg[0])), to_tuple(lin.arg[0])))])
+  # advance the put/write pointers past the packet
+  bump_put_ptr = put_ptr.index(zero).store(put + len(pkt))
+  bump_wptr = wptr.index(zero).store(put + len(pkt))
+  flush = UOp.barrier(write_pkt, bump_put_ptr, bump_wptr)
+  return doorbell.after(flush).index(zero).store(put + len(pkt))
+
+pm_pm4_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"), pm4_submit)])

 # *****************
 # SDMA
@@ -180,26 +180,27 @@ class SDMAOps(FastEnum): COPY = auto(); POLL_REGMEM = auto(); FENCE = auto(); TR
 def sdma_copy(ctx, call):
  sz = call.src[2].max_numel() * call.src[2].dtype.itemsize
  src_addr, dst_addr = call.src[2].getaddr(ctx.devs), call.src[1].getaddr(ctx.devs)
-  return call.ins(SDMAOps.COPY, src=tuple(UOp.const(dtypes.uint32, x) for off in range(0, sz, ctx.max_copy_size) for x in (
+  return call.ins(SDMAOps.COPY, src=tuple(UOp.const(x, dtypes.uint32) for off in range(0, sz, ctx.max_copy_size) for x in (
    ctx.sdma.SDMA_OP_COPY | ctx.sdma.SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_COPY_LINEAR),
-    ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz-off, ctx.max_copy_size)-1), 0, *data64_le(src_addr+off), *data64_le(dst_addr+off))))
+    ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz-off, ctx.max_copy_size)-1), 0,
+    *data64_le(src_addr+UOp.const(off, dtypes.uint64)), *data64_le(dst_addr+UOp.const(off, dtypes.uint64)))))

 def sdma_wait(ctx, ins, dst, val):
  op = ctx.sdma.SDMA_OP_POLL_REGMEM | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_FUNC(WAIT_REG_MEM_FUNCTION_GEQ) \
     | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)
-  return ins.ins(SDMAOps.POLL_REGMEM, src=tuple(UOp.const(dtypes.uint32, x) for x in (
+  return ins.ins(SDMAOps.POLL_REGMEM, src=tuple(UOp.const(x, dtypes.uint32) for x in (
    op, *data64_le(dst.getaddr(ctx.devs)), val, 0xffffffff,
    ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(0x04) | ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff))))

 def sdma_store(ctx, ins, dst, val):
  op = ctx.sdma.SDMA_OP_FENCE | (ctx.sdma.SDMA_PKT_FENCE_HEADER_MTYPE(3) if ctx.target[0] != 9 else 0)
  return UOp(Ops.LINEAR, src=(
-    ins.ins(SDMAOps.FENCE, src=tuple(UOp.const(dtypes.uint32, x) for x in (op, *data64_le(dst.getaddr(ctx.devs)), val))),
-    ins.ins(SDMAOps.TRAP, src=tuple(UOp.const(dtypes.uint32, x) for x in (ctx.sdma.SDMA_OP_TRAP, 0)))))
+    ins.ins(SDMAOps.FENCE, src=tuple(UOp.const(x, dtypes.uint32) for x in (op, *data64_le(dst.getaddr(ctx.devs)), val))),
+    ins.ins(SDMAOps.TRAP, src=tuple(UOp.const(x, dtypes.uint32) for x in (ctx.sdma.SDMA_OP_TRAP, 0)))))

 def sdma_timestamp(ctx, ins, dst):
  op = ctx.sdma.SDMA_OP_TIMESTAMP | ctx.sdma.SDMA_PKT_TIMESTAMP_GET_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_TIMESTAMP_GET_GLOBAL)
-  return ins.ins(SDMAOps.TIMESTAMP, src=tuple(UOp.const(dtypes.uint32, x) for x in (op, *data64_le(dst.getaddr(ctx.devs)))))
+  return ins.ins(SDMAOps.TIMESTAMP, src=tuple(UOp.const(x, dtypes.uint32) for x in (op, *data64_le(dst.getaddr(ctx.devs)))))

 pm_sdma_opsel = PatternMatcher([
  (UPat(Ops.CALL, src=(UPat(Ops.COPY),), name="call", allow_any_len=True), sdma_copy),
@@ -212,7 +213,7 @@ pm_sdma_opsel = PatternMatcher([

 def sdma_submit(cmdbuf, devs):
  # the cmdbuf to submit + the patch writes that fill it
-  size_dw, zero = cmdbuf.nbytes() // dtypes.uint32.itemsize, UOp.const(dtypes.int, 0)
+  size_dw, zero = cmdbuf.nbytes() // dtypes.uint32.itemsize, UOp.const(0, dtypes.int)

  # the sdma queue's ring and its host-side ring/write/put pointers
  for d in devs: q = Device[d].sdma_queue(0)
@@ -228,8 +229,8 @@ def sdma_submit(cmdbuf, devs):

  # zero the wrapped tail, then copy the cmdbuf into the ring
  zi = UOp.range(zero_amt_dw, 0, dtype=dtypes.int, src=(cmdbuf,))
-  zero_tail = ring.index(tail_off_dw + zi).store(UOp.const(dtypes.uint32, 0)).end(zi)
-  i = UOp.range(UOp.const(dtypes.int, size_dw), 0, dtype=dtypes.int, src=(cmdbuf,))
+  zero_tail = ring.index(tail_off_dw + zi).store(UOp.const(0, dtypes.uint32)).end(zi)
+  i = UOp.range(UOp.const(size_dw, dtypes.int), 0, dtype=dtypes.int, src=(cmdbuf,))
  copy_to_ring = ring.index(start_dw + i).store(cmdbuf.index(i).load()).end(i)

  # advance the put/write pointers past the zeroed tail and the cmdbuf
@@ -242,7 +243,7 @@ def sdma_submit(cmdbuf, devs):
  return doorbell.after(flush).index(zero).store(next_put_b)

 pm_sdma_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"),
-  lambda lin: sdma_submit(make_cmdbuf(lin, to_tuple(lin.arg[0])), to_tuple(lin.arg[0])))])
+  lambda ctx, lin: sdma_submit(make_cmdbuf(lin, ctx.devs), ctx.devs))])

@dataclass(frozen=True)
 class AMDEncodeCtx:  # encode-time constants for one queue: devs (every cmdbuf address resolves into these) + gfx version + packet/ip modules
@@ -253,7 +254,7 @@ def encode_queue(q:UOp) -> UOp|None:
  d = Device[(devs:=to_tuple(q.arg[0]))[0]]
  ctx = AMDEncodeCtx(devs, d.target, d.pm4, d.sdma, d.soc, d.gc, d.nbio, d.xccs, d.max_copy_size, d.tmpring_size)
  opsel, submit = (pm_pm4_opsel, pm_pm4_submit) if q.arg[1].startswith("COMPUTE") else (pm_sdma_opsel, pm_sdma_submit)
-  return submit.rewrite(graph_rewrite(q, opsel + pm_flatten_linear, walk=True, ctx=ctx, name=f"{q.arg[1]} opsel"))
+  return submit.rewrite(graph_rewrite(q, opsel + pm_flatten_linear, walk=True, ctx=ctx, name=f"{q.arg[1]} opsel"), ctx)

@dataclass(frozen=True)
 class AMDProgramData:
@@ -507,11 +508,12 @@ class PCIIface(PCIIfaceBase):
    if drain_only: d.iface.dev_impl.ih.drain()
    else: d.iface.dev_impl.ih.interrupt_handler()

-    if reset and d.iface.dev_impl.recover():
+    if reset and d.iface.dev_impl.recover(force=True):
      cq = d.compute_queue
      for b in (cq.put_value, cq.read_ptr, cq.write_ptr): b._buf.view.view(fmt='Q')[0] = 0
      d.iface.dev_impl.gfx.setup_ring(*cq.params)
-      d.timeline_signal()._buf.cpu_view().mv.cast('Q')[0] = d.timeline_value().as_memoryview(force_zero_copy=True).cast('Q')[0] - 1
+      d.signal('timeline')._buf.cpu_view().mv.cast('Q')[0] = \
+        d.signal('value', 1).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] - 1

  def sleep(self, timeout):
    if hasattr(self.pci_dev, 'irq_poller') and self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
@@ -537,9 +539,12 @@ 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

@@ -631,9 +636,6 @@ class AMDDevice(HCQ2Compiled):
    qname = f"{'COPY' if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA else 'COMPUTE'}:{idx}"
    self.pm_bufferize = PatternMatcher([
      (UPat(Ops.PARAM, tag=f"{qname}_{name}"), lambda ctx, b=getattr(queue, name): b) for name in ["ring", "write_ptr", "doorbell", "put_value"]
-    ] + [
-      (UPat(Ops.PARAM, tag=f"{qname}_timeline_signal"), lambda ctx, q=qname: ctx[0].timeline_signal(q)),
-      (UPat(Ops.PARAM, tag=f"{qname}_timeline_value"), lambda ctx, q=qname: ctx[0].timeline_value(q)),
    ]) + self.pm_bufferize

    return queue
@@ -692,7 +694,7 @@ 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
@@ -31,12 +31,12 @@ def dname_of(device) -> str:

 def alloc_like(shape, dtype, device, axis=None) -> Tensor:
  if isinstance(device, tuple) and axis is not None:
-    return Tensor(Tensor.invalids(*shard_shape(shape, axis, len(device)), dtype=dtype, device=device).uop.multi(axis), device=device)
+    return Tensor(Tensor.invalids(*shard_shape(shape, axis, len(device)), dtype=dtype, device=device).uop.unshard(axis), device=device)
  return Tensor.invalids(*shape, dtype=dtype, device=device)

 def alloc_local(shape, dtype, device, axis=None) -> Tensor:
  if isinstance(device, tuple) and axis is not None:
-    return Tensor(Tensor.invalids(*shape, dtype=dtype, device=device).uop.multi(0), device=device)
+    return Tensor(Tensor.invalids(*shape, dtype=dtype, device=device).uop.unshard(0), device=device)
  return Tensor.invalids(*shape, dtype=dtype, device=device)

 def compile_hip(src:str, defines:list[str]):
@@ -48,7 +48,7 @@ def _fused_ce_loss_bwd(gradient:UOp, kernel:UOp, label_smoothing:float):
    axis = logits_u.axis
    ndev = len(device)
    local_shape = tuple(s//ndev if i == axis else s for i,s in enumerate((MBS, SEQ, VOCAB)))
-    d_logits = Tensor(Tensor.invalids(*local_shape, dtype=dtypes.bfloat16, device=device).uop.multi(axis), device=device)
+    d_logits = Tensor(Tensor.invalids(*local_shape, dtype=dtypes.bfloat16, device=device).uop.unshard(axis), device=device)
    rows_per_dev = local_shape[0] * local_shape[1]
    seq_per_dev = local_shape[1]
  else:
@@ -74,11 +74,11 @@ def fused_ce_loss(logits:Tensor, targets:Tensor, label_smoothing:float=0.1) -> T
    axis = logits.uop.axis
    assert axis in (0, 1), f"unsupported sharding axis={axis} for CE loss"
    ndev = len(logits.device)
-    loss_out = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.multi(0),
+    loss_out = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.unshard(0),
                      device=logits.device)
-    max_out  = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.multi(0),
+    max_out  = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.unshard(0),
                      device=logits.device)
-    lse_out  = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.multi(0),
+    lse_out  = Tensor(Tensor.invalids(rows // ndev, dtype=dtypes.float32, device=logits.device).uop.unshard(0),
                      device=logits.device)
    local_shape = tuple(s//ndev if i == axis else s for i,s in enumerate(logits.shape))
    rows_per_dev = local_shape[0] * local_shape[1]
@@ -16,7 +16,7 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_out:UOp, x:UOp, amax_state:

  wg = UOp.range(NUM_WG, 0, AxisType.GLOBAL)
  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
-  it = UOp.range((n_elems // VEC) // (NUM_WG * THREADS_PER_WG), 2, AxisType.LOOP)
+  it = UOp.range((n_elems // VEC) // (NUM_WG * THREADS_PER_WG), 2, AxisType.WEAK)
  lane = UOp.range(VEC, 3, AxisType.UNROLL)

  idx = (((it * NUM_WG + wg) * THREADS_PER_WG + tid) * VEC) + lane
@@ -36,19 +36,19 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_out:UOp, x:UOp, amax_state:
  lmax_val = lmax.after(lmax_store.end(it))[0]

  lds = UOp.placeholder((THREADS_PER_WG,), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL)
-  lds = lds.after(lds[tid].store(lmax_val).barrier())
+  lds = lds.after(lds[tid].store(lmax_val))

  step = THREADS_PER_WG // 2
  while step:
    active = tid < step
    other = lds[(tid + step).valid(active)].load()
-    lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)).barrier())
+    lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)))
    step //= 2

  device = device[0].split(":")[0] if isinstance(device, tuple) else device.split(":")[0]
  if device in {"AMD", "NULL"}: atomic_arg = "if ({2} > {3}) __hip_atomic_fetch_max((int*){0}, {1}, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_AGENT);"
  else: raise NotImplementedError(f"no atomic max for device {device}")
-  amax_idx = amax_out.reshape((1,)).index(UOp.const(dtypes.weakint, 0))
+  amax_idx = amax_out.reshape((1,)).index(UOp.const(0))
  max_val = lds[0].load()
  atomic = UOp(Ops.CUSTOM, dtypes.void, (amax_idx, max_val.bitcast(dtypes.int32), max_val, amax_idx.load()), arg=atomic_arg)
  return atomic.end(tid, wg).sink(arg=KernelInfo(f"quantize_fp8_with_amax_{n_elems}", opts_to_apply=()))
@@ -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])
@@ -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();
+      }
+    }
+  }
+}
@@ -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]
@@ -4,7 +4,7 @@ from hexdump import hexdump
 from copy import deepcopy
 import pathlib, sys
 from tinygrad.helpers import to_mv, getenv
-from tinygrad.runtime.autogen import adreno
+from tinygrad.runtime.autogen import mesa
 sys.path.append(pathlib.Path(__file__).parent.parent.parent.as_posix())

 IOCTL = getenv("IOCTL", 0)
@@ -23,7 +23,7 @@ for child in xml.getroot():
 CAPTURED_STATE = {}

 REGS = {}
-for k, v in adreno.__dict__.items():
+for k, v in mesa.__dict__.items():
  if k.startswith("REG_") and isinstance(v, int) and v > 1024: REGS[v] = k

 from extra.qcom_gpu_driver import msm_kgsl
@@ -42,7 +42,7 @@ def get_struct(argp, stype):

 def format_struct(s):
  sdats = []
-  for field_name, *_ in s._real_fields_:
+  for field_name, *_ in s._fields_:
    if field_name in {"__pad", "PADDING_0"}: continue
    dat = getattr(s, field_name)
    if isinstance(dat, int): sdats.append(f"{field_name}:0x{dat:X}")
@@ -96,9 +96,9 @@ def parse_cmd_buf(dat):
        CAPTURED_STATE['LOAD_FRAGS'].append((state_block, state_type, num_unit, dst_off))

        if state_block == SB6_CS_SHADER:
-          from extra.disassemblers.adreno import disasm_raw
+          from tinygrad.runtime.support.compiler_mesa import disas_adreno
          if state_type == ST6_SHADER and IOCTL > 3:
-            disasm_raw(get_mem(((vals[2] << 32) | vals[1]), num_unit * 128))
+            disas_adreno(get_mem(((vals[2] << 32) | vals[1]), num_unit * 128))
          if state_type == ST6_CONSTANTS:
            x = get_mem(((vals[2] << 32) | vals[1]), num_unit*4)
            CAPTURED_STATE['constants'] = x[:]
@@ -142,7 +142,7 @@ def parse_cmd_buf(dat):
      vals = struct.unpack("I"*size, dat[ptr+4:ptr+4+4*size])
      if IOCTL > 0: print(f"{ptr:3X} -- typ 4: {size=:3d}, {reg_name}", hprint(vals))
      for vi,v in enumerate(vals): CAPTURED_STATE[offset+vi] = v
-      if offset == adreno.REG_A6XX_SP_CS_CONFIG:
+      if offset == mesa.REG_A6XX_SP_CS_CONFIG:
        val = vals[0]
        if IOCTL > 0:
          print(f"\tBINDLESS_TEX={(val >> 0) & 0b1}")
@@ -215,79 +215,3 @@ def install_hook(c_function, python_function):

 libc = ctypes.CDLL(ctypes.util.find_library("libc"))
 install_hook(libc.ioctl, ioctl)
-
-def before_launch():
-  global CAPTURED_STATE
-  CAPTURED_STATE.clear()
-def collect_last_launch_state():
-  global CAPTURED_STATE
-  return deepcopy(CAPTURED_STATE)
-def compare_launch_state(state, good_state):
-  cmp = [
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_NTEX__MASK),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_NSAMP__MASK),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_NIBO__MASK),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_ENABLED),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_TEX),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_SAMP),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_IBO),
-    (adreno.REG_A6XX_SP_CS_CONFIG, adreno.A6XX_SP_CS_CONFIG_BINDLESS_UBO),
-
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_HALFREGFOOTPRINT__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_BRANCHSTACK__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_THREADMODE__MASK),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_EARLYPREAMBLE),
-    (adreno.REG_A6XX_SP_CS_CTRL_REG0, adreno.A6XX_SP_CS_CTRL_REG0_MERGEDREGS),
-
-    (adreno.REG_A6XX_SP_CS_PVT_MEM_PARAM, adreno.A6XX_SP_CS_PVT_MEM_PARAM_MEMSIZEPERITEM__MASK),
-    (adreno.REG_A6XX_SP_CS_PVT_MEM_PARAM, adreno.A6XX_SP_CS_PVT_MEM_PARAM_HWSTACKSIZEPERTHREAD__MASK),
-
-    (adreno.REG_A6XX_SP_CS_UNKNOWN_A9B1, adreno.A6XX_SP_CS_UNKNOWN_A9B1_UNK5),
-    (adreno.REG_A6XX_SP_CS_UNKNOWN_A9B1, adreno.A6XX_SP_CS_UNKNOWN_A9B1_UNK6),
-
-    (adreno.REG_A6XX_SP_CS_BRANCH_COND, 0xffffffff),
-
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_KERNELDIM__MASK),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEX__MASK),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEY__MASK),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_0, adreno.A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEZ__MASK),
-
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_1, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_2, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_3, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_4, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_5, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_NDRANGE_6, 0xffffffff),
-
-    (adreno.REG_A6XX_HLSQ_CS_CNTL_0, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_CNTL_1, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_KERNEL_GROUP_X, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_KERNEL_GROUP_Y, 0xffffffff),
-    (adreno.REG_A6XX_HLSQ_CS_KERNEL_GROUP_Z, 0xffffffff),
-  ]
-
-  for x,m in cmp:
-    print(f"Field {REGS[x]}, mask: 0x{m:X} cmp: {state.get(x, 0) & m} vs {good_state.get(x, 0) & m}")
-    if state.get(x, 0) & m != good_state.get(x, 0) & m:
-      return False, f"Field {REGS[x]}, mask: 0x{m:X} mismatch: {state.get(x, 0) & m} vs {good_state.get(x, 0) & m}"
-
-  for n in ['descriptors', 'ibos']:
-    if n not in good_state: continue
-    mv1, mv2 = state.get(n), good_state.get(n)
-
-    if len(mv1) != len(mv2): return False, f"{n}: len mismatch {len(mv1)} != {len(mv2)}"
-    mv1 = memoryview(bytearray(mv1)).cast('I')
-    mv2 = memoryview(bytearray(mv2)).cast('I')
-    for i in range(len(mv2)):
-      if i % 8 == 5 or i % 8 == 4: continue # addresses
-      if mv1[i]!=mv2[i]: return False, f"{n}: content mismatch {i} {mv1[i]} {mv2[i]}"
-
-  for n in ['samplers']:
-    if n not in good_state: continue
-    mv1, mv2 = state.get(n), good_state.get(n)
-    if len(mv1) != len(mv2): return False, f"{n}: len mismatch {len(mv1)} != {len(mv2)}"
-    if any(mv1[i]!=mv2[i] for i in range(len(mv1))): return False, f"{n}: content mismatch"
-
-  return True, "PASS"
@@ -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
@@ -14,7 +14,7 @@ def _sharded_empty(shape:Tensor, ref:Tensor, axis:int|None, dtype:DTypeLike|None
  shard_axis = ref.uop.axis if axis is None else axis
  shape = tuple(s // len(ref.device) if i == shard_axis else s for i, s in enumerate(shape))
  axis = ref.uop.axis if axis is None else axis
-  return Tensor(Tensor.invalids(*shape, dtype=dtype, device=ref.device).uop.multi(axis), dtype=dtype, device=ref.device)
+  return Tensor(Tensor.invalids(*shape, dtype=dtype, device=ref.device).uop.unshard(axis), dtype=dtype, device=ref.device)

@functools.cache
 def custom_fused_qkv_rope_forward(q:UOp, k:UOp, v:UOp, xqkv:UOp, freqs_cis:UOp,
@@ -206,10 +206,11 @@ 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)))
@@ -236,10 +237,11 @@ 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)))
@@ -268,10 +270,11 @@ 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 +301,11 @@ 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)))
@@ -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);
@@ -533,6 +549,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -638,6 +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);
@@ -791,6 +816,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -895,6 +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);
@@ -1048,6 +1083,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1151,6 +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);
@@ -1303,6 +1349,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1428,6 +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);
@@ -1582,6 +1637,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1689,6 +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);
@@ -1842,6 +1906,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1946,6 +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);
@@ -2099,6 +2173,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -2202,6 +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);
@@ -2354,6 +2439,8 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
+        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
+        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -2471,6 +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);
@@ -2732,6 +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);
@@ -2988,6 +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);
@@ -3244,6 +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);
@@ -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);
        }
    }
@@ -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");
+    }
+}
@@ -0,0 +1,74 @@
+#include "kittens.cuh"
+
+using namespace kittens;
+
+#ifndef MATVEC_N
+#define MATVEC_N 1536
+#endif
+#ifndef MATVEC_K
+#define MATVEC_K 7168
+#endif
+
+constexpr int SPLIT_WAVES = 8;
+
+template<int W>
+__device__ __forceinline__ float run_split(const bf16 *A_ptr, const bf16 *B_ptr, int out_base,
+                                           st_bf<16, 32, st_16x32_s> &As,
+                                           st_bf<16, 32, st_16x32_s> &Bs) {
+  constexpr int K = MATVEC_K;
+  rt_bf<16, 32, row_l, rt_16x32_s> A;
+  rt_bf<16, 32, row_l, rt_16x32_s> B;
+  rt_fl<16, 16, col_l, rt_16x16_s> C;
+  zero(C);
+  const int lane = laneid();
+  constexpr int k_begin = W * (K / SPLIT_WAVES), k_end = k_begin + K / SPLIT_WAVES;
+  #pragma unroll 1
+  for (int k = k_begin; k < k_end; k += 32) {
+    #pragma unroll
+    for (int idx = lane; idx < 16 * 32; idx += 64) {
+      const int row = idx / 32, col = idx % 32;
+      *reinterpret_cast<bf16 *>(reinterpret_cast<char *>(&As.data[0]) + As.swizzle({row, col})) = A_ptr[k + col];
+      *reinterpret_cast<bf16 *>(reinterpret_cast<char *>(&Bs.data[0]) + Bs.swizzle({row, col})) =
+        B_ptr[(out_base + row) * K + k + col];
+    }
+    asm volatile("s_waitcnt lgkmcnt(0)");
+    load(A, As);
+    load(B, Bs);
+    asm volatile("s_waitcnt lgkmcnt(0)");
+    mma_ABt(C, A, B, C);
+  }
+  return C.tiles[0][0].data[0].x;
+}
+
+// Eight waves split K for one 16-channel output tile.  Each wave uses MFMA on
+// a repeated activation row, then wave zero reduces the eight FP32 partials.
+__global__ __launch_bounds__(64 * SPLIT_WAVES, 1)
+void hk_bf16_matvec_splitk(bf16 *C_ptr, const bf16 *A_ptr, const bf16 *B_ptr, bf16 *unused) {
+  constexpr int N = MATVEC_N, K = MATVEC_K;
+  static_assert(N % 16 == 0 && K % (32 * SPLIT_WAVES) == 0);
+  __shared__ st_bf<16, 32, st_16x32_s> As[SPLIT_WAVES];
+  __shared__ st_bf<16, 32, st_16x32_s> Bs[SPLIT_WAVES];
+  __shared__ float partial[SPLIT_WAVES][16];
+  const int tid = threadIdx.x, wave = tid / 64, lane = tid & 63;
+  const int out_base = blockIdx.x * 16;
+  float result = 0.0f;
+  switch (wave) {
+    case 0: result = run_split<0>(A_ptr, B_ptr, out_base, As[0], Bs[0]); break;
+    case 1: result = run_split<1>(A_ptr, B_ptr, out_base, As[1], Bs[1]); break;
+    case 2: result = run_split<2>(A_ptr, B_ptr, out_base, As[2], Bs[2]); break;
+    case 3: result = run_split<3>(A_ptr, B_ptr, out_base, As[3], Bs[3]); break;
+    case 4: result = run_split<4>(A_ptr, B_ptr, out_base, As[4], Bs[4]); break;
+    case 5: result = run_split<5>(A_ptr, B_ptr, out_base, As[5], Bs[5]); break;
+    case 6: result = run_split<6>(A_ptr, B_ptr, out_base, As[6], Bs[6]); break;
+    case 7: result = run_split<7>(A_ptr, B_ptr, out_base, As[7], Bs[7]); break;
+  }
+  if (lane < 16) partial[wave][lane] = result;
+  asm volatile("s_waitcnt lgkmcnt(0)");
+  __builtin_amdgcn_s_barrier();
+  if (wave == 0 && lane < 16) {
+    float total = 0.0f;
+    #pragma unroll
+    for (int i = 0; i < SPLIT_WAVES; i++) total += partial[i][lane];
+    C_ptr[out_base + lane] = static_cast<bf16>(total);
+  }
+}
@@ -16,7 +16,7 @@ def _sharded_empty(shape:Tensor, ref:Tensor, axis:int|None) -> Tensor:
  if not isinstance(ref.device, tuple): return Tensor.empty(*shape, dtype=ref.dtype, device=ref.device)
  shape = tuple(s // len(ref.device) if i == ref.uop.axis else s for i, s in enumerate(shape))
  axis = ref.uop.axis if axis is None else axis
-  return Tensor(Tensor.empty(*shape, dtype=ref.dtype, device=ref.device).uop.multi(axis), dtype=ref.dtype, device=ref.device)
+  return Tensor(Tensor.empty(*shape, dtype=ref.dtype, device=ref.device).uop.unshard(axis), dtype=ref.dtype, device=ref.device)

 def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
  return _sharded_empty(ref.shape, ref, axis)
@@ -224,7 +224,7 @@ class Group:

      # store to shared memory
      red_local_store = red_local[self.laneid].store(red_reg[0])
-      red_local = red_local.after(red_local_store.barrier()).reshape(red_local.shape)
+      red_local = red_local.after(red_local_store).reshape(red_local.shape)

      # reduce from shared memory
      for inner in self.ker.range(3, axis_type=AxisType.REDUCE, track=False):
@@ -258,7 +258,7 @@ class Group:

      # store to shared memory
      red_local_store = red_local[self.laneid].store(red_reg[0])
-      red_local = red_local.after(red_local_store.barrier()).reshape(red_local.shape)
+      red_local = red_local.after(red_local_store).reshape(red_local.shape)

      # reduce from shared memory
      for inner in self.ker.range(3, axis_type=AxisType.REDUCE, track=False):
@@ -342,7 +342,7 @@ class Group:
          if src.dtype != dst.dtype:
            src_load = src_load.cast(dst.dtype)
          dst_store = dst[*dst_idxs, height, width, srow, scol].store(src_load)
-          dst_store = dst_store.end(height, width, outer, inner).barrier()
+          dst_store = dst_store.end(height, width, outer, inner)
    elif dst.addrspace == AddrSpace.REG and src.addrspace == AddrSpace.GLOBAL and isinstance(dst, RT):
      srcf = src.flatten()
      row_stride = prod(src.shape[axis+1:])
@@ -48,14 +48,14 @@ class Kernel(AbstractContextManager):
  @property
  def warpgroup(self): return self.group(4)

-  def range(self, start:int, end:int=0, step:int=1, axis_type:AxisType=AxisType.LOOP, track:bool=True):
+  def range(self, start:int, end:int=0, step:int=1, axis_type:AxisType=AxisType.WEAK, track:bool=True):
    if end == 0: start, end = 0, start
    rng = _tk_range(start, end, step, axis_type, self.range_id)
    self.range_id += 1
    if track: self.range_stack.append(rng)
    return rng

-  def raw_range(self, end:int=0, axis_type:AxisType=AxisType.LOOP):
+  def raw_range(self, end:int=0, axis_type:AxisType=AxisType.WEAK):
    rng = UOp.range(end, self.range_id, axis_type=axis_type)
    self.range_id += 1
    return rng
@@ -3,7 +3,7 @@
 # A002 Function argument `input` is shadowing a Python builtin
 # A006 Lambda argument `input` is shadowing a Python builtin
 from tinygrad import Tensor, dtypes, Device
-from tinygrad.uop.ops import Ops
+from tinygrad.uop.ops import Ops, GroupOp
 from tinygrad.helpers import getenv, prod, strides_for_shape, argfix
 import torch.lib
 TORCH_DEBUG = getenv("TORCH_DEBUG")
@@ -18,12 +18,16 @@ def _to_torch_device(device: str): return torch.device("tiny", int(device.partit

 import torch.utils.cpp_extension
 mod = torch.utils.cpp_extension.load(name="custom_device_extension", sources=[str(pathlib.Path(__file__).parent / "wrapped_tensor.cpp")])
+# TODO: this assumes a contiguous source, so PERMUTE/EXPAND/PAD/FLIP are wrong. UOp.contiguous_view_offset does it
+# properly, but it needs a device (these are deviceless)
+alias_ops = GroupOp.Movement | {Ops.BITCAST, Ops.DETACH, Ops.AFTER}
 def calculate_storage_offset(x: Tensor) -> int:
-  offset = 0
-  for u in x.uop.toposort():
-    if u.op == Ops.SHRINK:
+  offset, u = 0, x.uop
+  while u.op in alias_ops:
+    if u.op is Ops.SHRINK:
      u_strides = strides_for_shape(u.src[0].shape)
      for i, (start, _) in enumerate(u.marg): offset += start * u_strides[i]
+    u = u.src[0]
  return offset
 def wrap(x: Tensor, dev: torch.device|None=None) -> torch.Tensor:
  x._strides = strides_for_shape(x.shape) # always recalculate
@@ -220,7 +224,7 @@ def _as_strided(tensor:Tensor, size, stride, storage_offset=0):

@torch.library.impl("aten::as_strided", "privateuseone")
 def as_strided(tensor:torch.Tensor, size, stride, storage_offset=None):
-  storage_offset = storage_offset or tensor.storage_offset()
+  if storage_offset is None: storage_offset = tensor.storage_offset()
  return _as_strided(tensor, size, stride, storage_offset)

@torch.library.impl("aten::_reshape_alias", "privateuseone")
@@ -228,16 +232,16 @@ def _reshape_alias(tensor:torch.Tensor, size, stride):
  return _as_strided(tensor, size, stride)

@torch.library.impl("aten::empty_strided", "privateuseone")
-def empty_strided(size, stride, dtype, layout=None, device=None, pin_memory=False):
+def empty_strided(size, stride, dtype=None, layout=None, device=None, pin_memory=False):
  if TORCH_DEBUG: print(f"empty_strided {size=} {stride=} {dtype=} {layout=} {device=} {pin_memory=}")
-  ret = Tensor.empty(*size, dtype=_from_torch_dtype(dtype), device=_from_torch_device(device)).contiguous()
+  ret = Tensor.empty(*size, dtype=_from_torch_dtype(dtype or torch.get_default_dtype()), device=_from_torch_device(device))
  # TODO: should return with requested strides
  return wrap(ret)

@torch.library.impl("aten::empty.memory_format", "privateuseone")
 def empty_memory_format(size, dtype=None, layout=None, device=None, pin_memory=False, memory_format=None):
  if TORCH_DEBUG: print(f"empty.memory_format {size=} {dtype=} {layout=} {device=} {pin_memory=} {memory_format=}")
-  ret = Tensor.empty(*size, dtype=_from_torch_dtype(dtype or torch.get_default_dtype()), device=_from_torch_device(device)).contiguous()
+  ret = Tensor.empty(*size, dtype=_from_torch_dtype(dtype or torch.get_default_dtype()), device=_from_torch_device(device))
  return wrap(ret)

@torch.library.impl("aten::max_pool2d_with_indices", "privateuseone")
@@ -551,6 +555,8 @@ def wrap_out(f):
    assert out.shape == assigned.shape, f"shape mismatch: {assigned.shape} -> {out.shape}"
    assert out.device == assigned.device or out.device is None or assigned.device is None, f"device mismatch: {assigned.device} -> {out.device}"
    assert out.dtype == assigned.dtype, f"dtype mismatch: {assigned.dtype} -> {out.dtype}"
+    # an out= that is a view has to be written through its base, and _apply_inplace gives a deviceless base its buffer first
+    if canonical_base(out) is not out: return _apply_inplace(out, assigned) or out
    if out.device is None and assigned.device is not None: out.replace(out.empty_like(device=assigned.device))
    return out.assign(assigned)
  return _wrap_out
@@ -652,35 +658,6 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
  "aten.unfold": Tensor.unfold,
 }}

-# operations that need inplace treatment (use _inplace_op instead of wrap_fxn) AKA return original tensor
-inplace_ops = {
-  "aten.zero_",
-  "aten.fill_.Scalar",
-  "aten.fill_.Tensor",
-  "aten.add_.Tensor",
-  "aten.add_.Scalar",
-  "aten.mul_.Tensor",
-  "aten.mul_.Scalar",
-  "aten.floor_divide_.Tensor",
-  "aten.__ilshift__.Scalar",
-  "aten.__irshift__.Scalar",
-  "aten.relu_",
-  "aten.random_",
-  "aten.random_.from",
-  "aten.uniform_",
-  "aten.normal_",
-  "aten.logical_or_",
-  "aten.masked_fill_.Scalar",
-  "aten.masked_fill_.Tensor",
-}
-
-inplace_view_ops = {
-  "aten.squeeze_.dim",
-  "aten.unsqueeze_",
-  "aten.transpose_",
-  "aten.t_",
-}
-
 def wrap_fxn(k,f):
  def nf(*args, **kwargs):
    if TORCH_DEBUG:
@@ -694,7 +671,7 @@ def wrap_fxn(k,f):
    else: raise RuntimeError(f"unknown output type {type(out)}")
  return nf

-def wrap_inplace(k,f):
+def wrap_inplace(f):
  def nf(*args, **kwargs):
    orig = args[0]
    args, kwargs = unwrap_args(args, kwargs)
@@ -702,7 +679,7 @@ def wrap_inplace(k,f):
    return orig
  return nf

-def wrap_inplace_view_op(k,f):
+def wrap_inplace_view_op(f):
  def nf(*args, **kwargs):
    orig = args[0]
    args, kwargs = unwrap_args(args, kwargs)
@@ -735,11 +712,17 @@ def wrap_inplace_view_op(k,f):
    return orig
  return nf

+# the aten schema says how an op is called: an inplace view retargets the view, a writable first arg is inplace,
+# and a writable out arg must have come from tiny_backend_out so that wrap_out was applied
 for k,v in tiny_backend.items():
-  if k in inplace_view_ops: wrapper = wrap_inplace_view_op
-  elif k in inplace_ops: wrapper = wrap_inplace
-  else: wrapper = wrap_fxn
-  torch.library.impl(k.replace("aten.", "aten::"), "privateuseone")(wrapper(k,v))
+  name, _, overload = k.removeprefix("aten.").partition(".")
+  op = getattr(getattr(aten, name), overload or "default")
+  writes = [a.name for a in op._schema.arguments if a.alias_info is not None and a.alias_info.is_write]
+  if torch.Tag.inplace_view in op.tags: fxn = wrap_inplace_view_op(v)
+  elif writes == [op._schema.arguments[0].name] and op._schema.returns: fxn = wrap_inplace(v)
+  elif not writes or (writes == ["out"] and k in tiny_backend_out): fxn = wrap_fxn(k, v)
+  else: raise RuntimeError(f"{k} writes {writes}: expected an inplace first arg, or an out arg with {k} in tiny_backend_out")
+  torch.library.impl(k.replace("aten.", "aten::"), "privateuseone")(fxn)

@torch.library.impl("aten::equal", "privateuseone")
 def equal(x: torch.Tensor, y: torch.Tensor): return (x==y).all().item()
@@ -775,21 +758,17 @@ def native_batch_norm(input, weight, bias, running_mean, running_var, training,
@torch.library.impl("aten::native_batch_norm_backward", "privateuseone")
 def native_batch_norm_backward(grad_out, input, weight, running_mean, running_var, save_mean, save_invstd, train, eps, output_mask):
  grad_out_t, input_t = unwrap(grad_out), unwrap(input)
-  weight_t = unwrap(weight) if weight is not None else None
-  save_mean_t = unwrap(save_mean)
-  save_invstd_t = unwrap(save_invstd)
-  out = input_t.batchnorm(weight_t, None, save_mean_t, save_invstd_t)
-  targets = [t for t, m in zip([input_t, weight_t], output_mask[:2]) if t is not None and m]
-  if targets:
-    grads = out.gradient(*targets, gradient=grad_out_t)
-    grad_input = grads.pop(0) if output_mask[0] else None
-    grad_weight = grads.pop(0) if output_mask[1] and weight_t is not None else None
-  else:
-    grad_input, grad_weight = None, None
-  grad_bias = grad_out_t.sum(axis=tuple(x for x in range(grad_out_t.ndim) if x != 1)) if output_mask[2] else None
-  return (wrap(grad_input) if grad_input is not None else None,
-          wrap(grad_weight) if grad_weight is not None else None,
-          wrap(grad_bias) if grad_bias is not None else None)
+  dims, shape = tuple(x for x in range(input_t.ndim) if x != 1), (1, -1) + (1,)*(input_t.ndim-2)
+  # training differentiates the batch stats it was given, eval treats the running stats as constants
+  if train: mean, invstd = unwrap(save_mean), unwrap(save_invstd)
+  else: mean, invstd = unwrap(running_mean), unwrap(running_var).add(eps).rsqrt()
+  xhat = (input_t - mean.reshape(shape)) * invstd.reshape(shape)
+  grad_bias, grad_weight = grad_out_t.sum(axis=dims), (grad_out_t * xhat).sum(axis=dims)
+  grad_input = grad_out_t if not train else \
+    grad_out_t - (grad_bias.reshape(shape) + xhat * grad_weight.reshape(shape)) / (input_t.numel() // input_t.shape[1])
+  grad_input = grad_input * invstd.reshape(shape) * (unwrap(weight).reshape(shape) if weight is not None else 1)
+  return (wrap(grad_input) if output_mask[0] else None, wrap(grad_weight) if output_mask[1] else None,
+          wrap(grad_bias) if output_mask[2] else None)

 # _pad_circular is not CompositeImplicitAutograd (unlike reflect/replicate pad)
 # we need torch.autograd.Function with explicit AutogradPrivateUse1 registration
@@ -71,6 +71,48 @@ class TestTorchBackend(unittest.TestCase):
    a = a.as_strided((1,1,5,5), (50,50,7,1), storage_offset=21)
    np.testing.assert_equal(a.cpu().numpy().sum(-1), [[[115,150,185,220,255]]])

+  def test_storage_offset_of_computed_tensor(self):
+    # a computed result owns its storage, so a slice anywhere in its history must not shift the offset
+    a = torch.arange(8., device=device)
+    self.assertEqual((a[3:]+1).storage_offset(), 0)
+
+  def test_storage_offset_through_aliases(self):
+    a = torch.arange(8., device=device)[3:]
+    self.assertEqual(a.detach().storage_offset(), 3)
+    self.assertEqual(a.view(torch.int32).storage_offset(), 3)
+    torch.add(torch.ones(5, device=device), torch.ones(5, device=device), out=a)
+    self.assertEqual(a.detach().storage_offset(), 3)
+
+  @unittest.expectedFailure  # TODO: storage offset assumes a contiguous source, use UOp.contiguous_view_offset
+  def test_storage_offset_non_contiguous_source(self):
+    a = torch.arange(12., device=device).reshape(3,4)
+    self.assertEqual(a.permute(1,0)[1:].storage_offset(), 1)
+    self.assertEqual(a.flatten()[3:].flip(0).storage_offset(), 0)
+
+  def test_as_strided_explicit_zero_offset(self):
+    # storage_offset=0 is a real offset, not "unspecified": it must not fall back to the input's own offset
+    a = torch.arange(6., device=device)
+    np.testing.assert_equal(a[3:].as_strided((2,), (1,), 0).cpu().numpy(), [0,1])
+    np.testing.assert_equal(a[3:].as_strided((2,), (1,)).cpu().numpy(), [3,4])
+
+  def test_empty_strided_default_dtype(self):
+    self.assertEqual(torch.empty_strided((2,3), (1,2), device=device).dtype, torch.get_default_dtype())
+
+  @unittest.expectedFailure  # TODO: empty_strided ignores the requested strides, the backend treats everything as contiguous
+  def test_empty_strided_honors_strides(self):
+    self.assertEqual(tuple(torch.empty_strided((2,3), (1,2), device=device).stride()), (1,2))
+
+  @unittest.expectedFailure  # TODO: torch refuses an out= that overlaps an input, we compute silently
+  def test_out_overlapping_input_is_rejected(self):
+    x = torch.arange(6., device=device)
+    with self.assertRaises(RuntimeError): torch.add(x[:-1], 10, out=x[1:])
+
+  def test_out_disjoint_input_is_allowed(self):
+    # torch permits an out= that shares a base with an input as long as they do not overlap
+    x, xc = torch.arange(6., device=device), torch.arange(6.)
+    torch.add(x[:3], 10, out=x[3:]); torch.add(xc[:3], 10, out=xc[3:])
+    np.testing.assert_equal(x.cpu().numpy(), xc.numpy())
+
  def test_plus_inplace(self):
    a = torch.ones(4, device=device)
    b = torch.ones(4, device=device)
@@ -316,6 +358,21 @@ class TestTorchBackend(unittest.TestCase):
    assert b.shape == (4, 2, 3)
    np.testing.assert_equal(b.cpu().numpy(), a.cpu().numpy().transpose(2, 0, 1))

+  def test_batchnorm_backward_realized_stats(self):
+    # the saved stats are a function of input in training, so grad_input must flow through them even when handed in realized.
+    # the backward eps is unused in training: torch differentiates the save_invstd it was given
+    x0, g0 = torch.randn(8, 4, 3, 3), torch.randn(8, 4, 3, 3)
+    def run(dev, bwd_eps):
+      x, go = x0.to(dev), g0.to(dev)
+      w, b = torch.linspace(0.5, 2.0, 4).to(dev), torch.zeros(4, device=dev)
+      rm, rv = torch.zeros(4, device=dev), torch.ones(4, device=dev)
+      out, sm, si = torch.ops.aten.native_batch_norm(x, w, b, rm, rv, True, 0.1, 1e-5)
+      grads = torch.ops.aten.native_batch_norm_backward(go, x, w, rm, rv, sm.clone().detach(), si.clone().detach(),
+                                                        True, bwd_eps, [True,True,True])
+      return [t.cpu().numpy() for t in grads]
+    for bwd_eps in [1e-5, 0.3]:
+      for got, want in zip(run(device, bwd_eps), run("cpu", bwd_eps)): np.testing.assert_allclose(got, want, atol=1e-4, rtol=1e-3)
+
  def test_batchnorm_unsqueeze(self):
    bn = torch.nn.BatchNorm2d(4).to(device)
    x = torch.randn(8, 4, 3, 3, device=device)
@@ -742,6 +799,20 @@ class TestTorchBackend(unittest.TestCase):

 from tinygrad import Tensor
 class TestBackendHelpers(unittest.TestCase):
+  def test_unwrap_rejects_foreign_tensor(self):
+    # unwrap casts to the tiny impl, so a tensor from another backend must be refused rather than reinterpreted
+    with self.assertRaises(RuntimeError): extra.torch_backend.backend.unwrap(torch.ones(4))
+
+  def test_update_metadata_rejects_foreign_tensor(self):
+    # resizing a tensor we don't own would expose memory past its allocation
+    t = torch.ones(4)
+    with self.assertRaises(RuntimeError): extra.torch_backend.backend.mod.update_metadata(t, [8], [1], 0)
+    self.assertEqual(t.shape, (4,))
+
+  def test_unwrap_parameter_and_detached(self):
+    # nn.Parameter and detach rebuild the base OpaqueTensorImpl, which unwrap still has to accept
+    extra.torch_backend.backend.unwrap(torch.nn.Parameter(torch.ones(4, device="tiny")))
+    extra.torch_backend.backend.unwrap(torch.ones(4, device="tiny").detach())

  def test_calculate_storage_offset_no_shrink(self):
    t = Tensor.ones(3, 4)
@@ -124,16 +124,21 @@ at::Tensor wrap_tensor(py::object &py_obj, c10::ScalarType dtype, c10::DeviceInd
    sizes, strides, storage_offset);
 }

+// shallow_copy_and_detach (nn.Parameter, aten.detach) rebuilds the base OpaqueTensorImpl, so that is the type every tiny tensor has
+at::OpaqueTensorImpl<std::shared_ptr<c10::SafePyObject>> *tiny_impl(const at::Tensor &tensor) {
+  auto* impl = dynamic_cast<at::OpaqueTensorImpl<std::shared_ptr<c10::SafePyObject>>*>(tensor.unsafeGetTensorImpl());
+  TORCH_CHECK(impl != nullptr, "expected a tiny tensor, got a ", tensor.device().str(), " one. move it with .to(\"tiny\") first");
+  return impl;
+}
+
 py::object unwrap_tensor(const at::Tensor &tensor) {
-  auto* impl = tensor.unsafeGetTensorImpl();
-  auto* opaque_impl = static_cast<at::TinyOpaqueTensorImpl<std::shared_ptr<c10::SafePyObject>>*>(impl);
-  std::shared_ptr<c10::SafePyObject> tiny = opaque_impl->opaque_handle();
+  std::shared_ptr<c10::SafePyObject> tiny = tiny_impl(tensor)->opaque_handle();
  return py::reinterpret_borrow<py::object>(tiny->ptr(getPyInterpreter()));
 }

 void update_metadata(const at::Tensor &tensor, const std::vector<int64_t> &sizes,
                     const std::vector<int64_t> &strides, int64_t storage_offset) {
-  auto* impl = tensor.unsafeGetTensorImpl();
+  auto* impl = tiny_impl(tensor);
  impl->set_allow_tensor_metadata_change(true);
  impl->set_sizes_and_strides(sizes, strides, storage_offset);
 }
@@ -29,6 +29,7 @@
 \definecolor{axbrred}{HTML}{E53935}      % GROUP_REDUCE
 \definecolor{axyellow}{HTML}{F9A825}     % UPCAST
 \definecolor{axmagenta}{HTML}{7B1FA2}    % UNROLL
+\definecolor{axgreen}{HTML}{2E7D32}      % DEVICE

 \title{tinygrad: a single dialect from Tensor programs to Command Buffers}
 \author{tinygrad, Corp. \\ \texttt{[email protected]}}
@@ -79,6 +80,7 @@ All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op
 \op{Index} & $(T, i_0, i_1, \ldots)$ & --- & Index from left. $()$-shaped $i$ removes dim; $(k,)$-shaped makes it $k$. \\
 \op{Stack} & $(T_0, T_1, \ldots)$ & --- & Join along a newly created leading axis. All shapes must match. \\
 \op{Bitcast} & $(T,)$ & dtype & Reinterpret storage as target dtype; preserve total bytes. \\
+\op{Unshard} & $(T, R_0, R_1, \ldots)$ & axes $(a_0, a_1, \ldots)$ & Concatenate shards of \op{Range} $R_k$ along axis $a_k$; $R_k$ is outer. \\
 \bottomrule
 \end{tabular}

@@ -258,6 +260,7 @@ Every UOp has a \textbf{dtype}, \textbf{shape}, \textbf{device}, \textbf{addrspa
 \op{Const} & from arg & $()$ & \textsc{null} & $[v, v]$ \\
 \op{Param} & from arg & from $\mathrm{src}[0]$ & from arg & from src or dtype range \\[3pt]
 Movement ops & $\mathrm{src}[0].\mathrm{dtype}$ & (see op) & $\mathrm{src}[0].\mathrm{device}$ & $\mathrm{src}[0]$ \\
+\op{Unshard} & $\mathrm{src}[0].\mathrm{dtype}$ & $\mathrm{src}[0]$, each $a_k \times n_k$ & $\mathrm{src}[0].\mathrm{device}$ & $\mathrm{src}[0]$ \\
 \op{Reduce} & $\mathrm{src}[0].\mathrm{dtype}$ & remove first $n$ axes & $\mathrm{src}[0].\mathrm{device}$ & dtype range \\[3pt]
 \op{Cast} & from arg & $\mathrm{src}[0].\mathrm{shape}$ & $\mathrm{src}[0].\mathrm{device}$ & clamped to dtype \\
 \op{Bitcast} & from arg & $\mathrm{src}[0].\mathrm{shape}$ & $\mathrm{src}[0].\mathrm{device}$ & dtype range \\
@@ -283,9 +286,9 @@ $[a,A]$, $[b,B]$, $[c,C]$ denote min\_max of $\mathrm{src}[0]$, $\mathrm{src}[1]
 Default \emph{dtype range}: $[\mathrm{dtype\_min},\, \mathrm{dtype\_max}]$.

 \medskip
-\textbf{axis} tracks the multi-device sharding dimension. \op{Buffer} with $n$-tuple device: axis $= 0$ (device dim).
-\op{Reshape} remaps axis to preserve the shard boundary. \op{Permute} follows the permutation. \op{Expand} shifts axis right by $|\mathbf{n}|$.
-\op{Reduce} on the shard axis $\to$ \textsc{null} (shard axis is among the first $n$ axes). \op{Replicated} on the shard axis $\to$ \textsc{null}. \op{Copy} $\to$ \textsc{null}. ALU ops inherit from sources. Default: \textsc{null}.
+\textbf{sharding} tracks multi-device sharding as a set of (axis, \op{Range}) pairs. \op{Unshard} defines it: arg is the tuple of sharded axes, one \op{Range} in src per axis (positional: the $k$-th \op{Range} shards the $k$-th axis). \op{Buffer} with $n$-tuple device: sharded on axis $0$ (device dim). The single-axis convenience \textbf{axis} is \textsc{null} unless exactly one axis is sharded.
+\op{Reshape} remaps each sharded axis to preserve its shard boundary. \op{Permute} follows the permutation. \op{Expand} shifts all sharded axes right by $|\mathbf{n}|$.
+\op{Reduce} on a sharded axis drops it. \op{Replicated} on the shard axis $\to$ \textsc{null}. \op{Copy} $\to$ \textsc{null}. ALU ops inherit from sources. Default: \textsc{null}.

 %% ============================================================
 \subsection*{Kernel Optimizations (OptOps) \normalfont\small--- schedule-level transforms on kernel ranges}
@@ -297,6 +300,7 @@ Each kernel's iteration space is a set of \op{Range} axes. Every range has an \t
 \toprule
 \textbf{AxisType} & \textbf{Letter} & \textbf{Split from} & \textbf{Direction} & \textbf{Semantics} \\
 \midrule
+{\color{axgreen}\texttt{DEVICE}} & \texttt{d} & --- & --- & Multi-device sharding dimension. \\
 {\color{axblue}\texttt{GLOBAL}} & \texttt{g} & --- & --- & GPU global workgroup dimension. \\
 {\color{axcyan}\texttt{LOCAL}} & \texttt{l} & g, L & inner & Workgroup local dimension (shared memory). \\
 {\color{axbrcyan}\texttt{WARP}} & \texttt{w} & \multicolumn{2}{l}{(created by \op{TC})} & Warp-level lanes for tensor cores. \\
@@ -378,6 +382,8 @@ def scatter_add(T, idx, val):
 Let $D = (d_0, \ldots, d_{n-1})$ be an $n$-tuple device.
 \op{Copy} to an $n$-tuple device reshards with axis $= 0$. \op{Copy} never changes shape.

+\textbf{Sharding} splits a tensor along an axis across $n$ devices. It opens a \op{Range} of type \texttt{DEVICE} (a symbolic per-device index $d$), shrinks each device's view to its piece, then closes the range with \op{Unshard}$(T, R, a)$. The result is a logical tensor whose shape along axis $a$ is the full size; each device holds $1/n$ of it. \op{Unshard} is the inverse of sharding --- it marks the boundary between per-device computation and the logical multi-device tensor. The range need not be \texttt{DEVICE}; e.g.\ a \texttt{WARP} range closes the same way, concatenating per-lane shards along $a$ with the range as the outer factor. A tensor may be sharded along several axes at once: \op{Unshard}$(T, R_0, R_1, \ldots;\; a_0, a_1, \ldots)$ carries one \op{Range} per sharded axis, and every movement op maps each sharded axis independently.
+
 \begin{lstlisting}
 # T has shape (s,) on a single device.

@@ -169,11 +169,10 @@ def run_program_emu(instructions: list, n_lanes: int = 1) -> WaveState:
  return parse_output(bytes(out_buf), n_lanes)

 def run_program_hw(instructions: list, n_lanes: int = 1) -> WaveState:
-  """Run instructions on real AMD hardware via HIPCompiler and AMDProgram."""
-  from tinygrad.device import Device
-  from tinygrad.runtime.ops_amd import AMDProgram
+  """Run instructions on real AMD hardware via HIPCompiler and the AMD runtime."""
+  from tinygrad.device import Device, TinyELF
  from tinygrad.runtime.support.compiler_amd import HIPCompiler
-  from tinygrad.helpers import flat_mv
+  from tinygrad.helpers import Target, flat_mv

  dev = Device["AMD"]
  compiler = HIPCompiler(dev.arch)  # type: ignore[attr-defined]
@@ -223,7 +222,7 @@ amdhsa.kernels:
 """

  lib = compiler.compile(asm_src)
-  prg = AMDProgram(dev, "test", lib)  # type: ignore[arg-type]
+  prg = dev.runtime(TinyELF(lib, "test", Target("AMD", arch=dev.arch), ()))

  buf_sz = _out_bytes(n_lanes)
  out_gpu = dev.allocator.alloc(buf_sz)
@@ -5,7 +5,7 @@ gfx950 hardware when USE_HW=1.
 """
 import ctypes, struct, unittest
 import tinygrad.runtime.autogen.amd.cdna.ins as cdna
-from tinygrad.helpers import flat_mv
+from tinygrad.helpers import Target, flat_mv
 from tinygrad.renderer.amd.dsl import NULL
 from test.amd.hw.helpers import USE_HW, assemble
 from test.mockgpu.amd.emu import run_asm
@@ -42,8 +42,7 @@ def _run_emu(instructions: list, out_reg: int = 2) -> int:
  return out_buf[0]

 def _run_hw(instructions: list, out_reg: int = 2) -> int:
-  from tinygrad.device import Device
-  from tinygrad.runtime.ops_amd import AMDProgram
+  from tinygrad.device import Device, TinyELF
  from tinygrad.runtime.support.compiler_amd import HIPCompiler

  dev = Device["AMD"]
@@ -86,7 +85,7 @@ amdhsa.kernels:
 ...
 .end_amdgpu_metadata
 """
-  prg = AMDProgram(dev, "test", HIPCompiler(dev.arch).compile(asm_src))
+  prg = dev.runtime(TinyELF(HIPCompiler(dev.arch).compile(asm_src), "test", Target("AMD", arch=dev.arch), ()))
  prg(global_size=(1, 1, 1), local_size=(LANES, 1, 1), wait=True)
  out = bytearray(LANES * 4)
  dev.allocator._copyout(flat_mv(memoryview(out)), out_gpu)
@@ -6,7 +6,7 @@ when USE_HW=1.
 """
 import ctypes, unittest
 from tinygrad.runtime.autogen.amd.rdna3.ins import *
-from tinygrad.helpers import flat_mv
+from tinygrad.helpers import Target, flat_mv
 from test.amd.hw.helpers import USE_HW, assemble
 from test.mockgpu.amd.emu import run_asm

@@ -37,8 +37,7 @@ def _run_wave64_emu(instructions: list, out_reg: int = 1) -> list[int]:
  return list(out_buf)

 def _run_wave64_hw(instructions: list, out_reg: int = 1) -> list[int]:
-  from tinygrad.device import Device
-  from tinygrad.runtime.ops_amd import AMDProgram
+  from tinygrad.device import Device, TinyELF
  from tinygrad.runtime.support.compiler_amd import HIPCompiler

  dev = Device["AMD"]
@@ -84,7 +83,7 @@ amdhsa.kernels:
 .end_amdgpu_metadata
 """
  lib = compiler.compile(asm_src)
-  prg = AMDProgram(dev, "test", lib)  # type: ignore[arg-type]
+  prg = dev.runtime(TinyELF(lib, "test", Target("AMD", arch=dev.arch), ()))
  out_gpu = dev.allocator.alloc(WAVE64 * 4)
  prg(out_gpu, global_size=(1, 1, 1), local_size=(WAVE64, 1, 1), wait=True)
  out = bytearray(WAVE64 * 4)
@@ -5,7 +5,7 @@ real RDNA4 hardware when USE_HW=1.
 """
 import ctypes, unittest
 import tinygrad.runtime.autogen.amd.rdna4.ins as r4
-from tinygrad.helpers import flat_mv
+from tinygrad.helpers import Target, flat_mv
 from tinygrad.renderer.amd.dsl import NULL
 from test.amd.hw.helpers import USE_HW, assemble
 from test.mockgpu.amd.emu import run_asm
@@ -36,8 +36,7 @@ def _run_emu(instructions: list, out_reg: int = 2) -> list[int]:
  return list(out_buf)

 def _run_hw(instructions: list, out_reg: int = 2) -> list[int]:
-  from tinygrad.device import Device
-  from tinygrad.runtime.ops_amd import AMDProgram
+  from tinygrad.device import Device, TinyELF
  from tinygrad.runtime.support.compiler_amd import HIPCompiler

  dev = Device['AMD']
@@ -85,7 +84,7 @@ amdhsa.kernels:
 .end_amdgpu_metadata
 """
  lib = compiler.compile(asm_src)
-  prg = AMDProgram(dev, 'test', lib)
+  prg = dev.runtime(TinyELF(lib, "test", Target("AMD", arch=dev.arch), ()))
  out_gpu = dev.allocator.alloc(LANES * 4)
  prg(out_gpu, global_size=(1, 1, 1), local_size=(LANES, 1, 1), wait=True)
  out = bytearray(LANES * 4)
@@ -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).

@@ -36,7 +36,7 @@ def custom_add_var(A:UOp, B:UOp) -> UOp:
  A,B = A.flatten(), B.flatten()
  assert A.dtype == dtypes.uint32, f"buffer dtype must be uint32, got {A.dtype}"
  threads = UOp.special(A.numel(), "lidx0")
-  var = UOp.param(2, dtypes.weakint, vmin_vmax=(0, 10), name="var", addrspace=AddrSpace.ALU)
+  var = UOp.param(2, dtypes.int, vmin_vmax=(0, 10), name="var", addrspace=AddrSpace.ALU)
  insts = [
    s_load_b128(s[4:7], s[0:1]),
    s_load_b32(s[8], s[0:1], offset=0x10), # all threads load the same variable
@@ -11,8 +11,8 @@ from tinygrad.runtime.autogen.amd.rdna3.enum import VOP1Op, VOP2Op, SOP2Op, DSOp

 def _srcs():
  """Create minimal source variables for pcode parsing."""
-  def u32(v=0): return UOp.const(dtypes.uint32, v)
-  return {'S0': u32(), 'S1': u32(), 'S2': u32(), 'SCC': u32(), 'VCC': UOp.const(dtypes.uint64, 0), 'laneId': u32()}
+  def u32(v=0): return UOp.const(v, dtypes.uint32)
+  return {'S0': u32(), 'S1': u32(), 'S2': u32(), 'SCC': u32(), 'VCC': UOp.const(0, dtypes.uint64), 'laneId': u32()}

 class TestBasicParsing(unittest.TestCase):
  """Test basic pcode parsing for common instruction patterns."""
@@ -44,8 +44,8 @@ class TestWithSources(unittest.TestCase):

  def test_v_add_f32_with_sources(self):
    """Test V_ADD_F32 with actual float constants."""
-    s0 = UOp.const(dtypes.uint32, 0x3f800000)  # 1.0f
-    s1 = UOp.const(dtypes.uint32, 0x40000000)  # 2.0f
+    s0 = UOp.const(0x3f800000, dtypes.uint32)  # 1.0f
+    s1 = UOp.const(0x40000000, dtypes.uint32)  # 2.0f
    _, assigns = parse_pcode(PCODE[VOP2Op.V_ADD_F32_E32], {'S0': s0, 'S1': s1})
    self.assertEqual(len(assigns), 1)
    dest, val = assigns[0]
@@ -55,8 +55,8 @@ class TestWithSources(unittest.TestCase):

  def test_v_mul_f32_with_sources(self):
    """Test V_MUL_F32 with actual float constants."""
-    s0 = UOp.const(dtypes.uint32, 0x40000000)  # 2.0f
-    s1 = UOp.const(dtypes.uint32, 0x40400000)  # 3.0f
+    s0 = UOp.const(0x40000000, dtypes.uint32)  # 2.0f
+    s1 = UOp.const(0x40400000, dtypes.uint32)  # 3.0f
    _, assigns = parse_pcode(PCODE[VOP2Op.V_MUL_F32_E32], {'S0': s0, 'S1': s1})
    self.assertEqual(len(assigns), 1)
    dest, val = assigns[0]
@@ -67,36 +67,36 @@ class TestParseExpr(unittest.TestCase):

  def test_integer_literals(self):
    """Test parsing integer literals."""
-    self.assertEqual(parse_expr('0', {}).arg, 0)
-    self.assertEqual(parse_expr('42', {}).arg, 42)
-    self.assertEqual(parse_expr('42U', {}).arg, 42)
+    self.assertEqual(parse_expr('0', {}).val, 0)
+    self.assertEqual(parse_expr('42', {}).val, 42)
+    self.assertEqual(parse_expr('42U', {}).val, 42)

  def test_negative_integers(self):
    """Test parsing negative integer literals."""
    result = parse_expr('-1', {})
-    self.assertEqual(result.arg, -1)
+    self.assertEqual(result.val, -1)
    self.assertEqual(result.dtype, dtypes.int)

  def test_float_literals(self):
    """Test parsing float literals."""
    result = parse_expr('1.0F', {})
-    self.assertEqual(result.arg, 1.0)
+    self.assertEqual(result.val, 1.0)
    self.assertEqual(result.dtype, dtypes.float32)

  def test_hex_literals(self):
    """Test parsing hex literals."""
    result = parse_expr('0xFF', {})
-    self.assertEqual(result.arg, 255)
+    self.assertEqual(result.val, 255)

  def test_variable_lookup(self):
    """Test variable lookup in parse_expr."""
-    vrs = {'x': UOp.const(dtypes.uint32, 42)}
+    vrs = {'x': UOp.const(42, dtypes.uint32)}
    result = parse_expr('x', vrs)
-    self.assertEqual(result.arg, 42)
+    self.assertEqual(result.val, 42)

  def test_binary_ops(self):
    """Test parsing binary operations."""
-    vrs = {'a': UOp.const(dtypes.uint32, 10), 'b': UOp.const(dtypes.uint32, 5)}
+    vrs = {'a': UOp.const(10, dtypes.uint32), 'b': UOp.const(5, dtypes.uint32)}

    # Addition
    result = parse_expr('a + b', vrs)
@@ -105,11 +105,11 @@ class TestParseExpr(unittest.TestCase):
    # Subtraction with constant folding
    result = parse_expr('10 - 5', {})
    self.assertEqual(result.op, Ops.CONST)
-    self.assertEqual(result.arg, 5)
+    self.assertEqual(result.val, 5)

  def test_ternary(self):
    """Test parsing ternary expressions."""
-    vrs = {'cond': UOp.const(dtypes.bool, True), 'a': UOp.const(dtypes.uint32, 1), 'b': UOp.const(dtypes.uint32, 0)}
+    vrs = {'cond': UOp.const(True), 'a': UOp.const(1, dtypes.uint32), 'b': UOp.const(0, dtypes.uint32)}
    result = parse_expr('cond ? a : b', vrs)
    self.assertEqual(result.op, Ops.WHERE)

@@ -127,7 +127,7 @@ class TestForLoopParsing(unittest.TestCase):
  def test_clz_parsing(self):
    """Test CLZ pcode parsing produces correct structure."""
    pcode = PCODE[VOP1Op.V_CLZ_I32_U32_E32]
-    S0 = UOp.const(dtypes.uint32, 0xFFFFFFFF)  # All ones - CLZ should be 0
+    S0 = UOp.const(0xFFFFFFFF, dtypes.uint32)  # All ones - CLZ should be 0
    _vrs, assigns = parse_pcode(pcode, {'S0': S0})

    self.assertEqual(len(assigns), 1)
@@ -139,7 +139,7 @@ class TestForLoopParsing(unittest.TestCase):
  def test_clz_with_zero(self):
    """Test CLZ with input 0 - should return -1."""
    pcode = PCODE[VOP1Op.V_CLZ_I32_U32_E32]
-    S0 = UOp.const(dtypes.uint32, 0)
+    S0 = UOp.const(0, dtypes.uint32)
    _vrs, assigns = parse_pcode(pcode, {'S0': S0})

    # Check that the innermost value (default) is -1 (may be wrapped in CAST)
@@ -150,7 +150,7 @@ class TestForLoopParsing(unittest.TestCase):
    # Unwrap CAST if present
    while val.op == Ops.CAST:
      val = val.src[0]
-    self.assertEqual(val.arg, -1)
+    self.assertEqual(val.val, -1)

  def test_ctz_parsing(self):
    """Test CTZ pcode parsing."""
@@ -158,7 +158,7 @@ class TestForLoopParsing(unittest.TestCase):
    if pcode is None:
      self.skipTest("V_CTZ_I32_B32_E32 pcode not available")

-    S0 = UOp.const(dtypes.uint32, 1)  # LSB set - CTZ should be 0
+    S0 = UOp.const(1, dtypes.uint32)  # LSB set - CTZ should be 0
    _vrs, assigns = parse_pcode(pcode, {'S0': S0})
    self.assertEqual(len(assigns), 1)

@@ -169,8 +169,8 @@ class TestDSPcodePatterns(unittest.TestCase):
    """Test GLOBAL_ATOMIC_ADD_F32 keeps memory values in float dtype."""
    vmem = UOp.param(2, dtypes.uint32, (1024,))
    srcs = {
-      'ADDR': UOp.const(dtypes.uint64, 0),
-      'DATA': UOp.const(dtypes.uint32, 0x3f800000),
+      'ADDR': UOp.const(0, dtypes.uint64),
+      'DATA': UOp.const(0x3f800000, dtypes.uint32),
      '_vmem': vmem,
    }

@@ -199,8 +199,8 @@ class TestDSPcodePatterns(unittest.TestCase):
    """Test MEM[addr].type read expression parsing."""
    # Create a mock LDS buffer
    lds = UOp.param(3, dtypes.uint32, (16384,))
-    addr = UOp.const(dtypes.uint32, 0)
-    vrs = {'_lds': lds, 'ADDR': addr, 'OFFSET': UOp.const(dtypes.uint32, 0)}
+    addr = UOp.const(0, dtypes.uint32)
+    vrs = {'_lds': lds, 'ADDR': addr, 'OFFSET': UOp.const(0, dtypes.uint32)}

    result = parse_expr('MEM[ADDR + OFFSET].b32', vrs)
    # Should be an INDEX operation into LDS
@@ -212,13 +212,13 @@ class TestDSPcodePatterns(unittest.TestCase):
    self.assertIsNotNone(pcode)
    assert pcode is not None
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 0),
-      'OFFSET0': UOp.const(dtypes.uint32, 0),
-      'OFFSET1': UOp.const(dtypes.uint32, 1),
-      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
-      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+      'ADDR': UOp.const(0, dtypes.uint32),
+      'OFFSET0': UOp.const(0, dtypes.uint32),
+      'OFFSET1': UOp.const(1, dtypes.uint32),
+      'DATA': UOp.const(0xAAAAAAAA, dtypes.uint32),
+      'DATA2': UOp.const(0xBBBBBBBB, dtypes.uint32),
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # Should have 2 MEM write assignments
    self.assertEqual(len(assigns), 2)
@@ -235,12 +235,12 @@ class TestDSPcodePatterns(unittest.TestCase):
    assert pcode is not None
    lds = UOp.param(3, dtypes.uint32, (16384,))
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 0),
-      'OFFSET0': UOp.const(dtypes.uint32, 0),
-      'OFFSET1': UOp.const(dtypes.uint32, 1),
+      'ADDR': UOp.const(0, dtypes.uint32),
+      'OFFSET0': UOp.const(0, dtypes.uint32),
+      'OFFSET1': UOp.const(1, dtypes.uint32),
      '_lds': lds,
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # Should have 2 RETURN_DATA assignments
    self.assertEqual(len(assigns), 2)
@@ -252,36 +252,36 @@ class TestDSPcodePatterns(unittest.TestCase):
    pcode = PCODE.get(DSOp.DS_STORE_2ADDR_B32)
    assert pcode is not None
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 100),
-      'OFFSET0': UOp.const(dtypes.uint32, 2),
-      'OFFSET1': UOp.const(dtypes.uint32, 5),
-      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
-      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+      'ADDR': UOp.const(100, dtypes.uint32),
+      'OFFSET0': UOp.const(2, dtypes.uint32),
+      'OFFSET1': UOp.const(5, dtypes.uint32),
+      'DATA': UOp.const(0xAAAAAAAA, dtypes.uint32),
+      'DATA2': UOp.const(0xBBBBBBBB, dtypes.uint32),
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # Check addresses: 100 + 2*4 = 108, 100 + 5*4 = 120
    # assigns[i][1] is (addr, val) tuple for MEM writes; mypy sees UOp
-    self.assertEqual(assigns[0][1][0].simplify().arg, 108)  # type: ignore[index]
-    self.assertEqual(assigns[1][1][0].simplify().arg, 120)  # type: ignore[index]
+    self.assertEqual(assigns[0][1][0].simplify().val, 108)  # type: ignore[index]
+    self.assertEqual(assigns[1][1][0].simplify().val, 120)  # type: ignore[index]

  def test_ds_store_data_values(self):
    """Test DS_STORE_2ADDR_B32 uses correct data values."""
    pcode = PCODE.get(DSOp.DS_STORE_2ADDR_B32)
    assert pcode is not None
    srcs = {
-      'ADDR': UOp.const(dtypes.uint32, 0),
-      'OFFSET0': UOp.const(dtypes.uint32, 0),
-      'OFFSET1': UOp.const(dtypes.uint32, 1),
-      'DATA': UOp.const(dtypes.uint32, 0xAAAAAAAA),
-      'DATA2': UOp.const(dtypes.uint32, 0xBBBBBBBB),
+      'ADDR': UOp.const(0, dtypes.uint32),
+      'OFFSET0': UOp.const(0, dtypes.uint32),
+      'OFFSET1': UOp.const(1, dtypes.uint32),
+      'DATA': UOp.const(0xAAAAAAAA, dtypes.uint32),
+      'DATA2': UOp.const(0xBBBBBBBB, dtypes.uint32),
    }
-    srcs['laneId'] = UOp.const(dtypes.uint32, 0)
+    srcs['laneId'] = UOp.const(0, dtypes.uint32)
    _, assigns = parse_pcode(pcode, srcs)
    # assigns[i][1] is (addr, val) tuple for MEM writes; mypy sees UOp
    # DATA[31:0] should preserve the value
-    self.assertEqual(assigns[0][1][1].simplify().arg, 0xAAAAAAAA)  # type: ignore[index]
-    self.assertEqual(assigns[1][1][1].simplify().arg, 0xBBBBBBBB)  # type: ignore[index]
+    self.assertEqual(assigns[0][1][1].simplify().val, 0xAAAAAAAA)  # type: ignore[index]
+    self.assertEqual(assigns[1][1][1].simplify().val, 0xBBBBBBBB)  # type: ignore[index]

 class TestConditionalParsing(unittest.TestCase):
  """Test conditional (if/elsif/else) pcode parsing."""
@@ -290,9 +290,9 @@ class TestConditionalParsing(unittest.TestCase):
    """Test parsing ternary expression (which becomes WHERE)."""
    # S_CSELECT_B32: D0.u32 = SCC ? S0.u32 : S1.u32
    pcode = PCODE[SOP2Op.S_CSELECT_B32]
-    s0 = UOp.const(dtypes.uint32, 10)
-    s1 = UOp.const(dtypes.uint32, 20)
-    scc = UOp.const(dtypes.uint32, 1)
+    s0 = UOp.const(10, dtypes.uint32)
+    s1 = UOp.const(20, dtypes.uint32)
+    scc = UOp.const(1, dtypes.uint32)
    _vrs, assigns = parse_pcode(pcode, {'S0': s0, 'S1': s1, 'SCC': scc})
    self.assertEqual(len(assigns), 1)
    dest, val = assigns[0]
@@ -305,26 +305,26 @@ class TestConcatWidthParsing(unittest.TestCase):

  def test_permlanex16_altrow_concat(self):
    for row, expected in [(0, 1), (1, 0), (2, 3), (3, 2)]:
-      parsed = parse_expr('{ row[1], ~row[0] }', {'row': UOp.const(dtypes.uint32, row)})
-      self.assertEqual(parsed.simplify().arg, expected)
+      parsed = parse_expr('{ row[1], ~row[0] }', {'row': UOp.const(row, dtypes.uint32)})
+      self.assertEqual(parsed.simplify().val, expected)

  def test_permlane64_altlane_concat(self):
    for lane, expected in [(0, 32), (1, 33), (31, 63), (32, 0), (63, 31)]:
-      parsed = parse_expr('{ ~lane[5], lane[4:0] }', {'lane': UOp.const(dtypes.uint32, lane)})
-      self.assertEqual(parsed.simplify().arg, expected)
+      parsed = parse_expr('{ ~lane[5], lane[4:0] }', {'lane': UOp.const(lane, dtypes.uint32)})
+      self.assertEqual(parsed.simplify().val, expected)

  def test_permlane64_wave64_pcode_indices(self):
    vgpr = UOp.param(0, dtypes.uint32, (256,))
    srcs = {
-      'SRC0': UOp.const(dtypes.uint32, 0),
-      'VDST': UOp.const(dtypes.uint32, 1),
-      'EXEC_LO': UOp.const(dtypes.uint32, 0xFFFFFFFF),
-      'EXEC': UOp.const(dtypes.uint64, 0xFFFFFFFFFFFFFFFF),
+      'SRC0': UOp.const(0, dtypes.uint32),
+      'VDST': UOp.const(1, dtypes.uint32),
+      'EXEC_LO': UOp.const(0xFFFFFFFF, dtypes.uint32),
+      'EXEC': UOp.const(0xFFFFFFFFFFFFFFFF, dtypes.uint64),
      '_vgpr': vgpr,
      '_wave_size': 64,
-      'S0': UOp.const(dtypes.uint32, 0),
-      'S1': UOp.const(dtypes.uint32, 0),
-      'S2': UOp.const(dtypes.uint32, 0),
+      'S0': UOp.const(0, dtypes.uint32),
+      'S1': UOp.const(0, dtypes.uint32),
+      'S2': UOp.const(0, dtypes.uint32),
    }

    def load_idx(v: UOp) -> int:
@@ -333,12 +333,12 @@ class TestConcatWidthParsing(unittest.TestCase):
      self.assertEqual(simp.src[0].op, Ops.INDEX)
      idx = simp.src[0].src[1].simplify()
      self.assertEqual(idx.op, Ops.CONST)
-      return idx.arg
+      return idx.val

    _, assigns = parse_pcode(PCODE[VOP1Op.V_PERMLANE64_B32_E32], srcs)
    self.assertEqual(len(assigns), 64)
    for lane, (dst_idx, src_idx) in {0: (64, 32), 31: (95, 63), 32: (96, 0), 63: (127, 31)}.items():
-      self.assertEqual(assigns[lane][1][0].simplify().arg, dst_idx)  # type: ignore[index]
+      self.assertEqual(assigns[lane][1][0].simplify().val, dst_idx)  # type: ignore[index]
      self.assertEqual(load_idx(assigns[lane][1][1]), src_idx)  # type: ignore[index]

 class TestAllPcode(unittest.TestCase):
@@ -346,7 +346,7 @@ class TestAllPcode(unittest.TestCase):

  def _make_srcs(self):
    """Create dummy source variables for pcode parsing."""
-    u32, u64 = lambda v=0: UOp.const(dtypes.uint32, v), lambda v=0: UOp.const(dtypes.uint64, v)
+    u32, u64 = lambda v=0: UOp.const(v, dtypes.uint32), lambda v=0: UOp.const(v, dtypes.uint64)
    lds = UOp.param(3, dtypes.uint32, (16384,))
    return {'laneId': u32(), 'laneID': u32(), 'S0': u32(), 'S1': u32(), 'S2': u32(), 'S3': u32(), 'SRC0': u32(),
            'D0': u32(), 'D1': u32(), 'DST': u32(), 'VDST': u32(), 'SDST': u32(),
@@ -358,7 +358,7 @@ class TestAllPcode(unittest.TestCase):
            'M0': u32(), 'PC': u64(), 'DENORM': u32(1), 'ROUND_MODE': u32(), 'ROUND_TOWARD_ZERO': u32(),
            'ROUND_NEAREST_EVEN': u32(), 'WAVE_STATUS': u32(),
            'MAX_FLOAT_F32': u32(0x7f7fffff), 'Unsigned': u32(1), 'clampedLOD': u32(),
-            '_lds': lds, '_vmem': lds, '_active': UOp.const(dtypes.bool, True)}
+            '_lds': lds, '_vmem': lds, '_active': UOp.const(True)}

  def _parse_all_pcode(self, pcode_dict, arch: str, min_pct: float):
    """Parse all pcode. RuntimeError = parser limitation (ok), other exceptions = real bugs."""
@@ -7,15 +7,16 @@ class TestMockGPUInvalidInstruction(unittest.TestCase):
    """Test that unsupported instructions raise immediately through the full MOCKGPU stack."""
    test_code = '''
 import struct
+from dataclasses import replace
 from tinygrad import Device, Tensor
 from tinygrad.engine.realize import compile_linear
-from tinygrad.runtime.ops_amd import AMDProgram

 dev = Device["AMD"]
 a = Tensor([1.0]).realize()
 b = a + 1
 linear = compile_linear(b.schedule_linear())
-lib = bytearray(linear.src[-1].src[0].src[3].arg)
+compiled_prg = linear.src[-1].src[0]
+lib = bytearray(compiled_prg.src[3].arg)

 # Find s_endpgm (0xBFB00000) and replace with V_MOVRELD_B32 (op=66) which has no pcode
 # VOP1 encoding: bits[31:25]=0x7E, op=bits[16:9], so op=66 -> 66<<9 = 0x8400
@@ -27,7 +28,7 @@ for i in range(0, len(lib) - 4, 4):
    break
 assert found, "s_endpgm not found"

-patched_prg = AMDProgram(dev, "patched", bytes(lib))
+patched_prg = dev.runtime(replace(compiled_prg.to_elf(), name="patched", lib=bytes(lib)))
 b.uop.buffer.allocate()
 patched_prg(b.uop.buffer._buf, a.uop.buffer._buf, global_size=(1,1,1), local_size=(1,1,1))
 dev.synchronize()
@@ -4,13 +4,13 @@ from tinygrad import Tensor, GlobalCounters, dtypes, nn, Device, Variable
 from tinygrad.helpers import Context, getenv, DEV
 from tinygrad.engine.realize import run_linear, estimate_uop, compile_linear
 from tinygrad.renderer.ptx import PTXRenderer
-from test.helpers import needs_second_gpu
+from test.helpers import needs_second_gpu, check_schedule, assert_kernel_count, KernelCountException

 class TestArange(unittest.TestCase):
  def _get_flops(self, tensor, desired):
    GlobalCounters.reset()
    linear = compile_linear(tensor.schedule_linear())
-    self.assertEqual(len(linear.src), 1)
+    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
    run_linear(linear)
    np.testing.assert_equal(tensor.numpy(), desired)
    return estimate_uop(linear.src[-1]).ops
@@ -19,6 +19,12 @@ class TestArange(unittest.TestCase):
    self.assertLess(self._get_flops(Tensor.arange(256).clone(), np.arange(256)), 256*4)
    self.assertLess(self._get_flops(Tensor.arange(2560).clone(), np.arange(2560)), 2560*4)

+  def test_cat_complexity(self):
+    x = Tensor.arange(2**10) + Tensor.empty((), dtype=dtypes.uint32)
+    out = x.cat(x).cat(Tensor.empty(1, dtype=dtypes.uint32))
+    linear = compile_linear(out.schedule_linear())
+    self.assertLessEqual(estimate_uop(linear.src[-1]).ops, out.numel()*20)
+
  @unittest.skipIf(Device.DEFAULT == "CL", "flaky in CI")
  def test_arange_cumsum(self):
    np.testing.assert_equal(Tensor.arange(513).cumsum(0).numpy(), np.arange(513).cumsum())
@@ -49,8 +55,7 @@ class TestIndexing(unittest.TestCase):
    with Context(NOOPT=1):
      GlobalCounters.reset()
      out = ((Tensor.arange(1,16385)-1)*needle).sum()
-      linear, var_vals = out.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(out, 1)
      run_linear(linear, var_vals)
    self.assertEqual(out.item(), 1337)

@@ -66,8 +71,7 @@ class TestIndexing(unittest.TestCase):
      reshape_dataset = dataset.T.reshape(1, DDIM, DSET, 1).expand(4, DDIM, DSET, 1)
      full = (rng==idxs).where(reshape_dataset, Tensor.zeros(4, DDIM, DSET, 1, buffer=False))
      X = full.sum(axis=(2,3))
-      linear, var_vals = X.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(X, 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -92,8 +96,7 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      X = dataset[idxs]
      assert X.shape == (4,DDIM)
-      linear, var_vals = X.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(X, 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -107,8 +110,7 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      X = dataset[idxs]
      assert X.shape == (4,DDIM)
-      linear, var_vals = X.linear_with_vars()
-      self.assertEqual(len(linear.src), 1)
+      linear, var_vals = check_schedule(X, 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops} != {4*DSET}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -151,7 +153,7 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      z = emb(x).realize()
      self.assertLessEqual(GlobalCounters.global_ops, op_limit)
-      self.assertEqual(GlobalCounters.kernel_count, 2)
+      assert_kernel_count(2)
    if getenv("CHECK", 1):
      import torch
      with torch.no_grad():
@@ -251,7 +253,7 @@ class TestIndexing(unittest.TestCase):
    xq_rope, _ = apply_rotary_emb(xq, xq, freqs_cis)
    xq_rope.sum().backward()
    linear = compile_linear(wq.grad.schedule_linear())
-    assert len(linear.src) == 1, f"expected one kernel for backward, got: {len(linear.src)}"
+    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
    bwd_ops = estimate_uop(linear.src[0]).ops
    expected_ops = bs*seqlen*dim*dim*ops_scale
    print(f"rope matmul bwd ({dtype}): {GlobalCounters.kernel_count} kernels, {bwd_ops:,} ops")
@@ -150,6 +150,44 @@ class TestAsmGEMM(unittest.TestCase):
    with self.assertRaisesRegex(AssertionError, "not a multiple"):
      verify_asm_gemm(1, 256, 1000, 256)

+class TestMXFP4(unittest.TestCase):
+  def setUp(self):
+    if not is_cdna4() or DEV.interface.startswith("MOCK"):
+      self.skipTest("requires real amd machine")
+
+  def test_quantize(self):
+    import numpy as np
+    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
+    M = N = K = 256
+    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().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)
+    a = Tensor.empty(M, K, dtype=dtypes.bfloat16)
+    b = Tensor.empty(N, K, dtype=dtypes.bfloat16)
+    asm_gemm(a, b.T, mxfp4=True).realize()
+
 # test the Asm GEMM with Llama shapes, only run on the real machine for speed

@unittest.skipUnless(has_hipcc(), "requires hipcc to compile")
@@ -6,11 +6,11 @@ from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.uop.ops import KernelInfo

 def call_out_kernel(F:UOp, C:UOp) -> UOp:
-  call = F[0].load().call(UOp.const(dtypes.int, 3), C[0], ret_dtype=dtypes.void)
+  call = F[0].load().call(UOp.const(3).cast(dtypes.int), C[0], ret_dtype=dtypes.void)
  return C.after(call)[1].store(C.after(call)[0].load() + 1).sink(arg=KernelInfo(name="call_out"))

 def call_ret_kernel(F:UOp, C:UOp) -> UOp:
-  val = F[0].load().call(UOp.const(dtypes.int, 21), ret_dtype=dtypes.int)
+  val = F[0].load().call(UOp.const(21).cast(dtypes.int), ret_dtype=dtypes.int)
  return C[0].store(val * 2).sink(arg=KernelInfo(name="call_ret"))

@unittest.skipUnless(isinstance(Device["CPU"].renderer, CStyleLanguage), "TODO: CALL is rendered in C style only")
@@ -16,7 +16,7 @@ def _check_ast_count(desired_count:int, t:Tensor):

 class TestMovedConstFolding(unittest.TestCase):
  def test_contiguous_deviceless_const(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0)).contiguous()
+    t = Tensor(UOp.const(2.0, dtypes.float)).contiguous()
    self.assertIs(t.uop.op, Ops.CONST)
    self.assertIsNone(t.uop.device)

@@ -169,7 +169,7 @@ class TestMultiConstFolding(unittest.TestCase):
 class TestThreefryConstFolding(unittest.TestCase):
  def test_threefry(self):
    # THREEFRY(const,const) folds to a const once decomposed
-    x = threefry2x32(UOp.const(dtypes.uint64, 5), UOp.const(dtypes.uint64, 10))
+    x = threefry2x32(UOp.const(5, dtypes.uint64), UOp.const(10, dtypes.uint64))
    self.assertIs(x.simplify().op, Ops.CONST)

 class TestTautologicalCompare(unittest.TestCase):
@@ -1,7 +1,10 @@
 import unittest
 from tinygrad import Tensor, UOp, GlobalCounters, Context, Device
+import numpy as np
 from tinygrad.dtype import AddrSpace, dtypes, Invalid
 from tinygrad.uop.ops import KernelInfo, AxisType, Ops
+from tinygrad.renderer.ptx import PTXRenderer
+from test.helpers import assert_kernel_count

 # **** kernels ****

@@ -55,7 +58,7 @@ def flip_contract_kernel(dest:UOp, src:UOp):
  return store.end(i, j).sink(arg=KernelInfo(name=f"flip_contract_{dest.numel()}", opts_to_apply=()))

 def slice_sum_kernel(dest:UOp, src:UOp):
-  G = UOp.range(src.shape[0], 0)
+  G = UOp.range(src.shape[0], 0, dtype=dtypes.int)
  slice_src = src[G, :]
  reg = UOp.placeholder((1,), dest.dtype, 0, addrspace=AddrSpace.REG)
  reg = reg.after(G)[0].set(0)
@@ -117,13 +120,20 @@ class TestCustomKernel(unittest.TestCase):
    out = c.flatten().tolist()
    assert all(x == 2 for x in out), "all 2"

+  def test_duplicate_call_arg(self):
+    x = Tensor.arange(4).clone().realize()
+    x = Tensor.custom_kernel(x, x, fxn=custom_add_one_kernel)[0]
+    # webgpu silently errors when a kernel has duplicate buffer args, so the list stays the same.
+    # https://gpuweb.github.io/gpuweb/#abstract-opdef-encoder-bind-groups-alias-a-writable-resource
+    self.assertEqual(x.tolist(), [1, 2, 3, 4] if Device.DEFAULT != "WEBGPU" else [0, 1, 2, 3])
+
  def test_simple_sharded(self):
    devs = ("CPU:0", "CPU:1")

    a = Tensor.ones(16, 16).contiguous().shard(devs, axis=0)
    b = Tensor.ones(16, 16).contiguous().shard(devs, axis=0)
    # ugly construction to get a sharded empty tensor
-    c = Tensor(Tensor.empty(8, 16, device=devs).uop.multi(0), device=devs)
+    c = Tensor(Tensor.empty(8, 16, device=devs).uop.unshard(0), device=devs)
    c = Tensor.custom_kernel(c,a,b, fxn=custom_elementwise_add_kernel)[0]
    out = c.flatten().tolist()
    assert all(x == 2 for x in out), "all 2"
@@ -132,7 +142,7 @@ class TestCustomKernel(unittest.TestCase):
    # PYTHON backend explicitly checks for OOB access for wrong multi shape regression
    devs = ("PYTHON:0", "PYTHON:1")
    a = Tensor.ones(4, 4).contiguous().shard(devs, axis=0)
-    c = Tensor(Tensor.empty(2, 4, device=devs).uop.multi(0), device=devs)
+    c = Tensor(Tensor.empty(2, 4, device=devs).uop.unshard(0), device=devs)
    c = Tensor.custom_kernel(c, a, fxn=custom_add_one_kernel)[0]
    assert (c == 2).all().item()

@@ -180,6 +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)
@@ -206,7 +222,7 @@ class TestCustomKernel(unittest.TestCase):
    N = 16
    a = Tensor.randn(N, N).shard_(devs, axis=0)
    b = Tensor.randn(N, N).to(devs)
-    c = Tensor(Tensor.empty(N//2, N, device=devs).uop.multi(0), device=devs)
+    c = Tensor(Tensor.empty(N//2, N, device=devs).uop.unshard(0), device=devs)
    tst = Tensor.custom_kernel(c, a, b, fxn=custom_gemm)[0]
    self.assertTrue(tst.allclose(a@b, atol=1e-3).item())

@@ -267,7 +283,7 @@ class TestCustomKernel(unittest.TestCase):

    GlobalCounters.reset()
    out.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 5)
+    assert_kernel_count(5)

  def test_simple_reshape(self):
    a = Tensor.ones(2,3,4).realize()
@@ -277,7 +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()}"
-    self.assertEqual(GlobalCounters.kernel_count, 3)
+    assert_kernel_count(2)

  def test_multi_after_schedule_order(self):
    """Test correct scheduling order when custom_kernel has multiple outputs.
@@ -323,16 +339,16 @@ class TestCustomKernel(unittest.TestCase):
  def test_multi_invalids_custom_kernel_no_copy(self):
    devs = ("CPU:0", "CPU:1")
    a = Tensor.ones(4, 4).shard(devs, axis=0).realize()
-    c = Tensor(UOp.const(dtypes.float, Invalid, shape=(2, 4)).clone(device=devs).multi(0), device=devs)
+    c = Tensor(Tensor.invalids(2, 4, dtype=dtypes.float, device=devs).uop.unshard(0), device=devs)
    c = Tensor.custom_kernel(c, a, fxn=custom_add_one_kernel)[0]
    GlobalCounters.reset()
    c.realize()
-    self.assertEqual(GlobalCounters.kernel_count, len(devs))
+    assert_kernel_count(len(devs))
    self.assertTrue((c == 2).all().item())

  def test_partial_invalid_store_keeps_uncovered_reads(self):
    x = Tensor([10., 20., 30., 40.])
-    after = x.uop.after(x.uop.shrink(((0, 2),)).store(UOp.const(dtypes.float, Invalid, shape=(2,))))
+    after = x.uop.after(x.uop.shrink(((0, 2),)).store(Invalid))
    self.assertEqual(Tensor(after).contiguous().tolist(), [10., 20., 30., 40.])

  def test_multi_after_invalid_store_dep_removed(self):
@@ -388,17 +404,15 @@ class TestCustomKernel(unittest.TestCase):
    y = Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
    if use_custom:
      z = Tensor.empty_like(x)
-      z = Tensor.custom_kernel(y, y.T.T, fxn=custom_add_one_kernel)[0]
+      z = Tensor.custom_kernel(z, y.T.T, fxn=custom_add_one_kernel)[0]
    else: z = y.T.T+1
    GlobalCounters.reset()
    z.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    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
@@ -409,24 +423,213 @@ class TestCustomKernel(unittest.TestCase):
    GlobalCounters.reset()
    y = run(x[0]).realize()
    # it's copying the input and the output
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    # TODO: subbuffer usage has runtime specific behavior, this will be fixed after the removal of SLICE.
+    assert_kernel_count(2 if y.device in ("CL", "WEBGPU") else 1)
    self.assertEqual(y.tolist(), [1, 2, 3, 4])

  @Context(DEV="CPU")
  def test_simple_from_source(self):
-    a = Tensor([0., 1., 2.]).realize()
-
-    src = "void test_src(float* restrict a) { a[0] = 1.0; }"
+    a = Tensor.arange(4).clone().realize()
+    src = "void test_src(int* restrict a) { a[0] = 1; }"
    # 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)))
+      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).clone(), a.reshape(2, 2).T, fxn=custom_src_kernel)[0]
+    self.assertEqual(a.tolist(), [[1, 1], [2, 3]])

-    a = Tensor.custom_kernel(a, fxn=custom_src_kernel)[0]
-    self.assertEqual(a.tolist(), [1., 1., 2.])
+  @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)
+      col = UOp.range(A.shape[1], 1)
+      return A[row, col].store(A[row].max(axis=0)).end(col).end(row).sink(arg=KernelInfo(name=f"assign_row_max_{A.numel()}"))
+    a = Tensor.arange(4).clone().realize()
+    a = Tensor.custom_kernel(a.reshape(2, 2).T, fxn=custom_assign_row_max_kernel)[0]
+    self.assertEqual(a.flatten().tolist(), [2, 2, 3, 3])
+    self.assertEqual(a.shape, (2, 2))
+
+class TestCustomKernelInput(unittest.TestCase):
+  def _test_mop(self, mop_fxn, max_kernels):
+    # default: input is BUFFER
+    x = mop_fxn(Tensor.arange(32).clone("CPU").realize())
+    y = Tensor.custom_kernel(Tensor.empty_like(x), x, fxn=custom_add_one_kernel)[0]
+    GlobalCounters.reset()
+    y.realize()
+    kernel_count = GlobalCounters.kernel_count
+    self.assertEqual(y.tolist(), x.add(1).tolist())
+    self.assertLessEqual(kernel_count, max_kernels)
+    # same test with @function, input is PARAM
+    from tinygrad import function
+    x0 = Tensor.arange(32).clone("CPU").realize()
+    @function(precompile=True)
+    def run(a:Tensor) -> Tensor:
+      xv = mop_fxn(a)
+      y = Tensor.invalids(*xv.shape, dtype=xv.dtype, device=a.device)
+      return Tensor.custom_kernel(y, xv, fxn=custom_add_one_kernel)[0]
+    GlobalCounters.reset()
+    y = run(x0).realize()
+    kernel_count = GlobalCounters.kernel_count
+    self.assertEqual(y.tolist(), mop_fxn(x0).add(1).tolist())
+    self.assertLessEqual(kernel_count, max_kernels)
+
+  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=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)
+  def test_2d_shrink(self): self._test_mop(lambda x: x.reshape(4, 8)[:, 2:6], max_kernels=3)
+  def test_expand(self): self._test_mop(lambda x: x.reshape(16, 2)[:, :1].expand(16, 2), max_kernels=3)
+
+class TestUnshardIndex(unittest.TestCase):
+  """Regression tests for INDEX on UNSHARD (fragment) resolution in schedule/multi.py.
+
+  A fragment is a per-thread REG buffer wrapped in UNSHARD over LOCAL thread ranges.
+  index_multi must resolve an INDEX on the UNSHARD view into an INDEX on the per-thread
+  shard. Two ownership patterns must work:
+    contiguous: idx = rng*shard_sz + local   (thread rng owns [rng*shard_sz, ...))
+    strided:    idx = rng + ir*shard_sz      (thread rng owns {rng, rng+shard_sz, ...})
+  """
+  def _run(self, kernel, shape=(8, 8)):
+    c = Tensor.empty(*shape)
+    out = Tensor.custom_kernel(c, fxn=kernel)[0]
+    try: return out.numpy()
+    except RuntimeError as e:
+      if isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and "dynamic register indexing" in str(e):
+        self.skipTest("PTX does not support dynamic register indexing")
+      raise
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_contiguous_fragment_index(self):
+    # thread ty owns rows [ty*8, ty*8+8) of a 64-row fragment -- contiguous ownership.
+    # This is the pre-existing case that index_multi always handled.
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      ir = UOp.range(8, 1, AxisType.LOOP)
+      j = UOp.range(8, 2, AxisType.LOOP)
+      # 8x8 fragment, 8 threads -> 64x8 full tile. thread ty owns rows [ty*8, ty*8+8).
+      frag = UOp.placeholder((8, 8), dtypes.float32, 0, AddrSpace.REG).unshard((0,), (ty,))
+      return C[ty*8 + ir, j].store(frag[ty*8 + ir, j]).end(j, ir, ty).sink(arg=KernelInfo(name="contig_frag"))
+    out = self._run(kernel, (64, 8))
+    assert out.shape == (64, 8)
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_strided_fragment_index(self):
+    # thread ty owns rows {ty, ty+8, ty+16, ty+24, ..., ty+56} of a 64-row fragment --
+    # strided ownership. idx = ty + ir*8 where shard_sz=8 (8 threads, shard rows=8).
+    # The contiguous check (idx - rng*shard_sz) fails; the strided check
+    # (idx-rng) % shard_sz == 0 must succeed. This is the pattern the index_multi fix adds.
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      ir = UOp.range(8, 1, AxisType.LOOP)
+      j = UOp.range(8, 2, AxisType.LOOP)
+      # 8x8 fragment, 8 threads -> 64x8 full tile. thread ty owns rows {ty, ty+8, ..., ty+56}.
+      frag = UOp.placeholder((8, 8), dtypes.float32, 0, AddrSpace.REG).unshard((0,), (ty,))
+      return C[ty + ir*8, j].store(frag[ty + ir*8, j]).end(j, ir, ty).sink(arg=KernelInfo(name="strided_frag"))
+    out = self._run(kernel, (64, 8))
+    assert out.shape == (64, 8)
+
+  def test_fragment_index_cannot_shard(self):
+    # thread ty indexing rows [ty, ty+8) overlaps with other threads' rows -- this matches neither
+    # the contiguous nor the strided ownership pattern, so index_multi must raise.
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      ir = UOp.range(8, 1, AxisType.LOOP)
+      j = UOp.range(8, 2, AxisType.LOOP)
+      frag = UOp.placeholder((8, 8), dtypes.float32, 0, AddrSpace.REG).unshard((0,), (ty,))
+      return C[ty + ir, j].store(frag[ty + ir, j]).end(j, ir, ty).sink(arg=KernelInfo(name="bad_frag"))
+    with self.assertRaisesRegex(RuntimeError, "cannot shard index"):
+      self._run(kernel, (64, 8))
+
+def _run_fragment_kernel(testcase, kernel, out_shape, inputs=()):
+  c = Tensor.empty(*out_shape)
+  out = Tensor.custom_kernel(c, *inputs, fxn=kernel)[0]
+  try: return out.numpy()
+  except RuntimeError as e:
+    if isinstance(Device[Device.DEFAULT].renderer, PTXRenderer) and "dynamic register indexing" in str(e):
+      testcase.skipTest("PTX does not support dynamic register indexing")
+    raise
+
+class TestUnshardAlu(unittest.TestCase):
+  """Tests for ALU on (fragment) UNSHARD values in schedule/multi.py's alu_multi.
+
+  An ALU with UNSHARD srcs lowers to per-shard ops when every src is one of:
+    same sharding:  peel the UNSHARD, keep the layout
+    scalar:         broadcast to every shard
+    whole unsharded same-shape value: takes its per-shard sub-view (shard_subview)
+  """
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_alu_scalar_broadcast(self):
+    # scalar srcs broadcast to every shard: frag*2.0 where frag is 1.5 per thread -> 3.0 everywhere
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      # 8 values per thread, 8 threads -> 64-value full view
+      frag = UOp.placeholder((8,), dtypes.float32, 0, AddrSpace.LOCAL).unshard((0,), (ty,))
+      v = frag.after(frag.store(1.5)) * 2.0
+      return C.store(v).end(ty).sink(arg=KernelInfo(name="alu_scalar", opts_to_apply=()))
+    out = _run_fragment_kernel(self, kernel, (64,))
+    np.testing.assert_allclose(out, 3.0)
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_alu_whole_value_subview(self):
+    # UNSHARD + whole unsharded same-shape value: each shard adds its own sub-view of A.
+    def kernel(C:UOp, A:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      frag = UOp.placeholder((8,), dtypes.float32, 0, AddrSpace.LOCAL).unshard((0,), (ty,))
+      v = frag.after(frag.store(0.0)) + A
+      return C.store(v).end(ty).sink(arg=KernelInfo(name="alu_subview", opts_to_apply=()))
+    a = Tensor(np.arange(64, dtype=np.float32))
+    out = _run_fragment_kernel(self, kernel, (64,), inputs=(a,))
+    np.testing.assert_allclose(out, a.numpy(), atol=1e-4)
+
+class TestUnshardStore(unittest.TestCase):
+  """Tests for STORE of a sharded value into an unsharded dest (store_value_multi in schedule/multi.py).
+
+  Every shard stores its value into its own contiguous sub-view of the dest, one SHRINK per sharded axis.
+  """
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_store_unshard_value(self):
+    # single-axis: 8 threads each own 8 values of the 64-value output tile
+    def kernel(C:UOp) -> UOp:
+      ty = UOp.range(8, 0, AxisType.LOCAL)
+      frag = UOp.placeholder((8,), dtypes.float32, 0, AddrSpace.LOCAL).unshard((0,), (ty,))
+      v = frag.after(frag.store(0.0)) + 2.5
+      return C.store(v).end(ty).sink(arg=KernelInfo(name="store_unshard", opts_to_apply=()))
+    out = _run_fragment_kernel(self, kernel, (64,))
+    np.testing.assert_allclose(out, 2.5)
+
+  @unittest.skipIf(not Device[Device.DEFAULT].renderer.has_local, "fragment tests need LOCAL ranges")
+  def test_store_unshard_value_2axis(self):
+    # two sharded axes (the gemm fragment layout): thread (ty, tx) owns the (2, 1, 1, 2) sub-view of the
+    # (2, 4, 2, 2) output tile; the store must SHRINK dest on both sharded axes
+    def kernel(C:UOp, A:UOp) -> UOp:
+      ty = UOp.range(4, 0, AxisType.LOCAL)
+      tx = UOp.range(2, 1, AxisType.LOCAL)
+      frag = UOp.placeholder((2, 1, 1, 2), dtypes.float32, 0, AddrSpace.REG).unshard((1, 2), (ty, tx))
+      v = frag.after(frag.store(0.0)) + A
+      return C.store(v).end(tx, ty).sink(arg=KernelInfo(name="store_unshard_2axis", opts_to_apply=()))
+    a = Tensor(np.arange(32, dtype=np.float32).reshape(2, 4, 2, 2))
+    out = _run_fragment_kernel(self, kernel, (2, 4, 2, 2), inputs=(a,))
+    np.testing.assert_allclose(out, a.numpy(), atol=1e-4)

 class TestUOpReduce(unittest.TestCase):
  def test_uop_sum(self):
@@ -124,6 +124,7 @@ class TestFp8sConversions(unittest.TestCase):
  def test_float_to_fp8e4m3(self, x):
    np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3), torch.tensor(x, dtype=torch.float8_e4m3fn).view(torch.uint8).item())

+  @unittest.skip("fp8 overflow semantics are inconsistent")
  def test_float_to_fp8e4m3_extreme_values(self):
    for x in [FP8E4M3_MAX, FP8E4M3_MAX*1.01, -FP8E4M3_MAX, -FP8E4M3_MAX*1.01, math.inf, -math.inf, math.nan, -math.nan]:
      np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3), torch.tensor(x, dtype=torch.float8_e4m3fn).view(torch.uint8).item())
@@ -168,6 +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)
@@ -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())
@@ -7,7 +7,7 @@ from tinygrad.renderer.isa.x86 import X86Renderer, X86Ops
 from tinygrad.renderer.isa import IselContext

 # INDEX on a register value with a constant index extracts a single element (the old GEP)
-def lane(y:UOp, i:int) -> UOp: return y.index(UOp.const(dtypes.int, i), dtype=y.dtype.scalar())
+def lane(y:UOp, i:int) -> UOp: return y.index(UOp.const(i, dtypes.int), dtype=y.dtype.scalar())

@unittest.skipUnless(isinstance(Device[Device.DEFAULT].renderer, X86Renderer), "only x86")
 class TestIselX86(unittest.TestCase):
@@ -49,7 +49,7 @@ class TestIselX86(unittest.TestCase):
    load = UOp.param(0, dtypes.int32, (16,)).index(a + 1).load()
    n = self.isel_rewrite(load)
    # displacement is the constant in "a" scaled to the buffer element size, dtype is int8 when the value fits otherwise int32
-    self.assertTrue(n.src[2].op is Ops.CONST and n.src[2].dtype is dtypes.int8 and n.src[2].arg == 4)
+    self.assertTrue(n.src[2].op is Ops.CONST and n.src[2].dtype is dtypes.int8 and n.src[2].val == 4)

 if __name__ == "__main__":
  unittest.main()
@@ -2,7 +2,7 @@
 import unittest
 import numpy as np

-from test.helpers import assert_jit_cache_len, call_is_graph, not_support_multi_device, needs_second_gpu
+from test.helpers import assert_jit_cache_len, call_is_graph, not_support_multi_device, needs_second_gpu, KernelCountException
 from test.unit.test_jit import _simple_test
 from tinygrad import Tensor, Variable, TinyJit, Device, dtypes
 from tinygrad.engine.jit import graph_class
@@ -97,7 +97,7 @@ class TestJit(unittest.TestCase):
      prev = o

    # Checking that 2 graphs are inited.
-    assert len(jf.captured.linear.src) == 2
+    if len(jf.captured.linear.src) != 2: raise KernelCountException(2, len(jf.captured.linear.src))
    for si in jf.captured.linear.src:
      assert call_is_graph(si)

@@ -12,7 +12,8 @@ from tinygrad.dtype import DType, dtypes, AddrSpace
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.cstyle import CUDARenderer
 from tinygrad.renderer.isa import ISARenderer
-from test.helpers import replace_opts
+from test.helpers import replace_opts, check_schedule
+from test.backend.test_softmax_fusion import single_kernel_softmax
 MOCKGPU = DEV.interface.startswith("MOCK")

 from tinygrad.uop.render import print_uops # noqa: F401 # pylint: disable=unused-import
@@ -267,9 +268,9 @@ class TestLinearizer(unittest.TestCase):
    uops = tuple(to_program(replace_opts(ast, []), renderer=Device[Device.DEFAULT].renderer).src[1].src)
    idxs = dedup([uop for uop in uops if uop.op is Ops.SPECIAL])
    idxs = sorted(idxs, key=lambda uop: uop.arg)
-    assert (idxs[0].arg, idxs[0].src[0].arg) == ('gidx0', 6), idxs[0]
-    assert (idxs[1].arg, idxs[1].src[0].arg) == ('gidx1', 5), idxs[1].arg
-    assert (idxs[2].arg, idxs[2].src[0].arg) == ('gidx2', 4), idxs[2].arg
+    assert (idxs[0].arg, idxs[0].src[0].val) == ('gidx0', 6), idxs[0]
+    assert (idxs[1].arg, idxs[1].src[0].val) == ('gidx1', 5), idxs[1].arg
+    assert (idxs[2].arg, idxs[2].src[0].val) == ('gidx2', 4), idxs[2].arg

  def test_sum_collapse(self):
    t = Tensor([2]).reshape(1, 1).expand(256, 256).sum()
@@ -292,8 +293,7 @@ class TestLinearizer(unittest.TestCase):
    a = Tensor.ones(4, 4).contiguous().realize()
    b = a.shrink(((1, 2), None)).pad(((1, 2), None)).bool()
    a.assign(b.where(2, a))
-    linear, var_vals = a.linear_with_vars()
-    assert len(linear.src) == 1
+    linear, var_vals = check_schedule(a, 1)
    run_linear(linear, var_vals)
    np.testing.assert_equal(a.flatten().numpy(), [1.,1.,1.,1.,2.,2.,2.,2.,1.,1.,1.,1.,1.,1.,1.,1.])
    program = to_program(replace_opts(linear.src[-1].src[0], []), renderer=Device[Device.DEFAULT].renderer)
@@ -392,6 +392,16 @@ class TestLinearizer(unittest.TestCase):
    # the global store doesn't change
    assert stores[1].src[1].dtype == dtypes.float

+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  def test_two_grouped_stores_local(self):
+    # GROUP on both reduces puts two LOCAL buffers in one kernel, and the store to each needs its own barrier
+    a = Tensor.rand(32, 32).realize()
+    opts = [Opt(OptOps.GROUP, 1, 4), Opt(OptOps.GROUP, 2, 4)]
+    ast = helper_linearizer_opt(single_kernel_softmax(a), [opts])
+    uops = to_program(replace_opts(ast, opts), renderer=Device[Device.DEFAULT].renderer).src[1].src
+    self.assertEqual(len([u for u in uops if u.op is Ops.BARRIER]), 2)
+
 # *** helpers ***

 def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
@@ -12,18 +12,18 @@ class TestLinearizerFailure(unittest.TestCase):
  @unittest.skipUnless(Device.DEFAULT == "METAL", "only tested on METAL")
  def test_failure_beam_mnist(self):
    c0 = UOp.param(0, dtypes.uchar, (4014080,))
-    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 0, AxisType.GLOBAL)
-    c2 = UOp.range(UOp.const(dtypes.weakint, 784), 1, AxisType.GLOBAL)
-    c3 = UOp.range(UOp.const(dtypes.weakint, 10), 3, AxisType.GLOBAL)
+    c1 = UOp.range(UOp.const(512), 0, AxisType.GLOBAL)
+    c2 = UOp.range(UOp.const(784), 1, AxisType.GLOBAL)
+    c3 = UOp.range(UOp.const(10), 3, AxisType.GLOBAL)
    c4 = UOp.param(1, dtypes.int, (512,))
-    c5 = c4.index(c1.valid(UOp.const(dtypes.bool, True)))
-    c6 = UOp.range(UOp.const(dtypes.weakint, 6000), 1004, AxisType.REDUCE)
-    c7 = UOp.range(UOp.const(dtypes.weakint, 3750), 2006, AxisType.REDUCE)
-    c8 = UOp.range(UOp.const(dtypes.weakint, 16), 2007, AxisType.GROUP_REDUCE)
+    c5 = c4.index(c1.valid(UOp.const(True)))
+    c6 = UOp.range(UOp.const(6000), 1004, AxisType.REDUCE)
+    c7 = UOp.range(UOp.const(3750), 2006, AxisType.REDUCE)
+    c8 = UOp.range(UOp.const(16), 2007, AxisType.GROUP_REDUCE)
    c9 = UOp.param(2, dtypes.uchar, (47040000,))
-    c10 = c9.index((((c3*UOp.const(dtypes.weakint, 4704000))+c2)+(c6*UOp.const(dtypes.weakint, 784))).valid(UOp.const(dtypes.bool, True)))
-    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.weakint, 6000))+c6)+((c7*UOp.const(dtypes.weakint, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.weakint, 59999)).where(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(dtypes.int, -1))).where(UOp.const(dtypes.uchar, 0), c10).reduce(c6, arg=Ops.ADD)
-    c12 = c0.index((((c1*UOp.const(dtypes.weakint, 7840))+(c2*UOp.const(dtypes.weakint, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
+    c10 = c9.index((((c3*UOp.const(4704000))+c2)+(c6*UOp.const(784))).valid(UOp.const(True)))
+    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(6000))+c6)+((c7*UOp.const(16))+c8)).alu(Ops.CMPLT, UOp.const(59999)).where(UOp.const(0).cast(dtypes.int), UOp.const(1).cast(dtypes.int)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(-1).cast(dtypes.int))).where(UOp.const(0).cast(dtypes.uchar), c10).reduce(c6, arg=Ops.ADD)
+    c12 = c0.index((((c1*UOp.const(7840))+(c2*UOp.const(10)))+c3).valid(UOp.const(True))).store(c11).end(c1, c2, c3)
    ast = c12.sink(arg=KernelInfo(name='test', axis_types=(), dont_use_locals=False, applied_opts=(Opt(op=OptOps.GROUP, axis=1, arg=16),), opts_to_apply=None))
    _ = to_program(ast, Device["METAL"].renderer)

@@ -5,9 +5,10 @@ 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
+from test.helpers import needs_second_gpu, assert_kernel_count
 from test.backend.test_asm_gemm import has_hipcc

 def run_fused_ce(bs:int, seqlen:int, vocab:int, label_smoothing:float=0.0) -> None:
@@ -80,7 +81,7 @@ class TestQuantizeFP8(unittest.TestCase):
  @needs_second_gpu
  def test_multi(self):
    devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(8))
-    x = Tensor.empty(2048*8, 1024, dtype=dtypes.bfloat16, device=devs).uop.multi(0)
+    x = Tensor.empty(2048*8, 1024, dtype=dtypes.bfloat16, device=devs).uop.unshard(0)
    x = Tensor(x, device=devs)
    amax_state = Tensor.full((), 2.0, dtype=dtypes.float32, device=devs).contiguous()
    amax_out = Tensor.zeros((), dtype=dtypes.float32, device=devs).realize()
@@ -95,7 +96,7 @@ class TestLocalAmax(unittest.TestCase):
    x = Tensor.arange(16).reshape(4, 4).cast(dtypes.float).clone(devices[0]).realize().shard(devices, axis=0).realize()
    GlobalCounters.reset()
    out = (x * local_abs_max(x)).clone().realize()
-    self.assertEqual(GlobalCounters.kernel_count, 2)
+    assert_kernel_count(2)
    self.assertEqual(out.tolist(), [[0., 7., 14., 21.], [28., 35., 42., 49.], [120., 135., 150., 165.], [180., 195., 210., 225.]])

@unittest.skipUnless(has_hipcc() and Device.DEFAULT == "AMD", "requires hipcc to compile and amd device to run")
@@ -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()
@@ -1,12 +1,12 @@
 import unittest, random
 from tinygrad import Tensor, Device, nn, GlobalCounters, TinyJit, dtypes, Variable
-from tinygrad.uop.ops import Ops, UOp
+from tinygrad.uop.ops import Ops, UOp, AxisType
 from tinygrad.helpers import getenv, prod, Context
 from tinygrad.nn.state import get_parameters
 from tinygrad.engine.realize import run_linear, compile_linear
 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
+from test.helpers import not_support_multi_device, needs_second_gpu, slow, call_is_graph, check_schedule, assert_kernel_count

 settings.register_profile("my_profile", max_examples=200, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
 settings.load_profile("my_profile")
@@ -52,15 +52,17 @@ class TestMultiTensor(unittest.TestCase):
  def test_shard(self):
    X = Tensor.ones(256).contiguous().realize()
    X.shard_(devices_2, 0)
-    for lb in X.uop.src:
-      assert lb.shape == (128,)
+    assert X.uop.src[0].shape == (128,)
+    # the MULTI carries and ends the DEVICE range as its second src
+    assert X.uop.src[1].op is Ops.RANGE and X.uop.src[1].arg[-1] is AxisType.DEVICE
+    assert X.uop.ended_ranges == X.uop.src[1:]
    (X + X).realize()

  @unittest.expectedFailure # TODO: fix
  def test_shard_empty(self):
    GlobalCounters.reset()
    X = Tensor.empty(256).shard(devices_2, 0).realize()
-    assert GlobalCounters.kernel_count == 0
+    assert_kernel_count(0)
    (X + X).realize()

  # TODO: fix this to not copy on the src device
@@ -74,6 +76,13 @@ class TestMultiTensor(unittest.TestCase):
    run_linear(linear)
    self.assertEqual(len(set(names)), 1, "function was relinearized")

+  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, ())
+
  def test_shard_same_device(self):
    X = Tensor.ones(256).contiguous().realize()
    X.shard_((d1, X.device), 0)
@@ -346,8 +355,7 @@ class TestMultiTensor(unittest.TestCase):
  def test_const_like_shrink_on_shard_axis(self):
    t = Tensor.ones(16, 16, dtype=dtypes.int).shard(devices_2, axis=0)
    out = t.const_like(2)[:, :8]
-    linear, var_vals = out.linear_with_vars()
-    self.assertEqual(len(linear.src), 0)
+    linear, var_vals = check_schedule(out, 0)
    run_linear(linear, var_vals)
    self.assertEqual(out.tolist(), [[2]*8]*16)

@@ -423,6 +431,49 @@ class TestMultiBufferView(unittest.TestCase):
    run_linear(linear, var_vals)
    np.testing.assert_equal(out.numpy(), ref[5].numpy())

+@unittest.skipIf(not_support_multi_device(), "need multi")
+class Test2DShard(unittest.TestCase):
+  def setUp(self):
+    self.devices_4 = tuple(f"{Device.DEFAULT}:{i}" for i in range(4))
+    self.rng = UOp.range(4, -1, AxisType.DEVICE)
+    self.rng0, self.rng1 = self.rng // 2, self.rng % 2
+
+  def _shard_2d(self, t:Tensor) -> Tensor:
+    u = t.uop.copy_to_device(self.devices_4)._shard(0, self.rng0)._shard(1, self.rng1).unshard((0, 1), (self.rng0, self.rng1))
+    return Tensor(u)
+
+  def test_2d_shard_basic(self):
+    ref = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = t.contiguous().realize()
+    np.testing.assert_equal(out.numpy(), ref.numpy())
+
+  def test_2d_shard_elementwise(self):
+    ref = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = (t + 1).contiguous().realize()
+    np.testing.assert_equal(out.numpy(), ref.numpy() + 1)
+
+  def test_2d_shard_sum_all(self):
+    ref = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = t.sum().contiguous().realize()
+    np.testing.assert_equal(out.numpy(), np.array(ref.numpy().sum()))
+
+  def test_2d_shard_sum_non_sharded_axis(self):
+    ref = Tensor.arange(4*4*2).reshape(4, 4, 2).contiguous().realize()
+    t = self._shard_2d(ref)
+    out = t.sum(axis=2).contiguous().realize()
+    np.testing.assert_equal(out.numpy(), ref.numpy().sum(axis=2))
+
+  def test_2d_shard_matmul(self):
+    a = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    b = Tensor.arange(16).reshape(4, 4).contiguous().realize()
+    a_s = self._shard_2d(a)
+    b_s = self._shard_2d(b)
+    out = (a_s @ b_s).contiguous().realize()
+    np.testing.assert_equal(out.numpy(), a.numpy() @ b.numpy())
+
@unittest.skipIf(not_support_multi_device(), "need multi")
 class TestMultiTransformer(unittest.TestCase):
  @needs_second_gpu
@@ -3,11 +3,11 @@ import unittest
 import numpy as np
 import torch
 from tinygrad import Tensor, Device, TinyJit, dtypes
-from tinygrad.uop.ops import Ops
 from tinygrad.helpers import GlobalCounters, Context
 from tinygrad.nn import Conv1d, ConvTranspose1d, Conv2d, ConvTranspose2d, Linear, Embedding
 from tinygrad.nn import BatchNorm, LayerNorm, LayerNorm2d, GroupNorm, InstanceNorm, RMSNorm, LSTMCell
 from tinygrad.nn.state import load_state_dict
+from test.helpers import check_schedule
 from tinygrad.engine.realize import run_linear
 from test.helpers import not_support_multi_device, needs_second_gpu, slow

@@ -428,18 +428,14 @@ class TestNN(unittest.TestCase):
    a = Tensor([[1, 5, 9, 11],
                [12, 19, 8, 1]])
    result = layer(a)
-    linear, var_vals = result.linear_with_vars()
-    self.assertEqual(len([call for call in linear.src if call.src[0].op is Ops.SINK]), kcount,
-                     "first run realizes weight and embedding")
+    linear, var_vals = check_schedule(result, kcount)
    run_linear(linear, var_vals)

    b = Tensor([[1, 2, 3],
                [4, 5, 6],
                [7, 8, 9]])
    result = layer(b)
-    linear, var_vals = result.linear_with_vars()
-    self.assertEqual(1, len([call for call in linear.src if call.src[0].op is Ops.SINK]),
-                     "second run realizes embedding only")
+    linear, var_vals = check_schedule(result, 1)
    run_linear(linear, var_vals)
    print(f"Embedding used {GlobalCounters.global_ops} ops")
    self.assertLessEqual(GlobalCounters.global_ops, ops)
@@ -340,11 +340,11 @@ class TestOps(unittest.TestCase):

  def test_where(self):
    helper_test_op([], lambda: torch.where(torch.tensor([True, False]), 1, 3).type(torch.int32),
-                   lambda: Tensor([True, False]).where(1, 3), forward_only=True)
+                   lambda: Tensor([True, False]).where(1, 3).clone(), forward_only=True)
    helper_test_op(
      [(100,)],
      lambda x: torch.where(x > 0.5, 4, 2).type(torch.int32),
-      lambda x: (x > 0.5).where(4, 2), forward_only=True)
+      lambda x: (x > 0.5).where(4, 2).clone(), forward_only=True)

    for shps in [[(8,),(1,),(1,)], [(10,10),(10,),(10,)], [(100,)]*3, [(10,10)]*3]:
      helper_test_op(
@@ -356,7 +356,7 @@ class TestOps(unittest.TestCase):
    helper_test_op(
      [(5, 5)],
      lambda x: torch.where(x > 0.5, 4, 2).type(torch.int32).permute((1, 0)),
-      lambda x: (x > 0.5).where(4, 2).permute((1, 0)), forward_only=True)
+      lambda x: (x > 0.5).where(4, 2).clone().permute((1, 0)), forward_only=True)

  def _test_cmp(self, fxn, reverse=True):
    # test different dtypes
@@ -636,9 +636,9 @@ class TestOps(unittest.TestCase):
        helper_test_op(None, lambda x,y: x%y, forward_only=True, vals=[va, vb])
        helper_test_op(None, lambda x: x%2, forward_only=True, vals=[va])
        helper_test_op(None, lambda x: x%3, forward_only=True, vals=[va])
-        helper_test_op(None, lambda x: x%3.5, forward_only=True, vals=[va])
+        helper_test_op(None, lambda x: x%3.5, lambda x: (x%3.5).clone(), forward_only=True, vals=[va])
        helper_test_op(None, lambda x: 100%x, forward_only=True, vals=[va])
-        helper_test_op(None, lambda x: 100.5%x, forward_only=True, vals=[va])
+        helper_test_op(None, lambda x: 100.5%x, lambda x: (100.5%x).clone(), forward_only=True, vals=[va])

  def test_fmod(self):
    a = [-4, 7, 5, 4, -7, 8, -9]
@@ -649,7 +649,7 @@ class TestOps(unittest.TestCase):
        vb = [float(bi) for bi in b] if float_b else b
        helper_test_op(None, lambda x,y: x.fmod(y), forward_only=True, vals=[va, vb])
        helper_test_op(None, lambda x: x.fmod(2), forward_only=True, vals=[va])
-        helper_test_op(None, lambda x: x.fmod(3.5), forward_only=True, vals=[va])
+        helper_test_op(None, lambda x: x.fmod(3.5), lambda x: x.fmod(3.5).clone(), forward_only=True, vals=[va])

  def test_mul_naninf(self):
    helper_test_op([(45,65)], lambda x: x*math.inf)
@@ -706,7 +706,7 @@ class TestOps(unittest.TestCase):
    helper_test_op(None, lambda x: 0.7**x, vals=[[-2.,-1,0,1,2,3]])
    helper_test_op(None, lambda x: (-2)**x, vals=[[-2.,-1,0,1,2,3]])
    # float to power of int
-    helper_test_op(None, lambda x: 0.7**x, vals=[[-2,-1,0,1,2,3]], forward_only=True)
+    helper_test_op(None, lambda x: 0.7**x, lambda x: (0.7**x).clone(), vals=[[-2,-1,0,1,2,3]], forward_only=True)

  @unittest.skipIf(COMPILE_ONLY, "test requires runtime")
  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, NIRRenderer), "TODO: broken in LVP")
@@ -728,6 +728,17 @@ class TestOps(unittest.TestCase):
        else:
          self.assertAlmostEqual(tiny_out, torch_out, msg=f"{x}, {c}")

+  def test_pow_neg_inf_frac_exponent(self):
+    # pow(-inf, 0.3) is +inf, so the gradient 0.3*pow(-inf, -0.7) is 0, never nan
+    helper_test_op(None, lambda x: x**0.3, vals=[[-math.inf]])
+    # is_odd truncates, so it calls 3.3 odd: only the non_int guard keeps pow(-inf, 3.3) from negating to -inf
+    helper_test_op(None, lambda x: x**3.3, vals=[[-math.inf]])
+
+  def test_pow_zero_exponent(self):
+    # x ** 0 is the constant 1 for every x, so the gradient with respect to the base is 0, never nan
+    # TODO: nan ** 0, failed on WEBGPU
+    helper_test_op(None, lambda x,y: x**y, vals=[[-math.inf, math.inf, 0.0], [0.0, 0.0, 0.0]])
+
  def test_pow_zero_tensor(self):
    helper_test_op(None, lambda x,y: x**y, vals=[[0.0], [0.0]])
    # TODO: fix WEBGPU
@@ -775,7 +786,7 @@ class TestOps(unittest.TestCase):
  def test_pow_int_base_float_exponent(self):
    for exponent in (0.5, 1.5, 2.0, -1.0, 0.0):
      helper_test_op([], lambda: torch.tensor([1, 2, 3, 4], dtype=torch.int) ** exponent,
-                         lambda: Tensor([1, 2, 3, 4], dtype=dtypes.int32) ** exponent, forward_only=True)
+                         lambda: (Tensor([1, 2, 3, 4], dtype=dtypes.int32) ** exponent).clone(), forward_only=True)

  def test_sqrt(self):
    helper_test_op([(45,65)], lambda x: x.sqrt())
@@ -1524,6 +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))
@@ -1728,6 +1741,8 @@ class TestOps(unittest.TestCase):
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(0), atol=1e-7, grad_atol=1e-7)
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(1), atol=1e-7, grad_atol=1e-7)
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(2), atol=1e-7, grad_atol=1e-7)
+  def test_softmin(self):
+    helper_test_op([(45,65)], torch.nn.Softmin(dim=1), Tensor.softmin, atol=1e-7, grad_atol=1e-7)

  def test_normalize(self):
    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x), lambda x: x.normalize(), atol=1e-7, grad_atol=1e-7)
@@ -143,12 +143,11 @@ class TestOptim(unittest.TestCase):

  @unittest.skipUnless(dtypes.half in Device[Device.DEFAULT].renderer.supported_dtypes(), "need half")
  def test_mixed_precision(self):
-    old_default_float, dtypes.default_float = dtypes.default_float, dtypes.half
+    self.enterContext(Context(DEFAULT_FLOAT=dtypes.half))
    # weight update would overflow without upcasting
    self._test_sgd(10, {'lr': 1e10}, 1e-6, 3e-4)
    self._test_adam(1, {'lr': 1e10}, 1e-4, 1e-4)
    self._test_adamw(1, {'lr': 1e10}, 1e-4, 1e-4)
-    dtypes.default_float = old_default_float

  def test_assert_tensor_train(self):
    t = Tensor.ones((1,1))
@@ -13,7 +13,7 @@ class TestPickle(unittest.TestCase):

  def test_pickle_pattern_matcher(self):
    pm = PatternMatcher([(UPat.cvar('x'), lambda x: x*2)])
-    sink = UOp.const(dtypes.int, 2)
+    sink = UOp.const(2)
    tt = pm.rewrite(sink)
    pm_str = pickle.dumps(pm)
    pm2 = pickle.loads(pm_str)
@@ -1,7 +1,7 @@
 import unittest, math

 from tinygrad import dtypes, Tensor, Device
-from tinygrad.helpers import getenv, DEV
+from tinygrad.helpers import getenv, DEV, Context
 from tinygrad.codegen import to_program

 from tinygrad.uop.ops import Ops
@@ -232,16 +232,14 @@ class TestRandomness(unittest.TestCase):
  @given(strat.sampled_from([dtypes.float, dtypes.float16, dtypes.bfloat16]))
  def test_randn_finite(self, default_float):
    if default_float not in Device[Device.DEFAULT].renderer.supported_dtypes(): return
-    old_default_float = dtypes.default_float
    # low precision can result in inf from randn
-    dtypes.default_float = default_float
+    self.enterContext(Context(DEFAULT_FLOAT=default_float))
    t = Tensor.randn(64, 64)
    mx = t.max().numpy().item()
    mn = t.min().numpy().item()
    print(f"testing with {default_float=}")
    assert math.isfinite(mx), mx
    assert math.isfinite(mn), mn
-    dtypes.default_float = old_default_float

  def test_random_counter_overflow(self):
    device = Device.DEFAULT
@@ -8,6 +8,7 @@ from tinygrad.helpers import prod
 from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.wgsl import WGSLRenderer
+from test.helpers import check_schedule
 from tinygrad.runtime.ops_python import PythonRenderer
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, python_alu
 from tinygrad.tensor import Tensor
@@ -24,7 +25,7 @@ def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
  dtype = alu_src_uops[0].dtype
  a = UOp.param(0, dtype, (1,))
  b = UOp.param(1, dtype, (1,))
-  idx = UOp.const(dtypes.int, 0)
+  idx = UOp.const(0)
  ld = b.index(idx).load()
  alu = ld.alu(alu_op, *alu_src_uops)
  store = UOp.store(a.index(idx), alu)
@@ -35,7 +36,7 @@ class TestRendererFailures(unittest.TestCase):
  def test_gated_store_with_alu(self):
    a = UOp.param(0, dtypes.int, (4,))
    gate_alu = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0.valid(gate_alu)), UOp.const(dtypes.int, 1)))
+    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0.valid(gate_alu)), UOp.const(1).cast(dtypes.int)))
    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
    ret = _test_uop_result([], sink, local_size=[4, 1, 1])[0]
    np.testing.assert_equal(ret, [0, 1, 1, 1])
@@ -45,7 +46,7 @@ class TestRendererFailures(unittest.TestCase):
    a = UOp.param(0, dtypes.int, (8,))
    gate_alu_0 = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
    gate_alu_1 = (lidx1:=UOp.special(2, 'lidx1')).ne(0)
-    gated_alu_store = UOp(Ops.STORE, src=(a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(dtypes.int, 1)))
+    gated_alu_store = UOp(Ops.STORE, src=(a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(1).cast(dtypes.int)))
    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
    ret = _test_uop_result([], sink, local_size=[4, 2, 1])[0]
    np.testing.assert_equal(ret, [0, 0, 0, 0, 0, 1, 1, 1])
@@ -54,15 +55,14 @@ class TestRendererFailures(unittest.TestCase):
 class TestCStyleFailures(unittest.TestCase):
  def test_inline_const_alu(self):
    # CPU doesn't use the max function
-    ret = _setup_and_test_alu(Ops.MAX, 1, UOp.const(dtypes.int, dtypes.int.min+1))
+    ret = _setup_and_test_alu(Ops.MAX, 1, UOp.const(dtypes.int.min+1).cast(dtypes.int))
    self.assertEqual(ret[0], 1)

  def _test_src_strip_paren(self, op: Ops, should_strip_paren:bool=True):
    dtype = "bool" if op in (Ops.OR, Ops.XOR, Ops.AND) else None
    ret = Tensor.empty(1, dtype=dtype)
    for _ in range(5): ret = python_alu[op](ret, Tensor.empty(1, dtype=dtype))
-    linear = ret.schedule_linear()
-    assert len(linear.src) == 1
+    linear, _ = check_schedule(ret, 1)
    src = to_program(linear.src[0].src[0], Device[Device.DEFAULT].renderer).src[2].arg
    self.assertEqual("("*5 not in src, should_strip_paren)

@@ -80,7 +80,7 @@ class TestWGSLFailures(unittest.TestCase):
  def test_multiply_infinity(self):
    # multiplying a positive constant by infinity should return infinity
    # WGSL pipelines do not handle this reliably, some of which return zero, unless infinity always comes from a read on a dynamic buffer
-    ret = _setup_and_test_alu(Ops.MUL, 5.0, UOp.const(dtypes.float32, float("inf")))
+    ret = _setup_and_test_alu(Ops.MUL, 5.0, UOp.const(float("inf")).cast(dtypes.float32))
    self.assertEqual(ret[0], float("inf"))

  # WGSL has a specific select(alt, val, gate) ternary operator instead of gate?val:alt
@@ -104,7 +104,7 @@ class TestPTXFailures(unittest.TestCase):
  def test_gated_store_with_if(self):
    a = UOp.param(0, dtypes.int, (4,))
    gate_alu = (lidx0:=UOp.special(4, 'lidx0')).ne(0)
-    val = UOp.const(dtypes.int, 1)
+    val = UOp.const(1).cast(dtypes.int)
    if_uop = UOp(Ops.IF, src=(gate_alu,))
    gated_alu_store = UOp(Ops.STORE, src=(a.index(lidx0, if_uop), val))
    sink = UOp(Ops.SINK, src=(gated_alu_store,), arg=KernelInfo())
@@ -6,34 +6,13 @@ import unittest, time
 import numpy as np

 from tinygrad import nn, dtypes, Device, Tensor, Variable
-from tinygrad.uop.ops import UOp, Ops, UPat
-from tinygrad.helpers import DEBUG, DEV, GlobalCounters, Context, all_same, temp
-from tinygrad.engine.realize import compile_linear, run_linear
+from tinygrad.uop.ops import Ops, UPat
+from tinygrad.helpers import DEV, GlobalCounters, Context, all_same, temp
+from tinygrad.engine.realize import run_linear
+from test.helpers import check_schedule, assert_kernel_count

 supported_dtypes = Device[Device.DEFAULT].renderer.supported_dtypes()

-class KernelCountException(Exception): pass
-def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Tensor]|None=None, filter_sink=True):
-  if to_prerealize:
-    with Context(DEBUG=0, TRACK_MATCH_STATS=0): Tensor.realize(*to_prerealize)
-  if isinstance(t, Tensor): linear, var_vals = t.linear_with_vars()
-  elif isinstance(t, list) and isinstance(t[0], Tensor): linear, var_vals = Tensor.linear_with_vars(*t)
-  else:
-    assert isinstance(t, UOp), f"can't schedule {t}"
-    linear, var_vals = Tensor(t).linear_with_vars()
-  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:
-    print(f"SCHEDULE ISSUE, expecting {allowed} got {kernel_cnt}")
-    if DEBUG >= 3:
-      for i,call in enumerate(linear.src):
-        print("kernel", i+1)
-        print(call.src[0])
-    raise KernelCountException(f"{kernel_cnt} != {allowed}")
-  # test compiling the linear
-  compile_linear(linear)
-  return linear, var_vals
-
 def _realize_weights(m):
  for p in nn.state.get_parameters(m): p.realize()

@@ -113,11 +92,9 @@ class TestSchedule(unittest.TestCase):
    a2 = mop(a)
    expected = (a+a2).tolist()
    a.assign(a+a2)
-    linear, var_vals = a.linear_with_vars()
-    kcount = len(linear.src)
+    linear, var_vals = check_schedule(a, expected_kcount)
    run_linear(linear, var_vals)
    self.assertListEqual(a.tolist(), expected)
-    self.assertEqual(kcount, expected_kcount)
  def test_setitem_permuted_sched(self): self.test_setitem_sched(lambda x: x.T, 2)
  def test_setitem_paddded_sched(self): self.test_setitem_sched(lambda x: x.shrink_to(4, 1).pad_to(4, 4), 1)

@@ -126,9 +103,9 @@ class TestSchedule(unittest.TestCase):
    a = Tensor.arange(16).clone().realize()
    GlobalCounters.reset()
    a[4] = 3
-    self.assertEqual(GlobalCounters.kernel_count, 0)
+    assert_kernel_count(0)
    a.realize()
-    self.assertEqual(GlobalCounters.kernel_count, 1)
+    assert_kernel_count(1)
    self.assertListEqual(a.tolist(), [0, 1, 2, 3, 3, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15])

  def test_no_extra_contiguous_on_setitem_assign_back(self):
@@ -363,7 +340,7 @@ class TestCopyFolding(unittest.TestCase):

  def test_one_hot_with_copy(self):
    y = Tensor([1, 2, 3]).to("CPU")
-    x = y.one_hot(10)
+    x = y.one_hot(10).int()
    check_schedule(x, 3, filter_sink=False)

  @unittest.skip("no longer supported")
@@ -4,6 +4,7 @@ from tinygrad import Tensor, GlobalCounters, Context, Device
 from tinygrad.dtype import DTypeLike, dtypes
 from tinygrad.engine.realize import run_linear
 from tinygrad.helpers import DEBUG, get_single_element
+from test.helpers import check_schedule

 def single_kernel_softmax(x_in:Tensor, axis=-1, dtype:DTypeLike|None=None) -> Tensor:
  # only support axis =-1
@@ -103,8 +104,7 @@ class TestFuse(unittest.TestCase):
    k = (x @ wk).contiguous()
    v = (x @ wv).contiguous()
    attn = q.scaled_dot_product_attention(k, v)
-    s = attn.schedule_linear()
-    self.assertEqual(len(s.src), 4) # 3 matmul and 1 attention
+    check_schedule(attn, 4) # 3 matmul and 1 attention

  @unittest.skip("needs RANGEIFY>1")
  def test_flash_attention(self):
@@ -6,6 +6,15 @@ from examples.gpt2 import Attention
 import numpy as np

 class TestSymbolicOps(unittest.TestCase):
+  def test_negative_slice(self):
+    a = Tensor.rand(3, 10, 4)
+    for i in range(3, 10):
+      vi = Variable("i", 1, 10).bind(i)
+      # negative int bounds against a symbolic dim must resolve against the size, like slice.indices
+      np.testing.assert_allclose(a[:, :vi][:, -3:-1].numpy(), a[:, :i][:, -3:-1].numpy(), atol=1e-6, rtol=1e-6)
+      np.testing.assert_allclose(a[:, :vi][:, -1:].numpy(), a[:, :i][:, -1:].numpy(), atol=1e-6, rtol=1e-6)
+      np.testing.assert_allclose(a[:, :vi][:, -1].numpy(), a[:, :i][:, -1].numpy(), atol=1e-6, rtol=1e-6)
+
  def test_plus1(self):
    def f(a): return (a+1).realize()
    a = Tensor.rand(3, 10)
@@ -24,13 +24,13 @@ class TestTinygrad(unittest.TestCase):
    self.assertEqual(Tensor(3.14).shape, ())

  def test_deviceless_const_construct_device_repr(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0))
+    t = Tensor(UOp.const(2.0).cast(dtypes.float))
    self.assertIsNone(t.uop.device)
    self.assertIsNone(t.device)
    self.assertIn("<UOp None", repr(t))

  def test_deviceless_const_realize_noop(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0))
+    t = Tensor(UOp.const(2.0).cast(dtypes.float))
    uop = t.uop
    t.realize()
    self.assertIs(t.uop, uop)
@@ -728,12 +728,12 @@ class TestZeroShapeTensor(unittest.TestCase):
    self.assertIsNot(a.uop.base.buffer, b.uop.base.buffer)

  def test_clone_deviceless_const(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0)).clone()
+    t = Tensor(UOp.const(2.0).cast(dtypes.float)).clone()
    np.testing.assert_equal(t.numpy(), 2.0)
    self.assertTrue(t.uop.has_buffer_identity())

  def test_numpy_deviceless_const(self):
-    np.testing.assert_equal(Tensor(UOp.const(dtypes.float, 2.0)).numpy(), 2.0)
+    np.testing.assert_equal(Tensor(UOp.const(2.0).cast(dtypes.float)).numpy(), 2.0)

  def test_clone_with_shrink(self):
    a = Tensor.rand(16, 16)
@@ -756,7 +756,7 @@ class TestZeroShapeTensor(unittest.TestCase):
    np.testing.assert_allclose(a.grad.numpy(), b.grad.numpy())

  def test_clone_deviceless_const_to_cpu(self):
-    t = Tensor(UOp.const(dtypes.float, 2.0)).clone(device="CPU")
+    t = Tensor(UOp.const(2.0).cast(dtypes.float)).clone(device="CPU")
    self.assertEqual(t.device, "CPU")
    np.testing.assert_equal(t.numpy(), 2.0)

@@ -1,6 +1,6 @@
 import unittest
 import numpy as np
-from tinygrad import Tensor, Variable, dtypes
+from tinygrad import Device, Tensor, Variable, TinyJit, dtypes
 from tinygrad.helpers import CHECK_OOB

 class TestTensorVariable(unittest.TestCase):
@@ -18,13 +18,24 @@ class TestTensorVariable(unittest.TestCase):
      self.assertListEqual((vv * t).tolist(), [2, 2, 2])
    except RuntimeError: pass

+  @unittest.skipUnless(dtypes.long in Device[Device.DEFAULT].renderer.supported_dtypes(), "requires long support")
  def test_large_range_variable(self):
-    vv = Variable("b", 0, 2**40).bind(2**35)
-    # TODO: pm_lower_index_dtype lowers ALU PARAM to int32 unconditionally
-    try:
-      self.assertEqual(Tensor(vv).item(), 2**35)
-    except AssertionError:
-      pass
+    self.assertEqual(Tensor(Variable("b", 0, 2**40, dtype=dtypes.long).bind(2**35)).clone(Device.DEFAULT).item(), 2**35)
+
+  @unittest.skipUnless(dtypes.long in Device[Device.DEFAULT].renderer.supported_dtypes(), "requires long support")
+  def test_large_range_variable_jit(self):
+    @TinyJit
+    def f(a,b): return (Tensor(a+b).clone(Device.DEFAULT) * 2).realize()
+    for i in range(3):
+      a = Variable("a", 0, 2**10, dtype=dtypes.int).bind(i)
+      b = Variable("b", 0, 2**40, dtype=dtypes.long).bind(2**35)
+      self.assertEqual(f(a,b).item(), (2**35 + i) * 2)
+
+  def test_variable_defers_like_a_literal(self):
+    vv = Variable("a", 1, 10).bind(2)
+    self.assertEqual(Tensor(vv).dtype, dtypes.weakint)
+    self.assertEqual((Tensor(vv) + Tensor([1], dtype=dtypes.int8)).dtype, dtypes.int8)  # takes the concrete side, no widening
+    self.assertEqual(Tensor(vv).item(), 2)                                              # a read commits at default_int

  def test_variable_tensor_dtype_arg(self):
    vv = Variable("a", 1, 10).bind(2)
@@ -2,7 +2,7 @@ import unittest
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.helpers import Context, getenv, DEV, OSX
-from test.backend.test_schedule import check_schedule
+from test.helpers import check_schedule
 from test.backend.test_dtype_alu import ht, dtypes_float
 import numpy as np
 import math
@@ -19,7 +19,7 @@ def run_uops(uops_list:list[UOp], bufs:list[Buffer]):
  run_linear(UOp(Ops.LINEAR, src=(UOp.sink(*uops_list, arg=KernelInfo()).call(*buf_uops),)))

 def uop(uops:list[UOp], op:Ops, dtype:Optional[DType], src:tuple[UOp, ...], arg:Any=None) -> UOp:
-  if op is Ops.CONST: uops.append(UOp.const(dtype, arg))
+  if op is Ops.CONST: uops.append(UOp.const(arg).cast(dtype))
  elif op is Ops.PARAM: uops.append(UOp.param(arg, dtype, shape=(1,)))
  else: uops.append(UOp(op, dtype, tuple(src), arg))
  return uops[-1]
@@ -43,7 +43,7 @@ def _test_single_value_const(vals, op, dts):
  buf_store = uop(uops, Ops.PARAM, output_dtype, (), 0)
  loads = (uop(uops, Ops.CONST, dtype, [], a) for a,dtype in zip(vals, dts))
  alu = uop(uops, op, output_dtype, loads)
-  out = buf_store[UOp.const(dtypes.int32, 0)].store(alu)
+  out = buf_store[UOp.const(0).cast(dtypes.int32)].store(alu)
  buf = Buffer(Device.DEFAULT, 1, output_dtype).allocate()
  run_uops([out], [buf])
  return np.frombuffer(buf.as_memoryview(), _to_np_dtype(output_dtype))[0]
@@ -221,12 +221,12 @@ class TestAssembly(unittest.TestCase):
  def test_bitshift_left(self):
    g1 = UOp.param(0, dtypes.int32, shape=(3,))
    out = UOp.param(1, dtypes.int32, shape=(2,))
-    c1 = UOp.const(dtypes.int, 2)
-    c2 = UOp.const(dtypes.int, 3)
+    c1 = UOp.const(2)
+    c2 = UOp.const(3)
    l1 = g1.index(c1)
    a1 = UOp(Ops.MUL, src=(l1, c1))
    a2 = UOp(Ops.MUL, src=(l1, c2))
-    uops = to_uops_list([out.index(UOp.const(dtypes.int, 0)).store(a1), out.index(UOp.const(dtypes.int, 1)).store(a2)],
+    uops = to_uops_list([out.index(UOp.const(0)).store(a1), out.index(UOp.const(1)).store(a2)],
                        ren=Device[Device.DEFAULT].renderer)
    Device[Device.DEFAULT].renderer.render(uops)
    ops = [x.op for x in uops]
@@ -249,16 +249,16 @@ class TestAssembly(unittest.TestCase):

  def test_mulacc_shl(self):
    g1 = UOp.param(0, dtypes.int32, shape=(2,))
-    c1 = UOp.const(dtypes.int, 0)
-    c2 = UOp.const(dtypes.int, 1)
-    expr = g1.index(c1) * UOp.const(dtypes.int, 4096) + g1.index(c2)
+    c1 = UOp.const(0)
+    c2 = UOp.const(1)
+    expr = g1.index(c1) * UOp.const(4096) + g1.index(c2)
    uops = to_uops_list([expr], ren=Device[Device.DEFAULT].renderer)
    Device[Device.DEFAULT].renderer.render(uops)
    self.assertIn(Ops.MULACC, [x.op for x in uops])

  def test_use_cmpeq(self):
    g = UOp.param(0, dtypes.uint32, shape=(8,))
-    c = UOp.const(dtypes.uint, 7)
+    c = UOp.const(7)
    comp = g.index(c).ne(c).ne(True)
    uops = to_uops_list([comp], ren=Device[Device.DEFAULT].renderer)
    Device[Device.DEFAULT].renderer.render(uops)
@@ -330,6 +330,7 @@ class TestHCQ(unittest.TestCase):
  # Test profile api
  def test_speed_exec_time(self):
    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
+    st = time.perf_counter()
    TestHCQ.d0.hw_compute_queue_t().timestamp(sig_st) \
                                   .exec(TestHCQ.runtime, TestHCQ.kernargs_ba_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size) \
                                   .timestamp(sig_en) \
@@ -337,11 +338,13 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
+    host_us = (time.perf_counter() - st) * 1e6

    et = float(sig_en.timestamp - sig_st.timestamp)

    print(f"exec kernel time: {et:.2f} us")
-    assert 0.1 <= et <= (3000000 if MOCKGPU or Device.DEFAULT in {"CPU"} else 100)
+    # emulated devices are only bounded by the host window around submit+wait
+    assert 0.1 <= et <= (host_us if MOCKGPU or Device.DEFAULT in {"CPU"} else 100)

  def test_speed_copy_bandwidth(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -352,6 +355,7 @@ class TestHCQ(unittest.TestCase):
    b = Buffer(Device.DEFAULT, SZ, dtypes.uint8, options=BufferSpec(nolru=True)).allocate()

    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
+    st = time.perf_counter()
    TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
                                .copy(a._buf, b._buf, SZ) \
                                .timestamp(sig_en) \
@@ -359,13 +363,14 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
+    host_ms = (time.perf_counter() - st) * 1e3

-    et = float(sig_en.timestamp - sig_st.timestamp)
-    et_ms = et / 1e3
+    et_ms = float(sig_en.timestamp - sig_st.timestamp) / 1e3
+    assert 0 < et_ms <= host_ms # timestamps are in us and cover only the copy

    gb_s = ((SZ / 1e9) / et_ms) * 1e3
    print(f"same device copy:  {et_ms:.2f} ms, {gb_s:.2f} GB/s")
-    assert (0.2 if MOCKGPU else 10) <= gb_s <= 1000
+    assert (0 if MOCKGPU else 10) <= gb_s <= 1000

  def test_speed_cross_device_copy_bandwidth(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -379,6 +384,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.allocator._map(b._buf)

    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
+    st = time.perf_counter()
    TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
                                .copy(a._buf, b._buf, SZ) \
                                .timestamp(sig_en) \
@@ -386,13 +392,14 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
+    host_ms = (time.perf_counter() - st) * 1e3

-    et = float(sig_en.timestamp - sig_st.timestamp)
-    et_ms = et / 1e3
+    et_ms = float(sig_en.timestamp - sig_st.timestamp) / 1e3
+    assert 0 < et_ms <= host_ms # timestamps are in us and cover only the copy

    gb_s = ((SZ / 1e9) / et_ms) * 1e3
    print(f"cross device copy: {et_ms:.2f} ms, {gb_s:.2f} GB/s")
-    assert (0.2 if MOCKGPU else 2) <= gb_s <= 100
+    assert (0 if MOCKGPU else 2) <= gb_s <= 100

  def test_timeline_signal_rollover(self):
    for queue_type in [TestHCQ.d0.hw_compute_queue_t, TestHCQ.d0.hw_copy_queue_t]:
@@ -1,7 +1,8 @@
 import unittest
-from tinygrad.device import CompileError, Device, BufferSpec
+from tinygrad.device import CompileError, Device, BufferSpec, TinyELF
+from tinygrad.helpers import Target
 if Device.DEFAULT=="METAL":
-  from tinygrad.runtime.ops_metal import MetalDevice, MetalCompiler, MetalProgram
+  from tinygrad.runtime.ops_metal import MetalDevice, MetalCompiler
@unittest.skipIf(Device.DEFAULT!="METAL", "Metal support required")
 class TestMetal(unittest.TestCase):
  def test_alloc_oom(self):
@@ -48,7 +49,7 @@ kernel void r_5(device int* data0, const device int* data1, uint3 gid [[threadgr
 """)
    with self.assertRaises(RuntimeError):
      compiled = compiled[:40] # corrupt the compiled program
-      MetalProgram(device, "r_5", compiled)
+      device.runtime(TinyELF(compiled, "r_5", Target("METAL"), ()))

  def test_free(self):
    size = 2**16
@@ -1,18 +1,20 @@
 import unittest
 from unittest.mock import patch
 from tinygrad import Device
-from tinygrad.device import Buffer
+from tinygrad.device import Buffer, TinyELF
 from tinygrad.dtype import dtypes
-from tinygrad.runtime.ops_cl import CLDevice, CLAllocator, CLCompiler, CLProgram
+from tinygrad.helpers import Target
+from tinygrad.runtime.ops_cl import CLDevice, CLAllocator, CLCompiler

@unittest.skipUnless(Device.DEFAULT == "CL", "Runs only on OpenCL")
 class TestCLCompileCache(unittest.TestCase):
  def test_compile_cached(self):
    device = Device[Device.DEFAULT]
    src = "__kernel void cached_test(__global int* a) { a[0] = 1; }"
-    CLProgram(device, name="cached_test", lib=src.encode())
+    obj = TinyELF(src.encode(), "cached_test", Target("CL"), ())
+    device.runtime(obj)
    with patch.object(CLCompiler, 'compile', side_effect=RuntimeError("compile should not be called on cache hit")):
-      CLProgram(device, name="cached_test", lib=src.encode())
+      device.runtime(obj)

@unittest.skipUnless(Device.DEFAULT == "CL", "Runs only on OpenCL")
 class TestCLError(unittest.TestCase):
@@ -27,7 +29,7 @@ class TestCLError(unittest.TestCase):
  def test_invalid_kernel_name(self):
    device = Device[Device.DEFAULT]
    with self.assertRaises(RuntimeError) as err:
-      CLProgram(device, name="", lib="__kernel void test(__global int* a) { a[0] = 1; }".encode())
+      device.runtime(TinyELF(b"__kernel void test(__global int* a) { a[0] = 1; }", "", Target("CL"), ()))
    assert str(err.exception) == "OpenCL Error -46: CL_INVALID_KERNEL_NAME"

  def test_unaligned_copy(self):
--- a/Show More
+++ b/Show More