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
253 changed files with 7807 additions and 12236 deletions
@@ -4,7 +4,7 @@ inputs:
  python-version:
    description: 'Python version to use'
    required: false
-    default: '3.14'
+    default: '' # if you don't set a version, the native python version will be used
  key:
    description: 'Key for the python cache'
    required: false
@@ -41,12 +41,12 @@ inputs:
    description: "Install LLVM?"
    required: false
    default: 'false'
-  qemu:
-    description: "Install qemu?"
+  tinydreno:
+    description: "Install tinydreno"
    required: false
    default: 'false'
-  ninja:
-    description: "Install ninja?"
+  qemu:
+    description: "Install qemu"
    required: false
    default: 'false'
 runs:
@@ -59,18 +59,18 @@ runs:
        echo "OMP_NUM_THREADS=1" >> "$GITHUB_ENV"
        # no buffers should be over 300MB in CI
        echo "MAX_BUFFER_SIZE=300000000" >> "$GITHUB_ENV"
-        if [[ "$RUNNER_OS" == "Linux" ]]; then
-          echo "VIRTUAL_ENV=/opt/venv/${{ inputs.python-version }}" >> "$GITHUB_ENV"
-          echo "UV_PYTHON_INSTALL_DIR=/opt/python" >> "$GITHUB_ENV"
-        else
-          echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
-        fi

    - name: Set up uv
      uses: astral-sh/setup-uv@08807647e7069bb48b6ef5acd8ec9567f424441b
      with:
        enable-cache: 'false' # see below for manual caching

+    - name: Set up Python ${{ inputs.python-version }}
+      uses: actions/setup-python@v6
+      if: inputs.python-version != ''
+      with:
+        python-version: ${{ inputs.python-version }}
+
    # **** Caching packages ****

    - name: Cache Python packages (PR)
@@ -109,15 +109,15 @@ runs:
      if: inputs.deps != ''
      shell: bash
      run: |
-        uv venv --allow-existing --python ${{ inputs.python-version }} "$VIRTUAL_ENV"
+        uv venv .venv
        DEPS="${{ inputs.deps }}"
-        uv pip install --python "$VIRTUAL_ENV" -e ".[${DEPS// /,}]" ${{ inputs.pydeps }} --torch-backend cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
+        uv pip install --python .venv -e ".[${DEPS// /,}]" ${{ inputs.pydeps }} --torch-backend cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
    - name: Install dependencies in venv (without extra)
      if: inputs.deps == ''
      shell: bash
      run: |
-        uv venv --allow-existing --python ${{ inputs.python-version }} "$VIRTUAL_ENV"
-        uv pip install --python "$VIRTUAL_ENV" -e . ${{ inputs.pydeps }}
+        uv venv .venv
+        uv pip install --python .venv -e . ${{ inputs.pydeps }}
    - name: Prune uv cache
      if: github.event_name != 'pull_request'
      shell: bash
@@ -125,15 +125,16 @@ runs:
    - name: Configure venv
      shell: bash
      run: |
+        echo "VIRTUAL_ENV=${{ github.workspace }}/.venv" >> "$GITHUB_ENV"
        if [[ "$RUNNER_OS" == "Windows" ]]; then
-          echo "$VIRTUAL_ENV/Scripts" >> "$GITHUB_PATH"
+          echo "${{ github.workspace }}/.venv/Scripts" >> "$GITHUB_PATH"
        else
-          echo "$VIRTUAL_ENV/bin" >> "$GITHUB_PATH"
+          echo "${{ github.workspace }}/.venv/bin" >> "$GITHUB_PATH"
        fi

    # ******************* apt *******************
    - name: Setup apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true')
+      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
@@ -161,7 +162,7 @@ runs:
        echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list

    - name: Compute Package List + Hash
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true')
      id: apt-pkgs
      shell: bash
      run: |
@@ -186,29 +187,25 @@ runs:
        if [[ "${{ inputs.qemu }}" == "true" ]]; then
          pkgs+=" qemu-user-static"
        fi
-        # **** ninja ****
-        if [[ "${{ inputs.ninja }}" == "true" ]]; then
-          pkgs+=" ninja-build"
-        fi

        echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
        echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"

    - name: Cache apt (PR)
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && github.event_name == 'pull_request'
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true') && github.event_name == 'pull_request'
      uses: actions/cache/restore@v5
      with:
        path: /var/cache/apt/archives/
        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
    - name: Cache apt
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true') && github.event_name != 'pull_request'
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true') && github.event_name != 'pull_request'
      uses: actions/cache@v5
      with:
        path: /var/cache/apt/archives/
        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}

    - name: Run apt Update + Install
-      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true' || inputs.ninja == 'true')
+      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true' || inputs.qemu == 'true')
      shell: bash
      run: |
        sudo apt -qq update || true
@@ -280,6 +277,12 @@ runs:
      shell: bash
      run: brew install llvm@20

+    # *** tinydreno ***
+    - name: Install tinydreno (linux)
+      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
+
    # *** OpenCL ***
    - name: Install rusticl
      if: inputs.opencl == 'true'
@@ -35,15 +35,15 @@ jobs:
        key: 'autogen'
        amd: 'true'
        llvm: 'true'
-        deps: 'autogen'
+        pydeps: 'pyyaml mako'
    - name: Install autogen support packages
      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev libdrm-dev liburing-dev
    - name: Regenerate autogen files
      run: |
-        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
+        find tinygrad/runtime/autogen -type f -name "*.py" -not -path "*/amd/*" -not -name "__init__.py" -not -name "comgr.py" -not -name "metal.py" -not -name "iokit.py" -not -name "corefoundation.py" -not -name "libclang.py" -delete
        python3 -c "from tinygrad.runtime.autogen import opencl"
        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv_580, nv_610, nv"
-        python3 -c "from tinygrad.runtime.autogen import comgr, comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
+        python3 -c "from tinygrad.runtime.autogen import comgr_3, hsa, hip, amd_gpu, sqtt, rocprof, amdgpu_kd, amdgpu_drm"
        python3 -c "from tinygrad.runtime.autogen.am import *"
        python3 -c "from tinygrad.runtime.autogen.nv_regs import *"
        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, pci, vfio"
@@ -54,7 +54,7 @@ jobs:
        python3 -c "from tinygrad.runtime.autogen import mesa"
        python3 -c "from tinygrad.runtime.autogen import avcodec"
        python3 -c "from tinygrad.runtime.autogen import llvm_qcom"
-        python3 -c "from tinygrad.runtime.autogen import mlx5, bnxt"
+        python3 -c "from tinygrad.runtime.autogen import mlx5"
        python3 -c "from tinygrad.runtime.autogen import ggml_common"
        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
    - name: Check for differences
@@ -102,3 +102,42 @@ jobs:
      with:
        name: autogen-macos-patch
        path: autogen-macos.patch
+
+  autogen-comgr-2:
+    name: In-tree Autogen (comgr 2)
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: 'autogen-comgr'
+    - name: Install autogen support packages
+      run: |
+        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+        sudo tee /etc/apt/sources.list.d/rocm.list <<EOF
+        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.2 $(lsb_release -cs) main
+        EOF
+        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
+        sudo apt -qq update || true
+        sudo apt-get install -y --no-install-recommends libclang-20-dev comgr
+    - name: Regenerate autogen files
+      run: |
+        rm tinygrad/runtime/autogen/comgr.py
+        python3 -c "from tinygrad.runtime.autogen import comgr"
+    - name: Check for differences
+      run: |
+        if ! git diff --quiet; then
+          git diff
+          git diff > autogen-comgr2.patch
+          echo "Autogen mismatch detected. Patch available at: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}#artifacts"
+          exit 1
+        fi
+    - name: Upload patch artifact
+      if: failure()
+      uses: actions/upload-artifact@v7
+      with:
+        name: autogen-comgr2-patch
+        path: autogen-comgr2.patch
@@ -88,7 +88,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 30
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -102,11 +102,16 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/hcq/hcq_smi.py amd rmmod --expect
-        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
+        ./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: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p weights
+        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -116,14 +121,18 @@ 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 qwen3.8
-      # qwen3.8:27b 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=qwen38_27b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.8:27b --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' }}
      run: BENCHMARK_LOG=olmoe JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m olmoe --benchmark --warmup
+    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
+      # only run on machines with multiple gpus
+      if: ${{ matrix.dev != 'METAL' }}
+      run: BENCHMARK_LOG=llama3_beam_4gpu JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -134,7 +143,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 10
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -148,11 +157,12 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/hcq/hcq_smi.py amd rmmod --expect
-        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
+        ./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: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
+      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: |
@@ -172,6 +182,10 @@ jobs:
      # slow on metal
      if: ${{ matrix.dev != 'METAL' }}
      run: time BENCHMARK_LOG=cifar DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run full CIFAR training steps w 6 GPUS
+      # only run on machines with multiple gpus
+      if: ${{ matrix.dev != 'METAL' }}
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -182,7 +196,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['AMD', 'NV']
-    timeout-minutes: 5
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -196,11 +210,12 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/hcq/hcq_smi.py amd rmmod --expect
-        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
+        ./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: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
+      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p extra/datasets
@@ -212,8 +227,15 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
+    - name: Run MLPerf resnet eval on training data
+      run: time BENCHMARK_LOG=resnet_eval MODEL=resnet python3 examples/mlperf/model_eval.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
      run: BENCHMARK_LOG=resnet_10steps DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+      run: BENCHMARK_LOG=resnet_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf Bert training steps (6 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -224,7 +246,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 15
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -238,11 +260,12 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/hcq/hcq_smi.py amd rmmod --expect
-        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
+        ./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: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
+      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: |
@@ -262,59 +285,6 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

-  multigpubenchmark:
-    name: Multi-GPU Benchmarks (DEV=${{ matrix.dev }})
-    runs-on: [self-hosted, "${{ matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
-    strategy:
-      fail-fast: false
-      matrix:
-        dev: ['AMD', 'NV']
-    timeout-minutes: 20
-    defaults:
-      run:
-        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
-      uses: actions/checkout@v6
-    - name: Setup (AMD)
-      if: ${{ matrix.dev == 'AMD' }}
-      run: |
-        ./extra/hcq/hcq_smi.py amd rmmod --expect
-        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
-    - name: Setup (NV)
-      if: ${{ matrix.dev == 'NV' }}
-      run: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        mkdir -p extra/datasets
-        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
-        ln -s /raid/datasets/imagenet extra/datasets/imagenet
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: reset process replay
-      run: python3 test/external/process_replay/reset.py
-    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run MLPerf resnet eval on training data
-      run: time BENCHMARK_LOG=resnet_eval MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf Bert training steps (6 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
  tests:
    name: Tests (DEV=${{ matrix.dev }})
    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
@@ -322,7 +292,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['METAL', 'AMD', 'NV']
-    timeout-minutes: 10
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -335,11 +305,12 @@ jobs:
    - name: Setup (AMD)
      if: ${{ matrix.dev == 'AMD' }}
      run: |
-        ./extra/hcq/hcq_smi.py amd rmmod --expect
-        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
+        ./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: lsof -tQ /dev/nvidia* | { xargs -r kill -9 || true; }
+      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: |
@@ -416,7 +387,7 @@ jobs:
  testusbgpu:
    name: UsbGPU Benchmark
    runs-on: [self-hosted, macOS]
-    timeout-minutes: 3
+    timeout-minutes: 10
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -431,70 +402,32 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: Kill stale pids
      run: |
-        ./extra/hcq/hcq_smi.py amd kill_pids --sudoless
-        ./extra/hcq/hcq_smi.py nv kill_pids --sudoless
+        PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
+        PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
+    # since sudo is required for usbgpu on macos, do not write bytecode, as some of the files are owned by root
    - name: UsbGPU boot time
-      run: GMMU=0 DEBUG=2 AM_RESET=1 DEV=USB+AMD time python3.11 test/test_tiny.py TestTiny.test_plus
+      run: sudo -E PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. GMMU=0 DEBUG=2 AM_RESET=1 DEV=USB+AMD time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU tiny tests
-      run: GMMU=0 DEV=USB+AMD python3.11 test/test_tiny.py
+      run: sudo -E PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/test_tiny.py
    - name: UsbGPU copy speeds
-      run: SIZE=64000000 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+      run: sudo -E PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+    #- name: UsbGPU openpilot test
+    #  run: sudo -E PYTHONPATH=. GMMU=0 DEV=USB+AMD GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
    - name: UsbGPU (USB4/TB) install script
-      run: sh extra/setup_tinygpu_osx.sh
+      run: PYTHONPATH=. sh extra/setup_tinygpu_osx.sh
    - name: UsbGPU (USB4/TB) boot time
-      run: DEBUG=3 DEV=PCI+NV:NAK time python3.11 test/test_tiny.py TestTiny.test_plus
+      run: PYTHONPATH=. DEBUG=3 DEV=PCI+NV:NAK time python3.11 test/test_tiny.py TestTiny.test_plus
    - name: UsbGPU (USB4/TB) tiny tests
-      run: DEV=PCI+NV:NAK python3.11 test/test_tiny.py
+      run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py

-  testcomma:
-    strategy:
-      matrix:
-        dev: ['QCOM', 'QCOM:IR3']
-        version: ['0.11.0', '0.11.2']
-        model: ['vision', 'policy', 'supercombo', 'dmonitoring']
-        # exclude non-existent models
-        exclude: [{ version: '0.11.0', model: supercombo }, { version: '0.11.2', model: vision }, { version: '0.11.2', model: policy }]
-        include:
-          - version: '0.11.0'
-            model: vision
-            url: https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
-            timing: 17
-          - version: '0.11.0'
-            model: policy
-            url: https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
-            timing: 3.2
-          - version: '0.11.0'
-            model: dmonitoring
-            url: https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
-            timing: 11
-          - version: '0.11.2'
-            model: supercombo
-            url: https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/433f85f956837606ad1f1cbee4aa7e2158ad23c768dea914b20436c97232741b
-            timing: 26
-          - dev: QCOM:IR3
-            version: '0.11.2'
-            model: supercombo
-            timing: 41
-          - version: '0.11.2'
-            model: dmonitoring
-            url: https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/3e7b31dfbc0a5234f1baf196513b77fc6af12204b8a8ffe8ee0417e48352f316
-            timing: 11
-          # IR3 dmonitoring is slightly slower
-          - dev: QCOM:IR3
-            model: dmonitoring
-            timing: 12
-      fail-fast: false
-    name: openpilot ${{ matrix.version }} compile3 ${{ matrix.model }} (DEV=${{ matrix.dev }})
+  testcommalatest:
+    name: comma Benchmark (0.11.2)
    runs-on: [self-hosted, Linux, comma]
-    timeout-minutes: 5
+    timeout-minutes: 12
    defaults:
      run:
        shell: bash -e -o pipefail {0}
    if: github.repository_owner == 'tinygrad'
-    env:
-      DEV: ${{ matrix.dev }}
-      ASSERT_MIN_STEP_TIME: ${{ matrix.timing }}
-      BENCHMARK_LOG: ${{ matrix.dev == 'QCOM:IR3' && 'ir3_' || '' }}openpilot_${{ matrix.version }}_${{ matrix.model }}
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
@@ -505,10 +438,45 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: compile
-      run: FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py ${{ matrix.url }}
-    - name: run pickle
-      run: BENCHMARK_LOG="${BENCHMARK_LOG}_run_pickle" RUN_PICKLE=1 taskset -c 4-7 python3 examples/openpilot/compile3.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
+    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 0.11.0 driving_vision
+      run: BENCHMARK_LOG=openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.11.0 driving_vision (from pickle)
+      run: BENCHMARK_LOG=openpilot_0_11_0_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM taskset -c 4-7 python3 examples/openpilot/compile3.py
+    - name: IR3 openpilot compile3 0.11.0 driving_vision
+      run: BENCHMARK_LOG=ir3_openpilot_0_11_0_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=18 DEV=QCOM:IR3 FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.11.0 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_11_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=3.2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.11.0/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.11.0 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_11_0_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/v0.11.0/selfdrive/modeld/models/dmonitoring_model.onnx
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

@@ -521,6 +489,15 @@ jobs:
        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: Checkout Code
      uses: actions/checkout@v6
    - name: setup staging db
@@ -544,7 +521,7 @@ jobs:
  testcommausbgpubenchmark:
    name: UsbGPU Benchmark (comma)
    runs-on: [self-hosted, Linux, comma4]
-    timeout-minutes: 10
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -564,7 +541,7 @@ jobs:
    - 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=64000000 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+      run: SIZE=64e6 PYTHONPATH=. GMMU=0 DEV=USB+AMD python3 test/external/external_test_usb_asm24.py TestDevCopySpeeds

  driverbenchmarks:
    name: PCI Driver Benchmark (DEV=${{ matrix.dev }})
@@ -573,7 +550,7 @@ jobs:
      fail-fast: false
      matrix:
        dev: ['AMD', 'NV']
-    timeout-minutes: 5
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -585,8 +562,9 @@ jobs:
      uses: actions/checkout@v6
    - name: Setup
      run: |
-        ./extra/hcq/hcq_smi.py ${{ matrix.dev }} rmmod --expect
-        ./extra/hcq/hcq_smi.py ${{ matrix.dev }} kill_pids --sudoless
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} rmmod
+        ./extra/hcq/hcq_smi.py ${{ matrix.dev == 'AMD' && 'amd' || 'nv' }} kill_pids
        mkdir -p extra/datasets
        ln -s /raid/datasets/imagenet extra/datasets/imagenet
    - name: setup staging db
@@ -621,9 +599,6 @@ jobs:
      run: |
        GRAPH_ONE_KERNEL=1 NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        GRAPH_ONE_KERNEL=1 NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
-    - name: HEVC Decode Benchmark
-      if: ${{ matrix.dev == 'NV' }}
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 PYTHONPATH=. python3 extra/hevc/decode.py
    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
      if: ${{ matrix.dev == 'NV' }}
      run: BENCHMARK_LOG=resnet_10steps MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
@@ -646,7 +621,7 @@ jobs:
  llvmspeed:
    name: LLVM Speed
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 5
+    timeout-minutes: 20
    if: github.repository_owner == 'tinygrad'
    steps:
    - name: Checkout Code
@@ -8,7 +8,7 @@ permissions:
  contents: write
 jobs:
  deploy:
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
      - name: Configure Git Credentials
@@ -166,7 +166,7 @@ jobs:
        uses: ./.github/actions/setup-tinygrad
        with:
          key: windows-${{ matrix.dev }}-minimal
-          deps: testing_minimal
+          deps: testing_unit
          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
      - name: Set env
        shell: bash
@@ -179,3 +179,35 @@ jobs:
      - 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
@@ -10,7 +10,7 @@ on:

 jobs:
  deploy:
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    steps:
    - uses: actions/checkout@v6
    - name: Set up Python
@@ -10,7 +10,7 @@ concurrency:
 jobs:
  checkbranch:
    name: Check PR Branch status
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    outputs:
      branchstat: ${{ steps.brstat.outputs.stat}}
    steps:
@@ -44,7 +44,7 @@ jobs:
    permissions:
      contents: read
      pull-requests: write
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    needs: checkbranch
    if: needs.checkbranch.outputs.branchstat == 'false'
    steps:
@@ -87,7 +87,7 @@ jobs:
    name: Core Library Line Difference
    permissions:
      pull-requests: write
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-latest
    needs: checkbranch
    if: needs.checkbranch.outputs.branchstat == 'true'
    steps:
@@ -21,7 +21,7 @@ concurrency:
 jobs:
  docs:
    name: Docs
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: &linux ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 10
    env:
      CHECK_OOB: 0
@@ -31,7 +31,8 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        deps: "docs testing_minimal"
+        deps: docs
+        pydeps: "capstone torch"
    - name: Build wheel and show size
      run: |
        uv build --wheel
@@ -60,7 +61,7 @@ jobs:

  torchbackend:
    name: Torch Backend Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -72,7 +73,10 @@ jobs:
        deps: testing_unit
        pydeps: "pillow torchvision expecttest"
        llvm: 'true'
-        ninja: 'true'
+    - name: Install ninja
+      run: |
+        sudo apt update || true
+        sudo apt install -y --no-install-recommends ninja-build
    - name: Test ResNet-18
      run: DEBUG=2 python3 extra/torch_backend/example.py
    - name: Test one op in torch tests
@@ -82,26 +86,9 @@ jobs:
    - name: Custom tests
      run: DEV=CPU:LLVM GPUS=4 TINY_BACKEND=1 python3 -m pytest -nauto extra/torch_backend/test.py extra/torch_backend/test_inplace.py extra/torch_backend/test_multigpu.py extra/torch_backend/test_kernel_fusion.py --durations=20

-  torchbackendtrain:
-    name: Torch Backend Training
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
-        llvm: 'true'
-        ninja: 'true'
-    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 DEV=CPU TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
-
  bepython:
    name: Python Backend
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -139,7 +126,7 @@ jobs:

  linter:
    name: Linters
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 10

    steps:
@@ -170,7 +157,7 @@ jobs:

  nulltest:
    name: Null Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15

    steps:
@@ -204,7 +191,7 @@ jobs:

  unittest:
    name: Unit Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15

    steps:
@@ -233,7 +220,7 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
    - name: Repo line count <= 26000 lines
-      run: MAX_LINE_COUNT=26500 python sz.py
+      run: MAX_LINE_COUNT=26000 python sz.py

  spec:
    strategy:
@@ -241,7 +228,7 @@ jobs:
      matrix:
        group: [1, 2]
    name: SPEC=2 (${{ matrix.group }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -257,7 +244,7 @@ jobs:

  fuzzing:
    name: Fuzzing
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 10
    steps:
    - name: Checkout Code
@@ -273,7 +260,7 @@ jobs:

  testopenclimage:
    name: CL IMAGE Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -293,7 +280,7 @@ jobs:

  testopenpilot:
    name: openpilot Compile Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -322,7 +309,7 @@ jobs:

  testonnxcpu:
    name: ONNX (CPU) Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20

    steps:
@@ -341,7 +328,7 @@ jobs:

  testoptim:
    name: Optimization Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    steps:
      - name: Checkout Code
@@ -373,7 +360,7 @@ jobs:

  testllm:
    name: Test LLM
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    env:
      CHECK_OOB: 0
@@ -400,7 +387,7 @@ jobs:

  testmodels:
    name: Models
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -420,7 +407,7 @@ jobs:

  testdsp:
    name: Linux (DSP)
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -448,7 +435,7 @@ jobs:
          - 'WEBGPU'

    name: Linux (DEV=${{ matrix.dev }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    steps:
    - name: Checkout Code
@@ -474,7 +461,7 @@ jobs:

  testamdasm:
    name: AMD ASM IDE
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    env:
      DEV: MOCKKFD+AMD
@@ -504,7 +491,7 @@ jobs:
      - name: Run AMD renderer tests (AMD:LLVM)
        run: DEV=MOCKKFD+AMD:LLVM python -m pytest -n=auto test/amd/ --durations 20
      - name: Run SQTT profiling tests
-        run: VIZ=-2 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
+        run: PROFILE=1 SQTT=1 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
      - name: Run AMD emulated tests on NULL backend
        env:
          AMD: 0
@@ -520,7 +507,7 @@ jobs:

  hcq2:
    name: hcq2
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 5
    steps:
      - name: Checkout Code
@@ -534,15 +521,16 @@ jobs:
      - name: Run HCQ2 tests
        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/test_tiny.py
      - name: Run HCQ2 multi-device tests
-        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest -n=auto test/backend/test_multitensor.py
+        run: |
+          HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_multitensor.py \
+            TestMultiTensor.test_simple_add TestMultiTensor.test_shard_reduce \
+            TestMultiTensor.test_backward_sum TestMultiTensor.test_matmul_shard_0_0
      - name: Run HCQ2 JIT tests
        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_jit.py
-      - name: Run HCQ2 unit tests
-        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest test/device/test_hcq2.py

  testmockam:
    name: Linux (am)
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    env:
      DEV: MOCKPCI+AMD
@@ -578,7 +566,7 @@ jobs:
        arch: [gfx1100, gfx1201, gfx950]

    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    env:
      DEV: MOCKKFD+AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
@@ -601,7 +589,7 @@ jobs:
        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
@@ -616,7 +604,7 @@ jobs:
        backend: [ptx, nv]

    name: Linux (${{ matrix.backend }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    env:
      FORWARD_ONLY: 1
@@ -650,17 +638,10 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        dev:
-          - 'NULL:IR3:a630'
-          - 'NULL:QCOMCL:a630'
-          - 'NULL:NAK:sm_120'
-    name: Compile-only (DEV=${{ matrix.dev }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+        backend: [ir3, nak]
+    name: Compile-only (${{ matrix.backend }})
+    runs-on: *linux
    timeout-minutes: 15
-    env:
-      NULL_ALLOW_COPYOUT: 1
-      DEV: ${{ matrix.dev }}${{ contains(matrix.dev, 'a630') && ',IMAGE_PITCH_ALIGNMENT=64' || '' }}
-      IMAGE: ${{ contains(matrix.dev, 'a630') && '1' || '0' }}
    steps:
      - name: Checkout Code
        uses: actions/checkout@v6
@@ -669,15 +650,21 @@ jobs:
        with:
          key: compile-${{ matrix.backend }}
          deps: "testing_unit mesa"
-          qemu: ${{ contains(matrix.dev, 'QCOMCL') }}
-      - name: Test IMAGE
+      - name: Set env
        shell: bash
-        if: contains(matrix.dev, 'a630')
-        run: DEBUG=7 python3 test/backend/test_ops.py TestOps.test_gemm | grep isam
+        run: printf "NULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'DEV=NULL:IR3:a630' || matrix.backend == 'nak' && 'DEV=NULL:NAK:sm_120' }}" >> $GITHUB_ENV
      - name: Run test_ops
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          # QCOMCL compiles in qemu, too slow for parallel workers
-          ${{ contains(matrix.dev, 'QCOMCL') && 'PARALLEL=0' || '' }} python -m pytest -n=auto test/backend/test_ops.py --durations=20
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+      - name: Run test_ops (IMAGE)
+        if: matrix.backend == 'ir3'
+        shell: bash
+        env:
+          IMAGE: 1
+          DEV: "NULL:IR3:a630,IMAGE_PITCH_ALIGNMENT=64"
+        run: |
+          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_gemm | grep image_load
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20
@@ -69,4 +69,3 @@ mutants
 dagre/
 graphlib/
 uv.lock
-pi_session_window0.jsonl
@@ -28,7 +28,7 @@ repos:
        pass_filenames: false
      - id: tests
        name: comprehensive test suite
-        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/backend/test_ops.py test/backend/test_schedule.py test/backend/test_assign.py test/backend/test_tensor.py test/backend/test_jit.py test/unit/test_schedule_cache.py test/null/test_pattern_matcher.py test/null/test_uop_symbolic.py test/unit/test_helpers.py
+        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/backend/test_ops.py test/backend/test_schedule.py test/unit/test_assign.py test/backend/test_tensor.py test/backend/test_jit.py test/unit/test_schedule_cache.py test/null/test_pattern_matcher.py test/null/test_uop_symbolic.py test/unit/test_helpers.py
        language: system
        always_run: true
        pass_filenames: false
@@ -4,4 +4,3 @@
 - 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
- Do not do amend commits. Always do a new commit if a force push to origin would be required.
@@ -140,7 +140,7 @@ Documentation along with a quick start guide can be found on the [docs website](
 ```python
 from tinygrad import Tensor

-x = Tensor.eye(3).clone()  # clone to make it a buffer
+x = Tensor.eye(3)
 y = Tensor([[2.0,0,-2.0]])
 z = y.matmul(x).sum()
 z.backward()
@@ -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.
@@ -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)
@@ -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()
@@ -1674,7 +1674,8 @@ def train_gptoss():
  config = {}
  BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4-8b/"))
  BS                 = config["BS"]                     = getenv("BS", 16)
-  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS
+  grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
+  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
  SEED               = config["SEED"]                   = getenv("SEED", 5760)
  DATA_SEED          = config["DATA_SEED"]              = getenv("DATA_SEED", SEED)
  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
@@ -1736,13 +1737,13 @@ def train_gptoss():
  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=1, 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=1, device=optim_device),
+    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:
-    p.grad = p.zeros_like(dtype=dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype).contiguous()
-    if getattr(p, "_zero2", False): p.grad = optim.optimizers[0]._zero_shard(p.grad)
+    grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
+    p.grad = p.zeros_like(dtype=grad_dtype).contiguous()
  grads = [p.grad for p in optim.params]

  from extra.gemm.cdna_asm_gemm import _mx_block_scale
@@ -1769,34 +1770,31 @@ def train_gptoss():

  @TinyJit
  @Context(TRAINING=1)
-  def step(tokens:Tensor):
+  def minibatch(tokens:Tensor):
    if is_dp: tokens = tokens.to(None).shard(device, 0)
    if not is_sharding: tokens = tokens.to(None)
-
    logits:Tensor = model(tokens[:, :-1], save=True)
-    if getenv("FUSED_CE", 0):
-      from extra.llama_kernels.fused_ce import fused_ce_loss
-      loss = fused_ce_loss(logits.cast(dtypes.bfloat16), tokens[:, 1:], label_smoothing=0.0)
-    else:
-      loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
+    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])

    for g, new_g in zip(grads, loss.gradient(*optim.params)):
      apply_grad(g, new_g.uop)

-    Tensor.realize(loss, *grads)
+    loss_cpu = loss.flatten().float().to("CPU")
+    return loss_cpu.realize(*grads)

-    grad_norm = clip_grads(grads, 1, 1.0)
+  @TinyJit
+  def optim_step():
+    grad_norm = clip_grads(grads, grad_acc, 1.0)
    optim.fstep(grads, grad_norm)
    scheduler.step()

    for g in grads: g.assign(0)

-    loss_cpu = loss.flatten().float().to("CPU")
    lr_cpu = optim.lr.float().to("CPU")
    grad_norm_cpu = grad_norm.float().to("CPU")
-    Tensor.realize(loss_cpu, lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales)
+    Tensor.realize(lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales)

-    return loss_cpu, lr_cpu, grad_norm_cpu
+    return lr_cpu, grad_norm_cpu

  @TinyJit
  @Context(TRAINING=0)
@@ -1845,20 +1843,30 @@ def train_gptoss():
      profile_marker(f"train @ {i}")
      st = time.perf_counter()

-      ist = time.perf_counter()
+      stopped = False
+      losses, data_time, dev_time = [], 0, 0
+      for _ in range(grad_acc if i >= 2 else 1):
+        ist = time.perf_counter()
+        try: tokens = next(train_iter)
+        except StopIteration:
+          stopped = True
+          break
+        mst = time.perf_counter()
+        data_time += mst - ist
+        losses.append(minibatch(tokens).item())
+        dev_time += time.perf_counter() - mst
+      if stopped: break

-      try: tokens = next(train_iter)
-      except StopIteration: break
-      mst = time.perf_counter()
-      data_time = mst - ist
-
-      ret = step(tokens)
-      dev_time = time.perf_counter() - mst
-
-      loss, lr, grad_norm = ret[0].item(), ret[1].item(), ret[2].item()
+      gt = time.perf_counter()
+      ret = optim_step()
+      lr, grad_norm = ret[0].item(), ret[1].item()
      et = time.perf_counter()

+      loss = sum(losses) / len(losses)
+      optim_time = et - gt
+      dev_time += optim_time
      step_time = et - st
+      gbs_time = gt - st
      if BENCHMARK: step_times.append(step_time)

      i += 1
@@ -1868,7 +1876,7 @@ def train_gptoss():
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 100
      tqdm.write(
-          f"{i:5} {step_time:.3f} s step, {dev_time:.3f} s dev, {data_time:.3f} s data, {loss:.4f} loss, " \
+          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")
      if DEBUG >= 1: tqdm.write("  mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))

@@ -1878,6 +1886,8 @@ def train_gptoss():
          "train/lr": lr,
          "train/grad_norm": grad_norm,
          "train/step_time": step_time,
+          "train/gbs_time": gbs_time,
+          "train/optim_time": optim_time,
          "train/dev_time": dev_time,
          "train/data_time": data_time,
          "train/mem": mem_gb,
@@ -12,7 +12,7 @@ from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
 from tinygrad.uop.ops import Ops, UOp
 from extra.models.llama import apply_rotary_emb
 from extra.llama_kernels.rmsnorm import rmsnorm
-from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8, asm_gemm, can_use_asm_gemm
+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

@@ -146,7 +146,6 @@ class GPTOSS:
      return w_q, w_e8.is_param_(False)
    if moe:
      qs = [_one(*shape[1:]) for _ in range(shape[0])]
-      for q in qs: q[0]._zero2 = True  # grad arrives sharded on the expert axis under ZeRO-2 (moe_gemm)
      return [q[0] for q in qs], [q[1] for q in qs]
    return _one(*shape)

@@ -183,14 +182,12 @@ class GPTOSS:
    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)  # (B,N,H,D)/(B,N,KV,D)

-    fa_saves = []
-    if getenv("HK_FLASH_ATTENTION"):
-      from extra.thunder.amd.fa import flash_attention
-      attn, _, l_vec = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks, window=self.sliding_window if sliding else 0)
-      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
-      fa_saves = [xq, xk, xv, l_vec]
-    elif sliding:
+    if sliding:
      attn = self._sliding_attention(xq, xk, xv, sinks)
+    elif getenv("HK_FLASH_ATTENTION"):
+      from extra.thunder.amd.fa import flash_attention
+      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks)
+      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
    else:
      xqm = xq.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
      xkm, xvm = xk.permute(0, 2, 1, 3).unsqueeze(2), xv.permute(0, 2, 1, 3).unsqueeze(2)
@@ -202,7 +199,7 @@ class GPTOSS:
      attn = (w @ xvm).permute(0, 3, 1, 2, 4).reshape(bsz, seqlen, self.n_heads * self.head_dim)

    out = matmul_mx(attn, wo, wo_scale) + wo_bias
-    return out, [x_normed, rrms, attn] + fa_saves
+    return out, [x_normed, rrms, attn]

  def feed_forward(self, x:Tensor, *, ffn_norm:Tensor, gate:Tensor, gate_bias:Tensor,
                   w_gate_up:Tensor, w_gate_up_scale:Tensor, w_gate_up_bias:Tensor,
@@ -223,7 +220,6 @@ class GPTOSS:
      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)
-      return out, [x_normed, rrms, xg, h, y, z, r.weights, r.dest_row, r.off]
    else:
      thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
      weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
@@ -267,11 +263,7 @@ class GPTOSS:
                        w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
      h, *_ = self.run_layer(h, freqs_cis, mask_full, i % 2 == 0, attn_kwargs, ffn_kwargs, save=save)

-    h_normed = self.norm(h)
-    pad = (-self.dim) % 256
-    h_padded, w_padded = h_normed.pad((None, None, (0, pad))), self.output.pad(((0, 0), (0, pad)))
-    if ASM_GEMM and can_use_asm_gemm(h_padded, w_padded.T): logits = asm_gemm(h_padded, w_padded.T)
-    else: logits = h_normed @ self.output.T
+    logits = self.norm(h) @ self.output.T
    return logits

 def _get_pads(uop:UOp) -> list[UOp]:
@@ -15,7 +15,7 @@ def stochastic_round_bf16(x:Tensor) -> Tensor:
  bits = x.bitcast(dtypes.uint32)
  if isinstance(x.device, tuple):
    shape = x.uop.shard_shape if x.uop.axis is not None else x.shape
-    noise = Tensor(UOp(Ops.MSTACK, src=tuple(Tensor.rand(*shape, device=d).uop for d in x.device)))
+    noise = Tensor(UOp(Ops.MSTACK, dtypes.default_float, tuple(Tensor.rand(*shape, device=d).uop for d in x.device)))
  else:
    noise = x.rand_like()
  noise = (noise * 0xFFFF).cast(dtypes.uint32)
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -35,7 +35,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=2048 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -44,7 +44,7 @@ export SEED=${SEED:-5760}
 export DATA_SEED=${DATA_SEED:-5760}

 export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1

 export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
@@ -1,8 +1,8 @@
 #!/usr/bin/env bash

 export PYTHONPATH="."
-export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
-export PATH="$ROCM_PATH/bin:$PATH"
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -35,7 +35,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=2048 MAX_STEPS=1200000
+export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -44,6 +44,6 @@ export SEED=${SEED:-$RANDOM}
 export DATA_SEED=${DATA_SEED:-5760}

 export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1

 python3 examples/mlperf/model_train.py
@@ -17,7 +17,7 @@ export USE_ATOMICS=1
 export ASM_GEMM=1
 export WQKV=1
 export MASTER_WEIGHTS=1
-export MXFP4=1
+export FP8=1
 export ALLREDUCE_CAST=1
 export FAST_CE=1
 export FUSED_INPUT_QUANTIZE=1
@@ -26,7 +26,7 @@ export FUSED_ADD_NORM_MUL_QUANTIZE=1
 export FUSED_SILU_W13=1
 export SPLIT_W13=0

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=8 MP=1 BS=16 EVAL_BS=8 GRADIENT_ACC_STEPS=2
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -44,7 +44,7 @@ export SEED=$RANDOM
 export DATA_SEED=$SEED

 export JITBEAM=3
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
+export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1

 export LOGMLPERF=1

@@ -11,7 +11,6 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -11,7 +11,6 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -107,21 +107,14 @@ def compile(onnx_file):
  return inputs, test_val

 def test_vs_compile(run, inputs, test_val=None):
-  if (log:=bool(getenv("BENCHMARK_LOG", ""))): from extra.bench_log import WallTimeEvent, BenchEvent

  # run 20 times
  step_times = []
  for _ in range(20):
    st = time.perf_counter()
-    if log:
-      with WallTimeEvent(BenchEvent.STEP):
-        out = run(**inputs)
-        mt = time.perf_counter()
-        val = out.numpy()
-    else:
-      out = run(**inputs)
-      mt = time.perf_counter()
-      val = out.numpy()
+    out = run(**inputs)
+    mt = time.perf_counter()
+    val = out.numpy()
    et = time.perf_counter()
    step_times.append((et-st)*1e3)
    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {step_times[-1]:6.2f} ms")
@@ -167,6 +160,12 @@ def test_vs_onnx(new_inputs, test_val, onnx_file, tol):
  print("test vs onnx passed")
  return timings

+def bench(run, inputs):
+  from extra.bench_log import WallTimeEvent, BenchEvent
+  for _ in range(10):
+    with WallTimeEvent(BenchEvent.STEP):
+      run(**inputs).numpy()
+
 if __name__ == "__main__":
  if getenv("RUN_PICKLE"):
    with open(OUTPUT, "rb") as f: pickle_loaded = load_pickle(f)
@@ -182,3 +181,6 @@ if __name__ == "__main__":
    test_vs_compile(pickle_loaded, inputs, outputs)
    if getenv("SELFTEST"):
      test_vs_onnx(inputs, outputs, onnx_file, 1e-4)
+
+  if getenv("BENCHMARK_LOG", ""):
+    bench(pickle_loaded, inputs)
@@ -84,8 +84,7 @@ class AMSMI(AMDev):
    with open(f"/sys/bus/pci/devices/{self.pcibus}/power_state", "r") as f: return f.read().strip().rstrip()

 class SMICtx:
-  def __init__(self, dev_filter=None):
-    self.dev_filter = dev_filter
+  def __init__(self):
    self.devs = []
    self.opened_pcidevs = []
    self.opened_pci_resources = {}
@@ -136,7 +135,6 @@ class SMICtx:
    pattern = os.path.join('/tmp', 'am_*.lock')
    for d in [f[8:-5] for f in glob.glob(pattern)]:
      if d.startswith("usb"): continue
-      if self.dev_filter is not None and d != self.dev_filter: continue
      if d not in self.opened_pcidevs:
        self._open_am_device(d)

@@ -408,7 +406,7 @@ if __name__ == "__main__":

  try:
    if not args.list: os.system('clear')
-    smi_ctx = SMICtx(args.dev)
+    smi_ctx = SMICtx()
    while True:
      smi_ctx.rescan_devs()
      smi_ctx.draw(args.list)
@@ -35,7 +35,7 @@ class WallTimeEvent:
    return self
  def __exit__(self, *_):
    self.time = time.monotonic() - self.start
-    _events[self.event]["wall"].append((self.time, BENCHMARK_LOG.value))
+    _events[self.event]["wall"].append(self.time)
    return False

 class KernelTimeEvent:
@@ -47,19 +47,19 @@ class KernelTimeEvent:
    self.start = GlobalCounters.time_sum_s
    return self
  def __exit__(self, *_):
-    _events[self.event]["kernel"].append((GlobalCounters.time_sum_s - self.start, BENCHMARK_LOG.value))
+    _events[self.event]["kernel"].append(GlobalCounters.time_sum_s - self.start)
    return False

 def log_event_instant(event:InstantBenchEvent, value:float):
-  _events[event].append((value, BENCHMARK_LOG.value))
+  _events[event].append(value)

 if BENCHMARK_LOG:
  INFLUXDB_HOST = getenv("INFLUXDB_HOST", "")
  INFLUXDB_ORG = getenv("INFLUXDB_ORG", "tiny")
  INFLUXDB_TOKEN = getenv("INFLUXDB_TOKEN", "")

-  def _create_point(run_id, i, attempt, ref, commit, name, value, log_name, run):
-    point = Point(log_name.replace(':', '_').replace('.', '_')).tag("id", run_id).tag("index", i)
+  def _create_point(run_id, i, attempt, ref, commit, name, value, run):
+    point = Point(BENCHMARK_LOG.value).tag("id", run_id).tag("index", i)
    point = point.tag("device", Device.DEFAULT)
    point = point.tag("attempt", attempt).tag("ref", ref).tag("commit", commit)
    point = point.field(name, value).field("x", run)
@@ -91,12 +91,12 @@ if BENCHMARK_LOG:
      run_id = str(uuid.uuid4())
      if isinstance(event, BenchEvent):
        for event_type, values in _events[event].items():
-          for i, (value, log_name) in enumerate(values):
-            point = _create_point(run_id, i, attempt, ref, commit, f"{event.value}_{event_type}", value, log_name, run)
+          for i, value in enumerate(values):
+            point = _create_point(run_id, i, attempt, ref, commit, f"{event.value}_{event_type}", value, run)
            points.append(point)
      else:
-        for i, (value, log_name) in enumerate(_events[event]):
-          point = _create_point(run_id, i, attempt, ref, commit, event.value, value, log_name, run)
+        for i, value in enumerate(_events[event]):
+          point = _create_point(run_id, i, attempt, ref, commit, event.value, value, run)
          points.append(point)

    write_options = WriteOptions(write_type=WriteType.synchronous, retry_interval=5000, max_retries=5, max_retry_delay=30000, exponential_base=2)
@@ -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()
@@ -1,238 +0,0 @@
-import ctypes, struct
-from tinygrad.helpers import ceildiv, getenv, wait_cond, DEBUG
-from tinygrad.runtime.autogen import bnxt, pci
-from tinygrad.runtime.support.system import PCIDevice, System, ipv4_to_gid
-
-BNXT_DEBUG = getenv("BNXT_DEBUG", 0)
-BNXT_ACCESS, BNXT_INIT_MASK, BNXT_RTR_MASK, BNXT_RTS_MASK = 3, 0xd, 0x41515ad, 0xae005
-BNXT_CHIMP_COMM, BNXT_CHIMP_COMM_TRIGGER = 0x0, 0x100
-BNXT_BACKING_STORE = ((0, 2), (1, 0), (2, 2), (3, 0), (4, 2), (5, 0), (6, 0), (14, 2), (15, 0))
-
-def db_value(xid, typ, index, epoch):
-  return (xid & bnxt.DBC_DBC_XID_MASK | bnxt.DBC_DBC_PATH_ROCE | typ | bnxt.BNXT_QPLIB_DBR_VALID) << 32 | \
-         index & bnxt.DBC_DBC_INDEX_MASK | epoch << bnxt.BNXT_QPLIB_DBR_EPOCH_SHIFT
-
-def _pbl(dev, paddrs, queue=False):
-  if len(paddrs) == 1: return 0, paddrs[0]
-  values = [p | bnxt.PTU_PTE_VALID for p in paddrs]
-  if queue:
-    values[-1] |= bnxt.PTU_PTE_LAST
-    if len(values) > 1: values[-2] |= bnxt.PTU_PTE_NEXT_TO_LAST
-  table, table_paddrs = dev.pci_dev.alloc_sysmem(ceildiv(len(values), 512) * 0x1000)
-  table[:len(values) * 8] = struct.pack(f"<{len(values)}Q", *values)
-  if len(table_paddrs) == 1: return 1, table_paddrs[0]
-  top, top_paddrs = dev.pci_dev.alloc_sysmem(0x1000)
-  top[:len(table_paddrs) * 8] = struct.pack(f"<{len(table_paddrs)}Q", *(p | bnxt.PTU_PTE_VALID for p in table_paddrs))
-  return 2, top_paddrs[0]
-
-def _queue(dev, stride:int=16, aux=False):
-  mem, paddrs = dev.pci_dev.alloc_sysmem(0x1000 + aux * 0x400)
-  level, base = _pbl(dev, paddrs, queue=True)
-  return {"mem":mem, "paddrs":paddrs, "stride":stride, "prod":0, "cons":0, "level":level, "base":base}
-
-def _qread(q, i):
-  off = (i & 15) * q["stride"]
-  return q["mem"][off:off + q["stride"]]
-
-def _qwrite(q, i, data, aux=False):
-  off = 0x1000 + i % 128 * 8 if aux else (i & 15) * q["stride"]
-  q["mem"][off:off + len(data)] = data
-
-class BNXTDev:
-  def __init__(self, pci_dev:PCIDevice, ip:str=getenv("BNXT_IP", "10.0.0.1")):
-    self.pci_dev, self.devfmt = pci_dev, pci_dev.pcibus
-    self.bar0, self.db = pci_dev.map_bar(0, fmt='I'), pci_dev.map_bar(2, fmt='Q')
-    pci_dev.write_config(pci.PCI_COMMAND, pci_dev.read_config(pci.PCI_COMMAND, 2) | pci.PCI_COMMAND_MASTER, 2)
-    self.resp, self.resp_pa = pci_dev.alloc_sysmem(0x1000)
-    self.seq = 0
-
-    ver = self.hwrm("ver_get")
-    if DEBUG >= 2: print(f"bnxt {self.devfmt}: firmware {ver.hwrm_fw_maj_8b}.{ver.hwrm_fw_min_8b}.{ver.hwrm_fw_bld_8b}")
-    self.hwrm("func_reset", timeout_ms=40000)
-    caps = self.hwrm("func_qcaps", fid=0xffff)
-    self.mac, self.port_id = int.from_bytes(bytes(caps.mac_address), 'big'), caps.port_id
-    self.hwrm("func_drv_rgtr")
-    self.db_off = self.hwrm("func_qcfg", fid=0xffff).legacy_l2_db_size_kb * 1024
-
-    self.setup_backing_store()
-    self._open_rcfw()
-    self._open_l2()
-    self.local_gid = ipv4_to_gid(ip)
-    gids, mac = (ctypes.c_uint32 * 4)(*(int.from_bytes(self.local_gid[i:i + 4], 'big') for i in (12, 8, 4, 0))), self.mac.to_bytes(6, 'big')
-    smac = (ctypes.c_uint16 * 3)(*(int.from_bytes(mac[i:i + 2], 'big') for i in (0, 2, 4)))
-    self.gid_id = self.rcfw("add_gid", gid=gids, src_mac=smac).xid
-
-    if DEBUG >= 2: print(f"bnxt {self.devfmt}: booted mac={self.mac.to_bytes(6, 'big').hex(':')} gid={self.local_gid.hex()}")
-
-  def hwrm(self, name, timeout_ms=10000, **fields):
-    inp, out = getattr(bnxt, f"struct_hwrm_{name}_input"), getattr(bnxt, f"struct_hwrm_{name}_output")
-    opcode = getattr(bnxt, f"HWRM_{name.upper()}")
-    self.seq = (self.seq + 1) & 0xffff
-    data = bytes(inp(req_type=opcode, cmpl_ring=bnxt.BNXT_HWRM_NO_CMPL_RING, seq_id=self.seq, target_id=bnxt.BNXT_HWRM_TARGET,
-                     resp_addr=self.resp_pa[0], **fields))
-    self.resp[:] = bytes(len(self.resp))
-    System.memory_barrier()
-    for i, w in enumerate(memoryview(bytearray(data.ljust(bnxt.HWRM_MAX_REQ_LEN, b'\0'))).cast('I')):
-      self.bar0[BNXT_CHIMP_COMM // 4 + i] = w
-    self.bar0[BNXT_CHIMP_COMM_TRIGGER // 4] = 1
-    def hdr(): return bnxt.struct_hwrm_resp_hdr.from_buffer_copy(bytes(self.resp[:8]))
-    wait_cond(lambda: (n := hdr().resp_len) and hdr().seq_id == self.seq and self.resp[n - 1], timeout_ms=timeout_ms, msg=f"HWRM {name}")
-    ret = out.from_buffer_copy(bytes(self.resp[:ctypes.sizeof(out)]))
-    assert ret.error_code == 0, f"HWRM {name}: {ret.error_code}"
-    return ret
-
-  def setup_backing_store(self):
-    counts: dict[int, int] = {}
-    for typ, extra in BNXT_BACKING_STORE:
-      caps = self.hwrm("func_backing_store_qcaps_v2", type=typ)
-      size, splits = caps.entry_size, tuple(getattr(caps, f"split_entry_{j}") for j in range(caps.subtype_valid_cnt))
-      counts[typ] = n = counts[0] if typ == 15 else max(caps.min_num_entries, sum(splits) + extra)
-      # a zero bitmap means the type has a single instance 0
-      for instance in [i for i in range(8) if caps.instance_bit_map >> i & 1] or [0]:
-        mem, paddrs = self.pci_dev.alloc_sysmem(ceildiv(n * size, 0x1000) * 0x1000)
-        if caps.ctx_init_value:
-          for off in range(caps.ctx_init_offset, len(mem), size): mem[off] = caps.ctx_init_value
-        lvl, base = _pbl(self, paddrs)
-        self.hwrm("func_backing_store_cfg_v2", type=typ, instance=instance, entry_size=size, num_entries=n, page_dir=base,
-          page_size_pbl_level=lvl, subtype_valid_cnt=len(splits),
-          flags=bnxt.FUNC_BACKING_STORE_CFG_V2_REQ_FLAGS_BS_CFG_ALL_DONE if typ == 15 else 0,
-          **{f"split_entry_{j}": v for j, v in enumerate(splits)})
-
-  def _open_rcfw(self):
-    self.rcfw_first = True
-
-    self.creq = _queue(self)
-    self.creq_id = self.hwrm("ring_alloc", ring_type=bnxt.RING_ALLOC_REQ_RING_TYPE_NQ, page_tbl_addr=self.creq["base"],
-      page_size=12, page_tbl_depth=self.creq["level"], length=16, int_mode=bnxt.RING_ALLOC_REQ_INT_MODE_MSIX).ring_id
-
-    self.cmdq = _queue(self)
-    self.doorbell(self.creq_id, bnxt.DBC_DBC_TYPE_NQ_ARM, 0, 0)
-    init = bnxt.struct_cmdq_init(cmdq_pbl=self.cmdq["base"], creq_ring_id=self.creq_id,
-                                 cmdq_size_cmdq_lvl=16 << bnxt.CMDQ_INIT_CMDQ_SIZE_SFT)
-
-    System.memory_barrier()
-    for i, w in enumerate(memoryview(bytearray(bytes(init))).cast('I')): self.bar0[bnxt.RCFW_COMM_BASE_OFFSET // 4 + i] = w
-
-    _, p = self.pci_dev.alloc_sysmem(0x1000)
-    self.rcfw("initialize_fw", stat_ctx_id=self.hwrm("stat_ctx_alloc", stats_dma_addr=p[0], stats_dma_length=176).stat_ctx_id,
-      flags=bnxt.CMDQ_INITIALIZE_FW_FLAGS_HW_REQUESTER_RETX_SUPPORTED)
-
-    # RoCE notification ring: never armed or serviced, but CQ and L2 ring allocation require one
-    nq = _queue(self)
-    self.nq_id = self.hwrm("ring_alloc", ring_type=bnxt.RING_ALLOC_REQ_RING_TYPE_NQ, page_tbl_addr=nq["base"],
-      page_size=12, page_tbl_depth=nq["level"], length=16, logical_id=1, int_mode=bnxt.RING_ALLOC_REQ_INT_MODE_MSIX).ring_id
-
-  def rcfw(self, name, timeout_ms=20000, **fields):
-    req_t, resp_t = getattr(bnxt, f"struct_cmdq_{name}"), getattr(bnxt, f"struct_creq_{name}_resp")
-    op = getattr(bnxt, f"CMDQ_BASE_OPCODE_{name.upper()}")
-    data = bytes(req_t(opcode=op, cmd_size=(slots := ceildiv(ctypes.sizeof(req_t), 16)), **fields)).ljust(slots * 16, b'\0')
-    for i in range(slots): _qwrite(self.cmdq, self.cmdq["prod"] + i, data[i * 16:(i + 1) * 16])
-
-    self.cmdq["prod"] += slots
-    prod = self.cmdq["prod"] & 0xffff
-    if self.rcfw_first: prod, self.rcfw_first = prod | 1 << bnxt.FIRMWARE_FIRST_FLAG, False
-
-    System.memory_barrier()
-
-    self.bar0[(bnxt.RCFW_COMM_BASE_OFFSET + bnxt.RCFW_PF_VF_COMM_PROD_OFFSET) // 4] = prod
-    self.bar0[(bnxt.RCFW_COMM_BASE_OFFSET + bnxt.RCFW_COMM_TRIG_OFFSET) // 4] = bnxt.RCFW_CMDQ_TRIG_VAL
-
-    def poll():
-      h = bnxt.struct_creq_base.from_buffer_copy(bytes(_qread(self.creq, self.creq["cons"])))
-      return bool(h.v & bnxt.CREQ_BASE_V) != bool((self.creq["cons"] // 16) & 1)
-    wait_cond(poll, timeout_ms=timeout_ms, msg=f"RCFW {name}")
-
-    ret = resp_t.from_buffer_copy(bytes(_qread(self.creq, self.creq["cons"])))
-    self.creq["cons"] += 1
-
-    # NQ_ARM also publishes the CREQ consumer index, which is what frees ring space for the next command
-    self.doorbell(self.creq_id, bnxt.DBC_DBC_TYPE_NQ_ARM, self.creq["cons"] & 15, (self.creq["cons"] // 16) & 1)
-    assert ret.status == 0, f"RCFW {name}: {ret.status}"
-
-    if BNXT_DEBUG >= 1: print(f"bnxt {self.devfmt}: rcfw {name} xid={getattr(ret, 'xid', 0):#x}")
-    return ret
-
-  def doorbell(self, xid, typ, index, epoch):
-    System.memory_barrier()
-    self.db[self.db_off // 8] = db_value(xid, typ, index, epoch)
-
-  # L2 receive path, required for RoCE ingress even though no ethernet receive buffers are posted
-  def _open_l2(self):
-    cq = _queue(self)
-    ci = self.hwrm("ring_alloc", enables=bnxt.RING_ALLOC_REQ_ENABLES_NQ_RING_ID_VALID, ring_type=bnxt.RING_ALLOC_REQ_RING_TYPE_L2_CMPL,
-      page_tbl_addr=cq["base"], page_size=12, page_tbl_depth=cq["level"], length=16, nq_ring_id=self.nq_id).ring_id
-    rx = _queue(self)
-    ri = self.hwrm("ring_alloc", enables=bnxt.RING_ALLOC_REQ_ENABLES_NQ_RING_ID_VALID |
-      bnxt.RING_ALLOC_REQ_ENABLES_RX_BUF_SIZE_VALID, ring_type=bnxt.RING_ALLOC_REQ_RING_TYPE_RX, page_tbl_addr=rx["base"],
-      page_size=12, page_tbl_depth=rx["level"], length=16, rx_buf_size=640, nq_ring_id=self.nq_id).ring_id
-    vi = self.hwrm("vnic_alloc").vnic_id
-    self.hwrm("vnic_cfg", enables=bnxt.VNIC_CFG_REQ_ENABLES_MRU | bnxt.VNIC_CFG_REQ_ENABLES_DEFAULT_RX_RING_ID |
-      bnxt.VNIC_CFG_REQ_ENABLES_DEFAULT_CMPL_RING_ID, vnic_id=vi, mru=9018,
-      default_rx_ring_id=ri, default_cmpl_ring_id=ci)
-    self.hwrm("cfa_l2_filter_alloc", flags=bnxt.CFA_L2_FILTER_ALLOC_REQ_FLAGS_PATH_RX,
-      enables=bnxt.CFA_L2_FILTER_ALLOC_REQ_ENABLES_L2_ADDR | bnxt.CFA_L2_FILTER_ALLOC_REQ_ENABLES_L2_ADDR_MASK |
-      bnxt.CFA_L2_FILTER_ALLOC_REQ_ENABLES_DST_ID, l2_addr=tuple(self.mac.to_bytes(6, 'big')), l2_addr_mask=(0xff,) * 6, dst_id=vi)
-
-  def register_mem(self, paddrs:list[int], size:int, log_page_size:int=12) -> int:
-    level, base = _pbl(self, paddrs[:ceildiv(size, 1 << log_page_size)])
-    return self.rcfw("register_mr", flags=bnxt.CMDQ_REGISTER_MR_FLAGS_ALLOC_MR,
-      log2_pg_size_lvl=level << bnxt.CMDQ_REGISTER_MR_LVL_SFT | log_page_size << bnxt.CMDQ_REGISTER_MR_LOG2_PG_SIZE_SFT,
-      access=bnxt.CMDQ_REGISTER_MR_ACCESS_LOCAL_WRITE | bnxt.CMDQ_REGISTER_MR_ACCESS_REMOTE_WRITE,
-      log2_pbl_pg_size=12, pbl=base, va=paddrs[0], mr_size=size).xid
-
-class BNXTQP:
-  def __init__(self, dev:BNXTDev):
-    self.dev, self.sq_psn, self.msn = dev, 0, 0
-
-    self.cqq = _queue(dev, ctypes.sizeof(bnxt.struct_cq_base))
-    self.cq_id = dev.rcfw("create_cq", cq_size=16, pbl=self.cqq["base"],
-                          pg_size_lvl=self.cqq["level"], cq_fco_cnq_id=dev.nq_id).xid
-
-    self.sq = _queue(dev, aux=True)
-    self.qpn = dev.rcfw("create_qp", type=bnxt.CMDQ_CREATE_QP_TYPE_RC,
-      sq_size=16, sq_fwo_sq_sge=1, scq_cid=self.cq_id, rcq_cid=self.cq_id,
-      sq_pbl=self.sq["base"], sq_pg_size_sq_lvl=self.sq["level"]).xid
-    self.qp_op(1, BNXT_INIT_MASK, access=BNXT_ACCESS, pkey=0xffff)
-
-  def qp_op(self, state, mask, network_type=0, **fields):
-    self.dev.rcfw("modify_qp", qp_cid=self.qpn, modify_mask=mask,
-      network_type_en_sqd_async_notify_new_state=state | network_type, **fields)
-
-  def connect(self, qpn:int, gid:bytes, mac:int):
-    network_type = bnxt.CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4
-    dgid = (ctypes.c_uint32 * 4)(*(int.from_bytes(gid[i:i + 4], 'little') for i in (0, 4, 8, 12)))
-    dmac = (ctypes.c_uint16 * 3)(*(int.from_bytes(mac.to_bytes(6, 'big')[i:i + 2], 'little') for i in (0, 2, 4)))
-
-    self.qp_op(2, BNXT_RTR_MASK, network_type=network_type, qp_type=bnxt.CMDQ_MODIFY_QP_QP_TYPE_RC, access=BNXT_ACCESS,
-      pkey=0xffff, dgid=dgid, sgid_index=self.dev.gid_id, hop_limit=64, dest_mac=dmac,
-      path_mtu_pingpong_push_enable=bnxt.CMDQ_MODIFY_QP_PATH_MTU_MTU_1024, max_dest_rd_atomic=4,
-      dest_qp_id=qpn)
-    self.qp_op(3, BNXT_RTS_MASK, network_type=network_type, qp_type=bnxt.CMDQ_MODIFY_QP_QP_TYPE_RC, access=BNXT_ACCESS,
-      max_rd_atomic=1)
-
-    if BNXT_DEBUG >= 1: print(f"bnxt: QP {self.qpn:#x} connected (remote={qpn:#x})")
-
-  def _poll(self, timeout):
-    def poll():
-      base = bnxt.struct_cq_base.from_buffer_copy(bytes(_qread(self.cqq, self.cqq["cons"])))
-      return bool(base.cqe_type_toggle & bnxt.CQ_BASE_TOGGLE) == (not bool((self.cqq["cons"] // 16) & 1))
-    wait_cond(poll, timeout_ms=timeout, msg="BNXT CQ")
-    raw = bytes(_qread(self.cqq, self.cqq["cons"]))
-    self.cqq["cons"] += 1
-    self.dev.doorbell(self.cq_id, bnxt.DBC_DBC_TYPE_CQ, self.cqq["cons"] & 15, (self.cqq["cons"] // 16) & 1)
-    return raw
-
-  def rdma_write(self, rva, rkey, lva, lkey, size, timeout_ms=20000):
-    start = self.sq["prod"] & 15
-    hdr = bytes(bnxt.struct_sq_rdma_hdr(wqe_type=bnxt.SQ_RDMA_HDR_WQE_TYPE_WRITE_WQE,
-                flags=bnxt.SQ_SEND_FLAGS_SIGNAL_COMP, wqe_size=3, length=size, remote_va=rva, remote_key=rkey))
-    for i, data in enumerate((hdr[:16], hdr[16:32], bytes(bnxt.struct_sq_sge(va_or_pa=lva, l_key=lkey, size=size)))):
-      _qwrite(self.sq, start + i, data)
-    nxt = (self.sq_psn + max(1, ceildiv(size, 1024))) & 0xffffff
-    value = start << bnxt.SQ_MSN_SEARCH_START_IDX_SFT | nxt << bnxt.SQ_MSN_SEARCH_NEXT_PSN_SFT | self.sq_psn
-    _qwrite(self.sq, self.msn, struct.pack("<Q", value), aux=True)
-
-    self.msn, self.sq_psn, self.sq["prod"] = (self.msn + 1) % 128, nxt, self.sq["prod"] + 3
-    self.dev.doorbell(self.qpn, bnxt.DBC_DBC_TYPE_SQ, self.sq["prod"] & 15, (self.sq["prod"] // 16) & 1)
-    cqe = bnxt.struct_cq_req.from_buffer_copy(self._poll(timeout_ms))
-    assert cqe.status == 0
@@ -1,118 +0,0 @@
-#!/usr/bin/env python3
-"""Send and validate one RDMA WRITE between two Broadcom BNXT hosts.
-
-This follows ``extra/mlx_driver/connect.py``: sync the driver, start the remote
-endpoint over SSH, exchange QP/GID/MAC/MR metadata, move both RC QPs to RTS,
-write bytes into the remote MR, and verify the bytes on the remote host.
-
-Both PCI functions must be unbound from bnxt_en/bnxt_re first.
-"""
-import json
-import os
-import subprocess
-import sys
-from typing import Any, IO
-
-TINYGRAD = os.path.normpath(os.path.join(os.path.dirname(os.path.abspath(__file__)), "../.."))
-sys.path.insert(0, TINYGRAD)
-
-from extra.bnxt_driver.bnxtdev import BNXTDev, BNXTQP
-from tinygrad.runtime.support.system import PCIDevice
-
-REMOTE_HOST = os.getenv("REMOTE_HOST", "192.168.52.213")
-REMOTE_USER = os.getenv("REMOTE_USER", "nimlgen")
-LOCAL_PCI = os.getenv("BNXT_PCI", "0000:41:00.0")
-REMOTE_PCI = os.getenv("REMOTE_PCI", "0000:41:00.0")
-LOCAL_IP = os.getenv("LOCAL_IP", "10.0.200.5")
-REMOTE_IP = os.getenv("REMOTE_IP", "10.0.200.6")
-MESSAGE = os.getenv("RDMA_MESSAGE", "Test message, rdma works!").encode()
-REMOTE = f"{REMOTE_USER}@{REMOTE_HOST}"
-SSH = ["ssh", "-o", "BatchMode=yes", "-o", "ConnectTimeout=10", "-o", "StrictHostKeyChecking=accept-new", REMOTE]
-SYNC_FILES = ("tinygrad/runtime/autogen/bnxt.py", "tinygrad/runtime/support/system.py",
-              "extra/bnxt_driver/bnxtdev.py", "extra/bnxt_driver/connect.py")
-
-def read_json(stream:IO[str], what:str) -> dict[str, Any]:
-  for line in iter(stream.readline, ""):
-    print(f"  [remote] {line}", end="")
-    try: value = json.loads(line)
-    except json.JSONDecodeError: continue
-    if isinstance(value, dict): return value
-  raise RuntimeError(f"remote exited before publishing {what}")
-
-def wait_line(stream:IO[str], text:str) -> str:
-  for line in iter(stream.readline, ""):
-    print(f"  [remote] {line}", end="")
-    if text in line: return line
-  raise RuntimeError(f"remote exited before reporting {text!r}")
-
-def send_line(stream:IO[str], value:str|dict[str, Any]):
-  stream.write((json.dumps(value) if isinstance(value, dict) else value) + "\n")
-  stream.flush()
-
-def qp_info(dev:BNXTDev, qp:BNXTQP) -> dict[str, Any]:
-  return {"qpn":qp.qpn, "mac":dev.mac.to_bytes(6, "big").hex(), "gid":dev.local_gid.hex()}
-
-def server():
-  dev = BNXTDev(PCIDevice("bnxt", os.getenv("BNXT_PCI", "0000:41:00.0")), ip=os.getenv("BNXT_IP", REMOTE_IP))
-  qp = BNXTQP(dev)
-  print(json.dumps(qp_info(dev, qp)), flush=True)
-
-  peer = json.loads(sys.stdin.readline())
-  qp.connect(peer["qpn"], bytes.fromhex(peer["gid"]), int(peer["mac"], 16))
-  print("connected", flush=True)
-
-  target, target_paddrs = dev.pci_dev.alloc_sysmem(0x1000)
-  target[:0x1000] = bytes(0x1000)
-  rkey = dev.register_mem(target_paddrs, 0x1000)
-  print(json.dumps({"target_addr":target_paddrs[0], "rkey":rkey}), flush=True)
-
-  assert sys.stdin.readline().strip() == "done"
-  received = bytes(target).rstrip(b"\0")
-  print(f"AS TEXT: {received.decode(errors='replace')!r}", flush=True)
-  print(json.dumps({"data":received.hex()}), flush=True)
-
-def sync_remote():
-  if os.getenv("SYNC", "1") == "0": return
-  print("syncing BNXT driver to remote")
-  subprocess.run(["rsync", "-azR", *SYNC_FILES, f"{REMOTE}:~/tinygrad/"], cwd=TINYGRAD, check=True)
-
-def start_remote() -> subprocess.Popen[str]:
-  print("booting remote")
-  command = (f"cd ~/tinygrad && sudo env PYTHONPATH=. PYTHONUNBUFFERED=1 BNXT_DEBUG={os.getenv('BNXT_DEBUG', '0')} "
-             f"BNXT_PCI={REMOTE_PCI} BNXT_IP={REMOTE_IP} python3 extra/bnxt_driver/connect.py --server")
-  return subprocess.Popen(SSH + [command], stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=sys.stderr, text=True)
-
-def client():
-  assert 0 < len(MESSAGE) <= 0x1000
-  sync_remote()
-  remote = start_remote()
-  assert remote.stdin is not None and remote.stdout is not None
-  remote_info = read_json(remote.stdout, "QP information")
-  print("booting local")
-  dev = BNXTDev(PCIDevice("bnxt", LOCAL_PCI), ip=LOCAL_IP)
-  qp = BNXTQP(dev)
-
-  send_line(remote.stdin, qp_info(dev, qp))
-  wait_line(remote.stdout, "connected")
-  qp.connect(remote_info["qpn"], bytes.fromhex(remote_info["gid"]), int(remote_info["mac"], 16))
-  print("both QPs in RTS")
-
-  remote_target = read_json(remote.stdout, "MR information")
-  source, source_paddrs = dev.pci_dev.alloc_sysmem(0x1000)
-  source[:len(MESSAGE)] = MESSAGE
-  lkey = dev.register_mem(source_paddrs, 0x1000)
-  print(f"RDMA WRITE {len(MESSAGE)}B to remote phys 0x{remote_target['target_addr']:x}")
-  qp.rdma_write(remote_target["target_addr"], remote_target["rkey"], source_paddrs[0], lkey, len(MESSAGE))
-
-  send_line(remote.stdin, "done")
-  wait_line(remote.stdout, "AS TEXT")
-  result = read_json(remote.stdout, "RDMA result")
-  assert bytes.fromhex(result["data"]) == MESSAGE
-  print("RDMA WRITE data verified")
-
-  remote.stdin.close()
-  assert remote.wait() == 0
-  print("RDMA WRITE test complete")
-
-if __name__ == "__main__":
-  server() if "--server" in sys.argv else client()
@@ -1,46 +0,0 @@
-#!/usr/bin/env python3
-"""Local BNXT RoCEv2 RDMA WRITE loopback using the firmware's PHY loopback mode.
-
-The kernel bnxt_en/bnxt_re modules must be unloaded first.
-
-  sudo PYTHONPATH=. BNXT_PCI=0000:41:00.0 BNXT_IP=10.0.200.5 python3 extra/bnxt_driver/loopback.py
-"""
-import os
-import sys
-import time
-
-sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "../.."))
-
-from extra.bnxt_driver.bnxtdev import BNXTDev, BNXTQP
-from tinygrad.runtime.autogen import bnxt
-from tinygrad.runtime.support.system import PCIDevice
-
-BUF_SIZE = 0x1000
-BNXT_PCI = os.getenv("BNXT_PCI", "0000:41:00.0")
-BNXT_IP = os.getenv("BNXT_IP", "10.0.200.5")
-
-if __name__ == "__main__":
-  print(f"[init] BNXT at {BNXT_PCI}")
-  dev = BNXTDev(PCIDevice("bnxt", BNXT_PCI), ip=BNXT_IP)
-  tx_qp, rx_qp = BNXTQP(dev), BNXTQP(dev)
-  print(f"[init] loopback-connect TX QP 0x{tx_qp.qpn:x} <-> RX QP 0x{rx_qp.qpn:x}")
-  tx_qp.connect(rx_qp.qpn, dev.local_gid, dev.mac)
-  rx_qp.connect(tx_qp.qpn, dev.local_gid, dev.mac)
-
-  src, src_paddrs = dev.pci_dev.alloc_sysmem(BUF_SIZE)
-  dst, dst_paddrs = dev.pci_dev.alloc_sysmem(BUF_SIZE)
-  message = b"Hello from BNXT RoCE PHY loopback!"
-  src[:BUF_SIZE], dst[:BUF_SIZE] = bytes(BUF_SIZE), bytes(BUF_SIZE)
-  src[:len(message)] = message
-  lkey = dev.register_mem(src_paddrs, BUF_SIZE)
-  rkey = dev.register_mem(dst_paddrs, BUF_SIZE)
-
-  print("[loopback] enabling local PHY loopback")
-  dev.hwrm("port_phy_cfg", port_id=dev.port_id, enables=bnxt.PORT_PHY_CFG_REQ_ENABLES_LPBK, lpbk=bnxt.PORT_PHY_CFG_REQ_LPBK_LOCAL)
-  time.sleep(1)
-  tx_qp.rdma_write(dst_paddrs[0], rkey, src_paddrs[0], lkey, len(message))
-  got = bytes(dst[:len(message)])
-  print(f"[result] {got!r}")
-  assert got == message
-  print("BNXT RoCE PHY loopback RDMA WRITE passed")
-  dev.hwrm("port_phy_cfg", port_id=dev.port_id, enables=bnxt.PORT_PHY_CFG_REQ_ENABLES_LPBK, lpbk=bnxt.PORT_PHY_CFG_REQ_LPBK_NONE)
@@ -35,7 +35,7 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li
    return name

  for call in iter_kernel_calls(linear):
-    arg_uops = [b for b in call.src[1:] if not b.is_bound_var]
+    arg_uops = [b for b in call.src[1:] if b.op is not Ops.BIND]
    prg = to_program(call.src[0], Device[arg_uops[0].device].renderer)
    info = prg.arg
    functions[info.function_name] = prg.src[2].arg
@@ -122,8 +122,7 @@ def custom_mxfp4_gemm(C:UOp, A:UOp, B:UOp, scale_a:UOp, scale_b:UOp, *extra:UOp,
  groups_x, groups_y = UOp.special(ceildiv(N, tile_n), "gidx0"), UOp.special(ceildiv(M, tile_m), "gidx1")
  lds = UOp.placeholder((163840,), dtypes.uint8, 0, AddrSpace.LOCAL)
  sink = UOp.sink(C.base, A.base, B.base, scale_a.base, scale_b.base, *(x.base for x in extra), lds, threads, groups_x, groups_y,
-                  arg=KernelInfo(f"mxfp4_gemm_{M}_{N}_{K}",
-                                 estimates=Estimates(ops=2*M*N*K, mem=(M*half_k+N*half_k)*A.dtype.itemsize+M*N*C.dtype.itemsize)))
+                  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))))

@@ -1,32 +1,10 @@
 import functools, pathlib
 from tinygrad import Tensor, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
-from tinygrad.helpers import getenv
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.renderer import Estimates
 from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from extra.gemm.cdna_asm_gemm import quantize_mxfp8, _mx_block_scale, _mx_block_scale_3d

-ZERO_OPTIM = getenv("ZERO_OPTIM", 0)
-
-def reduce_scatter_devaxis(out:Tensor, shard_axis:int=0) -> Tensor:
-  # out: sharded on the device axis, shape (ndev, *rest); return the device-axis sum left sharded on shard_axis.
-  u = out.uop
-  devs, rest = u.device, u.shape[1:]
-  assert rest[shard_axis] % len(devs) == 0, f"reduce_scatter needs even shards: {rest[shard_axis]} % {len(devs)}"
-  # reach the raw per-device buffer below the UNSHARD, keeping the AFTERs so reads stay ordered after the kernel writes
-  node, barriers = u, []
-  while node.op is not Ops.UNSHARD:
-    if node.op is Ops.AFTER: barriers += node.src[1:]
-    node = node.src[0]
-  mbuf = node.src[0].after(*barriers) if barriers else node.src[0]
-  sz = rest[shard_axis] // len(devs)
-  shards = []
-  for i in range(len(devs)):
-    bounds = tuple((0,s) if a != shard_axis else (i*sz,(i+1)*sz) for a,s in enumerate(rest))
-    contribs = [mbuf.mselect(j).reshape(rest).shrink(bounds).copy_to_device(devs[i]) for j in range(len(devs))]
-    shards.append(functools.reduce(lambda a,b: a.alu(Ops.ADD, b), contribs))
-  return Tensor(UOp.mstack(*shards).unshard(shard_axis, UOp.range(len(devs), -1, AxisType.DEVICE)), device=devs)
-
@functools.cache
 def custom_hk_grouped_mxfp8_gemm(C:UOp, A:UOp, B:UOp, scale_A:UOp, scale_B:UOp, *extra:UOp, dname:str, n_experts:int) -> UOp:
  M, K = A.shape
@@ -80,8 +58,7 @@ def grouped_mx_wgrad(g:Tensor, xg:Tensor, expert_off:Tensor, n_experts:int) -> T
  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]
-  if is_multi and ZERO_OPTIM: out = reduce_scatter_devaxis(out, 0)
-  else: out = out.sum(0) if is_multi else out.squeeze(0)
+  out = out.sum(0) if is_multi else out.squeeze(0)
  return out.reshape(n_experts, N, K)

 def mx_pack_3d(e8:Tensor) -> Tensor:
@@ -53,7 +53,7 @@ def _ggather_bwd(gradient:UOp, kernel:UOp) -> tuple:
    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, src=(gtab.index(g, row, j), val), arg=(atomic_str, dtypes.void))
+    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)
@@ -79,7 +79,7 @@ if __name__ == "__main__":
  linear, var_vals = C.linear_with_vars()
  last_call = linear.src[-1]
  ast = last_call.src[0]
-  bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]
+  bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]

  src = compiled.asm["ptx"]
  # specify the shared memory here so we don't need to do it dynamically
@@ -16,12 +16,9 @@ def _do_reset_device(pci_bus): os.system(f"sudo sh -c 'echo 1 > /sys/bus/pci/dev
 def _is_module_loaded(name: str) -> bool: return os.path.isdir(f"/sys/module/{name}")

 def cmd_remove_module(args):
-  modules = ["nvidia_drm", "nvidia_modeset", "nvidia_uvm", "nvidia"] if args.backend == "nv" else ["amdgpu"]
+  modules = ["nvidia_drm", "nvidia_modeset", "nvidia_uvm", "nvidia", "ast"] if args.backend == "nv" else ["amdgpu"]
  to_unload = [m for m in modules if _is_module_loaded(m)]
  if not to_unload: print("Kernel modules are not loaded")
-  elif getattr(args, "expect", False):
-    print(f"Kernel modules are loaded: {to_unload}")
-    sys.exit(1)
  else:
    print("Removing kernel modules:", ", ".join(to_unload))
    try: subprocess.run(["sudo", "modprobe", "-r", *to_unload], check=True)
@@ -63,19 +60,17 @@ def cmd_show_pids(args):

 def cmd_kill_pids(args):
  devs = scan_devs_based_on_lock(prefix:={"amd":"am", "nv":"nv"}[args.backend], args)
-  use_sudo = not getattr(args, "sudoless", False)

  for dev in devs:
    for i in range(128):
      if i > 0: time.sleep(0.2)

      try:
-        try: pid = subprocess.check_output((['sudo'] if use_sudo else []) +
-                                           ['lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
+        try: pid = subprocess.check_output(['sudo', 'lsof', temp(f'{prefix}_{dev}.lock')]).decode('utf-8').strip().split('\n')[1].split()[1]
        except subprocess.CalledProcessError: break

        print(f"Killing process {pid} (which uses {dev})")
-        subprocess.run((['sudo'] if use_sudo else []) + ['kill', '-9', pid], check=True)
+        subprocess.run(['sudo', 'kill', '-9', pid], check=True)
      except subprocess.CalledProcessError as e:
        print(f"Failed to kill process for device {dev}: {e}", file=sys.stderr)

@@ -84,7 +79,6 @@ def add_common_commands(parent_subparsers):
  p_insmod.set_defaults(func=cmd_insert_module)

  p_rmmod = parent_subparsers.add_parser("rmmod", help="Remove a kernel module")
-  p_rmmod.add_argument("--expect", action="store_true", help="Just assert that module is already unloaded")
  p_rmmod.set_defaults(func=cmd_remove_module)

  p_reset = parent_subparsers.add_parser("reset", help="Reset a device")
@@ -97,20 +91,17 @@ def add_common_commands(parent_subparsers):

  p_reset = parent_subparsers.add_parser("kill_pids", help="Kill pids of processes using the device")
  p_reset.add_argument("--pci_bus", default="", help="PCI bus ID of the device")
-  p_reset.add_argument("--sudoless", action="store_true", help="Do not use sudo when detecting or killing pids")
  p_reset.set_defaults(func=cmd_kill_pids)

 if __name__ == "__main__":
  parser = argparse.ArgumentParser()
  backend_subparsers = parser.add_subparsers(dest="backend", required=True, metavar="{nv,amd}", help="Hardware backend to target")

-  nv_parser = backend_subparsers.add_parser("nv", aliases=["NV"], help="NVIDIA GPUs")
-  nv_parser.set_defaults(backend="nv")
+  nv_parser = backend_subparsers.add_parser("nv", help="NVIDIA GPUs")
  nv_commands = nv_parser.add_subparsers(dest="command", required=True)
  add_common_commands(nv_commands)

-  amd_parser = backend_subparsers.add_parser("amd", aliases=["AMD"], help="AMD GPUs")
-  amd_parser.set_defaults(backend="amd")
+  amd_parser = backend_subparsers.add_parser("amd", help="AMD GPUs")
  amd_commands = amd_parser.add_subparsers(dest="command", required=True)
  add_common_commands(amd_commands)

@@ -3,7 +3,7 @@ from typing import cast, Any, Callable
 import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit
 assert sys.platform != 'win32'
 from dataclasses import dataclass
-from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, encode_kernargs_clike, make_cmdbuf
+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
@@ -19,7 +19,7 @@ from tinygrad.runtime.support.hcq import FileIOInterface, HCQBuffer, MMIOInterfa
 from tinygrad.runtime.support.am.amdev import AMDev, AMMemoryManager
 from tinygrad.runtime.support.amd import AMDReg, AMDIP, import_module, import_soc, import_pmc
 from tinygrad.runtime.support.system import PCIIfaceBase, PCIAllocationMeta, USBPCIDevice, MAP_FIXED, MAP_NORESERVE
-from tinygrad.runtime.support.usb import USB3, usb_ib, usb_push, usb_arm_bytes, pm_usb_stage, pm_usb_hostio, pm_usb_bufferize
+from tinygrad.runtime.support.usb import USB3
 from tinygrad.runtime.support.memory import AddrSpace, BumpAllocator
 from tinygrad.runtime.ops_amd import SQTT, SQTT_ITRACE_SE_MASK, SQTT_LIMIT_SE, SQTT_SIMD_SEL, SQTT_TOKEN_EXCLUDE, PMC
 from tinygrad.runtime.ops_amd import EVENT_INDEX_PARTIAL_FLUSH, WAIT_REG_MEM_FUNCTION_EQ, WAIT_REG_MEM_FUNCTION_NEQ, WAIT_REG_MEM_FUNCTION_GEQ
@@ -87,7 +87,7 @@ def release_mem(ctx, address=0x0, value=0, data_sel=0, int_sel=2, ctxid=0, cache

 def memory_barrier(ctx):
  pf = '' if ctx.nbio.version[0] == 2 else '0' if ctx.nbio.version[:2] != (7, 11) else '1'
-  return UOp(Ops.LINEAR, src=(
+  return UOp(Ops.LINEAR, dtypes.void, (
    wait_reg_mem(ctx, reg=getattr(ctx.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_REQ').addr[0],
                 reg_done=getattr(ctx.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_DONE').addr[0], value=0xffffffff),
    acquire_mem(ctx)))
@@ -135,7 +135,7 @@ def pm4_program(ctx, call, prg):
    wreg(ctx, ctx.gc.regCOMPUTE_START_X, 0, 0, 0, *(info.local_size or (1, 1, 1)), 0, 0),
    pkt3(ctx, PM4Ops.DISPATCH_DIRECT, *info.global_size, dispatch_init),
    pkt3(ctx, PM4Ops.EVENT_WRITE, ctx.pm4.EVENT_TYPE(ctx.soc.CS_PARTIAL_FLUSH) | ctx.pm4.EVENT_INDEX(EVENT_INDEX_PARTIAL_FLUSH))]
-  return UOp(Ops.LINEAR, src=tuple(ins))
+  return UOp(Ops.LINEAR, dtypes.void, tuple(ins))

 pm_pm4_opsel = PatternMatcher([
  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="prg"),), name="call", allow_any_len=True), pm4_program),
@@ -146,14 +146,11 @@ pm_pm4_opsel = PatternMatcher([
  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), pm4_store),
 ])

-def queue_ptrs(devs, qname:str, q:AMDQueueDesc) -> tuple[UOp, ...]:
-  return tuple(UOp.placeholder((b.size,), b.dtype, 0, device=devs).rtag(f"{qname}_{n}")
-               for n, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))
-
 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 = queue_ptrs(devs, "COMPUTE:0", q)
+  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)))

  # 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)
@@ -182,10 +179,11 @@ 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
-  hdr = ctx.sdma.SDMA_OP_COPY | ctx.sdma.SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_COPY_LINEAR)
-  return call.ins(SDMAOps.COPY, src=tuple(x for off in range(0, sz, ctx.max_copy_size) for x in (
-    *(UOp.const(v, dtypes.uint32) for v in (hdr, ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz-off, ctx.max_copy_size)-1), 0)),
-    *(a + UOp.const(off, dtypes.uint64) if off else a for a in (call.src[2].getaddr(ctx.devs), call.src[1].getaddr(ctx.devs))))))
+  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(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+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) \
@@ -207,7 +205,7 @@ def sdma_timestamp(ctx, ins, dst):
 pm_sdma_opsel = PatternMatcher([
  (UPat(Ops.CALL, src=(UPat(Ops.COPY),), name="call", allow_any_len=True), sdma_copy),

-  (UPat(Ops.INS, arg="barrier"), lambda: UOp(Ops.NOOP)),
+  (UPat(Ops.INS, arg="barrier"), lambda: UOp(Ops.NOOP, dtypes.void, ())),
  (UPat(Ops.INS, arg="wait", src=(UPat(name="dst"), UPat(name="val")), name="ins"), sdma_wait),
  (UPat(Ops.INS, arg="timestamp", src=(UPat(name="dst"),), name="ins"), sdma_timestamp),
  (UPat(Ops.INS, arg="store", src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val")), name="ins"), sdma_store),
@@ -219,7 +217,8 @@ def sdma_submit(cmdbuf, devs):

  # the sdma queue's ring and its host-side ring/write/put pointers
  for d in devs: q = Device[d].sdma_queue(0)
-  ring, wptr, doorbell, put_ptr = queue_ptrs(devs, "COPY:0", q)
+  ring, wptr, doorbell, put_ptr = (UOp.placeholder((b.size,), b.dtype, 0, device=devs).rtag(f"COPY:0_{name}")
+    for name, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))

  # sdma needs the cmdbuf contiguous: if it won't fit before the ring end, restart at 0 and zero the tail
  put_b = put_ptr.index(zero)
@@ -246,32 +245,15 @@ def sdma_submit(cmdbuf, devs):
 pm_sdma_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"),
  lambda ctx, lin: sdma_submit(make_cmdbuf(lin, ctx.devs), ctx.devs))])

-# *****************
-# USB submit
-
-def amd_usb_submit(ctx, lin):
-  for d in ctx.devs: q = Device[d].compute_queue if (comp:=ctx.qname.startswith("COMPUTE")) else Device[d].sdma_queue(0)
-
-  if nb:=usb_arm_bytes(ctx.pre, Device[ctx.devs[0]].iface.usb_sram):
-    poke = (ctx.sdma.SDMA_OP_WRITE, *data64_le(Device[ctx.devs[0]].iface.cq_buf.va_addr + 12), 0, 0)
-    lin = lin.replace(src=lin.src + (UOp(Ops.INS, arg="poke", src=tuple(UOp.const(x, dtypes.uint32) for x in poke)),))
-
-  ib_host, ib_gpu, pkt_dw = usb_ib(ctx.devs, lin, 32 if comp else 0x100, nb)
-  pkt = (ctx.pm4.PACKET3(ctx.pm4.PACKET3_INDIRECT_BUFFER,2),*data64_le(ib_gpu.getaddr(ctx.devs)),pkt_dw|ctx.pm4.INDIRECT_BUFFER_VALID) if comp else ()
-  return usb_push(ctx.devs, *queue_ptrs(ctx.devs, ctx.qname, q), ib_host, ib_gpu, pkt, 4 if comp else 1)
-
-pm_usb_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"), amd_usb_submit)])
-
@dataclass(frozen=True)
 class AMDEncodeCtx:  # encode-time constants for one queue: devs (every cmdbuf address resolves into these) + gfx version + packet/ip modules
  devs: tuple[str, ...]; target: tuple[int, ...]; pm4: Any; sdma: Any; soc: Any  # noqa: E702
-  gc: AMDIP; nbio: AMDIP; xccs: int; max_copy_size: int; tmpring_size: Callable; qname: str; pre: UOp # pre: the queue before opsel
+  gc: AMDIP; nbio: AMDIP; xccs: int; max_copy_size: int; tmpring_size: Callable  # noqa: E702

 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, q.arg[1], q)
-  opsel = pm_pm4_opsel if (comp:=q.arg[1].startswith("COMPUTE")) else pm_sdma_opsel
-  submit = d.pm_submit if d.pm_submit is not None else (pm_pm4_submit if comp else pm_sdma_submit)
+  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"), ctx)

@dataclass(frozen=True)
@@ -280,11 +262,10 @@ class AMDProgramData:
  private_segment_size:int; kernargs_segment_size:int; kernargs_alloc_size:int
  enable_dispatch_ptr:int; enable_private_segment_sgpr:int

-_amd_program_cache:dict[tuple[bytes, tuple[str, ...]], UOp] = {}
+_amd_program_cache:dict[tuple[bytes,str], tuple[AMDProgramData,bytes]] = {}
 def amd_build_program(prg:UOp) -> UOp:
  dev = Device[to_tuple(prg.device)[0]] # TODO: rm this
-  # key on the full device tuple: the same lib can be built for different device sets, each needs its own program buffer
-  if (cached:=_amd_program_cache.get(key:=(lib:=prg.src[3].arg, to_tuple(prg.device)))) is None:
+  if (cached:=_amd_program_cache.get(key:=(lib:=prg.src[3].arg, dev.device))) is None:
    image, sections, relocs = elf_loader(lib)
    rodata = next(sh.header.sh_addr for sh in sections if sh.name == ".rodata")
    for off, sym, typ, addent in relocs:
@@ -301,23 +282,20 @@ def amd_build_program(prg:UOp) -> UOp:
      wave32=bool(desc.kernel_code_properties & 0x400), private_segment_size=desc.private_segment_fixed_size, kernargs_segment_size=desc.kernarg_size,
      kernargs_alloc_size=desc.kernarg_size + (ctypes.sizeof(hsa.hsa_kernel_dispatch_packet_t) if edp else 0), enable_dispatch_ptr=edp,
      enable_private_segment_sgpr=desc.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER)
-    image = bytes(image).ljust(round_up(len(image), 4), b"\x00") # the program is uploaded as whole dwords
    buf = UOp.placeholder((len(image),), dtypes.uint8, next(UOp.unique_num), device=prg.device).rtag("program")
-    cached = _amd_program_cache[key] = prg.replace(src=(buf.after(make_binary_patch(buf, image)),), arg=(data, prg.arg))
+    cached = _amd_program_cache[key] = prg.replace(src=(buf.after(make_binary_patch(buf, bytes(image))),), arg=(data, prg.arg))
  return cached

 class AMDAllocator(HCQAllocator['AMDDevice']):
  def __init__(self, dev:AMDDevice):
-    super().__init__(dev, supports_copy_from_disk=dev.has_copy_queue, supports_transfer=dev.has_copy_queue and not dev.is_usb)
+    super().__init__(dev, supports_copy_from_disk=dev.has_sdma_queue, supports_transfer=dev.has_sdma_queue and not dev.is_usb())

-  def _alloc(self, size:int, options:BufferSpec) -> HCQBuffer:
-    return self.dev.iface.alloc(size, host=options.host, uncached=options.uncached, cpu_access=options.cpu_access or not self.dev.has_copy_queue)
+  def _alloc(self, size:int, options:BufferSpec) -> HCQ2Buffer:
+    return self.dev.iface.alloc(size, host=options.host, uncached=options.uncached, cpu_access=options.cpu_access or not self.dev.has_sdma_queue)

  def _do_free(self, opaque, options:BufferSpec): self.dev.iface.free(opaque)

-  def _do_map(self, buf:HCQBuffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)
-
-  def _do_unmap(self, buf:HCQBuffer): self.dev.iface.unmap(buf)
+  def _do_map(self, buf:HCQ2Buffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)

@dataclass
 class AMDQueueDesc:
@@ -410,24 +388,15 @@ class KFDIface:
    return hcqbuf

  def free(self, mem):
-    self._unmap(mem)
-    if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
-    kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
-
-  def unmap(self, mem):
-    self._unmap(mem)
-    if getattr(mem, '_owns_kfd_handle', False): kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
-
-  def _unmap(self, mem):
    gpus = (ctypes.c_int32 * 1)(self.gpu_id)
    stm = kfd.AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(gpus), n_devices=1)
    assert stm.n_success == 1
+    if mem.owner == self.dev:
+      if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
+      kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)

  def map(self, mem):
-    if mem.owner is not None and mem.owner._is_cpu():
-      mapped = self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)
-      mapped._owns_kfd_handle = True
-      return mapped
+    if mem.owner is not None and mem.owner._is_cpu(): return self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)

    c_gpus = (ctypes.c_int32 * 1)(self.gpu_id)
    stm = kfd.AMDKFD_IOC_MAP_MEMORY_TO_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(c_gpus), n_devices=1)
@@ -499,7 +468,6 @@ class PCIIface(PCIIfaceBase):

  def require_profile_mode(self): return True
  def is_wgp_active(self, xcc, se, sa, wgp) -> bool: return True # TODO: account for WGP disablement on some asics.
-  def unmap(self, mem): self.free(mem)

  def _compute_props(self):
    self.ip_versions = self.dev_impl.ip_ver
@@ -544,7 +512,8 @@ class PCIIface(PCIIfaceBase):
      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.signal('timeline')._buf.cpu_view().view(fmt='Q')[0] = d.signal('value', 1, device="CPU")._buf.cpu_view().view(fmt='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))):
@@ -558,32 +527,6 @@ class PCIIface(PCIIfaceBase):

  def device_fini(self): self.dev_impl.fini()

-class USBIface(PCIIface):
-  def __init__(self, dev, dev_id): # pylint: disable=super-init-not-called
-    if dev_id >= len(visible:=hcq_filter_visible_devices(USB3.list_devices(0xADD1, 0x0001) + USB3.list_devices(0x3801, 0x0001), "AMD")):
-      raise RuntimeError(f"AMD:{dev_id} does not exist ({pluralize('device', len(visible))} available)")
-    self.dev, self.pci_dev, self.vram_bar, self.count = dev, USBPCIDevice("AM", *visible[dev_id]), 0, len(visible)
-    self.dev_impl = AMDev(self.pci_dev)
-    self._compute_props()
-    self.sram = self._dma_region(ctrl_addr=0xf000, sys_addr=0x200000, size=0x80000)
-    self.cq_buf = self._dma_region(ctrl_addr=0xb800, sys_addr=0x822000, size=0x1000) # +12 is the dword that releases an armed read
-    self.usb_handle = unwrap(ctypes.cast(self.pci_dev.usb.usb.handle, ctypes.c_void_p).value)
-
-  def _dma_region(self, ctrl_addr, sys_addr, size):
-    region = self.dev_impl.mm.map_range(vaddr:=self.dev_impl.mm.alloc_vaddr(size=size), size, [(sys_addr, size)], aspace=AddrSpace.SYS, uncached=True)
-    return HCQBuffer(vaddr, size, meta=PCIAllocationMeta(region, has_cpu_mapping=False), view=self.pci_dev.dma_view(ctrl_addr, size), owner=self.dev)
-
-  def alloc(self, size:int, host=False, uncached=False, cpu_access=False, contiguous=False, force_devmem=False, **kwargs) -> HCQBuffer:
-    # everything, even host-style signals, lives in vram: gpu writes into the bridge's own memory collide with an armed 0xF2 read stream
-    return super().alloc(size, host=False, uncached=uncached, cpu_access=cpu_access or host, contiguous=contiguous, force_devmem=True, **kwargs)
-
-  def sleep(self, timeout): pass
-
-  # we don't own the sram region, so the buffer never frees it
-  @functools.cached_property
-  def usb_sram(self) -> Buffer:
-    return Buffer(self.dev.device, (b:=self.sram).size, dtypes.uint8, options=BufferSpec(external_ptr=b.va_addr, nolru=True)).allocate(opaque=b)
-
 def _mock(iface, name=None): return type(name or f"MOCK{iface.__name__}", (iface,), {})

 class AMDDevice(HCQ2Compiled):
@@ -594,21 +537,21 @@ class AMDDevice(HCQ2Compiled):
    # encoding of cmdbuf
    (UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),)), encode_queue),
  ])
-  pm_submit: PatternMatcher|None = None

  timestamp_divider = 100.0  # AMD GPU clock: ticks/us
  max_scratch_psize = 0

-  ifaces = [KFDIface, PCIIface, USBIface, _mock(KFDIface, "MOCKIface"), _mock(KFDIface), _mock(PCIIface), _mock(USBIface)]
+  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

  def __init__(self, device:str=""):
-    self.iface = self._select_iface(device)
-    self.is_usb = isinstance(self.iface, USBIface)
-    if self.is_usb: self.rt_nbytes = 4 << 20
+    self.device_id = int(device.split(":")[1]) if ":" in device else 0
+
+    self.iface = self._select_iface()

    self.target:tuple[int, ...] = ((trgt:=self.iface.props['gfx_target_version']) // 10000, (trgt // 100) % 100, trgt % 100)
    self.arch = "gfx%d%x%x" % self.target
@@ -633,12 +576,12 @@ class AMDDevice(HCQ2Compiled):

    self.is_aql = getenv("AMD_AQL", int(self.xccs > 1))
    if self.is_aql:
-      self.pm4_ibs = self.iface.alloc(0x2000 if self.is_usb else (16 << 20), uncached=True, cpu_access=True)
+      self.pm4_ibs = self.iface.alloc(0x2000 if self.is_usb() else (16 << 20), uncached=True, cpu_access=True)
      self.pm4_ib_alloc = BumpAllocator(self.pm4_ibs.size, wrap=True)

    self.max_copy_size = 0x40000000 if self.iface.ip_versions[am.SDMA0_HWIP][0] >= 5 else 0x400000
    self.sdma_queues:dict = {}
-    self.has_copy_queue = not getenv("AMD_DISABLE_SDMA")
+    self.has_sdma_queue = True # self.sdma_queue(0) is not None, TODO: think of this

    super().__init__(device, AMDAllocator(self), [HIPRenderer, AMDLLVMRenderer, HIPCCRenderer], None, can_recover=self.is_am(), arch=self.arch)

@@ -646,10 +589,6 @@ class AMDDevice(HCQ2Compiled):
    self.max_private_segment_size = 0
    self.pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, tag="scratch", name="b"), lambda ctx, b: ctx[0].scratch_buffer(b.max_numel()))]) + self.pm_bufferize

-    if self.is_usb:
-      self.pm_bufferize = pm_usb_bufferize + self.pm_bufferize
-      self.pm_stage_copy, self.pm_host_lower, self.pm_submit = pm_usb_stage, pm_usb_hostio, pm_usb_submit
-
    self.pmc_enabled:bool = PROFILE > 0 and PMC > 0
    if self.pmc_enabled:
      self.iface.require_profile_mode()
@@ -710,7 +649,7 @@ class AMDDevice(HCQ2Compiled):
    wg_data_size = round_up((vgpr_size_per_cu + sgrp_size_per_cu + lds_size_per_cu + hwreg_size_per_cu) * self.cu_cnt, mmap.PAGESIZE)
    ctl_stack_size = round_up((12 if self.target[0] != 9 else 8) * self.wave_cnt + 8 + 40, mmap.PAGESIZE)
    return self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL if self.is_aql else kfd.KFD_IOC_QUEUE_TYPE_COMPUTE,
-      0x2000 if self.is_usb else (16 << 20), eop_buffer_size=0x1000,
+      0x2000 if self.is_usb() else (16 << 20), eop_buffer_size=0x1000,
      ctx_save_restore_size=0 if self.is_am() else wg_data_size + ctl_stack_size, ctl_stack_size=ctl_stack_size,
      debug_memory_size=round_up(self.wave_cnt * 32, 64))

@@ -718,7 +657,7 @@ class AMDDevice(HCQ2Compiled):
    if getenv("AMD_DISABLE_SDMA"): return None
    if idx in self.sdma_queues: return self.sdma_queues[idx]
    with contextlib.suppress(OSError):
-      self.sdma_queues[idx] = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_SDMA, 0x2000 if self.is_usb else (16 << 20), idx=idx)
+      self.sdma_queues[idx] = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_SDMA, 0x200 if self.is_usb() else (16 << 20), idx=idx)
    return self.sdma_queues.get(idx, None)

  def tmpring_size(self, private_segment_size):
@@ -50,7 +50,7 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_out:UOp, x:UOp, amax_state:
  else: raise NotImplementedError(f"no atomic max for device {device}")
  amax_idx = amax_out.reshape((1,)).index(UOp.const(0))
  max_val = lds[0].load()
-  atomic = UOp(Ops.CUSTOM, src=(amax_idx, max_val.bitcast(dtypes.int32), max_val, amax_idx.load()), arg=(atomic_arg, dtypes.void))
+  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=()))

@functools.cache
@@ -12,7 +12,7 @@ def _custom_quantize_mxfp4(row_fp4:UOp, row_scale:UOp, col_fp4:UOp, col_scale:UO
  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, src=(o.base.index(0),), arg=("", dtypes.void)) for o in outputs),
+                  *(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()
@@ -5,9 +5,9 @@ from tinygrad.helpers import getenv, DEBUG

 # https://github.com/facebookresearch/llama/blob/1076b9c51c77ad06e9d7ba8a4c6df775741732bd/llama/model.py#L47
 def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0) -> Tensor:
-  freqs = 1.0 / (theta ** (Tensor.arange(0, dim, 2, dtype=dtypes.float32)[:(dim // 2)] / dim))
-  freqs = Tensor.arange(end, dtype=dtypes.float32).unsqueeze(dim=1) * freqs.unsqueeze(dim=0)
-  return Tensor.stack(freqs.cos(), freqs.sin(), dim=-1).cast(dtypes.default_float).reshape(1, end, 1, dim//2, 2)
+  freqs = 1.0 / (theta ** (Tensor.arange(0, dim, 2)[:(dim // 2)] / dim))
+  freqs = Tensor.arange(end).unsqueeze(dim=1) * freqs.unsqueeze(dim=0)
+  return Tensor.stack(freqs.cos(), freqs.sin(), dim=-1).reshape(1, end, 1, dim//2, 2)

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

@@ -23,8 +23,5 @@ if __name__ == "__main__":
      # AM_RESET=1 gets a clear trace, does not work on mi300 machines
      subprocess.run([sys.executable, *shlex.split(test)], cwd=EXAMPLES_DIR.parent.parent.parent,
                     env={**os.environ, "DEV":"AMD", "AM_RESET":"1" if not arch.startswith("gfx9") else "0", "VIZ":"-2", "PYTHONPATH":"."})
-      with open(PROFILE_PATH, "rb") as f: events = pickle.load(f)
-      with open(PROFILE_PATH, "wb") as f:
-        pickle.dump([e for e in events if type(e).__name__ in {"ProfilePMCEvent", "ProfileSQTTEvent", "ProfileProgramEvent"}], f)
      PROFILE_PATH.rename(dest:=EXAMPLES_DIR/arch/f"profile_{name}_run_{i}.pkl")
      print(f"saved SQTT trace to {dest}")
@@ -192,7 +192,7 @@ def unpack_insts(viz_data, i:int, j:int, data:dict) -> dict:
    prev_instr = max(prev_instr, e.time + e.dur)
  summary = [{"label":"Total Cycles", "value":w.end_time-w.begin_time}, {"label":"SE", "value":w.se}, {"label":"CU", "value":w.cu},
             {"label":"SIMD", "value":w.simd}, {"label":"Wave ID", "value":w.wave_id}, {"label":"Run number", "value":data["run_number"]}]
-  return {"rows":[tuple(v.values()) for v in rows.values()], "cols":columns, "metadata":[summary],"ref":viz_data.ref_map.get(data["prg"].profile_key)}
+  return {"rows":[tuple(v.values()) for v in rows.values()], "cols":columns, "metadata":[summary], "ref":viz_data.ref_map.get(data["prg"].name)}

 def print_data(data:dict) -> None:
  from tabulate import tabulate
@@ -19,33 +19,16 @@ def _sharded_empty(shape:Tensor, ref:Tensor, axis:int|None, dtype:DTypeLike|None
@functools.cache
 def custom_fused_qkv_rope_forward(q:UOp, k:UOp, v:UOp, xqkv:UOp, freqs_cis:UOp,
                                  device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
-  group_size = H // H_KV
-  q, k, v = q.reshape(B, N, H, D), k.reshape(B, N, H_KV, D), v.reshape(B, N, H_KV, D)
-  xqkv = xqkv.reshape(B, N, H_KV, group_size + 2, D)
-  b, n = UOp.range(B, 0), UOp.range(N, 1)
-  pair = UOp.range(D // 2, 2)
-  even = pair * 2
-  c = freqs_cis[0, n, 0, pair, 0].cast(dtypes.float)
-  s = freqs_cis[0, n, 0, pair, 1].cast(dtypes.float)
-  ordered:UOp|None = None
-  for kvh in range(H_KV):
-    q_out, k_out, v_out = (x.after(ordered) if ordered is not None else x for x in (q, k, v))
-    x_in = xqkv.after(ordered) if ordered is not None else xqkv
-    stores:list[UOp] = []
-    for rep in range(group_size):
-      a = x_in[b, n, kvh, rep, even].cast(dtypes.float)
-      bb = x_in[b, n, kvh, rep, even + 1].cast(dtypes.float)
-      h = kvh * group_size + rep
-      stores += [q_out[b, n, h, even].store((a * c - bb * s).cast(q.dtype)), q_out[b, n, h, even + 1].store((a * s + bb * c).cast(q.dtype))]
-    a = x_in[b, n, kvh, group_size, even].cast(dtypes.float)
-    bb = x_in[b, n, kvh, group_size, even + 1].cast(dtypes.float)
-    stores += [k_out[b, n, kvh, even].store((a * c - bb * s).cast(k.dtype)),
-               k_out[b, n, kvh, even + 1].store((a * s + bb * c).cast(k.dtype)),
-               v_out[b, n, kvh, even].store(x_in[b, n, kvh, group_size + 1, even]),
-               v_out[b, n, kvh, even + 1].store(x_in[b, n, kvh, group_size + 1, even + 1])]
-    ordered = UOp.group(*stores)
-  assert ordered is not None
-  return ordered.end(pair, n, b).sink(arg=KernelInfo(name="fused_qkv_rope_forward"))
+  code = (pathlib.Path(__file__).parent / "fused_qkv_rope.cpp").read_text()
+  threads = 256
+  thread_idx = UOp.special(threads, "lidx0")
+  block_idx_x, block_idx_y = UOp.special(B, "gidx0"), UOp.special(N, "gidx1")
+  sink = UOp.sink(q.base, k.base, v.base, xqkv.base, freqs_cis.base, thread_idx, block_idx_x, block_idx_y,
+                  arg=KernelInfo(name="fused_qkv_rope_forward"))
+  compile_args = ["-std=c++20", "-ffast-math", f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}",
+                  f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DTHREADS_PER_BLOCK={threads}"]
+  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
 def custom_fused_qkv_rope_backward(dxqkv:UOp, dq:UOp, dk:UOp, dv:UOp, freqs_cis:UOp,
@@ -126,7 +109,8 @@ def fused_qkv_rope(xqkv:Tensor, freqs_cis:Tensor, n_heads:int, n_kv_heads:int, h
 def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
  return _sharded_empty(ref.shape, ref, axis)

-def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=0):
+@functools.cache
+def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink):
  def grad(dou:UOp, ker:UOp) -> tuple:
    do = Tensor(dou, device=dou.device)
    attn = Tensor(ker.src[1].after(ker), device=ker.src[1].device)
@@ -145,7 +129,7 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
    delta_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
    delta_vec, dq = Tensor.custom_kernel(delta_vec, dq, attn, do, fxn=functools.partial(custom_fa_backward_pre, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:2]

-    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, window=window))[:3]
+    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:3]

    if D == 64:
      dq = dq.reshape(B, H, N//16, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2).permute(0, 1, 2, 8, 9, 10, 11, 3, 4, 6, 7, 5, 12).reshape(B, H, N, D).transpose(1, 2)
@@ -165,7 +149,7 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
  return grad

 # TODO: remove write_flat once scheduler can remove reshapes between custom_kernel. TestCustomKernel.test_simple_reshape
-def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None, window:int=0):
+def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None):
  assert attn_mask is None, "attn_mask not supported"
  assert is_causal, "only causal attention supported"

@@ -192,18 +176,18 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
  attn = _sharded_empty((B, N, H * D), xq, axis=shard_axis) if write_flat else _sharded_empty_like(xq, axis=shard_axis)
  l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)

-  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=window)
+  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink)

  fwd_inputs = (attn, l_vec, xq, xk, xv) + ((sinks,) if has_sink else ())
-  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink, window=window), grad_fxn=grad)[:2]
+  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink), grad_fxn=grad)[:2]

  return attn, attn, l_vec

@functools.cache
-def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True, window:int=0):
+def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True):
  code = (pathlib.Path(__file__).parent / "fa_fwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}", f"-DWINDOW={window}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}"]

  Q_BLOCK_SIZE = 32
  NUM_WARPS = 8
@@ -263,10 +247,10 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
-def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, window:int=0):
+def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
  code = (pathlib.Path(__file__).parent / "fa_bwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DWINDOW={window}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}"]

  BLOCK_SIZE_KV = 256
  GROUP_SIZE = H // H_KV
@@ -269,9 +269,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p

    qo_tile<D, float> q_reg_fl;
    load<1, qo_tile<D, float>, _gl_QKVO>(q_reg_fl, g.Qg, {batch_idx, tile_idx, head_idx, 0});
-    #if !WINDOW
    mul(q_reg_fl, q_reg_fl, TEMPERATURE_SCALE);  // Use sqrtf for clarity
-    #endif
    copy(q_reg, q_reg_fl);
    transpose(q_reg_transposed, q_reg);

@@ -290,9 +288,6 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    zero(att_block[0]);
    transpose(k_reg_transposed, k_reg);
    mma_AtB(att_block[0], k_reg_transposed, q_reg_transposed, att_block[0]);
-    #if WINDOW
-    mul(att_block[0], att_block[0], TEMPERATURE_SCALE);
-    #endif
    __builtin_amdgcn_sched_barrier(0);
    if constexpr (causal) {
        const int kv_end_pos = (min_tile + 1) * KV_BLOCK_SIZE;
@@ -342,9 +337,6 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
        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
-        mul(att_block[1], att_block[1], TEMPERATURE_SCALE);
-        #endif
 #if WINDOW
        // window masks interior tiles that causal skips
        mask_kv_tile(att_block[1], tile_idx, j - 2, neg_inf_v, lane);
@@ -409,9 +401,6 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
        zero(att_block[0]);
        transpose(k_reg_transposed, k_reg);
        mma_AtB(att_block[0], k_reg_transposed, q_reg_transposed, att_block[0]);
-        #if WINDOW
-        mul(att_block[0], att_block[0], TEMPERATURE_SCALE);
-        #endif
        //      Finish softmax for QK1
        exp2(att_block[1].tiles[1][0], att_block[1].tiles[1][0]);
        mul(norm_vec, norm_vec, scale_vec);
@@ -480,9 +469,6 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    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
-    mul(att_block[1], att_block[1], TEMPERATURE_SCALE);
-    #endif
    //      Finish softmax for QK2
    exp2(att_block[0].tiles[1][0], att_block[0].tiles[1][0]);
    mul(norm_vec, norm_vec, scale_vec);
@@ -549,9 +535,6 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    zero(att_block[0]);
    transpose(k_reg_transposed, k_reg);
    mma_AtB(att_block[0], k_reg_transposed, q_reg_transposed, att_block[0]);
-    #if WINDOW
-    mul(att_block[0], att_block[0], TEMPERATURE_SCALE);
-    #endif
    //      Finish softmax for QK3
    exp2(att_block[1].tiles[1][0], att_block[1].tiles[1][0]);
    mul(norm_vec, norm_vec, scale_vec);
@@ -614,9 +597,6 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    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
-    mul(att_block[1], att_block[1], TEMPERATURE_SCALE);
-    #endif
    //      Finish softmax for QK4
    exp2(att_block[0].tiles[1][0], att_block[0].tiles[1][0]);
    mul(norm_vec, norm_vec, scale_vec);
@@ -0,0 +1,69 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_bf16.h>
+
+#ifndef ATTN_B
+#define ATTN_B 2
+#endif
+#ifndef ATTN_N
+#define ATTN_N 8192
+#endif
+#ifndef ATTN_H
+#define ATTN_H 32
+#endif
+#ifndef ATTN_H_KV
+#define ATTN_H_KV 8
+#endif
+#ifndef ATTN_D
+#define ATTN_D 128
+#endif
+#ifndef THREADS_PER_BLOCK
+#define THREADS_PER_BLOCK 256
+#endif
+
+constexpr int GROUP_SIZE = ATTN_H / ATTN_H_KV;
+constexpr int HALF_D = ATTN_D / 2;
+constexpr int PACKED_D = (GROUP_SIZE + 2) * ATTN_D;
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_BLOCK) void
+fused_qkv_rope_forward(
+    __hip_bfloat16*       __restrict__ q,
+    __hip_bfloat16*       __restrict__ k,
+    __hip_bfloat16*       __restrict__ v,
+    const __hip_bfloat16* __restrict__ xqkv,
+    const __hip_bfloat16* __restrict__ freqs_cis) {
+  const int b = blockIdx.x;
+  const int n = blockIdx.y;
+  const int bn = b * ATTN_N + n;
+  const int packed_bn = bn * ATTN_H_KV * PACKED_D;
+  const int q_bn = bn * ATTN_H * ATTN_D;
+  const int kv_bn = bn * ATTN_H_KV * ATTN_D;
+
+  if (threadIdx.x < HALF_D) {
+    const int pair = threadIdx.x;
+    const int even = pair << 1;
+    const float c = static_cast<float>(freqs_cis[((n * HALF_D + pair) * 2) + 0]);
+    const float s = static_cast<float>(freqs_cis[((n * HALF_D + pair) * 2) + 1]);
+
+    for (int kvh = 0; kvh < ATTN_H_KV; kvh++) {
+      const int base = packed_bn + kvh * PACKED_D;
+
+      for (int rep = 0; rep < GROUP_SIZE; rep++) {
+        const int qbase = base + rep * ATTN_D;
+        const int h = kvh * GROUP_SIZE + rep;
+        const float a = static_cast<float>(xqkv[qbase + even]);
+        const float bb = static_cast<float>(xqkv[qbase + even + 1]);
+        const int out = q_bn + h * ATTN_D + even;
+        q[out] = static_cast<__hip_bfloat16>(a * c - bb * s);
+        q[out + 1] = static_cast<__hip_bfloat16>(a * s + bb * c);
+      }
+
+      const float a = static_cast<float>(xqkv[base + GROUP_SIZE * ATTN_D + even]);
+      const float bb = static_cast<float>(xqkv[base + GROUP_SIZE * ATTN_D + even + 1]);
+      const int out = kv_bn + kvh * ATTN_D + even;
+      k[out] = static_cast<__hip_bfloat16>(a * c - bb * s);
+      k[out + 1] = static_cast<__hip_bfloat16>(a * s + bb * c);
+      v[out] = xqkv[base + (GROUP_SIZE + 1) * ATTN_D + even];
+      v[out + 1] = xqkv[base + (GROUP_SIZE + 1) * ATTN_D + even + 1];
+    }
+  }
+}
@@ -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);
+  }
+}
@@ -209,7 +209,7 @@ class ST:
    return cls(uop, rows, cols, layout, base_shape, ker)

  def swizzle(self, row, col):
-    swizzled_offset = self.base_shape.swizzle(row, col, self._uop.dtype)
+    swizzled_offset = self.base_shape.swizzle(row, col, self._uop.dtype.scalar())

    row = swizzled_offset // self.base_shape.cols
    col = swizzled_offset % self.base_shape.cols
@@ -4,7 +4,7 @@
 # A006 Lambda argument `input` is shadowing a Python builtin
 from tinygrad import Tensor, dtypes, Device
 from tinygrad.uop.ops import Ops, GroupOp
-from tinygrad.helpers import getenv, prod, strides_for_shape
+from tinygrad.helpers import getenv, prod, strides_for_shape, argfix
 import torch.lib
 TORCH_DEBUG = getenv("TORCH_DEBUG")
 import torch, pathlib, operator, functools, weakref
@@ -73,12 +73,6 @@ def wrap_view_op(fn):
    return wrap(ret)
  return _wrap

-# NOTE: list assignment raises IndexError on an out of range dim, and the index must be a tuple: a list of all ints is one advanced index
-def _index_dim(self, dim, idx):
-  idxs = [slice(None)] * self.ndim
-  idxs[dim] = idx
-  return self[tuple(idxs)]
-
 view_ops = {
  "aten.view": Tensor.reshape,
  "aten._unsafe_view": Tensor.reshape,  # when are views unsafe, and do we care?
@@ -88,13 +82,15 @@ view_ops = {
  "aten.transpose.int": Tensor.transpose,
  "aten.squeeze.dim": Tensor.squeeze,
  "aten.unsqueeze": Tensor.unsqueeze,
-  "aten.select.int": _index_dim,
+  "aten.select.int": lambda self, dim, idx: self[(slice(None),) * (dim%self.ndim) + (idx,)],
  "aten.permute": Tensor.permute,
  "aten.alias": lambda self: self,
  "aten.diagonal": Tensor.diagonal,
-  "aten.slice.Tensor": lambda self, dim=0, start=None, end=None, step=1: _index_dim(self, dim, slice(start, end, step)),
  }

+# torch 2.10 handles this natively
+if tuple(map(int, torch.__version__.split('.')[:2])) < (2, 10): view_ops.update({"aten.detach": Tensor.detach})
+
 for k,v in view_ops.items(): torch.library.impl(k.replace("aten.", "aten::"), "privateuseone")(wrap_view_op(v))

 def _get_view_ops(view): return getattr(view, "_view_ops", [])
@@ -103,21 +99,46 @@ def _apply_view_ops(target, ops):
  for fn, args, kwargs in ops: target = fn(target, *args, **kwargs)
  return target

-# a chain of reshapes is undone by reshaping the value back to the base
+# similar to https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/InferSize.h
+def _reshape_target_shape(shape:tuple[int, ...], args) -> tuple[int, ...]|None:
+  if not (req := argfix(*args)): return None
+  new_shape, infer_idx = [], -1
+  for i, s in enumerate(req):
+    if s is None: s = shape[i] if i < len(shape) else None
+    if not isinstance(s, int): return None
+    if s == -1:
+      if infer_idx != -1: return None
+      infer_idx = len(new_shape)
+    new_shape.append(s)
+  total = prod(shape)
+  if infer_idx != -1:
+    known = prod(x for x in new_shape if x != -1)
+    if known == 0:
+      if total != 0: return None
+      new_shape[infer_idx] = 0
+    else: new_shape[infer_idx] = total // known
+  return tuple(new_shape) if prod(new_shape) == total else None
+
+# TODO: can we get rid of this? only for test_flatten_reshape_add
 def _try_simple_reshape_view_write(base: Tensor, view: Tensor, val: Tensor) -> bool:
  if not (ops := _get_view_ops(view)): return False
-  if any(fn is not Tensor.reshape for fn, _, _ in ops): return False
-  base.assign(val.reshape(base.shape))
+  shapes = [base.shape]
+  for fn, args, _ in ops:
+    if fn is Tensor.reshape:
+      if not (next_shape := _reshape_target_shape(shapes[-1], args)): return False
+      shapes.append(next_shape)
+  if shapes[-1] != view.shape: return False
+  for s in reversed(shapes[:-1]): val = val.reshape(s)
+  base.assign(val)
  return True

 def _view_write(base: Tensor, view: Tensor, value: Tensor) -> None:
  val = value if value.dtype == base.dtype else value.cast(base.dtype)
+  if view.shape == base.shape: return base.assign(val)
  if _try_simple_reshape_view_write(base, view, val): return
  idx_base = Tensor.arange(base.numel(), dtype=dtypes.int32).reshape(base.shape)
  idx_view = _apply_view_ops(idx_base, _get_view_ops(view)).reshape(-1)
-  # clone, not contiguous: contiguous() on a base that already owns its buffer returns the base itself, and scattering
-  # into that is an in-place write to a buffer other tensors still hold, which setitem refuses
-  flat_base = base.reshape(base.numel()).clone()
+  flat_base = base.reshape(base.numel()).contiguous()
  flat_base[idx_view] = val.reshape(-1)
  base.assign(flat_base.reshape(base.shape))

@@ -145,6 +166,11 @@ def _index_put_impl_(self, indices, values, accumulate=False, unsafe=False):
 def index_put(self, indices, values, accumulate=False):
  return aten.index_put(self.cpu(), [z.cpu() if isinstance(z, torch.Tensor) else None for z in indices], values.clone().cpu(), accumulate).tiny()

+@torch.library.impl("aten::isin.Tensor_Tensor_out", "privateuseone")
+def isin_tensor_tensor_out(x, y, *, assume_unique=False, invert=False, out=None):
+  result = (unwrap(x).unsqueeze(-1) == unwrap(y).flatten()).any(-1)
+  return out.copy_(wrap(~result if invert else result))
+
@torch.library.impl("aten::randperm.generator_out", "privateuseone")
 def randperm_generator(n, generator=None, out=None):
  if generator is not None: raise NotImplementedError("tinygrad torch backend does not support torch.Generator for randperm")
@@ -205,6 +231,49 @@ def as_strided(tensor:torch.Tensor, size, stride, storage_offset=None):
 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=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 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))
+  return wrap(ret)
+
+@torch.library.impl("aten::max_pool2d_with_indices", "privateuseone")
+def max_pool2d_with_indices(self:torch.Tensor, kernel_size:tuple[int, ...], stride=None, padding=0, dilation=1, ceil_mode=False):
+  # TODO: supprt stride [] in tinygrad?
+  if stride is not None and len(stride) == 0: stride = None
+  ret, idx = unwrap(self).max_pool2d(kernel_size, stride, dilation, padding, ceil_mode, return_indices=True)
+  return (wrap(ret), wrap(idx.cast(dtypes.int64)))
+
+@torch.library.impl("aten::max_pool2d_with_indices_backward", "privateuseone")
+def max_pool2d_with_indices_backward(grad_out:torch.Tensor, self:torch.Tensor, kernel_size:tuple[int, ...], stride=None, padding=0, dilation=1, ceil_mode=False, indices=None):
+  return wrap(Tensor.max_unpool2d(unwrap(grad_out), unwrap(indices), output_size=unwrap(self).shape))
+
+@torch.library.impl("aten::max_unpool2d", "privateuseone")
+def max_unpool2d(self:torch.Tensor, indices:torch.Tensor, output_size):
+  return wrap(unwrap(self).max_unpool2d(unwrap(indices), output_size=output_size))
+
+@torch.library.impl("aten::arange", "privateuseone")
+def arange(end, dtype=None, device=None, pin_memory=None):
+  has_float = isinstance(end, float)
+  return wrap(Tensor.arange(0, end, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
+
+@torch.library.impl("aten::arange.start", "privateuseone")
+def arange_start(start, end, dtype=None, device=None, pin_memory=None):
+  has_float = any(isinstance(x, float) for x in (start, end))
+  return wrap(Tensor.arange(start, end, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
+
+@torch.library.impl("aten::arange.start_step", "privateuseone")
+def arange_start_step(start, end, step, dtype=None, device=None, pin_memory=None):
+  has_float = any(isinstance(x, float) for x in (start, end, step))
+  return wrap(Tensor.arange(start, end, step, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
+
@torch.library.impl("aten::convolution_overrideable", "privateuseone")
 def convolution_overrideable(input, weight, bias, stride, padding, dilation, transposed, output_padding, groups):
  if TORCH_DEBUG >= 1:
@@ -225,27 +294,12 @@ def convolution_backward_overrideable(grad_out, input, weight, stride, padding,
  grads = out.gradient(*[t for t,m in zip([input, weight, bias], output_mask) if m], gradient=grad_out)
  return tuple([wrap(grads.pop(0)) if m else None for m in output_mask])

-# the functional scatters. without an impl aten falls back to a path that assumes a real storage: "self.has_storage() INTERNAL ASSERT FAILED"
-def _scatter_into(self, src, dim, index):
-  out = unwrap(self).clone()
-  slices = [slice(None)] * out.ndim
-  slices[dim] = index
-  out[slices] = unwrap(src).cast(out.dtype)  # torch casts src to self's dtype, tinygrad setitem demands they already match
-  return wrap(out)
-
-@torch.library.impl("aten::slice_scatter", "privateuseone")
-def slice_scatter(self, src, dim=0, start=None, end=None, step=1): return _scatter_into(self, src, dim, slice(start, end, step))
-
-@torch.library.impl("aten::select_scatter", "privateuseone")
-def select_scatter(self, src, dim, index): return _scatter_into(self, src, dim, index)
-
-@torch.library.impl("aten::diagonal_scatter", "privateuseone")
-def diagonal_scatter(self, src, offset=0, dim1=0, dim2=1):
-  # a diagonal is not one axis, so scatter through the flat indices it picks out
-  base, out = unwrap(self), unwrap(self).clone().reshape(-1)
-  idx = Tensor.arange(base.numel(), dtype=dtypes.int32).reshape(base.shape).diagonal(offset, dim1, dim2).reshape(-1)
-  out[idx] = unwrap(src).cast(base.dtype).reshape(-1)
-  return wrap(out.reshape(base.shape))
+@torch.library.impl("aten::slice.Tensor", "privateuseone")
+@wrap_view_op
+def slice_tensor(self, dim=0, start=None, end=None, step=1):
+  slices = [slice(None)] * self.ndim
+  slices[dim] = slice(start, end, step)
+  return self[slices]

@torch.library.impl("aten::slice_backward", "privateuseone")
 def slice_backward(grad_out, input_sizes, dim, start, end, step):
@@ -287,14 +341,19 @@ for dim in [1, 2, 3]:
    torch.library.impl(f"aten::{pad_type}_pad{dim}d", "privateuseone")(functools.partial(pad_forward, mode=mode))
    torch.library.impl(f"aten::{pad_type}_pad{dim}d_backward", "privateuseone")(functools.partial(pad_backward, mode=mode))

-# the schemas are all positional: (self, output_size, align_corners, *scales) for linear, (self, output_size, *scales) for nearest.
-def upsample(self, size, *args, mode=None):
-  return wrap(Tensor.interpolate(unwrap(self), size, mode=mode, align_corners=args[0] if mode == "linear" else False))
+def upsample(self, size, align_corners=False, mode=None): return wrap(Tensor.interpolate(unwrap(self), size, mode=mode, align_corners=align_corners))
 for i,pre in enumerate(["", "bi", "tri"]):
  torch.library.impl(f"aten::upsample_{pre}linear{i+1}d", "privateuseone")(functools.partial(upsample, mode="linear"))
  torch.library.impl(f"aten::upsample_nearest{i+1}d", "privateuseone")(functools.partial(upsample, mode="nearest"))
  torch.library.impl(f"aten::_upsample_nearest_exact{i+1}d", "privateuseone")(functools.partial(upsample, mode="nearest-exact"))

+@torch.library.impl("aten::scatter_add.out", "privateuseone")
+def scatter_add(self, dim, index, src, out):
+  self, index, src, out_unwrapped = unwrap(self), unwrap(index), unwrap(src), unwrap(out)
+  if self.shape == (): _apply_inplace(out_unwrapped, src)
+  else: _apply_inplace(out_unwrapped, Tensor.scatter_reduce(self, dim, index, src, reduce='sum'))
+  return out
+
 def _copy_between_devices(src, dest, cast_dtype, to_device, non_blocking=False):
  if src.is_tiny and dest.is_tiny:
    src_t, dest_t = unwrap(src), unwrap(dest)
@@ -345,11 +404,15 @@ def sort_values(input, dim=-1, descending=False, stable=True, values=None, indic
  _apply_inplace(unwrap(indices), out_indices.cast(dtypes.int64))
  return values, indices

+@torch.library.impl("aten::_linalg_svd", "privateuseone")
+def _linalg_svd(self, full_matrices=False):
+  U, S, Vh = unwrap(self).svd(full_matrices)
+  return wrap(U), wrap(S), wrap(Vh)
+
 # register some decompositions
 from torch._decomp import get_decompositions
 decomps = [
  aten.native_layer_norm_backward,
-  aten.native_group_norm_backward,
  aten.linalg_cross,
  aten.addmm,
  aten.addcmul,
@@ -384,20 +447,12 @@ decomps = [
  aten._softmax_backward_data, aten.embedding_dense_backward,
  aten.linalg_vector_norm,
  aten.binary_cross_entropy, aten.binary_cross_entropy_backward,
-  # the C++ mse/smooth_l1 kernels resize their out tensor, and a tiny tensor has no storage to resize
-  aten.mse_loss, aten.mse_loss_backward,
-  aten.smooth_l1_loss, aten.smooth_l1_loss_backward,
  aten.upsample_nearest2d.out,
-  # NOTE: only the "out" overload, the "vec" one is CompositeImplicitAutograd and overriding it loses the autograd kernel
-  aten.upsample_bicubic2d.out,
-  aten._adaptive_avg_pool2d,
  # activations
  aten.hardswish, aten.hardswish_backward,
  aten.hardtanh, aten.hardtanh_backward,
  aten.gelu, aten.gelu_backward,
-  # NOTE: no aten.logical_or here, its decomposition reaches aten.bitwise_or through a path that checks aliasing by
-  # reading storage, which a tiny tensor has none of. it gets a direct impl below instead
-  aten.logical_and, aten.logical_xor,
+  aten.logical_and,
  aten.randint,
  aten.eye,
  aten.hardsigmoid_backward,
@@ -440,7 +495,7 @@ simple_tensor_methods = [
  # reduce
  "all", "any", "argmax", "argmin", "cumsum", "cumprod",
  # complex
-  "linspace"]
+  "avg_pool2d", "linspace"]

 tiny_backend_out = {**{f"aten.{x}.out":getattr(Tensor,x) for x in simple_tensor_methods}, **{
  "aten.add.out": lambda input,other,alpha=1: input+alpha*other,
@@ -485,8 +540,6 @@ tiny_backend_out = {**{f"aten.{x}.out":getattr(Tensor,x) for x in simple_tensor_
  "aten.where.self_out": Tensor.where,
  "aten.prod.int_out": Tensor.prod,
  "aten.scatter.src_out": Tensor.scatter,
-  "aten.scatter_add.out": lambda self,dim,index,src: src if self.shape == () else Tensor.scatter_reduce(self, dim, index, src, reduce="sum"),
-  "aten.isin.Tensor_Tensor_out": lambda x,y,assume_unique=False,invert=False: (x.unsqueeze(-1)==y.flatten()).any(-1) != invert,
  # NOTE: axis=[] in torch means all, change tinygrad?
  "aten.sum.IntList_out": lambda self,axis,keepdim=False,dtype=None:
    self.sum(axis if axis is None or len(axis) else None, keepdim,
@@ -502,9 +555,10 @@ 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}"
-    # writing out= is an in-place write like any other: through the base if it is a view, refreshing any derived views
-    _apply_inplace(out, assigned)
-    return out
+    # 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

 def _inplace_op(t, new_value):
@@ -512,14 +566,7 @@ def _inplace_op(t, new_value):
  else: _apply_inplace(t, new_value)
  return t

-# the three arange overloads are one function at different arity, and dtype/layout/device/pin_memory are keyword only in all of them
-def _arange(*args, dtype=None, **_):
-  return Tensor.arange(*args, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if any(isinstance(x, float) for x in args) else torch.int64)))
-
-def _empty(size, dtype=None, device=None, **_):
-  return Tensor.empty(*size, dtype=_from_torch_dtype(dtype or torch.get_default_dtype()), device=_from_torch_device(device))
-
-tiny_backend = {**tiny_backend_out, **{
+tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
  "aten.remainder.Scalar_Tensor": lambda x,y: x%y,
  "aten.floor_divide": lambda x,y: x//y,
  "aten.floor_divide_.Tensor": lambda x,y: x//y,
@@ -532,8 +579,8 @@ tiny_backend = {**tiny_backend_out, **{
  # inplace ops using replace for fusion
  "aten.zero_": lambda x: x.const_like(0),
  "aten.fill_.Scalar": lambda x, y: x.const_like(y),
-  "aten.add_.Tensor": lambda self, other, alpha=1: self + other * alpha,
-  "aten.add_.Scalar": lambda self, other, alpha=1: self + other * alpha,
+  "aten.add_.Tensor": lambda self, other, alpha=1.0: self + other * alpha,
+  "aten.add_.Scalar": lambda self, other, alpha=1.0: self + other * alpha,
  "aten.mul_.Tensor": lambda self, other: self * other,
  "aten.mul_.Scalar": lambda self, other: self * other,
  # relu doesn't have an out form?
@@ -566,9 +613,7 @@ tiny_backend = {**tiny_backend_out, **{
  # these don't work in out form, they have size 0
  "aten.abs": Tensor.abs,
  "aten.logical_not": Tensor.logical_not,
-  # compare against zero first: logical_* is bool-valued for any input dtype, while | is bitwise
-  "aten.logical_or": lambda x, y: (x != 0) | (y != 0),
-  "aten.logical_or_": lambda x, y: (x != 0) | (y != 0),
+  "aten.logical_or_": lambda x, y: x | y,
  "aten.multinomial": Tensor.multinomial,
  "aten.masked_fill_.Scalar": lambda self, mask, value: self.masked_fill(mask, value),
  "aten.masked_fill_.Tensor": lambda self, mask, value: self.masked_fill(mask, value),
@@ -577,7 +622,14 @@ tiny_backend = {**tiny_backend_out, **{
  "aten.masked_select": Tensor.masked_select,
  "aten.all": Tensor.all,
  "aten.sgn": Tensor.sign,
+  "aten.acos": Tensor.acos,
  "aten.any": Tensor.any,
+  "aten.bitwise_not": Tensor.bitwise_not,
+  "aten.argmax": Tensor.argmax,
+  "aten.argmin": Tensor.argmin,
+  "aten.asinh": Tensor.asinh,
+  "aten.mul": Tensor.mul,
+  "aten.atanh": Tensor.atanh,
  "aten.fill_.Tensor": lambda self, value: self.const_like(value.reshape(()).item()),
  "aten.flip": Tensor.flip,
  "aten.scatter_reduce.two": Tensor.scatter_reduce,
@@ -588,22 +640,10 @@ tiny_backend = {**tiny_backend_out, **{
  "aten.add.Tensor": lambda input,other,alpha=1: input+alpha*other,
  "aten.linspace": lambda start, stop, steps, dtype=None, **kwargs:
    Tensor.linspace(start, stop, steps, **({"dtype": _from_torch_dtype(dtype)} if dtype is not None else {})),
-  # the functional copy_. without an impl the fallback segfaults on a tensor with no storage
-  "aten.copy": lambda self,src,non_blocking=False: src.cast(self.dtype).to(self.device).expand(self.shape),
-  "aten.arange": lambda end, **kwargs: _arange(0, end, **kwargs),
-  "aten.arange.start": _arange,
-  "aten.arange.start_step": _arange,
-  # empty_strided takes the strides and drops them: we always allocate contiguous
-  "aten.empty_strided": lambda size, stride, **kwargs: _empty(size, **kwargs),
-  "aten.empty.memory_format": _empty,
-  # TODO: supprt stride [] in tinygrad?
-  "aten.max_pool2d_with_indices": lambda self,kernel_size,stride=None,padding=0,dilation=1,ceil_mode=False: ((r:=Tensor.max_pool2d(self, kernel_size, stride or None, dilation, padding, ceil_mode, return_indices=True))[0], r[1].cast(dtypes.int64)),
-  "aten.max_pool2d_with_indices_backward": lambda grad_out,self,kernel_size,stride=None,padding=0,dilation=1,ceil_mode=False,indices=None: Tensor.max_unpool2d(grad_out, indices, output_size=self.shape),
-  "aten.max_unpool2d": lambda self,indices,output_size: Tensor.max_unpool2d(self, indices, output_size=output_size),
-  "aten._linalg_svd": lambda self,full_matrices=False: Tensor.svd(self, full_matrices),
  "aten.topk": Tensor.topk,
  "aten.constant_pad_nd": lambda self, padding, value=0.0: self.pad(padding, mode="constant", value=value).contiguous(),
-  "aten.cumsum": lambda self, dim: self.cumsum(dim),
+  # TODO: input contiguous is needed to prevent CFGContext circular dependency assertion for shapes >512 (see test_cumsum_arange_large)
+  "aten.cumsum": lambda self, dim: self.contiguous().cumsum(dim),
  "aten.logsumexp": lambda self, axis, keepdim=False: self.logsumexp(axis[0], keepdim=keepdim),
  "aten.roll": Tensor.roll,
  "aten.logcumsumexp": Tensor.logcumsumexp,
@@ -612,7 +652,6 @@ tiny_backend = {**tiny_backend_out, **{
    self.ones_like(**{k: v for k, v in {"dtype": _from_torch_dtype(dtype) if dtype else None,
                                        "device": _from_torch_device(device) if device else None}.items() if v is not None}),
  "aten.max.dim": lambda self, dim, keepdim=False: (self.max(dim, keepdim), self.argmax(dim, keepdim).cast(dtype=dtypes.int64)),
-  "aten.min.dim": lambda self, dim, keepdim=False: (self.min(dim, keepdim), self.argmin(dim, keepdim).cast(dtype=dtypes.int64)),
  "aten.cummax": lambda self, dim: ((r := self.cummax(dim))[0], r[1].cast(dtypes.int64)),
  "aten.cummin": lambda self, dim: ((r := self.cummin(dim))[0], r[1].cast(dtypes.int64)),
  "aten.nonzero": Tensor.nonzero,
@@ -674,16 +713,15 @@ def wrap_inplace_view_op(f):
  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 gets wrap_out's dtype cast, shape assert, and view write-through
+# 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():
  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: fxn = wrap_fxn(k, v)
-  elif writes == ["out"]: fxn = wrap_fxn(k, wrap_out(v))
-  else: raise RuntimeError(f"{k} writes {writes}: unhandled writable arg in schema")
+  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")
@@ -83,12 +83,6 @@ class TestTorchBackend(unittest.TestCase):
    torch.add(torch.ones(5, device=device), torch.ones(5, device=device), out=a)
    self.assertEqual(a.detach().storage_offset(), 3)

-  def test_out_refreshes_views_of_base(self):
-    a = torch.zeros(4, device=device)
-    v = a[2:]
-    torch.add(torch.ones(4, device=device), torch.ones(4, device=device), out=a)
-    np.testing.assert_equal(v.cpu().numpy(), [2., 2.])
-
  @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)
@@ -172,15 +166,6 @@ class TestTorchBackend(unittest.TestCase):
    expected = np.array([[1.5, 5.2, 9.0], [13.2, 17.1, 18.4]], dtype=np.float32)
    np.testing.assert_equal(y3.cpu().numpy(), expected)

-  def test_argmax_argmin(self):
-    a = torch.arange(12, dtype=torch.float32, device=device).reshape(3, 4)
-    c = a.cpu()
-    for got, want in [(a.argmax(), c.argmax()), (a.argmin(0), c.argmin(0)), (a.argmax(1, keepdim=True), c.argmax(1, keepdim=True)),
-                      (torch.min(a, 1).indices, torch.min(c, 1).indices), (torch.max(a, 1).indices, torch.max(c, 1).indices),
-                      (torch.min(a, 1).values, torch.min(c, 1).values), (torch.min(a, 1, keepdim=True).indices, torch.min(c, 1, keepdim=True).indices)]:
-      self.assertEqual(got.dtype, want.dtype)  # torch's arg reduces are int64, tinygrad's are int32
-      np.testing.assert_equal(got.cpu().numpy(), want.numpy())
-
  def test_isfinite(self):
    a = torch.ones(4, device=device)
    np.testing.assert_equal(torch.isfinite(a).cpu().numpy(), [True, True, True, True])
@@ -388,22 +373,6 @@ class TestTorchBackend(unittest.TestCase):
    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_groupnorm_backward(self):
-    def run(dev):
-      x = torch.arange(24., device=dev).reshape(2, 4, 3).requires_grad_()
-      w = torch.linspace(0.5, 2.0, 4).to(dev).requires_grad_()
-      torch.nn.functional.group_norm(x, 2, w, torch.zeros(4, device=dev)).square().sum().backward()
-      return x.grad.cpu().numpy(), w.grad.cpu().numpy()
-    for got, want in zip(run(device), run("cpu")): np.testing.assert_allclose(got, want, atol=1e-4, rtol=1e-3)
-
-  def test_mse_smooth_l1_loss_backward(self):
-    def run(dev, loss):
-      x = torch.arange(4., device=dev).requires_grad_()
-      loss(x, torch.ones(4, device=dev)).backward()
-      return x.grad.cpu().numpy()
-    for loss in [torch.nn.functional.mse_loss, torch.nn.functional.smooth_l1_loss]:
-      np.testing.assert_allclose(run(device, loss), run("cpu", loss), atol=1e-6)
-
  def test_batchnorm_unsqueeze(self):
    bn = torch.nn.BatchNorm2d(4).to(device)
    x = torch.randn(8, 4, 3, 3, device=device)
@@ -547,15 +516,6 @@ class TestTorchBackend(unittest.TestCase):
    cpu_res = torch.arange(20, dtype=torch.float32)[::2][1:4].numpy()
    np.testing.assert_equal(torch_res, cpu_res)

-  def test_select_out_of_range_dim(self):
-    a = torch.arange(12, dtype=torch.int32, device=device).reshape(3, 4)
-    with self.assertRaises(IndexError): a.select(5, 0)
-
-  def test_select_collapses_the_only_dim(self):
-    a = torch.arange(3, dtype=torch.int32, device=device)
-    self.assertEqual(a.select(0, 1).shape, ())
-    np.testing.assert_equal(a.select(0, 1).cpu().numpy(), 1)
-
  def test_slice_negative_dim(self):
    a = torch.arange(13, dtype=torch.int32, device=device).repeat(8, 1)
    torch_chunks = a.chunk(3, -1)
@@ -836,86 +796,6 @@ class TestTorchBackend(unittest.TestCase):
    np.testing.assert_allclose(w_tiny.grad.cpu().numpy(), w_cpu.grad.numpy(), atol=1e-4, rtol=1e-3)
    np.testing.assert_allclose(b_tiny.grad.cpu().numpy(), b_cpu.grad.numpy(), atol=1e-4, rtol=1e-3)

-  def test_write_through_detach_of_unrealized(self):
-    a = torch.empty(4, device=device)
-    a.detach().fill_(3)
-    np.testing.assert_equal(a.cpu().numpy(), [3, 3, 3, 3])
-
-  def test_square_transpose_inplace(self):
-    # a same-shape transpose is not a reshape: writing the transposed values straight back would scramble the base
-    a = torch.tensor([[0., 1., 2.], [3., 4., 5.], [6., 7., 8.]], device=device)
-    a.transpose(0, 1).add_(100)
-    np.testing.assert_equal(a.cpu().numpy(), [[100., 101., 102.], [103., 104., 105.], [106., 107., 108.]])
-
-  def test_interpolate(self):
-    a = torch.arange(4, dtype=torch.float32, device=device).reshape(1, 1, 2, 2)
-    nearest = torch.nn.functional.interpolate(a, scale_factor=2.0)
-    np.testing.assert_equal(nearest.cpu().numpy()[0, 0], [[0, 0, 1, 1], [0, 0, 1, 1], [2, 2, 3, 3], [2, 2, 3, 3]])
-    linear = torch.nn.functional.interpolate(a, size=(4, 4), mode="bilinear", align_corners=False)
-    ref = torch.nn.functional.interpolate(a.cpu(), size=(4, 4), mode="bilinear", align_corners=False)
-    np.testing.assert_allclose(linear.cpu().numpy(), ref.numpy(), rtol=1e-5)
-
-  def test_interpolate_bicubic_area(self):
-    a = torch.arange(32, dtype=torch.float32, device=device).reshape(1, 2, 4, 4)
-    for mode, scale in [("bicubic", 2.0), ("area", 0.5)]:
-      ref = torch.nn.functional.interpolate(a.cpu(), scale_factor=scale, mode=mode)
-      np.testing.assert_allclose(torch.nn.functional.interpolate(a, scale_factor=scale, mode=mode).cpu().numpy(), ref.numpy(), atol=1e-4)
-
-  @unittest.expectedFailure
-  def test_interpolate_bicubic_backward(self):
-    # the forward comes from a decomposition, but aten::upsample_bicubic2d_backward has none (nor does
-    # aten::_adaptive_avg_pool2d_backward, for area), so training through these modes needs a real kernel
-    x = torch.arange(32., dtype=torch.float32, device=device).reshape(1, 2, 4, 4).requires_grad_()
-    torch.nn.functional.interpolate(x, scale_factor=2.0, mode="bicubic").sum().backward()
-
-  @unittest.expectedFailure
-  def test_interpolate_inexact_scale(self):
-    # torch forwards the raw scale_factor, Tensor.interpolate recomputes it from output_size, and they disagree here
-    a = torch.arange(6, dtype=torch.float32, device=device).reshape(1, 1, 2, 3)
-    tiny = torch.nn.functional.interpolate(a, scale_factor=2.5, mode="bilinear")
-    ref = torch.nn.functional.interpolate(a.cpu(), scale_factor=2.5, mode="bilinear")
-    np.testing.assert_allclose(tiny.cpu().numpy(), ref.numpy(), rtol=1e-5)
-
-  def test_logical_or_xor(self):
-    a = torch.tensor([True, True, False, False], device=device)
-    b = torch.tensor([True, False, True, False], device=device)
-    np.testing.assert_equal(torch.logical_or(a, b).cpu().numpy(), [True, True, True, False])
-    np.testing.assert_equal(torch.logical_xor(a, b).cpu().numpy(), [False, True, True, False])
-    # bool-valued whatever the input dtype, so this is not | and ^
-    i, j = torch.tensor([2, 0, 5, 0], device=device), torch.tensor([0, 0, 1, 1], device=device)
-    np.testing.assert_equal(torch.logical_or(i, j).cpu().numpy(), [True, False, True, True])
-    np.testing.assert_equal(torch.logical_xor(i, j).cpu().numpy(), [True, False, False, True])
-
-  def test_slice_scatter(self):
-    # the scatters are functional: they return a new tensor and must leave the one they were given alone
-    a = torch.arange(12, dtype=torch.float32, device=device).reshape(3, 4)
-    out = torch.slice_scatter(a, torch.ones(1, 4, device=device), 0, 0, 1)
-    np.testing.assert_equal(out.cpu().numpy(), [[1, 1, 1, 1], [4, 5, 6, 7], [8, 9, 10, 11]])
-    np.testing.assert_equal(a.cpu().numpy(), np.arange(12, dtype=np.float32).reshape(3, 4))
-
-  def test_slice_scatter_casts_src(self):
-    a = torch.zeros(3, 4, device=device)
-    out = torch.slice_scatter(a, torch.ones(1, 4, dtype=torch.int32, device=device), 0, 0, 1)
-    self.assertEqual(out.dtype, torch.float32)
-    np.testing.assert_equal(out.cpu().numpy()[0], np.ones(4, dtype=np.float32))
-
-  def test_select_scatter(self):
-    a = torch.arange(12, dtype=torch.float32, device=device).reshape(3, 4)
-    out = torch.select_scatter(a, torch.ones(4, device=device), 0, 1)
-    np.testing.assert_equal(out.cpu().numpy(), [[0, 1, 2, 3], [1, 1, 1, 1], [8, 9, 10, 11]])
-
-  def test_diagonal_scatter(self):
-    a = torch.zeros(3, 3, device=device)
-    out = torch.diagonal_scatter(a, torch.arange(3, dtype=torch.float32, device=device))
-    np.testing.assert_equal(out.cpu().numpy(), np.diag([0., 1., 2.]))
-    np.testing.assert_equal(a.cpu().numpy(), np.zeros((3, 3), dtype=np.float32))
-
-  def test_copy_functional(self):
-    # without an impl this segfaults rather than fails: a regression here takes the whole run down
-    a = torch.arange(4, dtype=torch.float32, device=device)
-    out = torch.ops.aten.copy(a, torch.zeros(4, device=device))
-    np.testing.assert_equal(out.cpu().numpy(), [0., 0., 0., 0.])
-    np.testing.assert_equal(a.cpu().numpy(), [0., 1., 2., 3.])

 from tinygrad import Tensor
 class TestBackendHelpers(unittest.TestCase):
@@ -1,6 +1,6 @@
 [project]
 name = "tinygrad"
-version = "0.14.0"
+version = "0.13.0"
 description = "You like pytorch? You like micrograd? You love tinygrad! <3"
 authors = [{ name = "George Hotz" }]

@@ -84,7 +84,7 @@ testing = [
  "pillow",
  "onnx==1.19.0",
  "onnx2torch",
-  "onnxruntime==1.24.1",
+  "onnxruntime",
  "opencv-python",
  "transformers",
  "sentencepiece",
@@ -111,10 +111,6 @@ docs = [
  "numpy",
 ]
 mesa = ["tinymesa==25.2.7.2"]
-autogen = [
-  "pyyaml",
-  "mako",
-]


 [tool.mutmut]
@@ -1002,39 +1002,6 @@ class TestBarrier(unittest.TestCase):
    for tid in range(64):
      self.assertEqual(st.vgpr[tid][0], tid + 100 + 1000, f"tid={tid}")

-class TestSMaxMinSCCRegressions(unittest.TestCase):
-  """Regression test: S_MAX sets SCC only on strict inequality (equal operands -> SCC=0)."""
-
-  def test_s_max_i32_equal_scc(self):
-    st = run_program([s_mov_b32(s[4], 64), s_mov_b32(s[5], 64), s_max_i32(s[6], s[4], s[5])], n_lanes=1)
-    self.assertEqual(st.scc, 0)
-    self.assertEqual(st.sgpr[6], 64)
-    st = run_program([s_mov_b32(s[4], 65), s_mov_b32(s[5], 64), s_max_i32(s[6], s[4], s[5])], n_lanes=1)
-    self.assertEqual(st.scc, 1)  # still set when strictly greater
-
-  def test_s_max_u32_equal_scc(self):
-    st = run_program([s_mov_b32(s[4], 64), s_mov_b32(s[5], 64), s_max_u32(s[6], s[4], s[5])], n_lanes=1)
-    self.assertEqual(st.scc, 0)
-
-class TestAbsdiffOverflowRegressions(unittest.TestCase):
-  """Regression test: S_ABSDIFF_I32 computes abs on the WRAPPED 32-bit difference (found by random difftest vs hardware)."""
-
-  def test_s_absdiff_wrapped(self):
-    # |45 - (-2147483647)| overflows int32; hardware takes abs of the wrapped 32-bit difference
-    instructions = [s_mov_b32(s[4], 45), s_mov_b32(s[5], 0x80000001), s_absdiff_i32(s[6], s[4], s[5])]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.sgpr[6], 0x7FFFFFD4)
-    self.assertEqual(st.scc, 1)
-    # INT_MIN - 1 wraps to +2147483647, already positive
-    instructions = [s_mov_b32(s[4], 0x80000000), s_mov_b32(s[5], 1), s_absdiff_i32(s[6], s[4], s[5])]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.sgpr[6], 0x7FFFFFFF)
-    # equality -> 0 and SCC=0
-    instructions = [s_mov_b32(s[4], 7), s_mov_b32(s[5], 7), s_absdiff_i32(s[6], s[4], s[5])]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.sgpr[6], 0)
-    self.assertEqual(st.scc, 0)
-

 if __name__ == '__main__':
  unittest.main()
@@ -1629,66 +1629,5 @@ class TestSwap(unittest.TestCase):
    self.assertEqual(st.vgpr[0][1], 0x55555555)


-class TestCvtFrexpRegressions(unittest.TestCase):
-  """Regression tests for float<->int conversion and FREXP corner cases (found by random difftest vs hardware)."""
-
-  def test_cvt_i32_f32_nan_is_zero(self):
-    """v_cvt_i32_f32 of NaN is 0, not INT_MIN (x86 cvttss2si returns INT_MIN)."""
-    for nan in (0x7FC00000, 0xFFC00000, 0x7F800001):
-      st = run_program([v_mov_b32_e32(v[0], nan), v_cvt_i32_f32_e32(v[1], v[0])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][1], 0, f"nan=0x{nan:08x}")
-
-  def test_cvt_i32_f32_positive_overflow(self):
-    """v_cvt_i32_f32 saturates positive overflow/inf to INT_MAX, not INT_MIN."""
-    for bits in (0x7F800000, 0x4F000000, 0x4F800000):  # +inf, 2^31, ~2^32
-      st = run_program([v_mov_b32_e32(v[0], bits), v_cvt_i32_f32_e32(v[1], v[0])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][1], 0x7FFFFFFF, f"bits=0x{bits:08x}")
-
-  def test_cvt_i32_f32_negative_overflow(self):
-    """v_cvt_i32_f32 saturates negative overflow/-inf to INT_MIN."""
-    for bits in (0xFF800000, 0xCF000001):  # -inf, below -2^31
-      st = run_program([v_mov_b32_e32(v[0], bits), v_cvt_i32_f32_e32(v[1], v[0])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][1], 0x80000000, f"bits=0x{bits:08x}")
-
-  def test_cvt_u32_f32_nan_is_zero(self):
-    """v_cvt_u32_f32 of NaN is 0, not UINT_MAX."""
-    for nan in (0x7FC00000, 0xFFC00000, 0x7F800001):
-      st = run_program([v_mov_b32_e32(v[0], nan), v_cvt_u32_f32_e32(v[1], v[0])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][1], 0, f"nan=0x{nan:08x}")
-
-  def test_cvt_i32_f64_nan_and_overflow(self):
-    """v_cvt_i32_f64: NaN -> 0, positive overflow/+inf -> INT_MAX."""
-    st = run_program([v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0x7FF80000), v_cvt_i32_f64_e32(v[2], v[0:1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0)
-    st = run_program([v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0x41F00000), v_cvt_i32_f64_e32(v[2], v[0:1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x7FFFFFFF)  # 2^32 -> INT_MAX
-
-  def test_frexp_f32_denormal(self):
-    """v_frexp_exp/mant_f32 of denormal/zero inputs is (0, signed zero) on hardware."""
-    for bits in (0x00000001, 0x007FFFFF, 0x00000000):
-      st = run_program([v_mov_b32_e32(v[0], bits), v_frexp_exp_i32_f32_e32(v[1], v[0]), v_frexp_mant_f32_e32(v[2], v[0])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][1] & 0xFFFFFFFF, 0, f"exp bits=0x{bits:08x}")
-      self.assertEqual(st.vgpr[0][2], bits & 0x80000000, f"mant bits=0x{bits:08x}")
-    # negative denormal: mant is -0.0
-    st = run_program([v_mov_b32_e32(v[0], 0x80000001), v_frexp_mant_f32_e32(v[2], v[0])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x80000000)
-
-  def test_frexp_f64_denormal(self):
-    """v_frexp_exp_f64 of a denormal returns the normalized exponent (-1073 for min-denormal); zero -> 0."""
-    st = run_program([v_mov_b32_e32(v[0], 1), v_mov_b32_e32(v[1], 0), v_frexp_exp_i32_f64_e32(v[2], v[0:1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2] & 0xFFFFFFFF, 0xFFFFFBCF)  # -1073
-    st = run_program([v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0), v_frexp_exp_i32_f64_e32(v[2], v[0:1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0)
-
-  def test_frexp_exp_inf_nan(self):
-    """v_frexp_exp of +/-inf and NaN is 0 on hardware (host frexp gives 129/1024), for both f32 and f64."""
-    for bits in (0x7F800000, 0xFF800000, 0x7FC00000):
-      st = run_program([v_mov_b32_e32(v[0], bits), v_frexp_exp_i32_f32_e32(v[1], v[0])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][1] & 0xFFFFFFFF, 0, f"f32 bits=0x{bits:08x}")
-    for lo, hi in ((0, 0x7FF00000), (0, 0xFFF00000), (0, 0x7FF80000), (1, 0x7FF00000)):
-      st = run_program([v_mov_b32_e32(v[0], lo), v_mov_b32_e32(v[1], hi), v_frexp_exp_i32_f64_e32(v[2], v[0:1])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][2] & 0xFFFFFFFF, 0, f"f64 bits=0x{hi:08x}{lo:08x}")
-
-
 if __name__ == '__main__':
  unittest.main()
@@ -989,53 +989,6 @@ class TestCarryOps(unittest.TestCase):
    self.assertEqual(st.vgpr[0][0], 0)  # 0xFFFFFFFF + 1 + 0 = 0 (overflow)
    self.assertEqual(st.vcc, 0xDEADBEEF)  # VCC unchanged - carry was discarded

-class TestSelectFlushRegressions(unittest.TestCase):
-  """Regression tests: f32 MIN/MAX flush denormal inputs to signed zero (select-style ops propagate inputs bitwise)."""
-
-  def test_v_min_f32_denormal_flush(self):
-    """min(denormal, 1.0) is +0, min(-denormal, -1.0) is -0."""
-    st = run_program([v_mov_b32_e32(v[0], 0x00000001), v_mov_b32_e32(v[1], 0x3F800000), v_min_f32_e32(v[2], v[0], v[1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x00000000)
-    # flush(-denormal) = -0.0 > -1.0, so the result is -1.0 (both operand orders)
-    st = run_program([v_mov_b32_e32(v[0], 0x80000001), v_mov_b32_e32(v[1], 0xBF800000), v_min_f32_e32(v[2], v[0], v[1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0xBF800000)
-    st = run_program([v_mov_b32_e32(v[1], 0xBF800000), v_mov_b32_e32(v[2], 0x80000001), v_min_f32_e32(v[3], v[1], v[2])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][3], 0xBF800000)
-
-  def test_v_max_f32_denormal_flush(self):
-    """max(-denormal, -1.0) is -0; max(+denormal, -0) is +0."""
-    st = run_program([v_mov_b32_e32(v[0], 0x80000001), v_mov_b32_e32(v[1], 0xBF800000), v_max_f32_e32(v[2], v[0], v[1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x80000000)
-    st = run_program([v_mov_b32_e32(v[0], 0x00000001), v_mov_b32_e32(v[1], 0x80000000), v_max_f32_e32(v[2], v[0], v[1])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x00000000)
-
-
-class TestCarryExecRegressions(unittest.TestCase):
-  """Regression tests: per-lane VCC writes (carry ops) zero inactive lane bits - VCC = mask & EXEC, never preserved."""
-
-  def test_co_ci_e32_vcc_masked_by_exec(self):
-    """v_sub_co_ci_u32_e32 with EXEC=0xFFFF0000: hw clears inactive VCC bits instead of preserving them."""
-    instructions = [
-      s_mov_b32(EXEC_LO, 0xFFFF0000),
-      s_mov_b32(VCC_LO, 0xFFFFFFFF),  # preset all bits
-      v_mov_b32_e32(v[0], 0xFFFFFFFE), v_mov_b32_e32(v[1], 0x80000000),
-      v_sub_co_ci_u32_e32(v[2], v[0], v[1]),  # active lanes: no borrow
-    ]
-    st = run_program(instructions, n_lanes=32)
-    self.assertEqual(st.vcc, 0x00000000)
-
-  def test_co_ci_e32_vcc_masked_by_exec_ones(self):
-    """Same with all-ones carry: VCC = borrow_mask & EXEC."""
-    instructions = [
-      s_mov_b32(EXEC_LO, 0x0F0F0F0F),
-      s_mov_b32(VCC_LO, 0),
-      v_mov_b32_e32(v[0], 0xFFFFFFFF), v_mov_b32_e32(v[1], 1),
-      v_add_co_ci_u32_e32(v[2], v[0], v[1]),  # all lanes would carry if active
-    ]
-    st = run_program(instructions, n_lanes=32)
-    self.assertEqual(st.vcc, 0x0F0F0F0F)
-    self.assertEqual(st.vgpr[31][2], 0)  # 0xFFFFFFFF + 1 wraps to 0 in active lanes
-

 if __name__ == '__main__':
  unittest.main()
@@ -4,7 +4,6 @@ Includes: v_fma_f32, v_div_scale_f32, v_div_fmas_f32, v_div_fixup_f32,
          v_alignbit_b32, v_bfe_i32, v_mad_u64_u32, v_readlane_b32, v_writelane_b32
 """
 import unittest
-from tinygrad.helpers import OSX
 from test.amd.hw.helpers import *

 class TestFMA(unittest.TestCase):
@@ -3265,23 +3264,6 @@ class TestVOP3ClampMAD(unittest.TestCase):
    # 0xFFFF * 2 = 0x1FFFE, low 16 bits = 0xFFFE
    self.assertEqual(st.vgpr[0][3] & 0xFFFF, 0xFFFE, f"expected 0xFFFE, got 0x{st.vgpr[0][3] & 0xFFFF:04x}")

-class TestMadNarrowClampRegressions(unittest.TestCase):
-  """Regression tests: mad i16/i24 with clamp saturate to narrow output range (found by random difftest vs hardware)."""
-
-  def test_mad_i16_clamp_sat_max(self):
-    # neg/src-floggled 16-bit mul operands are sign-extended after toggling bit15; sum > INT_MAX saturates
-    instructions = [s_mov_b32(s[4], 1232348160), v_mov_b32_e32(v[3], 0x80000000),
-                    v_mov_b32_e32(v[1], 0x7F7FFFFF), v_mad_i32_i16(v[0], s[4], v[3], v[1], 0, 3, 5, 1)]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vgpr[0][0], 0x7FFFFFFF)
-
-  def test_mad_i24_clamp_sat_min(self):
-    # sext24(-6344704) * sext24(+4210688) << -2^31 saturates to INT_MIN
-    instructions = [s_mov_b32(s[7], 4290772992), v_mov_b32_e32(v[1], 1077936128),
-                    v_mad_i32_i24(v[0], s[7], v[1], v[1], 1, 0, 0, 1)]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vgpr[0][0], 0x80000000)
-

 class TestCvtPkF16(unittest.TestCase):
  """Tests for V_CVT_PK_RTZ_F16_F32 - pack two f32 to f16 with round toward zero."""
@@ -3669,80 +3651,6 @@ class TestPermlane(unittest.TestCase):
    self.assertEqual(st.vgpr[21][1], 5)
    self.assertEqual(st.vgpr[31][1], 15)

-class TestClampLdExpRegressions(unittest.TestCase):
-  """Regression tests for f32 clamp (-0 -> +0) and ldexp input passthrough."""
-
-  def test_clamp_negative_zero(self):
-    """clmp=1 maps -0.0 to +0.0 (found by random difftest vs hardware)."""
-    instructions = [
-      v_mov_b32_e32(v[0], 0x80000000), v_mov_b32_e32(v[1], 0x80000000),
-      v_add_f32_e64(v[2], v[0], v[1], clmp=1),  # -0 + -0 = -0, clamp -> +0
-    ]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x00000000)
-    instructions = [
-      v_mov_b32_e32(v[0], 0x3F800000), v_mov_b32_e32(v[1], 0x80000000),
-      v_min_f32_e64(v[2], v[0], v[1], clmp=1),  # min(1.0, -0) = -0, clamp -> +0
-    ]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x00000000)
-
-  def test_ldexp_special_inputs(self):
-    """v_ldexp_f32 of 0/-0/inf/NaN propagates the input instead of computing val * 2**exp (0*inf = NaN on host)."""
-    # -0.0 * 2^INT_MIN = -0.0 (src1 as integer exponent; huge negative)
-    instructions = [v_mov_b32_e32(v[0], 0x80000000), v_mov_b32_e32(v[1], 0x80000000), v_ldexp_f32(v[2], v[0], v[1])]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x80000000)
-    # inf stays inf even with negative exponent
-    instructions = [v_mov_b32_e32(v[0], 0x7F800000), v_mov_b32_e32(v[1], 0xFFFFFF80), v_ldexp_f32(v[2], v[0], v[1])]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x7F800000)
-
-  def test_ldexp_denormal_flush(self):
-    """v_ldexp_f32/f64 flush denormal inputs to signed zero (found by random difftest vs hardware)."""
-    # ldexp(+denorm, 1) = +0, ldexp(-denorm, 250) = -0
-    for src, exp_val, want in [(0x00000001, 1, 0x00000000), (0x80000001, 250, 0x80000000)]:
-      st = run_program([v_mov_b32_e32(v[0], src), v_mov_b32_e32(v[1], exp_val), v_ldexp_f32(v[2], v[0], v[1])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][2], want)
-
-  def test_v_mul_neg_modifier_nan_sign(self):
-    """neg modifier is a pure sign-bit toggle on a NaN operand; result keeps that sign (found by random difftest)."""
-    # mul(normal, NEG(ABS(qNaN))): NaN payload negated in the operand stays negative qNaN
-    instructions = [v_mov_b32_e32(v[0], 0xC96CF47F), v_mov_b32_e32(v[1], 0x7FC00000),
-                    v_mul_f32_e64(v[2], v[0], v[1], s[0], 0, 7, 6)]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0xFFC00000)
-    # plain neg modifier still applies to non-NaN values: mul(-1.0, NEG(2.0)) = +2.0
-    st = run_program([v_mov_b32_e32(v[0], 0xBF800000), v_mov_b32_e32(v[1], 0x40000000),
-                      v_mul_f32_e64(v[2], v[0], v[1], s[0], 0, 2, 0)], n_lanes=1)
-    self.assertEqual(st.vgpr[0][2], 0x40000000)
-
-
-class TestNaNPropagationRegressions(unittest.TestCase):
-  """Regression tests: float arithmetic propagates a NaN from the FIRST NaN operand, quieted with its own sign/payload."""
-
-  @unittest.skipIf(OSX, "broken on mac, TODO: why?")
-  def test_mul_nan_priority(self):
-    # first NaN operand wins (sign+payload), not x86's second-source propagation
-    for a, b, want in [(0x7FC00001, 0x7F800003, 0x7FC00001), (0xFFC00005, 0x7F800003, 0xFFC00005),
-                       (0x7F800001, 0xFFC00005, 0x7FC00001), (0xFF9F1800, 0x7F800001, 0xFFDF1800)]:
-      st = run_program([v_mov_b32_e32(v[0], a), v_mov_b32_e32(v[1], b),
-                        v_mul_f32_e32(v[2], v[0], v[1])], n_lanes=1)
-      self.assertEqual(st.vgpr[0][2], want, f"mul({a:#x}, {b:#x})")
-
-class TestMinMaxFlushE64Regressions(unittest.TestCase):
-  """Regression tests: f32 min/max/median flush denormal inputs to signed zero (e64 forms)."""
-
-  def test_v_min3_f32_denormal_flush(self):
-    st = run_program([v_mov_b32_e32(v[0], 0x00000001), v_mov_b32_e32(v[1], 0x3F800000), v_mov_b32_e32(v[2], 0x40000000),
-                      v_min3_f32(v[3], v[0], v[1], v[2])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][3], 0x00000000)  # min(+denorm, 1, 2) = +0
-
-  def test_v_med3_f32_denormal_flush(self):
-    st = run_program([v_mov_b32_e32(v[0], 0x80000001), v_mov_b32_e32(v[1], 0x3F800000), v_mov_b32_e32(v[2], 0x40000000),
-                      v_med3_f32(v[3], v[0], v[1], v[2])], n_lanes=1)
-    self.assertEqual(st.vgpr[0][3], 0x3F800000)  # med(-0, 1, 2) = 1
-

 if __name__ == '__main__':
  unittest.main()
@@ -973,71 +973,6 @@ class TestCmpxPartialWavefront(unittest.TestCase):
    self.assertEqual(st.sgpr[EXEC_LO.offset] & 0xFFFFFFFF, 0x4,
                     "Only lane 2 should be active after v_cmpx_eq_u32_e64")

-class TestClassDenormalRegressions(unittest.TestCase):
-  """Regression tests: V_CMP_CLASS classifies denormals as DENORMAL (raw bits), not as zero class."""
-
-  def test_class_pos_denormal(self):
-    for bits in (0x00000001, 0x007FFFFF):
-      instructions = [v_mov_b32_e32(v[0], bits), v_mov_b32_e32(v[1], 0x80), v_cmp_class_f32_e64(VCC_LO, v[0], v[1])]
-      st = run_program(instructions, n_lanes=1)
-      self.assertEqual(st.vcc, 1, f"bits=0x{bits:08x}")  # n_lanes=1
-      # ...and it is not the zero class
-      instructions = [v_mov_b32_e32(v[0], bits), v_mov_b32_e32(v[1], 0x40), v_cmp_class_f32_e64(VCC_LO, v[0], v[1])]
-      st = run_program(instructions, n_lanes=1)
-      self.assertEqual(st.vcc, 0, f"bits=0x{bits:08x}")
-
-  def test_class_neg_denormal(self):
-    instructions = [v_mov_b32_e32(v[0], 0x80000001), v_mov_b32_e32(v[1], 0x10), v_cmp_class_f32_e64(VCC_LO, v[0], v[1])]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc, 1)  # n_lanes=1
-    instructions = [v_mov_b32_e32(v[0], 0x80000001), v_mov_b32_e32(v[1], 0x20), v_cmp_class_f32_e64(VCC_LO, v[0], v[1])]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc, 0)  # not the negative-zero class
-
-
-class TestIntCmpModRegressions(unittest.TestCase):
-  """Regression tests: int compares (i32/u32) honor abs/neg as bit-level sign clear/flip (not integer abs/negate)."""
-
-  def test_cmp_i32_abs_neg_bit_level(self):
-    # abs(0x80000001) = 1 -> 1 > 1 is false (integer abs would give 2147483647 > 1)
-    instructions = [v_mov_b32_e32(v[0], 0x80000001), v_mov_b32_e32(v[1], 1), v_cmp_gt_i32_e64(VCC_LO, v[0], v[1], abs=1)]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc, 0)
-    # neg(0x80000001) flips the sign bit -> 1 > 2 is false (integer negate would give 2147483647 > 2)
-    instructions = [v_mov_b32_e32(v[0], 0x80000001), v_mov_b32_e32(v[1], 2), v_cmp_gt_i32_e64(VCC_LO, v[0], v[1], neg=1)]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc, 0)
-
-  def test_cmp_u32_abs_bit_level(self):
-    # abs(0x80000000) = 0 -> 0 < 1 is true
-    instructions = [v_mov_b32_e32(v[0], 0x80000000), v_mov_b32_e32(v[1], 1), v_cmp_lt_u32_e64(VCC_LO, v[0], v[1], abs=1)]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc, 1)  # n_lanes=1
-
-
-class TestCmpxSdstRegressions(unittest.TestCase):
-  """Regression tests: V_CMPX_*_E64 writes EXEC only, never SDST (hardware verified)."""
-
-  def test_cmpx_e64_no_sdst(self):
-    instructions = [
-      s_mov_b32(VCC_LO, 0),  # preset VCC to 0
-      v_mov_b32_e32(v[0], 0x3F800000), v_mov_b32_e32(v[1], 0x40000000),
-      v_cmpx_lt_f32_e64(VCC_LO, v[0], v[1]),  # 1.0 < 2.0
-    ]
-    st = run_program(instructions, n_lanes=32)
-    self.assertEqual(st.sgpr[EXEC_LO.offset], 0xFFFFFFFF)  # EXEC updated
-    self.assertEqual(st.vcc, 0)  # but VCC untouched
-
-  def test_cmpx_e64_partial_exec(self):
-    instructions = [
-      s_mov_b32(EXEC_LO, 0x0F0F0F0F),
-      s_mov_b32(VCC_LO, 0xFFFFFFFF),
-      v_mov_b32_e32(v[0], 0), v_mov_b32_e32(v[1], 0x3F800000),
-      v_cmpx_lt_f32_e64(VCC_LO, v[0], v[1]),
-    ]
-    st = run_program(instructions, n_lanes=32)
-    self.assertEqual(st.sgpr[EXEC_LO.offset], 0x0F0F0F0F)  # EXEC = computed & old EXEC
-

 if __name__ == '__main__':
  unittest.main()
@@ -67,26 +67,32 @@ class TestParseExpr(unittest.TestCase):

  def test_integer_literals(self):
    """Test parsing integer literals."""
-    self.assertIs(parse_expr('0', {}), UOp.const(0, dtypes.uint32))
-    self.assertIs(parse_expr('42', {}), UOp.const(42, dtypes.uint32))
-    self.assertIs(parse_expr('42U', {}), UOp.const(42, dtypes.uint32))
+    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."""
-    self.assertIs(parse_expr('-1', {}), UOp.const(-1, dtypes.int))
+    result = parse_expr('-1', {})
+    self.assertEqual(result.val, -1)
+    self.assertEqual(result.dtype, dtypes.int)

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

  def test_hex_literals(self):
    """Test parsing hex literals."""
-    self.assertIs(parse_expr('0xFF', {}), UOp.const(255, dtypes.uint32))
+    result = parse_expr('0xFF', {})
+    self.assertEqual(result.val, 255)

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

  def test_binary_ops(self):
    """Test parsing binary operations."""
@@ -97,7 +103,9 @@ class TestParseExpr(unittest.TestCase):
    self.assertEqual(result.op, Ops.ADD)

    # Subtraction with constant folding
-    self.assertIs(parse_expr('10 - 5', {}), UOp.const(5, dtypes.uint32))
+    result = parse_expr('10 - 5', {})
+    self.assertEqual(result.op, Ops.CONST)
+    self.assertEqual(result.val, 5)

  def test_ternary(self):
    """Test parsing ternary expressions."""
@@ -134,8 +142,15 @@ class TestForLoopParsing(unittest.TestCase):
    S0 = UOp.const(0, dtypes.uint32)
    _vrs, assigns = parse_pcode(pcode, {'S0': S0})

-    # every cond folds (S0 is a const), leaving the default branch: -1 in the destination dtype
-    self.assertIs(assigns[0][1].simplify(), UOp.const(-1, dtypes.uint32))
+    # Check that the innermost value (default) is -1 (may be wrapped in CAST)
+    val = assigns[0][1]
+    # Traverse to innermost WHERE
+    while val.op == Ops.WHERE:
+      val = val.src[2]  # false branch
+    # Unwrap CAST if present
+    while val.op == Ops.CAST:
+      val = val.src[0]
+    self.assertEqual(val.val, -1)

  def test_ctz_parsing(self):
    """Test CTZ pcode parsing."""
@@ -247,8 +262,8 @@ class TestDSPcodePatterns(unittest.TestCase):
    _, 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.assertIs(assigns[0][1][0].simplify(), UOp.const(108, dtypes.uint32))  # type: ignore[index]
-    self.assertIs(assigns[1][1][0].simplify(), UOp.const(120, dtypes.uint32))  # 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."""
@@ -265,8 +280,8 @@ class TestDSPcodePatterns(unittest.TestCase):
    _, 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.assertIs(assigns[0][1][1].simplify(), UOp.const(0xAAAAAAAA, dtypes.uint32))  # type: ignore[index]
-    self.assertIs(assigns[1][1][1].simplify(), UOp.const(0xBBBBBBBB, dtypes.uint32))  # 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."""
@@ -291,12 +306,12 @@ 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(row, dtypes.uint32)})
-      self.assertIs(parsed.simplify(), UOp.const(expected, 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(lane, dtypes.uint32)})
-      self.assertIs(parsed.simplify(), UOp.const(expected, dtypes.uint32))
+      self.assertEqual(parsed.simplify().val, expected)

  def test_permlane64_wave64_pcode_indices(self):
    vgpr = UOp.param(0, dtypes.uint32, (256,))
@@ -312,17 +327,19 @@ class TestConcatWidthParsing(unittest.TestCase):
      'S2': UOp.const(0, dtypes.uint32),
    }

-    def check_load_idx(v: UOp, expected: int):
+    def load_idx(v: UOp) -> int:
      simp = v.simplify()
      self.assertEqual(simp.op, Ops.LOAD)
      self.assertEqual(simp.src[0].op, Ops.INDEX)
-      self.assertIs(simp.src[0].src[1].simplify(), UOp.const(expected, dtypes.uint32))
+      idx = simp.src[0].src[1].simplify()
+      self.assertEqual(idx.op, Ops.CONST)
+      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.assertIs(assigns[lane][1][0].simplify(), UOp.const(dst_idx, dtypes.uint32))  # type: ignore[index]
-      check_load_idx(assigns[lane][1][1], src_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):
  """Test that all pcode from all architectures can be parsed."""
@@ -88,6 +88,7 @@ def run_rocprof_decoder(blobs: list[bytes], lib: bytes, base: int, target: str):
  if t.is_alive(): raise RuntimeError("rocprof decoder timeout")
  return occupancy_records, wave_insts

+@unittest.skip("TODO: fix to not require unpickling UOps.")
 class SQTTExamplesTestBase(unittest.TestCase):
  target: str
  examples: dict
@@ -1,28 +1,30 @@
 import unittest, contextlib
 from tinygrad import Device, Tensor, Context, TinyJit
-from tinygrad.device import Compiled, ProfileProgramEvent
+from tinygrad.device import Compiled, ProfileProgramEvent, ProfileDeviceEvent
 from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
 from tinygrad.viz.serve import load_amd_counters, VizData

@contextlib.contextmanager
 def save_sqtt():
-  Device[Device.DEFAULT].synchronize()
-  profile_start = len(Compiled.profile_events)
  data = VizData()
  yield data.ctxs
  Device[Device.DEFAULT].synchronize()
  Device[Device.DEFAULT]._at_profile_finalize()
-  load_amd_counters(data, [e for e in Compiled.profile_events[:profile_start] if isinstance(e, ProfileProgramEvent)] +
-                          Compiled.profile_events[profile_start:])
+  load_amd_counters(data, Compiled.profile_events)
  data.ctxs[:] = [r for r in data.ctxs if r["name"].startswith("SQTT")]

@unittest.skipUnless(Device.DEFAULT == "AMD", "only runs on AMD")
 class TestSQTTProfiler(unittest.TestCase):
+  # TODO: can we enable SQTT profiling in context?
  @classmethod
  def setUpClass(cls):
    if not Device[Device.DEFAULT].sqtt_enabled: raise unittest.SkipTest("device must be in SQTT profiling mode")

+  def setUp(self):
+    Device[Device.DEFAULT].synchronize()
+    Compiled.profile_events[:] = [e for e in Compiled.profile_events if isinstance(e, (ProfileProgramEvent, ProfileDeviceEvent))]
+
  def test_simple(self):
    t = Tensor.empty(1) + 1
    with save_sqtt() as sqtt:
@@ -1,5 +1,4 @@
 import unittest
-import functools
 from tinygrad import Tensor, Device, dtypes, Context
 from tinygrad.helpers import getenv, system, DEV
 from extra.gemm.cdna_asm_gemm import asm_gemm, hk_bf16_atb_gemm
@@ -10,7 +9,6 @@ from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8, FP8_MAX
 # Use DEV=NULL:HIP:gfx950 to also test the assembly
 def is_cdna4(): return Device[Device.DEFAULT].renderer.target.arch.startswith("gfx950")

-@functools.cache
 def has_hipcc():
  try: system("hipcc --version")
  except Exception: return False
@@ -188,7 +186,7 @@ class TestMXFP4(unittest.TestCase):
    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)
-    for _ in range(getenv("CNT", 1)): asm_gemm(a, b.T, mxfp4=True).realize()
+    asm_gemm(a, b.T, mxfp4=True).realize()

 # test the Asm GEMM with Llama shapes, only run on the real machine for speed

@@ -1,7 +1,7 @@
 import unittest, math
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.dtype import DTYPES_DICT
-from tinygrad.uop.ops import Ops, UOp, GroupOp
+from tinygrad.uop.ops import Ops, UOp
 from tinygrad.codegen.decomp.op import threefry2x32
 import numpy as np
 from test.helpers import not_support_multi_device
@@ -17,7 +17,7 @@ def _check_ast_count(desired_count:int, t:Tensor):
 class TestMovedConstFolding(unittest.TestCase):
  def test_contiguous_deviceless_const(self):
    t = Tensor(UOp.const(2.0, dtypes.float)).contiguous()
-    self.assertIs(t.uop, UOp.const(2.0, dtypes.float))
+    self.assertIs(t.uop.op, Ops.CONST)
    self.assertIsNone(t.uop.device)

  def test_add_shrunk_zero(self):
@@ -169,8 +169,8 @@ 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(5, dtypes.uint64), UOp.const(10, dtypes.uint64)).simplify()
-    self.assertEqual([u.op for u in x.toposort() if u.op in GroupOp.ALU], [])
+    x = threefry2x32(UOp.const(5, dtypes.uint64), UOp.const(10, dtypes.uint64))
+    self.assertIs(x.simplify().op, Ops.CONST)

 class TestTautologicalCompare(unittest.TestCase):
  # without const folding, these would have triggered -Wtautological-compare in clang
@@ -188,6 +188,7 @@ class TestTautologicalCompare(unittest.TestCase):
    np.testing.assert_equal((Tensor(True) < Tensor(False)).numpy(), False)
    np.testing.assert_equal((Tensor(True) < Tensor(True)).numpy(), False)

+  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "WEBGPU doesn't support NaN comparison correctly")
  def test_a_eq_a(self):
    # self eq is always true for int or bool
    a = Tensor([1, 2, 3])
@@ -4,7 +4,7 @@ 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, KernelCountException
+from test.helpers import assert_kernel_count

 # **** kernels ****

@@ -422,8 +422,9 @@ class TestCustomKernel(unittest.TestCase):
      return Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
    GlobalCounters.reset()
    y = run(x[0]).realize()
-    # backends that support contiguous views don't launch extra kernels
-    assert_kernel_count(2 if x[0].uop.contiguous_view() is None else 1)
+    # it's copying the input and the output
+    # 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")
@@ -474,7 +475,7 @@ class TestCustomKernelInput(unittest.TestCase):
    y.realize()
    kernel_count = GlobalCounters.kernel_count
    self.assertEqual(y.tolist(), x.add(1).tolist())
-    if kernel_count > max_kernels: raise KernelCountException(max_kernels, kernel_count)
+    self.assertLessEqual(kernel_count, max_kernels)
    # same test with @function, input is PARAM
    from tinygrad import function
    x0 = Tensor.arange(32).clone("CPU").realize()
@@ -487,7 +488,7 @@ class TestCustomKernelInput(unittest.TestCase):
    y = run(x0).realize()
    kernel_count = GlobalCounters.kernel_count
    self.assertEqual(y.tolist(), mop_fxn(x0).add(1).tolist())
-    if kernel_count > max_kernels: raise KernelCountException(max_kernels, kernel_count)
+    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)
@@ -340,9 +340,6 @@ class TestUint64DType(TestDType):
  DTYPE = dtypes.uint64
  def test_uint64_load(self):
    assert Tensor(2**64 - 1, dtype=dtypes.uint64).numpy() == 2**64 - 1
-  @unittest.skipIf(dtypes.double not in supported_dtypes, "needs float64")
-  def test_uint64_cast_double(self):
-    assert Tensor([2**32 + 1], dtype=dtypes.uint64).cast(dtypes.double).numpy() == 2**32 + 1

@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
 class TestEmulatedUInt64DType(TestUint64DType):
@@ -6,8 +6,6 @@ from tinygrad.tensor import _to_np_dtype
 from tinygrad.runtime.ops_python import from_storage_scalar
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
-from tinygrad.renderer.llvmir import CPULLVMRenderer
-from tinygrad.renderer.isa.x86 import X86Renderer
 from tinygrad.uop import Ops
 import numpy as np
 import pytest
@@ -66,8 +64,6 @@ ht.fp8e5m2fnuz = ht.uint8
 def universal_test(a, b, dtype, op):
  if not isinstance(op, tuple): op = (op, op)
  if op[0] == operator.mod and b == 0: return
-  # TODO: throws floating point exception
-  if isinstance(Device[Device.DEFAULT].renderer, (X86Renderer, CPULLVMRenderer)) and op[0] == operator.mod and a == dtype.min and b == -1: return
  # lt and max with nan is undefined in tinygrad
  if op[0] in (operator.lt, Tensor.maximum) and (math.isnan(a) or math.isnan(b)): return
  ta, tb = Tensor([a], dtype=dtype), Tensor([b], dtype=dtype)
@@ -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.cconst(i, dtypes.int))
+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):
@@ -46,10 +46,10 @@ class TestIselX86(unittest.TestCase):
  # complex address is [base + index*scale + displacement]
  def test_complex_address(self):
    a = UOp.variable("a", 0, 0, dtypes.int32)
-    load = UOp.param(0, dtypes.int32, (16,)).index(a + UOp.cconst(1, dtypes.int32)).load()
+    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].dtype is dtypes.int8 and n.src[2].src[0].op is Ops.CONST and n.src[2].src[0].val == 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()
@@ -360,7 +360,7 @@ class TestJitGraphSplit(unittest.TestCase):
    self.expect(f, inp, inp_cpu,
      graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
      multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
-      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()]) # cpu is hcq2 now, it does not join hcq graphs
+      hcqgraph=[self.ji_graph(4)])

  def test_jit_cpu_several(self):
    if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")
@@ -377,9 +377,9 @@ class TestJitGraphSplit(unittest.TestCase):
    inp = Tensor.randn(10, 10, device=Device.DEFAULT).realize()
    inp_cpu = Tensor.randn(10, 10, device="CPU").realize()
    self.expect(f, inp, inp_cpu,
-      graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
-      multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
-      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()])
+      graph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
+      multigraph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
+      hcqgraph=[self.ji_graph(5)])

  def test_jit_multidev(self):
    if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")
@@ -16,6 +16,8 @@ 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
+
@unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, ISARenderer), "isa backends don't preserve the op spec when lowering")
 class TestLinearizer(unittest.TestCase):
  def test_arg_dedup(self):
@@ -28,7 +30,7 @@ class TestLinearizer(unittest.TestCase):
    c = ((a.shrink(((0, 2),)) - a.shrink(((2, 4),))) - (b.shrink(((0, 2),)) - b.shrink(((2, 4),))))
    linear = c.schedule_linear()
    run_linear(linear)
-    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if not s.is_bound_var]
+    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if s.op is not Ops.BIND]
    assert len(rawbufs) == 3 and set(rawbufs[1:]) == {a.uop.base.realized, b.uop.base.realized}
    np_c = (np_a[:2] - np_a[2:]) - (np_b[:2] - np_b[2:])
    np.testing.assert_allclose(np_c, c.numpy(), atol=1e-4, rtol=1e-4)
@@ -246,10 +248,11 @@ class TestLinearizer(unittest.TestCase):
    uops = tuple(to_program(replace_opts(ast, opt), renderer=Device[Device.DEFAULT].renderer).src[1].src)
    begin_range = [i for i, x in enumerate(uops) if x.op is Ops.RANGE][-1]
    end_range = [i for i, x in enumerate(uops) if x.op is Ops.END][0]
+    for i,u in enumerate(uops): print(i, u.op, [uops.index(s) for s in u.src], u.arg, u.dtype)
    for u in uops:
      if u.op is Ops.STORE and u.src[0].addrspace is AddrSpace.REG:
        if uops.index(u) < begin_range:
-          assert u.src[1].op not in GroupOp.ALU
+          assert u.src[1].op is Ops.CONST
        else:
          assert u.src[1].op in GroupOp.ALU
          assert begin_range < uops.index(u) < end_range
@@ -265,9 +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].src[0].val) == ('gidx0', 6), idxs[0]
-    assert (idxs[1].arg, idxs[1].src[0].src[0].val) == ('gidx1', 5), idxs[1].arg
-    assert (idxs[2].arg, idxs[2].src[0].src[0].val) == ('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()
@@ -408,7 +411,7 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
  last_call = linear.src[-1]
  ast = last_call.src[0]
  assert ast.op is Ops.SINK, f"helper_realized_ast expects a SINK {last_call}"
-  last_bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]
+  last_bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]
  # now all input buffers in last_call should be realized
  # create fresh buffers for the outputs
  bufs = [Buffer(x.device, x.size, x.dtype).allocate() if i < len(ast.src) else x for i,x in enumerate(last_bufs)]
@@ -434,7 +437,7 @@ def reset_bufs(bufs:list[Buffer]):
  for buf in bufs: buf.copy_from(Buffer("PYTHON", buf.size, buf.dtype, opaque=memoryview(bytearray(buf.nbytes))))

 def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[],
-                               apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[], check_default_opt=True):
+                               apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[]):
  outbufs = real_bufs[:len(realized_ast.src)]
  wanna_output = [np.array(x).flatten() for x in wanna_output]
  buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in real_bufs]
@@ -456,7 +459,9 @@ def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[]
    for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)

  # Check correctness of handcoded optimiztions.
-  if check_default_opt: check_opt(None)
+  reset_bufs(outbufs)
+  run_prg(opts=None)
+  for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)
  for x in opts: # Check custom transformations if any.
    check_opt(([Opt(OptOps.TC, 0, (TC_SELECT.value, TC_OPT.value, 1))] if apply_tc else [])+x)

@@ -9,7 +9,7 @@ from extra.llama_kernels.swiglu import swiglu
 from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
 from extra.thunder.amd.fa import custom_fused_qkv_rope_backward, fused_qkv_rope
 from test.helpers import needs_second_gpu, assert_kernel_count
-from test.backend.test_asm_gemm import has_hipcc, is_cdna4
+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:
  Tensor.manual_seed(0)
@@ -99,20 +99,22 @@ class TestLocalAmax(unittest.TestCase):
    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")
 class TestFusedQKVRoPE(unittest.TestCase):
  SHAPE = (2, 8192, 32, 8, 128)

-  def setUp(self):
-    if dtypes.bfloat16 not in Device[Device.DEFAULT].renderer.supported_dtypes(): self.skipTest("test uses bf16 inputs")
-
  def rand_bf16(self, *shape:int) -> Tensor:
    return (Tensor.randn(*shape) * 0.1).cast(dtypes.bfloat16).contiguous().realize()

-  def test_forward(self):
+  def freqs_cis(self) -> Tensor:
+    _, N, _, _, D = self.SHAPE
+    return precompute_freqs_cis(D, N * 2).cast(dtypes.bfloat16).clone().realize()
+
+  def test_llama31_8b_forward(self):
    Tensor.manual_seed(0)
-    B, N, H, H_KV, D = 1, 32, 8, 2, 16
+    B, N, H, H_KV, D = self.SHAPE
    GROUP = H // H_KV
-    freqs_cis = (Tensor.randn(1, N * 2, 1, D // 2, 2) * 0.1).cast(dtypes.bfloat16).contiguous().realize()
+    freqs_cis = self.freqs_cis()

    x = self.rand_bf16(B, N, H_KV * (GROUP + 2) * D)
    q, k, v = fused_qkv_rope(x, freqs_cis, H, H_KV, D)
@@ -129,13 +131,12 @@ class TestFusedQKVRoPE(unittest.TestCase):
      self.assertTrue(k.allclose(k_ref, atol=2e-2, rtol=0).item(), "K forward mismatch")
      self.assertTrue(v.allclose(v_ref, atol=0, rtol=0).item(), "V forward mismatch")

-  @unittest.skipUnless(has_hipcc() and is_cdna4(), "backward kernel requires hipcc to compile")
-  def test_llama31_8b(self):
+  def test_llama31_8b_backward(self):
    Tensor.manual_seed(1)
    B, N, H, H_KV, D = self.SHAPE
    PARTIALS = 2
    GROUP = H // H_KV
-    freqs_cis = precompute_freqs_cis(D, N * 2).cast(dtypes.bfloat16).clone().realize()
+    freqs_cis = self.freqs_cis()
    dq = self.rand_bf16(B, N, H, D)
    dk_partial = self.rand_bf16(B * PARTIALS, N, H_KV, D)
    dv_partial = self.rand_bf16(B * PARTIALS, N, H_KV, D)
@@ -1,12 +1,12 @@
 import unittest, random
 from tinygrad import Tensor, Device, nn, GlobalCounters, TinyJit, dtypes, Variable
-from tinygrad.uop.ops import Ops, UOp, AxisType, graph_rewrite
+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, lower_and_compile, pm_beam
+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, check_schedule, assert_kernel_count, KernelCountException
+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")
@@ -58,11 +58,6 @@ class TestMultiTensor(unittest.TestCase):
    assert X.uop.ended_ranges == X.uop.src[1:]
    (X + X).realize()

-  def test_shard_invalids_contiguous(self):
-    # every store is Invalid, so none of them should become a (empty) kernel
-    t = Tensor.invalids(8).shard(devices_2, axis=0).contiguous()
-    self.assertEqual(len([c for c in t.schedule_linear().src if c.src[0].op is Ops.SINK]), 1)
-
  @unittest.expectedFailure # TODO: fix
  def test_shard_empty(self):
    GlobalCounters.reset()
@@ -77,16 +72,16 @@ class TestMultiTensor(unittest.TestCase):
    X.shard_(devices_2, 0)
    out = (X + X)
    linear = compile_linear(out.schedule_linear())
-    uops = [call.src[0].src[0] for call in linear.src if call.src[0].op is Ops.PROGRAM]
+    names = [call.src[0].src[0].arg.name for call in linear.src if call.src[0].op is Ops.PROGRAM]
    run_linear(linear)
-    self.assertEqual(len(set(uops)), 1, "function was relinearized")
+    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()
-    lin = UOp(Ops.LINEAR, src=(src.to(cpu_2[::-1]).schedule_linear().src[0],))
-    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): call = lower_and_compile(graph_rewrite(lin, pm_beam, ctx=1, walk=True)).src[0]
-    self.assertNotEqual(call.src[0].src[0].arg.applied_opts, ())
+    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()
@@ -400,7 +395,7 @@ class TestMultiBufferView(unittest.TestCase):
    linear, var_vals = b_multi.linear_with_vars()
    if all(not d.startswith(("WEBGPU", "CL")) for d in b_multi.device):
      compiled = [call for call in linear.src if call.src[0].op is Ops.SINK]
-      if len(compiled) != 0: raise KernelCountException(0, len(compiled))
+      self.assertEqual(len(compiled), 0, f"expected zero compiled kernels, got {len(compiled)}")
    run_linear(linear, var_vals)
    np.testing.assert_equal(b_multi.numpy(), b_ref.numpy())

@@ -6,7 +6,6 @@ from tinygrad.helpers import getenv, DEBUG, DEV, IMAGE, Context
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.renderer.nir import NIRRenderer
-from tinygrad.renderer.isa.x86 import X86Renderer

 TINY_BACKEND = getenv("TINY_BACKEND")
 if TINY_BACKEND:
@@ -721,11 +720,10 @@ class TestOps(unittest.TestCase):
      return torch.autograd.grad(t ** c, t)[0].item()
    for x in [-math.inf, 0, 1, math.inf]:
      for c in [-1, 0, 0.3, 1, 2]:
-        torch_out = get_torch_gradient(x, c)
-        # the pow backward routes through exp2/log2, whose 0/inf behavior is undefined on WEBGPU
-        if Device.DEFAULT == "WEBGPU" and not math.isfinite(torch_out): continue
        tiny_out = get_tiny_gradient(x, c)
+        torch_out = get_torch_gradient(x, c)
        if math.isnan(tiny_out):
+          if Device.DEFAULT == "WEBGPU": continue # TODO: WEBGPU issue with nan
          assert math.isnan(torch_out)
        else:
          self.assertAlmostEqual(tiny_out, torch_out, msg=f"{x}, {c}")
@@ -751,7 +749,6 @@ class TestOps(unittest.TestCase):
  def test_exp2_log2_zero_times_negative(self):
    # gallivm's exp2/log2 have "undefined behavior with infs, 0s and nans", so exp2(log2(0)*y) returns 0 instead of inf
    helper_test_op(None, lambda x,y: (x.log2()*y).exp2(), lambda x,y: (x.log2()*y).exp2(), vals=[[0.0], [-0.7]], forward_only=True)
-  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "pow at 0 routes through exp2/log2, whose 0/inf behavior is undefined on WEBGPU")
  def test_pow_zero_const(self):
    helper_test_op(None, lambda x: x**0.3, vals=[[0.0]])
    helper_test_op(None, lambda x: x**0.0, vals=[[0.0]])
@@ -809,8 +806,6 @@ class TestOps(unittest.TestCase):
    helper_test_op([], lambda: tor^0x1337, lambda: ten^0x1337, forward_only=True)
    helper_test_op([], lambda: 0x1337^tor, lambda: 0x1337^ten, forward_only=True)

-  # TODO: x86 PARAM dtype fails SPEC=2
-  @Context(SPEC=1 if isinstance(Device[Device.DEFAULT].renderer, X86Renderer) else 2)
  def test_and(self):
    data = [[1,-8,1],[32,1,6]]
    tor = torch.tensor(data, dtype=torch.int)
@@ -825,10 +820,6 @@ class TestOps(unittest.TestCase):
    helper_test_op([], lambda: tor0&tor1, lambda: ten0&ten1, forward_only=True)

    helper_test_op(None, lambda x: (1 < x) & (x < 2), forward_only=True, vals=[[1.2, 1.2, 1.2, 3.2]])
-    helper_test_op([(3000,)]*10, lambda *xs: (sum(xs[1:], xs[0]) > 5) & (xs[0] < 0.9), forward_only=True)
-
-    if not COMPILE_ONLY:
-      np.testing.assert_equal((Tensor(2**64-1, dtype=dtypes.uint64) & 0xFFFFFFFF).numpy(), 0xFFFFFFFF)

  def test_or(self):
    data = [[1,-8,1],[32,1,6]]
@@ -868,9 +859,9 @@ class TestOps(unittest.TestCase):
                   lambda: (ten << Tensor([0,2,4], dtype=dtypes.uint32)).cast(dtypes.int32), forward_only=True)
    helper_test_op([], lambda: tor.__lshift__(2), lambda: ten.__lshift__(2).cast(dtypes.int32), forward_only=True)
    helper_test_op([], lambda: tor.bitwise_left_shift(2), lambda: ten.lshift(2).cast(dtypes.int32), forward_only=True)
-    self.helper_test_exception([], lambda: torch.tensor([1.0]) << 2, lambda: (Tensor([1.0]) << 2).realize(), expected=RuntimeError)
-    self.helper_test_exception([], lambda: tor << torch.tensor([1.0]), lambda: (ten << Tensor([1.0])).realize(), expected=RuntimeError)
-    self.helper_test_exception([], lambda: tor << 1.0, lambda: (ten << 1.0).realize(), expected=RuntimeError)
+    self.helper_test_exception([], lambda: torch.tensor([1.0]) << 2, lambda: Tensor([1.0]) << 2, expected=RuntimeError)
+    self.helper_test_exception([], lambda: tor << torch.tensor([1.0]), lambda: ten << Tensor([1.0]), expected=RuntimeError)
+    self.helper_test_exception([], lambda: tor << 1.0, lambda: ten << 1.0, expected=RuntimeError)

  def test_rshift(self):
    data = [[0,1,2],[1<<8,1<<16,1<<31-1]]
@@ -884,8 +875,8 @@ class TestOps(unittest.TestCase):
                   lambda: (ten >> Tensor([0,2,4], dtype=dtypes.uint32)).cast(dtypes.int32), forward_only=True)
    helper_test_op([], lambda: tor.__rshift__(2), lambda: ten.__rshift__(2).cast(dtypes.int32), forward_only=True)
    helper_test_op([], lambda: tor.bitwise_right_shift(2), lambda: ten.rshift(2).cast(dtypes.int32), forward_only=True)
-    self.helper_test_exception([], lambda: torch.tensor([4.0]) >> 1, lambda: (Tensor([4.0]) >> 1).realize(), expected=RuntimeError)
-    self.helper_test_exception([], lambda: tor >> torch.tensor([1.0]), lambda: (ten >> Tensor([1.0])).realize(), expected=RuntimeError)
+    self.helper_test_exception([], lambda: torch.tensor([4.0]) >> 1, lambda: Tensor([4.0]) >> 1, expected=RuntimeError)
+    self.helper_test_exception([], lambda: tor >> torch.tensor([1.0]), lambda: ten >> Tensor([1.0]), expected=RuntimeError)

  def test_lshift_signed(self):
    data = [[-1, -3, 1, 7], [0, -2147483648, 2147483647, -1]]
@@ -1810,10 +1801,9 @@ class TestOps(unittest.TestCase):
      helper_test_op([()], lambda x: torch.nn.functional.hardtanh(x, -val, val), lambda x: x.hardtanh(-val, val), grad_atol=1e-6)
  def test_asinh(self):
    helper_test_op([(45,65)], lambda x: x.asinh(), grad_atol=1e-6)
-    helper_test_op([(45,65)], lambda x: x.asinh(), grad_atol=1e-6, low=-300, high=-297)
+    # TODO: this one has larger tol?
+    helper_test_op([(45,65)], lambda x: x.asinh(), atol=1e-2, rtol=2e-2, grad_rtol=2e-2, low=-300, high=-297)
    helper_test_op([(45,65)], lambda x: x.asinh(), grad_atol=1e-6, low=300, high=303)
-    helper_test_op([(45,65)], lambda x: x.asinh(), grad_atol=1e-6, low=-1e10, high=-1e9)
-    helper_test_op(None, lambda x: x.asinh(), grad_atol=1e-6, vals=[[-1.0, 0.0, 1.0]])
  def test_acosh(self):
    helper_test_op([(45,65)], lambda x: x.acosh(), grad_atol=1e-6)
    helper_test_op([(45,65)], lambda x: x.acosh(), grad_atol=1e-3, grad_rtol=1e-2, low=-300, high=-297)
@@ -2172,10 +2162,6 @@ class TestOps(unittest.TestCase):
  def test_roll(self):
    helper_test_op([(2, 4)], lambda x: x.roll(1))
    helper_test_op([(2, 4)], lambda x: x.roll((1,)))
-    helper_test_op([(0,)], lambda x: x.roll(1, 0))
-    helper_test_op([(2, 0, 3)], lambda x: x.roll(1, 0))
-    helper_test_op([(2, 0, 3)], lambda x: x.roll(1, 1))
-    helper_test_op([(2, 0, 3)], lambda x: x.roll(1))
    self.helper_test_exception([(2, 4)], lambda x: x.roll((1, 2)), expected=RuntimeError)
    helper_test_op([(2, 4)], lambda x: x.roll(1, 0))
    helper_test_op([(2, 4)], lambda x: x.roll(-1, 0))
@@ -87,8 +87,7 @@ class TestOptim(unittest.TestCase):
  def test_muon(self): self._test_muon(1, {'lr': 0.001}, 1e-3, 0)
  # TODO: disabled due to big atol
  # def test_muon_high_lr(self): self._test_muon(1, {'lr': 10}, 1e-6, 3e-4)
-  # NOTE: big weight_decay so a missing wd would be way over atol
-  def test_muon_wd(self): self._test_muon(1, {'lr': 0.001, 'weight_decay': 10}, 1e-3, 3e-4)
+  def test_muon_wd(self): self._test_muon(1, {'lr': 0.001, 'weight_decay': 0.01}, 1e-3, 3e-4)
  # TODO: disabled due to big atol
  # def test_muon_high_lr_wd(self): self._test_muon(1, {'lr': 10, 'weight_decay': 0.01}, 1e-6, 5e-4)

@@ -2,8 +2,7 @@ import unittest, pickle, types, tracemalloc
 import numpy as np
 from tinygrad import Tensor, Device, TinyJit, Variable, dtypes
 from tinygrad.helpers import GlobalCounters, ContextVar, Context, DEV
-from tinygrad.uop.ops import PatternMatcher, UPat, UOp, deconstruct_function
-from test.helpers import KernelCountException
+from tinygrad.uop.ops import PatternMatcher, UPat, UOp

 class TestPickle(unittest.TestCase):
  def test_pickle_code_object(self):
@@ -12,11 +11,6 @@ class TestPickle(unittest.TestCase):
    fxn = types.FunctionType(pickle.loads(code_str), globals())
    self.assertEqual(fxn(2), 4)

-  def test_deconstruct_function_nested_comprehension(self):
-    # pre PEP 709, each comprehension is its own code object, so dtypes here is referenced two code objects deep
-    def fxn(): return [[dtypes.int for _ in range(2)] for _ in range(2)]
-    self.assertEqual(types.FunctionType(*deconstruct_function(fxn))(), fxn())
-
  def test_pickle_pattern_matcher(self):
    pm = PatternMatcher([(UPat.cvar('x'), lambda x: x*2)])
    sink = UOp.const(2)
@@ -42,7 +36,7 @@ class TestPickle(unittest.TestCase):
    t2:Tensor = pickle.loads(st)
    np.testing.assert_equal(t_values, t2.numpy())
    # expect at most one COPY kernel
-    if GlobalCounters.kernel_count > 1: raise KernelCountException(1, GlobalCounters.kernel_count)
+    self.assertLessEqual(GlobalCounters.kernel_count, 1)

  def test_pickle_realized_tensor_alt(self):
    print("** init")
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	a0d4a2971b	kimi linear working	2026-08-10 15:37:20 +00:00
geohot	951aaf893b	llm: reuse exact recurrent prefixes	2026-08-10 14:06:15 +00:00
geohot	a455e17539	llm: fully warm recurrent serving at startup	2026-08-10 14:06:15 +00:00
nimlgenandgeohot	778f8aee59	fix hevc (#17477 ) * hevc tests * x	2026-08-10 14:06:15 +00:00
qazalandgeohot	9b347cc3e7	late loss.to("CPU") in llama (#17476 ) * late loss.to("CPU") in llama * acc = 0	2026-08-10 14:06:15 +00:00
qazalandgeohot	32b9149040	llama: custom silu kernels (#17462 ) * start by copying the C * uop kernel * cleanup tests * estimates is part of SPEC	2026-08-10 14:06:15 +00:00
George Hotzandgeohot	bbd4a77351	move platform tests to platform.yml (#17475 ) * ci: split mac/windows/qcom-cl tests into platform.yml Move the 6 jobs that don't run on Linux (4 macos, 1 windows, 1 QCOM CL compile test on arm) out of test.yml into a separate Platform Tests workflow so they run (and can be gated/runners-matched) independently. * ci: gate platform tests to the upstream repo Skip mac/windows/qcom-cl jobs anywhere but tinygrad/tinygrad, so the Platform Tests workflow is disabled on the gitea fork (and any fork). * ci: revert repo gate on platform tests Job-level if is only evaluated by gitea when a runner with matching labels fetches the task; with no mac/windows/arm runners the jobs queue forever. Disable the workflow on the instance instead.	2026-08-10 14:06:15 +00:00
geohot	e103421a12	document Kimi K3 optimization handoff	2026-08-10 06:14:30 -07:00
geohot	2b1b8c22a9	add fast exact-shape Kimi K3 benchmark	2026-08-10 06:13:24 -07:00
geohot	f53f0e7e79	document final Kimi K3 MI350 load profiling	2026-08-10 06:13:24 -07:00
geohot	224bac0318	speed up Kimi K3 decode projections on MI350X	2026-08-10 06:13:24 -07:00
geohot	1d86204718	optimize Kimi K3 serving on MI350X	2026-08-10 06:13:24 -07:00
geohot	c6ac4961d7	speed up Kimi K3 safetensor loading	2026-08-10 06:13:24 -07:00
geohot	1b3732a6ed	speed up Kimi K3 TP8 loading on MI350X	2026-08-10 06:13:24 -07:00
geohot	553bdf68e6	llm: defer recurrent server warmup	2026-08-10 06:59:49 +00:00
geohot	4e1c0166f8	llm: stabilize recurrent serving across requests	2026-08-10 06:43:27 +00:00
geohot	d28f5f261b	llm: keep recurrent single-token shapes static	2026-08-10 05:45:20 +00:00
geohot	14595b9ae8	llm: prepare direct Kimi K3 serving on MI350X	2026-08-10 05:28:18 +00:00
geohot	eaf7822239	llm: accelerate Kimi serving on gfx11	2026-08-10 03:58:44 +00:00
geohot	77e5be99bc	llm: prepare Kimi K3 and accelerate recurrent prefill	2026-08-10 02:03:14 +00:00
geohot	6edb5f9698	get kimi-linear running on 4x7900XTX (codex slop)	2026-08-10 00:38:38 +00:00