fix flaky

* remove pm_render

* move merge_reduce_ends

* ext devectorize

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

@@ -10,7 +10,7 @@ inputs:
     required: false
     default: ''  # if you don't set a key, it doesn't cache
   deps:
-    description: 'Extra dependency groups (comma separated)'
+    description: 'Extra dependency groups (space separated)'
     required: false
     default: ''
   pydeps:
@@ -41,10 +41,6 @@ inputs:
     description: "Install LLVM?"
     required: false
     default: 'false'
-  mesa:
-    description: "Install mesa (true, false, cpu)"
-    required: false
-    default: 'false'
   tinydreno:
     description: "Install tinydreno"
     required: false
@@ -80,7 +76,7 @@ runs:
     - name: Cache Python packages (PR)
       if: github.event_name == 'pull_request'
       id: restore-venv-pr
-      uses: actions/cache/restore@v4
+      uses: actions/cache/restore@v5
       with:
         path: /tmp/.uv-cache
         key: uv-${{ runner.os }}-${{ runner.arch }}-python-${{ inputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
@@ -96,7 +92,7 @@ runs:
 
     - name: Cache downloads (PR)
       if: inputs.key != '' && github.event_name == 'pull_request'
-      uses: actions/cache/restore@v4
+      uses: actions/cache/restore@v5
       with:
         path: ${{ runner.os == 'Linux' && '~/.cache/tinygrad/downloads/' || '~/Library/Caches/tinygrad/downloads/' }}
         key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
@@ -114,7 +110,8 @@ runs:
       shell: bash
       run: |
         uv venv .venv
-        uv pip install --python .venv -e ".[${{ inputs.deps }}]" ${{ inputs.pydeps }} --torch-backend cpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/
+        DEPS="${{ inputs.deps }}"
+        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
@@ -146,11 +143,6 @@ runs:
         echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
         echo 'Binary::apt::APT::Keep-Downloaded-Packages "true";' | sudo tee -a /etc/apt/apt.conf.d/99keep-debs
 
-    - name: Add OpenCL Repo
-      if: inputs.opencl == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: echo "deb [ allow-insecure=yes ] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneAPI.list
-
     - name: Add AMD Repo (Linux)
       if: inputs.amd == 'true' && runner.os == 'Linux'
       shell: bash
@@ -176,10 +168,7 @@ runs:
         pkgs=""
         # **** OpenCL ****
         if [[ "${{ inputs.opencl }}" == "true" ]]; then
-          pkgs+=" opencl-headers \
-            intel-oneapi-runtime-openmp=2023.2.1-16 intel-oneapi-runtime-compilers-common=2023.2.1-16 intel-oneapi-runtime-compilers=2023.2.1-16 \
-            intel-oneapi-runtime-dpcpp-sycl-opencl-cpu=2023.2.1-16 intel-oneapi-runtime-tbb-common=2021.10.0-49541 \
-            intel-oneapi-runtime-tbb=2021.10.0-49541 intel-oneapi-runtime-opencl=2023.2.1-16"
+          pkgs+=" ocl-icd-opencl-dev"
         fi
         # **** AMD ****
         if [[ "${{ inputs.amd }}" == "true" ]]; then
@@ -203,7 +192,7 @@ runs:
 
     - name: Cache apt (PR)
       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@v4
+      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 }}
@@ -286,18 +275,18 @@ runs:
       shell: bash
       run: brew install llvm@20
 
-    # **** mesa ****
-    - name: Install mesa (linux)
-      if: inputs.mesa != 'false' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinymesa/releases/download/v1/libtinymesa${{ inputs.mesa == 'cpu' && '_cpu' || '' }}-mesa-25.2.7-linux-amd64.so -o /usr/lib/libtinymesa${{ inputs.mesa == 'cpu' && '_cpu' || '' }}.so
-    - name: Install mesa (macOS)
-      if: inputs.mesa != 'false' && runner.os == 'macOS'
-      shell: bash
-      run: brew install sirhcm/tinymesa/tinymesa${{ inputs.mesa == 'cpu' && '_cpu' || '' }}
-
     # *** tinydreno ***
     - name: Install tinydreno (linux)
       if: inputs.tinydreno == 'true' && runner.os == 'Linux'
       shell: bash
       run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
+
+    # *** OpenCL ***
+    - name: Install rusticl
+      if: inputs.opencl == 'true'
+      shell: bash
+      run: |
+        sudo curl -fL https://github.com/sirhcm/tinymesa/releases/download/rusticl-v1/libRusticlOpenCL.so.1.0.0 -o /usr/lib/libRusticlOpenCL.so
+        sudo mkdir -p /etc/OpenCL/vendors
+        echo "/usr/lib/libRusticlOpenCL.so" | sudo tee /etc/OpenCL/vendors/rusticl.icd
+        echo "RUSTICL_ENABLE=llvmpipe" >> "$GITHUB_ENV"

.github/workflows/benchmark.yml

@@ -81,103 +81,285 @@ jobs:
   #      source /tmp/tinygrad_pytest_ci/bin/activate
   #      pytest -nauto --durations=20
 
-  testmacbenchmark:
-    name: Mac Benchmark
-    runs-on: [self-hosted, macOS]
+  llmbenchmark:
+    name: LLM (DEV=${{ matrix.dev }})
+    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['METAL', 'AMD', 'NV']
     timeout-minutes: 60
     defaults:
       run:
         shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
     if: github.repository_owner == 'tinygrad'
     steps:
     - name: Checkout Code
       uses: actions/checkout@v6
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
+      run: |
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py amd rmmod
+        ./extra/hcq/hcq_smi.py amd kill_pids
     - name: Symlink models and datasets
       run: |
         mkdir -p weights
-        mkdir -p extra/disassemblers
-        ln -s ~/tinygrad/extra/disassemblers/applegpu extra/disassemblers/applegpu
-        ln -s ~/tinygrad/weights/sd-v1-4.ckpt weights/sd-v1-4.ckpt
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
+        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
     - 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: python3.11 test/external/process_replay/reset.py
-    - name: Print macOS version
-      run: sw_vers
+      run: python3 test/external/process_replay/reset.py
+    - name: Run llama3.2
+      run: BENCHMARK_LOG=llama32_3b-f16 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m llama3.2:3b-f16 --benchmark --warmup
+    - name: Run qwen3.5
+      # qwen3.5:35b-a3b doesn't fit on mac
+      if: ${{ matrix.dev != 'METAL' }}
+      run: BENCHMARK_LOG=qwen35_35b-a3b JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 -m tinygrad.llm -m qwen3.5:35b-a3b --benchmark --warmup
+    - 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: ${{ matrix.dev != 'METAL' }}
+      run: BENCHMARK_LOG=llama3_beam_4gpu JITBEAM=2 IGNORE_BEAM_CACHE=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 4 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  cifarbenchmark:
+    name: HLB-CIFAR10 (DEV=${{ matrix.dev }})
+    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['METAL', 'AMD', 'NV']
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
+      run: |
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py amd rmmod
+        ./extra/hcq/hcq_smi.py amd kill_pids
+    - 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 10 CIFAR training steps
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'NV' && '130' || matrix.dev == 'AMD' && '200' || '3000' }}
+      run: BENCHMARK_LOG=cifar_10steps STEPS=10 python3 examples/hlb_cifar10.py
+    - name: Run 10 CIFAR training steps w HALF
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'NV' && '120' || matrix.dev == 'AMD' && '235' || '3000' }}
+      run: BENCHMARK_LOG=cifar_10steps_half STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+    - name: Run full CIFAR training w 1 GPU
+      # 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: ${{ matrix.dev != 'METAL' }}
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  mlperfbenchmark:
+    name: MLPerf (${{ matrix.dev }})
+    runs-on: [self-hosted, Linux, "${{ matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['AMD', 'NV']
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
+      run: |
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py amd rmmod
+        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: Symlink models and datasets
+      run: |
+        mkdir -p extra/datasets
+        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" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: test/external/process_replay/reset.py
+    - name: Run MLPerf resnet eval on training data
+      run: time BENCHMARK_LOG=resnet_eval MODEL=resnet python3 examples/mlperf/model_eval.py
+    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+      run: BENCHMARK_LOG=resnet_10steps DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+      run: BENCHMARK_LOG=resnet_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py
+    - name: Run 10 MLPerf Bert training steps (6 gpu)
+      # TODO: remove BERT_LAYERS once scheduler is fast
+      run: BENCHMARK_LOG=bert_10steps_6gpu CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  sdbenchmark:
+    name: Stable Diffusion (DEV=${{ matrix.dev }})
+    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['METAL', 'AMD', 'NV']
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
+      run: |
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py amd rmmod
+        ./extra/hcq/hcq_smi.py amd kill_pids
+    - 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 Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
-    - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 ASSERT_MIN_STEP_TIME=720 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing
-    - name: Run Stable Diffusion v2
-      # TODO: very slow step time
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=4500 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing
-    # process replay can't capture this, the graph is too large
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'METAL' && '720' || matrix.dev == 'AMD' && '550' || '0' }}
+      run: BENCHMARK_LOG=stable_diffusion python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
     - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=5000 CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing
-    - name: Run model inference benchmark
-      run: DEV=METAL NOCLANG=1 python3.11 test/external/external_model_benchmark.py
-    - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py
+      if: ${{ matrix.dev != 'NV' }}
+      env:
+        ASSERT_MIN_STEP_TIME: ${{ matrix.dev == 'METAL' && '5000' || matrix.dev == 'AMD' && '3200' || '2000' }}
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  tests:
+    name: Tests (DEV=${{ matrix.dev }})
+    runs-on: [self-hosted, "${{ matrix.dev == 'METAL' && 'macOS' || matrix.dev == 'AMD' && 'tinybox' || 'tinyboxgreen' }}"]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['METAL', 'AMD', 'NV']
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup (AMD)
+      if: ${{ matrix.dev == 'AMD' }}
+      run: |
+        ./extra/amdpci/setup_python_cap.sh
+        ./extra/hcq/hcq_smi.py amd rmmod
+        ./extra/hcq/hcq_smi.py amd kill_pids
+    - name: setup staging db
+      if: github.ref == 'refs/heads/update_benchmark_staging'
+      run: |
+        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
+        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
+    - name: reset process replay
+      run: python3 test/external/process_replay/reset.py
+    - name: Test tiny
+      run: |
+        DEBUG=2 python -m pytest -rA test/test_tiny.py
+        if [[ "${{ matrix.dev }}" == "NV" ]]; then
+          DEBUG=2 DEV=CUDA python -m pytest -rA test/test_tiny.py
+        fi
     - name: Test tensor cores
-      run: DEV=METAL python3.11 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (float)
-      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (half)
-      run: DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3.11 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (bfloat16)
-      run: DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3.11 extra/gemm/simple_matmul.py
-    - name: Fuzz Padded Tensor Core GEMM
-      run: DEV=METAL M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3.11 ./extra/gemm/fuzz_matmul.py
-    - name: Run LLaMA
       run: |
-        BENCHMARK_LOG=llama_nojit JIT=0 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama JIT=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run quantized LLaMA
-      run: |
-        BENCHMARK_LOG=llama_int8 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize int8
-        BENCHMARK_LOG=llama_nf4 python3.11 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing --quantize nf4
-    - name: Run quantized LLaMA3
-      run: |
-        BENCHMARK_LOG=llama3_int8 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize int8
-        BENCHMARK_LOG=llama3_nf4 python3.11 examples/llama3.py --size 8B --temperature 0 --benchmark --quantize nf4
-    #- name: Run LLaMA 7B on 4 (virtual) GPUs
-    #  run: python3.11 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run GPT2
-      run: |
-        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run OLMoE
-      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
+        if [[ "${{ matrix.dev }}" == "METAL" ]]; then
+          python3 test/opt/test_tensor_cores.py
+          DEBUG=2 SHOULD_USE_TC=1 python3 extra/gemm/simple_matmul.py
+          DEBUG=2 SHOULD_USE_TC=1 HALF=1 python3 extra/gemm/simple_matmul.py
+          DEBUG=2 SHOULD_USE_TC=1 BFLOAT16=1 python3 extra/gemm/simple_matmul.py
+          M_START=6 M_STOP=10 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=6 K_STOP=24 K_STEP=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
+        elif [[ "${{ matrix.dev }}" == "NV" ]]; then
+          ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+          DEV=NV:PTX ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+          DEV=CUDA SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          DEV=CUDA SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          DEV=CUDA SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+          DEV=CUDA SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          DEV=NV:PTX SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          # TODO: too slow
+          # M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+          # DEV=NV:PTX M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
+        else
+          python3 test/opt/test_tensor_cores.py
+          # TODO: this is flaky
+          # DEV=AMD:LLVM python3 test/opt/test_tensor_cores.py
+          SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+          # TODO: AMD compiler bug causes this to fail
+          # HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
+        fi
+    - name: Run model inference benchmark
+      # TODO: unstable on AMD
+      if: ${{ matrix.dev != 'AMD' }}
+      run: CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
+    - name: Test speed vs torch
+      # TODO: unstable on AMD
+      if: ${{ matrix.dev != 'AMD' }}
+      env:
+        HALF: ${{ matrix.dev == 'NV' && '1' || '0' }}
+      run: CAPTURE_PROCESS_REPLAY=0 BIG=2 ${{ matrix.dev == 'METAL' && 'MPS=1' || 'TORCHCUDA=1' }} python3 test/speed/external_test_speed_v_torch.py
+    - name: Test speed vs theoretical
+      # no targets for METAL
+      if: ${{ matrix.dev != 'METAL' }}
+      run: IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
     - name: Train MNIST
-      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py
-
-    # NOTE: this is failing in CI. it is not failing on my machine and I don't really have a way to debug it
-    # the error is "RuntimeError: Internal Error (0000000e:Internal Error)"
-    #- name: Run 10 CIFAR training steps
-    #  run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=3000 STEPS=10 python3.11 examples/hlb_cifar10.py
-    #- name: Run 10 CIFAR training steps w HALF
-    #  run: BENCHMARK_LOG=cifar_10steps_half JIT=2 ASSERT_MIN_STEP_TIME=3000 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py
-
-    #- name: Run 10 CIFAR training steps w BF16
-    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py
-    # TODO: too slow
-    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py
+      run: time TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
+    - name: Test benchmark allreduce
+      if: ${{ matrix.dev == 'NV' }}
+      run: python test/external/external_benchmark_multitensor_allreduce.py
+    - 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
     - uses: actions/upload-artifact@v7
+      if: ${{ matrix.dev != 'AMD' }}
       with:
-        name: Speed (Mac)
+        name: Speed (${{ matrix.dev }})
         path: |
           onnx_inference_speed.csv
     - name: Run process replay tests
@@ -219,381 +401,6 @@ jobs:
     - name: UsbGPU (USB4/TB) tiny tests
       run: PYTHONPATH=. DEV=PCI+NV:NAK python3.11 test/test_tiny.py
 
-  testnvidiabenchmark:
-    name: tinybox green Benchmark
-    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Print nvidia-smi
-      run: nvidia-smi
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
-        mkdir -p extra/datasets
-        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: test/external/process_replay/reset.py
-    - name: Run model inference benchmark
-      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
-    - name: Test speed vs torch
-      run: DEV=NV CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
-    - name: Test speed vs theoretical
-      run: DEV=NV IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test benchmark allreduce
-      run: DEV=NV python test/external/external_benchmark_multitensor_allreduce.py
-    - name: Test tensor cores
-      run: |
-        DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-        DEV=NV:PTX ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (CUDA)
-      run: |
-        DEV=CUDA SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        DEV=CUDA SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        DEV=CUDA SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-        DEV=CUDA SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (PTX)
-      run: DEV=NV:PTX SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (NV)
-      run: DEV=NV SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Test DEV=NV
-      run: DEBUG=2 DEV=NV python -m pytest -rA test/test_tiny.py
-    - name: Test DEV=CUDA
-      run: DEBUG=2 DEV=CUDA python -m pytest -rA test/test_tiny.py
-    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion DEV=NV python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
-    # TODO: too slow
-    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing
-    - name: Run LLaMA
-      run: |
-        BENCHMARK_LOG=llama_nojit DEV=NV JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama DEV=NV JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run LLaMA with BEAM
-      run: BENCHMARK_LOG=llama_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
-    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=NV 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 quantized LLaMA3
-      run: BENCHMARK_LOG=llama3_fp8 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --temperature 0 --benchmark --quantize fp8
-    # - name: Run LLaMA-3 8B on 6 GPUs
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    # - name: Run LLaMA-2 70B
-    #   run: DEV=NV CAPTURE_PROCESS_REPLAY=0 MAX_CONTEXT=256 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral DEV=NV CAPTURE_PROCESS_REPLAY=0 python3 examples/mixtral.py --temperature 0 --count 10 --timing
-    - name: Run GPT2
-      run: |
-        BENCHMARK_LOG=gpt2_nojit DEV=NV JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 DEV=NV JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half DEV=NV HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam DEV=NV HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
-    - uses: actions/upload-artifact@v7
-      with:
-        name: Speed (NVIDIA)
-        path: |
-          onnx_inference_speed.csv
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmorenvidiabenchmark:
-    name: tinybox green Training Benchmark
-    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        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: test/external/process_replay/reset.py
-    # TODO: too slow
-    # - name: Fuzz Padded Tensor Core GEMM (NV)
-    #   run: DEV=NV M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
-    # TODO: too slow
-    # - name: Fuzz Padded Tensor Core GEMM (PTX)
-    #   run: DEV=NV:PTX M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
-    - name: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 DEV=NV PYTHONPATH=. python3 extra/hevc/decode.py
-    - name: Train MNIST
-      run: time PYTHONPATH=. DEV=NV TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
-    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=130 DEV=NV STEPS=10 python3 examples/hlb_cifar10.py
-    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 DEV=NV STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
-    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 DEV=NV WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEV=NV 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
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 DEV=NV 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 DEV=NV MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps DEV=NV 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 DEV=NV 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 DEV=NV 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
-
-  testamdbenchmark:
-    name: tinybox red Benchmark
-    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    #- name: Insert amdgpu
-    #  run: sudo modprobe amdgpu
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        ln -s /raid/weights/LLaMA-3 weights/LLaMA-3
-        mkdir -p extra/datasets
-        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: test/external/process_replay/reset.py
-    #- name: setup perflevel
-    #  run: |
-    #    examples/mlperf/training_submission_v4.1/tinycorp/benchmarks/bert/implementations/tinybox_red/setup.sh
-    #    rocm-smi
-    #- name: Show off tinybox
-    #  run: /opt/rocm/bin/rocm-bandwidth-test
-    # TODO: unstable on AMD
-    #- name: Run model inference benchmark
-    #  run: LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 NOCLANG=1 python3 test/external/external_model_benchmark.py
-    # TODO: unstable on AMD
-    #- name: Test speed vs torch
-    #  run: |
-    #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py
-    - name: Test speed vs theoretical
-      run: DEV=AMD IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
-    - name: Test tensor cores (no LLVM)
-      run: DEV=AMD python3 test/opt/test_tensor_cores.py
-    # TODO: this is flaky
-    # - name: Test tensor cores AMD:LLVM
-    #   run: DEV=AMD:LLVM python3 test/opt/test_tensor_cores.py
-    - name: Run Tensor Core GEMM (AMD)
-      run: |
-        DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-        DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-    - name: Test DEV=AMD
-      run: DEBUG=2 DEV=AMD python -m pytest -rA test/test_tiny.py
-    #- name: Test HIP=1
-    #  run: DEBUG=2 HIP=1 python -m pytest -rA test/test_tiny.py
-    # TODO: AMD compiler bug causes this to fail
-    #- name: Fuzz Padded Tensor Core GEMM
-    #  run: HSA=1 M_START=12 M_STOP=20 M_STEP=1 N_START=12 N_STOP=20 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 DEBUG=2 python3 ./extra/gemm/fuzz_matmul.py
-    #- name: Remove amdgpu
-    #  run: sleep 10 && sudo rmmod amdgpu # sleep a bit to let the driver unload the prev pid.
-    - name: Test AM cold start time
-      run: time DEV=AMD AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
-    - name: Test AM warm start time
-      run: time DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
-    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 DEV=AMD python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing
-    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 DEV=AMD python3 examples/sdxl.py --seed 0 --noshow --timing
-    - name: Run LLaMA 7B
-      run: |
-        BENCHMARK_LOG=llama_nojit DEV=AMD JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=llama DEV=AMD JIT=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run LLaMA 7B with BEAM
-      run: BENCHMARK_LOG=llama_beam DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing
-    # - name: Run LLaMA 7B on 4 GPUs
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 10 --temperature 0  --timing
-    # - name: Run LLaMA 7B on 6 GPUs
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 1 --size 7B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run LLaMA-3 8B BEAM
-      run: BENCHMARK_LOG=llama3_beam DEV=AMD JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
-      run: BENCHMARK_LOG=llama3_beam_4gpu DEV=AMD 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 LLaMA-3 8B on 6 GPUs
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0
-    #- name: Restore amdgpu
-    #  run: sudo modprobe amdgpu
-    # - name: Run LLaMA-2 70B
-    #   run: DEV=AMD CAPTURE_PROCESS_REPLAY=0 python3 examples/llama.py --gen 2 --size 70B --shard 6 --prompt "Hello." --count 10 --temperature 0  --timing
-    - name: Run Mixtral 8x7B
-      run: time BENCHMARK_LOG=mixtral DEV=AMD python3 examples/mixtral.py --temperature 0 --count 10 --timing
-    - name: Run GPT2
-      run: |
-        BENCHMARK_LOG=gpt2_nojit DEV=AMD JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-        BENCHMARK_LOG=gpt2 DEV=AMD JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half DEV=AMD HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run GPT2 w HALF/BEAM
-      run: BENCHMARK_LOG=gpt2_half_beam DEV=AMD HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmoreamdbenchmark:
-    name: tinybox red Training Benchmark
-    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        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: test/external/process_replay/reset.py
-    - name: Test GPU crash recovery
-      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
-    - name: Train MNIST
-      run: time PYTHONPATH=. DEV=AMD TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
-    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 DEV=AMD STEPS=10 python3 examples/hlb_cifar10.py
-    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 DEV=AMD STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    # - name: Run 10 CIFAR training steps w BF16
-    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 DEV=AMD STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
-    # TODO: too slow
-    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 DEV=AMD WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEV=AMD 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
-      run: time BENCHMARK_LOG=cifar_6gpu DEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # TODO: broken on some of the machines
-    #- name: Test full tinyfs load
-    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmlperfamdbenchmark:
-    name: tinybox red MLPerf Benchmark
-    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        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: test/external/process_replay/reset.py
-    - name: Run MLPerf resnet eval
-      run: time BENCHMARK_LOG=resnet_eval DEV=AMD MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps DEV=AMD 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 DEV=AMD 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 DEV=AMD 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
-
   testcommalatest:
     name: comma Benchmark (0.11.0)
     runs-on: [self-hosted, Linux, comma]
@@ -617,7 +424,7 @@ jobs:
     - 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=17 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
+      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
@@ -715,31 +522,28 @@ jobs:
     - name: openpilot run_pickle 0.10.1 driving_vision
       run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_run_pickle RUN_PICKLE=1 PYTHONPATH="." GMMU=0 DEV=USB+AMD ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py
 
-  testreddriverbenchmark:
-    name: AM Benchmark
+  driverbenchmarks:
+    name: PCI Driver Benchmark (DEV=${{ matrix.dev }})
     runs-on: [self-hosted, Linux, tinyboxrandom]
+    strategy:
+      fail-fast: false
+      matrix:
+        dev: ['AMD', 'NV']
     timeout-minutes: 20
     defaults:
       run:
         shell: bash -e -o pipefail {0}
+    env:
+      DEV: ${{ matrix.dev }}
     if: github.repository_owner == 'tinygrad'
     steps:
     - name: Checkout Code
       uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove amd modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Symlink models and datasets
+    - name: Setup
       run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
+        ./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
@@ -750,104 +554,45 @@ jobs:
     - name: reset process replay
       run: test/external/process_replay/reset.py
     - name: Test driver cold start time
-      run: time DEBUG=3 DEV=AMD AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
+      run: time DEBUG=3 AM_RESET=1 python3 test/test_tiny.py TestTiny.test_plus
     - name: Test driver warm start time
-      run: time DEBUG=3 DEV=AMD python3 test/test_tiny.py TestTiny.test_plus
+      if: ${{ matrix.dev == 'AMD' }}
+      run: time DEBUG=3 python3 test/test_tiny.py TestTiny.test_plus
     - name: Test GPU crash recovery
-      run: DEV=AMD python3 -m pytest -rA test/external/external_test_gpu_crash.py
-    # Fails on 9070
-    # - name: Test tensor cores
-    #   run: |
-    #     DEV=AMD python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     DEV=AMD:LLVM python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     DEV=AMD SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
-    - name: Run Tensor Core GEMM (AMD)
-      run: DEV=AMD SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
-    - name: Test DEV=AMD
-      run: DEBUG=2 DEV=AMD python -m pytest -rA test/test_tiny.py
+      if: ${{ matrix.dev == 'AMD' }}
+      run: python3 -m pytest -rA test/external/external_test_gpu_crash.py
+    - name: Test tensor cores
+      run: |
+        if [[ "${{ matrix.dev }}" == "AMD" ]]; then
+          # Fails on 9070
+          # python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+          # DEV=AMD:LLVM python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+          # SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
+          SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py
+        else
+          ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+        fi
     - name: Test DISK copy time
-      run: DEV=AMD TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
+      run: TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
     - name: Test CPU copy time
       run: |
-        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        DEV=AMD GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEV=AMD DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-    #   run: BENCHMARK_LOG=resnet_10steps DEV=AMD MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
+        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: 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
     - name: Run 10 MLPerf Bert training steps (1 gpu)
       # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps DEV=AMD CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: BENCHMARK_LOG=bert_10steps CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
     - name: Remote
       run: |
         pkill -f 'extra/remote/serve.py' || true
         PYTHONPATH=. python3 extra/remote/serve.py 6482 &
         sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD python3 test/test_tiny.py
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 DEV=PCI+AMD AMD_AQL=1 python3 test/test_tiny.py
-        pkill -f 'extra/remote/serve.py' || true
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testgreendriverbenchmark:
-    name: NV Benchmark
-    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setcap to python
-      run: ./extra/amdpci/setup_python_cap.sh
-    - name: Remove nv modules
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv rmmod
-    - name: Kill stale pids
-      run: PYTHONPATH=. ./extra/hcq/hcq_smi.py nv kill_pids
-    - name: Symlink models and datasets
-      run: |
-        mkdir -p weights
-        ln -s ~/tinygrad/weights/bpe_simple_vocab_16e6.txt.gz weights/bpe_simple_vocab_16e6.txt.gz
-        ln -s ~/tinygrad/weights/LLaMA weights/LLaMA
-        ln -s ~/tinygrad/extra/datasets/cifar-10-python.tar.gz extra/datasets/cifar-10-python.tar.gz
-        ln -s /raid/weights/mixtral-8x7b-32kseqlen weights/mixtral-8x7b-32kseqlen
-        ln -s /raid/weights/LLaMA-2 weights/LLaMA-2
-        mkdir -p extra/datasets
-        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: test/external/process_replay/reset.py
-    - name: Test driver start time
-      run: time DEBUG=3 DEV=NV python3 test/test_tiny.py TestTiny.test_plus
-    - name: Test tensor cores
-      run: DEV=NV ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-    - name: Test DISK copy time
-      run: DEV=NV TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
-    - name: Test CPU copy time
-      run: |
-        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        DEV=NV GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
-    - name: Test LLAMA-3
-      run: BENCHMARK_LOG=llama3_beam DEV=NV JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0
-    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar DEV=NV DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps DEV=NV MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py
-    - name: Run 10 MLPerf Bert training steps (1 gpu)
-      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps DEV=NV CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
-    - name: Remote
-      run: |
-        pkill -f 'extra/remote/serve.py' || true
-        PYTHONPATH=. python3 extra/remote/serve.py 6483 &
-        sleep 1
-        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6483 DEV=NV python3 test/test_tiny.py
+        DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 python3 test/test_tiny.py
+        if [[ "${{ matrix.dev }}" == "AMD" ]]; then
+          DEBUG=2 PYTHONPATH=. REMOTE=127.0.0.1:6482 AM_RESET=1 AMD_AQL=1 python3 test/test_tiny.py
+        fi
         pkill -f 'extra/remote/serve.py' || true
     - name: Run process replay tests
       uses: ./.github/actions/process-replay

.github/workflows/mlperf.yml

@@ -1,8 +1,8 @@
 name: Run MLPerf Training
 
 on:
-  schedule:
-    - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
+  #schedule:
+  #  - cron: '5 8 * * *'  # Runs at 08:05 UTC (12:05 AM Pacific Time)
   push:
     branches:
       - update_mlperf

.github/workflows/test.yml

@@ -77,45 +77,14 @@ jobs:
       run: |
         sudo apt update || true
         sudo apt install -y --no-install-recommends ninja-build
-    - name: Test one op
-      run: FORWARD_ONLY=1 TINY_BACKEND=1 python3 test/test_tiny.py TestTiny.test_plus
     - name: Test ResNet-18
       run: DEBUG=2 python3 extra/torch_backend/example.py
-    - name: custom tests
-      run: python3 -m pytest -n auto extra/torch_backend/test.py --durations=20
     - name: Test one op in torch tests
       run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
     - name: Test Ops with TINY_BACKEND
       run: DEV=CPU:LLVM LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/backend/test_ops.py --durations=20
-    - name: Test in-place operations on views
-      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
-    - name: Test multi-gpu
-      run: DEV=CPU:LLVM GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
-    - name: Test kernel fusion
-      run: python3 extra/torch_backend/test_kernel_fusion.py
-
-
-  torchbackendmore:
-    name: Torch Backend Tests More
-    runs-on: *linux
-    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'
-    - name: Install ninja
-      run: |
-        sudo apt update || true
-        sudo apt install -y --no-install-recommends ninja-build
-    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 DEV=CPU TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
-    - name: Test some torch tests (expect failure)
-      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true
+    - 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
 
   bepython:
     name: Python Backend
@@ -133,46 +102,26 @@ jobs:
       run: SKIP_SLOW_TEST=1 DEV=PYTHON python3 -m pytest -n=auto test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_ops.py test/backend/test_uops.py test/backend/test_symbolic_ops.py test/backend/test_renderer_failures.py::TestRendererFailures --durations=20
     - name: Test IMAGE support
       run: IMAGE=1 DEV=PYTHON python3 test/backend/test_ops.py TestOps.test_gemm TestOps.test_simple_conv2d
-    - name: Test emulated METAL tensor cores
+    - name: Test emulated tensor cores
       env:
-        DEV: 'PYTHON::METAL'
+        DEBUG: 2
+        N: 64
+        CNT: 1
+        SHOULD_USE_TC: 1
       run: |
-        DEBUG=2 python3 test/backend/test_ops.py TestOps.test_big_gemm
-        python3 -m pytest -nauto test/opt/test_tensor_cores.py
-    - name: Test emulated AMD tensor cores
-      env:
-        DEV: 'PYTHON::gfx1100'
+        parallel -k --link --tagstring '[{1}]' '{2} python3 ./extra/gemm/simple_matmul.py' \
+          ::: metal gfx950 gfx1100 gfx1100_acchalf gfx1201 gfx1201_acchalf sm_75 sm_80_half sm_80_tf32 \
+          ::: 'DEV=PYTHON::METAL' 'DEV=PYTHON::gfx950 HALF=1 ACC_HALF=0' \
+              'DEV=PYTHON::gfx1100 HALF=1 ACC_HALF=0' 'DEV=PYTHON::gfx1100 HALF=1 ACC_HALF=1 ATOL=1e-3' \
+              'DEV=PYTHON::gfx1201 HALF=1 ACC_HALF=0' 'DEV=PYTHON::gfx1201 HALF=1 ACC_HALF=1 ATOL=1e-3' \
+              'DEV=PYTHON::sm_75 HALF=1' 'DEV=PYTHON::sm_80 HALF=1' 'DEV=PYTHON::sm_80 ALLOW_TF32=1'
+    - name: Run additional tensor core tests
       run: |
-        DEBUG=2 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        python3 -m pytest -nauto test/opt/test_tensor_cores.py
-    - name: Test emulated AMD MFMA tensor cores
-      env:
-        DEV: 'PYTHON::gfx950'
-      run: |
-        DEBUG=2 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        python3 -m pytest -nauto test/opt/test_tensor_cores.py
-    - name: Test emulated AMD RDNA4 tensor cores
-      env:
-        DEV: 'PYTHON::gfx1201'
-      run: |
-        DEBUG=2 N=16 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 N=64 HALF=1 ACC_HALF=0 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 N=16 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        DEBUG=2 N=64 HALF=1 ACC_HALF=1 ATOL=1e-3 python3 ./extra/gemm/simple_matmul.py
-        python3 -m pytest -nauto test/opt/test_tensor_cores.py
-    - name: Test emulated CUDA tensor cores
-      run: |
-        DEBUG=2 DEV=PYTHON::sm_80 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 ALLOW_TF32=1 DEV=PYTHON::sm_80 python3 test/backend/test_ops.py TestOps.test_gemm
-        DEBUG=2 DEV=PYTHON::sm_75 python3 test/backend/test_ops.py TestOps.test_gemm_fp16
+        DEV=PYTHON::METAL python3 -m pytest -nauto test/opt/test_tensor_cores.py test/null/test_uops_stats.py::TestUOpsStatsMatmulHalf
+        DEV=PYTHON::gfx1100 python3 -m pytest -nauto test/opt/test_tensor_cores.py test/null/test_uops_stats.py::TestUOpsStatsMatmulHalf
+        DEV=PYTHON::gfx950 python3 -m pytest -nauto test/opt/test_tensor_cores.py
+        DEV=PYTHON::gfx1201 python3 -m pytest -nauto test/opt/test_tensor_cores.py
         ALLOW_TF32=1 DEV=PYTHON::sm_89 python3 -m pytest -nauto test/opt/test_tensor_cores.py
-    - name: Test device flop counts
-      run: |
-        DEBUG=2 DEV=PYTHON::METAL python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
-        DEBUG=2 DEV=PYTHON::gfx1100 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
         DEBUG=2 DEV=PYTHON::sm_80 python3 ./test/null/test_uops_stats.py TestUOpsStatsMatmulHalf
 
   linter:
@@ -218,14 +167,15 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
         deps: testing_unit
         llvm: 'true'
         amd: 'true'
     - name: Run NULL backend tests
       run: DEV=NULL python -m pytest -n=auto test/null/ --durations=20
     - name: Run targeted tests on NULL backend
-      run: DEV=NULL python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+      run: |
+        DEV=NULL python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+        DEV=NULL VIZ=1 python3 -m pytest -n=auto test/null/test_viz.py
     # TODO: too slow
     # - name: Run SDXL on NULL backend
     #   run: DEV=NULL DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
@@ -249,7 +199,7 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
+        pydeps: "pre-commit"
         deps: testing_unit
         llvm: 'true'
     - name: Run pre-commit test hooks
@@ -266,13 +216,6 @@ jobs:
       run: python3 test/external/external_benchmark_schedule.py
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
-    - name: Regen dataset on test_tiny
-      run: |
-        test/external/process_replay/reset.py
-        CAPTURE_PROCESS_REPLAY=1 python test/test_tiny.py TestTiny.test_plus
-        python extra/optimization/extract_dataset.py
-        gzip -c /tmp/sops > extra/datasets/sops.gz
-        #DEBUG=1 MIN_ASTS=1 python extra/optimization/get_action_space.py
     - name: Repo line count < 25000 lines
       run: MAX_LINE_COUNT=25000 python sz.py
 
@@ -308,14 +251,9 @@ jobs:
       with:
         key: fuzzing-unit
         deps: testing_unit
-    - name: Fuzz Test symbolic
-      run: python test/external/fuzz_symbolic.py
-    - name: Fuzz Test symbolic (symbolic divisors)
-      run: python test/external/fuzz_symbolic_symbolic_div.py
-    - name: Fuzz Test fast idiv
-      run: python test/external/fuzz_fast_idiv.py
-    - name: Fuzz Test shape ops
-      run: python test/external/fuzz_shape_ops.py
+    - name: Fuzz Tests
+      run: |
+        parallel --tagstring '[{}]' 'python test/external/fuzz_{}.py' ::: symbolic symbolic_div fast_idiv shape_ops
 
   testopenclimage:
     name: CL IMAGE Tests
@@ -337,31 +275,6 @@ jobs:
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
-  testgpumisc:
-    name: CL Misc tests
-    runs-on: *linux
-    timeout-minutes: 10
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: gen-dataset
-          deps: testing
-          opencl: 'true'
-      - name: Generate Dataset
-        run: DEV=CL extra/optimization/generate_dataset.sh
-      - name: Run Kernel Count Test
-        run: DEV=CL python -m pytest -n=auto test/external/external_test_opt.py
-      - name: Run fused optimizer tests
-        run: DEV=CL FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/backend/test_optim.py -k "not muon"
-      - name: Upload artifact
-        uses: actions/upload-artifact@v7
-        with:
-          name: sops.gz
-          path: /tmp/sops.gz
-
   testopenpilot:
     name: openpilot Compile Tests
     runs-on: *linux
@@ -378,7 +291,8 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1468 ALLOWED_GATED_READ_IMAGE=10 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=55 FLOAT16=1 DEV="CL::IMAGE_PITCH_ALIGNMENT=64" IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          # IMAGE_PITCH_ALIGNMENT=64 matches adreno 630
       - name: Test openpilot CL compile fp32 (test correctness)
         run: |
           DEV=CL IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
@@ -421,7 +335,6 @@ jobs:
         with:
           key: optim
           deps: testing
-          pydeps: "tensorflow==2.19"
           opencl: 'true'
       #- name: Test Optimization Helpers
       #  run: DEBUG=1 python3 extra/optimization/test_helpers.py
@@ -430,7 +343,7 @@ jobs:
       - name: Test Beam Search
         run: DEV=CL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
       - name: Test MLPerf stuff
-        run: DEV=CL python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_lr_schedule.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
       - name: DEV=NULL beautiful_mnist_multigpu
         run: DEV=NULL NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
       - name: Test Bert training
@@ -468,7 +381,7 @@ jobs:
 # ****** Models Tests ******
 
   testmodels:
-    name: Models (llvm+cpu+gpu)
+    name: Models
     runs-on: *linux
     timeout-minutes: 15
     steps:
@@ -479,34 +392,12 @@ jobs:
         with:
           key: models
           deps: testing
-          opencl: 'true'
           llvm: 'true'
-      - name: Test models (llvm)
-        run: DEV=CPU:LLVM python -m pytest -n=auto test/models --durations=20
-      - name: Test models (opencl)
-        run: DEV=CL python -m pytest -n=auto test/models --durations=20
       - name: Test models (cpu)
         run: DEV=CPU python -m pytest -n=auto test/models --durations=20
       - name: Run process replay tests
         uses: ./.github/actions/process-replay
 
-  testmetalmodels:
-    name: Models (metal)
-    runs-on: &macos macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: metal
-        deps: testing
-    - name: Test models (Metal)
-      run: DEV=METAL python -m pytest -n=auto test/models --durations=20
-    - name: Test LLaMA compile speed
-      run: DEV=METAL python test/external/external_test_speed_llama.py
-
 # ****** Feature Tests ******
 
   testdsp:
@@ -548,9 +439,8 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: linux-${{ matrix.dev }}
-        deps: testing_unit
+        deps: "testing_unit${{ contains(matrix.dev, 'LVP') && ' mesa' || '' }}"
         llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') || contains(matrix.dev, 'CLANG') }}
-        mesa: ${{ contains(matrix.dev, 'LVP') && 'cpu' || 'false' }}
         webgpu: ${{ matrix.dev == 'WEBGPU' }}
         opencl: ${{ matrix.dev == 'CL' }}
     - name: Set env
@@ -602,7 +492,7 @@ jobs:
           AMD: 0
         run: |
           PYTHONPATH=. DEV=NULL:HIP:gfx1100 python extra/mmapeak/mmapeak.py
-          PYTHONPATH=. DEV=NULL:HIP:gfx950 python3 -m pytest -n=auto test/testextra/test_tk.py test/backend/test_asm_gemm.py
+          PYTHONPATH=. DEV=NULL:HIP:gfx950 python3 -m pytest -n=auto test/testextra/test_tk.py
       - name: Run matmul on MOCKKFD
         run: |
           PYTHONPATH="." DEV=MOCKKFD+AMD N=256 python3 extra/gemm/amd_asm_matmul.py
@@ -715,7 +605,7 @@ jobs:
 
   unittestmacos:
     name: MacOS (unit)
-    runs-on: *macos
+    runs-on: &macos macos-26
     timeout-minutes: 20
     steps:
     - name: Checkout Code
@@ -758,6 +648,33 @@ jobs:
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
+  testmetal:
+    strategy:
+      fail-fast: false
+      matrix:
+        group: [1, 2]
+    name: MacOS (DEV=METAL) (${{ matrix.group }})
+    runs-on: *macos
+    timeout-minutes: 20
+    env:
+      DEV: METAL
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: macos-metal
+        deps: testing_unit
+    - name: Check Device.DEFAULT and print some source
+      run: |
+        python -c "from tinygrad import Device; assert Device.DEFAULT == 'METAL'"
+        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+    - name: Run backend tests
+      run: python -m pytest -n=auto test/backend --durations=20 --splits 2 --group ${{ matrix.group }}
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
   testmacos:
     strategy:
       fail-fast: false
@@ -766,7 +683,6 @@ jobs:
           - 'CPU:CLANG'
           - 'CPU:LLVM'
           - 'CPU:LVP'
-          - 'METAL'
           - 'WEBGPU'
 
     name: MacOS (DEV=${{ matrix.dev }})
@@ -779,9 +695,8 @@ jobs:
       uses: ./.github/actions/setup-tinygrad
       with:
         key: macos-${{ matrix.dev }}
-        deps: testing_unit
+        deps: "testing_unit${{ contains(matrix.dev, 'LVP') && ' mesa' || '' }}"
         llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') }}
-        mesa: ${{ contains(matrix.dev, 'LVP') && 'cpu' || 'false' }}
         webgpu: ${{ matrix.dev == 'WEBGPU' }}
     - name: Set env
       run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
@@ -789,8 +704,8 @@ jobs:
       run: |
         python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
         DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run backend tests
-      run: python -m pytest -n=auto test/backend --durations=20
+    - name: Run test_tiny
+      run: python -m pytest -n=auto test/test_tiny.py --durations=20
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 
@@ -847,8 +762,7 @@ jobs:
         uses: ./.github/actions/setup-tinygrad
         with:
           key: compile-${{ matrix.backend }}
-          deps: testing_unit
-          mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
+          deps: "testing_unit mesa"
       - name: Set env
         shell: bash
         run: printf "NULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'DEV=NULL:IR3:a630' || matrix.backend == 'nak' && 'DEV=NULL:NAK:sm_120' }}" >> $GITHUB_ENV

README.md

@@ -72,7 +72,7 @@ As it turns out, 90% of what you need for neural networks are a decent autograd/
 Throw in an optimizer, a data loader, and some compute, and you have all you need.
 
 ```python
-from tinygrad import Tensor, nn
+from tinygrad import Tensor, nn, Context
 
 class LinearNet:
   def __init__(self):
@@ -86,7 +86,7 @@ optim = nn.optim.Adam([model.l1, model.l2], lr=0.001)
 
 x, y = Tensor.rand(4, 1, 28, 28), Tensor([2,4,3,7])  # replace with real mnist dataloader
 
-with Tensor.train():
+with Context(TRAINING=1):
   for i in range(10):
     optim.zero_grad()
     loss = model(x).sparse_categorical_crossentropy(y).backward()

docs/abstractions3.py

@@ -11,7 +11,7 @@ X_train -= X_train.mean()
 # *****
 # 1. Define an MNIST model.
 
-from tinygrad import Tensor
+from tinygrad import Tensor, Context
 
 l1 = Tensor.kaiming_uniform(128, 784)
 l2 = Tensor.kaiming_uniform(10, 128)
@@ -24,11 +24,11 @@ l1n, l2n = l1.numpy(), l2.numpy()
 from tinygrad.nn.optim import SGD
 optim = SGD([l1, l2])
 
-Tensor.training = True
-X, Y = X_train[(samples:=Tensor.randint(128, high=X_train.shape[0]))], Y_train[samples]
-optim.zero_grad()
-model(X).sparse_categorical_crossentropy(Y).backward()
-optim.schedule_step()   # this will step the optimizer without running realize
+with Context(TRAINING=1):
+  X, Y = X_train[(samples:=Tensor.randint(128, high=X_train.shape[0]))], Y_train[samples]
+  optim.zero_grad()
+  model(X).sparse_categorical_crossentropy(Y).backward()
+  optim.schedule_step()   # this will step the optimizer without running realize
 
 # *****
 # 3. Create a schedule (linear uop).

docs/abstractions4.py

@@ -67,8 +67,7 @@ def example_2_hip(a:Tensor, correct):
     # the sink specifies the GLOBAL and LOCAL sizes, along with the input buffers and name
     sink = UOp.sink(UOp.special(GLOBALS, 'gidx0'), UOp.special(THREADS, 'lidx0'), out, buf,
                     arg=KernelInfo(name="hip_reduce_sum_kernel"))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
-                UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
   eval_harness("HIP kernel", a, lambda x: Tensor.empty(GLOBALS).custom_kernel(x, fxn=hip_reduce_sum)[0].sum(), check=correct)
 
 def example_3_custom_uop(a:Tensor, correct):
@@ -123,8 +122,7 @@ def example_5_custom_assembly(a:Tensor, correct):
         offset_dwords = (self.labels[inst._target] - inst._pos - inst.size()) // 4
         if not -32768 <= offset_dwords <= 32767: raise ValueError(f"branch to '{inst._target}' offset {offset_dwords} exceeds simm16 range")
         inst.simm16 = offset_dwords
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT),
-                                   UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in self.instructions]))))
+      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in self.instructions]))))
 
   CU_COUNT = 32
   LANES = 64

docs/mnist.md

@@ -24,7 +24,7 @@ You will see `CUDA` here on a GPU instance, or `CPU` here on a CPU instance.
 We'll use the model from [the Keras tutorial](https://keras.io/examples/vision/mnist_convnet/).
 
 ```python
-from tinygrad import Tensor, nn
+from tinygrad import Tensor, nn, Context
 
 class Model:
   def __init__(self):
@@ -74,8 +74,8 @@ We'll use the Adam optimizer. The `nn.state.get_parameters` will walk the model
 ```python
 optim = nn.optim.Adam(nn.state.get_parameters(model))
 batch_size = 128
+@Context(TRAINING=1)
 def step():
-  Tensor.training = True  # makes dropout work
   samples = Tensor.randint(batch_size, high=X_train.shape[0])
   X, Y = X_train[samples], Y_train[samples]
   optim.zero_grad()
@@ -143,7 +143,6 @@ Since we are just randomly sampling from the dataset, there's no real concept of
 for step in range(7000):
   loss = jit_step()
   if step%100 == 0:
-    Tensor.training = False
     acc = (model(X_test).argmax(axis=1) == Y_test).mean().item()
     print(f"step {step:4d}, loss {loss.item():.2f}, acc {acc*100.:.2f}%")
 ```

docs/quickstart.md

@@ -165,13 +165,14 @@ from extra.datasets import fetch_mnist
 Now we have everything we need to start training our neural network.
 We will be training for 1000 steps with a batch size of 64.
 
-We use `with Tensor.train()` to set the internal flag `Tensor.training` to `True` during training.
+We use `with Context(TRAINING=1)` to enable training mode.
 Upon exit, the flag is restored to its previous value by the context manager.
 
 ```python
+from tinygrad import Context
 X_train, Y_train, X_test, Y_test = fetch_mnist()
 
-with Tensor.train():
+with Context(TRAINING=1):
   for step in range(1000):
     # random sample a batch
     samp = np.random.randint(0, X_train.shape[0], size=(64))

examples/anthropic_challenge.py

@@ -100,7 +100,7 @@ class VLIWRenderer(Renderer):
       assert u.dtype.count in (1,8), "dtype count must be 1 or 8"
 
       # dumb register allocator
-      if u.op not in {Ops.STORE, Ops.SINK, Ops.GEP}:
+      if u.op not in {Ops.STORE, Ops.SINK, Ops.INDEX}:
         r[u] = reg
         reg += u.dtype.count
 
@@ -110,9 +110,9 @@ class VLIWRenderer(Renderer):
           inst.append({"flow": [("halt",)]})
         case Ops.CONST:
           inst.append({"load": [("const", r[u], u.arg)]})
-        case Ops.GEP:
-          # a GEP is just an alias to a special register in the vector
-          r[u] = r[u.src[0]] + u.arg[0]
+        case Ops.INDEX:
+          # an INDEX is just an alias to a special register in the vector
+          r[u] = r[u.src[0]] + u.src[1].arg
         case Ops.STACK:
           if all(s == u.src[0] for s in u.src):
             # if all sources are the same, we can broadcast
@@ -182,7 +182,7 @@ if __name__ == "__main__":
   # *** run on Machine and compare ***
 
   # NOTE: the scratch size needs to be reduced to 1536 when you have a register allocator
-  src = eval(prg.src[3].arg)
+  src = eval(prg.src[2].arg)
   max_regs = max(t[1] for instr in src for v in instr.values() for t in v if len(t) > 1) + 8
   print(f"{max_regs:5d} regs used" + ("" if max_regs <= 1536 else "       <-- WARNING: TOO MANY REGISTERS, MUST BE <= 1536"))
   machine = problem.Machine(mem, src, problem.DebugInfo(scratch_map={}), n_cores=1, trace=False, scratch_size=max_regs)

examples/beautiful_cartpole.py

@@ -1,6 +1,6 @@
 from typing import Tuple
 import time
-from tinygrad import Tensor, TinyJit, nn
+from tinygrad import Tensor, TinyJit, nn, Context
 import gymnasium as gym
 from tinygrad.helpers import trange
 import numpy as np  # TODO: remove numpy import
@@ -55,7 +55,7 @@ if __name__ == "__main__":
 
   @TinyJit
   def train_step(x:Tensor, selected_action:Tensor, reward:Tensor, old_log_dist:Tensor) -> Tuple[Tensor, Tensor, Tensor]:
-    with Tensor.train():
+    with Context(TRAINING=1):
       log_dist, value = model(x)
       action_mask = (selected_action.reshape(-1, 1) == Tensor.arange(log_dist.shape[1]).reshape(1, -1).expand(selected_action.shape[0], -1)).float()
 

examples/beautiful_cifar.py

@@ -122,7 +122,7 @@ if __name__ == "__main__":
     return ret.mul(hyp['opt']['loss_scale_scaler']*loss_batchsize_scaler).sum().div(hyp['opt']['loss_scale_scaler'])
 
   @TinyJit
-  @Tensor.train()
+  @Context(TRAINING=1)
   def train_step(idxs:Tensor) -> Tensor:
     X, Y = X_train[idxs], Y_train[idxs]
     if len(GPUS) > 1:

examples/beautiful_mnist.py

@@ -1,6 +1,6 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, function, Context
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist
 
@@ -19,7 +19,7 @@ class Model:
   def __call__(self, x:Tensor) -> Tensor: return x.sequential(self.layers)
 
   @TinyJit
-  @Tensor.train()
+  @Context(TRAINING=1)
   def train_step(self, X_train:Tensor, Y_train:Tensor) -> Tensor:
     opt.zero_grad()
     samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])

examples/beautiful_mnist_multigpu.py

@@ -1,6 +1,6 @@
 # model based off https://towardsdatascience.com/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
 from typing import List, Callable
-from tinygrad import Tensor, TinyJit, nn, GlobalCounters, Device
+from tinygrad import Tensor, TinyJit, nn, GlobalCounters, Device, Context
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist
 
@@ -31,7 +31,7 @@ if __name__ == "__main__":
 
   @TinyJit
   def train_step() -> Tensor:
-    with Tensor.train():
+    with Context(TRAINING=1):
       opt.zero_grad()
       samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
       Xt, Yt = X_train[samples].shard_(GPUS, axis=0), Y_train[samples].shard_(GPUS, axis=0)  # we shard the data on axis 0

examples/gradaccum_mnist.py

@@ -1,6 +1,6 @@
 import itertools
 from typing import Callable
-from tinygrad import nn, Tensor, dtypes, Device, TinyJit
+from tinygrad import nn, Tensor, dtypes, Device, TinyJit, Context
 from tinygrad.helpers import getenv, trange, partition
 
 class Model:
@@ -59,7 +59,7 @@ if __name__ == "__main__":
   Tensor.realize(*params, *buffers, *adam_params, loss, grads)
 
   @TinyJit
-  @Tensor.train()
+  @Context(TRAINING=1)
   def microbatch():
     samples = Tensor.randint(BS // ACC_STEPS, high=X_train.shape[0])
     for t in params: t.grad = None

examples/hlb_cifar10.py

@@ -10,7 +10,7 @@ from extra.lr_scheduler import OneCycleLR
 from tinygrad import nn, dtypes, Tensor, Device, GlobalCounters, TinyJit, Variable
 from tinygrad.nn.state import get_state_dict
 from tinygrad.nn import optim
-from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod
+from tinygrad.helpers import Context, BEAM, WINO, getenv, colored, prod, TRAINING
 from extra.bench_log import BenchEvent, WallTimeEvent
 
 cifar_mean = [0.4913997551666284, 0.48215855929893703, 0.4465309133731618]
@@ -44,7 +44,7 @@ class UnsyncedBatchNorm:
     return ret.reshape(x.shape).cast(x.dtype)
 
   def calc_stats(self, x:Tensor):
-    if Tensor.training:
+    if TRAINING:
       # This requires two full memory accesses to x
       # https://github.com/pytorch/pytorch/blob/c618dc13d2aa23625cb0d7ada694137532a4fa33/aten/src/ATen/native/cuda/Normalization.cuh
       # There's "online" algorithms that fix this, like https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Welford's_Online_algorithm
@@ -359,11 +359,11 @@ def train_cifar():
   i = 0
   eval_acc_pct = 0.0
   batcher = fetch_batches(X_train, Y_train, BS=BS, is_train=True)
-  with Tensor.train():
+  with Context(TRAINING=1):
     st = time.monotonic()
     while i <= STEPS:
       if i % getenv("EVAL_STEPS", STEPS) == 0 and i > 1 and not getenv("DISABLE_BACKWARD"):
-        # Use Tensor.training = False here actually bricks batchnorm, even with track_running_stats=True
+        # Using Context(TRAINING=0) here actually bricks batchnorm, even with track_running_stats=True
         corrects = []
         corrects_ema = []
         losses = []

examples/llm.c/export.py

@@ -3,7 +3,7 @@ import os
 if "NOOPT" not in os.environ: os.environ["NOOPT"] = "1"
 from tinygrad import Device, nn, Tensor, dtypes
 from train_gpt2 import GPT, GPTConfig
-from tinygrad.helpers import DEV, dedup, flatten, getenv, GlobalCounters, to_function_name
+from tinygrad.helpers import DEV, dedup, flatten, getenv, GlobalCounters, to_function_name, Context
 from tinygrad.engine.realize import get_kernel
 from tinygrad.schedule.memory import memory_planner
 from tinygrad.uop.ops import Ops
@@ -23,23 +23,23 @@ if __name__ == "__main__":
   #B, T = Variable("B", 1, 128).bind(4), 64 #Variable("T", 1, 1024).bind(64)
   B, T = 4, 64
 
-  Tensor.training = True
   optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=1e-4)
   warmup_count = getenv("WARMUP", 3)
-  for i in range(warmup_count):  # TODO: why does it take three and not two to stabilize
-    GlobalCounters.reset()
-    X = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
-    Y = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
-    _, loss = model(X, Y)
-    optimizer.zero_grad()
-    if getenv("BACKWARD", 1):
-      loss.backward()
-      tensors = optimizer.schedule_step()
-    else:
-      tensors = []
-    sched = loss.schedule(*tensors)
-    print(f"calls {i}:", len(sched))
-    #run_schedule(sched[:])
+  with Context(TRAINING=1):
+    for i in range(warmup_count):  # TODO: why does it take three and not two to stabilize
+      GlobalCounters.reset()
+      X = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
+      Y = Tensor.empty(4, 64, dtype=dtypes.int).reshape(B, T)
+      _, loss = model(X, Y)
+      optimizer.zero_grad()
+      if getenv("BACKWARD", 1):
+        loss.backward()
+        tensors = optimizer.schedule_step()
+      else:
+        tensors = []
+      sched = loss.schedule(*tensors)
+      print(f"calls {i}:", len(sched))
+      #run_schedule(sched[:])
   sched = memory_planner(sched)
   ast_dedup = dedup([si.ast for si in sched if si.ast.op is Ops.SINK])
   srcs = {}

examples/llm.c/train_gpt2.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 import os, math, time
 import numpy as np
-from tinygrad import Tensor, nn, fetch, Device, TinyJit, GlobalCounters
+from tinygrad import Tensor, nn, fetch, Device, TinyJit, GlobalCounters, Context
 from dataclasses import dataclass
 
 @dataclass
@@ -177,7 +177,7 @@ if __name__ == "__main__":
   if args.gpus > 1: x, y = x.shard(GPUS, axis=0), y.shard(GPUS, axis=0)
 
   @TinyJit
-  @Tensor.train()
+  @Context(TRAINING=1)
   def step(x:Tensor, y:Tensor) -> Tensor:
     _, loss = model(x, y)
     optimizer.zero_grad()
@@ -204,4 +204,3 @@ if __name__ == "__main__":
     top_k = 40
     y = model.generate(x, max_new_tokens, temperature=temperature, top_k=top_k)
     print(decode(y[0].tolist()))
-

examples/minrf.py

@@ -1,5 +1,5 @@
 # much taken from https://github.com/cloneofsimo/minRF
-from tinygrad import Tensor, nn, GlobalCounters, TinyJit
+from tinygrad import Tensor, nn, GlobalCounters, TinyJit, Context
 from tinygrad.helpers import getenv, trange
 from extra.models.llama import Attention, FeedForward, precompute_freqs_cis
 
@@ -135,7 +135,7 @@ if __name__ == "__main__":
   optimizer = nn.optim.Adam(nn.state.get_parameters(model), lr=5e-4)
 
   @TinyJit
-  @Tensor.train()
+  @Context(TRAINING=1)
   def train_step():
     if getenv("OVERFIT"): samples = Tensor.zeros(getenv("BS", 256), dtype='int')
     else: samples = Tensor.randint(getenv("BS", 256), high=X_train.shape[0])

examples/mlperf/initializers.py

@@ -2,7 +2,7 @@ import math
 from typing import Union
 
 from tinygrad import Tensor, nn, dtypes
-from tinygrad.helpers import prod, argfix, Context
+from tinygrad.helpers import prod, argfix, Context, TRAINING
 from tinygrad.nn.state import get_parameters
 from extra.models.unet import UNetModel
 
@@ -85,7 +85,7 @@ class FrozenBatchNorm2dRetinaNet(nn.BatchNorm2d):
 
   def __call__(self, x:Tensor) -> Tensor:
     batch_mean, batch_var = super().calc_stats(x.cast(dtypes.float32))
-    if self.track_running_stats and Tensor.training:
+    if self.track_running_stats and TRAINING:
       self.running_mean.assign((1-self.momentum) * self.running_mean + self.momentum * batch_mean.detach().cast(self.running_mean.dtype))
       self.running_var.assign((1-self.momentum) * self.running_var + self.momentum * x.numel()/(x.numel()-x.shape[1]) * batch_var.detach().cast(self.running_var.dtype))
       self.num_batches_tracked += 1

examples/mlperf/model_eval.py

@@ -358,7 +358,7 @@ def eval_stable_diffusion():
       batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
     return batch, unpadded_bs 
 
-  @Tensor.train(mode=False)
+  @Context(TRAINING=0)
   def eval_unet(eval_inputs:list[dict], unet:UNetModel, cond_stage:FrozenOpenClipEmbedder, first_stage:AutoencoderKL,
                 inception:FidInceptionV3, clip:OpenClipEncoder) -> tuple[float, float]:
     # Eval is divided into 5 jits, one per model
@@ -498,11 +498,10 @@ def eval_stable_diffusion():
 
 if __name__ == "__main__":
   # inference only
-  Tensor.training = False
-
   models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert").split(",")
-  for m in models:
-    nm = f"eval_{m}"
-    if nm in globals():
-      print(f"eval {m}")
-      globals()[nm]()
+  with Context(TRAINING=0):
+    for m in models:
+      nm = f"eval_{m}"
+      if nm in globals():
+        print(f"eval {m}")
+        globals()[nm]()

examples/mlperf/model_spec.py

@@ -1,6 +1,6 @@
 # load each model here, quick benchmark
 from tinygrad import Tensor, GlobalCounters
-from tinygrad.helpers import getenv
+from tinygrad.helpers import getenv, Context
 import numpy as np
 
 def test_model(model, *inputs):
@@ -59,11 +59,10 @@ def spec_mrcnn():
 
 if __name__ == "__main__":
   # inference only for now
-  Tensor.training = False
-
-  for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
-    nm = f"spec_{m}"
-    if nm in globals():
-      print(f"testing {m}")
-      globals()[nm]()
+  with Context(TRAINING=0):
+    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(","):
+      nm = f"spec_{m}"
+      if nm in globals():
+        print(f"testing {m}")
+        globals()[nm]()
 

examples/mlperf/model_train.py

@@ -2,7 +2,7 @@ import os, time, math, functools, random, contextlib
 from pathlib import Path
 import multiprocessing
 
-from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
+from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes, Context
 from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling, profile_marker, DEBUG
 from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW
@@ -157,6 +157,7 @@ def train_resnet():
   # input_std = Tensor([0.229, 0.224, 0.225], device=GPUS, dtype=dtypes.float32).reshape(1, -1, 1, 1)
   def normalize(x): return (x.permute([0, 3, 1, 2]) - input_mean).cast(dtypes.default_float)
   @TinyJit
+  @Context(TRAINING=1)
   def train_step(X, Y):
     optimizer_group.zero_grad()
     X = normalize(X)
@@ -170,6 +171,7 @@ def train_resnet():
     return loss.realize(), top_1.realize()
 
   @TinyJit
+  @Context(TRAINING=0)
   def eval_step(X, Y):
     X = normalize(X)
     out = model.forward(X)
@@ -192,7 +194,6 @@ def train_resnet():
     # ** train loop **
     if MLLOGGER and RUNMLPERF:
       MLLOGGER.start(key=mllog_constants.EPOCH_START, value=e+1, metadata=dict(epoch_num=e+1))
-    Tensor.training = True
     BEAM.value = TRAIN_BEAM
 
     if INITMLPERF:
@@ -271,7 +272,6 @@ def train_resnet():
       eval_loss = 0.0
       eval_top_1 = 0
       eval_num_samples = 0
-      Tensor.training = False
       BEAM.value = EVAL_BEAM
 
       if INITMLPERF:
@@ -614,7 +614,7 @@ def train_retinanet():
 
       if getenv("RESET_STEP", 1): _train_step.reset()
 
-      with Tensor.train(mode=False):
+      with Context(TRAINING=0):
         if not RUNMLPERF:
           i, proc = 0, _fake_data_get(EVAL_BS, val=(val:=True))
         else:
@@ -784,7 +784,7 @@ def train_unet3d():
     return x.shard(GPUS, axis=0).realize(), y.shard(GPUS, axis=0), cookie
 
   @TinyJit
-  @Tensor.train()
+  @Context(TRAINING=1)
   def train_step(model, x, y):
     optim.zero_grad()
 
@@ -795,7 +795,7 @@ def train_unet3d():
     optim.step()
     return loss.realize()
 
-  @Tensor.train(mode=False)
+  @Context(TRAINING=0)
   def eval_step(model, x, y):
     y_hat, y = sliding_window_inference(model, x, y, gpus=GPUS)
     y_hat, y = Tensor(y_hat), Tensor(y)
@@ -919,6 +919,7 @@ def train_rnnt():
   pass
 
 @TinyJit
+@Context(TRAINING=0)
 def eval_step_bert(model, input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor, masked_positions:Tensor, masked_lm_ids:Tensor,
                    masked_lm_weights:Tensor, next_sentence_labels:Tensor, GPUS):
   for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
@@ -1106,6 +1107,7 @@ def train_bert():
       MLLOGGER.start(key=mllog_constants.EPOCH_START, value=i*GBS, metadata={"epoch_num": i*GBS})
 
   @TinyJit
+  @Context(TRAINING=1)
   def train_step_bert(input_ids:Tensor, segment_ids:Tensor, attention_mask:Tensor,
                       masked_positions:Tensor, masked_lm_ids:Tensor, masked_lm_weights:Tensor, next_sentence_labels:Tensor):
     for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
@@ -1133,7 +1135,6 @@ def train_bert():
 
   while train_data is not None and i < train_steps and not achieved:
     if getenv("TRAIN", 1):
-      Tensor.training = True
       BEAM.value = TRAIN_BEAM
       st = time.perf_counter()
       GlobalCounters.reset()
@@ -1186,7 +1187,6 @@ def train_bert():
       eval_lm_accs = []
       eval_clsf_accs = []
       eval_times = []
-      Tensor.training = False
       BEAM.value = EVAL_BEAM
 
       for j in tqdm(range(max_eval_steps), desc="Evaluating", total=max_eval_steps, disable=BENCHMARK):
@@ -1490,7 +1490,7 @@ def train_llama3():
     return lr_cpu, grad_norm_cpu
 
   @TinyJit
-  @Tensor.train(False)
+  @Context(TRAINING=0)
   def eval_step(tokens:Tensor):
     if is_dp: tokens = tokens.to(None).shard(device, 0)
     if is_mp: tokens = tokens.shard(device)
@@ -1803,7 +1803,7 @@ if __name__ == "__main__":
   elif getenv("RUNMLPERF"): bench_log_manager = WallTimeEvent(BenchEvent.MLPERF_RUN)
   else: bench_log_manager = contextlib.nullcontext()
 
-  with Tensor.train():
+  with Context(TRAINING=1):
     for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
       nm = f"train_{m}"
       if nm in globals():

examples/mlperf/models/flat_llama.py

@@ -38,7 +38,7 @@ def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
 
 def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
            x_fp8:Tensor|None=None, x_new_amax:Tensor|None=None,
-           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
+           grad_amax_state:Tensor|None=None, x_prequant_mx:tuple|None=None) -> tuple[Tensor,...]:
   if not fp8:
     if ASM_GEMM:
       from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
@@ -47,12 +47,14 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
   assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
   if MXFP8:
     from extra.gemm.cdna_asm_gemm import asm_gemm, quantize_mxfp8, mx_pack, can_use_asm_gemm, _mx_block_scale
-    x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
+    if x_prequant_mx is not None: x_q, x_e8, x_si = x_prequant_mx       # fused producer already quantized (2d)
+    else: x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
+    l_shape = x.shape[:-1] if x is not None else x_q.shape[:-1]
     if can_use_asm_gemm(x_q, w.T):
       out = asm_gemm(x_q, w.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(w_inv_scale), w_inv_scale),
-                     mx_w_stored=True).reshape(*x.shape[:-1], w.shape[0])
+                     mx_w_stored=True).reshape(*l_shape, w.shape[0])
     else:
-      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*x.shape[:-1], x.shape[-1])
+      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*l_shape, x_q.shape[-1])
       out = x_phys @ (w.cast(dtypes.bfloat16) * _mx_block_scale(w_inv_scale)).T
     return out, (amax_x.detach() if amax_x is not None else None), x_q
   if x_fp8 is None:
@@ -126,10 +128,8 @@ class FlatTransformer:
 
     # FeedForward
     if SPLIT_W13:
-      if getenv("ZEROS"): w13_raw = Tensor.zeros(2, self.n_layers, hidden_dim, dim)
-      else: w13_raw = Tensor.normal(2, self.n_layers, hidden_dim, dim, mean=0.0, std=0.02)
-      self.w1, s_1 = self.lin_per_layer(dim, hidden_dim, w=w13_raw[0])
-      self.w3, s_3 = self.lin_per_layer(dim, hidden_dim, w=w13_raw[1])
+      self.w1, s_1 = self.lin_per_layer(dim, hidden_dim)
+      self.w3, s_3 = self.lin_per_layer(dim, hidden_dim)
     else:
       self.w13, s_13 = self.lin_per_layer(dim, hidden_dim * 2)
     self.w2, s_2 = self.lin_per_layer(hidden_dim, dim, std=scaled_std)
@@ -160,7 +160,7 @@ class FlatTransformer:
   def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02, w:Tensor|None=None):
     if w is None:
       if getenv("ZEROS"): w = Tensor.zeros(self.n_layers, out_features, in_features)
-      else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std).realize()
+      else: w = Tensor.normal(self.n_layers, out_features, in_features, mean=0.0, std=std)
     if MXFP8:
       from extra.gemm.cdna_asm_gemm import quantize_mxfp8
       w_q, w_e8, _ = quantize_mxfp8(w.reshape(self.n_layers * out_features, in_features))
@@ -216,8 +216,15 @@ class FlatTransformer:
       x_w3, new_amax, *s = matmul(inp, kwargs["w3"], amax_x=kwargs["amax_x3"], w_inv_scale=kwargs["s_3"], grad_amax_state=kwargs["grad_amax_xw3"])
       amaxs.append(new_amax)
       saves.extend([*s, x_w3])
-      out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
-                                 grad_amax_state=kwargs["grad_amax_xout"])
+      if FUSED_SILU_W13 and MXFP8:
+        from extra.llama_kernels.fused_silu_mul_quantize_mxfp8 import fused_silu_mul_quantize_mxfp8
+        aq, ae8, asi = fused_silu_mul_quantize_mxfp8(x_w1.reshape(-1, x_w1.shape[-1]), x_w3.reshape(-1, x_w3.shape[-1]))
+        out, new_amax, *s = matmul(None, kwargs["w2"], x_prequant_mx=(aq, ae8, asi), amax_x=kwargs["amax_x2"],
+                                   w_inv_scale=kwargs["s_2"], grad_amax_state=kwargs["grad_amax_xout"])
+        out = out.reshape(*x_w1.shape[:-1], kwargs["w2"].shape[0])
+      else:
+        out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
+                                   grad_amax_state=kwargs["grad_amax_xout"])
       amaxs.append(new_amax)
       saves.extend([*s, out])
     else:
@@ -247,20 +254,30 @@ class FlatTransformer:
       for v in get_parameters(self): v.shard_(device, axis=None)
     else:
       # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
-      def _shard_fp8(name:str, axis:int):
-        getattr(self, name).shard_(device, axis=axis)
-        scale_axis = axis if MXFP8 else (1 if axis == 1 else None) if COLUMNWISE_WEIGHT_SCALE else None
-        self._fp8_inv_scale[name] = self._fp8_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
-        self._fp8_next_inv_scale[name] = self._fp8_next_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
-        Tensor.realize(getattr(self, name), self._fp8_inv_scale[name], self._fp8_next_inv_scale[name])
+      def _shard_fp8(name:str, axis:int, std:float=0.02):
+        w = getattr(self, name)
+        if MXFP8:
+          from extra.gemm.cdna_asm_gemm import quantize_mxfp8
+          w_bf16 = Tensor.empty(self.n_layers, w.shape[1], w.shape[2], dtype=dtypes.bfloat16).shard(device, axis=axis).randn_like() * std
+          w_q, w_e8, _ = quantize_mxfp8(w_bf16)
+          w.replace(w_q)
+          self._fp8_inv_scale[name].replace(w_e8.contiguous()).is_param_(False)
+          self._fp8_next_inv_scale[name].replace(w_e8.contiguous()).is_param_(False)
+        else:
+          w.shard_(device, axis=axis)
+          scale_axis = (1 if axis == 1 else None) if COLUMNWISE_WEIGHT_SCALE else None
+          self._fp8_inv_scale[name] = self._fp8_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
+          self._fp8_next_inv_scale[name] = self._fp8_next_inv_scale[name].shard(device, axis=scale_axis).contiguous().is_param_(False)
+        Tensor.realize(w, self._fp8_inv_scale[name], self._fp8_next_inv_scale[name])
+      sstd = 0.02 / math.sqrt(2 * self.n_layers)
       _shard_fp8("wqkv", 1)          # (n_layers, out, dim) shard out
-      _shard_fp8("wo", 2)            # (n_layers, dim, in) shard in
+      _shard_fp8("wo", 2, sstd)      # (n_layers, dim, in) shard in
       if SPLIT_W13:
         _shard_fp8("w1", 1)
         _shard_fp8("w3", 1)
       else:
         _shard_fp8("w13", 1)         # (n_layers, hidden*2, dim) shard out
-      _shard_fp8("w2", 2)            # (n_layers, dim, hidden) shard in
+      _shard_fp8("w2", 2, sstd)      # (n_layers, dim, hidden) shard in
       self.attention_norm.shard_(device, axis=None).realize()
       self.ffn_norm.shard_(device, axis=None).realize()
       self.norm.weight.shard_(device, axis=None).realize()

examples/mlperf/models/gpt_oss.py

@@ -0,0 +1,285 @@
+import math, os, functools
+if __name__ == "__main__":
+  os.environ["DEFAULT_FLOAT"] = "bfloat16"
+  os.environ["OPTIM_DTYPE"] = "bfloat16"
+  if "DEV" not in os.environ: os.environ["DEV"] = "NULL::gfx950"
+  # CDNA
+  os.environ["DEVICE_IN_FUNCTION_BUG"] = "1"
+  os.environ["ALL2ALL"] = "1"
+  os.environ["USE_ATOMICS"] = "1"
+from tinygrad import Tensor, nn, function, getenv, dtypes, TinyJit
+from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
+from tinygrad.uop.ops import Ops, UOp
+from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
+from extra.llama_kernels.rmsnorm import rmsnorm
+from extra.gemm.cdna_asm_gemm import _mx_block_scale
+
+FP8_DTYPE = dtypes.fp8e4m3
+FP8_MAX = 448.0
+INIT_STD = 0.008
+
+def quantize_mx(x:Tensor) -> tuple[Tensor, Tensor]:
+  *batch, K = x.shape
+  scale_K = K // 32
+  amax = x.detach().float().reshape(*batch, scale_K, 32).abs().max(axis=-1)
+  e8 = (amax.maximum(1e-38).log2().floor() + 127).clamp(0, 254).cast(dtypes.uint8)
+  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(*batch, scale_K, 1).expand(*batch, scale_K, 32).reshape(*batch, K)
+  x_scaled = x.float() * qscale
+  x_clamped = x_scaled + (x_scaled.detach().clamp(-FP8_MAX, FP8_MAX) - x_scaled.detach())  # STE
+  return x_clamped.cast(FP8_DTYPE), e8
+
+def _quant_dequant_fwd(x:Tensor) -> Tensor:
+  # x (2d bf16) -> bf16 value after an mxfp8 round-trip (1x32 block scaling on the last axis)
+  M, K = x.shape
+  scale_K = K // 32
+  amax = x.float().reshape(M, scale_K, 32).abs().max(axis=-1)
+  e8 = (amax.maximum(1e-38).log2().floor() + 127).clamp(0, 254).cast(dtypes.uint8)
+  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(M, scale_K, 1).expand(M, scale_K, 32).reshape(M, K)
+  x_fp8 = (x.float() * qscale).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE).cast(dtypes.float32)
+  return (x_fp8 * _mx_block_scale(e8)).cast(dtypes.bfloat16)
+
+@functools.cache
+def _quant_dequant_fwd_fxn(x_p, device):
+  return _quant_dequant_fwd(Tensor(x_p, device=device))
+
+def _quant_dequant_bwd(grad:UOp, call:UOp) -> tuple:
+  return (Tensor(grad).cast(dtypes.bfloat16).uop,)
+
+def quant_dequant_mx(x:Tensor) -> Tensor:
+  fxn = _quant_dequant_fwd_fxn(x.as_param(0).uop, x.device)
+  return Tensor(UOp.maketuple(fxn.uop).call(x.uop, grad_fxn=_quant_dequant_bwd).gettuple(0))
+
+def _dequant_fwd(w_q:Tensor, w_scale:Tensor) -> Tensor:
+  return w_q.cast(dtypes.bfloat16) * _mx_block_scale(w_scale)
+
+@functools.cache
+def _dequant_fwd_fxn(wq_p, ws_p, device):
+  return _dequant_fwd(Tensor(wq_p, device=device), Tensor(ws_p, device=device))
+
+def _dequant_bwd(grad:UOp, call:UOp) -> tuple:
+  w_scale = Tensor(call.src[2])
+  return ((Tensor(grad).cast(dtypes.bfloat16) * _mx_block_scale(w_scale).cast(dtypes.bfloat16)).uop, None)
+
+def dequant_weight(w_q:Tensor, w_scale:Tensor) -> Tensor:
+  fxn = _dequant_fwd_fxn(w_q.as_param(0).uop, w_scale.as_param(1).uop, w_q.device)
+  call = UOp.maketuple(fxn.uop).call(w_q.uop, w_scale.uop, grad_fxn=_dequant_bwd)
+  return Tensor(call.gettuple(0))
+
+def matmul_mx(x:Tensor, w_q:Tensor, w_scale:Tensor) -> Tensor:
+  l_shape = x.shape[:-1]
+  x_phys = quant_dequant_mx(x.reshape(-1, x.shape[-1])).reshape(*l_shape, x.shape[-1])
+  w_phys = dequant_weight(w_q, w_scale)
+  return (x_phys @ w_phys.T).cast(dtypes.bfloat16)
+
+def swiglu(x:Tensor, limit:float=7.0, alpha:float=1.702) -> Tensor:
+  x_glu, x_linear = x[..., ::2], x[..., 1::2]
+  x_glu = x_glu.clamp(max_=limit)
+  x_linear = x_linear.clamp(-limit, limit)
+  return (x_glu * (alpha * x_glu).sigmoid()) * (x_linear + 1)
+
+class GPTOSS:
+  def __init__(self, dim:int, n_layers:int, n_heads:int, n_kv_heads:int, head_dim:int, n_experts:int, experts_per_tok:int,
+               intermediate_size:int, vocab_size:int, norm_eps:float=1e-5, rope_theta:int=150000, sliding_window:int=128,
+               swiglu_limit:float=7.0, max_context:int=8192):
+    self.dim, self.n_layers, self.n_heads, self.n_kv_heads, self.head_dim = dim, n_layers, n_heads, n_kv_heads, head_dim
+    self.n_rep = n_heads // n_kv_heads
+    self.n_experts, self.experts_per_tok, self.intermediate_size = n_experts, experts_per_tok, intermediate_size
+    self.vocab_size, self.norm_eps, self.sliding_window, self.swiglu_limit = vocab_size, norm_eps, sliding_window, swiglu_limit
+    self.sm_scale = 1.0 / math.sqrt(head_dim)
+
+    scaled_std = INIT_STD / math.sqrt(2 * n_layers)
+    q_dim, qkv_dim = n_heads * head_dim, head_dim * (n_heads + 2 * n_kv_heads)
+
+    # attn
+    self.wqkv, self.wqkv_scale = self._quant_weight(n_layers, qkv_dim, dim)
+    self.wqkv_bias = Tensor.zeros(n_layers, qkv_dim, dtype=dtypes.bfloat16).contiguous()
+    self.wo, self.wo_scale = self._quant_weight(n_layers, dim, q_dim, std=scaled_std)
+    self.wo_bias = Tensor.zeros(n_layers, dim, dtype=dtypes.bfloat16).contiguous()
+    self.sinks = Tensor.zeros(n_layers, n_heads, dtype=dtypes.bfloat16).contiguous()
+    self.attention_norm = Tensor.ones(n_layers, dim).contiguous()
+
+    # moe ffn
+    self.ffn_norm = Tensor.ones(n_layers, dim).contiguous()
+    self.gate = Tensor.normal(n_layers, n_experts, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
+    self.gate_bias = Tensor.zeros(n_layers, n_experts, dtype=dtypes.bfloat16).contiguous()
+    self.w_gate_up, self.w_gate_up_scale = self._quant_weight(n_layers, n_experts, intermediate_size * 2, dim)
+    self.w_gate_up_bias = Tensor.zeros(n_layers, n_experts, intermediate_size * 2, dtype=dtypes.bfloat16).contiguous()
+    self.w_down, self.w_down_scale = self._quant_weight(n_layers, n_experts, dim, intermediate_size, std=scaled_std)
+    self.w_down_bias = Tensor.zeros(n_layers, n_experts, dim, dtype=dtypes.bfloat16).contiguous()
+
+    # output
+    self.norm = nn.RMSNorm(dim, norm_eps)
+    self.tok_embeddings = nn.Embedding(vocab_size, dim)
+    self.tok_embeddings.weight = Tensor.normal(vocab_size, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
+    self.output = Tensor.normal(vocab_size, dim, mean=0.0, std=INIT_STD, dtype=dtypes.bfloat16)
+    self.freqs_cis = precompute_freqs_cis(head_dim, max_context * 2, rope_theta).contiguous().is_param_(False)
+
+  def _quant_weight(self, *shape:int, std:float=INIT_STD):
+    w = Tensor.zeros(*shape) if getenv("ZEROS") else Tensor.normal(*shape, mean=0.0, std=std)
+    w_q, w_e8 = quantize_mx(w)
+    return w_q, w_e8.is_param_(False)
+
+  def _attn_mask(self, seqlen:int, sliding:bool, dtype) -> Tensor:
+    i, j = Tensor.arange(seqlen).reshape(seqlen, 1), Tensor.arange(seqlen).reshape(1, seqlen)
+    allowed = j <= i
+    if sliding: allowed = allowed & (i - j < self.sliding_window)
+    return allowed.where(0.0, -1e30).cast(dtype).contiguous()
+
+  def attention(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, *, attention_norm:Tensor, wqkv:Tensor, wqkv_scale:Tensor,
+                wqkv_bias:Tensor, wo:Tensor, wo_scale:Tensor, wo_bias:Tensor, sinks:Tensor):
+    bsz, seqlen, _ = x.shape
+    x_normed, rrms = rmsnorm(x, self.norm_eps)
+    qkv = matmul_mx(x_normed * attention_norm, wqkv, wqkv_scale) + wqkv_bias
+    qkv = qkv.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep + 2, self.head_dim)
+    xq = qkv[:, :, :, :self.n_rep].reshape(bsz, seqlen, self.n_heads, self.head_dim)
+    xk, xv = qkv[:, :, :, self.n_rep], qkv[:, :, :, self.n_rep + 1]
+    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
+
+    xq = xq.cast(dtypes.bfloat16).reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
+    xk = xk.cast(dtypes.bfloat16).permute(0, 2, 1, 3).unsqueeze(2)
+    xv = xv.cast(dtypes.bfloat16).permute(0, 2, 1, 3).unsqueeze(2)
+    scores = ((xq @ xk.transpose(-2, -1)).float() * self.sm_scale + mask).contiguous()
+    sink = sinks.reshape(1, self.n_kv_heads, self.n_rep, 1, 1).float()
+    m = scores.max(-1, keepdim=True).maximum(sink)
+    e = (scores - m).exp()
+    w = (e / (e.sum(-1, keepdim=True) + (sink - m).exp())).cast(dtypes.bfloat16).contiguous()
+    attn = (w @ xv).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]
+
+  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,
+                   w_down:Tensor, w_down_scale:Tensor, w_down_bias:Tensor):
+    x_normed, rrms = rmsnorm(x, self.norm_eps)
+    inp = x_normed * ffn_norm
+
+    logits = inp.float() @ gate.float().T + gate_bias.float()
+    thresh = logits.topk(self.experts_per_tok)[0][..., -1:]
+    weights = (logits >= thresh).where(logits, -float("inf")).softmax(-1)
+
+    out = None
+    for e in range(self.n_experts):
+      gate_up = (matmul_mx(inp.contiguous_backward(), w_gate_up[e], w_gate_up_scale[e]) + w_gate_up_bias[e]).contiguous()
+      y = (matmul_mx(swiglu(gate_up, self.swiglu_limit), w_down[e], w_down_scale[e]) + w_down_bias[e]).contiguous()
+      contrib = (weights[..., e:e+1].cast(y.dtype) * y).contiguous()
+      out = contrib if out is None else out + contrib
+    return out, [x_normed, rrms]
+
+  @function(precompile=True, precompile_backward=True)
+  def run_layer(self, x:Tensor, freqs_cis:Tensor, mask:Tensor, attn_kwargs:dict, ffn_kwargs:dict, save:bool=True):
+    attn, attn_saves = self.attention(x, freqs_cis, mask, **attn_kwargs)
+    h = x + attn
+    ffn, ffn_saves = self.feed_forward(h, **ffn_kwargs)
+    h = h + ffn
+    if save: return (h, *attn_saves, *ffn_saves)
+    return (h,)
+
+  def shard(self, device:tuple[str, ...], mp:bool=False):
+    assert not mp, "MP not supported"
+    from tinygrad.nn.state import get_parameters
+    for v in get_parameters(self): v.shard_(device, axis=None)
+    Tensor.realize(*get_parameters(self))
+
+  def __call__(self, tokens:Tensor, save:bool=True):
+    h = self.tok_embeddings(tokens)
+    bsz, seqlen = tokens.shape
+    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :seqlen, :, :, :]
+    mask_full = self._attn_mask(seqlen, False, dtypes.float32)
+    mask_sliding = self._attn_mask(seqlen, True, dtypes.float32)
+    for i in range(self.n_layers):
+      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wqkv_scale=self.wqkv_scale[i],
+                         wqkv_bias=self.wqkv_bias[i], wo=self.wo[i], wo_scale=self.wo_scale[i], wo_bias=self.wo_bias[i],
+                         sinks=self.sinks[i])
+      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], gate=self.gate[i], gate_bias=self.gate_bias[i],
+                        w_gate_up=self.w_gate_up[i], w_gate_up_scale=self.w_gate_up_scale[i], w_gate_up_bias=self.w_gate_up_bias[i],
+                        w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
+      mask = mask_sliding if i % 2 == 0 else mask_full
+      h, *_ = self.run_layer(h, freqs_cis, mask, attn_kwargs, ffn_kwargs, save=save)
+
+    logits = self.norm(h) @ self.output.T
+    return logits
+
+def _get_pads(uop:UOp) -> list[UOp]:
+  if uop.op == Ops.ADD: return _get_pads(uop.src[0]) + _get_pads(uop.src[1])
+  return [uop]
+
+def apply_grad(grad_buf:Tensor, new_grad:UOp):
+  pads = _get_pads(new_grad)
+  if len(pads) <= 1:
+    new_grad = new_grad.cast(grad_buf.dtype)
+    grad_buf.uop = grad_buf.uop.after(grad_buf.uop.store(grad_buf.uop + new_grad))
+    return
+  cur = grad_buf.uop
+  for pad in sorted(pads, key=lambda p: p.marg[0][0] if p.op == Ops.PAD else 0, reverse=True):
+    if pad.op == Ops.PAD:
+      grad_shrink = tuple([(p[0], s+p[0]) for s,p in zip(pad.src[0].shape, pad.marg)])
+      buf_slice = cur.shrink(grad_shrink)
+      cur = cur.after(buf_slice.store(buf_slice + pad.src[0].cast(cur.dtype)))
+    else:
+      cur = cur.after(cur.store(cur + pad.cast(cur.dtype)))
+  grad_buf.uop = cur
+
+GPT_OSS_20B = dict(dim=2880, n_layers=24, n_heads=64, n_kv_heads=8, head_dim=64, n_experts=32, experts_per_tok=4,
+                   intermediate_size=2880, vocab_size=128256, norm_eps=1e-5, rope_theta=150000, sliding_window=128,
+                   swiglu_limit=7.0)
+
+if __name__ == "__main__":
+  config = {}
+  BS      = config["BS"]      = getenv("BS", 16)
+  SEQLEN  = config["SEQLEN"]  = getenv("SEQLEN", 8192)
+
+  model_params = GPT_OSS_20B
+  real_vocab_size = model_params["vocab_size"]
+  if (layers := getenv("LAYERS")) != 0: model_params["n_layers"] = layers
+
+  model = GPTOSS(**model_params, max_context=SEQLEN)
+
+  state = nn.state.get_state_dict(model)
+  print("tensor count:", len(state))
+
+  from tinygrad import Device
+  is_dp = (DP := getenv("DP", 1)) > 1
+  device_count = DP
+  device = tuple(f"{Device.DEFAULT}:{i}" for i in range(device_count))
+
+  if is_dp: model.shard(device)
+
+  # preallocate all the grad buffers and zero them out
+  grad_dtype = lambda x: dtypes.bfloat16 if x.dtype in dtypes.fp8s else x.dtype
+  grads = {x:x.zeros_like(dtype=grad_dtype(x)).contiguous() for x in state.values() if x.is_param}
+
+  # print model size
+  sz = 0
+  for k,v in state.items():
+    print(f"{colored(k, 'green' if v in grads else 'white'):30s} {str(v.shape):30s} {str(v.dtype):20s} {v.device}  {v.nbytes()/1e9:.2f} GB")
+    sz += v.nbytes()
+  print(f"total sz: {sz/1e9:.2f} GB")
+
+  with Timing("fake data: "): tokens = Tensor.randint(BS, SEQLEN+1, low=0, high=real_vocab_size, dtype=dtypes.int)
+  with Timing("realize weights/grads/data: "): Tensor.realize(*state.values(), *grads.values(), tokens)
+  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
+  if is_dp: tokens = tokens.shard(device, axis=0)
+
+  @TinyJit
+  def fwd_bwd(tokens:Tensor):
+    with Timing("python forward: "):
+      logits = model(tokens[:, :-1], save=True)
+      loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
+    with Timing("python backward: "):
+      for t,g in zip(grads, loss.gradient(*grads)):
+        apply_grad(grads[t], g.uop)
+    with Timing("run fwd_bwd: "): loss.realize(*grads.values())
+
+  @TinyJit
+  def optim_step():
+    for g in grads.values(): g.assign(g.zeros_like())
+    Tensor.realize(*grads.values())
+
+  for i in range(6):
+    GlobalCounters.reset()
+    profile_marker(f"step {i}")
+    with Timing(colored(f"*** step {i}: ", "red")):
+      fwd_bwd(tokens)
+      optim_step()
+  print("mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))

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

@@ -14,7 +14,6 @@ export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export USE_HK_BF16_GEMM=${USE_HK_BF16_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
 export FP8=${FP8:-1}

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

@@ -14,7 +14,6 @@ export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export USE_HK_BF16_GEMM=${USE_HK_BF16_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
 export FP8=${FP8:-1}

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

@@ -14,7 +14,6 @@ export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export USE_HK_BF16_GEMM=${USE_HK_BF16_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
 export FP8=${FP8:-1}

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

@@ -14,7 +14,6 @@ export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export USE_HK_BF16_GEMM=${USE_HK_BF16_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
 export FP8=${FP8:-1}

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

@@ -1,6 +1,6 @@
 #!/bin/bash
-export BENCHMARK=5
+export BENCHMARK=${BENCHMARK:-5}
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
 SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
-python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
+[ "$BENCHMARK" -le 3 ] || python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"

examples/mnist_gan.py

@@ -3,7 +3,7 @@ import torch
 from torchvision.utils import make_grid, save_image
 from tinygrad.nn.state import get_parameters
 from tinygrad.tensor import Tensor
-from tinygrad.helpers import trange
+from tinygrad.helpers import trange, Context
 from tinygrad.nn import optim
 from tinygrad.nn.datasets import mnist
 
@@ -86,7 +86,7 @@ if __name__ == "__main__":
   optim_g = optim.Adam(get_parameters(generator), lr=0.0002, b1=0.5)  # 0.0002 for equilibrium!
   optim_d = optim.Adam(get_parameters(discriminator), lr=0.0002, b1=0.5)
   # training loop
-  with Tensor.train():
+  with Context(TRAINING=1):
     for epoch in (t := trange(epochs)):
       loss_g, loss_d = 0.0, 0.0
       for _ in range(n_steps):

examples/openpilot/compile3.py

@@ -42,7 +42,7 @@ def compile(onnx_file):
   kernel_calls = [u for u in run_onnx_jit.captured.linear.toposort(gate=lambda x: x.op not in kernel_asts)
                   if u.op is Ops.CALL and u.src[0].op in kernel_asts]
   print(f"captured {len(kernel_calls)} kernels")
-  np.testing.assert_equal(test_val, ret, "JIT run failed")
+  if getenv("TEST", 1): np.testing.assert_equal(test_val, ret, "JIT run failed")
   print("jit run validated")
 
   # check gated read_image usage
@@ -50,7 +50,7 @@ def compile(onnx_file):
   read_image_count = 0
   gated_read_image_count = 0
   for call in kernel_calls:
-    _, _, _, source, _ = call.src[0].src
+    _, _, source, _ = call.src[0].src
     src = source.arg
     kernel_count += 1
     read_image_count += src.count("read_image")

examples/stunning_mnist.py

@@ -5,7 +5,7 @@
 #  - symbolic removal
 
 from examples.beautiful_mnist import Model
-from tinygrad import Tensor, nn, getenv, GlobalCounters, Variable
+from tinygrad import Tensor, nn, getenv, GlobalCounters, Variable, Context
 from tinygrad.nn.datasets import mnist
 from tinygrad.helpers import trange
 
@@ -26,7 +26,7 @@ if __name__ == "__main__":
   X_samp, Y_samp = X_train[samples], Y_train[samples]
   print("*** got samples")
 
-  with Tensor.train():
+  with Context(TRAINING=1):
     """
     i = UOp.range(samples.shape[0])  # TODO: fix range function on UOp
     losses = model(X_samp[i]).sparse_categorical_crossentropy(Y_samp[i]).backward().contract(i)

extra/export_model.py

@@ -1,5 +1,5 @@
 from typing import Tuple, Dict, List, Optional
-from tinygrad.dtype import DType, dtypes
+from tinygrad.dtype import DType, dtypes, AddrSpace
 from tinygrad.tensor import Tensor
 from tinygrad.device import Device, Buffer
 from tinygrad.engine.jit import TinyJit
@@ -38,8 +38,8 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li
     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[3].arg
-    cargs = [name_of(bu, i == 0) for i, bu in enumerate(arg_uops)] + [v for v in info.vars if v.op is Ops.DEFINE_VAR]
+    functions[info.function_name] = prg.src[2].arg
+    cargs = [name_of(bu, i == 0) for i, bu in enumerate(arg_uops)] + list(info.vars)
     statements.append((info.function_name, cargs, info.global_size, info.local_size))
 
   return functions, statements, {name:(size, dtype, key) for name, size, dtype, key in bufs.values()}, bufs_to_save
@@ -253,17 +253,18 @@ def export_model(model, target:str, *inputs, model_name: Optional[str] = "model"
   symbolic_vars = OrderedDict()
   for i, (_, args, global_size, _) in enumerate(statements):
     for j, var in enumerate(args):
-      if getattr(var, "op", None) is Ops.DEFINE_VAR and isinstance(getattr(var, "arg", None), tuple) and isinstance(var.arg[0], str):
+      if getattr(var, "op", None) is Ops.PARAM and var.addrspace is AddrSpace.ALU and var.arg.name is not None:
         if var not in symbolic_vars:
-          symbolic_vars[var] = var.arg[0]
+          symbolic_vars[var] = var.expr
           bufs[symbolic_vars[var]] = (var.dtype.itemsize, var.dtype, symbolic_vars[var])
         statements[i][1][j] = symbolic_vars[var]
 
     if global_size:
       for j, dim in enumerate(global_size):
-        if getattr(dim, "op", None) is Ops.ADD and len(dim.src) == 2 and {dim.src[0].op, dim.src[1].op} == {Ops.DEFINE_VAR, Ops.CONST}:
+        if getattr(dim, "op", None) is Ops.ADD and len(dim.src) == 2 and \
+           any(s.op is Ops.PARAM and s.addrspace is AddrSpace.ALU for s in dim.src) and any(s.op is Ops.CONST for s in dim.src):
           name, val = dim.src if dim.src[1].op is Ops.CONST else reversed(dim.src)
-          global_size[j] = f"_{name.arg[0]}[0] + {val.arg}"
+          global_size[j] = f"_{name.expr}[0] + {val.arg}"
 
   prg = ""
   if target == "clang":

extra/gemm/amd_asm_matmul.py

@@ -458,10 +458,11 @@ def test_matmul():
   def asm_kernel(A:UOp, B:UOp, C:UOp) -> UOp:
     gidxs = [UOp.special(n, f"gidx{i}") for i,n in enumerate(grid)]
     lidxs = [UOp.special(n, f"lidx{i}") for i,n in enumerate(local)]
-    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=max(LDS_SIZE, 65536//getenv("LIMIT_OCC", 65536)), addrspace=AddrSpace.LOCAL), (), 'lds')
+    lds_size = max(LDS_SIZE, 65536//getenv("LIMIT_OCC", 65536))
+    lds = UOp.placeholder((lds_size,), dtypes.uint8, 0, AddrSpace.LOCAL)
     sink = UOp.sink(A.base, B.base, C.base, lds, *gidxs, *lidxs, arg=KernelInfo(name=colored("kernel", "cyan"),
                                                                                   estimates=Estimates(ops=N*N*N*2, mem=N*N*4*3)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
   c = Tensor.custom_kernel(a, b, c, fxn=asm_kernel)[2]
   linear = c.schedule_linear()
 

extra/gemm/amd_copy_matmul.py

@@ -66,7 +66,7 @@ def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
 
   # accumulator (unified: both paths use (TM, TN) with scalar dtypes.float)
   acc = UOp.placeholder((TM, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  acc = acc.after(acc.store(acc.zeros_like()))
+  acc = acc.after(acc.store(acc.zeros_like(buffer=False)))
 
   if use_wmma:
     k = UOp.range(BLOCK_K // WMMA_K, 101, AxisType.REDUCE)

extra/gemm/amd_flash_attention.py

@@ -126,7 +126,7 @@ def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
   P_lds = QP_lds[:, :BLOCK_N]
   P_write = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TN, LANES_PER_WAVE_N)
   P_write = P_write.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TN)
-  # TODO: P_write[tid].store(S_reg.cast(dtypes.half)) — shaped store fails due to RESHAPE(DEFINE_LOCAL) surviving linearization
+  # TODO: P_write[tid].store(S_reg.cast(dtypes.half)) -- shaped store fails due to RESHAPE(local BUFFER) surviving linearization
   rw1 = UOp.range(TM, 296, AxisType.LOOP)
   rw2 = UOp.range(TN, 297, AxisType.LOOP)
   P_store = P_write[tid, rw1, rw2].store(S_reg[rw1, rw2].cast(dtypes.half)).end(rw1, rw2)

extra/gemm/amd_matmul.py

@@ -17,7 +17,7 @@ def make_matmul_kernel(name:str, src:str, local_size:int):
     wg_y = UOp.special(N//128, "gidx1")
     sink = UOp.sink(a.base, b.base, c.base, threads, wg_x, wg_y, arg=KernelInfo(name, estimates=Estimates(ops=2*N**3, mem=3*N*N*4)))
     lib = Device[Device.DEFAULT].compiler.compile_cached(src)
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                  UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
   return fxn
 

extra/gemm/metal_uop_matmul.py

@@ -33,7 +33,7 @@ def hand_spec_tc_cores():
   acc_load = UOp.vectorize(acc.after(gk)[0], acc.after(gk)[1])
   out = UOp(Ops.WMMA, dtypes.float.vec(2), (a_tc, b_tc, acc_load), arg=wmma_arg)
 
-  end_loop = UOp.group(*[acc[i].store(out.gep(i)) for i in range(2)]).end(gk)
+  end_loop = UOp.group(*[acc[i].store(out.index(i)) for i in range(2)]).end(gk)
 
   sink = UOp.group(*[mat_idx(c.after(end_loop), gx, gy, warp, i).store(acc[i]) for i in range(2)])
   return sink.sink(arg=KernelInfo(name="custom_metal_matmul", opts_to_apply=())).simplify()

extra/gemm/mi350x_uop_matmul.py

@@ -180,7 +180,7 @@ def custom_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
   # store the acc into gmem
   cp_i, cp_j = UOp.range(BLOCK_M//TC_M//WARPGROUP_SIZE, 10004), UOp.range(BLOCK_N//TC_N, 10005)
   c_load = lambda i: C[gx, cp_i*TC_M*WARPGROUP_SIZE + warpgroup*TC_M + (warp//16)*4+i, gy, cp_j*TC_N + warp%16]
-  store = UOp.group(*[c_load(i).store(acc[cp_j, cp_i].gep(i)) for i in range(4)])
+  store = UOp.group(*[c_load(i).store(acc[cp_j, cp_i].index(i)) for i in range(4)])
   store = store.end(cp_i, cp_j)
 
   return store.sink(arg=KernelInfo(name="custom_gemm", opts_to_apply=())).simplify()
@@ -197,7 +197,7 @@ wmma_arg = ('WMMA_16_16_32_half_float', (16, 16, 32), dtypes.half, dtypes.float,
 out = UOp(Ops.WMMA, dtypes.float.vec(4), (A_in, B_in, acc_load), arg=wmma_arg)
 
 # store back the acc
-acc = acc.after(UOp.group(*[acc[i].store(out.gep(i)) for i in range(4)]).end(K_loop))
+acc = acc.after(UOp.group(*[acc[i].store(out.index(i)) for i in range(4)]).end(K_loop))
 
 # store the acc into gmem
 store = UOp.group(*[C[gx, (warp//16)*4+i, gy, warp%16].store(acc[i]) for i in range(4)])

extra/gemm/rdna4_asm_matmul.py

@@ -219,10 +219,11 @@ def test_matmul():
   def asm_kernel(A, B, C):
     gidxs = [UOp.special(n, f"gidx{i}") for i,n in enumerate(grid)]
     lidxs = [UOp.special(THREADS, "lidx0")]
-    lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=max(LDS_SIZE, 65536//getenv("LIMIT_OCC",2)), addrspace=AddrSpace.LOCAL), (), 'lds')
+    lds_size = max(LDS_SIZE, 65536//getenv("LIMIT_OCC",2))
+    lds = UOp.placeholder((lds_size,), dtypes.uint8, 0, AddrSpace.LOCAL)
     sink = UOp.sink(A.base, B.base, C.base, lds, *gidxs, *lidxs,
                     arg=KernelInfo(name=colored("kernel","cyan"), estimates=Estimates(ops=N*N*N*2, mem=N*N*2*3)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
 
   c = Tensor.custom_kernel(a, b, c, fxn=asm_kernel)[2]
   linear = c.schedule_linear()

extra/gemm/triton_nv_matmul.py

@@ -93,7 +93,7 @@ if __name__ == "__main__":
   info = ProgramInfo(name="matmul_kernel",
                      global_size=(M//BLOCK_SIZE_M, N//BLOCK_SIZE_N, 1), local_size=(32*compiled.metadata.num_warps, 1, 1))
   sink = UOp.sink(arg=KernelInfo(name="matmul_kernel"))
-  prg_uop = to_program(UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR), UOp(Ops.SOURCE, arg=src)), arg=info),
+  prg_uop = to_program(UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR), UOp(Ops.SOURCE, arg=src)), arg=info),
                        Device.default.renderer)
   rt = get_runtime(Device.DEFAULT, prg_uop)
   all_bufs = [x.ensure_allocated() for x in bufs]

extra/hcq2/hcq2.py

@@ -2,12 +2,12 @@ from __future__ import annotations
 from typing import cast, Callable, TypeVar, Generic, Any
 import struct, functools, time, collections, importlib, itertools, weakref
 from dataclasses import replace, dataclass, field
-from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, DEBUG, dedup, flatten, pluralize
-from tinygrad.helpers import to_tuple, round_up
+from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, DEBUG, dedup, pluralize
+from tinygrad.helpers import to_tuple, round_up, partition
 from tinygrad.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer
 from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, track_rewrites, GroupOp
 from tinygrad.uop.symbolic import symbolic_simple, symbolic
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, AddrSpace, truncate
 from tinygrad.runtime.support.hcq import MMIOInterface
 from tinygrad.renderer import Renderer, Estimates
 from tinygrad.engine.realize import to_program, get_call_arg_uops, get_call_name, get_call_outs_ins, estimate_uop, pm_flatten_linear
@@ -23,11 +23,12 @@ class HCQ2Compiled(Compiled):
 
     # default pm bufferize
     self.pm_bufferize = PatternMatcher([
-      (UPat(Ops.BUFFER, tag="timeline_signal"), lambda ctx: ctx.timeline_signal()),
-      (UPat(Ops.BUFFER, tag="timeline_value"), lambda ctx: ctx.timeline_value()),
-      (UPat(Ops.BUFFER, tag="sentinel_signal"), lambda ctx: ctx.timeline_signal("sentinel", (1 << 64) - 1)),
-      (UPat(Ops.BUFFER, name="b"), lambda ctx, b:
-        Buffer(ctx.device, b.arg, b.dtype, options=BufferSpec(host=False, uncached=True, cpu_access=True, nolru=True))), # TODO: remove nolru
+      (UPat(Ops.PARAM, tag="timeline_signal"), lambda ctx: ctx.timeline_signal()),
+      (UPat(Ops.PARAM, tag="timeline_value"), lambda ctx: ctx.timeline_value()),
+      (UPat(Ops.PARAM, tag="sentinel_signal"), lambda ctx: ctx.timeline_signal("sentinel", (1 << 64) - 1)),
+      (UPat(Ops.PARAM, name="b"), lambda ctx, b:
+        Buffer(ctx.device, b.max_numel(), b.dtype.base, options=BufferSpec(host=False, uncached=True, cpu_access=True, nolru=True))
+        if b.tag is not None else None), # TODO: remove nolru
     ])
 
     super().__init__(device, allocator, compilers, lambda *a, **kw: None, None, arch=arch)
@@ -44,10 +45,6 @@ class HCQ2Compiled(Compiled):
     buf.as_memoryview(force_zero_copy=True).cast('Q')[0] = init_value
     return buf
 
-  @functools.cached_property
-  def timestamps_buf(self) -> Buffer:
-    return Buffer(self.device, 0x1000, dtypes.uint8, options=BufferSpec(cpu_access=True), preallocate=True)
-
   def synchronize(self, timeout:int|None=None):
     if not hasattr(self, 'iface'): return
     sig = self.timeline_signal()._buf.cpu_view().mv.cast('Q')
@@ -132,32 +129,6 @@ class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
 
   # def _as_buffer(self, buf): return buf.cpu_view().mv
 
-def unwrap_after(uop):
-  while uop.op is Ops.AFTER: uop = uop.src[0]
-  return uop
-
-def make_getaddr(u, device=None):
-  if unwrap_after(u).op not in (Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT): return u
-  return UOp(Ops.GETADDR, dtypes.uint64, src=(u, UOp(Ops.DEVICE, arg=device or to_tuple(u.device)[0])))
-
-def make_ins(op, *srcs):
-  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)
-
-def make_cmdbuf(lin, devs, tag):
-  blob, patches = b'', []
-  for s in (s for ins in lin.src for s in ins.src):
-    if s.op is not Ops.CONST: patches.append((len(blob), s))
-    blob += struct.pack(f'<{s.dtype.fmt}', s.arg if s.op is Ops.CONST else 0x0)
-  buf = UOp.new_buffer(devs, len(blob), dtypes.uint8).rtag(tag)
-  stores = [buf.index(UOp.const(dtypes.int, off), dtype=buf.dtype.ptr()).cast(s.dtype.ptr()).store(s) for off, s in patches]
-  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *stores)
-
-def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))
-
-def make_signal(devs, queue=None, sentinel=False):
-  return UOp.new_buffer(devs, 1, dtypes.uint64).rtag("sentinel_signal" if sentinel else (queue, "timeline_signal") if queue else "timeline_signal")
-def make_signal_value(devs, queue=None): return UOp.new_buffer(devs, 1, dtypes.uint64).rtag((queue, "timeline_value") if queue else "timeline_value")
-
 # *****************
 # 0. helpers
 
@@ -166,95 +137,92 @@ HCQ_P2P_DEVS = HCQ_DEVS | frozenset(("CPU",))
 
 def all_devices_in(d:Any, c:frozenset[str]) -> bool: return {x.split(":")[0] for x in to_tuple(d)} <= c
 
+def unwrap_after(uop):
+  while uop.op is Ops.AFTER: uop = uop.src[0]
+  return uop
+
+def make_getaddr(u, device=None):
+  if unwrap_after(u).op not in (Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT, Ops.PARAM): return u
+  return UOp(Ops.GETADDR, dtypes.uint64, src=(u,), arg=device or to_tuple(u.device)[0])
+
+def make_ins(op, *srcs):
+  return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)
+
+def make_placeholder(devs, size:int, dtype, name=None, unique=True) -> UOp:
+  return UOp.param(next(UOp.unique_num) if unique else 0, dtype.ptr(size), device=devs).rtag(name or "buf")
+
+def make_patch(buf:UOp, off:sint, val:UOp, dtype=None) -> UOp:
+  dt = dtype or val.dtype
+  return UOp(Ops.SHRINK, buf.dtype.base, (buf, UOp.const(dtypes.int, off), UOp.const(dtypes.int, dt.itemsize))).bitcast(dt).store(val.cast(dt))
+
+def make_cmdbuf(lin, devs):
+  blob, patches = b'', []
+  for s in (s for ins in lin.src for s in ins.src):
+    if s.op is not Ops.CONST: patches.append((len(blob), s))
+    blob += struct.pack(f'<{s.dtype.fmt}', s.arg if s.op is Ops.CONST else 0x0)
+  buf = make_placeholder(devs, len(blob), dtypes.uint8)
+
+  # pull patches to cmdbuf
+  afters = dedup(u for _, s in patches for u in s.toposort() if u.op is Ops.AFTER)
+  deps = tuple(d for p in afters for d in p.src[1:])
+  cmdbuf = buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *[make_patch(buf, off, s) for off, s in patches], *deps)
+  return cmdbuf.substitute({p: p.src[0] for p in afters}) if afters else cmdbuf
+
+def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))
+
+def make_signal(devs, queue=None, sentinel=False):
+  return make_placeholder(devs, 1, dtypes.uint64, "sentinel_signal" if sentinel else (queue, "timeline_signal") if queue else "timeline_signal", unique=False)
+def make_signal_value(devs, queue=None):
+  return make_placeholder(devs, 1, dtypes.uint64, (queue, "timeline_value") if queue else "timeline_value", unique=False)
+
+def make_submit(*cmds, devs:str|tuple[str, ...], queue:str) -> UOp:
+  return UOp.custom_function("submit_cmdbuf", UOp(Ops.LINEAR, dtypes.void, src=tuple(cmds), arg=(to_tuple(devs), queue)))
+def get_submit(ast:UOp) -> UOp: return next(u for u in ast.toposort() if u.op is Ops.CUSTOM_FUNCTION and u.arg == "submit_cmdbuf")
+
 @dataclass(frozen=True)
 class HCQInfo:
-  name:str = ""
-  estimates:Estimates = Estimates()
-  outs:tuple[int, ...] = ()
-  devs:tuple[str, ...] = ()
+  name:str
+  estimates:Estimates
+  device:tuple[str, ...]
+  queue:str
 
-  params:tuple[int, ...] = ()
+  input_idxs:tuple[int, ...] = () # indexes into input_uops used by this call
   inputs:int|None = None
 
-  @staticmethod
-  def from_call(call:UOp) -> HCQInfo: return HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), get_call_outs_ins(call)[0])
+# *****************
+# 0.1. prep: replace buffers with params
+
+def replace_call_buffers(ctx:list[UOp], call:UOp) -> UOp|None:
+  ctx += [s for s in dedup(call.src[1:]) if s not in ctx and s.op not in (Ops.PARAM, Ops.BIND)]
+  return call.replace(src=call.src[:1] + tuple(s if s.op in (Ops.PARAM, Ops.BIND) else s.param_like(ctx.index(s)) for s in call.src[1:]))
+pm_replace_buffers = PatternMatcher([(UPat(Ops.CALL, name="call"), replace_call_buffers)])
 
 # *****************
-# 1.1. prep runtimes: staging copies
+# 1.1. prep: staging copies
 
 def _need_staging(a, b): return all_devices_in(a.device, HCQ_DEVS) and not all_devices_in(b.device, HCQ_P2P_DEVS)
 
 def stage_copy(dst:UOp, src:UOp) -> UOp|None:
   if not (_need_staging(src, dst) or _need_staging(dst, src)): return None
 
-  stage = UOp.new_buffer("CPU", src.buffer.nbytes, dtypes.uint8)
+  stage = UOp.new_buffer("CPU", src.max_numel() * src.dtype.base.itemsize, dtypes.uint8)
   return UOp(Ops.LINEAR, dtypes.void, (src.copy_to_device("CPU").call(stage, src), stage.copy_to_device(dst.device).call(dst, stage)))
 pm_insert_copy_staging = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src"))), stage_copy)])
 
 # *****************
-# 1.2. prep runtimes: programs/kernargs
+# 2.1. tag hcq calls
 
-@functools.cache
-def get_pm_prep_program(name:str) -> PatternMatcher|None:
-  try:
-    importlib.import_module(f'tinygrad.runtime.ops_{name.lower()}') # TODO: remove that
-    return importlib.import_module(f'extra.hcq2.ops_{name.lower()}2').pm_prep_program
-  except ImportError: return None
+def tag_hcq_call(ctx:itertools.count, call:UOp) -> UOp:
+  if (hcq_devs:=next((b.device for b in call.src[1:] if all_devices_in(b.device, HCQ_DEVS)), None)) is None: return call
 
-def prep_program(call:UOp, prg:UOp) -> UOp|None:
-  dev = call.src[1].device
-  if (pm:=get_pm_prep_program(to_tuple(dev)[0].split(":")[0])) is None or (lowered:=pm.rewrite(prg)) is None: return None
-
-  data, image_bytes = lowered
-  buf = UOp.new_buffer(dev, len(image_bytes), dtypes.uint8).rtag("program")
-  blob = UOp(Ops.BINARY, dtypes.void, src=(), arg=image_bytes)
-  return prg.replace(src=(buf.after(buf.store(blob)),), arg=(data, prg.arg)).call(*call.src[1:], aux=HCQInfo.from_call(call))
-
-def prep_kernargs(call:UOp, prg:UOp) -> UOp:
-  data, info = prg.arg
-  patches = [(i*dtypes.uint64.itemsize, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], UOp(Ops.DEVICE, arg=call.src[1+gi].device))),
-              dtypes.uint64) for i,gi in enumerate(info.globals)] \
-          + [(len(info.globals)*dtypes.uint64.itemsize + i*dtypes.uint32.itemsize, v, dtypes.uint32) for i,v in enumerate(info.vars)]
-
-  buf = UOp.new_buffer(call.src[1].device, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
-  kernargs = buf.after(*tuple(buf.index(UOp.const(dtypes.int, o), dtype=buf.dtype.ptr()).cast(dt.ptr()).store(val.cast(dt)) for o, val, dt in patches))
-
-  return call.replace(src=(prg.replace(src=prg.src + (kernargs,), arg=(data, info)),) + call.src[1:])
-
-pm_prep_runtime = PatternMatcher([
-  # bind generic PROGRAM device to the call's actual dev(s), then run device-specific lowering
-  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(), UPat(), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"),),
-    name="call", allow_any_len=True), prep_program),
-
-  # lower kernargs (PROGRAM.src[0] is now AFTER(BUFFER, COPY) — the lowered program image)
-  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(Ops.BUFFER).or_after(),), name="prg"),), name="call", allow_any_len=True), prep_kernargs),
-])
+  queue = "COMPUTE:0" if call.src[0].op is Ops.PROGRAM else "COPY:0"
+  info = HCQInfo(get_call_name(call, get_call_arg_uops(call)), estimate_uop(call), to_tuple(hcq_devs), queue)
+  return call.replace(arg=replace(call.arg, aux=info)).rtag(next(ctx))
+pm_tag_hcq_calls = PatternMatcher([(UPat(Ops.LINEAR, name="linear"),
+  lambda ctx, linear: linear.replace(src=tuple(tag_hcq_call(ctx, s) for s in linear.src)))])
 
 # *****************
-# 2. lowering to hcq ir
-
-def make_submit(*cmds, devs:str|tuple[str, ...], queue:str) -> UOp:
-  devs:tuple[str, ...] = to_tuple(devs)
-  return UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(UOp(Ops.LINEAR, dtypes.void, src=tuple(cmds), arg=(devs, queue)),), arg="submit")
-
-def lower_program(call:UOp, prg:UOp) -> UOp:
-  return make_submit(prg, devs=call.src[1].device, queue="COMPUTE:0").sink().call(*call.src[1:], aux=call.arg.aux).rtag("hcq")
-
-def lower_copy(call:UOp, copy:UOp) -> UOp|None:
-  dst, src = call.src[1], call.src[2]
-  if (hcq_dev:=next((b.device for b in (dst, src) if b.device.split(":")[0] in HCQ_DEVS), None)) is None: return None
-
-  cp_op = UOp(Ops.COPY, dtypes.void, src=(dst, src), arg=src.buffer.nbytes)
-  return make_submit(cp_op, devs=hcq_dev, queue="COPY:0").sink().call(*call.src[1:], aux=HCQInfo.from_call(call)).rtag("hcq")
-
-pm_lower_ops = PatternMatcher([
-  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(Ops.BUFFER).or_after(), UPat(Ops.BUFFER).or_after()), name="prg"),),
-    name="call", allow_any_len=True), lower_program),
-  (UPat(Ops.CALL, src=(UPat(Ops.COPY, name="copy"),), name="call", allow_any_len=True), lower_copy),
-])
-
-# *****************
-# 3.1. deps tracking
+# 2.2. deps tracking
 # device.timeline_signal/value are the per-device schedule epoch. Before a schedule queue accesses memory owned by device N for the first time,
 # it waits for device[N].timeline_signal >= device[N].timeline_value - 1. This orders the schedule after all prior schedules that touched device N.
 #
@@ -267,129 +235,95 @@ pm_lower_ops = PatternMatcher([
 #
 # C programs reserve and bump timeline values, then patch command buffers with the concrete wait/signal values.
 
-@dataclass
-class DepsCtx:
-  deps:DepsTracker = field(default_factory=DepsTracker)
-  opid:itertools.count = field(default_factory=lambda: itertools.count(0))
-  last_per_queue:weakref.WeakValueDictionary[tuple[Any, str], UOp] = field(default_factory=weakref.WeakValueDictionary)
-  params:dict[tuple[int, int], Buffer] = field(default_factory=dict)
+class HCQDepsTracker(DepsTracker):
+  @staticmethod
+  def _key(buf:Any) -> tuple[Any, int, int]:
+    return (buf.arg.slot, 0, buf.max_numel() * buf.dtype.base.itemsize) if isinstance(buf, UOp) else DepsTracker._key(buf)
 
-def get_dep_buf(ctx:DepsCtx, u:UOp, lane:int) -> Buffer:
-  # TODO: should this be a part of DepsTracker?
-  if u.op is Ops.PARAM: return ctx.params.setdefault((u.arg.slot, lane), Buffer("NULL", u.max_numel(), u.dtype.base))
-  if u.op is Ops.MSTACK: return get_dep_buf(ctx, u.src[lane], 0)
-  if u.op in (Ops.SLICE, Ops.MSELECT): return get_dep_buf(ctx, u.src[0], u.arg if u.op is Ops.MSELECT else lane)
-  return b.bufs[lane] if isinstance(b:=u.buffer, MultiBuffer) else b
+def make_deps(u:UOp, dep_lanes:list[tuple[UOp, int, int]], nlanes:int) -> UOp:
+  deps:dict[UOp, list[int|None]] = collections.defaultdict(lambda: [None]*nlanes)
+  for dep, dlane, lane in dep_lanes: deps[dep][lane] = dlane
+  return u.after(*deps, arg=tuple(tuple(v) for v in deps.values()))
 
-def schedule_inner_sync(ctx:DepsCtx, linear:UOp) -> UOp:
-  new_src = []
-  for call in linear.src:
-    if call.tag != "hcq":
-      new_src.append(call)
-      continue
+def sched_sync(ctx:DepsTracker, call:UOp) -> UOp|None:
+  if not isinstance(call.arg.aux, HCQInfo): return None
 
-    q = get_submit(call.src[0]).src[0]
-    new_q = ctx.last_per_queue[q.arg] = q.rtag(next(ctx.opid))
+  refs = get_call_arg_uops(call)
+  outs, _ = get_call_outs_ins(call)
+  devices, queue = call.arg.aux.device, call.arg.aux.queue
 
-    refs = get_call_arg_uops(call)
-    deps = dedup(flatten(ctx.deps.access_resources([get_dep_buf(ctx, b, l) for b in refs], call.arg.aux.outs, new_q) for l in range(len(q.arg[0]))))
+  dep_lanes:list[tuple[UOp, int, int]] = []
+  for lane, d in enumerate(devices):
+    lane_refs = [b if b.op is Ops.PARAM else mb.bufs[lane] if isinstance(mb:=b.buffer, MultiBuffer) else mb for b in refs]
+    for dep, dlane in ctx.access_resources(lane_refs, outs, (call, lane)): dep_lanes.append((dep, dlane, lane))
 
-    # optims: keep only the max wait per queue, and drop self-queue waits when the queue self-orders
-    deps = {dep.arg:dep for dep in sorted(deps, key=lambda x: x.tag)}
-    if to_tuple(new_q.arg[0])[0].split(":")[0] in {"AMD", "QCOM"} or new_q.arg[1].startswith("COPY"):
-      deps.pop(new_q.arg, None)
+  if devices[0].split(":")[0] in {"AMD", "QCOM"} or queue.startswith("COPY"):
+    dep_lanes = [(dep, dlane, lane) for dep, dlane, lane in dep_lanes if (dep.arg.aux.device[dlane], dep.arg.aux.queue) != (devices[lane], queue)]
 
-    new_q = new_q.after(*deps.values()).rtag("deps") if deps else new_q
-    new_src.append(call.replace(src=(call.src[0].substitute({q:new_q}), *call.src[1:])))
-  return linear.replace(src=tuple(new_src))
-pm_schedule_inner_sync = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), schedule_inner_sync)])
+  # keep latest dep per (dep device, queue, cur lane)
+  latest = {((dep.arg.aux.device[dlane], dep.arg.aux.queue), lane): (dep, dlane) for dep, dlane, lane in sorted(dep_lanes, key=lambda x: x[0].tag)}
+  return make_deps(call, [(dep, dlane, lane) for (_, lane), (dep, dlane) in latest.items()], len(devices))
+pm_sched_sync = PatternMatcher([(UPat(Ops.CALL, name="call"), sched_sync)])
 
 # *****************
-# 3.2. finalizer
+# 2.3. merge into queues
 
-def make_finalizer(queues:list[UOp], nbump:int) -> UOp:
-  devs = tuple(dedup([d for q in queues for d in to_tuple(q.arg[0])]))
-  zero = UOp.const(dtypes.int, 0)
-  tl = make_signal_value(devs)
-
-  # queue is inc with deps
-  submit = make_submit(make_signal(devs).store(tl.index(zero)), devs=devs, queue="COMPUTE:0")
-  submit = submit.replace(src=(submit.src[0].after(*queues).rtag("deps"),))
-
-  upd = [(tl, 1)] + [(make_signal_value(devs, queue=qn), nbump) for qn in dedup([q.arg[1] for q in queues])]
-  patches = [s.after(submit).index(zero, dtype=s.dtype.ptr()).store(s.index(zero) + inc) for s, inc in upd]
-  return UOp.barrier(*patches).sink().call(aux=HCQInfo("hcq finalizer")).rtag("hcq")
-
-def add_finalizer(ctx:DepsCtx, linear:UOp) -> UOp:
-  parts:dict[str, list[UOp]] = collections.defaultdict(list)
-  for d, q in ctx.last_per_queue.items(): parts[to_tuple(d[0])[0].split(':')[0]].append(q)
-
-  nbump = next(ctx.opid)
-  return linear.replace(src=linear.src + tuple([make_finalizer(queues, nbump) for queues in parts.values()]))
-pm_add_finalizer = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), add_finalizer)])
-
-# *****************
-# 3.3. lower loads/stores
-
-def add_loads(ctx:set[int], deps:UOp) -> UOp:
-  cur_devs = to_tuple((cur:=deps.src[0]).arg[0])
-
-  waits = []
-  for dep in deps.src[1:]:
-    devs, queue = dep.arg
-    ctx.add(dep.tag) # mark op to update signal.
-
-    sig = make_mstack([make_signal(d, queue=queue, sentinel=d not in devs) for d in cur_devs])
-    val = make_signal_value(cur_devs, queue=queue).index(UOp.const(dtypes.int, 0))
-    waits.append(sig.wait(val + dep.tag))
-  return cur.replace(src=tuple(waits) + cur.src)
-pm_add_inner_loads = PatternMatcher([(UPat(Ops.AFTER, tag="deps", name="deps"), add_loads)])
-
-def add_stores(ctx:set[int], submit:UOp, q:UOp) -> UOp|None:
-  if q.tag not in ctx: return None
-  devs, queue = q.arg
-  src = q.src + (make_signal(devs, queue=queue).store(make_signal_value(devs, queue=queue).index(UOp.const(dtypes.int, 0)) + q.tag),)
-  return submit.replace(src=(q.replace(src=src, tag=None),))
-pm_add_inner_stores = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_stores)])
-
-# *****************
-# 4.1. merge queues
-
-def get_submit(ast:UOp) -> UOp: return next(u for u in ast.toposort() if u.op is Ops.CUSTOM_FUNCTION and u.arg == "submit")
-
-def merge_sink(sinks:list[UOp]) -> UOp:
-  if len(sinks) == 1: return sinks[0]
-  submits = [get_submit(sink) for sink in sinks]
-  queues = [submit.src[0] for submit in submits]
-  anchor = submits[-1].replace(src=(queues[-1].replace(src=tuple(x for q in queues for x in q.src)),))
-  for sink, submit in zip(sinks[:-1], submits[:-1]):
-    if sink.src[0] is not submit: anchor = sink.src[0].substitute({submit: anchor}, walk=True)
-  return sinks[-1].substitute({submits[-1]: anchor}, walk=True)
+def _merged_hcq_call(calls:list[UOp]):
+  info = replace(unwrap_after(calls[0]).arg.aux, estimates=sum((unwrap_after(c).arg.aux.estimates for c in calls), start=Estimates()))
+  cmdbuf = make_submit(*calls, devs=info.device, queue=info.queue)
+  return UOp.custom_function("hcq", cmdbuf.sink()).call(name="hcq", aux=info)
 
 def merge_queues(linear:UOp) -> UOp:
   new_src:list[UOp] = []
-  opened_qs:dict[tuple[tuple[str, ...], str], tuple[list[UOp], HCQInfo]] = {} # (devs, queue) -> (sinks, aux), kept in submit order
+  opened_qs:dict[tuple[tuple[str, ...], str], list[UOp]] = {} # (devs, queue) -> list of calls, kept in submit order
 
   for call in linear.src:
-    if call.tag != "hcq":
-      new_src += [merge_sink((sa:=opened_qs.pop(k))[0]).call(aux=sa[1]).rtag("hcq") for k in list(opened_qs)] + [call]
+    if not isinstance(unwrap_after(call).arg.aux, HCQInfo):
+      new_src += [_merged_hcq_call(opened_qs.pop(k)) for k in list(opened_qs)] + [call]
       continue
 
-    devs, queue = get_submit(new_sink:=call.src[0]).src[0].arg
-    new_rec = ([new_sink], call.arg.aux)
-    if (old:=opened_qs.pop((devs, queue), None)) is not None:
-      new_rec = (old[0] + [new_sink], replace(new_rec[1], name=f"{queue.lower()} submit", estimates=old[1].estimates + new_rec[1].estimates))
+    devices, queue = unwrap_after(call).arg.aux.device, unwrap_after(call).arg.aux.queue
+
+    if (old:=opened_qs.pop((devices, queue), None)) is not None: new_rec = old + [call]
     else:
       # no such queue opened: close every open submit on this queue that shares a device, so submit order is kept
-      closing = [k for k in opened_qs if k[1] == queue and set(k[0]) & set(devs)]
-      new_src += [merge_sink((sa:=opened_qs.pop(k))[0]).call(aux=sa[1]).rtag("hcq") for k in closing]
-    opened_qs[(devs, queue)] = new_rec
-
-  return linear.replace(src=tuple(new_src + [merge_sink(sinks).call(aux=aux).rtag("hcq") for sinks, aux in opened_qs.values()]))
+      closing = [k for k in opened_qs if k[1] == queue and set(k[0]) & set(devices)]
+      new_src += [_merged_hcq_call(opened_qs.pop(k)) for k in closing]
+      new_rec = [call]
+    opened_qs[(devices, queue)] = new_rec
+  return linear.replace(src=tuple(new_src + [_merged_hcq_call(c) for c in opened_qs.values()]))
 pm_merge_queues = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), merge_queues)])
 
 # *****************
-# 4.2. global sync
+# 2.4. finalizer
+
+def add_finalizer(ctx:itertools.count, linear:UOp) -> UOp:
+  # collect by device type
+  parts:dict[str, list[UOp]] = collections.defaultdict(list)
+  for call in linear.src:
+    if (c:=unwrap_after(call)).src[0].op is not Ops.CUSTOM_FUNCTION or c.src[0].arg != "hcq": continue
+    parts[c.arg.aux.device[0].split(':')[0]].append(unwrap_after(get_submit(call).src[0].src[0]))
+
+  nbump = next(ctx)
+  finalizers = []
+  for calls in parts.values():
+    devs = tuple(dedup(d for call in calls for d in unwrap_after(call).arg.aux.device))
+    zero = UOp.const(dtypes.int, 0)
+    tl = make_signal_value(devs)
+
+    # split each (multi-device) call into per-device deps, then store the device timeline value into the device signal after them
+    dep_lanes = [(call, dlane, devs.index(d)) for call in calls for dlane, d in enumerate(unwrap_after(call).arg.aux.device)]
+    store = make_deps(make_signal(devs).store(tl.index(zero)), dep_lanes, len(devs))
+    submit = make_submit(store, devs=devs, queue="COMPUTE:0")
+
+    upd = [(tl, 1)] + [(make_signal_value(devs, queue=qn), nbump) for qn in dedup([unwrap_after(call).arg.aux.queue for call in calls])]
+    patches = [s.after(submit).index(zero, dtype=s.dtype).store(s.index(zero) + inc) for s, inc in upd]
+    finalizers.append(UOp.custom_function("hcq", UOp.barrier(*patches).sink()).call(aux=HCQInfo("hcq finalizer", Estimates(), devs, "COMPUTE:0")))
+  return linear.replace(src=linear.src + tuple(finalizers))
+pm_add_finalizer = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), add_finalizer)])
+
+# *****************
+# 2.4. global sync
 
 def add_global_sync(ctx:set[tuple[str, ...]], submit:UOp, q:UOp) -> UOp|None:
   if (devs:=q.arg[0]) in ctx: return None
@@ -398,94 +332,143 @@ def add_global_sync(ctx:set[tuple[str, ...]], submit:UOp, q:UOp) -> UOp|None:
   # some devices from a command buffer might be used for the first time this schedule, so we wait for their global timeline epoch.
   wait = make_signal(devs).wait(make_signal_value(devs).index(UOp.const(dtypes.int, 0)) - 1)
   return submit.replace(src=(q.replace(src=(UOp(Ops.BARRIER, dtypes.void), wait, *q.src)),))
-pm_add_global_sync = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_global_sync)])
+pm_add_global_sync = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_global_sync)])
 
 # *****************
-# 4.3. annotate exec devs
+# 3.1. lower loads/stores
 
-pm_annotate_devs = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"),
-  lambda call: call.replace(arg=replace(call.arg, aux=replace(call.arg.aux, devs=get_submit(call.src[0]).src[0].arg[0]))))])
+def add_loads(ctx:set[int], submit:UOp, q:UOp) -> UOp|None:
+  cur_devs = q.arg[0]
+  new_src:list[UOp] = []
+  for s in q.src:
+    if s.op is Ops.AFTER:
+      for lanes, dep in zip(s.arg, s.src[1:]):
+        devs, queue = dep.arg.aux.device, dep.arg.aux.queue
+        ctx.add(dep.tag) # mark op to update signal.
+
+        sig = make_mstack([make_signal(d if dl is None else devs[dl], queue=queue, sentinel=dl is None) for dl, d in zip(lanes, cur_devs)])
+        val = make_mstack([make_signal_value(d if dl is None else devs[dl], queue=queue) for dl, d in zip(lanes, cur_devs)]).index(UOp.const(dtypes.int, 0))
+        new_src.append(sig.wait(val + dep.tag))
+      s = s.src[0]
+    new_src.append(s)
+  return submit.replace(src=(q.replace(src=tuple(new_src)),))
+pm_add_inner_loads = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_loads)])
+
+def add_stores(ctx:set[int], submit:UOp, q:UOp) -> UOp|None:
+  devs, queue = q.arg
+  new_src:list[UOp] = []
+  for op in q.src:
+    new_src.append(op)
+    if (sigval:=unwrap_after(op).tag) in ctx:
+      new_src.append(make_signal(devs, queue=queue).store(make_signal_value(devs, queue=queue).index(UOp.const(dtypes.int, 0)) + sigval))
+  return submit.replace(src=(q.replace(src=tuple(new_src)),))
+pm_add_inner_stores = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),), name="submit"), add_stores)])
 
 # *****************
-# 4.4. replace params with per-submit input address loads
+# 2.1. hcq lowering: programs
 
-def replace_params(call:UOp) -> UOp|None:
-  if not (params:={u:u.arg.slot for u in call.src[0].toposort() if u.op is Ops.PARAM and u.addrspace is not None}): return None
-
-  # fill new info
-  hcqinfo = replace(call.arg.aux, params=tuple(sorted(set(params.values()))), inputs=len(get_call_arg_uops(call)))
-
-  inputs = UOp.new_buffer(get_submit(call.src[0]).src[0].arg[0], len(hcqinfo.params), dtypes.uint64).rtag("inputs")
-
-  slot2idx = {s:i for i,s in enumerate(hcqinfo.params)}
-  body = call.src[0].substitute({u:inputs.index(UOp.const(dtypes.int, slot2idx[s])).load() for u,s in params.items()})
-
-  return call.replace(src=(body, *call.src[1:], inputs), arg=replace(call.arg, aux=hcqinfo))
-pm_replace_params = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), replace_params)])
+def encode_kernargs_clike(call:UOp, prg:UOp, devs:str|tuple[str, ...]) -> UOp:
+  data, info = prg.arg
+  call_args = get_call_arg_uops(call)
+  buf = make_placeholder(devs, data.kernargs_alloc_size, dtypes.uint8, name="kernargs")
+  patches = [make_patch(buf, i*8, make_getaddr(call_args[gi], devs)) for i,gi in enumerate(info.globals)] \
+          + [make_patch(buf, len(info.globals)*8 + i*4, v, dtypes.uint32) for i,v in enumerate(info.vars)]
+  return buf.after(*patches)
 
 # *****************
-# 5.1. encode cmdbufs
-
-@functools.cache
-def get_pm_lower(name:str) -> PatternMatcher|None:
-  try:
-    importlib.import_module(f'tinygrad.runtime.ops_{name.lower()}') # TODO: remove that
-    return importlib.import_module(f'extra.hcq2.ops_{name.lower()}2').pm_lower
-  except ImportError: return None
+# 2.2. hcq lowering: ops to ir
 
 def encode_cmdbuf(submit:UOp, lin:UOp) -> UOp|None:
-  if (pm:=get_pm_lower(to_tuple(lin.arg[0])[0].split(":")[0])) is None: return None
-  return pm.rewrite(submit)
-pm_encode_cmdbufs = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit", src=(UPat(Ops.LINEAR, name="lin"),), name="submit"), encode_cmdbuf)])
+  if (pm:=Device.get_class(lin.arg[0][0]).pm_lower) is None: return None
+  return graph_rewrite(submit, pm, name=f"encode {lin.arg[0]}", enter_calls=True)
+pm_encode_cmdbufs = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="lin"),), name="submit"), encode_cmdbuf)])
 
 # *****************
-# 5.2. lift patches to the command buffer (root)
 
-def lift_patches_to_cmdbuf(cmdbuf:UOp) -> UOp|None:
-  if not (patches:=dedup(u for store in cmdbuf.src[1:] for u in store.toposort() if u.op is Ops.AFTER)): return None
-  deps = tuple(d for p in patches for d in p.src[1:])
-  return cmdbuf.replace(src=cmdbuf.src + deps).substitute({p: p.src[0] for p in patches})
-pm_lift_patches_to_cmdbuf = PatternMatcher([
-  (UPat(Ops.AFTER, src=(UPat(Ops.BUFFER, tag={"compute", "copy"}),), allow_any_len=True, name="cmdbuf"), lift_patches_to_cmdbuf),
+pm_early_simplify = PatternMatcher([
+  # getaddr(slice(base, off)) -> getaddr(base) + byte offset
+  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, src=(UPat.var("base"), UPat.cvar("off"))),), name="g"),
+   lambda g, base, off: g.replace(src=(base,)) + UOp.const(dtypes.uint64, off.arg * base.dtype.itemsize)),
 ])
 
+# *****************
+
+def replace_params(call:UOp) -> UOp|None:
+  gaddrs = [u for u in call.src[0].toposort(enter_calls=False) if u.op is Ops.GETADDR and u.src[0].op is Ops.PARAM and u.src[0].tag is None]
+  if not gaddrs: return None
+
+  idxs:dict[int, int] = {}
+  for g in gaddrs: idxs.setdefault(g.src[0].arg.slot, len(idxs))
+
+  inputs = make_placeholder(call.arg.aux.device, len(idxs), dtypes.uint64, "inputs")
+  body = call.src[0].substitute({g: inputs.index(UOp.const(dtypes.int, idxs[g.src[0].arg.slot])).load() for g in gaddrs})
+  return call.replace(src=(body, *call.src[1:], inputs), arg=replace(call.arg, aux=replace(call.arg.aux, input_idxs=tuple(idxs))))
+pm_replace_params = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), replace_params)])
+
+# *****************
+
+def changes_per_submit(u:UOp) -> bool: return u.op in (Ops.LOAD, Ops.INDEX) or (u.op is Ops.PARAM and u.tag is None)
+def is_placeholder(b:UOp) -> bool: return (b.op is Ops.PARAM and b.tag is not None) or (b.op is Ops.MSTACK and all(is_placeholder(x) for x in b.src))
+def is_link_patch(s:UOp) -> bool: return is_placeholder(s.buf_uop) and not any(changes_per_submit(u) for u in s.backward_slice)
+
+def trim_link_patches(ctx:list[UOp], a:UOp) -> UOp|None:
+  links, kept = partition(a.src[1:], is_link_patch)
+  ctx += links
+  return a.src[0].after(*kept) if links else None
+pm_trim_link_patches = PatternMatcher([(UPat(Ops.AFTER, src=(UPat((Ops.PARAM, Ops.MSTACK)),), allow_any_len=True, name="a"), trim_link_patches)])
+
+def split_patches(call:UOp) -> UOp|None:
+  # trim link-time patches
+  body = graph_rewrite(call.src[0], pm_trim_link_patches, ctx=(lt_patches:=[]), name=f"trim link-time patches ({call.arg.aux.name})")
+
+  units:dict[UOp, None] = {}
+  def unit_gate(u:UOp) -> bool:
+    if (is_plc:=is_placeholder(u)): units[u] = None
+    return not is_plc
+  body.toposort(gate=unit_gate)
+
+  srcs = dedup(list(call.src[1:]) + list(units) + [s.buf_uop for s in lt_patches])
+  param_sub = {u: UOp.param(i, u.dtype, device=u.device) for i,u in enumerate(srcs)}
+  for b in dedup(s.buf_uop for s in lt_patches):
+    idx = param_sub[b].arg.slot
+    srcs[idx] = srcs[idx].after(*dedup(s for s in lt_patches if s.buf_uop is b))
+
+  param_sub |= {v: v.replace(arg=replace(v.arg, slot=-1)) for v in body.variables() if v.op is Ops.PARAM}
+
+  info = replace(call.arg.aux, inputs=next((i for i,u in enumerate(srcs) if unwrap_after(u).tag == "inputs"), None))
+  return call.replace(src=(body.substitute(param_sub), *srcs), arg=replace(call.arg, aux=info))
+pm_split_patches = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), split_patches)])
+
 # *****************
 # 5.3. pack placeholders buffers
 
 def pack_hcq_placeholders(call:UOp) -> UOp|None:
-  bufs = [b for b in call.src[0].toposort() if b.op is Ops.BUFFER and b.tag in (maxtags:={"scratch"}) | (sumtags:={"program", "kernargs"})]
+  bufs = [b for b in call.src[0].toposort() if b.op is Ops.PARAM and b.tag in (maxtags:={"scratch"}) | (sumtags:={"program", "kernargs"})]
 
   off_per_buf:dict[UOp, int] = {}
   size_per_tag:dict[str, int] = {}
   for b in bufs:
-    if b.tag in maxtags: size_per_tag[b.tag] = max(size_per_tag.get(b.tag, 0), b.arg)
+    bsz = b.max_numel()
+    if b.tag in maxtags: size_per_tag[b.tag] = max(size_per_tag.get(b.tag, 0), bsz)
     elif b.tag in sumtags:
       off_per_buf[b] = round_up(size_per_tag.get(b.tag, 0), {"program": 0x1000}.get(b.tag, 128))
-      size_per_tag[b.tag] = off_per_buf[b] + b.arg
+      size_per_tag[b.tag] = off_per_buf[b] + bsz
 
   count_per_tag = collections.Counter(b.tag for b in bufs)
   ref_bufs = {b.tag:b for b in bufs if count_per_tag[b.tag] > 1}
-  bases = {tag:UOp.new_buffer(b.src[1].arg, size_per_tag[tag], b.dtype).rtag(tag) for tag,b in ref_bufs.items()}
-  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.weakint, off_per_buf.get(b, 0))), b.arg) for b in bufs if b.tag in bases}
+  bases = {tag:UOp.new_buffer(b.device, size_per_tag[tag], b.dtype).rtag(tag) for tag,b in ref_bufs.items()}
+  subs = {b:UOp(Ops.SLICE, b.dtype, (bases[b.tag], UOp.const(dtypes.weakint, off_per_buf.get(b, 0))), b.max_numel()) for b in bufs if b.tag in bases}
   return call.replace(src=(call.src[0].substitute(subs, walk=True), *call.src[1:])) if subs else None
-pm_pack_placeholders = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), pack_hcq_placeholders)])
-
-# *****************
-# 5.4. capture buffers reachable from each hcq call as BIND, so we don't drop their refs
-
-def hold_call_buffers(call:UOp) -> UOp|None:
-  if not (bufs:=tuple(dedup(u for u in call.src[0].toposort() if u.op is Ops.BUFFER and u not in call.src))): return None
-  return call.replace(src=call.src + (UOp(Ops.BIND, dtypes.void, src=bufs),))
-pm_hold_call_buffers = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), hold_call_buffers)])
+pm_pack_placeholders = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="hcq"),), name="call", allow_any_len=True), pack_hcq_placeholders)])
 
 # *****************
 # 6. bufferize placeholders: replace placeholders with real buffers.
 
 def bufferize_buf(buf:UOp) -> UOp|None:
   if buf.tag is None: return None
-  uops = tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=dv), "CPU") for dev in to_tuple(buf.src[1].arg))
+  uops = tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=dv), "CPU") for dev in to_tuple(buf.device))
   return make_mstack(uops)
-pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, name="buf"), bufferize_buf)])
+pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, name="buf"), bufferize_buf)])
 
 # *****************
 # 7. resolve patches
@@ -499,84 +482,79 @@ def fold_blob_store(buf:UOp, blob:UOp) -> UOp:
 
 def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
   for b, v in zip((bs:=mb.bufs if isinstance((mb:=buf.buffer), MultiBuffer) else (mb,)), val.src if val.op is Ops.STACK else (val,)*len(bs)):
-    struct.pack_into(f'<{v.dtype.fmt}', b.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * buf.dtype.base.itemsize, v.arg)
+    struct.pack_into(f'<{v.dtype.fmt}', b.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * buf.dtype.base.itemsize, truncate[v.dtype](v.arg))
   return UOp(Ops.NOOP)
 
 def resolve_getaddr(buf:UOp, g:UOp) -> UOp:
   if buf.op not in (Ops.BUFFER, Ops.MSTACK, Ops.MSELECT): return buf
-  if isinstance(b:=buf.buffer, Buffer): return UOp.const(dtypes.uint64, b.get_buf(g.src[1].arg).va_addr)
-  return UOp(Ops.STACK, dtypes.uint64.vec(len(b.bufs)), tuple(UOp.const(dtypes.uint64, x.ensure_allocated()._buf.va_addr) for x in b.bufs))
+  devs, b = to_tuple(g.arg), buf.buffer
+  bufs = tuple(cast(Buffer, x.buffer) for x in buf.src) if buf.op is Ops.MSTACK else tuple(b.bufs if isinstance(b, MultiBuffer) else (b,)*len(devs))
+  assert len(bufs) == len(devs), f"can't resolve {len(bufs)} buffers on {len(devs)} devices"
+  addrs = tuple(UOp.const(dtypes.uint64, x.get_buf(d).va_addr) for x, d in zip(bufs, devs))
+  return addrs[0] if len(addrs) == 1 else UOp(Ops.STACK, dtypes.uint64.vec(len(addrs)), addrs)
 
-def resolve_getaddr_slice(bv:UOp, dev:UOp) -> UOp:
+def resolve_getaddr_slice(bv:UOp, g:UOp) -> UOp:
   itemsize = bv.src[0].dtype.itemsize if unwrap_after(bv.src[0]).op in (Ops.BUFFER, Ops.SLICE, Ops.MSTACK, Ops.MSELECT) else bv.dtype.itemsize
-  return UOp(Ops.GETADDR, dtypes.uint64, src=(bv.src[0], dev)) + UOp.const(dtypes.uint64, bv.src[1].arg * itemsize)
+  return UOp(Ops.GETADDR, dtypes.uint64, src=(bv.src[0],), arg=g.arg) + UOp.const(dtypes.uint64, bv.src[1].arg * itemsize)
 
 pm_resolve_patches = PatternMatcher([
   # multi
   (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
   (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),
 
-  # index on slice is index
-  (UPat(Ops.INDEX, src=(UPat(Ops.SLICE, name="bv"), UPat()), name="idx", allow_any_len=True),
-    lambda idx, bv: idx.replace(src=(bv.src[0], idx.src[1] + bv.src[1].cast(idx.src[1].dtype), *idx.src[2:]))),
+  # shrink on slice is shrink on base at offset
+  (UPat(Ops.SHRINK, src=(UPat(Ops.SLICE, name="bv"), UPat(), UPat()), name="shr"),
+    lambda shr, bv: shr.replace(src=(bv.src[0], shr.src[1] + bv.src[1].cast(shr.src[1].dtype), shr.src[2]))),
 
   # getaddr
-  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"), UPat(Ops.DEVICE, name="dev"))), resolve_getaddr_slice), # getaddr(slice(x)) -> offset+getaddr(x)
-  (UPat(Ops.GETADDR, src=(UPat(name="buf"), UPat(Ops.DEVICE)), name="g"), resolve_getaddr),
+  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"),), name="g"), resolve_getaddr_slice), # getaddr(slice(x)) -> offset+getaddr(x)
+  (UPat(Ops.GETADDR, src=(UPat(name="buf"),), name="g"), resolve_getaddr),
 
   # folders
   (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
-  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").index(UPat.cvar("off")).or_casted().store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))),
-    fold_const_store),
-]) + symbolic_simple
+  (UPat(Ops.SHRINK, src=(UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf"), UPat.cvar("off"), UPat(Ops.CONST))).bitcast()
+    .store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))), fold_const_store),
+])
 
 # *****************
 # 8. callify hcq programs
 
-def to_param(bufs:list[UOp], ref:UOp) -> UOp:
-  if ref not in bufs: bufs.append(ref)
-  return UOp.placeholder((ref.buffer.size,), ref.dtype, bufs.index(ref))
-pm_to_param = PatternMatcher([(UPat({Ops.MSELECT, Ops.MSTACK, Ops.BUFFER}, name="r"), lambda ctx, r: to_param(ctx, r))])
+pm_callify_hcq = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, arg="hcq", src=(UPat(Ops.SINK),), name="cf"),
+  lambda cf: cf.replace(src=(to_program(cf.src[0].replace(arg=KernelInfo("hcq_submit"), tag=1), Device["CPU"].renderer),)))])
 
-def parametrize_host_buffers(call:UOp) -> UOp:
-  # preserve original order of args
-  body = graph_rewrite(call.src[0], pm_to_param, ctx=(bufs:=list(get_call_arg_uops(call))), bottom_up=True, name="parametrize host buffers")
+hcq_compile_cache:dict[bytes, tuple[UOp, tuple[UOp, ...]]] = {}
 
-  # move vars to new slots
-  var_slots = {nm:len(bufs)+i for i,nm in enumerate(sorted({v.expr for v in body.variables() if v.op is Ops.PARAM}))}
-  body = body.substitute({v:v.replace(arg=replace(v.arg, slot=var_slots[v.expr])) for v in body.variables() if v.op is Ops.PARAM})
+@track_rewrites(lambda linear,input_uops,ret: f"HCQ Compile {pluralize('Kernel', len(ret.src))}")
+def hcq_compile(linear:UOp, input_uops:list[UOp]|None=None) -> UOp:
+  if input_uops is not None: linear = graph_rewrite(linear, pm_replace_buffers, ctx=input_uops, walk=True, enter_calls=True, name="replace buffer")
 
-  return call.replace(src=(body, *bufs) + tuple(x for x in call.src[1:] if x.op is Ops.BIND))
-pm_parametrize_host_buffers = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), parametrize_host_buffers)])
+  if (final_linear:=(hcq_compile_cache.get(cache_key:=linear.key))) is None:
+    # schedule
+    linear = linear.substitute(back_map:={s.param_like(i): s for i,s in enumerate(input_uops)} if input_uops is not None else {}, walk=True)
+    linear = graph_rewrite(linear, pm_insert_copy_staging + pm_flatten_linear, name="insert copy staging")
+    linear = graph_rewrite(linear, pm_tag_hcq_calls, ctx=(enumerator:=itertools.count(0)), walk=True, name="tag hcq calls")
+    linear = graph_rewrite(linear, pm_sched_sync, ctx=HCQDepsTracker(), walk=True, name="schedule sync")
+    linear = linear.substitute({s: p for p, s in back_map.items()}, walk=True)
+    linear = graph_rewrite(linear, pm_merge_queues, walk=True, name="merge queues")
+    linear = graph_rewrite(linear, pm_add_finalizer, ctx=enumerator, walk=True, name="add finalizer")
+    linear = graph_rewrite(linear, pm_add_global_sync, ctx=set(), walk=True, name="add global sync", enter_calls=True)
 
-def callify_hcq(call:UOp) -> UOp:
-  prg = to_program(call.src[0].sink(arg=KernelInfo("hcq_submit"), tag=1), Device["CPU"].renderer)
-  return UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(prg,), arg="hcq").call(*call.src[1:], aux=call.arg.aux)
-pm_callify_hcq = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), callify_hcq)])
+    # lowering to hcq ir
+    linear = graph_rewrite(linear, pm_add_inner_loads, ctx=(waited:=set()), walk=True, name="add loads", enter_calls=True)
+    linear = graph_rewrite(linear, pm_add_inner_stores, ctx=waited, walk=True, name="add stores", enter_calls=True)
+    linear = graph_rewrite(linear, pm_encode_cmdbufs, walk=True, name="encode cmdbufs", enter_calls=True)
 
-@track_rewrites(lambda _,ret: f"HCQ Schedule {pluralize('Kernel', len(ret.src))}")
-def hcq_schedule(linear:UOp) -> UOp:
-  linear = graph_rewrite(linear, pm_insert_copy_staging + pm_flatten_linear, name="insert copy staging")
-  linear = graph_rewrite(linear, pm_prep_runtime, name="prepare runtime")
+    # pie
+    linear = graph_rewrite(linear, pm_early_simplify + symbolic, bottom_up=False, name="early simplify patches")
+    linear = graph_rewrite(linear, pm_replace_params, walk=True, name="replace params")
+    linear = graph_rewrite(linear, pm_split_patches, walk=True, name="split patches")
 
-  linear = graph_rewrite(linear, pm_lower_ops, name="lower ops into hcq ir")
-  linear = graph_rewrite(linear, pm_schedule_inner_sync, ctx=(deps_ctx:=DepsCtx()), walk=True, name="schedule inner sync")
-  linear = graph_rewrite(linear, pm_add_finalizer, ctx=deps_ctx, walk=True, name="add finalizer")
-  linear = graph_rewrite(linear, pm_add_inner_loads, ctx=(waited:=set()), walk=True, name="add loads", enter_calls=True)
-  linear = graph_rewrite(linear, pm_add_inner_stores, ctx=waited, walk=True, name="add stores", enter_calls=True)
-  linear = graph_rewrite(linear, pm_merge_queues, name="merge queues")
-  linear = graph_rewrite(linear, pm_add_global_sync, ctx=set(), walk=True, name="add global sync", enter_calls=True)
-  linear = graph_rewrite(linear, pm_annotate_devs, name="annotate devs")
-  linear = graph_rewrite(linear, pm_replace_params, name="replace params")
-  linear = graph_rewrite(linear, pm_encode_cmdbufs, walk=True, name="encode cmdbufs", enter_calls=True)
-  linear = graph_rewrite(linear, pm_lift_patches_to_cmdbuf, name="lift patches to cmdbuf", enter_calls=True)
-  linear = graph_rewrite(linear, pm_pack_placeholders, walk=True, name="pack placeholders")
-  linear = graph_rewrite(linear, pm_hold_call_buffers, walk=True, name="hold call buffers")
+    # and compile it
+    final_linear = hcq_compile_cache[cache_key] = graph_rewrite(linear, pm_callify_hcq, name="callify hcq", enter_calls=True)
 
-  # realize starts from here
-  linear = graph_rewrite(linear, pm_bufferize, bottom_up=True, walk=True, name="bufferize placeholders", enter_calls=True)
-  linear = graph_rewrite(linear, pm_resolve_patches, bottom_up=False, name="simplify patches", enter_calls=True)
-  linear = graph_rewrite(linear, pm_parametrize_host_buffers, walk=True, name="parametrize host buffers")
-  linear = graph_rewrite(linear, pm_callify_hcq, name="callify hcq")
+  return final_linear
 
-  return linear
+@track_rewrites(lambda _,ret: f"HCQ Link {pluralize('Kernel', len(ret.src))}")
+def hcq_link(linear:UOp) -> UOp:
+  linear = graph_rewrite(linear, pm_bufferize, bottom_up=True, walk=True, name="bufferize placeholders")
+  return graph_rewrite(linear, pm_resolve_patches + symbolic, bottom_up=False, name="simplify patches")

extra/hcq2/ops_amd2.py

@@ -1,9 +1,9 @@
 from __future__ import annotations
-from typing import cast
+from typing import cast, Any, Callable
 import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit
 assert sys.platform != 'win32'
 from dataclasses import dataclass
-from extra.hcq2.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, make_getaddr, make_ins, make_cmdbuf
+from extra.hcq2.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_getaddr, make_ins, make_cmdbuf, make_placeholder
 from tinygrad.uop.ops import sint, UOp
 from tinygrad.device import Compiled, BufferSpec, Buffer, Device
 from tinygrad.dtype import dtypes
@@ -89,25 +89,26 @@ def memory_barrier(ctx):
                  reg_done=getattr(ctx.nbio, f'regBIF_BX_PF{pf}_GPU_HDP_FLUSH_DONE').addr[0], value=0xffffffff),
     acquire_mem(ctx)))
 
-def pm4_wait(ctx, dst, val): return wait_reg_mem(ctx, val, mem=make_getaddr(dst, ctx.device))
+def pm4_wait(ctx, dst, val): return wait_reg_mem(ctx, val, mem=make_getaddr(dst, ctx.devs))
 
 def pm4_barrier(ctx): return memory_barrier(ctx)
 
 def pm4_store(ctx, dst, val):
   if val.op is Ops.BINARY: return None
-  return release_mem(ctx, make_getaddr(dst, ctx.device), val, ctx.pm4.data_sel__mec_release_mem__send_32_bit_low,
+  return release_mem(ctx, make_getaddr(dst, ctx.devs), val, ctx.pm4.data_sel__mec_release_mem__send_32_bit_low,
                      ctx.pm4.int_sel__mec_release_mem__send_interrupt_after_write_confirm, cache_flush=True)
 
 def pm4_timestamp(ctx, dst):
-  return release_mem(ctx, make_getaddr(dst, ctx.device), 0, ctx.pm4.data_sel__mec_release_mem__send_gpu_clock_counter,
+  return release_mem(ctx, make_getaddr(dst, ctx.devs), 0, ctx.pm4.data_sel__mec_release_mem__send_gpu_clock_counter,
                      ctx.pm4.int_sel__mec_release_mem__none)
 
-def pm4_program(ctx, prg):
+def pm4_program(ctx, call, prg):
   data, info = prg.arg
-  lib_gpu, args = prg.src
-  prog_addr = make_getaddr(lib_gpu, ctx.device) + data.entry_point_offset
-  scratch_addr = make_getaddr(UOp.new_buffer(lib_gpu.device, data.private_segment_size, dtypes.uint8).rtag("scratch"), ctx.device)
-  args_addr = make_getaddr(args, ctx.device)
+  lib_gpu = prg.src[0]
+  args = encode_kernargs_clike(call, prg, ctx.devs)
+  prog_addr = make_getaddr(lib_gpu, ctx.devs) + data.entry_point_offset
+  scratch_addr = make_getaddr(make_placeholder(ctx.devs, data.private_segment_size, dtypes.uint8, "scratch", unique=False), ctx.devs)
+  args_addr = make_getaddr(args, ctx.devs)
 
   user_regs = []
   if data.enable_private_segment_sgpr:
@@ -134,19 +135,20 @@ def pm4_program(ctx, prg):
   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),
+
   (UPat(Ops.WAIT, src=(UPat(name="dst"), UPat(name="val"))), pm4_wait),
   (UPat(Ops.BARRIER), pm4_barrier),
-  (UPat(Ops.PROGRAM, name="prg"), pm4_program),
   (UPat(Ops.CUSTOM_FUNCTION, arg="timestamp", src=(UPat(name="dst"),)), pm4_timestamp),
   (UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), pm4_store),
 ])
 
 def pm4_submit(cmdbuf, devs):
-  size, zero = UOp.const(dtypes.uint32, cmdbuf.src[0].arg // dtypes.uint32.itemsize), UOp.const(dtypes.int, 0)
+  size, zero = UOp.const(dtypes.uint32, cmdbuf.nbytes() // dtypes.uint32.itemsize), UOp.const(dtypes.int, 0)
 
   # the compute queue's ring and its host-side ring/write/put pointers (placeholders, resolved in pm_bufferize)
   for d in devs: q = Device[d].compute_queue
-  ring, wptr, doorbell, put_ptr = (UOp.new_buffer(devs, b.size, b.dtype).rtag(("COMPUTE:0", name))
+  ring, wptr, doorbell, put_ptr = (make_placeholder(devs, b.size, b.dtype, ("COMPUTE:0", name), unique=False)
     for name, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))
 
   # place the cmdbuf at the ring's write offset, wrapping the ring
@@ -156,60 +158,63 @@ def pm4_submit(cmdbuf, devs):
   ring_idx = ((put + i.cast(put.dtype)) % q.ring.size).cast(dtypes.int)
 
   # copy the cmdbuf into the ring and advance the put/write pointers
-  copy_to_ring = ring.index(ring_idx, dtype=ring.dtype.ptr()).store(
-    cmdbuf.index(i*4, dtype=cmdbuf.dtype.ptr()).cast(dtypes.uint32.ptr()).load()).end(i)
-  bump_put_ptr = put_ptr.index(zero, dtype=put_ptr.dtype.ptr()).store(next_put)
-  bump_wptr = wptr.index(zero, dtype=wptr.dtype.ptr()).store(next_put)
+  copy_to_ring = ring.index(ring_idx, ptr=True).store(
+    cmdbuf.index(i*4, ptr=True).cast(dtypes.uint32.ptr()).load()).end(i)
+  bump_put_ptr = put_ptr.index(zero, ptr=True).store(next_put)
+  bump_wptr = wptr.index(zero, ptr=True).store(next_put)
 
   # ring the doorbell once the copy and pointer bumps have landed
   flush = UOp.barrier(copy_to_ring, bump_put_ptr, bump_wptr)
-  return doorbell.after(flush).index(zero, dtype=doorbell.dtype.ptr()).store(next_put)
+  return doorbell.after(flush).index(zero, ptr=True).store(next_put)
 
 pm_pm4_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"),
-  lambda lin: pm4_submit(make_cmdbuf(lin, to_tuple(lin.arg[0]), "compute"), to_tuple(lin.arg[0])))])
+  lambda lin: pm4_submit(make_cmdbuf(lin, to_tuple(lin.arg[0])), to_tuple(lin.arg[0])))])
 
 # *****************
 # SDMA
 
 class SDMAOps(FastEnum): COPY = auto(); POLL_REGMEM = auto(); FENCE = auto(); TRAP = auto(); TIMESTAMP = auto()  # noqa: E702
 
-def sdma_copy(ctx, dst, src, copy):
-  src_addr, dst_addr = make_getaddr(src, ctx.device), make_getaddr(dst, ctx.device)
+def sdma_copy(ctx, call):
+  dst, src = call.src[1], call.src[2]
+  sz = src.max_numel() * src.dtype.base.itemsize
+  src_addr, dst_addr = make_getaddr(src, ctx.devs), make_getaddr(dst, ctx.devs)
   return UOp(Ops.LINEAR, dtypes.void, tuple([make_ins(SDMAOps.COPY,
      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(copy.arg - off, ctx.max_copy_size) - 1), 0,
-     *data64_le(src_addr + off), *data64_le(dst_addr + off)) for off in range(0, copy.arg, ctx.max_copy_size)]))
+     ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz - off, ctx.max_copy_size) - 1), 0,
+     *data64_le(src_addr + off), *data64_le(dst_addr + off)) for off in range(0, sz, ctx.max_copy_size)]))
 
 def sdma_wait(ctx, dst, val):
   op = ctx.sdma.SDMA_OP_POLL_REGMEM | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_FUNC(WAIT_REG_MEM_FUNCTION_GEQ) \
      | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_MEM_POLL(1)
-  return make_ins(SDMAOps.POLL_REGMEM, op, *data64_le(make_getaddr(dst, ctx.device)), val, 0xffffffff,
+  return make_ins(SDMAOps.POLL_REGMEM, op, *data64_le(make_getaddr(dst, ctx.devs)), val, 0xffffffff,
     ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_INTERVAL(0x04) | ctx.sdma.SDMA_PKT_POLL_REGMEM_DW5_RETRY_COUNT(0xfff))
 
 def sdma_store(ctx, dst, val):
   op = ctx.sdma.SDMA_OP_FENCE | (ctx.sdma.SDMA_PKT_FENCE_HEADER_MTYPE(3) if ctx.target[0] != 9 else 0)
   return UOp(Ops.LINEAR, dtypes.void, (
-    make_ins(SDMAOps.FENCE, op, *data64_le(make_getaddr(dst, ctx.device)), val), make_ins(SDMAOps.TRAP, ctx.sdma.SDMA_OP_TRAP, 0)))
+    make_ins(SDMAOps.FENCE, op, *data64_le(make_getaddr(dst, ctx.devs)), val), make_ins(SDMAOps.TRAP, ctx.sdma.SDMA_OP_TRAP, 0)))
 
 def sdma_timestamp(ctx, dst):
   op = ctx.sdma.SDMA_OP_TIMESTAMP | ctx.sdma.SDMA_PKT_TIMESTAMP_GET_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_TIMESTAMP_GET_GLOBAL)
-  return make_ins(SDMAOps.TIMESTAMP, op, *data64_le(make_getaddr(dst, ctx.device)))
+  return make_ins(SDMAOps.TIMESTAMP, op, *data64_le(make_getaddr(dst, ctx.devs)))
 
 pm_sdma_opsel = PatternMatcher([
+  (UPat(Ops.CALL, src=(UPat(Ops.COPY),), name="call", allow_any_len=True), sdma_copy),
+
   (UPat(Ops.BARRIER), lambda: UOp(Ops.NOOP, dtypes.void, ())),
   (UPat(Ops.WAIT, src=(UPat(name="dst"), UPat(name="val"))), sdma_wait),
-  (UPat(Ops.COPY, src=(UPat(name="dst"), UPat(name="src")), name="copy"), sdma_copy),
   (UPat(Ops.CUSTOM_FUNCTION, arg="timestamp", src=(UPat(name="dst"),)), sdma_timestamp),
   (UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM), name="dst"), UPat(name="val"))), sdma_store),
 ])
 
 def sdma_submit(cmdbuf, devs):
   # the cmdbuf to submit + the patch writes that fill it
-  size_dw, zero = cmdbuf.src[0].arg // dtypes.uint32.itemsize, UOp.const(dtypes.int, 0)
+  size_dw, zero = cmdbuf.nbytes() // dtypes.uint32.itemsize, UOp.const(dtypes.int, 0)
 
   # 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 = (UOp.new_buffer(devs, b.size, b.dtype).rtag(("COPY:0", name))
+  ring, wptr, doorbell, put_ptr = (make_placeholder(devs, b.size, b.dtype, ("COPY:0", name), unique=False)
     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
@@ -221,22 +226,33 @@ def sdma_submit(cmdbuf, devs):
 
   # zero the wrapped tail, then copy the cmdbuf into the ring
   zi = UOp.range(zero_amt_dw, 0, dtype=dtypes.int, src=(cmdbuf,))
-  zero_tail = ring.index(tail_off_dw + zi, dtype=ring.dtype.ptr()).store(UOp.const(dtypes.uint32, 0)).end(zi)
+  zero_tail = ring.index(tail_off_dw + zi, ptr=True).store(UOp.const(dtypes.uint32, 0)).end(zi)
   i = UOp.range(UOp.const(dtypes.int, size_dw), 0, dtype=dtypes.int, src=(cmdbuf,))
-  copy_to_ring = ring.index(start_dw + i, dtype=ring.dtype.ptr()).store(
-    cmdbuf.index(i*4, dtype=cmdbuf.dtype.ptr()).cast(dtypes.uint32.ptr()).load()).end(i)
+  copy_to_ring = ring.index(start_dw + i, ptr=True).store(
+    cmdbuf.index(i*4, ptr=True).cast(dtypes.uint32.ptr()).load()).end(i)
 
   # advance the put/write pointers past the zeroed tail and the cmdbuf
   next_put_b = put_b + ((zero_amt_dw + size_dw) * 4).cast(put_b.dtype)
-  bump_put_ptr = put_ptr.index(zero, dtype=put_ptr.dtype.ptr()).store(next_put_b)
-  bump_wptr = wptr.index(zero, dtype=wptr.dtype.ptr()).store(next_put_b)
+  bump_put_ptr = put_ptr.index(zero, ptr=True).store(next_put_b)
+  bump_wptr = wptr.index(zero, ptr=True).store(next_put_b)
 
   # ring the doorbell once the writes have landed
   flush = UOp.barrier(zero_tail, copy_to_ring, bump_put_ptr, bump_wptr)
-  return doorbell.after(flush).index(zero, dtype=doorbell.dtype.ptr()).store(next_put_b)
+  return doorbell.after(flush).index(zero, ptr=True).store(next_put_b)
 
 pm_sdma_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"),
-  lambda lin: sdma_submit(make_cmdbuf(lin, to_tuple(lin.arg[0]), "copy"), to_tuple(lin.arg[0])))])
+  lambda lin: sdma_submit(make_cmdbuf(lin, to_tuple(lin.arg[0])), to_tuple(lin.arg[0])))])
+
+@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  # 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)
+  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"))
 
 @dataclass(frozen=True)
 class AMDProgramData:
@@ -245,10 +261,9 @@ class AMDProgramData:
   enable_dispatch_ptr:int; enable_private_segment_sgpr:int
 
 _amd_program_cache:dict[tuple[bytes,str], tuple[AMDProgramData,bytes]] = {}
-
 def amd_build_program(prg:UOp) -> UOp:
-  dev = Device[prg.src[1].arg] # TODO: rm this
-  if (cached:=_amd_program_cache.get(key:=(lib:=prg.src[4].arg, dev.device))) is None:
+  dev = Device[to_tuple(prg.device)[0]] # TODO: rm this
+  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:
@@ -258,22 +273,17 @@ def amd_build_program(prg:UOp) -> UOp:
     if (lds:=((desc.group_segment_fixed_size+511)//512)&0x1FF) > (dev.iface.props['lds_size_in_kb']*1024)//512:
       raise RuntimeError("Too many resources requested: group_segment_size")
     edp = desc.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_DISPATCH_PTR
-    cached = _amd_program_cache[key] = (AMDProgramData(
-      entry_point_offset=rodata + desc.kernel_code_entry_byte_offset,
+
+    data = AMDProgramData(entry_point_offset=rodata + desc.kernel_code_entry_byte_offset,
       rsrc1=desc.compute_pgm_rsrc1 | ((1<<20) if dev.target[0]==11 else 0),  # priv=1 on gfx11 for cwsr
       rsrc2=desc.compute_pgm_rsrc2 | (lds<<15), rsrc3=desc.compute_pgm_rsrc3,
-      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), bytes(image))
+      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)
+    buf = make_placeholder(prg.device, len(image), dtypes.uint8, "program")
+    cached = _amd_program_cache[key] = prg.replace(src=(buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=bytes(image)))),), arg=(data, prg.arg))
   return cached
 
-pm_prep_program = PatternMatcher([
-  (UPat(Ops.PROGRAM, src=(UPat(), UPat(Ops.DEVICE, arg="AMD"), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"), amd_build_program),
-])
-
 class AMDAllocator(HCQAllocator['AMDDevice']):
   def __init__(self, dev:AMDDevice):
     super().__init__(dev, supports_copy_from_disk=dev.has_sdma_queue, supports_transfer=dev.has_sdma_queue and not dev.is_usb())
@@ -516,17 +526,15 @@ class PCIIface(PCIIfaceBase):
 
 def _mock(iface, name=None): return type(name or f"MOCK{iface.__name__}", (iface,), {})
 
-def encode_queue(q:UOp) -> UOp|None:
-  if not (isinstance(q.arg, tuple) and len(q.arg) == 2 and isinstance(q.arg[1], str) and q.arg[1].startswith(("COMPUTE", "COPY"))): return None
-  devs = to_tuple(q.arg[0])
-  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=Device[devs[0]], name=f"{q.arg[1]} opsel"))
-
-pm_lower = PatternMatcher([
-  (UPat(Ops.CUSTOM_FUNCTION, arg="submit", src=(UPat(Ops.LINEAR, name="q"),)), encode_queue),
-])
-
 class AMDDevice(HCQ2Compiled):
+  pm_lower = PatternMatcher([
+    # prep program
+    (UPat(Ops.PROGRAM, src=(UPat(), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"), amd_build_program),
+
+    # encoding of cmdbuf
+    (UPat(Ops.CUSTOM_FUNCTION, arg="submit_cmdbuf", src=(UPat(Ops.LINEAR, name="q"),)), encode_queue),
+  ])
+
   timestamp_divider = 100.0  # AMD GPU clock: ticks/us
 
   ifaces = [KFDIface, PCIIface]
@@ -573,7 +581,7 @@ class AMDDevice(HCQ2Compiled):
 
     # Scratch setup
     self.max_private_segment_size = 0
-    self.pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, tag="scratch", name="b"), lambda ctx, b: ctx.scratch_buffer(b.arg))]) + self.pm_bufferize
+    self.pm_bufferize = PatternMatcher([(UPat(Ops.PARAM, tag="scratch", name="b"), lambda ctx, b: ctx.scratch_buffer(b.max_numel()))]) + self.pm_bufferize
 
     self.pmc_enabled:bool = PROFILE > 0 and PMC > 0
     if self.pmc_enabled:
@@ -621,10 +629,10 @@ class AMDDevice(HCQ2Compiled):
 
     qname = f"{'COPY' if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA else 'COMPUTE'}:{idx}"
     self.pm_bufferize = PatternMatcher([
-      (UPat(Ops.BUFFER, tag={(qname, name)}), lambda ctx, b=getattr(queue, name): b) for name in ["ring", "write_ptr", "doorbell", "put_value"]
+      (UPat(Ops.PARAM, tag={(qname, name)}), lambda ctx, b=getattr(queue, name): b) for name in ["ring", "write_ptr", "doorbell", "put_value"]
     ] + [
-      (UPat(Ops.BUFFER, tag={(qname, "timeline_signal")}), lambda ctx, q=qname: ctx.timeline_signal(q)),
-      (UPat(Ops.BUFFER, tag={(qname, "timeline_value")}), lambda ctx, q=qname: ctx.timeline_value(q)),
+      (UPat(Ops.PARAM, tag={(qname, "timeline_signal")}), lambda ctx, q=qname: ctx.timeline_signal(q)),
+      (UPat(Ops.PARAM, tag={(qname, "timeline_value")}), lambda ctx, q=qname: ctx.timeline_value(q)),
     ]) + self.pm_bufferize
 
     return queue

extra/llama_kernels/cast_amax/__init__.py

@@ -21,7 +21,7 @@ def _custom_fused_bwd_w13(grad_xw13_fp8:UOp, grad_amax_buf:UOp,
                   xw13.base, grad_x2.base, amax_state.base, grad_amax_state.base, threads, workgroups,
                   arg=KernelInfo(f"fused_silu_mul_bwd_w13_{n_elems}", estimates=Estimates(ops=10*n_elems, mem=mem)))
   src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_bwd_w13.cpp", n_elems, hidden)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
 
 @functools.cache
@@ -34,7 +34,7 @@ def _custom_fused_cast_amax_w13(fp8_out:UOp, amax_buf:UOp, xw13:UOp, amax_state:
   sink = UOp.sink(fp8_out.base, amax_buf.base, xw13.base, amax_state.base, threads, workgroups,
                   arg=KernelInfo(f"fused_silu_mul_cast_amax_w13_{n_elems}", estimates=Estimates(ops=5*n_elems, mem=mem)))
   src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_fwd_w13.cpp", n_elems, hidden)
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
 
 def _fused_quantize_bwd_w13(gradient:UOp, kernel:UOp):

extra/llama_kernels/fp8_transpose/__init__.py

@@ -1,41 +0,0 @@
-from __future__ import annotations
-import functools, pathlib
-from tinygrad import Tensor, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.renderer import Estimates
-from extra.llama_kernels import THREADS_PER_WG, alloc_like, dname_of, compile_hip
-
-TILE = 64
-
-@functools.cache
-def _custom_fp8_transpose(out:UOp, inp:UOp, dname:str) -> UOp:
-  M, N = inp.shape
-  num_wg = (M // TILE) * (N // TILE)
-  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(num_wg, "gidx0")
-  mem = M * N * 2  # one byte read + one byte write per element
-  sink = UOp.sink(out.base, inp.base, threads, workgroups,
-                  arg=KernelInfo(f"fp8_transpose_{M}_{N}",
-                                 estimates=Estimates(ops=M*N, mem=mem)))
-  src = (pathlib.Path(__file__).parent/"fp8_transpose.cpp").read_text()
-  defines = [f"-DM_DIM={M}", f"-DN_DIM={N}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
-                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
-
-def fast_fp8_transpose(t:Tensor) -> Tensor:
-  assert t.ndim == 2, f"fast_fp8_transpose needs 2D input, got shape {t.shape}"
-  assert t.dtype in dtypes.fp8s, f"fast_fp8_transpose needs fp8 dtype, got {t.dtype}"
-  M, N = t.shape
-  assert M % TILE == 0 and N % TILE == 0, f"M={M}, N={N} must be multiples of {TILE}"
-
-  device = t.device
-  axis = t.uop.axis if isinstance(device, tuple) else None
-  out_axis = None
-  if axis == 0: out_axis = 1
-  elif axis == 1: out_axis = 0
-  elif axis is not None:
-    raise ValueError(f"fast_fp8_transpose: unsupported axis {axis}")
-
-  out = alloc_like((N, M), t.dtype, device, out_axis)
-  fxn = functools.partial(_custom_fp8_transpose, dname=dname_of(device))
-  out, _ = Tensor.custom_kernel(out, t, fxn=fxn)
-  return out

extra/llama_kernels/fp8_transpose/fp8_transpose.cpp

@@ -1,74 +0,0 @@
-#include <hip/hip_runtime.h>
-
-// LDS-staged 64x64 fp8 transpose.
-//   in : (M_DIM, N_DIM) fp8 contiguous
-//   out: (N_DIM, M_DIM) fp8 contiguous, out[c][r] = in[r][c]
-//
-// One WG processes one 64x64 output tile.  Each thread reads one uint4 (16 fp8) coalesced
-// from input rows, stages into LDS, then writes one uint4 coalesced to the output (whose
-// 16 fp8 come from 16 different input rows via in-LDS gather).
-//
-// LDS layout: lds[64][LDS_STRIDE] with LDS_STRIDE=65 (1 byte pad) to mitigate bank conflicts
-// during the column-direction read of the write phase.
-
-#ifndef M_DIM
-#define M_DIM 16384
-#endif
-#ifndef N_DIM
-#define N_DIM 28672
-#endif
-#ifndef THREADS_PER_WG
-#define THREADS_PER_WG 256
-#endif
-
-constexpr int TILE = 64;
-constexpr int VEC = 16;                    // fp8 per uint4 (128-bit) load/store
-constexpr int LDS_PAD = 1;
-constexpr int LDS_STRIDE = TILE + LDS_PAD; // 65 fp8 per row
-
-static_assert(THREADS_PER_WG * VEC == TILE * TILE, "256 threads * 16 fp8 = 64*64");
-static_assert(M_DIM % TILE == 0, "M_DIM must be a multiple of 64");
-static_assert(N_DIM % TILE == 0, "N_DIM must be a multiple of 64");
-
-constexpr int N_TILES_N = N_DIM / TILE;
-
-struct alignas(16) fp8x16 { uint8_t v[16]; };
-
-extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
-fp8_transpose(uint8_t* __restrict__ out,         // (N_DIM, M_DIM)
-              const uint8_t* __restrict__ in)    // (M_DIM, N_DIM)
-{
-  __shared__ uint8_t lds[TILE * LDS_STRIDE];
-
-  const int tid     = threadIdx.x;
-  const int wg_id   = blockIdx.x;
-  const int tile_r  = wg_id / N_TILES_N;       // tile index along M dim of input
-  const int tile_c  = wg_id % N_TILES_N;       // tile index along N dim of input
-
-  const int a = tid / (TILE / VEC);            // 0..63 (row within tile during read; col within tile during write)
-  const int b = tid % (TILE / VEC);            // 0..3
-  const int b16 = b * VEC;                     // 0,16,32,48
-
-  // ---- Read phase: input rows -> LDS rows
-  {
-    const long long src = (long long)(tile_r * TILE + a) * (long long)N_DIM
-                        + (long long)(tile_c * TILE + b16);
-    fp8x16 v = *reinterpret_cast<const fp8x16*>(&in[src]);
-    *reinterpret_cast<fp8x16*>(&lds[a * LDS_STRIDE + b16]) = v;
-  }
-  __syncthreads();
-
-  // ---- Write phase: LDS columns (gathered) -> output rows
-  // out[(tile_c*TILE + a)][(tile_r*TILE + b16 + i)] = in[(tile_r*TILE + b16 + i)][(tile_c*TILE + a)]
-  //                                                  = lds[b16 + i][a]
-  {
-    fp8x16 v;
-    #pragma unroll
-    for (int i = 0; i < VEC; ++i) {
-      v.v[i] = lds[(b16 + i) * LDS_STRIDE + a];
-    }
-    const long long dst = (long long)(tile_c * TILE + a) * (long long)M_DIM
-                        + (long long)(tile_r * TILE + b16);
-    *reinterpret_cast<fp8x16*>(&out[dst]) = v;
-  }
-}

extra/llama_kernels/fused_rmsnorm_mul_quantize_fp8/__init__.py

@@ -22,7 +22,7 @@ def _custom_fwd(fp8_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax_buf:UOp,
   defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
              f"-DEPS_LITERAL={eps_val}f"]
   src = _src()
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
 
 @functools.cache
@@ -39,7 +39,7 @@ def _custom_fwd_add(fp8_out:UOp, h_out:UOp, x_normed_out:UOp, rrms_out:UOp, amax
   defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}",
              f"-DEPS_LITERAL={eps_val}f", f"-DHAS_RESIDUAL=1"]
   src = _src()
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
 
 @functools.cache
@@ -55,7 +55,7 @@ def _custom_bwd(grad_x:UOp, grad_weight_partial:UOp,
                                  estimates=Estimates(ops=8*n_elems, mem=mem)))
   defines = [f"-DN_ELEMS={n_elems}", f"-DHIDDEN={HIDDEN}", f"-DNUM_WG={NUM_WG}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
   src = _src_bwd()
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)),
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
                                UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
 
 def _bwd_common(fp8_grad_u, h_grad_u, x_u, x_normed_u, rrms_u, weight_u, amax_state_u, kernel:UOp):

extra/llama_kernels/fused_silu_mul_quantize_mxfp8/__init__.py

@@ -0,0 +1,104 @@
+import functools
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from extra.llama_kernels import FP8_MAX, THREADS_PER_WG, alloc_like
+
+BLK = 32
+PACK = 4
+LOG2E = 1.4426950408889634
+
+@functools.cache
+def _custom_silu_mul_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x_w1:UOp, x_w3:UOp) -> UOp:
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  n_elems = rows * K
+  n_super = n_elems // (BLK * PACK)
+  sk4 = scale_K // PACK
+  assert n_super % THREADS_PER_WG == 0, f"{n_super=} must divide over {THREADS_PER_WG=}"
+  nwg = n_super // THREADS_PER_WG
+
+  x_w1, x_w3 = x_w1.reshape(n_elems), x_w3.reshape(n_elems)
+  fp8_out = fp8_out.reshape(n_elems)
+  e8_out = e8_out.reshape(rows * scale_K)
+  si_out = si_out.reshape(sk4 * rows)
+
+  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  sb = UOp.range(PACK, 2, AxisType.UNROLL)
+  lane = UOp.range(BLK, 3, AxisType.UNROLL)
+
+  super_idx = wg * THREADS_PER_WG + tid
+  idx = super_idx * (BLK * PACK) + sb * BLK + lane
+
+  w1 = x_w1[idx].cast(dtypes.float)
+  w3 = x_w3[idx].cast(dtypes.float)
+  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
+  act = w1 * sig * w3
+  abs_a = (act < 0.0).where(-act, act)
+  blk_max = abs_a.reduce(lane, arg=Ops.MAX)
+  e8f = (blk_max.maximum(1e-38).log2().floor() + 127.0).maximum(0.0).minimum(254.0)
+  qscale = (127.0 - e8f).exp2()
+  scaled = (act * qscale).maximum(-FP8_MAX).minimum(FP8_MAX)
+  e8u8 = e8f.cast(dtypes.uint8)
+
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  e8_store = e8_out.after(fp8_store)[super_idx * PACK + sb].store(e8u8)
+  packed = (e8u8.cast(dtypes.uint32) << (sb.cast(dtypes.uint32) * 8)).reduce(sb, arg=Ops.ADD)
+  row, col4 = super_idx // sk4, super_idx % sk4
+  si_store = si_out.after(e8_store.end(sb))[col4 * rows + row].store(packed)
+  return si_store.end(tid, wg).sink(arg=KernelInfo(f"silu_mul_quantize_mxfp8_{n_elems}", opts_to_apply=()))
+
+@functools.cache
+def _custom_silu_mul_bwd_mxfp8(gx1_out:UOp, gx3_out:UOp, x_w1:UOp, x_w3:UOp, grad_aq:UOp, e8:UOp) -> UOp:
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  n_elems = rows * K
+  VEC = 8
+  assert n_elems % (THREADS_PER_WG * VEC) == 0, f"{n_elems=} must divide {THREADS_PER_WG*VEC=}"
+  nwg = n_elems // (THREADS_PER_WG * VEC)
+  x_w1, x_w3, grad_aq = x_w1.reshape(n_elems), x_w3.reshape(n_elems), grad_aq.reshape(n_elems)
+  gx1_out, gx3_out, e8 = gx1_out.reshape(n_elems), gx3_out.reshape(n_elems), e8.reshape(rows * scale_K)
+
+  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  lane = UOp.range(VEC, 2, AxisType.UNROLL)
+  idx = (wg * THREADS_PER_WG + tid) * VEC + lane
+
+  e8v = e8[idx // BLK].cast(dtypes.float)
+  qscale = (127.0 - e8v).exp2()
+  ga = grad_aq[idx].cast(dtypes.float) * qscale
+  w1 = x_w1[idx].cast(dtypes.float)
+  w3 = x_w3[idx].cast(dtypes.float)
+  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
+  s = w1 * sig
+  sprime = sig * (1.0 + w1 * (1.0 - sig))
+  gx1 = gx1_out[idx].store((ga * sprime * w3).cast(gx1_out.dtype.base))
+  gx3 = gx3_out.after(gx1)[idx].store((ga * s).cast(gx3_out.dtype.base))
+  return gx3.end(lane, tid, wg).sink(arg=KernelInfo(f"silu_mul_bwd_mxfp8_{n_elems}", opts_to_apply=()))
+
+def _silu_mul_quantize_mxfp8_bwd(gradient:UOp, kernel:UOp):
+  _, e8_out, _, x_w1, x_w3 = kernel.src[1:]
+  device = x_w1.device
+  rows, K = x_w1.shape
+  axis = x_w1.axis if isinstance(device, tuple) else None
+  gx1 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
+  gx3 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
+  gx1, gx3, *_ = Tensor.custom_kernel(gx1, gx3, Tensor(x_w1, device=device), Tensor(x_w3, device=device),
+                                      Tensor(gradient, device=device).cast(dtypes.bfloat16), Tensor(e8_out.after(kernel), device=device),
+                                      fxn=_custom_silu_mul_bwd_mxfp8)
+  return (None, None, None, gx1.uop, gx3.uop)
+
+def fused_silu_mul_quantize_mxfp8(x_w1:Tensor, x_w3:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  assert x_w1.shape == x_w3.shape, f"{x_w1.shape} != {x_w3.shape}"
+  assert x_w1.dtype == dtypes.bfloat16 and x_w3.dtype == dtypes.bfloat16
+  assert x_w1.ndim == 2, f"expected 2d, got {x_w1.shape}"
+  from extra.gemm.cdna_asm_gemm import FP8_DTYPE
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  axis = x_w1.uop.axis if isinstance(x_w1.device, tuple) else None
+  fp8_out = alloc_like((rows, K), FP8_DTYPE, x_w1.device, axis)
+  e8_out = alloc_like((rows, scale_K), dtypes.uint8, x_w1.device, axis)
+  si_out = alloc_like((scale_K // PACK, rows), dtypes.uint32, x_w1.device, None if axis is None else (1 if axis == 0 else 0))
+  fp8_out, e8_out, si_out, *_ = Tensor.custom_kernel(fp8_out, e8_out, si_out, x_w1, x_w3,
+                                                     fxn=_custom_silu_mul_quantize_mxfp8, grad_fxn=_silu_mul_quantize_mxfp8_bwd)
+  return fp8_out, e8_out, si_out

extra/llama_kernels/quantize_fp8_delayed/__init__.py

@@ -42,8 +42,8 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_st
   step = THREADS_PER_WG // 2
   while step:
     active = tid < step
-    other = lds[tid + step].load(UOp.const(dtypes.float, 0.0), active)
-    lds = lds.after(lds[tid].store(lds[tid].maximum(other), gate=active).barrier())
+    other = lds[(tid + step).valid(active)].load()
+    lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)).barrier())
     step //= 2
 
   amax_store = amax_partial[tid.eq(0).where(wg, UOp.invalid())].store(lds[0])

extra/llama_kernels/quantize_mxfp8_fused/__init__.py

@@ -0,0 +1,71 @@
+import functools
+from tinygrad import Tensor, dtypes
+from tinygrad.helpers import prod
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from extra.llama_kernels import FP8_MAX, THREADS_PER_WG, alloc_like
+
+BLK = 32
+PACK = 4
+
+@functools.cache
+def _custom_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x:UOp) -> UOp:
+  rows, K = x.shape
+  scale_K = K // BLK
+  n_elems = rows * K
+  n_super = n_elems // (BLK * PACK)
+  sk4 = scale_K // PACK
+  assert n_super % THREADS_PER_WG == 0, f"{n_super=} must divide over {THREADS_PER_WG=}"
+  nwg = n_super // THREADS_PER_WG
+
+  x = x.reshape(n_elems)
+  fp8_out = fp8_out.reshape(n_elems)
+  e8_out = e8_out.reshape(rows * scale_K)
+  si_out = si_out.reshape(sk4 * rows)
+
+  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  sb = UOp.range(PACK, 2, AxisType.UNROLL)
+  lane = UOp.range(BLK, 3, AxisType.UNROLL)
+
+  super_idx = wg * THREADS_PER_WG + tid
+  idx = super_idx * (BLK * PACK) + sb * BLK + lane
+
+  x_f = x[idx].cast(dtypes.float)
+  abs_x = (x_f < 0.0).where(-x_f, x_f)
+  blk_max = abs_x.reduce(lane, arg=Ops.MAX)
+  e8f = (blk_max.maximum(1e-38).log2().floor() + 127.0).maximum(0.0).minimum(254.0)
+  qscale = (127.0 - e8f).exp2()
+  scaled = (x_f * qscale).maximum(-FP8_MAX).minimum(FP8_MAX)
+  e8u8 = e8f.cast(dtypes.uint8)
+
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  e8_store = e8_out.after(fp8_store)[super_idx * PACK + sb].store(e8u8)
+
+  # pack the 4 e8 of this super-block into one uint32 (little-endian: byte sb), write transposed (sk4, row)
+  packed = (e8u8.cast(dtypes.uint32) << (sb.cast(dtypes.uint32) * 8)).reduce(sb, arg=Ops.ADD)
+  row, col4 = super_idx // sk4, super_idx % sk4
+  si_store = si_out.after(e8_store.end(sb))[col4 * rows + row].store(packed)
+  return si_store.end(tid, wg).sink(arg=KernelInfo(f"quantize_mxfp8_{n_elems}", opts_to_apply=()))
+
+def _quantize_mxfp8_fused_bwd(gradient:UOp, kernel:UOp):
+  _, e8_out, _, x = kernel.src[1:]
+  device = x.device
+  rows, K = x.shape
+  scale_K = K // BLK
+  e8 = Tensor(e8_out, device=device).reshape(rows, scale_K)
+  qscale = (127.0 - e8.cast(dtypes.float32)).exp2().reshape(rows, scale_K, 1).expand(rows, scale_K, BLK).reshape(rows, K)
+  grad_x = (Tensor(gradient, device=device).float() * qscale).cast(dtypes.bfloat16)
+  return (None, None, None, grad_x.uop)
+
+def quantize_mxfp8_fused(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  assert x.dtype == dtypes.bfloat16, f"expected bf16, got {x.dtype}"
+  assert x.ndim == 2, f"expected 2d (rows, K), got {x.shape}"
+  from extra.gemm.cdna_asm_gemm import FP8_DTYPE
+  rows, K = x.shape
+  scale_K = K // BLK
+  axis = x.uop.axis if isinstance(x.device, tuple) else None
+  fp8_out = alloc_like((rows, K), FP8_DTYPE, x.device, axis)
+  e8_out = alloc_like((rows, scale_K), dtypes.uint8, x.device, axis)
+  si_out = alloc_like((scale_K // PACK, rows), dtypes.uint32, x.device, None if axis is None else (1 if axis == 0 else 0))
+  fp8_out, e8_out, si_out, *_ = Tensor.custom_kernel(fp8_out, e8_out, si_out, x, fxn=_custom_quantize_mxfp8, grad_fxn=_quantize_mxfp8_fused_bwd)
+  return fp8_out, e8_out, si_out

extra/llama_kernels/transpose_quantize_mxfp8/__init__.py

@@ -0,0 +1,37 @@
+from __future__ import annotations
+import functools, pathlib
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+from tinygrad.renderer import Estimates
+from extra.llama_kernels import THREADS_PER_WG, alloc_like, dname_of, compile_hip
+
+TILE_N = THREADS_PER_WG   # 256
+BLK = 32
+
+@functools.cache
+def _custom_transpose_quantize_mxfp8(q:UOp, e8:UOp, g:UOp, dname:str) -> UOp:
+  M, N = g.shape
+  num_wg = (M // BLK) * (N // TILE_N)
+  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(num_wg, "gidx0")
+  mem = M * N * 2 + M * N + (M // BLK) * N   # read bf16, write fp8 + e8
+  sink = UOp.sink(q.base, e8.base, g.base, threads, workgroups,
+                  arg=KernelInfo(f"transpose_quantize_mxfp8_{M}_{N}", estimates=Estimates(ops=M*N, mem=mem)))
+  src = (pathlib.Path(__file__).parent/"transpose_quantize_mxfp8.cpp").read_text()
+  defines = [f"-DM_DIM={M}", f"-DN_DIM={N}", f"-DTHREADS_PER_WG={THREADS_PER_WG}"]
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)),
+                               UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=compile_hip(src, defines))))
+
+def transpose_quantize_mxfp8(g:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  # fused g.T quantize: returns (q, e8, si) == quantize_mxfp8(g.T) — q (N,M) fp8, e8 (N, M/32), si packed (M/128, N)
+  assert g.ndim == 2 and g.dtype == dtypes.bfloat16, f"{g.shape} {g.dtype}"
+  from extra.gemm.cdna_asm_gemm import FP8_DTYPE, mx_pack
+  M, N = g.shape
+  assert M % BLK == 0 and N % TILE_N == 0, f"M={M} must%{BLK}, N={N} must%{TILE_N}"
+  device = g.device
+  axis = g.uop.axis if isinstance(device, tuple) else None
+  out_axis = None if axis is None else (1 if axis == 0 else 0)
+  q = alloc_like((N, M), FP8_DTYPE, device, out_axis)
+  e8 = alloc_like((N, M // BLK), dtypes.uint8, device, out_axis)
+  fxn = functools.partial(_custom_transpose_quantize_mxfp8, dname=dname_of(device))
+  q, e8, *_ = Tensor.custom_kernel(q, e8, g, fxn=fxn)
+  return q, e8, mx_pack(e8)

extra/llama_kernels/transpose_quantize_mxfp8/transpose_quantize_mxfp8.cpp

@@ -0,0 +1,62 @@
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp8.h>
+#include <hip/hip_bf16.h>
+
+#ifndef M_DIM
+#define M_DIM 8192
+#endif
+#ifndef N_DIM
+#define N_DIM 14336
+#endif
+#ifndef THREADS_PER_WG
+#define THREADS_PER_WG 256
+#endif
+
+constexpr int BLK = 32;
+constexpr int TILE_M = BLK;                 // one mxfp8 block along M per tile
+constexpr int TILE_N = THREADS_PER_WG;      // 256, one output column per thread
+constexpr int LDS_STRIDE = TILE_N + 1;      // +1 pad: stride 257 ≡ 1 (mod 32) -> conflict-free column reads
+constexpr int N_TILES_N = N_DIM / TILE_N;
+constexpr float FP8_MAX = 448.0f;
+
+static_assert(M_DIM % TILE_M == 0, "M_DIM must be a multiple of 32");
+static_assert(N_DIM % TILE_N == 0, "N_DIM must be a multiple of 256");
+
+extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
+transpose_quantize_mxfp8(__hip_fp8_storage_t* __restrict__ q,        // (N_DIM, M_DIM)
+                         uint8_t* __restrict__ e8_out,               // (N_DIM, M_DIM/32)
+                         const __hip_bfloat16* __restrict__ g)       // (M_DIM, N_DIM)
+{
+  __shared__ __hip_bfloat16 lds[TILE_M * LDS_STRIDE];
+  const int tid    = threadIdx.x;
+  const int tile_m = blockIdx.x / N_TILES_N;     // which 32-block along M
+  const int tile_n = blockIdx.x % N_TILES_N;
+
+  #pragma unroll
+  for (int mm = 0; mm < TILE_M; mm++)
+    lds[mm * LDS_STRIDE + tid] = g[(long long)(tile_m * TILE_M + mm) * N_DIM + (tile_n * TILE_N + tid)];
+  __syncthreads();
+
+  float vals[TILE_M];
+  float amax = 0.0f;
+  #pragma unroll
+  for (int mm = 0; mm < TILE_M; mm++) {
+    float v = (float)lds[mm * LDS_STRIDE + tid];
+    vals[mm] = v;
+    amax = fmaxf(amax, fabsf(v));
+  }
+  int e8 = (int)floorf(log2f(fmaxf(amax, 1e-38f))) + 127;
+  e8 = max(0, min(254, e8));
+  float qscale = exp2f((float)(127 - e8));
+
+  const long long n = tile_n * TILE_N + tid;
+  __hip_fp8_storage_t out[TILE_M];
+  #pragma unroll
+  for (int mm = 0; mm < TILE_M; mm++)
+    out[mm] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, vals[mm] * qscale)), __HIP_SATFINITE, __HIP_E4M3);
+  // 32 contiguous fp8 along M -> two 16-byte vector stores
+  long long obase = n * M_DIM + (long long)(tile_m * TILE_M);
+  *reinterpret_cast<uint4*>(&q[obase])      = *reinterpret_cast<uint4*>(&out[0]);
+  *reinterpret_cast<uint4*>(&q[obase + 16]) = *reinterpret_cast<uint4*>(&out[16]);
+  e8_out[n * (M_DIM / BLK) + tile_m] = (uint8_t)e8;
+}

extra/mmapeak/mmapeak.py

@@ -37,7 +37,7 @@ def launchBenchmark(instruction, vgprIndices, dense=True, accum=False, **kwargs)
     gidx = UOp.special(NUM_WORKGROUPS, "gidx0")
     FLOPs = FLOPS_PER_MATMUL * NUM_WAVES * NUM_WORKGROUPS * INTERNAL_LOOP * INSTRUCTIONS_PER_LOOP
     sink = UOp.sink(A.base, threads, gidx, arg=KernelInfo(inst.op.name.lower(), estimates=Estimates(ops=FLOPs, mem=0)))
-    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=Device.DEFAULT), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+    return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
   dummy = Tensor.zeros(1).contiguous().realize()
   out = Tensor.custom_kernel(dummy, fxn=fxn)[0]
   linear = out.schedule_linear()

extra/models/convnext.py

@@ -1,6 +1,6 @@
 from tinygrad.tensor import Tensor
 from tinygrad.nn import Conv2d, LayerNorm, LayerNorm2d, Linear
-from tinygrad.helpers import fetch, get_child
+from tinygrad.helpers import fetch, get_child, Context
 
 class Block:
   def __init__(self, dim):
@@ -58,7 +58,6 @@ if __name__ == "__main__":
   from test.models.test_efficientnet import chicken_img, preprocess, _LABELS
   img = Tensor(preprocess(chicken_img))
 
-  Tensor.training = False
-
-  out = model(img).numpy()
+  with Context(TRAINING=0):
+    out = model(img).numpy()
   print(_LABELS[out.argmax()])

extra/models/mask_rcnn.py

@@ -5,7 +5,7 @@ import numpy as np
 from pathlib import Path
 from tinygrad import nn, Tensor, dtypes
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.helpers import get_child, fetch
+from tinygrad.helpers import get_child, fetch, TRAINING
 from tinygrad.nn.state import torch_load
 from examples.mlperf.helpers import BoxCoder
 from extra.models.resnet import ResNet
@@ -1069,7 +1069,7 @@ class RoIBoxHead:
   def __call__(self, features, proposals, targets=None):
     x = self.feature_extractor(features, proposals)
     class_logits, box_regression = self.predictor(x)
-    if not Tensor.training:
+    if not TRAINING:
       result = self.post_processor((class_logits, box_regression), proposals)
       return x, result, {}
 
@@ -1111,7 +1111,7 @@ class Mask:
     x = self.feature_extractor(features, proposals)
     if x is not None:
       mask_logits = self.predictor(x)
-      if not Tensor.training:
+      if not TRAINING:
         result = self.post_processor(mask_logits, proposals)
         return x, result, {}
     return x, [], {}

extra/models/retinanet.py

@@ -1,6 +1,6 @@
 import math
 from tinygrad import Tensor, dtypes
-from tinygrad.helpers import flatten, get_child
+from tinygrad.helpers import flatten, get_child, TRAINING
 from examples.mlperf.helpers import generate_anchors, BoxCoder
 from examples.mlperf.losses import sigmoid_focal_loss, l1_loss
 from extra.models.resnet import ResNet
@@ -141,7 +141,7 @@ class ClassificationHead:
     out = [self.cls_logits(feat.sequential(self.conv)).permute(0, 2, 3, 1).reshape(feat.shape[0], -1, self.num_classes) for feat in x]
     out = out[0].cat(*out[1:], dim=1)
 
-    if Tensor.training:
+    if TRAINING:
       assert labels is not None and matches is not None, "labels and matches should be passed in when training"
       return self._compute_loss(out.cast(dtypes.float32), labels, matches)
 
@@ -167,7 +167,7 @@ class RegressionHead:
     out = [self.bbox_reg(feat.sequential(self.conv)).permute(0, 2, 3, 1).reshape(feat.shape[0], -1, 4) for feat in x]
     out = out[0].cat(*out[1:], dim=1)
 
-    if Tensor.training:
+    if TRAINING:
       assert bboxes is not None and matches is not None and anchors is not None, "bboxes, matches, and anchors should be passed in when training"
       return self._compute_loss(out, bboxes, matches, anchors)
 
@@ -187,7 +187,7 @@ class RetinaHead:
     self.regression_head = RegressionHead(in_channels, num_anchors)
 
   def __call__(self, x:Tensor, **kwargs) -> Tensor|dict[str, Tensor]:
-    if Tensor.training:
+    if TRAINING:
       return {
         "classification_loss": self.classification_head(x, labels=kwargs["labels"], matches=kwargs["matches"]),
         "regression_loss": self.regression_head(x, bboxes=kwargs["bboxes"], matches=kwargs["matches"], anchors=kwargs["anchors"])

extra/thunder/amd/fa.py

@@ -111,7 +111,7 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:st
   lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             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_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
@@ -141,7 +141,7 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
   lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             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):
@@ -171,7 +171,7 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
   lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             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_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
@@ -201,4 +201,4 @@ def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int,
   lib = bytes(lib)
 
   return UOp(Ops.PROGRAM,
-             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
+             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

extra/thunder/amd/gemm_mxfp8.cpp

@@ -26,9 +26,7 @@ constexpr int REG_N      = BLOCK_COL / WARPS_COL / 2;
 
 using G = kittens::group<NUM_WARPS>;
 
-__global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3 *A_ptr, fp8e4m3 *B_ptr,
-    const uint32_t *__restrict__ scale_A_iter,
-    const uint32_t *__restrict__ scale_B_iter,
+__global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3 *A_ptr, fp8e4m3 *B_ptr, fp8e8m0 *scale_A_ptr, fp8e8m0 *scale_B_ptr,
     const uint8_t *__restrict__ a_e8_unused,
     const uint8_t *__restrict__ b_e8_unused) {
     constexpr int M = GEMM_M, N = GEMM_N, K = GEMM_K;
@@ -40,15 +38,19 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
     constexpr int k_iters      = K / BLOCK_K;
     constexpr int NUM_THREADS  = NUM_WARPS * WARP_THREADS;
 
-    using ST_A = st_fp8e4m3<HALF_ROW, BLOCK_K, st_16x128_s>;
-    using ST_B = st_fp8e4m3<HALF_COL, BLOCK_K, st_16x128_s>;
-    using RT_A = rt_fp8e4m3<REG_M, BLOCK_K>;
-    using RT_B = rt_fp8e4m3<REG_N, BLOCK_K>;
-    using RT_C = rt_fl<REG_M, REG_N, col_l, rt_16x16_s>;
+    kittens::gl<fp8e8m0, k_iters * (M / BLOCK_ROW), 1, 16, 64> scale_A_gl{scale_A_ptr, nullptr, nullptr, nullptr, nullptr};
+    kittens::gl<fp8e8m0, k_iters * (N / BLOCK_COL), 1, 16, 64> scale_B_gl{scale_B_ptr, nullptr, nullptr, nullptr, nullptr};
+
+    using ST_A     = st_fp8e4m3<HALF_ROW, BLOCK_K, st_16x128_s>;
+    using ST_B     = st_fp8e4m3<HALF_COL, BLOCK_K, st_16x128_s>;
+    using ST_Scale = st<fp8e8m0, 16, 64, st_16x64_s>;
+    using RT_A     = rt_fp8e4m3<REG_M, BLOCK_K>;
+    using RT_B     = rt_fp8e4m3<REG_N, BLOCK_K>;
+    using RT_C     = rt_fl<REG_M, REG_N, col_l, rt_16x16_s>;
 
     __shared__ ST_A As[2][2];
     __shared__ ST_B Bs[2][2];
-    __shared__ uint8_t smem_scales[2048];
+    __shared__ ST_Scale scale_A_smem[2], scale_B_smem[2];
 
     RT_A a;
     RT_B b0, b1;
@@ -81,54 +83,87 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
     G::prefill_swizzled_offsets(As[0][0], A, sw_A);
     G::prefill_swizzled_offsets(Bs[0][0], B, sw_B);
 
+    const T *a_base = (const T *)&A[{0, 0, 0, 0}];
+    const T *b_base = (const T *)&B[{0, 0, 0, 0}];
+    const int a_row_stride = A.template stride<2>() * sizeof(T);
+    const int b_row_stride = B.template stride<2>() * sizeof(T);
+    i32x4 a_srd = make_srsrc(a_base, M * a_row_stride, a_row_stride);
+    i32x4 b_srd = make_srsrc(b_base, N * b_row_stride, b_row_stride);
+
+    const int wid = warpid() % NUM_WARPS;
+    constexpr int elem_per_warp = (16 / sizeof(T)) * kittens::WARP_THREADS;
+    uint32_t a_lds_00 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[0][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t a_lds_01 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[0][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t a_lds_10 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[1][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t a_lds_11 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&As[1][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_00 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[0][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_01 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[0][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_10 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[1][0].data[0]) + wid * elem_per_warp * sizeof(T)));
+    uint32_t b_lds_11 = __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(reinterpret_cast<uintptr_t>(&Bs[1][1].data[0]) + wid * elem_per_warp * sizeof(T)));
+
     int a_row_h0 = warp_m * REG_M;
     int a_row_h1 = HALF_ROW + warp_m * REG_M;
     int b_row_h0 = warp_n * REG_N;
     int b_row_h1 = HALF_COL + warp_n * REG_N;
 
     int tic = 0, toc = 1;
+    int tic_scales = 0, toc_scales = 1;
 
-    G::load(Bs[tic][0], B, {0, 0, block_col * 2, 0}, sw_B);
-    G::load(As[tic][0], A, {0, 0, block_row * 2, 0}, sw_A);
-    G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, 0}, sw_B);
-    G::load(As[tic][1], A, {0, 0, block_row * 2 + 1, 0}, sw_A);
+    uint32_t a_lds[2][2] = {{a_lds_00, a_lds_01}, {a_lds_10, a_lds_11}};
+    uint32_t b_lds[2][2] = {{b_lds_00, b_lds_01}, {b_lds_10, b_lds_11}};
+
+    if constexpr (k_iters >= 6 && (k_iters % 2 == 0)) {
+    G::load(Bs[tic][0], B, {0, 0, block_col * 2, 0}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][0]));
+    G::load(As[tic][0], A, {0, 0, block_row * 2, 0}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[tic][0]));
+    G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, 0}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][1]));
+    G::load(As[tic][1], A, {0, 0, block_row * 2 + 1, 0}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[tic][1]));
 
     if (warp_m == 1) __builtin_amdgcn_s_barrier();
     asm volatile("s_waitcnt vmcnt(4)");
     __builtin_amdgcn_s_barrier();
 
-    G::load(As[toc][0], A, {0, 0, block_row * 2, 1}, sw_A);
-    G::load(Bs[toc][0], B, {0, 0, block_col * 2, 1}, sw_B);
-    G::load(Bs[toc][1], B, {0, 0, block_col * 2 + 1, 1}, sw_B);
-
+    G::load(As[toc][0], A, {0, 0, block_row * 2, 1}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[toc][0]));
+    G::load(Bs[toc][0], B, {0, 0, block_col * 2, 1}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[toc][0]));
+    G::load(Bs[toc][1], B, {0, 0, block_col * 2 + 1, 1}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[toc][1]));
     asm volatile("s_waitcnt vmcnt(6)");
     __builtin_amdgcn_s_barrier();
 
-    #pragma unroll 2
-    for (int k = 0; k < k_iters - 2; k++, tic ^= 1, toc ^= 1) {
-        load_scales_to_lds(smem_scales, scale_A_iter, scale_B_iter, block_m, block_n, k, M, N);
+    G::load(scale_A_smem[0], scale_A_gl, {0 * tiles_M + block_row, 0, 0, 0});
+    G::load(scale_B_smem[0], scale_B_gl, {0 * tiles_N + block_col, 0, 0, 0});
+    asm volatile("s_waitcnt vmcnt(0)");
+    asm volatile("s_waitcnt lgkmcnt(0)");
+    __builtin_amdgcn_s_barrier();
 
+    int sa_idx = block_row, sb_idx = block_col;
+
+    #pragma unroll 2
+    for (int k = 0; k < k_iters - 2; k++, tic ^= 1, toc ^= 1, tic_scales ^= 1, toc_scales ^= 1) {
+        if (k + 1 < k_iters) {
+            sa_idx += tiles_M; sb_idx += tiles_N;
+            G::load(scale_A_smem[toc_scales], scale_A_gl, {sa_idx, 0, 0, 0});
+            G::load(scale_B_smem[toc_scales], scale_B_gl, {sb_idx, 0, 0, 0});
+        }
         auto bs0 = subtile_inplace<REG_N, BLOCK_K>(Bs[tic][0], {warp_n, 0});
         load(b0, bs0);
         auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[tic][0], {warp_m, 0});
         load(a, as0);
-        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A);
+        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[toc][1]));
         asm volatile("s_waitcnt lgkmcnt(8)");
         __builtin_amdgcn_s_barrier();
 
-        fp8e8m0_4 sa_h0 = pack_scales(smem_scales, 0, a_row_h0);
-        fp8e8m0_4 sb_h0 = pack_scales(smem_scales, 1024, b_row_h0);
-
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[tic_scales].data, a_row_h0);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[tic_scales].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[tic_scales].data, b_row_h1);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[tic_scales].data, a_row_h1);
         __builtin_amdgcn_s_setprio(2);
         mma_ABt_scaled(cA, a, b0, cA, &sa_h0, &sb_h0);
         __builtin_amdgcn_s_setprio(0);
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
 
-        fp8e8m0_4 sb_h1 = pack_scales(smem_scales, 1024, b_row_h1);
         auto bs1 = subtile_inplace<REG_N, BLOCK_K>(Bs[tic][1], {warp_n, 0});
         load(b1, bs1);
-        G::load(As[tic][0], A, {0, 0, block_row * 2, k + 2}, sw_A);
+        G::load(As[tic][0], A, {0, 0, block_row * 2, k + 2}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[tic][0]));
         asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
 
@@ -137,10 +172,9 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_s_setprio(0);
         __builtin_amdgcn_s_barrier();
 
-        fp8e8m0_4 sa_h1 = pack_scales(smem_scales, 0, a_row_h1);
         auto as1 = subtile_inplace<REG_M, BLOCK_K>(As[tic][1], {warp_m, 0});
         load(a, as1);
-        G::load(Bs[tic][0], B, {0, 0, block_col * 2, k + 2}, sw_B);
+        G::load(Bs[tic][0], B, {0, 0, block_col * 2, k + 2}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][0]));
         asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
 
@@ -150,7 +184,7 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_s_barrier();
         __builtin_amdgcn_sched_barrier(0);
 
-        G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, k + 2}, sw_B);
+        G::load(Bs[tic][1], B, {0, 0, block_col * 2 + 1, k + 2}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[tic][1]));
         asm volatile("s_waitcnt vmcnt(6)");
         __builtin_amdgcn_s_barrier();
 
@@ -162,19 +196,24 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
 
     { // Epilogue k = k_iters - 2
         int k = k_iters - 2;
-        load_scales_to_lds(smem_scales, scale_A_iter, scale_B_iter, block_m, block_n, k, M, N);
+        if (k + 1 < k_iters) {
+            sa_idx += tiles_M; sb_idx += tiles_N;
+            G::load(scale_A_smem[toc_scales], scale_A_gl, {sa_idx, 0, 0, 0});
+            G::load(scale_B_smem[toc_scales], scale_B_gl, {sb_idx, 0, 0, 0});
+        }
+        asm volatile("s_waitcnt vmcnt(0)");
+        asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
-
-        fp8e8m0_4 sa_h0 = pack_scales(smem_scales, 0, a_row_h0);
-        fp8e8m0_4 sa_h1 = pack_scales(smem_scales, 0, a_row_h1);
-        fp8e8m0_4 sb_h0 = pack_scales(smem_scales, 1024, b_row_h0);
-        fp8e8m0_4 sb_h1 = pack_scales(smem_scales, 1024, b_row_h1);
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[tic_scales].data, a_row_h0);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[tic_scales].data, a_row_h1);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[tic_scales].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[tic_scales].data, b_row_h1);
 
         auto bs0 = subtile_inplace<REG_N, BLOCK_K>(Bs[tic][0], {warp_n, 0});
         load(b0, bs0);
         auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[tic][0], {warp_m, 0});
         load(a, as0);
-        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A);
+        G::load(As[toc][1], A, {0, 0, block_row * 2 + 1, k + 1}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[toc][1]));
         __builtin_amdgcn_s_barrier();
 
         asm volatile("s_waitcnt lgkmcnt(0)");
@@ -217,17 +256,17 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_sched_barrier(0);
 
         tic ^= 1; toc ^= 1;
+        tic_scales ^= 1; toc_scales ^= 1;
     }
 
     { // Epilogue k = k_iters - 1 (final)
-        int k = k_iters - 1;
-        load_scales_to_lds(smem_scales, scale_A_iter, scale_B_iter, block_m, block_n, k, M, N);
+        asm volatile("s_waitcnt vmcnt(0)");
+        asm volatile("s_waitcnt lgkmcnt(0)");
         __builtin_amdgcn_s_barrier();
-
-        fp8e8m0_4 sa_h0 = pack_scales(smem_scales, 0, a_row_h0);
-        fp8e8m0_4 sa_h1 = pack_scales(smem_scales, 0, a_row_h1);
-        fp8e8m0_4 sb_h0 = pack_scales(smem_scales, 1024, b_row_h0);
-        fp8e8m0_4 sb_h1 = pack_scales(smem_scales, 1024, b_row_h1);
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[tic_scales].data, a_row_h0);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[tic_scales].data, a_row_h1);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[tic_scales].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[tic_scales].data, b_row_h1);
 
         auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[tic][0], {warp_m, 0});
         load(a, as0);
@@ -263,6 +302,38 @@ __global__ __launch_bounds__(512, 2) void mxfp8_gemm_kernel(bf16 *C_ptr, fp8e4m3
         __builtin_amdgcn_s_barrier();
     }
 
+    } else {
+    #pragma unroll 1
+    for (int kk = 0; kk < k_iters; kk++) {
+        G::load(As[0][0], A, {0, 0, block_row * 2,     kk}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[0][0]));
+        G::load(As[0][1], A, {0, 0, block_row * 2 + 1, kk}, sw_A, a_srd, a_base, __builtin_amdgcn_readfirstlane(a_lds[0][1]));
+        G::load(Bs[0][0], B, {0, 0, block_col * 2,     kk}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[0][0]));
+        G::load(Bs[0][1], B, {0, 0, block_col * 2 + 1, kk}, sw_B, b_srd, b_base, __builtin_amdgcn_readfirstlane(b_lds[0][1]));
+        G::load(scale_A_smem[0], scale_A_gl, {kk * tiles_M + block_row, 0, 0, 0});
+        G::load(scale_B_smem[0], scale_B_gl, {kk * tiles_N + block_col, 0, 0, 0});
+        asm volatile("s_waitcnt vmcnt(0)");
+        asm volatile("s_waitcnt lgkmcnt(0)");
+        __builtin_amdgcn_s_barrier();
+
+        fp8e8m0_4 sa_h0 = pack_scales(scale_A_smem[0].data, a_row_h0);
+        fp8e8m0_4 sa_h1 = pack_scales(scale_A_smem[0].data, a_row_h1);
+        fp8e8m0_4 sb_h0 = pack_scales(scale_B_smem[0].data, b_row_h0);
+        fp8e8m0_4 sb_h1 = pack_scales(scale_B_smem[0].data, b_row_h1);
+
+        auto bs0 = subtile_inplace<REG_N, BLOCK_K>(Bs[0][0], {warp_n, 0}); load(b0, bs0);
+        auto bs1 = subtile_inplace<REG_N, BLOCK_K>(Bs[0][1], {warp_n, 0}); load(b1, bs1);
+        auto as0 = subtile_inplace<REG_M, BLOCK_K>(As[0][0], {warp_m, 0}); load(a, as0);
+        asm volatile("s_waitcnt lgkmcnt(0)");
+        mma_ABt_scaled(cA, a, b0, cA, &sa_h0, &sb_h0);
+        mma_ABt_scaled(cB, a, b1, cB, &sa_h0, &sb_h1);
+        auto as1 = subtile_inplace<REG_M, BLOCK_K>(As[0][1], {warp_m, 0}); load(a, as1);
+        asm volatile("s_waitcnt lgkmcnt(0)");
+        mma_ABt_scaled(cC, a, b0, cC, &sa_h1, &sb_h0);
+        mma_ABt_scaled(cD, a, b1, cD, &sa_h1, &sb_h1);
+        __builtin_amdgcn_s_barrier();
+    }
+    }
+
     store(C, cA, {0, 0, block_row * WARPS_ROW * 2 + warp_m, block_col * WARPS_COL * 2 + warp_n});
     store(C, cB, {0, 0, block_row * WARPS_ROW * 2 + warp_m, block_col * WARPS_COL * 2 + WARPS_COL + warp_n});
     store(C, cC, {0, 0, block_row * WARPS_ROW * 2 + WARPS_ROW + warp_m, block_col * WARPS_COL * 2 + warp_n});

extra/thunder/amd/include/cdna4/common/macros.cuh

@@ -438,6 +438,35 @@ __device__ __forceinline__ T buffer_load_dwordx2(
   return result;
 }
 
+template<int GPR_START>
+__device__ __forceinline__ void buffer_load_dwordx3(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
+  if constexpr (GPR_START >= 256) {
+    asm volatile("buffer_load_dwordx3 a[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START - 256), "n"(GPR_START + 2 - 256), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  } else {
+    asm volatile("buffer_load_dwordx3 v[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START), "n"(GPR_START + 2), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  }
+}
+
+// Returns a u32x3 (ext_vector_type(3)). Note sizeof(u32x3) == 16 due to vector
+// alignment padding, but the compiler still allocates exactly 3 consecutive VGPRs
+// for the "=v" operand and emits buffer_load_dwordx3 v[N:N+2].
+template<typename T = u32x3>
+__device__ __forceinline__ T buffer_load_dwordx3(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  T result;
+  asm volatile("buffer_load_dwordx3 %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
 template<int GPR_START>
 __device__ __forceinline__ void buffer_load_dwordx4(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   if constexpr (GPR_START >= 256) {
@@ -465,6 +494,116 @@ __device__ __forceinline__ T buffer_load_dwordx4(
   return result;
 }
 
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_ubyte(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_ubyte %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_sbyte(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_sbyte %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_ushort(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_ushort %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ T buffer_load_sshort(
+  const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  T result;
+  asm volatile("buffer_load_sshort %0, %1, %2, %3 offen offset:%4"
+    : "=v"(result)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+  return result;
+}
+
+// d16 variants merge the loaded sub-dword into the low (or high, for _hi) 16 bits
+// of the destination register, leaving the other half untouched: read-modify-write.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_ubyte_d16(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_ubyte_d16 %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_ubyte_d16_hi(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_ubyte_d16_hi %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_sbyte_d16(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_sbyte_d16 %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_sbyte_d16_hi(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_sbyte_d16_hi %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_short_d16(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_short_d16 %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_load_short_d16_hi(
+  T& dst, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_load_short_d16_hi %0, %1, %2, %3 offen offset:%4"
+    : "+v"(dst)
+    : "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
 template<int GPR>
 __device__ __forceinline__ void buffer_store_dword(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   // AGPRS
@@ -519,6 +658,32 @@ __device__ __forceinline__ void buffer_store_dwordx2(
     : "memory");
 }
 
+template<int GPR_START>
+__device__ __forceinline__ void buffer_store_dwordx3(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
+  // AGPRS
+  if constexpr (GPR_START >= 256) {
+    asm volatile("buffer_store_dwordx3 a[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START - 256), "n"(GPR_START + 2 - 256), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  // VGPRS
+  } else {
+    asm volatile("buffer_store_dwordx3 v[%0:%1], %2, %3, %4 offen offset:%5"
+      :
+      : "n"(GPR_START), "n"(GPR_START + 2), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+      : "memory");
+  }
+}
+
+template<typename T = u32x3>
+__device__ __forceinline__ void buffer_store_dwordx3(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  asm volatile("buffer_store_dwordx3 %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
 template<int GPR_START>
 __device__ __forceinline__ void buffer_store_dwordx4(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   // AGPRS
@@ -546,6 +711,50 @@ __device__ __forceinline__ void buffer_store_dwordx4(
     : "memory");
 }
 
+// Stores the low 8 bits of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_byte(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_byte %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+// Stores the high 16 bits' low 8 bits (bits [23:16]) of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_byte_d16_hi(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_byte_d16_hi %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+// Stores the low 16 bits of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_short(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_short %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
+// Stores the high 16 bits of the value register.
+template<typename T = uint32_t>
+__device__ __forceinline__ void buffer_store_short_d16_hi(
+  const T& value, const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const uint32_t i_offset = 0) {
+  static_assert(sizeof(T) == sizeof(uint32_t));
+  asm volatile("buffer_store_short_d16_hi %0, %1, %2, %3 offen offset:%4"
+    :
+    : "v"(value), "v"(v_offset), "s"(*(const i32x4*)&br), "s"(s_offset), "i"(i_offset)
+    : "memory");
+}
+
 template<int GPR>
 __device__ __forceinline__ void buffer_atomic_pk_add_bf16(const buffer_resource& br, const uint32_t v_offset, const uint32_t s_offset = 0, const int i_offset = 0) {
   if constexpr (GPR >= 256) {

extra/thunder/amd/include/cdna4/common/util.cuh

@@ -52,6 +52,8 @@ __device__ __forceinline__ int laneid() { return threadIdx.x & 0x3f; }
 
 using i32x2 = int32_t __attribute__((ext_vector_type(2)));
 using u32x2 = uint32_t __attribute__((ext_vector_type(2)));
+using i32x3 = int32_t __attribute__((ext_vector_type(3)));
+using u32x3 = uint32_t __attribute__((ext_vector_type(3)));
 using i32x4 = int32_t __attribute__((ext_vector_type(4)));
 using u32x4 = uint32_t __attribute__((ext_vector_type(4)));
 

extra/thunder/amd/include/cdna4/includes.cuh

@@ -0,0 +1,10 @@
+/**
+ * @file
+ * @brief Aggregate header for CDNA4 ThunderKittens core include tree.
+ */
+
+#pragma once
+
+#include "common/common.cuh"
+#include "types/types.cuh"
+#include "ops/ops.cuh"

extra/thunder/amd/include/cdna4/ops/warp/memory/tile/shared_to_register.cuh

@@ -248,7 +248,7 @@ __device__ inline static void load(RT &dst, const ST &src) {
                                         "ds_read_b64_tr_b16 %0, %2 offset:%3\n"
                                         "ds_read_b64_tr_b16 %1, %2 offset:%4\n"
                                         // "s_waitcnt lgkmcnt(0)\n"
-                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])), 
+                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])),
                                         "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx + 2]))
                                         : "v"(addr), "i"(offset), "i"(offset + 4 * ST::underlying_subtile_row_bytes)
                                         : "memory"
@@ -259,7 +259,7 @@ __device__ inline static void load(RT &dst, const ST &src) {
                                         "ds_read_b64_tr_b16 %0, %2 offset:%4\n"
                                         "ds_read_b64_tr_b16 %1, %3 offset:%4\n"
                                         // "s_waitcnt lgkmcnt(0)\n"
-                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])), 
+                                        : "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx])),
                                         "=v"(*reinterpret_cast<float2*>(&dst.tiles[register_row][register_col].data[idx + 2]))
                                         : "v"(addr), "v"(next_addr), "i"(offset)
                                         : "memory"
@@ -688,4 +688,88 @@ __device__ inline static void store(ST &dst, const RT &src) {
     }
 }
 
+template<ducks::rt::col_layout RT, ducks::st::all ST>
+__device__ inline static void load(RT &dst, const ST &src, int col_offset) {
+    static_assert(RT::cols == ST::rows,
+                  "col_layout load with col_offset: ST.rows must equal RT::cols");
+    static_assert(RT::width == 1, "col_layout load with col_offset: only width==1 supported");
+
+    using T2 = RT::dtype;
+    using U  = ST::dtype;
+    using U2 = base_types::packing<U>::packed_type;
+
+    const int laneid = kittens::laneid();
+    const uint32_t src_ptr = (uint32_t)(uintptr_t)(&src.data[0]);
+    constexpr int subtile_bytes = ST::underlying_subtile_bytes;
+
+    if constexpr (std::is_same_v<T2, fp8e4m3_4>) {
+        const int block_id = laneid / 16;
+        const int l_within = laneid % 16;
+        const int tr_k_grp = l_within / 2;
+        const int m_half   = l_within & 1;
+
+        #pragma unroll
+        for (int i = 0; i < RT::height; i++) {
+            const int m_col = col_offset + i * RT::base_tile_rows + m_half * 8;
+
+            const int subtile_base = block_id;
+            uint32_t addr = src_ptr
+                          + (uint32_t)(subtile_base * subtile_bytes)
+                          + src.swizzle({tr_k_grp, m_col});
+
+            asm volatile(
+                "ds_read_b64_tr_b8 %0, %2 offset:0\n"
+                "ds_read_b64_tr_b8 %1, %2 offset:%3\n"
+                : "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[0])),
+                  "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[4]))
+                : "v"(addr), "i"(4 * subtile_bytes)
+                : "memory"
+            );
+
+            addr ^= 1088u;
+
+            asm volatile(
+                "ds_read_b64_tr_b8 %0, %2 offset:0\n"
+                "ds_read_b64_tr_b8 %1, %2 offset:%3\n"
+                : "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[2])),
+                  "=&v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[6]))
+                : "v"(addr), "i"(4 * subtile_bytes)
+                : "memory"
+            );
+        }
+    } else if constexpr (std::is_same_v<U2, bf16_2> || std::is_same_v<U2, half_2>) {
+        const int row_offset = ((laneid % 16) / 4) + ((laneid / 16) * 4);
+        const int col_offset_lane = (laneid % 4) * 4;
+
+        #pragma unroll
+        for (int i = 0; i < RT::height; i++) {
+            const int m_col = col_offset + i * RT::base_tile_rows + col_offset_lane;
+            const uint32_t addr = src_ptr + src.swizzle({row_offset, m_col});
+            const uint32_t next_addr = src_ptr + src.swizzle({row_offset + 4, m_col});
+
+            if constexpr (RT::base_tile_stride == 8) {
+                asm volatile(
+                    "ds_read_b64_tr_b16 %0, %2 offset:0\n"
+                    "ds_read_b64_tr_b16 %1, %3 offset:0\n"
+                    : "=v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[0])),
+                      "=v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[2]))
+                    : "v"(addr), "v"(next_addr)
+                    : "memory"
+                );
+            } else if constexpr (RT::base_tile_stride == 4) {
+                asm volatile(
+                    "ds_read_b64_tr_b16 %0, %1 offset:0\n"
+                    : "=v"(*reinterpret_cast<float2*>(&dst.tiles[i][0].data[0]))
+                    : "v"(addr)
+                    : "memory"
+                );
+            } else {
+                static_assert(false, "Unsupported stride for col_layout load with col_offset");
+            }
+        }
+    } else {
+        static_assert(false, "Unsupported type for col_layout load with col_offset");
+    }
+}
+
 } // namespace kittens

extra/thunder/amd/include/cdna4/ops/warp/register/tile/mma.cuh

@@ -201,10 +201,10 @@ __device__ static inline void mma_AB_base(rt_base<float, ducks::rt_layout::col,
  * @param[in] b The second input rt_base<bf16_2, row_layout> matrix in row-major mode.
  * @param[in] c The input rt_base<float2, row_layout> accumulator matrix.
  */
-template<ducks::rt_shape::all D_shape, ducks::rt_shape::all A_shape, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T=bf16>
+template<ducks::rt_shape::all D_shape, ducks::rt_layout::all A_layout, ducks::rt_shape::all A_shape, ducks::rt_layout::all B_layout, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T=bf16>
 __device__ static inline void mma_ABt_base(rt_base<float, ducks::rt_layout::col, D_shape> &d,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, A_shape> &a,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, B_shape> &b, // in row-major mode
+    const rt_base<MM_Operand_T, A_layout, A_shape> &a,
+    const rt_base<MM_Operand_T, B_layout, B_shape> &b,
     const rt_base<float, ducks::rt_layout::col, C_shape> &c) {
 
     static_assert(std::is_same_v<D_shape, C_shape>, "D and C must have the same shape");
@@ -243,6 +243,32 @@ __device__ static inline void mma_ABt_base(rt_base<float, ducks::rt_layout::col,
     }
 }
 
+/**
+ * @brief Pack 4 E8M0 scale bytes from LDS into one fp8e8m0_4 register.
+ *
+ * @param smem_scales LDS pointer to scale tile data.
+ * @param row_offset Starting row within the scale region (warp's tile offset).
+ * @return fp8e8m0_4 with 4 scale bytes packed for MFMA opsel.
+ */
+__device__ __forceinline__ fp8e8m0_4 pack_scales(
+    const fp8e8m0 *smem_scales, int row_offset) {
+    int lid   = laneid();
+    int r16   = lid % 16;
+    int k_sub = lid / 16;
+
+    const fp8e8m0_4 *s4 = (const fp8e8m0_4 *)smem_scales;
+    fp8e8m0_4 w0 = s4[row_offset + 0 * 16 + r16];
+    fp8e8m0_4 w1 = s4[row_offset + 1 * 16 + r16];
+    fp8e8m0_4 w2 = s4[row_offset + 2 * 16 + r16];
+    fp8e8m0_4 w3 = s4[row_offset + 3 * 16 + r16];
+
+    fp8e8m0_4 sel = 0x0C0C0000u | (k_sub << 8) | (4u + k_sub);
+    fp8e8m0_4 lo  = __builtin_amdgcn_perm(w0, w1, sel);
+    fp8e8m0_4 hi  = __builtin_amdgcn_perm(w2, w3, sel);
+
+    return lo | (hi << 16);
+}
+
 /**
  * @brief Base dot product operation for row layout.
  *
@@ -254,10 +280,10 @@ __device__ static inline void mma_ABt_base(rt_base<float, ducks::rt_layout::col,
  * @param[in] b The second input rt_base<Operand_T, row_layout> matrix.
  * @param[in] c The input rt_base<float, col_layout> accumulator matrix.
  */
-template<int opsel_a, int opsel_b, int cbsz = 0, int blgp = 0, ducks::rt_shape::all D_shape, ducks::rt_shape::all A_shape, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T>
+template<int opsel_a, int opsel_b, int cbsz = 0, int blgp = 0, ducks::rt_shape::all D_shape, ducks::rt_layout::all A_layout, ducks::rt_shape::all A_shape, ducks::rt_layout::all B_layout, ducks::rt_shape::all B_shape, ducks::rt_shape::all C_shape, typename MM_Operand_T>
 __device__ static inline void mma_ABt_base_scaled(rt_base<float, ducks::rt_layout::col, D_shape> &d,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, A_shape> &a,
-    const rt_base<MM_Operand_T, ducks::rt_layout::row, B_shape> &b,
+    const rt_base<MM_Operand_T, A_layout, A_shape> &a,
+    const rt_base<MM_Operand_T, B_layout, B_shape> &b,
     const rt_base<float, ducks::rt_layout::col, C_shape> &c,
     const fp8e8m0_4 *scale_a,
     const fp8e8m0_4 *scale_b) {
@@ -441,7 +467,7 @@ __device__ static inline void mma_AB(D &d,
  * @param[in] b The second input rt_bf<M, K, row_layout> matrix in row-major mode.
  * @param[in] c The input rt_fl<N, M, row_layout> accumulator matrix.
  */
-template<ducks::rt::col_layout D, ducks::rt::row_layout A, ducks::rt::row_layout B, ducks::rt::col_layout C>
+template<ducks::rt::col_layout D, ducks::rt::all A, ducks::rt::all B, ducks::rt::col_layout C>
 __device__ static inline void mma_ABt(D &d,
                                 const A &a,
                                 const B &b, // notice row and (M, K) instead of col and (K, M)
@@ -499,7 +525,7 @@ __device__ static inline void mma_ABt(D &d,
  * @param[in] scale_a Pointer to the packed E8M0 scale for the A matrix.
  * @param[in] scale_b Pointer to the packed E8M0 scale for the B matrix.
  */
-template<int cbsz = 0, int blgp = 0, ducks::rt::col_layout D, ducks::rt::row_layout A, ducks::rt::row_layout B, ducks::rt::col_layout C>
+template<int cbsz = 0, int blgp = 0, ducks::rt::col_layout D, ducks::rt::all A, ducks::rt::all B, ducks::rt::col_layout C>
 __device__ static inline void mma_ABt_scaled(D &d,
                                 const A &a,
                                 const B &b,

extra/thunder/amd/include/cdna4/ops/warp/register/tile/tile.cuh

@@ -9,6 +9,5 @@
 #include "maps.cuh"
 #include "reductions.cuh"
 #include "mma.cuh"
-#include "scales.cuh"
 
 #include "assembly/tile.cuh"