not a part of regen

fix torch frontend
imports
2025-09-05 15:41:34 -07:00 · 2025-09-05 15:39:52 -07:00 · 2025-09-05 15:34:49 -07:00 · 2025-09-05 15:31:39 -07:00 · 2025-09-05 15:15:05 -07:00 · 2025-09-05 15:13:36 -07:00
792 changed files with 222710 additions and 208305 deletions
@@ -1,3 +0,0 @@
-[run]
-source = tinygrad
-branch = True
@@ -41,10 +41,6 @@ inputs:
    description: "Install LLVM?"
    required: false
    default: 'false'
-  mesa:
-    description: "Install mesa"
-    required: false
-    default: 'false'
 runs:
  using: "composite"
  steps:
@@ -61,7 +57,7 @@ runs:
      uses: actions/cache@v4
      with:
        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ hashFiles('**/setup.py') }}-${{ env.PYTHON_CACHE_VERSION }}

    # **** Caching downloads ****

@@ -70,13 +66,13 @@ runs:
      uses: actions/cache@v4
      with:
        path: ~/.cache/tinygrad/downloads/
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
+        key: downloads-cache-${{ inputs.key }}-${{ env.DOWNLOAD_CACHE_VERSION }}
    - name: Cache downloads (macOS)
      if: inputs.key != '' && runner.os == 'macOS'
      uses: actions/cache@v4
      with:
        path: ~/Library/Caches/tinygrad/downloads/
-        key: downloads-${{ github.job }}-${{ inputs.key }}-${{ env.CACHE_VERSION }}
+        key: osx-downloads-cache-${{ inputs.key }}-${{ env.DOWNLOAD_CACHE_VERSION }}

    # **** Python deps ****

@@ -187,7 +183,7 @@ runs:
      uses: actions/cache@v4
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.APT_CACHE_VERSION }}

    - name: Run apt Update + Install
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -221,7 +217,7 @@ runs:
        sudo mkdir -p /usr/local/lib
        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/amdcomgr_dylib/releases/latest | \
          jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
-          sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib
+          sudo xargs curl -L -o /usr/local/lib/libamd_comgr.dylib
        cargo build --release --manifest-path ./extra/remu/Cargo.toml

    # **** gpuocelot ****
@@ -247,7 +243,7 @@ runs:
        cache-name: cache-gpuocelot-build-1
      with:
        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}
    - name: Clone/compile gpuocelot
      if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
      shell: bash
@@ -257,13 +253,8 @@ runs:
        git checkout b16039dc940dc6bc4ea0a98380495769ff35ed99
        mkdir build
        cd build
-
-        CMAKE_ARGS="-Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5"
-        if [[ "${{ runner.os }}" == "macOS" ]]; then
-          CMAKE_ARGS="$CMAKE_ARGS -DBoost_INCLUDE_DIR=$(brew --prefix boost)/include -DBoost_LIBRARY_DIR=$(brew --prefix boost)/lib"
-        fi
-
-        cmake .. $CMAKE_ARGS
+        cmake .. -Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF \
+          -DBoost_INCLUDE_DIR=$(brew --prefix boost)/include -DBoost_LIBRARY_DIR=$(brew --prefix boost)/lib -DCMAKE_POLICY_VERSION_MINIMUM=3.5
        ninja
    - name: Install gpuocelot
      if: inputs.ocelot == 'true'
@@ -278,7 +269,7 @@ runs:
      if: inputs.webgpu == 'true' && runner.os == 'Linux'
      shell: bash
      run: |
-        sudo curl -fL https://github.com/wpmed92/pydawn/releases/download/v0.1.6/libwebgpu_dawn.so -o /usr/local/lib/libwebgpu_dawn.so
+        sudo curl -L https://github.com/wpmed92/pydawn/releases/download/v0.1.6/libwebgpu_dawn.so -o /usr/local/lib/libwebgpu_dawn.so
        sudo ldconfig
    - name: Install WebGPU dawn (macOS)
      if: inputs.webgpu == 'true' && runner.os == 'macOS'
@@ -293,13 +284,3 @@ runs:
      if: inputs.llvm == 'true' && runner.os == 'macOS'
      shell: bash
      run: brew install llvm@20
-
-    # **** mesa ****
-    - name: Install mesa (linux)
-      if: inputs.mesa == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinymesa/releases/download/v1/libtinymesa_cpu-mesa-25.2.7-linux-amd64.so -o /usr/lib/libtinymesa_cpu.so
-    - name: Install mesa (macOS)
-      if: inputs.mesa == 'true' && runner.os == 'macOS'
-      shell: bash
-      run: brew install sirhcm/tinymesa/tinymesa_cpu
@@ -1,176 +0,0 @@
-name: Autogen
-env:
-  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '13'
-  CAPTURE_PROCESS_REPLAY: 1
-  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-  PYTHONPATH: ${{ github.workspace }}
-
-on:
-  push:
-    branches:
-      - master
-  pull_request:
-    paths:
-    - 'tinygrad/runtime/autogen/**/*'
-    - 'tinygrad/runtime/support/autogen.py'
-  workflow_dispatch:
-    paths:
-    - 'tinygrad/runtime/autogen/**/*'
-    - 'tinygrad/runtime/support/autogen.py'
-
-jobs:
-  autogen:
-    name: In-tree Autogen
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        opencl: 'true'
-        amd: 'true'
-        cuda: 'true'
-        llvm: 'true'
-        webgpu: 'true'
-        mesa: 'true'
-        pydeps: 'pyyaml mako'
-    - name: Install autogen support packages
-      run: sudo apt-get install -y --no-install-recommends libclang-20-dev llvm-20-dev hip-dev libusb-1.0-0-dev
-    - name: Verify OpenCL autogen
-      run: |
-        mv tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
-        python3 -c "from tinygrad.runtime.autogen import opencl"
-        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
-    - name: Verify CUDA autogen
-      run: |
-        mv tinygrad/runtime/autogen/cuda.py /tmp/cuda.py.bak
-        mv tinygrad/runtime/autogen/nvrtc.py /tmp/nvrtc.py.bak
-        mv tinygrad/runtime/autogen/nvjitlink.py /tmp/nvjitlink.py.bak
-        mv tinygrad/runtime/autogen/nv_570.py /tmp/nv_570.py.bak
-        mv tinygrad/runtime/autogen/nv.py /tmp/nv.py.bak
-        python3 -c "from tinygrad.runtime.autogen import cuda, nvrtc, nvjitlink, nv_570, nv"
-        diff /tmp/cuda.py.bak tinygrad/runtime/autogen/cuda.py
-        diff /tmp/nvrtc.py.bak tinygrad/runtime/autogen/nvrtc.py
-        diff /tmp/nvjitlink.py.bak tinygrad/runtime/autogen/nvjitlink.py
-        diff /tmp/nv_570.py.bak tinygrad/runtime/autogen/nv_570.py
-        diff /tmp/nv.py.bak tinygrad/runtime/autogen/nv.py
-    - name: Verify AMD autogen
-      run: |
-        mv tinygrad/runtime/autogen/comgr.py /tmp/comgr.py.bak
-        mv tinygrad/runtime/autogen/hsa.py /tmp/hsa.py.bak
-        mv tinygrad/runtime/autogen/hip.py /tmp/hip.py.bak
-        mv tinygrad/runtime/autogen/amd_gpu.py /tmp/amd_gpu.py.bak
-        mv tinygrad/runtime/autogen/sqtt.py /tmp/sqtt.py.bak
-        mv tinygrad/runtime/autogen/rocprof.py /tmp/rocprof.py.bak
-        mv tinygrad/runtime/autogen/am/am.py /tmp/am_am.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_soc15.py /tmp/am_pm4_soc15.py.bak
-        mv tinygrad/runtime/autogen/am/pm4_nv.py /tmp/am_pm4_nv.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_4_0_0.py /tmp/am_sdma_4_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_5_0_0.py /tmp/am_sdma_5_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/sdma_6_0_0.py /tmp/am_sdma_6_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v13_0_0.py /tmp/am_smu_v13_0_0.py.bak
-        mv tinygrad/runtime/autogen/am/smu_v14_0_2.py /tmp/am_smu_v14_0_2.py.bak
-        python3 -c "from tinygrad.runtime.autogen import comgr, hsa, hip, amd_gpu, sqtt, rocprof; from tinygrad.runtime.autogen.am import am, pm4_soc15, pm4_nv, sdma_4_0_0, sdma_5_0_0, sdma_6_0_0, smu_v13_0_0, smu_v14_0_2"
-        diff /tmp/comgr.py.bak tinygrad/runtime/autogen/comgr.py
-        diff /tmp/hsa.py.bak tinygrad/runtime/autogen/hsa.py
-        diff /tmp/hip.py.bak tinygrad/runtime/autogen/hip.py
-        diff /tmp/amd_gpu.py.bak tinygrad/runtime/autogen/amd_gpu.py
-        diff /tmp/sqtt.py.bak tinygrad/runtime/autogen/sqtt.py
-        diff /tmp/rocprof.py.bak tinygrad/runtime/autogen/rocprof.py
-        diff /tmp/am_am.py.bak tinygrad/runtime/autogen/am/am.py
-        diff /tmp/am_pm4_soc15.py.bak tinygrad/runtime/autogen/am/pm4_soc15.py
-        diff /tmp/am_pm4_nv.py.bak tinygrad/runtime/autogen/am/pm4_nv.py
-        diff /tmp/am_sdma_4_0_0.py.bak tinygrad/runtime/autogen/am/sdma_4_0_0.py
-        diff /tmp/am_sdma_5_0_0.py.bak tinygrad/runtime/autogen/am/sdma_5_0_0.py
-        diff /tmp/am_sdma_6_0_0.py.bak tinygrad/runtime/autogen/am/sdma_6_0_0.py
-        diff /tmp/am_smu_v13_0_0.py.bak tinygrad/runtime/autogen/am/smu_v13_0_0.py
-        diff /tmp/am_smu_v14_0_2.py.bak tinygrad/runtime/autogen/am/smu_v14_0_2.py
-    - name: Verify Linux autogen
-      run: |
-        mv tinygrad/runtime/autogen/libc.py /tmp/libc.py.bak
-        mv tinygrad/runtime/autogen/kfd.py /tmp/kfd.py.bak
-        mv tinygrad/runtime/autogen/io_uring.py /tmp/io_uring.py.bak
-        mv tinygrad/runtime/autogen/ib.py /tmp/ib.py.bak
-        mv tinygrad/runtime/autogen/pci.py /tmp/pci.py.bak
-        mv tinygrad/runtime/autogen/vfio.py /tmp/vfio.py.bak
-        python3 -c "from tinygrad.runtime.autogen import libc, kfd, io_uring, ib, pci, vfio"
-        diff /tmp/libc.py.bak tinygrad/runtime/autogen/libc.py
-        diff /tmp/kfd.py.bak tinygrad/runtime/autogen/kfd.py
-        diff /tmp/io_uring.py.bak tinygrad/runtime/autogen/io_uring.py
-        diff /tmp/ib.py.bak tinygrad/runtime/autogen/ib.py
-        diff /tmp/pci.py.bak tinygrad/runtime/autogen/pci.py
-        diff /tmp/vfio.py.bak tinygrad/runtime/autogen/vfio.py
-    - name: Verify LLVM autogen
-      run: |
-        mv tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
-        python3 -c "from tinygrad.runtime.autogen import llvm"
-        diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
-    - name: Verify WebGPU autogen
-      run: |
-        mv tinygrad/runtime/autogen/webgpu.py /tmp/webgpu.py.bak
-        python3 -c "from tinygrad.runtime.autogen import webgpu"
-        diff /tmp/webgpu.py.bak tinygrad/runtime/autogen/webgpu.py
-    - name: Verify Qualcomm autogen
-      run: |
-        mv tinygrad/runtime/autogen/kgsl.py /tmp/kgsl.py.bak
-        mv tinygrad/runtime/autogen/qcom_dsp.py /tmp/qcom_dsp.py.bak
-        python3 -c "from tinygrad.runtime.autogen import kgsl, qcom_dsp"
-        diff /tmp/kgsl.py.bak tinygrad/runtime/autogen/kgsl.py
-        diff /tmp/qcom_dsp.py.bak tinygrad/runtime/autogen/qcom_dsp.py
-    - name: Verify libusb autogen
-      run: |
-        mv tinygrad/runtime/autogen/libusb.py /tmp/libusb.py.bak
-        python3 -c "from tinygrad.runtime.autogen import libusb"
-        diff /tmp/libusb.py.bak tinygrad/runtime/autogen/libusb.py
-    - name: Verify mesa autogen
-      run: |
-        mv tinygrad/runtime/autogen/mesa.py /tmp/mesa.py.bak
-        python3 -c "from tinygrad.runtime.autogen import mesa"
-        diff /tmp/mesa.py.bak tinygrad/runtime/autogen/mesa.py
-    - name: Verify libclang autogen
-      run: |
-        cp tinygrad/runtime/autogen/libclang.py /tmp/libclang.py.bak
-        REGEN=1 python3 -c "from tinygrad.runtime.autogen import libclang"
-        diff /tmp/libclang.py.bak tinygrad/runtime/autogen/libclang.py
-  autogen-mac:
-    name: In-tree Autogen (macos)
-    runs-on: macos-14
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        llvm: 'true'
-    - name: Verify macos autogen
-      run: |
-        mv tinygrad/runtime/autogen/metal.py /tmp/metal.py.bak
-        LIBCLANG_PATH=/opt/homebrew/opt/llvm@20/lib/libclang.dylib python3 -c "from tinygrad.runtime.autogen import metal"
-        diff /tmp/metal.py.bak tinygrad/runtime/autogen/metal.py
-  autogen-comgr-3:
-    name: In-tree Autogen (comgr 3)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-    - name: Install autogen support packages
-      run: |
-        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-        sudo tee /etc/apt/sources.list.d/rocm.list <<EOF
-        deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.4 $(lsb_release -cs) main
-        EOF
-        echo -e 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600' | sudo tee /etc/apt/preferences.d/rocm-pin-600
-        sudo apt -qq update || true
-        sudo apt-get install -y --no-install-recommends libclang-20-dev comgr
-    - name: Verify comgr (3) autogen
-      run: |
-        mv tinygrad/runtime/autogen/comgr_3.py /tmp/comgr_3.py.bak
-        python3 -c "from tinygrad.runtime.autogen import comgr_3"
-        diff /tmp/comgr_3.py.bak tinygrad/runtime/autogen/comgr_3.py
@@ -14,6 +14,12 @@ on:
      - update_benchmark
      - update_benchmark_staging
  workflow_dispatch:
+    inputs:
+      run_process_replay:
+        description: "Run process replay tests"
+        required: false
+        default: false
+        type: boolean

 jobs:
  testmacbenchmark:
@@ -22,7 +28,7 @@ jobs:
      # since sudo is required for usbgpu on macos, move the cache to a new location, as some of the files are owned by root
      PYTHONPYCACHEPREFIX: /tmp/tiny_python_pycache
    runs-on: [self-hosted, macOS]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -33,7 +39,6 @@ jobs:
    - 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
@@ -46,30 +51,27 @@ jobs:
        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
    - 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 | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion JIT=1 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - 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 | tee sd_no_fp16.txt
+      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
    - 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 | tee sdv2.txt
+      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
    # process replay can't capture this, the graph is too large
    - 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 | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run model inference benchmark
-      run: METAL=1 NOCLANG=1 python3.11 test/external/external_model_benchmark.py
+      run: METAL=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 | tee torch_speed.txt
    - name: Test tensor cores
      run: METAL=1 python3.11 test/opt/test_tensor_cores.py
    - name: Test AMX tensor cores
      run: |
-        DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_tensor_cores.py
-        DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_tensor_cores.py
-        DEBUG=2 CPU=1 CPU_LLVM=0 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
-        DEBUG=2 CPU=1 CPU_LLVM=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
+        DEBUG=2 CPU=1 AMX=1 python3.11 test/opt/test_tensor_cores.py
+        DEBUG=2 LLVM=1 AMX=1 python3.11 test/opt/test_tensor_cores.py
+        DEBUG=2 CPU=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
+        DEBUG=2 LLVM=1 AMX=1 python3.11 test/opt/test_gen_float4.py TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
    - name: Run Tensor Core GEMM (float)
      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
    - name: Run Tensor Core GEMM (half)
@@ -97,7 +99,7 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 JIT=1 ASSERT_MIN_STEP_TIME=13 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 JIT=1 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
      run: BENCHMARK_LOG=gpt2_half HALF=1 python3.11 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
@@ -106,19 +108,22 @@ jobs:
      run: BENCHMARK_LOG=olmoe python3.11 examples/olmoe.py
    - name: Train MNIST
      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
-
-    # 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 | tee train_cifar.txt
-    #- 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 | tee train_cifar_half.txt
-
+    - name: Run 10 CIFAR training steps
+      run: BENCHMARK_LOG=cifar_10steps JIT=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+    - name: Run 10 CIFAR training steps w HALF
+      run: BENCHMARK_LOG=cifar_10steps_half JIT=2 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
    #- name: Run 10 CIFAR training steps w BF16
    #  run: STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3.11 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
-    # 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 | tee train_cifar_wino.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+    - name: UsbGPU boot time
+      run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
+    - name: UsbGPU tiny tests
+      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
+    - name: UsbGPU copy speeds
+      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
+    #- name: UsbGPU openpilot test
+    #  run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB NOLOCALS=0 IMAGE=0 GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (Mac)
@@ -152,41 +157,10 @@ jobs:
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3.11 process_replay.py

-  testusbgpu:
-    name: UsbGPU Benchmark
-    env:
-      PYTHONPYCACHEPREFIX: /tmp/tiny_python_pycache
-    runs-on: [self-hosted, macOS]
-    timeout-minutes: 10
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - 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: UsbGPU boot time
-      run: sudo -E PYTHONPATH=. DEBUG=2 AM_RESET=1 AMD=1 AMD_IFACE=USB time python3.11 test/test_tiny.py TestTiny.test_plus
-    - name: UsbGPU tiny tests
-      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/test_tiny.py
-    - name: UsbGPU copy speeds
-      run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB python3.11 test/external/external_test_usb_asm24.py TestDevCopySpeeds
-    #- name: UsbGPU openpilot test
-    #  run: sudo -E PYTHONPATH=. AMD=1 AMD_IFACE=USB GRAPH_ONE_KERNEL=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
-    - name: UsbGPU (USB4/TB) boot time
-      run: PYTHONPATH=. DEBUG=3 NV=1 NV_IFACE=PCI NV_NAK=1 time python3.11 test/test_tiny.py TestTiny.test_plus
-    - name: UsbGPU (USB4/TB) tiny tests
-      run: PYTHONPATH=. NV=1 NV_IFACE=PCI NV_NAK=1 python3.11 test/test_tiny.py
-
  testnvidiabenchmark:
    name: tinybox green Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
+    timeout-minutes: 30
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -217,21 +191,20 @@ jobs:
    - name: Test speed vs torch
      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test speed vs theoretical
-      run: NV=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      run: NV=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
    - name: Test benchmark allreduce
      run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
    - name: Test tensor cores
      run: |
        NV=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
-        NV=1 NV_PTX=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+        PTX=1 ALLOW_TF32=1 NV=1 python3 test/opt/test_tensor_cores.py
    - name: Run Tensor Core GEMM (CUDA)
      run: |
        CUDA=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul.txt
        CUDA=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_bfloat16.txt
        CUDA=1 SHOULD_USE_TC=1 ALLOW_TF32=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_tf32.txt
-        CUDA=1 SHOULD_USE_TC=1 FP8E4M3=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_fp8.txt
    - name: Run Tensor Core GEMM (PTX)
-      run: NV=1 NV_PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
+      run: NV=1 PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
    - name: Run Tensor Core GEMM (NV)
      run: NV=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_nv.txt
    - name: Test NV=1
@@ -240,9 +213,8 @@ jobs:
      run: DEBUG=2 CUDA=1 python -m pytest -rA test/test_tiny.py
    - name: Run Stable Diffusion
      run: BENCHMARK_LOG=stable_diffusion NV=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
-    # TODO: too slow
-    # - name: Run SDXL
-    #   run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+    - name: Run SDXL
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA
      run: |
        BENCHMARK_LOG=llama_nojit NV=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
@@ -257,8 +229,6 @@ jobs:
      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_beam.txt
    - name: Run LLaMA-3 8B on 4 GPUs with BEAM
      run: BENCHMARK_LOG=llama3_beam_4gpu NV=1 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 | tee llama3_four_gpu.txt
-    - 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 | tee llama3_fp8.txt
    # - name: Run LLaMA-3 8B on 6 GPUs
    #   run: NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/llama3.py --size 8B --shard 6 --model weights/LLaMA-3/8B-SF-DPO/ --benchmark --temperature 0 | tee llama3_six_gpu.txt
    # - name: Run LLaMA-2 70B
@@ -268,9 +238,9 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit NV=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 NV=1 JIT=1 ASSERT_MIN_STEP_TIME=4 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 NV=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 ASSERT_MIN_STEP_TIME=6 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
      run: BENCHMARK_LOG=gpt2_half_beam NV=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
@@ -292,7 +262,6 @@ jobs:
          llama3_beam.txt
          llama3_four_gpu.txt
          llama3_six_gpu.txt
-          llama3_fp8.txt
          llama_2_70B.txt
          mixtral.txt
          gpt2_unjitted.txt
@@ -305,7 +274,7 @@ jobs:
  testmorenvidiabenchmark:
    name: tinybox green Training Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -330,37 +299,33 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    # TODO: too slow
-    # - name: Fuzz Padded Tensor Core GEMM (NV)
-    #   run: NV=1 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: NV=1 NV_PTX=1 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 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+    - name: Fuzz Padded Tensor Core GEMM (NV)
+      run: NV=1 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: Fuzz Padded Tensor Core GEMM (PTX)
+      run: NV=1 PTX=1 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: Train MNIST
      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      run: BENCHMARK_LOG=cifar_10steps_half NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
-    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=350 NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_bf16 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: BENCHMARK_LOG=cifar_10steps_half_wino NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+      run: time BENCHMARK_LOG=cifar_6gpu CAPTURE_PROCESS_REPLAY=0 NV=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
    - name: Run MLPerf resnet eval on training data
      run: time BENCHMARK_LOG=resnet_eval NV=1 MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps NV=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA Training)
@@ -381,7 +346,7 @@ jobs:
  testamdbenchmark:
    name: tinybox red Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -428,7 +393,7 @@ jobs:
    #    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 | tee torch_speed.txt
    - name: Test speed vs theoretical
-      run: AMD=1 IGNORE_BEAM_CACHE=1 CCACHE=0 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
+      run: AMD=1 IGNORE_BEAM_CACHE=1 BEAM_DEBUG=1 DEBUG=1 python -m pytest -rA test/external/speed_v_theoretical.py --durations=20
    - name: Test tensor cores
      run: |
        AMD=1 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
@@ -450,9 +415,9 @@ jobs:
    - name: Test AM warm start time
      run: time AMD=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion ASSERT_MIN_STEP_TIME=550 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      run: BENCHMARK_LOG=stable_diffusion AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl ASSERT_MIN_STEP_TIME=3200 CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run LLaMA 7B
      run: |
        BENCHMARK_LOG=llama_nojit AMD=1 JIT=0 python3 examples/llama.py --gen 1 --prompt "Hello." --count 10 --temperature 0 --timing | tee llama_unjitted.txt
@@ -478,9 +443,9 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit AMD=1 JIT=0 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
+        BENCHMARK_LOG=gpt2 AMD=1 JIT=1 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.txt
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 ASSERT_MIN_STEP_TIME=5 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
    - name: Run GPT2 w HALF/BEAM
      run: BENCHMARK_LOG=gpt2_half_beam AMD=1 HALF=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half_beam.txt
    - uses: actions/upload-artifact@v4
@@ -511,7 +476,7 @@ jobs:
  testmoreamdbenchmark:
    name: tinybox red Training Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 30
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -543,18 +508,19 @@ jobs:
    - name: Train MNIST
      run: time PYTHONPATH=. AMD=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
-    # - name: Run 10 CIFAR training steps w BF16
-    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
-    # TODO: too slow
-    # - name: Run 10 CIFAR training steps w winograd
-    #   run: BENCHMARK_LOG=cifar_10steps_half_wino ASSERT_MIN_STEP_TIME=66 AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_half AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+    - name: Run 10 CIFAR training steps w BF16
+      run: BENCHMARK_LOG=cifar_10steps_bf16 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+    - name: Run 10 CIFAR training steps w winograd
+      run: BENCHMARK_LOG=cifar_10steps_half_wino AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
-    - name: Run full CIFAR training steps w 6 GPUS
-      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_one_gpu.txt
+    #- name: Run full CIFAR training steps w 6 GPUS
+    #  run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu.txt
+    #- name: Run full CIFAR training steps w 6 GPUS (REMOTE)
+    #  run: time BENCHMARK_LOG=cifar_6gpu_remote REMOTE=1 REMOTEDEV=AMD DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee train_cifar_six_gpu_remote.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD Training)
@@ -566,13 +532,14 @@ jobs:
          train_cifar_wino.txt
          train_cifar_one_gpu.txt
          train_cifar_six_gpu.txt
+          train_cifar_six_gpu_remote.txt
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

  testmlperfamdbenchmark:
    name: tinybox red MLPerf Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 60
+    timeout-minutes: 30
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -603,13 +570,13 @@ jobs:
      run: test/external/process_replay/reset.py
    - name: Run MLPerf resnet eval
      run: time BENCHMARK_LOG=resnet_eval AMD=1 MODEL=resnet python3 examples/mlperf/model_eval.py
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
-    - name: Run 10 MLPerf ResNet50 training steps (6 gpu)
-      run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps AMD=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet_one_gpu.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (6 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=1536 GPUS=6 MODEL=resnet python3 examples/mlperf/model_train.py | tee train_resnet.txt
    - name: Run 10 MLPerf Bert training steps (6 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=72 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
+      run: BENCHMARK_LOG=bert_10steps_6gpu AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=6 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD MLPerf)
@@ -638,35 +605,47 @@ jobs:
        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
    - name: reset process replay
      run: test/external/process_replay/reset.py
-    - name: openpilot compile3 0.10.0 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_0_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.10.0 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_0_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=11 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.10.0/selfdrive/modeld/models/dmonitoring_model.onnx
-    - name: DEBUG=2 openpilot compile3 0.10.1 driving_vision
-      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: DEBUG=2 IMAGE=1 openpilot compile3 0.10.1 driving_vision
-      run: PYTHONPATH="." DEBUG=2 DEV=QCOM FLOAT16=1 IMAGE=1 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_10_1_vision PYTHONPATH="." ASSERT_MIN_STEP_TIME=17 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot compile3 0.10.1 driving_policy
-      run: BENCHMARK_LOG=openpilot_0_10_1_policy PYTHONPATH="." ASSERT_MIN_STEP_TIME=4 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_policy.onnx
-    - name: openpilot compile3 0.10.1 dmonitoring
-      run: BENCHMARK_LOG=openpilot_0_10_1_dmonitoring PYTHONPATH="." ASSERT_MIN_STEP_TIME=10 DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: benchmark openpilot 0.9.9 driving_vision
+      run: BENCHMARK_LOG=openpilot_0_9_9_vision PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+    - name: benchmark openpilot 0.9.9 driving_policy
+      run: BENCHMARK_LOG=openpilot_0_9_9_policy PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+    - name: benchmark openpilot 0.9.9 dmonitoring
+      run: BENCHMARK_LOG=openpilot_0_9_9_dmonitoring PYTHONPATH=. NOLOCALS=1 FLOAT16=1 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 0.9.9 driving_vision
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_vision.onnx
+    - name: openpilot compile3 0.9.9 driving_policy
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/driving_policy.onnx
+    - name: openpilot compile3 0.9.9 dmonitoring
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.9/selfdrive/modeld/models/dmonitoring_model.onnx
+    - name: openpilot compile3 Space Lab policy + vision
+      run: |
+        PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/22aec22a10ce09384d4a4af2a0bbff08d54af7e0c888503508f356fae4ff0e29
+        PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/c824f68646a3b94f117f01c70dc8316fb466e05fbd42ccdba440b8a8dc86914b
    - name: benchmark MobileNetV2 on DSP
      run: |
        # generate quantized weights
        ln -s /data/home/tiny/tinygrad/extra/datasets/imagenet extra/datasets/imagenet
        ln -s /data/home/tiny/tinygrad/testsig-*.so .
-        PYTHONPATH=. CC=clang-19 CPU=1 CPU_LLVM=0 QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
+        PYTHONPATH=. CC=clang-19 CPU=1 QUANT=1 CNT=0 python3 examples/test_onnx_imagenet.py https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx /tmp/model.quant.onnx
        # benchmark on DSP with NOOPT=1, the devectorizer has issues
-        PYTHONPATH=. CC=clang-19 DSP=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+        PYTHONPATH=. CC=clang-19 DSP=1 DONT_REALIZE_EXPAND=1 NOOPT=1 CNT=2 DEBUG=2 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py
+    - uses: actions/upload-artifact@v4
+      with:
+        name: Speed (comma)
+        path: |
+          openpilot_compile_0_9_4.txt
+          openpilot_compile_0_9_7.txt
+          openpilot_0_9_4.txt
+          openpilot_0_9_7.txt
+          openpilot_image_0_9_4.txt
+          openpilot_image_0_9_7.txt

  testreddriverbenchmark:
    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
+    timeout-minutes: 15
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -716,7 +695,8 @@ jobs:
        AMD=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
        AMD=1 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 AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee am_train_cifar_one_gpu.txt
+    # TODO: enable
    # - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
    #   run: BENCHMARK_LOG=resnet_10steps AMD=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee am_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
@@ -736,7 +716,7 @@ jobs:
  testgreendriverbenchmark:
    name: NV Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
+    timeout-minutes: 15
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -778,9 +758,9 @@ jobs:
    - name: Test LLAMA-3
      run: BENCHMARK_LOG=llama3_beam NV=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 python3 examples/llama3.py --size 8B --benchmark --temperature 0 | tee nv_llama3_beam.txt
    - name: Run full CIFAR training w 1 GPU
-      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
-    - name: Run 10 MLPerf ResNet50 training steps (1 gpu)
-      run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
+      run: time BENCHMARK_LOG=cifar NV=1 DEFAULT_FLOAT=HALF LATEWINO=1 STEPS=1000 TARGET_EVAL_ACC_PCT=93.2 python3 examples/hlb_cifar10.py | tee nv_train_cifar_one_gpu.txt
+    #- name: Run 10 MLPerf ResNet50 training steps (1 gpu)
+    #  run: BENCHMARK_LOG=resnet_10steps NV=1 MNISTMOCK=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=256 GPUS=1 MODEL=resnet python3 examples/mlperf/model_train.py | tee nv_train_resnet_one_gpu.txt
    - name: Run 10 MLPerf Bert training steps (1 gpu)
      # TODO: remove BERT_LAYERS once scheduler is fast
      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
@@ -22,13 +22,13 @@ jobs:
    - name: Run SDXL with new search
      # TODO: GCVM_L2_PROTECTION_FAULT_STATUS with llvm19
      run: |
-        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 CCACHE=0 python examples/sdxl.py --noshow --timing --seed 0
+        BENCHMARK_LOG=search_sdxl PYTHONPATH=. AMD=1 JITBEAM=2 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run SDXL with cached search
      run: |
        BENCHMARK_LOG=search_sdxl_cached PYTHONPATH=. AMD=1 JITBEAM=2 python examples/sdxl.py --noshow --timing --seed 0
    - name: Run winograd cifar with new search
      run: |
-        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 IGNORE_BEAM_CACHE=1 CCACHE=0 BS=1024 STEPS=500 python examples/hlb_cifar10.py
+        BENCHMARK_LOG=search_wino_cifar WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 IGNORE_BEAM_CACHE=1 DISABLE_COMPILER_CACHE=1 BS=1024 STEPS=500 python examples/hlb_cifar10.py
    - name: Run winograd cifar with cached search
      run: |
        BENCHMARK_LOG=search_wino_cifar_cached WINO=1 DEFAULT_FLOAT=HALF JITBEAM=4 BS=1024 STEPS=500 python examples/hlb_cifar10.py
@@ -12,7 +12,7 @@ jobs:
  run_script_job:
    runs-on: [self-hosted, Linux, tinybox]
    if: github.repository_owner == 'tinygrad'
-    timeout-minutes: 720
+    timeout-minutes: 360

    steps:
    - name: Checkout Code
@@ -27,4 +27,4 @@ jobs:
      run: |
        rm "~/.cache/tinygrad/cache_mlperf.db" || true
        BENCHMARK_LOG=mlpert_train_resnet LOGMLPERF=0 CACHEDB="~/.cache/tinygrad/cache_mlperf.db" examples/mlperf/training_submission_v5.1/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-        rm "~/.cache/tinygrad/cache_mlperf.db"
+        rm "~/.cache/tinygrad/cache_mlperf.db"
@@ -20,11 +20,11 @@ jobs:
    - name: Install dependencies
      run: |
        python -m pip install --upgrade pip
-        pip install setuptools wheel build twine
+        pip install setuptools wheel twine
    - name: Build and publish
      env:
        TWINE_USERNAME: ${{ secrets.PYPI_USERNAME }}
        TWINE_PASSWORD: ${{ secrets.PYPI_PASSWORD }}
      run: |
-        python -m build
+        python setup.py sdist bdist_wheel
        twine upload dist/*
@@ -56,15 +56,15 @@ jobs:
        uses: actions/checkout@v4
        with:
          path: base
-      - name: Set up Python 3.12
+      - name: Set up Python 3.10
        uses: actions/setup-python@v5
        with:
-          python-version: '3.12'
+          python-version: '3.10'
      - name: Count Line Diff
        run: |
+          pip install tabulate
          BASE="$GITHUB_WORKSPACE/base"
          PR="$GITHUB_WORKSPACE/pr"
-          pip install tabulate $BASE
          cp "$BASE/sz.py" .
          echo "loc_content<<EOF" >> "$GITHUB_ENV"
          python sz.py "$BASE" "$PR" >> "$GITHUB_ENV"
@@ -1,7 +1,10 @@
 name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '15'
+  DOWNLOAD_CACHE_VERSION: '12'
+  PYTHON_CACHE_VERSION: '3'
+  APT_CACHE_VERSION: '1'
+  BUILD_CACHE_VERSION: '1'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
@@ -27,10 +30,12 @@ jobs:
        key: llvm-speed
        deps: testing_minimal
        llvm: 'true'
+    - name: External Benchmark Schedule
+      run: PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
    - name: Speed Test
-      run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: LLVM=1 python3 test/speed/external_test_speed_v_torch.py
    - name: Speed Test (BEAM=2)
-      run: BEAM=2 CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
+      run: BEAM=2 LLVM=1 python3 test/speed/external_test_speed_v_torch.py

  docs:
    name: Docs
@@ -43,7 +48,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        deps: docs
-        pydeps: "capstone torch"
+        pydeps: "capstone"
    - name: Build wheel and show size
      run: |
        pip install build
@@ -67,27 +72,98 @@ jobs:
        source venv/bin/activate
        pip install $GITHUB_WORKSPACE
        cp $GITHUB_WORKSPACE/examples/beautiful_mnist.py .
-        BS=2 STEPS=10 python beautiful_mnist.py
+        PYTHONPATH=$GITHUB_WORKSPACE BS=2 STEPS=10 python beautiful_mnist.py
    - name: Test Docs Build
      run: python -m mkdocs build --strict
    - name: Test Docs
-      run: python docs/abstractions3.py
-    - name: Test README
-      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
+      run: |
+        python docs/abstractions2.py
+        python docs/abstractions3.py
    - name: Test Quickstart
      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' docs/quickstart.md > quickstart.py &&  python quickstart.py
    - name: Test DEBUG
      run: DEBUG=100 python3 -c "from tinygrad import Tensor; N = 1024; a, b = Tensor.rand(N, N), Tensor.rand(N, N); c = (a.reshape(N, 1, N) * b.T.reshape(1, N, N)).sum(axis=2); print((c.numpy() - (a.numpy() @ b.numpy())).mean())"
    - name: Compile EfficientNet to C and test it
      run: |
-        CPU=1 CPU_LLVM=0 python examples/compile_efficientnet.py > recognize.c
+        CPU=1 PYTHONPATH="." python examples/compile_efficientnet.py > recognize.c
        clang -O2 recognize.c -lm -o recognize
        cat test/models/efficientnet/Chicken.jpg | ./recognize | grep cock

+  autogen:
+    name: Autogen
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        opencl: 'true'
+        amd: 'true'
+        cuda: 'true'
+        webgpu: 'true'
+        llvm: 'true'
+    - name: Install autogen support packages
+      run: sudo apt-get install -y --no-install-recommends llvm-14-dev libclang-14-dev
+    - name: Verify OpenCL autogen
+      run: |
+        cp tinygrad/runtime/autogen/opencl.py /tmp/opencl.py.bak
+        ./autogen_stubs.sh opencl
+        diff /tmp/opencl.py.bak tinygrad/runtime/autogen/opencl.py
+    - name: Verify CUDA autogen
+      run: |
+        cp tinygrad/runtime/autogen/cuda.py /tmp/cuda.py.bak
+        cp tinygrad/runtime/autogen/nv_gpu.py /tmp/nv_gpu.py.bak
+        ./autogen_stubs.sh cuda
+        ./autogen_stubs.sh nv
+        diff /tmp/cuda.py.bak tinygrad/runtime/autogen/cuda.py
+        diff /tmp/nv_gpu.py.bak tinygrad/runtime/autogen/nv_gpu.py
+    - name: Verify AMD autogen
+      run: |
+        cp tinygrad/runtime/autogen/hsa.py /tmp/hsa.py.bak
+        cp tinygrad/runtime/autogen/kfd.py /tmp/kfd.py.bak
+        cp tinygrad/runtime/autogen/comgr.py /tmp/comgr.py.bak
+        cp tinygrad/runtime/autogen/amd_gpu.py /tmp/amd_gpu.py.bak
+        cp tinygrad/runtime/autogen/sqtt.py /tmp/sqtt.py.bak
+        ./autogen_stubs.sh hsa
+        ./autogen_stubs.sh kfd
+        ./autogen_stubs.sh comgr
+        ./autogen_stubs.sh amd
+        ./autogen_stubs.sh sqtt
+        diff /tmp/hsa.py.bak tinygrad/runtime/autogen/hsa.py
+        diff /tmp/kfd.py.bak tinygrad/runtime/autogen/kfd.py
+        diff /tmp/comgr.py.bak tinygrad/runtime/autogen/comgr.py
+        diff /tmp/amd_gpu.py.bak tinygrad/runtime/autogen/amd_gpu.py
+        diff /tmp/sqtt.py.bak tinygrad/runtime/autogen/sqtt.py
+    - name: Verify Linux autogen
+      run: |
+        cp tinygrad/runtime/autogen/libc.py /tmp/libc.py.bak
+        cp tinygrad/runtime/autogen/io_uring.py /tmp/io_uring.py.bak
+        cp tinygrad/runtime/autogen/ib.py /tmp/ib.py.bak
+        ./autogen_stubs.sh libc
+        ./autogen_stubs.sh io_uring
+        ./autogen_stubs.sh ib
+        diff /tmp/libc.py.bak tinygrad/runtime/autogen/libc.py
+        diff /tmp/io_uring.py.bak tinygrad/runtime/autogen/io_uring.py
+        diff /tmp/ib.py.bak tinygrad/runtime/autogen/ib.py
+    - name: Verify WebGPU autogen
+      run: |
+        cp tinygrad/runtime/autogen/webgpu.py /tmp/webgpu.py.bak
+        ./autogen_stubs.sh webgpu
+        diff /tmp/webgpu.py.bak tinygrad/runtime/autogen/webgpu.py
+    - name: Verify LLVM autogen
+      run: |
+        cp tinygrad/runtime/autogen/llvm.py /tmp/llvm.py.bak
+        ./autogen_stubs.sh llvm
+        diff /tmp/llvm.py.bak tinygrad/runtime/autogen/llvm.py
+
  torchbackend:
    name: Torch Backend Tests
    runs-on: ubuntu-latest
    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -115,19 +191,18 @@ jobs:
    - name: Test one op in torch tests
      run: DEBUG=2 python3 extra/torch_backend/torch_tests.py TestTinyBackendPRIVATEUSE1.test_unary_log_tiny_float32
    - name: Test Ops with TINY_BACKEND
-      run: CPU=1 CPU_LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
+      run: LLVM=1 LLVMOPT=0 TINY_BACKEND=1 python3 -m pytest -n auto test/test_ops.py --durations=20
    - name: Test in-place operations on views
      run: TORCH_DEBUG=1 python3 extra/torch_backend/test_inplace.py
    - name: Test multi-gpu
-      run: CPU=1 CPU_LLVM=1 GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py
-    - name: Test kernel fusion
-      run: python3 extra/torch_backend/test_kernel_fusion.py
-
+      run: LLVM=1 GPUS=4 TORCH_DEBUG=1 python3 extra/torch_backend/test_multigpu.py

  torchbackendmore:
    name: Torch Backend Tests More
    runs-on: ubuntu-latest
    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -142,32 +217,24 @@ jobs:
        sudo apt update || true
        sudo apt install -y --no-install-recommends ninja-build
    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 CPU=1 TARGET_EVAL_ACC_PCT=90.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
+      run: SPLIT_REDUCEOP=0 FUSE_ARANGE=1 LLVM=1 TARGET_EVAL_ACC_PCT=96.0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
    - name: Test some torch tests (expect failure)
      run: python3 -m pytest extra/torch_backend/torch_tests.py -v --tb=no || true

-  bepython:
-    name: Python Backend
+  tc:
+    name: Tensor Core tests
    runs-on: ubuntu-latest
-    timeout-minutes: 15
+    timeout-minutes: 10
+    env:
+      IGNORE_OOB: 0
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: be-minimal
+        key: uops-minimal
        deps: testing_minimal
-    - name: Test dtype with Python emulator
-      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/test_dtype.py test/test_dtype_alu.py
-    - name: Test ops with Python emulator
-      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/test_ops.py --durations=20
-    - name: Test uops with Python emulator
-      run: PYTHON=1 python3 -m pytest test/test_uops.py --durations=20
-    - name: Test symbolic with Python emulator
-      run: PYTHON=1 python3 test/test_symbolic_ops.py
-    - name: test_renderer_failures with Python emulator
-      run: PYTHON=1 python3 -m pytest -rA test/test_renderer_failures.py::TestRendererFailures
    - name: Test IMAGE=2 support
      run: |
        IMAGE=2 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
@@ -201,7 +268,7 @@ jobs:
        DEBUG=2 EMULATE=CUDA FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
        DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
        DEBUG=2 EMULATE=CUDA_SM75 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
-        DEBUG=2 EMULATE=CUDA_SM89 ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
+        PYTHONPATH="." DEBUG=2 EMULATE=CUDA ALLOW_TF32=1 FORWARD_ONLY=1 PYTHON=1 python3 test/opt/test_tensor_cores.py
    - name: Test emulated INTEL OpenCL tensor cores
      run: DEBUG=2 EMULATE=INTEL FORWARD_ONLY=1 PYTHON=1 HALF=1 N=64 python3 ./extra/gemm/simple_matmul.py
    - name: Test emulated AMX tensor cores
@@ -214,6 +281,31 @@ jobs:
        DEBUG=2 EMULATE=INTEL PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStatsMatmulHalf
        DEBUG=2 AMX=1 EMULATE=AMX PYTHON=1 python3 ./test/test_uops_stats.py TestUOpsStats.test_simple_matmul

+  bepython:
+    name: Python Backend
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    env:
+      IGNORE_OOB: 0
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v4
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: be-minimal
+        deps: testing_minimal
+    - name: Test dtype with Python emulator
+      run: DEBUG=1 PYTHON=1 python3 -m pytest -n=auto test/test_dtype.py test/test_dtype_alu.py
+    - name: Test ops with Python emulator
+      run: DEBUG=2 SKIP_SLOW_TEST=1 PYTHON=1 python3 -m pytest -n=auto test/test_ops.py --durations=20
+    - name: Test uops with Python emulator
+      run: PYTHON=1 python3 -m pytest test/test_uops.py --durations=20
+    - name: Test symbolic with Python emulator
+      run: PYTHON=1 python3 test/test_symbolic_ops.py
+    - name: test_renderer_failures with Python emulator
+      run: PYTHON=1 python3 -m pytest -rA test/test_renderer_failures.py::TestRendererFailures
+
  linter:
    name: Linters
    runs-on: ubuntu-latest
@@ -227,7 +319,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: linting-only
-        python-version: '3.11'
+        python-version: '3.10'
        deps: linting
    - name: Lint bad-indentation and trailing-whitespace with pylint
      run: python -m pylint --disable=all -e W0311 -e C0303 --jobs=0 --indent-string='  ' --recursive=y .
@@ -236,14 +328,14 @@ jobs:
        pip3 install --upgrade --force-reinstall ruff==0.11.0
        python3 -m ruff check .
        python3 -m ruff check examples/mlperf/ --ignore E501
-        python3 -m ruff check extra/thunder/tiny/ --ignore E501 --ignore F841 --ignore E722
+    - name: Lint tinygrad with pylint
+      run: python -m pylint tinygrad/
    - name: Run mypy
      run: |
        python -m mypy --strict-equality --lineprecision-report .
        cat lineprecision.txt
-    # broken because of UPatAny
-    #- name: Run TYPED=1
-    #  run: TYPED=1 python -c "import tinygrad"
+    - name: Run TYPED=1
+      run: TYPED=1 python -c "import tinygrad"

  unittest:
    name: Unit Tests
@@ -257,30 +349,21 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: unittest-12
-        pydeps: "pillow numpy ftfy regex"
+        pydeps: "pillow"
        deps: testing_unit
-    - name: Check Device.DEFAULT
-      run: python -c "from tinygrad import Device; assert Device.DEFAULT == 'CPU', Device.DEFAULT"
+    - name: Test README
+      run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py &&  python README.py
    - name: Run unit tests
-      run: |
-        CPU=1 python test/unit/test_device.py TestRunAsModule.test_module_runs
-        CPU=1 python -m pytest -n=auto test/unit/ --durations=20 --deselect=test/unit/test_device.py::TestRunAsModule::test_module_runs
+      run: PYTHONPATH="." python -m pytest -n=auto test/unit/ --durations=20
    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 -m unittest test.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step test/device/test_null.py
-    # TODO: too slow
-    # - name: Run SDXL on NULL backend
-    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
-    - name: Run Clip tests for SD MLPerf on NULL backend
-      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
-    - name: Run AMD emulated BERT training on NULL backend
-      run: EMULATE=AMD_RDNA4 NULL=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+      run: PYTHONPATH="." NULL=1 python3 test/test_multitensor.py TestMultiTensor.test_data_parallel_resnet_train_step
+    - name: Run SDXL on NULL backend
+      run: MAX_BUFFER_SIZE=0 PYTHONPATH="." NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
    # TODO: support fake weights
    #- name: Run LLaMA 7B on 4 fake devices
    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
    - name: Run GC tests
-      run: python test/external/external_uop_gc.py
-    - name: External Benchmark Schedule
-      run: python3 test/external/external_benchmark_schedule.py
+      run: PYTHONPATH="." python test/external/external_uop_gc.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
    - name: Regen dataset on test_tiny
@@ -290,28 +373,8 @@ jobs:
        python extra/optimization/extract_dataset.py
        gzip -c /tmp/sops > extra/datasets/sops.gz
        #DEBUG=1 MIN_ASTS=1 python extra/optimization/get_action_space.py
-    - name: Repo line count < 20000 lines
-      run: MAX_LINE_COUNT=20000 python sz.py
-
-  spec:
-    strategy:
-      fail-fast: false
-      matrix:
-        group: [1, 2]
-    name: SPEC=2 (${{ matrix.group }})
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: spec-unit
-        deps: testing_unit
-        python-version: '3.14'
-    - name: Test SPEC=2
-      run: IGNORE_OOB=0 SPEC=2 PYTHONPATH="." pytest --maxfail=10 -n auto --durations=30 --ignore=test/models --ignore test/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" --splits 2 --group ${{ matrix.group }}
+    - name: Repo line count < 18000 lines
+      run: MAX_LINE_COUNT=18000 python sz.py

  fuzzing:
    name: Fuzzing
@@ -327,17 +390,21 @@ jobs:
        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 shapetracker
+      run: |
+        PYTHONPATH="." python test/external/fuzz_shapetracker.py
+        PYTHONPATH="." python test/external/fuzz_shapetracker_math.py
    - name: Fuzz Test shape ops
      run: python test/external/fuzz_shape_ops.py

-  testopenclimage:
-    name: CL IMAGE Tests
+  testgpuimage:
+    name: 'GPU IMAGE Tests'
    runs-on: ubuntu-22.04
-    timeout-minutes: 15
+    timeout-minutes: 10
+    env:
+      IGNORE_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -347,18 +414,25 @@ jobs:
          key: gpu-image
          deps: testing_minimal
          opencl: 'true'
-      - name: Test CL IMAGE=2 ops
+      - name: Run Kernel Count Test
+        run: PYTHONPATH="." GPU=1 python -m pytest -n=auto test/external/external_test_opt.py
+      - name: Test WINO=1
+        run: GPU=1 DEBUG=2 WINO=1 python3 test/test_ops.py TestOps.test_simple_conv2d
+      - name: Test GPU IMAGE=2 ops + training
        run: |
-          CL=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
-        # TODO: training is broken
-        # CL=1 IMAGE=2 python test/models/test_end2end.py TestEnd2End.test_linear_mnist
+          PYTHONPATH="." GPU=1 IMAGE=2 python -m pytest -n=auto test/test_ops.py --durations=20
+          PYTHONPATH="." GPU=1 IMAGE=2 python3 test/models/test_end2end.py TestEnd2End.test_linear_mnist
+      - name: Run fused optimizer tests
+        run: PYTHONPATH="." GPU=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

-  testgpumisc:
-    name: CL Misc tests
+  testgendataset:
+    name: 'GPU Generate Kernel Dataset'
    runs-on: ubuntu-22.04
    timeout-minutes: 10
+    env:
+      IGNORE_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -369,11 +443,7 @@ jobs:
          deps: testing_minimal
          opencl: 'true'
      - name: Generate Dataset
-        run: CL=1 extra/optimization/generate_dataset.sh
-      - name: Run Kernel Count Test
-        run: CL=1 python -m pytest -n=auto test/external/external_test_opt.py
-      - name: Run fused optimizer tests
-        run: CL=1 FUSE_OPTIM=1 python -m pytest -n=auto test/models/test_mnist.py test/test_optim.py -k "not muon"
+        run: PYTHONPATH="." extra/optimization/generate_dataset.sh
      - name: Upload artifact
        uses: actions/upload-artifact@v4
        with:
@@ -381,9 +451,11 @@ jobs:
          path: /tmp/sops.gz

  testopenpilot:
-    name: openpilot Compile Tests
+    name: 'openpilot Compile Tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -396,22 +468,26 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1397 ALLOWED_GATED_READ_IMAGE=94 FLOAT16=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
-      - name: Test openpilot CL compile fp16
-        run: FLOAT16=1 DEBUGCL=1 CL=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
-      - name: Test openpilot CL compile fp32 (test correctness)
-        run: DEBUGCL=1 CL=1 IMAGE=2 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
-      - name: Test openpilot LLVM compile fp16
-        run: FLOAT16=1 CPU=1 CPU_LLVM=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          PYTHONPATH="." ALLOWED_KERNEL_COUNT=208 ALLOWED_READ_IMAGE=2175 ALLOWED_GATED_READ_IMAGE=16 FLOAT16=0 GPU=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx
+      - name: Test openpilot alt model correctness (float32)
+        run: PYTHONPATH="." FLOAT16=0 DEBUGCL=1 GPU=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/3799fe46b3a629e491d4b8498b8ae83e4c88c304/selfdrive/modeld/models/supercombo.onnx
+      - name: Test openpilot fastvits model correctness (float32)
+        run: PYTHONPATH="." FLOAT16=0 DEBUGCL=1 GPU=1 IMAGE=2 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+      # - name: Test openpilot simple_plan vision model correctness (float32)
+      #   run: PYTHONPATH="." FLOAT16=0 DEBUGCL=1 GPU=1 IMAGE=2 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/35ff4f4577002f2685e50c8346addae33fe8da27a41dd4d6a0f14d1f4b1af81b
+      - name: Test openpilot LLVM compile
+        run: PYTHONPATH="." LLVM=1 LLVMOPT=1 JIT=2 BEAM=0 IMAGE=0 python examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/9118973ed03c1ae1d40cf69a29507ec2cc78efd7/selfdrive/modeld/models/supercombo.onnx
+      - name: Test openpilot compile4
+        run: PYTHONPATH="." NOLOCALS=1 GPU=1 IMAGE=2 FLOAT16=1 DEBUG=2 python3 examples/openpilot/compile4.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

-# ****** ONNX Tests ******
-
  testonnxcpu:
-    name: ONNX (CPU) Tests
+    name: 'ONNX (CPU) Tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 20
+    env:
+      IGNORE_OOB: 0

    steps:
      - name: Checkout Code
@@ -424,22 +500,25 @@ jobs:
          python-version: '3.11'
          llvm: 'true'
      - name: Test ONNX (CPU)
-        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+        run: CPU=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      - name: Test ONNX (LLVM)
-        run: CPU=1 CPU_LLVM=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+        run: LLVM=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      - name: Test ONNX Runner (CPU)
-        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_runner.py
+        run: CPU=1 python3 test/external/external_test_onnx_runner.py
      - name: Test Additional ONNX Ops (CPU)
-        run: CPU=1 CPU_LLVM=0 python3 test/external/external_test_onnx_ops.py
+        run: CPU=1 python3 test/external/external_test_onnx_ops.py
      - name: Test Quantize ONNX
-        run: CPU=1 CPU_LLVM=0 python3 test/test_quantize_onnx.py
+        run: CPU=1 python3 test/test_quantize_onnx.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

  testopencl:
-    name: ONNX (CL)+Optimization Tests
+    name: 'ONNX (GPU)+Optimization Tests'
    runs-on: ubuntu-22.04
    timeout-minutes: 20
+    env:
+      IGNORE_OOB: 0
+
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -448,25 +527,21 @@ jobs:
        with:
          key: onnxoptl
          deps: testing
-          pydeps: "tensorflow==2.19"
+          pydeps: "tensorflow==2.15.1 tensorflow_addons"
          python-version: '3.11'
          opencl: 'true'
-      - name: Test ONNX (CL)
-        run: CL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
+      - name: Test ONNX (GPU)
+        run: GPU=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
      #- name: Test Optimization Helpers
      #  run: DEBUG=1 python3 extra/optimization/test_helpers.py
      #- name: Test Action Space
-      #  run: DEBUG=1 CL=1 python3 extra/optimization/get_action_space.py
+      #  run: DEBUG=1 GPU=1 python3 extra/optimization/get_action_space.py
      - name: Test Beam Search
-        run: CL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
+        run: GPU=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
      - name: Test MLPerf stuff
-        run: CL=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
-      - name: NULL=1 beautiful_mnist_multigpu
-        run: NULL=1 python examples/beautiful_mnist_multigpu.py
-      - name: Test Bert training
-        run: NULL=1 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=24 GPUS=4 BERT_LAYERS=2 MODEL=bert python3 examples/mlperf/model_train.py
+        run: GPU=1 python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
      - name: Test llama 3 training
-        run: NULL=1 SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=1 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
+        run: MAX_BUFFER_SIZE=0 DEV=NULL SAMPLES=300 BS=8 SEQLEN=512 GRADIENT_ACC_STEPS=8 FAKEDATA=1 DEFAULT_FLOAT=bfloat16 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B MODEL=llama3 python3 examples/mlperf/model_train.py
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

@@ -484,12 +559,12 @@ jobs:
      - name: Test 1B LLM
        run: echo "What's a male chicken called? Answer with only one word." | MAX_BUFFER_SIZE=0 python3 -m tinygrad.apps.llm | grep -i rooster

-# ****** Models Tests ******
-
  testmodels:
    name: Models (llvm+cpu+gpu)
    runs-on: ubuntu-22.04
    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -501,38 +576,47 @@ jobs:
          opencl: 'true'
          llvm: 'true'
      - name: Test models (llvm)
-        run: CPU=1 CPU_LLVM=1 python -m pytest -n=auto test/models --durations=20
-      - name: Test models (opencl)
-        run: CL=1 python -m pytest -n=auto test/models --durations=20
+        run: LLVM=1 python -m pytest -n=auto test/models --durations=20
+      - name: Test models (gpu)
+        run: GPU=1 python -m pytest -n=auto test/models --durations=20
      - name: Test models (cpu)
-        run: CPU=1 CPU_LLVM=0 python -m pytest -n=auto test/models --durations=20
+        run: CPU=1 python -m pytest -n=auto test/models --durations=20
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

-  testmetalmodels:
-    name: Models (metal)
-    runs-on: macos-14
-    timeout-minutes: 20
+  testrangeify:
+    name: Linux (rangeify)
+    runs-on: ubuntu-24.04
+    timeout-minutes: 15
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: metal
-        deps: testing
-        python-version: '3.11'
-    - name: Test models (Metal)
-      run: METAL=1 python -m pytest -n=auto test/models --durations=20
-    - name: Test LLaMA compile speed
-      run: METAL=1 python test/external/external_test_speed_llama.py
-
-# ****** Feature Tests ******
+        key: rangeify-minimal-llvm
+        deps: testing_minimal
+        llvm: "true"
+    - name: Test CPU=1 RANGEIFY=1
+      # TODO: add more passing tests here
+      # test_symbolic_arange_sym_step is passing now
+      # test_threefry_doesnt_use_long is because there's a contig after the long now
+      run: |
+        CPU=1 RANGEIFY=1 python3 -m pytest -n auto --durations 20 \
+          -k "not test_symbolic_arange_sym_step and not test_threefry_doesnt_use_long" \
+          test/test_tiny.py test/test_rangeify.py test/test_ops.py test/test_tensor_variable.py \
+          test/test_outerworld_range.py test/test_sample.py test/test_randomness.py
+    - name: Test CPU=1 RANGEIFY=2
+      run: CPU=1 RANGEIFY=2 python3 -m pytest -n auto test/test_tiny.py test/test_rangeify.py test/test_ops.py --durations 20
+    - name: Test LLVM=1 RANGEIFY=1 (slow tests)
+      run: LLVM=1 RANGEIFY=1 python3 -m pytest -n auto test/models/test_mnist.py --durations 20

  testdevectorize:
    name: Linux (devectorize)
    runs-on: ubuntu-24.04
    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -544,16 +628,18 @@ jobs:
        pydeps: "pillow"
        llvm: "true"
    - name: Test LLVM=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
+      run: LLVM=1 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"
    - name: Test LLVM=1 DEVECTORIZE=0 for model
-      run: CPU=1 CPU_LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
+      run: LLVM=1 DEVECTORIZE=0 python3 test/models/test_efficientnet.py
    - name: Test CPU=1 DEVECTORIZE=0
-      run: CPU=1 CPU_LLVM=0 DEVECTORIZE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py
+      run: CPU=1 DEVECTORIZE=0 FUSE_ARANGE=0 python3 -m pytest -n auto test/test_tiny.py test/test_ops.py -k "not test_avg_pool3d_failure"

  testdsp:
    name: Linux (DSP)
    runs-on: ubuntu-24.04
    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
@@ -620,6 +706,7 @@ jobs:
    runs-on: ubuntu-22.04
    timeout-minutes: 20
    env:
+      IGNORE_OOB: 0
      AMD: 1
      MOCKGPU: 1
      FORWARD_ONLY: 1
@@ -649,8 +736,7 @@ jobs:
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
      - name: Run TestOps.test_add with SQTT
        run: |
-          VIZ=1 PMC=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
-          VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
+          PROFILE=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run process replay tests
        uses: ./.github/actions/process-replay
@@ -678,7 +764,7 @@ jobs:
          cuda: 'true'
          ocelot: 'true'
      - name: Set env
-        run: printf "${{ matrix.backend == 'PTX' && 'CUDA=1\nCUDA_PTX=1' || matrix.backend == 'nv' && 'NV=1\nSKIP_SLOW_TEST=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'PTX' && 'CUDA=1\nPTX=1' || matrix.backend == 'nv' && 'NV=1\nSKIP_SLOW_TEST=1' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CUDA','NV'], Device.DEFAULT"
@@ -689,15 +775,18 @@ jobs:
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

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

    name: Linux (${{ matrix.backend }})
    runs-on: ubuntu-22.04
    timeout-minutes: 20
+    env:
+      IGNORE_OOB: 0
+
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -706,16 +795,15 @@ jobs:
        with:
          key: ${{ matrix.backend }}-minimal
          deps: testing_minimal
-          opencl: ${{ matrix.backend == 'opencl' && 'true' }}
-          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
+          opencl: ${{ matrix.backend == 'gpu' && 'true' }}
+          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'CL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'LLVM=1' || matrix.backend == 'cpu' && 'CPU=1' || matrix.backend == 'gpu' && 'GPU=1' }}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
-          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
+          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['LLVM','CPU','GPU'], Device.DEFAULT"
          DEBUG=5 FORWARD_ONLY=1 python3 test/test_ops.py TestOps.test_add
-      - name: Run pytest (${{ matrix.backend }})
+      - name: Run pytest (not cuda)
        run: python -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
      - name: Run TRANSCENDENTAL math
        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
@@ -724,42 +812,48 @@ jobs:

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

-  testmetal:
+  testmetal2:
    name: MacOS (unit)
    runs-on: macos-14
    timeout-minutes: 20
+    env:
+      IGNORE_OOB: 0
+
    steps:
    - name: Checkout Code
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: metal
+        key: metal2
        deps: testing
        python-version: '3.11'
        amd: 'true'
        cuda: 'true'
        ocelot: 'true'
        llvm: 'true'
-    - name: Run unit tests
-      run: METAL=1 python -m pytest -n=auto test/unit/ --durations=20
+    - name: Run real world test
+      run: METAL=1 python -m pytest -n=auto test/models/test_real_world.py --durations=20
+    - name: Test models (Metal)
+      run: METAL=1 python -m pytest -n=auto test/models -v --durations=20
    - name: Run ONNX
      run: METAL=1 python -m pytest -n=auto test/external/external_test_onnx_backend.py --durations=20
    - name: Test tensor core ops (fake)
-      run: METAL=1 DEBUG=3 TC=2 python test/test_ops.py TestOps.test_gemm
+      run: TC=2 METAL=1 DEBUG=3 python test/test_ops.py TestOps.test_gemm
    - name: Test tensor core ops (real)
      run: METAL=1 DEBUG=3 python test/test_ops.py TestOps.test_big_gemm
+    - name: Test LLaMA compile speed
+      run: METAL=1 python test/external/external_test_speed_llama.py
    - name: Test Beam Search
      run: METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
    #- name: Fuzz Test linearizer
    #  run: METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
    - name: Run TRANSCENDENTAL math
-      run: METAL=1 TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
+      run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
    - name: Run pytest (amd)
      env:
        MOCKGPU: 1
        AMD: 1
-        AMD_LLVM: 0
        FORWARD_ONLY: 1
      run: |
        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
@@ -767,14 +861,13 @@ jobs:
      env:
        MOCKGPU: 1
        AMD: 1
-        AMD_LLVM: 1
        FORWARD_ONLY: 1
      run: |
        python -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py test/device/test_amd_llvm.py --durations=20
    - name: Run pytest (ptx)
      env:
        MOCKGPU: 1
-        NV_PTX: 1
+        PTX: 1
        NV: 1
        FORWARD_ONLY: 1
      run: |
@@ -819,14 +912,105 @@ jobs:
    - name: Test ONNX Runner (WEBGPU)
      run: WEBGPU=1 python3 test/external/external_test_onnx_runner.py

+  osxremote:
+   name: MacOS (remote metal)
+   runs-on: macos-15
+   timeout-minutes: 10
+   env:
+     REMOTE: 1
+     REMOTEDEV: METAL
+   steps:
+     - name: Checkout Code
+       uses: actions/checkout@v4
+     - name: Setup Environment
+       uses: ./.github/actions/setup-tinygrad
+       with:
+         key: macos-remote
+         deps: testing_minimal
+     - name: Check Device.DEFAULT and print some source
+       run: |
+         python -c "from tinygrad import Device; assert Device.DEFAULT == 'REMOTE', Device.DEFAULT"
+         python -c "from tinygrad import Device; assert Device.default.properties.real_device == 'METAL', Device.default.properties.real_device"
+         DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
+     - name: Run REMOTE=1 Test
+       run: |
+         python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_tensor_variable.py
+
+  amdremote:
+    name: Linux (remote)
+    runs-on: ubuntu-22.04
+    timeout-minutes: 20
+    env:
+      REMOTE: 1
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v4
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: linux-remote
+          deps: testing_minimal
+          amd: 'true'
+          llvm: 'true'
+          opencl: 'true'
+      - name: Start remote server
+        run: |
+          start_server() {
+            systemd-run --user \
+              --unit="$1" \
+              --setenv=REMOTEDEV="$2" \
+              --setenv=MOCKGPU=1 \
+              --setenv=PYTHONPATH=. \
+              --setenv=PORT="$3" \
+              --working-directory="$(pwd)" \
+              python tinygrad/runtime/ops_remote.py
+          }
+
+          start_server "remote-server-amd-1" "AMD" 6667
+          start_server "remote-server-amd-2" "AMD" 6668
+          start_server "remote-server-gpu" "GPU" 7667
+          start_server "remote-server-cpu" "CPU" 8667
+      - name: Check Device.DEFAULT and print some source
+        env:
+          HOST: 127.0.0.1:6667*6,127.0.0.1:6668*6
+        run: |
+          python -c "from tinygrad import Device; assert Device.DEFAULT == 'REMOTE', Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.default.properties.real_device == 'AMD', Device.default.properties.real_device"
+          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
+      - name: Run REMOTE=1 Test (AMD)
+        env:
+          HOST: 127.0.0.1:6667*6,127.0.0.1:6668*6
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_subbuffer.py test/test_graph.py test/test_multitensor.py test/test_remote.py test/test_tensor_variable.py --durations 20
+      - name: Run REMOTE=1 Test (GPU)
+        env:
+          HOST: 127.0.0.1:7667*6
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_image_dtype.py test/test_jit.py --durations 20
+          IMAGE=2 python3 -m pytest test/test_tiny.py test/test_image_dtype.py
+      - name: Run REMOTE=1 Test (CPU)
+        env:
+          HOST: 127.0.0.1:8667*6
+        run: |
+          python3 -m pytest test/test_tiny.py test/test_jit.py test/test_multitensor.py --durations 20
+      - name: Show remote server logs
+        if: always()
+        run: |
+          journalctl --user -u remote-server-amd-1 --no-pager
+          journalctl --user -u remote-server-amd-2 --no-pager
+          journalctl --user -u remote-server-gpu --no-pager
+          journalctl --user -u remote-server-cpu --no-pager
+
  osxtests:
    strategy:
      fail-fast: false
      matrix:
-        backend: [metal, llvm, cpu, lvp]
+        backend: [metal, llvm, cpu]
    name: MacOS (${{ matrix.backend }})
    runs-on: macos-15
    timeout-minutes: 20
+    env:
+      IGNORE_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -836,13 +1020,12 @@ jobs:
          key: macos-${{ matrix.backend }}-minimal
          deps: testing_minimal
          pydeps: "capstone"
-          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'true' }}
+          llvm: ${{ matrix.backend == 'llvm' && 'true' }}
      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'metal' && 'METAL=1' || matrix.backend == 'lvp' && 'CPU=1\nCPU_LVP=1' }}" >> $GITHUB_ENV
+        run: printf "${{ matrix.backend == 'llvm' && 'LLVM=1' || matrix.backend == 'cpu' && 'CPU=1' || matrix.backend == 'metal' && 'METAL=1'}}" >> $GITHUB_ENV
      - name: Check Device.DEFAULT and print some source
        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.DEFAULT == '${{ matrix.backend }}'.upper(), Device.DEFAULT"
          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
      - name: Run pytest (${{ matrix.backend }})
        run: python3 -m pytest -n=auto test/ --ignore=test/models --ignore=test/unit --durations=20
@@ -863,6 +1046,8 @@ jobs:
    name: Windows (${{ matrix.backend }})
    runs-on: windows-latest
    timeout-minutes: 15
+    env:
+      IGNORE_OOB: 0
    steps:
      - name: Checkout Code
        uses: actions/checkout@v4
@@ -874,7 +1059,7 @@ jobs:
          pydeps: ${{ matrix.backend == 'webgpu' && 'dawn-python' || '' }}
      - name: Set env
        shell: bash
-        run:  printf "${{ matrix.backend == 'llvm' && 'CPU=1\nCPU_LLVM=1' || matrix.backend == 'cpu' && 'CPU=1\nCPU_LLVM=0\nCPU_COUNT=2' || matrix.backend == 'webgpu' && 'WEBGPU=1'}}" >> $GITHUB_ENV
+        run:  printf "${{ matrix.backend == 'llvm' && 'LLVM=1' || matrix.backend == 'cpu' && 'CPU=1' || matrix.backend == 'webgpu' && 'WEBGPU=1'}}" >> $GITHUB_ENV
      - name: Run unit tests
        if: matrix.backend=='llvm'
        # test_newton_schulz hits RecursionError
@@ -882,35 +1067,5 @@ jobs:
      - name: Run pytest (${{ matrix.backend }})
        shell: bash
        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
+          python -c "from tinygrad import Device; assert Device.DEFAULT == '${{ matrix.backend }}'.upper(), Device.DEFAULT"
          python -m pytest -n=auto test/test_tiny.py test/test_ops.py --durations=20
-
-# ****** Compile-only Tests ******
-
-  compiletests:
-    strategy:
-      fail-fast: false
-      matrix:
-        backend: [ir3, nak]
-    name: Compile-only (${{ matrix.backend }})
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-${{ matrix.backend }}
-          deps: testing_minimal
-          mesa: ${{ (matrix.backend == 'ir3' || matrix.backend == 'nak') && 'true' }}
-          python-version: '3.14'
-      - name: Set env
-        shell: bash
-        run: printf "NULL=1\n${{ matrix.backend == 'ir3' && 'NULL_IR3=1' || matrix.backend == 'nak' && 'NULL_NAK=1' }}" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/test_ops.py --durations=20
@@ -38,7 +38,6 @@ extra/huggingface_onnx/models/*
 extra/huggingface_onnx/*.yaml
 extra/weights
 venv
-venv_sd_mlperf
 examples/**/net.*[js,json]
 examples/**/*.safetensors
 node_modules
@@ -63,5 +62,3 @@ profile_stats
 *.log
 target
 .mypy_cache
-mutants
-.mutmut-cache
@@ -20,15 +20,21 @@ repos:
        language: system
        always_run: true
        pass_filenames: false
+      - id: tests
+        name: subset of tests
+        entry: env PYTHONPATH="." python3 -m pytest -n=4 test/test_ops.py test/test_dtype.py test/test_schedule.py test/test_assign.py
+        language: system
+        always_run: true
+        pass_filenames: false
      - id: example
-        name: test all devices
+        name: multi device tests
        entry: python3 test/external/external_test_example.py
        language: system
        always_run: true
        pass_filenames: false
-      - id: tests
-        name: comprehensive test suite
-        entry: env OMP_NUM_THREADS=1 SKIP_SLOW_TEST=1 PYTHONPATH="." python3 -m pytest -n=6 test/test_ops.py test/test_schedule.py test/test_assign.py test/test_tensor.py test/test_jit.py test/unit/test_schedule_cache.py test/unit/test_pattern_matcher.py test/unit/test_uop_symbolic.py test/unit/test_helpers.py
+      - id: pylint
+        name: pylint
+        entry: python3 -m pylint tinygrad/
        language: system
        always_run: true
-        pass_filenames: false
+        pass_filenames: false
@@ -30,6 +30,10 @@ persistent=yes
 # Specify a configuration file.
 #rcfile=

+# When enabled, pylint would attempt to guess common misconfiguration and emit
+# user-friendly hints instead of false-positive error messages
+suggestion-mode=yes
+
 # Allow loading of arbitrary C extensions. Extensions are imported into the
 # active Python interpreter and may run arbitrary code.
 unsafe-load-any-extension=no
@@ -1,227 +0,0 @@
-# Claude Code Guide for tinygrad
-
-## Architecture Overview
-
-tinygrad compiles tensor operations into optimized kernels. The pipeline:
-
-1. **Tensor** (`tensor.py`) - User-facing API, creates UOp graph
-2. **UOp** (`uop/ops.py`) - Unified IR for all operations (both tensor and kernel level)
-3. **Schedule** (`engine/schedule.py`, `schedule/`) - Converts tensor UOps to kernel UOps
-4. **Codegen** (`codegen/`) - Converts kernel UOps to device code
-5. **Runtime** (`runtime/`) - Device-specific execution
-
-## Key Concepts
-
-### UOp (Universal Operation)
-Everything is a UOp - tensors, operations, buffers, kernels. Key properties:
- `op`: The operation type (Ops enum)
- `dtype`: Data type
- `src`: Tuple of source UOps
- `arg`: Operation-specific argument
- `tag`: Optional tag for graph transformations
-
-UOps are **immutable and cached** - creating the same UOp twice returns the same object (ucache).
-
-### PatternMatcher
-Used extensively for graph transformations:
-```python
-pm = PatternMatcher([
-  (UPat(Ops.ADD, src=(UPat.cvar("x"), UPat.cvar("x"))), lambda x: x * 2),
-])
-result = graph_rewrite(uop, pm)
-```
-
-### Schedule Cache
-Schedules are cached by graph structure. BIND nodes (variables with bound values) are unbound before cache key computation so different values hit the same cache.
-
-## Directory Structure
-
-```
-tinygrad/
-├── tensor.py          # Tensor class, user API
-├── device.py          # Buffer, device management
-├── dtype.py           # Data types
-├── helpers.py         # Utilities, environment vars
-├── uop/
-│   ├── ops.py         # UOp class, Ops enum, PatternMatcher
-│   ├── spec.py        # UOp type verification
-│   └── symbolic.py    # Symbolic math simplification
-├── engine/
-│   ├── schedule.py    # Schedule creation, caching
-│   ├── realize.py     # Tensor realization
-│   ├── jit.py         # JIT compilation
-│   └── memory.py      # Memory planning
-├── schedule/
-│   ├── rangeify.py    # Convert movements to ranges
-│   └── indexing.py    # Index calculations
-├── codegen/
-│   ├── kernel.py      # Kernel optimization
-│   └── uopgraph.py    # UOp graph transformations
-├── renderer/          # Code generation (CUDA, Metal, etc.)
-└── runtime/           # Device backends
-```
-
-## Testing
-
-```bash
-# Run specific test
-python -m pytest test/unit/test_schedule_cache.py -xvs
-
-# Run with timeout
-python -m pytest test/test_symbolic_ops.py -x --timeout=60
-
-# Debug with print
-DEBUG=2 python -m pytest test/test_schedule.py::test_name -xvs
-
-# Visualize UOp graphs
-VIZ=1 python -c "from tinygrad import Tensor; Tensor.ones(10).sum().realize()"
-```
-
-## Common Environment Variables
-
- `DEBUG=1-4` - Increasing verbosity
- `VIZ=1` - Enable graph visualization
- `SPEC=1` - Enable UOp spec verification
- `NOOPT=1` - Disable optimizations
- `DEVICE=CPU/CUDA/AMD/METAL` - Set default device
-
-## Debugging Tips
-
-1. **Print UOp graphs**: `print(tensor.uop)` or `print(tensor.uop.sink())`
-2. **Check schedule**: `tensor.schedule()` returns list of ExecItems
-3. **Trace graph rewrites**: Use `VIZ=1` or add print in PatternMatcher callbacks
-4. **Find UOps by type**: `[u for u in uop.toposort() if u.op is Ops.SOMETHING]`
-
-## Workflow Rules
-
- **NEVER commit without explicit user approval** - always show the diff and wait for approval
- **NEVER amend commits** - always create a new commit instead
- Run `pre-commit run --all-files` before committing to catch linting/type errors
- Run tests before proposing commits
- Test with `SPEC=2` when modifying UOp-related code
-
-## Style Notes
-
- 2-space indentation, 150 char line limit
- PatternMatchers should be defined at module level (slow to construct)
- Prefer `graph_rewrite` over manual graph traversal
- UOp methods like `.replace()` preserve tags unless explicitly changed
- Use `.rtag(value)` to add tags to UOps
-
-## Lessons Learned
-
-### UOp ucache Behavior
-UOps are cached by their contents - creating a UOp with identical (op, dtype, src, arg) returns the **same object**. This means:
- `uop.replace(tag=None)` on a tagged UOp returns the original untagged UOp if it exists in cache
- Two UOps with same structure are identical (`is` comparison works)
-
-### Spec Validation
-When adding new UOp patterns, update `tinygrad/uop/spec.py`. Test with:
-```bash
-SPEC=2 python3 test/unit/test_something.py
-```
-Spec issues appear as `RuntimeError: SPEC ISSUE None: UOp(...)`.
-
-### Schedule Cache Key Normalization
-The schedule cache strips values from BIND nodes so different bound values (e.g., KV cache positions) hit the same cache entry:
- `pm_pre_sched_cache`: BIND(DEFINE_VAR, CONST) → BIND(DEFINE_VAR) for cache key
- `pm_post_sched_cache`: restores original BIND from context
- When accessing `bind.src[1]`, check `len(bind.src) > 1` first (might be stripped)
- Extract var_vals from `input_buffers` dict after graph_rewrite (avoids extra toposort)
-
-### Avoiding Extra Work
- Use ctx dict from graph_rewrite to collect info during traversal instead of separate toposort
- Only extract var_vals when schedule is non-empty (no kernels = no vars needed)
- PatternMatchers are slow to construct - define at module level, not in functions
-
-### Readability Over Speed
-Don't add complexity for marginal performance gains. Simpler code that's slightly slower is often better:
-```python
-# BAD: "optimized" with extra complexity
-if has_afters:  # skip toposort if no AFTERs
-  after_map = [(u, u.buf_uop) for u in big_sink.toposort() if u.op is Ops.AFTER]
-
-# GOOD: simple, always works
-after_map = [(u, u.buf_uop) for u in big_sink.toposort() if u.op is Ops.AFTER]
-```
-The conditional check adds complexity, potential bugs, and often negligible speedup. Only optimize when profiling shows a real bottleneck.
-
-### Testing LLM Changes
-```bash
-# Quick smoke test
-echo "Hello" | DEBUG=1 python tinygrad/apps/llm.py --model "llama3.2:1b"
-
-# Check cache hits (should see "cache hit" after warmup)
-echo "Hello world" | DEBUG=1 python tinygrad/apps/llm.py --model "llama3.2:1b" 2>&1 | grep cache
-
-# Test with beam search
-echo "Hello" | BEAM=2 python tinygrad/apps/llm.py --model "llama3.2:1b"
-```
-
-## Common Patterns
-
-### Graph Transformation
-```python
-def my_transform(ctx, x):
-  # Return new UOp or None to skip
-  return x.replace(arg=new_arg)
-
-pm = PatternMatcher([
-  (UPat(Ops.SOMETHING, name="x"), my_transform),
-])
-result = graph_rewrite(input_uop, pm, ctx={})
-```
-
-### Finding Variables
-```python
-# Get all variables in a UOp graph
-variables = uop.variables()
-
-# Get bound variable values
-var, val = bind_uop.unbind()
-```
-
-### Shape Handling
-```python
-# Shapes can be symbolic (contain UOps)
-shape = tensor.shape  # tuple[sint, ...] where sint = int | UOp
-```
-
-## Performance Optimization
-
-When optimizing tinygrad internals:
-
-1. **Measure wall time, not just call counts** - Reducing `graph_rewrite` calls doesn't always improve wall time. The overhead of conditional checks can exceed the cost of the operation being skipped.
-
-2. **Profile each optimization individually** - Run benchmarks with and without each change to measure actual impact. Use `test/external/external_benchmark_schedule.py` for schedule/rewrite timing.
-
-3. **Early exits in hot paths are effective** - Simple checks like `if self.op is Ops.CONST: return self` in `simplify()` can eliminate many unnecessary `graph_rewrite` calls.
-
-4. **`graph_rewrite` is expensive** - Each call has overhead even for small graphs. Avoid calling it when the result is trivially known (e.g., simplifying a CONST returns itself).
-
-5. **Beware iterator overhead** - Checks like `all(x.op is Ops.CONST for x in self.src)` can be slower than just running the operation, especially for small sequences.
-
-6. **Verify cache hit rates before adding/keeping caches** - Measure actual hit rates with real workloads. A cache with 0% hit rate is pure overhead (e.g., `pm_cache` was removed because the algorithm guarantees each UOp is only passed to `pm_rewrite` once).
-
-7. **Use `TRACK_MATCH_STATS=2` to profile pattern matching** - This shows match rates and time per pattern. Look for patterns with 0% match rate that still cost significant time - these are pure overhead for that workload.
-
-8. **Cached properties beat manual traversal** - `backward_slice` uses `@functools.cached_property`. A DFS with early-exit sounds faster but is actually slower because it doesn't benefit from caching. The cache hit benefit often outweighs algorithmic improvements.
-
-9. **Avoid creating intermediate objects in hot paths** - For example, `any(x.op in ops for x in self.backward_slice)` is faster than `any(x.op in ops for x in {self:None, **self.backward_slice})` because it avoids dict creation.
-
-## Pattern Matching Profiling
-
-Use `TRACK_MATCH_STATS=2` to identify expensive patterns:
-
-```bash
-TRACK_MATCH_STATS=2 PYTHONPATH="." python3 test/external/external_benchmark_schedule.py
-```
-
-Output format: `matches / attempts -- match_time / total_time ms -- location`
-
-Key patterns to watch (from ResNet50 benchmark):
- `split_load_store`: ~146ms, 31% match rate - does real work
- `simplify_valid`: ~75ms, 0% match rate in this workload - checks AND ops for INDEX in backward slice
- `vmin==vmax folding`: ~55ms, 0.33% match rate - checks 52K ops but rarely matches
-
-Patterns with 0% match rate are workload-specific overhead. They may be useful in other workloads, so don't remove them without understanding their purpose.
@@ -21,38 +21,17 @@ tinygrad: For something between [PyTorch](https://github.com/pytorch/pytorch) an

 ---

-tinygrad is an end-to-end deep learning stack:
+Despite tinygrad's size, it is a fully featured deep learning framework.

- **Tensor library** with autograd
- **IR and compiler** that fuse and lower kernels
- **JIT + graph execution**
- **nn / optim / datasets** for real training
+Due to its extreme simplicity, it is the easiest framework to add new accelerators to, with support for both inference and training. If XLA is CISC, tinygrad is RISC.

-It’s inspired by PyTorch (ergonomics), JAX (functional transforms and IR-based AD), and TVM (scheduling and codegen), but stays intentionally tiny and hackable.
+tinygrad is now beta software, we [raised some money](https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html) to make it good. Someday, we will tape out chips.

---
+## Features

-## How tinygrad compares
+### LLaMA and Stable Diffusion

-**PyTorch**
-
- ✅ Similar: eager `Tensor` API, autograd, `optim`, basic datasets and layers.
- ✅ You can write familiar training loops.
- 🔁 Unlike PyTorch, the entire compiler and IR are visible and hackable.
-
-**JAX**
-
- ✅ IR-based autodiff over primitives (like JAXPR + XLA).
- ✅ Function-level JIT (`TinyJit`) that captures and replays kernels.
- 🔁 Fewer functional transforms (no full `vmap`/`pmap` yet), but far easier to read.
-
-**TVM**
-
- ✅ Multiple lowering passes, scheduling, and BEAM search over kernels.
- ✅ Device “graphs” for batched execution.
- 🔁 tinygrad also ships the **front-end framework** (tensors, nn, optim), not just the compiler.
-
---
+tinygrad can run [LLaMA](/docs/showcase.md#llama) and [Stable Diffusion](/docs/showcase.md#stable-diffusion)!

 ### Laziness

@@ -100,8 +79,9 @@ See [examples/beautiful_mnist.py](examples/beautiful_mnist.py) for the full vers

 tinygrad already supports numerous accelerators, including:

- [x] [OpenCL](tinygrad/runtime/ops_cl.py)
- [x] [CPU](tinygrad/runtime/ops_cpu.py)
+- [x] [GPU (OpenCL)](tinygrad/runtime/ops_gpu.py)
+- [x] [CPU (C Code)](tinygrad/runtime/ops_cpu.py)
+- [x] [LLVM](tinygrad/runtime/ops_llvm.py)
 - [x] [METAL](tinygrad/runtime/ops_metal.py)
 - [x] [CUDA](tinygrad/runtime/ops_cuda.py)
 - [x] [AMD](tinygrad/runtime/ops_amd.py)
@@ -0,0 +1,470 @@
+#!/bin/bash -e
+
+# setup instructions for clang2py
+if [[ ! $(clang2py -V) ]]; then
+  pushd .
+  cd /tmp
+  sudo apt-get install -y --no-install-recommends clang
+  pip install --upgrade pip setuptools
+  pip install clang==14.0.6
+  git clone https://github.com/nimlgen/ctypeslib.git
+  cd ctypeslib
+  pip install .
+  clang2py -V
+  popd
+fi
+
+BASE=tinygrad/runtime/autogen/
+
+fixup() {
+  sed -i '1s/^/# mypy: ignore-errors\n/' $1
+  sed -i 's/ *$//' $1
+  grep FIXME_STUB $1 || true
+}
+
+patch_dlopen() {
+  path=$1; shift
+  name=$1; shift
+  cat <<EOF | sed -i "/import ctypes.*/r /dev/stdin" $path
+PATHS_TO_TRY = [
+$(for p in "$@"; do echo "  $p,"; done)
+]
+def _try_dlopen_$name():
+  library = ctypes.util.find_library("$name")
+  if library: return ctypes.CDLL(library)
+  for candidate in PATHS_TO_TRY:
+    try: return ctypes.CDLL(candidate)
+    except OSError: pass
+  return None
+EOF
+}
+
+generate_opencl() {
+  clang2py /usr/include/CL/cl.h -o $BASE/opencl.py -l /usr/lib/x86_64-linux-gnu/libOpenCL.so.1 -k cdefstum
+  fixup $BASE/opencl.py
+  # hot patches
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/opencl.py
+  sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libOpenCL.so.1')\ctypes.CDLL(ctypes.util.find_library('OpenCL'))\g" $BASE/opencl.py
+  python3 -c "import tinygrad.runtime.autogen.opencl"
+}
+
+generate_hip() {
+  clang2py /opt/rocm/include/hip/hip_ext.h /opt/rocm/include/hip/hiprtc.h \
+  /opt/rocm/include/hip/hip_runtime_api.h /opt/rocm/include/hip/driver_types.h \
+  --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/hip.py -l /opt/rocm/lib/libamdhip64.so
+  echo "hipDeviceProp_t = hipDeviceProp_tR0600" >> $BASE/hip.py
+  echo "hipGetDeviceProperties = hipGetDevicePropertiesR0600" >> $BASE/hip.py
+  fixup $BASE/hip.py
+  # we can trust HIP is always at /opt/rocm/lib
+  #sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/hip.py
+  #sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhiprtc.so')\ctypes.CDLL(ctypes.util.find_library('hiprtc'))\g" $BASE/hip.py
+  #sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamdhip64.so')\ctypes.CDLL(ctypes.util.find_library('amdhip64'))\g" $BASE/hip.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/hip.py
+  sed -i "s\'/opt/rocm/\os.getenv('ROCM_PATH', '/opt/rocm/')+'/\g" $BASE/hip.py
+  python3 -c "import tinygrad.runtime.autogen.hip"
+}
+
+generate_comgr() {
+  clang2py /opt/rocm/include/amd_comgr/amd_comgr.h \
+  --clang-args="-D__HIP_PLATFORM_AMD__ -I/opt/rocm/include -x c++" -o $BASE/comgr.py -l /opt/rocm/lib/libamd_comgr.so
+  fixup $BASE/comgr.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/comgr.py
+  patch_dlopen $BASE/comgr.py amd_comgr "'/opt/rocm/lib/libamd_comgr.so'" "os.getenv('ROCM_PATH', '')+'/lib/libamd_comgr.so'" "'/usr/local/lib/libamd_comgr.dylib'" "'/opt/homebrew/lib/libamd_comgr.dylib'"
+  sed -i "s\ctypes.CDLL('/opt/rocm/lib/libamd_comgr.so')\_try_dlopen_amd_comgr()\g" $BASE/comgr.py
+  python3 -c "import tinygrad.runtime.autogen.comgr"
+}
+
+generate_kfd() {
+  clang2py /usr/include/linux/kfd_ioctl.h -o $BASE/kfd.py -k cdefstum
+
+  fixup $BASE/kfd.py
+  sed -i "s/import ctypes/import ctypes, os/g" $BASE/kfd.py
+  sed -i "s/import fcntl, functools/import functools/g" $BASE/kfd.py
+  sed -i "/import functools/a from tinygrad.runtime.support.hcq import FileIOInterface" $BASE/kfd.py
+  sed -i "s/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd, \*\*kwargs):/def _do_ioctl(__idir, __base, __nr, __user_struct, __fd:FileIOInterface, \*\*kwargs):/g" $BASE/kfd.py
+  sed -i "s/fcntl.ioctl(__fd, (__idir<<30)/__fd.ioctl((__idir<<30)/g" $BASE/kfd.py
+  sed -i "s/!!/not not /g" $BASE/kfd.py
+  python3 -c "import tinygrad.runtime.autogen.kfd"
+}
+
+generate_cuda() {
+  clang2py /usr/include/cuda.h --clang-args="-D__CUDA_API_VERSION_INTERNAL" -o $BASE/cuda.py -l /usr/lib/x86_64-linux-gnu/libcuda.so
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/cuda.py
+  sed -i "s\ctypes.CDLL('/usr/lib/x86_64-linux-gnu/libcuda.so')\ctypes.CDLL(ctypes.util.find_library('cuda'))\g" $BASE/cuda.py
+  fixup $BASE/cuda.py
+  python3 -c "import tinygrad.runtime.autogen.cuda"
+}
+
+generate_nvrtc() {
+  clang2py /usr/local/cuda/include/nvrtc.h /usr/local/cuda/include/nvJitLink.h -o $BASE/nvrtc.py -l /usr/local/cuda/lib64/libnvrtc.so -l /usr/local/cuda/lib64/libnvJitLink.so
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" $BASE/nvrtc.py
+  sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvrtc.so')\ctypes.CDLL(ctypes.util.find_library('nvrtc'))\g" $BASE/nvrtc.py
+  sed -i "s\ctypes.CDLL('/usr/local/cuda/lib64/libnvJitLink.so')\ctypes.CDLL(ctypes.util.find_library('nvJitLink'))\g" $BASE/nvrtc.py
+  fixup $BASE/nvrtc.py
+  python3 -c "import tinygrad.runtime.autogen.nvrtc"
+}
+
+generate_nv() {
+  NVKERN_COMMIT_HASH=81fe4fb417c8ac3b9bdcc1d56827d116743892a5
+  NVKERN_SRC=/tmp/open-gpu-kernel-modules-$NVKERN_COMMIT_HASH
+  if [ ! -d "$NVKERN_SRC" ]; then
+    git clone https://github.com/NVIDIA/open-gpu-kernel-modules $NVKERN_SRC
+    pushd .
+    cd $NVKERN_SRC
+    git reset --hard $NVKERN_COMMIT_HASH
+    popd
+  fi
+
+  clang2py -k cdefstum \
+    extra/nv_gpu_driver/clc6c0qmd.h \
+    extra/nv_gpu_driver/clcec0qmd.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0000.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl0080.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl2080_notification.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc56f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc86f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc96f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc761.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/cl83de.h \
+    $NVKERN_SRC/src/nvidia/generated/g_allclasses.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clc6c0.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/class/clcdc0.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/clc6b5.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/clc9b5.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_ioctl.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/uvm_linux_ioctl.h \
+    $NVKERN_SRC/kernel-open/nvidia-uvm/hwref/ampere/ga100/dev_fault.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv_escape.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numbers.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-ioctl-numa.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include/nv-unix-nvos-params-wrappers.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/alloc/alloc_channel.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/nvos.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0000/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl0080/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl2080/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl83de/*.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlc36f.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrlcb33.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrla06c.h \
+    $NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl/ctrl90f1.h \
+    --clang-args="-include $NVKERN_SRC/src/common/sdk/nvidia/inc/nvtypes.h -I$NVKERN_SRC/src/common/inc -I$NVKERN_SRC/kernel-open/nvidia-uvm -I$NVKERN_SRC/kernel-open/common/inc -I$NVKERN_SRC/src/common/sdk/nvidia/inc -I$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include -I$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl" \
+    -o $BASE/nv_gpu.py
+  fixup $BASE/nv_gpu.py
+  sed -i "s\(0000000001)\1\g" $BASE/nv_gpu.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/nv_gpu.py
+  sed -i 's/#\?\s\([A-Za-z0-9_]\+\) = MW ( \([0-9]\+\) : \([0-9]\+\) )/\1 = (\2 , \3)/' $BASE/nv_gpu.py # NVC6C0_QMDV03_00 processing
+  sed -i 's/#\sdef NVC6C0_QMD\([A-Za-z0-9_()]\+\):/def NVC6C0_QMD\1:/' $BASE/nv_gpu.py
+  sed -i 's/#\sdef NVCEC0_QMD\([A-Za-z0-9_()]\+\):/def NVCEC0_QMD\1:/' $BASE/nv_gpu.py
+  sed -E -i -n '/^def (NVCEC0_QMDV05_00_RELEASE)(_ENABLE)\(i\):/{p;s//\1'"0"'\2=\1\2(0)\n\1'"1"'\2=\1\2(1)/;H;b};p;${x;s/^\n//;p}' "$BASE/nv_gpu.py"
+  sed -i 's/#\s*return MW(\([0-9i()*+]\+\):\([0-9i()*+]\+\))/    return (\1 , \2)/' $BASE/nv_gpu.py
+  sed -i 's/#\?\s*\(.*\)\s*=\s*\(NV\)\?BIT\(32\)\?\s*(\s*\([0-9]\+\)\s*)/\1 = (1 << \4)/' $BASE/nv_gpu.py # name = BIT(x) -> name = (1 << x)
+  sed -i "s/UVM_\([A-Za-z0-9_]\+\) = \['i', '(', '\([0-9]\+\)', ')'\]/UVM_\1 = \2/" $BASE/nv_gpu.py # UVM_name = ['i', '(', '<num>', ')'] -> UVM_name = <num>
+
+  # Parse status codes
+  sed -n '1i\
+nv_status_codes = {}
+/^NV_STATUS_CODE/ { s/^NV_STATUS_CODE(\([^,]*\), *\([^,]*\), *"\([^"]*\)") *.*$/\1 = \2\nnv_status_codes[\1] = "\3"/; p }' $NVKERN_SRC/src/common/sdk/nvidia/inc/nvstatuscodes.h >> $BASE/nv_gpu.py
+  python3 -c "import tinygrad.runtime.autogen.nv_gpu"
+
+  clang2py -k cdefstum \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/fsp/kern_fsp_cot_payload.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gspifpub.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gsp_fw_wpr_meta.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/gsp/gsp_fw_sr_meta.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/gsp/gsp_init_args.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/gpu/gsp/gsp_init_args.h \
+    $NVKERN_SRC/src/common/uproc/os/common/include/libos_init_args.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/rmRiscvUcode.h \
+    $NVKERN_SRC/src/common/shared/msgq/inc/msgq/msgq_priv.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/vgpu/rpc_headers.h \
+    $NVKERN_SRC/src/nvidia/inc/kernel/vgpu/rpc_global_enums.h \
+    $NVKERN_SRC/src/nvidia/generated/g_rpc-structures.h \
+    $NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc/fsp/fsp_nvdm_format.h \
+    extra/nv_gpu_driver/g_rpc-message-header.h \
+    extra/nv_gpu_driver/gsp_static_config.h \
+    extra/nv_gpu_driver/vbios.h \
+    --clang-args="-DRPC_MESSAGE_STRUCTURES -DRPC_STRUCTURES -include $NVKERN_SRC/src/common/sdk/nvidia/inc/nvtypes.h -I$NVKERN_SRC/src/nvidia/generated -I$NVKERN_SRC/src/common/inc -I$NVKERN_SRC/src/nvidia/inc -I$NVKERN_SRC/src/nvidia/interface/ -I$NVKERN_SRC/src/nvidia/inc/kernel -I$NVKERN_SRC/src/nvidia/inc/libraries -I$NVKERN_SRC/src/nvidia/arch/nvalloc/common/inc -I$NVKERN_SRC/kernel-open/nvidia-uvm -I$NVKERN_SRC/kernel-open/common/inc -I$NVKERN_SRC/src/common/sdk/nvidia/inc -I$NVKERN_SRC/src/nvidia/arch/nvalloc/unix/include -I$NVKERN_SRC/src/common/sdk/nvidia/inc/ctrl" \
+    -o $BASE/nv/nv.py
+
+  fixup $BASE/nv/nv.py
+  python3 -c "import tinygrad.runtime.autogen.nv.nv"
+}
+
+generate_amd() {
+  # clang2py broken when pass -x c++ to prev headers
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    extra/hip_gpu_driver/nvd.h \
+    extra/hip_gpu_driver/gc_11_0_0_offset.h \
+    extra/hip_gpu_driver/sienna_cichlid_ip_offset.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/amd_gpu.py
+
+  fixup $BASE/amd_gpu.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/amd_gpu.py
+  python3 -c "import tinygrad.runtime.autogen.amd_gpu"
+}
+
+generate_hsa() {
+  clang2py \
+    /opt/rocm/include/hsa/hsa.h \
+    /opt/rocm/include/hsa/hsa_ext_amd.h \
+    /opt/rocm/include/hsa/amd_hsa_signal.h \
+    /opt/rocm/include/hsa/amd_hsa_queue.h \
+    /opt/rocm/include/hsa/amd_hsa_kernel_code.h \
+    /opt/rocm/include/hsa/hsa_ext_finalize.h /opt/rocm/include/hsa/hsa_ext_image.h \
+    /opt/rocm/include/hsa/hsa_ven_amd_aqlprofile.h \
+    --clang-args="-I/opt/rocm/include" \
+    -o $BASE/hsa.py -l /opt/rocm/lib/libhsa-runtime64.so
+
+  fixup $BASE/hsa.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/hsa.py
+  sed -i "s\ctypes.CDLL('/opt/rocm/lib/libhsa-runtime64.so')\ctypes.CDLL(os.getenv('ROCM_PATH')+'/lib/libhsa-runtime64.so' if os.getenv('ROCM_PATH') else ctypes.util.find_library('hsa-runtime64'))\g" $BASE/hsa.py
+  python3 -c "import tinygrad.runtime.autogen.hsa"
+}
+
+generate_io_uring() {
+  clang2py -k cdefstum \
+    /usr/include/liburing.h \
+    /usr/include/linux/io_uring.h \
+    -o $BASE/io_uring.py
+
+  sed -r '/^#define __NR_io_uring/ s/^#define __(NR_io_uring[^ ]+) (.*)$/\1 = \2/; t; d' /usr/include/asm-generic/unistd.h >> $BASE/io_uring.py # io_uring syscalls numbers
+  fixup $BASE/io_uring.py
+}
+
+generate_ib() {
+  clang2py -k cdefstum \
+    /usr/include/infiniband/verbs.h \
+    /usr/include/infiniband/verbs_api.h \
+    /usr/include/infiniband/ib_user_ioctl_verbs.h \
+    /usr/include/rdma/ib_user_verbs.h \
+    -o $BASE/ib.py
+
+  sed -i "s\import ctypes\import ctypes, ctypes.util\g" "$BASE/ib.py"
+  sed -i "s\FIXME_STUB\libibverbs\g" "$BASE/ib.py"
+  sed -i "s\FunctionFactoryStub()\ctypes.CDLL(ctypes.util.find_library('ibverbs'), use_errno=True)\g" "$BASE/ib.py"
+
+  fixup $BASE/ib.py
+}
+
+generate_libc() {
+  clang2py -k cdefstum \
+    $(dpkg -L libc6-dev | grep sys/mman.h) \
+    $(dpkg -L libc6-dev | grep sys/syscall.h) \
+    /usr/include/string.h \
+    /usr/include/elf.h \
+    /usr/include/unistd.h \
+    /usr/include/asm-generic/mman-common.h \
+    -o $BASE/libc.py
+
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libc.py
+  sed -i "s\FIXME_STUB\libc\g" $BASE/libc.py
+  sed -i "s\FunctionFactoryStub()\None if (libc_path := ctypes.util.find_library('c')) is None else ctypes.CDLL(libc_path, use_errno=True)\g" $BASE/libc.py
+
+  fixup $BASE/libc.py
+}
+
+generate_llvm() {
+  INC="$(llvm-config-14 --includedir)"
+  clang2py -k cdefstum \
+    $(find "$INC/llvm-c/" -type f -name '*.h' | sort) \
+    "$INC/llvm/Config/Targets.def" \
+    "$INC/llvm/Config/AsmPrinters.def" \
+    "$INC/llvm/Config/AsmParsers.def" \
+    "$INC/llvm/Config/Disassemblers.def" \
+    --clang-args="$(llvm-config-14 --cflags)" \
+    -o "$BASE/llvm.py"
+
+  sed -i "s\import ctypes\import ctypes, tinygrad.runtime.support.llvm as llvm_support\g" "$BASE/llvm.py"
+  sed -i "s\FIXME_STUB\llvm\g" "$BASE/llvm.py"
+  sed -i "s\FunctionFactoryStub()\ctypes.CDLL(llvm_support.LLVM_PATH)\g" "$BASE/llvm.py"
+
+  fixup "$BASE/llvm.py"
+}
+
+generate_kgsl() {
+  clang2py extra/qcom_gpu_driver/msm_kgsl.h -o $BASE/kgsl.py -k cdefstum
+  fixup $BASE/kgsl.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/kgsl.py
+  sed -nE 's/#define ([A-Za-z0-9_]+)_SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1_SHIFT) \& \1_MASK/p' extra/qcom_gpu_driver/msm_kgsl.h >> $BASE/kgsl.py
+  sed -i "s\fcntl.ioctl(__fd, (__idir<<30)\__fd.ioctl((__idir<<30)\g" $BASE/kgsl.py
+  python3 -c "import tinygrad.runtime.autogen.kgsl"
+}
+
+generate_adreno() {
+  clang2py extra/qcom_gpu_driver/a6xx.xml.h -o $BASE/adreno.py -k cestum
+  sed -nE 's/#define ([A-Za-z0-9_]+)__SHIFT\s*[^\S\r\n]*[0-9]*$/def \1(val): return (val << \1__SHIFT) \& \1__MASK/p' extra/qcom_gpu_driver/a6xx.xml.h >> $BASE/adreno.py
+  fixup $BASE/adreno.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/adreno.py
+  python3 -c "import tinygrad.runtime.autogen.adreno"
+}
+
+generate_qcom() {
+  clang2py -k cdefstum \
+    extra/dsp/include/ion.h \
+    extra/dsp/include/msm_ion.h \
+    extra/dsp/include/adsprpc_shared.h \
+    extra/dsp/include/remote_default.h \
+    extra/dsp/include/apps_std.h \
+    -o $BASE/qcom_dsp.py
+
+  fixup $BASE/qcom_dsp.py
+  python3 -c "import tinygrad.runtime.autogen.qcom_dsp"
+}
+
+generate_pci() {
+  clang2py -k cdefstum \
+    /usr/include/linux/pci_regs.h \
+    -o $BASE/pci.py
+  fixup $BASE/pci.py
+}
+
+generate_vfio() {
+  clang2py -k cdefstum \
+    /usr/include/linux/vfio.h \
+    -o $BASE/vfio.py
+  fixup $BASE/vfio.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/vfio.py
+  sed -i "s\import fcntl, functools\import functools" $BASE/vfio.py
+  sed -i "s\import ctypes,os\a from tinygrad.runtime.support import FileIOInterface\g" $BASE/vfio.py
+  sed -i "s\fcntl.ioctl(__fd, (__idir<<30)\return __fd.ioctl((__idir<<30)\g" $BASE/vfio.py
+}
+
+generate_am() {
+  AMKERN_COMMIT_HASH=ceb12c04e2b5b53ec0779362831f5ee40c4921e4
+  AMKERN_SRC=/tmp/ROCK-Kernel-Driver-$AMKERN_COMMIT_HASH
+  if [ ! -d "$AMKERN_SRC" ]; then
+    git clone https://github.com/ROCm/ROCK-Kernel-Driver $AMKERN_SRC --depth 1
+  fi
+  AMKERN_AMD=$AMKERN_SRC/drivers/gpu/drm/amd/
+  AMKERN_INC=$AMKERN_AMD/include/
+
+  clang2py -k cdefstum \
+    extra/amdpci/headers/v11_structs.h \
+    extra/amdpci/headers/v12_structs.h \
+    extra/amdpci/headers/amdgpu_vm.h \
+    extra/amdpci/headers/discovery.h \
+    extra/amdpci/headers/amdgpu_ucode.h \
+    extra/amdpci/headers/psp_gfx_if.h \
+    extra/amdpci/headers/amdgpu_psp.h \
+    extra/amdpci/headers/amdgpu_irq.h \
+    extra/amdpci/headers/amdgpu_doorbell.h \
+    $AMKERN_INC/soc15_ih_clientid.h \
+    --clang-args="-include stdint.h" \
+    -o $BASE/am/am.py
+  fixup $BASE/am/am.py
+  sed -i "s\(int64_t)\ \g" $BASE/am/am.py
+  sed -i "s\AMDGPU_PTE_MTYPE_VG10(2)\AMDGPU_PTE_MTYPE_VG10(0, 2)\g" $BASE/am/am.py # incorrect parsing (TODO: remove when clang2py is gone).
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/amdkfd/kfd_pm4_headers_ai.h \
+    $AMKERN_AMD/amdgpu/soc15d.h \
+    -o $BASE/am/pm4_soc15.py
+  fixup $BASE/am/pm4_soc15.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/amdkfd/kfd_pm4_headers_ai.h \
+    $AMKERN_AMD/amdgpu/nvd.h \
+    -o $BASE/am/pm4_nv.py
+  fixup $BASE/am/pm4_nv.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/vega10_sdma_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_4_0_0.py
+  fixup $BASE/am/sdma_4_0_0.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/navi10_sdma_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_5_0_0.py
+  fixup $BASE/am/sdma_5_0_0.py
+
+  clang2py -k cdefstum \
+    extra/hip_gpu_driver/sdma_registers.h \
+    $AMKERN_AMD/amdgpu/sdma_v6_0_0_pkt_open.h \
+    --clang-args="-I/opt/rocm/include -x c++" \
+    -o $BASE/am/sdma_6_0_0.py
+  fixup $BASE/am/sdma_6_0_0.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v13_0_0_ppsmc.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu13_driver_if_v13_0_0.h \
+    extra/amdpci/headers/amdgpu_smu.h \
+    -o $BASE/am/smu_v13_0_0.py
+  fixup $BASE/am/smu_v13_0_0.py
+
+  clang2py -k cdefstum \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v14_0_0_pmfw.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu_v14_0_2_ppsmc.h \
+    $AMKERN_AMD/pm/swsmu/inc/pmfw_if/smu14_driver_if_v14_0.h \
+    extra/amdpci/headers/amdgpu_smu.h \
+    --clang-args="-include stdint.h" \
+    -o $BASE/am/smu_v14_0_2.py
+  fixup $BASE/am/smu_v14_0_2.py
+}
+
+generate_sqtt() {
+  clang2py -k cdefstum \
+    extra/sqtt/sqtt.h \
+    -o $BASE/sqtt.py
+
+  fixup $BASE/sqtt.py
+  sed -i "s\import ctypes\import ctypes, os\g" $BASE/sqtt.py
+  python3 -c "import tinygrad.runtime.autogen.sqtt"
+}
+
+generate_webgpu() {
+  clang2py extra/webgpu/webgpu.h -o $BASE/webgpu.py
+  fixup $BASE/webgpu.py
+  sed -i "s/FIXME_STUB/webgpu/g" "$BASE/webgpu.py"
+  sed -i "s/FunctionFactoryStub()/ctypes.CDLL(webgpu_support.WEBGPU_PATH)/g" "$BASE/webgpu.py"
+  sed -i "s/import ctypes/import ctypes, tinygrad.runtime.support.webgpu as webgpu_support/g" "$BASE/webgpu.py"
+  python3 -c "import tinygrad.runtime.autogen.webgpu"
+}
+
+generate_libusb() {
+  clang2py -k cdefstum \
+    /usr/include/libusb-1.0/libusb.h \
+    -o $BASE/libusb.py
+
+  fixup $BASE/libusb.py
+  sed -i "s\import ctypes\import ctypes, ctypes.util, os\g" $BASE/libusb.py
+  sed -i "s/FIXME_STUB/libusb/g" "$BASE/libusb.py"
+  sed -i "s/libusb_le16_to_cpu = libusb_cpu_to_le16//g" "$BASE/libusb.py"
+  sed -i "s/FunctionFactoryStub()/None if (lib_path:=os.getenv('LIBUSB_PATH', ctypes.util.find_library('usb-1.0'))) is None else ctypes.CDLL(lib_path)/g" "$BASE/libusb.py"
+  python3 -c "import tinygrad.runtime.autogen.libusb"
+}
+
+if [ "$1" == "opencl" ]; then generate_opencl
+elif [ "$1" == "hip" ]; then generate_hip
+elif [ "$1" == "comgr" ]; then generate_comgr
+elif [ "$1" == "cuda" ]; then generate_cuda
+elif [ "$1" == "nvrtc" ]; then generate_nvrtc
+elif [ "$1" == "hsa" ]; then generate_hsa
+elif [ "$1" == "kfd" ]; then generate_kfd
+elif [ "$1" == "nv" ]; then generate_nv
+elif [ "$1" == "amd" ]; then generate_amd
+elif [ "$1" == "am" ]; then generate_am
+elif [ "$1" == "nvdrv" ]; then generate_nvdrv
+elif [ "$1" == "sqtt" ]; then generate_sqtt
+elif [ "$1" == "qcom" ]; then generate_qcom
+elif [ "$1" == "io_uring" ]; then generate_io_uring
+elif [ "$1" == "ib" ]; then generate_ib
+elif [ "$1" == "libc" ]; then generate_libc
+elif [ "$1" == "llvm" ]; then generate_llvm
+elif [ "$1" == "kgsl" ]; then generate_kgsl
+elif [ "$1" == "adreno" ]; then generate_adreno
+elif [ "$1" == "pci" ]; then generate_pci
+elif [ "$1" == "vfio" ]; then generate_vfio
+elif [ "$1" == "webgpu" ]; then generate_webgpu
+elif [ "$1" == "libusb" ]; then generate_libusb
+elif [ "$1" == "all" ]; then generate_opencl; generate_hip; generate_comgr; generate_cuda; generate_nvrtc; generate_hsa; generate_kfd; generate_nv; generate_amd; generate_io_uring; generate_libc; generate_am; generate_webgpu
+else echo "usage: $0 <type>"
+fi
@@ -0,0 +1,137 @@
+# tinygrad is a tensor library, and as a tensor library it has multiple parts
+# 1. a "runtime". this allows buffer management, compilation, and running programs
+# 2. a "Device" that uses the runtime but specifies compute in an abstract way for all
+# 3. a "UOp" that fuses the compute into kernels, using memory only when needed
+# 4. a "Tensor" that provides an easy to use frontend with autograd ".backward()"
+
+
+print("******** first, the runtime ***********")
+
+from tinygrad.runtime.ops_cpu import ClangJITCompiler, CPUDevice, CPUProgram
+
+cpu = CPUDevice()
+
+# allocate some buffers
+out = cpu.allocator.alloc(4)
+a = cpu.allocator.alloc(4)
+b = cpu.allocator.alloc(4)
+
+# load in some values (little endian)
+cpu.allocator._copyin(a, memoryview(bytearray([2,0,0,0])))
+cpu.allocator._copyin(b, memoryview(bytearray([3,0,0,0])))
+
+# compile a program to a binary
+lib = ClangJITCompiler().compile("void add(int *out, int *a, int *b) { out[0] = a[0] + b[0]; }")
+
+# create a runtime for the program
+fxn = cpu.runtime("add", lib)
+
+# run the program
+fxn(out, a, b)
+
+# check the data out
+print(val := cpu.allocator._as_buffer(out).cast("I").tolist()[0])
+assert val == 5
+
+
+print("******** second, the Device ***********")
+
+DEVICE = "CPU"   # NOTE: you can change this!
+
+import struct
+from tinygrad.dtype import dtypes
+from tinygrad.device import Buffer, Device
+from tinygrad.uop.ops import UOp, Ops
+from tinygrad.shape.shapetracker import ShapeTracker
+
+# allocate some buffers + load in values
+out = Buffer(DEVICE, 1, dtypes.int32).allocate()
+a = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struct.pack("I", 2))))
+b = Buffer(DEVICE, 1, dtypes.int32).allocate().copyin(memoryview(bytearray(struct.pack("I", 3))))
+# NOTE: a._buf is the same as the return from cpu.allocator.alloc
+
+# describe the computation
+buf_1 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 1)
+buf_2 = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 2)
+ld_1 = UOp(Ops.LOAD, dtypes.int32, (buf_1.view(ShapeTracker.from_shape((1,))),))
+ld_2 = UOp(Ops.LOAD, dtypes.int32, (buf_2.view(ShapeTracker.from_shape((1,))),))
+alu = ld_1 + ld_2
+output_buf = UOp(Ops.DEFINE_GLOBAL, dtypes.int32.ptr(), (), 0)
+st_0 = UOp(Ops.STORE, dtypes.void, (output_buf.view(ShapeTracker.from_shape((1,))), alu))
+s = UOp(Ops.SINK, dtypes.void, (st_0,))
+
+# convert the computation to a "linearized" format (print the format)
+from tinygrad.engine.realize import get_program, CompiledRunner
+program = get_program(s, Device[DEVICE].renderer)
+
+# compile a program (and print the source)
+fxn = CompiledRunner(program)
+print(fxn.p.src)
+# NOTE: fxn.clprg is the CPUProgram
+
+# run the program
+fxn.exec([out, a, b])
+
+# check the data out
+assert out.as_buffer().cast('I')[0] == 5
+
+
+print("******** third, the UOp ***********")
+
+from tinygrad.engine.realize import run_schedule
+from tinygrad.engine.schedule import create_schedule_with_vars
+from tinygrad.schedule.kernelize import get_kernelize_map
+
+# allocate some values + load in values
+a = UOp.new_buffer(DEVICE, 1, dtypes.int32)
+b = UOp.new_buffer(DEVICE, 1, dtypes.int32)
+a.buffer.allocate().copyin(memoryview(bytearray(struct.pack("I", 2))))
+b.buffer.allocate().copyin(memoryview(bytearray(struct.pack("I", 3))))
+
+# describe the computation
+out = a + b
+s = UOp(Ops.SINK, dtypes.void, (out,))
+
+# group the computation into kernels
+becomes_map = get_kernelize_map(s)
+
+# the compute maps to an assign
+assign = becomes_map[a+b]
+
+# the first source is the output buffer (data)
+assert assign.src[0].op is Ops.BUFFER
+# the second source is the kernel (compute)
+assert assign.src[1].op is Ops.KERNEL
+
+# schedule the kernel graph in a linear list
+s = UOp(Ops.SINK, dtypes.void, (assign,))
+sched, _ = create_schedule_with_vars(s)
+assert len(sched) == 1
+
+# DEBUGGING: print the compute ast
+print(sched[-1].ast)
+# NOTE: sched[-1].ast is the same as st_0 above
+
+# the output will be stored in a new buffer
+out = assign.buf_uop
+assert out.op is Ops.BUFFER and not out.buffer.is_allocated()
+print(out)
+
+# run that schedule
+run_schedule(sched)
+
+# check the data out
+assert out.is_realized and out.buffer.as_buffer().cast('I')[0] == 5
+
+
+print("******** fourth, the Tensor ***********")
+
+from tinygrad import Tensor
+
+a = Tensor([2], dtype=dtypes.int32, device=DEVICE)
+b = Tensor([3], dtype=dtypes.int32, device=DEVICE)
+out = a + b
+
+# check the data out
+print(val:=out.item())
+assert val == 5
@@ -38,19 +38,25 @@ optim.schedule_step()   # this will step the optimizer without running realize
 # The weight Tensors have been assigned to, but not yet realized. Everything is still lazy at this point
 # l1.uop and l2.uop define a computation graph

-from tinygrad.engine.schedule import ExecItem
-schedule: List[ExecItem] = Tensor.schedule(l1, l2)
+from tinygrad.engine.schedule import ScheduleItem
+schedule: List[ScheduleItem] = Tensor.schedule(l1, l2)

 print(f"The schedule contains {len(schedule)} items.")
 for si in schedule: print(str(si)[:80])

 # *****
-# 4. Lower and run the schedule.
+# 4. Lower a schedule.

-for si in tqdm(schedule): si.run()
+from tinygrad.engine.realize import lower_schedule_item, ExecItem
+lowered: List[ExecItem] = [lower_schedule_item(si) for si in tqdm(schedule)]

 # *****
-# 5. Print the weight change
+# 5. Run the schedule
+
+for ei in tqdm(lowered): ei.run()
+
+# *****
+# 6. Print the weight change

 print("first weight change\n", l1.numpy()-l1n)
 print("second weight change\n", l2.numpy()-l2n)
@@ -13,19 +13,19 @@ There's also a [doc describing speed](../developer/speed.md)

 Everything in [Tensor](../tensor/index.md) is syntactic sugar around constructing a graph of [UOps](../developer/uop.md).

-The `UOp` graph specifies the compute in terms of low level tinygrad ops. Not all UOps will actually become realized. There's two types of UOps, base and view. base contains compute into a contiguous buffer, and view is a view. Inputs to a base can be either base or view, inputs to a view can only be a single base.
+The `UOp` graph specifies the compute in terms of low level tinygrad ops. Not all UOps will actually become realized. There's two types of UOps, base and view. base contains compute into a contiguous buffer, and view is a view (specified by a ShapeTracker). Inputs to a base can be either base or view, inputs to a view can only be a single base.

 ## Scheduling

-The [scheduler](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/schedule.py) converts the graph of UOps into a list of `ExecItem`. One `ExecItem` is one kernel on the GPU, and the scheduler is responsible for breaking the large compute graph into subgraphs that can fit in a kernel. `ast` specifies what compute to run, and `bufs` specifies what buffers to run it on.
+The [scheduler](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/schedule.py) converts the graph of UOps into a list of `ScheduleItem`. One `ScheduleItem` is one kernel on the GPU, and the scheduler is responsible for breaking the large compute graph into subgraphs that can fit in a kernel. `ast` specifies what compute to run, and `bufs` specifies what buffers to run it on.

-::: tinygrad.engine.schedule.ExecItem
+::: tinygrad.engine.schedule.ScheduleItem

 ## Lowering

-The code in [realize](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/realize.py) lowers `ExecItem` by populating its `prg` field with
+The code in [realize](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/engine/realize.py) lowers `ScheduleItem` to `ExecItem` with

-::: tinygrad.engine.realize.run_schedule
+::: tinygrad.engine.realize.lower_schedule

 There's a ton of complexity hidden behind this, see the `codegen/` directory.

@@ -0,0 +1,109 @@
+# Kernel Creation
+
+Tinygrad lazily builds up a graph of Tensor operations. The Tensor graph includes a mix of:
+
+- Buffer and Assignment Ops: `BUFFER`, `BUFFER_VIEW`, `COPY`, `ASSIGN`
+- Movement Ops: `RESHAPE`, `EXPAND`, `PERMUTE`, `PAD`, `SHRINK`, `FLIP`
+- Compute Ops: `ADD`, `MUL`, `REDUCE_AXIS`, ...
+
+`Tensor.kernelize` creates the kernels and buffers needed to realize the output Tensor(s).
+
+## Kernelize flow
+
+Let's see how a multiply add Tensor graph becomes a fused elementwise kernel.
+
+```py
+# initialize 3 input buffers on the device
+a = Tensor([1]).realize()
+b = Tensor([2]).realize()
+c = Tensor([3]).realize()
+
+# create the Tensor graph
+mul = a*b
+out = mul+c
+
+print(mul) # <Tensor <UOp METAL (1,) int (<Ops.MUL: 48>, None)> on METAL with grad None>
+print(out) # <Tensor <UOp METAL (1,) int (<Ops.ADD: 52>, None)> on METAL with grad None>
+
+out.kernelize()
+
+print(mul) # <Tensor <UOp METAL (1,) int (<Ops.MUL: 48>, None)> on METAL with grad None>
+print(out) # <Tensor <UOp METAL (1,) int (<Ops.ASSIGN: 66>, None)> on METAL with grad None>
+```
+
+The multiply Tensor stays the same because it is fused. The output Tensor's UOp becomes a new ASSIGN UOp:
+
+```py
+print(out.uop)
+```
+
+The first source is the output BUFFER:
+
+```
+UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),
+  UOp(Ops.UNIQUE, dtypes.void, arg=6, src=()),))
+```
+
+And the second source is the KERNEL and its 4 buffer edges (output_buffer, a, b, c):
+
+```
+UOp(Ops.KERNEL, dtypes.void, arg=<Kernel 12 SINK(<Ops.STORE: 45>,) (__add__, __mul__)>, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+    x1:=UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),
+    UOp(Ops.UNIQUE, dtypes.void, arg=6, src=()),)),
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+     x1,
+    UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),)),
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+     x1,
+    UOp(Ops.UNIQUE, dtypes.void, arg=3, src=()),)),
+  UOp(Ops.BUFFER, dtypes.int, arg=1, src=(
+     x1,
+    UOp(Ops.UNIQUE, dtypes.void, arg=5, src=()),)),))
+```
+
+KERNEL describes the compute AST, metadata and memory dependencies.
+
+BUFFER holds a reference to the device memory where the output will be stored.
+
+Once a Tensor is kernelized, all children will LOAD its BUFFER, instead of fusing it:
+
+```py
+child = out+2
+child.kernelize()
+print(child.uop.src[1].arg.ast)
+```
+
+```
+UOp(Ops.SINK, dtypes.void, arg=None, src=(
+  UOp(Ops.STORE, dtypes.void, arg=None, src=(
+    UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=0, src=()),
+    x2:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(1,), strides=(0,), offset=0, mask=None, contiguous=True),)), src=()),
+    UOp(Ops.ADD, dtypes.int, arg=None, src=(
+      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
+        UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(1), arg=1, src=()),
+         x2,)),
+      UOp(Ops.CONST, dtypes.int, arg=2, src=(
+         x2,)),)),)),))
+```
+
+`Tensor.realize` will execute the kernels and write outputs to memory:
+
+```py
+Tensor.realize(out)
+print(out)        # <Tensor <UOp METAL (1,) int (<Ops.BUFFER: 23>, <buf real:True device:METAL size:1 dtype:dtypes.int offset:0>)> on METAL with grad None>
+print(out.item()) # 5
+```
+
+<hr />
+
+**Summary**
+
+- The large Tensor graph is built from a mix of data, compute and movement Ops.
+
+- `Tensor.kernelize` splits the Tensor graph into data (BUFFER), compute (KERNEL) and links dependencies with ASSIGN.
+
+- `Tensor.realize` executes KERNELs on device and replaces the Tensor graph with just a BUFFER.
+
+- Kernelize can be called multiple times on a Tensor. This allows for incrementally building the kernel fusion layout of a large Tensor graph, without having to call `realize` or `schedule`.
@@ -10,7 +10,7 @@ Directories are listed in order of how they are processed.

 Group UOps into kernels.

-::: tinygrad.schedule.rangeify.get_rangeify_map
+::: tinygrad.schedule.kernelize.get_kernelize_map
    options:
        members: false
        show_labels: false
@@ -3,7 +3,7 @@
 This is a list of environment variable that control the runtime behavior of tinygrad and its examples.
 Most of these are self-explanatory, and are usually used to set an option at runtime.

-Example: `CL=1 DEBUG=4 python3 -m pytest`
+Example: `GPU=1 DEBUG=4 python3 -m pytest`

 However you can also decorate a function to set a value only inside that function.

@@ -31,17 +31,20 @@ These control the behavior of core tinygrad even when used as a library.
 Variable | Possible Value(s) | Description
 ---|---|---
 DEBUG               | [1-7]      | enable debugging output (operations, timings, speed, generated code and more)
-CL                  | [1]        | enable OpenCL backend
+GPU                 | [1]        | enable the GPU (OpenCL) backend
 CUDA                | [1]        | enable CUDA backend
 AMD                 | [1]        | enable AMD backend
 NV                  | [1]        | enable NV backend
 METAL               | [1]        | enable Metal backend (for Mac M1 and after)
-CPU                 | [1]        | enable CPU backend
+CPU                 | [1]        | enable CPU (Clang) backend
+LLVM                | [1]        | enable LLVM backend
 BEAM                | [#]        | number of beams in kernel beam search
 DEFAULT_FLOAT       | [HALF, ...]| specify the default float dtype (FLOAT32, HALF, BFLOAT16, FLOAT64, ...), default to FLOAT32
 IMAGE               | [1-2]      | enable 2d specific optimizations
 FLOAT16             | [1]        | use float16 for images instead of float32
-HCQ_VISIBLE_DEVICES | [list[int]]| restricts the HCQ devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
+PTX                 | [1]        | enable the specialized [PTX](https://docs.nvidia.com/cuda/parallel-thread-execution/) assembler for Nvidia GPUs. If not set, defaults to generic CUDA codegen backend.
+PROFILE             | [1]        | enable profiling. This feature is supported in NV, AMD, QCOM and METAL backends.
+VISIBLE_DEVICES     | [list[int]]| restricts the NV/AMD devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
 JIT                 | [0-2]      | 0=disabled, 1=[jit enabled](quickstart.md#jit) (default), 2=jit enabled, but graphs are disabled
 VIZ                 | [1]        | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
 ALLOW_TF32          | [1]        | enable TensorFloat-32 tensor cores on Ampere or newer GPUs.
@@ -131,7 +131,7 @@ timeit.repeat(jit_step, repeat=5, number=1)

 1.0 ms is 75x faster! Note that we aren't syncing the GPU, so GPU time may be slower.

-The first two runs of the function execute normally, with the JIT capturing the kernels. Starting from the third run, only the tinygrad operations are replayed, removing the overhead by skipping Python code execution. So be aware that any non-tinygrad Python values affecting the kernels will be "frozen" from the second run. Note that `Tensor` randomness functions work as expected.
+The slowness the first two times is the JIT capturing the kernels. And this JIT will not run any Python in the function, it will just replay the tinygrad kernels that were run, so be aware that non tinygrad Python operations won't work. Randomness functions work as expected.

 Unlike other JITs, we JIT everything, including the optimizer. Think of it as a dumb replay on different data.

@@ -0,0 +1,293 @@
+#!/usr/bin/env python3
+
+# this file is a "ramp" for people new to tinygrad to think about how to approach it
+# it is runnable and editable.
+# whenever you see stuff like DEBUG=2 or CPU=1 discussed, these are environment variables
+# in a unix shell like bash `DEBUG=2 CPU=1 python docs/ramp.py`
+
+# this pip installs tinygrad master for the system
+# the -e allows you to edit the tinygrad folder and update system tinygrad
+# tinygrad is pure Python, so you are encouraged to do this
+# git pull in the tinygrad directory will also get you the latest
+"""
+git clone https://github.com/tinygrad/tinygrad.git
+cd tinygrad
+python3 -m pip install -e .
+"""
+
+# %% ********
+print("******* PART 1 *******")
+
+# we start with a Device.
+# a Device is where Tensors are stored and compute is run
+# tinygrad autodetects the best device on your system and makes it the DEFAULT
+from tinygrad import Device
+print(Device.DEFAULT)  # on Mac, you can see this prints METAL
+
+# now, lets create a Tensor
+from tinygrad import Tensor, dtypes
+t = Tensor([1,2,3,4])
+
+# you can see this Tensor is on the DEFAULT device with int dtype and shape (4,)
+assert t.device == Device.DEFAULT
+assert t.dtype == dtypes.int
+assert t.shape == (4,)
+
+# unlike in torch, if we print it, it doesn't print the contents
+# this is because tinygrad is lazy
+# this Tensor has not been computed yet
+print(t)
+# <Tensor <UOp METAL (4,) int (<Ops.COPY: 7>, None)> on METAL with grad None>
+
+# the ".uop" property on Tensor contains the specification of how to compute it
+print(t.uop)
+"""
+UOp(Ops.COPY, dtypes.int, arg=None, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=0, src=()),
+    UOp(Ops.DEVICE, dtypes.void, arg='PYTHON', src=()),)),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# as you can see, it's specifying a copy from PYTHON device
+# which is where the [1,2,3,4] array lives
+
+# UOps are the specification language in tinygrad
+# they are immutable and form a DAG
+# they have a "Ops", a "dtype", a tuple of srcs (parents), and an arg
+
+t.realize()
+# if we want to "realize" a tensor, we can with the "realize" method
+# now when we look at the uop, it's changed
+print(t.uop)
+"""
+UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+  UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# the copy was actually run, and now the "uop" of the Tensor is just a BUFFER
+# if you run this script with DEBUG=2 in the environment, you can see the copy happen
+# *** METAL      1 copy       16,   METAL <- PYTHON ...
+
+# now let's do some compute
+# we look at the uop to see the specification of the compute
+t_times_2 = t * 2
+print(t_times_2.uop)
+"""
+UOp(Ops.MUL, dtypes.int, arg=None, src=(
+  UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+    x2:=UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),)),
+  UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+    UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+      UOp(Ops.CONST, dtypes.int, arg=2, src=(
+        UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(), strides=(), offset=0, mask=None, contiguous=True),)), src=(
+           x2,)),)),)),)),))
+"""
+# the BUFFER from above is being multiplied by a CONST 2
+# it's RESHAPEd and EXPANDed to broadcast the CONST to the BUFFER
+
+# we can check the result with
+assert t_times_2.tolist() == [2, 4, 6, 8]
+
+# UOps are both immutable and globally unique
+# if i multiply the Tensor by 4 twice, these result Tensors will have the same uop specification
+t_times_4_try_1 = t * 4
+t_times_4_try_2 = t * 4
+assert t_times_4_try_1.uop is t_times_4_try_2.uop
+# the specification isn't just the same, it's the exact same Python object
+assert t_times_4_try_1 is not t_times_4_try_2
+# the Tensor is a different Python object
+
+# if we realize `t_times_4_try_1` ...
+t_times_4_try_1.realize()
+print(t_times_4_try_2.uop)
+"""
+UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+  UOp(Ops.UNIQUE, dtypes.void, arg=4, src=()),
+  UOp(Ops.DEVICE, dtypes.void, arg='METAL', src=()),))
+"""
+# ... `t_times_4_try_2` also becomes the same BUFFER
+assert t_times_4_try_1.uop is t_times_4_try_2.uop
+# so this print doesn't require any computation, just a copy back to the CPU so we can print it
+print("** only the copy start")
+print(t_times_4_try_2.tolist())  # [4, 8, 12, 16]
+print("** only the copy end")
+# you can confirm this with DEBUG=2, seeing what's printed in between the "**" prints
+
+# tinygrad has an auto differentiation engine that operates according to these same principles
+# the derivative of "log(x)" is "1/x", and you can see this on line 20 of gradient.py
+t_float = Tensor([3.0])
+t_log = t_float.log()
+t_log_grad, = t_log.sum().gradient(t_float)
+# due to how log is implemented, this gradient contains a lot of UOps
+print(t_log_grad.uop)
+# ...not shown here...
+# but if you run with DEBUG=4 (CPU=1 used here for simpler code), you can see the generated code
+"""
+void E_(float* restrict data0, float* restrict data1) {
+  float val0 = *(data1+0);
+  *(data0+0) = (1/val0);
+}
+"""
+# the derivative is close to 1/3
+assert (t_log_grad.item() - 1/3) < 1e-6
+
+# %% ********
+print("******* PART 2 *******")
+
+# we redefine the same t here so this cell can run on it's own
+from tinygrad import Tensor
+t = Tensor([1,2,3,4])
+
+# what's above gives you enough of an understanding to go use tinygrad as a library
+# however, a lot of the beauty of tinygrad is in how easy it is to interact with the internals
+# NOTE: the APIs here are subject to change
+
+t_plus_3_plus_4 = t + 3 + 4
+print(t_plus_3_plus_4.uop)
+"""
+UOp(Ops.ADD, dtypes.int, arg=None, src=(
+  UOp(Ops.ADD, dtypes.int, arg=None, src=(
+    UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+      UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+      x3:=UOp(Ops.DEVICE, dtypes.void, arg='CPU', src=()),)),
+    UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+      UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+        UOp(Ops.CONST, dtypes.int, arg=3, src=(
+          x7:=UOp(Ops.VIEW, dtypes.void, arg=ShapeTracker(views=(View(shape=(), strides=(), offset=0, mask=None, contiguous=True),)), src=(
+             x3,)),)),)),)),)),
+  UOp(Ops.EXPAND, dtypes.int, arg=(4,), src=(
+    UOp(Ops.RESHAPE, dtypes.int, arg=(1,), src=(
+      UOp(Ops.CONST, dtypes.int, arg=4, src=(
+         x7,)),)),)),))
+"""
+# you can see it's adding both 3 and 4
+
+# but by the time we are actually running the code, it's adding 7
+# `kernelize` will simplify and group the operations in the graph into kernels
+t_plus_3_plus_4.kernelize()
+print(t_plus_3_plus_4.uop)
+"""
+UOp(Ops.ASSIGN, dtypes.int, arg=None, src=(
+  x0:=UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+    UOp(Ops.UNIQUE, dtypes.void, arg=7, src=()),
+    x2:=UOp(Ops.DEVICE, dtypes.void, arg='CPU', src=()),)),
+  UOp(Ops.KERNEL, dtypes.void, arg=<Kernel 12 SINK(<Ops.STORE: 48>,) (__add__,)>, src=(
+     x0,
+    UOp(Ops.BUFFER, dtypes.int, arg=4, src=(
+      UOp(Ops.UNIQUE, dtypes.void, arg=1, src=()),
+       x2,)),)),))
+"""
+# ASSIGN has two srcs, src[0] is the BUFFER that's assigned to, and src[1] is the thing to assign
+# src[1] is the GPU Kernel that's going to be run
+# we can get the ast of the Kernel as follows
+kernel_ast = t_plus_3_plus_4.uop.src[1].arg.ast
+
+# almost everything in tinygrad functions as a rewrite of the UOps
+# the codegen rewrites the ast to a simplified form ready for "rendering"
+from tinygrad.codegen import full_rewrite_to_sink
+rewritten_ast = full_rewrite_to_sink(kernel_ast)
+print(rewritten_ast)
+"""
+UOp(Ops.SINK, dtypes.void, arg=None, src=(
+  UOp(Ops.STORE, dtypes.void, arg=None, src=(
+    UOp(Ops.INDEX, dtypes.int.ptr(4), arg=None, src=(
+      UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(4), arg=0, src=()),
+      x3:=UOp(Ops.SPECIAL, dtypes.int, arg=('gidx0', 4), src=()),)),
+    UOp(Ops.ADD, dtypes.int, arg=None, src=(
+      UOp(Ops.LOAD, dtypes.int, arg=None, src=(
+        UOp(Ops.INDEX, dtypes.int.ptr(4), arg=None, src=(
+          UOp(Ops.DEFINE_GLOBAL, dtypes.int.ptr(4), arg=1, src=()),
+           x3,)),)),
+      UOp(Ops.CONST, dtypes.int, arg=7, src=()),)),)),))
+"""
+# you can see at this point we are adding 7, not 3 and 4
+
+# with DEBUG=4, we can see the code.
+# since optimizations are on, it UPCASTed the operation, explicitly writing out all 4 +7s
+t_plus_3_plus_4.realize()
+"""
+void E_4n2(int* restrict data0, int* restrict data1) {
+  int val0 = *(data1+0);
+  int val1 = *(data1+1);
+  int val2 = *(data1+2);
+  int val3 = *(data1+3);
+  *(data0+0) = (val0+7);
+  *(data0+1) = (val1+7);
+  *(data0+2) = (val2+7);
+  *(data0+3) = (val3+7);
+}
+"""
+# the function name E_4n2 is "E" for elementwise op (as opposed to "r" for reduce op)
+# "4" for the size, and "n2" for name deduping (it's the 3rd function with the same E and 4 in this session)
+# when you print the name with DEBUG=2, you'll see the 4 is yellow, meaning that it's upcasted
+# if you run with NOOPT=1 ...
+"""
+void E_4n2(int* restrict data0, int* restrict data1) {
+  for (int ridx0 = 0; ridx0 < 4; ridx0++) {
+    int val0 = *(data1+ridx0);
+    *(data0+ridx0) = (val0+7);
+  }
+}
+"""
+# ... you get this unoptimized code with a loop and the 4 is blue (for global). the color code is in kernel.py
+
+# %% ********
+print("******* PART 3 *******")
+
+# now, we go even lower and understand UOps better and how the graph rewrite engine works.
+# it's much simpler than what's in LLVM or MLIR
+
+from tinygrad import dtypes
+from tinygrad.uop.ops import UOp, Ops
+
+# first, we'll construct some const UOps
+a = UOp(Ops.CONST, dtypes.int, arg=2)
+b = UOp(Ops.CONST, dtypes.int, arg=2)
+
+# if you have been paying attention, you should know these are the same Python object
+assert a is b
+
+# UOps support normal Python math operations, so a_plus_b expresses the spec for 2 + 2
+a_plus_b = a + b
+print(a_plus_b)
+"""
+UOp(Ops.ADD, dtypes.int, arg=None, src=(
+  x0:=UOp(Ops.CONST, dtypes.int, arg=2, src=()),
+   x0,))
+"""
+
+# we could actually render this 2+2 into a language like c and run it
+# or, we can use tinygrad's graph rewrite engine to "constant fold"
+
+from tinygrad.uop.ops import graph_rewrite, UPat, PatternMatcher
+
+# a `PatternMatcher` is a list of tuples. for each element in the list:
+# [0] is the pattern to match, and [1] is the function to run.
+# this function can return either a UOp to replace the pattern with, or None to not replace
+simple_pm = PatternMatcher([
+  (UPat(Ops.ADD, src=(UPat(Ops.CONST, name="c1"), UPat(Ops.CONST, name="c2"))),
+   lambda c1,c2: UOp(Ops.CONST, dtype=c1.dtype, arg=c1.arg+c2.arg)),
+])
+# this pattern matches the addition of two CONST and rewrites it into a single CONST UOp
+
+# to actually apply the pattern to a_plus_b, we use graph_rewrite
+a_plus_b_simplified = graph_rewrite(a_plus_b, simple_pm)
+print(a_plus_b_simplified)
+"""
+UOp(Ops.CONST, dtypes.int, arg=4, src=())
+"""
+# 2+2 is in fact, 4
+
+# we can also use syntactic sugar to write the pattern nicer
+simpler_pm = PatternMatcher([
+  (UPat.cvar("c1")+UPat.cvar("c2"), lambda c1,c2: c1.const_like(c1.arg+c2.arg))
+])
+assert graph_rewrite(a_plus_b, simple_pm) is graph_rewrite(a_plus_b, simpler_pm)
+# note again the use of is, UOps are immutable and globally unique
+
+# %% ********
+
+# that brings you to an understanding of the most core concepts in tinygrad
+# you can run this with VIZ=1 to use the web based graph rewrite explorer
+# hopefully now you understand it. the nodes in the graph are just UOps
@@ -2,17 +2,17 @@

 tinygrad supports various runtimes, enabling your code to scale across a wide range of devices. The default runtime can be automatically selected based on the available hardware, or you can force a specific runtime to be default using environment variables (e.g., `CPU=1`).

-| Runtime | Description | Compiler Options | Requirements |
-|---------|-------------|------------------|--------------|
-| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`NV_PTX=1`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via `NV_IFACE=(NVK\|PCI)`. See [NV interfaces](#nv-interfaces) for details. |
-| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`AMD_LLVM=1`)<br>HIP/COMGR (`AMD_HIP=1`) | RDNA2 or newer GPUs.<br>You can select an interface via `AMD_IFACE=(KFD\|PCI\|USB)`. See [AMD interfaces](#amd-interfaces) for details. |
-| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
-| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for `bfloat` support |
-| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`CUDA_PTX=1`) | NVIDIA GPU with CUDA support |
-| [CL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | - | OpenCL 2.0 compatible device |
-| [CPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang or llvm compiler | Clang JIT (default)<br>LLVM IR (`CPU_LLVM=1`) | `clang` compiler in system `PATH` |
-| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | - | Dawn library installed and discoverable. Binaries: [pydawn v0.3.0](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0) |
-
+| Runtime | Description | Requirements |
+|---------|-------------|--------------|
+| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | Ampere/Ada series GPUs |
+| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | RDNA2/RDNA3/RDNA4 series GPUs. You can select one of the interfaces for communication by setting `AMD_IFACE=(KFD|PCI)`. See [AMD interfaces](#amd-interfaces) for more details. |
+| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | 6xx series GPUs |
+| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | M1+ Macs; Metal 3.0+ for `bfloat` support |
+| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | NVIDIA GPU with CUDA support |
+| [GPU (OpenCL)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_gpu.py) | Accelerates computations using OpenCL on GPUs | OpenCL 2.0 compatible device |
+| [CPU (C Code)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang compiler | `clang` compiler in system `PATH` |
+| [LLVM (LLVM IR)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_llvm.py) | Runs on CPU using the LLVM compiler infrastructure | llvm libraries installed and findable |
+| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | Dawn library installed and findable. Download binaries [here](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0). |

 ## Interoperability

@@ -70,12 +70,5 @@ AMD backend supports several interfaces for communicating with devices:

 * `KFD`: uses the amdgpu driver
 * `PCI`: uses the [AM driver](developer/am.md)
-* `USB`: USB3 interafce for asm24xx chips.

 You can force an interface by setting `AMD_IFACE` to one of these values. In the case of `AMD_IFACE=PCI`, this may unbind your GPU from the amdgpu driver.
-
-## NV Interfaces
-NV backend supports several interfaces for communicating with devices:
-
-* `NVK`: uses the nvidia driver
-* `PCI`: uses the [NV driver](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/support/nv/nvdev.py)
@@ -41,7 +41,7 @@ The BMC also has a web interface you can use if you find that easier.
 It is recommended that you change the BMC password after setting up the box, as the password on the screen is only the initial password.

 If you do decide to change the BMC password and no longer want the initial password to be displayed, remove the `/root/.bmc_password` file.
-Reboot after making these changes or restart the `tinybox-display.service` service.
+Reboot after making these changes or restart the `displayservice.service` service.

 ## What do I use it for?

@@ -0,0 +1,9 @@
+import globals from "globals";
+import pluginJs from "@eslint/js";
+import pluginHtml from "eslint-plugin-html";
+
+export default [
+  {files: ["**/*.html"], plugins: {html: pluginHtml}, rules:{"max-len": ["error", {"code": 150}]}},
+  {languageOptions: {globals: globals.browser}},
+  pluginJs.configs.recommended,
+];
@@ -10,7 +10,7 @@ GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]

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

 # from https://github.com/tysam-code/hlb-CIFAR10/blob/main/main.py
 batchsize = getenv("BS", 1024)
@@ -21,7 +21,7 @@ if __name__ == "__main__":
  X_train, Y_train, X_test, Y_test = mnist(fashion=getenv("FASHION"))

  model = Model()
-  opt = (nn.optim.Muon if getenv("MUON") else nn.optim.SGD if getenv("SGD") else nn.optim.Adam)(nn.state.get_parameters(model))
+  opt = (nn.optim.Adam if not getenv("MUON") else nn.optim.Muon)(nn.state.get_parameters(model))

  @TinyJit
  @Tensor.train()
@@ -1,6 +1,6 @@
 import sys, time
 from tinygrad import TinyJit, GlobalCounters, fetch, getenv
-from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner
 from extra.onnx_helpers import get_example_inputs, validate

 def load_onnx_model(onnx_file):
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+import os, sys, traceback
+sys.path.append(os.getcwd())
+
+from io import StringIO
+from contextlib import redirect_stdout
+from tinygrad import Tensor, nn
+from tinygrad.helpers import Timing, colored, getenv, fetch
+from extra.models.llama import Transformer, convert_from_huggingface, fix_bf16
+from sentencepiece import SentencePieceProcessor
+
+def create_fixed_tokenizer(output_file):
+  print("creating fixed tokenizer")
+  import extra.junk.sentencepiece_model_pb2 as spb2
+  mp = spb2.ModelProto()
+  mp.ParseFromString(fetch("https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/resolve/main/tokenizer.model?download=true").read_bytes())
+  mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_end|>", score=0))
+  mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_start|>", score=0))
+  with open(output_file, "wb") as f:
+    f.write(mp.SerializeToString())
+
+# example:
+# echo -en "write 2+2\nwrite hello world\ny\n" | TEMP=0 python3 examples/coder.py
+
+if __name__ == "__main__":
+  # https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/config.json
+  with Timing("create model: "):
+    model = Transformer(4096, 14336, n_heads=32, n_layers=32, norm_eps=1e-5, vocab_size=32002, n_kv_heads=8, max_context=4096, jit=getenv("JIT", 1))
+
+  with Timing("download weights: "):
+    part1 = nn.state.torch_load(fetch("https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/resolve/main/pytorch_model-00001-of-00002.bin?download=true"))
+    part2 = nn.state.torch_load(fetch("https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/resolve/main/pytorch_model-00002-of-00002.bin?download=true"))
+
+  with Timing("weights -> model: "):
+    nn.state.load_state_dict(model, fix_bf16(convert_from_huggingface(part1, 32, 32, 8)), strict=False)
+    nn.state.load_state_dict(model, fix_bf16(convert_from_huggingface(part2, 32, 32, 8)), strict=False)
+
+  if not os.path.isfile("/tmp/tokenizer.model"): create_fixed_tokenizer("/tmp/tokenizer.model")
+  spp = SentencePieceProcessor(model_file="/tmp/tokenizer.model")
+
+  # https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/blob/main/tokenizer_config.json
+  #   "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+  IM_END = 32000
+  IM_START = 32001
+  def encode_prompt(k, v): return [IM_START]+spp.encode(f"{k}\n{v}")+[IM_END]+spp.encode("\n")
+  def start_prompt(k): return [IM_START]+spp.encode(f"{k}\n")
+  def output(outputted, toks, color):
+    cur = spp.decode(toks)[len(outputted):]
+    sys.stdout.write(colored(cur, color))
+    sys.stdout.flush()
+    outputted += cur
+    return outputted
+
+  # *** app below this line ***
+
+  toks = [spp.bos_id()] + encode_prompt("system", "You are Quentin. Quentin is a useful assistant who writes Python code to answer questions. He keeps the code as short as possible and doesn't read from user input")
+
+  PROMPT = getenv("PROMPT", 1)
+  temperature = getenv("TEMP", 0.7)
+
+  start_pos = 0
+  outputted = output("", toks, "green")
+  turn = True
+  while 1:
+    if PROMPT:
+      toks += encode_prompt("user", input("Q: ")) + start_prompt("assistant")
+    else:
+      toks += start_prompt("user" if turn else "assistant")
+      turn = not turn
+    old_output_len = len(outputted)
+    while 1:
+      tok = model(Tensor([toks[start_pos:]]), start_pos, temperature).item()
+      start_pos = len(toks)
+      toks.append(tok)
+      outputted = output(outputted, toks, "blue" if not turn else "cyan")
+      if tok == IM_END: break
+      if tok == spp.eos_id(): break
+      new_output = outputted[old_output_len:]
+
+      if new_output.endswith("```") and '```python\n' in new_output:
+        python_code = new_output.split('```python\n')[1].split("```")[0]
+        # AI safety. Warning to user. Do not press y if the AI is trying to do unsafe things.
+        if input(colored(f" <-- PYTHON DETECTED, RUN IT? ", "red")).lower() == 'y':
+          my_stdout = StringIO()
+          try:
+            with redirect_stdout(my_stdout): exec(python_code)
+            result = my_stdout.getvalue()
+          except Exception as e:
+            result = ''.join(traceback.format_exception_only(e))
+          toks += spp.encode(f"\nOutput:\n```\n{result}```")
+          outputted = output(outputted, toks, "yellow")
+          old_output_len = len(outputted)
+    print("")
@@ -8,7 +8,7 @@ import numpy as np
 import subprocess
 import tensorflow as tf
 import tf2onnx
-from tinygrad.nn.onnx import OnnxRunner
+from tinygrad.frontend.onnx import OnnxRunner
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import to_mv
 from extra.export_model import export_model_clang, compile_net, jit_model
@@ -0,0 +1,341 @@
+import argparse
+import multiprocessing as mp
+import os
+import re
+import sys
+import time
+from contextlib import contextmanager
+from pathlib import Path
+
+import numpy as np
+import pyaudio
+import yaml
+from llama import LLaMa
+from vits import MODELS as VITS_MODELS
+from vits import Y_LENGTH_ESTIMATE_SCALARS, HParams, Synthesizer, TextMapper, get_hparams_from_file, load_model
+from whisper import init_whisper, transcribe_waveform
+from sentencepiece import SentencePieceProcessor
+
+from tinygrad.helpers import Timing, fetch
+from tinygrad import Tensor, dtypes
+
+# Whisper constants
+RATE = 16000
+CHUNK = 1600
+
+# LLaMa constants
+IM_START = 32001
+IM_END = 32002
+
+
+# Functions for encoding prompts to chatml md
+def encode_prompt(spp, k, v): return [IM_START]+spp.encode(f"{k}\n{v}")+[IM_END]+spp.encode("\n")
+def start_prompt(spp, k): return [IM_START]+spp.encode(f"{k}\n")
+
+def chunks(lst, n):
+  for i in range(0, len(lst), n): yield lst[i:i + n]
+
+def create_fixed_tokenizer():
+  """Function needed for extending tokenizer with additional chat tokens"""
+  import extra.junk.sentencepiece_model_pb2 as spb2
+  tokenizer_path = fetch("https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.4/resolve/main/tokenizer.model")
+  if SentencePieceProcessor(model_file=str(tokenizer_path)).vocab_size() != 32003:
+    print("creating fixed tokenizer")
+    mp = spb2.ModelProto()
+    mp.ParseFromString(tokenizer_path.read_bytes())
+    # https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.4/blob/main/added_tokens.json
+    mp.pieces.append(spb2.ModelProto.SentencePiece(piece="[PAD]", score=0))
+    mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_start|>", score=0))
+    mp.pieces.append(spb2.ModelProto.SentencePiece(piece="<|im_end|>", score=0))
+    tokenizer_path.write_bytes(mp.SerializeToString())
+  return tokenizer_path
+
+def llama_prepare(llama: LLaMa, temperature: float, pre_prompt_path: Path) -> tuple[list[int], str, str, str]:
+  """Prepares a llama model from a specified pre-prompt file"""
+  with open(str(pre_prompt_path)) as f:
+    config = yaml.safe_load(f.read())
+  toks = [llama.tokenizer.bos_id()] + encode_prompt(llama.tokenizer, "system", config["pre_prompt"].replace("\n", " "))
+  for i in config["examples"]:
+    toks += encode_prompt(llama.tokenizer, config["user_delim"], i["user_prompt"])
+    toks += encode_prompt(llama.tokenizer, config["resp_delim"], i["resp_prompt"])
+  llama.model(Tensor([toks]), 0, temperature).realize()  # NOTE: outputs are not used
+  return toks, config["user_delim"], config["resp_delim"], len(toks), llama.tokenizer.decode(toks)
+
+def llama_generate(
+  llama: LLaMa,
+  toks: list[int],
+  outputted: str,
+  prompt: str,
+  start_pos: int,
+  user_delim: str,
+  resp_delim: str,
+  temperature=0.7,
+  max_tokens=1000
+):
+  """Generates an output for the specified prompt"""
+  toks += encode_prompt(llama.tokenizer, user_delim, prompt)
+  toks += start_prompt(llama.tokenizer, resp_delim)
+
+  outputted = llama.tokenizer.decode(toks)
+  init_length = len(outputted)
+  for _ in range(max_tokens):
+    token = llama.model(Tensor([toks[start_pos:]]), start_pos, temperature).item()
+    start_pos = len(toks)
+    toks.append(token)
+
+    cur = llama.tokenizer.decode(toks)
+
+    # Print is just for debugging
+    sys.stdout.write(cur[len(outputted):])
+    sys.stdout.flush()
+    outputted = cur
+    if toks[-1] == IM_END: break
+  else:
+    toks.append(IM_END)
+  print() # because the output is flushed
+  return outputted, start_pos, outputted[init_length:].replace("<|im_end|>", "")
+
+def tts(
+  text_to_synthesize: str,
+  synth: Synthesizer,
+  hps: HParams,
+  emotion_embedding: Path,
+  speaker_id: int,
+  model_to_use: str,
+  noise_scale: float,
+  noise_scale_w: float,
+  length_scale: float,
+  estimate_max_y_length: bool,
+  text_mapper: TextMapper,
+  model_has_multiple_speakers: bool,
+  pad_length=600,
+  vits_pad_length=1000
+):
+  if model_to_use == "mmts-tts": text_to_synthesize = text_mapper.filter_oov(text_to_synthesize.lower())
+
+  # Convert the input text to a tensor.
+  stn_tst = text_mapper.get_text(text_to_synthesize, hps.data.add_blank, hps.data.text_cleaners)
+  init_shape = stn_tst.shape
+  assert init_shape[0] < pad_length, "text is too long"
+  x_tst, x_tst_lengths = stn_tst.pad(((0, pad_length - init_shape[0]),), value=1).unsqueeze(0), Tensor([init_shape[0]], dtype=dtypes.int64)
+  sid = Tensor([speaker_id], dtype=dtypes.int64) if model_has_multiple_speakers else None
+
+  # Perform inference.
+  audio_tensor = synth.infer(x_tst, x_tst_lengths, sid, noise_scale, length_scale, noise_scale_w, emotion_embedding=emotion_embedding,
+                             max_y_length_estimate_scale=Y_LENGTH_ESTIMATE_SCALARS[model_to_use] if estimate_max_y_length else None, pad_length=vits_pad_length)[0, 0]
+  # Save the audio output.
+  audio_data = (np.clip(audio_tensor.numpy(), -1.0, 1.0) * 32767).astype(np.int16)
+  return audio_data
+
+def init_vits(
+  model_to_use: str,
+  emotion_path: Path,
+  speaker_id: int,
+  seed: int,
+):
+  model_config = VITS_MODELS[model_to_use]
+
+  # Load the hyperparameters from the config file.
+  hps = get_hparams_from_file(fetch(model_config[0]))
+
+  # If model has multiple speakers, validate speaker id and retrieve name if available.
+  model_has_multiple_speakers = hps.data.n_speakers > 0
+  if model_has_multiple_speakers:
+    if speaker_id >= hps.data.n_speakers: raise ValueError(f"Speaker ID {speaker_id} is invalid for this model.")
+    if hps.__contains__("speakers"): # maps speaker ids to names
+      speakers = hps.speakers
+      if isinstance(speakers, list): speakers = {speaker: i for i, speaker in enumerate(speakers)}
+
+  # Load emotions if any. TODO: find an english model with emotions, this is untested atm.
+  emotion_embedding = None
+  if emotion_path is not None:
+    if emotion_path.endswith(".npy"): emotion_embedding = Tensor(np.load(emotion_path), dtype=dtypes.int64).unsqueeze(0)
+    else: raise ValueError("Emotion path must be a .npy file.")
+
+  # Load symbols, instantiate TextMapper and clean the text.
+  if hps.__contains__("symbols"): symbols = hps.symbols
+  elif model_to_use == "mmts-tts": symbols = [x.replace("\n", "") for x in fetch("https://huggingface.co/facebook/mms-tts/raw/main/full_models/eng/vocab.txt").open(encoding="utf-8").readlines()]
+  else: symbols = ['_'] + list(';:,.!?¡¿—…"«»“” ') + list('ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz') + list("ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ")
+  text_mapper = TextMapper(apply_cleaners=True, symbols=symbols)
+
+  # Load the model.
+  if seed is not None:
+    Tensor.manual_seed(seed)
+    np.random.seed(seed)
+  net_g = load_model(text_mapper.symbols, hps, model_config)
+
+  return net_g, emotion_embedding, text_mapper, hps, model_has_multiple_speakers
+
+@contextmanager
+def output_stream(num_channels: int, sample_rate: int):
+  try:
+    p = pyaudio.PyAudio()
+    stream = p.open(format=pyaudio.paInt16, channels=num_channels, rate=sample_rate, output=True)
+    yield stream
+  except KeyboardInterrupt: pass
+  finally:
+    stream.stop_stream()
+    stream.close()
+    p.terminate()
+
+@contextmanager
+def log_writer():
+  try:
+    logs = []
+    yield logs
+  finally:
+    sep = "="*os.get_terminal_size()[1]
+    print(f"{sep[:-1]}\nCHAT LOG")
+    print(*logs, sep="\n")
+    print(sep)
+
+def listener(q: mp.Queue, event: mp.Event):
+  try:
+    p = pyaudio.PyAudio()
+    stream = p.open(format=pyaudio.paInt16, channels=1, rate=RATE, input=True, frames_per_buffer=CHUNK)
+    did_print = False
+    while True:
+      data = stream.read(CHUNK) # read data to avoid overflow
+      if event.is_set():
+        if not did_print:
+          print("listening")
+          did_print = True
+        q.put(((np.frombuffer(data, np.int16)/32768).astype(np.float32)*3))
+      else:
+        did_print = False
+  finally:
+    stream.stop_stream()
+    stream.close()
+    p.terminate()
+
+def mp_output_stream(q: mp.Queue, counter: mp.Value, num_channels: int, sample_rate: int):
+  with output_stream(num_channels, sample_rate) as stream:
+    while True:
+      try:
+        stream.write(q.get())
+        counter.value += 1
+      except KeyboardInterrupt:
+        break
+
+if __name__ == "__main__":
+  import nltk
+  nltk.download("punkt")
+  # Parse CLI arguments
+  parser = argparse.ArgumentParser("Have a tiny conversation with tinygrad")
+
+  # Whisper args
+  parser.add_argument("--whisper_model_name", type=str, default="tiny.en")
+
+  # LLAMA args
+  parser.add_argument("--llama_pre_prompt_path", type=Path, default=Path(__file__).parent / "conversation_data" / "pre_prompt_stacy.yaml", help="Path to yaml file which contains all pre-prompt data needed. ")
+  parser.add_argument("--llama_count", type=int, default=1000, help="Max number of tokens to generate")
+  parser.add_argument("--llama_temperature", type=float, default=0.7, help="Temperature in the softmax")
+  parser.add_argument("--llama_quantize", type=str, default=None, help="Quantize the weights to int8 or nf4 in memory")
+  parser.add_argument("--llama_model", type=Path, default=None, help="Folder with the original weights to load, or single .index.json, .safetensors or .bin file")
+  parser.add_argument("--llama_gen", type=str, default="tiny", required=False, help="Generation of the model to use")
+  parser.add_argument("--llama_size", type=str, default="1B-Chat", required=False, help="Size of model to use")
+  parser.add_argument("--llama_tokenizer", type=Path, default=None, required=False, help="Path to llama tokenizer.model")
+
+  # vits args
+  parser.add_argument("--vits_model_to_use", default="vctk", help="Specify the model to use. Default is 'vctk'.")
+  parser.add_argument("--vits_speaker_id", type=int, default=12, help="Specify the speaker ID. Default is 6.")
+  parser.add_argument("--vits_noise_scale", type=float, default=0.667, help="Specify the noise scale. Default is 0.667.")
+  parser.add_argument("--vits_noise_scale_w", type=float, default=0.8, help="Specify the noise scale w. Default is 0.8.")
+  parser.add_argument("--vits_length_scale", type=float, default=1, help="Specify the length scale. Default is 1.")
+  parser.add_argument("--vits_seed", type=int, default=None, help="Specify the seed (set to None if no seed). Default is 1337.")
+  parser.add_argument("--vits_num_channels", type=int, default=1, help="Specify the number of audio output channels. Default is 1.")
+  parser.add_argument("--vits_sample_width", type=int, default=2, help="Specify the number of bytes per sample, adjust if necessary. Default is 2.")
+  parser.add_argument("--vits_emotion_path", type=Path, default=None, help="Specify the path to emotion reference.")
+  parser.add_argument("--vits_estimate_max_y_length", type=str, default=False, help="If true, overestimate the output length and then trim it to the correct length, to prevent premature realization, much more performant for larger inputs, for smaller inputs not so much. Default is False.")
+  parser.add_argument("--vits_vocab_path", type=Path, default=None, help="Path to the TTS vocabulary.")
+
+  # conversation args
+  parser.add_argument("--max_sentence_length", type=int, default=20, help="Max words in one sentence to pass to vits")
+
+  args = parser.parse_args()
+
+  # Init models
+  model, enc = init_whisper(args.whisper_model_name)
+  synth, emotion_embedding, text_mapper, hps, model_has_multiple_speakers = init_vits(args.vits_model_to_use, args.vits_emotion_path, args.vits_speaker_id, args.vits_seed)
+
+  # Download tinyllama chat as a default model
+  if args.llama_model is None:
+    args.llama_model = fetch("https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v0.4/resolve/main/model.safetensors", "tinyllamachat.safetensors")
+    args.llama_gen = "tiny"
+    args.llama_size = "1B-Chat"
+  # Add 3 more tokens to the tokenizer
+  if args.llama_gen == "tiny" and args.llama_size.endswith("Chat"): args.llama_tokenizer = create_fixed_tokenizer()
+  tokenizer_path = args.llama_tokenizer or args.llama_model.parent / "tokenizer.model"
+  llama = LLaMa.build(args.llama_model, tokenizer_path, args.llama_gen, args.llama_size, args.llama_quantize)
+  toks, user_delim, resp_delim, start_pos, outputted = llama_prepare(llama, args.llama_temperature, args.llama_pre_prompt_path)
+
+  # Start child process for mic input
+  q = mp.Queue()
+  is_listening_event = mp.Event()
+  p = mp.Process(target=listener, args=(q, is_listening_event,))
+  p.daemon = True
+  p.start()
+
+  # Start child process for speaker output
+  out_q = mp.Queue()
+  out_counter = mp.Value("i", 0)
+  out_p = mp.Process(target=mp_output_stream, args=(out_q, out_counter, args.vits_num_channels, hps.data.sampling_rate,))
+  out_p.daemon = True
+  out_p.start()
+
+  # JIT tts
+  for i in ["Hello, I'm a chat bot", "I am capable of doing a lot of things"]:
+    tts(
+      i, synth, hps, emotion_embedding,
+      args.vits_speaker_id, args.vits_model_to_use, args.vits_noise_scale,
+      args.vits_noise_scale_w, args.vits_length_scale,
+      args.vits_estimate_max_y_length, text_mapper, model_has_multiple_speakers
+    )
+
+  # Start the pipeline
+  with log_writer() as log:
+    while True:
+      tokens = [enc._special_tokens["<|startoftranscript|>"], enc._special_tokens["<|notimestamps|>"]]
+      total = np.array([])
+      out_counter.value = 0
+
+      s = time.perf_counter()
+      is_listening_event.set()
+      prev_text = None
+      while True:
+        for _ in range(RATE // CHUNK): total = np.concatenate([total, q.get()])
+        txt = transcribe_waveform(model, enc, [total], truncate=True)
+        print(txt, end="\r")
+        if txt == "[BLANK_AUDIO]" or re.match(r"^\([\w+ ]+\)$", txt.strip()): continue
+        if prev_text is not None and prev_text == txt:
+          is_listening_event.clear()
+          break
+        prev_text = txt
+      print() # to avoid llama printing on the same line
+      log.append(f"{user_delim.capitalize()}: {txt}")
+
+      # Generate with llama
+      with Timing("llama generation: "):
+        outputted, start_pos, response = llama_generate(
+          llama, toks, outputted, txt, start_pos,
+          user_delim=user_delim, resp_delim=resp_delim, temperature=args.llama_temperature,
+          max_tokens=args.llama_count
+        )
+        log.append(f"{resp_delim.capitalize()}: {response}")
+
+      # Convert to voice
+      with Timing("tts: "):
+        sentences = nltk.sent_tokenize(response.replace('"', ""))
+        for i in sentences:
+          total = np.array([], dtype=np.int16)
+          for j in chunks(i.split(), args.max_sentence_length):
+            audio_data = tts(
+              " ".join(j), synth, hps, emotion_embedding,
+              args.vits_speaker_id, args.vits_model_to_use, args.vits_noise_scale,
+              args.vits_noise_scale_w, args.vits_length_scale,
+              args.vits_estimate_max_y_length, text_mapper, model_has_multiple_speakers
+            )
+            total = np.concatenate([total, audio_data])
+          out_q.put(total.tobytes())
+      while out_counter.value < len(sentences): continue
+      log.append(f"Total: {time.perf_counter() - s}")
@@ -0,0 +1,89 @@
+# load weights from
+# https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b0-355c32eb.pth
+# a rough copy of
+# https://github.com/lukemelas/EfficientNet-PyTorch/blob/master/efficientnet_pytorch/model.py
+import sys
+import ast
+import time
+import numpy as np
+from PIL import Image
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import getenv, fetch, Timing
+from tinygrad.engine.jit import TinyJit
+from extra.models.efficientnet import EfficientNet
+np.set_printoptions(suppress=True)
+
+# TODO: you should be able to put these in the jitted function
+bias = Tensor([0.485, 0.456, 0.406])
+scale = Tensor([0.229, 0.224, 0.225])
+
+@TinyJit
+def _infer(model, img):
+  img = img.permute((2,0,1))
+  img = img / 255.0
+  img = img - bias.reshape((1,-1,1,1))
+  img = img / scale.reshape((1,-1,1,1))
+  return model.forward(img).realize()
+
+def infer(model, img):
+  # preprocess image
+  aspect_ratio = img.size[0] / img.size[1]
+  img = img.resize((int(224*max(aspect_ratio,1.0)), int(224*max(1.0/aspect_ratio,1.0))))
+
+  img = np.array(img)
+  y0,x0=(np.asarray(img.shape)[:2]-224)//2
+  retimg = img = img[y0:y0+224, x0:x0+224]
+
+  # if you want to look at the image
+  """
+  import matplotlib.pyplot as plt
+  plt.imshow(img)
+  plt.show()
+  """
+
+  # run the net
+  out = _infer(model, Tensor(img.astype("float32"))).numpy()
+
+  # if you want to look at the outputs
+  """
+  import matplotlib.pyplot as plt
+  plt.plot(out[0])
+  plt.show()
+  """
+  return out, retimg
+
+if __name__ == "__main__":
+  # instantiate my net
+  model = EfficientNet(getenv("NUM", 0))
+  model.load_from_pretrained()
+
+  # category labels
+  lbls = ast.literal_eval(fetch("https://gist.githubusercontent.com/yrevar/942d3a0ac09ec9e5eb3a/raw/238f720ff059c1f82f368259d1ca4ffa5dd8f9f5/imagenet1000_clsidx_to_labels.txt").read_text())
+
+  # load image and preprocess
+  url = sys.argv[1] if len(sys.argv) >= 2 else "https://raw.githubusercontent.com/tinygrad/tinygrad/master/docs/showcase/stable_diffusion_by_tinygrad.jpg"
+  if url == 'webcam':
+    import cv2
+    cap = cv2.VideoCapture(0)
+    cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
+    while 1:
+      _ = cap.grab() # discard one frame to circumvent capture buffering
+      ret, frame = cap.read()
+      img = Image.fromarray(frame[:, :, [2,1,0]])
+      lt = time.monotonic_ns()
+      out, retimg = infer(model, img)
+      print(f"{(time.monotonic_ns()-lt)*1e-6:7.2f} ms", np.argmax(out), np.max(out), lbls[np.argmax(out)])
+      SCALE = 3
+      simg = cv2.resize(retimg, (224*SCALE, 224*SCALE))
+      retimg = cv2.cvtColor(simg, cv2.COLOR_RGB2BGR)
+      cv2.imshow('capture', retimg)
+      if cv2.waitKey(1) & 0xFF == ord('q'):
+        break
+    cap.release()
+    cv2.destroyAllWindows()
+  else:
+    img = Image.open(fetch(url))
+    for i in range(getenv("CNT", 1)):
+      with Timing("did inference in "):
+        out, _ = infer(model, img)
+        print(np.argmax(out), np.max(out), lbls[np.argmax(out)])
@@ -0,0 +1,498 @@
+# pip3 install sentencepiece
+
+# This file incorporates code from the following:
+# Github Name                    | License | Link
+# black-forest-labs/flux         | Apache  | https://github.com/black-forest-labs/flux/tree/main/model_licenses
+
+from tinygrad import Tensor, nn, dtypes, TinyJit
+from tinygrad.nn.state import safe_load, load_state_dict
+from tinygrad.helpers import fetch, tqdm, colored
+from sdxl import FirstStage
+from extra.models.clip import FrozenClosedClipEmbedder
+from extra.models.t5 import T5Embedder
+import numpy as np
+
+import math, time, argparse, tempfile
+from typing import List, Dict, Optional, Union, Tuple, Callable
+from dataclasses import dataclass
+from pathlib import Path
+from PIL import Image
+
+urls:dict = {
+  "flux-schnell": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/flux1-schnell.safetensors",
+  "flux-dev": "https://huggingface.co/camenduru/FLUX.1-dev/resolve/main/flux1-dev.sft",
+  "ae": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/ae.safetensors",
+  "T5_1_of_2": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/text_encoder_2/model-00001-of-00002.safetensors",
+  "T5_2_of_2": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/text_encoder_2/model-00002-of-00002.safetensors",
+  "T5_tokenizer": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/tokenizer_2/spiece.model",
+  "clip": "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/text_encoder/model.safetensors"
+}
+
+def tensor_identity(x:Tensor) -> Tensor: return x
+
+class AutoEncoder:
+  def __init__(self, scale_factor:float, shift_factor:float):
+    self.decoder = FirstStage.Decoder(128, 3, 3, 16, [1, 2, 4, 4], 2, 256)
+    self.scale_factor = scale_factor
+    self.shift_factor = shift_factor
+
+  def decode(self, z:Tensor) -> Tensor:
+    z = z / self.scale_factor + self.shift_factor
+    return self.decoder(z)
+
+# Conditioner
+class ClipEmbedder(FrozenClosedClipEmbedder):
+  def __call__(self, texts:Union[str, List[str], Tensor]) -> Tensor:
+    if isinstance(texts, str): texts = [texts]
+    assert isinstance(texts, (list,tuple)), f"expected list of strings, got {type(texts).__name__}"
+    tokens = Tensor.cat(*[Tensor(self.tokenizer.encode(text)) for text in texts], dim=0)
+    return self.transformer.text_model(tokens.reshape(len(texts),-1))[:, tokens.argmax(-1)]
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/math.py
+def attention(q:Tensor, k:Tensor, v:Tensor, pe:Tensor) -> Tensor:
+  q, k = apply_rope(q, k, pe)
+  x = Tensor.scaled_dot_product_attention(q, k, v)
+  return x.rearrange("B H L D -> B L (H D)")
+
+def rope(pos:Tensor, dim:int, theta:int) -> Tensor:
+  assert dim % 2 == 0
+  scale = Tensor.arange(0, dim, 2, dtype=dtypes.float32, device=pos.device) / dim # NOTE: this is torch.float64 in reference implementation
+  omega = 1.0 / (theta**scale)
+  out = Tensor.einsum("...n,d->...nd", pos, omega)
+  out = Tensor.stack(Tensor.cos(out), -Tensor.sin(out), Tensor.sin(out), Tensor.cos(out), dim=-1)
+  out = out.rearrange("b n d (i j) -> b n d i j", i=2, j=2)
+  return out.float()
+
+def apply_rope(xq:Tensor, xk:Tensor, freqs_cis:Tensor) -> Tuple[Tensor, Tensor]:
+  xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
+  xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+  xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
+  xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
+  return xq_out.reshape(*xq.shape).cast(xq.dtype), xk_out.reshape(*xk.shape).cast(xk.dtype)
+
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/modules/layers.py
+class EmbedND:
+  def __init__(self, dim:int, theta:int, axes_dim:List[int]):
+    self.dim = dim
+    self.theta = theta
+    self.axes_dim = axes_dim
+
+  def __call__(self, ids:Tensor) -> Tensor:
+    n_axes = ids.shape[-1]
+    emb = Tensor.cat(*[rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)], dim=-3)
+    return emb.unsqueeze(1)
+
+class MLPEmbedder:
+  def __init__(self, in_dim:int, hidden_dim:int):
+    self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
+    self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)
+
+  def __call__(self, x:Tensor) -> Tensor:
+    return self.out_layer(self.in_layer(x).silu())
+
+class QKNorm:
+  def __init__(self, dim:int):
+    self.query_norm = nn.RMSNorm(dim)
+    self.key_norm = nn.RMSNorm(dim)
+
+  def __call__(self, q:Tensor, k:Tensor) -> Tuple[Tensor, Tensor]:
+    return self.query_norm(q), self.key_norm(k)
+
+class SelfAttention:
+  def __init__(self, dim:int, num_heads:int = 8, qkv_bias:bool = False):
+    self.num_heads = num_heads
+    head_dim = dim // num_heads
+
+    self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+    self.norm = QKNorm(head_dim)
+    self.proj = nn.Linear(dim, dim)
+
+  def __call__(self, x:Tensor, pe:Tensor) -> Tensor:
+    qkv = self.qkv(x)
+    q, k, v = qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    q, k = self.norm(q, k)
+    x = attention(q, k, v, pe=pe)
+    return self.proj(x)
+
+@dataclass
+class ModulationOut:
+  shift:Tensor
+  scale:Tensor
+  gate:Tensor
+
+class Modulation:
+  def __init__(self, dim:int, double:bool):
+    self.is_double = double
+    self.multiplier = 6 if double else 3
+    self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
+
+  def __call__(self, vec:Tensor) -> Tuple[ModulationOut, Optional[ModulationOut]]:
+    out = self.lin(vec.silu())[:, None, :].chunk(self.multiplier, dim=-1)
+    return ModulationOut(*out[:3]), ModulationOut(*out[3:]) if self.is_double else None
+
+class DoubleStreamBlock:
+  def __init__(self, hidden_size:int, num_heads:int, mlp_ratio:float, qkv_bias:bool = False):
+    mlp_hidden_dim = int(hidden_size * mlp_ratio)
+    self.num_heads = num_heads
+    self.hidden_size = hidden_size
+    self.img_mod = Modulation(hidden_size, double=True)
+    self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+    self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.img_mlp = [nn.Linear(hidden_size, mlp_hidden_dim, bias=True), Tensor.gelu, nn.Linear(mlp_hidden_dim, hidden_size, bias=True)]
+
+    self.txt_mod = Modulation(hidden_size, double=True)
+    self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+    self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.txt_mlp = [nn.Linear(hidden_size, mlp_hidden_dim, bias=True), Tensor.gelu, nn.Linear(mlp_hidden_dim, hidden_size, bias=True)]
+
+  def __call__(self, img:Tensor, txt:Tensor, vec:Tensor, pe:Tensor) -> tuple[Tensor, Tensor]:
+    img_mod1, img_mod2 = self.img_mod(vec)
+    txt_mod1, txt_mod2 = self.txt_mod(vec)
+    assert img_mod2 is not None and txt_mod2 is not None
+    # prepare image for attention
+    img_modulated = self.img_norm1(img)
+    img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+    img_qkv = self.img_attn.qkv(img_modulated)
+    img_q, img_k, img_v = img_qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    img_q, img_k = self.img_attn.norm(img_q, img_k)
+
+    # prepare txt for attention
+    txt_modulated = self.txt_norm1(txt)
+    txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+    txt_qkv = self.txt_attn.qkv(txt_modulated)
+    txt_q, txt_k, txt_v = txt_qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k)
+
+    # run actual attention
+    q = Tensor.cat(txt_q, img_q, dim=2)
+    k = Tensor.cat(txt_k, img_k, dim=2)
+    v = Tensor.cat(txt_v, img_v, dim=2)
+
+    attn = attention(q, k, v, pe=pe)
+    txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+    # calculate the img bloks
+    img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+    img = img + img_mod2.gate * ((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift).sequential(self.img_mlp)
+
+    # calculate the txt bloks
+    txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn)
+    txt = txt + txt_mod2.gate * ((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift).sequential(self.txt_mlp)
+    return img, txt
+
+
+class SingleStreamBlock:
+  """
+  A DiT block with parallel linear layers as described in
+  https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+  """
+
+  def __init__(self,hidden_size:int, num_heads:int, mlp_ratio:float=4.0, qk_scale:Optional[float]=None):
+    self.hidden_dim = hidden_size
+    self.num_heads = num_heads
+    head_dim = hidden_size // num_heads
+    self.scale = qk_scale or head_dim**-0.5
+
+    self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+    # qkv and mlp_in
+    self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
+    # proj and mlp_out
+    self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
+
+    self.norm = QKNorm(head_dim)
+
+    self.hidden_size = hidden_size
+    self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+    self.mlp_act = Tensor.gelu
+    self.modulation = Modulation(hidden_size, double=False)
+
+  def __call__(self, x:Tensor, vec:Tensor, pe:Tensor) -> Tensor:
+    mod, _ = self.modulation(vec)
+    x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+    qkv, mlp = Tensor.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+    q, k, v = qkv.rearrange("B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+    q, k = self.norm(q, k)
+
+    # compute attention
+    attn = attention(q, k, v, pe=pe)
+    # compute activation in mlp stream, cat again and run second linear layer
+    output = self.linear2(Tensor.cat(attn, self.mlp_act(mlp), dim=2))
+    return x + mod.gate * output
+
+
+class LastLayer:
+  def __init__(self, hidden_size:int, patch_size:int, out_channels:int):
+    self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+    self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+    self.adaLN_modulation:List[Callable[[Tensor], Tensor]] = [Tensor.silu, nn.Linear(hidden_size, 2 * hidden_size, bias=True)]
+
+  def __call__(self, x:Tensor, vec:Tensor) -> Tensor:
+    shift, scale = vec.sequential(self.adaLN_modulation).chunk(2, dim=1)
+    x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+    return self.linear(x)
+
+def timestep_embedding(t:Tensor, dim:int, max_period:int=10000, time_factor:float=1000.0) -> Tensor:
+  """
+  Create sinusoidal timestep embeddings.
+  :param t: a 1-D Tensor of N indices, one per batch element.
+                    These may be fractional.
+  :param dim: the dimension of the output.
+  :param max_period: controls the minimum frequency of the embeddings.
+  :return: an (N, D) Tensor of positional embeddings.
+  """
+  t = time_factor * t
+  half = dim // 2
+  freqs = Tensor.exp(-math.log(max_period) * Tensor.arange(0, stop=half, dtype=dtypes.float32) / half).to(t.device)
+
+  args = t[:, None].float() * freqs[None]
+  embedding = Tensor.cat(Tensor.cos(args), Tensor.sin(args), dim=-1)
+  if dim % 2:  embedding = Tensor.cat(*[embedding, Tensor.zeros_like(embedding[:, :1])], dim=-1)
+  if Tensor.is_floating_point(t):  embedding = embedding.cast(t.dtype)
+  return embedding
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/model.py
+class Flux:
+  """
+  Transformer model for flow matching on sequences.
+  """
+
+  def __init__(
+      self,
+      guidance_embed:bool,
+      in_channels:int = 64,
+      vec_in_dim:int = 768,
+      context_in_dim:int = 4096,
+      hidden_size:int = 3072,
+      mlp_ratio:float = 4.0,
+      num_heads:int = 24,
+      depth:int = 19,
+      depth_single_blocks:int = 38,
+      axes_dim:Optional[List[int]] = None,
+      theta:int = 10_000,
+      qkv_bias:bool = True,
+      ):
+
+    axes_dim = axes_dim or [16, 56, 56]
+    self.guidance_embed = guidance_embed
+    self.in_channels = in_channels
+    self.out_channels = self.in_channels
+    if hidden_size % num_heads != 0:
+      raise ValueError(f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}")
+    pe_dim = hidden_size // num_heads
+    if sum(axes_dim) != pe_dim:
+      raise ValueError(f"Got {axes_dim} but expected positional dim {pe_dim}")
+    self.hidden_size = hidden_size
+    self.num_heads = num_heads
+    self.pe_embedder = EmbedND(dim=pe_dim, theta=theta, axes_dim=axes_dim)
+    self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=True)
+    self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
+    self.vector_in = MLPEmbedder(vec_in_dim, self.hidden_size)
+    self.guidance_in:Callable[[Tensor], Tensor] = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if guidance_embed else tensor_identity
+    self.txt_in = nn.Linear(context_in_dim, self.hidden_size)
+
+    self.double_blocks = [DoubleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias) for _ in range(depth)]
+    self.single_blocks = [SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=mlp_ratio) for _ in range(depth_single_blocks)]
+    self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels)
+
+  def __call__(self, img:Tensor, img_ids:Tensor, txt:Tensor, txt_ids:Tensor, timesteps:Tensor, y:Tensor, guidance:Optional[Tensor] = None) -> Tensor:
+    if img.ndim != 3 or txt.ndim != 3:
+      raise ValueError("Input img and txt tensors must have 3 dimensions.")
+    # running on sequences img
+    img = self.img_in(img)
+    vec = self.time_in(timestep_embedding(timesteps, 256))
+    if self.guidance_embed:
+      if guidance is None:
+        raise ValueError("Didn't get guidance strength for guidance distilled model.")
+      vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+    vec = vec + self.vector_in(y)
+    txt = self.txt_in(txt)
+    ids = Tensor.cat(txt_ids, img_ids, dim=1)
+    pe = self.pe_embedder(ids)
+    for double_block in self.double_blocks:
+      img, txt = double_block(img=img, txt=txt, vec=vec, pe=pe)
+
+    img = Tensor.cat(txt, img, dim=1)
+    for single_block in self.single_blocks:
+      img = single_block(img, vec=vec, pe=pe)
+
+    img = img[:, txt.shape[1] :, ...]
+
+    return self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/util.py
+def load_flow_model(name:str, model_path:str):
+  # Loading Flux
+  print("Init model")
+  model = Flux(guidance_embed=(name != "flux-schnell"))
+  if not model_path: model_path = fetch(urls[name])
+  state_dict = {k.replace("scale", "weight"): v for k, v in safe_load(model_path).items()}
+  load_state_dict(model, state_dict)
+  return model
+
+def load_T5(max_length:int=512):
+  # max length 64, 128, 256 and 512 should work (if your sequence is short enough)
+  print("Init T5")
+  T5 = T5Embedder(max_length, fetch(urls["T5_tokenizer"]))
+  pt_1 = fetch(urls["T5_1_of_2"])
+  pt_2 = fetch(urls["T5_2_of_2"])
+  load_state_dict(T5.encoder, safe_load(pt_1) | safe_load(pt_2), strict=False)
+  return T5
+
+def load_clip():
+  print("Init Clip")
+  clip = ClipEmbedder()
+  load_state_dict(clip.transformer, safe_load(fetch(urls["clip"])))
+  return clip
+
+def load_ae() -> AutoEncoder:
+  # Loading the autoencoder
+  print("Init AE")
+  ae = AutoEncoder(0.3611, 0.1159)
+  load_state_dict(ae, safe_load(fetch(urls["ae"])))
+  return ae
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/sampling.py
+def prepare(T5:T5Embedder, clip:ClipEmbedder, img:Tensor, prompt:Union[str, List[str]]) -> Dict[str, Tensor]:
+  bs, _, h, w = img.shape
+  if bs == 1 and not isinstance(prompt, str):
+    bs = len(prompt)
+
+  img = img.rearrange("b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+  if img.shape[0] == 1 and bs > 1:
+    img = img.expand((bs, *img.shape[1:]))
+
+  img_ids = Tensor.zeros(h // 2, w // 2, 3).contiguous()
+  img_ids[..., 1] = img_ids[..., 1] + Tensor.arange(h // 2)[:, None]
+  img_ids[..., 2] = img_ids[..., 2] + Tensor.arange(w // 2)[None, :]
+  img_ids = img_ids.rearrange("h w c -> 1 (h w) c")
+  img_ids = img_ids.expand((bs, *img_ids.shape[1:]))
+
+  if isinstance(prompt, str):
+    prompt = [prompt]
+  txt = T5(prompt).realize()
+  if txt.shape[0] == 1 and bs > 1:
+    txt = txt.expand((bs, *txt.shape[1:]))
+  txt_ids = Tensor.zeros(bs, txt.shape[1], 3)
+
+  vec = clip(prompt).realize()
+  if vec.shape[0] == 1 and bs > 1:
+    vec = vec.expand((bs, *vec.shape[1:]))
+
+  return {"img": img, "img_ids": img_ids.to(img.device), "txt": txt.to(img.device), "txt_ids": txt_ids.to(img.device), "vec": vec.to(img.device)}
+
+
+def get_schedule(num_steps:int, image_seq_len:int, base_shift:float=0.5, max_shift:float=1.15, shift:bool=True) -> List[float]:
+  # extra step for zero
+  step_size = -1.0 / num_steps
+  timesteps = Tensor.arange(1, 0 + step_size, step_size)
+
+  # shifting the schedule to favor high timesteps for higher signal images
+  if shift:
+    # estimate mu based on linear estimation between two points
+    mu = 0.5 + (max_shift - base_shift) * (image_seq_len - 256) / (4096 - 256)
+    timesteps = math.exp(mu) / (math.exp(mu) + (1 / timesteps - 1))
+  return timesteps.tolist()
+
+@TinyJit
+def run(model, *args): return model(*args).realize()
+
+def denoise(model, img:Tensor, img_ids:Tensor, txt:Tensor, txt_ids:Tensor, vec:Tensor, timesteps:List[float], guidance:float=4.0) -> Tensor:
+  # this is ignored for schnell
+  guidance_vec = Tensor((guidance,), device=img.device, dtype=img.dtype).expand((img.shape[0],))
+  for t_curr, t_prev in tqdm(list(zip(timesteps[:-1], timesteps[1:])), "Denoising"):
+    t_vec = Tensor((t_curr,), device=img.device, dtype=img.dtype).expand((img.shape[0],))
+    pred = run(model, img, img_ids, txt, txt_ids, t_vec, vec, guidance_vec)
+    img = img + (t_prev - t_curr) * pred
+
+  return img
+
+def unpack(x:Tensor, height:int, width:int) -> Tensor:
+  return x.rearrange("b (h w) (c ph pw) -> b c (h ph) (w pw)", h=math.ceil(height / 16), w=math.ceil(width / 16), ph=2, pw=2)
+
+# https://github.com/black-forest-labs/flux/blob/main/src/flux/cli.py
+if __name__ == "__main__":
+  default_prompt = "bananas and a can of coke"
+  parser = argparse.ArgumentParser(description="Run Flux.1", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+
+  parser.add_argument("--name",       type=str,   default="flux-schnell", help="Name of the model to load")
+  parser.add_argument("--model_path", type=str,   default="",             help="path of the model file")
+  parser.add_argument("--width",      type=int,   default=512,            help="width of the sample in pixels (should be a multiple of 16)")
+  parser.add_argument("--height",     type=int,   default=512,            help="height of the sample in pixels (should be a multiple of 16)")
+  parser.add_argument("--seed",       type=int,   default=None,           help="Set a seed for sampling")
+  parser.add_argument("--prompt",     type=str,   default=default_prompt, help="Prompt used for sampling")
+  parser.add_argument('--out',        type=str,   default=Path(tempfile.gettempdir()) / "rendered.png", help="Output filename")
+  parser.add_argument("--num_steps",  type=int,   default=None,           help="number of sampling steps (default 4 for schnell, 50 for guidance distilled)") #noqa:E501
+  parser.add_argument("--guidance",   type=float, default=3.5,            help="guidance value used for guidance distillation")
+  parser.add_argument("--output_dir", type=str,   default="output",       help="output directory")
+  args = parser.parse_args()
+
+  if args.name not in ["flux-schnell", "flux-dev"]:
+    raise ValueError(f"Got unknown model name: {args.name}, chose from flux-schnell and flux-dev")
+
+  if args.num_steps is None:
+    args.num_steps = 4 if args.name == "flux-schnell" else 50
+
+  # allow for packing and conversion to latent space
+  height = 16 * (args.height // 16)
+  width = 16 * (args.width // 16)
+
+  if args.seed is None: args.seed = Tensor._seed
+  else: Tensor.manual_seed(args.seed)
+
+  print(f"Generating with seed {args.seed}:\n{args.prompt}")
+  t0 = time.perf_counter()
+
+  # prepare input noise
+  x = Tensor.randn(1, 16, 2 * math.ceil(height / 16), 2 * math.ceil(width / 16), dtype="bfloat16")
+
+  # load text embedders
+  T5 = load_T5(max_length=256 if args.name == "flux-schnell" else 512)
+  clip = load_clip()
+
+  # embed text to get inputs for model
+  inp = prepare(T5, clip, x, prompt=args.prompt)
+  timesteps = get_schedule(args.num_steps, inp["img"].shape[1], shift=(args.name != "flux-schnell"))
+
+  # done with text embedders
+  del T5, clip
+
+  # load model
+  model = load_flow_model(args.name, args.model_path)
+
+  # denoise initial noise
+  x = denoise(model, **inp, timesteps=timesteps, guidance=args.guidance)
+
+  # done with model
+  del model, run
+
+  # load autoencoder
+  ae = load_ae()
+
+  # decode latents to pixel space
+  x = unpack(x.float(), height, width)
+  x = ae.decode(x).realize()
+
+  t1 = time.perf_counter()
+  print(f"Done in {t1 - t0:.1f}s. Saving {args.out}")
+
+  # bring into PIL format and save
+  x = x.clamp(-1, 1)
+  x = x[0].rearrange("c h w -> h w c")
+  x = (127.5 * (x + 1.0)).cast("uint8")
+
+  img = Image.fromarray(x.numpy())
+
+  img.save(args.out)
+
+  # validation!
+  if args.prompt == default_prompt and args.name=="flux-schnell" and args.seed == 0 and args.width == args.height == 512:
+    ref_image = Tensor(np.array(Image.open("examples/flux1_seed0.png")))
+    distance = (((x.cast(dtypes.float) - ref_image.cast(dtypes.float)) / ref_image.max())**2).mean().item()
+    assert distance < 4e-3, colored(f"validation failed with {distance=}", "red")
+    print(colored(f"output validated with {distance=}", "green"))
@@ -26,8 +26,8 @@ class Attention:
      start_pos = start_pos.val

    if HALF: x = x.half()
-    xqkv = self.c_attn(x).reshape(None, None, 3, self.n_heads, self.head_dim)
-    xq, xk, xv = [xqkv[:, :, i, :, :] for i in range(3)]
+    xqkv = self.c_attn(x)
+    xq, xk, xv = [xqkv.shrink((None, None, (i*self.dim, (i+1)*self.dim))).reshape(None, None, self.n_heads, self.head_dim) for i in range(3)]
    bsz, seqlen, _, _ = xq.shape

    # create kv cache
@@ -35,11 +35,11 @@ class Attention:
      self.cache_kv = Tensor.zeros(2, bsz, MAX_CONTEXT, self.n_heads, self.head_dim, dtype=x.dtype).contiguous().realize()

    # update the cache
-    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()
+    self.cache_kv.shrink((None, None,(start_pos,start_pos+seqlen),None,None)).assign(Tensor.stack(xk, xv)).realize()

    if start_pos > 0:
-      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
-      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
+      keys = self.cache_kv[0].shrink((None, (0, start_pos+seqlen), None, None))
+      values = self.cache_kv[1].shrink((None, (0, start_pos+seqlen), None, None))
    else:
      keys = xk
      values = xv
@@ -64,7 +64,7 @@ class TransformerBlock:

  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]):
    h = x + self.attn(self.ln_1(x), start_pos, mask).float()
-    return (h + self.mlp(self.ln_2(h))).contiguous()
+    return (h + self.mlp(self.ln_2(h)))

 class Transformer:
  def __init__(self, dim, n_heads, n_layers, norm_eps, vocab_size, max_seq_len=1024):
@@ -181,7 +181,6 @@ class GPT2:
    self.tokenizer = tokenizer

  def generate(self, prompt:str, max_length:int, temperature:float, timing:bool=False, batch_size:int=1):
-    step_times = []
    prompt_tokens = self.tokenizer.encode(prompt, allowed_special={"<|endoftext|>"})
    toks = [prompt_tokens[:] for _ in range(batch_size)]
    start_pos = 0
@@ -189,7 +188,7 @@ class GPT2:
      GlobalCounters.reset()
      if timing: print("")
      st = GlobalCounters.time_sum_s
-      with Timing("ran model in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
+      with Timing("ran model in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
                  f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
                  (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=timing):
        with WallTimeEvent(BenchEvent.STEP):
@@ -198,13 +197,8 @@ class GPT2:
          else:
            tokens = Tensor([x[start_pos:] for x in toks])
          tok = self.model(tokens, Variable("start_pos", 1 if start_pos else 0, MAX_CONTEXT-1).bind(start_pos), temperature).tolist()
-      step_times.append((GlobalCounters.time_sum_s-st)*1e3)
      start_pos = len(toks[0])
      for i,t in enumerate(tok): toks[i].append(t)
-
-    if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-      min_time = min(step_times)
-      assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
    return [self.tokenizer.decode(x) for x in toks]

 # **** main code ****
@@ -232,7 +226,7 @@ if __name__ == "__main__":
  gpt2 = GPT2.build_gguf(args.model_size) if args.model_size.startswith("gpt2_gguf_") else GPT2.build(args.model_size)

  if args.benchmark != -1:
-    gpt2.model(Tensor.randint(args.batch_size, args.benchmark), Variable("a", 0, MAX_CONTEXT).bind(0)).realize()
+    gpt2.model(Tensor.rand(args.batch_size, args.benchmark), Variable("a", 0, MAX_CONTEXT).bind(0)).realize()
  else:
    texts = gpt2.generate(args.prompt, args.count, args.temperature, timing=args.timing, batch_size=args.batch_size)
    if not args.noshow:
@@ -1,108 +0,0 @@
-import itertools
-from typing import Callable
-from tinygrad import nn, Tensor, dtypes, Device, TinyJit
-from tinygrad.helpers import getenv, trange, partition
-
-class Model:
-  def __init__(self):
-    self.layers: list[Callable[[Tensor], Tensor]] = [
-      nn.Conv2d(1, 32, 5), Tensor.relu,
-      nn.Conv2d(32, 32, 5), Tensor.relu,
-      nn.BatchNorm(32), Tensor.max_pool2d,
-      nn.Conv2d(32, 64, 3), Tensor.relu,
-      nn.Conv2d(64, 64, 3), Tensor.relu,
-      nn.BatchNorm(64), Tensor.max_pool2d,
-      lambda x: x.flatten(1), nn.Linear(576, 10)]
-
-  def __call__(self, x:Tensor) -> Tensor: return x.sequential(self.layers)
-
-# TODO: refactor this into optim/onnx
-def functional_adam(g:Tensor, m:Tensor, v:Tensor, b1_t:Tensor, b2_t:Tensor, lr=0.001, b1=0.9, b2=0.999, eps=1e-6) -> Tensor:
-  b1_t *= b1
-  b2_t *= b2
-  m.assign(b1 * m + (1.0 - b1) * g)
-  v.assign(b2 * v + (1.0 - b2) * (g * g))
-  m_hat = m / (1.0 - b1_t)
-  v_hat = v / (1.0 - b2_t)
-  return lr * (m_hat / (v_hat.sqrt() + eps))
-
-if __name__ == "__main__":
-  BS = getenv("BS", 512)
-  ACC_STEPS = getenv("ACC_STEPS", 8)
-
-  X_train, Y_train, X_test, Y_test = nn.datasets.mnist()
-  model = Model()
-
-  params = nn.state.get_parameters(model)
-
-  # init params, set requires grad on the ones we need gradients of
-  for x in params:
-    if x.requires_grad is None: x.requires_grad_()
-    x.replace(x.contiguous())
-  Tensor.realize(*params)
-
-  # split params (with grads) and buffers (without)
-  params, buffers = partition(params, lambda x: x.requires_grad)
-  print(f"params: {len(params)} buffers: {len(buffers)}")
-
-  # optim params
-  pos_params = list(itertools.accumulate(params, lambda x,y: x+y.numel(), initial=0))
-  adam_m = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
-  adam_v = Tensor.zeros(pos_params[-1], device="CPU").contiguous()
-  adam_b1_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False).contiguous()
-  adam_b2_t = Tensor.ones((1,), dtype=dtypes.float32, device="CPU", requires_grad=False).contiguous()
-  adam_params = [adam_m, adam_v, adam_b1_t, adam_b2_t]
-
-  # create loss and grads. init all state so the JIT works on microbatch
-  for x in params: x.assign(x.detach())
-  loss = Tensor.zeros(tuple()).contiguous()
-  grads = Tensor.zeros(pos_params[-1]).contiguous()
-  Tensor.realize(*params, *buffers, *adam_params, loss, grads)
-
-  @TinyJit
-  @Tensor.train()
-  def microbatch():
-    samples = Tensor.randint(BS // ACC_STEPS, high=X_train.shape[0])
-    for t in params: t.grad = None
-    # divide by ACC_STEPS at the loss
-    uloss = (model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]) / ACC_STEPS).backward()
-    ugrads = Tensor.cat(*[t.grad.contiguous().flatten() for t in params], dim=0)
-    for t in params: t.grad = None
-    # concat the grads and assign them
-    loss.assign(loss + uloss)
-    grads.assign(grads + ugrads)
-    Tensor.realize(*params, *buffers, loss, grads)
-
-  @TinyJit
-  def optimizer():
-    # run optimizer (on CPU, where adam params live)
-    delta = functional_adam(grads.to("CPU"), adam_m, adam_v, adam_b1_t, adam_b2_t)
-
-    # update the params, copying back the delta one at a time to avoid OOM
-    # NOTE: the scheduler is ordering things poorly, all the copies are happening before the adds
-    for j,tt in enumerate(params):
-      tt.assign(tt.detach() - delta[pos_params[j]:pos_params[j+1]].reshape(tt.shape).to(Device.DEFAULT))
-
-    # realize everything, zero out loss and grads
-    loss.assign(Tensor.zeros_like(loss))
-    grads.assign(Tensor.zeros_like(grads))
-    Tensor.realize(*params, *adam_params, loss, grads)
-
-  @TinyJit
-  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100
-
-  test_acc = float('nan')
-  for i in (t:=trange(getenv("STEPS", 70))):
-    # microbatch sets the gradients
-    for _ in range(ACC_STEPS): microbatch()
-
-    # get the loss before the optimizer clears it
-    # this is already realized so this isn't a schedule
-    loss_item = loss.item()
-
-    # run the optimizer
-    optimizer()
-
-    # eval
-    if i%10 == 9: test_acc = get_test_acc().item()
-    t.set_description(f"loss: {loss_item:6.2f} test_accuracy: {test_acc:5.2f}%")
@@ -145,6 +145,7 @@ hyp = {
  },
 }

+@Context(FUSE_ARANGE=getenv("FUSE_ARANGE", 1))
 def train_cifar():

  def set_seed(seed):
@@ -228,8 +229,7 @@ def train_cifar():
    if getenv("RANDOM_CROP", 1):
      X = random_crop(X, crop_size=32)
    if getenv("RANDOM_FLIP", 1):
-      # NOTE: RANGEIFY=1 needs this contiguous or the X[perms] is very slow
-      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X).contiguous() # flip LR
+      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X) # flip LR
    X, Y = X[perms], Y[perms]
    return X, Y, *cutmix(X, Y, perms, mask_size=hyp['net']['cutmix_size'])

@@ -355,7 +355,7 @@ def train_cifar():

  # https://www.anandtech.com/show/16727/nvidia-announces-geforce-rtx-3080-ti-3070-ti-upgraded-cards-coming-in-june
  # 136 TFLOPS is the theoretical max w float16 on 3080 Ti
-  step_times = []
+
  model_ema: Optional[modelEMA] = None
  projected_ema_decay_val = hyp['ema']['decay_base'] ** hyp['ema']['every_n_steps']
  i = 0
@@ -413,17 +413,12 @@ def train_cifar():
          model_ema.update(model, Tensor([projected_ema_decay_val*(i/STEPS)**hyp['ema']['decay_pow']]))

      cl = time.monotonic()
-      step_times.append((cl-st)*1000.0)
      device_str = loss.device if isinstance(loss.device, str) else f"{loss.device[0]} * {len(loss.device)}"
      #  53  221.74 ms run,    2.22 ms python,  219.52 ms CL,  803.39 loss, 0.000807 LR, 4.66 GB used,   3042.49 GFLOPS,    674.65 GOPS
      print(f"{i:3d} {(cl-st)*1000.0:7.2f} ms run, {(et-st)*1000.0:7.2f} ms python, {(cl-et)*1000.0:7.2f} ms {device_str}, {loss_cpu:7.2f} loss, {opt_non_bias.lr.numpy()[0]:.6f} LR, {GlobalCounters.mem_used/1e9:.2f} GB used, {GlobalCounters.global_ops*1e-9/(cl-st):9.2f} GFLOPS, {GlobalCounters.global_ops*1e-9:9.2f} GOPS")
      st = cl
      i += 1

-  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-    min_time = min(step_times)
-    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
-
  # verify eval acc
  if target := getenv("TARGET_EVAL_ACC_PCT", 0.0):
    if eval_acc_pct >= target:
@@ -478,7 +478,7 @@ After you are done speaking, output [EOS]. You are not Chad.
      with Profiling(enabled=args.profile):
        with Timing("total ", enabled=args.timing, on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
          with WallTimeEvent(BenchEvent.STEP):
-            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
                        f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
                        (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=args.timing):
              tok_tensor = llama.model(next_tok, start_pos, args.temperature)
@@ -1,6 +1,8 @@
 from pathlib import Path
 from typing import List
 import json, argparse, random, time, os
+import tiktoken
+from tiktoken.load import load_tiktoken_bpe
 from extra.models.llama import Transformer, convert_from_huggingface, convert_from_gguf, fix_bf16
 from tinygrad.nn.state import safe_load, torch_load, load_state_dict, get_parameters, gguf_load
 from tinygrad import Tensor, dtypes, nn, Context, Device, GlobalCounters
@@ -10,8 +12,6 @@ from extra.bench_log import BenchEvent, WallTimeEvent
 class Tokenizer:
  pat_str = r"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"
  def __init__(self, model_path: str):
-    import tiktoken
-    from tiktoken.load import load_tiktoken_bpe
    mergeable_ranks = load_tiktoken_bpe(model_path)
    self.num_base_tokens = len(mergeable_ranks)
    special_tokens = [
@@ -145,41 +145,6 @@ def NF4Linear(block_size):
      return new_state_dict
  return _NF4Linear

-def quantize_to_fp8(x: Tensor, dtype=dtypes.fp8e4m3):
-  fp8_min = -448.0 if dtype == dtypes.fp8e4m3 else -57344.0
-  fp8_max = 448.0 if dtype == dtypes.fp8e4m3 else 57344.0
-  scale = fp8_max / x.abs().max()
-  x_scl_sat = (x * scale).clamp(fp8_min, fp8_max)
-  return x_scl_sat.cast(dtype), scale.float().reciprocal()
-
-class FP8Linear:
-  def __init__(self, in_features, out_features, bias=True):
-    self.weight = Tensor.empty(out_features, in_features, dtype=dtypes.fp8e4m3)
-    self.bias = Tensor.empty(out_features, dtype=dtypes.float16) if bias else None
-    self.weight_scale = Tensor.empty((), dtype=dtypes.float16)
-
-  def __call__(self, x:Tensor):
-    y = x.dot(self.weight.T.cast(dtypes.float32)) * self.weight_scale
-    if self.bias is not None: y = y + self.bias.cast(y.dtype)
-    return y.cast(x.dtype)
-
-  @staticmethod
-  def quantize(tensors, device, scale_dtype=dtypes.float16, quantize_embeds=False):
-    assert not quantize_embeds
-    new_tensors = {}
-    for name,v in tensors.items():
-      if "feed_forward" in name or "attention.w" in name:
-        assert "weight" in name, name
-        fp8_weight, scale = quantize_to_fp8(v)
-        new_tensors[name] = fp8_weight
-        new_tensors[name.replace('weight', 'weight_scale')] = scale.cast(scale_dtype)
-        if isinstance(device, tuple):
-          new_tensors[name].shard_(device, axis=-1)
-          new_tensors[name.replace('weight', 'weight_scale')].shard_(device, axis=None)
-      else:
-        new_tensors[name] = v
-    return new_tensors
-
 MODEL_PARAMS = {
  "1B": {
    "args": {"dim": 2048, "n_heads": 32, "n_kv_heads": 8, "n_layers": 16, "norm_eps": 1e-5, "rope_theta": 500000, "vocab_size": 128256, "hidden_dim": 8192},
@@ -202,7 +167,6 @@ def build_transformer(model_path: Path, model_size="8B", quantize=None, scale_dt
  # build model
  if quantize == "int8": linear, embedding, quantize_embeds = Int8Linear, Int8Embedding, True
  elif quantize == "nf4": linear, embedding, quantize_embeds = NF4Linear(64), nn.Embedding, False
-  elif quantize == "fp8": linear, embedding, quantize_embeds = FP8Linear, nn.Embedding, False
  else: linear, embedding, quantize_embeds = nn.Linear, nn.Embedding, False
  model = Transformer(**MODEL_PARAMS[model_size]["args"], linear=linear, embedding=embedding, max_context=max_context, jit=True)

@@ -278,7 +242,7 @@ if __name__ == "__main__":
  parser.add_argument("--model", type=Path, help="Model path")
  parser.add_argument("--size", choices=["1B", "8B", "70B", "405B"], default="1B", help="Model size")
  parser.add_argument("--shard", type=int, default=1, help="Shard the model across multiple devices")
-  parser.add_argument("--quantize", choices=["int8", "nf4", "float16", "fp8"], help="Quantization method")
+  parser.add_argument("--quantize", choices=["int8", "nf4", "float16"], help="Quantization method")
  parser.add_argument("--no_api", action="store_true", help="Disable the api and run a cli test interface")
  parser.add_argument("--host", type=str, default="0.0.0.0", help="Web server bind address")
  parser.add_argument("--port", type=int, default=7776, help="Web server port")
@@ -477,7 +441,7 @@ if __name__ == "__main__":
      with Profiling(enabled=args.profile):
        with Timing("total ", on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
          with WallTimeEvent(BenchEvent.STEP):
-            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
                        f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
                        (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None):
              tok = model(Tensor([[last_tok]], device=device), start_pos, TEMPERATURE, TOP_K, TOP_P, ALPHA_F, ALPHA_P)
@@ -515,7 +479,7 @@ if __name__ == "__main__":
        st = GlobalCounters.time_sum_s
        with Profiling(enabled=args.profile):
          with Timing("total ", enabled=args.timing, on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
-            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "")+
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "")+
                        f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB"+
                        (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=args.timing):

@@ -279,15 +279,9 @@ def generate(model, tokenizer, prompt: str, n_tokens_to_gen: int = 10, temp: boo
  # Loading in the prompt tokens
  logits = model.forward(Tensor([tks]))[:, -1, :]
  for _ in tqdm(range(n_tokens_to_gen), desc="Speed Gen"):
+    # TODO: topk
    if sample:
-      scaled_logits = logits / temp
-      if top_k is not None:
-        topk_values, topk_indices = scaled_logits.topk(top_k)
-        filtered_logits = Tensor.full_like(scaled_logits, -float("inf"))
-        filtered_logits = filtered_logits.scatter(dim=-1, index=topk_indices, src=topk_values)
-        tok_Tens = filtered_logits.softmax().multinomial()
-      else:
-        tok_Tens = scaled_logits.softmax().multinomial()
+      tok_Tens = (logits/temp).softmax().multinomial()
    else:
      tok_Tens = logits.argmax(axis=-1).unsqueeze(0)
    tok = tok_Tens.item()
@@ -304,7 +298,6 @@ if __name__ == "__main__":
  parser.add_argument("--size", type=str, default="370m",
                      help=f"Size of model to use [{', '.join([k for k in MODELS.keys()])}]")
  parser.add_argument("--n_tokens", type=int, default=10, help="Number of tokens to generate")
-  parser.add_argument("--top_k", type=int, help="Limit sampling to the top k most likely tokens")
  parser.add_argument("--sample", dest="sample", action="store_true", help="Sample flag")
  parser.add_argument("--temp", type=float, default=1.0, help="Sampling temp has to be <=1.0")
  args = parser.parse_args()
@@ -315,9 +308,8 @@ if __name__ == "__main__":
  num_toks = args.n_tokens
  sample = args.sample
  temp = args.temp
-  top_k = args.top_k
  s = time.time()
-  tinyoutput = generate(model, tokenizer, prompt, n_tokens_to_gen=num_toks, sample=sample, temp=temp, top_k=top_k)
+  tinyoutput = generate(model, tokenizer, prompt, n_tokens_to_gen=num_toks, sample=sample, temp=temp)
  print(tinyoutput)
  print('TIME: ', time.time() - s)
  TORCHOUTPUT = "Why is gravity \nso important?\nBecause it's the only"
@@ -0,0 +1,299 @@
+from extra.models.mask_rcnn import MaskRCNN
+from extra.models.resnet import ResNet
+from extra.models.mask_rcnn import BoxList
+from torch.nn import functional as F
+from torchvision import transforms as T
+from torchvision.transforms import functional as Ft
+import random
+from tinygrad.tensor import Tensor
+from PIL import Image
+import numpy as np
+import torch
+import argparse
+import cv2
+
+
+class Resize:
+  def __init__(self, min_size, max_size):
+    if not isinstance(min_size, (list, tuple)):
+      min_size = (min_size,)
+    self.min_size = min_size
+    self.max_size = max_size
+
+  # modified from torchvision to add support for max size
+  def get_size(self, image_size):
+    w, h = image_size
+    size = random.choice(self.min_size)
+    max_size = self.max_size
+    if max_size is not None:
+      min_original_size = float(min((w, h)))
+      max_original_size = float(max((w, h)))
+      if max_original_size / min_original_size * size > max_size:
+        size = int(round(max_size * min_original_size / max_original_size))
+
+      if (w <= h and w == size) or (h <= w and h == size):
+        return (h, w)
+
+      if w < h:
+        ow = size
+        oh = int(size * h / w)
+      else:
+        oh = size
+        ow = int(size * w / h)
+
+      return (oh, ow)
+
+  def __call__(self, image):
+    size = self.get_size(image.size)
+    image = Ft.resize(image, size)
+    return image
+
+
+class Normalize:
+  def __init__(self, mean, std, to_bgr255=True):
+    self.mean = mean
+    self.std = std
+    self.to_bgr255 = to_bgr255
+
+  def __call__(self, image):
+    if self.to_bgr255:
+      image = image[[2, 1, 0]] * 255
+    else:
+      image = image[[0, 1, 2]] * 255
+    image = Ft.normalize(image, mean=self.mean, std=self.std)
+    return image
+
+transforms = lambda size_scale: T.Compose(
+  [
+    Resize(int(800*size_scale), int(1333*size_scale)),
+    T.ToTensor(),
+    Normalize(
+      mean=[102.9801, 115.9465, 122.7717], std=[1., 1., 1.], to_bgr255=True
+    ),
+  ]
+)
+
+def expand_boxes(boxes, scale):
+  w_half = (boxes[:, 2] - boxes[:, 0]) * .5
+  h_half = (boxes[:, 3] - boxes[:, 1]) * .5
+  x_c = (boxes[:, 2] + boxes[:, 0]) * .5
+  y_c = (boxes[:, 3] + boxes[:, 1]) * .5
+
+  w_half *= scale
+  h_half *= scale
+
+  boxes_exp = torch.zeros_like(boxes)
+  boxes_exp[:, 0] = x_c - w_half
+  boxes_exp[:, 2] = x_c + w_half
+  boxes_exp[:, 1] = y_c - h_half
+  boxes_exp[:, 3] = y_c + h_half
+  return boxes_exp
+
+
+def expand_masks(mask, padding):
+  N = mask.shape[0]
+  M = mask.shape[-1]
+  pad2 = 2 * padding
+  scale = float(M + pad2) / M
+  padded_mask = mask.new_zeros((N, 1, M + pad2, M + pad2))
+  padded_mask[:, :, padding:-padding, padding:-padding] = mask
+  return padded_mask, scale
+
+
+def paste_mask_in_image(mask, box, im_h, im_w, thresh=0.5, padding=1):
+  # TODO: remove torch
+  mask = torch.tensor(mask.numpy())
+  box = torch.tensor(box.numpy())
+  padded_mask, scale = expand_masks(mask[None], padding=padding)
+  mask = padded_mask[0, 0]
+  box = expand_boxes(box[None], scale)[0]
+  box = box.to(dtype=torch.int32)
+
+  TO_REMOVE = 1
+  w = int(box[2] - box[0] + TO_REMOVE)
+  h = int(box[3] - box[1] + TO_REMOVE)
+  w = max(w, 1)
+  h = max(h, 1)
+
+  mask = mask.expand((1, 1, -1, -1))
+
+  mask = mask.to(torch.float32)
+  mask = F.interpolate(mask, size=(h, w), mode='bilinear', align_corners=False)
+  mask = mask[0][0]
+
+  if thresh >= 0:
+    mask = mask > thresh
+  else:
+    mask = (mask * 255).to(torch.uint8)
+
+  im_mask = torch.zeros((im_h, im_w), dtype=torch.uint8)
+  x_0 = max(box[0], 0)
+  x_1 = min(box[2] + 1, im_w)
+  y_0 = max(box[1], 0)
+  y_1 = min(box[3] + 1, im_h)
+
+  im_mask[y_0:y_1, x_0:x_1] = mask[
+                              (y_0 - box[1]): (y_1 - box[1]), (x_0 - box[0]): (x_1 - box[0])
+                              ]
+  return im_mask
+
+
+class Masker:
+  def __init__(self, threshold=0.5, padding=1):
+    self.threshold = threshold
+    self.padding = padding
+
+  def forward_single_image(self, masks, boxes):
+    boxes = boxes.convert("xyxy")
+    im_w, im_h = boxes.size
+    res = [
+      paste_mask_in_image(mask[0], box, im_h, im_w, self.threshold, self.padding)
+      for mask, box in zip(masks, boxes.bbox)
+    ]
+    if len(res) > 0:
+      res = torch.stack(*res, dim=0)[:, None]
+    else:
+      res = masks.new_empty((0, 1, masks.shape[-2], masks.shape[-1]))
+    return Tensor(res.numpy())
+
+  def __call__(self, masks, boxes):
+    if isinstance(boxes, BoxList):
+      boxes = [boxes]
+
+    results = []
+    for mask, box in zip(masks, boxes):
+      result = self.forward_single_image(mask, box)
+      results.append(result)
+    return results
+
+
+masker = Masker(threshold=0.5, padding=1)
+
+def select_top_predictions(predictions, confidence_threshold=0.9):
+  scores = predictions.get_field("scores").numpy()
+  keep = [idx for idx, score in enumerate(scores) if score > confidence_threshold]
+  return predictions[keep]
+
+def compute_prediction(original_image, model, confidence_threshold, size_scale=1.0):
+  image = transforms(size_scale)(original_image).numpy()
+  image = Tensor(image, requires_grad=False)
+  predictions = model(image)
+  prediction = predictions[0]
+  prediction = select_top_predictions(prediction, confidence_threshold)
+  width, height = original_image.size
+  prediction = prediction.resize((width, height))
+
+  if prediction.has_field("mask"):
+    masks = prediction.get_field("mask")
+    masks = masker([masks], [prediction])[0]
+    prediction.add_field("mask", masks)
+  return prediction
+
+def compute_prediction_batched(batch, model, size_scale=1.0):
+  imgs = []
+  for img in batch:
+    imgs.append(transforms(size_scale)(img).numpy())
+  image = [Tensor(image, requires_grad=False) for image in imgs]
+  predictions = model(image)
+  del image
+  return predictions
+
+palette = np.array([2 ** 25 - 1, 2 ** 15 - 1, 2 ** 21 - 1])
+
+def findContours(*args, **kwargs):
+  if cv2.__version__.startswith('4'):
+    contours, hierarchy = cv2.findContours(*args, **kwargs)
+  elif cv2.__version__.startswith('3'):
+    _, contours, hierarchy = cv2.findContours(*args, **kwargs)
+  return contours, hierarchy
+
+def compute_colors_for_labels(labels):
+  l = labels[:, None]
+  colors = l * palette
+  colors = (colors % 255).astype("uint8")
+  return colors
+
+def overlay_mask(image, predictions):
+  image = np.asarray(image)
+  masks = predictions.get_field("mask").numpy()
+  labels = predictions.get_field("labels").numpy()
+
+  colors = compute_colors_for_labels(labels).tolist()
+
+  for mask, color in zip(masks, colors):
+    thresh = mask[0, :, :, None]
+    contours, hierarchy = findContours(
+        thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE
+    )
+    image = cv2.drawContours(image, contours, -1, color, 3)
+
+  composite = image
+
+  return composite
+
+CATEGORIES = [
+    "__background", "person", "bicycle", "car", "motorcycle", "airplane", "bus", "train", "truck", "boat", "traffic light",
+    "fire hydrant", "stop sign", "parking meter", "bench", "bird", "cat", "dog", "horse", "sheep", "cow", "elephant",
+    "bear", "zebra", "giraffe", "backpack", "umbrella", "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard",
+    "sports ball", "kite", "baseball bat", "baseball glove", "skateboard", "surfboard", "tennis racket", "bottle",
+    "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple", "sandwich", "orange", "broccoli",
+    "carrot", "hot dog", "pizza", "donut", "cake", "chair", "couch", "potted plant", "bed", "dining table",
+    "toilet", "tv", "laptop", "mouse", "remote", "keyboard", "cell phone", "microwave", "oven", "toaster",
+    "sink", "refrigerator", "book", "clock", "vase", "scissors", "teddy bear", "hair drier", "toothbrush",
+]
+
+def overlay_boxes(image, predictions):
+  labels = predictions.get_field("labels").numpy()
+  boxes = predictions.bbox
+  image = np.asarray(image)
+  colors = compute_colors_for_labels(labels).tolist()
+
+  for box, color in zip(boxes, colors):
+    box = torch.tensor(box.numpy())
+    box = box.to(torch.int64)
+    top_left, bottom_right = box[:2].tolist(), box[2:].tolist()
+    image = cv2.rectangle(
+        image, tuple(top_left), tuple(bottom_right), tuple(color), 1
+    )
+
+  return image
+
+def overlay_class_names(image, predictions):
+  scores = predictions.get_field("scores").numpy().tolist()
+  labels = predictions.get_field("labels").numpy().tolist()
+  labels = [CATEGORIES[int(i)] for i in labels]
+  boxes = predictions.bbox.numpy()
+  image = np.asarray(image)
+  template = "{}: {:.2f}"
+  for box, score, label in zip(boxes, scores, labels):
+    x, y = box[:2]
+    s = template.format(label, score)
+    x, y = int(x), int(y)
+    cv2.putText(
+        image, s, (x, y), cv2.FONT_HERSHEY_SIMPLEX, .5, (255, 255, 255), 1
+    )
+
+  return image
+
+
+if __name__ == '__main__':
+  parser = argparse.ArgumentParser(description='Run MaskRCNN', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+  parser.add_argument('--image', type=str, help="Path of the image to run")
+  parser.add_argument('--threshold', type=float, default=0.7, help="Detector threshold")
+  parser.add_argument('--size_scale', type=float, default=1.0, help="Image resize multiplier")
+  parser.add_argument('--out', type=str, default="/tmp/rendered.png", help="Output filename")
+  args = parser.parse_args()
+
+  resnet = ResNet(50, num_classes=None, stride_in_1x1=True)
+  model_tiny = MaskRCNN(resnet)
+  model_tiny.load_from_pretrained()
+  img = Image.open(args.image)
+  top_result_tiny = compute_prediction(img, model_tiny, confidence_threshold=args.threshold, size_scale=args.size_scale)
+  bbox_image = overlay_boxes(img, top_result_tiny)
+  mask_image = overlay_mask(bbox_image, top_result_tiny)
+  final_image = overlay_class_names(mask_image, top_result_tiny)
+
+  im = Image.fromarray(final_image)
+  print(f"saving {args.out}")
+  im.save(args.out)
+  im.show()
@@ -511,33 +511,6 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
      # happens with BENCHMARK set
      pass

-# stable diffusion callbacks to match mlperf ref; declared here because they're pickled
-def filter_dataset(sample:dict): return {k:v for k,v in sample.items() if k in {'npy', 'txt'}}
-def collate(batch:list[dict]):
-  ret = {"npy": [], "txt": [], "__key__": []}
-  for sample in batch:
-    for k,v in sample.items():
-      ret[k].append(v)
-  return ret
-def collate_fn(batch): return batch
-
-# Reference (code): https://github.com/mlcommons/training/blob/2f4a93fb4888180755a8ef55f4b977ef8f60a89e/stable_diffusion/ldm/data/webdatasets.py, Line 55
-# Reference (params): https://github.com/mlcommons/training/blob/ab4ae1ca718d7fe62c369710a316dff18768d04b/stable_diffusion/configs/train_01x08x08.yaml, Line 107
-def batch_load_train_stable_diffusion(urls:str, BS:int):
-  import webdataset
-  dataset = webdataset.WebDataset(urls=urls, resampled=True, cache_size=-1, cache_dir=None)
-  dataset = dataset.shuffle(size=1000)
-  dataset = dataset.decode()
-  dataset = dataset.map(filter_dataset)
-  dataset = dataset.batched(BS, partial=False, collation_fn=collate)
-  dataset = webdataset.WebLoader(dataset, batch_size=None, shuffle=False, num_workers=1, persistent_workers=True, collate_fn=collate_fn)
-
-  for x in dataset:
-    assert isinstance(x, dict) and all(isinstance(k, str) for k in x.keys()) and all(isinstance(v, list) for v in x.values())
-    assert all(isinstance(moment_mean_logvar, np.ndarray) and moment_mean_logvar.shape==(1,8,64,64) for moment_mean_logvar in x["npy"])
-    assert all(isinstance(caption, str) for caption in x["txt"])
-    yield x
-
 # llama3

 class BinIdxDataset:
@@ -223,13 +223,13 @@ def get_mlperf_bert_model():

 def get_fake_data_bert(BS:int):
  return {
-    "input_ids": Tensor.zeros((BS, 512), dtype=dtypes.int32, device="CPU").contiguous(),
-    "input_mask": Tensor.zeros((BS, 512), dtype=dtypes.int32, device="CPU").contiguous(),
-    "segment_ids": Tensor.zeros((BS, 512), dtype=dtypes.int32, device="CPU").contiguous(),
-    "masked_lm_positions": Tensor.zeros((BS, 76), dtype=dtypes.int32, device="CPU").contiguous(),
-    "masked_lm_ids": Tensor.zeros((BS, 76), dtype=dtypes.int32, device="CPU").contiguous(),
-    "masked_lm_weights": Tensor.zeros((BS, 76), dtype=dtypes.float32, device="CPU").contiguous(),
-    "next_sentence_labels": Tensor.zeros((BS, 1), dtype=dtypes.int32, device="CPU").contiguous(),
+    "input_ids": Tensor.empty((BS, 512), dtype=dtypes.int32, device="CPU"),
+    "input_mask": Tensor.empty((BS, 512), dtype=dtypes.int32, device="CPU"),
+    "segment_ids": Tensor.empty((BS, 512), dtype=dtypes.int32, device="CPU"),
+    "masked_lm_positions": Tensor.empty((BS, 76), dtype=dtypes.int32, device="CPU"),
+    "masked_lm_ids": Tensor.empty((BS, 76), dtype=dtypes.int32, device="CPU"),
+    "masked_lm_weights": Tensor.empty((BS, 76), dtype=dtypes.float32, device="CPU"),
+    "next_sentence_labels": Tensor.empty((BS, 1), dtype=dtypes.int32, device="CPU"),
  }

 def find_matches(match_quality_matrix:np.ndarray, high_threshold:float=0.5, low_threshold:float=0.4, allow_low_quality_matches:bool=False) -> np.ndarray:
@@ -2,9 +2,7 @@ import math
 from typing import Union

 from tinygrad import Tensor, nn, dtypes
-from tinygrad.helpers import prod, argfix, Context
-from tinygrad.nn.state import get_parameters
-from extra.models.unet import UNetModel
+from tinygrad.helpers import prod, argfix

 # rejection sampling truncated randn
 def rand_truncn(*shape, dtype=None, truncstds=2, **kwargs) -> Tensor:
@@ -19,10 +17,6 @@ def he_normal(*shape, a: float = 0.00, **kwargs) -> Tensor:
  std = math.sqrt(2.0 / (1 + a ** 2)) / math.sqrt(prod(argfix(*shape)[1:])) / 0.87962566103423978
  return std * rand_truncn(*shape, **kwargs)

-# Stable Diffusion v2 training uses default torch gelu, which doesn't use tanh approximation
-def gelu_erf(x:Tensor) -> Tensor:
-  return 0.5 * x * (1.0 + (x / 1.4142135623730951).erf())
-
 class Conv2dHeNormal(nn.Conv2d):
  def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
    super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
@@ -59,7 +53,9 @@ class EmbeddingBert(nn.Embedding):
    arange_shp, weight_shp, big_shp = (1, 1, self.vocab_sz, 1), (1, 1, self.vocab_sz, self.embed_sz), idx.shape+(self.vocab_sz, self.embed_sz,)
    if not hasattr(self, 'arange'): self.arange = Tensor.arange(self.vocab_sz, requires_grad=False, device=self.weight.device).reshape(arange_shp)
    arange, idx, vals = self.arange.expand(big_shp), idx.reshape(idx.shape+(1, 1,)).expand(big_shp), self.weight.cast(dtypes.default_float).reshape(weight_shp).expand(big_shp)
-    return (arange == idx).where(vals, 0).sum(2, dtype=vals.dtype)
+    # TODO: contiguous() here because the embedding dropout creates different asts on each device, and search becomes very slow.
+    # Should fix with fixing random ast on multi device, and fuse arange to make embedding fast.
+    return (arange == idx).mul(vals).sum(2, dtype=vals.dtype).contiguous()

 class LayerNormBert:
  def __init__(self, normalized_shape:Union[int, tuple[int, ...]], eps:float=1e-12, elementwise_affine:bool=True):
@@ -131,59 +127,3 @@ class Conv2dRetinaNet(nn.Conv2d):
  def __call__(self, x:Tensor) -> Tensor:
    return x.conv2d(self.weight.cast(dtypes.default_float), self.bias.cast(dtypes.default_float) if self.bias is not None else None,
                    groups=self.groups, stride=self.stride, dilation=self.dilation, padding=self.padding)
-
-# copy torch AMP: isolate mixed precision to just the below autocast ops, instead of using dtypes.default_float which affects all new Tensors
-class AutocastLinear(nn.Linear):
-  cast_dtype=dtypes.bfloat16 # enable monkeypatching of the mixed precision dtype
-  def __call__(self, x:Tensor) -> Tensor:
-    dtype = type(self).cast_dtype
-    return x.cast(dtype).linear(self.weight.cast(dtype).transpose(), self.bias.cast(dtype) if self.bias is not None else None)
-
-class AutocastConv2d(nn.Conv2d):
-  cast_dtype=dtypes.bfloat16
-  def __call__(self, x:Tensor) -> Tensor:
-    dtype = type(self).cast_dtype
-    return x.cast(dtype).conv2d(self.weight.cast(dtype), self.bias.cast(dtype), self.groups, self.stride, self.dilation, self.padding)
-
-# copy torch AMP: upcast to float32 before GroupNorm and LayerNorm
-class AutocastGroupNorm(nn.GroupNorm):
-  def __call__(self, x:Tensor) -> Tensor:
-    return super().__call__(x.cast(dtypes.float32))
-
-class AutocastLayerNorm(nn.LayerNorm):
-  def __call__(self, x:Tensor) -> Tensor:
-    return super().__call__(x.cast(dtypes.float32))
-
-def zero_module(module):
-  for p in get_parameters(module): p.assign(Tensor.zeros_like(p).contiguous())
-
-# Stable Diffusion mlperf reference doesn't call scaled_dot_product_attention
-# copy torch AMP: upcast to float32 before softmax on CUDA
-def attn_f32_softmax(q:Tensor, k:Tensor, v:Tensor) -> Tensor:
-  return (q.matmul(k.transpose(-2,-1), dtype=dtypes.float32) / math.sqrt(q.shape[-1])).softmax(-1).cast(q.dtype) @ v
-
-def init_stable_diffusion(version:str, pretrained:str, devices:list[str]):
-  from examples.stable_diffusion import StableDiffusion
-  from tinygrad.nn.state import safe_load, safe_save, load_state_dict, get_state_dict
-  from tempfile import TemporaryDirectory
-  model = StableDiffusion(version=version, pretrained=pretrained)
-  unet:UNetModel = model.model.diffusion_model
-
-  # this prevents extra consumption of memory, enabling much larger BS
-  Tensor.realize(*get_parameters(unet))
-  with TemporaryDirectory(prefix="unet_init") as tmp:
-    safe_save(get_state_dict(unet), init_fn:=f"{tmp}/init_model.safetensors")
-    load_state_dict(unet, safe_load(init_fn))
-
-  sqrt_alphas_cumprod = model.alphas_cumprod.sqrt().realize()
-  sqrt_one_minus_alphas_cumprod = (1 - model.alphas_cumprod).sqrt().realize()
-
-  if len(devices) > 1:
-    to_move = [sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod]
-    if version == "v2-mlperf-train": to_move += get_parameters(unet) + get_parameters(model.cond_stage_model)
-    for p in to_move:
-      p.to_(devices)
-    with Context(BEAM=0):
-      Tensor.realize(*to_move)
-
-  return model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod
@@ -1,9 +1,8 @@
 import math
-from tinygrad import dtypes, Tensor
+from tinygrad import dtypes
 from tinygrad.nn.optim import Optimizer

 from extra.lr_scheduler import LR_Scheduler
-from typing import Callable

 # https://github.com/mlcommons/training/blob/e237206991d10449d9675d95606459a3cb6c21ad/image_classification/tensorflow2/lars_util.py
 class PolynomialDecayWithWarmup(LR_Scheduler):
@@ -37,24 +36,4 @@ class CosineAnnealingLRWithWarmup(LR_Scheduler):
  def get_lr(self):
    warmup_lr = ((self.epoch_counter+1) / self.warmup_steps) * self.base_lr
    decay_lr = self.end_lr + 0.5 * (self.base_lr-self.end_lr) * (1 + (((self.epoch_counter+1-self.warmup_steps)/self.decay_steps) * math.pi).cos())
-    return (self.epoch_counter < self.warmup_steps).where(warmup_lr, decay_lr).cast(self.optimizer.lr.dtype)
-
-# Reference: https://github.com/mlcommons/training/blob/64b14a9abc74e08779a175abca7d291f8c957632/stable_diffusion/ldm/lr_scheduler.py, Lines 36-97
-class LambdaLinearScheduler:
-  def __init__(self, warm_up_steps:int, f_min:float, f_max:float, f_start:float, cycle_lengths:int):
-    self.lr_warm_up_steps, self.f_min, self.f_max, self.f_start, self.cycle_lengths = warm_up_steps, f_min, f_max, f_start, cycle_lengths
-
-  def schedule(self, n:Tensor) -> Tensor:
-    warm_up = (n < self.lr_warm_up_steps)
-    f_warm_up = (self.f_max - self.f_start) / self.lr_warm_up_steps * n + self.f_start
-    return warm_up.where(f_warm_up, self.f_min + (self.f_max - self.f_min) * (self.cycle_lengths - n) / (self.cycle_lengths))
-
-# based on torch.optim.lr_scheduler.LambdaLR
-class LambdaLR(LR_Scheduler):
-  def __init__(self, optimizer:Optimizer, base_lr:Tensor, lr_lambda:Callable):
-    super().__init__(optimizer)
-    self.base_lr, self.lr_lambda = base_lr, lr_lambda
-    self.step()
-
-  def get_lr(self):
-    return self.base_lr * self.lr_lambda(self.epoch_counter - 1)
+    return (self.epoch_counter < self.warmup_steps).where(warmup_lr, decay_lr).cast(self.optimizer.lr.dtype)
@@ -1,10 +1,10 @@
-import time, math, os
+import time, math
 start = time.perf_counter()
 from pathlib import Path
 import numpy as np
 from tinygrad import Tensor, Device, dtypes, GlobalCounters, TinyJit
 from tinygrad.nn.state import get_parameters, load_state_dict, safe_load
-from tinygrad.helpers import getenv, Context, prod
+from tinygrad.helpers import getenv
 from extra.bench_log import BenchEvent, WallTimeEvent
 def tlog(x): print(f"{x:25s}  @ {time.perf_counter()-start:5.2f}s")

@@ -204,6 +204,43 @@ def eval_bert():

    st = time.perf_counter()

+def eval_mrcnn():
+  from tqdm import tqdm
+  from extra.models.mask_rcnn import MaskRCNN
+  from extra.models.resnet import ResNet
+  from extra.datasets.coco import BASEDIR, images, convert_prediction_to_coco_bbox, convert_prediction_to_coco_mask, accumulate_predictions_for_coco, evaluate_predictions_on_coco, iterate
+  from examples.mask_rcnn import compute_prediction_batched, Image
+  mdl = MaskRCNN(ResNet(50, num_classes=None, stride_in_1x1=True))
+  mdl.load_from_pretrained()
+
+  bbox_output = '/tmp/results_bbox.json'
+  mask_output = '/tmp/results_mask.json'
+
+  accumulate_predictions_for_coco([], bbox_output, rm=True)
+  accumulate_predictions_for_coco([], mask_output, rm=True)
+
+  #TODO: bs > 1 not as accurate
+  bs = 1
+
+  for batch in tqdm(iterate(images, bs=bs), total=len(images)//bs):
+    batch_imgs = []
+    for image_row in batch:
+      image_name = image_row['file_name']
+      img = Image.open(BASEDIR/f'val2017/{image_name}').convert("RGB")
+      batch_imgs.append(img)
+    batch_result = compute_prediction_batched(batch_imgs, mdl)
+    for image_row, result in zip(batch, batch_result):
+      image_name = image_row['file_name']
+      box_pred = convert_prediction_to_coco_bbox(image_name, result)
+      mask_pred = convert_prediction_to_coco_mask(image_name, result)
+      accumulate_predictions_for_coco(box_pred, bbox_output)
+      accumulate_predictions_for_coco(mask_pred, mask_output)
+    del batch_imgs
+    del batch_result
+
+  evaluate_predictions_on_coco(bbox_output, iou_type='bbox')
+  evaluate_predictions_on_coco(mask_output, iou_type='segm')
+
 def eval_llama3():
  from extra.models.llama import Transformer
  from examples.llama3 import MODEL_PARAMS, load, convert_from_huggingface
@@ -250,261 +287,11 @@ def eval_llama3():
  log_perplexity = np.mean(losses)
  print(f"Log Perplexity: {log_perplexity}")

-# NOTE: BEAM hangs on 8xmi300x with DECODE_BS=384 in final realize below; function is declared here for external testing
-@TinyJit
-def vae_decode(x:Tensor, vae, disable_beam=False) -> Tensor:
-  from examples.stable_diffusion import AutoencoderKL
-  assert isinstance(vae, AutoencoderKL)
-  x = vae.post_quant_conv(1./0.18215 * x)
-
-  x = vae.decoder.conv_in(x)
-  x = vae.decoder.mid(x)
-  for i, l in enumerate(vae.decoder.up[::-1]):
-    print("decode", x.shape)
-    for b in l['block']: x = b(x)
-    if 'upsample' in l:
-      bs,c,py,px = x.shape
-      x = x.reshape(bs, c, py, 1, px, 1).expand(bs, c, py, 2, px, 2).reshape(bs, c, py*2, px*2)
-      x = l['upsample']['conv'](x)
-    if i == len(vae.decoder.up) - 1 and disable_beam:
-      with Context(BEAM=0): x.realize()
-    else: x.realize()
-  x = vae.decoder.conv_out(vae.decoder.norm_out(x).swish())
-
-  x = ((x + 1.0) / 2.0).clip(0.0, 1.0)
-  return x
-
-def eval_stable_diffusion():
-  import csv, PIL, sys
-  from tqdm import tqdm
-  from examples.mlperf.initializers import init_stable_diffusion, gelu_erf
-  from examples.stable_diffusion import AutoencoderKL
-  from extra.models.unet import UNetModel
-  from tinygrad.nn.state import load_state_dict, torch_load
-  from tinygrad.helpers import BEAM
-  from extra.models import clip
-  from extra.models.clip import FrozenOpenClipEmbedder
-  from extra.models.clip import OpenClipEncoder
-  from extra.models.inception import FidInceptionV3
-
-  config = {}
-  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
-  for x in GPUS: Device[x]
-  print(f"running eval on {GPUS}")
-  seed               = config["seed"]                   = getenv("SEED", 12345)
-  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
-  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
-  CONTEXT_BS         = config["CONTEXT_BS"]             = getenv("CONTEXT_BS", 1 * len(GPUS))
-  DENOISE_BS         = config["DENOISE_BS"]             = getenv("DENOISE_BS", 1 * len(GPUS))
-  DECODE_BS          = config["DECODE_BS"]              = getenv("DECODE_BS", 1 * len(GPUS))
-  INCEPTION_BS       = config["INCEPTION_BS"]           = getenv("INCEPTION_BS", 1 * len(GPUS))
-  CLIP_BS            = config["CLIP_BS"]                = getenv("CLIP_BS", 1 * len(GPUS))
-  EVAL_CKPT_DIR      = config["EVAL_CKPT_DIR"]          = getenv("EVAL_CKPT_DIR", "")
-  STOP_IF_CONVERGED  = config["STOP_IF_CONVERGED"]      = getenv("STOP_IF_CONVERGED", 0)
-
-  if (WANDB := getenv("WANDB", "")):
-    import wandb
-    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
-
-  assert EVAL_CKPT_DIR != "", "provide a directory with checkpoints to be evaluated"
-  print(f"running eval on checkpoints in {EVAL_CKPT_DIR}\nSEED={seed}")
-  eval_queue:list[tuple[int, Path]] = []
-  for p in Path(EVAL_CKPT_DIR).iterdir():
-    if p.name.endswith(".safetensors"):
-      ckpt_iteration = p.name.split(".safetensors")[0]
-      assert ckpt_iteration.isdigit(), f"invalid checkpoint name: {p.name}, expected <digits>.safetensors"
-      eval_queue.append((int(ckpt_iteration), p))
-  assert len(eval_queue), f'no files ending with ".safetensors" were found in {EVAL_CKPT_DIR}'
-  print(sorted(eval_queue, reverse=True))
-
-  Tensor.manual_seed(seed)  # seed for weight initialization
-  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-eval", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
-
-  # load prompts for generating images for validation; 2 MB of data total
-  with open(DATADIR / "coco2014" / "val2014_30k.tsv") as f:
-    reader = csv.DictReader(f, delimiter="\t")
-    eval_inputs:list[dict] = [{"image_id": int(row["image_id"]), "id": int(row["id"]), "caption": row["caption"]} for row in reader]
-  assert len(eval_inputs) == 30_000
-  # NOTE: the clip weights are the same between model.cond_stage_model and clip_encoder
-  eval_timesteps = list(reversed(range(1, 1000, 20)))
-
-  original_device, Device.DEFAULT = Device.DEFAULT, "CPU"
-  # The choice of alphas_prev[0] = alphas_cumprod[0] seems arbitrary, but it's how the mlperf ref does it:
-  #   alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
-  eval_alphas_prev = model.alphas_cumprod[0:1].cat(model.alphas_cumprod[list(range(1, 1000, 20))[:-1]]).to(GPUS).realize()
-  inception = FidInceptionV3().load_from_pretrained(CKPTDIR / "inception" / "pt_inception-2015-12-05-6726825d.pth")
-  vision_cfg = {'width': 1280, 'layers': 32, 'd_head': 80, 'image_size': 224, 'patch_size': 14}
-  text_cfg = {'width': 1024, 'n_heads': 16, 'layers': 24, 'vocab_size': 49408, 'ctx_length': 77}
-  clip.gelu = gelu_erf
-  clip_encoder = OpenClipEncoder(1024, text_cfg, vision_cfg)
-  loaded = torch_load(CKPTDIR / "clip" / "open_clip_pytorch_model.bin")
-  loaded.update({"attn_mask": clip_encoder.attn_mask, "mean": clip_encoder.mean, "std": clip_encoder.std})
-  load_state_dict(clip_encoder, loaded)
-  Device.DEFAULT=original_device
-
-  @TinyJit
-  def denoise_step(x:Tensor, x_x:Tensor, t_t:Tensor, uc_c:Tensor, sqrt_alphas_cumprod_t:Tensor, sqrt_one_minus_alphas_cumprod_t:Tensor,
-                    alpha_prev:Tensor, unet:UNetModel, GPUS) -> Tensor:
-    out_uncond, out = unet(x_x, t_t, uc_c).to("CPU").reshape(-1, 2, 4, 64, 64).chunk(2, dim=1)
-    out_uncond = out_uncond.squeeze(1).shard(GPUS,axis=0)
-    out = out.squeeze(1).shard(GPUS,axis=0)
-    v_t = out_uncond + 8.0 * (out - out_uncond)
-    e_t = sqrt_alphas_cumprod_t * v_t + sqrt_one_minus_alphas_cumprod_t * x
-    pred_x0 = sqrt_alphas_cumprod_t * x - sqrt_one_minus_alphas_cumprod_t * v_t
-    dir_xt = (1. - alpha_prev).sqrt() * e_t
-    x_prev = alpha_prev.sqrt() * pred_x0 + dir_xt
-    return x_prev.realize()
-
-  def shard_tensor(t:Tensor) -> Tensor: return t.shard(GPUS, axis=0) if len(GPUS) > 1 else t.to(GPUS[0])
-  def get_batch(whole:Tensor, i:int, bs:int) -> tuple[Tensor, int]:
-    batch = whole[i: i + bs].to("CPU")
-    if (unpadded_bs:=batch.shape[0]) < bs:
-      batch = batch.cat(batch[-1:].expand(bs - unpadded_bs, *batch[-1].shape))
-    return batch, unpadded_bs 
-
-  @Tensor.train(mode=False)
-  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
-    # It doesn't make sense to merge these jits, e.g. unet repeats 50 times in isolation; images fork to separate inception/clip
-    # We're generating and scoring 30,000 images per eval, and all the data can flow through one jit at a time
-    # To maximize throughput for each jit, we have only one model/jit on the GPU at a time, and pool outputs from each jit off-GPU
-    for model in (unet, first_stage, inception, clip):
-      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
-
-    uc_written = False
-    models = (cond_stage, unet, first_stage, inception, clip)
-    jits = (jit_context:=TinyJit(cond_stage.embed_tokens), denoise_step, vae_decode, jit_inception:=TinyJit(inception),
-            jit_clip:=TinyJit(clip.get_clip_score))
-    all_bs = (CONTEXT_BS, DENOISE_BS, DECODE_BS, INCEPTION_BS, CLIP_BS)
-    if (EVAL_SAMPLES:=getenv("EVAL_SAMPLES", 0)) and EVAL_SAMPLES > 0:
-      eval_inputs = eval_inputs[0:EVAL_SAMPLES]
-    output_shapes = [(ns:=len(eval_inputs),77), (ns,77,1024), (ns,4,64,64), (ns,3,512,512), (ns,2048), (ns,)]
-    # Writing progress to disk lets us resume eval if we crash
-    stages = ["tokens", "embeds", "latents", "imgs", "inception", "clip"]
-    disk_tensor_names, disk_tensor_shapes = stages + ["end", "uc"], output_shapes + [(6,), (1,77,1024)]
-    if not all(os.path.exists(f"{EVAL_CKPT_DIR}/{name}.bytes") for name in disk_tensor_names):
-      for name, shape in zip(disk_tensor_names, disk_tensor_shapes):
-        file = Path(f"{EVAL_CKPT_DIR}/{name}.bytes")
-        file.unlink(missing_ok=True)
-        with file.open("wb") as f: f.truncate(prod(shape) * 4)
-    progress = {name: Tensor.empty(*shape, device=f"disk:{EVAL_CKPT_DIR}/{name}.bytes", dtype=dtypes.int if name in {"tokens", "end"} else dtypes.float)
-                for name, shape in zip(disk_tensor_names, disk_tensor_shapes)}
-
-    def embed_tokens(tokens:Tensor) -> Tensor:
-      nonlocal uc_written
-      if not uc_written:
-        with Context(BEAM=0): progress["uc"].assign(cond_stage.embed_tokens(cond_stage.tokenize("").to(GPUS)).to("CPU").realize()).realize()
-        uc_written = True
-      return jit_context(shard_tensor(tokens))
-
-    def generate_latents(embeds:Tensor) -> Tensor:
-      uc_c = Tensor.stack(progress["uc"].to("CPU").expand(bs, 77, 1024), embeds, dim=1).reshape(-1, 77, 1024)
-      uc_c = shard_tensor(uc_c)
-      x = shard_tensor(Tensor.randn(bs,4,64,64))
-      for step_idx, timestep in enumerate(tqdm(eval_timesteps)):
-        reversed_idx = Tensor([50 - step_idx - 1], device=GPUS)
-        alpha_prev = eval_alphas_prev[reversed_idx]
-        ts = Tensor.full(bs, fill_value=timestep, dtype=dtypes.int, device="CPU")
-        ts_ts = shard_tensor(ts.cat(ts))
-        ts = shard_tensor(ts)
-        sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
-        sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[ts].reshape(bs, 1, 1, 1)
-        x_x = shard_tensor(Tensor.stack(x.to("CPU"), x.to("CPU"), dim=1).reshape(-1, 4, 64, 64))
-        x.assign(denoise_step(x, x_x, ts_ts, uc_c, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t, alpha_prev, unet, GPUS)).realize()
-      return x
-
-    def decode_latents(latents:Tensor) -> Tensor: return vae_decode(shard_tensor(latents), first_stage, disable_beam=True)
-    def generate_inception(imgs:Tensor) -> Tensor: return jit_inception(shard_tensor(imgs))[:,:,0,0]
-
-    def calc_clip_scores(batch:Tensor, batch_tokens:Tensor) -> Tensor:
-      # Tensor.interpolate does not yet support bicubic, so we use PIL
-      batch = (batch.to(GPUS[0]).permute(0,2,3,1) * 255).clip(0, 255).cast(dtypes.uint8).numpy()
-      batch = [np.array(PIL.Image.fromarray(batch[i]).resize((224,224), PIL.Image.BICUBIC)) for i in range(bs)]
-      batch = shard_tensor(Tensor(np.stack(batch, axis=0).transpose(0,3,1,2), device="CPU").realize())
-      batch = batch.cast(dtypes.float) / 255
-      batch = (batch - model.mean) / model.std
-      batch = jit_clip(shard_tensor(batch_tokens), batch)
-      return batch
-
-    callbacks = (embed_tokens, generate_latents, decode_latents, generate_inception, calc_clip_scores)
-
-    # save every forward pass output to disk; NOTE: this needs ~100 GB disk space because 30k images are large
-    def stage_progress(stage_idx:int) -> int: return progress["end"].to("CPU")[stage_idx].item()
-    if stage_progress(0) < len(eval_inputs):
-      tokens = []
-      for i in tqdm(range(0, len(eval_inputs), CONTEXT_BS)):
-        subset = [cond_stage.tokenize(row["caption"], device="CPU") for row in eval_inputs[i: i+CONTEXT_BS]]
-        tokens.append(Tensor.cat(*subset, dim=0).realize())
-      progress["tokens"].assign(Tensor.cat(*tokens, dim=0).realize()).realize()
-      progress["end"][0:1].assign(Tensor([len(eval_inputs)], dtype=dtypes.int)).realize()
-    prev_stage = "tokens"
-    tokens = progress["tokens"]
-
-    # wrapper code for every model
-    for stage_idx, model, jit, bs, callback in zip(range(1,6), models, jits, all_bs, callbacks):
-      stage = stages[stage_idx]
-      if stage_progress(stage_idx) >= len(eval_inputs):
-        prev_stage = stage
-        continue # use cache
-      t0 = time.perf_counter()
-      print(f"starting eval with model: {model}")
-      if stage_idx == 1: inputs = tokens
-      elif stage_idx == 5: inputs = progress["imgs"]
-      else: inputs = progress[prev_stage]
-
-      Tensor.realize(*[p.to_(GPUS) for p in get_parameters(model)])
-      for batch_idx in tqdm(range(stage_progress(stage_idx), inputs.shape[0], bs)):
-        t1 = time.perf_counter()
-        batch, unpadded_bs = get_batch(inputs, batch_idx, bs)
-        if isinstance(model, OpenClipEncoder): batch = callback(batch, get_batch(tokens, batch_idx, bs)[0].realize())
-        else: batch = callback(batch)
-        # to(GPUS[0]) is necessary for this to work, without that the result is still on GPUS, probably due to a bug
-        batch = batch.to(GPUS[0]).to("CPU")[0:unpadded_bs].realize()
-        progress[stage][batch_idx: batch_idx + bs].assign(batch).realize()
-        # keep track of what our last output was, so we can resume from there if we crash in this loop
-        progress["end"][stage_idx: stage_idx + 1].assign(Tensor([batch_idx + bs], dtype=dtypes.int)).realize()
-        print(f"model: {model}, batch_idx: {batch_idx}, elapsed: {(time.perf_counter() - t1):.2f}")
-      del batch
-        
-      jit.reset()
-      Tensor.realize(*[p.to_("CPU") for p in get_parameters(model)])
-      print(f"done with model: {model}, elapsed: {(time.perf_counter() - t0):.2f}")
-      prev_stage = stage
-
-    inception_stats_fn = str(DATADIR / "coco2014" / "val2014_30k_stats.npz")
-    fid_score = inception.compute_score(progress["inception"].to("CPU"), inception_stats_fn)
-    clip_score = progress["clip"].to(GPUS[0]).mean().item()
-    for name in disk_tensor_names:
-      Path(f"{EVAL_CKPT_DIR}/{name}.bytes").unlink(missing_ok=True)
-    
-    if EVAL_SAMPLES and BEAM:
-      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-      sys.exit() # Don't eval additional models; we don't care about clip/fid scores when running BEAM on eval sample subset
-
-    return clip_score, fid_score
-
-  # evaluate checkpoints in reverse chronological order
-  for ckpt_iteration, p in sorted(eval_queue, reverse=True):
-    unet_ckpt = safe_load(p)
-    load_state_dict(unet, unet_ckpt)
-    clip_score, fid_score = eval_unet(eval_inputs, unet, model.cond_stage_model, model.first_stage_model, inception, clip_encoder)
-    converged = True if clip_score >= 0.15 and fid_score <= 90 else False
-    print(f"eval results for {EVAL_CKPT_DIR}/{p.name}: clip={clip_score}, fid={fid_score}, converged={converged}")
-    if WANDB:
-      wandb.log({"eval/ckpt_iteration": ckpt_iteration, "eval/clip_score": clip_score, "eval/fid_score": fid_score})
-    if converged and STOP_IF_CONVERGED:
-      print(f"Convergence detected, exiting early before evaluating other checkpoints due to STOP_IF_CONVERGED={STOP_IF_CONVERGED}")
-      sys.exit()
-
-  # for testing
-  return clip_score, fid_score, ckpt_iteration
-
 if __name__ == "__main__":
  # inference only
  Tensor.training = False

-  models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert").split(",")
+  models = getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,mrcnn").split(",")
  for m in models:
    nm = f"eval_{m}"
    if nm in globals():
@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing

 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
-from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, Profiling
+from tinygrad.helpers import getenv, BEAM, WINO, round_up, diskcache_clear, FUSE_CONV_BW, Profiling
 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

@@ -252,10 +252,6 @@ def train_resnet():
        print(f"epoch global_ops: {steps_in_train_epoch * GlobalCounters.global_ops:_}, "
              f"epoch global_mem: {steps_in_train_epoch * GlobalCounters.global_mem:_}")
        # if we are doing beam search, run the first eval too
-        if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-          min_time = min(step_times)
-          assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
-
        if (TRAIN_BEAM or EVAL_BEAM) and e == start_epoch: break
        return
    if MLLOGGER and RUNMLPERF:
@@ -348,8 +344,6 @@ def train_resnet():
        print(f"saving ckpt to {fn}")
        safe_save(get_training_state(model, optimizer_group, scheduler_group), fn)

-
-
 def train_retinanet():
  from contextlib import redirect_stdout
  from examples.mlperf.dataloader import batch_load_retinanet
@@ -707,7 +701,7 @@ def train_unet3d():
  ```BASEDIR=<folder_path> ./examples/mlperf/scripts/setup_kits19_dataset.sh```

  2) To start training the model, run the following:
-  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
+  ```time PYTHONPATH=. WANDB=1 TRAIN_BEAM=3 FUSE_CONV_BW=1 GPUS=6 BS=6 MODEL=unet3d python3 examples/mlperf/model_train.py```
  """
  from examples.mlperf.losses import dice_ce_loss
  from examples.mlperf.metrics import dice_score
@@ -749,6 +743,7 @@ def train_unet3d():
    "train_beam": TRAIN_BEAM,
    "eval_beam": EVAL_BEAM,
    "wino": WINO.value,
+    "fuse_conv_bw": FUSE_CONV_BW.value,
    "gpus": GPUS,
    "default_float": dtypes.default_float.name
  }
@@ -918,6 +913,40 @@ def train_rnnt():
  # TODO: RNN-T
  pass

+@TinyJit
+def train_step_bert(model, optimizer, scheduler, loss_scaler:float, GPUS, grad_acc:int, **kwargs):
+  optimizer.zero_grad()
+
+  for i in range(grad_acc):
+    input_ids, segment_ids = kwargs[f"input_ids{i}"], kwargs[f"segment_ids{i}"]
+    # NOTE: these two have different names
+    attention_mask, masked_positions = kwargs[f"input_mask{i}"], kwargs[f"masked_lm_positions{i}"]
+    masked_lm_ids, masked_lm_weights, next_sentence_labels = kwargs[f"masked_lm_ids{i}"], kwargs[f"masked_lm_weights{i}"], kwargs[f"next_sentence_labels{i}"]
+
+    for t in [input_ids, segment_ids, attention_mask, masked_positions, masked_lm_ids, masked_lm_weights, next_sentence_labels]:
+      if len(GPUS) > 1: t.shard_(GPUS, axis=0)
+      else: t.to_(GPUS[0])
+
+    lm_logits, seq_relationship_logits = model(input_ids, attention_mask, masked_positions, segment_ids)
+    loss = model.loss(lm_logits, seq_relationship_logits, masked_lm_ids, masked_lm_weights, next_sentence_labels)
+    (loss * loss_scaler).backward()
+    # TODO: OOM without this realize with large grad_acc
+    Tensor.realize(*[p.grad for p in optimizer.params])
+
+  global_norm = Tensor(0.0, dtype=dtypes.float32, device=optimizer[0].device)
+  for p in optimizer.params:
+    p.grad = p.grad / loss_scaler
+    global_norm += p.grad.float().square().sum()
+  global_norm = global_norm.sqrt().contiguous()
+  for p in optimizer.params:
+    p.grad = (global_norm > 1.0).where((p.grad/global_norm).cast(p.grad.dtype), p.grad)
+
+  optimizer.step()
+  scheduler.step()
+  # TODO: no to("CPU") here because it blocks and messes the python time
+  Tensor.realize(loss, global_norm, optimizer.optimizers[0].lr)
+  return loss, global_norm, optimizer.optimizers[0].lr
+
@TinyJit
 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):
@@ -980,8 +1009,7 @@ def train_bert():
  # ** hyperparameters **
  BS                 = config["BS"]                     = getenv("BS", 11 * len(GPUS) if dtypes.default_float in (dtypes.float16, dtypes.bfloat16) else 8 * len(GPUS))
  grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
-  # TODO: implement grad accumulation + mlperf logging
-  assert grad_acc == 1
+  # TODO: mlperf logging
  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 1 * len(GPUS))
  max_lr             = config["OPT_BASE_LEARNING_RATE"] = getenv("OPT_BASE_LEARNING_RATE", 0.000175 * math.sqrt(GBS/96))
@@ -1040,8 +1068,8 @@ def train_bert():

  # ** Optimizer **
  parameters_no_wd = [v for k, v in get_state_dict(model).items() if "bias" in k or "LayerNorm" in k]
-  parameters_wd = [x for x in parameters if x not in set(parameters_no_wd)]
-  optimizer_wd = LAMB(parameters_wd, lr=max_lr, b1=opt_lamb_beta_1, b2=opt_lamb_beta_2, eps=epsilon, weight_decay=decay, adam=False)
+  parameters = [x for x in parameters if x not in set(parameters_no_wd)]
+  optimizer_wd = LAMB(parameters, lr=max_lr, b1=opt_lamb_beta_1, b2=opt_lamb_beta_2, eps=epsilon, weight_decay=decay, adam=False)
  optimizer_no_wd = LAMB(parameters_no_wd, lr=max_lr, b1=opt_lamb_beta_1, b2=opt_lamb_beta_2, eps=epsilon, weight_decay=0.0, adam=False)
  optimizer_group = OptimizerGroup(optimizer_wd, optimizer_no_wd)

@@ -1098,38 +1126,12 @@ def train_bert():
  # ** train loop **
  wc_start = time.perf_counter()

-  i, train_data = start_step, next(train_it)
+  i, train_data = start_step, [next(train_it) for _ in range(grad_acc)]

  if RUNMLPERF:
    if MLLOGGER:
      MLLOGGER.start(key=mllog_constants.EPOCH_START, value=i*GBS, metadata={"epoch_num": i*GBS})

-  @TinyJit
-  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]:
-      if len(GPUS) > 1: t.shard_(GPUS, axis=0)
-      else: t.to_(GPUS[0])
-    optimizer_group.zero_grad()
-
-    lm_logits, seq_relationship_logits = model(input_ids, attention_mask, masked_positions, segment_ids)
-    loss = model.loss(lm_logits, seq_relationship_logits, masked_lm_ids, masked_lm_weights, next_sentence_labels)
-    (loss * loss_scaler).backward()
-
-    global_norm = Tensor(0.0, dtype=dtypes.float32, device=optimizer_group[0].device)
-    for p in optimizer_group.params:
-      p.grad = p.grad / loss_scaler
-      global_norm += p.grad.float().square().sum()
-    global_norm = global_norm.sqrt().contiguous()
-    for p in optimizer_group.params:
-      p.grad = (global_norm > 1.0).where((p.grad/global_norm).cast(p.grad.dtype), p.grad)
-
-    optimizer_group.step()
-    scheduler_group.step()
-    # TODO: no to("CPU") here because it blocks and messes the python time
-    Tensor.realize(loss, global_norm, optimizer_group.optimizers[0].lr)
-    return loss, global_norm, optimizer_group.optimizers[0].lr
-
  while train_data is not None and i < train_steps and not achieved:
    if getenv("TRAIN", 1):
      Tensor.training = True
@@ -1137,12 +1139,16 @@ def train_bert():
      st = time.perf_counter()
      GlobalCounters.reset()
      with WallTimeEvent(BenchEvent.STEP):
-        loss, global_norm, lr = train_step_bert(
-          train_data["input_ids"], train_data["segment_ids"], train_data["input_mask"], train_data["masked_lm_positions"], \
-          train_data["masked_lm_ids"], train_data["masked_lm_weights"], train_data["next_sentence_labels"])
+        data = {f"{k}{i}":v for i,d in enumerate(train_data) for k,v in d.items()}
+        loss, global_norm, lr = train_step_bert(model, optimizer_group, scheduler_group, loss_scaler, GPUS, grad_acc, **data)

        pt = time.perf_counter()
-        next_data = next(train_it)
+
+        try:
+          next_data = [next(train_it) for _ in range(grad_acc)]
+        except StopIteration:
+          next_data = None
+
        dt = time.perf_counter()

        device_str = parameters[0].device if isinstance(parameters[0].device, str) else f"{parameters[0].device[0]} * {len(parameters[0].device)}"
@@ -1177,9 +1183,7 @@ def train_bert():
      if MLLOGGER and RUNMLPERF:
        MLLOGGER.start(key=mllog_constants.EVAL_START, value=None, metadata={"epoch_num": i*GBS, "step_num": i})
      if getenv("RESET_STEP"): train_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE") and train_step_bert.captured is not None:
-        # TODO: this hangs on tiny green after 90 minutes of training
-        train_step_bert.captured.free_intermediates()
+      elif getenv("FREE_INTERMEDIATE", 1) and train_step_bert.captured is not None: train_step_bert.captured.free_intermediates()
      eval_lm_losses = []
      eval_clsf_losses = []
      eval_lm_accs = []
@@ -1213,7 +1217,7 @@ def train_bert():
          return

      if getenv("RESET_STEP"): eval_step_bert.reset()
-      elif getenv("FREE_INTERMEDIATE") and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()
+      elif getenv("FREE_INTERMEDIATE", 1) and eval_step_bert.captured is not None: eval_step_bert.captured.free_intermediates()

      del eval_data
      avg_lm_loss = sum(eval_lm_losses) / len(eval_lm_losses)
@@ -1289,7 +1293,6 @@ def train_llama3():
  BASEDIR            = config["BASEDIR"]                = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
  BS                 = config["BS"]                     = getenv("BS", 16)
  grad_acc           = config["GRADIENT_ACC_STEPS"]     = getenv("GRADIENT_ACC_STEPS", 1)
-  assert grad_acc == 1, f"{grad_acc=} is not supported"
  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
  SEED               = config["SEED"]                   = getenv("SEED", 5760)
  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
@@ -1300,7 +1303,7 @@ def train_llama3():
  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)

-  # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 JITBEAM=2 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=512 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py
+  # LR=1e-4 TRAIN_ON_VAL=1 DEFAULT_FLOAT=bfloat16 FUSE_ARANGE=1 JITBEAM=2 OPTIM_DTYPE=bfloat16 LLAMA3_SIZE=1B WARMUP_STEPS=36 DECAY_STEPS=360 SEQLEN=512 PYTHONPATH=. AMD=1 AMD_LLVM=0 MODEL=llama3 python3 examples/mlperf/model_train.py
  # trains to 7

  opt_adamw_beta_1 = 0.9
@@ -1314,14 +1317,12 @@ def train_llama3():
  opt_base_learning_rate = getenv("LR", 8e-5 * GBS / 1152)  # NOTE: cannot change for benchmark
  opt_end_learning_rate = getenv("END_LR", 8e-7)

-  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
+  # TODO: confirm weights are in bf16
  # vocab_size from the mixtral tokenizer
-  if not SMALL: model_params |= {"vocab_size": 32000}
-  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
-  model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
-  params = get_parameters(model)
-  # weights are all bfloat16 for now
-  assert params and all(p.dtype == dtypes.bfloat16 for p in params)
+  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
+  params = params | {"vocab_size": 32000} if not SMALL else params
+  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: params['n_layers'] = llama_layers
+  model = Transformer(**params, max_context=SEQLEN, jit=False, disable_kv_cache=True)

  if getenv("FAKEDATA"):
    for v in get_parameters(model):
@@ -1366,17 +1367,20 @@ def train_llama3():

  @TinyJit
  @Tensor.train()
-  def train_step(model, tokens:Tensor):
+  def train_step(model, tokens:Tensor, grad_acc:int):
    optim.zero_grad()
-    if (DP := getenv("DP", 1)) > 1:
-      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
-      tokens = tokens.shard(device, 0)
-    if (MP := getenv("MP", 1)) > 1:
-      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
-      tokens = tokens.shard(device)
-    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
-    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    loss.backward()
+    # grad acc
+    for batch in tokens.split(tokens.shape[0]//grad_acc):
+      if (DP := getenv("DP", 1)) > 1:
+        device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
+        batch = batch.shard(device, 0)
+      if (MP := getenv("MP", 1)) > 1:
+        device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
+        batch = batch.shard(device)
+      logits:Tensor = model(batch[:, :-1], start_pos=0, temperature=math.nan)
+      loss = logits.sparse_categorical_crossentropy(batch[:, 1:])
+      loss.backward()
+      Tensor.realize(*[p.grad for p in optim.params])
    # L2 norm grad clip
    # https://github.com/NVIDIA/NeMo/blob/3368c3fc0b4a186ab33a1d68a504315100c0b2a6/nemo/collections/nlp/modules/common/megatron/clip_grads.py#L57
    # https://docs.pytorch.org/docs/stable/generated/torch.nn.utils.clip_grad_norm_.html
@@ -1411,18 +1415,18 @@ def train_llama3():
  # ** data iters **
  def fake_data(bs, samples):
    for _ in range(samples // bs):
-      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=model_params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)
+      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)

  def get_train_iter():
    if getenv("FAKEDATA", 0):
-      return fake_data(BS, SAMPLES)
+      return fake_data(GBS, SAMPLES)
    else:
      if SMALL:
        from examples.mlperf.dataloader import batch_load_llama3_small
-        return batch_load_llama3_small(BS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
+        return batch_load_llama3_small(GBS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
      else:
        from examples.mlperf.dataloader import batch_load_llama3
-        return batch_load_llama3(BS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
+        return batch_load_llama3(GBS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))

  def get_eval_iter():
    if getenv("FAKEDATA", 0):
@@ -1438,34 +1442,28 @@ def train_llama3():
  iter = get_train_iter()
  i, sequences_seen = resume_ckpt, 0
  for tokens in tqdm(iter, total=SAMPLES//GBS):
+    t = time.perf_counter()
    GlobalCounters.reset()
-    if getenv("TRAIN", 1):
-      t = time.perf_counter()
-      loss, lr = train_step(model, tokens)
-      loss = loss.float().item()
-      lr = lr.item()
+    loss, lr = train_step(model, tokens, grad_acc)
+    loss = loss.float().item()

-      i += 1
-      sequences_seen += tokens.shape[0]
+    i += 1
+    sequences_seen += tokens.shape[0]

-      sec = time.perf_counter()-t
-      tqdm.write(
-        f"{i:5} {sec:.2f} s run, {loss:.4f} loss, {lr:.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, "
-        f"{GlobalCounters.global_ops * 1e-9 / sec:9.2f} GFLOPS")
+    tqdm.write(f"{loss:.4f} loss, {lr.item():.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, {time.perf_counter()-t:.2f} s")
+    if (fname:=getenv("LOSS_FILE", "")):
+      with open(fname, "a") as f:
+        f.write(f"{i} {loss:.4f} {lr.item():.12f} {GlobalCounters.mem_used / 1e9:.2f}\n")

-      if (fname:=getenv("LOSS_FILE", "")):
-        with open(fname, "a") as f:
-          f.write(f"{i} {loss:.4f} {lr.item():.12f} {GlobalCounters.mem_used / 1e9:.2f}\n")
+    if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
+      tqdm.write("saving checkpoint")
+      if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
+      fn = f"{ckpt_dir}/llama3_{i}.safe"
+      safe_save(get_state_dict(model), fn)

-      if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
-        tqdm.write("saving checkpoint")
-        if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
-        fn = f"{ckpt_dir}/llama3_{i}.safe"
-        safe_save(get_state_dict(model), fn)
-
-        tqdm.write("saving optim checkpoint")
-        fn = f"{ckpt_dir}/llama3_{i}_optim.safe"
-        safe_save(get_state_dict(scheduler), fn)
+      tqdm.write("saving optim checkpoint")
+      fn = f"{ckpt_dir}/llama3_{i}_optim.safe"
+      safe_save(get_state_dict(scheduler), fn)

    if sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1):
      tqdm.write(f"evaluating after {sequences_seen} sequences")
@@ -1489,144 +1487,6 @@ def train_llama3():
          safe_save(get_state_dict(model), fn)
        break

-def train_stable_diffusion():
-  from extra.models.unet import UNetModel
-  from examples.mlperf.dataloader import batch_load_train_stable_diffusion
-  from examples.mlperf.lr_schedulers import LambdaLR, LambdaLinearScheduler
-  from examples.mlperf.initializers import init_stable_diffusion
-  from examples.mlperf.helpers import get_training_state
-  import numpy as np
-
-  config = {}
-  GPUS               = config["GPUS"]                   = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 1))]
-  seed               = config["seed"]                   = getenv("SEED", 12345)
-  # ** hyperparameters **
-  BS                 = config["BS"]                     = getenv("BS", 1 * len(GPUS))
-  BASE_LR            = config["LEARNING_RATE"]          = getenv("LEARNING_RATE", 2.5e-7)
-  # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
-  # "Checkpoint must be collected every 512,000 images. CEIL(512000 / global_batch_size) if 512000 is not divisible by GBS."
-  # NOTE: It's inferred that "steps" is the unit for the output of the CEIL formula, based on all other cases of CEIL in the rules
-  CKPT_STEP_INTERVAL = config["CKPT_STEP_INTERVAL"]     = getenv("CKPT_STEP_INTERVAL", math.ceil(512_000 / BS))
-  CKPTDIR            = config["CKPTDIR"]                = Path(getenv("CKPTDIR", "./checkpoints"))
-  DATADIR            = config["DATADIR"]                = Path(getenv("DATADIR", "./datasets"))
-  UNET_CKPTDIR       = config["UNET_CKPTDIR"]           = Path(getenv("UNET_CKPTDIR", "./checkpoints"))
-  TOTAL_CKPTS        = config["TOTAL_CKPTS"]            = getenv("TOTAL_CKPTS", 0)
-
-  print(f"training on {GPUS}")
-  lr = BS * BASE_LR
-  print(f"BS={BS}, BASE_LR={BASE_LR}, lr={lr}")
-  print(f"CKPT_STEP_INTERVAL = {CKPT_STEP_INTERVAL}")
-  for x in GPUS: Device[x]
-  if (WANDB := getenv("WANDB", "")):
-    import wandb
-    wandb.init(config=config, project="MLPerf-Stable-Diffusion")
-
-  Tensor.manual_seed(seed)  # seed for weight initialization
-  model, unet, sqrt_alphas_cumprod, sqrt_one_minus_alphas_cumprod = init_stable_diffusion("v2-mlperf-train", CKPTDIR / "sd" / "512-base-ema.ckpt", GPUS)
-
-  optimizer = AdamW(get_parameters(unet))
-  lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
-  lr_scheduler = LambdaLR(optimizer, Tensor(lr, dtype=dtypes.float, device=optimizer.device), lambda_lr_callback)
-
-  @TinyJit
-  def train_step(mean:Tensor, logvar:Tensor, tokens:Tensor, unet:UNetModel, optimizer:LAMB, lr_scheduler:LambdaLR) -> Tensor:
-    optimizer.zero_grad()
-
-    timestep = Tensor.randint(BS, low=0, high=model.alphas_cumprod.shape[0], dtype=dtypes.int, device=GPUS[0])
-    latent_randn = Tensor.randn(*mean.shape, device=GPUS[0])
-    noise = Tensor.randn(*mean.shape, device=GPUS[0])
-    for t in (mean, logvar, tokens, timestep, latent_randn, noise):
-      t.shard_(GPUS, axis=0)
-
-    std = Tensor.exp(0.5 * logvar.clamp(-30.0, 20.0))
-    latent = (mean + std * latent_randn) * 0.18215
-
-    sqrt_alphas_cumprod_t = sqrt_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
-    sqrt_one_minus_alphas_cumprod_t = sqrt_one_minus_alphas_cumprod[timestep].reshape(timestep.shape[0], 1, 1, 1)
-    latent_with_noise = sqrt_alphas_cumprod_t * latent + sqrt_one_minus_alphas_cumprod_t * noise
-    v_true = sqrt_alphas_cumprod_t * noise - sqrt_one_minus_alphas_cumprod_t * latent
-
-    context = model.cond_stage_model.embed_tokens(tokens)
-
-    out = unet(latent_with_noise, timestep, context)
-    loss = ((out - v_true) ** 2).mean()
-    del mean, logvar, std, latent, noise, sqrt_alphas_cumprod_t, sqrt_one_minus_alphas_cumprod_t
-    del out, v_true, context, latent_randn, tokens, timestep
-    loss.backward()
-
-    optimizer.step()
-    lr_scheduler.step()
-    loss, out_lr = loss.detach().to("CPU"), optimizer.lr.to("CPU")
-    Tensor.realize(loss, out_lr)
-    return loss, out_lr
-    
-  # checkpointing takes ~9 minutes without this, and ~1 minute with this
-  @TinyJit
-  def ckpt_to_cpu():
-    ckpt = get_training_state(unet, optimizer, lr_scheduler)
-    # move to CPU first so more GPU bufs aren't created (can trigger OOM)
-    for k,v in ckpt.items(): ckpt[k] = v.detach().to("CPU")
-    Tensor.realize(*[v for v in ckpt.values()])
-    for k,v in ckpt.items(): ckpt[k] = v.cast(v.dtype.base).contiguous()
-    Tensor.realize(*[v for v in ckpt.values()])
-    return ckpt
-
-  # training loop
-  dl = batch_load_train_stable_diffusion(f'{DATADIR}/laion-400m/webdataset-moments-filtered/{{00000..00831}}.tar', BS)
-  # for tests
-  saved_checkpoints = []
-
-  train_start_time = time.perf_counter()
-  t0 = t6 = time.perf_counter()
-  for i, batch in enumerate(dl, start=1):
-    loop_time = time.perf_counter() - t0
-    t0 = time.perf_counter()
-    dl_time = t0 - t6
-    GlobalCounters.reset()
-
-    mean, logvar = np.split(np.concatenate(batch["npy"], axis=0), 2, axis=1)
-    mean, logvar = Tensor(mean, dtype=dtypes.float32, device="CPU"), Tensor(logvar, dtype=dtypes.float32, device="CPU")
-    tokens = []
-    for text in batch['txt']: tokens += model.cond_stage_model.tokenizer.encode(text, pad_with_zeros=True)
-    tokens = Tensor(tokens, dtype=dtypes.int32, device="CPU").reshape(-1, 77)
-
-    t1 = time.perf_counter()
-    loss, lr = train_step(mean, logvar, tokens, unet, optimizer, lr_scheduler)
-    loss_item, lr_item = loss.item(), lr.item()
-    t2 = time.perf_counter()
-
-    if i == 3:
-      for _ in range(3): ckpt_to_cpu() # do this at the beginning of run to prevent OOM surprises when checkpointing
-      print("BEAM COMPLETE", flush=True) # allows wrapper script to detect BEAM search completion and retry if it failed
-      
-    total_train_time = time.perf_counter() - train_start_time
-    if WANDB:
-      wandb.log({"train/loss": loss_item, "train/lr": lr_item, "train/loop_time_prev": loop_time, "train/dl_time": dl_time, "train/step": i,
-                    "train/GFLOPS": GlobalCounters.global_ops * 1e-9 / (t2-t1), "train/input_prep_time": t1-t0,
-                    "train/train_step_time": t2-t1, "train/total_time": total_train_time})
-
-      if i == 1 and wandb.run is not None:
-        with open(f"{UNET_CKPTDIR}/wandb_run_id_{wandb.run.id}", "w") as f:
-          f.write(f"wandb.run.id = {wandb.run.id}")
-
-    if i % CKPT_STEP_INTERVAL == 0:
-      # https://github.com/mlcommons/training_policies/blob/cfa99da479b8d5931f7a3c67612d021dfb47510a/training_rules.adoc#benchmark_specific_rules
-      # "evaluation is done offline, the time is not counted towards the submission time."
-      fn = f"{UNET_CKPTDIR}/{i}.safetensors"
-      print(f"saving unet checkpoint at {fn}")
-      saved_checkpoints.append(fn)
-      safe_save({k.replace("model.", ""):v for k,v in ckpt_to_cpu().items() if k.startswith("model.")}, fn)
-      if TOTAL_CKPTS and i == TOTAL_CKPTS * CKPT_STEP_INTERVAL:
-        print(f"ending run after {i} steps ({TOTAL_CKPTS} checkpoints collected)")
-        return saved_checkpoints
-
-    t3 = time.perf_counter()
-    print(f"""step {i}: {GlobalCounters.global_ops * 1e-9 / (t2-t1):9.2f} GFLOPS, mem_used: {GlobalCounters.mem_used / 1e9:.2f} GB,
-  loop_time_prev: {loop_time:.2f}, dl_time: {dl_time:.2f}, input_prep_time: {t1-t0:.2f}, train_step_time: {t2-t1:.2f},
-  t3-t2: {t3-t2:.4f}, loss:{loss_item:.5f}, lr:{lr_item:.3e}, total_train_time:{total_train_time:.2f}
-  """)
-    t6 = time.perf_counter()
-
 if __name__ == "__main__":
  multiprocessing.set_start_method('spawn')

@@ -1635,7 +1495,7 @@ if __name__ == "__main__":
  else: bench_log_manager = contextlib.nullcontext()

  with Tensor.train():
-    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
+    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn").split(","):
      nm = f"train_{m}"
      if nm in globals():
        print(f"training {m}")
@@ -1,57 +0,0 @@
-#!/usr/bin/env bash
-# adapted from https://github.com/mlcommons/training/blob/4bdf5c8ed218ad76565a2ba1ac27c919ccc6d689/stable_diffusion/README.md
-
-# setup dirs
-
-DATA=/raid/datasets/stable_diffusion
-
-LAION=$DATA/laion-400m/webdataset-moments-filtered 
-COCO=$DATA/coco2014
-mkdir -p $LAION $COCO
-
-CKPT=/raid/weights/stable_diffusion
-mkdir -p $CKPT/clip $CKPT/sd $CKPT/inception
-
-# download data
-
-# if rclone isn't installed system-wide / in your PATH, put the executable path in quotes below
-#RCLONE=""
-RCLONE="rclone"
-
-## VAE-encoded image latents, from 6.1M image subset of laion-400m
-## about 1 TB for whole download
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/ ${LAION} --include="*.tar" -P
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/laion-400m/moments-webdataset-filtered/sha512sums.txt ${LAION} -P
-cd $LAION && grep -E '\.tar$' sha512sums.txt | sha512sum -c --quiet - && \
-  echo "All .tar files verified" || { echo "Checksum failure when validating downloaded Laion moments"; exit 1; }
-
-## prompts and FID statistics from 30k image subset of coco2014
-## 33 MB
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k.tsv ${COCO} -P
-
-$RCLONE config create mlc-training s3 provider=Cloudflare access_key_id=76ea42eadb867e854061a1806220ee1e secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-$RCLONE copy mlc-training:mlcommons-training-wg-public/stable_diffusion/datasets/coco2014/val2014_30k_stats.npz ${COCO} -P
-
-# download checkpoints
-
-## clip (needed for text and vision encoders for validation)
-CLIP_WEIGHTS_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/resolve/main/open_clip_pytorch_model.bin"
-CLIP_WEIGHTS_SHA256="9a78ef8e8c73fd0df621682e7a8e8eb36c6916cb3c16b291a082ecd52ab79cc4"
-CLIP_CONFIG_URL="https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/raw/main/open_clip_config.json"
-wget -N -P ${CKPT}/clip ${CLIP_WEIGHTS_URL}
-wget -N -P ${CKPT}/clip ${CLIP_CONFIG_URL}
-echo "${CLIP_WEIGHTS_SHA256}  ${CKPT}/clip/open_clip_pytorch_model.bin"                    | sha256sum -c
-
-## sd (needed for latent->image decoder for validation, also has clip text encoder for training)
-SD_WEIGHTS_URL='https://huggingface.co/stabilityai/stable-diffusion-2-base/resolve/main/512-base-ema.ckpt'
-SD_WEIGHTS_SHA256="d635794c1fedfdfa261e065370bea59c651fc9bfa65dc6d67ad29e11869a1824"
-wget -N -P ${CKPT}/sd ${SD_WEIGHTS_URL}
-echo "${SD_WEIGHTS_SHA256}  ${CKPT}/sd/512-base-ema.ckpt"                    | sha256sum -c
-
-## inception (needed for validation)
-FID_WEIGHTS_URL='https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth'
-FID_WEIGHTS_SHA1="bd836944fd6db519dfd8d924aa457f5b3c8357ff"
-wget -N -P ${CKPT}/inception ${FID_WEIGHTS_URL}
-echo "${FID_WEIGHTS_SHA1}  ${CKPT}/inception/pt_inception-2015-12-05-6726825d.pth"                    | sha1sum -c
@@ -1,72 +0,0 @@
-#!/usr/bin/env bash
-
-DATETIME=${2:-$(date "+%m%d%H%M")}
-LOGFILE="${HOME}/logs/sd_mi300x_${DATETIME}.log"
-# UNET_CKPTDIR must be set: training saves checkpoints to this path, then a separate eval process scans this path to know which checkpoints to eval
-export UNET_CKPTDIR="${HOME}/stable_diffusion/training_checkpoints/${DATETIME}"
-mkdir -p "${HOME}/logs" "$UNET_CKPTDIR"
-
-# run this script in isolation when using the --bg flag
-if [[ "${1:-}" == "--bg" ]]; then
-  echo "logging output to $LOGFILE"
-  echo "saving UNet checkpoints to $UNET_CKPTDIR"
-  script_path="$(readlink -f "${BASH_SOURCE[0]}")"
-  nohup bash "$script_path" run "$DATETIME" >"$LOGFILE" 2>&1 & disown $!
-  exit 0
-fi
-
-# venv management
-if [[ -d .venv-sd-mlperf ]]; then
-  . .venv-sd-mlperf/bin/activate
-else
-  python3 -m venv .venv-sd-mlperf && . .venv-sd-mlperf/bin/activate
-  pip install --index-url https://download.pytorch.org/whl/cpu torch && pip install tqdm numpy ftfy regex pillow scipy wandb webdataset
-fi
-pip list
-apt list --installed | grep amdgpu
-rocm-smi --version
-modinfo amdgpu | grep version
-
-export BEAM=2 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 IGNORE_JIT_FIRST_BEAM=1 HCQDEV_WAIT_TIMEOUT_MS=300000
-export AMD_LLVM=0 # bf16 seems to require this
-export DATADIR="/raid/datasets/stable_diffusion"
-export CKPTDIR="/raid/weights/stable_diffusion"
-export EVAL_CKPT_DIR=$UNET_CKPTDIR
-export MODEL="stable_diffusion" PYTHONPATH="."
-export GPUS=8 BS=304
-export CONTEXT_BS=816 DENOISE_BS=600 DECODE_BS=384 INCEPTION_BS=560 CLIP_BS=240
-export WANDB=1
-export PARALLEL=4
-export PYTHONUNBUFFERED=1
-sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setperfdeterminism 1500 || exit 1
-
-# Retry BEAM search if script fails before BEAM COMPLETE is printed, but don't retry after that
-run_retry(){ local try=0 max=5 code tmp py pgid kids
-  while :; do
-    tmp=$(mktemp)
-    setsid bash -c 'exec env "$@"' _ "$@" > >(tee -a "$LOGFILE" | tee "$tmp") 2>&1 &
-    py=$!; pgid=$(ps -o pgid= -p "$py" | tr -d ' ')
-    wait "$py"; code=$?
-    [[ -n "$pgid" ]] && { kill -TERM -"$pgid" 2>/dev/null; sleep 1; kill -KILL -"$pgid" 2>/dev/null; }
-    kids=$(pgrep -P "$py" || true)
-    while [[ -n "$kids" ]]; do
-      kill -TERM $kids 2>/dev/null; sleep 0.5
-      kids=$(for k in $kids; do pgrep -P "$k" || true; done)
-    done
-    grep -q 'BEAM COMPLETE' "$tmp" && { rm -f "$tmp"; return 1; }
-    rm -f "$tmp"
-    ((code==0)) && return 0
-    ((try>=max)) && return 2
-    ((try++)); sleep 90; echo "try = ${try}"
-  done
-}
-
-# Power limiting to 400W is only needed if GPUs fall out of sync (causing 2.2x increased train time) at higher power, which has been observed at 450W
-sudo rocm-smi -d 0 1 2 3 4 5 6 7 --setpoweroverdrive 750 && \
-run_retry TOTAL_CKPTS=7 python3 examples/mlperf/model_train.py; (( $? == 2 )) && { echo "training failed before BEAM completion"; exit 2; }
-sleep 90
-
-run_retry EVAL_SAMPLES=600 python3 examples/mlperf/model_eval.py; (( $? == 2 )) && { echo "eval failed before BEAM completion"; exit 2; }
-# Checkpoints will be evaluated in reverse chronological order, even if above training crashed early
-# STOP_IF_CONVERGED=1: Stop the eval after the first time convergence is detected; no more checkpoints will be evaluated after that.
-STOP_IF_CONVERGED=1 python3 examples/mlperf/model_eval.py
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=1 BS=128 EVAL_BS=128
-
-export IGNORE_OOB=1
-
-export BEAM=3 BEAM_UOPS_MAX=4000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-# export BEAM_LOG_SURPASS_MAX=1
-# export BASEDIR="/raid/datasets/wiki"
-
-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export BENCHMARK=10 BERT_LAYERS=2
-
-python3 examples/mlperf/model_train.py
@@ -1,20 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_8xMI300X"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_8xMI300x_${DATETIME}_${SEED}.log"
-
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_8xMI350X"
-export DEFAULT_FLOAT="HALF" GPUS=8 BS=1024 EVAL_BS=1024
-
-# similar to https://github.com/mlcommons/training_results_v3.1/blob/d06288b2bd675a9d88e0e6181f5bb5626b71ec19/Quanta_Cloud_Technology/results/D54U-3U/bert/result_1.txt#L54
-export OPT_BASE_LEARNING_RATE=0.0011 OPT_LAMB_BETA_1=0.60466 OPT_LAMB_BETA_2=0.85437 DECAY=0.1
-export TRAIN_STEPS=3900
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=5000000
-
-export BEAM=3 BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1 FREE_INTERMEDIATE=0
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_8xMI350x_${DATETIME}_${SEED}.log"
-
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,17 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
-export BASEDIR="/raid/datasets/wiki"
-
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,16 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,28 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." NV=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=72 EVAL_BS=72
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=8 BEAM_UOPS_MAX=10000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,69 +0,0 @@
-# 1. Problem
-
-This problem uses BERT for NLP.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-Also install gdown (for dataset), numpy, tqdm and tensorflow.
-```
-pip install gdown numpy tqdm tensorflow
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download and verify data
-
-### 1. Download raw data
-
-```
-BASEDIR="/raid/datasets/wiki" WIKI_TRAIN=1 VERIFY_CHECKSUM=1 python3 extra/datasets/wikipedia_download.py
-```
-
-### 2. Preprocess train and validation data
-
-Note: The number of threads used for preprocessing is limited by available memory. With 128GB of RAM, a maximum of 16 threads is recommended. 
-
-#### Training:
-```
-BASEDIR="/raid/datasets/wiki" NUM_WORKERS=16 python3 extra/datasets/wikipedia.py pre-train all
-```
-
-Generating a specific topic (Between 0 and 499)
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-train 42
-```
-
-#### Validation:
-```
-BASEDIR="/raid/datasets/wiki" python3 extra/datasets/wikipedia.py pre-eval
-```
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_green/run_and_time.sh
-```
-
-### tinybox_red
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_red/run_and_time.sh
-```
-### tinybox_8xMI300X
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/bert/implementations/tinybox_8xMI300X/run_and_time.sh
-```
@@ -1,18 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BEAM_LOG_SURPASS_MAX=1
-export BASEDIR="/raid/datasets/wiki"
-
-export RESET_STEP=1
-export BENCHMARK=10 BERT_LAYERS=2 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,16 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-export WANDB=1 PARALLEL=0
-
-RUNMLPERF=1 python3 examples/mlperf/model_train.py
@@ -1,31 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="bert"
-export SUBMISSION_PLATFORM="tinybox_red"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export IGNORE_OOB=1
-export REWRITE_STACK_LIMIT=500000
-
-export BEAM=5 BEAM_UOPS_MAX=8000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/wiki"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="bert_red_${DATETIME}_${SEED}.log"
-
-export HCQDEV_WAIT_TIMEOUT_MS=100000  # prevents hang?
-
-# init
-sleep 5 && sudo rmmod amdgpu || true
-BENCHMARK=10 INITMLPERF=1 BERT_LAYERS=2 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,50 +0,0 @@
-# 1. Problem
-
-This problem uses the ResNet-50 CNN to do image classification.
-
-## Requirements
-
-Install tinygrad and mlperf-logging from master.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-### tinybox_red
-Disable cwsr
-This is the default on production tinybox red.
-```
-sudo vi /etc/modprobe.d/amdgpu.conf
-cat <<EOF > /etc/modprobe.d/amdgpu.conf
-options amdgpu cwsr_enable=0
-EOF
-sudo update-initramfs -u
-sudo reboot
-
-# validate
-sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
-```
-
-# 2. Directions
-
-## Steps to download and verify data
-
-```
-IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
-```
-
-## Steps for one time setup
-
-### tinybox_red
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
-```
-
-## Steps to run benchmark
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-```
@@ -1,13 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-export BENCHMARK=10 DEBUG=2
-
-python3 examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-export EVAL_START_EPOCH=3 EVAL_FREQ=4
-
-export WANDB=1 PARALLEL=0
-
-python3 examples/mlperf/model_train.py
@@ -1,25 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." NV=1
-export MODEL="resnet"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=10 BEAM_PADTO=0
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="resnet_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,50 +0,0 @@
-# 1. Problem
-
-This problem uses the ResNet-50 CNN to do image classification.
-
-## Requirements
-
-Install tinygrad and mlperf-logging from master.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-### tinybox_red
-Disable cwsr
-This is the default on production tinybox red.
-```
-sudo vi /etc/modprobe.d/amdgpu.conf
-cat <<EOF > /etc/modprobe.d/amdgpu.conf
-options amdgpu cwsr_enable=0
-EOF
-sudo update-initramfs -u
-sudo reboot
-
-# validate
-sudo cat /sys/module/amdgpu/parameters/cwsr_enable #= 0
-```
-
-# 2. Directions
-
-## Steps to download and verify data
-
-```
-IMGNET_TRAIN=1 python3 extra/datasets/imagenet_download.py
-```
-
-## Steps for one time setup
-
-### tinybox_red
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/setup.sh
-```
-
-## Steps to run benchmark
-```
-examples/mlperf/training_submission_v4.0/tinycorp/benchmarks/resnet/implementations/tinybox_red/run_and_time.sh
-```
@@ -1,13 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=10 DEBUG=${DEBUG:-2}
-
-python3 examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="resnet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export EVAL_START_EPOCH=3 EVAL_FREQ=4
-
-export WANDB=1 PARALLEL=0
-
-python3 examples/mlperf/model_train.py
@@ -1,26 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." AMD=1
-export MODEL="resnet"
-export SUBMISSION_PLATFORM="tinybox_red"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=1536 EVAL_BS=192
-
-export RESET_STEP=0
-
-export TRAIN_BEAM=4 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=2000 BEAM_UPCAST_MAX=96 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=${LOGMLPERF:-1}
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="resnet_red_${DATETIME}_${SEED}.log"
-
-# init
-sleep 5 && sudo rmmod amdgpu || true
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 EVAL_START_EPOCH=3 EVAL_FREQ=4 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,8 +0,0 @@
-#!/bin/bash
-
-rocm-smi --setprofile compute
-rocm-smi --setmclk 3
-rocm-smi --setperflevel high
-
-# power cap to 350W
-echo "350000000" | sudo tee /sys/class/drm/card{1..6}/device/hwmon/hwmon*/power1_cap
@@ -1,38 +0,0 @@
-# 1. Problem
-
-This problem uses RetinaNet for SSD.
-
-## Requirements
-
-Install tinygrad and mlperf-logging (uncomment mlperf from setup.py) from branch mlperf_training_v5.0.
-```
-git clone https://github.com/tinygrad/tinygrad.git
-python3 -m pip install -e ".[mlperf]"
-```
-
-Also install the following dependencies:
-```
-pip install tqdm numpy pycocotools boto3 pandas torch torchvision
-```
-
-### tinybox_green
-Install the p2p driver per [README](https://github.com/tinygrad/open-gpu-kernel-modules/blob/550.54.15-p2p/README.md)
-This is the default on production tinybox green.
-
-# 2. Directions
-
-## Steps to download data
-
-Run the following:
-```
-BASEDIR=/raid/datasets/openimages python3 extra/datasets/openimages.py
-```
-
-## Running
-
-### tinybox_green
-
-#### Steps to run benchmark
-```
-examples/mlperf/training_submission_v5.0/tinycorp/benchmarks/retinanet/implementations/tinybox_green/run_and_time.sh
-```
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -1,25 +0,0 @@
-#!/bin/bash
-set -e  # Exit on any error
-set -o pipefail  # Make pipeline fail if any command fails
-
-export PYTHONPATH="." NV=1
-export MODEL="retinanet"
-export SUBMISSION_PLATFORM="tinybox_green"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-
-export TRAIN_BEAM=2 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-export IGNORE_JIT_FIRST_BEAM=1
-export BASEDIR="/raid/datasets/openimages"
-
-# pip install -e ".[mlperf]"
-export LOGMLPERF=1
-
-export SEED=$RANDOM
-DATETIME=$(date "+%m%d%H%M")
-LOGFILE="retinanet_green_${DATETIME}_${SEED}.log"
-
-# init
-BENCHMARK=10 INITMLPERF=1 python3 examples/mlperf/model_train.py | tee $LOGFILE
-
-# run
-PARALLEL=0 RUNMLPERF=1 python3 examples/mlperf/model_train.py | tee -a $LOGFILE
@@ -1,14 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export BENCHMARK=5 DEBUG=2
-
-python examples/mlperf/model_train.py
@@ -1,15 +0,0 @@
-#!/bin/bash
-
-export PYTHONPATH="." AMD=1
-export MODEL="retinanet"
-export DEFAULT_FLOAT="HALF" GPUS=6 BS=96 EVAL_BS=96
-export BASEDIR="/raid/datasets/openimages"
-
-# export RESET_STEP=0
-
-export TRAIN_BEAM=2 IGNORE_JIT_FIRST_BEAM=1 BEAM_UOPS_MAX=1500 BEAM_UPCAST_MAX=64 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=0
-
-export WANDB=1 PARALLEL=0
-export RUNMLPERF=1
-
-python examples/mlperf/model_train.py
@@ -1,38 +0,0 @@
-{
-    "submitter": "tinycorp",
-    "division": "closed",
-    "status": "Available on-premise",
-    "system_name": "tinybox 8xMI300X",
-    "number_of_nodes": "1",
-    "host_processors_per_node": "2",
-    "host_processor_model_name": "AMD EPYC 9354",
-    "host_processor_core_count": "32",
-    "host_processor_vcpu_count": "64",
-    "host_processor_frequency": "",
-    "host_processor_caches": "",
-    "host_processor_interconnect": "",
-    "host_memory_capacity": "2304GB",
-    "host_storage_type": "NVMe SSD",
-    "host_storage_capacity": "3x 4TB raid array",
-    "host_networking": "",
-    "host_networking_topology": "",
-    "host_memory_configuration": "24x 96GB DDR5",
-    "accelerators_per_node": "8",
-    "accelerator_model_name": "AMD Instinct MI300X 192GB HBM3",
-    "accelerator_host_interconnect": "PCIe 5.0 x16",
-    "accelerator_frequency": "",
-    "accelerator_on-chip_memories": "",
-    "accelerator_memory_configuration": "HBM3",
-    "accelerator_memory_capacity": "192GB",
-    "accelerator_interconnect": "",
-    "accelerator_interconnect_topology": "",
-    "cooling": "air",
-    "hw_notes": "",
-    "framework": "tinygrad, branch mlperf_training_v5.0",
-    "other_software_stack": {
-        "python": "3.10.16",
-        "ROCm": "3.0.0+94441cb"
-    },
-    "operating_system": "Ubuntu 24.04.1 LTS",
-    "sw_notes": ""
-  }
@@ -1,38 +0,0 @@
-{
-  "submitter": "tinycorp",
-  "division": "closed",
-  "status": "Available on-premise",
-  "system_name": "tinybox green",
-  "number_of_nodes": "1",
-  "host_processors_per_node": "1",
-  "host_processor_model_name": "AMD EPYC 7532",
-  "host_processor_core_count": "32",
-  "host_processor_vcpu_count": "64",
-  "host_processor_frequency": "",
-  "host_processor_caches": "",
-  "host_processor_interconnect": "",
-  "host_memory_capacity": "128GB",
-  "host_storage_type": "NVMe SSD",
-  "host_storage_capacity": "4 TB raid array + 1 TB boot",
-  "host_networking": "",
-  "host_networking_topology": "",
-  "host_memory_configuration": "8x 16GB DDR4",
-  "accelerators_per_node": "6",
-  "accelerator_model_name": "NVIDIA GeForce RTX 4090",
-  "accelerator_host_interconnect": "PCIe 4.0 x16",
-  "accelerator_frequency": "",
-  "accelerator_on-chip_memories": "",
-  "accelerator_memory_configuration": "GDDR6X",
-  "accelerator_memory_capacity": "24GB",
-  "accelerator_interconnect": "",
-  "accelerator_interconnect_topology": "",
-  "cooling": "air",
-  "hw_notes": "",
-  "framework": "tinygrad, branch mlperf_training_v5.0",
-  "other_software_stack": {
-    "python": "3.10.12",
-    "CUDA": "12.4"
-  },
-  "operating_system": "Ubuntu 22.04.4",
-  "sw_notes": ""
-}
@@ -1,37 +0,0 @@
-{
-  "submitter": "tinycorp",
-  "division": "closed",
-  "status": "Available on-premise",
-  "system_name": "tinybox red",
-  "number_of_nodes": "1",
-  "host_processors_per_node": "1",
-  "host_processor_model_name": "AMD EPYC 7532",
-  "host_processor_core_count": "32",
-  "host_processor_vcpu_count": "64",
-  "host_processor_frequency": "",
-  "host_processor_caches": "",
-  "host_processor_interconnect": "",
-  "host_memory_capacity": "128GB",
-  "host_storage_type": "NVMe SSD",
-  "host_storage_capacity": "4 TB raid array + 1 TB boot",
-  "host_networking": "",
-  "host_networking_topology": "",
-  "host_memory_configuration": "8x 16GB DDR4",
-  "accelerators_per_node": "6",
-  "accelerator_model_name": "AMD Radeon RX 7900 XTX",
-  "accelerator_host_interconnect": "PCIe 4.0 x16",
-  "accelerator_frequency": "",
-  "accelerator_on-chip_memories": "",
-  "accelerator_memory_configuration": "GDDR6",
-  "accelerator_memory_capacity": "24GB",
-  "accelerator_interconnect": "",
-  "accelerator_interconnect_topology": "",
-  "cooling": "air",
-  "hw_notes": "",
-  "framework": "tinygrad, branch mlperf_training_v5.0",
-  "other_software_stack": {
-    "python": "3.10.12"
-  },
-  "operating_system": "Ubuntu 22.04.4",
-  "sw_notes": ""
-}
@@ -0,0 +1,118 @@
+import json, pprint
+from tinygrad import fetch, nn, Tensor
+from tinygrad.helpers import DEBUG
+
+class FeedForward:
+  def __init__(self, model_dim, intermediate_dim):
+    self.proj_1 = nn.Linear(model_dim, 2*intermediate_dim, bias=False)
+    self.proj_2 = nn.Linear(intermediate_dim, model_dim, bias=False)
+
+  def __call__(self, x):
+    y_12 = self.proj_1(x)
+    y_1, y_2 = y_12.chunk(2, dim=-1)
+    return self.proj_2(y_1.silu() * y_2)
+
+# NOTE: this RoPE doesn't match LLaMA's?
+def _rotate_half(x: Tensor) -> Tensor:
+  x1, x2 = x.chunk(2, dim=-1)
+  return Tensor.cat(-x2, x1, dim=-1)
+
+def _apply_rotary_pos_emb(x: Tensor, pos_sin: Tensor, pos_cos: Tensor) -> Tensor:
+  return (x * pos_cos) + (_rotate_half(x) * pos_sin)
+
+class Attention:
+  def __init__(self, model_dim, num_query_heads, num_kv_heads, head_dim):
+    self.qkv_proj = nn.Linear(model_dim, (num_query_heads + num_kv_heads*2) * head_dim, bias=False)
+    self.num_query_heads, self.num_kv_heads = num_query_heads, num_kv_heads
+    self.head_dim = head_dim
+    self.q_norm = nn.RMSNorm(head_dim)
+    self.k_norm = nn.RMSNorm(head_dim)
+    self.out_proj = nn.Linear(num_query_heads * head_dim, model_dim, bias=False)
+
+  def __call__(self, x:Tensor) -> Tensor:
+    batch_size, seq_len, embed_dim = x.shape
+    qkv = self.qkv_proj(x)
+    qkv = qkv.reshape(batch_size, seq_len, self.num_query_heads+self.num_kv_heads*2, self.head_dim).transpose(1, 2)
+    xq,xk,xv = qkv.split([self.num_query_heads, self.num_kv_heads, self.num_kv_heads], dim=1)
+    xq = self.q_norm(xq)
+    xk = self.k_norm(xk)
+
+    # add positional embedding (how many kernels is this?)
+    freq_constant = 10000
+    inv_freq = 1.0 / (freq_constant ** (Tensor.arange(0, self.head_dim, 2) / self.head_dim))
+    pos_index_theta = Tensor.einsum("i,j->ij", Tensor.arange(seq_len), inv_freq)
+    emb = Tensor.cat(pos_index_theta, pos_index_theta, dim=-1)
+    cos_emb, sin_emb = emb.cos()[None, None, :, :], emb.sin()[None, None, :, :]
+    xq = _apply_rotary_pos_emb(xq, sin_emb, cos_emb)
+    xk = _apply_rotary_pos_emb(xk, sin_emb, cos_emb)
+
+    # grouped-query attention
+    num_groups = self.num_query_heads // self.num_kv_heads
+    xk = xk.repeat_interleave(num_groups, dim=1)
+    xv = xv.repeat_interleave(num_groups, dim=1)
+
+    # masked attention
+    #start_pos = 0
+    #mask = Tensor.full((1, 1, seq_len, start_pos+seq_len), float("-inf"), dtype=xq.dtype, device=xq.device).triu(start_pos+1)
+    #attn_output = xq.scaled_dot_product_attention(xk, xv, mask).transpose(1, 2)
+
+    # causal is fine, no mask needed
+    attn_output = xq.scaled_dot_product_attention(xk, xv, is_causal=True).transpose(1, 2)
+    return self.out_proj(attn_output.reshape(batch_size, seq_len, self.num_query_heads * self.head_dim))
+
+class Layer:
+  def __init__(self, model_dim, intermediate_dim, num_query_heads, num_kv_heads, head_dim):
+    self.ffn = FeedForward(model_dim, intermediate_dim)
+    self.attn = Attention(model_dim, num_query_heads, num_kv_heads, head_dim)
+    self.ffn_norm = nn.RMSNorm(model_dim)
+    self.attn_norm = nn.RMSNorm(model_dim)
+
+  def __call__(self, x:Tensor) -> Tensor: # (batch, seq_len, embed_dim)
+    x = x + self.attn(self.attn_norm(x))
+    x = x + self.ffn(self.ffn_norm(x))
+    return x
+
+# stupidly complex
+def make_divisible(v, divisor):
+  new_v = max(divisor, int(v + divisor / 2) // divisor * divisor)
+  if new_v < 0.9 * v: new_v += divisor
+  return new_v
+
+class Transformer:
+  def __init__(self, cfg):
+    if DEBUG >= 3: pprint.pp(cfg)
+    self.layers = [Layer(cfg['model_dim'], make_divisible(int(cfg["model_dim"] * cfg['ffn_multipliers'][i]), cfg['ffn_dim_divisor']),
+                         cfg['num_query_heads'][i], cfg['num_kv_heads'][i], cfg['head_dim']) for i in range(cfg['num_transformer_layers'])]
+    self.norm = nn.RMSNorm(cfg['model_dim'])
+    self.token_embeddings = nn.Embedding(cfg['vocab_size'], cfg['model_dim'])
+
+  def __call__(self, tokens:Tensor):
+    # _bsz, seqlen = tokens.shape
+    x = self.token_embeddings(tokens)
+    for l in self.layers: x = l(x)
+    return self.norm(x) @ self.token_embeddings.weight.T
+
+if __name__ == "__main__":
+  #model_name = "OpenELM-270M-Instruct"
+  model_name = "OpenELM-270M"  # this is fp32
+  model = Transformer(json.loads(fetch(f"https://huggingface.co/apple/{model_name}/resolve/main/config.json?download=true").read_bytes()))
+  weights = nn.state.safe_load(fetch(f"https://huggingface.co/apple/{model_name}/resolve/main/model.safetensors?download=true"))
+  if DEBUG >= 3:
+    for k, v in weights.items(): print(k, v.shape)
+  nn.state.load_state_dict(model, {k.removeprefix("transformer."):v for k,v in weights.items()})
+
+  from sentencepiece import SentencePieceProcessor
+  tokenizer = SentencePieceProcessor(fetch("https://github.com/karpathy/llama2.c/raw/master/tokenizer.model").as_posix())
+  toks = [tokenizer.bos_id()] + tokenizer.encode("Some car brands include")
+  for i in range(100):
+    ttoks = Tensor([toks])
+    out = model(ttoks).realize()
+    t0 = out[0].argmax(axis=-1).tolist()
+    toks.append(t0[-1])
+    # hmmm...passthrough still doesn't match (it shouldn't, it outputs the most likely)
+    print(tokenizer.decode(toks))
+    #print(toks)
+    #print(tokenizer.decode(t0))
+    #print(t0)
+
+
@@ -1,11 +1,16 @@
-import os, sys, pickle, time, re
+import os, sys, pickle, time
 import numpy as np
+if "FLOAT16" not in os.environ: os.environ["FLOAT16"] = "1"
+if "IMAGE" not in os.environ: os.environ["IMAGE"] = "2"
+if "NOLOCALS" not in os.environ: os.environ["NOLOCALS"] = "1"
 if "JIT_BATCH_SIZE" not in os.environ: os.environ["JIT_BATCH_SIZE"] = "0"

 from tinygrad import fetch, Tensor, TinyJit, Context, GlobalCounters, Device, dtypes
 from tinygrad.helpers import DEBUG, getenv
 from tinygrad.engine.realize import CompiledRunner
-from tinygrad.nn.onnx import OnnxRunner
+
+import onnx
+from tinygrad.frontend.onnx import OnnxRunner

 OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
 OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
@@ -16,14 +21,11 @@ def compile(onnx_file):

  input_shapes = {name: spec.shape for name, spec in run_onnx.graph_inputs.items()}
  input_types = {name: spec.dtype for name, spec in run_onnx.graph_inputs.items()}
-
  # Float inputs and outputs to tinyjits for openpilot are always float32
-  # TODO this seems dumb
  input_types = {k:(dtypes.float32 if v is dtypes.float16 else v) for k,v in input_types.items()}
  Tensor.manual_seed(100)
-  inputs = {k:Tensor(Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize().numpy(), device='NPY') for k,shp in sorted(input_shapes.items())}
-  if not getenv("NPY_IMG"):
-    inputs = {k:Tensor(v.numpy(), device=Device.DEFAULT).realize() if 'img' in k else v for k,v in inputs.items()}
+  new_inputs = {k:Tensor.randn(*shp, dtype=input_types[k]).mul(8).realize() for k,shp in sorted(input_shapes.items())}
+  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
  print("created tensors")

  run_onnx_jit = TinyJit(lambda **kwargs:
@@ -31,6 +33,8 @@ def compile(onnx_file):
  for i in range(3):
    GlobalCounters.reset()
    print(f"run {i}")
+    inputs = {**{k:v.clone() for k,v in new_inputs.items() if 'img' in k},
+              **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}
    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1)):
      ret = run_onnx_jit(**inputs).numpy()
    # copy i == 1 so use of JITBEAM is okay
@@ -39,7 +43,7 @@ def compile(onnx_file):
  np.testing.assert_equal(test_val, ret, "JIT run failed")
  print("jit run validated")

-  # check gated read_image usage
+  # checks from compile2
  kernel_count = 0
  read_image_count = 0
  gated_read_image_count = 0
@@ -48,8 +52,6 @@ def compile(onnx_file):
      kernel_count += 1
      read_image_count += ei.prg.p.src.count("read_image")
      gated_read_image_count += ei.prg.p.src.count("?read_image")
-      for v in [m.group(1) for m in re.finditer(r'(val\d+)\s*=\s*read_imagef\(', ei.prg.p.src)]:
-        if len(re.findall(fr'[\?\:]{v}\.[xyzw]', ei.prg.p.src)) > 0: gated_read_image_count += 1
  print(f"{kernel_count=},  {read_image_count=}, {gated_read_image_count=}")
  if (allowed_kernel_count:=getenv("ALLOWED_KERNEL_COUNT", -1)) != -1:
    assert kernel_count == allowed_kernel_count, f"different kernels! {kernel_count=}, {allowed_kernel_count=}"
@@ -65,77 +67,85 @@ def compile(onnx_file):
  print(f"mdl size is {mdl_sz/1e6:.2f}M")
  print(f"pkl size is {pkl_sz/1e6:.2f}M")
  print("**** compile done ****")
-  return inputs, test_val
+  return test_val

-def test_vs_compile(run, inputs, test_val=None):
+def test_vs_compile(run, new_inputs, test_val=None):
+  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
+
+  # create fake "from_blob" tensors for the inputs, and wrapped NPY tensors for the numpy inputs (these have the same underlying memory)
+  inputs = {**{k:v for k,v in new_inputs.items() if 'img' in k},
+            **{k:Tensor(v, device="NPY").realize() for k,v in new_inputs_numpy.items() if 'img' not in k}}

  # run 20 times
-  step_times = []
  for _ in range(20):
    st = time.perf_counter()
    out = run(**inputs)
    mt = time.perf_counter()
    val = out.numpy()
    et = time.perf_counter()
-    step_times.append((et-st)*1e3)
-    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {step_times[-1]:6.2f} ms")
-
-  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-    min_time = min(step_times)
-    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
-
+    print(f"enqueue {(mt-st)*1e3:6.2f} ms -- total run {(et-st)*1e3:6.2f} ms")
+  print(out, val.shape, val.dtype)
  if test_val is not None: np.testing.assert_equal(test_val, val)
  print("**** test done ****")

  # test that changing the numpy changes the model outputs
-  inputs_2x = {k: Tensor(v.numpy()*2, device=v.device) for k,v in inputs.items()}
-  out = run(**inputs_2x)
-  changed_val = out.numpy()
-  np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, val, changed_val)
+  if any([x.device == 'NPY' for x in inputs.values()]):
+    for v in new_inputs_numpy.values(): v *= 2
+    out = run(**inputs)
+    changed_val = out.numpy()
+    np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, val, changed_val)
  return val

-def test_vs_onnx(new_inputs, test_val, onnx_file, tol):
-  import onnx
-  import onnxruntime as ort
-  
-  onnx_inputs = {k:v.numpy() for k,v in new_inputs.items()}
+def test_vs_onnx(new_inputs, test_val, onnx_file, ort=False):
+  new_inputs_numpy = {k:v.numpy() for k,v in new_inputs.items()}
  onnx_model = onnx.load(onnx_file)

-  ORT_TO_NP_DTYPES: dict[str, np.dtype] = {
-    'tensor(float)': np.dtype('float32'),
-    'tensor(float16)': np.dtype('float16'),
-    'tensor(uint8)': np.dtype('uint8'),
-  }
-
  timings = []
-  onnx_session = ort.InferenceSession(onnx_file)
-  onnx_types = {x.name: ORT_TO_NP_DTYPES[x.type] for x in onnx_session.get_inputs()}
-  onnx_inputs = {k:onnx_inputs[k].astype(onnx_types[k]) for k in onnx_inputs}
+  if ort:
+    # test with onnxruntime
+    import onnxruntime as ort
+    onnx_session = ort.InferenceSession(onnx_file)
+    for _ in range(1 if test_val is not None else 5):
+      st = time.perf_counter()
+      onnx_output = onnx_session.run([onnx_model.graph.output[0].name], {k:v.astype(np.float16) for k,v in new_inputs_numpy.items()})
+      timings.append(time.perf_counter() - st)
+    new_torch_out = onnx_output[0]
+  else:
+    # test with torch
+    import torch
+    from onnx2torch import convert
+    inputs = {k.name:new_inputs_numpy[k.name] for k in onnx_model.graph.input}
+    torch_model = convert(onnx_model).float()
+    with torch.no_grad():
+      for _ in range(1 if test_val is not None else 5):
+        st = time.perf_counter()
+        torch_out = torch_model(*[torch.tensor(x) for x in inputs.values()])
+        timings.append(time.perf_counter() - st)
+      new_torch_out = torch_out.numpy()

-  for _ in range(1 if test_val is not None else 5):
-    st = time.perf_counter()
-    onnx_output = onnx_session.run([onnx_model.graph.output[0].name], onnx_inputs)
-    timings.append(time.perf_counter() - st)
-
-  np.testing.assert_allclose(onnx_output[0].reshape(test_val.shape), test_val, atol=tol, rtol=tol)
-  print("test vs onnx passed")
+  if test_val is not None:
+    np.testing.assert_allclose(new_torch_out.reshape(test_val.shape), test_val, atol=1e-4, rtol=1e-2)
+    print("test vs onnx passed")
  return timings

-def bench(run, inputs):
-  from extra.bench_log import WallTimeEvent, BenchEvent
-  for _ in range(10):
-    with WallTimeEvent(BenchEvent.STEP):
-      run(**inputs).numpy()
-
 if __name__ == "__main__":
  onnx_file = fetch(OPENPILOT_MODEL)
-  inputs, outputs = compile(onnx_file)
+  test_val = compile(onnx_file) if not getenv("RUN") else None

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

-  test_vs_compile(pickle_loaded, inputs, outputs)
-  if getenv("SELFTEST"):
-    test_vs_onnx(inputs, outputs, onnx_file, 1e-4)
+  # same randomness as compile
+  Tensor.manual_seed(100)
+  new_inputs = {nm:Tensor.randn(*st.shape, dtype=dtype).mul(8).realize() for nm, (st, _, dtype, _) in
+                sorted(zip(pickle_loaded.captured.expected_names, pickle_loaded.captured.expected_st_vars_dtype_device))}
+
+  test_val = test_vs_compile(pickle_loaded, new_inputs, test_val)
+  if getenv("BENCHMARK"):
+    for be in ["torch", "ort"]:
+      try:
+        timings = test_vs_onnx(new_inputs, None, onnx_file, be=="ort")
+        print(f"timing {be}: {min(timings)*1000:.2f} ms")
+      except Exception as e:
+        print(f"{be} fail with {e}")
+  if not getenv("FLOAT16"): test_vs_onnx(new_inputs, test_val, onnx_file, getenv("ORT"))

-  if getenv("BENCHMARK_LOG", ""):
-    bench(pickle_loaded, inputs)
@@ -0,0 +1,47 @@
+import sys
+from tinygrad import Tensor, fetch, GlobalCounters, dtypes
+from tinygrad.uop.ops import UOp
+from tinygrad.frontend.onnx import OnnxRunner
+from tinygrad.schedule.kernelize import get_kernelize_map
+from tinygrad.engine.schedule import create_schedule_with_vars
+from tinygrad.engine.realize import run_schedule
+
+# NOLOCALS=1 GPU=1 IMAGE=2 FLOAT16=1 VIZ=1 DEBUG=2 python3 examples/openpilot/compile4.py
+
+OPENPILOT_MODEL = sys.argv[1] if len(sys.argv) > 1 else "https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx"
+OUTPUT = sys.argv[2] if len(sys.argv) > 2 else "/tmp/openpilot.pkl"
+
+if __name__ == "__main__":
+  onnx_file = fetch(OPENPILOT_MODEL)
+  run_onnx = OnnxRunner(onnx_file)
+
+  inputs = run_onnx.get_empty_input_data("npy", dtypes.float32)
+  out: Tensor = next(iter(run_onnx({k:v.to(None) for k,v in inputs.items()}).values())).to('cpu')
+  root = out.uop
+  targets = [x.uop for x in inputs.values()]
+  print(targets)
+
+  # TODO: abstract this from gradient?
+
+  # compute the target path (top down)
+  in_target_path: dict[UOp, bool] = {}
+  for u in root.toposort(): in_target_path[u] = any(x in targets or in_target_path[x] for x in u.src)
+  independent_set = {}
+  for u in root.toposort():
+    if in_target_path[u]:
+      for s in u.src:
+        if not in_target_path[s]:
+          independent_set[s] = None
+  independent = UOp.sink(*independent_set.keys())
+  kernelized = get_kernelize_map(independent)
+  independent = independent.substitute(kernelized)
+  schedule, var_vals = create_schedule_with_vars(independent)
+  run_schedule(schedule)
+
+  print("**** real ****")
+  GlobalCounters.reset()
+  out.uop = root.substitute(kernelized)
+  out.kernelize()
+
+  # realize
+  out.realize()
@@ -0,0 +1,55 @@
+from tinygrad.helpers import trange
+from tinygrad.nn.datasets import mnist
+import mlx.core as mx
+import mlx.nn as nn
+import mlx.optimizers as optim
+from functools import partial
+
+class Model(nn.Module):
+  def __init__(self):
+    super().__init__()
+    self.c1 = nn.Conv2d(1, 32, 5)
+    self.c2 = nn.Conv2d(32, 32, 5)
+    self.bn1 = nn.BatchNorm(32)
+    self.m1 = nn.MaxPool2d(2)
+    self.c3 = nn.Conv2d(32, 64, 3)
+    self.c4 = nn.Conv2d(64, 64, 3)
+    self.bn2 = nn.BatchNorm(64)
+    self.m2 = nn.MaxPool2d(2)
+    self.lin = nn.Linear(576, 10)
+  def __call__(self, x):
+    x = mx.maximum(self.c1(x), 0)
+    x = mx.maximum(self.c2(x), 0)
+    x = self.m1(self.bn1(x))
+    x = mx.maximum(self.c3(x), 0)
+    x = mx.maximum(self.c4(x), 0)
+    x = self.m2(self.bn2(x))
+    return self.lin(mx.flatten(x, 1))
+
+if __name__ == "__main__":
+  X_train, Y_train, X_test, Y_test = mnist()
+  X_train = mx.array(X_train.float().permute((0,2,3,1)).numpy())
+  Y_train = mx.array(Y_train.numpy())
+  X_test = mx.array(X_test.float().permute((0,2,3,1)).numpy())
+  Y_test = mx.array(Y_test.numpy())
+
+  model = Model()
+  optimizer = optim.Adam(1e-3)
+  def loss_fn(model, x, y): return nn.losses.cross_entropy(model(x), y).mean()
+
+  state = [model.state, optimizer.state]
+  @partial(mx.compile, inputs=state, outputs=state)
+  def step(samples):
+    # Compiled functions will also treat any inputs not in the parameter list as constants.
+    X,Y = X_train[samples], Y_train[samples]
+    loss_and_grad_fn = nn.value_and_grad(model, loss_fn)
+    loss, grads = loss_and_grad_fn(model, X, Y)
+    optimizer.update(model, grads)
+    return loss
+
+  test_acc = float('nan')
+  for i in (t:=trange(70)):
+    samples = mx.random.randint(0, X_train.shape[0], (512,))  # putting this in JIT didn't work well
+    loss = step(samples)
+    if i%10 == 9: test_acc = ((model(X_test).argmax(axis=-1) == Y_test).sum() * 100 / X_test.shape[0]).item()
+    t.set_description(f"loss: {loss.item():6.2f} test_accuracy: {test_acc:5.2f}%")
@@ -27,7 +27,7 @@ class Model(nn.Module):

 if __name__ == "__main__":
  if getenv("TINY_BACKEND"):
-    import tinygrad.nn.torch  # noqa: F401
+    import tinygrad.frontend.torch  # noqa: F401
    device = torch.device("tiny")
  else:
    device = torch.device({"METAL":"mps","NV":"cuda"}.get(Device.DEFAULT, "cpu"))
@@ -8,7 +8,7 @@ from typing import Dict, Union

 from extra.models.llama import Transformer, convert_from_huggingface, fix_bf16
 from examples.llama3 import load
-from tinygrad import nn, Tensor, Device
+from tinygrad import nn, Tensor
 from tinygrad.helpers import fetch, colored, GlobalCounters, Timing, DEBUG
 from tinygrad.nn.state import load_state_dict, get_parameters

@@ -80,7 +80,7 @@ if __name__ == "__main__":
    st = GlobalCounters.time_sum_s
    next_tok = Tensor([toks[start_pos:]]) if tok_tensor is None or (len(toks)-start_pos) > 1 else tok_tensor.reshape(1, 1)
    with Timing("total ", enabled=args.timing, on_exit=lambda x: f", {1e9/x:.2f} tok/s, {GlobalCounters.global_mem/x:.2f} GB/s, param {param_bytes/x:.2f} GB/s"):
-      with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" if DEBUG>=2 else "") +
+      with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU" if DEBUG>=2 else "") +
                  f", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB" +
                  (f", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s, param {param_bytes*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s" if DEBUG>=2 else "")) if DEBUG else None, enabled=args.timing):
        tok_tensor = transformer(next_tok, start_pos, args.temperature)
@@ -0,0 +1,45 @@
+import gymnasium as gym
+import numpy as np
+from gymnasium.envs.registration import register
+
+# a very simple game
+# one of <size> lights will light up
+# take the action of the lit up light
+# in <hard_mode>, you act differently based on the step number and need to track this
+
+class PressTheLightUpButton(gym.Env):
+  metadata = {"render_modes": []}
+  def __init__(self, render_mode=None, size=2, game_length=10, hard_mode=False):
+    self.size, self.game_length = size, game_length
+    self.observation_space = gym.spaces.Box(0, 1, shape=(self.size,), dtype=np.float32)
+    self.action_space = gym.spaces.Discrete(self.size)
+    self.step_num = 0
+    self.done = True
+    self.hard_mode = hard_mode
+
+  def _get_obs(self):
+    obs = [0]*self.size
+    if self.step_num < len(self.state):
+      obs[self.state[self.step_num]] = 1
+    return np.array(obs, dtype=np.float32)
+
+  def reset(self, seed=None, options=None):
+    super().reset(seed=seed)
+    self.state = np.random.randint(0, self.size, size=self.game_length)
+    self.step_num = 0
+    self.done = False
+    return self._get_obs(), {}
+
+  def step(self, action):
+    target = ((action + self.step_num) % self.size) if self.hard_mode else action
+    reward = int(target == self.state[self.step_num])
+    self.step_num += 1
+    if not reward:
+      self.done = True
+    return self._get_obs(), reward, self.done, self.step_num >= self.game_length, {}
+
+register(
+  id="PressTheLightUpButton-v0",
+  entry_point="examples.rl.lightupbutton:PressTheLightUpButton",
+  max_episode_steps=None,
+)
@@ -99,7 +99,6 @@ if __name__ == "__main__":
  parser.add_argument('--timing',      action='store_true', help="Print timing per step")
  parser.add_argument('--noshow',      action='store_true', help="Don't show the image")
  parser.add_argument('--fp16',        action='store_true', help="Cast the weights to float16")
-  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
  args = parser.parse_args()

  N = 1
@@ -113,22 +112,19 @@ if __name__ == "__main__":

  model = StableDiffusionV2(**params)

-  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    if not args.fakeweights:
-      default_weights_url = 'https://huggingface.co/sd2-community/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
-      weights_fn = args.weights_fn
-      if not weights_fn:
-        weights_url = args.weights_url if args.weights_url else default_weights_url
-        weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))
+  default_weights_url = 'https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-ema-pruned.safetensors'
+  weights_fn = args.weights_fn
+  if not weights_fn:
+    weights_url = args.weights_url if args.weights_url else default_weights_url
+    weights_fn  = fetch(weights_url, os.path.basename(str(weights_url)))

-        load_state_dict(model, safe_load(weights_fn), strict=False)
+  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
+    load_state_dict(model, safe_load(weights_fn), strict=False)

    if args.fp16:
      for k,v in get_state_dict(model).items():
        if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16))
-
-    Tensor.realize(*get_state_dict(model).values())
+          v.replace(v.cast(dtypes.float16).realize())

  c  = { "crossattn": model.cond_stage_model(args.prompt) }
  uc = { "crossattn": model.cond_stage_model("") }
@@ -6,7 +6,7 @@
 from tinygrad import Tensor, TinyJit, dtypes, GlobalCounters
 from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import safe_load, load_state_dict
-from tinygrad.helpers import fetch, trange, colored, Timing, getenv
+from tinygrad.helpers import fetch, trange, colored, Timing
 from extra.models.clip import Embedder, FrozenClosedClipEmbedder, FrozenOpenClipEmbedder
 from extra.models.unet import UNetModel, Upsample, Downsample, timestep_embedding
 from extra.bench_log import BenchEvent, WallTimeEvent
@@ -14,7 +14,7 @@ from examples.stable_diffusion import ResnetBlock, Mid
 import numpy as np

 from typing import Dict, List, Callable, Optional, Any, Set, Tuple, Union, Type
-import argparse, tempfile, time
+import argparse, tempfile
 from abc import ABC, abstractmethod
 from pathlib import Path
 from PIL import Image
@@ -342,13 +342,11 @@ class DPMPP2MSampler:
    sigmas = self.discretization(num_steps).to(x.device)
    x *= Tensor.sqrt(1.0 + sigmas[0] ** 2.0)
    num_sigmas = len(sigmas)
-    step_times = []

    old_denoised = None
    for i in trange(num_sigmas - 1):
      with Timing("step in ", enabled=timing, on_exit=lambda _: f", using {GlobalCounters.mem_used/1e9:.2f} GB"):
        GlobalCounters.reset()
-        st = time.perf_counter_ns()
        with WallTimeEvent(BenchEvent.STEP):
          x, old_denoised = self.sampler_step(
            old_denoised=old_denoised,
@@ -360,13 +358,8 @@ class DPMPP2MSampler:
            c=c,
            uc=uc,
          )
-          step_times.append(t:=(time.perf_counter_ns() - st)*1e-6)
          x.realize(old_denoised)

-    if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-      min_time = min(step_times)
-      assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
-
    return x


@@ -437,8 +430,8 @@ if __name__ == "__main__":
    im.show()

  # validation!
-  is_default = args.prompt == default_prompt and args.steps == 10 and args.seed == 0 and args.guidance == 6.0 and args.width == args.height == 1024
-  if is_default and not args.weights and not args.fakeweights:
+  if args.prompt == default_prompt and args.steps == 10 and args.seed == 0 and args.guidance == 6.0 and args.width == args.height == 1024 \
+    and not args.weights:
    ref_image = Tensor(np.array(Image.open(Path(__file__).parent / "sdxl_seed0.png")))
    distance = (((x.cast(dtypes.float) - ref_image.cast(dtypes.float)) / ref_image.max())**2).mean().item()
    assert distance < 4e-3, colored(f"validation failed with {distance=}", "red")
@@ -0,0 +1,48 @@
+import os, pathlib, argparse
+from examples.llama3 import Tokenizer
+from tabulate import tabulate
+from tinygrad import fetch
+from tinygrad.helpers import flatten
+
+# llama 3 tokenizer
+tokenizer = Tokenizer(fetch("https://huggingface.co/bofenghuang/Meta-Llama-3-8B/resolve/main/original/tokenizer.model").as_posix())
+
+def read_code(base_path):
+  ret = []
+  for path, _, files in os.walk(os.path.join(base_path, "tinygrad")):
+    for name in files:
+      if not name.endswith(".py"): continue
+      if 'tinygrad/runtime/autogen' in path.replace('\\', '/'): continue
+      fullpath = os.path.join(path, name)
+      code = pathlib.Path(fullpath).read_text()
+      ret.append(("### " + fullpath.split("tinygrad/", 1)[1], code))
+  return ret
+
+def write_code_to_file(filename, code_list):
+  """Writes the combined code to a specified file."""
+  with open(filename, 'w') as f:
+    f.write('\n'.join(flatten(code_list)))
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(description="Analyze and optionally save tinygrad code.")
+  parser.add_argument("--output", help="Output file to write the combined code to.")
+  args = parser.parse_args()
+
+  ret = read_code(".")
+
+  table = []
+  for name,code in ret:
+    table.append([name, len(tokenizer.encode(name+"\x00"+code))])
+  print(tabulate([["name", "llm tokens"]]+sorted(table, key=lambda x: -x[1]), headers="firstrow"))
+
+  code_str = '\x00'.join(flatten(ret))
+  print(f"code has {len(code_str)} chars")
+  newline_count = code_str.count('\n')
+  print(f"code has {newline_count} newlines")
+
+  encoded = tokenizer.encode(code_str)
+  print(f"code has {len(encoded)} tokens")
+
+  if args.output:
+    write_code_to_file(args.output, ret)
+    print(f"Combined code written to {args.output}")
@@ -0,0 +1,136 @@
+#!/usr/bin/env python
+#inspired by https://github.com/Matuzas77/MNIST-0.17/blob/master/MNIST_final_solution.ipynb
+import sys
+import numpy as np
+from tinygrad.nn.state import get_parameters
+from tinygrad.tensor import Tensor
+from tinygrad.nn import BatchNorm2d, optim
+from tinygrad.helpers import getenv
+from extra.datasets import fetch_mnist
+from extra.augment import augment_img
+from extra.training import train, evaluate
+GPU = getenv("GPU")
+QUICK = getenv("QUICK")
+DEBUG = getenv("DEBUG")
+
+class SqueezeExciteBlock2D:
+  def __init__(self, filters):
+    self.filters = filters
+    self.weight1 = Tensor.scaled_uniform(self.filters, self.filters//32)
+    self.bias1 = Tensor.scaled_uniform(1,self.filters//32)
+    self.weight2 = Tensor.scaled_uniform(self.filters//32, self.filters)
+    self.bias2 = Tensor.scaled_uniform(1, self.filters)
+
+  def __call__(self, input):
+    se = input.avg_pool2d(kernel_size=(input.shape[2], input.shape[3])) #GlobalAveragePool2D
+    se = se.reshape(shape=(-1, self.filters))
+    se = se.dot(self.weight1) + self.bias1
+    se = se.relu()
+    se = se.dot(self.weight2) + self.bias2
+    se = se.sigmoid().reshape(shape=(-1,self.filters,1,1)) #for broadcasting
+    se = input.mul(se)
+    return se
+
+class ConvBlock:
+  def __init__(self, h, w, inp, filters=128, conv=3):
+    self.h, self.w = h, w
+    self.inp = inp
+    #init weights
+    self.cweights = [Tensor.scaled_uniform(filters, inp if i==0 else filters, conv, conv) for i in range(3)]
+    self.cbiases = [Tensor.scaled_uniform(1, filters, 1, 1) for i in range(3)]
+    #init layers
+    self._bn = BatchNorm2d(128)
+    self._seb = SqueezeExciteBlock2D(filters)
+
+  def __call__(self, input):
+    x = input.reshape(shape=(-1, self.inp, self.w, self.h))
+    for cweight, cbias in zip(self.cweights, self.cbiases):
+      x = x.pad(padding=[1,1,1,1]).conv2d(cweight).add(cbias).relu()
+    x = self._bn(x)
+    x = self._seb(x)
+    return x
+
+class BigConvNet:
+  def __init__(self):
+    self.conv = [ConvBlock(28,28,1), ConvBlock(28,28,128), ConvBlock(14,14,128)]
+    self.weight1 = Tensor.scaled_uniform(128,10)
+    self.weight2 = Tensor.scaled_uniform(128,10)
+
+  def parameters(self):
+    if DEBUG: #keeping this for a moment
+      pars = [par for par in get_parameters(self) if par.requires_grad]
+      no_pars = 0
+      for par in pars:
+        print(par.shape)
+        no_pars += np.prod(par.shape)
+      print('no of parameters', no_pars)
+      return pars
+    else:
+      return get_parameters(self)
+
+  def save(self, filename):
+    with open(filename+'.npy', 'wb') as f:
+      for par in get_parameters(self):
+        #if par.requires_grad:
+        np.save(f, par.numpy())
+
+  def load(self, filename):
+    with open(filename+'.npy', 'rb') as f:
+      for par in get_parameters(self):
+        #if par.requires_grad:
+        try:
+          par.numpy()[:] = np.load(f)
+          if GPU:
+            par.gpu()
+        except:
+          print('Could not load parameter')
+
+  def forward(self, x):
+    x = self.conv[0](x)
+    x = self.conv[1](x)
+    x = x.avg_pool2d(kernel_size=(2,2))
+    x = self.conv[2](x)
+    x1 = x.avg_pool2d(kernel_size=(14,14)).reshape(shape=(-1,128)) #global
+    x2 = x.max_pool2d(kernel_size=(14,14)).reshape(shape=(-1,128)) #global
+    xo = x1.dot(self.weight1) + x2.dot(self.weight2)
+    return xo
+
+
+if __name__ == "__main__":
+  lrs = [1e-4, 1e-5] if QUICK else [1e-3, 1e-4, 1e-5, 1e-5]
+  epochss = [2, 1] if QUICK else [13, 3, 3, 1]
+  BS = 32
+
+  lmbd = 0.00025
+  lossfn = lambda out,y: out.sparse_categorical_crossentropy(y) + lmbd*(model.weight1.abs() + model.weight2.abs()).sum()
+  X_train, Y_train, X_test, Y_test = fetch_mnist()
+  X_train = X_train.reshape(-1, 28, 28).astype(np.uint8)
+  X_test = X_test.reshape(-1, 28, 28).astype(np.uint8)
+  steps = len(X_train)//BS
+  np.random.seed(1337)
+  if QUICK:
+    steps = 1
+    X_test, Y_test = X_test[:BS], Y_test[:BS]
+
+  model = BigConvNet()
+
+  if len(sys.argv) > 1:
+    try:
+      model.load(sys.argv[1])
+      print('Loaded weights "'+sys.argv[1]+'", evaluating...')
+      evaluate(model, X_test, Y_test, BS=BS)
+    except:
+      print('could not load weights "'+sys.argv[1]+'".')
+
+  if GPU:
+    params = get_parameters(model)
+    [x.gpu_() for x in params]
+
+  for lr, epochs in zip(lrs, epochss):
+    optimizer = optim.Adam(model.parameters(), lr=lr)
+    for epoch in range(1,epochs+1):
+      #first epoch without augmentation
+      X_aug = X_train if epoch == 1 else augment_img(X_train)
+      train(model, X_aug, Y_train, optimizer, steps=steps, lossfn=lossfn, BS=BS)
+      accuracy = evaluate(model, X_test, Y_test, BS=BS)
+      model.save(f'examples/checkpoint{accuracy * 1e6:.0f}')
@@ -0,0 +1,17 @@
+from tinygrad.tensor import Tensor
+from tinygrad.nn import Conv2d, BatchNorm2d
+from tinygrad.nn.state import get_parameters
+
+if __name__ == "__main__":
+  with Tensor.train():
+
+    BS, C1, H, W = 4, 16, 224, 224
+    C2, K, S, P = 64, 7, 2, 1
+
+    x = Tensor.uniform(BS, C1, H, W)
+    conv = Conv2d(C1, C2, kernel_size=K, stride=S, padding=P)
+    bn = BatchNorm2d(C2, track_running_stats=False)
+    for t in get_parameters([x, conv, bn]): t.realize()
+
+    print("running network")
+    x.sequential([conv, bn]).numpy()
@@ -0,0 +1,669 @@
+# original implementation: https://github.com/svc-develop-team/so-vits-svc
+from __future__ import annotations
+import sys, logging, time, io, math, argparse, operator, numpy as np
+from functools import partial, reduce
+from pathlib import Path
+from typing import Tuple, Optional, Type
+from tinygrad import nn, dtypes, Tensor
+from tinygrad.helpers import getenv, fetch
+from tinygrad.nn.state import torch_load
+from examples.vits import ResidualCouplingBlock, PosteriorEncoder, Encoder, ResBlock1, ResBlock2, LRELU_SLOPE, sequence_mask, split, get_hparams_from_file, load_checkpoint, weight_norm, HParams
+from examples.sovits_helpers import preprocess
+import soundfile
+
+DEBUG = getenv("DEBUG")
+
+F0_BIN = 256
+F0_MAX = 1100.0
+F0_MIN = 50.0
+F0_MEL_MIN = 1127 * np.log(1 + F0_MIN / 700)
+F0_MEL_MAX = 1127 * np.log(1 + F0_MAX / 700)
+
+class SpeechEncoder:
+  def __init__(self, hidden_dim, model:ContentVec): self.hidden_dim, self.model = hidden_dim, model
+  def encode(self, ): raise NotImplementedError("implement me")
+  @classmethod
+  def load_from_pretrained(cls, checkpoint_path:str, checkpoint_url:str) -> ContentVec:
+    contentvec = ContentVec.load_from_pretrained(checkpoint_path, checkpoint_url)
+    return cls(contentvec)
+
+class ContentVec256L9(SpeechEncoder):
+  def __init__(self, model:ContentVec): super().__init__(hidden_dim=256, model=model)
+  def encode(self, wav: Tensor):
+    feats = wav
+    if len(feats.shape) == 2:  # double channels
+      feats = feats.mean(-1)
+    assert len(feats.shape) == 1, feats.dim()
+    feats = feats.reshape(1, -1)
+    padding_mask = Tensor.zeros_like(feats).cast(dtypes.bool)
+    logits = self.model.extract_features(feats.to(wav.device), padding_mask=padding_mask.to(wav.device), output_layer=9)
+    feats = self.model.final_proj(logits[0])
+    return feats.transpose(1,2)
+
+class ContentVec768L12(SpeechEncoder):
+  def __init__(self, model:ContentVec): super().__init__(hidden_dim=768, model=model)
+  def encode(self, wav: Tensor):
+    feats = wav
+    if len(feats.shape) == 2:  # double channels
+      feats = feats.mean(-1)
+    assert len(feats.shape) == 1, feats.dim()
+    feats = feats.reshape(1, -1)
+    padding_mask = Tensor.zeros_like(feats).cast(dtypes.bool)
+    logits = self.model.extract_features(feats.to(wav.device), padding_mask=padding_mask.to(wav.device), output_layer=12)
+    return logits[0].transpose(1,2)
+
+# original code for contentvec: https://github.com/auspicious3000/contentvec/
+class ContentVec:
+  # self.final_proj dims are hardcoded and depend on fairseq.data.dictionary Dictionary in the checkpoint. This param can't yet be loaded since there is no pickle for it. See with DEBUG=2.
+  # This means that the ContentVec only works with the hubert weights used in all SVC models
+  def __init__(self, cfg: HParams):
+    self.feature_grad_mult, self.untie_final_proj = cfg.feature_grad_mult, cfg.untie_final_proj
+    feature_enc_layers = eval(cfg.conv_feature_layers)
+    self.embed = feature_enc_layers[-1][0]
+    final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim
+    self.feature_extractor = ConvFeatureExtractionModel(conv_layers=feature_enc_layers, dropout=0.0, mode=cfg.extractor_mode, conv_bias=cfg.conv_bias)
+    self.post_extract_proj = nn.Linear(self.embed, cfg.encoder_embed_dim) if self.embed != cfg.encoder_embed_dim else None
+    self.encoder = TransformerEncoder(cfg)
+    self.layer_norm = nn.LayerNorm(self.embed)
+    self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim * 1) if self.untie_final_proj else nn.Linear(cfg.encoder_embed_dim, final_dim)
+    self.mask_emb = Tensor.uniform(cfg.encoder_embed_dim, dtype=dtypes.float32)
+    self.label_embs_concat = Tensor.uniform(504, final_dim, dtype=dtypes.float32)
+  def forward_features(self, source, padding_mask):
+    if self.feature_grad_mult > 0:
+      features = self.feature_extractor(source, padding_mask)
+      if self.feature_grad_mult != 1.0: pass  # training: GradMultiply.forward(features, self.feature_grad_mult)
+    else:
+      features = self.feature_extractor(source, padding_mask)
+    return features
+  def forward_padding_mask(self, features, padding_mask):  # replaces original forward_padding_mask for batch inference
+    lengths_org = tilde(padding_mask.cast(dtypes.bool)).cast(dtypes.int64).sum(1)  # ensure its bool for tilde
+    lengths = (lengths_org - 400).float().div(320).floor().cast(dtypes.int64) + 1  # intermediate float to divide
+    padding_mask = lengths_to_padding_mask(lengths)
+    return padding_mask
+  def extract_features(self, source: Tensor, spk_emb:Tensor=None, padding_mask=None, ret_conv=False, output_layer=None, tap=False):
+    features = self.forward_features(source, padding_mask)
+    if padding_mask is not None:
+      padding_mask = self.forward_padding_mask(features, padding_mask)
+    features = features.transpose(1, 2)
+    features = self.layer_norm(features)
+    if self.post_extract_proj is not None:
+      features = self.post_extract_proj(features)
+    x, _ = self.encoder(features, spk_emb, padding_mask=padding_mask, layer=(None if output_layer is None else output_layer - 1), tap=tap)
+    res = features if ret_conv else x
+    return res, padding_mask
+  @classmethod
+  def load_from_pretrained(cls, checkpoint_path:str, checkpoint_url:str) -> ContentVec:
+    fetch(checkpoint_url, checkpoint_path)
+    cfg = load_fairseq_cfg(checkpoint_path)
+    enc = cls(cfg.model)
+    _ = load_checkpoint_enc(checkpoint_path, enc, None)
+    logging.debug(f"{cls.__name__}: Loaded model with cfg={cfg}")
+    return enc
+
+class TransformerEncoder:
+  def __init__(self, cfg: HParams):
+    def make_conv() -> nn.Conv1d:
+      layer = nn.Conv1d(self.embedding_dim, self.embedding_dim, kernel_size=cfg.conv_pos, padding=cfg.conv_pos // 2, groups=cfg.conv_pos_groups)
+      std = std = math.sqrt(4 / (cfg.conv_pos * self.embedding_dim))
+      layer.weight, layer.bias = (Tensor.normal(*layer.weight.shape, std=std)), (Tensor.zeros(*layer.bias.shape))
+      # for training: layer.weights need to be weight_normed
+      return layer
+    self.dropout, self.embedding_dim, self.layer_norm_first, self.layerdrop, self.num_layers, self.num_layers_1 = cfg.dropout, cfg.encoder_embed_dim, cfg.layer_norm_first, cfg.encoder_layerdrop, cfg.encoder_layers, cfg.encoder_layers_1
+    self.pos_conv, self.pos_conv_remove = [make_conv()], (1 if cfg.conv_pos % 2 == 0 else 0)
+    self.layers = [
+      TransformerEncoderLayer(self.embedding_dim, cfg.encoder_ffn_embed_dim, cfg.encoder_attention_heads, self.dropout, cfg.attention_dropout, cfg.activation_dropout, cfg.activation_fn, self.layer_norm_first, cond_layer_norm=(i >= cfg.encoder_layers))
+      for i in range(cfg.encoder_layers + cfg.encoder_layers_1)
+      ]
+    self.layer_norm = nn.LayerNorm(self.embedding_dim)
+    self.cond_layer_norm = CondLayerNorm(self.embedding_dim) if cfg.encoder_layers_1 > 0 else None
+    # training: apply init_bert_params
+  def __call__(self, x, spk_emb, padding_mask=None, layer=None, tap=False):
+    x, layer_results = self.extract_features(x, spk_emb, padding_mask, layer, tap)
+    if self.layer_norm_first and layer is None:
+      x = self.cond_layer_norm(x, spk_emb) if (self.num_layers_1 > 0) else self.layer_norm(x)
+    return x, layer_results
+  def extract_features(self, x: Tensor, spk_emb: Tensor, padding_mask=None, tgt_layer=None, tap=False):
+    if tgt_layer is not None:  # and not self.training
+      assert tgt_layer >= 0 and tgt_layer < len(self.layers)
+    if padding_mask is not None:
+      # x[padding_mask] = 0
+      assert padding_mask.shape == x.shape[:len(padding_mask.shape)]  # first few dims of x must match padding_mask
+      tmp_mask = padding_mask.unsqueeze(-1).repeat((1, 1, x.shape[-1]))
+      tmp_mask = tilde(tmp_mask.cast(dtypes.bool))
+      x = tmp_mask.where(x, 0)
+    x_conv = self.pos_conv[0](x.transpose(1,2))
+    if self.pos_conv_remove > 0: x_conv = x_conv[:, :, : -self.pos_conv_remove]
+    x_conv = x_conv.gelu().transpose(1, 2)
+    x = (x + x_conv).transpose(0, 1)  # B x T x C -> T x B x C
+    if not self.layer_norm_first: x = self.layer_norm(x)
+    x = x.dropout(p=self.dropout)
+    layer_results = []
+    r = None
+    for i, layer in enumerate(self.layers):
+      if i < self.num_layers:  # if (not self.training or (dropout_probability > self.layerdrop)) and (i < self.num_layers):
+        assert layer.cond_layer_norm == False
+        x = layer(x, self_attn_padding_mask=padding_mask, need_weights=False)
+        if tgt_layer is not None or tap:
+          layer_results.append(x.transpose(0, 1))
+      if i>= self.num_layers:
+        assert layer.cond_layer_norm == True
+        x = layer(x, emb=spk_emb, self_attn_padding_mask=padding_mask, need_weights=False)
+      if i == tgt_layer:
+        r = x
+        break
+    if r is not None:
+      x = r
+    x = x.transpose(0, 1)  # T x B x C -> B x T x C
+    return x, layer_results
+
+class TransformerEncoderLayer:
+  def __init__(self, embedding_dim=768.0, ffn_embedding_dim=3072.0, num_attention_heads=8.0, dropout=0.1, attention_dropout=0.1, activation_dropout=0.1, activation_fn="relu", layer_norm_first=False, cond_layer_norm=False):
+    def get_activation_fn(activation):
+      if activation == "relu": return Tensor.relu
+      if activation == "gelu": return Tensor.gelu
+      else: raise RuntimeError(f"activation function={activation} is not forseen")
+    self.embedding_dim, self.dropout, self.activation_dropout, self.layer_norm_first, self.num_attention_heads, self.cond_layer_norm, self.activation_fn = embedding_dim, dropout, activation_dropout, layer_norm_first, num_attention_heads, cond_layer_norm, get_activation_fn(activation_fn)
+    self.self_attn = MultiHeadAttention(self.embedding_dim, self.num_attention_heads)
+    self.self_attn_layer_norm = nn.LayerNorm(self.embedding_dim) if not cond_layer_norm else CondLayerNorm(self.embedding_dim)
+    self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
+    self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
+    self.final_layer_norm = nn.LayerNorm(self.embedding_dim) if not cond_layer_norm else CondLayerNorm(self.embedding_dim)
+  def __call__(self, x:Tensor, self_attn_mask:Tensor=None, self_attn_padding_mask:Tensor=None, emb:Tensor=None, need_weights=False):
+    #self_attn_padding_mask = self_attn_padding_mask.reshape(x.shape[0], 1, 1, self_attn_padding_mask.shape[1]).expand(-1, self.num_attention_heads, -1, -1).reshape(x.shape[0] * self.num_attention_heads, 1, self_attn_padding_mask.shape[1]) if self_attn_padding_mask is not None else None
+    assert self_attn_mask is None and self_attn_padding_mask is not None
+    residual = x
+    if self.layer_norm_first:
+      x = self.self_attn_layer_norm(x) if not self.cond_layer_norm else self.self_attn_layer_norm(x, emb)
+      x = self.self_attn(x=x, mask=self_attn_padding_mask)
+      x = x.dropout(self.dropout)
+      x = residual + x
+      x = self.final_layer_norm(x) if not self.cond_layer_norm else self.final_layer_norm(x, emb)
+      x = self.activation_fn(self.fc1(x))
+      x = x.dropout(self.activation_dropout)
+      x = self.fc2(x)
+      x = x.dropout(self.dropout)
+      x = residual + x
+    else:
+      x = self.self_attn(x=x, mask=self_attn_padding_mask)
+      x = x.dropout(self.dropout)
+      x = residual + x
+      x = self.self_attn_layer_norm(x) if not self.cond_layer_norm else self.self_attn_layer_norm(x, emb)
+      residual = x
+      x = self.activation_fn(self.fc1(x))
+      x = x.dropout(self.activation_dropout)
+      x = self.fc2(x)
+      x = x.dropout(self.dropout)
+      x = residual + x
+      x = self.final_layer_norm(x) if not self.cond_layer_norm else self.final_layer_norm(x, emb)
+    return x
+
+class MultiHeadAttention:
+  def __init__(self, n_state, n_head):
+    self.n_state, self.n_head = n_state, n_head
+    self.q_proj, self.k_proj, self.v_proj, self.out_proj = [nn.Linear(n_state, n_state) for _ in range(4)]
+  def __call__(self, x:Tensor, xa:Optional[Tensor]=None, mask:Optional[Tensor]=None):
+    x = x.transpose(0,1)  # TxBxC -> BxTxC
+    q, k, v = self.q_proj(x), self.k_proj(xa or x), self.v_proj(xa or x)
+    q, k, v = [x.reshape(*q.shape[:2], self.n_head, -1) for x in (q, k, v)]
+    wv = Tensor.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), None).transpose(1, 2).reshape(*x.shape[:2], -1)
+    ret =  self.out_proj(wv).transpose(0,1)  # BxTxC -> TxBxC
+    return ret
+
+class ConvFeatureExtractionModel:
+  def __init__(self, conv_layers, dropout=.0, mode="default", conv_bias=False):
+    assert mode in {"default", "group_norm_masked", "layer_norm"}
+    def block(n_in, n_out, k, stride, is_layer_norm=False, is_group_norm=False, conv_bias=False):
+      def make_conv():
+        conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias)
+        conv.weight = Tensor.kaiming_normal(*conv.weight.shape)
+        return conv
+      assert (is_layer_norm and is_group_norm) == False, "layer norm and group norm are exclusive"
+      if is_layer_norm:
+        return [make_conv(), partial(Tensor.dropout, p=dropout),[partial(Tensor.transpose, dim0=-2, dim1=-1), nn.LayerNorm(dim, elementwise_affine=True), partial(Tensor.transpose, dim0=-2, dim1=-1)], Tensor.gelu]
+      elif is_group_norm and mode == "default":
+        return [make_conv(), partial(Tensor.dropout, p=dropout), nn.GroupNorm(dim, dim, affine=True), Tensor.gelu]
+      elif is_group_norm and mode == "group_norm_masked":
+        return [make_conv(), partial(Tensor.dropout, p=dropout), GroupNormMasked(dim, dim, affine=True), Tensor.gelu]
+      else:
+        return [make_conv(), partial(Tensor.dropout, p=dropout), Tensor.gelu]
+    in_d, self.conv_layers, self.mode = 1, [], mode
+    for i, cl in enumerate(conv_layers):
+      assert len(cl) == 3, "invalid conv definition: " + str(cl)
+      (dim, k, stride) = cl
+      if i == 0: self.cl = cl
+      self.conv_layers.append(block(in_d, dim, k, stride, is_layer_norm=(mode == "layer_norm"), is_group_norm=((mode == "default" or mode == "group_norm_masked") and i == 0), conv_bias=conv_bias))
+      in_d = dim
+  def __call__(self, x:Tensor, padding_mask:Tensor):
+    x = x.unsqueeze(1)  # BxT -> BxCxT
+    if self.mode == "group_norm_masked":
+      if padding_mask is not None:
+        _, k, stride = self.cl
+        lengths_org = tilde(padding_mask.cast(dtypes.bool)).cast(dtypes.int64).sum(1)  # ensure padding_mask is bool for tilde
+        lengths = (((lengths_org - k) / stride) + 1).floor().cast(dtypes.int64)
+        padding_mask = tilde(lengths_to_padding_mask(lengths)).cast(dtypes.int64)  # lengths_to_padding_mask returns bool tensor
+      x = self.conv_layers[0][0](x)  # padding_mask is numeric
+      x = self.conv_layers[0][1](x)
+      x = self.conv_layers[0][2](x, padding_mask)
+      x = self.conv_layers[0][3](x)
+    else:
+      x = x.sequential(self.conv_layers[0])  # default
+    for _, conv in enumerate(self.conv_layers[1:], start=1):
+      conv = reduce(lambda a,b: operator.iconcat(a,b if isinstance(b, list) else [b]), conv, [])  # flatten
+      x = x.sequential(conv)
+    return x
+
+class CondLayerNorm:  # https://github.com/auspicious3000/contentvec/blob/main/contentvec/modules/cond_layer_norm.py#L10
+  def __init__(self, dim_last, eps=1e-5, dim_spk=256, elementwise_affine=True):
+    self.dim_last, self.eps, self.dim_spk, self.elementwise_affine = dim_last, eps, dim_spk, elementwise_affine
+    if self.elementwise_affine:
+      self.weight_ln = nn.Linear(self.dim_spk, self.dim_last, bias=False)
+      self.bias_ln = nn.Linear(self.dim_spk, self.dim_last, bias=False)
+      self.weight_ln.weight, self.bias_ln.weight = (Tensor.ones(*self.weight_ln.weight.shape)), (Tensor.zeros(*self.bias_ln.weight.shape))
+  def __call__(self, x: Tensor, spk_emb: Tensor):
+    axis = tuple(-1-i for i in range(len(x.shape[1:])))
+    x = x.layernorm(axis=axis, eps=self.eps)
+    if not self.elementwise_affine: return x
+    weights, bias = self.weight_ln(spk_emb), self.bias_ln(spk_emb)
+    return weights * x + bias
+
+class GroupNormMasked:  # https://github.com/auspicious3000/contentvec/blob/d746688a32940f4bee410ed7c87ec9cf8ff04f74/contentvec/modules/fp32_group_norm.py#L16
+  def __init__(self, num_groups, num_channels, eps=1e-5, affine=True):
+    self.num_groups, self.num_channels, self.eps, self.affine = num_groups, num_channels, eps, affine
+    self.weight, self.bias = (Tensor.ones(num_channels)), (Tensor.zeros(num_channels)) if self.affine else (None, None)
+  def __call__(self, x:Tensor, mask:Tensor):
+    bsz, n_c, length = x.shape
+    assert n_c % self.num_groups == 0
+    x = x.reshape(bsz, self.num_groups, n_c // self.num_groups, length)
+    if mask is None: mask = Tensor.ones_like(x)
+    else: mask = mask.reshape(bsz, 1, 1, length)
+    x = x * mask
+    lengths = mask.sum(axis=3, keepdim=True)
+    assert x.shape[2] == 1
+    mean_ = x.mean(dim=3, keepdim=True)
+    mean = mean_ * length / lengths
+    var = (((x.std(axis=3, keepdim=True) ** 2) + mean_**2) * length / lengths - mean**2) + self.eps
+    return x.add(-mean).div(var.sqrt()).reshape(bsz, n_c, length).mul(self.weight.reshape(1,-1,1)).add(self.bias.reshape(1,-1,1))
+
+class Synthesizer:
+  def __init__(self, spec_channels, segment_size, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels, ssl_dim, n_speakers, sampling_rate=44100, vol_embedding=False, n_flow_layer=4, **kwargs):
+    self.spec_channels, self.inter_channels, self.hidden_channels, self.filter_channels, self.n_heads, self.n_layers, self.kernel_size, self.p_dropout, self.resblock, self.resblock_kernel_sizes, self.resblock_dilation_sizes, self.upsample_rates, self.upsample_initial_channel, self.upsample_kernel_sizes, self.segment_size, self.n_speakers, self.gin_channels, self.vol_embedding = spec_channels, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, segment_size, n_speakers, gin_channels, vol_embedding
+    self.emb_g = nn.Embedding(n_speakers, gin_channels)
+    if vol_embedding: self.emb_vol = nn.Linear(1, hidden_channels)
+    self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
+    self.enc_p = TextEncoder(inter_channels, hidden_channels, kernel_size, n_layers, filter_channels=filter_channels, n_heads=n_heads, p_dropout=p_dropout)
+    self.dec = Generator(sampling_rate, inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels)
+    self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
+    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, n_flow_layer, gin_channels=gin_channels)
+    self.emb_uv = nn.Embedding(vocab_size=2, embed_size=hidden_channels)
+  def infer(self, c:Tensor, f0:Tensor, uv:Tensor, g:Tensor=None, noise_scale=0.35, seed=52468, vol=None) -> Tuple[Tensor, Tensor]:
+    Tensor.manual_seed(getenv('SEED', seed))
+    c_lengths = (Tensor.ones([c.shape[0]]) * c.shape[-1]).to(c.device)
+    if len(g.shape) == 1: g = g.unsqueeze(0)
+    g = self.emb_g(g).transpose(1, 2)
+    x_mask = sequence_mask(c_lengths, c.shape[2]).unsqueeze(1).cast(c.dtype)
+    vol = self.emb_vol(vol[:,:,None]).transpose(1,2) if vol is not None and self.vol_embedding else 0
+    x = self.pre(c) * x_mask + self.emb_uv(uv.cast(dtypes.int64)).transpose(1, 2) + vol
+    z_p, _, _, c_mask = self.enc_p.forward(x, x_mask, f0=self._f0_to_coarse(f0), noise_scale=noise_scale)
+    z = self.flow.forward(z_p, c_mask, g=g, reverse=True)
+    o = self.dec.forward(z * c_mask, g=g, f0=f0)
+    return o,f0
+  def _f0_to_coarse(self, f0 : Tensor):
+    f0_mel = 1127 * (1 + f0 / 700).log()
+    a = (F0_BIN - 2) / (F0_MEL_MAX - F0_MEL_MIN)
+    b = F0_MEL_MIN * a - 1.
+    f0_mel = (f0_mel > 0).where(f0_mel * a - b, f0_mel)
+    f0_coarse = f0_mel.ceil().cast(dtype=dtypes.int64)
+    f0_coarse = f0_coarse * (f0_coarse > 0)
+    f0_coarse = f0_coarse + ((f0_coarse < 1) * 1)
+    f0_coarse = f0_coarse * (f0_coarse < F0_BIN)
+    f0_coarse = f0_coarse + ((f0_coarse >= F0_BIN) * (F0_BIN - 1))
+    return f0_coarse
+  @classmethod
+  def load_from_pretrained(cls, config_path:str, config_url:str, weights_path:str, weights_url:str) -> Synthesizer:
+    fetch(config_url, config_path)
+    hps = get_hparams_from_file(config_path)
+    fetch(weights_url, weights_path)
+    net_g = cls(hps.data.filter_length // 2 + 1, hps.train.segment_size // hps.data.hop_length, **hps.model)
+    _ = load_checkpoint(weights_path, net_g, None, skip_list=["f0_decoder"])
+    logging.debug(f"{cls.__name__}:Loaded model with hps: {hps}")
+    return net_g, hps
+
+class TextEncoder:
+  def __init__(self, out_channels, hidden_channels, kernel_size, n_layers, gin_channels=0, filter_channels=None, n_heads=None, p_dropout=None):
+    self.out_channels, self.hidden_channels, self.kernel_size, self.n_layers, self.gin_channels = out_channels, hidden_channels, kernel_size, n_layers, gin_channels
+    self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+    self.f0_emb = nn.Embedding(256, hidden_channels)  # n_vocab = 256
+    self.enc_ = Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
+  def forward(self, x, x_mask, f0=None, noise_scale=1):
+    x = x + self.f0_emb(f0).transpose(1, 2)
+    x = self.enc_.forward(x * x_mask, x_mask)
+    stats = self.proj(x) * x_mask
+    m, logs = split(stats, self.out_channels, dim=1)
+    z = (m + randn_like(m) * logs.exp() * noise_scale) * x_mask
+    return z, m, logs, x_mask
+
+class Upsample:
+  def __init__(self, scale_factor):
+    assert scale_factor % 1 == 0, "Only integer scale factor allowed."
+    self.scale = int(scale_factor)
+  def forward(self, x:Tensor):
+    repeats = tuple([1] * len(x.shape) + [self.scale])
+    new_shape = (*x.shape[:-1], x.shape[-1] * self.scale)
+    return x.unsqueeze(-1).repeat(repeats).reshape(new_shape)
+
+class SineGen:
+  def __init__(self, samp_rate, harmonic_num=0, sine_amp=0.1, noise_std=0.003, voice_threshold=0, flag_for_pulse=False):
+    self.sine_amp, self.noise_std, self.harmonic_num, self.sampling_rate, self.voiced_threshold, self.flag_for_pulse = sine_amp, noise_std, harmonic_num, samp_rate, voice_threshold, flag_for_pulse
+    self.dim = self.harmonic_num + 1
+  def _f02uv(self, f0): return (f0 > self.voiced_threshold).float()  #generate uv signal
+  def _f02sine(self, f0_values):
+    def padDiff(x : Tensor): return (x.pad((0,0,-1,1)) - x).pad((0,0,0,-1))
+    def mod(x: Tensor, n: int) -> Tensor: return x - n * x.div(n).floor()  # this is what the % operator does in pytorch.
+    rad_values = mod((f0_values / self.sampling_rate) , 1)  # convert to F0 in rad
+    rand_ini = Tensor.rand(f0_values.shape[0], f0_values.shape[2], device=f0_values.device)  # initial phase noise
+
+    #rand_ini[:, 0] = 0
+    m = Tensor.ones(f0_values.shape[0]).unsqueeze(1).pad((0,f0_values.shape[2]-1,0,0)).cast(dtypes.bool)
+    m = tilde(m)
+    rand_ini = m.where(rand_ini, 0)
+
+    #rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+    tmp = rad_values[:, 0, :] + rand_ini
+    m = Tensor.ones(tmp.shape).pad((0,0,0,rad_values.shape[1]-1,0)).cast(dtypes.bool)
+    m = tilde(m)
+    tmp = tmp.unsqueeze(1).pad((0,0,0,rad_values.shape[1]-1,0))
+    rad_values = m.where(rad_values, tmp)
+
+    tmp_over_one = mod(rad_values.cumsum(1), 1)
+    tmp_over_one_idx = padDiff(tmp_over_one) < 0
+    cumsum_shift = Tensor.zeros_like(rad_values)
+
+    #cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+    tmp_over_one_idx = (tmp_over_one_idx * -1.0).pad((0,0,1,0))
+    cumsum_shift = tmp_over_one_idx
+
+    sines = ((rad_values + cumsum_shift).cumsum(1) * 2 * np.pi).sin()
+    return sines
+  def forward(self, f0, upp=None):
+    fn = f0.mul(Tensor([[range(1, self.harmonic_num + 2)]], dtype=dtypes.float32).to(f0.device))
+    sine_waves = self._f02sine(fn) * self.sine_amp  #generate sine waveforms
+    uv = self._f02uv(f0)  # generate uv signal
+    noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+    noise = noise_amp * randn_like(sine_waves)
+    sine_waves = sine_waves * uv + noise
+    return sine_waves, uv, noise
+
+class SourceHnNSF:
+  def __init__(self, sampling_rate, harmonic_num=0, sine_amp=0.1, add_noise_std=0.003, voiced_threshold=0):
+    self.sine_amp, self.noise_std = sine_amp, add_noise_std
+    self.l_sin_gen = SineGen(sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshold)
+    self.l_linear = nn.Linear(harmonic_num + 1, 1)
+  def forward(self, x, upp=None):
+    sine_waves, uv, _ = self.l_sin_gen.forward(x, upp)
+    sine_merge = self.l_linear(sine_waves.cast(self.l_linear.weight.dtype)).tanh()
+    noise = randn_like(uv) * self.sine_amp / 3
+    return sine_merge, noise, uv
+
+# most of the hifigan in standard vits is reused here, but need to upsample and construct harmonic source from f0
+class Generator:
+  def __init__(self, sampling_rate, inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels):
+    self.sampling_rate, self.inter_channels, self.resblock, self.resblock_kernel_sizes, self.resblock_dilation_sizes, self.upsample_rates, self.upsample_initial_channel, self.upsample_kernel_sizes, self.gin_channels = sampling_rate, inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels
+    self.num_kernels, self.num_upsamples = len(resblock_kernel_sizes), len(upsample_rates)
+    self.conv_pre = nn.Conv1d(inter_channels, upsample_initial_channel, 7, 1, padding=3)
+    self.f0_upsamp = Upsample(scale_factor=np.prod(upsample_rates))
+    self.m_source = SourceHnNSF(sampling_rate, harmonic_num=8)
+    resblock = ResBlock1 if resblock == '1' else ResBlock2
+    self.ups, self.noise_convs, self.resblocks = [], [], []
+    for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+      c_cur = upsample_initial_channel//(2**(i+1))
+      self.ups.append(nn.ConvTranspose1d(upsample_initial_channel//(2**i), c_cur, k, u, padding=(k-u)//2))
+      stride_f0 = int(np.prod(upsample_rates[i + 1:]))
+      self.noise_convs.append(nn.Conv1d(1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2) if (i + 1 < len(upsample_rates)) else nn.Conv1d(1, c_cur, kernel_size=1))
+    for i in range(len(self.ups)):
+      ch = upsample_initial_channel // (2 ** (i + 1))
+      for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+        self.resblocks.append(resblock(ch, k, d))
+    self.conv_post = nn.Conv1d(ch, 1, 7, 1, padding=3)
+    if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+    self.upp = np.prod(upsample_rates)
+  def forward(self, x, f0, g=None):
+    f0 = self.f0_upsamp.forward(f0[:, None]).transpose(1, 2)  # bs,n,t
+    har_source, _, _ = self.m_source.forward(f0, self.upp)
+    har_source = har_source.transpose(1, 2)
+    x = self.conv_pre(x)
+    if g is not None:  x = x + self.cond(g)
+    for i in range(self.num_upsamples):
+      x, xs = self.ups[i](x.leaky_relu(LRELU_SLOPE)), None
+      x_source = self.noise_convs[i](har_source)
+      x = x + x_source
+      for j in range(self.num_kernels):
+        if xs is None: xs = self.resblocks[i * self.num_kernels + j].forward(x)
+        else: xs += self.resblocks[i * self.num_kernels + j].forward(x)
+      x = xs / self.num_kernels
+    return self.conv_post(x.leaky_relu()).tanh()
+
+# **** helpers ****
+
+def randn_like(x:Tensor) -> Tensor: return Tensor.randn(*x.shape, dtype=x.dtype).to(device=x.device)
+
+def tilde(x: Tensor) -> Tensor:
+  if x.dtype == dtypes.bool: return (1 - x).cast(dtypes.bool)
+  return (x + 1) * -1  # this seems to be what the ~ operator does in pytorch for non bool
+
+def lengths_to_padding_mask(lens:Tensor) -> Tensor:
+  bsz, max_lens = lens.shape[0], lens.max().numpy().item()
+  mask = Tensor.arange(max_lens).to(lens.device).reshape(1, max_lens)
+  mask = mask.expand(bsz, -1) >= lens.reshape(bsz, 1).expand(-1, max_lens)
+  return mask.cast(dtypes.bool)
+
+def repeat_expand_2d_left(content, target_len): # content : [h, t]
+  src_len = content.shape[-1]
+  temp = np.arange(src_len+1) * target_len / src_len
+  current_pos, cols = 0, []
+  for i in range(target_len):
+    if i >= temp[current_pos+1]:
+      current_pos += 1
+    cols.append(content[:, current_pos])
+  return Tensor.stack(*cols).transpose(0, 1)
+
+def load_fairseq_cfg(checkpoint_path):
+  assert Path(checkpoint_path).is_file()
+  state = torch_load(checkpoint_path)
+  cfg = state["cfg"] if ("cfg" in state and state["cfg"] is not None) else None
+  if cfg is None: raise RuntimeError(f"No cfg exist in state keys = {state.keys()}")
+  return HParams(**cfg)
+
+def load_checkpoint_enc(checkpoint_path, model: ContentVec, optimizer=None, skip_list=[]):
+  assert Path(checkpoint_path).is_file()
+  start_time = time.time()
+  checkpoint_dict = torch_load(checkpoint_path)
+  saved_state_dict = checkpoint_dict['model']
+  weight_g, weight_v, parent = None, None, None
+  for key, v in saved_state_dict.items():
+    if any(layer in key for layer in skip_list): continue
+    try:
+      obj, skip = model, False
+      for k in key.split('.'):
+        if k.isnumeric(): obj = obj[int(k)]
+        elif isinstance(obj, dict): obj = obj[k]
+        else:
+          if k in ["weight_g", "weight_v"]:
+            parent, skip = obj, True
+            if k == "weight_g": weight_g = v
+            else: weight_v = v
+          if not skip:
+            parent = obj
+            obj = getattr(obj, k)
+      if weight_g and weight_v:
+        setattr(obj, "weight_g", weight_g.numpy())
+        setattr(obj, "weight_v", weight_v.numpy())
+        obj, v = getattr(parent, "weight"), weight_norm(weight_v, weight_g, 0)
+        weight_g, weight_v, parent, skip = None, None, None, False
+      if not skip and obj.shape == v.shape:
+        if "feature_extractor" in key and (isinstance(parent, (nn.GroupNorm, nn.LayerNorm))):  # cast
+          obj.assign(v.to(obj.device).float())
+        else:
+          obj.assign(v.to(obj.device))
+      elif not skip: logging.error(f"MISMATCH SHAPE IN {key}, {obj.shape} {v.shape}")
+    except Exception as e: raise e
+  logging.info(f"Loaded checkpoint '{checkpoint_path}' in {time.time() - start_time:.4f}s")
+  return model, optimizer
+
+def pad_array(arr, target_length):
+  current_length = arr.shape[0]
+  if current_length >= target_length: return arr
+  pad_width = target_length - current_length
+  pad_left = pad_width // 2
+  pad_right = pad_width - pad_left
+  padded_arr = np.pad(arr, (pad_left, pad_right), 'constant', constant_values=(0, 0))
+  return padded_arr
+
+def split_list_by_n(list_collection, n, pre=0):
+  for i in range(0, len(list_collection), n):
+    yield list_collection[i-pre if i-pre>=0 else i: i + n]
+
+def get_sid(spk2id:HParams, speaker:str) -> Tensor:
+  speaker_id = spk2id[speaker]
+  if not speaker_id and type(speaker) is int:
+    if len(spk2id.__dict__) >= speaker: speaker_id = speaker
+  if speaker_id is None: raise RuntimeError(f"speaker={speaker} not in the speaker list")
+  return Tensor([int(speaker_id)], dtype=dtypes.int64).unsqueeze(0)
+
+def get_encoder(ssl_dim) -> Type[SpeechEncoder]:
+  if ssl_dim == 256: return ContentVec256L9
+  if ssl_dim == 768: return ContentVec768L12
+
+#########################################################################################
+# CODE: https://github.com/svc-develop-team/so-vits-svc
+#########################################################################################
+# CONTENTVEC:
+#   CODE: https://github.com/auspicious3000/contentvec
+#   PAPER: https://arxiv.org/abs/2204.09224
+#########################################################################################
+# INSTALLATION: dependencies are for preprocessing and loading/saving audio.
+# pip3 install soundfile librosa praat-parselmouth
+#########################################################################################
+# EXAMPLE USAGE:
+# python3 examples/so_vits_svc.py --model tf2spy --file ~/recording.wav
+#########################################################################################
+# DEMO USAGE (uses audio sample from LJ-Speech):
+# python3 examples/so_vits_svc.py --model saul_goodman
+#########################################################################################
+SO_VITS_SVC_PATH = Path(__file__).parents[1] / "weights/So-VITS-SVC"
+VITS_MODELS = { # config_path, weights_path, config_url, weights_url
+  "saul_goodman" : (SO_VITS_SVC_PATH / "config_saul_gman.json", SO_VITS_SVC_PATH / "pretrained_saul_gman.pth", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/Saul_Goodman_80000/config.json", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/Saul_Goodman_80000/G_80000.pth"),
+  "drake" : (SO_VITS_SVC_PATH / "config_drake.json", SO_VITS_SVC_PATH / "pretrained_drake.pth", "https://huggingface.co/jaspa/so-vits-svc/resolve/main/aubrey/config_aubrey.json", "https://huggingface.co/jaspa/so-vits-svc/resolve/main/aubrey/pretrained_aubrey.pth"),
+  "cartman" : (SO_VITS_SVC_PATH / "config_cartman.json", SO_VITS_SVC_PATH / "pretrained_cartman.pth", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/EricCartman/config.json", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/EricCartman/G_10200.pth"),
+  "tf2spy" : (SO_VITS_SVC_PATH / "config_tf2spy.json", SO_VITS_SVC_PATH / "pretrained_tf2spy.pth", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_spy_60k/config.json", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_spy_60k/G_60000.pth"),
+  "tf2heavy" : (SO_VITS_SVC_PATH / "config_tf2heavy.json", SO_VITS_SVC_PATH / "pretrained_tf2heavy.pth", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_heavy_100k/config.json", "https://huggingface.co/Amo/so-vits-svc-4.0_GA/resolve/main/ModelsFolder/TF2_heavy_100k/G_100000.pth"),
+  "lady_gaga" : (SO_VITS_SVC_PATH / "config_gaga.json", SO_VITS_SVC_PATH / "pretrained_gaga.pth", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/LadyGaga/config.json", "https://huggingface.co/marcoc2/so-vits-svc-4.0-models/resolve/main/LadyGaga/G_14400.pth")
+}
+ENCODER_MODELS = { # weights_path, weights_url
+  "contentvec": (SO_VITS_SVC_PATH / "contentvec_checkpoint.pt", "https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt")
+}
+ENCODER_MODEL = "contentvec"
+DEMO_PATH, DEMO_URL = Path(__file__).parents[1] / "temp/LJ037-0171.wav", "https://keithito.com/LJ-Speech-Dataset/LJ037-0171.wav"
+if __name__=="__main__":
+  logging.basicConfig(stream=sys.stdout, level=(logging.INFO if DEBUG < 1 else logging.DEBUG))
+  parser = argparse.ArgumentParser()
+  parser.add_argument("-m", "--model", default=None, help=f"Specify the model to use. All supported models: {VITS_MODELS.keys()}", required=True)
+  parser.add_argument("-f", "--file", default=DEMO_PATH, help=f"Specify the path of the input file")
+  parser.add_argument("--out_dir", default=str(Path(__file__).parents[1] / "temp"), help="Specify the output path.")
+  parser.add_argument("--out_path", default=None, help="Specify the full output path. Overrides the --out_dir and --name parameter.")
+  parser.add_argument("--base_name", default="test", help="Specify the base of the output file name. Default is 'test'.")
+  parser.add_argument("--speaker", default=None, help="If not specified, the first available speaker is chosen. Usually there is only one speaker per model.")
+  parser.add_argument("--noise_scale", default=0.4)
+  parser.add_argument("--tran", default=0.0, help="Pitch shift, supports positive and negative (semitone) values. Default 0.0")
+  parser.add_argument("--pad_seconds", default=0.5)
+  parser.add_argument("--lg_num", default=0.0)
+  parser.add_argument("--clip_seconds", default=0.0)
+  parser.add_argument("--slice_db", default=-40)
+  args = parser.parse_args()
+
+  vits_model = args.model
+  encoder_location, vits_location = ENCODER_MODELS[ENCODER_MODEL], VITS_MODELS[vits_model]
+
+  Tensor.training = False
+  # Get Synthesizer and ContentVec
+  net_g, hps = Synthesizer.load_from_pretrained(vits_location[0], vits_location[2], vits_location[1], vits_location[3])
+  Encoder = get_encoder(hps.model.ssl_dim)
+  encoder = Encoder.load_from_pretrained(encoder_location[0], encoder_location[1])
+
+  # model config args
+  target_sample, spk2id, hop_length, target_sample = hps.data.sampling_rate, hps.spk, hps.data.hop_length, hps.data.sampling_rate
+  vol_embedding = hps.model.vol_embedding if hasattr(hps.data, "vol_embedding") and hps.model.vol_embedding is not None else False
+
+  # args
+  slice_db, clip_seconds, lg_num, pad_seconds, tran, noise_scale, audio_path = args.slice_db, args.clip_seconds, args.lg_num, args.pad_seconds, args.tran, args.noise_scale, args.file
+  speaker = args.speaker if args.speaker is not None else list(hps.spk.__dict__.keys())[0]
+
+  ### Loading audio and slicing ###
+  if audio_path == DEMO_PATH: fetch(DEMO_URL, DEMO_PATH)
+  assert Path(audio_path).is_file() and Path(audio_path).suffix == ".wav"
+  chunks = preprocess.cut(audio_path, db_thresh=slice_db)
+  audio_data, audio_sr = preprocess.chunks2audio(audio_path, chunks)
+
+  per_size = int(clip_seconds * audio_sr)
+  lg_size = int(lg_num * audio_sr)
+
+  ### Infer per slice ###
+  global_frame = 0
+  audio = []
+  for (slice_tag, data) in audio_data:
+    print(f"\n====segment start, {round(len(data) / audio_sr, 3)}s====")
+    length = int(np.ceil(len(data) / audio_sr * target_sample))
+
+    if slice_tag:
+      print("empty segment")
+      _audio = np.zeros(length)
+      audio.extend(list(pad_array(_audio, length)))
+      global_frame += length // hop_length
+      continue
+
+    datas = [data] if per_size == 0 else split_list_by_n(data, per_size, lg_size)
+
+    for k, dat in enumerate(datas):
+      per_length = int(np.ceil(len(dat) / audio_sr * target_sample)) if clip_seconds!=0 else length
+      pad_len = int(audio_sr * pad_seconds)
+      dat = np.concatenate([np.zeros([pad_len]), dat, np.zeros([pad_len])])
+      raw_path = io.BytesIO()
+      soundfile.write(raw_path, dat, audio_sr, format="wav")
+      raw_path.seek(0)
+
+      ### Infer START ###
+      wav, sr = preprocess.load_audiofile(raw_path)
+      wav = preprocess.sinc_interp_resample(wav, sr, target_sample)[0]
+      wav16k, f0, uv = preprocess.get_unit_f0(wav, tran, hop_length, target_sample)
+      sid = get_sid(spk2id, speaker)
+      n_frames = f0.shape[1]
+
+      # ContentVec infer
+      start = time.time()
+      c = encoder.encode(wav16k)
+      c = repeat_expand_2d_left(c.squeeze(0).realize(), f0.shape[1])  # interpolate speech encoding to match f0
+      c = c.unsqueeze(0).realize()
+      enc_time = time.time() - start
+
+      # VITS infer
+      vits_start = time.time()
+      out_audio, f0 = net_g.infer(c, f0=f0, uv=uv, g=sid, noise_scale=noise_scale, vol=None)
+      out_audio = out_audio[0,0].float().realize()
+      vits_time = time.time() - vits_start
+
+      infer_time = time.time() - start
+      logging.info("total infer time:{:.2f}s, speech_enc time:{:.2f}s, vits time:{:.2f}s".format(infer_time, enc_time, vits_time))
+      ### Infer END ###
+
+      out_sr, out_frame = out_audio.shape[-1], n_frames
+      global_frame += out_frame
+      _audio = out_audio.numpy()
+      pad_len = int(target_sample * pad_seconds)
+      _audio = _audio[pad_len:-pad_len]
+      _audio = pad_array(_audio, per_length)
+      audio.extend(list(_audio))
+
+  audio = np.array(audio)
+  out_path = Path(args.out_path or Path(args.out_dir)/f"{args.model}{f'_spk_{speaker}'}_{args.base_name}.wav")
+  out_path.parent.mkdir(parents=True, exist_ok=True)
+  soundfile.write(out_path, audio, target_sample, format="flac")
+  logging.info(f"Saved audio output to {out_path}")
@@ -0,0 +1,204 @@
+import math
+from typing import Optional, Tuple
+from tinygrad import Tensor, dtypes
+import librosa
+import soundfile
+import numpy as np
+import parselmouth
+
+class PMF0Predictor:  # from https://github.com/svc-develop-team/so-vits-svc/
+  def __init__(self,hop_length=512,f0_min=50,f0_max=1100,sampling_rate=44100):
+    self.hop_length, self.f0_min, self.f0_max, self.sampling_rate, self.name = hop_length, f0_min, f0_max, sampling_rate, "pm"
+  def interpolate_f0(self,f0):
+    vuv_vector = np.zeros_like(f0, dtype=np.float32)
+    vuv_vector[f0 > 0.0] = 1.0
+    vuv_vector[f0 <= 0.0] = 0.0
+    nzindex = np.nonzero(f0)[0]
+    data = f0[nzindex]
+    nzindex = nzindex.astype(np.float32)
+    time_org = self.hop_length / self.sampling_rate * nzindex
+    time_frame = np.arange(f0.shape[0]) * self.hop_length / self.sampling_rate
+    if data.shape[0] <= 0: return np.zeros(f0.shape[0], dtype=np.float32),vuv_vector
+    if data.shape[0] == 1: return np.ones(f0.shape[0], dtype=np.float32) * f0[0],vuv_vector
+    f0 = np.interp(time_frame, time_org, data, left=data[0], right=data[-1])
+    return f0,vuv_vector
+  def compute_f0(self,wav,p_len=None):
+    x = wav
+    if p_len is None: p_len = x.shape[0]//self.hop_length
+    else: assert abs(p_len-x.shape[0]//self.hop_length) < 4, "pad length error"
+    time_step = self.hop_length / self.sampling_rate * 1000
+    f0 = parselmouth.Sound(x, self.sampling_rate) \
+                    .to_pitch_ac(time_step=time_step / 1000, voicing_threshold=0.6,pitch_floor=self.f0_min, pitch_ceiling=self.f0_max) \
+                    .selected_array['frequency']
+    pad_size=(p_len - len(f0) + 1) // 2
+    if(pad_size>0 or p_len - len(f0) - pad_size>0):
+      f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
+    f0,uv = self.interpolate_f0(f0)
+    return f0
+  def compute_f0_uv(self,wav,p_len=None):
+    x = wav
+    if p_len is None: p_len = x.shape[0]//self.hop_length
+    else: assert abs(p_len-x.shape[0]//self.hop_length) < 4, "pad length error"
+    time_step = self.hop_length / self.sampling_rate * 1000
+    f0 = parselmouth.Sound(x, self.sampling_rate).to_pitch_ac(
+      time_step=time_step / 1000, voicing_threshold=0.6,
+      pitch_floor=self.f0_min, pitch_ceiling=self.f0_max).selected_array['frequency']
+    pad_size=(p_len - len(f0) + 1) // 2
+    if(pad_size>0 or p_len - len(f0) - pad_size>0):
+      f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
+    f0,uv = self.interpolate_f0(f0)
+    return f0,uv
+
+class Slicer:  # from https://github.com/svc-develop-team/so-vits-svc/
+  def __init__(self, sr: int, threshold: float = -40., min_length: int = 5000, min_interval: int = 300, hop_size: int = 20, max_sil_kept: int = 5000):
+    if not min_length >= min_interval >= hop_size:
+      raise ValueError('The following condition must be satisfied: min_length >= min_interval >= hop_size')
+    if not max_sil_kept >= hop_size:
+      raise ValueError('The following condition must be satisfied: max_sil_kept >= hop_size')
+    min_interval = sr * min_interval / 1000
+    self.threshold = 10 ** (threshold / 20.)
+    self.hop_size = round(sr * hop_size / 1000)
+    self.win_size = min(round(min_interval), 4 * self.hop_size)
+    self.min_length = round(sr * min_length / 1000 / self.hop_size)
+    self.min_interval = round(min_interval / self.hop_size)
+    self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+  def _apply_slice(self, waveform, begin, end):
+    if len(waveform.shape) > 1: return waveform[:, begin * self.hop_size: min(waveform.shape[1], end * self.hop_size)]
+    else: return waveform[begin * self.hop_size: min(waveform.shape[0], end * self.hop_size)]
+  def slice(self, waveform):
+    samples = librosa.to_mono(waveform) if len(waveform.shape) > 1 else waveform
+    if samples.shape[0] <= self.min_length: return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
+    rms_list = librosa.feature.rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
+    sil_tags, silence_start, clip_start = [], None, 0
+    for i, rms in enumerate(rms_list):
+      if rms < self.threshold:  # Keep looping while frame is silent.
+        if silence_start is None:  # Record start of silent frames.
+          silence_start = i
+        continue
+      if silence_start is None: continue  # Keep looping while frame is not silent and silence start has not been recorded.
+      # Clear recorded silence start if interval is not enough or clip is too short
+      is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+      need_slice_middle = i - silence_start >= self.min_interval and i - clip_start >= self.min_length
+      if not is_leading_silence and not need_slice_middle:
+        silence_start = None
+        continue
+      if i - silence_start <= self.max_sil_kept:  # Need slicing. Record the range of silent frames to be removed.
+        pos = rms_list[silence_start: i + 1].argmin() + silence_start
+        sil_tags.append((0, pos) if silence_start == 0 else (pos, pos))
+        clip_start = pos
+      elif i - silence_start <= self.max_sil_kept * 2:
+        pos = rms_list[i - self.max_sil_kept: silence_start + self.max_sil_kept + 1].argmin()
+        pos += i - self.max_sil_kept
+        pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
+        pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
+        if silence_start == 0:
+          sil_tags.append((0, pos_r))
+          clip_start = pos_r
+        else:
+          sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+          clip_start = max(pos_r, pos)
+      else:
+        pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
+        pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
+        sil_tags.append((0, pos_r) if silence_start == 0 else (pos_l, pos_r))
+        clip_start = pos_r
+      silence_start = None
+    total_frames = rms_list.shape[0]
+    if silence_start is not None and total_frames - silence_start >= self.min_interval:  # Deal with trailing silence.
+      silence_end = min(total_frames, silence_start + self.max_sil_kept)
+      pos = rms_list[silence_start: silence_end + 1].argmin() + silence_start
+      sil_tags.append((pos, total_frames + 1))
+    if len(sil_tags) == 0: return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}  # Apply and return slices.
+    chunks = []
+    if sil_tags[0][0]:
+      chunks.append({"slice": False, "split_time": f"0,{min(waveform.shape[0], sil_tags[0][0] * self.hop_size)}"})
+    for i in range(0, len(sil_tags)):
+      if i: chunks.append({"slice": False, "split_time": f"{sil_tags[i - 1][1] * self.hop_size},{min(waveform.shape[0], sil_tags[i][0] * self.hop_size)}"})
+      chunks.append({"slice": True, "split_time": f"{sil_tags[i][0] * self.hop_size},{min(waveform.shape[0], sil_tags[i][1] * self.hop_size)}"})
+    if sil_tags[-1][1] * self.hop_size < len(waveform):
+      chunks.append({"slice": False, "split_time": f"{sil_tags[-1][1] * self.hop_size},{len(waveform)}"})
+    chunk_dict = {}
+    for i in range(len(chunks)): chunk_dict[str(i)] = chunks[i]
+    return chunk_dict
+
+# sinc_interp_hann audio resampling
+class Resample:
+  def __init__(self, orig_freq:int=16000, new_freq:int=16000, lowpass_filter_width:int=6, rolloff:float=0.99, beta:Optional[float]=None, dtype:Optional[dtypes]=None):
+    self.orig_freq, self.new_freq, self.lowpass_filter_width, self.rolloff, self.beta = orig_freq, new_freq, lowpass_filter_width, rolloff, beta
+    self.gcd = math.gcd(int(self.orig_freq), int(self.new_freq))
+    self.kernel, self.width = self._get_sinc_resample_kernel(dtype) if self.orig_freq != self.new_freq else (None, None)
+  def __call__(self, waveform:Tensor) -> Tensor:
+    if self.orig_freq == self.new_freq: return waveform
+    return self._apply_sinc_resample_kernel(waveform)
+  def _apply_sinc_resample_kernel(self, waveform:Tensor):
+    if not waveform.is_floating_point(): raise TypeError(f"Waveform tensor expected to be of type float, but received {waveform.dtype}.")
+    orig_freq, new_freq = (int(self.orig_freq) // self.gcd), (int(self.new_freq) // self.gcd)
+    shape = waveform.shape
+    waveform = waveform.reshape(-1, shape[-1])  # pack batch
+    num_wavs, length = waveform.shape
+    target_length = int(math.ceil(new_freq * length / orig_freq))
+    waveform = waveform.pad((self.width, self.width + orig_freq))
+    resampled = waveform[:, None].conv2d(self.kernel, stride=orig_freq)
+    resampled = resampled.transpose(1, 2).reshape(num_wavs, -1)
+    resampled = resampled[..., :target_length]
+    resampled = resampled.reshape(shape[:-1] + resampled.shape[-1:])  # unpack batch
+    return resampled
+  def _get_sinc_resample_kernel(self, dtype=None):
+    orig_freq, new_freq = (int(self.orig_freq) // self.gcd), (int(self.new_freq) // self.gcd)
+    if self.lowpass_filter_width <= 0: raise ValueError("Low pass filter width should be positive.")
+    base_freq = min(orig_freq, new_freq)
+    base_freq *= self.rolloff
+    width = math.ceil(self.lowpass_filter_width * orig_freq / base_freq)
+    idx = Tensor.arange(-width, width + orig_freq, dtype=(dtype if dtype is not None else dtypes.float32))[None, None] / orig_freq
+    t = Tensor.arange(0, -new_freq, -1, dtype=dtype)[:, None, None] / new_freq + idx
+    t *= base_freq
+    t = t.clip(-self.lowpass_filter_width, self.lowpass_filter_width)
+    window = (t * math.pi / self.lowpass_filter_width / 2).cos() ** 2
+    t *= math.pi
+    scale = base_freq / orig_freq
+    kernels = Tensor.where(t == 0, Tensor(1.0, dtype=t.dtype).to(t.device), t.sin() / t)
+    kernels *= window * scale
+    if dtype is None: kernels = kernels.cast(dtype=dtypes.float32)
+    return kernels, width
+
+def sinc_interp_resample(x:Tensor, orig_freq:int=16000, new_freq:int=1600, lowpass_filter_width:int=6, rolloff:float=0.99, beta:Optional[float]=None):
+  resamp = Resample(orig_freq, new_freq, lowpass_filter_width, rolloff, beta, x.dtype)
+  return resamp(x)
+
+def cut(audio_path, db_thresh=-30, min_len=5000):
+  audio, sr = librosa.load(audio_path, sr=None)
+  slicer = Slicer(sr=sr, threshold=db_thresh, min_length=min_len)
+  chunks = slicer.slice(audio)
+  return chunks
+
+def chunks2audio(audio_path, chunks):
+  chunks = dict(chunks)
+  audio, sr = load_audiofile(audio_path)
+  if len(audio.shape) == 2 and audio.shape[1] >= 2:
+    audio = audio.mean(0).unsqueeze(0)
+  audio = audio.numpy()[0]
+  result = []
+  for k, v in chunks.items():
+    tag = v["split_time"].split(",")
+    if tag[0] != tag[1]:
+      result.append((v["slice"], audio[int(tag[0]):int(tag[1])]))
+  return result, sr
+
+def load_audiofile(filepath:str, frame_offset:int=0, num_frames:int=-1, channels_first:bool=True):
+  with soundfile.SoundFile(filepath, "r") as file_:
+    frames = file_._prepare_read(frame_offset, None, num_frames)
+    waveform = file_.read(frames, "float32", always_2d=True)
+    sample_rate = file_.samplerate
+  waveform = Tensor(waveform)
+  if channels_first: waveform = waveform.transpose(0, 1)
+  return waveform, sample_rate
+
+def get_unit_f0(wav:Tensor, tran, hop_length, target_sample, f0_filter=False) -> Tuple[Tensor,Tensor,Tensor]:
+  f0_predictor = PMF0Predictor(hop_length, sampling_rate=target_sample)
+  f0, uv = f0_predictor.compute_f0_uv(wav.numpy())
+  if f0_filter and sum(f0) == 0: raise RuntimeError("No voice detected")
+  f0 = Tensor(f0.astype(np.float32)).float()
+  f0 = (f0 * 2 ** (tran / 12)).unsqueeze(0)
+  uv = Tensor(uv.astype(np.float32)).float().unsqueeze(0)
+  wav16k = sinc_interp_resample(wav[None,:], target_sample, 16000)[0]
+  return wav16k.realize(), f0.realize(), uv.realize()
@@ -2,20 +2,18 @@
 # https://github.com/ekagra-ranjan/huggingface-blog/blob/main/stable_diffusion.md
 import tempfile
 from pathlib import Path
-import argparse, time
+import argparse
 from collections import namedtuple
 from typing import Dict, Any

 from PIL import Image
 import numpy as np
 from tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit
-from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten, profile_marker
+from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm
 from tinygrad.nn import Conv2d, GroupNorm
 from tinygrad.nn.state import torch_load, load_state_dict, get_state_dict
-from extra.models.clip import Closed, Tokenizer, FrozenOpenClipEmbedder
-from extra.models import unet, clip
+from extra.models.clip import Closed, Tokenizer
 from extra.models.unet import UNetModel
-from examples.mlperf.initializers import AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm, zero_module, attn_f32_softmax, gelu_erf
 from extra.bench_log import BenchEvent, WallTimeEvent

 class AttnBlock:
@@ -156,46 +154,12 @@ unet_params: Dict[str,Any] = {
  "use_linear": False,
 }

-mlperf_params: Dict[str,Any] = {"adm_in_ch": None, "in_ch": 4, "out_ch": 4, "model_ch": 320, "attention_resolutions": [4, 2, 1], "num_res_blocks": 2,
-                                "channel_mult": [1, 2, 4, 4], "d_head": 64, "transformer_depth": [1, 1, 1, 1], "ctx_dim": 1024, "use_linear": True,
-                                "num_groups":16, "st_norm_eps":1e-6}
-
 class StableDiffusion:
-  def __init__(self, version:str|None=None, pretrained:str|None=None):
+  def __init__(self):
    self.alphas_cumprod = get_alphas_cumprod()
-    if version != "v2-mlperf-train":
-      self.first_stage_model = AutoencoderKL() # only needed for decoding generated latents to images; not needed in mlperf training from preprocessed moments
-
-    if not version:
-      self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))
-      unet_init_params = unet_params
-    elif version in {"v2-mlperf-train", "v2-mlperf-eval"}:
-      unet_init_params = mlperf_params
-      clip.gelu = gelu_erf
-      self.cond_stage_model = FrozenOpenClipEmbedder(**{"dims": 1024, "n_heads": 16, "layers": 24, "return_pooled": False, "ln_penultimate": True,
-                                                        "clip_tokenizer_version": "sd_mlperf_v5_0"})
-      unet.Linear, unet.Conv2d, unet.GroupNorm, unet.LayerNorm = AutocastLinear, AutocastConv2d, AutocastGroupNorm, AutocastLayerNorm
-      unet.attention, unet.gelu, unet.mixed_precision_dtype = attn_f32_softmax, gelu_erf, dtypes.bfloat16
-      if pretrained:
-        print("loading text encoder")
-        weights: dict[str,Tensor] = {k.replace("cond_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("cond_stage_model.")}
-        weights["model.attn_mask"] = Tensor.full((77, 77), fill_value=float("-inf")).triu(1)
-        load_state_dict(self.cond_stage_model, weights)
-        # only the eval model needs the decoder
-        if version == "v2-mlperf-eval":
-          print("loading image latent encoder")
-          weights = {k.replace("first_stage_model.", "", 1):v for k,v in torch_load(pretrained)["state_dict"].items() if k.startswith("first_stage_model.")}
-          load_state_dict(self.first_stage_model, weights)
-
-    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_init_params))
-    if version == "v2-mlperf-train":
-      # the mlperf reference inits certain weights as zeroes
-      for bb in flatten(self.model.diffusion_model.input_blocks) + self.model.diffusion_model.middle_block + flatten(self.model.diffusion_model.output_blocks):
-        if isinstance(bb, unet.ResBlock):
-          zero_module(bb.out_layers[3])
-        elif isinstance(bb, unet.SpatialTransformer):
-          zero_module(bb.proj_out)
-      zero_module(self.model.diffusion_model.out[2])
+    self.model = namedtuple("DiffusionModel", ["diffusion_model"])(diffusion_model = UNetModel(**unet_params))
+    self.first_stage_model = AutoencoderKL()
+    self.cond_stage_model = namedtuple("CondStageModel", ["transformer"])(transformer = namedtuple("Transformer", ["text_model"])(text_model = Closed.ClipTextTransformer()))

  def get_x_prev_and_pred_x0(self, x, e_t, a_t, a_prev):
    temperature = 1
@@ -263,34 +227,25 @@ if __name__ == "__main__":
  parser.add_argument('--timing', action='store_true', help="Print timing per step")
  parser.add_argument('--seed', type=int, help="Set the random latent seed")
  parser.add_argument('--guidance', type=float, default=7.5, help="Prompt strength")
-  parser.add_argument('--fakeweights', action='store_true', help="Skip loading checkpoints and use fake weights")
  args = parser.parse_args()

-  profile_marker("create model")
  model = StableDiffusion()

-  profile_marker("load in weights")
+  # load in weights
  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):
-    if not args.fakeweights:
-      model_bin = fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt')
-      state_dict = torch_load(model_bin)['state_dict']
-      profile_marker("state dict loaded")
-      load_state_dict(model, state_dict, verbose=False, strict=False, realize=False)
+    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], strict=False)

    if args.fp16:
      for k,v in get_state_dict(model).items():
        if k.startswith("model"):
-          v.replace(v.cast(dtypes.float16))
+          v.replace(v.cast(dtypes.float16).realize())

-    Tensor.realize(*get_state_dict(model).values())
-
-  profile_marker("run clip (conditional)")
+  # run through CLIP to get context
  tokenizer = Tokenizer.ClipTokenizer()
  prompt = Tensor([tokenizer.encode(args.prompt)])
  context = model.cond_stage_model.transformer.text_model(prompt).realize()
  print("got CLIP context", context.shape)

-  profile_marker("run clip (unconditional)")
  prompt = Tensor([tokenizer.encode("")])
  unconditional_context = model.cond_stage_model.transformer.text_model(prompt).realize()
  print("got unconditional CLIP context", unconditional_context.shape)
@@ -311,39 +266,30 @@ if __name__ == "__main__":
  def run(model, *x): return model(*x).realize()

  # this is diffusion
-  step_times = []
  with Context(BEAM=getenv("LATEBEAM")):
    for index, timestep in (t:=tqdm(list(enumerate(timesteps))[::-1])):
-      profile_marker(f"step {len(timesteps)-index-1}")
      GlobalCounters.reset()
-      st = time.perf_counter_ns()
      t.set_description("%3d %3d" % (index, timestep))
      with Timing("step in ", enabled=args.timing, on_exit=lambda _: f", using {GlobalCounters.mem_used/1e9:.2f} GB"):
        with WallTimeEvent(BenchEvent.STEP):
          tid = Tensor([index])
          latent = run(model, unconditional_context, context, latent, Tensor([timestep]), alphas[tid], alphas_prev[tid], Tensor([args.guidance]))
          if args.timing: Device[Device.DEFAULT].synchronize()
-      step_times.append((time.perf_counter_ns() - st)*1e-6)
-    # done with diffusion model
    del run
-    del model.model

-  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
-    min_time = min(step_times)
-    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"
-  profile_marker("run decoder") # upsample latent space to image with autoencoder
-  x = model.decode(latent).realize()
+  # upsample latent space to image with autoencoder
+  x = model.decode(latent)
  print(x.shape)

-  profile_marker("save image")
+  # save image
  im = Image.fromarray(x.numpy())
  print(f"saving {args.out}")
  im.save(args.out)
  # Open image.
  if not args.noshow: im.show()

+  # validation!
  if args.prompt == default_prompt and args.steps == 6 and args.seed == 0 and args.guidance == 7.5:
-    profile_marker("validate")
    ref_image = Tensor(np.array(Image.open(Path(__file__).parent / "stable_diffusion_seed0.png")))
    distance = (((x.cast(dtypes.float) - ref_image.cast(dtypes.float)) / ref_image.max())**2).mean().item()
    assert distance < 3e-3, colored(f"validation failed with {distance=}", "red")  # higher distance with WINO
@@ -19,8 +19,8 @@ from tinygrad.helpers import fetch, getenv

 # QUANT=1 python3 examples/test_onnx_imagenet.py
 #   https://github.com/xamcat/mobcat-samples/raw/refs/heads/master/onnx_runtime/InferencingSample/InferencingSample/mobilenetv2-7.onnx
-# python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
-# VIZ=1 python3 examples/benchmark_onnx.py /tmp/model.quant.onnx
+# DONT_REALIZE_EXPAND=1 python3 examples/test_onnx_imagenet.py /tmp/model.quant.onnx
+# VIZ=1 DONT_REALIZE_EXPAND=1 python3 examples/benchmark_onnx.py /tmp/model.quant.onnx

 def imagenet_dataloader(cnt=0):
  input_mean = Tensor([0.485, 0.456, 0.406]).reshape(1, -1, 1, 1)
@@ -1,34 +0,0 @@
-#!/usr/bin/env python3
-from tinygrad import Tensor, Device, GlobalCounters, Context, dtypes
-from tinygrad.helpers import getenv, colored
-
-SZ = 8_000_000_000
-GPUS = getenv("GPUS", 4) # TODO: expose a way in tinygrad to access this
-
-if __name__ == "__main__":
-  # create tensors
-  tens = [Tensor.ones(SZ, dtype=dtypes.uint8, device=f"{Device.DEFAULT}:{i}").contiguous() for i in range(GPUS)]
-  Tensor.realize(*tens)
-
-  bw = [[0.0]*GPUS for _ in range(GPUS)]
-  for i in range(GPUS):
-    for j in range(GPUS):
-      GlobalCounters.reset()
-      with Context(DEBUG=2):
-        if i == j:
-          # this copy would be optimized out, just add 1
-          (tens[i]+1).realize()
-        else:
-          tens[i].to(f"{Device.DEFAULT}:{j}").realize()
-      t = max(GlobalCounters.time_sum_s, 1e-9)
-      bw[i][j] = SZ / t / 1e9  # GB/s
-
-  def fmt(x):
-    c = "green" if x > 50 else "yellow" if x > 20 else "red"
-    return colored(f"{x:6.1f}", c)
-
-  # header
-  print(" " * 8 + " ".join(f"{'d'+str(j):>6}" for j in range(GPUS)))
-  # rows
-  for i in range(GPUS):
-    print(f"{'s'+str(i):>6} -> " + " ".join(fmt(x) for x in bw[i]))
@@ -1,16 +0,0 @@
-from tinygrad import Tensor, Device, TinyJit, dtypes
-from tinygrad.helpers import getenv
-
-GPUS = getenv("GPUS", 4) # TODO: expose a way in tinygrad to access this
-N = 6144
-
-@TinyJit
-def many_matmul(A, B):
-  out = A
-  for _ in range(8): out = out@B
-  return out
-
-if __name__ == "__main__":
-  A = Tensor.ones(GPUS, N, N, dtype=dtypes.half).shard(devices=tuple([f"{Device.DEFAULT}:{i}" for i in range(GPUS)]), axis=0).contiguous()
-  B = Tensor.ones(GPUS, N, N, dtype=dtypes.half).shard(devices=tuple([f"{Device.DEFAULT}:{i}" for i in range(GPUS)]), axis=0).contiguous()
-  while 1: many_matmul(A, B)
@@ -0,0 +1,104 @@
+import traceback
+import time
+from multiprocessing import Process, Queue
+import numpy as np
+from tinygrad.nn.state import get_parameters
+from tinygrad.nn import optim
+from tinygrad.helpers import getenv, trange
+from tinygrad.tensor import Tensor
+from extra.datasets import fetch_cifar
+from extra.models.efficientnet import EfficientNet
+
+class TinyConvNet:
+  def __init__(self, classes=10):
+    conv = 3
+    inter_chan, out_chan = 8, 16   # for speed
+    self.c1 = Tensor.uniform(inter_chan,3,conv,conv)
+    self.c2 = Tensor.uniform(out_chan,inter_chan,conv,conv)
+    self.l1 = Tensor.uniform(out_chan*6*6, classes)
+
+  def forward(self, x):
+    x = x.conv2d(self.c1).relu().max_pool2d()
+    x = x.conv2d(self.c2).relu().max_pool2d()
+    x = x.reshape(shape=[x.shape[0], -1])
+    return x.dot(self.l1)
+
+if __name__ == "__main__":
+  IMAGENET = getenv("IMAGENET")
+  classes = 1000 if IMAGENET else 10
+
+  TINY = getenv("TINY")
+  TRANSFER = getenv("TRANSFER")
+  if TINY:
+    model = TinyConvNet(classes)
+  elif TRANSFER:
+    model = EfficientNet(getenv("NUM", 0), classes, has_se=True)
+    model.load_from_pretrained()
+  else:
+    model = EfficientNet(getenv("NUM", 0), classes, has_se=False)
+
+  parameters = get_parameters(model)
+  print("parameter count", len(parameters))
+  optimizer = optim.Adam(parameters, lr=0.001)
+
+  BS, steps = getenv("BS", 64 if TINY else 16), getenv("STEPS", 2048)
+  print(f"training with batch size {BS} for {steps} steps")
+
+  if IMAGENET:
+    from extra.datasets.imagenet import fetch_batch
+    def loader(q):
+      while 1:
+        try:
+          q.put(fetch_batch(BS))
+        except Exception:
+          traceback.print_exc()
+    q = Queue(16)
+    for i in range(2):
+      p = Process(target=loader, args=(q,))
+      p.daemon = True
+      p.start()
+  else:
+    X_train, Y_train, _, _ = fetch_cifar()
+    X_train = X_train.reshape((-1, 3, 32, 32))
+    Y_train = Y_train.reshape((-1,))
+
+  with Tensor.train():
+    for i in (t := trange(steps)):
+      if IMAGENET:
+        X, Y = q.get(True)
+      else:
+        samp = np.random.randint(0, X_train.shape[0], size=(BS))
+        X, Y = X_train.numpy()[samp], Y_train.numpy()[samp]
+
+      st = time.time()
+      out = model.forward(Tensor(X.astype(np.float32), requires_grad=False))
+      fp_time = (time.time()-st)*1000.0
+
+      y = np.zeros((BS,classes), np.float32)
+      y[range(y.shape[0]),Y] = -classes
+      y = Tensor(y, requires_grad=False)
+      loss = out.log_softmax().mul(y).mean()
+
+      optimizer.zero_grad()
+
+      st = time.time()
+      loss.backward()
+      bp_time = (time.time()-st)*1000.0
+
+      st = time.time()
+      optimizer.step()
+      opt_time = (time.time()-st)*1000.0
+
+      st = time.time()
+      loss = loss.numpy()
+      cat = out.argmax(axis=1).numpy()
+      accuracy = (cat == Y).mean()
+      finish_time = (time.time()-st)*1000.0
+
+      # printing
+      t.set_description("loss %.2f accuracy %.2f -- %.2f + %.2f + %.2f + %.2f = %.2f" %
+        (loss, accuracy,
+        fp_time, bp_time, opt_time, finish_time,
+        fp_time + bp_time + opt_time + finish_time))
+
+      del out, y, loss
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	652e395efd	not a part of regen	2025-09-05 15:41:34 -07:00
geohot	13e4cbd508	fix torch frontend	2025-09-05 15:39:52 -07:00
geohot	f6b661eb3c	imports	2025-09-05 15:34:49 -07:00
George HotzandGitHub	3e8dabbffc	Merge branch 'master' into delete_kernel_py	2025-09-05 15:31:39 -07:00
geohot	5bac311df4	don't test search	2025-09-05 15:15:05 -07:00
George HotzandGitHub	9a6d7f9a1e	Merge branch 'master' into delete_kernel_py	2025-09-05 15:13:36 -07:00
geohot	8f254d6c48	rip and tear	2025-09-05 14:56:58 -07:00
geohot	9d07e6ae03	delete that file	2025-09-05 14:53:48 -07:00
geohot	645b3c5ae3	delete kernel.py	2025-09-05 14:51:29 -07:00