fix fuse unique

2025-08-04 19:07:49 -07:00
343 changed files with 229680 additions and 13451 deletions
@@ -121,7 +121,7 @@ runs:
        echo 'Acquire::GzipIndexes "true";' | sudo tee /etc/apt/apt.conf.d/gzip
        echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
        echo 'Binary::apt::APT::Keep-Downloaded-Packages "true";' | sudo tee -a /etc/apt/apt.conf.d/99keep-debs
-
+    
    - name: Add OpenCL Repo
      if: inputs.opencl == 'true' && runner.os == 'Linux'
      shell: bash
@@ -174,7 +174,7 @@ runs:
        if [[ "${{ inputs.llvm }}" == "true" ]]; then
          pkgs+=" libllvm20 clang-20 lld-20"
        fi
-
+        
        echo "pkgs=$pkgs" >> "$GITHUB_OUTPUT"
        echo "hash=$(echo -n "$pkgs" | sha256sum | cut -d' ' -f1)" >> "$GITHUB_OUTPUT"

@@ -183,21 +183,21 @@ runs:
      uses: actions/cache@v4
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.APT_CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}

    - name: Run apt Update + Install
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
      shell: bash
      run: |
        sudo apt -qq update || true
-
+        
        # ******** do install ********
        if [[ -n "${{ steps.apt-pkgs.outputs.pkgs }}" ]]; then
          sudo apt-get -y --allow-unauthenticated --no-install-recommends install ${{ steps.apt-pkgs.outputs.pkgs }}
        fi
-
+        
        sudo chown -R $USER:$USER /var/cache/apt/archives/
-
+    
    # **** AMD ****
    - name: Setup AMD (Linux)
      if: inputs.amd == 'true' && runner.os == 'Linux'
@@ -225,25 +225,16 @@ runs:
    - name: Install gpuocelot dependencies (MacOS)
      if: inputs.ocelot == 'true' && runner.os == 'macOS'
      shell: bash
-      run: |
-        pkgs=(cmake ninja llvm@15 zlib glew flex bison [email protected] zstd ncurses)
-        for f in "${pkgs[@]}"; do
-          brew ls --versions "$f" >/dev/null 2>&1 || brew install --quiet "$f"
-        done
-
-        # Fix boost 1.85 for gpuocelot
-        ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
-        ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
-        ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
+      run: brew install --quiet cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses
    - name: Cache gpuocelot
      if: inputs.ocelot == 'true'
      id: cache-build
      uses: actions/cache@v4
      env:
-        cache-name: cache-gpuocelot-build-1
+        cache-name: cache-gpuocelot-build
      with:
        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.BUILD_CACHE_VERSION }}
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-0
    - name: Clone/compile gpuocelot
      if: inputs.ocelot == 'true' && steps.cache-build.outputs.cache-hit != 'true'
      shell: bash
@@ -253,13 +244,7 @@ runs:
        git checkout b16039dc940dc6bc4ea0a98380495769ff35ed99
        mkdir build
        cd build
-
-        CMAKE_ARGS="-Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5"
-        if [[ "${{ 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 -DCMAKE_POLICY_VERSION_MINIMUM=3.5
        ninja
    - name: Install gpuocelot
      if: inputs.ocelot == 'true'
@@ -1,91 +0,0 @@
-name: Autogen
-env:
-  # increment this when downloads substantially change to avoid the internet
-  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 }}
-
-on:
-  push:
-    branches:
-      - master
-  pull_request:
-    paths: 
-    - 'tinygrad/runtime/autogen/**/*'
-  workflow_dispatch:
-    paths: 
-    - 'tinygrad/runtime/autogen/**/*'
-
-jobs:
-  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
@@ -28,7 +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}
@@ -52,26 +52,26 @@ jobs:
    - name: reset process replay
      run: python3.11 test/external/process_replay/reset.py
    - name: Run Stable Diffusion
-      run: BENCHMARK_LOG=stable_diffusion JIT=1 ASSERT_MIN_STEP_TIME=500 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=700 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
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 ASSERT_MIN_STEP_TIME=1600 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=3000 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 python3.11 test/external/external_model_benchmark.py
+    - name: Run huggingface_onnx test
+      run: METAL=1 python3.11 extra/huggingface_onnx/run_models.py test --debug FacebookAI/xlm-roberta-large
    - name: Test speed vs torch
-      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+      run: BIG=2 MPS=1 python3.11 test/test_speed_v_torch.py | tee torch_speed.txt
    - name: Test tensor cores
-      run: METAL=1 python3.11 test/opt/test_tensor_cores.py
+      run: METAL=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
    - 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/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops TestFloat4.test_float4_multidim_amx TestFloat4.test_float4_multidim_unaligned_load_amx
+        DEBUG=2 LLVM=1 AMX=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops 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)
@@ -99,7 +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=8 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
@@ -109,21 +109,21 @@ jobs:
    - name: Train MNIST
      run: time PYTHONPATH=. TARGET_EVAL_ACC_PCT=96.0 python3.11 examples/beautiful_mnist.py | tee beautiful_mnist.txt
    - name: Run 10 CIFAR training steps
-      run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=330 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+      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 ASSERT_MIN_STEP_TIME=385 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      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
    - 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
+      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
+    - 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)
@@ -160,7 +160,7 @@ jobs:
  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}
@@ -189,22 +189,22 @@ jobs:
    - name: Run model inference benchmark
      run: NV=1 CAPTURE_PROCESS_REPLAY=0 NOCLANG=1 python3 test/external/external_model_benchmark.py
    - 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
+      run: NV=1 CAPTURE_PROCESS_REPLAY=0 HALF=1 BIG=2 TORCHCUDA=1 python3 test/test_speed_v_torch.py | tee torch_speed.txt
    - name: Test speed vs theoretical
      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
+        NV=1 ALLOW_TF32=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
+        PTX=1 ALLOW_TF32=1 NV=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
    - 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
    - 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
@@ -214,7 +214,7 @@ jobs:
    - name: Run Stable Diffusion
      run: BENCHMARK_LOG=stable_diffusion NV=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=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      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
@@ -238,9 +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=5 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=5 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
@@ -274,7 +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}
@@ -302,30 +302,30 @@ jobs:
    - 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 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
+      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=85 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=68 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=75 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.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 ASSERT_MIN_STEP_TIME=35 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_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 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=93.2 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=66 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 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NVIDIA Training)
@@ -346,7 +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}
@@ -391,13 +391,13 @@ jobs:
    #- name: Test speed vs torch
    #  run: |
    #    python3 -c "import torch; print(torch.__version__)"
-    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
+    #    LD_PRELOAD="/opt/rocm/lib/libhsa-runtime64.so" HSA=1 BIG=2 TORCHCUDA=1 python3 test/test_speed_v_torch.py | tee torch_speed.txt
    - name: Test speed vs theoretical
      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
-        AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
+        AMD=1 AMD_LLVM=0 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded_amd TestLinearizer.test_tensor_cores_padded_uops
+        AMD=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded_amd TestLinearizer.test_tensor_cores_padded_uops
        AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (AMD)
      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee matmul_amd.txt
@@ -415,9 +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=450 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=1400 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
@@ -443,9 +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
@@ -476,7 +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}
@@ -508,19 +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=85 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=188 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
+      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 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_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 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
+    - 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)
@@ -539,7 +539,7 @@ jobs:
  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}
@@ -570,13 +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=66 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 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee train_bert.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AMD MLPerf)
@@ -605,18 +605,18 @@ 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: benchmark openpilot 0.9.9 driving_vision
-      run: BENCHMARK_LOG=openpilot_0_9_9_vision ASSERT_MIN_STEP_TIME=30 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 ASSERT_MIN_STEP_TIME=45 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 ASSERT_MIN_STEP_TIME=70 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: validate openpilot 0.9.7
+      run: PYTHONPATH=. FLOAT16=0 IMAGE=2 QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
+    - name: benchmark openpilot 0.9.7
+      run: BENCHMARK_LOG=openpilot_0_9_7 PYTHONPATH=. QCOM=1 taskset -c 4-7 python3 test/external/external_benchmark_openpilot.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_0_9_7.txt
+    - name: benchmark openpilot w IMAGE=2 0.9.7
+      run: BENCHMARK_LOG=openpilot_0_9_7_image 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/v0.9.7/selfdrive/modeld/models/supercombo.onnx | tee openpilot_image_0_9_7.txt
+    - name: openpilot compile3 0.9.7
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/selfdrive/modeld/models/supercombo.onnx
+    - name: openpilot compile3 0.9.7+ tomb raider
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/e8bea2c78ffa92685ece511e9b554122aaf1a79d/selfdrive/modeld/models/supercombo.onnx
+    - name: openpilot dmonitoring compile3 0.9.7
+      run: PYTHONPATH="." QCOM=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/v0.9.7/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
@@ -626,7 +626,7 @@ jobs:
        # 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 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
@@ -645,7 +645,7 @@ jobs:
  testreddriverbenchmark:
    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 20
+    timeout-minutes: 15
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -681,8 +681,8 @@ jobs:
    # Fails on 9070
    # - name: Test tensor cores
    #   run: |
-    #     AMD=1 AMD_LLVM=0 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
-    #     AMD=1 AMD_LLVM=1 python3 test/test_linearizer.py test/opt/test_tensor_cores.py
+    #     AMD=1 AMD_LLVM=0 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded_amd TestLinearizer.test_tensor_cores_padded_uops
+    #     AMD=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded_amd TestLinearizer.test_tensor_cores_padded_uops
    #     AMD=1 SHOULD_USE_TC=1 BFLOAT16=1 DEBUG=2 python3 extra/gemm/simple_matmul.py
    - name: Run Tensor Core GEMM (AMD)
      run: AMD=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 ATOL=2e-2 python3 extra/gemm/simple_matmul.py | tee am_matmul_amd.txt
@@ -690,18 +690,13 @@ jobs:
      run: DEBUG=2 AMD=1 python -m pytest -rA test/test_tiny.py
    - name: Test DISK copy time
      run: AMD=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
-    - name: Test CPU copy time
-      run: |
-        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 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 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)
      # TODO: remove BERT_LAYERS once scheduler is fast
-      run: BENCHMARK_LOG=bert_10steps AMD=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 am_train_bert_one_gpu.txt
+      run: BENCHMARK_LOG=bert_10steps AMD=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee am_train_bert_one_gpu.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (AM Driver)
@@ -716,7 +711,7 @@ 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}
@@ -748,22 +743,18 @@ jobs:
    - name: Test driver start time
      run: time DEBUG=3 NV=1 python3 test/test_tiny.py TestTiny.test_plus
    - name: Test tensor cores
-      run: NV=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
+      run: NV=1 ALLOW_TF32=1 python3 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
    - name: Test DISK copy time
      run: NV=1 TESTFILE=/raid/downloads/llama3-8b-sfr/model-00001-of-00004.safetensors python3 test/external/external_benchmark_disk_raw.py
-    - name: Test CPU copy time
-      run: |
-        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyDefaulttoCPUJit
-        NV=1 GRAPH_ONE_KERNEL=1 PYTHONPATH=. NSZ=8192 python3 test/speed/external_test_copy_speed.py TestCopySpeed.testCopyCPUtoDefaultJit
    - name: Test LLAMA-3
      run: BENCHMARK_LOG=llama3_beam 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 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
+      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 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
+      run: BENCHMARK_LOG=bert_10steps NV=1 CAPTURE_PROCESS_REPLAY=0 DEFAULT_FLOAT=HALF BENCHMARK=10 BS=66 GPUS=1 BERT_LAYERS=2 FUSE_ARANGE=1 FUSE_ARANGE_UINT=0 MODEL=bert python3 examples/mlperf/model_train.py | tee nv_train_bert_one_gpu.txt
    - uses: actions/upload-artifact@v4
      with:
        name: Speed (NV Driver)
@@ -20,6 +20,12 @@ repos:
        language: system
        always_run: true
        pass_filenames: false
+      - id: devicetests
+        name: select GPU tests
+        entry: env GPU=1 PYTHONPATH="." python3 -m pytest test/test_uops.py test/test_search.py
+        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
@@ -54,12 +54,11 @@ confidence=
 # --enable=similarities". If you want to run only the classes checker, but have
 # no Warning level messages displayed, use"--disable=all --enable=classes
 # --disable=W"
-disable=C,R,W0613,W0511,W0212,W0201,W0106,W0603,W0621,W0703,W1201,W1203,E1136,W1514,E1101,W0221,W0105,E0401,abstract-method,W0707
+disable=C,R,W0613,W0511,W0212,W0201,W0106,W0603,W0621,W0703,W1201,W1203,E1136,W1514,E1101,W0221,W0105,E0401,abstract-method
 # E1101 for function binding
 # W0221 for Function class
 # W0105 for comment strings
 # E0401 for missing imports
-# W0707 for not reraising

 # Enable the message, report, category or checker with the given id(s). You can
 # either give multiple identifier separated by comma (,) or put this option
@@ -79,8 +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)
@@ -198,7 +198,11 @@ generate_amd() {
  clang2py -k cdefstum \
    extra/hip_gpu_driver/sdma_registers.h \
    extra/hip_gpu_driver/nvd.h \
+    extra/hip_gpu_driver/kfd_pm4_headers_ai.h \
+    extra/hip_gpu_driver/soc21_enum.h \
+    extra/hip_gpu_driver/sdma_v6_0_0_pkt_open.h \
    extra/hip_gpu_driver/gc_11_0_0_offset.h \
+    extra/hip_gpu_driver/gc_10_3_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
@@ -372,6 +376,26 @@ generate_am() {
    -o $BASE/am/pm4_nv.py
  fixup $BASE/am/pm4_nv.py

+  clang2py -k cdefstum \
+    $AMKERN_INC/vega10_enum.h \
+    -o $BASE/am/vega10.py
+  fixup $BASE/am/vega10.py
+
+  clang2py -k cdefstum \
+    $AMKERN_INC/navi10_enum.h \
+    -o $BASE/am/navi10.py
+  fixup $BASE/am/navi10.py
+
+  clang2py -k cdefstum \
+    $AMKERN_INC/soc21_enum.h \
+    -o $BASE/am/soc21.py
+  fixup $BASE/am/soc21.py
+
+  clang2py -k cdefstum \
+    $AMKERN_INC/soc24_enum.h \
+    -o $BASE/am/soc24.py
+  fixup $BASE/am/soc24.py
+
  clang2py -k cdefstum \
    extra/hip_gpu_driver/sdma_registers.h \
    $AMKERN_AMD/amdgpu/vega10_sdma_pkt_open.h \
@@ -80,9 +80,7 @@ print("******** third, the UOp ***********")

 from tinygrad.engine.realize import run_schedule
 from tinygrad.engine.schedule import create_schedule_with_vars
-from tinygrad.helpers import RANGEIFY
 from tinygrad.schedule.kernelize import get_kernelize_map
-from tinygrad.schedule.rangeify import get_rangeify_map

 # allocate some values + load in values
 a = UOp.new_buffer(DEVICE, 1, dtypes.int32)
@@ -95,10 +93,10 @@ out = a + b
 s = UOp(Ops.SINK, dtypes.void, (out,))

 # group the computation into kernels
-becomes_map = get_rangeify_map(s) if RANGEIFY else get_kernelize_map(s)
+becomes_map = get_kernelize_map(s)

 # the compute maps to an assign
-assign = becomes_map[a+b].base
+assign = becomes_map[a+b]

 # the first source is the output buffer (data)
 assert assign.src[0].op is Ops.BUFFER
@@ -18,10 +18,16 @@ Group UOps into kernels.

 ---

-## tinygrad/codegen/opt
+## tinygrad/opt

 Transforms the ast into an optimized ast. This is where BEAM search and heuristics live.

+::: tinygrad.opt.get_optimized_ast
+    options:
+        members: false
+        show_labels: false
+        show_source: false
+
 ---

 ## tinygrad/codegen
@@ -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.

@@ -31,16 +31,19 @@ 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
+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)
@@ -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)
@@ -78,7 +78,6 @@ Elementwise ops operate on a per element basis. They don't change the shape of t
 ::: tinygrad.Tensor.minimum
 ::: tinygrad.Tensor.where
 ::: tinygrad.Tensor.copysign
-::: tinygrad.Tensor.logaddexp

 ## Casting Ops

@@ -6,7 +6,7 @@ If you don't have a tinybox and you want one, see [tinygrad.org](https://tinygra

 ## Welcome

-Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, the green box includes six 4090 GPUs, and the green v2 box includes four 5090 GPUs. Whether you bought a red one or a green one, we want you to love it.
+Welcome to your tinybox! The tinybox is the universal system purpose-built for all AI infrastructure and workloads, from training to inference. The red box includes six 7900XTX GPUs, and the green box includes six 4090 GPUs. Whether you bought a red one or a green one, we want you to love it.

 We don't have a stupid cloud service, you don't have to create a tiny account to set it up, and we aren't tracking how you use the box. We're just happy you bought one. This petaflop is your petaflop.

@@ -47,8 +47,8 @@ Reboot after making these changes or restart the `displayservice.service` servic

 The [default tinybox image](https://github.com/tinygrad/tinyos) ships with tinygrad and PyTorch. While we develop tinygrad, the box is universal hardware. Use whatever framework you desire, run notebooks, download demos, install more things, train, inference, live, laugh, love, you aren't paying per hour for this box so the only limit is your imagination.

-## Building the OS image
+## tinychat

-The OS image is built using `ubuntu-image` from <https://github.com/tinygrad/tinyos>.
+Since LLMs are so popular, we ship with a built in tinygrad based chatbot using a LLaMA-3 finetune. Visit the IP (not the BMC IP) of your tinybox in a web browser on your computer or phone, and you'll find a friendly looking chat interface. This chatbot also provides an OpenAI compatible LLM API on that port, so you can script it.

-After cloning, run `make green` or `make red` to build a tinybox green or tinybox red image respectively.
+The conversations you have with this chatbot are between you and your tinybox. Also, the history in the web app is saved on the client, not the tinybox.
@@ -2,6 +2,7 @@ import time
 start_tm = time.perf_counter()
 import math
 from typing import Tuple, cast
+import numpy as np
 from tinygrad import Tensor, nn, GlobalCounters, TinyJit, dtypes, Device
 from tinygrad.helpers import partition, trange, getenv, Context
 from extra.lr_scheduler import OneCycleLR
@@ -149,12 +150,13 @@ if __name__ == "__main__":
      acc.append((out.argmax(-1) == Y).sum() / eval_batchsize)
    return Tensor.stack(*loss).mean() / (batchsize*loss_batchsize_scaler), Tensor.stack(*acc).mean()

-  Tensor.manual_seed(1337)
-  num_train_samples = X_train.shape[0]
-
+  np.random.seed(1337)
  for epoch in range(math.ceil(hyp['misc']['train_epochs'])):
+    # TODO: move to tinygrad
    gst = time.perf_counter()
-    tidxs = Tensor.randperm(num_train_samples, dtype='int')[:num_steps_per_epoch*batchsize].reshape(num_steps_per_epoch, batchsize)
+    idxs = np.arange(X_train.shape[0])
+    np.random.shuffle(idxs)
+    tidxs = Tensor(idxs, dtype='int')[:num_steps_per_epoch*batchsize].reshape(num_steps_per_epoch, batchsize)  # NOTE: long doesn't fold
    train_loss:float = 0
    for epoch_step in (t:=trange(num_steps_per_epoch)):
      st = time.perf_counter()
@@ -1,12 +1,12 @@
 # model based off https://medium.com/data-science/going-beyond-99-mnist-handwritten-digits-recognition-cfff96337392
-from typing import Callable
+from typing import List, Callable
 from tinygrad import Tensor, TinyJit, nn, GlobalCounters
 from tinygrad.helpers import getenv, colored, trange
 from tinygrad.nn.datasets import mnist

 class Model:
  def __init__(self):
-    self.layers: list[Callable[[Tensor], Tensor]] = [
+    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,
@@ -21,15 +21,17 @@ if __name__ == "__main__":
  X_train, Y_train, X_test, Y_test = mnist(fashion=getenv("FASHION"))

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

  @TinyJit
  @Tensor.train()
  def train_step() -> Tensor:
    opt.zero_grad()
    samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0])
+    # TODO: this "gather" of samples is very slow. will be under 5s when this is fixed
    loss = model(X_train[samples]).sparse_categorical_crossentropy(Y_train[samples]).backward()
-    return loss.realize(*opt.schedule_step())
+    opt.step()
+    return loss

  @TinyJit
  def get_test_acc() -> Tensor: return (model(X_test).argmax(axis=1) == Y_test).mean()*100
@@ -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 ****
@@ -0,0 +1,134 @@
+from extra.models.resnet import ResNet50
+from extra.mcts_search import mcts_search
+from examples.mlperf.helpers import get_mlperf_bert_model
+from tinygrad import Tensor, Device, dtypes, nn
+from tinygrad.opt.kernel import Kernel
+from tinygrad.opt.heuristic import hand_coded_optimizations
+from tinygrad.uop.ops import Ops, sym_infer
+from tinygrad.device import Compiled
+from tinygrad.opt.search import beam_search, bufs_from_lin
+from tinygrad.helpers import DEBUG, ansilen, getenv, colored, TRACEMETA
+from extra.optimization.helpers import time_linearizer
+from tinygrad.engine.realize import get_program
+
+def get_sched_resnet():
+  mdl = ResNet50()
+  optim = (nn.optim.LARS if getenv("LARS") else nn.optim.SGD)(nn.state.get_parameters(mdl))
+  BS = getenv("BS", 64)
+
+  # run model twice to get only what changes, these are the kernels of the model
+  for _ in range(2):
+    out = mdl(Tensor.empty(BS, 3, 224, 224))
+    targets = [out]
+    if getenv("BACKWARD"):
+      optim.zero_grad()
+      out.sparse_categorical_crossentropy(Tensor.empty(BS, dtype=dtypes.int)).backward()
+      targets += [x for x in optim.schedule_step()]
+    sched = Tensor.schedule(*targets)
+    print(f"schedule length {len(sched)}")
+  return sched
+
+def get_sched_bert():
+  mdl = get_mlperf_bert_model()
+  optim = nn.optim.LAMB(nn.state.get_parameters(mdl))
+
+  # fake data
+  BS = getenv("BS", 9)
+  input_ids = Tensor.empty((BS, 512), dtype=dtypes.float32)
+  segment_ids = Tensor.empty((BS, 512), dtype=dtypes.float32)
+  attention_mask = Tensor.empty((BS, 512), dtype=dtypes.default_float)
+  masked_positions = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  masked_lm_ids = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  masked_lm_weights = Tensor.empty((BS, 76), dtype=dtypes.float32)
+  next_sentence_labels = Tensor.empty((BS, 1), dtype=dtypes.float32)
+
+  # run model twice to get only what changes, these are the kernels of the model
+  for _ in range(2):
+    lm_logits, seq_relationship_logits = mdl(input_ids, attention_mask, masked_positions, segment_ids)
+    targets = [lm_logits, seq_relationship_logits]
+    if getenv("BACKWARD"):
+      optim.zero_grad()
+      loss = mdl.loss(lm_logits, seq_relationship_logits, masked_lm_ids, masked_lm_weights, next_sentence_labels)
+      # ignore grad norm and loss scaler for now
+      loss.backward()
+      targets += [x for x in optim.schedule_step()]
+    sched = Tensor.schedule(*targets)
+    print(f"schedule length {len(sched)}")
+  return sched
+
+if __name__ == "__main__":
+  if getenv("HALF", 1):
+    dtypes.default_float = dtypes.half
+
+  # the device we are optimizing for
+  device: Compiled = Device[Device.DEFAULT]
+  if getenv("BACKWARD"): Tensor.training = True
+  print(f"optimizing for {Device.DEFAULT}")
+
+  sched = globals()[f"get_sched_{getenv('MODEL', 'resnet')}"]()
+  sched = [x for x in sched if x.ast.op is Ops.SINK]
+
+  # focus on one kernel
+  if getenv("KERNEL", -1) >= 0: sched = sched[getenv("KERNEL", -1):getenv("KERNEL", -1)+1]
+
+  # work with the schedule
+  total_tm = 0
+  running_gflops = 0
+  usage = {}
+  for i,si in enumerate(sched):
+    if DEBUG >= 3: print(si.ast)
+
+    rawbufs = bufs_from_lin(Kernel(si.ast))
+
+    # "linearize" the op into uops in different ways
+    lins: list[tuple[Kernel, str]] = []
+
+    # always try hand coded opt
+    lin = Kernel(si.ast, opts=device.renderer)
+    lin.apply_opts(hand_coded_optimizations(lin))
+    lins.append((lin, "HC"))
+
+    # maybe try tensor cores
+    lin = Kernel(si.ast, opts=device.renderer)
+    if lin.apply_tensor_cores():
+      lins.append((lin, "TC"))
+
+    # try a beam search
+    if beam:=getenv("BEAM"):
+      lin = Kernel(si.ast, opts=device.renderer)
+      lin = beam_search(lin, rawbufs, beam, bool(getenv("BEAM_ESTIMATE", 1)))
+      lins.append((lin, "BEAM"))
+
+    # try MCTS
+    if mcts:=getenv("MCTS"):
+      lin = Kernel(si.ast, opts=device.renderer)
+      lin = mcts_search(lin, rawbufs, mcts)
+      lins.append((lin, "MCTS"))
+
+    # benchmark the programs
+    choices = []
+    for lin, nm in lins:
+      tm = time_linearizer(lin, rawbufs, allow_test_size=False, cnt=10, disable_cache=True)
+      ops = (prg:=get_program(lin.get_optimized_ast(), lin.opts)).estimates.ops
+      gflops = sym_infer(ops, {k:k.min for k in lin.ast.variables()})*1e-9/tm
+      choices.append((tm, gflops, lin, prg, nm))
+
+    sorted_choices = sorted(choices, key=lambda x: x[0])
+    if DEBUG >= 1: # print all kernels
+      for tm, gflops, lin, prg, nm in choices:
+        print(f"                 kernel {i:2d} {lin.name+' '*(37-ansilen(lin.name))} {str(prg.global_size):18s} {str(prg.local_size):12s} takes {tm*1000:7.2f} ms, {gflops:6.0f} GFLOPS -- {colored(nm, 'green') if lin is sorted_choices[0][2] else nm}")
+
+    tm, gflops, lin, prg, nm = sorted_choices[0]
+    if getenv("SRC"):
+      print(si.ast)
+      print(lin.applied_opts)
+      print(get_program(lin.get_optimized_ast(), lin.opts).src)
+    total_tm += tm
+    running_gflops += gflops * tm
+    if (key := str([str(m) for m in si.metadata])) not in usage: usage[key] = (0, 0)
+    usage[key] = (usage[key][0] + tm, usage[key][1] + 1)
+    print(f"*** {total_tm*1000:7.2f} ms : kernel {i:2d} {lin.name+' '*(37-ansilen(lin.name))} {str(prg.global_size):18s} {str(prg.local_size):12s} takes {tm*1000:7.2f} ms, {gflops:6.0f} GFLOPS {[repr(m) if TRACEMETA >= 2 else str(m) for m in si.metadata]}")
+  print(f"******* total {total_tm*1000:.2f} ms, {running_gflops/total_tm:6.0f} GFLOPS")
+  print("usage:")
+  for k in sorted(usage, key=lambda x: -usage[x][0])[:10]:
+    print(f"{usage[k][0]*1000:.2f} ms: {k} ({usage[k][1]} times)")
@@ -118,7 +118,7 @@ class SpeedyResNet:
 # hyper-parameters were exactly the same as the original repo
 bias_scaler = 58
 hyp = {
-  'seed' : 201,
+  'seed' : 200,
  'opt': {
    'bias_lr':            1.76 * bias_scaler/512,
    'non_bias_lr':        1.76 / 512,
@@ -229,8 +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'])

@@ -356,7 +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
@@ -414,17 +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)
@@ -441,7 +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)
@@ -479,7 +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"
@@ -1,4 +1,6 @@
-import os, random, pickle, queue, struct, math, functools, hashlib, time
+import functools
+import hashlib
+import os, random, pickle, queue, struct, math
 from typing import List
 from pathlib import Path
 from multiprocessing import Queue, Process, shared_memory, connection, Lock, cpu_count
@@ -511,33 +513,6 @@ 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:
@@ -557,21 +532,21 @@ class BinIdxDataset:

    start = self.idx.tell()
    end = start + self.count * dtypes.int32.itemsize
-    self.sizes = self.idx_t[start:end].bitcast(dtypes.int32).numpy()
+    self.sizes = self.idx_t[start:end].bitcast(dtypes.int32)

    start = end
    end = start + self.count * dtypes.int64.itemsize
-    self.pointers = self.idx_t[start:end].bitcast(dtypes.int64).numpy()
+    self.pointers = self.idx_t[start:end].bitcast(dtypes.int64)

    start = end
    end = start + doc_count * dtypes.int64.itemsize
-    self.doc_idx = self.idx_t[start:end].bitcast(dtypes.int64).numpy()
+    self.doc_idx = self.idx_t[start:end].bitcast(dtypes.int64)

    # bin file
    self.bin_t = Tensor(base_path.with_name(f"{base_path.name}.bin"))

  def _index(self, idx) -> tuple[int, int]:
-    return int(self.pointers[idx]), int(self.sizes[idx])
+    return self.pointers[idx].item(), self.sizes[idx].item()

  def get(self, idx, offset:int=0, length:int|None=None):
    ptr, size = self._index(idx)
@@ -591,13 +566,10 @@ class GPTDataset:
    # check for cache
    cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
    cache_path = base_path.with_name(f"{base_path.name}.{cache_hash}.index_cache")
-    print(f"try loading GPTDataset from {cache_path}...")
    if cache_path.exists():
-      print("cache found, loading...")
      with open(cache_path, "rb") as f:
        self.doc_idx, self.sample_idx, self.shuffle_idx = pickle.load(f)
    else:
-      print("cache not found, building index...")
      self.doc_idx = self._build_doc_idx()
      self.sample_idx = self._build_sample_idx()
      self.shuffle_idx = self._build_shuffle_idx()
@@ -656,20 +628,14 @@ class GPTDataset:

  # https://github.com/NVIDIA/Megatron-LM/blob/94bd476bd840c2fd4c3ebfc7448c2af220f4832b/megatron/core/datasets/gpt_dataset.py#L558
  def _build_doc_idx(self):
-    print(f"building doc_idx for {self.num_epochs=}, {self.indexed_dataset.count=}")
-    st = time.perf_counter()
-    # doc_idx = np.mgrid[:self.num_epochs, :self.indexed_dataset.count][1]
-    doc_idx = np.arange(self.indexed_dataset.count).reshape(1, -1).repeat(self.num_epochs, axis=0).flatten()
+    doc_idx = np.mgrid[:self.num_epochs, :self.indexed_dataset.count][1]
+    doc_idx = doc_idx.reshape(-1)
    doc_idx = doc_idx.astype(np.int32)
-    at = time.perf_counter()
    if self.shuffle: self.rng.shuffle(doc_idx)
-    print(f"doc_idx built in {at - st:.3f}s, shuffled in {time.perf_counter() - at:.3f}s")
    return doc_idx

  def _build_sample_idx(self):
-    print(f"building sample_idx for {self.samples=}, {self.seqlen=}, {self.doc_idx.shape[0]=}")
-    sample_idx_max = max(self.doc_idx.shape[0], self.indexed_dataset.sizes.max())
-    sample_idx = np.empty((self.samples + 1, 2), dtype=np.int64 if sample_idx_max > dtypes.int32.max else np.int32)
+    sample_idx = np.empty((self.samples + 1, 2), dtype=np.int32)

    sample_idx_idx, doc_idx_idx, doc_offset = 0, 0, 0
    sample_idx[sample_idx_idx, 0], sample_idx[sample_idx_idx, 1] = doc_idx_idx, doc_offset
@@ -679,7 +645,7 @@ class GPTDataset:
      remaining_seqlen = self.seqlen + 1
      while remaining_seqlen > 0:
        doc_idx = int(self.doc_idx[doc_idx_idx])
-        doc_len = int(self.indexed_dataset.sizes[doc_idx]) - doc_offset
+        doc_len = self.indexed_dataset.sizes[doc_idx].item() - doc_offset
        remaining_seqlen -= doc_len
        if remaining_seqlen <= 0:
          doc_offset += remaining_seqlen + doc_len - 1
@@ -688,7 +654,7 @@ class GPTDataset:
          if doc_idx_idx == len(self.doc_idx) - 1:
            assert sample_idx_idx == self.samples
            doc_idx = int(self.doc_idx[doc_idx_idx])
-            doc_offset = int(self.indexed_dataset.sizes[doc_idx]) - 1
+            doc_offset = self.indexed_dataset.sizes[doc_idx].item() - 1
            break
          doc_idx_idx += 1
          doc_offset = 0
@@ -699,18 +665,13 @@ class GPTDataset:
    return sample_idx

  def _build_shuffle_idx(self):
-    print(f"building shuffle_idx for {self.samples=}")
-    st = time.perf_counter()
    shuffle_idx = np.arange(self.samples, dtype=np.int32)
-    at = time.perf_counter()
    if self.shuffle: self.rng.shuffle(shuffle_idx)
-    print(f"shuffle_idx built in {at - st:.3f}s, shuffled in {time.perf_counter() - at:.3f}s")
    return shuffle_idx

 class BlendedGPTDataset:
  def __init__(self, paths:list[Path], weights:list[float], samples:int, seqlen:int, seed:int, shuffle:bool):
-    self.shuffle = shuffle
-    self.rng = np.random.RandomState(seed)
+    self.seed = seed

    # normalize weights
    total_weight = sum(weights)
@@ -722,47 +683,10 @@ class BlendedGPTDataset:

    self.datasets = [GPTDataset(path, samples_per_blend[i], seqlen, seed + i, shuffle) for i,path in enumerate(paths)]

-    # check for cache
-    cache_hash = hashlib.sha256(f"{samples}:{seqlen}:{seed}:{shuffle}".encode()).hexdigest()
-    cache_path = paths[0].with_name(f"{paths[0].name}.{cache_hash}.blend_cache")
-    print(f"try loading BlendedGPTDataset from {cache_path}...")
-    if cache_path.exists():
-      print("cache found, loading...")
-      with open(cache_path, "rb") as f:
-        self.dataset_idx, self.dataset_sample_idx = pickle.load(f)
-    else:
-      print("cache not found, building index...")
-      self.dataset_idx, self.dataset_sample_idx = self._build_blend_idx()
-      # save cache
-      with open(cache_path, "wb") as f:
-        pickle.dump((self.dataset_idx, self.dataset_sample_idx), f)
-
  def get(self, idx:int):
-    tokens = self.datasets[self.dataset_idx[idx]][self.dataset_sample_idx[idx]]
+    tokens = self.datasets[0][idx]
    return tokens

-  def _build_blend_idx(self):
-    dataset_idx = np.zeros(self.samples, dtype=np.int16)
-    dataset_sample_idx = np.zeros(self.samples, dtype=np.int64)
-
-    unspent_datasets = set(range(len(self.datasets)))
-    dataset_sample_counts = [0] * len(self.datasets)
-
-    for i in tqdm(range(self.samples)):
-      error_argmax, error_max = 0, 0.0
-      for di in unspent_datasets:
-        error = self.weights[di] * max(i, 1) - dataset_sample_counts[di]
-        if error > error_max:
-          error_max = error
-          error_argmax = di
-
-      dataset_idx[i] = error_argmax
-      dataset_sample_idx[i] = dataset_sample_counts[error_argmax]
-
-      dataset_sample_counts[error_argmax] += 1
-
-    return dataset_idx, dataset_sample_idx
-
 def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
  if val:
    dataset = BlendedGPTDataset([
@@ -785,27 +709,6 @@ def batch_load_llama3(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0
      batch.append(tokens)
    yield Tensor.stack(batch, dim=0)

-def batch_load_llama3_small(bs:int, samples:int, seqlen:int, base_dir:Path, seed:int=0, val:bool=True):
-  if val:
-    dataset = BlendedGPTDataset([
-      base_dir / "c4-validation-91205-samples.en_text_document",
-    ], [
-      1.0
-    ], samples, seqlen, seed, False)
-  else:
-    dataset = BlendedGPTDataset([
-      base_dir / "c4-train.en_6_text_document",
-    ], [
-      1.0
-    ], samples, seqlen, seed, True)
-
-  for b in range(math.ceil(samples / bs)):
-    batch = []
-    for i in range(bs):
-      tokens = dataset.get(b * bs + i)
-      batch.append(tokens)
-    yield Tensor.stack(batch, dim=0)
-
 if __name__ == "__main__":
  def load_unet3d(val):
    assert not val, "validation set is not supported due to different sizes on inputs"
@@ -836,8 +739,8 @@ if __name__ == "__main__":

  def load_llama3(val):
    bs = 24
-    samples = 5760 if val else 1_200_000 * 1152
-    seqlen = 8192
+    samples = 5760 if val else 1_200_000
+    seqlen = 512

    max_, min_ = 0, math.inf
    for tokens in tqdm(batch_load_llama3(bs, samples, seqlen, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=5760, val=bool(val)), total=samples//bs):
@@ -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)
@@ -133,59 +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")

@@ -243,299 +243,31 @@ def eval_mrcnn():

 def eval_llama3():
  from extra.models.llama import Transformer
-  from examples.llama3 import MODEL_PARAMS, load, convert_from_huggingface
+  from examples.llama3 import MODEL_PARAMS
  from tinygrad.helpers import tqdm

-  BASEDIR = Path(getenv("BASEDIR", "/raid/datasets/c4/"))
-  BS = getenv("BS", 4)
-  SMALL = getenv("SMALL", 0)
-  SEQLEN = getenv("SEQLEN", 8192)
-  MODEL_PATH = Path(getenv("MODEL_PATH", "/raid/weights/llama31_8b/"))
+  bs = 4
+  sequence_length = 512

-  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)
-
-  # load weights
-  weights = load(str(MODEL_PATH / "model.safetensors.index.json"))
-  if "model.embed_tokens.weight" in weights:
-    print("converting from huggingface format")
-    weights = convert_from_huggingface(weights, params["n_layers"], params["n_heads"], params["n_kv_heads"])
-
-  load_state_dict(model, weights, strict=False, consume=True)
+  model = Transformer(**(MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]|{"vocab_size": 32000}), max_context=sequence_length, jit=False, disable_kv_cache=True)

  @TinyJit
  def eval_step(model, tokens):
    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
-    return loss.flatten().float()
+    return loss.flatten()

-  if SMALL:
-    from examples.mlperf.dataloader import batch_load_llama3_small
-    iter = batch_load_llama3_small(BS, 5760, SEQLEN, BASEDIR, val=True)
-  else:
-    from examples.mlperf.dataloader import batch_load_llama3
-    iter = batch_load_llama3(BS, 5760, SEQLEN, BASEDIR, val=True)
+  from examples.mlperf.dataloader import batch_load_llama3
+  iter = batch_load_llama3(bs, 5760, sequence_length, Path(getenv("BASEDIR", "/raid/datasets/c4/")), True)

  losses = []
-  for tokens in tqdm(iter, total=5760//BS):
+  for tokens in tqdm(iter, total=5760//bs):
    GlobalCounters.reset()
    losses += eval_step(model, tokens).tolist()
    tqdm.write(f"loss: {np.mean(losses)}")

-  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
+  log_perplexity = Tensor(losses).mean()
+  print(f"Log Perplexity: {log_perplexity.item()}")

 if __name__ == "__main__":
  # inference only
@@ -4,7 +4,7 @@ import multiprocessing

 from tinygrad import Device, GlobalCounters, Tensor, TinyJit, dtypes
 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.state import get_parameters, get_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW

 from extra.lr_scheduler import LRSchedulerGroup
@@ -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
@@ -1296,18 +1290,13 @@ def train_llama3():
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup

  config = {}
-  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)
  GBS                = config["GLOBAL_BATCH_SIZE"]      = BS * grad_acc
  SEED               = config["SEED"]                   = getenv("SEED", 5760)
  SEQLEN             = config["SEQLEN"]                 = getenv("SEQLEN", 8192)
  TRAIN_ON_VAL       = config["TRAIN_ON_VAL"]           = getenv("TRAIN_ON_VAL", 0)
-  SMALL              = config["SMALL"]                  = getenv("SMALL", 0)
-  SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000 * 1152)
-  EVAL_FREQ          = config["EVAL_FREQ"]              = getenv("EVAL_FREQ", 46080)
-  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 16)
-  EVAL_TARGET        = config["EVAL_TARGET"]            = getenv("EVAL_TARGET", 5.6)
+  SAMPLES            = config["SAMPLES"]                = getenv("SAMPLES", 5_760 if TRAIN_ON_VAL else 1_200_000)

  # 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
@@ -1319,58 +1308,44 @@ def train_llama3():

  opt_gradient_clip_norm = 1.0
  opt_learning_rate_warmup_steps = getenv("WARMUP_STEPS", math.ceil(8000 * 1152 / GBS))
-  opt_learning_rate_decay_steps = getenv("MAX_STEPS", math.ceil(1_200_000 * 1152 / GBS)) - opt_learning_rate_warmup_steps
+  opt_learning_rate_decay_steps = getenv("DECAY_STEPS", math.ceil(1_200_000 * 1152 / GBS) - opt_learning_rate_warmup_steps)
  opt_base_learning_rate = getenv("LR", 8e-5 * GBS / 1152)  # NOTE: cannot change for benchmark
-  opt_end_learning_rate = getenv("END_LR", 8e-7)
+  opt_end_learning_rate = 8e-7

  # TODO: confirm weights are in bf16
  # vocab_size from the mixtral tokenizer
-  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
-  params = params | {"vocab_size": 32000} if not SMALL else params
+  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]|{"vocab_size": 32000}
  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):
-      v = v.assign(Tensor.empty(v.shape))
-
  if (DP := getenv("DP", 1)) > 1:
    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
    for v in get_parameters(model):
      v.shard_(device, axis=None)

-  if (MP := getenv("MP", 1)) > 1:
-    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
-    for k,v in get_state_dict(model).items():
-      if 'scale' in k: v.shard_(device, axis=None)  # from quantized
-      elif '.attention.wq' in k: v.shard_(device, axis=0)
-      elif '.attention.wk' in k: v.shard_(device, axis=0)
-      elif '.attention.wv' in k: v.shard_(device, axis=0)
-      elif '.attention.wo' in k: v.shard_(device, axis=1)
-      elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
-      elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
-      elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
-      elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
-      elif 'output.weight' in k: v.shard_(device, axis=0)
-      else:
-        # attention_norm, ffn_norm, norm
-        v.shard_(device, axis=None)
-      # prevents memory spike on device 0
-      v.realize()
+  # TODO: MP
+  # if (GPUS := getenv("GPUS", 1)) > 1:
+  #   device = tuple(f"{Device.DEFAULT}:{i}" for i in range(GPUS))
+  #   for k,v in get_state_dict(model).items():
+  #     if 'scale' in k: v.shard_(device, axis=None)  # from quantized
+  #     # elif '.attention.wq' in k: v.shard_(device, axis=0)
+  #     # elif '.attention.wk' in k: v.shard_(device, axis=0)
+  #     # elif '.attention.wv' in k: v.shard_(device, axis=0)
+  #     # elif '.attention.wo' in k: v.shard_(device, axis=1)
+  #     # elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
+  #     # elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
+  #     # elif '.feed_forward.w3.' in k: v.shard_(device, axis=0)
+  #     # elif 'tok_embeddings.weight' in k: v.shard_(device, axis=0)
+  #     elif 'output.weight' in k: v.shard_(device, axis=0)  # 243.32
+  #     else:
+  #       # print(k)
+  #       # attention_norm, ffn_norm, norm
+  #       v.shard_(device, axis=None)

  optim = AdamW(get_parameters(model), lr=0.0,
                b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay)
  scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)

-  if resume_ckpt := getenv("RESUME_CKPT"):
-    fn = f"./ckpts/llama3_{resume_ckpt}.safe"
-    print(f"loading initial checkpoint from {fn}")
-    load_state_dict(model, safe_load(fn), realize=False)
-
-    fn = f"./ckpts/llama3_{resume_ckpt}_optim.safe"
-    print(f"loading optim checkpoint from {fn}")
-    load_state_dict(scheduler, safe_load(fn), realize=False)
-
  @TinyJit
  @Tensor.train()
  def train_step(model, tokens:Tensor, grad_acc:int):
@@ -1380,9 +1355,6 @@ def train_llama3():
      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()
@@ -1396,7 +1368,7 @@ def train_llama3():
        total_norm += p.grad.float().square().sum()
      total_norm = total_norm.sqrt().contiguous()
      for p in optim.params:
-        p.grad = p.grad * (opt_gradient_clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)
+        p.grad = p.grad * opt_gradient_clip_norm / (total_norm + 1e-6)

    optim.step()
    scheduler.step()
@@ -1405,231 +1377,32 @@ def train_llama3():
    loss.realize(lr)
    return loss, lr

-  @TinyJit
-  @Tensor.train(False)
-  def eval_step(model, tokens:Tensor):
-    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:])
-    return loss.flatten().float()
+  if getenv("FAKEDATA", 0):
+    def fake_data():
+      for _ in range(SAMPLES // GBS):
+        yield Tensor.randint(GBS, SEQLEN + 1, low=0, high=32000, dtype=dtypes.int32, device=Device.DEFAULT)
+    iter = fake_data()
+  else:
+    from examples.mlperf.dataloader import batch_load_llama3
+    iter = batch_load_llama3(GBS, SAMPLES, SEQLEN, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=SEED, val=bool(TRAIN_ON_VAL))

-  # ** data iters **
-  def fake_data(bs, samples):
-    for _ in range(samples // bs):
-      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(GBS, SAMPLES)
-    else:
-      if SMALL:
-        from examples.mlperf.dataloader import batch_load_llama3_small
-        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(GBS, SAMPLES, SEQLEN, BASEDIR, seed=SEED, val=bool(TRAIN_ON_VAL))
-
-  def get_eval_iter():
-    if getenv("FAKEDATA", 0):
-      return fake_data(EVAL_BS, 5760)
-    else:
-      if SMALL:
-        from examples.mlperf.dataloader import batch_load_llama3_small
-        return batch_load_llama3_small(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)
-      else:
-        from examples.mlperf.dataloader import batch_load_llama3
-        return batch_load_llama3(EVAL_BS, 5760, SEQLEN, BASEDIR, val=True)
-
-  iter = get_train_iter()
-  i, sequences_seen = resume_ckpt, 0
+  i = 0
  for tokens in tqdm(iter, total=SAMPLES//GBS):
    t = time.perf_counter()
    GlobalCounters.reset()
    loss, lr = train_step(model, tokens, grad_acc)
-    loss = loss.float().item()
-
-    i += 1
-    sequences_seen += tokens.shape[0]
-
-    tqdm.write(f"{loss:.4f} loss, {lr.item():.12f} LR, {GlobalCounters.mem_used / 1e9:.2f} GB used, {time.perf_counter()-t:.2f} s")
+    # above as tqdm.write f-string
+    tqdm.write(f"{loss.item():.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")
+        f.write(f"{i} {loss.item():.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)):
+    if getenv("CKPT") and (i % 200 == 0 or i == 10):
      tqdm.write("saving checkpoint")
      if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
-      fn = f"{ckpt_dir}/llama3_{i}.safe"
+      fn = f"{ckpt_dir}/{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)
-
-    if sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1):
-      tqdm.write(f"evaluating after {sequences_seen} sequences")
-
-      # run eval
-      eval_losses = []
-      eval_iter = get_eval_iter()
-      tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")
-
-      for tokens in tqdm(eval_iter, total=5760//EVAL_BS):
-        eval_losses += eval_step(model, tokens).tolist()
-      log_perplexity = Tensor(eval_losses).mean().float().item()
-
-      tqdm.write(f"eval log perplexity: {log_perplexity:.4f}")
-
-      if log_perplexity < EVAL_TARGET:
-        tqdm.write(f"target achieved after {sequences_seen} sequences")
-        if getenv("CKPT"):
-          if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
-          fn = f"{ckpt_dir}/llama3.safe"
-          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()
+    i += 1

 if __name__ == "__main__":
  multiprocessing.set_start_method('spawn')
@@ -1639,7 +1412,7 @@ 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
@@ -4,8 +4,6 @@ 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
@@ -5,8 +5,6 @@ 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 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"
@@ -8,8 +8,6 @@ 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 TRAIN_STEPS=3900

-export IGNORE_OOB=1
-
 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"
@@ -11,8 +11,6 @@ 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 TRAIN_STEPS=3900

-export IGNORE_OOB=1
-
 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"
@@ -2,9 +2,9 @@

 export PYTHONPATH="." NV=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0

 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
@@ -2,9 +2,9 @@

 export PYTHONPATH="." NV=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0

 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
@@ -5,9 +5,9 @@ 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=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0

 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
@@ -2,9 +2,9 @@

 export PYTHONPATH="." AMD=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0

 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
@@ -2,9 +2,9 @@

 export PYTHONPATH="." AMD=1
 export MODEL="bert"
-export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0

 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
@@ -5,9 +5,9 @@ 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=90 EVAL_BS=90
+export DEFAULT_FLOAT="HALF" SUM_DTYPE="HALF" GPUS=6 BS=96 EVAL_BS=96

-export IGNORE_OOB=1
+export FUSE_ARANGE=1 FUSE_ARANGE_UINT=0

 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
@@ -1,7 +1,8 @@
 # https://arxiv.org/pdf/2409.02060
-import time, functools
+import time
 import numpy as np
 np.set_printoptions(suppress=True, linewidth=1000)
+import functools
 from tinygrad import Tensor, nn, Device, GlobalCounters
 from tinygrad.helpers import Timing, getenv
 from extra.models.llama import Transformer, convert_from_huggingface
@@ -16,7 +17,7 @@ class MixtureFeedForward:
  def __call__(self, x:Tensor) -> Tensor:
    assert x.shape[0] == 1, "only BS=1"
    assert x.shape[1] == 1, "only length=1"
-    g = self.gate(x).softmax(-1)
+    g = self.gate(x).float().softmax(-1)

    g = g.squeeze() # (BS, length, num_experts) -> (num_experts,)
    probs, sel = g.topk(self.activated_experts)
@@ -24,7 +25,7 @@ class MixtureFeedForward:
    # run MoE
    x_up_gate = x.dot(self.gate_proj[sel].permute(0,2,1)).silu() * x.dot(self.up_proj[sel].permute(0,2,1))
    x_down = x_up_gate.dot(self.down_proj[sel].permute(0,2,1))
-    return (x_down * probs.reshape(self.activated_experts, 1, 1)).sum(axis=0)
+    return (x_down.float() * probs.reshape(self.activated_experts, 1, 1)).sum(axis=0)

 # model is bf16, 1.3B active, 6.9B total
 # M3 Max is 400 GB/s, so 400/2.6 = ~154 tok/s
@@ -6,7 +6,7 @@ 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 CL=1 IMAGE=2 FLOAT16=1 VIZ=1 DEBUG=2 python3 examples/openpilot/compile4.py
+# 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"
@@ -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)
@@ -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")
@@ -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
+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
@@ -302,23 +266,17 @@ 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])):
      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)
    del run

-  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"
  # upsample latent space to image with autoencoder
  x = model.decode(latent)
  print(x.shape)
@@ -109,7 +109,7 @@ class TextDecoder:

  def forward(self, x:Tensor, pos:Union[Variable, Literal[0]], encoded_audio:Tensor):
    seqlen = x.shape[-1]
-    x = self.token_embedding(x) + self.positional_embedding.shrink(((pos, pos+seqlen), None))
+    x = self.token_embedding(x) + self.positional_embedding.shrink(((pos, pos+seqlen), None, None))
    for block in self.blocks: x = block(x, xa=encoded_audio, mask=self.mask, len=pos)
    return self.output_tok(x)

@@ -71,8 +71,8 @@ def bbox_iou(box1, box2):
  # get the coordinates of the intersection rectangle
  inter_rect_x1 = np.maximum(b1_x1, b2_x1)
  inter_rect_y1 = np.maximum(b1_y1, b2_y1)
-  inter_rect_x2 = np.minimum(b1_x2, b2_x2)
-  inter_rect_y2 = np.minimum(b1_y2, b2_y2)
+  inter_rect_x2 = np.maximum(b1_x2, b2_x2)
+  inter_rect_y2 = np.maximum(b1_y2, b2_y2)
  #Intersection area
  inter_area = np.clip(inter_rect_x2 - inter_rect_x1 + 1, 0, 99999) * np.clip(inter_rect_y2 - inter_rect_y1 + 1, 0, 99999)
  #Union Area
@@ -297,13 +297,13 @@ class Darknet:
          # Get the number of weights of batchnorm
          num_bn_biases = math.prod(bn.bias.shape)
          # Load weights
-          bn_biases = Tensor(weights[ptr:ptr + num_bn_biases].astype(np.float32))
+          bn_biases = Tensor(weights[ptr:ptr + num_bn_biases])
          ptr += num_bn_biases
-          bn_weights = Tensor(weights[ptr:ptr+num_bn_biases].astype(np.float32))
+          bn_weights = Tensor(weights[ptr:ptr+num_bn_biases])
          ptr += num_bn_biases
-          bn_running_mean = Tensor(weights[ptr:ptr+num_bn_biases].astype(np.float32))
+          bn_running_mean = Tensor(weights[ptr:ptr+num_bn_biases])
          ptr += num_bn_biases
-          bn_running_var = Tensor(weights[ptr:ptr+num_bn_biases].astype(np.float32))
+          bn_running_var = Tensor(weights[ptr:ptr+num_bn_biases])
          ptr += num_bn_biases
          # Cast the loaded weights into dims of model weights
          bn_biases = bn_biases.reshape(shape=tuple(bn.bias.shape))
@@ -319,7 +319,7 @@ class Darknet:
          # load biases of the conv layer
          num_biases = math.prod(conv.bias.shape)
          # Load weights
-          conv_biases = Tensor(weights[ptr: ptr+num_biases].astype(np.float32))
+          conv_biases = Tensor(weights[ptr: ptr+num_biases])
          ptr += num_biases
          # Reshape
          conv_biases = conv_biases.reshape(shape=tuple(conv.bias.shape))
@@ -327,7 +327,7 @@ class Darknet:
          conv.bias = conv_biases
        # Load weighys for conv layers
        num_weights = math.prod(conv.weight.shape)
-        conv_weights = Tensor(weights[ptr:ptr+num_weights].astype(np.float32))
+        conv_weights = Tensor(weights[ptr:ptr+num_weights])
        ptr += num_weights
        conv_weights = conv_weights.reshape(shape=tuple(conv.weight.shape))
        conv.weight = conv_weights
@@ -371,7 +371,7 @@ class Darknet:
 if __name__ == "__main__":
  model = Darknet(fetch('https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov3.cfg').read_bytes())
  print("Loading weights file (237MB). This might take a while…")
-  model.load_weights('https://github.com/shadiakiki1986/yolov3.weights/releases/download/3.0.1/yolov3.weights')
+  model.load_weights('https://pjreddie.com/media/files/yolov3.weights')
  if len(sys.argv) > 1:
    url = sys.argv[1]
  else:
@@ -1,16 +1,6 @@
-import re, ctypes, sys, importlib
+import re, ctypes, sys

-from tinygrad.runtime.support.am.amdev import AMDev, AMRegister
-class AMDFake(AMDev):
-  def __init__(self, devfmt, vram, doorbell, mmio, dma_regions=None):
-    self.devfmt, self.vram, self.doorbell64, self.mmio, self.dma_regions = devfmt, vram, doorbell, mmio, dma_regions
-    self._run_discovery()
-    self._build_regs()
-
-amdev = importlib.import_module("tinygrad.runtime.support.am.amdev")
-amdev.AMDev = AMDFake
-
-from tinygrad.runtime.ops_amd import PCIIface
+from tinygrad.runtime.autogen.am import am, mp_11_0, mp_13_0_0, nbio_4_3_0, mmhub_3_0_0, gc_11_0_0, osssys_6_0_0

 def parse_amdgpu_logs(log_content, register_names=None):
  register_map = register_names
@@ -33,11 +23,16 @@ def parse_amdgpu_logs(log_content, register_names=None):
  return processed_log

 def main():
+  regs_offset = {13: {0: [3072, 37784576]}, 28: {0: [93184, 37754880], 1: [201327616, 201461760], 2: [209716224, 209850368], 3: [218104832, 218238976], 4: [226493440, 226627584], 5: [234882048, 235016192], 6: [243270656, 243404800]}, 21: {0: [28672, 12582912, 37795840, 130023424, 306184192], 1: [201326592, 201463808, 201465856, 204210176, 204472320], 2: [209715200, 209852416, 209854464, 212598784, 212860928], 3: [218103808, 218241024, 218243072, 220987392, 221249536], 4: [226492416, 226629632, 226631680, 229376000, 229638144], 5: [234881024, 235018240, 235020288, 237764608, 238026752], 6: [243269632, 243406848, 243408896, 246153216, 246415360]}, 22: {0: [18, 192, 13504, 36864, 37764096]}, 1: {0: [4704, 40960, 114688, 37760000]}, 2: {0: [3872, 37790720]}, 11: {0: [70656, 38103040]}, 12: {0: [106496, 37783552]}, 15: {0: [90112, 14417920, 14680064, 14942208, 38009856]}, 16: {0: [90112, 14417920, 14680064, 14942208, 38009856]}, 14: {0: [0, 20, 3360, 66560, 37859328, 67371008]}, 26: {0: [0, 20, 3360, 66560, 37859328, 67371008]}, 23: {0: [4256, 37789696]}, 33: {0: [0, 20, 3360, 66560, 37859328, 67371008]}, 25: {0: []}, 3: {0: [4704, 40960, 114688, 37760000]}, 4: {0: [4704, 40960, 114688, 37760000]}, 24: {0: [92160, 92672, 37752832, 54788096]}, 27: {0: [91648, 37751808], 1: [201339904, 201458176], 2: [209728512, 209846784], 3: [218117120, 218235392], 4: [226505728, 226624000], 5: [234894336, 235012608], 6: [243282944, 243401216]}, 29: {0: [201342976, 201344000, 205520896, 205537280], 1: [209731584, 209732608, 213909504, 213925888], 2: [218120192, 218121216, 222298112, 222314496], 3: [226508800, 226509824, 230686720, 230703104], 4: [234897408, 234898432, 239075328, 239091712], 5: [243286016, 243287040, 247463936, 247480320]}, 17: {0: [30720, 32256], 1: [31488, 73728]}}
+
  reg_names = {}
-  dev = PCIIface(None, 0)
-  for x, y in dev.dev_impl.__dict__.items():
-    if isinstance(y, AMRegister):
-      for inst, addr in y.addr.items(): reg_names[addr] = f"{x}, xcc={inst}"
+  def _prepare_registers(modules):
+    for base, m in modules:
+      for k, regval in m.__dict__.items():
+        if k.startswith("reg") and not k.endswith("_BASE_IDX") and (base_idx:=getattr(m, f"{k}_BASE_IDX", None)) is not None:
+          reg_names[regs_offset[am.__dict__.get(f"{base}_HWIP")][0][base_idx] + regval] = k
+
+  _prepare_registers([("MP0", mp_13_0_0), ("NBIO", nbio_4_3_0), ("MMHUB", mmhub_3_0_0), ("GC", gc_11_0_0), ("OSSSYS", osssys_6_0_0)])

  with open(sys.argv[1], 'r') as f:
    log_content = log_content_them = f.read()
@@ -1,7 +1,7 @@
 # copying the kernels from https://github.com/microsoft/ArchProbe into Python
 import numpy as np
 import pickle
-from tinygrad.runtime.ops_cl import CLProgram, CLBuffer
+from tinygrad.runtime.ops_gpu import CLProgram, CLBuffer
 from tinygrad import dtypes
 from tqdm import trange, tqdm
 from matplotlib import pyplot as plt
@@ -1,5 +1,5 @@
 from typing import Tuple, List, NamedTuple, Any, Dict, Optional, Union, DefaultDict, cast
-from tinygrad.codegen.opt.kernel import Ops, MemOp, UOp
+from tinygrad.opt.kernel import Ops, MemOp, UOp
 from tinygrad.uop.ops import BinaryOps, UnaryOps
 from tinygrad.dtype import DType, dtypes
 from tinygrad.helpers import DEBUG
@@ -3,7 +3,7 @@ from platform import system
 from typing import Tuple, Dict, List, Optional
 from tinygrad import dtypes
 from tinygrad.uop.ops import BinaryOps, UnaryOps, TernaryOps
-from tinygrad.codegen.opt.kernel import Ops, UOp
+from tinygrad.opt.kernel import Ops, UOp
 from tinygrad.helpers import CI
 from tinygrad.codegen.assembly import uops_to_asmstyle, AssemblyLanguage

@@ -1,7 +1,7 @@
 from typing import List
 import struct
 from tinygrad.codegen.assembly import uops_to_asmstyle, AssemblyLanguage
-from tinygrad.codegen.opt.kernel import Ops, UOp
+from tinygrad.opt.kernel import Ops, UOp
 from tinygrad import dtypes
 from tinygrad.uop.ops import BinaryOps, UnaryOps, TernaryOps
 from tinygrad.runtime.ops_cuda import arch
@@ -2,9 +2,9 @@ import yaml
 from typing import Tuple, Set, Dict
 from tinygrad import dtypes
 from tinygrad.codegen.assembly import AssemblyCodegen, Register
-from tinygrad.codegen.opt.kernel import Ops
+from tinygrad.opt.kernel import Ops
 from tinygrad.uop.ops import BinaryOps, UnaryOps, TernaryOps
-from tinygrad.runtime.ops_cl import ROCM_LLVM_PATH
+from tinygrad.runtime.ops_gpu import ROCM_LLVM_PATH

 # ugh, is this really needed?
 from extra.helpers import enable_early_exec
@@ -5,7 +5,7 @@ from tinygrad.helpers import colored
 from extra.helpers import enable_early_exec
 early_exec = enable_early_exec()

-from tinygrad.runtime.ops_cl import CLProgram, CLBuffer, ROCM_LLVM_PATH
+from tinygrad.runtime.ops_gpu import CLProgram, CLBuffer, ROCM_LLVM_PATH

 ENABLE_NON_ASM = False

@@ -10,13 +10,13 @@ from tinygrad.renderer.cstyle import ClangRenderer
 render_dtype = ClangRenderer().render_dtype

 class ClangGraph(GraphRunner):
-  def __init__(self, jit_cache: List[ExecItem], input_rawbuffers: List[Buffer], var_vals: Dict[str, int]):
+  def __init__(self, jit_cache: List[ExecItem], input_rawbuffers: List[Buffer], var_vals: Dict[Variable, int]):
    super().__init__(jit_cache, input_rawbuffers, var_vals)
    if not all(isinstance(ji.prg, CompiledRunner) for ji in jit_cache): raise GraphException

    prgs = '\n'.join(dedup([cast(CompiledRunner, ji.prg).p.src for ji in jit_cache]))
    args = [f"{render_dtype(x.dtype)}* arg{i}" for i,x in enumerate(input_rawbuffers)]
-    args += sorted([f"int {v}" for v in var_vals])
+    args += sorted([f"int {v.expr}" for v in var_vals])
    code = ["void batched("+','.join(args)+") {"]
    for ji in jit_cache:
      args = []
@@ -34,6 +34,6 @@ class ClangGraph(GraphRunner):
    assert compiler is not None
    self._prg = ClangProgram("batched", compiler.compile(prgs+"\n"+"\n".join(code))) # no point in caching the pointers

-  def __call__(self, rawbufs: List[Buffer], var_vals: Dict[str, int], wait=False):
+  def __call__(self, rawbufs: List[Buffer], var_vals: Dict[Variable, int], wait=False):
    return cpu_time_execution(
-    lambda: self._prg(*[x._buf for x in rawbufs], *[x[1] for x in sorted(var_vals.items(), key=lambda x: x[0])]), enable=wait)
+    lambda: self._prg(*[x._buf for x in rawbufs], *[x[1] for x in sorted(var_vals.items(), key=lambda x: x[0].expr)]), enable=wait)
@@ -26,7 +26,7 @@ class VirtAQLQueue(AQLQueue):
    self.available_packet_slots -= 1

 class HSAGraph(MultiGraphRunner):
-  def __init__(self, jit_cache: List[ExecItem], input_rawbuffers: List[Buffer], var_vals: Dict[str, int]):
+  def __init__(self, jit_cache: List[ExecItem], input_rawbuffers: List[Buffer], var_vals: Dict[Variable, int]):
    super().__init__(jit_cache, input_rawbuffers, var_vals)

    # Check all jit items are compatible.
@@ -53,7 +53,7 @@ class HSAGraph(MultiGraphRunner):
      self.ji_kargs_structs[j] = ji.prg._prg.args_struct_t.from_address(kernargs_ptrs[ji.prg.dev])
      kernargs_ptrs[ji.prg.dev] += round_up(ctypes.sizeof(ji.prg._prg.args_struct_t), 16)
      for i in range(len(ji.bufs)): self.ji_kargs_structs[j].__setattr__(f'f{i}', cast(Buffer, ji.bufs[i])._buf)
-      for i in range(len(ji.prg.p.vars)): self.ji_kargs_structs[j].__setattr__(f'v{i}', var_vals[ji.prg.p.vars[i].expr])
+      for i in range(len(ji.prg.p.vars)): self.ji_kargs_structs[j].__setattr__(f'v{i}', var_vals[ji.prg.p.vars[i]])

    # Build queues.
    self.virt_aql_queues: Dict[Compiled, VirtAQLQueue] = {dev:VirtAQLQueue(dev, 2*len(self.jit_cache)+16) for dev in self.devices}
@@ -106,7 +106,7 @@ class HSAGraph(MultiGraphRunner):
    for sig in self.signals_to_reset: hsa.hsa_signal_silent_store_relaxed(sig, 0)
    hsa.hsa_signal_silent_store_relaxed(self.finish_signal, 0)

-  def __call__(self, input_rawbuffers: List[Buffer], var_vals: Dict[str, int], wait=False) -> Optional[float]:
+  def __call__(self, input_rawbuffers: List[Buffer], var_vals: Dict[Variable, int], wait=False) -> Optional[float]:
    # Wait and restore signals
    hsa.hsa_signal_wait_scacquire(self.finish_signal, hsa.HSA_SIGNAL_CONDITION_LT, 1, (1 << 64) - 1, hsa.HSA_WAIT_STATE_ACTIVE)
    for sig in self.signals_to_reset: hsa.hsa_signal_silent_store_relaxed(sig, 1)
@@ -123,7 +123,7 @@ class HSAGraph(MultiGraphRunner):
    # Update var_vals
    for j in self.jc_idx_with_updatable_var_vals:
      for i,v in enumerate(cast(CompiledRunner, self.jit_cache[j].prg).p.vars):
-        self.ji_kargs_structs[j].__setattr__(f'v{i}', var_vals[v.expr])
+        self.ji_kargs_structs[j].__setattr__(f'v{i}', var_vals[v])

    # Update launch dims
    for j in self.jc_idx_with_updatable_launch_dims:
@@ -29,10 +29,10 @@ def uops_to_rdna(function_name:str, uops:UOpGraph) -> str:
  r: Dict[UOp, str] = {}
  for u in uops:
    if u.uop == UOps.SPECIAL:
-      if u.arg.startswith("lidx"):
-        r[u] = f'v{u.src[0].arg}'
-      elif u.arg.startswith("gidx"):
-        r[u] = f's{2+u.src[0].arg}'
+      if u.arg[1].startswith("lidx"):
+        r[u] = f'v{u.arg[0]}'
+      elif u.arg[1].startswith("gidx"):
+        r[u] = f's{2+u.arg[0]}'
      else:
        raise NotImplementedError
    elif u.uop == UOps.CONST:
@@ -2,7 +2,7 @@ from typing import Dict, List, Final, Callable, DefaultDict
 from collections import defaultdict
 from tinygrad.uop.ops import UnaryOps, BinaryOps, TernaryOps, Op
 from tinygrad.helpers import DType, PtrDType, dtypes, ImageDType, DEBUG, getenv
-from tinygrad.codegen.opt.kernel import  UOp, Ops
+from tinygrad.opt.kernel import  UOp, Ops
 from triton.compiler import compile as triton_compile
 import linecache
 import math
@@ -10,7 +10,7 @@ from tinygrad.uop.ops import Ops
 import json
 from collections import OrderedDict

-EXPORT_SUPPORTED_DEVICE = ["WEBGPU", "CPU", "CUDA", "CL"]
+EXPORT_SUPPORTED_DEVICE = ["WEBGPU", "CPU", "CUDA", "GPU"]

 def compile_net(run:TinyJit, special_names:Dict[int,str]) -> Tuple[Dict[str,str],List[Tuple[str,List[str],List[int]]],Dict[str,Tuple[int,DType,int]],Dict[str,Tensor]]:
  functions, bufs, bufs_to_save, statements, bufnum = {}, {}, {}, [], 0
@@ -67,12 +67,11 @@ def export_model_clang(functions:Dict[str,str], statements:Dict[str,Tuple[str,in
  forward_args = ",".join(f"{dtype}{'*' if name not in symbolic_vars.values() else ''} {name}" for name,dtype,_ in (outputs+inputs if wasm else inputs+outputs))

  if not wasm:
-    thread_id = 0 # NOTE: export does not support threading, thread_id is always 0
    for name,cl in bufs_to_save.items():
      weight = ''.join(["\\x%02X"%x for x in bytes(to_mv(cl._buf.va_addr, cl._buf.size))])
      cprog.append(f"unsigned char {name}_data[] = \"{weight}\";")
    cprog += [f"{dtype_map[dtype]} {name}[{len}];" if name not in bufs_to_save else f"{dtype_map[dtype]} *{name} = ({dtype_map[dtype]} *){name}_data;" for name,(len,dtype,_key) in bufs.items() if name not in input_names+output_names]
-    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)}, {thread_id});" for (name, args, _global_size, _local_size) in statements] + ["}"]
+    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)});" for (name, args, _global_size, _local_size) in statements] + ["}"]
    return '\n'.join(headers + cprog)
  else:
    if bufs_to_save:
@@ -240,9 +239,7 @@ export default {model_name};

 def export_model(model, target:str, *inputs, model_name: Optional[str] = "model", stream_weights=False):
  assert Device.DEFAULT in EXPORT_SUPPORTED_DEVICE, f"only {', '.join(EXPORT_SUPPORTED_DEVICE)} are supported"
-
-  # NOTE: CPU_COUNT=1, since export does not support threading
-  with Context(JIT=2, CPU_COUNT=1): run,special_names = jit_model(model, *inputs)
+  with Context(JIT=2): run,special_names = jit_model(model, *inputs)
  functions, statements, bufs, bufs_to_save = compile_net(run, special_names)
  state = get_state_dict(model)
  weight_names = {id(x.uop.base.realized): name for name, x in state.items()}
@@ -2,11 +2,11 @@ from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, graph_rewrite, AxisType, PatternMatcher, UPat
 from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
 from tinygrad.dtype import AddrSpace
+from tinygrad.schedule.kernelize import merge_views, view_left
 from tinygrad.helpers import getenv, colored, prod, unwrap
 from tinygrad.shape.shapetracker import ShapeTracker, View
 from tinygrad.shape.view import strides_for_shape
-from tinygrad.codegen.opt.kernel import axis_colors, Opt, OptOps
-from tinygrad.codegen.opt.swizzler import merge_views, view_left
+from tinygrad.opt.kernel import axis_colors

 def to_colored(full_shape, axis_types): return '_'.join([colored(str(s), axis_colors[at]) for s,at in zip(full_shape, axis_types)])

@@ -44,28 +44,13 @@ pm = PatternMatcher([
  (UPat(Ops.VIEW, src=(UPat(Ops.REDUCE_AXIS, src=(UPat.var("src"),), name="r"),), name="view"), swizzle_reduceop),
 ])

-def rangeify_kernel3():
-  a = Tensor.empty(N,N)
-  b = Tensor.empty(N,N)
-  c = a@b
-  #c = c.reshape((32,2,16,4,32,2,16,4)).contiguous()
-  with Context(RANGEIFY=1):
-    sink = c.schedule()[-1].ast
-  #print(sink)
-
-  opts  = [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UPCAST, 0, 2)]
-  opts += [Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.LOCAL, 1, 16), Opt(OptOps.UPCAST, 1, 2)]
-  opts += [Opt(OptOps.UNROLL, 0, 8)]
-
-  return sink.replace(arg=KernelInfo(opts_to_apply=tuple(opts)))
-
 def top_spec_kernel3():
  a = Tensor.empty(N,N)
  b = Tensor.empty(N,N)
  c = a@b
  sink = c.schedule()[-1].ast
  L = 16
-  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(N//BM, 0), 2:UOp.range(N//BN, 1)})
+  sink = sink.reshape((N//L, L, N//L, L)) #.lift({0:UOp.range(dtypes.int, N//BM, 0), 2:UOp.range(dtypes.int, N//BN, 1)})
  sink = graph_rewrite(sink, view_left+pm)
  axis_types = (AxisType.GLOBAL, AxisType.LOCAL, AxisType.GLOBAL, AxisType.LOCAL, AxisType.REDUCE)
  return sink.replace(arg=KernelInfo(name="top_"+to_colored(sink.full_shape, axis_types), axis_types=axis_types))
@@ -186,7 +171,7 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):

  c_regs = UOp(Ops.DEFINE_REG, dtypes.float.ptr(TM * nbIterWaveM * TN * nbIterWaveN), arg=2)

-  i = UOp.range(c_regs.dtype.size, 16)
+  i = UOp.range(dtypes.int, c_regs.dtype.size, 16)
  init_store = c_regs[i].store(UOp.const(dtypes.float, 0.0), i)

  if kernel4:
@@ -197,53 +182,53 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
    kId = 0

    # load from globals into locals
-    i = UOp.range(nbReadsB, 0)
+    i = UOp.range(dtypes.int, nbReadsB, 0)
    index_x = BN * blockIdx_x + rBIdx
    index_y = rBIdy + i * strideReadB + kId
    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)

-    i = UOp.range(nbReadsA, 1)
+    i = UOp.range(dtypes.int, nbReadsA, 1)
    index_x = rAIdx + kId
    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)

    # iterate over the middle chunk
-    kId_range = UOp.range(N//BK-1, 2)
+    kId_range = UOp.range(dtypes.int, N//BK-1, 2)
    kId = kId_range*BK

    barrier = UOp.barrier(As_store, Bs_store)

    # load from globals into registers (next round)
-    i = UOp.range(nbReadsB, 3)
+    i = UOp.range(dtypes.int, nbReadsB, 3)
    index_x = BN * blockIdx_x + rBIdx
    index_y = rBIdy + i * strideReadB + kId + BK
    regB_store = regB[i].store(b[N * index_y + index_x].load(), i)

-    i = UOp.range(nbReadsA, 4)
+    i = UOp.range(dtypes.int, nbReadsA, 4)
    index_x = rAIdx + kId + BK
    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
    regA_store = regA[i].store(a[N * index_y + index_x].load(), i)

    def inner_loop(first_range, inp_dep=()):
      # inner unroll
-      k = UOp.range(BK, first_range+0)
+      k = UOp.range(dtypes.int, BK, first_range+0)

      # load from locals into registers
-      iterWave = UOp.range(nbIterWaveN, first_range+1)
-      i = UOp.range(TN, first_range+2)
+      iterWave = UOp.range(dtypes.int, nbIterWaveN, first_range+1)
+      i = UOp.range(dtypes.int, TN, first_range+2)
      index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
      B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(*inp_dep), iterWave, i)

-      iterWave = UOp.range(nbIterWaveM, first_range+3)
-      i = UOp.range(TM, first_range+4)
+      iterWave = UOp.range(dtypes.int, nbIterWaveM, first_range+3)
+      i = UOp.range(dtypes.int, TM, first_range+4)
      index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
      A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(*inp_dep), iterWave, i)

      # do the GEMM math
-      iterWaveM = UOp.range(nbIterWaveM, first_range+5)
-      yt = UOp.range(TM, first_range+6)
-      iterWaveN = UOp.range(nbIterWaveN, first_range+7)
-      xt = UOp.range(TN, first_range+8)
+      iterWaveM = UOp.range(dtypes.int, nbIterWaveM, first_range+5)
+      yt = UOp.range(dtypes.int, TM, first_range+6)
+      iterWaveN = UOp.range(dtypes.int, nbIterWaveN, first_range+7)
+      xt = UOp.range(dtypes.int, TN, first_range+8)
      x = iterWaveN * TN + xt
      y = iterWaveM * TM + yt
      c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
@@ -256,12 +241,12 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
    sink = inner_loop(5, (barrier, regB_store, regA_store)).barrier()

    # load from registers into locals
-    i = UOp.range(nbReadsB, 14)
+    i = UOp.range(dtypes.int, nbReadsB, 14)
    index_x = BN * blockIdx_x + rBIdx
    index_y = rBIdy + i * strideReadB + kId + BK
    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(regB[i].load(sink), i, kId_range)

-    i = UOp.range(nbReadsA, 15)
+    i = UOp.range(dtypes.int, nbReadsA, 15)
    index_x = rAIdx + kId + BK
    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(regA[i].load(sink), i, kId_range)
@@ -269,40 +254,40 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
    # final iteration without the copy
    sink = inner_loop(16, (UOp.barrier(Bs_store, As_store),))
  else:
-    kId_range = UOp.range(N//BK, 0)
+    kId_range = UOp.range(dtypes.int, N//BK, 0)
    kId = kId_range*BK

    # load from globals into locals
-    i = UOp.range(nbReadsB, 1)
+    i = UOp.range(dtypes.int, nbReadsB, 1)
    index_x = BN * blockIdx_x + rBIdx
    index_y = rBIdy + i * strideReadB + kId
    Bs_store = Bs[(index_y % BK) * BN + index_x % BN].store(b[N * index_y + index_x].load(), i)

-    i = UOp.range(nbReadsA, 2)
+    i = UOp.range(dtypes.int, nbReadsA, 2)
    index_x = rAIdx + kId
    index_y = BM * blockIdx_y + rAIdy + i * strideReadA
    As_store = As[(index_x % BK) * BM_As_stride + index_y % BM].store(a[N * index_y + index_x].load(), i)

    barrier = UOp.barrier(As_store, Bs_store)

-    k = UOp.range(BK, 3)
+    k = UOp.range(dtypes.int, BK, 3)

    # load from locals into registers
-    iterWave = UOp.range(nbIterWaveN, 4)
-    i = UOp.range(TN, 5)
+    iterWave = UOp.range(dtypes.int, nbIterWaveN, 4)
+    i = UOp.range(dtypes.int, TN, 5)
    index = waveIdx * WN + iterWave * SUBWN + TN * idxInWave + i
    B_row_store = B_row[iterWave*TN + i].store(Bs[k*BN + index].load(barrier), iterWave, i)

-    iterWave = UOp.range(nbIterWaveM, 6)
-    i = UOp.range(TM, 7)
+    iterWave = UOp.range(dtypes.int, nbIterWaveM, 6)
+    i = UOp.range(dtypes.int, TM, 7)
    index = waveIdy * WM + iterWave * SUBWM + TM * idyInWave + i
    A_col_store = A_col[iterWave*TM + i].store(As[k*BM_As_stride + index].load(barrier), iterWave, i)

    # do the GEMM math
-    iterWaveM = UOp.range(nbIterWaveM, 8)
-    yt = UOp.range(TM, 9)
-    iterWaveN = UOp.range(nbIterWaveN, 10)
-    xt = UOp.range(TN, 12)
+    iterWaveM = UOp.range(dtypes.int, nbIterWaveM, 8)
+    yt = UOp.range(dtypes.int, TM, 9)
+    iterWaveN = UOp.range(dtypes.int, nbIterWaveN, 10)
+    xt = UOp.range(dtypes.int, TN, 12)
    x = iterWaveN * TN + xt
    y = iterWaveM * TM + yt
    c_regs_idx = c_regs[y * TN * nbIterWaveN + x]
@@ -310,10 +295,10 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):
                            iterWaveM, iterWaveN, yt, xt, k, kId_range)

  # store c_regs into c
-  iterWaveM = UOp.range(nbIterWaveM, 1000)
-  yt = UOp.range(TM, 1001)
-  iterWaveN = UOp.range(nbIterWaveN, 1002)
-  xt = UOp.range(TN, 1003)
+  iterWaveM = UOp.range(dtypes.int, nbIterWaveM, 1000)
+  yt = UOp.range(dtypes.int, TM, 1001)
+  iterWaveN = UOp.range(dtypes.int, nbIterWaveN, 1002)
+  xt = UOp.range(dtypes.int, TN, 1003)
  xOut = blockIdx_x * BN + waveIdx * WN + iterWaveN * SUBWN + TN * idxInWave
  yOut = blockIdx_y * BM + waveIdy * WM + iterWaveM * SUBWM + TM * idyInWave
  indexC = N * (yOut + yt) + xOut + xt
@@ -324,15 +309,10 @@ def hand_spec_kernel3(kernel4=getenv("K4", 0), kernel5=getenv("K5", 0)):

 if __name__ == "__main__":
  HL = getenv("HL")
-  if HL == 3: hprg = rangeify_kernel3()
-  elif HL == 2: hprg = top_spec_kernel3()
+  if HL == 2: hprg = top_spec_kernel3()
  elif HL == 1: hprg = hl_spec_kernel3()
  else: hprg = hand_spec_kernel3()
-  if HL == 3:
-    with Context(RANGEIFY=1, BLOCK_REORDER=0):
-      prg = get_program(hprg, Device.default.renderer)
-  else:
-    prg = get_program(hprg, Device.default.renderer)
+  prg = get_program(hprg, Device.default.renderer)
  print(prg.src)
  if getenv("SRC"): exit(0)
  hrunner = CompiledRunner(prg)
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 import numpy as np
-from tinygrad.runtime.ops_cl import CLProgram, CLCompiler
+from tinygrad.runtime.ops_gpu import CLProgram, CLCompiler
 from tinygrad import Device, dtypes
 from tinygrad.device import Buffer
 from hexdump import hexdump
@@ -11,7 +11,7 @@ from hexdump import hexdump
 # https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_subgroup_split_matrix_multiply_accumulate.html
 # https://hc34.hotchips.org/assets/program/conference/day1/GPU%20HPC/Intel_s%20Ponte%20Vecchio%20GPU%20-%20Architecture%20Systems%20and%20Software%20FINAL.pdf

-device = Device["CL"]
+device = Device["GPU"]

 # NOTE: only the subgroup type 8 ones work
 prog = CLProgram(device, "test", CLCompiler(device, "test").compile(f"""
@@ -26,9 +26,9 @@ __kernel void test(__global float* data0, const __global int* data1, const __glo
 """))
 #with open("/tmp/test.elf", "wb") as f: f.write(prog.lib)

-a = Buffer("CL", 8, dtypes.float32).allocate()
-b = Buffer("CL", 0x10, dtypes.float16).allocate()
-c = Buffer("CL", 8*0x10, dtypes.float16).allocate()
+a = Buffer("GPU", 8, dtypes.float32).allocate()
+b = Buffer("GPU", 0x10, dtypes.float16).allocate()
+c = Buffer("GPU", 8*0x10, dtypes.float16).allocate()

 row = np.array([1,2,3,4,5,6,7,8,1,2,3,4,5,6,7,8], np.float16)
 mat = np.random.random((8, 0x10)).astype(np.float16)
@@ -5,9 +5,9 @@ from typing import Optional, List, Tuple, cast, Dict, Final, DefaultDict, Self
 from tinygrad.engine.realize import get_program

 # for copied uops
-from tinygrad.codegen.opt.kernel import Kernel, KernelOptError
+from tinygrad.opt.kernel import Kernel, KernelOptError
 from tinygrad.uop.ops import UOp, Ops, BinaryOps, UnaryOps, TernaryOps, KernelInfo
-from tinygrad.codegen.opt.search import Opt, OptOps
+from tinygrad.opt.search import Opt, OptOps
 from tinygrad import Device, dtypes, Tensor
 from tinygrad.dtype import PtrDType, DType, DTYPES_DICT
 from tinygrad.shape.shapetracker import ShapeTracker
@@ -56,7 +56,7 @@ def randoms():
 def ast_to_cuda_prog(compiler, ast, opts):
  k = Kernel(ast)
  k.apply_opts(opts)
-  p = get_program(k.ast, k.opts, k.applied_opts)
+  p = get_program(k.get_optimized_ast(), k.opts)
  return CUDAProgram(device, p.function_name, compiler.compile(p.src))

 if __name__ == "__main__":
@@ -75,7 +75,7 @@ if __name__ == "__main__":

    if GEMM_VARIATION == "max" and (M%64)==0 and (N%128)==0 and (K%64)==0 and DTYPE_IN == dtypes.half and DTYPE_OUT == dtypes.float and DTYPE_ACC == dtypes.float:
      print("Using CUDA and triton-generated kernel")
-      # See nv_triton_gemm.annotated.ptx for PTX code which was generated from `PYTHONPATH=. DEBUG=6 CUDA=1 CUDA_PTX=1 python3 extra/gemm/triton_nv_matmul.py`
+      # See nv_triton_gemm.annotated.ptx for PTX code which was generated from `PYTHONPATH=. DEBUG=6 CUDA=1 PTX=1 python3 extra/gemm/triton_nv_matmul.py`
      # this kernel with M=N=K=4096 does 162TFLOPS, vs torch at 144TFLOPS and BEAM=8 tinygrad at 138TFLOPS.  theo max is 165TFLOPS.

      # WMMA element size is (M, N, K) = (16, 8, 16)
@@ -2,7 +2,7 @@ import numpy as np
 from tinygrad import dtypes, Tensor
 from tinygrad.helpers import getenv, get_single_element
 from tinygrad.dtype import _to_np_dtype
-from tinygrad.codegen.opt import OptOps
+from tinygrad.opt.kernel import OptOps
 from tinygrad.engine.realize import lower_schedule

 dtype_in = dtypes.half if getenv("HALF") else dtypes.bfloat16 if getenv("BFLOAT16") else dtypes.float
@@ -1,6 +1,6 @@
 from tinygrad import Tensor, dtypes, Device
 from tinygrad.helpers import getenv, DEBUG
-from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps
+from tinygrad.opt.kernel import Kernel, Opt, OptOps
 from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
 from dataclasses import replace

@@ -29,7 +29,7 @@ if __name__ == "__main__":
      Opt(op=OptOps.LOCAL, axis=0, amt=2),
    ]
  k.apply_opts(opts)
-  prg = get_program(k.ast, k.opts, k.applied_opts)
+  prg = get_program(k.get_optimized_ast(), k.opts)
  new_src = prg.src
  # can mod source here
  prg = replace(prg, src=new_src)
@@ -43,7 +43,7 @@ def matmul_kernel(c_ptr, a_ptr, b_ptr, BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N:
  c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
  tl.store(c_ptrs, c)

-# CUDA=1 CUDA_PTX=1 python3 extra/gemm/triton_nv_matmul.py
+# CUDA=1 PTX=1 python3 extra/gemm/triton_nv_matmul.py
 if __name__ == "__main__":
  BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K = 64, 128, 64
  M, N, K = 4096, 4096, 4096
@@ -37,7 +37,7 @@ B = Tensor.rand(K, N, device="CPU")
 C = (A.reshape(M, 1, K) * B.permute(1,0).reshape(1, N, K)).sum(axis=2)

 sched = C.schedule()
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.device import CompilerOptions
 lin = Kernel(sched[-1].ast, CompilerOptions(has_local=False, supports_float4=False))
 lin.to_program()
@@ -8,6 +8,7 @@ bert_train_params = {
  "BS": 96,
  "EVAL_BS": 96,
  "FUSE_ARANGE": 1,
+  "FUSE_ARANGE_UINT": 0,
  "BASEDIR": "/raid/datasets/wiki",
 }

@@ -1,61 +0,0 @@
-# HuggingFace ONNX
-
-Tool for discovering, downloading, and validating ONNX models from HuggingFace.
-
-## Extra Dependencies
-
-```bash
-pip install huggingface_hub pyyaml requests onnx onnxruntime numpy
-```
-
-## Huggingface Manager (discovering and downloading)
-
-The `huggingface_manager.py` script discovers top ONNX models from HuggingFace, collects metadata, and optionally downloads them.
-
-```bash
-# Download top 50 models sorted by downloads
-python huggingface_manager.py --limit 50 --download
-
-# Just collect metadata (no download)
-python huggingface_manager.py --limit 100
-
-# Sort by likes instead of downloads
-python huggingface_manager.py --limit 20 --sort likes --download
-
-# Custom output file
-python huggingface_manager.py --limit 10 --output my_models.yaml
-```
-
-### Output Format
-
-The tool generates a YAML file with the following structure:
-
-```yaml
-repositories:
-  "model-name":
-    url: "https://huggingface.co/model-name"
-    download_path: "/path/to/models/..."  # when --download used
-    files:
-      - file: "model.onnx"
-        size: "90.91MB"
-total_size: "2.45GB"
-created_at: "2024-01-15T10:30:00Z"
-```
-
-## Run Models (validation)
-
-The `run_models.py` script validates ONNX models against ONNX Runtime for correctness.
-
-```bash
-# Validate models from a YAML configuration file
-python run_models.py --validate huggingface_repos.yaml
-
-# Debug specific repository (downloads and validates all ONNX models)
-python run_models.py --debug sentence-transformers/all-MiniLM-L6-v2
-
-# Debug specific model file
-python run_models.py --debug sentence-transformers/all-MiniLM-L6-v2/onnx/model.onnx
-
-# Debug with model truncation for debugging and validating intermediate results
-DEBUGONNX=1 python run_models.py --debug sentence-transformers/all-MiniLM-L6-v2/onnx/model.onnx --truncate 10
-```
@@ -0,0 +1,85 @@
+import yaml, time, requests, argparse
+from pathlib import Path
+from huggingface_hub import list_models, HfApi
+from tinygrad.helpers import tqdm
+
+HUGGINGFACE_URL = "https://huggingface.co"
+SKIPPED_FILES = [
+  "fp16", "int8", "uint8", "quantized",      # numerical accuracy issues
+  "avx2", "arm64", "avx512", "avx512_vnni",  # numerical accuracy issues
+  "q4", "q4f16", "bnb4",                     # unimplemented quantization
+  "model_O4",                                # requires non cpu ort runner and MemcpyFromHost op
+  "merged",                                  # TODO implement attribute with graph type and Loop op
+]
+SKIPPED_REPO_PATHS = [
+  # Invalid model-index
+  "AdamCodd/vit-base-nsfw-detector",
+  # TODO: implement attribute with graph type and Loop op
+  "minishlab/potion-base-8M", "minishlab/M2V_base_output", "minishlab/potion-retrieval-32M",
+  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, GroupQueryAttention
+  "HuggingFaceTB/SmolLM2-360M-Instruct",
+  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, RotaryEmbedding, MultiHeadAttention
+  "HuggingFaceTB/SmolLM2-1.7B-Instruct",
+  # TODO: implmement RandomNormalLike
+  "stabilityai/stable-diffusion-xl-base-1.0", "stabilityai/sdxl-turbo", 'SimianLuo/LCM_Dreamshaper_v7',
+  # TODO: implement NonZero
+  "mangoapps/fb_zeroshot_mnli_onnx",
+  # TODO huge Concat in here with 1024 (1, 3, 32, 32) Tensors, and maybe a MOD bug with const folding
+  "briaai/RMBG-2.0",
+]
+
+def get_top_repos(n: int, sort: str) -> list[str]: # list["FacebookAI/xlm-roberta-large", ...]
+  print(f"** Getting top {n} models sorted by {sort} **")
+  repos = []
+  i = 0
+  for model in list_models(filter="onnx", sort=sort):
+    if model.id in SKIPPED_REPO_PATHS: continue
+    print(f"{i+1}/{n}: {model.id} ({getattr(model, sort)})")
+    repos.append(model.id)
+    i += 1
+    if i == n: break
+  return repos
+
+def get_metadata(repos:list[str]) -> dict:
+  api = HfApi()
+  repos_metadata = {"repositories": {}}
+  total_size = 0
+
+  # TODO: speed head requests up with async?
+  for repo in tqdm(repos, desc="Getting metadata"):
+    files_metadata = []
+    model_info = api.model_info(repo)
+
+    for file in model_info.siblings:
+      filename = file.rfilename
+      if not (filename.endswith('.onnx') or filename.endswith('.onnx_data')): continue
+      if any(skip_str in filename for skip_str in SKIPPED_FILES): continue
+      head = requests.head(f"{HUGGINGFACE_URL}/{repo}/resolve/main/{filename}", allow_redirects=True)
+      file_size = file.size or int(head.headers.get('Content-Length', 0))
+      files_metadata.append({"file": filename, "size": f"{file_size/1e6:.2f}MB"})
+      total_size += file_size
+
+    repos_metadata["repositories"][repo] = {
+      "url": f"{HUGGINGFACE_URL}/{repo}",
+      "download_path": None,
+      "files": files_metadata,
+    }
+  repos_metadata['total_size'] = f"{total_size/1e9:.2f}GB"
+  repos_metadata['created_at'] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+  return repos_metadata
+
+if __name__ == "__main__":
+  sort = "downloads" # recent 30 days downloads
+  huggingface_onnx_dir = Path(__file__).parent
+
+  parser = argparse.ArgumentParser(description="Produces a YAML file with metadata of top huggingface onnx models")
+  parser.add_argument("--limit", type=int, required=True, help="Number of top repositories to process (e.g., 100)")
+  parser.add_argument("--output", type=str, default="huggingface_repos.yaml", help="Output YAML file name to save the report")
+  args = parser.parse_args()
+
+  top_repos = get_top_repos(args.limit, sort)
+  metadata = get_metadata(top_repos)
+  yaml_path = huggingface_onnx_dir / args.output
+  with open(yaml_path, 'w') as f:
+    yaml.dump(metadata, f, sort_keys=False)
+    print(f"YAML saved to: {str(yaml_path)}")
@@ -0,0 +1,29 @@
+import yaml, argparse
+from pathlib import Path
+from huggingface_hub import snapshot_download
+
+def download_models(yaml_file: str, download_dir: str) -> None:
+  with open(yaml_file, 'r') as f: metadata = yaml.safe_load(f)
+  n = len(metadata["repositories"])
+
+  for i, (model_id, model_data) in enumerate(metadata["repositories"].items()):
+    print(f"Downloading {i+1}/{n}: {model_id}...")
+    allow_patterns = [file_info["file"] for file_info in model_data["files"]]
+    root_path = Path(snapshot_download(repo_id=model_id, allow_patterns=allow_patterns, cache_dir=download_dir))
+    # download configs too (the sizes are small)
+    snapshot_download(repo_id=model_id, allow_patterns=["*config.json"], cache_dir=download_dir)
+    print(f"Downloaded model files to: {root_path}")
+    model_data["download_path"] = str(root_path)
+
+  # Save the updated metadata back to the YAML file
+  with open(yaml_file, 'w') as f: yaml.dump(metadata, f, sort_keys=False)
+  print("Download completed according to YAML file.")
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(description="Download models from Huggingface Hub based on a YAML configuration file.")
+  parser.add_argument("input", type=str, help="Path to the input YAML configuration file containing model information.")
+  args = parser.parse_args()
+
+  models_folder = Path(__file__).parent / "models"
+  models_folder.mkdir(parents=True, exist_ok=True)
+  download_models(args.input, str(models_folder))
@@ -1,230 +0,0 @@
-import yaml
-import time
-import requests
-import argparse
-from pathlib import Path
-from huggingface_hub import list_models, HfApi, snapshot_download
-from tinygrad.helpers import _ensure_downloads_dir
-DOWNLOADS_DIR = _ensure_downloads_dir() / "models"
-from tinygrad.helpers import tqdm
-
-def snapshot_download_with_retry(*, repo_id: str, allow_patterns: list[str]|tuple[str, ...]|None=None, cache_dir: str|Path|None=None,
-                                 tries: int=2, **kwargs) -> Path:
-  for attempt in range(tries):
-    try:
-      return Path(snapshot_download(
-        repo_id=repo_id,
-        allow_patterns=allow_patterns,
-        cache_dir=str(cache_dir) if cache_dir is not None else None,
-        **kwargs
-      ))
-    except Exception as e:
-      if attempt == tries-1: raise
-      time.sleep(1)
-
-# Constants for filtering models
-HUGGINGFACE_URL = "https://huggingface.co"
-SKIPPED_FILES = [
-  "fp16", "int8", "uint8", "quantized",      # numerical accuracy issues
-  "avx2", "arm64", "avx512", "avx512_vnni",  # numerical accuracy issues
-  "q4", "q4f16", "bnb4",                     # unimplemented quantization
-  "model_O4",                                # requires non cpu ort runner and MemcpyFromHost op
-  "merged",                                  # TODO implement attribute with graph type and Loop op
-]
-
-SKIPPED_REPO_PATHS = [
-  # Invalid model-index
-  "AdamCodd/vit-base-nsfw-detector",
-  # TODO: implement attribute with graph type and Loop op
-  "minishlab/potion-base-8M", "minishlab/M2V_base_output", "minishlab/potion-retrieval-32M",
-  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, GroupQueryAttention
-  "HuggingFaceTB/SmolLM2-360M-Instruct",
-  # TODO: implement SimplifiedLayerNormalization, SkipSimplifiedLayerNormalization, RotaryEmbedding, MultiHeadAttention
-  "HuggingFaceTB/SmolLM2-1.7B-Instruct",
-  # TODO: implement RandomNormalLike
-  "stabilityai/stable-diffusion-xl-base-1.0", "stabilityai/sdxl-turbo", 'SimianLuo/LCM_Dreamshaper_v7',
-  # TODO: implement NonZero
-  "mangoapps/fb_zeroshot_mnli_onnx",
-  # TODO huge Concat in here with 1024 (1, 3, 32, 32) Tensors, and maybe a MOD bug with const folding
-  "briaai/RMBG-2.0",
-]
-
-
-class HuggingFaceONNXManager:
-  def __init__(self):
-    self.base_dir = Path(__file__).parent
-    self.models_dir = DOWNLOADS_DIR
-    self.api = HfApi()
-
-  def discover_models(self, limit: int, sort: str = "downloads") -> list[str]:
-    print(f"Discovering top {limit} ONNX models sorted by {sort}...")
-    repos = []
-    i = 0
-
-    for model in list_models(filter="onnx", sort=sort):
-      if model.id in SKIPPED_REPO_PATHS:
-        continue
-
-      print(f"  {i+1}/{limit}: {model.id} ({getattr(model, sort)})")
-      repos.append(model.id)
-      i += 1
-      if i == limit:
-        break
-
-    print(f"Found {len(repos)} suitable ONNX models")
-    return repos
-
-  def collect_metadata(self, repos: list[str]) -> dict:
-    print(f"Collecting metadata for {len(repos)} repositories...")
-    metadata = {"repositories": {}}
-    total_size = 0
-
-    for repo in tqdm(repos, desc="Collecting metadata"):
-      try:
-        files_metadata = []
-        model_info = self.api.model_info(repo)
-
-        for file in model_info.siblings:
-          filename = file.rfilename
-          if not (filename.endswith('.onnx') or filename.endswith('.onnx_data')):
-            continue
-          if any(skip_str in filename for skip_str in SKIPPED_FILES):
-            continue
-
-          # Get file size from API or HEAD request
-          try:
-            head = requests.head(
-              f"{HUGGINGFACE_URL}/{repo}/resolve/main/{filename}",
-              allow_redirects=True,
-              timeout=10
-            )
-            file_size = file.size or int(head.headers.get('Content-Length', 0))
-          except requests.RequestException:
-            file_size = file.size or 0
-
-          files_metadata.append({
-            "file": filename,
-            "size": f"{file_size/1e6:.2f}MB"
-          })
-          total_size += file_size
-
-        if files_metadata:  # Only add repos with valid ONNX files
-          metadata["repositories"][repo] = {
-            "url": f"{HUGGINGFACE_URL}/{repo}",
-            "download_path": None,
-            "files": files_metadata,
-          }
-
-      except Exception as e:
-        print(f"WARNING: Failed to collect metadata for {repo}: {e}")
-        continue
-
-    metadata['total_size'] = f"{total_size/1e9:.2f}GB"
-    metadata['created_at'] = time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
-
-    print(f"Collected metadata for {len(metadata['repositories'])} repositories")
-    print(f"Total estimated download size: {metadata['total_size']}")
-
-    return metadata
-
-  def download_models(self, metadata: dict) -> dict:
-    self.models_dir.mkdir(parents=True, exist_ok=True)
-
-    repos = metadata["repositories"]
-    n = len(repos)
-
-    print(f"Downloading {n} repositories to {self.models_dir}...")
-
-    for i, (model_id, model_data) in enumerate(repos.items()):
-      print(f"  Downloading {i+1}/{n}: {model_id}...")
-
-      try:
-        # Download ONNX model files
-        allow_patterns = [file_info["file"] for file_info in model_data["files"]]
-        root_path = snapshot_download_with_retry(
-          repo_id=model_id,
-          allow_patterns=allow_patterns,
-          cache_dir=str(self.models_dir)
-        )
-
-        # Download config files (usually small)
-        snapshot_download_with_retry(
-          repo_id=model_id,
-          allow_patterns=["*config.json"],
-          cache_dir=str(self.models_dir)
-        )
-
-        model_data["download_path"] = str(root_path)
-        print(f"    Downloaded to: {root_path}")
-
-      except Exception as e:
-        print(f"    ERROR: Failed to download {model_id}: {e}")
-        model_data["download_path"] = None
-        continue
-
-    successful_downloads = sum(1 for repo in repos.values() if repo["download_path"] is not None)
-    print(f"Successfully downloaded {successful_downloads}/{n} repositories")
-    print(f"All models saved to: {self.models_dir}")
-
-    return metadata
-
-  def save_metadata(self, metadata: dict, output_file: str):
-    yaml_path = self.base_dir / output_file
-    with open(yaml_path, 'w') as f:
-      yaml.dump(metadata, f, sort_keys=False)
-    print(f"Metadata saved to: {yaml_path}")
-
-  def discover_and_download(self, limit: int, output_file: str = "huggingface_repos.yaml",
-                          sort: str = "downloads", download: bool = True):
-    print(f"Starting HuggingFace ONNX workflow...")
-    print(f"   Limit: {limit} models")
-    print(f"   Sort by: {sort}")
-    print(f"   Download: {'Yes' if download else 'No'}")
-    print(f"   Output: {output_file}")
-    print("-" * 50)
-
-    repos = self.discover_models(limit, sort)
-
-    metadata = self.collect_metadata(repos)
-
-    if download:
-      metadata = self.download_models(metadata)
-
-    self.save_metadata(metadata, output_file)
-
-    print("-" * 50)
-    print("Workflow completed successfully!")
-    if download:
-      successful = sum(1 for repo in metadata["repositories"].values()
-                     if repo["download_path"] is not None)
-      print(f"{successful}/{len(metadata['repositories'])} models downloaded")
-
-    return metadata
-
-
-if __name__ == "__main__":
-  parser = argparse.ArgumentParser(
-    description="HuggingFace ONNX Model Manager - Discover, collect metadata, and download ONNX models",
-  )
-
-  parser.add_argument("--limit", type=int, help="Number of top repositories to process")
-  parser.add_argument("--output", type=str, default="huggingface_repos.yaml",
-                     help="Output YAML file name (default: huggingface_repos.yaml)")
-  parser.add_argument("--sort", type=str, default="downloads",
-                     choices=["downloads", "likes", "created", "modified"],
-                     help="Sort criteria for model discovery (default: downloads)")
-
-  parser.add_argument("--download", action="store_true", default=False,
-                     help="Download models after collecting metadata")
-
-  args = parser.parse_args()
-
-  if not args.limit: parser.error("--limit is required")
-
-  manager = HuggingFaceONNXManager()
-  manager.discover_and_download(
-    limit=args.limit,
-    output_file=args.output,
-    sort=args.sort,
-    download=args.download
-  )
@@ -1,11 +1,10 @@
-import onnx, yaml, tempfile, time, argparse, json
+import onnx, yaml, tempfile, time, collections, pprint, argparse, json
 from pathlib import Path
-from typing import Any
 from tinygrad.frontend.onnx import OnnxRunner
+from extra.onnx import get_onnx_ops
 from extra.onnx_helpers import validate, get_example_inputs
-from extra.huggingface_onnx.huggingface_manager import DOWNLOADS_DIR, snapshot_download_with_retry

-def get_config(root_path: Path) -> dict[str, Any]:
+def get_config(root_path: Path):
  ret = {}
  for path in root_path.rglob("*config.json"):
    config = json.load(path.open())
@@ -13,19 +12,19 @@ def get_config(root_path: Path) -> dict[str, Any]:
      ret.update(config)
  return ret

-def get_tolerances(file_name: str) -> tuple[float, float]:
+def run_huggingface_validate(onnx_model_path, config, rtol, atol):
+  onnx_runner = OnnxRunner(onnx_model_path)
+  inputs = get_example_inputs(onnx_runner.graph_inputs, config)
+  validate(onnx_model_path, inputs, rtol=rtol, atol=atol)
+
+def get_tolerances(file_name): # -> rtol, atol
  # TODO very high rtol atol
  if "fp16" in file_name: return 9e-2, 9e-2
  if any(q in file_name for q in ["int8", "uint8", "quantized"]): return 4, 4
  return 4e-3, 3e-2

-def run_huggingface_validate(onnx_model_path: str | Path, config: dict[str, Any], rtol: float, atol: float):
-  onnx_runner = OnnxRunner(onnx_model_path)
-  inputs = get_example_inputs(onnx_runner.graph_inputs, config)
-  validate(onnx_model_path, inputs, rtol=rtol, atol=atol)
-
 def validate_repos(models:dict[str, tuple[Path, Path]]):
-  print(f"** Validating {len(models)} models **")
+  print(f"** Validating {len(model_paths)} models **")
  for model_id, (root_path, relative_path) in models.items():
    print(f"validating model {model_id}")
    model_path = root_path / relative_path
@@ -37,6 +36,25 @@ def validate_repos(models:dict[str, tuple[Path, Path]]):
    et = time.time() - st
    print(f"passed, took {et:.2f}s")

+def retrieve_op_stats(models:dict[str, tuple[Path, Path]]) -> dict:
+  ret = {}
+  op_counter = collections.Counter()
+  unsupported_ops = collections.defaultdict(set)
+  supported_ops = get_onnx_ops()
+  print(f"** Retrieving stats from {len(model_paths)} models **")
+  for model_id, (root_path, relative_path) in models.items():
+    print(f"examining {model_id}")
+    model_path = root_path / relative_path
+    onnx_runner = OnnxRunner(model_path)
+    for node in onnx_runner.graph_nodes:
+      op_counter[node.op] += 1
+      if node.op not in supported_ops:
+        unsupported_ops[node.op].add(model_id)
+    del onnx_runner
+  ret["unsupported_ops"] = {k:list(v) for k, v in unsupported_ops.items()}
+  ret["op_counter"] = op_counter.most_common()
+  return ret
+
 def debug_run(model_path, truncate, config, rtol, atol):
  if truncate != -1:
    model = onnx.load(model_path)
@@ -53,9 +71,12 @@ def debug_run(model_path, truncate, config, rtol, atol):
    run_huggingface_validate(model_path, config, rtol, atol)

 if __name__ == "__main__":
-  parser = argparse.ArgumentParser(description="Huggingface ONNX Model Validator")
-  parser.add_argument("--validate", type=str, default="",
-                      help="Validate correctness of models from the specified YAML configuration file")
+  parser = argparse.ArgumentParser(description="Huggingface ONNX Model Validator and Ops Checker")
+  parser.add_argument("input", type=str, help="Path to the input YAML configuration file containing model information.")
+  parser.add_argument("--check_ops", action="store_true", default=False,
+                      help="Check support for ONNX operations in models from the YAML file")
+  parser.add_argument("--validate", action="store_true", default=False,
+                      help="Validate correctness of models from the YAML file")
  parser.add_argument("--debug", type=str, default="",
                      help="""Validates without explicitly needing a YAML or models pre-installed.
                      provide repo id (e.g. "minishlab/potion-base-8M") to validate all onnx models inside the repo
@@ -64,13 +85,13 @@ if __name__ == "__main__":
  parser.add_argument("--truncate", type=int, default=-1, help="Truncate the ONNX model so intermediate results can be validated")
  args = parser.parse_args()

-  if not (args.validate or args.debug):
-    parser.error("Please provide either --validate <yaml_file> or --debug <repo_id>.")
+  if not (args.check_ops or args.validate or args.debug):
+    parser.error("Please provide either --validate, --check_ops, or --debug.")
  if args.truncate != -1 and not args.debug:
    parser.error("--truncate and --debug should be used together for debugging")

-  if args.validate:
-    with open(args.validate, 'r') as f:
+  if args.check_ops or args.validate:
+    with open(args.input, 'r') as f:
      data = yaml.safe_load(f)
      assert all(repo["download_path"] is not None for repo in data["repositories"].values()), "please run `download_models.py` for this yaml"
      model_paths = {
@@ -80,16 +101,22 @@ if __name__ == "__main__":
        if model["file"].endswith(".onnx")
      }

-    validate_repos(model_paths)
+    if args.check_ops:
+      pprint.pprint(retrieve_op_stats(model_paths))
+
+    if args.validate:
+      validate_repos(model_paths)

  if args.debug:
+    from huggingface_hub import snapshot_download
+    download_dir = Path(__file__).parent / "models"
    path:list[str] = args.debug.split("/")
    if len(path) == 2:
      # repo id
      # validates all onnx models inside repo
      repo_id = "/".join(path)
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = Path(snapshot_download(repo_id=repo_id, allow_patterns=["*.onnx", "*.onnx_data"], cache_dir=download_dir))
+      snapshot_download(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=download_dir)
      config = get_config(root_path)
      for onnx_model in root_path.rglob("*.onnx"):
        rtol, atol = get_tolerances(onnx_model.name)
@@ -101,8 +128,8 @@ if __name__ == "__main__":
      onnx_model = path[-1]
      assert path[-1].endswith(".onnx")
      repo_id, relative_path = "/".join(path[:2]), "/".join(path[2:])
-      root_path = snapshot_download_with_retry(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=DOWNLOADS_DIR)
-      snapshot_download_with_retry(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=DOWNLOADS_DIR)
+      root_path = Path(snapshot_download(repo_id=repo_id, allow_patterns=[relative_path], cache_dir=download_dir))
+      snapshot_download(repo_id=repo_id, allow_patterns=["*config.json"], cache_dir=download_dir)
      config = get_config(root_path)
      rtol, atol = get_tolerances(onnx_model)
      print(f"validating {relative_path} with truncate={args.truncate}, {rtol=}, {atol=}")
@@ -4,9 +4,9 @@ import numpy as np
 np.set_printoptions(suppress=True)
 import math, functools, time, random, statistics
 from tinygrad.helpers import DEBUG, getenv, CACHELEVEL, diskcache_get, diskcache_put, colored, Profiling
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.device import Buffer, Device, CompileError
-from tinygrad.codegen.opt.search import _ensure_buffer_alloc, get_kernel_actions, _time_program
+from tinygrad.opt.search import _ensure_buffer_alloc, get_kernel_actions, _time_program
 from tinygrad.engine.realize import get_program

 class MCTSNode:
@@ -88,7 +88,7 @@ def mcts_search(lin:Kernel, rawbufs:List[Buffer], amt:int) -> Kernel:
    return ret

  rawbufs = _ensure_buffer_alloc(rawbufs)
-  var_vals = {k.expr:(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
+  var_vals = {k:(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
  dev = Device[lin.opts.device]
  root = MCTSNode(lin)

@@ -9,9 +9,6 @@ from PIL import Image
 import numpy as np
 import re, gzip

-# Allow for monkeypatching for mlperf.
-gelu = Tensor.gelu
-
@lru_cache()
 def default_bpe():
  # Clip tokenizer, taken from https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py (MIT license)
@@ -56,8 +53,8 @@ class Tokenizer:
    cs = [chr(n) for n in cs]
    return dict(zip(bs, cs))
  class ClipTokenizer:
-    def __init__(self, version=None):
-      self.byte_encoder, self.version = Tokenizer.bytes_to_unicode(), version
+    def __init__(self):
+      self.byte_encoder = Tokenizer.bytes_to_unicode()
      merges = gzip.open(default_bpe()).read().decode("utf-8").split('\n')
      merges = merges[1:49152-256-2+1]
      merges = [tuple(merge.split()) for merge in merges]
@@ -65,17 +62,11 @@ class Tokenizer:
      vocab = vocab + [v+'</w>' for v in vocab]
      for merge in merges:
        vocab.append(''.join(merge))
-      if self.version == "sd_mlperf_v5_0":
-        import regex
-        vocab.extend(['<start_of_text>', '<end_of_text>'])
-        self.cache = {'<start_of_text>': '<start_of_text>', '<end_of_text>': '<end_of_text>'}
-        self.pat = regex.compile(r"""<start_of_text>|<end_of_text>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", regex.IGNORECASE)
-      else:
-        vocab.extend(['<|startoftext|>', '<|endoftext|>'])
-        self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
-        self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[^\s]+""", re.IGNORECASE)
+      vocab.extend(['<|startoftext|>', '<|endoftext|>'])
      self.encoder = dict(zip(vocab, range(len(vocab))))
      self.bpe_ranks = dict(zip(merges, range(len(merges))))
+      self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
+      self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[^\s]+""", re.IGNORECASE)

    def bpe(self, token):
      if token in self.cache:
@@ -119,17 +110,8 @@ class Tokenizer:

    def encode(self, text:str, pad_with_zeros:bool=False) -> List[int]:
      bpe_tokens: List[int] = []
-      if self.version == "sd_mlperf_v5_0":
-        import regex, ftfy, html
-        text = ftfy.fix_text(text)
-        text = html.unescape(html.unescape(text)).strip()
-        text = Tokenizer.whitespace_clean(text).lower()
-        re_module = regex
-      else:
-        text = Tokenizer.whitespace_clean(text.strip()).lower()
-        re_module = re
-
-      for token in re_module.findall(self.pat, text):
+      text = Tokenizer.whitespace_clean(text.strip()).lower()
+      for token in re.findall(self.pat, text):
        token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
        bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
      # Truncation, keeping two slots for start and end tokens.
@@ -270,8 +252,10 @@ class Open:
      q,k,v = [y.reshape(T, B*self.n_heads, self.d_head).transpose(0, 1).reshape(B, self.n_heads, T, self.d_head) for y in proj.chunk(3)]

      attn_output = Tensor.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
-      attn_output = attn_output.permute(2, 0, 1, 3).reshape(T, B, C)
+      attn_output = attn_output.permute(2, 0, 1, 3).reshape(T*B, C)
+
      attn_output = self.out_proj(attn_output)
+      attn_output = attn_output.reshape(T, B, C)

      return attn_output

@@ -279,10 +263,9 @@ class Open:
    def __init__(self, dims, hidden_dims):
      self.c_fc   = Linear(dims, hidden_dims)
      self.c_proj = Linear(hidden_dims, dims)
-      self.gelu = gelu

    def __call__(self, x:Tensor) -> Tensor:
-      return x.sequential([self.c_fc, self.gelu, self.c_proj])
+      return x.sequential([self.c_fc, Tensor.gelu, self.c_proj])

  # https://github.com/mlfoundations/open_clip/blob/58e4e39aaabc6040839b0d2a7e8bf20979e4558a/src/open_clip/transformer.py#L210
  class ResidualAttentionBlock:
@@ -367,15 +350,15 @@ class Open:
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/encoders/modules.py#L396
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/encoders/modules.py#L498
 class FrozenOpenClipEmbedder(Embedder):
-  def __init__(self, dims:int, n_heads:int, layers:int, return_pooled:bool, ln_penultimate:bool=False, clip_tokenizer_version=None):
-    self.tokenizer = Tokenizer.ClipTokenizer(version=clip_tokenizer_version)
+  def __init__(self, dims:int, n_heads:int, layers:int, return_pooled:bool, ln_penultimate:bool=False):
+    self.tokenizer = Tokenizer.ClipTokenizer()
    self.model = Open.ClipTextTransformer(dims, n_heads, layers)
    self.return_pooled = return_pooled
    self.input_key = "txt"
    self.ln_penultimate = ln_penultimate

  def tokenize(self, text:str, device:Optional[str]=None) -> Tensor:
-    return Tensor(self.tokenizer.encode(text, pad_with_zeros=True), dtype=dtypes.int32, device=device).reshape(1,-1)
+    return Tensor(self.tokenizer.encode(text, pad_with_zeros=True), dtype=dtypes.int64, device=device).reshape(1,-1)

  def text_transformer_forward(self, x:Tensor, attn_mask:Optional[Tensor]=None):
    for r in self.model.transformer.resblocks:
@@ -466,7 +449,7 @@ class OpenClipEncoder:
    x = x + self.positional_embedding
    x = self.transformer(x, attn_mask=self.attn_mask)
    x = self.ln_final(x)
-    x = x[Tensor.arange(x.shape[0], device=x.device), tokens.argmax(axis=-1)]
+    x = x[:, tokens.argmax(axis=-1)]
    x = x @ self.text_projection
    return x

@@ -270,10 +270,8 @@ class FidInceptionV3:
    self.Mixed_7b = inception.Mixed_7b
    self.Mixed_7c = inception.Mixed_7c

-  def load_from_pretrained(self, path=None):
-    if path is None:
-      path = fetch("https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth", "pt_inception-2015-12-05-6726825d.pth")
-    state_dict = torch_load(str(path))
+  def load_from_pretrained(self):
+    state_dict = torch_load(str(fetch("https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth", "pt_inception-2015-12-05-6726825d.pth")))
    for k,v in state_dict.items():
      if k.endswith(".num_batches_tracked"):
        state_dict[k] = v.reshape(1)
@@ -181,11 +181,13 @@ class Transformer:
  def forward(self, tokens:Tensor, start_pos:Union[Variable,int], temperature:float, top_k:int, top_p:float, alpha_f:float, alpha_p:float):
    _bsz, seqlen = tokens.shape
    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, start_pos:start_pos+seqlen, :, :, :]
+
+    self.freqs_cis = self.freqs_cis.cast(h.dtype).contiguous()
+    freqs_cis = self.freqs_cis[:, start_pos:start_pos+seqlen, :, :, :]

    mask = Tensor.full((1, 1, seqlen, start_pos+seqlen), float("-inf"), dtype=h.dtype, device=h.device).triu(start_pos+1) if seqlen > 1 else None
    for layer in self.layers: h = layer(h, start_pos, freqs_cis, mask)
-    logits = self.output(self.norm(h))
+    logits = self.output(self.norm(h)).float()
    if math.isnan(temperature): return logits

    return sample(logits[:, -1, :].flatten(), temperature, top_k, top_p, alpha_f, alpha_p)
@@ -249,5 +251,8 @@ def convert_from_gguf(weights:dict[str, Tensor], n_layers:int):
  return sd

 def fix_bf16(weights:dict[Any, Tensor]):
-  # TODO: without casting to float16, 70B llama OOM on tinybox.
-  return {k:v.cast(dtypes.float32).cast(dtypes.float16) if v.dtype == dtypes.bfloat16 else v for k,v in weights.items()}
+  if getenv("SUPPORT_BF16", 1):
+    # TODO: without casting to float16, 70B llama OOM on tinybox.
+    return {k:v.cast(dtypes.float32).cast(dtypes.float16) if v.dtype == dtypes.bfloat16 else v for k,v in weights.items()}
+  # TODO: check if device supports bf16
+  return {k:v.llvm_bf16_cast(dtypes.half).to(v.device) if v.dtype == dtypes.bfloat16 else v for k,v in weights.items()}
@@ -1,24 +1,21 @@
-from tinygrad import Tensor, dtypes, nn
+from tinygrad import Tensor, dtypes
+from tinygrad.nn import Linear, Conv2d, GroupNorm, LayerNorm
 from tinygrad.device import is_dtype_supported
-from typing import Optional, Union, List, Any, Tuple, Callable
+from typing import Optional, Union, List, Any, Tuple
 import math

-# allow for monkeypatching
-Linear, Conv2d, GroupNorm, LayerNorm = nn.Linear, nn.Conv2d, nn.GroupNorm, nn.LayerNorm
-attention, gelu, mixed_precision_dtype = Tensor.scaled_dot_product_attention, Tensor.gelu, dtypes.float16
-
 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/diffusionmodules/util.py#L207
 def timestep_embedding(timesteps:Tensor, dim:int, max_period=10000):
  half = dim // 2
  freqs = (-math.log(max_period) * Tensor.arange(half, device=timesteps.device) / half).exp()
  args = timesteps.unsqueeze(1) * freqs.unsqueeze(0)
  out = Tensor.cat(args.cos(), args.sin(), dim=-1)
-  return out.cast(mixed_precision_dtype) if is_dtype_supported(mixed_precision_dtype) else out
+  return out.cast(dtypes.float16) if is_dtype_supported(dtypes.float16) else out

 class ResBlock:
-  def __init__(self, channels:int, emb_channels:int, out_channels:int, num_groups:int=32):
+  def __init__(self, channels:int, emb_channels:int, out_channels:int):
    self.in_layers = [
-      GroupNorm(num_groups, channels),
+      GroupNorm(32, channels),
      Tensor.silu,
      Conv2d(channels, out_channels, 3, padding=1),
    ]
@@ -27,7 +24,7 @@ class ResBlock:
      Linear(emb_channels, out_channels),
    ]
    self.out_layers = [
-      GroupNorm(num_groups, out_channels),
+      GroupNorm(32, out_channels),
      Tensor.silu,
      lambda x: x,  # needed for weights loading code to work
      Conv2d(out_channels, out_channels, 3, padding=1),
@@ -48,37 +45,35 @@ class CrossAttention:
    self.to_v = Linear(ctx_dim,   n_heads*d_head, bias=False)
    self.num_heads = n_heads
    self.head_size = d_head
-    self.attn = attention
    self.to_out = [Linear(n_heads*d_head, query_dim)]

  def __call__(self, x:Tensor, ctx:Optional[Tensor]=None) -> Tensor:
    ctx = x if ctx is None else ctx
    q,k,v = self.to_q(x), self.to_k(ctx), self.to_v(ctx)
    q,k,v = [y.reshape(x.shape[0], -1, self.num_heads, self.head_size).transpose(1,2) for y in (q,k,v)]
-    attention = self.attn(q, k, v).transpose(1,2)
+    attention = Tensor.scaled_dot_product_attention(q, k, v).transpose(1,2)
    h_ = attention.reshape(x.shape[0], -1, self.num_heads * self.head_size)
    return h_.sequential(self.to_out)

 class GEGLU:
  def __init__(self, dim_in:int, dim_out:int):
    self.proj = Linear(dim_in, dim_out * 2)
-    self.gelu = gelu
    self.dim_out = dim_out

  def __call__(self, x:Tensor) -> Tensor:
    x, gate = self.proj(x).chunk(2, dim=-1)
-    return x * self.gelu(gate)
+    return x * gate.gelu()

 class FeedForward:
  def __init__(self, dim:int, mult:int=4):
-    self.net: tuple[GEGLU, Callable, nn.Linear] = (
+    self.net = [
      GEGLU(dim, dim*mult),
      lambda x: x,  # needed for weights loading code to work
      Linear(dim*mult, dim)
-    )
+    ]

  def __call__(self, x:Tensor) -> Tensor:
-    return x.sequential(list(self.net))
+    return x.sequential(self.net)

 class BasicTransformerBlock:
  def __init__(self, dim:int, ctx_dim:int, n_heads:int, d_head:int):
@@ -97,13 +92,12 @@ class BasicTransformerBlock:

 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/attention.py#L619
 class SpatialTransformer:
-  def __init__(self, channels:int, n_heads:int, d_head:int, ctx_dim:Union[int,List[int]], use_linear:bool, depth:int=1,
-               norm_eps:float=1e-5):
+  def __init__(self, channels:int, n_heads:int, d_head:int, ctx_dim:Union[int,List[int]], use_linear:bool, depth:int=1):
    if isinstance(ctx_dim, int):
      ctx_dim = [ctx_dim]*depth
    else:
      assert isinstance(ctx_dim, list) and depth == len(ctx_dim)
-    self.norm = GroupNorm(32, channels, eps=norm_eps)
+    self.norm = GroupNorm(32, channels)
    assert channels == n_heads * d_head
    self.proj_in  = Linear(channels, channels) if use_linear else Conv2d(channels, channels, 1)
    self.transformer_blocks = [BasicTransformerBlock(channels, ctx_dim[d], n_heads, d_head) for d in range(depth)]
@@ -140,9 +134,7 @@ class Upsample:

 # https://github.com/Stability-AI/generative-models/blob/fbdc58cab9f4ee2be7a5e1f2e2787ecd9311942f/sgm/modules/diffusionmodules/openaimodel.py#L472
 class UNetModel:
-  def __init__(self, adm_in_ch:Optional[int], in_ch:int, out_ch:int, model_ch:int, attention_resolutions:List[int], num_res_blocks:int,
-               channel_mult:List[int], transformer_depth:List[int], ctx_dim:Union[int,List[int]], use_linear:bool=False, d_head:Optional[int]=None,
-               n_heads:Optional[int]=None, num_groups:int=32, st_norm_eps:float=1e-5):
+  def __init__(self, adm_in_ch:Optional[int], in_ch:int, out_ch:int, model_ch:int, attention_resolutions:List[int], num_res_blocks:int, channel_mult:List[int], transformer_depth:List[int], ctx_dim:Union[int,List[int]], use_linear:bool=False, d_head:Optional[int]=None, n_heads:Optional[int]=None):
    self.model_ch = model_ch
    self.num_res_blocks = [num_res_blocks] * len(channel_mult)

@@ -182,12 +174,12 @@ class UNetModel:
    for idx, mult in enumerate(channel_mult):
      for _ in range(self.num_res_blocks[idx]):
        layers: List[Any] = [
-          ResBlock(ch, time_embed_dim, model_ch*mult, num_groups),
+          ResBlock(ch, time_embed_dim, model_ch*mult),
        ]
        ch = mult * model_ch
        if ds in attention_resolutions:
          d_head, n_heads = get_d_and_n_heads(ch)
-          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx]))

        self.input_blocks.append(layers)
        input_block_channels.append(ch)
@@ -201,9 +193,9 @@ class UNetModel:

    d_head, n_heads = get_d_and_n_heads(ch)
    self.middle_block: List = [
-      ResBlock(ch, time_embed_dim, ch, num_groups),
-      SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[-1], norm_eps=st_norm_eps),
-      ResBlock(ch, time_embed_dim, ch, num_groups),
+      ResBlock(ch, time_embed_dim, ch),
+      SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[-1]),
+      ResBlock(ch, time_embed_dim, ch),
    ]

    self.output_blocks = []
@@ -211,13 +203,13 @@ class UNetModel:
      for i in range(self.num_res_blocks[idx] + 1):
        ich = input_block_channels.pop()
        layers = [
-          ResBlock(ch + ich, time_embed_dim, model_ch*mult, num_groups),
+          ResBlock(ch + ich, time_embed_dim, model_ch*mult),
        ]
        ch = model_ch * mult

        if ds in attention_resolutions:
          d_head, n_heads = get_d_and_n_heads(ch)
-          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx]))

        if idx > 0 and i == self.num_res_blocks[idx]:
          layers.append(Upsample(ch))
@@ -225,7 +217,7 @@ class UNetModel:
        self.output_blocks.append(layers)

    self.out = [
-      GroupNorm(num_groups, ch),
+      GroupNorm(32, ch),
      Tensor.silu,
      Conv2d(model_ch, out_ch, 3, padding=1),
    ]
@@ -238,10 +230,10 @@ class UNetModel:
      assert y.shape[0] == x.shape[0]
      emb = emb + y.sequential(self.label_emb[0])

-    if is_dtype_supported(mixed_precision_dtype):
-      emb = emb.cast(mixed_precision_dtype)
-      ctx = ctx.cast(mixed_precision_dtype)
-      x   = x  .cast(mixed_precision_dtype)
+    if is_dtype_supported(dtypes.float16):
+      emb = emb.cast(dtypes.float16)
+      ctx = ctx.cast(dtypes.float16)
+      x   = x  .cast(dtypes.float16)

    def run(x:Tensor, bb) -> Tensor:
      if isinstance(bb, ResBlock): x = bb(x, emb)
@@ -272,4 +272,4 @@ def compare_launch_state(states, good_states):

  return True, "PASS"

-# IOCTL=1 CUDA=1 CUDA_PTX=1 python3 test/test_ops.py TestOps.test_tiny_add
+# IOCTL=1 PTX=1 CUDA=1 python3 test/test_ops.py TestOps.test_tiny_add
@@ -0,0 +1,954 @@
+# mypy: disable-error-code="misc, list-item, assignment, operator, index, arg-type"
+from types import SimpleNamespace
+from typing import Any, Sequence, cast, Literal, Callable, get_args, NamedTuple
+import dataclasses, functools, io, math, types, warnings, pathlib, sys, enum
+from tinygrad.tensor import Tensor, _broadcast_shape, ReductionStr
+from tinygrad.helpers import getenv, DEBUG, all_same, prod, flatten, make_tuple, argsort, is_numpy_ndarray, get_single_element
+from tinygrad.dtype import DType, ConstType, dtypes, _from_np_dtype
+from tinygrad.device import is_dtype_supported, Device
+from extra.onnx_parser import onnx_load
+
+# https://github.com/onnx/onnx/blob/rel-1.17.0/onnx/onnx.proto3#L500-L544
+data_types: dict[int, DType] = {
+  1:dtypes.float32, 2:dtypes.uint8, 3:dtypes.int8, 4:dtypes.uint16, 5:dtypes.int16, 6:dtypes.int32, 7:dtypes.int64,
+  9:dtypes.bool, 10:dtypes.float16, 11:dtypes.double, 12:dtypes.uint32, 13:dtypes.uint64, 16:dtypes.bfloat16,
+}
+
+# https://github.com/onnx/onnx/blob/rel-1.17.0/onnx/onnx.proto3#L128-L145
+attribute_types: dict[int, Callable] = {
+  1: lambda a: float(a.f),
+  2: lambda a: int(a.i),
+  3: lambda a: a.s.data().tobytes().decode("utf8") if isinstance(a.s, Tensor) else a.s.decode("utf8"),
+  4: lambda a: buffer_parse(a.t),
+  6: lambda a: tuple(float(x) for x in a.floats),
+  7: lambda a: tuple(int(x) for x in a.ints),
+  8: lambda a: tuple(x.data().tobytes().decode("utf8") for x in a.strings)
+}
+
+# ***** protobuf parsing ******
+from onnx import AttributeProto, TensorProto, TypeProto
+
+def has_field(onnx_type: TypeProto|SimpleNamespace, field):
+  if isinstance(onnx_type, TypeProto): return onnx_type.HasField(field)
+  return hasattr(onnx_type, field)
+
+def dtype_parse(onnx_dtype: int, fallback_context: str | None = None) -> DType:
+  if onnx_dtype not in data_types: raise NotImplementedError(f"onnx dtype id {onnx_dtype} is not supported")
+  if is_dtype_supported(dtype := data_types[onnx_dtype]): return dtype
+  # if fallback_context is provided, we can fall back to a default dtype
+  if fallback_context is not None:
+    default_dtype = dtypes.default_int if dtypes.is_int(dtype) else dtypes.default_float
+    warnings.warn(f"dtype {dtype} on {Device.DEFAULT} from {fallback_context} is not supported, falling back to {default_dtype}")
+    assert is_dtype_supported(default_dtype), f"dtype {default_dtype} must be supported on {Device.DEFAULT}"
+    return default_dtype
+  raise RuntimeError(f"dtype {dtype} on device {Device.DEFAULT} is not supported")
+
+def attribute_parse(onnx_attribute: AttributeProto):
+  if onnx_attribute.type not in attribute_types: raise NotImplementedError(f"attribute type {onnx_attribute.type} is not supported")
+  return attribute_types[onnx_attribute.type](onnx_attribute)
+
+def buffer_parse(onnx_tensor: TensorProto) -> Tensor:
+  if onnx_tensor.string_data: raise NotImplementedError("Parsing for buffer with string data is not implemented.")
+  to_dtype, true_dtype = dtype_parse(onnx_tensor.data_type, "buffer parse"), data_types[onnx_tensor.data_type]
+  shape = tuple(onnx_tensor.dims)
+  keys = ['float_data', 'int32_data', 'int64_data', 'double_data', 'uint64_data', "raw_data"]
+  data = next((val for k in keys if (val := getattr(onnx_tensor, k)) is not None), None)
+  if data is None: raise RuntimeError("empty buffer")
+  if not isinstance(data, Tensor): return Tensor(data, dtype=to_dtype).reshape(shape)
+  assert data.dtype is dtypes.uint8, data.dtype
+  data = data.bitcast(true_dtype).reshape(shape)
+  data = data.to(Device.DEFAULT) if true_dtype is to_dtype else data.to("cpu").cast(to_dtype).to(Device.DEFAULT)
+  if shape == ():
+    if data.dtype is dtypes.float16 and sys.version_info < (3, 12): data = data.cast(dtypes.float32)
+    return Tensor(data.item(), dtype=to_dtype).reshape(shape)
+  return data
+
+def type_parse(onnx_type: TypeProto):
+  elem_type = onnx_type
+  if has_field(elem_type, "map_type") or has_field(elem_type, "sparse_tensor_type") or has_field(elem_type, "opaque_type"):
+    raise NotImplementedError("parsing for map_type, sparse_tensor_type and opaque_type are not implemented")
+  if is_optional := has_field(elem_type, "optional_type"): elem_type = elem_type.optional_type.elem_type
+  if is_sequence := has_field(elem_type, "sequence_type"): elem_type = elem_type.sequence_type.elem_type
+  if has_field(elem_type, "tensor_type"):
+    shape = tuple(getattr(d, "dim_param", None) or getattr(d, "dim_value") for d in elem_type.tensor_type.shape.dim) \
+      if has_field(elem_type.tensor_type, "shape") else None # test_identity_sequence_cpu
+    dtype = data_types[elem_type.tensor_type.elem_type]
+    return OnnxValue(shape, dtype, is_optional, is_sequence)
+  raise RuntimeError(f"TypeProto was not parsed properly: {onnx_type=}")
+
+# ***** onnx spec *****
+@dataclasses.dataclass(frozen=True)
+class OnnxValue:
+  shape: tuple[str|int, ...]
+  dtype: DType
+  is_optional: bool
+  is_sequence: bool
+
+class Domain(enum.Enum):
+  ONNX = "ai.onnx"
+  ONNX_ML = "ai.onnx.ml"
+  AI_ONNX_TRAINING = "ai.onnx.training"
+  AI_ONNX_PREVIEW_TRAINING = "ai.onnx.preview.training"
+  MICROSOFT_CONTRIB_OPS = "com.microsoft"
+  @classmethod
+  def from_onnx(cls, domain: str | None) -> "Domain": return cls.ONNX if domain is None or domain == "" else cls(domain)
+
+class OpSetId(NamedTuple):
+  domain: Domain
+  version: int
+
+@dataclasses.dataclass(frozen=True)
+class OnnxNode:
+  num: int
+  op: str
+  opset_id: OpSetId
+  inputs: tuple[str, ...]
+  outputs: tuple[str, ...]
+  opts: dict[str, Any]
+
+# ***** python const *****
+required_input_python_consts: dict[str, tuple[int, ...]] = {
+  "Tile": (1,), "Range": (0,1,2), "Expand": (1,), "Reshape": (1,), "Squeeze": (1,), "Unsqueeze": (1,), "Trilu": (1,), "ConstantOfShape": (0,),
+  "CumSum": (1,), "TopK": (1,), "Pad": (1,2,3), "MaxUnpool": (2,), "Dropout": (1,2), "CenterCropPad": (1,), "OneHot": (1,), "Compress": (1,),
+  "ImageDecoder": (0,), "AffineGrid": (1,), "Resize": (1,2,3), "Upsample": (1,), "Split": (1,), "Slice": (1,2,3,4),
+  **{"Reduce"+r: (1,) for r in ("Max", "Min", "Sum", "Mean", "SumSquare", "Prod", "L1", "L2", "LogSum", "LogSumExp")},
+  **{optim: (1,) for optim in ("Adam", "Adagrad", "Momentum")}
+}
+
+cache_misses = 0
+@functools.cache
+def _cached_to_python_const(t:Tensor):
+  if t.dtype is dtypes.uint8: return t.data().tobytes()
+  if 0 in t.shape: return []
+  return t.tolist()
+
+# Tensor -> python value cache for parameters
+def to_python_const(t:Any, op:str, idx:int) -> list[ConstType]|ConstType|bytes:
+  if idx not in required_input_python_consts.get(op, ()) or not isinstance(t, Tensor): return t
+  global cache_misses
+  ret = _cached_to_python_const(t)
+  if (info := _cached_to_python_const.cache_info()).misses > cache_misses and DEBUG >= 3:
+    print(f"Cache miss for {t}")
+    cache_misses = info.misses
+  return ret
+
+# ***** runner ******
+debug = int(getenv("DEBUGONNX", "0"))
+limit = int(getenv("ONNXLIMIT", "-1"))
+class OnnxRunner:
+  """
+  `OnnxRunner` executes an ONNX model using Tinygrad.
+
+  Args:
+    model_path: The ONNX model, provided as a file path (a string or Path object) or a Tensor.
+  """
+  def __init__(self, model_path: Tensor | str | pathlib.Path):
+    model = onnx_load(model_path)
+    self.is_training = any(n.domain in {Domain.AI_ONNX_TRAINING, Domain.AI_ONNX_PREVIEW_TRAINING} for n in model.graph.node)
+    self.old_training = Tensor.training
+    Tensor.training = True if self.is_training else False
+    self.graph_values = {"": None, **{x.name:buffer_parse(x) for x in model.graph.initializer}}
+    self.graph_inputs = {x.name:type_parse(x.type) for x in model.graph.input if x.name not in self.graph_values}
+    self.graph_outputs = tuple(x.name for x in model.graph.output)
+    opset_imports = {Domain.from_onnx(getattr(x, "domain", "")):x.version for x in model.opset_import}
+    self.graph_nodes = []
+    for num, n in enumerate(model.graph.node):
+      domain = Domain.from_onnx(n.domain)
+      opset_id = OpSetId(domain, opset_imports.get(domain, 1))
+      self.graph_nodes.append(OnnxNode(num, n.op_type, opset_id, tuple(n.input), tuple(n.output), {x.name:attribute_parse(x) for x in n.attribute}))
+    self.graph_nodes = tuple(self.graph_nodes)
+    self.variable_dims: dict[str, int] = {}
+
+    self.onnx_ops = onnx_ops
+
+  def _parse_input(self, name: str, value: Any, spec: OnnxValue):
+    if spec.is_optional and value is None: return None
+    if spec.is_sequence:
+      if not isinstance(value, Sequence): raise RuntimeError(f"input {name} received {value}, expected a sequence type")
+      sequence = [Tensor(v, dtype=spec.dtype, requires_grad=self.is_training) if not isinstance(v, Tensor) else v for v in value]
+      if not all_same(tuple(t.shape for t in sequence)): raise RuntimeError(f"Shapes for input {name} sequence must be homogeneous")
+      if not all(t.dtype is spec.dtype for t in sequence): warnings.warn(f"Dtypes for input {name} sequence aren't all {spec.dtype}")
+      return sequence
+    dtype = _from_np_dtype(value.dtype) if is_numpy_ndarray(value) else spec.dtype
+    tensor = Tensor(value, dtype=dtype, requires_grad=self.is_training) if not isinstance(value, Tensor) else value
+    if tensor.dtype is not spec.dtype: warnings.warn(f"input {name} has mismatch on dtype. Expected {spec.dtype}, received {tensor.dtype}.")
+    for dim, (onnx_dim, user_dim_input) in enumerate(zip(spec.shape, tensor.shape, strict=True)):
+      if isinstance(onnx_dim, str):
+        onnx_dim = self.variable_dims[onnx_dim] if onnx_dim in self.variable_dims else self.variable_dims.setdefault(onnx_dim, int(user_dim_input))
+      if user_dim_input != onnx_dim: raise RuntimeError(f"input {name} has mismatch on {dim=}. Expected {onnx_dim}, received {user_dim_input}.")
+    return tensor
+
+  def _select_op(self, op:str, required_opset:OpSetId) -> types.FunctionType:
+    if op not in self.onnx_ops: raise NotImplementedError(f"{op=} is not supported")
+    # return default implementation if no opset_id is specified
+    if isinstance(impl := self.onnx_ops[op], types.FunctionType): return impl
+    # match domain and select implementation with latest compatible version
+    eligible_ops = {impl_opset.version:impl_fxn for impl_opset,impl_fxn in impl.items()
+                    if impl_opset.domain == required_opset.domain and impl_opset.version <= required_opset.version}
+    if not eligible_ops: raise NotImplementedError(f"{op=} is not supported for domain {required_opset.domain} and version {required_opset.version}")
+    return eligible_ops[max(eligible_ops.keys())]
+
+  def get_empty_input_data(self, device:str|None=None, dtype:DType|None=None) -> dict[str, Tensor]:
+    return {name:Tensor.empty(*spec.shape, device=device, dtype=dtype or spec.dtype) for name, spec in self.graph_inputs.items()}
+
+  def to(self, device:str|None):
+    self.graph_values = {k:v.to(device) if isinstance(v, Tensor) else v for k,v in self.graph_values.items()}
+    self.graph_nodes = tuple(OnnxNode(n.num, n.op, n.opset_id, tuple(n.inputs), tuple(n.outputs),
+                                      {k:v.to(device) if isinstance(v, Tensor) else v for k,v in n.opts.items()}) for n in self.graph_nodes)
+    return self
+
+  def __call__(self, inputs:dict[str, Any], debug=debug):
+    for name, input_spec in self.graph_inputs.items():
+      if name not in inputs: raise RuntimeError(f"Please provide input data for {name}")
+      self.graph_values[name] = self._parse_input(name, inputs[name], input_spec)
+
+    for node in self.graph_nodes:
+      inps = [to_python_const(self.graph_values[name], node.op, i) for i,name in enumerate(node.inputs)]
+      opts = node.opts
+
+      # provide additional opts
+      if node.op == "Split" and 'num_outputs' not in opts: opts['num_outputs'] = len(node.outputs)
+      if node.op == "Gradient": opts['intermediate_tensors'] = self.graph_values
+
+      if debug >= 1: print(f"{node.num}: op '{node.op}' opt {opts}")
+      if debug >= 2 and node.inputs: print("\tinputs:\n" + "\n".join(f"\t\t{x} - {i!r}" for x,i in zip(node.inputs, inps)))
+      ret = self._select_op(node.op, node.opset_id)(*inps, **opts)
+      ret = ret if isinstance(ret, tuple) else (ret,)
+      if debug >= 2: print("\toutputs:\n" + "\n".join(f"\t\t{x} - {o!r}" for x,o in zip(node.outputs, ret)))
+
+      self.graph_values.update(dict(zip(node.outputs, ret[:len(node.outputs)], strict=True)))
+
+      if node.num == limit:
+        Tensor.training = self.old_training
+        return {name:self.graph_values[name] for name in node.outputs}
+    Tensor.training = self.old_training
+    return {name:self.graph_values[name] for name in self.graph_outputs}
+
+####################
+##### ONNX OPS #####
+####################
+def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionType]]:
+  # ***** helper functions *****
+  def _resolve_const(x: Sequence[ConstType]|ConstType): return x if isinstance(x, get_args(ConstType)) else get_single_element(x)
+
+  def _axes(axes, noop_with_empty_axes): return axes or ([] if noop_with_empty_axes else None)
+
+  # (padding_top, padding_left, ..., padding_bottom, padding_right, ...) -> (padding_left, padding_right, padding_top, padding_bottom, ...)
+  def _onnx_pads_to_tiny_pads(pads): return tuple(flatten(reversed(list(zip(pads, pads[len(pads)//2:])))))
+
+  AUTO_PAD_OPTIONS = Literal["NOTSET", "SAME_UPPER", "SAME_LOWER", "VALID"]
+  # (padding_height, padding_width) -> (padding_top, padding_left, padding_bottom, padding_right)
+  def _auto_pad(pads, auto_pad: AUTO_PAD_OPTIONS):
+    if auto_pad == "SAME_UPPER": return [pads[i]//2 for i in range(len(pads))] + [pads[i]-pads[i]//2 for i in range(len(pads))]
+    return [pads[i]-pads[i]//2 for i in range(len(pads))] + [pads[i]//2 for i in range(len(pads))]
+
+  def _resolve_pool_pads(x:Tensor, p_, k_, d_, s_, auto_pad:AUTO_PAD_OPTIONS):
+    if auto_pad == "VALID": return [0]*(len(k_)*2)
+    i_, (s_,d_,p_) = x.shape[-len(k_):], (make_tuple(x, len(k_)*2) for x in (s_, d_, p_))
+    if auto_pad == "NOTSET": return _onnx_pads_to_tiny_pads(p_ if len(p_)==len(k_)*2 else p_*2)
+    o_ = [((i - (1 if auto_pad in ("SAME_UPPER", "SAME_LOWER") else k)) // s + 1) for i,k,s in zip(i_, k_, s_)]
+    return _onnx_pads_to_tiny_pads(_auto_pad([(o-1)*s+k-i for o,i,k,s in zip(o_, i_, k_, s_)], auto_pad))
+
+  def _clamp_cast(x:Tensor, dtype:DType): return x.clamp(dtypes.min(dtype), dtypes.max(dtype)).cast(dtype)
+
+  def _prepare_quantize(x:Tensor, scale:Tensor, zero_point:Tensor|int, axis=1, block_size=0):
+    if axis < 0: axis += x.ndim
+    # https://github.com/onnx/onnx/blob/main/onnx/reference/ops/op_quantize_linear.py#L31
+    def reshape(val:Tensor):
+      if val.numel() == 1: return val
+      if block_size == 0: return val.reshape([val.shape[0] if dim == axis else 1 for dim in range(x.ndim)])
+      return val.repeat_interleave(block_size, axis)
+    return (reshape(scale), reshape(zero_point) if isinstance(zero_point, Tensor) else zero_point)
+
+  def _op_integer(op, inputs:list[Tensor], zero_points:list[Tensor], **opts):
+    adjusted_inputs = [inp.int() - zp for inp, zp in zip(inputs, zero_points)]
+    return op(*adjusted_inputs, **opts)
+
+  def _qlinearop_quantized(op, inputs:list[Tensor], zero_points:list[Tensor], scales:list[Tensor], out_scale:Tensor, out_zero_point:Tensor, **opts):
+    # op execution is done in quantized int
+    out = _op_integer(op, inputs, zero_points, **opts)
+    assert dtypes.is_int(out.dtype), "quantized op should've done math in int"
+    out_quantized = (out * prod(scales) / out_scale).round() + out_zero_point
+    return _clamp_cast(out_quantized, out_zero_point.dtype)
+
+  def _qlinearop_float(op, inputs:list[Tensor], zero_points:list[Tensor], scales:list[Tensor], out_scale:Tensor, out_zero_point:Tensor, **opts):
+    # op execution is done in float32
+    dequantized_inputs = [(inp.int() - zp) * scale for inp, zp, scale in zip(inputs, zero_points, scales)]
+    out = op(*dequantized_inputs, **opts)
+    assert dtypes.is_float(out.dtype), "op should've done math in float"
+    out_quantized = (out / out_scale).round() + out_zero_point
+    return _clamp_cast(out_quantized, out_zero_point.dtype)
+
+  def _onnx_training(input_group_size):
+    def __decorator(func):
+      def ___wrapper(R:Tensor, T:int, *inputs:Tensor, **kwargs):
+        R = R.detach()
+        groups = len(inputs) // input_group_size
+        ret = [func(R, T, *inps, **kwargs) for inps in (inputs[i::groups] for i in range(groups))]
+        return tuple(flatten(zip(*ret)))
+      return ___wrapper
+    return __decorator
+
+  # ***** Property/Graph Ops *****
+  def Identity(x:Tensor): return x
+  def Constant(sparse_value:Tensor|None=None, value:Tensor|None=None, value_float:float|None=None, value_floats:list[float]|None=None,
+              value_int:int|None=None, value_ints:list[int]|None=None, value_string:str|None=None, value_strings:list[str]|None=None):
+    if value is not None: return value
+    if value_float is not None: return Tensor(value_float, dtype=dtypes.float32, requires_grad=False)
+    if value_floats is not None: return Tensor(list(value_floats), dtype=dtypes.float32, requires_grad=False)
+    if value_int is not None: return Tensor(value_int, dtype=dtypes.int64, requires_grad=False)
+    if value_ints is not None: return Tensor(list(value_ints), dtype=dtypes.int64, requires_grad=False)
+    if value_string is not None or value_strings is not None and sparse_value is not None:
+      raise NotImplementedError('Constant OP not implemented for value_string, value_strings and sparse_value')
+
+  def Range(start:float|int|list[float|int], limit:float|int|list[float|int], delta:float|int|list[float|int]):
+    return Tensor.arange(start=_resolve_const(start), stop=_resolve_const(limit), step=_resolve_const(delta))
+
+  def ImageDecoder(encoded_stream:bytes, pixel_format="RGB"):
+    try: import PIL.Image
+    except ImportError as e: raise ImportError("Pillow must be installed for the ImageDecoder operator") from e
+    img = PIL.Image.open(io.BytesIO(encoded_stream))
+    if pixel_format == "BGR": return Tensor(img.tobytes(), dtype=dtypes.uint8).reshape(*img.size, 3).flip(-1)
+    if pixel_format == "RGB": return Tensor(img.tobytes(), dtype=dtypes.uint8).reshape(*img.size, 3)
+    if pixel_format == "Grayscale": return Tensor(img.convert("L").tobytes(), dtype=dtypes.uint8).reshape(*img.size, 1)
+    raise ValueError(f"pixel_format={pixel_format!r} is not supported.")
+
+  def EyeLike(x:Tensor, dtype:int|None=None, k:int=0):
+    ret = Tensor.eye(cast(int, min(x.shape)), dtype=dtype_parse(dtype, "EyeLike op") if dtype is not None else x.dtype)
+    return ret if x.size(0) == x.size(1) else ret.pad(tuple(None if d == ret.size(0) else (k, d-ret.shape[0]-k) for d in x.shape))
+
+  def OptionalHasElement(x:Tensor|None=None): return Tensor(x is not None and x.numel() > 0)
+  def OptionalGetElement(x:Tensor|None=None): return x if x is not None else Tensor([])
+  def ConstantOfShape(shape:list[int], value:Tensor|None=None):
+    if value is None: value = Tensor(0, dtype=dtypes.float32)
+    if shape == [0]: return Tensor([], dtype=value.dtype)
+    return value.expand(shape)
+
+  def Size(data:Tensor): return data.numel()
+  def Shape(data:Tensor, end:int|None=None, start:int=0): return Tensor(data.shape[start:end], dtype=dtypes.int64)
+
+  # ***** Unary Ops (math) *****
+  def Not(x:Tensor): return x.logical_not()
+  def Clip(x: Tensor, min:Tensor|None=None, max:Tensor|None=None): return x if min is None and max is None else x.clip(min, max)  # noqa: A002
+  def IsInf(x:Tensor, detect_negative:int=1, detect_positive:int=1): return x.isinf(bool(detect_positive), bool(detect_negative))
+
+  # ***** Unary Ops (activation) *****
+  def softmax_1(x:Tensor, axis:int=1): return x.softmax(axis)
+  def softmax_13(x:Tensor, axis:int=-1): return x.softmax(axis)
+  Softmax = {OpSetId(Domain.ONNX, 1):softmax_1, OpSetId(Domain.ONNX, 13):softmax_13}
+  def HardSigmoid(x:Tensor, alpha:float=0.2, beta:float=0.5): return (alpha*x + beta).clip(0, 1)
+  def Gelu(x:Tensor, approximate:str|None=None): return x.gelu() if approximate == "tanh" else 0.5 * x * (1 + (x/math.sqrt(2)).erf())
+  def BiasGelu(x: Tensor, bias: Tensor, approximate: str | None = None) -> Tensor: return Gelu(x + bias, approximate)
+  def FastGelu(x:Tensor, bias:Tensor|None=None): return (x + bias).gelu() if bias is not None else x.gelu() # this is tanh approximated
+  def PRelu(X:Tensor, slope:Tensor): return (X > 0).where(X, X * slope)
+  def LeakyRelu(X:Tensor, alpha:float=0.01): return X.leaky_relu(alpha)
+  def ThresholdedRelu(X:Tensor, alpha:float=1.0): return (X > alpha).where(X, 0)
+  def LogSoftmax(x: Tensor, axis:int=-1): return x.log_softmax(axis)
+  def Binarizer(x:Tensor, threshold:float=0.0): return (x > threshold).float()
+
+  # ***** Unary Ops (broadcasted) *****
+  def Add(x:Tensor,y:Tensor, broadcast=None, axis=None): return x + y
+  def Sub(x:Tensor|int,y:Tensor): return x - y # some test has input as int
+  def Div(x:Tensor,y:Tensor): return x.div(y, rounding_mode='trunc' if dtypes.is_int(x.dtype) else None)
+  def Less(x:Tensor,y:Tensor): return x < y
+  def LessOrEqual(x:Tensor,y:Tensor): return x <= y
+  def Greater(x:Tensor,y:Tensor): return x > y
+  def GreaterOrEqual(x:Tensor,y:Tensor): return x >= y
+  def Equal(x:Tensor,y:Tensor): return x == y
+  def And(x:Tensor,y:Tensor): return (x==y).where(x, False)
+  def Or(x:Tensor,y:Tensor): return (x==y).where(x, True)
+  def Xor(x:Tensor,y:Tensor): return x.bool().bitwise_xor(y.bool())
+  def BitwiseAnd(x:Tensor,y:Tensor): return x & y
+  def BitwiseOr(x:Tensor,y:Tensor): return x | y
+  def BitwiseXor(x:Tensor,y:Tensor): return x ^ y
+  def BitwiseNot(x:Tensor): return ~x
+  def Mod(x:Tensor,y:Tensor,fmod=0):
+    if fmod: return x - x.div(y, rounding_mode="trunc") * y
+    return x % y
+
+  # ***** Casting Ops *****
+  # TODO: saturate
+  def Cast(x:Tensor, to:int, saturate:int=1): return x.cast(dtype_parse(to, "Cast op"))
+  def CastLike(x:Tensor, target_type:Tensor, saturate:int=1): return x.cast(target_type.dtype)
+
+  # ***** Reduce Ops *****
+  def Max(*data_0:Tensor): return functools.reduce(Tensor.maximum, data_0)
+  def Min(*data_0:Tensor): return functools.reduce(Tensor.minimum, data_0)
+  def Sum(*data_0:Tensor): return functools.reduce(Tensor.add, data_0)
+  def Mean(*data_0:Tensor): return Sum(*data_0) / len(data_0)
+  def ReduceMax(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return data.max(_axes(axes, noop_with_empty_axes), keepdim=keepdims)
+  def ReduceMin(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return data.min(_axes(axes, noop_with_empty_axes), keepdim=keepdims)
+  def ReduceSum(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return data.sum(_axes(axes, noop_with_empty_axes), keepdim=keepdims)
+  def ReduceMean(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return data.mean(_axes(axes, noop_with_empty_axes), keepdim=keepdims)
+  def ReduceSumSquare(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return ReduceSum(data.square(), axes, keepdims, noop_with_empty_axes)
+  def ReduceProd(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return data.prod(_axes(axes, noop_with_empty_axes), keepdim=keepdims)
+  def ReduceL1(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return ReduceSum(data.abs(), axes, keepdims, noop_with_empty_axes)
+  def ReduceL2(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return ReduceSumSquare(data, axes, keepdims, noop_with_empty_axes).sqrt()
+  def ReduceLogSum(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return ReduceSum(data, axes, keepdims, noop_with_empty_axes).log()
+  def ReduceLogSumExp(data:Tensor, axes:list[int]|None=None, keepdims:int=1, noop_with_empty_axes:int=0):
+    return ReduceSum(data.exp(), axes, keepdims, noop_with_empty_axes).log()
+  def ArgMax(x:Tensor, axis:int=0, keepdims:int=1, select_last_index:int=0):
+    if select_last_index: return ((x.shape[axis]-1) - x.flip(axis).argmax(axis, keepdim=keepdims)).cast(dtypes.int64)
+    return x.argmax(axis, keepdim=keepdims).cast(dtypes.int64)
+  def ArgMin(x, axis:int=0, keepdims:int=1, select_last_index:int=0):
+    return ArgMax(-x, axis=axis, keepdims=keepdims, select_last_index=select_last_index)
+
+  # ***** Movement Ops *****
+  def Reshape(data:Tensor, shape:list[int], allowzero:int=0):
+    return data.reshape([x if x != 0 else (0 if allowzero else data.shape[i]) for i,x in enumerate(shape)])
+  def Flatten(x:Tensor, axis:int=1): return x.reshape(prod(x.shape[0:axis]), -1)
+  def Expand(x:Tensor, shape:list[int]): return x.expand(_broadcast_shape(x.shape, tuple(shape)))
+  def Shrink(x:Tensor, bias:float=0.0, lambd:float=0.5): return (x < -lambd)*(x+bias) + (x > lambd)*(x-bias)
+  def Transpose(x:Tensor, perm:list[int]|None=None): return x.permute(order=perm or list(range(x.ndim)[::-1]))
+
+  def Squeeze(data:Tensor, axes:list[int]|None=None):
+    return data.squeeze() if axes is None else functools.reduce(lambda d, dim: d.squeeze(dim), sorted(axes, reverse=True), data)
+  def Unsqueeze(data:Tensor, axes:list[int]): return functools.reduce(lambda d, dim: d.unsqueeze(dim), sorted(axes), data)
+
+  def Tile(x:Tensor, repeats:list[int]): return x.repeat(repeats)
+  def Concat(*xs:Tensor, axis:int): return Tensor.cat(*xs, dim=axis)
+  def Slice(data:Tensor, starts:list[int], ends:list[int], axes:list[int]|None=None, steps:list[int]|None=None):
+    axes = axes or list(range(data.ndim))
+    steps = steps or [1]*data.ndim
+    slices = [slice(0,x,1) for x in data.shape]
+    for i, axis in enumerate(axes): slices[axis] = slice(starts[i], ends[i], steps[i])
+    return data[tuple(slices)]
+
+  def Split(data:Tensor, split:list[int]|None=None, num_outputs:int=0, axis:int=0):
+    sz = data.shape[axis]
+    if split is None: split = [sz // num_outputs + (1 if i < sz % num_outputs else 0) for i in range(num_outputs)]
+    return data.split(split, axis)
+
+  def Pad(x:Tensor, pads:list[int], constant_value:ConstType|None=None, axes:list[int]|None=None,
+          mode:Literal["constant", "reflect", "edge", "wrap"]="constant", value=0):
+    value = constant_value or value
+    axes = axes or list(range(x.ndim))
+    real_pads = [0] * (x.ndim*2)
+    for i,axis in enumerate(axes): real_pads[axis%x.ndim], real_pads[axis%x.ndim+x.ndim] = pads[i], pads[i+len(axes)]
+    return x.pad(padding=_onnx_pads_to_tiny_pads(real_pads), mode={"edge":"replicate", "wrap":"circular"}.get(mode, mode), value=value)
+
+  def CenterCropPad(t:Tensor, shape:list[int], axes:list[int]|None=None):
+    shrink_arg:list[None|tuple[int,int]] = [None] * t.ndim
+    pad_arg:list[None|tuple[int,int]] = [None] * t.ndim
+    for s, x in zip(shape, axes or range(t.ndim)):
+      tx = t.shape[x]
+      if s < tx: shrink_arg[x] = (tx//2 - (s+1)//2, tx//2 + s//2)
+      elif s > tx: pad_arg[x] = ((s-tx)//2, (s-tx+1)//2)
+    return t.shrink(tuple(shrink_arg)).pad(tuple(pad_arg))
+
+  # ***** Processing Ops *****
+  def AveragePool(X: Tensor, kernel_shape:list[int], auto_pad:AUTO_PAD_OPTIONS="NOTSET", ceil_mode:int=0, count_include_pad:int=0,
+                  dilations:list[int]|int=1, pads:list[int]|int=0, strides:list[int]|int=1):
+    return X.avg_pool2d(kernel_shape, strides, dilations, _resolve_pool_pads(X, pads, kernel_shape, dilations, strides, auto_pad),
+                        ceil_mode=ceil_mode, count_include_pad=count_include_pad)
+
+  def MaxPool(X: Tensor, kernel_shape:list[int], auto_pad:AUTO_PAD_OPTIONS="NOTSET", ceil_mode:int=0, dilations:list[int]|int=1, pads:list[int]|int=0,
+              storage_order:int=0, strides:list[int]|int=1):
+    pads = _resolve_pool_pads(X, pads, kernel_shape, dilations, strides, auto_pad)
+    ret, idx = X.max_pool2d(kernel_shape, strides, dilations, pads, ceil_mode=ceil_mode, return_indices=True)
+    return ret, idx.transpose(-2, -1).cast(dtypes.int64) if storage_order else idx.cast(dtypes.int64)
+
+  def Conv(X: Tensor, W: Tensor, B:Tensor|None=None, auto_pad:AUTO_PAD_OPTIONS="NOTSET", dilations:list[int]|int=1, group:int=1,
+          kernel_shape:list[int]|None=None, pads:list[int]|int=0, strides:list[int]|int=1):
+    return X.conv2d(W, B, stride=strides, groups=group, dilation=dilations,
+                    padding=_resolve_pool_pads(X, pads, kernel_shape or W.shape[2:], dilations, strides, auto_pad))
+
+  def ConvTranspose(X: Tensor, W: Tensor, B:Tensor|None=None, auto_pad:AUTO_PAD_OPTIONS="NOTSET", dilations:list[int]|int=1, group:int=1,
+                    kernel_shape:list[int]|None=None, pads:list[int]|None=None, output_shape:list[int]|None=None, output_padding:list[int]|int=0,
+                    strides:list[int]|int=1):
+    input_shape, kernel_shape = X.shape[2:], (kernel_shape or W.shape[2:])
+    strides, dilations, output_padding = (make_tuple(x, len(input_shape)) for x in (strides, dilations, output_padding))
+    if output_shape is not None: # we pad according to output_shape
+      pads = _auto_pad([s*(i-1) + op + ((k-1)*d+1) - os for s,i,op,k,d,os in
+                        zip(strides, input_shape, output_padding, kernel_shape, dilations, output_shape)], auto_pad)
+    if pads is None: # we generate pads
+      output_shape = output_shape or [X.shape[i+2] * strides[i] for i in range(len(strides))]
+      pads = [strides[i]*(input_shape[i]-1)+output_padding[i]+((kernel_shape[i]-1)*dilations[i]+1)-output_shape[i] for i in range(len(input_shape))]
+      pads = _auto_pad(pads, auto_pad) if auto_pad != "NOTSET" else [0] * len(input_shape) * 2
+    pads = _onnx_pads_to_tiny_pads(pads)
+    return X.conv_transpose2d(W, B, stride=strides, groups=group, dilation=dilations, padding=pads, output_padding=output_padding)
+
+  def MaxUnpool(xT: Tensor, xI: Tensor, outshape: list[int]|None=None, kernel_shape:list[int]=None, pads:list[int]|int=0, strides:list[int]|int=1):
+    return Tensor.max_unpool2d(xT, xI, kernel_shape, strides, 1, pads, outshape if outshape is None else tuple(outshape))
+
+  def GlobalAveragePool(X:Tensor): return X.mean(axis=tuple(range(2, X.ndim)), keepdim=True)
+  def GlobalMaxPool(X:Tensor): return X.max(axis=tuple(range(2, X.ndim)), keepdim=True)
+
+  def Gemm(A:Tensor, B:Tensor, C:Tensor|None=None, alpha:float=1.0, beta:float=1.0, transA:int=0, transB:int=0, broadcast=0):
+    ret = alpha * (A.transpose(transA) @ B.transpose(transB))
+    if C is not None: ret = ret + beta * (C if broadcast == 0 else C.reshape([-1 if i < len(C.shape) else 1 for i in range(ret.ndim)][::-1]))
+    return ret
+
+  def Einsum(*Inputs:list[Tensor], equation:str): return Tensor.einsum(equation, *Inputs)
+
+  def CumSum(X:Tensor, axis:int|list[int], exclusive:int=0, reverse:int=0):
+    axis = X._resolve_dim(_resolve_const(axis))
+    if reverse: X = X.flip(axis)
+    if exclusive: X = X.pad(tuple((1,0) if i == axis else None for i in range(X.ndim)))\
+                        .shrink(tuple((0,X.shape[axis]) if i == axis else None for i in range(X.ndim)))
+    return X.cumsum(axis).flip(axis) if reverse else X.cumsum(axis)
+
+  def Trilu(x:Tensor, k:int|list[int]=0, upper:int=1):
+    k_ = _resolve_const(k)
+    return x.triu(k_) if upper else x.tril(k_)
+
+  def Resize(X:Tensor, roi:list[float]|None=None, scales:list[float]|None=None, sizes:list[int]|None=None, antialias:int=0,
+            axes:list[int]|None=None, coordinate_transformation_mode:str='half_pixel', cubic_coeff_a:float=-0.75, exclude_outside:int=0,
+            extrapolation_value:float=0.0, keep_aspect_ratio_policy:str='stretch', mode:str='nearest', nearest_mode:str='round_prefer_floor'):
+    def _apply_nearest_mode(index: Tensor, input_dim, mode: str):
+      if mode == "round_prefer_floor": index = (index - 0.5).ceil()
+      elif mode == "round_prefer_ceil": index = (index + 0.5).floor()
+      elif mode in ["floor", "ceil"]: index = getattr(index, mode)()
+      else: raise ValueError(f"invalid {nearest_mode=}")
+      return index.cast(dtypes.int32).clip(0, input_dim-1)
+    def _apply_transformation(index: Tensor, input_dim, scale_dim, mode):
+      # TODO: needs more testing, not confident in this
+      # NOTE: their reference implementation differ from the implementation in their reference docs
+      # https://github.com/onnx/onnx/blob/main/onnx/reference/ops/op_resize.py
+      # https://github.com/onnx/onnx/blob/main/docs/Operators.md#Resize
+      output_dim = scale_dim * input_dim
+      if mode == "half_pixel": index = (index + 0.5) / scale_dim - 0.5
+      elif mode == "align_corners": index = index * (input_dim - 1) / (output_dim - 1) if output_dim != 1 else Tensor([0])
+      elif mode == "asymmetric": index = index / scale_dim
+      elif mode == "pytorch_half_pixel": index = (index + 0.5) / scale_dim - 0.5 if output_dim != 1 else Tensor([-0.5])
+      elif mode == "half_pixel_symmetric": index = input_dim / 2 * (1 - int(output_dim) / output_dim) + (index + 0.5) / scale_dim - 0.5
+      else: raise NotImplementedError(f"invalid {coordinate_transformation_mode=}")
+      return index.clip(0, input_dim-1)
+
+    scales, sizes = (None if scales is None else scales[2-(X.ndim-len(scales)):]), (None if sizes is None else sizes[2-(X.ndim-len(sizes)):])
+    # we pre permute the axes and permute back after resize
+    axes, input_shape, = (axes or list(range(X.ndim))), cast(tuple[int, ...], X.shape[2:]),
+    perm = [a for a in range(len(X.shape)) if a not in axes] + list(axes)
+    X = X.permute(*perm)
+
+    if sizes is not None:
+      if keep_aspect_ratio_policy in ["not_larger", "not_smaller"]:
+        scale_fxn = min if keep_aspect_ratio_policy == "not_larger" else max
+        scales = [scale_fxn([sizes[i] / input_shape[i] for i in range(len(input_shape)) if i+2 in axes])] * 2
+        sizes = [int((scales[0] * input_shape[i]) + 0.5) if i+2 in axes else input_shape[i] for i in range(X.ndim-2)]
+      else:
+        scales = [size / input_shape for size, input_shape in zip(sizes, input_shape)]
+    else:
+      sizes = [int(sc*sh) for sc, sh in zip(scales, input_shape)]
+
+    # NOTE: this transformation makes it so that we can't just call Tensor.interpolate
+    # in Tensor.interpolate, we use indexes without any transformation
+    indexes = []
+    for shape, size, scale in zip(input_shape, sizes, scales):
+      indexes.append(_apply_transformation(Tensor.arange(size), shape, scale, coordinate_transformation_mode))
+
+    if mode == "nearest":
+      indexes = [_apply_nearest_mode(index, shape, nearest_mode) for (index, shape) in zip(indexes, input_shape)]
+      X = X[(..., *Tensor.meshgrid(*indexes))]
+    if mode == "linear":
+      expand = list(X.shape)
+      for i in range(-len(sizes), 0):
+        reshape, index = [1] * X.ndim, indexes[i]
+        reshape[i] = expand[i] = sizes[i]
+        low, high, perc = [y.reshape(reshape).expand(expand) for y in (index.floor().int(), index.ceil().int(), index - index.floor())]
+        X = X.gather(i, low).lerp(X.gather(i, high), perc)
+    if mode == "cubic": raise NotImplementedError("cubic interpolation is not implemented")
+    return X.permute(*argsort(perm)) if perm else X
+  def Upsample(X, scales, mode): return Resize(X=X, scales=scales, mode=mode)  # deprecated
+
+  def TopK(X:Tensor, K:int|list[int], axis:int=-1, largest:int=1, sorted:int=1):  # noqa: A002
+    val, idx = X.topk(_resolve_const(K), axis, largest, sorted)
+    return val, idx.cast(dtypes.int64)
+
+  # ***** Neural Network Ops *****
+  def BatchNormalization(X:Tensor, scale:Tensor, B:Tensor, input_mean:Tensor, input_var:Tensor, epsilon:float=1e-05, momentum:float=0.9,
+                        training_mode:int=0, spatial=1, is_test=0):
+    if training_mode:
+      x_detached = X.detach()
+      current_mean = x_detached.mean(axis=(0,2,3))
+      y = (x_detached - current_mean.reshape(shape=[1, -1, 1, 1]))
+      current_var = (y*y).mean(axis=(0,2,3))
+      current_invstd = current_var.add(epsilon).rsqrt()
+
+      running_mean = input_mean * momentum + current_mean * (1 - momentum)
+      running_var = input_var * momentum + current_var * (1 - momentum)
+
+      return X.batchnorm(scale, B, current_mean, current_invstd), running_mean, running_var
+    return X.batchnorm(scale, B, input_mean, (input_var + epsilon).rsqrt())
+  def GroupNormalization(x:Tensor, scale:Tensor, bias:Tensor, num_groups:int, epsilon:float=1e-05):
+    x = x.reshape(x.shape[0], num_groups, -1).layernorm(eps=epsilon).reshape(x.shape)
+    return x * scale.reshape(1, -1, *[1] * (x.ndim-2)) + bias.reshape(1, -1, *[1] * (x.ndim-2))
+  def InstanceNormalization(x:Tensor, scale:Tensor, bias:Tensor, epsilon:float=1e-05):
+    return GroupNormalization(x, scale, bias, num_groups=x.shape[1], epsilon=epsilon)
+  def LayerNormalization(x:Tensor, scale:Tensor, bias:Tensor, axis:int=-1, epsilon:float=1e-05, stash_type:int=1):
+    assert stash_type == 1, "only float32 is supported"
+    axes = tuple(i for i in range(axis if axis >= 0 else x.ndim + axis, x.ndim))
+    mean = x.mean(axis=axes, keepdim=True)
+    return x.layernorm(axes, epsilon).mul(scale).add(bias), mean, (x.sub(mean)).square().mean(axis=axes, keepdim=True).add(epsilon).rsqrt()
+  def SkipLayerNormalization(x:Tensor, skip:Tensor, gamma:Tensor, beta:Tensor|None=None, bias:Tensor|None=None, epsilon:float=1e-12):
+    x = x + skip
+    if bias is not None: x = x + bias
+    ret = x.layernorm(eps=epsilon) * gamma
+    if beta is not None: ret = ret + beta
+    return ret, None, None, x
+  def EmbedLayerNormalization(input_ids: Tensor, segment_ids:Tensor, word_embedding:Tensor, position_embedding:Tensor,
+                              segment_embedding:Tensor, gamma=None, beta=None, mask:Tensor|None=None,
+                              position_ids:Tensor|None=None, epsilon=1e-12, mask_index_type=0):
+    # https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.EmbedLayerNormalization
+    assert (segment_ids is None) is (segment_embedding is None)
+    assert mask is None and not mask_index_type, "functionality not supported yet"  # TODO
+    input_shape = input_ids.shape
+    seq_length = input_shape[1]
+    compute_seg_emb = (segment_embedding is not None and segment_ids is not None)
+    vocab_size, max_position_embeddings = word_embedding.shape[0], position_embedding.shape[0]
+    type_vocab_size  = (segment_embedding.shape[0] if compute_seg_emb else None)
+
+    def embedding(x:Tensor, vocab_size, weight:Tensor) -> Tensor:
+      return x.unsqueeze(-1).expand(*x.shape, vocab_size)._one_hot_along_dim(vocab_size) @ weight
+
+    # bert embedding layer
+    if position_ids is None: position_ids = Tensor.arange(seq_length, requires_grad=False).unsqueeze(0).expand(*input_shape)
+    wrd_embedding_res = embedding(input_ids, vocab_size, word_embedding)
+    pos_embedding_res = embedding(position_ids, max_position_embeddings, position_embedding)
+    seg_embedding_res = embedding(segment_ids, type_vocab_size, segment_embedding) if compute_seg_emb else None
+
+    embedding_sum = wrd_embedding_res + pos_embedding_res
+    if seg_embedding_res is not None: embedding_sum = embedding_sum + seg_embedding_res
+    out = embedding_sum.layernorm(eps=epsilon) * gamma + beta
+    return out, None, embedding_sum
+  def MeanVarianceNormalization(x:Tensor, axis:list[int]=[0,2,3]):
+    return (x - x.mean(axis, keepdim=True)) / (x.std(axis, keepdim=True, correction=0) + 1e-9)
+
+  def OneHot(indices:Tensor, depth:float|int|list[int|float], values:Tensor, axis:int=-1):
+    # Scalar or Rank 1 tensor containing exactly one element
+    depth = int(_resolve_const(depth))
+    indices = indices.int()
+    indices = (indices < 0).where(indices+depth, indices)
+    return indices.unsqueeze(axis)._one_hot_along_dim(depth, dim=axis).where(values[1], values[0])
+
+  def DepthToSpace(X:Tensor, blocksize:int, mode:str="DCR"):
+    return X.rearrange("b (c h1 w1) h w -> b c (h h1) (w w1)" if mode=="CRD" else "b (h1 w1 c) h w -> b c (h h1) (w w1)", h1=blocksize, w1=blocksize)
+  def SpaceToDepth(X:Tensor, blocksize:int):
+    return X.rearrange("b c (h h1) (w w1) -> b (h1 w1 c) h w", h1=blocksize, w1=blocksize)
+
+  # Reimplemented here because you need legacy RNG for passing ONNX tests.
+  def dropout_7(data:Tensor, ratio:float=0.5, training_mode:bool=False, seed:int|None=None):
+    import numpy as np
+    if not training_mode: return data, data.full_like(True, dtype=dtypes.bool)
+    if seed is not None:
+      rand = Tensor(np.random.RandomState(seed).random(cast(tuple[int,...], data.shape)), requires_grad=False, dtype=data.dtype, device=data.device)
+    else:
+      rand = data.rand_like(requires_grad=False)
+    mask = rand >= ratio
+    return data * mask / (1.0 - ratio), mask
+  # 6 with 'is_test' needed for https://github.com/MTlab/onnx2caffe/raw/refs/heads/master/model/MobileNetV2.onnx
+  def dropout_6(data:Tensor, ratio:float=0.5, is_test=0): return dropout_7(data, ratio, training_mode=not is_test)
+  Dropout = {OpSetId(Domain.ONNX, 6):dropout_6, OpSetId(Domain.ONNX, 7):dropout_7}
+
+  def LRN(x:Tensor, size:int, alpha:float=1e-4, beta:float=0.75, bias:float=1.0):
+    pooled_x = (x**2).rearrange('b c h w -> b 1 c (h w)').pad((0,0,(size-1)//2, size//2)).avg_pool2d((size, 1), 1)
+    return x / (pooled_x.reshape(x.shape) * alpha + bias).pow(beta)
+
+  def NegativeLogLikelihoodLoss(x:Tensor, target:Tensor, weight:Tensor|None=None, ignore_index:int|None=None, reduction:ReductionStr="mean"):
+    return x.nll_loss(target, weight, ignore_index, reduction)
+  def SoftmaxCrossEntropyLoss(scores:Tensor, labels:Tensor, weights:Tensor|None=None, ignore_index:int|None=None, reduction:ReductionStr="mean"):
+    log_probs = scores.log_softmax(1)
+    return log_probs.nll_loss(labels, weights, ignore_index, reduction), log_probs
+
+  def AffineGrid(theta:Tensor, size:list[int], align_corners:int=0):
+    N, _, *spatial_dims = size
+    def generate_grid(steps):
+      if align_corners: return Tensor.linspace(-1, 1, steps, device=theta.device)
+      return Tensor.linspace(-1+1/steps, 1-1/steps, steps, device=theta.device)
+    grids = Tensor.meshgrid(*(generate_grid(d) for d in spatial_dims))
+    base_grid = Tensor.stack(*reversed(grids), Tensor.ones_like(grids[0], device=theta.device), dim=-1)
+    base_grid = base_grid.reshape(1, prod(spatial_dims), len(grids)+1).expand(N, -1, -1)
+    return (base_grid @ theta.transpose(1, 2)).reshape(N, *spatial_dims, -1)
+
+  def attention_contrib(x:Tensor, weights:Tensor, bias:Tensor|None=None, mask_index:Tensor|None=None, past:Tensor|None=None,
+                        attention_bias:Tensor|None=None, past_sequence_length:Tensor|None=None,  do_rotary:int=0, mask_filter_value:float=-10000.0,
+                        num_heads:int|None=None, past_present_share_buffer:int|None=None, qkv_hidden_sizes:list[int]|None=None,
+                        rotary_embedding_dim:int|None=None, scale:float|None=None, unidirectional:int=0):
+    assert not do_rotary and not attention_bias, "TODO"
+    if qkv_hidden_sizes is None: qkv_hidden_sizes = [weights.shape[1] // 3] * 3
+    qkv = x.linear(weights, bias)
+    q, k, v = qkv.split(qkv_hidden_sizes, dim=2)
+
+    batch_size, seq_len, _ = x.shape
+    q_head_size, k_head_size, v_head_size = (sz // num_heads for sz in qkv_hidden_sizes)
+    q, k, v = (x.reshape(batch_size, seq_len, num_heads, hsz).transpose(1, 2) for x, hsz in zip((q, k, v), (q_head_size, k_head_size, v_head_size)))
+
+    present = None
+    if past is not None:
+      k, v = past[0].cat(k, dim=2), past[1].cat(v, dim=2)
+      present = k.stack(v)
+
+    if scale is None: scale = 1.0 / math.sqrt(q_head_size)
+    attn_scores = q @ k.transpose(-1, -2) * scale
+
+    if mask_index is not None:
+      assert 4 >= mask_index.ndim >= 1, f"{mask_index.ndim=}"
+      if mask_index.ndim != 1: mask = mask_index.bool()
+      else:
+        if mask_index.shape[0] == batch_size:
+          mask = Tensor.arange(attn_scores.shape[-1], requires_grad=False, device=mask_index.device).unsqueeze(0) < mask_index.unsqueeze(1)
+        elif mask_index.shape[0] == 2*batch_size:
+          end_positions = mask_index[:batch_size]
+          start_positions = mask_index[batch_size:]
+          arange = Tensor.arange(seq_len).unsqueeze(0)
+          mask = (arange < end_positions.unsqueeze(1)) & (arange >= start_positions.unsqueeze(1))
+        else: raise NotImplementedError("mask_index with shape (3 * batch_size + 2) is not implemented")
+      while mask.ndim < 4: mask = mask.unsqueeze(1)
+      attn_scores = mask.where(attn_scores, mask_filter_value)
+
+    if unidirectional:
+      causal_mask = Tensor.ones((seq_len, seq_len), dtype=dtypes.bool).tril()
+      attn_scores = causal_mask.where(attn_scores, mask_filter_value)
+
+    output = attn_scores.softmax(-1) @ v
+    output = output.transpose(1, 2).reshape(batch_size, seq_len, -1)
+    return output, present
+
+  def attention_onnx(Q:Tensor, K:Tensor, V:Tensor, attn_mask:Tensor|None=None, past_key:Tensor|None=None, past_value:Tensor|None=None,
+                     is_causal:int=0, kv_num_heads:int|None=None, q_num_heads:int|None=None, qk_matmul_output_mode:int=0, scale:float|None=None,
+                     softcap:float=0.0, softmax_precision:int|None=None):
+    input_shape_len = Q.ndim
+    if input_shape_len == 3:
+      assert q_num_heads is not None and kv_num_heads is not None
+      Q = Q.reshape(Q.shape[0], q_num_heads, Q.shape[1], -1)
+      K = K.reshape(K.shape[0], kv_num_heads, K.shape[1], -1)
+      V = V.reshape(V.shape[0], kv_num_heads, V.shape[1], -1)
+
+    if past_key is not None: K = past_key.cat(K, dim=2)
+    if past_value is not None: V = past_value.cat(V, dim=2)
+    present_key, present_value = K, V
+
+    _q_heads, _kv_heads = q_num_heads or Q.shape[1], kv_num_heads or K.shape[1]
+    if _q_heads != _kv_heads:
+      K = K.repeat((1, _q_heads // _kv_heads, 1, 1))
+      V = V.repeat((1, _q_heads // _kv_heads, 1, 1))
+
+    effective_scale = scale if scale is not None else 1.0 / (Q.shape[-1] ** 0.5)
+    scores = (Q @ K.transpose(-1, -2)) * effective_scale
+    qk_matmul_return_val = scores
+
+    if is_causal:
+      causal_mask = Tensor.ones(Q.shape[-2], K.shape[-2], device=Q.device, dtype=dtypes.bool, requires_grad=False).tril(0)
+      scores = scores.masked_fill(causal_mask.logical_not(), -float("inf"))
+
+    if attn_mask is not None:
+      mask_to_add = attn_mask.where(0, -float("inf")) if attn_mask.dtype == dtypes.bool else attn_mask
+      scores = scores + mask_to_add
+    if qk_matmul_output_mode == 1: qk_matmul_return_val = scores
+
+    if softcap > 0.0: scores = (scores / softcap).tanh() * softcap
+    if qk_matmul_output_mode == 2: qk_matmul_return_val = scores
+
+    if softmax_precision: scores = scores.cast({1: dtypes.float32, 10: dtypes.float16, 16: dtypes.bfloat16}[softmax_precision])
+    qk_softmax = scores.softmax(-1).cast(Q.dtype)
+    if qk_matmul_output_mode == 3: qk_matmul_return_val = qk_softmax
+
+    output = (qk_softmax @ V).cast(Q.dtype)
+    if input_shape_len == 3: output = output.permute(0, 2, 1, 3).reshape(Q.shape[0], Q.shape[2], -1)
+    return output, present_key, present_value, qk_matmul_return_val
+  Attention = {OpSetId(Domain.ONNX, 1): attention_onnx, OpSetId(Domain.MICROSOFT_CONTRIB_OPS, 1): attention_contrib}
+
+  def RMSNormalization(X:Tensor, scale:Tensor, axis:int=-1, epsilon:float=1e-5):
+    norm = X.square().mean(axis=tuple(range(axis + X.ndim if axis < 0 else axis, X.ndim)), keepdim=True).add(epsilon).rsqrt()
+    return X * norm * scale
+
+  def RotaryEmbedding(X:Tensor, cos_cache:Tensor, sin_cache:Tensor, position_ids:Tensor|None=None, interleaved:int=0, num_heads:int|None=None,
+                      rotary_embedding_dim:int=0):
+    original_input_shape = X.shape
+
+    if X.ndim == 4: X = X.permute(0, 2, 1, 3)
+    elif X.ndim == 3:
+      assert num_heads is not None, "num_heads must be provided for 3D input"
+      X = X.reshape(*X.shape[:-1], num_heads, X.shape[-1] // num_heads)
+
+    head_size = X.shape[-1]
+    rot_dim = rotary_embedding_dim or head_size
+    x_rotate, x_pass = X[..., :rot_dim], X[..., rot_dim:]
+
+    cos = cos_cache[position_ids] if position_ids is not None else cos_cache[:X.shape[1]]
+    sin = sin_cache[position_ids] if position_ids is not None else sin_cache[:X.shape[1]]
+    cos = cos[..., :rot_dim//2].unsqueeze(2)
+    sin = sin[..., :rot_dim//2].unsqueeze(2)
+
+    if interleaved:
+      x1, x2 = x_rotate[..., ::2], x_rotate[..., 1::2]
+      real = x1 * cos - x2 * sin
+      imag = x1 * sin + x2 * cos
+      x_rotated = Tensor.stack(real, imag, dim=-1).flatten(start_dim=-2)
+    else:
+      x1, x2 = x_rotate.chunk(2, dim=-1)
+      real = x1 * cos - x2 * sin
+      imag = x1 * sin + x2 * cos
+      x_rotated = real.cat(imag, dim=-1)
+
+    output = x_rotated.cat(x_pass, dim=-1)
+    return output.flatten(start_dim=2) if len(original_input_shape) == 3 else output.permute(0, 2, 1, 3)
+
+  # ***** Indexing Ops *****
+  def ArrayFeatureExtractor(x:Tensor, indices:Tensor): return x[..., indices]
+
+  def Gather(x:Tensor, indices:Tensor, axis:int=0):
+    if indices.numel() < 9: # NOTE lessor kernels for smaller indices but kernel number increases depending on size of indices
+      ret_shape = x.shape[:axis] + indices.shape + x.shape[axis+1:]
+      if indices.ndim > 1: indices = indices.flatten()
+      index_consts = [_cached_to_python_const(indices)] if indices.shape == () else _cached_to_python_const(indices)
+      index_consts = [x.shape[axis]+i if i<0 else i for i in index_consts]
+      args = [[(0,x) if j != axis else (i,i+1) for j, x in enumerate(x.shape)] for i in index_consts]
+      return x.shrink(arg=tuple(args[0])).cat(*[x.shrink(arg=tuple(arg)) for arg in args[1:]], dim=axis).reshape(ret_shape)
+    # NOTE faster gather, fixed number of kernels, but exceeds limited kernels for openpilot
+    return x[tuple([slice(None) if i != axis else indices for i in range(x.ndim)])]
+  def Scatter(*args, **kwargs): return ScatterElements(*args, **kwargs) # deprecated
+
+  def GatherND(x:Tensor, indices:Tensor, batch_dims:int=0):
+    if batch_dims == 0: return x[tuple(i.squeeze(-1) for i in indices.split(1, -1))]
+    x_shape, i_shape = x.shape, indices.shape
+    b = math.prod(x.shape[dim] for dim in range(batch_dims))
+    # NOTE: each batched dim of both input and indices are equal
+    x = x.reshape(b, *x.shape[batch_dims:])
+    indices = indices.reshape(b, *indices.shape[batch_dims:])
+    b_idx = Tensor.arange(b, device=x.device).reshape(b, *(1,)*(indices.ndim - 2)).expand(*indices.shape[:-1])
+    ret = x[(b_idx,) + tuple(i.squeeze(-1) for i in indices.split(1, -1))]
+    return ret.reshape(*x_shape[:batch_dims], *i_shape[batch_dims:-1], *ret.shape[indices.ndim-1:])
+  def ScatterND(x:Tensor, indices:Tensor, updates:Tensor, reduction:Literal["none", "add", "mul"]='none'):
+    assert updates.shape == indices.shape[:-1] + x.shape[cast(int, indices.shape[-1]):]
+    x = x.contiguous()
+    for index, u in zip(indices.split(1, 0), updates.split(1, 0)):
+      i = tuple(idx.squeeze(-1) for idx in index.squeeze(0).split(1, -1))
+      u = u.squeeze(0)
+      if reduction == "none": x[i] = u
+      elif reduction == "add": x[i] += u
+      elif reduction == "mul": x[i] *= u
+      else: raise NotImplementedError("reduction doesn't support max or min")
+    return x
+
+  def ScatterElements(x: Tensor, indices: Tensor, updates: Tensor, axis=0, reduction:Literal["none", "add", "mul", "min", "max"]="none"):
+    indices = (indices < 0).where(x.shape[axis], 0) + indices
+    if reduction == "none": return x.scatter(axis, indices, updates)
+    return x.scatter_reduce(axis, indices, updates, {"add": "sum", "mul": "prod", "min": "amin", "max": "amax"}.get(reduction))
+  def GatherElements(x:Tensor, indices:Tensor, axis:int):
+    indices = (indices < 0).where(x.shape[axis], 0) + indices
+    return x.gather(axis, indices)
+
+  def Compress(inp:Tensor, condition:list[bool], axis:int|None=None):
+    if axis is None:
+      inp = inp.flatten()
+      axis = 0
+    if axis < 0: axis += inp.ndim
+    con = Tensor([i for i,cond in enumerate(condition) if cond]) # compress in python
+    return inp[tuple(con if i == axis else slice(None) for i in range(inp.ndim))]
+
+  # ***** Quantization Ops *****
+  def QuantizeLinear(x:Tensor, y_scale:Tensor, y_zero_point:Tensor|int=0, axis:int=1, block_size:int=0, output_dtype:int=0, saturate=1):
+    if isinstance(y_zero_point, Tensor): out_dtype = y_zero_point.dtype
+    elif output_dtype != 0: out_dtype = dtype_parse(output_dtype, "QuantizeLinear op")
+    else: out_dtype = dtypes.uint8
+    y_scale, y_zero_point = _prepare_quantize(x, y_scale, y_zero_point, axis, block_size)
+    if out_dtype == dtypes.uchar:
+      # this appears to work in practice, at least for uchar out_dtype. it folds with the quantize stuff
+      ret = _clamp_cast((x / y_scale + 0.4999999 + y_zero_point).int(), out_dtype)
+    else:
+      ret = _clamp_cast(((x / y_scale).round() + y_zero_point), out_dtype)
+    return ret.contiguous()
+
+  def DynamicQuantizeLinear(x: Tensor):
+    # only support uint8
+    qmin, qmax = dtypes.min(dtypes.uint8), dtypes.max(dtypes.uint8)
+    scale = (x.max().maximum(0) + ((-x).max()).maximum(0)) / (qmax - qmin)
+    zero_point = _clamp_cast((qmin - x.min() / scale).round(), dtypes.uint8)
+    y = _clamp_cast((x / scale).round() + zero_point, dtypes.uint8)
+    return y, scale, zero_point
+
+  def DequantizeLinear(x:Tensor, x_scale:Tensor, x_zero_point:Tensor|int=0, axis:int=1, block_size:int=0):
+    x_scale, x_zero_point = _prepare_quantize(x, x_scale, x_zero_point, axis, block_size)
+    return ((x.int() - x_zero_point) * x_scale).cast(x_scale.dtype)
+
+  def QLinearConv(x:Tensor, x_scale:Tensor, x_zero_point:Tensor|int, w:Tensor, w_scale:Tensor, w_zero_point:Tensor|int, y_scale:Tensor,
+                  y_zero_point: Tensor|int, B:Tensor|None=None, **opts):
+    return _qlinearop_quantized(Conv, [x,w], [x_zero_point,w_zero_point], [x_scale,w_scale], y_scale, y_zero_point, **{"B":B, **opts})
+
+  def QLinearMatMul(a:Tensor, a_scale:Tensor, a_zero_point:Tensor|int, b:Tensor, b_scale:Tensor, b_zero_point:Tensor|int, y_scale:Tensor,
+                    y_zero_point:Tensor|int) -> Tensor:
+    return _qlinearop_quantized(Tensor.matmul, [a,b], [a_zero_point,b_zero_point], [a_scale,b_scale], y_scale, y_zero_point)
+
+  def QLinearAdd(a:Tensor, a_scale:Tensor, a_zero_point:Tensor, b:Tensor, b_scale:Tensor, b_zero_point:Tensor, c_scale:Tensor, c_zero_point:Tensor):
+    return _qlinearop_float(Tensor.add, [a,b], [a_zero_point,b_zero_point], [a_scale,b_scale], c_scale, c_zero_point)
+
+  def QLinearMul(a:Tensor, a_scale:Tensor, a_zero_point:Tensor, b:Tensor, b_scale:Tensor, b_zero_point:Tensor, c_scale:Tensor, c_zero_point:Tensor):
+    return _qlinearop_quantized(Tensor.mul, [a,b], [a_zero_point,b_zero_point], [a_scale,b_scale], c_scale, c_zero_point)
+
+  def QLinearGlobalAveragePool(X:Tensor, x_scale:Tensor, x_zero_point:Tensor, y_scale:Tensor, y_zero_point:Tensor, channels_last:int):
+    assert channels_last == 0, "TODO NHWC"
+    return _qlinearop_float(GlobalAveragePool, [X], [x_zero_point], [x_scale], y_scale, y_zero_point)
+
+  def ConvInteger(x: Tensor, w: Tensor, x_zero_point: Tensor | int = 0, w_zero_point: Tensor | int = 0, B: Tensor | None = None, **opts) -> Tensor:
+    return _op_integer(Conv, [x,w], [x_zero_point,w_zero_point], **{"B":B, **opts})
+
+  def MatMulInteger(A: Tensor, B: Tensor, a_zero_point: Tensor | int = 0, b_zero_point: Tensor | int = 0) -> Tensor:
+    return _op_integer(Tensor.matmul, [A,B], [a_zero_point,b_zero_point])
+
+  # ***** Training Ops *****
+  # NOTE: onnx training ops actually don't need the state for optim, all the ops work in a functional way, but we still can reuse optim.py code
+  @_onnx_training(3)
+  def Adagrad(R:Tensor, T:int, *inputs:Tensor, decay_factor:float=0.0, epsilon:float=0.0, norm_coefficient:float=0.0):
+    X, G, H = (i.detach() for i in inputs)
+    grad = norm_coefficient * X + G
+    H.assign(H + grad.square())
+    up = grad / (H.sqrt() + epsilon)
+    r = R / (1 + T * decay_factor)
+    X.assign(X.detach() - r * up)
+    return [X, H]
+
+  @_onnx_training(4)
+  def Adam(R:Tensor, T:int, *inputs:Tensor, alpha:float=0.9, beta:float=0.999, epsilon:float=0.0, norm_coefficient:float=0.0,
+          norm_coefficient_post:float=0.0):
+    from tinygrad.nn.optim import Adam as TinyAdam
+    X, G, V, H = inputs
+    G, V, H = G.detach(), V.detach(), H.detach()
+    X.grad = norm_coefficient * X.detach() + G
+    opt = TinyAdam([X], b1=alpha, b2=beta, eps=epsilon)
+    opt.m, opt.v, opt.lr = [V], [H], R
+    # need no-op for m_hat and v_hat if T == 0
+    if T == 0: opt.b1_t, opt.b2_t = opt.b1_t.zeros_like(), opt.b2_t.zeros_like()
+    else:
+      # `T-1` since it's applied again at the start of `_step`
+      opt.b1_t = Tensor([alpha**(T-1)], dtype=dtypes.float32, device=X.device, requires_grad=False)
+      opt.b2_t = Tensor([beta**(T-1)], dtype=dtypes.float32, device=X.device, requires_grad=False)
+    opt.step()
+    X = (1 - norm_coefficient_post) * X
+    return [X, V, H]
+
+  @_onnx_training(3)
+  def Momentum(R:Tensor, T:int, *inputs:Tensor, alpha:float, beta:float, mode:str, norm_coefficient:float):
+    X, G, V = (i.detach() for i in inputs)
+    grad = norm_coefficient * X + G
+    # NOTE: this beta_adjusted term makes it so we can't use SGD for nesterov
+    beta_adjusted = beta if T > 0 else 1
+    V.assign(alpha * V + grad * beta_adjusted)
+    X.assign(X - R * (V if mode == "standard" else (grad + alpha * V)))
+    return [X, V]
+
+  def Gradient(*inputs:Tensor, y:str, intermediate_tensors:dict[str, Tensor], **_):
+    intermediate_tensors[y].backward()
+    return tuple([t.grad for t in inputs])
+
+  return {
+    # Tensor ops
+    **{op: getattr(Tensor, op.lower()) for op in ("Neg", "Reciprocal", "Pow", "Sqrt", "Sign", "Abs", "Exp", "Log", "Mish", "Sin", "Cos", "Tan",
+    "Asin", "Acos", "Atan", "Relu", "Sigmoid", "MatMul", "Floor", "Ceil", "IsNaN", "Softplus", "HardSwish", "Where", "Mul", "Sinh", "Cosh",
+    "Tanh", "Softsign", "Asinh", "Acosh", "Atanh",  "Elu", "Celu", "Selu", "Round", "Erf")},
+    # Implemented ops
+    **{name:obj for name,obj in locals().items() if isinstance(obj, types.FunctionType) and not name.startswith("_") and name[0].isupper()},
+    # Version ops
+    **{name:obj for name,obj in locals().items() if isinstance(obj, dict)},
+  }
+
+onnx_ops = get_onnx_ops()
@@ -1,6 +1,7 @@
 from tinygrad import Tensor
 from tinygrad.tensor import _to_np_dtype
-from tinygrad.frontend.onnx import OnnxRunner, OnnxValue
+from tinygrad.frontend.onnx import OnnxRunner
+from extra.onnx import OnnxValue
 import numpy as np
 import onnxruntime as ort

@@ -0,0 +1,207 @@
+# https://github.com/onnx/onnx/blob/main/onnx/onnx.proto3
+
+import os, pathlib, struct
+from io import BufferedReader
+from types import SimpleNamespace
+from tinygrad.nn.state import TensorIO
+from tinygrad.tensor import Tensor
+
+# Protobuf Wire Types
+WIRETYPE_VARINT = 0; WIRETYPE_FIXED64 = 1; WIRETYPE_LENGTH_DELIMITED = 2; WIRETYPE_START_GROUP = 3; WIRETYPE_END_GROUP = 4; WIRETYPE_FIXED32 = 5 # noqa: E702
+
+# TensorProto.DataType
+class TensorDataType:
+  UNDEFINED = 0; FLOAT = 1; UINT8 = 2; INT8 = 3; UINT16 = 4; INT16 = 5; INT32 = 6; INT64 = 7 # noqa: E702
+  STRING = 8; BOOL = 9; FLOAT16 = 10; DOUBLE = 11; UINT32 = 12; UINT64 = 13; COMPLEX64 = 14; COMPLEX128 = 15; BFLOAT16 = 16 # noqa: E702
+
+# AttributeProto.AttributeType
+class AttributeType:
+  UNDEFINED = 0; FLOAT = 1; INT = 2; STRING = 3; TENSOR = 4; GRAPH = 5; SPARSE_TENSOR = 11; TYPE_PROTO = 13; FLOATS = 6; INTS = 7 # noqa: E702
+  STRINGS = 8; TENSORS = 9; GRAPHS = 10; SPARSE_TENSORS = 12; TYPE_PROTOS = 14 # noqa: E702
+
+class PBType: FLOAT = 1; INT = 2; STRING = 3; FLOATS = 4; INTS = 5; STRINGS = 6; BYTES = 7; SUB = 8 # noqa: E702
+
+PB_INFOS: dict[str, dict] = {
+  "OperatorSetIdProto": {1: ("domain", PBType.STRING), 2: ("version", PBType.INT)},
+  "StringStringEntryProto": {1: ("key", PBType.STRING), 2: ("value", PBType.STRING)},
+  "TensorProto": {1: ("dims", PBType.INT, True), 2: ("data_type", PBType.INT), 4: ("float_data", PBType.FLOATS),
+    13: ("external_data", PBType.SUB, True, "StringStringEntryProto"), 14: ("data_location", PBType.INT),
+    5: ("int32_data", PBType.INTS), 7: ("int64_data", PBType.INTS), 8: ("name", PBType.STRING), 9: ("raw_data", PBType.BYTES),
+    10: ("double_data", PBType.FLOATS), 11: ("uint64_data", PBType.INTS)},
+  "TensorShapeProtoDimension": {1: ("dim_value", PBType.INT), 2: ("dim_param", PBType.STRING)},
+  "TensorShapeProto": {1: ("dim", PBType.SUB, True, "TensorShapeProtoDimension")},
+  "ModelProto": {1: ("ir_version", PBType.INT), 5: ("model_version", PBType.INT),
+    2: ("producer_name", PBType.STRING), 3: ("producer_version", PBType.STRING), 4: ("domain", PBType.STRING), 6: ("doc_string", PBType.STRING),
+    7: ("graph", PBType.SUB, False, ("GraphProto", lambda: {"node": [], "initializer": [], "input": [], "output": [], "value_info": []})),
+    8: ("opset_import",PBType.SUB, True, "OperatorSetIdProto")},
+  "GraphProto": {2: ("name", PBType.STRING), 10: ("doc_string", PBType.STRING),
+    1: ("node", PBType.SUB, True, ("NodeProto", lambda: {"input": [], "output": [], "attribute": [], "domain": None})),
+    5: ("initializer", PBType.SUB, True, ("TensorProto", lambda: {"dims": [], "float_data": None, "int32_data": None, "string_data": None,
+                                                                  "int64_data": None, "double_data": None, "uint64_data": None, "raw_data": None})),
+    11: ("input", PBType.SUB, True, "ValueInfoProto"), 12: ("output", PBType.SUB, True, "ValueInfoProto")},
+  "NodeProto": { 1: ("input", PBType.STRING, True), 2: ("output", PBType.STRING, True), 3: ("name", PBType.STRING),
+    4: ("op_type", PBType.STRING), 6: ("doc_string", PBType.STRING), 7: ("domain", PBType.STRING),
+    5: ("attribute", PBType.SUB, True, ("AttributeProto", lambda: {"floats": [], "ints": [], "strings": []}))},
+  "AttributeProto": {1: ("name", PBType.STRING), 20: ("type", PBType.INT), 3: ("i", PBType.INT), 8: ("ints", PBType.INT, True),
+    2: ("f", PBType.FLOAT), 7: ("floats", PBType.FLOAT, True), 4: ("s", PBType.BYTES), 9: ("strings", PBType.BYTES, True),
+    5:("t", PBType.SUB, False, ("TensorProto", lambda: {"dims": [], "float_data": None, "int32_data": None, "string_data": None, "int64_data": None,
+                                                        "double_data": None, "uint64_data": None, "raw_data": None}))},
+  "ValueInfoProto": {1: ("name", PBType.STRING), 2: ("type", PBType.SUB, False, "TypeProto"), 3: ("doc_string", PBType.STRING)},
+  "TypeProto": {1: ("tensor_type", PBType.SUB, False, "TypeProtoTensor"), 4: ("sequence_type", PBType.SUB, False, "TypeProtoSequence"),
+    9: ("optional_type", PBType.SUB, False, "TypeProtoOptional"), 6: ("denotation", PBType.STRING)},
+  "TypeProtoSequence": {1: ("elem_type", PBType.SUB, False, "TypeProto")},
+  "TypeProtoOptional": {1: ("elem_type", PBType.SUB, False, "TypeProto")},
+  "TypeProtoTensor": {1: ("elem_type", PBType.INT), 2: ("shape", PBType.SUB, False, ("TensorShapeProto", lambda: {"dim": []}))},
+}
+
+def onnx_load(fn: Tensor|str|pathlib.Path, load_external_data: bool=True):
+  parser = OnnxParser(fn, load_external_data)
+  onnx_model = parser.parse()
+  model = dict_to_namespace(onnx_model)
+  return model
+
+def gen_result(obj: dict, key_name, val, repeated: bool):
+  if repeated: obj.setdefault(key_name, []).append(val)
+  else: obj[key_name] = val
+
+def dict_to_namespace(d):
+  if isinstance(d, dict): return SimpleNamespace(**{k: dict_to_namespace(v) for k, v in d.items()})
+  elif isinstance(d, list): return [dict_to_namespace(i) for i in d]
+  return d
+
+class OnnxParser:
+  def __init__(self, inp: Tensor|str|pathlib.Path, load_external_data: bool=True):
+    self.file_path: pathlib.Path|None = None
+    self.load_external_data = load_external_data
+    if not isinstance(inp, Tensor):
+      self.file_path = pathlib.Path(inp)
+      self.tensor = Tensor(self.file_path)
+    else: self.tensor = inp
+    self.attr_func_dict = { PBType.BYTES: self._handle_bytes, PBType.SUB: self._handle_sub_message, PBType.FLOATS: self._handle_packed_floats,
+      PBType.INT: self._handle_int64, PBType.INTS: self._handle_packed_int64s, PBType.STRING: self._handle_string, PBType.FLOAT: self._handle_float}
+    self.registered_handles = {}
+    for pb_name in PB_INFOS:
+      res = {}
+      for fid, config in PB_INFOS[pb_name].items():
+        parser_fn, repeated = None, False
+        if len(config) == 2: name, attr = config
+        elif len(config) == 3: name, attr, repeated = config
+        elif len(config) == 4: name, attr, repeated, parser_fn = config
+        handler_fn = self.attr_func_dict[attr]
+        def _wrapper_handler(obj, reader, wt, h=handler_fn, n=name, p=parser_fn, r=repeated): return h(obj, n, reader, wt, parser_func=p, repeated=r)
+        res[fid] = _wrapper_handler
+      self.registered_handles[pb_name] = res
+
+  def parse(self):
+    reader = BufferedReader(TensorIO(self.tensor))
+    return self._parse_message(reader, "ModelProto", lambda: {"opset_import": [], "domain": None, "graph": None})
+
+  def decode_varint(self, reader: BufferedReader) -> int:
+    result = 0
+    shift = 0
+    while True:
+      data = reader.read(1)
+      if data == b"": raise EOFError("decode_varint EOF")
+      result |= (data[0] & 0x7F) << shift
+      if not (data[0] & 0x80): return result
+      shift += 7
+      if shift >= 70: raise ValueError("Varint too long")
+
+  def skip_field_value(self, reader: BufferedReader, wire_type):
+    if wire_type == WIRETYPE_VARINT: self.decode_varint(reader)
+    elif wire_type == WIRETYPE_FIXED64: reader.seek(8, os.SEEK_CUR)
+    elif wire_type == WIRETYPE_FIXED32: reader.seek(4, os.SEEK_CUR)
+    elif wire_type == WIRETYPE_LENGTH_DELIMITED: reader.seek(self.decode_varint(reader), os.SEEK_CUR)
+    else: raise ValueError(f"Unknown wire type: {wire_type}")
+
+  def _parse_message(self, reader, message_field_handlers_name, initial_obj_factory=lambda: {}):
+    message_field_handlers = self.registered_handles[message_field_handlers_name]
+    obj = initial_obj_factory()
+    while True:
+      try:
+        tag_val = self.decode_varint(reader)
+        field_number = tag_val >> 3
+        wire_type = tag_val & 0x07
+        if handler := message_field_handlers.get(field_number):
+          handler(obj, reader, wire_type)
+        else: self.skip_field_value(reader, wire_type)
+      except EOFError: break
+    if message_field_handlers_name == "TensorProto" and self.load_external_data and obj.get("data_location", 0) == 1: self._parse_external_data(obj)
+    return obj
+
+  def _handle_delimited(self, reader:BufferedReader, use_tensor=False) -> Tensor|bytes:
+    str_len = self.decode_varint(reader)
+    if not use_tensor: return reader.read(str_len)
+    raw = reader.raw
+    assert isinstance(raw, TensorIO)
+    res = raw._tensor[reader.tell():(reader.tell()+str_len)]
+    reader.seek(str_len, os.SEEK_CUR)
+    return res
+
+  def _handle_string(self, obj, key_name, reader, wire_type, parser_func=None, repeated=False):
+    if wire_type != WIRETYPE_LENGTH_DELIMITED: raise ValueError(f"Expected length-delimited for string field '{key_name}'")
+    value = self._handle_delimited(reader)
+    assert isinstance(value, bytes)
+    gen_result(obj, key_name, value.decode("utf-8"), repeated)
+
+  def _handle_bytes(self, obj, key_name, reader, wire_type, parser_func=None, repeated=False):
+    if wire_type != WIRETYPE_LENGTH_DELIMITED: raise ValueError(f"Expected length-delimited for bytes field '{key_name}'")
+    value = self._handle_delimited(reader, use_tensor=True)
+    gen_result(obj, key_name, value, repeated)
+
+  def _handle_int64(self, obj, key_name, reader, wire_type, parser_func=None, repeated=False):
+    if wire_type != WIRETYPE_VARINT: raise ValueError(f"Expected varint for int64 field '{key_name}'")
+    val = self.decode_varint(reader)
+    gen_result(obj, key_name, val - 2**64 if val & (1 << 63) else val, repeated)
+
+  def _handle_float(self, obj, key_name, reader, wire_type, parser_func=None, repeated=False):
+    if wire_type != WIRETYPE_FIXED32: raise ValueError(f"Expected fixed32 for float field '{key_name}'")
+    val, = struct.unpack("<f", reader.read(4))
+    gen_result(obj, key_name, val, repeated)
+
+  def _handle_packed_int64s(self, obj, key_name, reader, wire_type, parser_func=None, repeated=False):
+    if wire_type != WIRETYPE_LENGTH_DELIMITED: raise ValueError("Packed int64s expected length_delimited")
+    total_bytes_len = self.decode_varint(reader)
+    old_pos = reader.tell()
+    values = []
+    while reader.tell() < total_bytes_len + old_pos:
+      val = self.decode_varint(reader) # need copy here because packed ints are varint
+      values.append(val - 2**64 if val & (1 << 63) else val)
+    obj[key_name] = values
+
+  def _handle_packed_floats(self, obj, key_name, reader, wire_type, parser_func=None, repeated=False):
+    if wire_type != WIRETYPE_LENGTH_DELIMITED: raise ValueError("Packed floats expected length_delimited")
+    value = self._handle_delimited(reader, use_tensor=True)
+    obj[key_name] = value
+
+  def _handle_sub_message(self, obj, key_name, reader, wire_type, parser_func=None, repeated=False):
+    if wire_type != WIRETYPE_LENGTH_DELIMITED: raise ValueError(f"Expected length-delimited for sub-message field '{key_name}'")
+    value = self._handle_delimited(reader, use_tensor=True)
+    assert isinstance(value, Tensor)
+    if isinstance(parser_func, str): sub_obj = self._parse_message(BufferedReader(TensorIO(value)), parser_func)
+    elif isinstance(parser_func, tuple): sub_obj = self._parse_message(BufferedReader(TensorIO(value)), parser_func[0], parser_func[1])
+    else: sub_obj = parser_func(BufferedReader(TensorIO(value)))
+    gen_result(obj, key_name, sub_obj, repeated)
+
+  def _parse_external_data(self, obj):
+    if "external_data" not in obj: raise ValueError("no external_data")
+    location = None
+    length = None
+    offset = 0
+    for kv in obj["external_data"]:
+      if kv["key"] == "location": location = kv["value"]
+      if kv["key"] == "offset": offset = int(kv["value"])
+      if kv["key"] == "length": length = int(kv["value"])
+    if location is None: raise ValueError("no location in external_data")
+    if self.file_path is None:
+      # get onnx file path from Tensor
+      if isinstance(self.tensor.device, str) and self.tensor.device.startswith("DISK:"):
+        self.file_path = pathlib.Path(self.tensor.device[5:])
+        if not (ext_path := self.file_path.parent.joinpath(location)).exists():
+          raise Exception(f"external location not exists: {ext_path}, may caused by symbolic link, try passing onnx file path to onnx_load")
+      else: raise Exception("onnx external_data need the origin file path, try passing onnx file path to onnx_load")
+    ext_path = self.file_path.parent.joinpath(location)
+    if not ext_path.exists(): raise Exception(f"external location not exists: {ext_path}")
+    ext_tensor = Tensor(ext_path)
+    obj["raw_data"] = ext_tensor[offset:offset+length] if length is not None else ext_tensor[offset:]
+    obj["data_location"] = 0
@@ -5,9 +5,9 @@ from tinygrad.nn import Linear
 from tinygrad.tensor import Tensor
 from tinygrad.nn.optim import Adam
 from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import actions
+from tinygrad.opt.search import actions
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, assert_same_lin
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tinygrad.helpers import getenv

 # stuff needed to unpack a kernel
@@ -17,7 +17,7 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.uop.ops import Variable
 inf, nan = float('inf'), float('nan')
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps

 INNER = 256
 class PolicyNet:
@@ -10,11 +10,11 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.uop.ops import Variable
 inf, nan = float('inf'), float('nan')
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps

 # more stuff
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import actions
+from tinygrad.opt.kernel import Kernel
+from tinygrad.opt.search import actions
 from extra.optimization.helpers import lin_to_feats
 from extra.optimization.pretrain_valuenet import ValueNet
 from tinygrad.nn.optim import Adam
@@ -7,7 +7,7 @@ rm $LOGOPS
 test/external/process_replay/reset.py

 CI=1 python3 -m pytest -n=auto test/test_ops.py test/test_nn.py test/test_winograd.py test/models/test_real_world.py --durations=20
-CL=1 python3 -m pytest test/test_tiny.py
+GPU=1 python3 -m pytest test/test_tiny.py

 # extract, sort and uniq
 extra/optimization/extract_dataset.py
@@ -1,8 +1,8 @@
 import random
 from extra.optimization.helpers import load_worlds, ast_str_to_lin
-from tinygrad.codegen.opt.search import actions
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.search import actions
+from tinygrad.opt.kernel import Kernel
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from tinygrad.helpers import tqdm

 tactions = set()
@@ -1,6 +1,6 @@
 # stuff needed to unpack a kernel
 from tinygrad import Variable
-from tinygrad.codegen.opt import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.dtype import dtypes, PtrDType
 from tinygrad.shape.shapetracker import ShapeTracker
@@ -11,7 +11,7 @@ inf, nan = float('inf'), float('nan')
 UOps = Ops

 # kernel unpacker
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 def ast_str_to_ast(ast_str:str) -> UOp: return eval(ast_str)
 def ast_str_to_lin(ast_str:str, opts=None): return Kernel(ast_str_to_ast(ast_str), opts=opts)
 def kern_str_to_lin(kern_str:str, opts=None):
@@ -103,7 +103,7 @@ def lin_to_feats(lin:Kernel, use_sts=True):
  return ret

 from tinygrad.device import Device, Buffer
-from tinygrad.codegen.opt.search import _ensure_buffer_alloc, _time_program
+from tinygrad.opt.search import _ensure_buffer_alloc, _time_program
 from tinygrad.helpers import to_function_name, CACHELEVEL, diskcache_get, diskcache_put

 def time_linearizer(lin:Kernel, rawbufs:list[Buffer], allow_test_size=True, max_global_size=65536, cnt=3, disable_cache=False, clear_l2=False) -> float:  # noqa: E501
@@ -115,7 +115,7 @@ def time_linearizer(lin:Kernel, rawbufs:list[Buffer], allow_test_size=True, max_
  assert dev.compiler is not None

  rawbufs = _ensure_buffer_alloc(rawbufs)
-  var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
+  var_vals: dict[Variable, int] = {k:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
  p = get_program(lin.get_optimized_ast(), lin.opts)
  tms = _time_program(p, dev.compiler.compile(p.src), var_vals, rawbufs,
                      max_global_size=max_global_size if allow_test_size else None, clear_l2=clear_l2, cnt=cnt, name=to_function_name(lin.name))
@@ -1,4 +1,4 @@
-from tinygrad.codegen.opt.kernel import Kernel
+from tinygrad.opt.kernel import Kernel
 from tqdm import tqdm, trange
 import math
 import random
@@ -14,7 +14,7 @@ from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
 from tinygrad.uop.ops import Variable
 inf, nan = float('inf'), float('nan')
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.opt.kernel import Opt, OptOps

 from extra.optimization.helpers import lin_to_feats, MAX_DIMS

@@ -3,7 +3,7 @@ import numpy as np
 import math, random
 from tinygrad.tensor import Tensor
 from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import actions, bufs_from_lin, get_kernel_actions
+from tinygrad.opt.search import actions, bufs_from_lin, get_kernel_actions
 from tinygrad.nn.optim import Adam
 from extra.optimization.extract_policynet import PolicyNet
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, time_linearizer
@@ -1,6 +1,6 @@
 from typing import List, Tuple
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import get_kernel_actions, actions
+from tinygrad.opt.kernel import Kernel
+from tinygrad.opt.search import get_kernel_actions, actions

 _net = None
 def beam_q_estimate(beam:List[Tuple[Kernel, float]]) -> List[Tuple[Kernel, float]]:
@@ -4,8 +4,8 @@ from extra.optimization.helpers import ast_str_to_lin, time_linearizer
 from tinygrad import dtypes
 from tinygrad.helpers import BEAM, getenv
 from tinygrad.device import Device, Compiled
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
+from tinygrad.opt.kernel import Kernel
+from tinygrad.opt.search import beam_search, bufs_from_lin


 if __name__ == '__main__':
@@ -16,9 +16,9 @@ class TestBeamSearch(unittest.TestCase):
    BEAM.value = self.old_beam

  def test_variable_ast_beam(self):
-    vi = Variable("a", 1, 10).bind(3)
-    a = rand(10, 3)[:vi]
-    a = (a+1).realize()
+    with Context(IGNORE_OOB=1):
+      a = rand(3, 3).reshape((Variable("a", 1, 10).bind(3), 3))
+      a = (a+1).realize()

  def test_big_prime_number(self):
    a = rand(367, 367)
@@ -42,16 +42,18 @@ class TestBeamSearch(unittest.TestCase):

  def test_variable_big_prime_number(self):
    v = Variable("v", 1, 400).bind(367)
-    a = rand(367, 400)
-    b = rand(400, 367)
-    c = (a[:, :v] @ b[:v, :]).realize()
-    np.testing.assert_allclose(c.numpy(), a[:, :367].numpy() @ b[:367, :].numpy(), atol=1e-4, rtol=1e-4)
+    a = rand(367, 367)
+    b = rand(367, 367)
+    with Context(IGNORE_OOB=1):
+      c = (a.reshape(367, v) @ b.reshape(v, 367)).realize()
+      np.testing.assert_allclose(c.numpy(), a.numpy() @ b.numpy(), atol=1e-4, rtol=1e-4)

  def test_variable_shrink_prime_number(self):
    v = Variable("v", 1, 400).bind(367)
    a = rand(400, 367)
-    b = (a.shrink(((0,v), None))+1)[:367,:367].realize()
-    np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)
+    with Context(IGNORE_OOB=1):
+      b = (a.shrink(((0,v), None))+1).reshape(367,367).realize()
+      np.testing.assert_allclose(b.numpy(), a.numpy()[:367]+1, atol=1e-4, rtol=1e-4)

  def test_no_mutate_rawbuffers(self):
    a = rand(3, 3).realize()
@@ -6,8 +6,8 @@ from copy import deepcopy
 from tinygrad.helpers import getenv, colored
 from tinygrad.tensor import Tensor
 from tinygrad.nn.state import get_parameters, get_state_dict, safe_save, safe_load, load_state_dict
-from tinygrad.codegen.opt.search import bufs_from_lin, actions, get_kernel_actions
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.search import bufs_from_lin, actions, get_kernel_actions
+from tinygrad.opt.heuristic import hand_coded_optimizations
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, lin_to_feats, time_linearizer
 from extra.optimization.extract_policynet import PolicyNet
 from extra.optimization.pretrain_valuenet import ValueNet
@@ -1,5 +1,5 @@
 from extra.optimization.helpers import load_worlds, ast_str_to_lin, time_linearizer
-from tinygrad.codegen.opt.search import bufs_from_lin, get_kernel_actions
+from tinygrad.opt.search import bufs_from_lin, get_kernel_actions

 if __name__ == "__main__":
  ast_strs = load_worlds()
@@ -1,6 +1,6 @@
 import sys, pickle, decimal, json
 from tinygrad.device import ProfileDeviceEvent, ProfileGraphEvent
-from tinygrad.helpers import tqdm, temp, ProfileEvent, ProfileRangeEvent, TracingKey
+from tinygrad.helpers import tqdm, temp, ProfileEvent, ProfileRangeEvent

 devices:dict[str, tuple[decimal.Decimal, decimal.Decimal, int]] = {}
 def prep_ts(device:str, ts:decimal.Decimal, is_copy): return int(decimal.Decimal(ts) + devices[device][is_copy])
@@ -11,14 +11,12 @@ def dev_ev_to_perfetto_json(ev:ProfileDeviceEvent):
          {"name": "thread_name", "ph": "M", "pid": dev_to_pid(ev.device)['pid'], "tid": 0, "args": {"name": "COMPUTE"}},
          {"name": "thread_name", "ph": "M", "pid": dev_to_pid(ev.device)['pid'], "tid": 1, "args": {"name": "COPY"}}]
 def range_ev_to_perfetto_json(ev:ProfileRangeEvent):
-  name = ev.name.display_name if isinstance(ev.name, TracingKey) else ev.name
-  return [{"name": name, "ph": "X", "ts": prep_ts(ev.device, ev.st, ev.is_copy), "dur": float(ev.en-ev.st), **dev_to_pid(ev.device, ev.is_copy)}]
+  return [{"name": ev.name, "ph": "X", "ts": prep_ts(ev.device, ev.st, ev.is_copy), "dur": float(ev.en-ev.st), **dev_to_pid(ev.device, ev.is_copy)}]
 def graph_ev_to_perfetto_json(ev:ProfileGraphEvent, reccnt):
  ret = []
  for i,e in enumerate(ev.ents):
    st, en = ev.sigs[e.st_id], ev.sigs[e.en_id]
-    name = e.name.display_name if isinstance(e.name, TracingKey) else e.name
-    ret += [{"name": name, "ph": "X", "ts": prep_ts(e.device, st, e.is_copy), "dur": float(en-st), **dev_to_pid(e.device, e.is_copy)}]
+    ret += [{"name": e.name, "ph": "X", "ts": prep_ts(e.device, st, e.is_copy), "dur": float(en-st), **dev_to_pid(e.device, e.is_copy)}]
    for dep in ev.deps[i]:
      d = ev.ents[dep]
      ret += [{"ph": "s", **dev_to_pid(d.device, d.is_copy), "id": reccnt+len(ret), "ts": prep_ts(d.device, ev.sigs[d.en_id], d.is_copy), "bp": "e"}]
@@ -26,8 +24,6 @@ def graph_ev_to_perfetto_json(ev:ProfileGraphEvent, reccnt):
  return ret
 def to_perfetto(profile:list[ProfileEvent]):
  # Start json with devices.
-  profile += [ProfileDeviceEvent("TINY")]
-
  prof_json = [x for ev in profile if isinstance(ev, ProfileDeviceEvent) for x in dev_ev_to_perfetto_json(ev)]
  for ev in tqdm(profile, desc="preparing profile"):
    if isinstance(ev, ProfileRangeEvent): prof_json += range_ev_to_perfetto_json(ev)
@@ -1,6 +1,6 @@
 import ctypes, array
 from hexdump import hexdump
-from tinygrad.runtime.ops_cl import CLDevice
+from tinygrad.runtime.ops_gpu import GPUDevice
 from tinygrad.helpers import getenv, to_mv, mv_address
 from tinygrad.dtype import dtypes
 from tinygrad import Tensor, TinyJit
@@ -8,7 +8,7 @@ from tinygrad.runtime.autogen import opencl as cl
 if getenv("IOCTL"): import extra.qcom_gpu_driver.opencl_ioctl  # noqa: F401  # pylint: disable=unused-import

 # create raw opencl buffer.
-gdev = CLDevice()
+gdev = GPUDevice()
 cl_buf = cl.clCreateBuffer(gdev.context, cl.CL_MEM_READ_WRITE, 0x100, None, status := ctypes.c_int32())
 assert status.value == 0

@@ -673,7 +673,6 @@ impl<'a> Thread<'a> {
                                        39 => f32::log2(s0),
                                        42 => 1.0 / s0,
                                        43 => 1.0 / s0,
-                                        46 => 1.0 / f32::sqrt(s0),
                                        51 => f32::sqrt(s0),
                                        _ => todo_instr!(instruction)?,
                                    }
@@ -930,7 +929,7 @@ impl<'a> Thread<'a> {

            let op = ((instr >> 16) & 0x3ff) as u32;
            match op {
-                764 | 765 | 288 | 289 | 290 | 766 | 767 | 768 | 769 => {
+                764 | 765 | 288 | 289 | 290 | 766 | 768 | 769 => {
                    let vdst = (instr & 0xff) as usize;
                    let sdst = ((instr >> 8) & 0x7f) as usize;
                    let f = |i: u32| -> usize { ((instr >> i) & 0x1ff) as usize };
@@ -944,16 +943,6 @@ impl<'a> Thread<'a> {
                    assert_eq!(clmp, 0);

                    let vcc = match op {
-                        767 => {
-                            let (s0, s1, s2): (u32, u32, u64) = (self.val(s0), self.val(s1), self.val(s2));
-                            let (mul_result, overflow_mul) = (s0 as i64).overflowing_mul(s1 as i64);
-                            let (ret, overflow_add) = mul_result.overflowing_add(s2 as i64);
-                            let overflowed = overflow_mul || overflow_add;
-                            if self.exec.read() {
-                                self.vec_reg.write64(vdst, ret as u64);
-                            }
-                            overflowed
-                        },
                        766 => {
                            let (s0, s1, s2): (u32, u32, u64) = (self.val(s0), self.val(s1), self.val(s2));
                            let (mul_result, overflow_mul) = (s0 as u64).overflowing_mul(s1 as u64);
@@ -1257,7 +1246,7 @@ impl<'a> Thread<'a> {
                            }

                            let ret = match op {
-                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 426 | 430 | 531 | 537 | 540 | 551 | 567 | 796 => {
+                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 426 | 531 | 537 | 540 | 551 | 567 | 796 => {
                                    let s0 = f32::from_bits(s0).negate(0, neg).absolute(0, abs);
                                    let s1 = f32::from_bits(s1).negate(1, neg).absolute(1, abs);
                                    let s2 = f32::from_bits(s2).negate(2, neg).absolute(2, abs);
@@ -1269,7 +1258,6 @@ impl<'a> Thread<'a> {
                                        272 => f32::max(s0, s1),
                                        299 => f32::mul_add(s0, s1, f32::from_bits(self.vec_reg[vdst])),
                                        426 => s0.recip(),
-                                        430 => 1.0 / f32::sqrt(s0),
                                        531 => f32::mul_add(s0, s1, s2),
                                        537 => f32::min(f32::min(s0, s1), s2),
                                        540 => f32::max(f32::max(s0, s1), s2),
@@ -2637,14 +2625,6 @@ mod test_vop1 {
        assert_eq!(thread.vec_reg[3], 1071644672);
    }

-    #[test]
-    fn test_v_rsq_f32() {
-        let mut thread = _helper_test_thread();
-        thread.vec_reg[0] = f32::to_bits(4.0);
-        r(&vec![0x7E005D00, END_PRG], &mut thread);
-        assert_eq!(f32::from_bits(thread.vec_reg[0]), 0.5);
-    }
-
    #[test]
    fn test_v_frexp_exp_i32_f64() {
        [(3573412790272.0, 42), (69.0, 7), (2.0, 2), (f64::NEG_INFINITY, 0)]
@@ -6,8 +6,8 @@ from tinygrad.helpers import getenv, BEAM
 from tinygrad.engine.jit import TinyJit
 from tinygrad.engine.realize import CompiledRunner, ExecItem, ScheduleItem, lower_schedule_item, get_program
 from tinygrad.renderer import ProgramSpec
-from tinygrad.codegen.opt.kernel import Kernel, Opt, OptOps
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
+from tinygrad.opt.kernel import Kernel, Opt, OptOps
+from tinygrad.opt.heuristic import hand_coded_optimizations
 import numpy as np

 def move_jit_captured_to_dev(captured, device="DSP"):
@@ -58,7 +58,7 @@ if __name__ == "__main__":
            GlobalCounters.kernel_count -= 1

        if not getenv("NOOPT"): k.apply_opts(hand_coded_optimizations(k))
-        p2 = get_program(k.ast, k.opts, k.applied_opts)
+        p2 = get_program(k.get_optimized_ast(), k.opts)
        new_ei = replace(ei, prg=CompiledRunner(p2))
        new_ei.run()
        new_jit.append(new_ei)
@@ -4,7 +4,7 @@

 Only supported on 7900XTX, requires either AM (`rmmod amdgpu`) or disabling power gating on AMD (`ppfeaturemask=0xffff3fff`, don't forget to rebuild initramfs)

-SQTT is implemented on top of normal tinygrad profiling, `VIZ=1 SQTT=1` to get profile pickle with sqtt data embedded in it.
+SQTT is implemented on top of normal tinygrad PROFILE=1, `PROFILE=1 SQTT=1` to get profile pickle with sqtt data embedded in it.

 `SQTT_BUFFER_SIZE=X` to change size of SQTT buffer (per shader engine, 6 SEs on 7900xtx) in megabytes, default 256.

@@ -1,40 +0,0 @@
-import time
-from extra.optimization.helpers import load_worlds, ast_str_to_ast
-from tinygrad import Device
-from tinygrad.codegen.lowerer import pm_lowerer, get_index
-from tinygrad.uop.ops import graph_rewrite
-from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.postrange import Scheduler
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from tinygrad.helpers import getenv
-
-if __name__ == "__main__":
-  renderer = Device.default.renderer
-  ast_strs = load_worlds()
-  if (n:=getenv("N", -1)) != -1: ast_strs = ast_strs[n:n+1]
-  good = 0
-  for i, ast_str in enumerate(ast_strs):
-    ast = ast_str_to_ast(ast_str)
-
-    st = time.perf_counter()
-    lin = Kernel(ast, renderer)
-    opt1 = hand_coded_optimizations(lin)
-    et_lin = time.perf_counter() - st
-
-    lowered = graph_rewrite(ast, pm_lowerer, ctx=get_index(ast), bottom_up=True)
-    st = time.perf_counter()
-    sch = Scheduler(lowered, renderer)
-    sch.convert_loop_to_global()
-    sch.simplify_merge_adjacent()
-    opt2 = hand_coded_optimizations(sch)
-    et_sch = time.perf_counter() - st
-
-    if opt1 != opt2:
-      print(f"******* {i:6d}")
-      print("Kernel:    ", lin.colored_shape(), "->", lin.apply_opts(opt1).colored_shape())
-      print("Scheduler: ", sch.colored_shape(), "->", sch.apply_opts(opt2).colored_shape())
-      print(opt1)
-      print(opt2)
-    else:
-      good += 1
-      print(f"******* {i:6d} MATCH {good/(i+1)*100:.2f}% -- {et_lin/et_sch:4.2f}x speedup")
@@ -1,20 +0,0 @@
-from extra.optimization.helpers import load_worlds, ast_str_to_ast
-from tinygrad.helpers import tqdm
-from tinygrad.uop.ops import pyrender, UOp, Ops
-from tinygrad import dtypes
-from tinygrad.shape.shapetracker import ShapeTracker, View
-inf, nan = float('inf'), float('nan')
-
-if __name__ == "__main__":
-  ast_strs = load_worlds()
-  for i, ast_str in enumerate(tqdm(ast_strs)):
-    good_ast = ast_str_to_ast(ast_str)
-    code = '\n'.join(pyrender(good_ast))
-    print("\n***************\n\n"+code)
-    exec(code)
-    if str(good_ast) != str(ast):
-      print(code)
-      print("MISMATCH")
-      print(good_ast)
-      print(ast)
-      break
@@ -4,13 +4,13 @@ import struct
 import json
 import traceback
 import numpy as np
-from tinygrad.runtime.ops_cl import CLProgram, compile_gpu
+from tinygrad.runtime.ops_gpu import CLProgram, compile_gpu
 from tinygrad.device import Device
 from tinygrad.helpers import DEBUG, getenv
 from collections import defaultdict
 import pyopencl as cl
-from tinygrad.runtime.ops_cl import OSX_TIMING_RATIO
-CL = Device["CL"]
+from tinygrad.runtime.ops_gpu import OSX_TIMING_RATIO
+CL = Device["GPU"]

 DEBUGCL = getenv("DEBUGCL", 0)
 FLOAT16 = getenv("FLOAT16", 0)
@@ -110,7 +110,7 @@ class Thneed:
    prgs = {}
    for o in jdat['binaries']:
      nptr = ptr + o['length']
-      prgs[o['name']] = CLProgram(Device["CL"], o['name'], weights[ptr:nptr])
+      prgs[o['name']] = CLProgram(Device["GPU"], o['name'], weights[ptr:nptr])
      ptr = nptr

    # populate the cl_cache
@@ -267,7 +267,7 @@ class Thneed:
    for prg, args in self.cl_cache:
      events.append(prg.clprg(CL.queue, *args))
    mt = time.monotonic()
-    Device["CL"].synchronize()
+    Device["GPU"].synchronize()
    et = time.monotonic() - st
    print(f"submit in {(mt-st)*1000.0:.2f} ms, total runtime is {et*1000.0:.2f} ms")

--- a/Show More
+++ b/Show More