split pm_substitute_recurse

remote: raise runtimeerror in checkz (#12453 )
add failing test after "pend substitutes for speed" (#12457 )
2025-10-06 09:22:01 +08:00 · 2025-10-05 21:22:53 +08:00 · 2025-10-05 16:10:04 +03:00 · 2025-10-05 08:45:01 -04:00 · 2025-10-05 07:56:05 -04:00 · 2025-10-05 18:09:32 +08:00
240 changed files with 7223 additions and 5429 deletions
@@ -226,10 +226,15 @@ runs:
      if: inputs.ocelot == 'true' && runner.os == 'macOS'
      shell: bash
      run: |
-        pkgs=(cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses)
+        pkgs=(cmake ninja llvm@15 zlib glew flex bison boost@1.85 zstd ncurses)
        for f in "${pkgs[@]}"; do
          brew ls --versions "$f" >/dev/null 2>&1 || brew install --quiet "$f"
        done
+
+        # Fix boost 1.85 for gpuocelot
+        ln -s /opt/homebrew/opt/[email protected] /opt/homebrew/opt/boost || true
+        ln -s /opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib /opt/homebrew/opt/boost/lib/libboost_atomic.dylib || true
+        ln -s /opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib /opt/homebrew/opt/boost/lib/libboost_thread.dylib || true
    - name: Cache gpuocelot
      if: inputs.ocelot == 'true'
      id: cache-build
@@ -248,7 +253,13 @@ runs:
        git checkout b16039dc940dc6bc4ea0a98380495769ff35ed99
        mkdir build
        cd build
-        cmake .. -Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5
+
+        CMAKE_ARGS="-Wno-dev -G Ninja -DOCELOT_BUILD_TOOLS=OFF -DCMAKE_BUILD_ALWAYS=0 -DBUILD_TESTS_CUDA=OFF -DCMAKE_POLICY_VERSION_MINIMUM=3.5"
+        if [[ "${{ 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
        ninja
    - name: Install gpuocelot
      if: inputs.ocelot == 'true'
@@ -0,0 +1,91 @@
+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: 20
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -52,24 +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 python3.11 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      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
    - name: Run Stable Diffusion without fp16
-      run: BENCHMARK_LOG=stable_diffusion_fp32 JIT=1 python3.11 examples/stable_diffusion.py --seed 0 --noshow --timing | tee sd_no_fp16.txt
+      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
    - name: Run Stable Diffusion v2
-      run: BENCHMARK_LOG=stable_diffusion_v2 JIT=1 python3.11 examples/sdv2.py --fp16 --seed 0 --noshow --timing | tee sdv2.txt
+      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
    # process replay can't capture this, the graph is too large
    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 JIT=1 python3.11 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      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
    - name: Run model inference benchmark
      run: METAL=1 python3.11 test/external/external_model_benchmark.py
    - name: Test speed vs torch
      run: BIG=2 MPS=1 python3.11 test/speed/external_test_speed_v_torch.py | tee torch_speed.txt
    - name: Test tensor cores
-      run: METAL=1 python3.11 test/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
+      run: METAL=1 python3.11 test/opt/test_tensor_cores.py
    - name: Test AMX tensor cores
      run: |
-        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
+        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
    - name: Run Tensor Core GEMM (float)
      run: DEBUG=2 SHOULD_USE_TC=1 python3.11 extra/gemm/simple_matmul.py | tee matmul.txt
    - name: Run Tensor Core GEMM (half)
@@ -97,7 +99,7 @@ jobs:
    - name: Run GPT2
      run: |
        BENCHMARK_LOG=gpt2_nojit JIT=0 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_unjitted.txt
-        BENCHMARK_LOG=gpt2 JIT=1 python3.11 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.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
    - 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
@@ -107,21 +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 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps JIT=1 ASSERT_MIN_STEP_TIME=330 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar.txt
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half JIT=2 STEPS=10 DEFAULT_FLOAT=HALF python3.11 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      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
    #- 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 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      run: BENCHMARK_LOG=cifar_10steps_wino JIT=1 ASSERT_MIN_STEP_TIME=150 WINO=1 STEPS=10 python3.11 examples/hlb_cifar10.py | tee train_cifar_wino.txt
    - name: 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)
@@ -158,7 +160,7 @@ jobs:
  testnvidiabenchmark:
    name: tinybox green Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 30
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -194,15 +196,15 @@ jobs:
      run: NV=1 python test/external/external_benchmark_multitensor_allreduce.py
    - name: Test tensor cores
      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
-        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
+        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
    - 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 PTX=1 SHOULD_USE_TC=1 HALF=1 DEBUG=2 python3 extra/gemm/simple_matmul.py | tee matmul_ptx.txt
+      run: NV=1 NV_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
@@ -212,7 +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 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 ASSERT_MIN_STEP_TIME=2000 CAPTURE_PROCESS_REPLAY=0 NV=1 CAPTURE_PROCESS_REPLAY=0 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
    - name: Run 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
@@ -236,9 +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 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.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
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half NV=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      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
    - 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
@@ -272,7 +274,7 @@ jobs:
  testmorenvidiabenchmark:
    name: tinybox green Training Benchmark
    runs-on: [self-hosted, Linux, tinyboxgreen]
-    timeout-minutes: 20
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -300,27 +302,27 @@ 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 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 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    - name: 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 NV=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=85 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 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_half.txt
+      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
    - name: Run 10 CIFAR training steps w BF16
-      run: BENCHMARK_LOG=cifar_10steps_bf16 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py | tee train_cifar_bf16.txt
+      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
    - name: Run 10 CIFAR training steps w winograd
-      run: BENCHMARK_LOG=cifar_10steps_half_wino NV=1 CAPTURE_PROCESS_REPLAY=0 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      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
    - 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
@@ -344,7 +346,7 @@ jobs:
  testamdbenchmark:
    name: tinybox red Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 20
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -394,8 +396,8 @@ jobs:
      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/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 AMD_LLVM=0 python3 test/opt/test_tensor_cores.py
+        AMD=1 AMD_LLVM=1 python3 test/opt/test_tensor_cores.py
        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
@@ -413,9 +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 AMD=1 python3 examples/stable_diffusion.py --fp16 --seed 0 --noshow --timing | tee sd.txt
+      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
    - name: Run SDXL
-      run: BENCHMARK_LOG=stable_diffusion_xl CAPTURE_PROCESS_REPLAY=0 AMD=1 python3 examples/sdxl.py --seed 0 --noshow --timing | tee sdxl.txt
+      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
    - 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
@@ -441,9 +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 python3 examples/gpt2.py --prompt "Hello." --count 10 --temperature 0 --timing | tee gpt2_jitted.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
    - name: Run GPT2 w HALF
-      run: BENCHMARK_LOG=gpt2_half AMD=1 HALF=1 python3 examples/gpt2.py --count 10 --temperature 0 --timing | tee gpt2_half.txt
+      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
    - 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
@@ -474,7 +476,7 @@ jobs:
  testmoreamdbenchmark:
    name: tinybox red Training Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 30
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -506,19 +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 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py | tee train_cifar.txt
+      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=85 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 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
+      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
    - name: Run 10 CIFAR training steps w winograd
-      run: BENCHMARK_LOG=cifar_10steps_half_wino AMD=1 WINO=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py | tee train_cifar_wino.txt
+      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
    - 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)
@@ -537,7 +539,7 @@ jobs:
  testmlperfamdbenchmark:
    name: tinybox red MLPerf Benchmark
    runs-on: [self-hosted, Linux, tinybox]
-    timeout-minutes: 30
+    timeout-minutes: 60
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -568,10 +570,10 @@ 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
@@ -604,11 +606,11 @@ jobs:
    - 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 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
+      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 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
+      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 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
+      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
@@ -624,7 +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 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 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
        # 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
@@ -643,7 +645,7 @@ jobs:
  testreddriverbenchmark:
    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 15
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -679,8 +681,8 @@ jobs:
    # Fails on 9070
    # - name: Test tensor cores
    #   run: |
-    #     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 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 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
@@ -688,6 +690,10 @@ 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
@@ -710,7 +716,7 @@ jobs:
  testgreendriverbenchmark:
    name: NV Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
-    timeout-minutes: 15
+    timeout-minutes: 20
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -742,15 +748,19 @@ 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/test_linearizer.py TestLinearizer.test_tensor_cores TestLinearizer.test_tensor_cores_padded TestLinearizer.test_tensor_cores_padded_uops
+      run: NV=1 ALLOW_TF32=1 python3 test/opt/test_tensor_cores.py
    - 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
+    #- 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
@@ -20,12 +20,6 @@ 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
@@ -54,11 +54,12 @@ 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
+disable=C,R,W0613,W0511,W0212,W0201,W0106,W0603,W0621,W0703,W1201,W1203,E1136,W1514,E1101,W0221,W0105,E0401,abstract-method,W0707
 # 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
@@ -79,9 +79,8 @@ See [examples/beautiful_mnist.py](examples/beautiful_mnist.py) for the full vers

 tinygrad already supports numerous accelerators, including:

- [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] [OpenCL](tinygrad/runtime/ops_cl.py)
+- [x] [CPU](tinygrad/runtime/ops_cpu.py)
 - [x] [METAL](tinygrad/runtime/ops_metal.py)
 - [x] [CUDA](tinygrad/runtime/ops_cuda.py)
 - [x] [AMD](tinygrad/runtime/ops_amd.py)
@@ -80,7 +80,9 @@ 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)
@@ -93,10 +95,10 @@ out = a + b
 s = UOp(Ops.SINK, dtypes.void, (out,))

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

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

 # the first source is the output buffer (data)
 assert assign.src[0].op is Ops.BUFFER
@@ -22,12 +22,6 @@ Group UOps into kernels.

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

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

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

@@ -31,19 +31,16 @@ 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)
-GPU                 | [1]        | enable the GPU (OpenCL) backend
+CL                  | [1]        | enable 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 (Clang) backend
-LLVM                | [1]        | enable LLVM backend
+CPU                 | [1]        | enable CPU 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 | 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). |
+| 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) |
+

 ## Interoperability

@@ -70,5 +70,12 @@ 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)
@@ -2,7 +2,6 @@ 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
@@ -150,13 +149,12 @@ 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()

-  np.random.seed(1337)
+  Tensor.manual_seed(1337)
+  num_train_samples = X_train.shape[0]
+
  for epoch in range(math.ceil(hyp['misc']['train_epochs'])):
-    # TODO: move to tinygrad
    gst = time.perf_counter()
-    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
+    tidxs = Tensor.randperm(num_train_samples, dtype='int')[:num_steps_per_epoch*batchsize].reshape(num_steps_per_epoch, batchsize)
    train_loss:float = 0
    for epoch_step in (t:=trange(num_steps_per_epoch)):
      st = time.perf_counter()
@@ -26,8 +26,8 @@ class Attention:
      start_pos = start_pos.val

    if HALF: x = x.half()
-    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)]
+    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)]
    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.shrink((None, None,(start_pos,start_pos+seqlen),None,None)).assign(Tensor.stack(xk, xv)).realize()
+    self.cache_kv[:, :, start_pos:start_pos+seqlen, :, :].assign(Tensor.stack(xk, xv)).realize()

    if start_pos > 0:
-      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))
+      keys = self.cache_kv[0][:, :start_pos+seqlen, :, :]
+      values = self.cache_kv[1][:, :start_pos+seqlen, :, :]
    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)))
+    return (h + self.mlp(self.ln_2(h))).contiguous()

 class Transformer:
  def __init__(self, dim, n_heads, n_layers, norm_eps, vocab_size, max_seq_len=1024):
@@ -181,6 +181,7 @@ 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
@@ -188,7 +189,7 @@ 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 GPU" if DEBUG>=2 else "")+
+      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 "")+
                  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):
@@ -197,8 +198,13 @@ 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 ****
@@ -229,7 +229,8 @@ def train_cifar():
    if getenv("RANDOM_CROP", 1):
      X = random_crop(X, crop_size=32)
    if getenv("RANDOM_FLIP", 1):
-      X = (Tensor.rand(X.shape[0],1,1,1) < 0.5).where(X.flip(-1), X) # flip LR
+      # 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, Y = X[perms], Y[perms]
    return X, Y, *cutmix(X, Y, perms, mask_size=hyp['net']['cutmix_size'])

@@ -355,7 +356,7 @@ def train_cifar():

  # https://www.anandtech.com/show/16727/nvidia-announces-geforce-rtx-3080-ti-3070-ti-upgraded-cards-coming-in-june
  # 136 TFLOPS is the theoretical max w float16 on 3080 Ti
-
+  step_times = []
  model_ema: Optional[modelEMA] = None
  projected_ema_decay_val = hyp['ema']['decay_base'] ** hyp['ema']['every_n_steps']
  i = 0
@@ -413,12 +414,17 @@ 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 GPU" if DEBUG>=2 else "")+
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" 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 GPU" if DEBUG>=2 else "")+
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" 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 GPU" if DEBUG>=2 else "")+
+            with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" 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):

@@ -279,9 +279,15 @@ 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:
-      tok_Tens = (logits/temp).softmax().multinomial()
+      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()
    else:
      tok_Tens = logits.argmax(axis=-1).unsqueeze(0)
    tok = tok_Tens.item()
@@ -298,6 +304,7 @@ 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()
@@ -308,8 +315,9 @@ 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)
+  tinyoutput = generate(model, tokenizer, prompt, n_tokens_to_gen=num_toks, sample=sample, temp=temp, top_k=top_k)
  print(tinyoutput)
  print('TIME: ', time.time() - s)
  TORCHOUTPUT = "Why is gravity \nso important?\nBecause it's the only"
@@ -511,6 +511,33 @@ 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:
@@ -758,6 +785,27 @@ 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"
@@ -2,7 +2,9 @@ import math
 from typing import Union

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

 # rejection sampling truncated randn
 def rand_truncn(*shape, dtype=None, truncstds=2, **kwargs) -> Tensor:
@@ -17,6 +19,10 @@ 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)
@@ -127,3 +133,59 @@ 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
@@ -1,8 +1,9 @@
 import math
-from tinygrad import dtypes
+from tinygrad import dtypes, Tensor
 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):
@@ -36,4 +37,24 @@ 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)
+    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)
@@ -1,10 +1,10 @@
-import time, math
+import time, math, os
 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
+from tinygrad.helpers import getenv, Context, prod
 from extra.bench_log import BenchEvent, WallTimeEvent
 def tlog(x): print(f"{x:25s}  @ {time.perf_counter()-start:5.2f}s")

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

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

-  bs = 4
-  sequence_length = 512
+  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/"))

-  model = Transformer(**(MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]|{"vocab_size": 32000}), max_context=sequence_length, jit=False, disable_kv_cache=True)
+  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)

  @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()
+    return loss.flatten().float()

-  from examples.mlperf.dataloader import batch_load_llama3
-  iter = batch_load_llama3(bs, 5760, sequence_length, Path(getenv("BASEDIR", "/raid/datasets/c4/")), True)
+  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)

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

 if __name__ == "__main__":
  # inference only
@@ -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, safe_load, safe_save
+from tinygrad.nn.state import get_parameters, get_state_dict, load_state_dict, safe_load, safe_save
 from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, Adam, AdamW

 from extra.lr_scheduler import LRSchedulerGroup
@@ -252,6 +252,10 @@ 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:
@@ -344,6 +348,8 @@ 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
@@ -1290,12 +1296,14 @@ 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)
@@ -1311,13 +1319,14 @@ 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("DECAY_STEPS", math.ceil(1_200_000 * 1152 / GBS) - opt_learning_rate_warmup_steps)
+  opt_learning_rate_decay_steps = getenv("MAX_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 = 8e-7
+  opt_end_learning_rate = getenv("END_LR", 8e-7)

  # TODO: confirm weights are in bf16
  # vocab_size from the mixtral tokenizer
-  params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]|{"vocab_size": 32000}
+  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)

@@ -1353,6 +1362,15 @@ def train_llama3():
                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):
@@ -1403,43 +1421,55 @@ def train_llama3():
  # ** data iters **
  def fake_data(bs, samples):
    for _ in range(samples // bs):
-      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=32000, dtype=dtypes.int32, device=Device.DEFAULT)
+      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)

  def get_train_iter():
    if getenv("FAKEDATA", 0):
      return fake_data(GBS, SAMPLES)
    else:
-      from examples.mlperf.dataloader import batch_load_llama3
-      return batch_load_llama3(GBS, SAMPLES, SEQLEN, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=SEED, val=bool(TRAIN_ON_VAL))
+      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:
-      from examples.mlperf.dataloader import batch_load_llama3
-      return batch_load_llama3(EVAL_BS, 5760, SEQLEN, Path(getenv("BASEDIR", "/raid/datasets/c4/")), seed=SEED, val=True)
+      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 = 0, 0
+  i, sequences_seen = resume_ckpt, 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()
-    # above as tqdm.write f-string
+
+    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")
    if (fname:=getenv("LOSS_FILE", "")):
      with open(fname, "a") as f:
        f.write(f"{i} {loss:.4f} {lr.item():.12f} {GlobalCounters.mem_used / 1e9:.2f}\n")

-    if getenv("CKPT") and (i % 200 == 0 or i == 10):
+    if (ckpt_freq := getenv("CKPT")) and (i % ckpt_freq == 0 and (i != 1 or ckpt_freq == 1)):
      tqdm.write("saving checkpoint")
      if not os.path.exists(ckpt_dir := "./ckpts"): os.mkdir(ckpt_dir)
      fn = f"{ckpt_dir}/llama3_{i}.safe"
      safe_save(get_state_dict(model), fn)

-    i += 1
-    sequences_seen += tokens.shape[0]
+      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")
@@ -1463,6 +1493,144 @@ def train_llama3():
          safe_save(get_state_dict(model), fn)
        break

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

@@ -1471,7 +1639,7 @@ if __name__ == "__main__":
  else: bench_log_manager = contextlib.nullcontext()

  with Tensor.train():
-    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn").split(","):
+    for m in getenv("MODEL", "resnet,retinanet,unet3d,rnnt,bert,maskrcnn,stable_diffusion").split(","):
      nm = f"train_{m}"
      if nm in globals():
        print(f"training {m}")
@@ -0,0 +1,57 @@
+#!/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}  ${CKPT}/inception/pt_inception-2015-12-05-6726825d.pth"                    | sha1sum -c
@@ -0,0 +1,72 @@
+#!/usr/bin/env bash
+
+DATETIME=${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
@@ -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 GPU=1 IMAGE=2 FLOAT16=1 VIZ=1 DEBUG=2 python3 examples/openpilot/compile4.py
+# NOLOCALS=1 CL=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
+from tinygrad import nn, Tensor, Device
 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 GPU" if DEBUG>=2 else "") +
+      with Timing("enqueue in ", on_exit=(lambda et: (f", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on {Device.DEFAULT}" 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
+from tinygrad.helpers import fetch, trange, colored, Timing, getenv
 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
+import argparse, tempfile, time
 from abc import ABC, abstractmethod
 from pathlib import Path
 from PIL import Image
@@ -342,11 +342,13 @@ 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,
@@ -358,8 +360,13 @@ 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


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

  # validation!
-  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:
+  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:
    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")
@@ -2,18 +2,20 @@
 # https://github.com/ekagra-ranjan/huggingface-blog/blob/main/stable_diffusion.md
 import tempfile
 from pathlib import Path
-import argparse
+import argparse, time
 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
+from tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm, flatten
 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
+from extra.models.clip import Closed, Tokenizer, FrozenOpenClipEmbedder
+from extra.models import unet, clip
 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:
@@ -154,12 +156,46 @@ 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):
+  def __init__(self, version:str|None=None, pretrained:str|None=None):
    self.alphas_cumprod = get_alphas_cumprod()
-    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()))
+    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])

  def get_x_prev_and_pred_x0(self, x, e_t, a_t, a_prev):
    temperature = 1
@@ -266,17 +302,23 @@ 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, None))
+    x = self.token_embedding(x) + self.positional_embedding.shrink(((pos, pos+seqlen), None))
    for block in self.blocks: x = block(x, xa=encoded_audio, mask=self.mask, len=pos)
    return self.output_tok(x)

@@ -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_gpu import CLProgram, CLBuffer
+from tinygrad.runtime.ops_cl import CLProgram, CLBuffer
 from tinygrad import dtypes
 from tqdm import trange, tqdm
 from matplotlib import pyplot as plt
@@ -4,7 +4,7 @@ from tinygrad import dtypes
 from tinygrad.codegen.assembly import AssemblyCodegen, Register
 from tinygrad.codegen.opt.kernel import Ops
 from tinygrad.uop.ops import BinaryOps, UnaryOps, TernaryOps
-from tinygrad.runtime.ops_gpu import ROCM_LLVM_PATH
+from tinygrad.runtime.ops_cl 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_gpu import CLProgram, CLBuffer, ROCM_LLVM_PATH
+from tinygrad.runtime.ops_cl 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[Variable, int]):
+  def __init__(self, jit_cache: List[ExecItem], input_rawbuffers: List[Buffer], var_vals: Dict[str, 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.expr}" for v in var_vals])
+    args += sorted([f"int {v}" 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[Variable, int], wait=False):
+  def __call__(self, rawbufs: List[Buffer], var_vals: Dict[str, 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].expr)]), 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])]), 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[Variable, int]):
+  def __init__(self, jit_cache: List[ExecItem], input_rawbuffers: List[Buffer], var_vals: Dict[str, 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]])
+      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])

    # 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[Variable, int], wait=False) -> Optional[float]:
+  def __call__(self, input_rawbuffers: List[Buffer], var_vals: Dict[str, 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])
+        self.ji_kargs_structs[j].__setattr__(f'v{i}', var_vals[v.expr])

    # 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[1].startswith("lidx"):
-        r[u] = f'v{u.arg[0]}'
-      elif u.arg[1].startswith("gidx"):
-        r[u] = f's{2+u.arg[0]}'
+      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}'
      else:
        raise NotImplementedError
    elif u.uop == UOps.CONST:
@@ -10,7 +10,7 @@ from tinygrad.uop.ops import Ops
 import json
 from collections import OrderedDict

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

 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
@@ -67,11 +67,12 @@ 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)});" for (name, args, _global_size, _local_size) in statements] + ["}"]
+    cprog += [f"void net({forward_args}) {{"] + [f"{name}({', '.join(args)}, {thread_id});" for (name, args, _global_size, _local_size) in statements] + ["}"]
    return '\n'.join(headers + cprog)
  else:
    if bufs_to_save:
@@ -239,7 +240,9 @@ 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"
-  with Context(JIT=2): run,special_names = jit_model(model, *inputs)
+
+  # NOTE: CPU_COUNT=1, since export does not support threading
+  with Context(JIT=2, CPU_COUNT=1): 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()}
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 import numpy as np
-from tinygrad.runtime.ops_gpu import CLProgram, CLCompiler
+from tinygrad.runtime.ops_cl 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["GPU"]
+device = Device["CL"]

 # 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("GPU", 8, dtypes.float32).allocate()
-b = Buffer("GPU", 0x10, dtypes.float16).allocate()
-c = Buffer("GPU", 8*0x10, dtypes.float16).allocate()
+a = Buffer("CL", 8, dtypes.float32).allocate()
+b = Buffer("CL", 0x10, dtypes.float16).allocate()
+c = Buffer("CL", 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)
@@ -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 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 CUDA_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.kernel import OptOps
+from tinygrad.codegen.opt 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
@@ -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 PTX=1 python3 extra/gemm/triton_nv_matmul.py
+# CUDA=1 CUDA_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
@@ -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:(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
+  var_vals = {k.expr:(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
  dev = Device[lin.opts.device]
  root = MCTSNode(lin)

@@ -9,6 +9,9 @@ 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)
@@ -53,8 +56,8 @@ class Tokenizer:
    cs = [chr(n) for n in cs]
    return dict(zip(bs, cs))
  class ClipTokenizer:
-    def __init__(self):
-      self.byte_encoder = Tokenizer.bytes_to_unicode()
+    def __init__(self, version=None):
+      self.byte_encoder, self.version = Tokenizer.bytes_to_unicode(), version
      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]
@@ -62,11 +65,17 @@ class Tokenizer:
      vocab = vocab + [v+'</w>' for v in vocab]
      for merge in merges:
        vocab.append(''.join(merge))
-      vocab.extend(['<|startoftext|>', '<|endoftext|>'])
+      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)
      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:
@@ -110,8 +119,17 @@ class Tokenizer:

    def encode(self, text:str, pad_with_zeros:bool=False) -> List[int]:
      bpe_tokens: List[int] = []
-      text = Tokenizer.whitespace_clean(text.strip()).lower()
-      for token in re.findall(self.pat, text):
+      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):
        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.
@@ -252,10 +270,8 @@ 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

@@ -263,9 +279,10 @@ 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, Tensor.gelu, self.c_proj])
+      return x.sequential([self.c_fc, self.gelu, self.c_proj])

  # https://github.com/mlfoundations/open_clip/blob/58e4e39aaabc6040839b0d2a7e8bf20979e4558a/src/open_clip/transformer.py#L210
  class ResidualAttentionBlock:
@@ -350,15 +367,15 @@ 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):
-    self.tokenizer = Tokenizer.ClipTokenizer()
+  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)
    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.int64, device=device).reshape(1,-1)
+    return Tensor(self.tokenizer.encode(text, pad_with_zeros=True), dtype=dtypes.int32, device=device).reshape(1,-1)

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

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

-  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")))
+  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))
    for k,v in state_dict.items():
      if k.endswith(".num_batches_tracked"):
        state_dict[k] = v.reshape(1)
@@ -249,8 +249,5 @@ def convert_from_gguf(weights:dict[str, Tensor], n_layers:int):
  return sd

 def fix_bf16(weights:dict[Any, Tensor]):
-  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()}
+  # 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()}
@@ -1,21 +1,24 @@
-from tinygrad import Tensor, dtypes
-from tinygrad.nn import Linear, Conv2d, GroupNorm, LayerNorm
+from tinygrad import Tensor, dtypes, nn
 from tinygrad.device import is_dtype_supported
-from typing import Optional, Union, List, Any, Tuple
+from typing import Optional, Union, List, Any, Tuple, Callable
 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(dtypes.float16) if is_dtype_supported(dtypes.float16) else out
+  return out.cast(mixed_precision_dtype) if is_dtype_supported(mixed_precision_dtype) else out

 class ResBlock:
-  def __init__(self, channels:int, emb_channels:int, out_channels:int):
+  def __init__(self, channels:int, emb_channels:int, out_channels:int, num_groups:int=32):
    self.in_layers = [
-      GroupNorm(32, channels),
+      GroupNorm(num_groups, channels),
      Tensor.silu,
      Conv2d(channels, out_channels, 3, padding=1),
    ]
@@ -24,7 +27,7 @@ class ResBlock:
      Linear(emb_channels, out_channels),
    ]
    self.out_layers = [
-      GroupNorm(32, out_channels),
+      GroupNorm(num_groups, out_channels),
      Tensor.silu,
      lambda x: x,  # needed for weights loading code to work
      Conv2d(out_channels, out_channels, 3, padding=1),
@@ -45,35 +48,37 @@ 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 = Tensor.scaled_dot_product_attention(q, k, v).transpose(1,2)
+    attention = self.attn(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 * gate.gelu()
+    return x * self.gelu(gate)

 class FeedForward:
  def __init__(self, dim:int, mult:int=4):
-    self.net = [
+    self.net: tuple[GEGLU, Callable, nn.Linear] = (
      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(self.net)
+    return x.sequential(list(self.net))

 class BasicTransformerBlock:
  def __init__(self, dim:int, ctx_dim:int, n_heads:int, d_head:int):
@@ -92,12 +97,13 @@ 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):
+  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):
    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)
+    self.norm = GroupNorm(32, channels, eps=norm_eps)
    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)]
@@ -134,7 +140,9 @@ 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):
+  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):
    self.model_ch = model_ch
    self.num_res_blocks = [num_res_blocks] * len(channel_mult)

@@ -174,12 +182,12 @@ 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),
+          ResBlock(ch, time_embed_dim, model_ch*mult, num_groups),
        ]
        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]))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))

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

    d_head, n_heads = get_d_and_n_heads(ch)
    self.middle_block: List = [
-      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),
+      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),
    ]

    self.output_blocks = []
@@ -203,13 +211,13 @@ 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),
+          ResBlock(ch + ich, time_embed_dim, model_ch*mult, num_groups),
        ]
        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]))
+          layers.append(SpatialTransformer(ch, n_heads, d_head, ctx_dim, use_linear, depth=transformer_depth[idx], norm_eps=st_norm_eps))

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

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

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

    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 PTX=1 CUDA=1 python3 test/test_ops.py TestOps.test_tiny_add
+# IOCTL=1 CUDA=1 CUDA_PTX=1 python3 test/test_ops.py TestOps.test_tiny_add
@@ -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
-GPU=1 python3 -m pytest test/test_tiny.py
+CL=1 python3 -m pytest test/test_tiny.py

 # extract, sort and uniq
 extra/optimization/extract_dataset.py
@@ -1,6 +1,6 @@
 # stuff needed to unpack a kernel
 from tinygrad import Variable
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.dtype import dtypes, PtrDType
 from tinygrad.shape.shapetracker import ShapeTracker
@@ -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[Variable, int] = {k:int(k.vmax+k.vmin)//2 for k in lin.ast.variables()}
+  var_vals: dict[str, int] = {k.expr: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))
@@ -50,7 +50,7 @@ class TestBeamSearch(unittest.TestCase):
  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).reshape(367,367).realize()
+    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)

  def test_no_mutate_rawbuffers(self):
@@ -1,6 +1,6 @@
 import ctypes, array
 from hexdump import hexdump
-from tinygrad.runtime.ops_gpu import GPUDevice
+from tinygrad.runtime.ops_cl import CLDevice
 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 = GPUDevice()
+gdev = CLDevice()
 cl_buf = cl.clCreateBuffer(gdev.context, cl.CL_MEM_READ_WRITE, 0x100, None, status := ctypes.c_int32())
 assert status.value == 0

@@ -673,6 +673,7 @@ 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)?,
                                    }
@@ -929,7 +930,7 @@ impl<'a> Thread<'a> {

            let op = ((instr >> 16) & 0x3ff) as u32;
            match op {
-                764 | 765 | 288 | 289 | 290 | 766 | 768 | 769 => {
+                764 | 765 | 288 | 289 | 290 | 766 | 767 | 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 };
@@ -943,6 +944,16 @@ 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);
@@ -1246,7 +1257,7 @@ impl<'a> Thread<'a> {
                            }

                            let ret = match op {
-                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 426 | 531 | 537 | 540 | 551 | 567 | 796 => {
+                                257 | 259 | 299 | 260 | 261 | 264 | 272 | 392 | 426 | 430 | 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);
@@ -1258,6 +1269,7 @@ 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),
@@ -2625,6 +2637,14 @@ 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)]
@@ -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 PROFILE=1, `PROFILE=1 SQTT=1` to get profile pickle with sqtt data embedded in it.
+SQTT is implemented on top of normal tinygrad profiling, `VIZ=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.

@@ -0,0 +1,40 @@
+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")
@@ -0,0 +1,20 @@
+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_gpu import CLProgram, compile_gpu
+from tinygrad.runtime.ops_cl 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_gpu import OSX_TIMING_RATIO
-CL = Device["GPU"]
+from tinygrad.runtime.ops_cl import OSX_TIMING_RATIO
+CL = Device["CL"]

 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["GPU"], o['name'], weights[ptr:nptr])
+      prgs[o['name']] = CLProgram(Device["CL"], 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["GPU"].synchronize()
+    Device["CL"].synchronize()
    et = time.monotonic() - st
    print(f"submit in {(mt-st)*1000.0:.2f} ms, total runtime is {et*1000.0:.2f} ms")

@@ -2,7 +2,6 @@ import itertools
 from enum import Enum, auto
 from collections import defaultdict
 from typing import List, Tuple, DefaultDict
-from extra.optimization.helpers import load_worlds, ast_str_to_ast
 from tinygrad.helpers import prod, tqdm
 from tinygrad.uop.ops import UOp, Ops
 from tinygrad.shape.shapetracker import ShapeTracker
@@ -36,7 +35,7 @@ def to_movement_ops(st: ShapeTracker) -> List[Tuple[MovementOps, Tuple]]:
  to_apply:List[Tuple[MovementOps, Tuple]] = []
  for i, v in enumerate(st.views):
    real_shape = tuple(y-x for x,y in v.mask) if v.mask else v.shape
-    offset = v.offset + sum(st*(s-1) for s,st in zip(real_shape, v.strides) if st<0)
+    offset = (v.offset or 0) + sum(st*(s-1) for s,st in zip(real_shape, v.strides) if st<0)
    real_offset = offset + (sum(x*st for (x,_),st in zip(v.mask, v.strides)) if v.mask else 0)
    real_real_shape = [s for s,st in zip(real_shape, v.strides) if st]
    strides: List[int] = [abs(st) if isinstance(st,int) else st for st in v.strides if st]
@@ -121,7 +120,7 @@ def st_equivalent(st1: ShapeTracker, st2: ShapeTracker):
    if i > 1000:
      print("WARNING: did not search all possible combinations")
      break
-    var_vals = {k:v for k,v in zip(vs, ranges)}
+    var_vals = {k.expr:v for k,v in zip(vs, ranges)}
    r1 = sym_infer(idx1, var_vals) if sym_infer(valid1, var_vals) else 0
    r2 = sym_infer(idx2, var_vals) if sym_infer(valid2, var_vals) else 0
    if r1 != r2: return False
@@ -147,6 +146,7 @@ def test_rebuild_bufferop_st(ast:UOp):
  for src in ast.src: test_rebuild_bufferop_st(src)

 if __name__ == "__main__":
+  from extra.optimization.helpers import load_worlds, ast_str_to_ast
  ast_strs = load_worlds(False, False, True)[:2000]
  for ast_str in tqdm(ast_strs):
    test_rebuild_bufferop_st(ast_str_to_ast(ast_str))
@@ -177,22 +177,28 @@ def cached_to_movement_ops(shape, st) -> list:

 from tinygrad.shape.shapetracker import ShapeTracker, View
 from extra.to_movement_ops import to_movement_ops, apply_mop, MovementOps
+
+@wrap_view_op
+def _as_strided(tensor:Tensor, size, stride, storage_offset=None):
+  # multiple as_strided do not compound
+  base = canonical_base(tensor)
+  # TODO: this is heavyweight
+  st = ShapeTracker(base.uop.st.views + (View.create(tuple(size), tuple(stride), storage_offset),))
+  ret = base
+  if TORCH_DEBUG >= 1: print("**** as_strided", tensor.shape, size, stride, st)
+  if prod(size) == 1: return ret.flatten()[storage_offset].reshape(size)
+  for mo in cached_to_movement_ops(tuple(base.shape), st): ret = apply_mop(ret, mo)
+  return ret
+
@torch.library.impl("aten::as_strided", "privateuseone")
 def as_strided(tensor:torch.Tensor, size, stride, storage_offset=None):
  storage_offset = storage_offset or tensor.storage_offset()
-  @wrap_view_op
-  def _as_strided(tensor:Tensor, size, stride, storage_offset=None):
-    # multiple as_strided do not compound
-    base = canonical_base(tensor)
-    # TODO: this is heavyweight
-    st = ShapeTracker(base.uop.st.views + (View.create(tuple(size), tuple(stride), storage_offset),))
-    ret = base
-    if TORCH_DEBUG >= 1: print("**** as_strided", tensor.shape, size, stride, st)
-    if prod(size) == 1: return ret.flatten()[storage_offset].reshape(size)
-    for mo in cached_to_movement_ops(tuple(base.shape), st): ret = apply_mop(ret, mo)
-    return ret
  return _as_strided(tensor, size, stride, storage_offset)

+@torch.library.impl("aten::_reshape_alias", "privateuseone")
+def _reshape_alias(tensor:torch.Tensor, size, stride):
+  return _as_strided(tensor, size, stride)
+
@torch.library.impl("aten::empty_strided", "privateuseone")
 def empty_strided(size, stride, dtype, layout=None, device=None, pin_memory=False):
  if TORCH_DEBUG: print(f"empty_strided {size=} {stride=} {dtype=} {layout=} {device=} {pin_memory=}")
@@ -1,2 +1,6 @@
 [pytest]
 norecursedirs = extra
+timeout = 240
+timeout_method = thread
+timeout_func_only = true
+testpaths = test
@@ -9,12 +9,12 @@ with open(directory / 'README.md', encoding='utf-8') as f:

 testing_minimal = [
  "numpy",
-  "torch==2.7.1",
+  "torch==2.8.0",
  "pytest",
  "pytest-xdist",
+  "pytest-timeout",
  "hypothesis",
  "z3-solver",
-  "ml_dtypes"
 ]

 setup(name='tinygrad',
@@ -59,7 +59,7 @@ setup(name='tinygrad',
        'triton': ["triton-nightly>=2.1.0.dev20231014192330"],
        'linting': [
            "pylint",
-            "mypy==1.13.0",
+            "mypy==1.18.1",
            "typing-extensions",
            "pre-commit",
            "ruff",
@@ -87,6 +87,7 @@ setup(name='tinygrad',
            "tiktoken",
            "blobfile",
            "librosa",
+            "numba>=0.55",  # librosa needs numba but uv ignores python upper bounds and some numba versions require <python3.10
            "networkx",
            "nibabel",
            "bottle",
@@ -6,7 +6,7 @@ from tinygrad.runtime.support.hcq import HCQCompiled, HCQBuffer
 from tinygrad.runtime.autogen import libc
 from tinygrad.runtime.support.system import PCIIfaceBase
 from tinygrad.engine.realize import get_runner, CompiledRunner, get_program
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad import Variable

 MOCKGPU = getenv("MOCKGPU")
@@ -52,12 +52,12 @@ class TestHCQ(unittest.TestCase):
      with self.subTest(name=str(queue_type)):
        q = queue_type().signal(virt_signal, virt_val)

-        var_vals = {virt_signal.base_buf.va_addr: TestHCQ.d0.timeline_signal.base_buf.va_addr, virt_val: TestHCQ.d0.timeline_value}
+        var_vals = {virt_signal.base_buf.va_addr.expr: TestHCQ.d0.timeline_signal.base_buf.va_addr, virt_val.expr: TestHCQ.d0.timeline_value}
        q.submit(TestHCQ.d0, var_vals)
        TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
        TestHCQ.d0.timeline_value += 1

-        var_vals = {virt_signal.base_buf.va_addr: TestHCQ.d0.timeline_signal.base_buf.va_addr, virt_val: TestHCQ.d0.timeline_value}
+        var_vals = {virt_signal.base_buf.va_addr.expr: TestHCQ.d0.timeline_signal.base_buf.va_addr, virt_val.expr: TestHCQ.d0.timeline_value}
        q.submit(TestHCQ.d0, var_vals)
        TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
        TestHCQ.d0.timeline_value += 1
@@ -75,7 +75,7 @@ class TestHCQ(unittest.TestCase):
        TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
        TestHCQ.d0.timeline_value += 1

-  @unittest.skipIf(MOCKGPU or Device.DEFAULT in {"CPU", "LLVM"}, "Can't handle async update on MOCKGPU for now")
+  @unittest.skipIf(MOCKGPU or Device.DEFAULT in {"CPU"}, "Can't handle async update on MOCKGPU for now")
  def test_wait_late_set(self):
    for queue_type in [TestHCQ.d0.hw_compute_queue_t, TestHCQ.d0.hw_copy_queue_t]:
      if queue_type is None: continue
@@ -106,7 +106,7 @@ class TestHCQ(unittest.TestCase):

        fake_signal.value = 0x30

-        q.submit(TestHCQ.d0, {virt_signal.base_buf.va_addr: fake_signal.base_buf.va_addr, virt_val: fake_signal.value})
+        q.submit(TestHCQ.d0, {virt_signal.base_buf.va_addr.expr: fake_signal.base_buf.va_addr, virt_val.expr: fake_signal.value})
        TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
        TestHCQ.d0.timeline_value += 1

@@ -131,13 +131,13 @@ class TestHCQ(unittest.TestCase):
     .signal(TestHCQ.d0.timeline_signal, virt_val)

    for _ in range(100):
-      q.submit(TestHCQ.d0, {virt_val: TestHCQ.d0.timeline_value})
+      q.submit(TestHCQ.d0, {virt_val.expr: TestHCQ.d0.timeline_value})
      TestHCQ.d0.timeline_value += 1

    val = TestHCQ.a.uop.buffer.as_buffer().cast("f")[0]
    assert val == 200.0, f"got val {val}"

-  @unittest.skipIf(Device.DEFAULT in {"CPU", "LLVM"}, "No globals/locals on LLVM/CPU")
+  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "No globals/locals on LLVM/CPU")
  def test_exec_update(self):
    sint_global = (Variable("sint_global", 0, 0xffffffff, dtypes.uint32),) + tuple(TestHCQ.runner.p.global_size[1:])
    sint_local = (Variable("sint_local", 0, 0xffffffff, dtypes.uint32),) + tuple(TestHCQ.runner.p.local_size[1:])
@@ -146,7 +146,7 @@ class TestHCQ(unittest.TestCase):
    q.exec(TestHCQ.runner._prg, TestHCQ.kernargs_ba_ptr, sint_global, sint_local) \
     .signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)

-    q.submit(TestHCQ.d0, {sint_global[0]: 1, sint_local[0]: 1})
+    q.submit(TestHCQ.d0, {sint_global[0].expr: 1, sint_local[0].expr: 1})
    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1

@@ -155,7 +155,7 @@ class TestHCQ(unittest.TestCase):
    val = TestHCQ.b.uop.buffer.as_buffer().cast("f")[1]
    assert val == 0.0, f"got val {val}, should not be updated"

-  @unittest.skipIf(Device.DEFAULT in {"CPU", "LLVM"}, "No globals/locals on LLVM/CPU")
+  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "No globals/locals on LLVM/CPU")
  def test_exec_update_fuzz(self):
    virt_val = Variable("sig_val", 0, 0xffffffff, dtypes.uint32)
    virt_local = [Variable(f"local_{i}", 0, 0xffffffff, dtypes.uint32) for i in range(3)]
@@ -181,7 +181,7 @@ class TestHCQ(unittest.TestCase):
        for z in range(1, 4):
          ctypes.memset(zt._buf.va_addr, 0, zb.nbytes)

-          q.submit(TestHCQ.d0, {virt_val: TestHCQ.d0.timeline_value, virt_local[0]: x, virt_local[1]: y, virt_local[2]: z})
+          q.submit(TestHCQ.d0, {virt_val.expr: TestHCQ.d0.timeline_value, virt_local[0].expr: x, virt_local[1].expr: y, virt_local[2].expr: z})
          TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
          TestHCQ.d0.timeline_value += 1

@@ -253,7 +253,7 @@ class TestHCQ(unittest.TestCase):
                                    .copy(virt_dest_addr, virt_src_addr, 8) \
                                    .signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)

-    q.submit(TestHCQ.d0, {virt_src_addr: TestHCQ.a.uop.buffer._buf.va_addr, virt_dest_addr: TestHCQ.b.uop.buffer._buf.va_addr})
+    q.submit(TestHCQ.d0, {virt_src_addr.expr: TestHCQ.a.uop.buffer._buf.va_addr, virt_dest_addr.expr: TestHCQ.b.uop.buffer._buf.va_addr})

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
@@ -276,7 +276,7 @@ class TestHCQ(unittest.TestCase):
                                    .copy(virt_dest_addr, virt_src_addr, sz) \
                                    .signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value)

-    q.submit(TestHCQ.d0, {virt_src_addr: buf2._buf.va_addr, virt_dest_addr: buf1._buf.va_addr})
+    q.submit(TestHCQ.d0, {virt_src_addr.expr: buf2._buf.va_addr, virt_dest_addr.expr: buf1._buf.va_addr})

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
@@ -299,7 +299,7 @@ class TestHCQ(unittest.TestCase):

        fake_signal.value = 0x30

-        q.submit(TestHCQ.d0, {virt_signal.base_buf.va_addr: fake_signal.base_buf.va_addr, virt_val: fake_signal.value})
+        q.submit(TestHCQ.d0, {virt_signal.base_buf.va_addr.expr: fake_signal.base_buf.va_addr, virt_val.expr: fake_signal.value})
        TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
        TestHCQ.d0.timeline_value += 1

@@ -336,7 +336,7 @@ class TestHCQ(unittest.TestCase):
    et = float(sig_en.timestamp - sig_st.timestamp)

    print(f"exec kernel time: {et:.2f} us")
-    assert 0.1 <= et <= (100000 if MOCKGPU or Device.DEFAULT in {"CPU", "LLVM"} else 100)
+    assert 0.1 <= et <= (100000 if MOCKGPU or Device.DEFAULT in {"CPU"} else 100)

  def test_speed_copy_bandwidth(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -0,0 +1,13 @@
+import unittest
+from tinygrad import dtypes, Device
+from tinygrad.device import is_dtype_supported
+
+@unittest.skipUnless(Device.DEFAULT=="NULL", "Don't run when testing non-NULL backends")
+class TestNULLSupportsDTypes(unittest.TestCase):
+  def test_null_supports_ints_floats_bool(self):
+    dts = dtypes.ints + dtypes.floats + (dtypes.bool,)
+    not_supported = [dt for dt in dts if not is_dtype_supported(dt, "NULL")]
+    self.assertFalse(not_supported, msg=f"expected these dtypes to be supported by NULL: {not_supported}")
+
+if __name__ == "__main__":
+  unittest.main()
@@ -3,9 +3,9 @@ from tinygrad import Device
 from tinygrad.device import Buffer
 from tinygrad.dtype import dtypes
 from tinygrad.helpers import CI
-from tinygrad.runtime.ops_gpu import CLDevice, CLAllocator, CLCompiler, CLProgram
+from tinygrad.runtime.ops_cl import CLDevice, CLAllocator, CLCompiler, CLProgram

-@unittest.skipUnless(Device.DEFAULT == "GPU", "Runs only on OpenCL (GPU)")
+@unittest.skipUnless(Device.DEFAULT == "CL", "Runs only on OpenCL")
 class TestCLError(unittest.TestCase):
  @unittest.skipIf(CI, "dangerous for CI, it allocates tons of memory")
  def test_oom(self):
@@ -24,7 +24,7 @@ class TestCLError(unittest.TestCase):
  def test_unaligned_copy(self):
    data = list(range(65))
    unaligned = memoryview(bytearray(data))[1:]
-    buffer = Buffer("GPU", 64, dtypes.uint8).allocate()
+    buffer = Buffer("CL", 64, dtypes.uint8).allocate()
    buffer.copyin(unaligned)
    result = memoryview(bytearray(len(data) - 1))
    buffer.copyout(result)
@@ -10,10 +10,11 @@ class TestQcom(unittest.TestCase):

    def __validate(imgdt, expected_pitch):
      img = dev.allocator.alloc(imgdt.shape[0] * imgdt.shape[1] * 16, options:=BufferSpec(image=imgdt))
-      pitch = (img.descriptor[2] & 0x1fffff80) >> 7
+      pitch = img.texture_info.pitch
      assert pitch == expected_pitch, f"Failed pitch for image: {imgdt}. Got 0x{pitch:X}, expected 0x{expected_pitch:X}"
      dev.allocator.free(img, imgdt.shape[0] * imgdt.shape[1] * 16, options)

+    # Match opencl pitches for perf
    __validate(dtypes.imageh((1, 201)), 0x680)
    __validate(dtypes.imageh((16, 216)), 0x700)
    __validate(dtypes.imageh((16, 9)), 0x80)
@@ -1,7 +1,7 @@
 import random, os
 from tinygrad.helpers import Timing
 from tinygrad.runtime.ops_hip import compile_hip, HIPDevice
-from tinygrad.runtime.ops_gpu import compile_cl, CLDevice
+from tinygrad.runtime.ops_cl import compile_cl, CLDevice

 # OMP_NUM_THREADS=1 strace -tt -f -e trace=file python3 test/external/external_benchmark_hip_compile.py
 # AMD_COMGR_REDIRECT_LOGS=stdout AMD_COMGR_EMIT_VERBOSE_LOGS=1 python3 test/external/external_benchmark_hip_compile.py
@@ -27,6 +27,7 @@ if __name__ == "__main__":

  # NOTE: the inputs to a JIT must be first level arguments
  run_onnx_jit = TinyJit(lambda **kwargs: run_onnx(kwargs), prune=True)
+  step_times = []
  for _ in range(20):
    GlobalCounters.reset()
    st = time.perf_counter_ns()
@@ -35,7 +36,12 @@ if __name__ == "__main__":
      inputs = {**{k:v for k,v in new_inputs_junk.items() if 'img' in k},
                **{k:Tensor(v) for k,v in new_inputs_junk_numpy.items() if 'img' not in k}}
      ret = next(iter(run_onnx_jit(**inputs).values())).cast(dtypes.float32).numpy()
-    print(f"jitted:  {(time.perf_counter_ns() - st)*1e-6:7.4f} ms")
+      step_times.append(t:=(time.perf_counter_ns() - st)*1e-6)
+    print(f"jitted:  {t:7.4f} ms")
+
+  if (assert_time:=getenv("ASSERT_MIN_STEP_TIME")):
+    min_time = min(step_times)
+    assert min_time < assert_time, f"Speed regression, expected min step time of < {assert_time} ms but took: {min_time} ms"

  suffix = ""
  if IMAGE.value < 2: suffix += f"_image{IMAGE.value}" # image=2 has no suffix for compatibility
@@ -1,4 +1,4 @@
-from tinygrad.runtime.ops_gpu import CLDevice, CLProgram, compile_cl
+from tinygrad.runtime.ops_cl import CLDevice, CLProgram, compile_cl

 if __name__ == "__main__":
  dev = CLDevice()
@@ -1,5 +1,5 @@
 # ugh, OS X OpenCL doesn't support half
-from tinygrad.runtime.ops_gpu import CLDevice, CLProgram, CLCompiler
+from tinygrad.runtime.ops_cl import CLDevice, CLProgram, CLCompiler

 src = """#pragma OPENCL EXTENSION cl_khr_fp16 : enable
 __kernel void max_half(__global half* data0, const __global half* data1) {
@@ -0,0 +1,56 @@
+# ruff: noqa: E501
+from tinygrad import dtypes
+from tinygrad.helpers import Timing, getenv
+from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.engine.realize import get_program, CompiledRunner
+from tinygrad.uop.ops import UOp, Ops, AxisType
+
+if __name__ == "__main__":
+  if getenv("TC", 0) == 0:
+    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1179648), arg=0, src=())
+    c1 = UOp.range(UOp.const(dtypes.int, 512), 0, AxisType.GLOBAL)
+    c2 = UOp.range(UOp.const(dtypes.int, 64), 1, AxisType.GLOBAL)
+    c3 = UOp.range(UOp.const(dtypes.int, 6), 2, AxisType.GLOBAL)
+    c4 = UOp.range(UOp.const(dtypes.int, 6), 3, AxisType.GLOBAL)
+    c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(2097152), arg=1, src=())
+    c6 = UOp.range(UOp.const(dtypes.int, 64), 1004, AxisType.REDUCE)
+    c7 = UOp.range(UOp.const(dtypes.int, 3), 1005, AxisType.REDUCE)
+    c8 = UOp.range(UOp.const(dtypes.int, 3), 1006, AxisType.REDUCE)
+    c9 = c5.index(((((((c1*UOp.const(dtypes.int, 4096))+(c3*UOp.const(dtypes.int, 8)))+c4)+(c6*UOp.const(dtypes.int, 64)))+(c7*UOp.const(dtypes.int, 8)))+c8), UOp.const(dtypes.bool, True)).load()
+    c10 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(36864), arg=2, src=())
+    c11 = c10.index(((((c2*UOp.const(dtypes.int, 576))+(c6*UOp.const(dtypes.int, 9)))+(c7*UOp.const(dtypes.int, 3)))+c8), UOp.const(dtypes.bool, True)).load()
+    c12 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(64), arg=3, src=())
+    c13 = c12.index(c2, UOp.const(dtypes.bool, True)).load()
+    c14 = ((c9*c11).reduce(c6, c7, c8, arg=Ops.ADD)+c13)
+    c15 = c0.index(((((c1*UOp.const(dtypes.int, 2304))+(c2*UOp.const(dtypes.int, 36)))+(c3*UOp.const(dtypes.int, 6)))+c4), UOp.const(dtypes.bool, True)).store(c14, c1, c2, c3, c4)
+    ast = c15.sink()
+
+    # this does have tons of locals
+    opts = [Opt(op=OptOps.LOCAL, axis=1, arg=16), Opt(op=OptOps.UPCAST, axis=3, arg=0),
+            Opt(op=OptOps.LOCAL, axis=0, arg=16), Opt(op=OptOps.UPCAST, axis=3, arg=2),
+            Opt(op=OptOps.GROUPTOP, axis=0, arg=16)]
+  else:
+    c0 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(10616832), arg=0, src=())
+    c1 = UOp.range(UOp.const(dtypes.int, 512), 0, AxisType.GLOBAL)
+    c2 = UOp.range(UOp.const(dtypes.int, 64), 1, AxisType.GLOBAL)
+    c3 = UOp.range(UOp.const(dtypes.int, 36), 2, AxisType.GLOBAL)
+    c4 = UOp.range(UOp.const(dtypes.int, 9), 3, AxisType.GLOBAL)
+    c5 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(36864), arg=1, src=())
+    c6 = UOp.range(UOp.const(dtypes.int, 64), 1004, AxisType.REDUCE)
+    c7 = c5.index((((c2*UOp.const(dtypes.int, 9))+c4)+(c6*UOp.const(dtypes.int, 576))), UOp.const(dtypes.bool, True)).load()
+    c8 = UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(1179648), arg=2, src=())
+    c9 = c8.index((((c1*UOp.const(dtypes.int, 2304))+c3)+(c6*UOp.const(dtypes.int, 36))), UOp.const(dtypes.bool, True)).load()
+    c10 = (c7*c9).reduce(c6, arg=Ops.ADD)
+    c11 = c0.index(((((c1*UOp.const(dtypes.int, 20736))+(c2*UOp.const(dtypes.int, 324)))+(c3*UOp.const(dtypes.int, 9)))+c4), UOp.const(dtypes.bool, True)).store(c10, c1, c2, c3, c4)
+    ast = c11.sink()
+
+    opts = [Opt(op=OptOps.TC, axis=0, arg=(0, 0, 1)), Opt(op=OptOps.UPCAST, axis=2, arg=4),
+            Opt(op=OptOps.UPCAST, axis=3, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=0)]
+
+  prg = get_program(ast, opts=opts)
+  print(prg.src)
+  for i in range(10):
+    with Timing(f"try {i}: "):
+      # NOTE: this doesn't even run the kernel
+      try: CompiledRunner(prg)
+      except RuntimeError: pass
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 # cd extra/disassemblers/ && git clone --recursive github.com:geohot/cuda_ioctl_sniffer.git
-# LD_PRELOAD=$PWD/extra/disassemblers/cuda_ioctl_sniffer/out/sniff.so GPU=1 python3 test/external/external_multi_gpu.py
+# LD_PRELOAD=$PWD/extra/disassemblers/cuda_ioctl_sniffer/out/sniff.so CL=1 python3 test/external/external_multi_gpu.py
 import numpy as np
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import colored, Timing, getenv
@@ -1,4 +1,4 @@
-from tinygrad.runtime.ops_gpu import CLProgram, CL, CLBuffer
+from tinygrad.runtime.ops_cl import CLProgram, CL, CLBuffer
 from tinygrad import dtypes
 import time

@@ -4,7 +4,7 @@ import unittest
 import numpy as np
 if 'IMAGE' not in os.environ:
  os.environ['IMAGE'] = '2'
-os.environ['GPU'] = '1'
+os.environ['CL'] = '1'
 os.environ['OPT'] = '2'
 from tinygrad.tensor import Tensor
 from tinygrad.nn import Conv2d
@@ -193,12 +193,12 @@ backend_test.exclude('test_adam_cpu')
 backend_test.exclude('test_gradient_of_add_and_mul_cpu')
 backend_test.exclude('test_gradient_of_add_cpu')

-if Device.DEFAULT in ['GPU', 'METAL']:
+if Device.DEFAULT in ['CL', 'METAL']:
  backend_test.exclude('test_resize_upsample_sizes_nearest_axes_2_3_cpu')
  backend_test.exclude('test_resize_upsample_sizes_nearest_axes_3_2_cpu')
  backend_test.exclude('test_resize_upsample_sizes_nearest_cpu')

-if Device.DEFAULT == "METAL" or (OSX and Device.DEFAULT == "GPU"):
+if Device.DEFAULT == "METAL" or (OSX and Device.DEFAULT == "CL"):
  # numerical inaccuracy
  backend_test.exclude('test_mish_cpu')
  backend_test.exclude('test_mish_expanded_cpu')
@@ -4,7 +4,7 @@ import numpy as np
 import torch

 from tinygrad import GlobalCounters, Tensor, Device
-from tinygrad.helpers import getenv, Context
+from tinygrad.helpers import getenv, Context, RANGEIFY
 from tinygrad.nn.state import get_parameters
 from tinygrad.engine.realize import capturing
 from tinygrad.tensor import _to_np_dtype
@@ -27,14 +27,14 @@ class CLCache:
    capturing.clear()
    print(f"cache: exiting with size {self.count}", f"allowed {self.allowed}" if self.allowed is not None else "")
    if self.allowed is not None:
-      assert self.count == self.allowed, f"{self.count} != {self.allowed}"
+      assert self.count <= self.allowed, f"{self.count} > {self.allowed}"

 from extra.models.convnext import ConvNeXt
 from extra.models.efficientnet import EfficientNet
 from extra.models.resnet import ResNet18
 from extra.models.vit import ViT

-@unittest.skipUnless(Device.DEFAULT == "GPU", "Not Implemented")
+@unittest.skipUnless(Device.DEFAULT == "CL", "Not Implemented")
 class TestInferenceMinKernels(unittest.TestCase):
  def setUp(self):
    self.training_old = Tensor.training
@@ -90,7 +90,7 @@ class TestInferenceMinKernels(unittest.TestCase):
    with CLCache(100):
      model(inp, 0).realize()

-@unittest.skipUnless(Device.DEFAULT == "GPU", "Not Implemented")
+@unittest.skipUnless(Device.DEFAULT == "CL", "Not Implemented")
 class TestOptBinOp(unittest.TestCase):
  def _test_no_binop_rerun(self, f1, f2=None, allowed=1):
    a = Tensor.randn(16, 16)
@@ -117,7 +117,7 @@ class TestOptBinOp(unittest.TestCase):
  #def test_no_binop_rerun_reduce(self): return self._test_no_binop_rerun(lambda a,b: (a*b).sum(), lambda a,b: (a*b).reshape(16, 16, 1).sum())
  #def test_no_binop_rerun_reduce_alt(self): return self._test_no_binop_rerun(lambda a,b: a.sum(1)+b[0], lambda a,b: a.sum(1).reshape(1,16)+b[0])

-@unittest.skipUnless(Device.DEFAULT == "GPU", "Not Implemented")
+@unittest.skipUnless(Device.DEFAULT == "CL", "Not Implemented")
 class TestOptReduceLoop(unittest.TestCase):
  def test_loop_left(self):
    a = Tensor.randn(16, 16)
@@ -139,7 +139,7 @@ class TestOptReduceLoop(unittest.TestCase):
      c.realize()
      assert cache.count == 2, "loop right fusion broken"

-@unittest.skipUnless(Device.DEFAULT == "GPU", "Not Implemented")
+@unittest.skipUnless(Device.DEFAULT == "CL", "Not Implemented")
 class TestOptWChild(unittest.TestCase):
  @unittest.skip("this no longer happens, use realize")
  def test_unrealized_child(self):
@@ -152,7 +152,7 @@ class TestOptWChild(unittest.TestCase):
      d.realize()
      assert cache.count == 2, "don't fuse if you have children"

-@unittest.skipUnless(Device.DEFAULT == "GPU", "Not Implemented")
+@unittest.skipUnless(Device.DEFAULT == "CL", "Not Implemented")
 class TestOpt(unittest.TestCase):
  def test_muladd(self):
    a,b,c = [Tensor.randn(2,2).realize() for _ in range(3)]
@@ -164,7 +164,7 @@ class TestOpt(unittest.TestCase):

  def test_permute_was_pushed(self):
    a = Tensor.randn(16, 16, 16)
-    with CLCache(2):
+    with CLCache(1 if RANGEIFY else 2):
      c = a.sum(2)
      d = c.permute(1,0).contiguous()
      d.realize()
@@ -172,7 +172,7 @@ class TestOpt(unittest.TestCase):

  def test_permute_was_pushed_through_contract_reshape(self):
    a = Tensor.randn(4, 4, 4, 4, 4)
-    with CLCache(2):
+    with CLCache(1 if RANGEIFY else 2):
      c = a.sum(-1)
      d = c.reshape(16,16).permute(1,0).contiguous()
      d.realize()
@@ -180,7 +180,7 @@ class TestOpt(unittest.TestCase):

  def test_permute_was_pushed_through_contractw1s_reshape(self):
    a = Tensor.randn(4, 4, 4, 4, 4)
-    with CLCache(2):
+    with CLCache(1 if RANGEIFY else 2):
      c = a.sum(-1)
      d = c.reshape(16,1,16).permute(2,1,0).contiguous()
      d.realize()
@@ -188,7 +188,7 @@ class TestOpt(unittest.TestCase):

  def test_permute_was_pushed_through_expand_reshape(self):
    a = Tensor.randn(16, 16, 16)
-    with CLCache(2):
+    with CLCache(1 if RANGEIFY else 2):
      c = a.sum(2)
      d = c.reshape(4,4,4,4).permute(2,3,0,1).contiguous()
      d.realize()
@@ -221,7 +221,7 @@ class TestOpt(unittest.TestCase):
      for axis in [0, 1]:
        for n in [4, 8, 16]:
          b = torch.ones(n, n).sum(axis).reshape(n, 1).expand(n, n).sum(axis)
-          with CLCache(allowed=2):
+          with CLCache(allowed=3 if RANGEIFY else 2):
            a = Tensor.ones(n, n).contiguous().sum(axis).reshape(n, 1).expand(n, n).sum(axis)
            a.realize()
          np.testing.assert_allclose(a.numpy(), b.numpy(), rtol=1e-3, atol=1e-5)
@@ -231,7 +231,7 @@ class TestOpt(unittest.TestCase):
      axis1, axis2 = 0, 1
      for n in [4, 8, 16]:
        b = torch.ones(n, n).sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
-        with CLCache(allowed=2):
+        with CLCache(allowed=3 if RANGEIFY else 2):
          a = Tensor.ones(n, n).contiguous().sum(axis1).reshape(n, 1).expand(n, n).sum(axis2)
          a.realize()
        np.testing.assert_allclose(a.numpy(), b.numpy(), rtol=1e-3, atol=1e-5)
@@ -11,7 +11,7 @@ from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, AdamW

 from test.external.mlperf_resnet.lars_optimizer import LARSOptimizer

-from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup, CosineAnnealingLRWithWarmup
+from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup, CosineAnnealingLRWithWarmup, LambdaLR, LambdaLinearScheduler
 from test.external.mlperf_resnet.lars_util import PolynomialDecayWithWarmup as PolynomialDecayWithWarmup_tf

 np.random.seed(1337)
@@ -192,5 +192,27 @@ class TestCosineAnnealingLRWithWarmup(unittest.TestCase):
  def test_lr_1(self): self._test_lr(3e-4, 8e-5, 10, 20)
  def test_lr_llama3(self): self._test_lr(8e-5, 8e-7, 20, 100)

+class TestLambdaLRLinearWarmup(unittest.TestCase):
+  def test_linear_lr_warmup(self):
+    BS, BASE_LR = 304, 2.5e-7
+    lr = BS * BASE_LR
+    # Use a dummy Tensor parameter for optimizer because the lr_scheduler only needs the optimizer's device and lr, the params aren't touched.
+    optimizer = AdamW([Tensor([1.])])
+    lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
+    lr_scheduler = LambdaLR(optimizer, Tensor(lr, device=optimizer.device), lambda_lr_callback)
+    lrs = {}
+
+    # with above settings, optimizer.lr should warm up to lr over 1000 steps linearly
+    for i in range(1200):
+      lr_scheduler.step()
+      if i in {0, 499, 998, 999, 1000, 1199}:
+        lrs[i] = optimizer.lr.item()
+
+    np.testing.assert_allclose(lr, lrs[999], rtol=0, atol=1e-11)
+    np.testing.assert_equal(lrs[999], lrs[1000])
+    np.testing.assert_equal(lrs[999], lrs[1199])
+    np.testing.assert_allclose(lrs[999] / lrs[0], 1000, rtol=0, atol=1)
+    np.testing.assert_allclose(lrs[999] / lrs[499], 2, rtol=0, atol=1e-5)
+
 if __name__ == '__main__':
  unittest.main()
@@ -20,7 +20,7 @@ class TestLLaMASpeed(unittest.TestCase):
  def test_llama_compile(self):
    backup_program = Device[Device.DEFAULT].runtime
    backup_allocator = Device[Device.DEFAULT].allocator
-    backup_compiler = Device[Device.DEFAULT].compiler
+    backup_compiler = Device[Device.DEFAULT].compiler.compile_cached
    Device[Device.DEFAULT].runtime = FakeProgram
    Device[Device.DEFAULT].allocator = FakeAllocator(Device.default)

@@ -44,14 +44,14 @@ class TestLLaMASpeed(unittest.TestCase):
    run_llama("codegen(1)")

    # test no compiler use for this
-    Device[Device.DEFAULT].compiler = None
+    Device[Device.DEFAULT].compiler.compile_cached = None
    run_llama("methodcache", False)
    with Profiling(sort='time', frac=0.1, fn="/tmp/llama.prof", ts=5):
      run_llama("profile", False)

    Device[Device.DEFAULT].runtime = backup_program
    Device[Device.DEFAULT].allocator = backup_allocator
-    Device[Device.DEFAULT].compiler = backup_compiler
+    Device[Device.DEFAULT].compiler.compile_cached = backup_compiler

 if __name__ == '__main__':
  TestLLaMASpeed().test_llama_compile()
@@ -2,7 +2,7 @@
 import unittest

 from tinygrad.uop.ops import UOp, Ops
-from tinygrad.codegen.opt.search import Opt, OptOps
+from .search import Opt, OptOps
 from tinygrad.dtype import dtypes
 from tinygrad.shape.shapetracker import ShapeTracker
 from tinygrad.shape.view import View
@@ -1,6 +1,6 @@
 import gc
 from tinygrad import Tensor, UOp, Device
-from tinygrad.shape.shapetracker import views_to_indexed_uops
+from tinygrad.shape.shapetracker import views_to_valid_uop
 from tinygrad.engine.realize import method_cache, get_program

 def uops_allocated(): return sum([isinstance(x, UOp) for x in gc.get_objects()])
@@ -60,9 +60,10 @@ if __name__ == "__main__":

    # these caches will keep uops alive
    method_cache.clear()
-    views_to_indexed_uops.cache_clear()
+    views_to_valid_uop.cache_clear()

    new_uops = uops_allocated()
+    print_uops()
    gc.collect()
    new_uops_gc = uops_allocated()
    print(f"{t.__name__:30s}: {new_uops:3d} -> {new_uops_gc:3d}")
@@ -11,7 +11,7 @@ if __name__ == "__main__":
  for i in range(10_000):
    if i % 1000 == 0:
      print(f"Progress: {i}")
-    dt = random.choice(dtypes.ints)
+    dt = random.choice(dtypes.ints + tuple(dt.vec(4) for dt in dtypes.ints))
    u = UOp.variable('x', random.randint(dt.min, 0), random.randint(1, dt.max), dtype=dt)
    d = random.randint(1, max(1, u.arg[2]))
    if d in powers_of_two: continue
@@ -16,13 +16,13 @@ if os.getenv("VALIDATE_HCQ", 0) != 0:
  try:
    import extra.qcom_gpu_driver.opencl_ioctl
    from tinygrad import Device
-    _, _ = Device["QCOM"], Device["GPU"]
+    _, _ = Device["QCOM"], Device["CL"]
  except Exception: pass

 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.codegen.opt.kernel import Kernel
-from tinygrad.codegen.opt.kernel import Opt, OptOps
+from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.codegen.opt.search import get_kernel_actions, bufs_from_lin
 from tinygrad.engine.realize import CompiledRunner
 from tinygrad.helpers import getenv, from_mv, prod, colored, Context, DEBUG, Timing
@@ -42,9 +42,9 @@ if getenv("VALIDATE_HCQ"):
    on_linearizer_did_run = extra.nv_gpu_driver.nv_ioctl.collect_last_launch_state
    compare_states = extra.nv_gpu_driver.nv_ioctl.compare_launch_state
  elif Device.DEFAULT == "QCOM":
-    print("VALIDATE_HCQ: Comparing QCOM to GPU")
+    print("VALIDATE_HCQ: Comparing QCOM to CL")
    import extra.qcom_gpu_driver.opencl_ioctl
-    validate_device = Device["GPU"]
+    validate_device = Device["CL"]
    on_linearizer_will_run = extra.qcom_gpu_driver.opencl_ioctl.before_launch
    on_linearizer_did_run = extra.qcom_gpu_driver.opencl_ioctl.collect_last_launch_state
    compare_states = extra.qcom_gpu_driver.opencl_ioctl.compare_launch_state
@@ -90,7 +90,7 @@ def get_fuzz_rawbuf_like(old_rawbuf, zero=False, copy=False, size=None, force_de

 def run_linearizer(lin: Kernel, rawbufs=None, var_vals=None) -> tuple[str, Any]: # (error msg, run state)
  if rawbufs is None: rawbufs = bufs_from_lin(lin)
-  if var_vals is None: var_vals = {v: v.min for v in lin.vars}
+  if var_vals is None: var_vals = {v.expr: v.min for v in lin.vars}

  # TODO: images needs required_optimization
  try:
@@ -129,7 +129,7 @@ def compare_linearizer(lin: Kernel, rawbufs=None, var_vals=None, ground_truth=No

  if var_vals is None:
    # TODO: handle symbolic max case
-    var_vals = {v: random.randint(v.vmin, v.vmax) for v in lin.ast.variables()}
+    var_vals = {v.expr: random.randint(v.vmin, v.vmax) for v in lin.ast.variables()}

  if ground_truth is None and not has_bf16:
    unoptimized = Kernel(lin.ast)
@@ -302,7 +302,7 @@ if __name__ == "__main__":
    for i, ast in enumerate(ast_strs[:getenv("FUZZ_N", len(ast_strs))]):
      if (nth := getenv("FUZZ_NTH", -1)) != -1 and i != nth: continue
      if getenv("FUZZ_IMAGEONLY") and "dtypes.image" not in ast: continue
-      if "dtypes.image" in ast and Device.DEFAULT not in {"GPU", "QCOM"}: continue  # IMAGE is only for GPU
+      if "dtypes.image" in ast and Device.DEFAULT not in {"CL", "QCOM"}: continue  # IMAGE is only for CL
      if ast in seen_ast_strs: continue
      seen_ast_strs.add(ast)

@@ -0,0 +1,59 @@
+import unittest, os
+import numpy as np
+from pathlib import Path
+from tempfile import TemporaryDirectory
+from tinygrad import Device, Tensor
+from tinygrad.helpers import getenv, Context
+from tinygrad.nn.state import safe_save, torch_load, get_parameters
+from examples.mlperf.model_eval import eval_stable_diffusion, vae_decode
+from examples.stable_diffusion import AutoencoderKL
+
+def set_eval_params():
+  # override these as needed from cli
+  for k,v in {"MODEL": "stable_diffusion", "GPUS": "8", "EVAL_SAMPLES": "600", "CONTEXT_BS": "816", "DENOISE_BS": "600", "DECODE_BS": "384",
+   "INCEPTION_BS": "560", "CLIP_BS": "240", "DATADIR": "/raid/datasets/stable_diffusion", "CKPTDIR": "/raid/weights/stable_diffusion",
+   "AMD_LLVM": "0"}.items():
+    os.environ[k] = getenv(k, v)
+
+class TestEval(unittest.TestCase):
+  def test_eval_ckpt(self):
+    set_eval_params()
+    with TemporaryDirectory(prefix="test-eval") as tmp:
+      os.environ["EVAL_CKPT_DIR"] = tmp
+      # NOTE Although this checkpoint has the original fully trained model from StabilityAI, we are using mlperf code that uses different
+      #   GroupNorm num_groups. Therefore, eval results may not reflect eval results on the original model.
+      # The purpose of using this checkpoint is to have reproducible eval outputs.
+      # Eval code expects file and weight names in a specific format, as .safetensors (not .ckpt), which is why we resave the checkpoint
+      sd_v2 = torch_load(Path(getenv("CKPTDIR", "")) / "sd" / "512-base-ema.ckpt")["state_dict"]
+      sd_v2 = {k.replace("model.diffusion_model.", "", 1): v for k,v in sd_v2.items() if k.startswith("model.diffusion_model.")}
+      safe_save(sd_v2, f"{tmp}/0.safetensors")
+      clip, fid, ckpt = eval_stable_diffusion()
+    assert ckpt == 0
+    if Device.DEFAULT == "NULL":
+      assert clip == 0
+      assert fid > 0 and fid < 1000
+    else:
+      # observed:
+      # clip=0.08369670808315277, fid=301.05236173709545 (if SEED=12345, commit=c01b2c93076e80ae6d1ebca64bb8e83a54dadba6)
+      # clip=0.08415728807449341, fid=300.3710877072948 (if SEED=12345, commit=179c7fcfe132f1a6344b57c9d8cef4eded586867)
+      # clip=0.0828116238117218, fid=301.241909555543 (if SEED=98765, commit=c01b2c93076e80ae6d1ebca64bb8e83a54dadba6)
+      np.testing.assert_allclose(fid, 301.147, rtol=0.1, atol=0)
+      np.testing.assert_allclose(clip, 0.08325, rtol=0.1, atol=0)
+
+  # only tested on 8xMI300x system
+  @unittest.skipUnless(getenv("HANG_OK"), "expected to hang")
+  def test_decoder_beam_hang(self):
+    set_eval_params()
+    for k,v in {"BEAM": "2", "HCQDEV_WAIT_TIMEOUT_MS": "300000", "BEAM_UOPS_MAX": "8000", "BEAM_UPCAST_MAX": "256", "BEAM_LOCAL_MAX": "1024",
+                "BEAM_MIN_PROGRESS": "5", "IGNORE_JIT_FIRST_BEAM": "1"}.items():
+      os.environ[k] = getenv(k, v)
+    with Context(BEAM=int(os.environ["BEAM"])): # necessary because helpers.py has already set BEAM=0 and cached getenv for "BEAM"
+      GPUS = [f"{Device.DEFAULT}:{i}" for i in range(getenv("GPUS", 8))]
+      vae = AutoencoderKL()
+      for p in get_parameters(vae): p.to_(GPUS).realize()
+      x = Tensor.zeros(48,4,64,64).contiguous().to(GPUS).realize()
+      x.uop = x.uop.multi(0)
+      for _ in range(2): vae_decode(x, vae)
+
+if __name__=="__main__":
+  unittest.main()
@@ -0,0 +1,114 @@
+import unittest
+import numpy as np
+from pathlib import Path
+from tinygrad import Tensor, dtypes, Device
+from tinygrad.helpers import getenv
+from tinygrad.nn.state import get_parameters
+from extra.models import clip
+from examples.mlperf.initializers import gelu_erf, init_stable_diffusion, attn_f32_softmax
+from typing import Literal
+
+clip_params = {"dims": 1024, "n_heads": 16, "layers": 24, "return_pooled": False, "ln_penultimate": True, "clip_tokenizer_version": "sd_mlperf_v5_0"}
+def get_cond_stage_model(GPUS:list[str]|None=None) -> clip.FrozenOpenClipEmbedder:
+  clip.gelu = gelu_erf
+  model = clip.FrozenOpenClipEmbedder(**clip_params)
+  if GPUS and len(GPUS) > 1:
+    for p in get_parameters(model): p.to_(GPUS)
+  return model
+def get_tokens(BS:int) -> Tensor: return Tensor([0] * 77 * BS, dtype=dtypes.int32).reshape(-1, 77)
+
+class TestOpenClip(unittest.TestCase):
+  def test_tokenizer(self):
+    prompt = "Beautiful is better than ugly.\nExplicit is better than implicit.\nSimple is better than complex.\nComplex is better than complicated."
+    model = get_cond_stage_model()
+    tokens = model.tokenizer.encode(prompt, pad_with_zeros=True)
+    expected = [49406, 1215, 533, 1539, 1126, 8159, 269, 33228, 533, 1539, 1126, 15269, 585, 269, 4129, 533, 1539, 1126, 6324, 269, 6324, 533,
+                1539, 1126, 16621, 269, 49407] + [0]*50
+    self.assertEqual(tokens, expected)
+
+  def test_clip_gelu_init(self):
+    for resblock in get_cond_stage_model().model.transformer.resblocks:
+      self.assertEqual(resblock.mlp.gelu, gelu_erf)
+
+  def test_multigpu_clip_embed(self):
+    BS = 304
+    GPUS = [f"{Device.DEFAULT}:{i}" for i in range(8)]
+    model = get_cond_stage_model(GPUS)
+    tokens = get_tokens(BS)
+    embeds = model.embed_tokens(tokens.shard(GPUS, axis=0)).realize()
+    self.assertEqual(embeds.shape, (BS, 77, 1024))
+    self.assertEqual(embeds.dtype, dtypes.float32)
+
+  def test_multigpu_clip_score(self):
+    BS = 240
+    GPUS = [f"{Device.DEFAULT}:{i}" for i in range(8)]
+    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 = clip.OpenClipEncoder(1024, text_cfg, vision_cfg)
+    for p in get_parameters(clip_encoder): p.to_(GPUS)
+    tokens = get_tokens(BS)
+    imgs = Tensor.zeros(BS,3,224,224).contiguous()
+    scores = clip_encoder.get_clip_score(tokens.shard(GPUS, axis=0), imgs.shard(GPUS, axis=0)).realize()
+    self.assertEqual(scores.shape, (BS,))
+    self.assertEqual(scores.dtype, dtypes.float32)
+
+class TestInitStableDiffusion(unittest.TestCase):
+  def setUp(self):
+    # NOTE: set env variable based on where checkpoints are on the system
+    self.CKPTDIR = Path(getenv("CKPTDIR", "/raid/weights/stable_diffusion"))
+
+  def helper_test_init(self, version:Literal["v2-mlperf-train", "v2-mlperf-eval"]):
+    model, unet, sqrt_acp, sqrt_omacp = init_stable_diffusion(version, self.CKPTDIR / "sd" / "512-base-ema.ckpt", ["CPU"])
+
+    with self.subTest("test that StableDiffusion has correct models"):
+      self.assertEqual(model.model.diffusion_model, unet)
+      has_encoder = True if version=="v2-mlperf-eval" else False
+      self.assertEqual(hasattr(model, "first_stage_model"), has_encoder, "only the eval model uses the encoder")
+      self.assertTrue(isinstance(model.cond_stage_model, clip.FrozenOpenClipEmbedder))
+
+    with self.subTest("test for mlperf unique attributes"):
+      self.assertEqual(model.cond_stage_model.tokenizer.version, 'sd_mlperf_v5_0')
+      self.assertEqual(unet.out[0].num_groups, 16)
+      self.assertEqual(unet.input_blocks[1][1].norm.eps, 1e-6)
+      self.assertEqual(unet.input_blocks[1][1].transformer_blocks[0].attn1.attn, attn_f32_softmax)
+
+    with self.subTest("test loaded clip parameters"):
+      sample = model.cond_stage_model.model.transformer.resblocks[8].mlp.c_fc.bias.flatten()[42:46].numpy()
+      expected = np.array([-0.49812260270118713, -0.3039605915546417, -0.40284937620162964, -0.45069342851638794], dtype=np.float32)
+      np.testing.assert_allclose(sample, expected, rtol=1e-7, atol=0, err_msg="loaded clip parameters are incorrect")
+
+    if version=="v2-mlperf-train":
+      with self.subTest("test that zero_module worked"):
+        self.assertTrue((unet.out[2].weight == 0).all().item(), "expected all zeroes")
+        self.assertTrue((unet.out[2].bias == 0).all().item(), "expected all zeroes")
+    elif version=="v2-mlperf-eval":
+      with self.subTest("test loaded vae parameters"):
+        sample = model.first_stage_model.decoder.up[0]['block'][1].conv2.weight.flatten()[42:46].numpy()
+        expected = np.array([0.08192943036556244, 0.040095631033182144, 0.07541035860776901, 0.1475081741809845], dtype=np.float32)
+        np.testing.assert_allclose(sample, expected, rtol=1e-7, atol=0, err_msg="loaded vae parameters are incorrect")
+
+    with self.subTest("check schedules"):
+      expected = np.array([0.9995748996734619, 0.06826484948396683], dtype=np.float32)
+      np.testing.assert_allclose(sqrt_acp[[0,-1]].numpy(), expected, rtol=1e-7, atol=0, err_msg="sqrt_acp is incorrect")
+      expected = np.array([0.029155133292078972, 0.9976672530174255], dtype=np.float32)
+      np.testing.assert_allclose(sqrt_omacp[[0,-1]].numpy(), expected, rtol=1e-7, atol=0, err_msg="sqrt_omacp is incorrect")
+
+    with self.subTest("check mixed precision"):
+      out = unet.input_blocks[2][1].proj_in(Tensor.randn(320, dtype=dtypes.float32))
+      self.assertEqual(out.dtype, dtypes.bfloat16, "expected float32 to be downcast to bfloat16 by Linear")
+      out = unet.out[2](Tensor.randn(304,320,64,64, dtype=dtypes.float32))
+      self.assertEqual(out.dtype, dtypes.bfloat16, "expected float32 to be downcast to bfloat16 by Conv2d")
+      out = unet.input_blocks[1][1].transformer_blocks[0].norm1(Tensor.randn(320, dtype=dtypes.bfloat16))
+      self.assertEqual(out.dtype, dtypes.float32, "expected bfloat16 to be upcast to float32 by LayerNorm")
+      out = unet.input_blocks[5][0].in_layers[0](Tensor.randn(304, 640, dtype=dtypes.bfloat16))
+      self.assertEqual(out.dtype, dtypes.float32, "expected bfloat16 to be upcast to float32 by GroupNorm")
+
+  def test_train_model(self):
+    self.helper_test_init("v2-mlperf-train")
+
+  def test_eval_model(self):
+    self.helper_test_init("v2-mlperf-eval")
+
+if __name__=="__main__":
+  unittest.main()
@@ -0,0 +1,23 @@
+import unittest, os
+from tempfile import TemporaryDirectory
+from tinygrad import Tensor
+from tinygrad.helpers import getenv
+from examples.mlperf.model_train import train_stable_diffusion
+
+class TestTrain(unittest.TestCase):
+  def test_train_to_ckpt(self):
+    # train for num_steps, save checkpoint, and stop training
+    num_steps = 42
+    os.environ.update({"MODEL": "stable_diffusion", "TOTAL_CKPTS": "1", "CKPT_STEP_INTERVAL": str(num_steps), "GPUS": "8", "BS": "304"})
+    # NOTE: update these based on where data/checkpoints are on your system
+    if not getenv("DATADIR", ""): os.environ["DATADIR"] = "/raid/datasets/stable_diffusion"
+    if not getenv("CKPTDIR", ""): os.environ["CKPTDIR"] = "/raid/weights/stable_diffusion"
+    with TemporaryDirectory(prefix="test-train") as tmp:
+      os.environ["UNET_CKPTDIR"] = tmp
+      with Tensor.train():
+        saved_ckpts = train_stable_diffusion()
+      expected_ckpt = f"{tmp}/{num_steps}.safetensors"
+      assert len(saved_ckpts) == 1 and saved_ckpts[0] == expected_ckpt
+
+if __name__=="__main__":
+  unittest.main()
@@ -12,7 +12,7 @@ try:
  from tinygrad.renderer import Renderer, ProgramSpec
  from tinygrad.engine.realize import get_program
  from tinygrad.uop.ops import UOp, Ops, KernelInfo
-  from tinygrad.codegen.opt.kernel import Opt
+  from tinygrad.codegen.opt import Opt
  from tinygrad.helpers import VERSION, Context, ContextVar, colored, db_connection, getenv, tqdm
  from tinygrad.device import Device
 except ImportError as e:
@@ -1,41 +0,0 @@
-from tinygrad import Device
-from tinygrad.helpers import getenv, DEBUG, BEAM
-from tinygrad.codegen.opt.search import beam_search, bufs_from_lin
-from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
-from extra.optimization.helpers import load_worlds, ast_str_to_lin, time_linearizer
-
-if __name__ == "__main__":
-  filter_reduce = bool(getenv("FILTER_REDUCE"))
-  ast_strs = load_worlds(filter_reduce=filter_reduce, filter_novariable=True)
-  dev = Device[Device.DEFAULT]
-
-  test_n = getenv("TEST_N", 10)
-  single = getenv("NUM", -1)
-  if single != -1: ast_strs = ast_strs[single:single+1]
-
-  beam_won, tested = 0, 0
-
-  for num, ast in enumerate(ast_strs[:test_n]):
-    def new_lin(): return ast_str_to_lin(ast, opts=dev.renderer)
-
-    k = new_lin()
-
-    if not (used_tensor_cores:=k.apply_tensor_cores(getenv("TC", 1))): k.apply_opts(hand_coded_optimizations(k))
-
-    assert BEAM > 0
-
-    lins = [(("tc" if used_tensor_cores else "hc"), k)]
-    if used_tensor_cores:
-      lins.append(("hc", new_lin()))
-      lins[-1][1].apply_opts(hand_coded_optimizations(lins[-1][1]))
-    kb = new_lin()
-    test_rawbuffers = bufs_from_lin(kb)    # allocate scratch buffers for optimization
-    lins.append((f"beam{BEAM.value}", beam_search(kb, test_rawbuffers, BEAM.value, bool(getenv("BEAM_ESTIMATE", 1)))))
-    timed = sorted([(nm, tk, time_linearizer(tk, test_rawbuffers, allow_test_size=False, clear_l2=True)) for nm, tk in lins], key=lambda x: x[2])
-    if DEBUG >= 1: print("  <  ".join(f"{nm:6s} : {lin.colored_shape(30, dense=True)} : {tm*1e6:8.2f} us" for nm, lin, tm in timed))
-
-    tested += 1
-    if timed[0][0].startswith("beam"):
-      beam_won += 1
-
-  print(f"{beam_won=} / {tested=} = {beam_won/tested:.3f}")
@@ -57,8 +57,8 @@ def eval_uop(uop:UOp, inputs:list[tuple[DType, list[Any]]]|None=None):
  return out_buf.cast(uop.dtype.fmt).tolist()[0]

 def not_support_multi_device():
-  # GPU and CUDA don't support multi device if in CI
-  return CI and REAL_DEV in ("GPU", "CUDA")
+  # CL and CUDA don't support multi device if in CI
+  return CI and REAL_DEV in ("CL", "CUDA")

 # NOTE: This will open REMOTE if it's the default device
 REAL_DEV = (Device.DEFAULT if Device.DEFAULT != "REMOTE" else Device['REMOTE'].properties.real_device)
@@ -128,7 +128,10 @@ def cuModuleUnload(hmod) -> int:
 def cuLaunchKernel(f, gx: int, gy: int, gz: int, lx: int, ly: int, lz: int, sharedMemBytes: int,
                   hStream: Any, kernelParams: Any, extra: Any) -> int:
  cargs = [ctypes.cast(getattr(extra, field[0]), ctypes.c_void_p) for field in extra._fields_]
-  gpuocelot_lib.ptx_run(ctypes.cast(f.value, ctypes.c_char_p), len(cargs), (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+  try: gpuocelot_lib.ptx_run(ctypes.cast(f.value, ctypes.c_char_p), len(cargs), (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+  except Exception as e:
+    print("Error in cuLaunchKernel:", e)
+    return orig_cuda.CUDA_ERROR_LAUNCH_FAILED
  return orig_cuda.CUDA_SUCCESS

 def cuDeviceComputeCapability(major, minor, dev: int) -> int:
@@ -97,7 +97,10 @@ class GPFIFO:
    cargs = [ctypes.cast(args[i], ctypes.c_void_p) for i in range(args_cnt)] + [ctypes.cast(vals[i], ctypes.c_void_p) for i in range(vals_cnt)]
    gx, gy, gz = qmd.cta_raster_width, qmd.cta_raster_height, qmd.cta_raster_depth
    lx, ly, lz = qmd.cta_thread_dimension0, qmd.cta_thread_dimension1, qmd.cta_thread_dimension2
-    gpuocelot_lib.ptx_run(ctypes.cast(prg_addr, ctypes.c_char_p), args_cnt+vals_cnt, (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+    try:
+      gpuocelot_lib.ptx_run(ctypes.cast(prg_addr, ctypes.c_char_p), args_cnt+vals_cnt,
+        (ctypes.c_void_p*len(cargs))(*cargs), lx, ly, lz, gx, gy, gz, 0)
+    except Exception as e: print("failed to execute:", e)
    if qmd.release0_enable:
      rel0 = to_mv(qmd.release0_address_lower + (qmd.release0_address_upper << 32), 0x10).cast('Q')
      rel0[0] = qmd.release0_payload_lower + (qmd.release0_payload_upper << 32)
@@ -1,14 +1,14 @@
 #!/usr/bin/env python
 import unittest
+from tinygrad import Tensor
 import numpy as np
-from tinygrad.tensor import Tensor
 import torch

 def get_question_samp(bsz, seq_len, vocab_size, seed):
  np.random.seed(seed)
  in_ids= np.random.randint(vocab_size, size=(bsz, seq_len))
  mask = np.random.choice([True, False], size=(bsz, seq_len))
-  seg_ids = np.random.randint(1, size=(bsz, seq_len))
+  seg_ids = np.random.randint(2, size=(bsz, seq_len))  # type_vocab_size
  return in_ids, mask, seg_ids

 def set_equal_weights(mdl, torch_mdl):
@@ -45,7 +45,7 @@ class TestBert(unittest.TestCase):

    seeds = (1337, 3141)
    bsz, seq_len = 1, 16
-    for _, seed in enumerate(seeds):
+    for seed in seeds:
      in_ids, mask, seg_ids = get_question_samp(bsz, seq_len, config['vocab_size'], seed)
      out = mdl(Tensor(in_ids), Tensor(mask), Tensor(seg_ids))
      torch_out = torch_mdl.forward(torch.from_numpy(in_ids).long(), torch.from_numpy(mask), torch.from_numpy(seg_ids).long())[:2]
@@ -1,12 +1,10 @@
-import ast
-import pathlib
-import unittest
+import ast, pathlib, unittest

 import numpy as np
 from PIL import Image

-from tinygrad.helpers import getenv
-from tinygrad.tensor import Tensor
+from tinygrad import Tensor
+from tinygrad.helpers import getenv, CI
 from extra.models.efficientnet import EfficientNet
 from extra.models.vit import ViT
 from extra.models.resnet import ResNet50
@@ -40,19 +38,13 @@ def preprocess(img, new=False):
    img /= np.array([0.229, 0.224, 0.225]).reshape((1, -1, 1, 1))
  return img

+def _infer(model: EfficientNet, img):
+  with Tensor.train(False):
+    out = model.forward(Tensor(img)).argmax(axis=-1)
+  return out.tolist()

-def _infer(model: EfficientNet, img, bs=1):
-  old_training = Tensor.training
-  Tensor.training = False
-  img = preprocess(img)
-  # run the net
-  if bs > 1: img = img.repeat(bs, axis=0)
-  out = model.forward(Tensor(img))
-  Tensor.training = old_training
-  return _LABELS[np.argmax(out.numpy()[0])]
-
-chicken_img = Image.open(pathlib.Path(__file__).parent / 'efficientnet/Chicken.jpg')
-car_img = Image.open(pathlib.Path(__file__).parent / 'efficientnet/car.jpg')
+chicken_img = preprocess(Image.open(pathlib.Path(__file__).parent / 'efficientnet/Chicken.jpg'))
+car_img = preprocess(Image.open(pathlib.Path(__file__).parent / 'efficientnet/car.jpg'))

 class TestEfficientNet(unittest.TestCase):
  @classmethod
@@ -64,17 +56,20 @@ class TestEfficientNet(unittest.TestCase):
  def tearDownClass(cls):
    del cls.model

+  @unittest.skipIf(CI, "covered by test_chicken_car")
  def test_chicken(self):
-    label = _infer(self.model, chicken_img)
-    self.assertEqual(label, "hen")
-
-  def test_chicken_bigbatch(self):
-    label = _infer(self.model, chicken_img, 2)
-    self.assertEqual(label, "hen")
+    labels = _infer(self.model, chicken_img)
+    self.assertEqual(_LABELS[labels[0]], "hen")

+  @unittest.skipIf(CI, "covered by test_chicken_car")
  def test_car(self):
-    label = _infer(self.model, car_img)
-    self.assertEqual(label, "sports car, sport car")
+    labels = _infer(self.model, car_img)
+    self.assertEqual(_LABELS[labels[0]], "sports car, sport car")
+
+  def test_chicken_car(self):
+    labels = _infer(self.model, np.concat([chicken_img, car_img], axis=0))
+    self.assertEqual(_LABELS[labels[0]], "hen")
+    self.assertEqual(_LABELS[labels[1]], "sports car, sport car")

 class TestViT(unittest.TestCase):
  @classmethod
@@ -87,12 +82,12 @@ class TestViT(unittest.TestCase):
    del cls.model

  def test_chicken(self):
-    label = _infer(self.model, chicken_img)
-    self.assertEqual(label, "cock")
+    labels = _infer(self.model, chicken_img)
+    self.assertEqual(_LABELS[labels[0]], "cock")

  def test_car(self):
-    label = _infer(self.model, car_img)
-    self.assertEqual(label, "racer, race car, racing car")
+    labels = _infer(self.model, car_img)
+    self.assertEqual(_LABELS[labels[0]], "racer, race car, racing car")

 class TestResNet(unittest.TestCase):
  @classmethod
@@ -105,12 +100,12 @@ class TestResNet(unittest.TestCase):
    del cls.model

  def test_chicken(self):
-    label = _infer(self.model, chicken_img)
-    self.assertEqual(label, "hen")
+    labels = _infer(self.model, chicken_img)
+    self.assertEqual(_LABELS[labels[0]], "hen")

  def test_car(self):
-    label = _infer(self.model, car_img)
-    self.assertEqual(label, "sports car, sport car")
+    labels = _infer(self.model, car_img)
+    self.assertEqual(_LABELS[labels[0]], "sports car, sport car")

 if __name__ == '__main__':
  unittest.main()
@@ -1,23 +1,12 @@
 #!/usr/bin/env python
-import os
-import time
 import unittest
 import numpy as np
-try:
-  import onnx
-except ModuleNotFoundError:
-  raise unittest.SkipTest("onnx not installed, skipping onnx test")
 from tinygrad.frontend.onnx import OnnxRunner
-from tinygrad.tensor import Tensor
 from tinygrad.device import Device
-from tinygrad.helpers import CI, fetch, temp, Context
+from tinygrad.helpers import fetch, Context

-try:
-  from extra.onnx_helpers import validate
-  from extra.huggingface_onnx.huggingface_manager import DOWNLOADS_DIR, snapshot_download_with_retry
-  HUGGINGFACE_AVAILABLE = True
-except ModuleNotFoundError:
-  HUGGINGFACE_AVAILABLE = False
+from extra.onnx_helpers import validate
+from extra.huggingface_onnx.huggingface_manager import DOWNLOADS_DIR, snapshot_download_with_retry

 def run_onnx_torch(onnx_model, inputs):
  import torch
@@ -27,86 +16,9 @@ def run_onnx_torch(onnx_model, inputs):
    torch_out = torch_model(*[torch.tensor(x) for x in inputs.values()])
  return torch_out

-OPENPILOT_MODEL = "https://github.com/commaai/openpilot/raw/v0.9.4/selfdrive/modeld/models/supercombo.onnx"
-
 np.random.seed(1337)

 class TestOnnxModel(unittest.TestCase):
-  @unittest.skip("this isn't a test, it can't fail")
-  def test_benchmark_openpilot_model(self):
-    onnx_model = fetch(OPENPILOT_MODEL)
-    run_onnx = OnnxRunner(onnx_model)
-    def get_inputs():
-      np_inputs = {
-        "input_imgs": np.random.randn(*(1, 12, 128, 256)),
-        "big_input_imgs": np.random.randn(*(1, 12, 128, 256)),
-        "desire": np.zeros((1, 100, 8)),
-        "traffic_convention": np.array([[1., 0.]]),
-        "nav_features": np.zeros((1, 256)),
-        "features_buffer": np.zeros((1, 99, 128)),
-    }
-      inputs = {k:Tensor(v.astype(np.float32), requires_grad=False) for k,v in np_inputs.items()}
-      return inputs
-
-    for _ in range(7):
-      inputs = get_inputs()
-      st = time.monotonic()
-      tinygrad_out = run_onnx(inputs)['outputs']
-      mt = time.monotonic()
-      tinygrad_out.realize()
-      mt2 = time.monotonic()
-      tinygrad_out = tinygrad_out.numpy()
-      et = time.monotonic()
-      if not CI:
-        print(f"ran openpilot model in {(et-st)*1000.0:.2f} ms, waited {(mt2-mt)*1000.0:.2f} ms for realize, {(et-mt2)*1000.0:.2f} ms for GPU queue")
-
-    if not CI:
-      import cProfile
-      import pstats
-      inputs = get_inputs()
-      pr = cProfile.Profile(timer=time.perf_counter_ns, timeunit=1e-6)
-      pr.enable()
-    tinygrad_out = run_onnx(inputs)['outputs']
-    tinygrad_out.realize()
-    tinygrad_out = tinygrad_out.numpy()
-    if not CI:
-      pr.disable()
-      stats = pstats.Stats(pr)
-      stats.dump_stats(temp("net.prof"))
-      os.system(f"flameprof {temp('net.prof')} > {temp('prof.svg')}")
-      ps = stats.sort_stats(pstats.SortKey.TIME)
-      ps.print_stats(30)
-
-  def test_openpilot_model(self):
-    onnx_model = fetch(OPENPILOT_MODEL)
-    run_onnx = OnnxRunner(onnx_model)
-    print("got run_onnx")
-    inputs = {
-      "input_imgs": np.random.randn(*(1, 12, 128, 256)),
-      "big_input_imgs": np.random.randn(*(1, 12, 128, 256)),
-      "desire": np.zeros((1, 100, 8)),
-      "traffic_convention": np.array([[1., 0.]]),
-      "nav_features": np.zeros((1, 256)),
-      "features_buffer": np.zeros((1, 99, 128)),
-    }
-    inputs = {k:v.astype(np.float32) for k,v in inputs.items()}
-
-    st = time.monotonic()
-    print("****** run onnx ******")
-    tinygrad_out = run_onnx(inputs)['outputs']
-    mt = time.monotonic()
-    print("****** realize ******")
-    tinygrad_out.realize()
-    mt2 = time.monotonic()
-    tinygrad_out = tinygrad_out.numpy()
-    et = time.monotonic()
-    print(f"ran openpilot model in {(et-st)*1000.0:.2f} ms, waited {(mt2-mt)*1000.0:.2f} ms for realize, {(et-mt2)*1000.0:.2f} ms for GPU queue")
-
-    onnx_model = onnx.load(fetch(OPENPILOT_MODEL))
-    torch_out = run_onnx_torch(onnx_model, inputs).numpy()
-    print(tinygrad_out, torch_out)
-    np.testing.assert_allclose(tinygrad_out, torch_out, atol=1e-4, rtol=1e-2)
-
  @unittest.skip("slow")
  def test_efficientnet(self):
    input_name, input_new = "images:0", True
@@ -146,7 +58,7 @@ class TestOnnxModel(unittest.TestCase):
    print(cls, _LABELS[cls])
    assert "car" in _LABELS[cls] or _LABELS[cls] == "convertible"

-@unittest.skipUnless(HUGGINGFACE_AVAILABLE and Device.DEFAULT == "METAL", "only run on METAL")
+@unittest.skipUnless(Device.DEFAULT == "METAL", "only run on METAL")
 class TestHuggingFaceOnnxModels(unittest.TestCase):
  @classmethod
  def setUpClass(cls):
@@ -53,8 +53,8 @@ class TestRealWorld(unittest.TestCase):
  @unittest.skipUnless(is_dtype_supported(dtypes.float16), "need dtypes.float16")
  def test_stable_diffusion(self):
    params = unet_params
-    params["model_ch"] = 16
-    params["ctx_dim"] = 16
+    params["model_ch"] = 8
+    params["ctx_dim"] = 8
    params["num_res_blocks"] = 1
    params["n_heads"] = 2
    model = UNetModel(**params)
@@ -94,7 +94,7 @@ class TestRealWorld(unittest.TestCase):
    @TinyJit
    def test(t, v):
      with Context(JIT=0): return model(t, v).realize()
-    helper_test("test_gpt2", lambda: (Tensor([[1,]]),Variable("pos", 1, 100).bind(1)), test, 0.23 if CI else 0.9, 137 if CI else 396, all_jitted=True)
+    helper_test("test_gpt2", lambda: (Tensor([[1,]]),Variable("pos", 1, 100).bind(1)), test, 0.23 if CI else 0.9, 160 if CI else 396, all_jitted=True)

  @unittest.skipIf(CI and Device.DEFAULT == "CPU", "slow")
  def test_train_mnist(self):
@@ -112,9 +112,19 @@ class TestRealWorld(unittest.TestCase):
        loss.backward()
        optimizer.step()

-      helper_test("train_mnist", lambda: (Tensor.randn(BS, 1, 28, 28),), train, 0.07, 93)
+      helper_test("train_mnist", lambda: (Tensor.randn(BS, 1, 28, 28),), train, 0.07, 102)

-  @unittest.skipIf(CI and Device.DEFAULT in {"CPU", "GPU", "LLVM"}, "slow")
+  @unittest.skipIf(CI and Device.DEFAULT in {"CPU", "CL"}, "slow")
+  def test_forward_cifar(self):
+    BS = 32
+    # with training batchnorm still though
+    with Tensor.train():
+      model = SpeedyResNet(Tensor.ones((12,3,2,2)))
+      @TinyJit
+      def run(X): return model(X)
+      helper_test("forward_cifar", lambda: (Tensor.randn(BS, 3, 32, 32),), run, (1.0/48)*BS, 126)
+
+  @unittest.skipIf(CI and Device.DEFAULT in {"CPU", "CL"}, "slow")
  def test_train_cifar(self):
    with Tensor.train():
      model = SpeedyResNet(Tensor.ones((12,3,2,2)))
@@ -144,6 +154,7 @@ class TestRealWorld(unittest.TestCase):
                                final_div_factor=1./(initial_div_factor*final_lr_ratio), total_steps=4)
      assert not np.isnan(lr_scheduler.min_lr), "lr too small or initial_div_facotr too big for half"

+  @unittest.skipIf(CI and Device.DEFAULT == "CPU", "slow")
  def test_bert(self):
    with Tensor.train():
      args_tiny = {"attention_probs_dropout_prob": 0.0, "hidden_dropout_prob": 0.0, "vocab_size": 30522, "type_vocab_size": 2,
@@ -165,11 +176,7 @@ class TestRealWorld(unittest.TestCase):
      for v in data.values(): v.to_(Device.DEFAULT)

      helper_test("train_bert", lambda: (data["input_ids"], data["segment_ids"], data["input_mask"], data["masked_lm_positions"], \
-          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.25, 347)
-
-  def test_bert_fuse_arange(self):
-    with Context(FUSE_ARANGE=1):
-      self.test_bert()
+          data["masked_lm_ids"], data["masked_lm_weights"], data["next_sentence_labels"]), train, 0.28, 357)

 if __name__ == '__main__':
  unittest.main()
@@ -1,8 +1,8 @@
 #!/usr/bin/env python
 import unittest
-import numpy as np
-from tinygrad.tensor import Tensor
+from tinygrad import Tensor
 from extra.models.rnnt import LSTM
+import numpy as np
 import torch

 class TestRNNT(unittest.TestCase):
@@ -1,9 +1,8 @@
-import unittest
-import time
+import unittest, time
 import numpy as np
+from tinygrad import Device
 from tinygrad.nn.state import get_parameters
 from tinygrad.nn import optim
-from tinygrad.tensor import Device
 from tinygrad.helpers import getenv, CI
 from extra.training import train
 from extra.models.convnext import ConvNeXt
@@ -27,7 +26,7 @@ def train_one_step(model,X,Y):
  print("done in %.2f ms" % (et*1000.))

 def check_gc():
-  if Device.DEFAULT == "GPU":
+  if Device.DEFAULT == "CL":
    from extra.introspection import print_objects
    assert print_objects() == 0

@@ -40,7 +39,6 @@ class TestTrain(unittest.TestCase):
    check_gc()

  @unittest.skipIf(CI, "slow")
-  @unittest.skipIf(Device.DEFAULT in ["METAL", "WEBGPU"], "too many buffers for webgpu and metal")
  def test_efficientnet(self):
    model = EfficientNet(0)
    X = np.zeros((BS,3,224,224), dtype=np.float32)
@@ -49,7 +47,6 @@ class TestTrain(unittest.TestCase):
    check_gc()

  @unittest.skipIf(CI, "slow")
-  @unittest.skipIf(Device.DEFAULT in ["METAL", "WEBGPU"], "too many buffers for webgpu and metal")
  def test_vit(self):
    model = ViT()
    X = np.zeros((BS,3,224,224), dtype=np.float32)
@@ -57,7 +54,7 @@ class TestTrain(unittest.TestCase):
    train_one_step(model,X,Y)
    check_gc()

-  @unittest.skipIf(Device.DEFAULT in ["METAL", "WEBGPU"], "too many buffers for webgpu and metal")
+  @unittest.skipIf(CI, "slow")
  def test_transformer(self):
    # this should be small GPT-2, but the param count is wrong
    # (real ff_dim is 768*4)
@@ -1,7 +1,7 @@
 import unittest
 import pathlib
 from examples.whisper import init_whisper, load_file_waveform, transcribe_file, transcribe_waveform
-from tinygrad.helpers import CI, fetch
+from tinygrad.helpers import CI, fetch, CPU_LLVM
 from tinygrad import Device, dtypes
 from tinygrad.device import is_dtype_supported

@@ -16,7 +16,7 @@ TRANSCRIPTION_2 = "a slightly longer audio file so that we can test batch transc
 TEST_FILE_3_URL = 'https://homepage.ntu.edu.tw/~karchung/miniconversations/mc45.mp3'
 TRANSCRIPTION_3 = "Just lie back and relax. Is the level of pressure about right? Yes, it's fine, and I'd like conditioner please. Sure. I'm going to start the second lathering now. Would you like some Q-tips? How'd you like it cut? I'd like my bangs and the back trimmed, and I'd like the rest thinned out a bit and layered. Where would you like the part? On the left, right about here. Here, have a look. What do you think? It's fine. Here's a thousand anti-dollars. It's 30-ant extra for the rants. Here's your change and receipt. Thank you, and please come again. So how do you like it? It could have been worse, but you'll notice that I didn't ask her for her card. Hmm, yeah. Maybe you can try that place over there next time."   # noqa: E501

-@unittest.skipIf(Device.DEFAULT in ["CPU", "LLVM"], "slow")
+@unittest.skipIf(Device.DEFAULT in ["CPU"], "slow")
@unittest.skipUnless(is_dtype_supported(dtypes.float16), "need float16 support")
 class TestWhisper(unittest.TestCase):
  @classmethod
@@ -33,11 +33,11 @@ class TestWhisper(unittest.TestCase):
  def test_transcribe_file1(self):
    self.assertEqual(transcribe_file(self.model, self.enc, TEST_FILE_1),  TRANSCRIPTION_1)

-  @unittest.skipIf(CI or Device.DEFAULT == "LLVM", "too many tests for CI")
+  @unittest.skipIf(CI or (Device.DEFAULT == "CPU" and CPU_LLVM), "too many tests for CI")
  def test_transcribe_file2(self):
    self.assertEqual(transcribe_file(self.model, self.enc, TEST_FILE_2),  TRANSCRIPTION_2)

-  @unittest.skipIf(CI or Device.DEFAULT == "LLVM", "too many tests for CI")
+  @unittest.skipIf(CI or (Device.DEFAULT == "CPU" and CPU_LLVM), "too many tests for CI")
  def test_transcribe_batch12(self):
    waveforms = [load_file_waveform(TEST_FILE_1), load_file_waveform(TEST_FILE_2)]
    transcriptions = transcribe_waveform(self.model, self.enc, waveforms)
@@ -52,13 +52,13 @@ class TestWhisper(unittest.TestCase):
    self.assertEqual(TRANSCRIPTION_2,  transcriptions[0])
    self.assertEqual(TRANSCRIPTION_1,  transcriptions[1])

-  @unittest.skipIf(CI or Device.DEFAULT == "LLVM", "too long for CI")
+  @unittest.skipIf(CI or (Device.DEFAULT == "CPU" and CPU_LLVM), "too long for CI")
  def test_transcribe_long(self):
    waveform = [load_file_waveform(fetch(TEST_FILE_3_URL))]
    transcription = transcribe_waveform(self.model, self.enc, waveform)
    self.assertEqual(TRANSCRIPTION_3, transcription)

-  @unittest.skipIf(CI or Device.DEFAULT == "LLVM", "too long for CI")
+  @unittest.skipIf(CI or (Device.DEFAULT == "CPU" and CPU_LLVM), "too long for CI")
  def test_transcribe_long_no_batch(self):
    waveforms = [load_file_waveform(fetch(TEST_FILE_3_URL)), load_file_waveform(TEST_FILE_1)]

@@ -0,0 +1,180 @@
+import unittest
+from tinygrad import Device, Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops
+from tinygrad.codegen.opt import Opt, OptOps
+from tinygrad.shape.shapetracker import ShapeTracker, View
+from tinygrad.engine.realize import get_program
+from tinygrad.helpers import AMX
+
+@unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "need backends that support float4")
+class TestFloat4(unittest.TestCase):
+  @staticmethod
+  def count_float4(uops: list[UOp], n=4):
+    return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype == dtypes.float.vec(n)]),
+            len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype == dtypes.float.vec(n)]))
+  @staticmethod
+  def count_half4(uops: list[UOp]):
+    return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype == dtypes.half.vec(4)]),
+            len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype == dtypes.half.vec(4)]))
+
+  def test_float4_basic(self):
+    a = Tensor.empty(2, 8).realize()
+    b = Tensor.empty(2, 8).realize()
+    c = a + b
+
+    s = c.schedule()[0]
+    realized_ast = s.ast
+    opts_to_apply = [Opt(op=OptOps.UPCAST, axis=0, arg=4)]
+    program = get_program(realized_ast, Device[Device.DEFAULT].renderer, opts=opts_to_apply)
+
+    assert TestFloat4.count_float4(program.uops) == (2, 1)
+
+  @unittest.skipIf(Device.DEFAULT in {"CPU"} and AMX, "CPU with AMX upcasts float up to size 16")
+  def test_float4_multidim(self):
+    a = Tensor.empty(2, 8).realize()
+    b = Tensor.empty(2, 8).realize()
+    c = a + b
+
+    s = c.schedule()[0]
+    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=2)]).uops
+    assert TestFloat4.count_float4(uops) == (4, 2)
+
+  @unittest.skipUnless(Device.DEFAULT in {"CPU"} and AMX, "Only CPU with AMX upcasts float up to size 16")
+  def test_float4_multidim_amx(self):
+    def kernel_for_shape(size, shift):
+      a = Tensor.empty(2, size).realize()
+      b = Tensor.empty(2, size).realize()
+      c = a + b
+
+      s = c.schedule()[0]
+      return get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=shift)]).uops
+
+    sizes = [12, 8, 16]
+    shifts = [3, 2, 4]
+    expected_upcast_size = [4, 8, 16]
+    expected_output = [(6,3), (2,1), (2,1)]
+
+    for i in range(len(sizes)):
+      assert TestFloat4.count_float4(kernel_for_shape(sizes[i], shifts[i]), expected_upcast_size[i]) == expected_output[i]
+
+  def test_float4_unaligned_load(self):
+    a = Tensor.empty(9).realize().shrink(((1, 9),))
+    b = Tensor.empty(9).realize().shrink(((1, 9),))
+    c = a + b
+
+    s = c.schedule()[0]
+    realized_ast = s.ast
+    opts_to_apply = [Opt(op=OptOps.UPCAST, axis=0, arg=4)]
+    program = get_program(realized_ast, Device[Device.DEFAULT].renderer, opts=opts_to_apply)
+
+    assert TestFloat4.count_float4(program.uops) == (0, 1)
+
+  @unittest.skipIf(Device.DEFAULT in {"CPU"} and AMX, "CPU with AMX upcasts float up to size 16")
+  def test_float4_multidim_unaligned_load(self):
+    a = Tensor.empty(2, 9).realize().shrink(((0, 2), (1, 9),))
+    b = Tensor.empty(2, 9).realize().shrink(((0, 2), (1, 9),))
+    c = a + b
+
+    s = c.schedule()[0]
+    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=2)]).uops
+
+    assert TestFloat4.count_float4(uops) == (0, 2)
+
+  @unittest.skipUnless(Device.DEFAULT in {"CPU"} and AMX, "Only CPU with AMX upcasts float up to size 16")
+  def test_float4_multidim_unaligned_load_amx(self):
+    def kernel_for_shape(size, shift):
+      a = Tensor.empty(2, size).realize().shrink(((0, 2), (1, size),))
+      b = Tensor.empty(2, size).realize().shrink(((0, 2), (1, size),))
+      c = a + b
+
+      s = c.schedule()[0]
+      return get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=1, arg=shift)]).uops
+
+    sizes = [13, 9, 17]
+    shifts = [3, 2, 4]
+    expected_upcast_size = [4, 8, 16]
+    expected_output = [(0,3), (0,1), (0,1)]
+
+    for i in range(len(sizes)):
+      assert TestFloat4.count_float4(kernel_for_shape(sizes[i], shifts[i]), expected_upcast_size[i]) == expected_output[i]
+
+  def test_float4_sometimes_unaligned(self):
+    a = Tensor.empty(1, 1, 8).realize()
+    b = Tensor.empty(1, 1, 5).realize().shrink(((0, 1), (0, 1), (1, 5)))
+    c = a.conv2d(b)
+    # only the first and last conv dot products are aligned in a, and b is never aligned, so no
+    # float4 should be emitted (the reduce axis of size 4 is the float4 axis here)
+
+    s = c.schedule()[0]
+    uops = get_program(s.ast, opts=[Opt(op=OptOps.UNROLL, axis=0, arg=4)]).uops
+
+    assert TestFloat4.count_float4(uops) == (0, 0)
+
+  def test_float4_multidim_sometimes_unaligned(self):
+    a = Tensor.empty(1, 1, 7).realize()
+    b = Tensor.empty(1, 1, 5).realize().shrink(((0, 1), (0, 1), (1, 5)))
+    c = a.conv2d(b)
+    # the first conv dot product is aligned in a. If we upcast the output and reduce
+    # dimension, then we could do float4 for only that one set of loads, but we currently
+    # don't.
+    # UPDATE: now we do this fusion
+
+    s = c.schedule()[0]
+    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.UNROLL, axis=0, arg=0)]).uops
+
+    assert TestFloat4.count_float4(uops) in {(0,1), (1,1)}
+
+  def test_float4_expand(self):
+    a = Tensor.empty(9).realize().shrink(((1, 9),))
+    b = Tensor.empty(2).realize().reshape((2, 1)).expand((2,4)).reshape((8,))
+    c = a + b
+
+    # we will upcast the top axis of sz 4. they should not be coalesced into float4,
+    # since the top axis is not contiguous.
+
+    s = c.schedule()[0]
+    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4)]).uops
+
+    assert TestFloat4.count_float4(uops) == (0, 1)
+
+  def test_float4_heterogeneous(self):
+    a = Tensor.empty(8).realize()
+    b = Tensor.empty(9).realize().shrink(((1, 9),))
+    c = a + b
+
+    # should float4 b but not a
+
+    s = c.schedule()[0]
+    uops = get_program(s.ast, opts=[Opt(op=OptOps.UPCAST, axis=0, arg=4)]).uops
+
+    assert TestFloat4.count_float4(uops) == (1, 1)
+
+  def test_half4_load_unrolled(self):
+    # from llama 7B shard 4 gpus
+    ast = UOp(Ops.SINK, dtypes.void, arg=None, src=(
+      UOp(Ops.STORE, dtypes.void, arg=None, src=(
+        UOp(Ops.VIEW, dtypes.float.ptr(96000), arg=ShapeTracker(views=(View(shape=(1, 3, 32000, 1), strides=(0, 32000, 1, 0), offset=0, mask=None, contiguous=True),)), src=( # noqa: E501
+          UOp(Ops.DEFINE_GLOBAL, dtypes.float.ptr(96000), arg=0, src=()),)),
+        UOp(Ops.REDUCE_AXIS, dtypes.float, arg=(Ops.ADD, (3,)), src=(
+          UOp(Ops.CAST, dtypes.float, arg=None, src=(
+            UOp(Ops.MUL, dtypes.half, arg=None, src=(
+              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
+                UOp(Ops.VIEW, dtypes.half.ptr(9216), arg=ShapeTracker(views=(View(shape=(1, 3, 32000, 1024), strides=(0, 4096, 0, 1), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
+                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(9216), arg=1, src=()),)),)),
+              UOp(Ops.LOAD, dtypes.half, arg=None, src=(
+                UOp(Ops.VIEW, dtypes.half.ptr(32768000), arg=ShapeTracker(views=(View(shape=(1, 3, 32000, 1024), strides=(0, 0, 1024, 1), offset=0, mask=None, contiguous=False),)), src=( # noqa: E501
+                  UOp(Ops.DEFINE_GLOBAL, dtypes.half.ptr(32768000), arg=2, src=()),)),)),)),)),)),)),))
+
+    # TODO: fix this, expected might change but should be positive
+    for expected, opts in [
+      ((7, 0), [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UPCAST, axis=0, arg=3), Opt(op=OptOps.UNROLL, axis=0, arg=4)]),
+      ((5, 0), [Opt(op=OptOps.UPCAST, axis=1, arg=4), Opt(op=OptOps.UNROLL, axis=0, arg=4)]),
+      ((2, 0), [Opt(op=OptOps.UNROLL, axis=0, arg=4)]),
+    ]:
+      program = get_program(ast, Device[Device.DEFAULT].renderer, opts=opts)
+
+      count = TestFloat4.count_half4(program.uops)
+      assert count == expected, f"{count=}, {expected=}"
+
+if __name__ == '__main__':
+  unittest.main()
@@ -0,0 +1,368 @@
+import unittest
+from tinygrad import Device, Tensor, dtypes
+from tinygrad.helpers import CI, RANGEIFY
+from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
+
+# TODO: write a clean version of this
+from test.test_linearizer import helper_linearizer_opt
+
+class TestKernelOpts(unittest.TestCase):
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  def test_local_and_grouped_reduce(self):
+    N = 128
+    Tensor.manual_seed(1882)
+    a = Tensor.rand(4, 4, N, N)
+    b = Tensor.rand(4, 4, N)
+    r = (b.sqrt() + ((a+1).sum(axis=3).exp()))
+    helper_linearizer_opt(r, [
+      [Opt(OptOps.LOCAL, 0, 2)],
+      [Opt(OptOps.LOCAL, 0, 8)],
+      [Opt(OptOps.LOCAL, 0, 16)], # Checking how it works with locals
+      [Opt(OptOps.GROUPTOP, 0, 2)],
+      [Opt(OptOps.GROUPTOP, 0, 32)],
+      [Opt(OptOps.GROUPTOP, 0, 64)], # Checking how it works with grouped reduce
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 2)],
+      [Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.GROUPTOP, 0, 16)],
+      [Opt(OptOps.LOCAL, 0, 32), Opt(OptOps.GROUPTOP, 0, 2)],
+      # Checking how it works with locals + grouped reduce
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 64)],
+      # Checking how it works with locals + grouped reduce + upcasts
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.UPCAST, 0, 8), Opt(OptOps.UNROLL, 1, 4)],
+      # many local + many group
+      [Opt(OptOps.GROUP, 0, 2)] * 4,
+      [Opt(OptOps.LOCAL, 0, 2)] * 4,
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUP, 0, 2)] * 4,
+    ])
+
+  def test_upcasts(self):
+    N = 16
+    Tensor.manual_seed(1772)
+    a = Tensor.rand(N, N)
+    b = Tensor.rand(N, N)
+    r = (a+b).sqrt() * ((a+1).exp())
+    helper_linearizer_opt(r, [
+      [Opt(OptOps.UPCAST, 0, 2)],
+      [Opt(OptOps.UPCAST, 0, 4)],
+      [Opt(OptOps.UPCAST, 0, 8)], # Checking how it works with upcasts
+    ])
+
+  def test_full_upcast(self):
+    Tensor.manual_seed(1772)
+    a = Tensor.rand(4)
+    b = Tensor.rand(4)
+    r = (a+b).sqrt() * ((a+1).exp())
+    helper_linearizer_opt(r, [
+      [Opt(OptOps.UPCAST, 0, 4)], # Checking how it works with upcasts
+    ])
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  def test_matmul(self):
+    N = 128
+    Tensor.manual_seed(1552)
+    a = Tensor.rand(N, N)
+    b = Tensor.rand(N, N)
+    r = a@b
+    helper_linearizer_opt(r, [
+      [Opt(OptOps.UPCAST, 0, 2)],
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4)], # Checking how it works with upcasts
+      [Opt(OptOps.LOCAL, 0, 2)],
+      [Opt(OptOps.LOCAL, 1, 32)],
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4)],
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 32)],
+      [Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.LOCAL, 1, 8)], # Checking how it works with locals
+      [Opt(OptOps.GROUPTOP, 0, 2)],
+      [Opt(OptOps.GROUPTOP, 0, 32)],
+      [Opt(OptOps.GROUPTOP, 0, 32), Opt(OptOps.UNROLL, 0, 4)], # Checking how it works with grouped_reduce
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.LOCAL, 1, 2), Opt(OptOps.GROUPTOP, 0, 32)],
+      [Opt(OptOps.LOCAL, 0, 8), Opt(OptOps.GROUPTOP, 0, 32)],
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 0, 8), Opt(OptOps.GROUPTOP, 0, 4)], # Checking how it works with local+grouped_reduce
+      # Checking all together
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 4),
+       Opt(OptOps.UPCAST, 1, 2)],
+      # Full global upcast + local
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UPCAST, 0, 8)],
+    ])
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  def test_double_reduce(self):
+    N = 128
+    Tensor.manual_seed(1552)
+    a = Tensor.rand(8, N, 8, N)
+    r = a.sum(axis=(1,3))
+    helper_linearizer_opt(r, [
+      # openCL / CL=1 is 256 max threads
+      [Opt(OptOps.GROUPTOP, 0, 2)], [Opt(OptOps.GROUPTOP, 0, 32)],
+      [Opt(OptOps.GROUPTOP, 1, 2)], [Opt(OptOps.GROUPTOP, 1, 32)], # Checking how it works with 1 grouped_reduce.
+      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 2)],
+      [Opt(OptOps.GROUPTOP, 0, 16), Opt(OptOps.GROUPTOP, 1, 2)],
+      [Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 64)], # Checking how it works with 2 grouped_reduces.
+      [Opt(OptOps.GROUPTOP, 0, 16), Opt(OptOps.GROUPTOP, 1, 2), Opt(OptOps.UNROLL, 0, 4)],
+      [Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 32), Opt(OptOps.UNROLL, 2, 4)], # Checking how it works with 2 grouped_reduces + upcasts.
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4), Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4)],
+      # Checking how it works with 2 grouped_reduces + upcasts + locals.
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4), Opt(OptOps.GROUPTOP, 0, 2), Opt(OptOps.GROUPTOP, 1, 32), Opt(OptOps.UNROLL, 1, 4)],
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.LOCAL, 1, 2), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UPCAST, 0, 2)],
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.LOCAL, 1, 2), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UPCAST, 0, 2),
+       Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.UNROLL, 1, 4)], # Checking how it works with 2 grouped_reduces + upcasts + locals.
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.LOCAL, 1, 4), Opt(OptOps.GROUPTOP, 0, 4), Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.UPCAST, 0, 2),
+       Opt(OptOps.UPCAST, 0, 2)], # No globals
+    ])
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  @unittest.skipUnless(any(tc.dtype_in == tc.dtype_out == dtypes.half for tc in Device[Device.DEFAULT].renderer.tensor_cores),
+                      "test requires tensor cores with accumulation in half") # testing with half suffices.
+  def test_tensor_core_opts(self):
+    N = 128
+    Tensor.manual_seed(1552)
+    a, b = Tensor.rand(N, N, dtype=dtypes.half), Tensor.rand(N, N, dtype=dtypes.half)
+    r = a.matmul(b, dtype=dtypes.half)
+    atol, rtol = 0.25, 0.01
+    helper_linearizer_opt(r, [
+      [],
+      [Opt(OptOps.UPCAST, 0, 4)],
+      [Opt(OptOps.UPCAST, 1, 4)],
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4)], # check upcasts
+      [Opt(OptOps.UNROLL, 0, 2)], # check unroll
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 2)], # check combo of unroll and local
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 2)],
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 4)],
+      [Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UPCAST, 0, 4)], # check permutations
+      [Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 0, 4)],
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 1, 4)],
+      [Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 4)],
+    ], apply_tc=True, atol=atol, rtol=rtol)
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  @unittest.skipUnless(any(tc.dtype_in == tc.dtype_out == dtypes.half for tc in Device[Device.DEFAULT].renderer.tensor_cores),
+                      "test requires tensor cores with accumulation in half") # testing with half suffices.
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  def test_tensor_core_opts_locals(self):
+    N = 128
+    Tensor.manual_seed(1552)
+    a, b = Tensor.rand(N, N, dtype=dtypes.half), Tensor.rand(N, N, dtype=dtypes.half)
+    r = a.matmul(b, dtype=dtypes.half)
+    atol, rtol = 0.25, 0.01
+    helper_linearizer_opt(r, [
+      [Opt(OptOps.UNROLL, 0, 0)], # check full unroll of reduce with locals
+      [Opt(OptOps.LOCAL, 0, 4)], # check local
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.LOCAL, 0, 2)],
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.UPCAST, 1, 4), Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 0, 4)],
+    ], apply_tc=True, atol=atol, rtol=rtol)
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared memory")
+  @unittest.skipUnless(any(tc.dtype_in == tc.dtype_out == dtypes.half for tc in Device[Device.DEFAULT].renderer.tensor_cores),
+                      "test requires tensor cores with accumulation in half") # testing with half suffices.
+  # NOTE: the METAL test is broken, likely due to a compiler bug. passes on CI with -O0 and with default opt level locally on M3
+  @unittest.skipIf(Device.DEFAULT == "METAL", "broken for METAL")
+  @unittest.skip("feature was removed")
+  def test_tensor_core_opts_group(self):
+    N = 128
+    Tensor.manual_seed(1552)
+    a, b = Tensor.rand(N, N, dtype=dtypes.half), Tensor.rand(N, N, dtype=dtypes.half)
+    r = a.matmul(b, dtype=dtypes.half)
+    atol, rtol = 0.25, 0.01
+    helper_linearizer_opt(r, [
+      [Opt(OptOps.GROUP, 0, 2)],
+      [Opt(OptOps.GROUPTOP, 0, 4)],
+      [Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.GROUP, 0, 2)],
+      [Opt(OptOps.LOCAL, 0, 4), Opt(OptOps.GROUP, 0, 2)],
+      [Opt(OptOps.UNROLL, 0, 4), Opt(OptOps.GROUP, 0, 2)],
+      [Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUP, 0, 2)],
+      [Opt(OptOps.LOCAL, 0, 2), Opt(OptOps.GROUPTOP, 0, 8), Opt(OptOps.UNROLL, 0, 2), Opt(OptOps.UPCAST, 1, 2)],
+    ], apply_tc=True, atol=atol, rtol=rtol)
+
+  def test_padto_matmul(self):
+    if (CI and Device.DEFAULT in ["AMD", "NV", "CUDA"]):
+      self.skipTest("super slow on CUDA and AMD because of the big grid dims")
+    N = 17 * 17
+    Tensor.manual_seed(289)
+    a = Tensor.rand(N, N)
+    b = Tensor.rand(N, N)
+    helper_linearizer_opt(a@b, [
+      [Opt(OptOps.PADTO, 0, 32)],
+      [Opt(OptOps.PADTO, 1, 32)],
+      [Opt(OptOps.PADTO, 2, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.PADTO, 1, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.PADTO, 1, 32), Opt(OptOps.PADTO, 2, 32)],
+      # can optimize further post PADTO
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.PADTO, 1, 32), Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.UPCAST, 1, 2),],
+    ])
+
+  def test_padto_upcasted_not_ok(self):
+    N = 4
+    a = Tensor.rand(N, N)
+    b = Tensor.rand(N, N)
+    helper_linearizer_opt(a@b, [
+      [Opt(OptOps.UPCAST, 0, 0)],
+      [Opt(OptOps.UPCAST, 1, 0)],
+      [Opt(OptOps.UNROLL, 0, 0)],
+      [Opt(OptOps.PADTO, 0, 8)],
+      [Opt(OptOps.PADTO, 1, 8)],
+      [Opt(OptOps.PADTO, 2, 8)],
+    ])
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(a@b, [[Opt(OptOps.UPCAST, 0, 0), Opt(OptOps.PADTO, 1, 8)]])
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(a@b, [[Opt(OptOps.UPCAST, 1, 0), Opt(OptOps.PADTO, 1, 8)]])
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(a@b, [[Opt(OptOps.UNROLL, 0, 0), Opt(OptOps.PADTO, 2, 8)]])
+
+  def test_padto_sum_ok(self):
+    N = 18 * 18
+    # NOTE: this setup prevents 17 * 17 contiguous merged into one dimension
+    a = Tensor.rand(N, N).realize().shrink(((0, 17), (0, 17))) * 100
+    b = (Tensor.rand(N, N) < 0.5).realize().shrink(((0, 17), (0, 17)))
+
+    helper_linearizer_opt(a.sum(0), [
+      [Opt(OptOps.PADTO, 0, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
+    ])
+    helper_linearizer_opt(a.sum(1), [
+      [Opt(OptOps.PADTO, 0, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
+    ])
+
+    # can pad sum reduce axis if there's no unsafe ops prior to sum
+    for axis in (0, 1):
+      helper_linearizer_opt(a.sum(), [[Opt(OptOps.PADTO, axis, 32)],])
+      helper_linearizer_opt(a.sum(0), [[Opt(OptOps.PADTO, axis, 32)],])
+      helper_linearizer_opt(b.sum(), [[Opt(OptOps.PADTO, axis, 32)],])
+      helper_linearizer_opt(b.sum(0), [[Opt(OptOps.PADTO, axis, 32)],])
+      helper_linearizer_opt(b.sum(dtype=dtypes.bool), [[Opt(OptOps.PADTO, axis, 32)],])
+      # TODO: why?
+      if Device.DEFAULT != "WEBGPU":
+        helper_linearizer_opt(b.sum(0, dtype=dtypes.bool), [[Opt(OptOps.PADTO, axis, 32)],])
+        helper_linearizer_opt(b.sum(1, dtype=dtypes.bool), [[Opt(OptOps.PADTO, axis, 32)],])
+
+    # having unsafe ops after sum is fine
+    helper_linearizer_opt(a.sum().exp(), [[Opt(OptOps.PADTO, 0, 32)],])
+    helper_linearizer_opt(a.sum(0).exp(), [[Opt(OptOps.PADTO, 1, 32)],])
+
+  def test_padto_sum_not_ok(self):
+    N = 18 * 18
+    # NOTE: this setup prevents 17 * 17 contiguous merged into one dimension
+    a = Tensor.rand(N, N).shrink(((0, 17), (0, 17))).exp()
+    # exp is not safe to pad
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(a.exp().sum(), [[Opt(OptOps.PADTO, 0, 32)],])
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(a.exp().sum(0), [[Opt(OptOps.PADTO, 1, 32)],])
+
+    b = a < 1
+    # lt is not safe to pad
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(b.sum(), [[Opt(OptOps.PADTO, 0, 32)],])
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(b.sum(0), [[Opt(OptOps.PADTO, 1, 32)],])
+
+  def test_padto_max(self):
+    N = 18 * 18
+    # NOTE: this setup prevents 17 * 17 contiguous merged into one axis
+    a = -Tensor.rand(N, N).shrink(((0, 17), (0, 17))) * 100
+
+    helper_linearizer_opt(a.max(0), [
+      [Opt(OptOps.PADTO, 0, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
+    ])
+    helper_linearizer_opt(a.max(1), [
+      [Opt(OptOps.PADTO, 0, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
+    ])
+
+    # cannot pad max kernel on reduce
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(a.max(), [[Opt(OptOps.PADTO, 0, 32)],])
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(a.max(0), [[Opt(OptOps.PADTO, 1, 32)],])
+
+  def test_padto_where(self):
+    Tensor.manual_seed(0)
+    N = 17 * 17
+    a = (Tensor.randn(N, N).realize().max(axis=0, keepdim=True) > 1).where(1, 0)
+    helper_linearizer_opt(a.max(0), [
+      [Opt(OptOps.PADTO, 0, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
+    ])
+
+  def test_padto_where_multioutput(self):
+    Tensor.manual_seed(0)
+    N = 17 * 17
+    r = Tensor.randn(N, N).realize().max(axis=0, keepdim=True) > 1
+    a0 = r.where(1, 0)
+    a1 = r.where(2, 0)
+    helper_linearizer_opt([a0.max(0), a1.max(0)], [
+      [Opt(OptOps.PADTO, 0, 32)],
+      [Opt(OptOps.PADTO, 0, 32), Opt(OptOps.UPCAST, 0, 8),],
+    ])
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  def test_color_shapes_with_local(self):
+    N = 32
+    Tensor.manual_seed(1552)
+    a = Tensor.rand(N, N)
+    b = Tensor.rand(N, N)
+    r = a@b
+    opts_shapes = [
+      ([Opt(OptOps.LOCAL, 0, 2)], [("blue",16),("blue",32),("cyan",2),("red",32)]),
+      ([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.GROUP, 0, 2)], [("blue",16),("blue",32),("cyan",2),("green",2),("red",16)]),
+      # check to ensure local_dims are stable for full UNROLL of the first reduce
+      ([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.UNROLL, 0, 0)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
+      ([Opt(OptOps.UNROLL, 0, 0),Opt(OptOps.LOCAL, 0, 2)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
+      # check behavior for full UNROLL on an existing GROUP
+      ([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.GROUP, 0, 0),Opt(OptOps.UNROLL, 0, 2)], [("blue",16),("blue",32),("cyan",2),("green",16),("magenta",2)]),
+      ([Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.GROUP, 0, 0),Opt(OptOps.UNROLL, 0, 0)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
+      ([Opt(OptOps.GROUP, 0, 0),Opt(OptOps.LOCAL, 0, 2),Opt(OptOps.UNROLL, 0, 0)], [("blue",16),("blue",32),("cyan",2),("magenta",32)]),
+      ([Opt(OptOps.GROUP, 0, 2),Opt(OptOps.UNROLL, 0, 0)], [("blue",32),("blue",32),("red",16),("magenta",2)]),
+    ]
+    helper_linearizer_opt(r, [x[0] for x in opts_shapes], color_sizes=[x[1] for x in opts_shapes])
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "test requires float4")
+  def test_arange_opts(self):
+    a = Tensor.arange(128)
+    # NOTE: arange no longer has reduce ops available for opt
+    helper_linearizer_opt(a, [
+      #[Opt(OptOps.GROUP, 0, 32)],
+      #[Opt(OptOps.GROUPTOP, 0, 32)],
+      [Opt(op=OptOps.LOCAL, axis=0, arg=8)],
+      [Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=0, arg=0)],
+      #[Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=8)],
+      #[Opt(op=OptOps.LOCAL, axis=0, arg=8), Opt(op=OptOps.UPCAST, axis=0, arg=0), Opt(op=OptOps.GROUP, axis=0, arg=8), Opt(op=OptOps.UNROLL, axis=1, arg=4)], # noqa: E501
+    ])
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_threads, "test requires threads")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.global_max is not None and
+                       Device[Device.DEFAULT].renderer.global_max[0] > 1, "test requires multicore")
+  def test_thread_opts(self):
+    a = Tensor.rand(4, 4, 4, 4)
+    b = Tensor.rand(4, 4, 4)
+    r = (b.sqrt() + ((a+1).sum(axis=3).exp()))
+    helper_linearizer_opt(r, [
+      [Opt(OptOps.THREAD, 0, 2)],
+      [Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.THREAD, 0, 2)],
+      [Opt(OptOps.UPCAST, 0, 2), Opt(OptOps.THREAD, 0, 2), Opt(OptOps.UNROLL, 0, 2)],
+    ] + [[Opt(OptOps.THREAD, 0, 4)] if Device[Device.DEFAULT].renderer.global_max[0] >= 4 else []]
+      + [[Opt(OptOps.THREAD, 0, 8)] if Device[Device.DEFAULT].renderer.global_max[0] >= 8 else []])
+
+  @unittest.skipUnless(RANGEIFY>=1, "Kernel only fuses with rangeify")
+  def test_double_sum_group(self):
+    a = Tensor.rand(4, 4, 4)
+    r = a.sum((1, 2)).sum()
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(r, [[Opt(OptOps.GROUPTOP, 0, 16)],])
+    r = a.sum((1, 2)).sum()
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(r, [[Opt(OptOps.UNROLL, 1, 4), Opt(OptOps.GROUPTOP, 0, 16)],])
+    r = a.sum((1, 2)).sum()
+    with self.assertRaises(KernelOptError):
+      helper_linearizer_opt(r, [[Opt(OptOps.GROUPTOP, 1, 4), Opt(OptOps.GROUPTOP, 0, 16)],])
+
+if __name__ == '__main__':
+  unittest.main()
@@ -0,0 +1,191 @@
+import numpy as np
+import unittest
+from dataclasses import replace
+
+from tinygrad import Device, Tensor, dtypes
+from tinygrad.tensor import _to_np_dtype
+from tinygrad.uop.ops import Ops
+from tinygrad.dtype import DType
+from tinygrad.device import is_dtype_supported
+from tinygrad.helpers import AMX, CI, AMD_LLVM, CPU_LLVM
+from tinygrad.engine.realize import CompiledRunner, get_program
+from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
+
+# TODO: write a clean version of this
+from test.test_linearizer import helper_realized_ast, helper_linearizer_opt
+
+def helper_tc_ensure_uops_and_opts_count(N: int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0,
+                                         ensure_triggered:bool=True):
+  a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
+  r = a.matmul(b, dtype=dtype_out)
+  sched = r.schedule()
+  realized_ast = sched[-1].ast
+  opts_to_apply = [Opt(OptOps.TC, axis, (tc_select, tc_opt, 1))]
+
+  if ensure_triggered:
+    program = get_program(realized_ast, Device[Device.DEFAULT].renderer, opts=opts_to_apply)
+    wmmas = len([uop for uop in program.uops if uop.op is Ops.WMMA])
+    tcs = len([x for x in program.applied_opts if x.op is OptOps.TC])
+    assert wmmas > 0, "tensor core not triggered"
+    assert tcs == 1, "tensor core opt not included"
+  else:
+    try:
+      program = get_program(realized_ast, Device[Device.DEFAULT].renderer, opts=opts_to_apply)
+      assert False, "OptOps.TC triggered, expected KernelOptError"
+    except KernelOptError: pass
+
+def helper_tc_allclose(N:int, M:int, K:int, dtype_in:DType, dtype_out:DType, axis:int=0, tc_select:int=-1, tc_opt:int=0, use_tensor_cores:int=1):
+  a, b = Tensor.rand(M, K, dtype=dtype_in), Tensor.rand(K, N, dtype=dtype_in)
+  np_a, np_b = a.numpy(), b.numpy()
+  r = a.matmul(b, dtype=dtype_out)
+  if dtype_in == dtypes.bfloat16: r = r.float()
+  realized_ast, bufs = helper_realized_ast(r)
+  opts = [Opt(op=OptOps.TC, axis=axis, arg=(tc_select, tc_opt, use_tensor_cores))]
+  prg = CompiledRunner(replace(get_program(realized_ast, opts=opts), device=Device.DEFAULT))
+  if use_tensor_cores == 1: assert len([uop for uop in prg.p.uops if uop.op is Ops.WMMA]) > 0, "wmma not triggered"
+  assert len([x for x in prg.p.uops[-1].arg.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
+  prg.exec(bufs)
+  if dtype_in == dtypes.half: tc_atol, tc_rtol = 1e-2, 1e-3
+  elif dtype_in == dtypes.bfloat16: tc_atol, tc_rtol = 1e-2, 1e-2
+  else: tc_atol, tc_rtol = 5e-3, 1e-4
+  c = bufs[0].numpy().reshape((M,N))
+  np.testing.assert_allclose(c, np_a @ np_b, atol=tc_atol, rtol=tc_rtol)
+
+class TestTensorCores(unittest.TestCase):
+  # TODO: don't skip bf16 for real device (METAL, AMD)
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  def test_tensor_cores(self):
+    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
+      if not is_dtype_supported(tc.dtype_in) or not is_dtype_supported(tc.dtype_out): continue
+      # for AMX, tc.dims[2] == 1 so reduceop is None thus tensor_cores are not triggered
+      helper_tc_allclose(tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)
+
+  @unittest.skipIf(Device.DEFAULT == "PYTHON", "not generated on EMULATED device")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  def test_tensor_cores_codegen(self):
+    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
+      if not is_dtype_supported(tc.dtype_in) or not is_dtype_supported(tc.dtype_out): continue
+      n, m, k = tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2]
+      a, b = Tensor.rand(m, k, dtype=tc.dtype_in), Tensor.rand(k, n, dtype=tc.dtype_in)
+      r = a.matmul(b, dtype=tc.dtype_out)
+      prg = get_program(r.schedule()[-1].ast, opts=[Opt(op=OptOps.TC, axis=0, arg=(-1, 2, 1))])
+      if Device.DEFAULT == "CPU" and CPU_LLVM:
+        assert "0x201000" in prg.src
+      elif Device.DEFAULT == "AMD" and AMD_LLVM:
+        assert "@llvm.amdgcn.wmma" in prg.src
+      elif Device[Device.DEFAULT].renderer.suffix == "PTX":
+        assert "mma.sync.aligned" in prg.src
+      else:
+        assert "__WMMA_" in prg.src
+
+  @unittest.skipIf((Device.DEFAULT == "AMD") or (Device.DEFAULT == "PYTHON" and Device.default.renderer.device == "AMD"), "broken for AMD")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  def test_tensor_cores_padded(self):
+    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
+      if not is_dtype_supported(tc.dtype_in) or not is_dtype_supported(tc.dtype_out): continue
+      helper_tc_allclose(tc.dims[0]+(pad:=1), tc.dims[1]+pad, tc.dims[2]+pad, tc.dtype_in, tc.dtype_out, tc_opt=2)
+
+  # AMD compiler bug: AMD miscompiles non-zero padded tc kernels with -O3, producing wrong results, nans or hang (see #9606)
+  # Internal bug: zero-stride dimensions combined with a mask may produce wrong index/valid for pad == 1 on AMD
+  @unittest.skipUnless((Device.DEFAULT == "AMD") or (Device.DEFAULT == "PYTHON" and Device.default.renderer.device == "AMD"), "test for AMD's tc")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  @unittest.skip("warp elements not duplicated properly across lanes")
+  def test_tensor_cores_padded_amd(self):
+    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
+      if not is_dtype_supported(tc.dtype_in) or not is_dtype_supported(tc.dtype_out): continue
+      helper_tc_allclose(tc.dims[0]+(pad:=1), tc.dims[1]+pad, tc.dims[2]+pad, tc.dtype_in, tc.dtype_out, tc_opt=2)
+
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  def test_tensor_cores_padded_uops(self):
+    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
+      pad = 1
+
+      # check that TC is triggered for TC_OPT=2
+      helper_tc_ensure_uops_and_opts_count(tc.dims[0]+pad, tc.dims[1]+pad, tc.dims[2]+pad,
+                                           tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=True)
+
+      # check that TC is not triggered for TC_OPT<2
+      helper_tc_ensure_uops_and_opts_count(tc.dims[0]+pad, tc.dims[1]+pad, tc.dims[2]+pad,
+                                           tc.dtype_in, tc.dtype_out, tc_opt=1, ensure_triggered=False)
+      helper_tc_ensure_uops_and_opts_count(tc.dims[0]+pad, tc.dims[1]+pad, tc.dims[2]+pad,
+                                           tc.dtype_in, tc.dtype_out, tc_opt=0, ensure_triggered=False)
+
+      # check excessive padding doesn't trigger padded TC in TC_OPT=2
+      helper_tc_ensure_uops_and_opts_count(tc.dims[0]//4, tc.dims[1], tc.dims[2], tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=False)
+      helper_tc_ensure_uops_and_opts_count(tc.dims[0], tc.dims[1]//4, tc.dims[2], tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=False)
+      if not AMX: # AMX tc.dims[2] == 1
+        helper_tc_ensure_uops_and_opts_count(tc.dims[0], tc.dims[1], tc.dims[2]//8, tc.dtype_in, tc.dtype_out, tc_opt=2, ensure_triggered=False)
+
+  @unittest.skipIf(Device.DEFAULT == "PYTHON", "not generated on EMULATED device")
+  @unittest.skipIf(CI and Device.DEFAULT in {"AMD"}, "AMD CI is really slow here")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  def test_tensor_cores_multi_reduce(self):
+    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
+      if not is_dtype_supported(tc.dtype_in) or not is_dtype_supported(tc.dtype_out): continue
+      if tc.dtype_in is dtypes.bfloat16: continue # <-- broken with numpy
+      # this will be a M=G16, N=G32, M=G16, M=G16, K=R16, K=R16, K=R16 with 9 choices of TC MNK axes
+      golden_result = None
+      for axis in range(9):
+        a = Tensor.rand(16, 16, 29, 29, dtype=tc.dtype_in).realize()
+        b = Tensor.rand(32, 16, 16, 16, dtype=tc.dtype_in).realize()
+        c = a.conv2d(b, padding=1, dtype=tc.dtype_out)
+        realized_ast, real_bufs = helper_realized_ast(c)
+
+        program = get_program(realized_ast, Device[Device.DEFAULT].renderer, opts=[Opt(OptOps.TC, axis, (-1, 2, 1))])
+        assert len([uop for uop in program.uops if uop.op is Ops.WMMA]) > 0, "tensor core not triggered"
+        assert len([x for x in program.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
+
+        prg = CompiledRunner(program)
+        # TODO: support this even if numpy doesn't
+        if _to_np_dtype(real_bufs[0].dtype) is None: continue
+        real_bufs[0].copyin(np.zeros((real_bufs[0].size, ), dtype=_to_np_dtype(real_bufs[0].dtype)).data) # Zero to check that all values are filled
+        prg.exec(real_bufs)
+        result = np.frombuffer(real_bufs[0].as_buffer(), _to_np_dtype(real_bufs[0].dtype))
+
+        # ensure the results for each choice of axis matches
+        if golden_result is None: golden_result = np.frombuffer(real_bufs[0].as_buffer(), _to_np_dtype(real_bufs[0].dtype))
+        np.testing.assert_allclose(result, golden_result, atol=0.1, rtol=0.2)
+
+  @unittest.skipIf(Device.DEFAULT == "PYTHON", "slow on EMULATED device")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  def test_tensor_cores_unroll_phi(self):
+    tc = Device[Device.DEFAULT].renderer.tensor_cores[0]
+    x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
+    r = x.matmul(y, dtype=tc.dtype_out)
+    opts = [Opt(OptOps.UNROLL, 0, 4)]
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
+    for u in get_program(ast, opts=opts).uops:
+      if u.op is Ops.WMMA:
+        assert u.src[-1].src[0].op != Ops.STORE
+
+  @unittest.skipIf(Device.DEFAULT == "PYTHON", "slow on EMULATED device")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "CPU does not support using a different type for accumulation")
+  def test_tensor_cores_unroll_casted_phi(self):
+    tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
+    x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
+    r = x.matmul(y, dtype=tc.dtype_out)
+    opts = [Opt(OptOps.UNROLL, 0, 4)]
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
+    for u in get_program(ast, opts=opts).uops:
+      if u.op is Ops.WMMA:
+        #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
+        assert u.src[-1].src[0].op != Ops.STORE
+
+  @unittest.skipIf(Device.DEFAULT == "PYTHON", "slow on EMULATED device")
+  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
+  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "CPU does not support using a different type for accumulation")
+  def test_tensor_cores_unroll_casted_phi_with_children(self):
+    # all STORE children are outside the loop
+    tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
+    x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
+    r = x.matmul(y, dtype=tc.dtype_out).relu()
+    opts = [Opt(OptOps.UNROLL, 0, 4)]
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
+    for u in get_program(ast, opts=opts).uops:
+      if u.op is Ops.WMMA:
+        #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
+        assert u.src[-1].src[0].op != Ops.STORE
+
+if __name__ == '__main__':
+  unittest.main()
@@ -1,9 +1,9 @@
 import unittest, numpy as np
 from tinygrad import Tensor, Device, TinyJit
-from tinygrad.helpers import Timing, CI, OSX
+from tinygrad.helpers import Timing, CI, OSX, getenv
 import multiprocessing.shared_memory as shared_memory

-N = 256
+N = getenv("NSZ", 256)
 class TestCopySpeed(unittest.TestCase):
  @classmethod
  def setUpClass(cls): Device[Device.DEFAULT].synchronize()
@@ -54,30 +54,32 @@ class TestCopySpeed(unittest.TestCase):
    @TinyJit
    def _do_copy(t): return t.to('CPU').realize()

-    t = Tensor.randn(N, N, 4).contiguous().realize()
+    t = Tensor.randn(N, N).contiguous().realize()
+    Device[Device.DEFAULT].synchronize()
    for _ in range(5):
-      with Timing("sync:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
+      with Timing(f"copy {Device.DEFAULT} -> CPU {t.nbytes()/(1024**2)}M:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
        x = _do_copy(t)
        Device[Device.DEFAULT].synchronize()
      np.testing.assert_equal(t.numpy(), x.numpy())

-  def testCopytoCPUtoDefaultJit(self):
+  def testCopyCPUtoDefaultJit(self):
    if Device.DEFAULT == "CPU": return unittest.skip("CPU to CPU copy is a no-op")

    @TinyJit
-    def _do_copy(x): return t.to(Device.DEFAULT).realize()
+    def _do_copy(x): return x.to(Device.DEFAULT).realize()

    for _ in range(5):
-      t = Tensor.randn(N, N, 4, device="CPU").contiguous().realize()
-      with Timing("sync:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
+      t = Tensor.randn(N, N, device="CPU").contiguous().realize()
+      Device["CPU"].synchronize()
+      with Timing(f"copy CPU -> {Device.DEFAULT} {t.nbytes()/(1024**2)}M:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s"):
        x = _do_copy(t)
        Device[Device.DEFAULT].synchronize()
      np.testing.assert_equal(t.numpy(), x.numpy())

  @unittest.skipIf(CI, "CI doesn't have 6 GPUs")
-  @unittest.skipIf(Device.DEFAULT != "GPU", "only test this on GPU")
+  @unittest.skipIf(Device.DEFAULT != "CL", "only test this on CL")
  def testCopyCPUto6GPUs(self):
-    from tinygrad.runtime.ops_gpu import CLDevice
+    from tinygrad.runtime.ops_cl import CLDevice
    if len(CLDevice.device_ids) != 6: raise unittest.SkipTest("computer doesn't have 6 GPUs")
    t = Tensor.ones(N, N, device="CPU").contiguous().realize()
    print(f"buffer: {t.nbytes()*1e-9:.2f} GB")
@@ -85,8 +87,8 @@ class TestCopySpeed(unittest.TestCase):
      with Timing("sync:  ", on_exit=lambda ns: f" @ {t.nbytes()/ns:.2f} GB/s ({t.nbytes()*6/ns:.2f} GB/s total)"):
        with Timing("queue: "):
          for g in range(6):
-            t.to(f"gpu:{g}").realize()
-        Device["gpu"].synchronize()
+            t.to(f"CL:{g}").realize()
+        Device["CL"].synchronize()

 if __name__ == '__main__':
  unittest.main()
@@ -1,75 +1,29 @@
-import unittest, contextlib
+import unittest
 import numpy as np
 from tinygrad import Tensor, GlobalCounters, dtypes, nn, Device, Variable
-from tinygrad.helpers import CI, Context, getenv
+from tinygrad.helpers import CI, Context, getenv, RANGEIFY
 from tinygrad.engine.realize import run_schedule
-from tinygrad.codegen.opt.kernel import Opt, OptOps, Kernel, KernelOptError
 from tinygrad.engine.realize import CompiledRunner, ExecItem, get_program
-from tinygrad.codegen.opt.search import get_kernel_actions
 from tinygrad.uop.ops import Ops
-from tinygrad.codegen import apply_rewrites, rewrites_for_views

 class TestArange(unittest.TestCase):
-  def _get_flops(self, N, opts=None):
+  def _get_flops(self, N):
    GlobalCounters.reset()
    tt = Tensor.arange(N)
    sched = tt.schedule()
    self.assertEqual(len(sched), 1)
-    p = get_program(sched[-1].ast, opts=opts)
-    print(p.name)
-    #print(p.src)
+    p = get_program(sched[-1].ast)
    ExecItem(CompiledRunner(p), [tt.uop.buffer]).run()
    np.testing.assert_equal(tt.numpy(), np.arange(N))
    return p.estimates.ops

-  def test_complexity(self, opts=None, limit=None):
-    f1 = self._get_flops(256, opts)
-    f2 = self._get_flops(2560, opts)
-    print(f"{f1=}, {f2=}")
-    # add 1 to avoid divide by 0. arange is 0 flops now!
-    assert (f1 < 6000 and f2 < 6000) or ((f2+1) / (f1+1) < 16), f"bad complexity, flops {(f2+1) / (f1+1):.1f}X while inputs 10X"
-    if limit is not None and not getenv("PTX"):
-      # PTX counts index ALU in flops
-      assert f1 <= limit, f"{f1=}, {limit=}"
+  def test_complexity(self):
+    self.assertEqual(self._get_flops(256), 0)
+    self.assertEqual(self._get_flops(2560), 0)

-  def test_complexity_w_upcast(self): return self.test_complexity([Opt(OptOps.UPCAST, 0, 4)], limit=0)
-  def test_complexity_w_unroll2(self): return self.test_complexity([Opt(OptOps.UNROLL, 0, 2)], limit=0)
-  def test_complexity_w_unroll4(self): return self.test_complexity([Opt(OptOps.UNROLL, 0, 4)], limit=0)
-  def test_complexity_w_unroll8(self): return self.test_complexity([Opt(OptOps.UNROLL, 0, 8)], limit=0)
-  def test_complexity_w_upcast_and_unroll(self): return self.test_complexity([Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 4)], limit=0)
-
-  if Device.default.renderer.has_local:
-    # TODO: fix limit
-    def test_complexity_w_group(self): return self.test_complexity([Opt(OptOps.GROUP, 0, 16)], limit=81920)
-    def test_complexity_w_group_top(self): return self.test_complexity([Opt(OptOps.GROUPTOP, 0, 16)], limit=106496)
-
-    def test_complexity_w_local(self): return self.test_complexity([Opt(OptOps.LOCAL, 0, 16)], limit=0)
-    @unittest.skip("doesn't work yet. TODO: this absolutely should work")
-    def test_complexity_w_local_unroll4(self): return self.test_complexity([Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.UNROLL, 0, 4)], limit=0)
-    @unittest.skip("doesn't work yet")
-    def test_complexity_w_local_and_padto(self): return self.test_complexity([Opt(OptOps.LOCAL, 0, 16), Opt(OptOps.PADTO, axis=1, arg=32)])
-
-  def test_all_opts(self, opts=None, exclude=None):
-    k = Kernel(apply_rewrites(Tensor.arange(256).schedule()[-1].ast, rewrites_for_views))
-    if opts is not None:
-      for o in opts: k.apply_opt(o)
-    all_opts_256 = [kk.applied_opts for kk in get_kernel_actions(k, include_0=False).values()]
-    k = Kernel(apply_rewrites(Tensor.arange(2560).schedule()[-1].ast, rewrites_for_views))
-    if opts is not None:
-      for o in opts: k.apply_opt(o)
-    all_opts_2560 = [kk.applied_opts for kk in get_kernel_actions(k, include_0=False).values()]
-    all_opts = [x for x in all_opts_256 if x in all_opts_2560]
-    for opts in all_opts:
-      if exclude is not None and opts[-1] in exclude: continue
-      print(opts)
-      self.test_complexity(opts)
-  def test_all_opts_w_local(self):
-    with contextlib.suppress(KernelOptError):
-      return self.test_all_opts([Opt(OptOps.LOCAL, 0, 16)], [Opt(op=OptOps.PADTO, axis=1, arg=32)])
-  def test_all_opts_w_upcast(self): return self.test_all_opts([Opt(OptOps.UPCAST, 0, 4)])
-  def test_all_opts_w_unroll(self): return self.test_all_opts([Opt(OptOps.UNROLL, 0, 4)], [Opt(op=OptOps.GROUP, axis=0, arg=0)])
-  def test_all_opts_w_upcast_and_unroll(self):
-    return self.test_all_opts([Opt(OptOps.UPCAST, 0, 4), Opt(OptOps.UNROLL, 0, 4)], [Opt(op=OptOps.GROUP, axis=0, arg=0)])
+  def test_arange_cat(self):
+    t = Tensor.arange(2, dtype=dtypes.int)+Tensor([3])
+    self.assertEqual(t.cat(t).tolist(), [3, 4, 3, 4])

 class TestRand(unittest.TestCase):
  def test_fused_rand_less_ops(self, noopt=1):
@@ -102,7 +56,6 @@ class TestIndexing(unittest.TestCase):
      run_schedule(sched)
    self.assertEqual(out.item(), 1337)

-  @unittest.skipIf(getenv("PTX"), "broken on ptx for some reason")
  def test_manual_index(self):
    dataset = Tensor.rand(DSET, DDIM).realize()
    idxs = Tensor([0,3,5,6]).realize()
@@ -158,7 +111,7 @@ class TestIndexing(unittest.TestCase):
      X = dataset[idxs]
      assert X.shape == (4,DDIM)
      sched = X.schedule()
-      self.assertEqual(len(sched), 2)
+      self.assertEqual(len(sched), 1 if RANGEIFY else 2)
      run_schedule(sched)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops} != {4*DSET}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -172,7 +125,6 @@ class TestIndexing(unittest.TestCase):
      X = dataset[idxs]
      np.testing.assert_equal(X.numpy(), 0)

-  @unittest.skipIf(getenv("PTX"), "broken on ptx for some reason")
  def test_index_mnist(self, noopt=1, op_limit=512*784*13, split_reduceop=0):
    # WEBGPU generates more ops due to bitpacking of < 4-byte dtypes
    if Device.DEFAULT == "WEBGPU": op_limit *= 15
@@ -191,7 +143,6 @@ class TestIndexing(unittest.TestCase):
  def test_index_mnist_split(self): self.test_index_mnist(1, split_reduceop=1)
  def test_index_mnist_opt_split(self): self.test_index_mnist(0, split_reduceop=1)

-  @unittest.skipIf(getenv("PTX"), "broken on ptx for some reason")
  def test_llama_embedding(self, noopt=1, op_limit=65536):
    # llama3 is 128256
    vocab_size, embed_size = (10, 3) if CI else (32000, 4096)
--- a/Show More
+++ b/Show More