fix assign copy

assign after copy shouldn't contig
2026-02-18 12:14:18 +08:00 · 2026-02-18 12:03:20 +08:00
237 changed files with 20222 additions and 10333 deletions
@@ -45,10 +45,6 @@ inputs:
    description: "Install mesa"
    required: false
    default: 'false'
-  tinydreno:
-    description: "Install tinydreno"
-    required: false
-    default: 'false'
 runs:
  using: "composite"
  steps:
@@ -66,14 +62,14 @@ runs:
      uses: actions/cache/restore@v4
      with:
        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
    - name: Cache Python packages
      if: github.event_name != 'pull_request'
      id: restore-venv
      uses: actions/cache@v4
      with:
        path: ${{ github.workspace }}/.venv
-        key: venv-${{ runner.os }}-${{ runner.arch }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}
+        key: venv-${{ runner.os }}-python-${{ steps.setup-python.outputs.python-version }}-${{ inputs.deps }}-${{ inputs.pydeps }}-${{ env.CACHE_VERSION }}

    # **** Caching downloads ****

@@ -199,13 +195,13 @@ runs:
      uses: actions/cache/restore@v4
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
    - name: Cache apt
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true') && github.event_name != 'pull_request'
      uses: actions/cache@v4
      with:
        path: /var/cache/apt/archives/
-        key: ${{ runner.os }}-${{ runner.arch }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-apt-${{ steps.apt-pkgs.outputs.hash }}-${{ env.CACHE_VERSION }}

    - name: Run apt Update + Install
      if: runner.os == 'Linux' && (inputs.opencl == 'true' || inputs.amd == 'true' || inputs.cuda == 'true' || inputs.webgpu == 'true' || inputs.llvm == 'true')
@@ -237,7 +233,7 @@ runs:
      shell: bash
      run: |
        sudo mkdir -p /usr/local/lib
-        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/tinygrad/amdcomgr_dylib/releases/latest | \
+        curl -s -H "Authorization: token $GH_TOKEN" curl -s https://api.github.com/repos/nimlgen/amdcomgr_dylib/releases/latest | \
          jq -r '.assets[] | select(.name == "libamd_comgr.dylib").browser_download_url' | \
          sudo xargs curl -fL -o /usr/local/lib/libamd_comgr.dylib
        cargo build --release --manifest-path ./extra/remu/Cargo.toml
@@ -330,9 +326,3 @@ runs:
      if: inputs.mesa == 'true' && runner.os == 'macOS'
      shell: bash
      run: brew install sirhcm/tinymesa/tinymesa_cpu
-
-    # *** tinydreno ***
-    - name: Install tinydreno (linux)
-      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
-      shell: bash
-      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
@@ -32,7 +32,6 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: 'autogen'
        opencl: 'true'
        amd: 'true'
        cuda: 'true'
@@ -82,7 +81,6 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: 'autogen-mac'
        llvm: 'true'
    - name: Regenerate autogen files
      run: |
@@ -112,8 +110,6 @@ jobs:
      uses: actions/checkout@v4
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: 'autogen-comgr'
    - name: Install autogen support packages
      run: |
        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
@@ -21,9 +21,6 @@ jobs:
  # the 3 minute timeout should not be raised
  testmacpytest:
    name: Mac pytest
-    env:
-      CI: ""
-      CAPTURE_PROCESS_REPLAY: "0"
    runs-on: [self-hosted, macOS]
    timeout-minutes: 3
    defaults:
@@ -44,14 +41,22 @@ jobs:
      run: |
        echo "CACHEDB=/tmp/pytest-db-ci.db" >> $GITHUB_ENV
        rm -f /tmp/pytest-db-ci*
+    # TODO: remove this step once all old caches are migrated
+    - name: Migrate old huggingface cache (symlinks break onnxruntime 1.24+)
+      run: |
+        cd ~/Library/Caches/tinygrad/downloads/models 2>/dev/null || exit 0
+        for old_dir in models--*; do
+          [ -d "$old_dir" ] || continue
+          repo_id=$(echo "$old_dir" | sed 's/models--//; s/--/\//g')
+          snapshot=$(ls -1 "$old_dir/snapshots" 2>/dev/null | head -1)
+          [ -n "$snapshot" ] || continue
+          mkdir -p "$repo_id"
+          cp -RLn "$old_dir/snapshots/$snapshot/"* "$repo_id/" 2>/dev/null || true
+        done
    - name: Run pytest -nauto
      run: |
        source /tmp/tinygrad_pytest_ci/bin/activate
        pytest -nauto --durations=20
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: FLOAT16=1 CL=1 IMAGE=2 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: IMAGE=1 openpilot compile3 0.10.1 driving_vision
-      run: FLOAT16=1 CL=1 IMAGE=1 python3.11 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx

  testmacbenchmark:
    name: Mac Benchmark
@@ -332,13 +337,13 @@ jobs:
    # - name: Fuzz Padded Tensor Core GEMM (PTX)
    #   run: NV=1 NV_PTX=1 M_START=12 M_STOP=20 M_STEP=1 N_START=6 N_STOP=10 N_STEP=1 K_START=28 K_STOP=36 K_STEP=1 HALF=1 TC_OPT=2 python3 ./extra/gemm/fuzz_matmul.py
    - name: HEVC Decode Benchmark
-      run: VALIDATE=1 MAX_FRAMES=100 ASSERT_FPS=1400 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
+      run: VALIDATE=1 MAX_FRAMES=100 JITBEAM=1 NV=1 PYTHONPATH=. python3 extra/hevc/decode.py
    - name: Train MNIST
      run: time PYTHONPATH=. NV=1 TARGET_EVAL_ACC_PCT=96.0 python3 examples/beautiful_mnist.py
    - name: Run 10 CIFAR training steps
      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 python3 examples/hlb_cifar10.py
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=110 NV=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
    - name: Run 10 CIFAR training steps w BF16
      run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=120 NV=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
    # - name: Run 10 CIFAR training steps w winograd
@@ -510,7 +515,7 @@ jobs:
    - name: Run 10 CIFAR training steps
      run: BENCHMARK_LOG=cifar_10steps ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 python3 examples/hlb_cifar10.py
    - name: Run 10 CIFAR training steps w HALF
-      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=230 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
+      run: BENCHMARK_LOG=cifar_10steps_half ASSERT_MIN_STEP_TIME=200 AMD=1 STEPS=10 DEFAULT_FLOAT=HALF python3 examples/hlb_cifar10.py
    # - name: Run 10 CIFAR training steps w BF16
    #   run: BENCHMARK_LOG=cifar_10steps_bf16 ASSERT_MIN_STEP_TIME=288 AMD=1 STEPS=10 DEFAULT_FLOAT=BFLOAT16 python3 examples/hlb_cifar10.py
    # TODO: too slow
@@ -520,9 +525,8 @@ jobs:
      run: time BENCHMARK_LOG=cifar AMD=1 DEFAULT_FLOAT=HALF STEPS=1000 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
    - name: Run full CIFAR training steps w 6 GPUS
      run: time BENCHMARK_LOG=cifar_6gpu AMD=1 DEFAULT_FLOAT=HALF STEPS=350 BS=1536 GPUS=6 TARGET_EVAL_ACC_PCT=93.0 python3 examples/hlb_cifar10.py
-    # TODO: broken on some of the machines
-    #- name: Test full tinyfs load
-    #  run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
+    - name: Test full tinyfs load
+      run: TINYFS_ENDPOINT=10.0.52.11:6767 PYTHONPATH=. python extra/tinyfs/fetch_file.py --hash d734f5e3be9f1e9d863bfaa4fc6c1ef2 --len 175866113 --dest mapping.json --check
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

@@ -617,27 +621,6 @@ jobs:
    - name: Run process replay tests
      run: cp test/external/process_replay/process_replay.py ./process_replay.py && git fetch origin master && git -c advice.detachedHead=false checkout origin/master && PYTHONPATH=. python3 process_replay.py

-  testcommausbgpubenchmark:
-    name: UsbGPU Benchmark (comma)
-    runs-on: [self-hosted, Linux, comma4]
-    timeout-minutes: 20
-    defaults:
-      run:
-        shell: bash -e -o pipefail {0}
-    if: github.repository_owner == 'tinygrad'
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: setup staging db
-      if: github.ref == 'refs/heads/update_benchmark_staging'
-      run: |
-        echo "CACHEDB=/tmp/staging.db" >> $GITHUB_ENV
-        rm -f /tmp/staging.db /tmp/staging.db-shm /tmp/staging.db-wal
-    - name: openpilot compile3 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision PYTHONPATH="." DEV=AMD AMD_LLVM=1 AMD_IFACE=USB ASSERT_MIN_STEP_TIME=50 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx
-    - name: openpilot load_pickle 0.10.1 driving_vision
-      run: BENCHMARK_LOG=usbgpu_openpilot_0_10_1_vision_load_pickle PYTHONPATH="." DEV=AMD AMD_IFACE=USB ASSERT_MIN_LOAD_TIME=15 python3 examples/openpilot/load_pickle.py
-
  testreddriverbenchmark:
    name: AM Benchmark
    runs-on: [self-hosted, Linux, tinyboxrandom]
@@ -1,7 +1,7 @@
 name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '18'
+  CACHE_VERSION: '16'
  CAPTURE_PROCESS_REPLAY: 1
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
@@ -244,37 +244,6 @@ jobs:
    - name: Run TYPED=1
      run: CHECK_OOB=0 DEV=CPU TYPED=1 python test/test_tiny.py

-  nulltest:
-    name: Null Tests
-    runs-on: ubuntu-latest
-    timeout-minutes: 15
-
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v4
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-13
-        pydeps: "pillow ftfy regex pre-commit"
-        deps: testing_unit
-        llvm: 'true'
-        amd: 'true'
-    - name: Run NULL backend tests
-      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
-    - name: Run targetted tests on NULL backend
-      run: NULL=1 python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
-    # TODO: too slow
-    # - name: Run SDXL on NULL backend
-    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
-    - name: Run Clip tests for SD MLPerf on NULL backend
-      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
-    - name: Run AMD emulated BERT training on NULL backend
-      run: EMULATE=AMD_RDNA4 NULL=1 NULL_ALLOW_COPYOUT=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
-    # TODO: support fake weights
-    #- name: Run LLaMA 7B on 4 fake devices
-    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
-
  unittest:
    name: Unit Tests
    runs-on: ubuntu-latest
@@ -299,6 +268,20 @@ jobs:
      run: |
        CPU=1 python test/null/test_device.py TestRunAsModule.test_module_runs
        CPU=1 python -m pytest -n=auto test/unit/ --durations=20
+    - name: Run NULL backend tests
+      run: NULL=1 python -m pytest -n=auto test/null/ --durations=20
+    - name: Run targetted tests on NULL backend
+      run: NULL=1 python3 -m unittest test.backend.test_multitensor.TestMultiTensor.test_data_parallel_resnet_train_step
+    # TODO: too slow
+    # - name: Run SDXL on NULL backend
+    #   run: NULL=1 DEBUG=1 python3 examples/sdxl.py --seed 0 --noshow --timing --fakeweights
+    - name: Run Clip tests for SD MLPerf on NULL backend
+      run: NULL=1 python -m pytest -n=auto test/external/mlperf_stable_diffusion/external_test_models.py::TestOpenClip --durations=20
+    - name: Run AMD emulated BERT training on NULL backend
+      run: EMULATE=AMD_RDNA4 NULL=1 NULL_ALLOW_COPYOUT=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
+    # TODO: support fake weights
+    #- name: Run LLaMA 7B on 4 fake devices
+    #  run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
    - name: Run GC tests
      run: python test/external/external_uop_gc.py
    - name: External Benchmark Schedule
@@ -661,13 +644,15 @@ jobs:
          sudo apt-get update
          sudo apt-get install llvm-21 llvm-21-tools cloc
      - name: Install rocprof-trace-decoder
-        run: sudo PYTHONPATH="." ./extra/sqtt/install_rocprof_decoder.py
+        run: sudo PYTHONPATH="." ./extra/sqtt/install_sqtt_decoder.py
      - name: Run AMD renderer tests
        run: AMD_LLVM=0 python -m pytest -n=auto test/amd/ --durations 20
      - name: Run AMD renderer tests (AMD_LLVM=1)
        run: AMD_LLVM=1 python -m pytest -n=auto test/amd/ --durations 20
-      - name: Run SQTT profiling tests
-        run: PROFILE=1 SQTT=1 python3 -m pytest -n=auto test/amd/test_sqtt_profiler.py
+      - name: Run TestOps.test_add with SQTT
+        run: |
+          VIZ=-2 DEBUG=5 python3 test/backend/test_ops.py TestOps.test_add
+          extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
      - name: Run AMD emulated tests on NULL backend
        env:
          AMD: 0
@@ -679,30 +664,6 @@ jobs:
      - name: Run LLVM test
        run: AMD_LLVM=1 python test/device/test_amd_llvm.py

-  testmockam:
-    name: Linux (am)
-    runs-on: ubuntu-24.04
-    timeout-minutes: 15
-    env:
-      AMD: 1
-      MOCKGPU: 1
-      AMD_IFACE: PCI
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: mockam
-          deps: testing_unit
-          amd: 'true'
-      - name: Run test_tiny on MOCKAM
-        run: python test/test_tiny.py
-      - name: Run test_tiny on MOCKAM USB
-        run: AMD_IFACE=USB python test/test_tiny.py
-      - name: Run test_hcq on MOCKAM
-        run: python -m pytest test/device/test_hcq.py
-
  testamd:
    strategy:
      fail-fast: false
@@ -841,8 +802,6 @@ jobs:
      run: METAL=1 DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
    - name: Test Beam Search
      run: METAL=1 IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
-    - name: Test Device Specific
-      run: METAL=1 python3 -m pytest test/device/test_metal.py
    #- name: Fuzz Test linearizer
    #  run: METAL=1 DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
    - name: Run TRANSCENDENTAL math
@@ -1011,26 +970,3 @@ jobs:
          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
          python -m pytest -n=auto test/backend/test_ops.py --durations=20
-  qcomclcompiletests:
-    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v4
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: compile-qcomcl
-          deps: testing_unit
-          tinydreno: 'true'
-          python-version: '3.12'
-      - name: Set env
-        shell: bash
-        run: printf "NULL=1\nNULL_ALLOW_COPYOUT=1\nNULL_QCOMCL=1" >> $GITHUB_ENV
-      - name: Run test_ops
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
-          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
-          python -m pytest -n=auto test/backend/test_ops.py --durations=20
@@ -66,5 +66,3 @@ target
 .mypy_cache
 mutants
 .mutmut-cache
-dagre/
-graphlib/
@@ -10,7 +10,7 @@ Directories are listed in order of how they are processed.

 Group UOps into kernels.

-::: tinygrad.schedule.rangeify.get_kernel_graph
+::: tinygrad.schedule.rangeify.get_rangeify_map
    options:
        members: false
        show_labels: false
@@ -19,8 +19,8 @@ cifar_std = [0.24703225141799082, 0.24348516474564, 0.26158783926049628]
 BS, STEPS = getenv("BS", 512), getenv("STEPS", 1000)
 EVAL_BS = getenv("EVAL_BS", BS)
 GPUS = [f'{Device.DEFAULT}:{i}' for i in range(getenv("GPUS", 1))]
-assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}"
-assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}"
+assert BS % len(GPUS) == 0, f"{BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"
+assert EVAL_BS % len(GPUS) == 0, f"{EVAL_BS=} is not a multiple of {len(GPUS)=}, uneven multi GPU is slow"

 class UnsyncedBatchNorm:
  def __init__(self, sz:int, eps=1e-5, affine=True, track_running_stats=True, momentum=0.1, num_devices=len(GPUS)):
@@ -65,7 +65,17 @@ def loader_process(q_in, q_out, X:Tensor, seed):
      else:
        # pad data with training mean
        img = np.tile(np.array([[[123.68, 116.78, 103.94]]], dtype=np.uint8), (224, 224, 1))
-      X[idx].flatten().assign(img.tobytes())
+
+      # broken out
+      #img_tensor = Tensor(img.tobytes(), device='CPU')
+      #storage_tensor = X[idx].contiguous().realize().lazydata.base.realized
+      #storage_tensor._copyin(img_tensor.numpy())
+
+      # faster
+      X[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = img.tobytes()
+
+      # ideal
+      #X[idx].assign(img.tobytes())   # NOTE: this is slow!
      q_out.put(idx)
    q_out.put(None)

@@ -254,8 +264,8 @@ def load_unet3d_data(preprocessed_dataset_dir, seed, queue_in, queue_out, X:Tens
      x = random_brightness_augmentation(x)
      x = gaussian_noise(x)

-    X[idx].flatten().assign(x.tobytes())
-    Y[idx].flatten().assign(y.tobytes())
+    X[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = x.tobytes()
+    Y[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = y.tobytes()

    queue_out.put(idx)
  queue_out.put(None)
@@ -369,12 +379,12 @@ def load_retinanet_data(base_dir:Path, val:bool, queue_in:Queue, queue_out:Queue
      clipped_match_idxs = np.clip(match_idxs, 0, None)
      clipped_boxes, clipped_labels = tgt["boxes"][clipped_match_idxs], tgt["labels"][clipped_match_idxs]

-      boxes[idx].flatten().assign(clipped_boxes.tobytes())
-      labels[idx].flatten().assign(clipped_labels.tobytes())
-      matches[idx].flatten().assign(match_idxs.tobytes())
-      anchors[idx].flatten().assign(anchor.tobytes())
+      boxes[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_boxes.tobytes()
+      labels[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = clipped_labels.tobytes()
+      matches[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = match_idxs.tobytes()
+      anchors[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = anchor.tobytes()

-    imgs[idx].flatten().assign(img.tobytes())
+    imgs[idx].contiguous().realize().uop.base.realized.as_memoryview(force_zero_copy=True)[:] = img.tobytes()

    queue_out.put(idx)
  queue_out.put(None)
@@ -396,7 +406,6 @@ def batch_load_retinanet(dataset, val:bool, base_dir:Path, batch_size:int=32, sh
      queue_in.put((idx, img, tgt))

  def _setup_shared_mem(shm_name:str, size:tuple[int, ...], dtype:dtypes) -> tuple[shared_memory.SharedMemory, Tensor]:
-    shm_name = f"{shm_name}_{os.getpid()}"
    if os.path.exists(f"/dev/shm/{shm_name}"): os.unlink(f"/dev/shm/{shm_name}")
    shm = shared_memory.SharedMemory(name=shm_name, create=True, size=prod(size))
    shm_tensor = Tensor.empty(*size, dtype=dtype, device=f"disk:/dev/shm/{shm_name}")
@@ -3,7 +3,7 @@ from pathlib import Path
 import multiprocessing

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

@@ -1282,7 +1282,7 @@ def train_bert():
        previous_step = i

 def train_llama3():
-  from examples.mlperf.models.llama import Transformer
+  from extra.models.llama import Transformer
  from examples.llama3 import MODEL_PARAMS
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
  from examples.mlperf.optim import GradAccClipAdamW
@@ -1335,16 +1335,10 @@ def train_llama3():
  model_params = MODEL_PARAMS[getenv("LLAMA3_SIZE", "8B")]["args"]
  # vocab_size from the mixtral tokenizer
  if not SMALL: model_params |= {"vocab_size": 32000}
-  real_vocab_size = model_params['vocab_size']
  if (llama_layers:=getenv("LLAMA_LAYERS")) != 0: model_params['n_layers'] = llama_layers
  print(f"model parameters: {model_params}")

-  # pad vocab
-  if (MP := getenv("MP", 1)) > 1: model_params['vocab_size'] = round_up(model_params['vocab_size'], 256 * MP)
-  vocab_mask:Tensor = Tensor.arange(model_params['vocab_size']).reshape(1, 1, -1) >= real_vocab_size
-
-  model = Transformer(**model_params, max_context=SEQLEN)
-
+  model = Transformer(**model_params, max_context=SEQLEN, jit=False, disable_kv_cache=True)
  params = get_parameters(model)
  # weights are all bfloat16 for now
  assert params and all(p.dtype == dtypes.bfloat16 for p in params)
@@ -1358,8 +1352,6 @@ def train_llama3():
    for v in get_parameters(model):
      v.shard_(device, axis=None)

-    vocab_mask.shard_(device, axis=None)
-
  if (MP := getenv("MP", 1)) > 1:
    device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
    for k,v in get_state_dict(model).items():
@@ -1367,7 +1359,6 @@ def train_llama3():
      elif '.attention.wq' in k: v.shard_(device, axis=0)
      elif '.attention.wk' in k: v.shard_(device, axis=0)
      elif '.attention.wv' in k: v.shard_(device, axis=0)
-      elif '.attention.wqkv' in k: v.shard_(device, axis=0)
      elif '.attention.wo' in k: v.shard_(device, axis=1)
      elif '.feed_forward.w1.' in k: v.shard_(device, axis=0)
      elif '.feed_forward.w2.' in k: v.shard_(device, axis=1)
@@ -1380,21 +1371,12 @@ def train_llama3():
      # prevents memory spike on device 0
      v.realize()

-    vocab_mask.shard_(device, axis=2).realize()
-
-  is_offload_optim = bool(getenv("OFFLOAD_OPTIM"))
-  is_fake_offload = Device.DEFAULT == "NULL"
-  optim_device = ("CPU" if not is_fake_offload else "NULL:99") if is_offload_optim else None
-  optim = GradAccClipAdamW(get_parameters(model), lr=0.0, b1=opt_adamw_beta_1, b2=opt_adamw_beta_2,
-                           eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc, device=optim_device)
+  optim = GradAccClipAdamW(get_parameters(model), lr=0.0,
+                b1=opt_adamw_beta_1, b2=opt_adamw_beta_2, eps=opt_adamw_epsilon, weight_decay=opt_adamw_weight_decay, grad_acc=grad_acc)

  # init grads
-  if is_offload_optim:
-    for p in optim.params:
-      p.grad = Tensor.zeros(p.shape, dtype=p.dtype, device=optim_device, requires_grad=False).contiguous().realize()
-  else:
-    for p in optim.params:
-      p.grad = p.zeros_like().contiguous().realize()
+  for p in optim.params:
+    p.grad = p.zeros_like().contiguous().realize()
  grads: list[Tensor] = [p.grad for p in optim.params]

  scheduler = CosineAnnealingLRWithWarmup(optim, opt_base_learning_rate, opt_end_learning_rate, opt_learning_rate_warmup_steps, opt_learning_rate_decay_steps)
@@ -1410,55 +1392,51 @@ def train_llama3():

  @TinyJit
  def minibatch(tokens:Tensor):
+    tokens = tokens.to(None)
    if (DP := getenv("DP", 1)) > 1:
      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
-      tokens = tokens.to(None).shard(device, 0)
+      tokens = tokens.shard(device, 0)
    if (MP := getenv("MP", 1)) > 1:
      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
      tokens = tokens.shard(device)
-    if DP == 1 and MP == 1: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1])
-    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
+    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
+    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
    loss.backward()
    assert all(p.grad is g for p,g in zip(optim.params, grads))
-    loss_cpu = loss.flatten().float().to("CPU")
-    Tensor.realize(loss_cpu, *grads)
-    return loss_cpu
+    Tensor.realize(loss, *grads)
+    return loss.flatten().float().to("CPU")

  @TinyJit
  def optim_step():
-    grad_norm = optim.fstep(grads)
+    optim.step()
    scheduler.step()

    for g in grads:
      g.assign(g.zeros_like())

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

-    return lr_cpu, grad_norm_cpu
+    return lr.float().to("CPU")

  @TinyJit
  @Tensor.train(False)
  def eval_step(tokens:Tensor):
+    tokens = tokens.to(None)
    if (DP := getenv("DP", 1)) > 1:
      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(DP))
-      tokens = tokens.to(None).shard(device, 0)
+      tokens = tokens.shard(device, 0)
    if (MP := getenv("MP", 1)) > 1:
      device = tuple(f"{Device.DEFAULT}:{i}" for i in range(MP))
      tokens = tokens.shard(device)
-    if DP == 1 and MP == 1: tokens = tokens.to(None)
-    logits:Tensor = model(tokens[:, :-1])
-    loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
+    logits:Tensor = model(tokens[:, :-1], start_pos=0, temperature=math.nan)
+    loss = logits.sparse_categorical_crossentropy(tokens[:, 1:])
    return loss.flatten().float().to("CPU")

  # ** data iters **
  def fake_data(bs, samples):
-    import numpy as np
    for _ in range(samples // bs):
-      fake_data_np = np.random.randint(0, model_params["vocab_size"], size=(bs, SEQLEN + 1), dtype=np.int32)
-      yield Tensor(fake_data_np, device="NPY")
+      yield Tensor.randint(bs, SEQLEN + 1, low=0, high=model_params["vocab_size"], dtype=dtypes.int32, device=Device.DEFAULT)

  def get_train_iter():
    if getenv("FAKEDATA", 0):
@@ -1485,53 +1463,49 @@ def train_llama3():
  step_times = []
  while i < MAX_STEPS:
    GlobalCounters.reset()
-    actual_gbs = GBS if i >= 2 else BS
    if getenv("TRAIN", 1):
      profile_marker(f"train @ {i}")
      st = time.perf_counter()

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

      gt = time.perf_counter()
-      ret = optim_step()
-      lr, grad_norm = ret[0].item(), ret[1].item()
-      et = time.perf_counter()
+      lr = optim_step()
+      ot = time.perf_counter()

-      loss = sum(losses) / len(losses)
-      optim_time = et - gt
-      dev_time += optim_time
+      loss = loss.float().item()
+      lr = lr.item()
+
+      et = time.perf_counter()
      step_time = et - st
      gbs_time = gt - st
+      optim_time = ot - gt
+      data_time = dt - ist
+      dev_time = step_time - data_time * grad_acc
      if BENCHMARK: step_times.append(step_time)

      i += 1
-      sequences_seen += actual_gbs
+      sequences_seen += GBS

      mem_gb = GlobalCounters.mem_used / 1e9
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * max(getenv("DP", 1), getenv("MP", 1)) * 2.3e15)) * 100
      tqdm.write(
          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
-          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
-      if DEBUG >= 1: tqdm.write("  mem per device: " + ', '.join(f"{dev}: {mem/1e9:.2f} GB" for dev, mem in sorted(GlobalCounters.mem_used_per_device.items())))
+          f"{lr:.12f} LR, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")

      if WANDB:
        wandb.log({
-          "train/loss": loss,
-          "train/lr": lr,
-          "train/grad_norm": grad_norm,
+          "lr": lr, "train/loss": loss,
          "train/step_time": step_time,
          "train/gbs_time": gbs_time,
          "train/optim_time": optim_time,
@@ -1560,7 +1534,7 @@ def train_llama3():
        print(f"epoch global_ops: {GlobalCounters.global_ops:_}, "
              f"epoch global_mem: {GlobalCounters.global_mem:_}")

-    if (sequences_seen // EVAL_FREQ != (sequences_seen - actual_gbs) // EVAL_FREQ and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
+    if (sequences_seen % EVAL_FREQ == 0 and (i != 1 or EVAL_FREQ == 1)) or (BENCHMARK and i == BENCHMARK):
      if EVAL_BS == 0: return
      tqdm.write(f"evaluating after {sequences_seen} sequences")
      profile_marker(f"eval @ {i}")
@@ -1568,7 +1542,7 @@ def train_llama3():
      # run eval
      eval_losses = []
      eval_iter = get_eval_iter()
-      tqdm.write(f"evaluating {EVAL_SAMPLES//EVAL_BS} batches of {EVAL_BS} sequences")
+      tqdm.write(f"evaluating {5760//EVAL_BS} batches of {EVAL_BS} sequences")

      for j,tokens in tqdm(enumerate(eval_iter), total=EVAL_SAMPLES//EVAL_BS):
        eval_losses += eval_step(tokens).tolist()
@@ -1,80 +0,0 @@
-from tinygrad import Tensor, nn
-from tinygrad.helpers import getenv
-from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
-
-class Attention:
-  def __init__(self, dim:int, n_heads:int, n_kv_heads:int|None=None, linear=nn.Linear):
-    self.n_heads = n_heads
-    self.n_kv_heads = n_kv_heads if n_kv_heads is not None else n_heads # n_kv_heads != n_heads implies MQA [arxiv/2307.09288, A.2.1]
-    self.head_dim = dim // n_heads
-    self.n_rep = self.n_heads // self.n_kv_heads
-
-    if getenv("WQKV"):
-      self.wqkv = linear(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2, bias=False)
-    else:
-      self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
-      self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-      self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-
-    self.wo = linear(self.n_heads * self.head_dim, dim, bias=False)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor) -> Tensor:
-    if getenv("WQKV"):
-      xqkv = self.wqkv(x)
-      xqkv = xqkv.reshape(xqkv.shape[0], xqkv.shape[1], self.n_kv_heads, self.n_rep + 2, self.head_dim)
-      xq = xqkv[:, :, :, :self.n_rep].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xk = xqkv[:, :, :, self.n_rep:self.n_rep+1].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xv = xqkv[:, :, :, self.n_rep+1:self.n_rep+2].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-    else:
-      xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
-
-    xq = xq.reshape(xq.shape[0], xq.shape[1], self.n_heads, self.head_dim)
-    xk = xk.reshape(xk.shape[0], xk.shape[1], self.n_kv_heads, self.head_dim)
-    xv = xv.reshape(xv.shape[0], xv.shape[1], self.n_kv_heads, self.head_dim)
-
-    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
-    bsz, seqlen, _, _ = xq.shape
-
-    xq, xk, xv = xq.transpose(1, 2), xk.transpose(1, 2), xv.transpose(1, 2)
-    attn = xq.scaled_dot_product_attention(xk, xv, is_causal=True, enable_gqa=True).transpose(1, 2)
-
-    attn = attn.reshape(bsz, seqlen, -1)
-    return self.wo(attn)
-
-class FeedForward:
-  def __init__(self, dim:int, hidden_dim:int, linear=nn.Linear):
-    self.w1 = linear(dim, hidden_dim, bias=False)
-    self.w2 = linear(hidden_dim, dim, bias=False)
-    self.w3 = linear(dim, hidden_dim, bias=False) # the gate in Gated Linear Unit
-
-  def __call__(self, x:Tensor) -> Tensor:
-    w1 = self.w1(x).silu()
-    w3 = self.w3(x)
-    return self.w2(w1 * w3)
-
-class TransformerBlock:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_kv_heads:int|None, norm_eps:float, linear=nn.Linear):
-    self.attention = Attention(dim, n_heads, n_kv_heads, linear)
-    self.feed_forward = FeedForward(dim, hidden_dim, linear)
-    self.attention_norm = nn.RMSNorm(dim, norm_eps)
-    self.ffn_norm = nn.RMSNorm(dim, norm_eps)
-
-  def __call__(self, x:Tensor, freqs_cis:Tensor):
-    h = x + self.attention(self.attention_norm(x), freqs_cis)
-    return h + self.feed_forward(self.ffn_norm(h))
-
-class Transformer:
-  def __init__(self, dim:int, hidden_dim:int, n_heads:int, n_layers:int, norm_eps:float, vocab_size:int, n_kv_heads:int|None=None,
-               rope_theta:int=10000, max_context:int=1024, linear=nn.Linear, embedding=nn.Embedding):
-    self.layers = [TransformerBlock(dim, hidden_dim, n_heads, n_kv_heads, norm_eps, linear) for _ in range(n_layers)]
-    self.norm = nn.RMSNorm(dim, norm_eps)
-    self.tok_embeddings = embedding(vocab_size, dim)
-    self.output = nn.Linear(dim, vocab_size, bias=False) if embedding == nn.Embedding else linear(dim, vocab_size, bias=False)
-    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).contiguous().requires_grad_(False)
-
-  def __call__(self, tokens:Tensor):
-    h = self.tok_embeddings(tokens)
-    freqs_cis = self.freqs_cis.cast(h.dtype)[:, :tokens.shape[1], :, :, :]
-    for layer in self.layers: h = layer(h, freqs_cis)
-    logits = self.output(self.norm(h))
-    return logits
@@ -1,57 +1,24 @@
 from tinygrad.tensor import Tensor
 from tinygrad.dtype import dtypes
-from tinygrad.nn.optim import Optimizer
+from tinygrad.nn.optim import LAMB
 from tinygrad.helpers import FUSE_OPTIM

-class GradAccClipAdamW(Optimizer):
-  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, device=None, fused=FUSE_OPTIM):
-    super().__init__(params, lr, device, fused)
-    self.b1, self.b2, self.eps, self.wd = b1, b2, eps, weight_decay
-    self.b1_t, self.b2_t = (Tensor.ones((1,), dtype=dtypes.float32, device=self.device, requires_grad=False) for _ in [b1, b2])
-    self.m = self._new_optim_param()
-    self.v = self._new_optim_param()
+class GradAccClipAdamW(LAMB):
+  def __init__(self, params:list[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-6, weight_decay=0.0, grad_acc=1, clip_norm=1.0, fused=FUSE_OPTIM):
+    super().__init__(params, lr, b1, b2, eps, weight_decay, adam=True, fused=FUSE_OPTIM)
    self.grad_acc, self.clip_norm = grad_acc, clip_norm

-  def fstep(self, grads:list[Tensor]):
-    if self.fused:
-      out, extra = self._step([], grads)
-      updates = [out[0][self.pos_params[i]:self.pos_params[i+1]].reshape(tt.shape) for i, tt in enumerate(self.params)]
-    else:
-      updates, extra = self._step([], grads)
-    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i]))
-    to_realize = extra+self.params+self.buffers
-
-    Tensor.realize(*to_realize)
-    return extra[-1]
-
  def _step(self, params:list[Tensor], grads:list[Tensor]) -> tuple[list[Tensor], list[Tensor]]:
-    for i in range(len(grads)):
-      if grads[i].device != self.m[i].device: grads[i].assign(grads[i].to(self.m[i].device))
-
    if self.fused:
-      grads[0].assign(grads[0] / self.grad_acc)
+      grads[0] = grads[0] / self.grad_acc
      total_norm = grads[0].float().square().sum().sqrt()
-      grads[0].assign((grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype))
+      grads[0] = (grads[0] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[0].dtype)
    else:
+      total_norm = Tensor.zeros((), dtype=dtypes.float32, device=self.device)
+      for g in grads:
+        total_norm += g.float().square().sum()
+      total_norm = total_norm.sqrt()
      for i in range(len(grads)):
-        grads[i].assign(grads[i] / self.grad_acc)
-      total_norm = Tensor.stack(*[g.float().square().sum() for g in grads]).sum().sqrt().contiguous()
-      for i in range(len(grads)):
-        grads[i].assign((grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype))
-
-    ret = []
-    self.b1_t *= self.b1
-    self.b2_t *= self.b2
-    for i, g in enumerate(grads):
-      self.m[i].assign((self.b1 * self.m[i] + (1.0 - self.b1) * g).cast(self.m[i].dtype))
-      self.v[i].assign((self.b2 * self.v[i] + (1.0 - self.b2) * (g * g)).cast(self.v[i].dtype))
-      m_hat = (self.m[i] / (1.0 - self.b1_t)).cast(self.m[i].dtype)
-      v_hat = (self.v[i] / (1.0 - self.b2_t)).cast(self.v[i].dtype)
-      up = m_hat / (v_hat.sqrt() + self.eps)
-      ret.append((self.lr * up).cast(g.dtype))
-    return ret, [self.b1_t, self.b2_t] + self.m + self.v + [total_norm]
-
-  def _apply_update(self, t:Tensor, up:Tensor) -> Tensor:
-    wd = self.wd if t.ndim >= 2 else 0.0
-    up = up.shard_like(t) + self.lr.to(t.device) * wd * t.detach()
-    return t.detach() - up.cast(t.dtype)
+        grads[i] = grads[i] / self.grad_acc
+        grads[i] = (grads[i] * (self.clip_norm / (total_norm + 1e-6)).clamp(max_=1.0)).cast(grads[i].dtype)
+    return super()._step(params, grads)
@@ -1,37 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -1,32 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8}
-export BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-1152}
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4/"
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"405B"}
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -5,17 +5,15 @@ export DEV=${DEV:-AMD}
 export EMULATE="AMD_CDNA4"
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}

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

 export MODEL="llama3"
@@ -23,7 +21,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -1,43 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-2}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-5760}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-export FAKEDATA=1 BENCHMARK=10
-if [ -z "$FULL_LAYERS" ]; then
-  export LLAMA_LAYERS=2
-fi
-
-python3 examples/mlperf/model_train.py
@@ -5,17 +5,15 @@ export DEV=${DEV:-AMD}
 export EMULATE="AMD_CDNA4"
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-0}

 export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-8} MP=${MP:-1} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-4}
+export DP=${DP:-8} BS=${BS:-8} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

 export MODEL="llama3"
@@ -23,7 +21,7 @@ export BASEDIR="/raid/datasets/c4-8b/"
 export SMALL=1
 export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
 export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
+export LR="4e-4" END_LR="4e-5" WARMUP_SAMPLES=256 MAX_STEPS=1200000
 export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
 export SAMPLES=$((MAX_STEPS * GBS))
 export SEQLEN=${SEQLEN:-8192}
@@ -1,38 +0,0 @@
-#!/usr/bin/env bash
-
-export PYTHONPATH="."
-export DEV=${DEV:-AMD}
-export EMULATE="AMD_CDNA4"
-export CHECK_OOB=0
-export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
-export DEVICE_IN_FUNCTION_BUG=1
-
-export DEBUG=${DEBUG:-0}
-export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ALL2ALL=${ALL2ALL:-1}
-export USE_ATOMICS=${USE_ATOMICS:-0}
-export ASM_GEMM=${ASM_GEMM:-1}
-export WQKV=${WQKV:-1}
-export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-1}
-
-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
-export DP=${DP:-1} MP=${MP:-8} BS=${BS:-1} EVAL_BS=${EVAL_BS:-1} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-32}
-export GBS=$((BS * GRADIENT_ACC_STEPS))
-
-export MODEL="llama3"
-export BASEDIR="/raid/datasets/c4-8b/"
-export SMALL=1
-export LLAMA3_SIZE=${LLAMA3_SIZE:-"8B"}
-export EVAL_TARGET=3.3 EVAL_FREQ=12288
-export LR="1e-3" END_LR="1e-4" WARMUP_SAMPLES=4096 MAX_STEPS=1200000
-export WARMUP_STEPS=$((WARMUP_SAMPLES / GBS))
-export SAMPLES=$((MAX_STEPS * GBS))
-export SEQLEN=${SEQLEN:-8192}
-
-export SEED=${SEED:-$RANDOM}
-export DATA_SEED=${DATA_SEED:-5760}
-
-export JITBEAM=${JITBEAM:-3}
-export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1
-
-python3 examples/mlperf/model_train.py
@@ -3,4 +3,4 @@ export BENCHMARK=5
 export EVAL_BS=0
 export VIZ=${VIZ:--1}
 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama8b/implementations/tinybox_8xMI350X/dev_run.sh
-extra/viz/cli.py --profile --device "AMD" --top 20
+PYTHONPATH="." extra/viz/cli.py --profile --device "AMD" --top 20
@@ -31,7 +31,7 @@ def compile(onnx_file):
  for i in range(3):
    GlobalCounters.reset()
    print(f"run {i}")
-    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1), OPENPILOT_HACKS=1):
+    with Context(DEBUG=max(DEBUG.value, 2 if i == 2 else 1)):
      ret = run_onnx_jit(**inputs).numpy()
    # copy i == 1 so use of JITBEAM is okay
    if i == 1: test_val = np.copy(ret)
@@ -1,16 +0,0 @@
-import sys, pickle
-from extra.bench_log import WallTimeEvent, BenchEvent
-from tinygrad.helpers import getenv
-
-PKL = sys.argv[1] if len(sys.argv) > 1 else "/tmp/openpilot.pkl"
-
-load_times = []
-
-for _ in range(10):
-  with WallTimeEvent(BenchEvent.STEP) as wte: pickle.load(open(PKL, 'rb'))
-  load_times.append(wte.time)
-  print(f"pickle load: {wte.time:6.2f} s")
-
-if (assert_time:=getenv("ASSERT_MIN_LOAD_TIME")):
-  min_time = min(load_times)
-  assert min_time < assert_time, f"Speed regression, expected min load time of < {assert_time} s but took: {min_time} s"
@@ -65,7 +65,7 @@ def get_bar0_size(pcibus):
 class AMSMI(AMDev):
  def __init__(self, pcibus, vram_bar:MMIOInterface, doorbell_bar:MMIOInterface, mmio_bar:MMIOInterface):
    self.pcibus = pcibus
-    self.vram, self.doorbell64, self.mmio = vram_bar, doorbell_bar, mmio_bar
+    self.vram, self.doorbell64, self.mmio, self.dma_regions = vram_bar, doorbell_bar, mmio_bar, None
    self.pci_state = self.read_pci_state()
    if self.pci_state == "D0": self._init_from_d0()

@@ -236,6 +236,8 @@ class SMICtx:
      case _: return metrics.SmuMetrics.AverageSocketPower, metrics.SmuMetrics.dGPU_W_MAX

  def get_mem_usage(self, dev):
+    return 0
+
    usage = 0
    pt_stack = [dev.mm.root_page_table]
    while len(pt_stack) > 0:
@@ -244,8 +246,8 @@ class SMICtx:
        entry = pt.entries[i]

        if (entry & am.AMDGPU_PTE_VALID) == 0: continue
-        if pt.lv < am.AMDGPU_VM_PDB0 and not dev.gmc.is_pte_huge_page(pt.lv, entry):
-          pt_stack.append(AMPageTableEntry(dev, dev.xgmi2paddr(entry & 0x0000FFFFFFFFF000), lv=pt.lv+1))
+        if pt.lv!=am.AMDGPU_VM_PTB and not dev.gmc.is_pte_huge_page(pt.lv, entry):
+          pt_stack.append(AMPageTableEntry(dev, entry & 0x0000FFFFFFFFF000, lv=pt.lv+1))
          continue
        if (entry & am.AMDGPU_PTE_SYSTEM) != 0: continue
        usage += (1 << ((9 * (3-pt.lv)) + 12))
@@ -7,8 +7,8 @@ class GFXFake:
  def __init__(self): self.xccs = 8

 class AMDFake(AMDev):
-  def __init__(self, pci_dev):
-    self.pci_dev, self.devfmt = pci_dev, pci_dev.pcibus
+  def __init__(self, pci_dev, dma_regions=None):
+    self.pci_dev, self.devfmt, self.dma_regions = pci_dev, pci_dev.pcibus, dma_regions
    self.vram, self.doorbell64, self.mmio = self.pci_dev.map_bar(0), self.pci_dev.map_bar(2, fmt='Q'), self.pci_dev.map_bar(5, fmt='I')
    self._run_discovery()
    self._build_regs()
@@ -34,8 +34,7 @@ class WallTimeEvent:
    self.start = time.monotonic()
    return self
  def __exit__(self, *_):
-    self.time = time.monotonic() - self.start
-    _events[self.event]["wall"].append(self.time)
+    _events[self.event]["wall"].append(time.monotonic() - self.start)
    return False

 class KernelTimeEvent:
@@ -1,84 +0,0 @@
-from tinygrad import UOp, getenv
-from tinygrad.uop.ops import AxisType, KernelInfo, Ops
-from tinygrad.dtype import AddrSpace, dtypes
-
-N = getenv("N", 4096)
-M = getenv("M", N)
-K = getenv("K", N)
-
-WARP_SIZE = 32
-BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 8
-TM, TN = 4, 4
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
-assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
-
-is_kernel5 = getenv("K5", 0)
-THREADS_PER_BLOCK = 128 if is_kernel5 else 256
-WAVES_PER_BLOCK_N = 1 if is_kernel5 else 2
-WAVES_PER_BLOCK_M = THREADS_PER_BLOCK // WARP_SIZE // WAVES_PER_BLOCK_N
-REG_TILES_PER_WAVE_N = BLOCK_N // (WAVES_PER_BLOCK_N * LANES_PER_WAVE_N * TN)
-REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
-
-assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
-assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"
-
-consts = {"wpb_m":WAVES_PER_BLOCK_M, "lpw_m":LANES_PER_WAVE_M, "rt_m":REG_TILES_PER_WAVE_M, "t_m": TM,
-          "wpb_n":WAVES_PER_BLOCK_N, "lpw_n":LANES_PER_WAVE_N, "rt_n":REG_TILES_PER_WAVE_N, "t_n": TN}
-
-# 128x128 out, kx128, kx128 in
-def block_128x128_gemm(c:UOp, a:UOp, b:UOp) -> UOp:
-  tid = UOp.range(THREADS_PER_BLOCK, 2, AxisType.LOCAL)
-
-  #tid = UOp.special(THREADS_PER_BLOCK, "lidx0")
-  warp, lane = tid // WARP_SIZE, tid % WARP_SIZE
-  wave_n, wave_m = warp % WAVES_PER_BLOCK_N, warp // WAVES_PER_BLOCK_N
-  lane_n, lane_m = lane % LANES_PER_WAVE_N, lane // LANES_PER_WAVE_N
-
-  # define locals
-  A_local = UOp.placeholder((BLOCK_K, BLOCK_M), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL)
-  B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
-
-  # open the main reduction range and copy in GLOBAL -> LOCAL
-  a = a.reshape(K // BLOCK_K, BLOCK_K, BLOCK_M)
-  b = b.reshape(K // BLOCK_K, BLOCK_K, BLOCK_N)
-  k_tile_range = UOp.range(K // BLOCK_K, 3, AxisType.REDUCE)
-  A_store = A_local.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(a[k_tile_range].reshape(-1, THREADS_PER_BLOCK)[:, tid])
-  B_store = B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid].store(b[k_tile_range].reshape(-1, THREADS_PER_BLOCK)[:, tid])
-  barrier = UOp.barrier(A_store, B_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
-
-  # define accumulator (128x128), but broadcast across tid
-  c_regs = UOp.placeholder((REG_TILES_PER_WAVE_M*TM, REG_TILES_PER_WAVE_N*TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
-  c_regs = c_regs.after(c_regs.store(UOp.const(dtypes.float, 0).reshape((1,)*len(c_regs.shape)).expand(c_regs.shape)))
-
-  # define registers (NOTE: the thread count is the device count for this multi, it's sharded across the THREADS_PER_BLOCK)
-  A_col = UOp.placeholder((REG_TILES_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  B_row = UOp.placeholder((REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-
-  # LOCAL -> REGS
-  k = UOp.range(BLOCK_K, 4, AxisType.REDUCE)
-  A_col = A_col.after(A_col.store(A_local[k].reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM)[wave_m, :, lane_m, :]))
-  B_row = B_row.after(B_row.store(B_local[k].reshape(WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)[wave_n, :, lane_n, :]))
-
-  # do FMA
-  A_col = A_col.reshape(REG_TILES_PER_WAVE_M*TM, 1).expand(REG_TILES_PER_WAVE_M*TM, REG_TILES_PER_WAVE_N*TN)
-  B_row = B_row.reshape(1, REG_TILES_PER_WAVE_N*TN).expand(REG_TILES_PER_WAVE_M*TM, REG_TILES_PER_WAVE_N*TN)
-  c_regs = c_regs.after(c_regs.store(c_regs.after(k) + (A_col * B_row)).end(k).barrier().end(k_tile_range))
-
-  # store back to c
-  c_store = c.rearrange("(wpb_m rt_m lpw_m t_m) (wpb_n rt_n lpw_n t_n) -> (wpb_m wpb_n lpw_m lpw_n) (rt_m t_m) (rt_n t_n)", **consts)
-  return c_store[tid].store(c_regs).end(tid)
-
-def amd_copy_matmul(c:UOp, a:UOp, b:UOp) -> UOp:
-  block_id_n = UOp.range(N // BLOCK_N, 0, AxisType.GLOBAL)
-  block_id_m = UOp.range(M // BLOCK_M, 1, AxisType.GLOBAL)
-
-  # index the output with the globals
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
-  a = a.T.reshape(K, M // BLOCK_M, BLOCK_M)[:, block_id_m, :]
-  b = b.reshape(K, N // BLOCK_N, BLOCK_N)[:, block_id_n, :]
-  return block_128x128_gemm(c, a, b).end(block_id_n, block_id_m).sink(arg=KernelInfo(opts_to_apply=()))
-
-if __name__ == "__main__":
-  from amd_uop_matmul import eval_custom_matmul
-  eval_custom_matmul(amd_copy_matmul)
@@ -1,74 +1,98 @@
-from tinygrad import Tensor, Context, GlobalCounters, dtypes
+import numpy as np
+from tinygrad import Tensor, Device, Context, GlobalCounters, dtypes
 from tinygrad.uop.ops import UOp, KernelInfo, sint, AxisType
+from tinygrad.engine.realize import ExecItem, get_runner
 from tinygrad.dtype import AddrSpace
-from tinygrad.helpers import DEBUG, getenv
+from tinygrad.helpers import getenv

 N = getenv("N", 4096)
-M = getenv("M", N)
-K = getenv("K", N)
-NUM_RUNS = getenv("CNT", 5)
+M = K = N
+run_count = getenv("CNT", 5)

 # ---------------------------
 # launch/config constants
 # ---------------------------

 WARP_SIZE = 32
-BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 8
-TM, TN = 4, 4
-LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8
-assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0
+
+# Threadblock tile sizes (block-level tile of C that a block computes)
+BLOCK_N = 128   # columns of C (N-dim) per block
+BLOCK_M = 128   # rows of C (M-dim) per block
+BLOCK_K = 8     # K-slice per block iteration
+
+# Register tile sizes (per-thread accumulator tile of C)
+TN = 4     # columns per thread
+TM = 4     # rows per thread

 is_kernel5 = getenv("K5", 0)
 THREADS_PER_BLOCK = 128 if is_kernel5 else 256
-WAVES_PER_BLOCK_N = 1 if is_kernel5 else 2
-WAVES_PER_BLOCK_M = THREADS_PER_BLOCK // WARP_SIZE // WAVES_PER_BLOCK_N
-REG_TILES_PER_WAVE_N = BLOCK_N // (WAVES_PER_BLOCK_N * LANES_PER_WAVE_N * TN)
-REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM)
+assert THREADS_PER_BLOCK % BLOCK_N == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_N"
+assert THREADS_PER_BLOCK % BLOCK_K == 0, "THREADS_PER_BLOCK must be divisible by BLOCK_K"
+assert (BLOCK_N * BLOCK_K) % THREADS_PER_BLOCK == 0
+assert (BLOCK_M * BLOCK_K) % THREADS_PER_BLOCK == 0

-assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong"
-assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong"
+WARPS_PER_BLOCK = THREADS_PER_BLOCK // WARP_SIZE
+WAVE_TILE_N = 128 if is_kernel5 else 64
+WAVE_TILE_M = BLOCK_N * BLOCK_M // WARPS_PER_BLOCK // WAVE_TILE_N
+assert BLOCK_N % WAVE_TILE_N == 0, "BN must be a multiple of WN"
+assert BLOCK_M % WAVE_TILE_M == 0, "BM must be a multiple of WM"
+WAVES_IN_BLOCK_X = BLOCK_N // WAVE_TILE_N
+WAVES_IN_BLOCK_Y = BLOCK_M // WAVE_TILE_M
+assert WAVES_IN_BLOCK_X * WAVES_IN_BLOCK_Y == WARPS_PER_BLOCK, "wave grid must match warps/block"
+
+LANES_PER_WAVE_X = 8
+LANES_PER_WAVE_Y = 4
+ITERS_PER_WAVE_N = WAVE_TILE_N // (LANES_PER_WAVE_X * TN)
+ITERS_PER_WAVE_M = WAVE_TILE_M // (LANES_PER_WAVE_Y * TM)
+assert WAVE_TILE_N % (LANES_PER_WAVE_X * TN) == 0, "WAVE_TILE_N must be divisible by LANES_PER_WAVE_X*TN"
+assert WAVE_TILE_M % (LANES_PER_WAVE_Y * TM) == 0, "WAVE_TILE_M must be divisible by LANES_PER_WAVE_Y*TM"

 def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.LOOP): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)]
-def copy(dest:UOp, src:UOp, rng:int, upcast=False):
+def copy(dest:UOp, src:UOp, rng:int, set=False, upcast=False):
  assert dest.shape == src.shape
  rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.LOOP)
-  return dest[*rngs].store(src[*rngs]).end(*rngs)
+  copy = dest[*rngs].store(src[*rngs]).end(*rngs)
+  return dest.after(copy) if set else copy

-def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
+def hand_spec_kernel3():
  # ---------------------------
-  # block indices
+  # block indices & placeholders
  # ---------------------------
-  block_id_n = UOp.special(N // BLOCK_N, "gidx0")
-  block_id_m = UOp.special(M // BLOCK_M, "gidx1")
+  blockIdx_x = UOp.special(N // BLOCK_N, "gidx0")
+  blockIdx_y = UOp.special(N // BLOCK_M, "gidx1")
+
+  a = UOp.placeholder((N, N), dtypes.float, slot=1)
+  b = UOp.placeholder((N, N), dtypes.float, slot=2)
+  c = UOp.placeholder((N, N), dtypes.float, slot=0)

  # index the output with the globals
-  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :]
+  c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[blockIdx_y, :, blockIdx_x, :]

  # open the main reduction range
-  k_tile_range = UOp.range(K // BLOCK_K, 0, AxisType.REDUCE)
-  a = a.reshape(M // BLOCK_M, BLOCK_M, K // BLOCK_K, BLOCK_K)[block_id_m, :, k_tile_range, :]
-  b = b.reshape(K // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, block_id_n, :]
+  k_tile_range = UOp.range(N // BLOCK_K, 0, AxisType.REDUCE)
+  a = a.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_K, BLOCK_K)[blockIdx_y, :, k_tile_range, :]
+  b = b.reshape(N // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, blockIdx_x, :]

  # globals are no longer used, they are already in the indexes
-  del block_id_m, block_id_n
+  del blockIdx_y, blockIdx_x

  # ---------------------------
-  # GLOBAL -> LOCAL (A_local, B_local)
+  # GLOBAL -> LOCAL (As, Bs)
  # ---------------------------
  tid = UOp.special(THREADS_PER_BLOCK, "lidx0")

  # A: read BM x BK tiles (permute on store into locals)
-  BM_A_local_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
-  A_local = UOp.placeholder((BLOCK_K, BM_A_local_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
-  A_local_store = copy(A_local.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)
+  BM_As_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M
+  As = UOp.placeholder((BLOCK_K, BM_As_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M))
+  As_store = copy(As.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100)

  # B: read BK x BN tiles
-  B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
-  B_local_store = copy(B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)
+  Bs = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL)
+  Bs_store = copy(Bs.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200)

  # TODO: can we automate barrier?
-  barrier = UOp.barrier(A_local_store, B_local_store)
-  A_local, B_local = A_local.after(barrier), B_local.after(barrier)
+  barrier = UOp.barrier(As_store, Bs_store)
+  As, Bs = As.after(barrier), Bs.after(barrier)

  # open inner k range
  k = UOp.range(BLOCK_K, 3, AxisType.REDUCE)
@@ -76,30 +100,31 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
  # ---------------------------
  # LOCAL -> REG (per-wave tiles)
  # ---------------------------
-  warp, lane = tid // WARP_SIZE, tid % WARP_SIZE
-  waveIdx, waveIdy = warp % WAVES_PER_BLOCK_N, warp // WAVES_PER_BLOCK_N
-  laneIdx, laneIdy = lane % LANES_PER_WAVE_N, lane // LANES_PER_WAVE_N
-  assert waveIdy.vmax+1 == WAVES_PER_BLOCK_M and laneIdy.vmax+1 == LANES_PER_WAVE_M
+  waveIdx = (tid // WARP_SIZE) % WAVES_IN_BLOCK_X
+  waveIdy = (tid // WARP_SIZE) // WAVES_IN_BLOCK_X
+  assert waveIdy.vmax+1 == WAVES_IN_BLOCK_Y

-  A_col = UOp.placeholder((REG_TILES_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
-  A_local_slice = A_local[k, :].reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM)[waveIdy, :, laneIdy, :]
-  A_col = A_col.after(copy(A_col, A_local_slice, 300, upcast=True))
+  laneIdx = (tid % WARP_SIZE) % LANES_PER_WAVE_X
+  laneIdy = (tid % WARP_SIZE) // LANES_PER_WAVE_X
+  assert laneIdy.vmax+1 == LANES_PER_WAVE_Y

-  B_row = UOp.placeholder((REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
-  B_local_slice = B_local[k, :].reshape(WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)[waveIdx, :, laneIdx, :]
-  B_row = B_row.after(copy(B_row, B_local_slice, 400, upcast=True))
+  A_col = UOp.placeholder((ITERS_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG)
+  A_col = copy(A_col, As[k, :].reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM)[waveIdy, :, laneIdy, :], 300, set=True, upcast=True)
+
+  B_row = UOp.placeholder((ITERS_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG)
+  B_row = copy(B_row, Bs[k, :].reshape(WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)[waveIdx, :, laneIdx, :], 400, set=True, upcast=True)

  # ---------------------------
  # FMA: c_regs += A_col * B_row
  # ---------------------------
-  c_regs = UOp.placeholder((REG_TILES_PER_WAVE_M, TM, REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  c_regs = UOp.placeholder((ITERS_PER_WAVE_M, TM, ITERS_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG)
  i = UOp.range(c_regs.size, 16)
  c_regs = c_regs.after(c_regs.flatten()[i].store(0.0).end(i))

  # TODO: why don't these work as upcast?
  # why if the ranges merge is it slow?!? (if you change the order on end, they will merge. big slowdown on METAL)
-  iter_m, t_m, iter_n, t_n = rngs = rngs_for_shape(c_regs.shape, 500)
-  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iter_m, t_m] * B_row[iter_n, t_n]).end(iter_m, iter_n, t_m, t_n)
+  iterWaveM, yt, iterWaveN, xt = rngs = rngs_for_shape(c_regs.shape, 500)
+  sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iterWaveM, yt] * B_row[iterWaveN, xt]).end(iterWaveM, iterWaveN, yt, xt)

  # Close k, sync, and close K tiles
  sink = sink.end(k).barrier().end(k_tile_range)
@@ -107,37 +132,38 @@ def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp:
  # ---------------------------
  # REG -> GLOBAL (epilogue)
  # ---------------------------
-  c = c.reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM,
-                WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)
+  c = c.reshape(WAVES_IN_BLOCK_Y, ITERS_PER_WAVE_M, LANES_PER_WAVE_Y, TM,
+                WAVES_IN_BLOCK_X, ITERS_PER_WAVE_N, LANES_PER_WAVE_X, TN)
  c = c[waveIdy, :, laneIdy, :,
        waveIdx, :, laneIdx, :]
  sink = copy(c, c_regs.after(sink), rng=600)

  return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify()

-def eval_custom_matmul(fxn):
-  a = Tensor.randn(M, K, dtype=dtypes.float)
-  b = Tensor.randn(K, N, dtype=dtypes.float)
-  c = Tensor.empty(M, N, dtype=dtypes.float)
-  with Context(DEBUG=0): Tensor.realize(a, b)
+def test_matmul(sink:UOp, dtype=dtypes.float32, N=N):
+  rng = np.random.default_rng()
+  a = Tensor(rng.random((N, N), dtype=np.float32)-0.5, dtype=dtype)
+  b = Tensor(rng.random((N, N), dtype=np.float32)-0.5, dtype=dtype)
+  hc = Tensor.empty(N, N, dtype=dtype)
+  Tensor.realize(a, b, hc)
+
+  ei = ExecItem(sink, [t.uop.buffer for t in [hc, a, b]], prg=get_runner(Device.DEFAULT, sink))

  ets = []
-  with Context(DEBUG=max(2, DEBUG.value)):
-    for _ in range(NUM_RUNS):
-      GlobalCounters.reset()
-      tst = Tensor.custom_kernel(c, a, b, fxn=fxn)[0].realize()
-      ets.append(GlobalCounters.time_sum_s)
-  print(f"REAL TFLOPS {M * N * K * 2 / min(ets) * 1e-12:.2f}")
+  with Context(DEBUG=2):
+    for _ in range(run_count):
+      ets.append(ei.run(wait=True))
+  print(f"REAL TFLOPS {N * N * N * 2 / min(ets) * 1e-12:.2f}")

  if getenv("VERIFY", 1):
    GlobalCounters.reset()
    with Context(DEBUG=2):
      tc = (a @ b).realize()
    with Context(DEBUG=0):
-      err = (tc - tst).square().mean().item()
+      err = (hc - tc).square().mean().item()
    print(f"mean squared error {err}")
    if err > 1e-06:
      raise RuntimeError("matmul is wrong!")

 if __name__ == "__main__":
-  eval_custom_matmul(hand_spec_kernel3)
+  test_matmul(hand_spec_kernel3(), N=N)
@@ -0,0 +1,101 @@
+import atexit, functools
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.dtype import AddrSpace
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from tinygrad.renderer import Estimates
+from tinygrad.helpers import getenv, all_same, dedup
+from extra.gemm.asm.cdna.asm import build_kernel, TILE_M, TILE_N, TILE_K, NUM_WG
+
+# ** CDNA4 assembly gemm
+
+WORKGROUP_SIZE = 256
+
+@functools.cache
+def custom_asm_gemm(C:UOp, A:UOp, B:UOp, dname:str, arch:str, wg:int) -> UOp:
+  batch, M, K = A.shape
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  lidx = UOp.special(WORKGROUP_SIZE, "lidx0")
+  gidx = UOp.special(wg, "gidx0")
+  insts = build_kernel(batch, M, N, K, A.dtype.base)
+  lds = UOp(Ops.DEFINE_LOCAL, dtypes.uint8.ptr(size=133_120, addrspace=AddrSpace.LOCAL), (), 'lds')
+  sink = UOp.sink(C.base, A.base, B.base, lds, lidx, gidx,
+                  arg=KernelInfo(name=f"gemm_{batch}_{M}_{N}_{K}", estimates=Estimates(ops=2*batch*M*N*K, mem=(batch*M*K + K*N + batch*M*N)*2)))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname),
+                                UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
+
+counters = {"used":0, "todos":[]}
+def todo(msg:str) -> bool: counters["todos"].append(msg); return False
+atexit.register(lambda: print(f'asm_gemm: {counters["used"]} used, {len(counters["todos"])} not used'))
+
+def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
+  if a.dtype != b.dtype: return todo(f"dtypes must match {a.dtype} != {b.dtype}")
+  if a.dtype not in {dtypes.bfloat16, dtypes.float16}: return todo(f"only bfloat16/float16, got {a.dtype}")
+  batch, M, K = (1, *a.shape) if a.ndim == 2 else a.shape
+  N = b.shape[1]
+  # only sharding on the batch or K is tested, others might work too
+  if isinstance(a.device, tuple):
+    if a.ndim == 2 and a.uop.axis == 1 and b.uop.axis == 0: K //= len(a.device)
+    elif a.ndim == 3 and a.uop.axis == 0 and b.uop.axis is None: batch //= len(a.device)
+    else: return todo(f"sharding mismatch a.ndim={a.ndim} a.uop.axis={a.uop.axis} b.uop.axis={b.uop.axis}")
+    dname = a.device[0]
+  else: dname = a.device
+  arch = getattr(Device[dname].renderer, "arch", "")
+  if batch not in {1, 2}: return todo(f"GEMM batch size {batch}")
+  if (M % TILE_M != 0 or N % TILE_N != 0 or K % TILE_K != 0) and arch == "gfx950":
+    return todo(f"GEMM shape ({M},{N},{K}) not a multiple of ({TILE_M},{TILE_N},{TILE_K})")
+  return True
+
+# ** UOp gemm to test Tensor.custom_kernel multi and backward correctness on non cdna4
+# note: this can be removed after we have GEMM on mixins
+
+def custom_uop_gemm(C:UOp, A:UOp, B:UOp) -> UOp:
+  M, K = A.shape[0]*A.shape[1], A.shape[2]
+  K2, N = B.shape[(1 if B.ndim == 3 else 0):]
+  assert K == K2
+  m = UOp.range(M, 1, AxisType.LOOP)
+  n = UOp.range(N, 2, AxisType.LOOP)
+  k = UOp.range(K, 0, AxisType.REDUCE)
+  mul = (A.index((m*UOp.const(dtypes.index, K)+k))*B.index((k*UOp.const(dtypes.index, N)+n))).cast(dtypes.float32)
+  red = mul.reduce(k, arg=Ops.ADD, dtype=dtypes.float32).cast(C.dtype.base)
+  store = C.index((m*UOp.const(dtypes.index, N)+n), ptr=True).store(red).end(m, n)
+  return store.sink(arg=KernelInfo(name=f'uop_gemm_{M}_{N}_{K}'))
+
+# ** backward gemm, might use the asm gemm
+
+def custom_gemm_bw(gradient:UOp, kernel:UOp):
+  out, a, b = kernel.src[1:]
+  assert all_same([gradient.device, a.device, b.device, out.device])
+  a_t, b_t, g_t = Tensor(a, device=a.device), Tensor(b, device=a.device), Tensor(gradient, device=a.device)
+  # TODO: this needs to be cleaned up and done properly, the batch dim of grad and a multi need to align
+  g_t = g_t[:a.shape[0]]
+  grad_a = (g_t @ b_t.T).uop
+  grad_b = (a_t.permute(2, 0, 1).reshape(a_t.shape[2], -1) @ g_t.reshape(-1, g_t.shape[-1])).uop
+  return (None, grad_a, grad_b)
+
+# ** main gemm function
+
+def asm_gemm(a:Tensor, b:Tensor) -> Tensor:
+  assert can_use_asm_gemm(a, b), f"{counters['todos'][-1]}"
+  counters["used"] += 1
+  squeeze = a.ndim == 2
+  if squeeze: a = a.unsqueeze(0)
+
+  batch, M, K = a.shape
+  N = b.shape[1]
+  is_multi = isinstance(a.device, tuple)
+  if (k_sharded:=is_multi and a.uop.axis == 2): K //= len(a.device)
+
+  if is_multi:
+    out = Tensor(Tensor.empty(batch//len(a.device) if a.uop.axis==0 else batch, M, N, dtype=a.dtype, device=a.device).uop.multi(0), device=a.device)
+  else:
+    out = Tensor.empty(batch, M, N, dtype=a.dtype, device=a.device)
+
+  renderer = Device[a.device[0] if is_multi else a.device].renderer
+  dname, arch = renderer.device, getattr(renderer, "arch", "")
+  if arch.startswith("gfx950") and getenv("USE_ASM", 1):
+    out = Tensor.custom_kernel(out, a, b, fxn=functools.partial(custom_asm_gemm, dname=dname, wg=NUM_WG, arch=arch), grad_fxn=custom_gemm_bw)[0]
+  else:
+    out = Tensor.custom_kernel(out, a, b, fxn=custom_uop_gemm, grad_fxn=custom_gemm_bw)[0]
+  if k_sharded: out = out.sum(0)
+  return out.squeeze(0) if squeeze else out
@@ -0,0 +1,76 @@
+.text
+.section	.text.
+.global	gemm
+.p2align	8
+.type	gemm,@function
+
+gemm:
+INSTRUCTIONS
+
+.section .rodata,"a",@progbits
+.p2align 6, 0x0
+.amdhsa_kernel gemm
+  # basic memory requirements
+  .amdhsa_group_segment_fixed_size 30336
+  .amdhsa_private_segment_fixed_size 0
+  .amdhsa_kernarg_size 32
+  # register usage (RSRC1)
+  .amdhsa_next_free_vgpr 256
+  .amdhsa_next_free_sgpr 100
+  # workgroup / workitem IDs (RSRC2)
+  .amdhsa_system_sgpr_workgroup_id_x 1
+  .amdhsa_system_sgpr_workgroup_id_y 1
+  .amdhsa_system_sgpr_workgroup_id_z 1
+  # user SGPRs: kernarg ptr in s[0:1]
+  .amdhsa_user_sgpr_kernarg_segment_ptr 1
+  .amdhsa_user_sgpr_count 2
+  # gfx10+ / gfx11 specifics (RSRC1[29..31])
+  .amdhsa_wavefront_size32 1
+  .amdhsa_workgroup_processor_mode 1
+  .amdhsa_memory_ordered 1
+  .amdhsa_forward_progress 1
+  # misc for gfx11
+  .amdhsa_dx10_clamp 1
+  .amdhsa_ieee_mode 1
+  .amdhsa_uses_dynamic_stack 0
+.end_amdhsa_kernel
+
+.amdgpu_metadata
+---
+amdhsa.kernels:
+  - .args:
+      - .address_space:  generic
+        .name:           C
+        .offset:         0
+        .size:           8
+        .value_kind:     global_buffer
+        .value_type:     f16
+      - .address_space:  generic
+        .name:           A
+        .offset:         8
+        .size:           8
+        .value_kind:     global_buffer
+        .value_type:     f16
+      - .address_space:  generic
+        .name:           B
+        .offset:         16
+        .size:           8
+        .value_kind:     global_buffer
+        .value_type:     f16
+    .group_segment_fixed_size: 30336
+    .kernarg_segment_align: 8
+    .kernarg_segment_size: 32
+    .max_flat_workgroup_size: 128
+    .name:           gemm
+    .private_segment_fixed_size: 0
+    .sgpr_count:     70
+    .sgpr_spill_count: 0
+    .symbol:         gemm.kd
+    .vgpr_count:     256
+    .vgpr_spill_count: 0
+    .wavefront_size: 32
+amdhsa.version:
+  - 1
+  - 1
+...
+.end_amdgpu_metadata
@@ -0,0 +1,30 @@
+import math, pathlib
+
+from tinygrad import Device, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
+
+from extra.gemm.amd_uop_matmul import test_matmul
+
+N = 4096
+TN = 96
+THREADS_PER_WG = 128
+NUM_WG = math.ceil(N / TN) * math.ceil(N / TN)
+
+dname:str = Device.DEFAULT
+template:str = (pathlib.Path(__file__).parent/"template.s").read_text()
+
+def asm_kernel() -> UOp:
+  lidx = UOp.special(THREADS_PER_WG, "lidx0")
+  gidx = UOp.special(NUM_WG, "gidx0")
+
+  a = UOp.placeholder((N*N,), dtypes.half, slot=1)
+  b = UOp.placeholder((N*N,), dtypes.half, slot=2)
+  c = UOp.placeholder((N*N,), dtypes.half, slot=0)
+
+  src = template.replace("INSTRUCTIONS", (pathlib.Path(__file__).parent/"gemm.s").read_text())
+
+  sink = UOp.sink(a, b, c, lidx, gidx, arg=KernelInfo(name="gemm"))
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg=dname), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src)))
+
+if __name__ == "__main__":
+  test_matmul(asm_kernel(), dtype=dtypes.half, N=N)
@@ -0,0 +1,179 @@
+# unpack the complete kernel descriptor of an amdgpu ELF
+# https://rocm.docs.amd.com/projects/llvm-project/en/latest/LLVM/llvm/html/AMDGPUUsage.html#code-object-v3-kernel-descriptor
+import struct, pathlib, sys
+from tinygrad.runtime.support.elf import elf_loader
+
+def bits(x, lo, hi): return (x >> lo) & ((1 << (hi - lo + 1)) - 1)
+def assert_zero(x, lo, hi): assert bits(x, lo, hi) == 0
+
+with open(sys.argv[1], "rb") as f:
+  lib = f.read()
+
+image, sections, relocs = elf_loader(lib)
+rodata_entry = next((sh.header.sh_addr for sh in sections if sh.name == ".rodata"))
+
+# rodata is exactly 64 bytes
+kd = image[rodata_entry:rodata_entry+64]
+desc = int.from_bytes(kd, byteorder="little")
+
+group_segment_fixed_size = bits(desc, 0, 31)
+private_segment_fixed_size = bits(desc, 32, 63)
+kernarg_size = bits(desc, 64, 95)
+reserved_127_96 = bits(desc, 96, 127)
+assert reserved_127_96 == 0
+
+print("GROUP_SEGMENT_FIXED_SIZE:", group_segment_fixed_size)
+print("PRIVATE_SEGMENT_FIXED_SIZE:", private_segment_fixed_size)
+print("KERNARG_SIZE:", kernarg_size)
+print("RESERVED 127:96:", reserved_127_96)
+
+entry_off = bits(desc, 128, 191)
+
+# sign-extend manually if needed
+if entry_off & (1 << 63):
+  entry_off -= 1 << 64
+
+print("KERNEL_CODE_ENTRY_BYTE_OFFSET:", entry_off)
+
+kd_addr = 0x1840
+entry_addr = kd_addr + entry_off
+
+print("Computed entry address: 0x%016x" % entry_addr)
+print("256B aligned:", entry_addr % 256 == 0)
+
+pgm_rsrc3 = bits(desc, 352, 383)
+pgm_rsrc1 = bits(desc, 384, 415)
+pgm_rsrc2 = bits(desc, 416, 447)
+
+print("COMPUTE_PGM_RSRC3: 0x%08x" % pgm_rsrc3)
+print("COMPUTE_PGM_RSRC1: 0x%08x" % pgm_rsrc1)
+print("COMPUTE_PGM_RSRC2: 0x%08x" % pgm_rsrc2)
+
+# rsrc 3 (gfx950)
+
+accum_offset_raw = bits(pgm_rsrc3, 0, 5)
+assert_zero(pgm_rsrc3, 6, 15)
+tg_split = bits(pgm_rsrc3, 16, 16)
+accum_offset_vgprs = (accum_offset_raw + 1) * 4
+print("RSRC3.ACCUM_OFFSET (AccVGPR index):", accum_offset_vgprs)
+print("RSRC3.TG_SPLIT:", tg_split)
+
+# rsrc 1
+
+vgpr_gran = bits(pgm_rsrc1, 0, 5)
+sgpr_gran = bits(pgm_rsrc1, 6, 9)
+assert_zero(pgm_rsrc1, 27, 28)
+
+# NOTE: this is vgprs + agprs
+vgprs_used = (vgpr_gran + 1) * 8
+assert 0 <= vgprs_used <= 512
+
+k = sgpr_gran // 2
+sgprs_used = (k + 1) * 16
+
+print("RSRC1.VGPRS:", vgprs_used)
+print("RSRC1.SGPRS:", sgprs_used)
+
+assert_zero(pgm_rsrc1, 10, 11)
+
+float_round_mode_32 = bits(pgm_rsrc1, 12, 13)
+float_round_mode_16_64 = bits(pgm_rsrc1, 15, 14)
+float_denorm_mode_32 = bits(pgm_rsrc1, 16, 17)
+float_denorm_mode_16_64 = bits(pgm_rsrc1, 18, 19)
+
+priv = bits(pgm_rsrc1, 20, 20)
+assert priv == 0
+enable_dx10_clamp_wg_rr_en = bits(pgm_rsrc1, 21, 21)
+debug_mode = bits(pgm_rsrc1, 22, 22)
+enable_ieee_mode = bits(pgm_rsrc1, 23, 23)
+bulky = bits(pgm_rsrc1, 24, 24)
+assert bulky == 0
+cdbg_user = bits(pgm_rsrc1, 25, 25)
+assert cdbg_user == 0
+fp16_ovfl = bits(pgm_rsrc1, 26, 26)
+assert_zero(pgm_rsrc1, 27, 28)  # reserved
+assert_zero(pgm_rsrc1, 29, 29)  # WGP_MODE (reserved on gfx9)
+assert_zero(pgm_rsrc1, 30, 30)  # MEM_ORDERED (reserved on gfx9)
+assert_zero(pgm_rsrc1, 31, 31)  # FWD_PROGRESS (reserved on gfx9)
+
+# rsrc 2
+
+enable_private_segment = bits(pgm_rsrc2, 0, 0)  # SCRATCH_EN
+user_sgpr_count = bits(pgm_rsrc2, 1, 5)         # USER_SGPR
+enable_trap_handler = bits(pgm_rsrc2, 6, 6)     # TRAP_PRESENT (must be 0 here)
+assert enable_trap_handler == 0
+
+enable_sgpr_workgroup_id_x = bits(pgm_rsrc2, 7, 7)
+enable_sgpr_workgroup_id_y = bits(pgm_rsrc2, 8, 8)
+enable_sgpr_workgroup_id_z = bits(pgm_rsrc2, 9, 9)
+enable_sgpr_workgroup_info = bits(pgm_rsrc2, 10, 10)
+
+enable_vgpr_workitem_id = bits(pgm_rsrc2, 11, 12)  # TIDIG_CMP_CNT enum (0..3)
+
+enable_exception_address_watch = bits(pgm_rsrc2, 13, 13)
+assert enable_exception_address_watch == 0
+enable_exception_memory = bits(pgm_rsrc2, 14, 14)
+assert enable_exception_memory == 0
+
+granulated_lds_size = bits(pgm_rsrc2, 15, 23)
+assert granulated_lds_size == 0  # spec: must be 0; CP uses dispatch packet rounding
+
+enable_exception_fp_invalid = bits(pgm_rsrc2, 24, 24)
+enable_exception_fp_denorm_src = bits(pgm_rsrc2, 25, 25)
+enable_exception_fp_div0 = bits(pgm_rsrc2, 26, 26)
+enable_exception_fp_overflow = bits(pgm_rsrc2, 27, 27)
+enable_exception_fp_underflow = bits(pgm_rsrc2, 28, 28)
+enable_exception_fp_inexact = bits(pgm_rsrc2, 29, 29)
+enable_exception_int_div0 = bits(pgm_rsrc2, 30, 30)
+
+assert_zero(pgm_rsrc2, 31, 31)
+
+print("RSRC2.ENABLE_PRIVATE_SEGMENT:", enable_private_segment)
+print("RSRC2.USER_SGPR_COUNT:", user_sgpr_count)
+print("RSRC2.ENABLE_SGPR_WORKGROUP_ID_X:", enable_sgpr_workgroup_id_x)
+print("RSRC2.ENABLE_SGPR_WORKGROUP_ID_Y:", enable_sgpr_workgroup_id_y)
+print("RSRC2.ENABLE_SGPR_WORKGROUP_ID_Z:", enable_sgpr_workgroup_id_z)
+print("RSRC2.ENABLE_SGPR_WORKGROUP_INFO:", enable_sgpr_workgroup_info)
+print("RSRC2.ENABLE_VGPR_WORKITEM_ID (enum):", enable_vgpr_workitem_id)
+
+print("RSRC2.EXC_FP_INVALID:", enable_exception_fp_invalid)
+print("RSRC2.EXC_FP_DENORM_SRC:", enable_exception_fp_denorm_src)
+print("RSRC2.EXC_FP_DIV0:", enable_exception_fp_div0)
+print("RSRC2.EXC_FP_OVERFLOW:", enable_exception_fp_overflow)
+print("RSRC2.EXC_FP_UNDERFLOW:", enable_exception_fp_underflow)
+print("RSRC2.EXC_FP_INEXACT:", enable_exception_fp_inexact)
+print("RSRC2.EXC_INT_DIV0:", enable_exception_int_div0)
+
+# user sgprs
+
+enable_sgpr_private_segment_buffer = bits(desc, 448, 448)
+enable_sgpr_dispatch_ptr = bits(desc, 449, 449)
+enable_sgpr_queue_ptr = bits(desc, 450, 450)
+enable_sgpr_kernarg_segment_ptr = bits(desc, 451, 451)
+enable_sgpr_dispatch_id = bits(desc, 452, 452)
+enable_sgpr_flat_scratch_init = bits(desc, 453, 453)
+enable_sgpr_private_segment_size = bits(desc, 454, 454)
+
+assert_zero(desc, 455, 457)
+
+print("DESC.ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER:", enable_sgpr_private_segment_buffer)
+print("DESC.ENABLE_SGPR_DISPATCH_PTR:", enable_sgpr_dispatch_ptr)
+print("DESC.ENABLE_SGPR_QUEUE_PTR:", enable_sgpr_queue_ptr)
+print("DESC.ENABLE_SGPR_KERNARG_SEGMENT_PTR:", enable_sgpr_kernarg_segment_ptr)
+print("DESC.ENABLE_SGPR_DISPATCH_ID:", enable_sgpr_dispatch_id)
+print("DESC.ENABLE_SGPR_FLAT_SCRATCH_INIT:", enable_sgpr_flat_scratch_init)
+print("DESC.ENABLE_SGPR_PRIVATE_SEGMENT_SIZE:", enable_sgpr_private_segment_size)
+
+assert_zero(desc, 458, 459)
+
+uses_dynamic_stack = bits(desc, 459, 460)
+print("DESC.USES_DYNAMIC_STACK:", uses_dynamic_stack)
+
+# gfx950 only
+assert_zero(desc, 460, 463)
+kernarg_preload_spec_length = bits(desc, 464, 470)
+print("DESC.KERNARG_PRELOAD_SPEC_LENGTH:", kernarg_preload_spec_length)
+kernarg_preload_spec_offset = bits(desc, 471, 479)
+print("DESC.KERNARG_PRELOAD_SPEC_OFFSET:", kernarg_preload_spec_offset)
+
+assert_zero(desc, 480, 511)
@@ -10,9 +10,9 @@ HEVC_ROUNDUP = getenv("DATA_ROUNDUP", 32)
@functools.cache
 def _hevc_jitted_decoder(out_image_size:tuple[int, int], max_hist:int, inplace:bool):
  def hevc_decode_frame(pos:Variable, hevc_tensor:Tensor, offset:Variable, sz:Variable, opaque:Tensor, i:Variable, *hist:Tensor, outbuf:Tensor|None=None):
-    x = hevc_tensor[offset:offset+sz*HEVC_ROUNDUP].decode_hevc_frame(pos, out_image_size, opaque[i], hist).realize()
+    x = hevc_tensor[offset:offset+sz*HEVC_ROUNDUP].decode_hevc_frame(pos, out_image_size, opaque[i], hist)
    if outbuf is not None: outbuf.assign(x).realize()
-    return x
+    return x.realize()
  return TinyJit(hevc_decode_frame)

 def hevc_decode(hevc_tensor:Tensor, opaque:Tensor, frame_info:list, luma_h:int, luma_w:int,
@@ -74,14 +74,10 @@ if __name__ == "__main__":
  Device.default.synchronize()

  # decode all frames using the iterator
-  tm = Timing("decoding whole file: ", on_exit=(lambda et: f", {len(frame_info)} frames, {len(frame_info)/(et/1e9):.2f} fps"))
-  with tm:
+  with Timing("decoding whole file: ", on_exit=(lambda et: f", {len(frame_info)} frames, {len(frame_info)/(et/1e9):.2f} fps")):
    images = list(hevc_decode(hevc_tensor, opaque_nv, frame_info, luma_h, luma_w, history=hist, preallocated_outputs=out_images))
    Device.default.synchronize()

-  fps = len(frame_info)/(tm.et/1e9)
-  assert fps >= getenv("ASSERT_FPS", 0), f"HEVC decode too slow: {fps:.2f} fps"
-
  # validation
  if getenv("VALIDATE", 0):
    import pickle
@@ -41,13 +41,9 @@ class Attention:
    self.n_rep = self.n_heads // self.n_kv_heads
    self.max_context = max_context

-    if getenv("WQKV"):
-      self.wqkv = linear(dim, self.n_heads * self.head_dim + self.n_kv_heads * self.head_dim * 2, bias=False)
-    else:
-      self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
-      self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-      self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
-
+    self.wq = linear(dim, self.n_heads * self.head_dim, bias=False)
+    self.wk = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
+    self.wv = linear(dim, self.n_kv_heads * self.head_dim, bias=False)
    self.wo = linear(self.n_heads * self.head_dim, dim, bias=False)

    self.q_norm = nn.RMSNorm(dim, qk_norm) if qk_norm is not None else None
@@ -55,11 +51,9 @@ class Attention:

  def __call__(self, x:Tensor, start_pos:Union[Variable,int], freqs_cis:Tensor, mask:Optional[Tensor]=None) -> Tensor:
    if getenv("WQKV"):
-      xqkv = self.wqkv(x)
-      xqkv = xqkv.reshape(xqkv.shape[0], xqkv.shape[1], self.n_kv_heads, self.n_rep + 2, self.head_dim)
-      xq = xqkv[:, :, :, :self.n_rep].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xk = xqkv[:, :, :, self.n_rep:self.n_rep+1].reshape(xqkv.shape[0], xqkv.shape[1], -1)
-      xv = xqkv[:, :, :, self.n_rep+1:self.n_rep+2].reshape(xqkv.shape[0], xqkv.shape[1], -1)
+      if not hasattr(self, 'wqkv'): self.wqkv = Tensor.cat(self.wq.weight, self.wk.weight, self.wv.weight)
+      xqkv = x @ self.wqkv.T
+      xq, xk, xv = xqkv.split([self.wq.weight.shape[0], self.wk.weight.shape[0], self.wv.weight.shape[0]], dim=2)
    else:
      xq, xk, xv = self.wq(x), self.wk(x.contiguous_backward()), self.wv(x)

@@ -2,13 +2,30 @@

 ## Getting SQ Thread Trace

-`VIZ=2` to enable SQTT profiling.
-
-`SQTT_ITRACE_SE_MASK=X` to select shader engines for instruction tracing, -1 = all, 0 = disabled, >0 = SE bitmask, default 0b11.
+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.

-## Viewing the traces
+`SQTT_ITRACE_SE_MASK=X` to select for which shader engines instruction tracing will be enabled, -1 is all, 0 is none (instruction tracing disabled), >0 is
+bitfield/mask for SEs to enable instruction tracing on. Masking shader engines will give smaller file sizes at a cost of less hits and kernels that
+don't have any wavefront on first simd of shader engine with instruction tracing enabled will not have instruction timings.
+The default is 2 (second shader engine only), only one for file size reasons, second instead of first because dispatch starts from it so there is
+greater chance that kernels with small global size will have instruction tracing data.
+ 
+Note that instruction tracing might not be available for kernels with small global dims, this is not a bug, but it can be improved with various hacks
+to the point where it can reliably trace a kernel consisting of a single wavefront (am only, not quite reliable under amdgpu due to waves sometimes
+being dispatched starting from different simds). More info in comments in ops_amd.py

- Web UI: `tinygrad/viz/serve.py`
- Command line: `python -m tinygrad.renderer.amd.sqtt`
+## Converting pickled profile with SQTT data into RGP file
+
+```bash
+extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
+```
+
+Then load gpu0.rgp into Radeon GPU Profiler. It works just fine both in wine (macos, native version available for linux) and via ssh X forwarding
+
+If multiple gpus are used you can select which one to export with `-d` like this:
+
+```bash
+extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -d 'AMD:5' -o /tmp/gpu5.rgp
+```
@@ -0,0 +1,152 @@
+import os
+os.environ["PYTHONPATH"] = "."
+os.environ["SQTT"] = "1"
+if "DEV" not in os.environ: os.environ["DEV"] = "AMD"
+os.environ["PROFILE"] = "1"
+os.environ["AMD_LLVM"] = "0"
+
+from dataclasses import replace
+import atexit, contextlib
+from tinygrad import Tensor
+from tinygrad.helpers import system, OSX
+from tinygrad.runtime.ops_amd import AMDProgram
+from extra.sqtt.roc import decode, WaveExec, ProfileSQTTEvent
+from tinygrad.device import Device
+
+from extra.sqtt.attempt_sqtt_parse import parse_sqtt_print_packets
+
+dev = Device["AMD"]
+
+@contextlib.contextmanager
+def save_sqtt():
+  # clear the old traces
+  dev.profile_events.clear()
+  sqtt:dict[str, list[WaveExec]] = {}
+  yield sqtt
+  events = dev.profile_events
+
+  #rctx = decode(events)
+  #assert len(rctx.inst_execs) > 0, "empty sqtt output"
+  #sqtt.update(rctx.inst_execs)
+
+  for e in events:
+    if isinstance(e, ProfileSQTTEvent):
+      print(replace(e, blob=b''))
+      if e.se == 0:
+        parse_sqtt_print_packets(e.blob)
+
+template = """.text
+.globl matmul
+.p2align 8
+.type matmul,@function
+matmul:
+  INSTRUCTION
+
+.rodata
+.p2align 6
+.amdhsa_kernel matmul
+  .amdhsa_kernarg_size 8
+  .amdhsa_user_sgpr_kernarg_segment_ptr 1
+  .amdhsa_next_free_vgpr .amdgcn.next_free_vgpr
+  .amdhsa_next_free_sgpr .amdgcn.next_free_sgpr
+  .amdhsa_wavefront_size32 1
+.end_amdhsa_kernel
+
+.amdgpu_metadata
+---
+amdhsa.version:
+  - 1
+  - 0
+amdhsa.kernels:
+  - .name: matmul
+    .symbol: matmul.kd
+    .group_segment_fixed_size: 0
+    .private_segment_fixed_size: 0
+    .wavefront_size: 32
+    .sgpr_count: 8
+    .vgpr_count: 8
+    .max_flat_workgroup_size: 1024
+    .kernarg_segment_align: 8
+    .kernarg_segment_size: 8
+    .args:
+      - .address_space:  global
+        .name:           a
+        .offset:         0
+        .size:           8
+        .type_name:      'float*'
+        .value_kind:     global_buffer
+...
+.end_amdgpu_metadata
+"""
+
+def run_asm(src, num_workgroups=1, num_waves=1):
+  WAVE_SIZE = 32
+  t = Tensor.empty(0x1000).realize()
+  buf = t.uop.buffer.ensure_allocated()
+  lib = dev.compiler.compile(template.replace("INSTRUCTION", '\n'.join(src)))
+  dev.compiler.disassemble(lib)
+  fxn = AMDProgram(dev, "matmul", lib)
+  fxn(buf._buf, global_size=(num_workgroups,1,1), local_size=(WAVE_SIZE*num_waves,1,1), wait=True)
+
+if __name__ == "__main__":
+  with save_sqtt() as sqtt:
+    run_asm([
+      "s_nop 100",
+      "s_nop 100",
+      "s_load_b64 s[0:1], s[0:1], null",
+      "s_waitcnt lgkmcnt(0)",
+      "s_nop 100",
+      "s_nop 100",
+      "s_add_i32 s2, s2, 10",
+      "s_add_i32 s2, s2, 10",
+      "s_nop 100",
+      "s_nop 100",
+      "v_mov_b32_e32 v0, 0",
+      "v_mov_b32_e32 v0, 0",
+      "s_nop 100",
+      "s_nop 100",
+      "v_dual_fmac_f32 v2, v48, v24 :: v_dual_fmac_f32 v9, v37, v51",
+      "v_dual_fmac_f32 v2, v48, v24 :: v_dual_fmac_f32 v9, v37, v51",
+      "s_nop 100",
+      "s_nop 100",
+      "global_load_b128 v[2:5], v0, s[0:1]",
+      "global_load_b128 v[2:5], v0, s[0:1]",
+      "s_nop 100",
+      "s_nop 100",
+      "s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)",
+      "s_endpgm",
+    ], num_workgroups=1, num_waves=1)
+  exit(0)
+
+  with save_sqtt() as sqtt:
+    #(Tensor.empty(16,16) @ Tensor.empty(16,16)).elu().realize()
+    #Tensor.empty(1, 64).sum(axis=1).realize()
+    Tensor.empty(1).log2().realize()
+  exit(0)
+
+  with save_sqtt() as sqtt:
+    # what's in v0?
+    run_asm([
+      "v_mov_b32_e32 v0, 0",
+      "v_mov_b32_e32 v1, 0",
+      "s_clause 0x1",
+      "s_load_b64 s[0:1], s[0:1], null",
+      "s_waitcnt lgkmcnt(0)",
+    ]+[
+      "global_load_b32 v1, v0, s[0:1]",
+    ]*10+[
+      "global_load_b32 v10, v1, s[0:1]",
+      "s_waitcnt vmcnt(0)",
+
+      #"v_rcp_f32 v1, v0"
+      #"v_add_f32_e32 v1 v0 v0",
+      #"v_add_f32_e32 v5 v4 v4",
+      #"v_add_f32_e32 v7 v6 v6",
+      #"v_add_f32_e32 v1 v0 v0",
+      #"v_add_f32_e32 v2 v1 v1",
+      #"s_nop 1"
+    ]*5+[
+      "v_add_f32_e32 v3 v2 v2",
+    ]*5+[
+      "v_mul_f32_e32 v3 v2 v2",
+    ]*7)
@@ -0,0 +1,548 @@
+import pickle, sys
+from tinygrad.helpers import getenv, Timing, colored
+from extra.sqtt.roc import decode, ProfileSQTTEvent
+
+# do these enums match fields in the packets?
+#from tinygrad.runtime.support.amd import import_soc
+#soc = import_soc([11])
+#perf_sel = {getattr(soc, k):k for k in dir(soc) if k.startswith("SQ_PERF_")}
+
+# Instruction packets (one per ISA op)
+# NOTE: these are bad guesses and may be wrong! feel free to update if you know better
+# some names were taken from SQ_TT_TOKEN_MASK_TOKEN_EXCLUDE_SHIFT
+
+# we see 18 opcodes
+# opcodes(18):  1  2  3  4  5  6  8  9  F 10 11 12 14 15 16 17 18 19
+# if you exclude everything, you are left with 6
+# opcodes( 6): 10 11 14 15 16 17
+# sometimes we see a lot of B, but not repeatable
+
+# not seen
+# 7 A C
+
+# NOTE: INST runs before EXEC
+
+OPCODE_COLORS = {
+  # dispatches are BLACK
+  0x1: "BLACK",
+  0x18: "BLACK",
+
+  # execs are yellow
+  0x2: "yellow",
+  0x3: "yellow",
+  0x4: "YELLOW",
+  0x5: "YELLOW",
+
+  # waves are blue
+  0x8: "blue",
+  0x9: "blue",
+  0x6: "cyan",
+  0xb: "cyan",
+}
+
+OPCODE_NAMES = {
+  # gated by SQ_TT_TOKEN_EXCLUDE_VALUINST_SHIFT (but others must be enabled for it to show)
+  0x01: "VALUINST",
+  # gated by SQ_TT_TOKEN_EXCLUDE_VMEMEXEC_SHIFT
+  0x02: "VMEMEXEC",
+  # gated by SQ_TT_TOKEN_EXCLUDE_ALUEXEC_SHIFT
+  0x03: "ALUEXEC",
+  # gated by SQ_TT_TOKEN_EXCLUDE_IMMEDIATE_SHIFT
+  0x04: "IMMEDIATE",
+  0x05: "IMMEDIATE_MASK",
+
+  # gated by SQ_TT_TOKEN_EXCLUDE_WAVERDY_SHIFT
+  0x06: "WAVERDY",
+  # gated by SQ_TT_TOKEN_EXCLUDE_WAVESTARTEND_SHIFT
+  0x08: "WAVEEND",
+  0x09: "WAVESTART",
+  # gated by SQ_TT_TOKEN_EXCLUDE_WAVEALLOC_SHIFT
+  0x0B: "WAVEALLOC",  # FFF00
+
+  # gated by NOT SQ_TT_TOKEN_EXCLUDE_PERF_SHIFT
+  0x0D: "PERF",
+  # gated by SQ_TT_TOKEN_EXCLUDE_EVENT_SHIFT
+  0x12: "EVENT",
+  0x13: "EVENT_BIG",  # FFFFF800
+  # some gated by SQ_TT_TOKEN_EXCLUDE_REG_SHIFT, some always there. something is broken with the timing on this
+  0x14: "REG",
+  # gated by SQ_TT_TOKEN_EXCLUDE_INST_SHIFT
+  0x18: "INST",
+  # gated by SQ_TT_TOKEN_EXCLUDE_UTILCTR_SHIFT
+  0x19: "UTILCTR",
+
+  # this is the first (8 byte) packet in the bitstream
+  0x17: "LAYOUT_HEADER",       # layout/mode/group + selectors A/B (reversed)
+
+  # pure time (no extra bits)
+  0x0F: "TS_DELTA_SHORT",
+  0x10: "NOP",
+  0x11: "TS_WAVE_STATE",     # almost pure time, has a small flag
+
+  # not a good name, but seen and understood mostly
+  0x15: "SNAPSHOT",          # small delta + 50-ish bits of snapshot
+  0x16: "TS_DELTA_OR_MARK",  # 36-bit long delta or 36-bit marker
+
+  # packets we haven't seen / rarely see 0x0b
+  0x07: "TS_DELTA_S8_W3_7",         # shift=8,  width=3  (small delta)
+  0x0A: "TS_DELTA_S5_W2_A",         # shift=5,  width=2
+  0x0C: "TS_DELTA_S5_W3_B",         # shift=5,  width=3 (different consumer)
+}
+
+#  SALU    =  0x0 / s_mov_b32
+#  SMEM    =  0x1 / s_load_b*
+#  JUMP    =  0x3 / s_cbranch_scc0
+#  NEXT    =  0x4 / s_cbranch_execz
+#  MESSAGE =  0x9 / s_sendmsg
+#  VALU    =  0xb / v_(exp,log)_f32_e32
+#  VALU    =  0xd / v_lshlrev_b64
+#  VALU    =  0xe / v_mad_u64_u32
+#  VMEM    = 0x21 / global_load_b32
+#  VMEM    = 0x22 / global_load_b32
+#  VMEM    = 0x24 / global_store_b32
+#  VMEM    = 0x25 / global_store_b64
+#  VMEM    = 0x27 / global_store
+#  VMEM    = 0x28 / global_store_b64
+#  LDS     = 0x29 / ds_load_b128
+#  LDS     = 0x2b / ds_store_b32
+#  LDS     = 0x2e / ds_store_b128
+#  ????    = 0x5a / hidden global_load  instruction
+#  ????    = 0x5b / hidden global_load  instruction
+#  ????    = 0x5c / hidden global_store instruction
+#  VALU    = 0x73 / v_cmpx_eq_u32_e32 (not normal VALUINST)
+OPNAME = {
+  0x0: "SALU",
+  0x1: "SMEM",
+  0x3: "JUMP",
+  0x4: "NEXT",
+  0x9: "MESSAGE",
+  0xb: "VALU",
+  0xd: "VALU",
+  0xe: "VALU",
+  0x21: "VMEM_LOAD",
+  0x22: "VMEM_LOAD",
+  0x24: "VMEM_STORE",
+  0x25: "VMEM_STORE",
+  0x26: "VMEM_STORE",
+  0x27: "VMEM_STORE",
+  0x28: "VMEM_STORE",
+  0x29: "LDS_LOAD",
+  0x2b: "LDS_STORE",
+  0x2e: "LDS_STORE",
+  0x50: "__SIMD_LDS_LOAD",
+  0x51: "__SIMD_LDS_LOAD",
+  0x54: "__SIMD_LDS_STORE",
+  0x5a: "__SIMD_VMEM_LOAD",
+  0x5b: "__SIMD_VMEM_LOAD",
+  0x5c: "__SIMD_VMEM_STORE",
+  0x5d: "__SIMD_VMEM_STORE",
+  0x5e: "__SIMD_VMEM_STORE",
+  0x5f: "__SIMD_VMEM_STORE",
+  0x72: "SALU_OR",
+  0x73: "VALU_CMPX",
+}
+
+ALUSRC = {
+  1: "SALU",
+  2: "VALU",
+  3: "VALU_SALU",
+}
+
+MEMSRC = {
+  0: "LDS",
+  1: "__LDS",
+  2: "VMEM",
+  3: "__VMEM",
+}
+
+
+# these tables are from rocprof trace decoder
+# rocprof_trace_decoder_parse_data-0x11c6a0
+# parse_sqtt_180 = b *rocprof_trace_decoder_parse_data-0x11c6a0+0x110040
+
+# ---------- 1. local_138: 256-byte state->opcode table ----------
+
+STATE_TO_OPCODE: bytes = bytes([
+  0x10, 0x16, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x17, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+  0x10, 0x07, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x19, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+  0x10, 0x00, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x11, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+  0x10, 0x12, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x15, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+  0x10, 0x16, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x17, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+  0x10, 0x07, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x19, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+  0x10, 0x00, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x11, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+  0x10, 0x13, 0x18, 0x01, 0x05, 0x0b, 0x0c, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x09, 0x04, 0x03, 0x02,
+  0x10, 0x15, 0x18, 0x01, 0x06, 0x08, 0x0d, 0x00, 0x0f, 0x14, 0x18, 0x01, 0x0a, 0x04, 0x03, 0x02,
+])
+
+# opcode mask (the bits used to determine the opcode, worked out by looking at the repeats in STATE_TO_OPCODE)
+
+opcode_mask = {
+  0x10: 0b1111,
+
+  0x16: 0b1111111,
+  0x17: 0b1111111,
+  0x07: 0b1111111,
+  0x19: 0b1111111,
+  0x11: 0b1111111,
+  0x12: 0b11111111,
+  0x13: 0b11111111,
+  0x15: 0b1111111,
+
+  0x18: 0b111,
+  0x1: 0b111,
+
+  0x5: 0b11111,
+  0x6: 0b11111,
+  0xb: 0b11111,
+  0x8: 0b11111,
+  0xc: 0b11111,
+  0xd: 0b11111,
+
+  0xf: 0b1111,
+  0x14: 0b1111,
+
+  0x9: 0b11111,
+  0xa: 0b11111,
+
+  0x4: 0b1111,
+  0x3: 0b1111,
+  0x2: 0b1111,
+}
+
+# ---------- 2. DAT_0012e280: nibble budget per opcode&0x1F ----------
+
+NIBBLE_BUDGET = [
+  0x08, 0x0C, 0x08, 0x08, 0x0C, 0x18, 0x18, 0x40, 0x14, 0x20, 0x30, 0x14, 0x34, 0x1C, 0x30, 0x08,
+  0x04, 0x18, 0x18, 0x20, 0x40, 0x40, 0x30, 0x40, 0x14, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+]
+
+# ---------- 3. delta_map from your hash nodes ----------
+
+# opcode -> (shift, width)
+DELTA_MAP_DEFAULT = {
+  0x01: (3,  3),   # shift=3,  end=6
+  0x02: (4,  2),   # shift=4,  end=6
+  0x03: (4,  2),   # shift=4,  end=6
+  0x04: (4,  3),   # shift=4,  end=7
+  0x05: (5,  3),   # shift=5,  end=8
+  0x06: (5,  3),   # shift=5,  end=8
+  0x07: (8,  3),   # shift=8,  end=11
+  0x08: (5,  3),   # shift=5,  end=8
+  0x09: (5,  2),   # shift=5,  end=7
+  0x0A: (5,  2),   # shift=5,  end=7
+  0x0B: (5,  3),   # shift=5,  end=8
+  0x0C: (5,  3),   # shift=5,  end=8
+  0x0D: (5,  3),   # shift=5,  end=8
+  # NOTE: 0x0e can never be decoded, it's not in the STATE_TO_OPCODE table
+  #0x0E: (7,  2),   # shift=7,  end=9
+  0x0F: (4,  4),   # shift=4,  end=8
+  0x10: (0,  0),   # shift=0,  end=0  (no delta)
+  0x11: (7,  9),   # shift=7,  end=16
+  0x12: (8,  3),   # shift=8,  end=11
+  0x13: (8,  3),   # shift=8,  end=11
+  0x14: (4,  3),   # shift=4,  end=7
+  0x15: (7,  3),   # shift=7,  end=10
+  0x16: (12, 36),  # shift=12, end=48 (36-bit field, matches the 0x16 special-case)
+  0x17: (0,  0),   # shift=0,  end=0  (no delta)
+  0x18: (4,  3),   # shift=4,  end=7
+  0x19: (7,  2),   # shift=7,  end=9
+}
+
+# ---------- 4. One-line-per-packet parser ----------
+
+def reg_mask(opcode):
+  nb_bits = NIBBLE_BUDGET[opcode & 0x1F]
+  shift, width = DELTA_MAP_DEFAULT[opcode]
+  delta_mask = ((1 << width) - 1) << shift
+  assert delta_mask & opcode_mask[opcode] == 0, "masks shouldn't overlap"
+  return ((1 << nb_bits) - 1) & ~(delta_mask | opcode_mask[opcode])
+
+def decode_packet_fields(opcode: int, reg: int) -> str:
+  """
+  Decode packet payloads conservatively, using:
+    - NIBBLE_BUDGET[opcode & 0x1F] to mask reg down to true width.
+    - DELTA_MAP_DEFAULT[opcode] to expose the "primary" field (often delta).
+    - Per-opcode layouts derived from rocprof's decompiled consumers.
+  """
+  # --- 0. Restrict to real packet bits not used in delta ---------------------------------
+  pkt = reg & reg_mask(opcode)
+  fields: list[str] = []
+
+  match opcode:
+    case 0x01: # VALUINST
+      # 6 bit field
+      flag = (pkt >> 6) & 1
+      wave = pkt >> 7
+      fields.append(f"wave={wave:x}")
+      if flag: fields.append("flag")
+    case 0x02: # VMEMEXEC
+      # 2 bit field (pipe is a guess)
+      src = pkt>>6
+      fields.append(f"src={src} [{MEMSRC.get(src, '')}]")
+    case 0x03: # ALUEXEC
+      # 2 bit field
+      src = pkt>>6
+      fields.append(f"src={src} [{ALUSRC.get(src, '')}]")
+    case 0x04: # IMMEDIATE_4
+      # 5 bit field (actually 4)
+      wave = pkt >> 7
+      fields.append(f"wave={wave:x}")
+    case 0x05: # IMMEDIATE_5
+      # 16 bit field
+      # 1 bit per wave
+      fields.append(f"mask={pkt>>8:016b}")
+    case 0x6:
+      # wave ready FFFF00
+      # 16 bit field
+      # 1 bit per wave
+      fields.append(f"mask={pkt>>8:016b}")
+    case 0x0d:
+      # 20 bit field
+      fields.append(f"arg = {pkt>>8:X}")
+    case 0x12:
+      fields.append(f"event = {pkt>>11:X}")
+    case 0x15:
+      fields.append(f"snap = {pkt>>10:X}")
+    case 0x19:
+      # wave end
+      fields.append(f"ctr = {pkt>>9:X}")
+    case 0xf:
+      extracted_delta = (reg >> 4) & 0xF
+      fields.append(f"strange_delta=0x{extracted_delta:x}")
+    case 0x11:
+      # DELTA_MAP_DEFAULT: shift=7, width=9 -> small delta.
+      # FF0000 is the mask
+      coarse    = pkt >> 16
+      fields.append(f"coarse=0x{coarse:02x}")
+      # From decomp:
+      #  - when layout<3 and coarse&1, it sets a "has interesting wave" flag
+      #  - when coarse&8, it marks all live waves as "terminated"
+      if coarse & 0x01:
+        fields.append("flag_wave_interest=1")
+      if coarse & 0x08:
+        fields.append("flag_terminate_all=1")
+    case 0x8:
+      # wave end, this is 20 bits (FFF00)
+      flag7   = (pkt >> 8) & 1
+      simd    = (pkt >> 9) & 3
+      cu      = ((pkt >> 11) & 0x7) | (flag7 << 3)
+      wave    = (pkt >> 15) & 0x1f
+      fields.append(f"wave={wave:x}")
+      fields.append(f"simd={simd}")
+      fields.append(f"cu={cu}")
+    case 0x9:
+      # From case 9 (WAVESTART) in multiple consumers:
+      #   flag7  = (w >> 7) & 1        (low bit of uVar41)
+      #   cls2   = (w >> 8) & 3        (class / group)
+      #   slot4  = (w >> 10) & 0xf     (slot / group index)
+      #   idx_lo = (w >> 0xd) & 0x1f   (low index, layout<4 path)
+      #   idx_hi = (w >> 0xf) & 0x1f   (high index, layout>=4 path)
+      #   id7    = (w >> 0x19) & 0x7f  (7-bit id)
+      flag7   = (pkt >> 7) & 1
+      simd    = (pkt >> 8) & 3
+      cu      = ((pkt >> 10) & 0x7) | (flag7 << 3)
+      wave    = (pkt >> 13) & 0x1F
+      id7     = (pkt >> 17)
+      fields.append(f"wave={wave:x}")
+      fields.append(f"simd={simd}")
+      fields.append(f"cu={cu}")
+      fields.append(f"id7=0x{id7:x}")
+    case 0x18:
+      # FFF88 is the mask
+      # From case 0x18:
+      #   low3   = w & 7
+      #   grp3   = (w >> 3) or (w >> 4) & 7   (layout-dependent)
+      #   flags  = bits 6 (B6) and 7 (B7)
+      #   hi8    = (w >> 0xc) & 0xff   (layout 4 path)
+      #   hi7    = (w >> 0xd) & 0x7f   (other layouts)
+      #   idx5   = (w >> 7) or (w >> 8) & 0x1f, used as wave index
+      flag1 = (pkt >> 3) & 1
+      flag2 = (pkt >> 7) & 1
+      wave = (pkt >> 8) & 0x1F
+      op = (pkt >> 13)
+      fields.append(f"wave={wave:x}")
+      fields.append(f"op=0x{op:02x} [{OPNAME.get(op, '')}]")
+      if flag1: fields.append("flag1")
+      if flag2: fields.append("flag2")
+    case 0x14:
+      subop   = (pkt >> 16) & 0xFFFF       # (short)(w >> 0x10)
+      val32   = (pkt >> 32) & 0xFFFFFFFF   # (uint)(w >> 0x20)
+      slot    = (pkt >> 7) & 0x7           # index in local_168[...] tables
+      hi_byte = (pkt >> 8) & 0xFF          # determines config vs marker
+
+      fields.append(f"subop=0x{subop:04x}")
+      fields.append(f"slot={slot}")
+      fields.append(f"val32=0x{val32:08x}")
+
+      if hi_byte & 0x80:
+        # Config flavour: writes config words into per-slot state arrays.
+        fields.append("kind=config")
+        if subop == 0x000C:
+          fields.append("slot=lo")
+        elif subop == 0x000D:
+          fields.append("slot=hi")
+      else:
+        # COR marker: subop 0xC342, payload "COR\0" → start of a COR region.
+        if subop == 0xC342:
+          fields.append("kind=cor_stream")
+          if val32 == 0x434F5200:
+            fields.append("cor_magic='COR\\0'")
+    case 0x16:
+      # Bits:
+      #   bit8  -> 0x100
+      #   bit9  -> 0x200
+      #   bits 12..47 -> 36-bit field used as delta or marker
+      bit8 = bool(pkt & 0x100)
+      bit9 = bool(pkt & 0x200)
+      if not bit9:
+        mode = "delta"
+      elif not bit8:
+        mode = "marker"
+      else:
+        mode = "other"
+      # need to use reg here
+      val36 = (reg >> 12) & ((1 << 36) - 1)
+      fields.append(f"mode={mode}")
+      if mode != "delta":
+        fields.append(f"val36=0x{val36:x}")
+    case 0x17:
+      # From decomp (two sites with identical logic):
+      #   layout = (w >> 7) & 0x3f
+      #   mode   = (w >> 0xd) & 3
+      #   group  = (w >> 0xf) & 7
+      #   sel_a  = (w >> 0x1c) & 0xf
+      #   sel_b  = (w >> 0x21) & 7
+      #   flag4  = (w >> 0x3b) & 1  (only meaningful when layout == 4)
+      layout = (pkt >> 7)  & 0x3F
+      simd   = (pkt >> 13) & 0x3  # you can change this by changing traced simd
+      group  = (pkt >> 15) & 0x7
+      sel_a  = (pkt >> 0x1C) & 0xF
+      sel_b  = (pkt >> 0x21) & 0x7
+      flag4  = (pkt >> 0x3B) & 0x1
+
+      fields.append(f"layout={layout}")
+      fields.append(f"group={group}")
+      fields.append(f"simd={simd}")
+      fields.append(f"sel_a={sel_a}")
+      fields.append(f"sel_b={sel_b}")
+      if layout == 4:
+        fields.append(f"layout4_flag={flag4}")
+    case _:
+      fields.append(f"{pkt:X} & {reg_mask(opcode):X}")
+  return ",".join(fields)
+
+FILTER_LEVEL = getenv("FILTER", 1)
+
+DEFAULT_FILTER: tuple[int, ...] = tuple()
+# NOP + pure time + "sample"
+if FILTER_LEVEL >= 0: DEFAULT_FILTER += (0x10, 0xf, 0x11)
+# reg + event + sample + marker
+# TODO: events are probably good
+if FILTER_LEVEL >= 1: DEFAULT_FILTER += (0x14, 0x12, 0x16)
+# instruction runs + valuinst
+if FILTER_LEVEL >= 2: DEFAULT_FILTER += (0x01, 0x02, 0x03)
+# instructions dispatch (inst, immed)
+if FILTER_LEVEL >= 3: DEFAULT_FILTER += (0x4, 0x5, 0x18)
+# waves
+if FILTER_LEVEL >= 4: DEFAULT_FILTER += (0x6, 0x8, 0x9)
+
+def parse_sqtt_print_packets(data: bytes, filter=DEFAULT_FILTER, verbose=True) -> None:
+  """
+  Minimal debug: print ONE LINE per decoded token (packet).
+
+  Now prints only the actual nibbles that belong to each packet, instead of
+  the full 64-bit shift register.
+  """
+  n = len(data)
+  time = 0
+  last_printed_time = 0
+  reg = 0          # shift register
+  offset = 0       # bit offset, in steps of 4 (one nibble)
+  nib_budget = 0x40
+  flags = 0
+  token_index = 0
+  opcodes_seen = set()
+
+  while (offset >> 3) < n:
+    # 1) Fill register with nibbles according to nib_budget
+    if nib_budget != 0:
+      target = offset + 4 + ((nib_budget - 1) & ~3)
+      while offset != target and (offset >> 3) < n:
+        byte = data[offset >> 3]
+        nib = (byte >> (offset & 4)) & 0xF
+        reg = ((reg >> 4) | (nib << 60)) & ((1 << 64) - 1)
+        offset += 4
+      if offset != target: break  # don't parse past the end
+
+    # 2) Decode token from low 8 bits
+    opcode = STATE_TO_OPCODE[reg & 0xFF]
+    opcodes_seen.add(opcode)
+
+    # 4) Set next nibble budget based on opcode
+    nib_budget = NIBBLE_BUDGET[opcode & 0x1F]
+
+    # 5) Get delta
+    shift, width = DELTA_MAP_DEFAULT[opcode]
+    delta = (reg >> shift) & ((1 << width) - 1)
+
+    # 6) Update time and handle special opcodes 0xF/0x16
+    if opcode == 0x16:
+      two_bits = (reg >> 8) & 0x3
+      if two_bits == 1:
+        flags |= 0x01
+
+      # Common 36-bit field at bits [12..47]
+      if (reg & 0x200) == 0:
+        # delta mode: add 36-bit delta to time
+        pass
+      elif (reg & 0x100) == 0:
+        # marker / other modes: no time advance
+        # real marker: bit9=1, bit8=0, non-zero payload
+        # "other" 0x16 variants, ignored for timing
+        delta = 0
+      else:
+        raise RuntimeError("unknown 0x16 delta")
+    elif opcode == 0x0F:
+      # opcode 0x0F has an offset of 4 to the delta
+      # update: it's actually computed to be 8 to match WAVESTART
+      delta = delta + 8
+
+    # Append extra decoded fields into the note string
+    note = decode_packet_fields(opcode, reg)
+
+    # this delta happens before the instruction
+    time += delta
+    token_index += 1
+
+    if verbose and (filter is None or opcode not in filter):
+      print(f"{time:8d} +{time-last_printed_time:8d} : "+colored(f"{OPCODE_NAMES[opcode]:18s} ", OPCODE_COLORS.get(opcode, "white"))+f"{note}")
+      last_printed_time = time
+
+  # Optional summary at the end
+  print(f"# done: tokens={token_index:_}, final_time={time}, flags=0x{flags:02x}")
+  if verbose:
+    print(f"opcodes({len(opcodes_seen):2d}):",
+          ' '.join([colored(f"{op:2X}", "WHITE" if op in opcodes_seen else "BLACK") for op in sorted(opcode_mask)]))
+
+
+def parse(fn:str):
+  with Timing(f"unpickle {fn}: "): dat = pickle.load(open(fn, "rb"))
+  #if getenv("ROCM", 0):
+  #  with Timing(f"decode {fn}: "): ctx = decode(dat)
+  dat_sqtt = [x for x in dat if isinstance(x, ProfileSQTTEvent)]
+  print(f"got {len(dat_sqtt)} SQTT events in {fn}")
+  return dat_sqtt
+
+if __name__ == "__main__":
+  fn = "extra/sqtt/examples/profile_gemm_run_0.pkl"
+  dat_sqtt = parse(sys.argv[1] if len(sys.argv) > 1 else fn)
+  for i,dat in enumerate(dat_sqtt):
+    with Timing(f"decode pkt {i} with len {len(dat.blob):_}: "):
+      parse_sqtt_print_packets(dat.blob, verbose=getenv("V", 1))
@@ -1,25 +1,23 @@
-import os, subprocess, sys, shlex
+import os, subprocess
 from pathlib import Path
 from tinygrad.helpers import temp

 EXAMPLES_DIR = Path(__file__).parent
 PROFILE_PATH = Path(temp("profile.pkl", append_user=True))

-EXAMPLES = {
-  "empty":"test/backend/test_custom_kernel.py TestCustomKernel.test_empty",
-  "plus":"test/test_tiny.py TestTiny.test_plus",
-  "gemm":"-c \"from tinygrad import Tensor; (Tensor.empty(N:=32, N)@Tensor.empty(N, N)).realize()\"",
-  "sync":"test/amd/test_custom_kernel.py TestCustomKernel.test_wave_sync",
-}
+EXAMPLES = [
+  "test.backend.test_custom_kernel.TestCustomKernel.test_empty",
+  "test.test_tiny.TestTiny.test_plus",
+  "test.test_tiny.TestTiny.test_gemm",
+]

 if __name__ == "__main__":
  arch = subprocess.check_output(["python", "-c", "from tinygrad import Device; print(Device['AMD'].arch)"], text=True,
                                 env={**os.environ, "DEBUG":"0"}).rstrip()
  (EXAMPLES_DIR/arch).mkdir(exist_ok=True)
-  for name,test in EXAMPLES.items():
+  for test in EXAMPLES:
    for i in range(2):
-      # AM_RESET=1 gets a clear trace, does not work on mi300 machines
-      subprocess.run([sys.executable, *shlex.split(test)], cwd=EXAMPLES_DIR.parent.parent.parent,
-                     env={**os.environ, "AMD":"1", "AM_RESET":"1" if not arch.startswith("gfx9") else "0", "VIZ":"-2", "PYTHONPATH":"."})
-      PROFILE_PATH.rename(dest:=EXAMPLES_DIR/arch/f"profile_{name}_run_{i}.pkl")
+      subprocess.run(["python", "-m", "unittest", test], cwd=EXAMPLES_DIR.parent.parent.parent,
+                     env={**os.environ, "AMD":"1", "SQTT_LIMIT_SE":"-1", "VIZ":"-2"}, check=True)
+      PROFILE_PATH.rename(dest:=EXAMPLES_DIR/arch/f"profile_{test.split('.')[-1].replace('test_', '')}_run_{i}.pkl")
      print(f"saved SQTT trace to {dest}")
@@ -4,8 +4,6 @@ from typing import Generator
 from tinygrad.helpers import temp, unwrap, DEBUG
 from tinygrad.runtime.ops_amd import ProfileSQTTEvent
 from tinygrad.runtime.autogen import rocprof
-from tinygrad.renderer.amd.dsl import Inst
-from test.amd.disasm import disasm

@dataclasses.dataclass(frozen=True)
 class InstExec:
@@ -46,8 +44,8 @@ class OccEvent(WaveSlot):
 RunKey = tuple[str, int]

 class _ROCParseCtx:
-  def __init__(self, sqtt_evs:list[ProfileSQTTEvent], disasms:dict[str, dict[int, Inst]]):
-    self.sqtt_evs, self.disasms = iter(sqtt_evs), {k:{k2:(disasm(v2), v2.size()) for k2,v2 in v.items()} for k,v in disasms.items()}
+  def __init__(self, sqtt_evs:list[ProfileSQTTEvent], disasms:dict[str, dict[int, tuple[str, int]]]):
+    self.sqtt_evs, self.disasms = iter(sqtt_evs), disasms
    self.inst_execs:dict[RunKey, list[WaveExec]] = {}
    self.occ_events:dict[RunKey, list[OccEvent]] = {}

@@ -73,7 +71,7 @@ class _ROCParseCtx:
    self.inst_execs.setdefault(unwrap(self.active_run), []).append(WaveExec(ev.wave_id, ev.cu, ev.simd, unwrap(self.active_se), ev.begin_time,
                                                                             ev.end_time, insts_blob))

-def decode(sqtt_evs:list[ProfileSQTTEvent], disasms:dict[str, dict[int, Inst]]) -> _ROCParseCtx:
+def decode(sqtt_evs:list[ProfileSQTTEvent], disasms:dict[str, dict[int, tuple[str, int]]]) -> _ROCParseCtx:
  ROCParseCtx = _ROCParseCtx(sqtt_evs, disasms)

  @rocprof.rocprof_trace_decoder_se_data_callback_t
@@ -118,7 +116,7 @@ def decode(sqtt_evs:list[ProfileSQTTEvent], disasms:dict[str, dict[int, Inst]])
    nonlocal exc
    try: rocprof.rocprof_trace_decoder_parse_data(copy_cb, trace_cb, isa_cb, None)
    except AttributeError as e:
-      exc = RuntimeError("Failed to find rocprof-trace-decoder. Run sudo ./extra/sqtt/install_rocprof_decoder.py to install")
+      exc = RuntimeError("Failed to find rocprof-trace-decoder. Run sudo ./extra/sqtt/install_sqtt_decoder.py to install")
      exc.__cause__ = e
  (t:=threading.Thread(target=worker, daemon=True)).start()
  t.join()
@@ -11,46 +11,13 @@ from tinygrad.uop.ops import UOp, Ops, KernelInfo
 def _sharded_empty(shape:Tensor, ref:Tensor, axis:int|None, dtype:DTypeLike|None=None) -> Tensor:
  dtype = dtype or ref.dtype
  if not isinstance(ref.device, tuple): return Tensor.empty(*shape, dtype=dtype, device=ref.device)
-  shard_axis = ref.uop.axis if axis is None else axis
-  shape = tuple(s // len(ref.device) if i == shard_axis else s for i, s in enumerate(shape))
+  shape = tuple(s // len(ref.device) if i == ref.uop.axis else s for i, s in enumerate(shape))
  axis = ref.uop.axis if axis is None else axis
  return Tensor(Tensor.empty(*shape, dtype=dtype, device=ref.device).uop.multi(axis), dtype=dtype, device=ref.device)

 def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
  return _sharded_empty(ref.shape, ref, axis)

-@functools.cache
-def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch):
-  def grad(dou:UOp, ker:UOp) -> tuple[None, None, UOp, UOp, UOp]:
-    do = Tensor(dou, device=dou.device)
-    attn = Tensor(ker.src[1].after(ker), device=ker.src[1].device)
-    l_vec = Tensor(ker.src[2].after(ker), device=ker.src[2].device)
-    xq = Tensor(ker.src[3], device=ker.src[3].device)
-    xk = Tensor(ker.src[4], device=ker.src[4].device)
-    xv = Tensor(ker.src[5], device=ker.src[5].device)
-
-    dq = _sharded_empty((B, H, N, D), xq, axis=shard_axis_t)
-    GROUP_SIZE = H_local // H_KV_local
-    dk_partial = _sharded_empty((B * GROUP_SIZE, N, H_KV, D), xk, axis=shard_axis)
-    dv_partial = _sharded_empty((B * GROUP_SIZE, N, H_KV, D), xv, axis=shard_axis)
-
-    # delta_vec = (do * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()
-    delta_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
-    delta_vec, dq = Tensor.custom_kernel(delta_vec, dq, attn, do, fxn=functools.partial(custom_fa_backward_pre, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:2]
-
-    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:3]
-
-    # unshuffle dq: atomic_pk_add_bf16_with_warpid creates a shuffled layout within each 16x128 tile
-    # decompose each tile into (j=4, a=2, b=2, d=4, e=4, k=4, c=2) and permute to (e, k, j, a, d, b, c) = standard row-major
-    dq = dq.reshape(B, H, N//16, 4, 2, 2, 4, 4, 4, 2).permute(0, 1, 2, 7, 8, 3, 4, 6, 5, 9).reshape(B, H, N, D).transpose(1, 2)
-
-    # reduce partial dK/dV across GROUP_SIZE query heads
-    dk = dk_partial.reshape(B, GROUP_SIZE, N, H_KV, D).sum(1)
-    dv = dv_partial.reshape(B, GROUP_SIZE, N, H_KV, D).sum(1)
-
-    return None, None, dq.uop, dk.uop, dv.uop
-  return grad
-
 def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False):
  assert attn_mask is None, "attn_mask not supported"
  assert is_causal, "only causal attention supported"
@@ -62,29 +29,42 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
  assert D == 128, "only D=128 supported"

  num_devices = len(xq.device) if isinstance(xq.device, tuple) else 1
-  is_dp = xq.uop.axis == 0
-  is_mp = xq.uop.axis == 2
-  B_local = B // num_devices if is_dp else B
-  H_local = H // num_devices if is_mp else H
-  H_KV_local = H_KV // num_devices if is_mp else H_KV
-  shard_axis = 0 if is_dp else 2 if is_mp else None
-  shard_axis_t = 0 if is_dp else 1 if is_mp else None
-  if DEBUG >= 2: print(f"Flash Attention {B=} {B_local=} {N=} {H=} {H_local=} {H_KV=} {H_KV_local=} {D=} on {num_devices} devices, {'DP' if is_dp else 'MP' if is_mp else 'no sharding'}")
+  B_local = B // num_devices
+  if DEBUG >= 2: print(f"Flash Attention {B=} {B_local=} {N=} {H=} {H_KV=} {D=}")

  single_device = xq.device[0] if isinstance(xq.device, tuple) else xq.device
  arch = Device[single_device].renderer.arch

-  attn = _sharded_empty_like(xq, axis=shard_axis)
-  l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
+  attn = _sharded_empty_like(xq, axis=0)
+  l_vec = _sharded_empty((B, H, 1, N), xq, axis=0, dtype=dtypes.float32)

-  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch)
+  def grad(dou:UOp, _) -> tuple[None, None, UOp, UOp, UOp]:
+    do = Tensor(dou, device=dou.device)
+    dq_in = _sharded_empty((B, H, N, D), xq, axis=0)
+    dq = _sharded_empty_like(xq, axis=0)
+    dk = _sharded_empty_like(xk, axis=0)
+    dv = _sharded_empty_like(xv, axis=0)

-  attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D), grad_fxn=grad)[:2]
+    # delta_vec = (do * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()
+    delta_vec = _sharded_empty((B, H, 1, N), xq, axis=0, dtype=dtypes.float32)
+    delta_vec, dq_in = Tensor.custom_kernel(delta_vec, dq_in, attn, do, fxn=functools.partial(custom_fa_backward_pre, device=single_device, arch=arch))[:2]
+
+    dq_in, dk, dv = Tensor.custom_kernel(dq_in, dk, dv, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch))[:3]
+
+    # unshuffle dq
+    dq = Tensor.custom_kernel(dq, dq_in, fxn=functools.partial(custom_fa_backward_post, device=single_device, arch=arch))[0]
+
+    return None, None, dq.uop, dk.uop, dv.uop
+
+  attn, l_vec = Tensor.custom_kernel(attn, l_vec, xq, xk, xv, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch), grad_fxn=grad)[:2]

  return attn.transpose(1, 2)

@functools.cache
-def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:str):
+  B, N, H, D = q.shape
+  H_KV = k.shape[2]
+
  code = (pathlib.Path(__file__).parent / "fa_fwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}"]
@@ -105,6 +85,7 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:st
                  arg=KernelInfo(name="custom_fa_forward", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
+
  lib = bytearray(lib)
  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
  struct.pack_into('<I', lib, rodata_off, 160000)
@@ -114,7 +95,9 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, device:str, arch:st
             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
-def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arch:str):
+  B, N, H, D = o.shape
+
  code = (pathlib.Path(__file__).parent / "fa_bwd_pre.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}"]
@@ -135,6 +118,7 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
                  arg=KernelInfo(name="custom_fa_backward_pre", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
+
  lib = bytearray(lib)
  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
  struct.pack_into('<I', lib, rodata_off, 160000)
@@ -144,7 +128,10 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
-def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str):
+  B, N, H, D = q.shape
+  H_KV = k.shape[2]
+
  code = (pathlib.Path(__file__).parent / "fa_bwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}"]
@@ -152,7 +139,7 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
  BLOCK_SIZE_KV = 256
  NUM_WARPS = 4
  NUM_THREADS = 64 * NUM_WARPS
-  gsz = (H, N // BLOCK_SIZE_KV, B)
+  gsz = (H_KV, N // BLOCK_SIZE_KV, B)
  lsz = (NUM_THREADS, 1, 1)
  threadIdx_x = UOp.special(lsz[0], "lidx0")
  blockIdx_x, blockIdx_y, blockIdx_z = UOp.special(gsz[0], "gidx0"), UOp.special(gsz[1], "gidx1"), UOp.special(gsz[2], "gidx2")
@@ -165,6 +152,7 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
                  arg=KernelInfo(name="custom_fa_backward", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
+
  lib = bytearray(lib)
  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
  struct.pack_into('<I', lib, rodata_off, 160000)
@@ -174,7 +162,9 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
             src=(sink, UOp(Ops.DEVICE, arg=device), UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
-def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
+def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str):
+  B, N, H, D = dq_out.shape
+
  code = (pathlib.Path(__file__).parent / "fa_bwd_post.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}"]
@@ -195,6 +185,7 @@ def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int,
                  arg=KernelInfo(name="custom_fa_backward_post", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
+
  lib = bytearray(lib)
  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
  struct.pack_into('<I', lib, rodata_off, 160000)
@@ -37,7 +37,7 @@ using namespace kittens;
 using _gl_QdO  = gl<bf16, ATTN_B, ATTN_N, ATTN_H, ATTN_D>;
 using _gl_KV   = gl<bf16, ATTN_B, ATTN_N, ATTN_H_KV, ATTN_D>;
 using _gl_dQ   = gl<bf16, ATTN_B, ATTN_H, ATTN_N, ATTN_D>;
-using _gl_dKV  = gl<bf16, ATTN_B * GROUP_SIZE, ATTN_N, ATTN_H_KV, ATTN_D>;
+using _gl_dKV  = gl<bf16, ATTN_B, ATTN_N, ATTN_H_KV, ATTN_D>;
 using _gl_Lvec = gl<float, ATTN_B, ATTN_H, 1, ATTN_N>;

 template<int D> struct attn_bwd_combined_globals {
@@ -47,7 +47,7 @@ template<int D> struct attn_bwd_combined_globals {
  _gl_dQ dQg;
  _gl_dKV dKg, dVg;
  _gl_Lvec L_vec, delta_vec;
-  dim3 grid() { return dim3(ATTN_H, (ATTN_N / BLOCK_SIZE_KV), ATTN_B); }
+  dim3 grid() { return dim3(ATTN_H_KV, (ATTN_N / BLOCK_SIZE_KV), ATTN_B); }
  dim3 block() { return dim3(NUM_THREADS); }
  size_t dynamic_shared_memory() { return MAX_SHARED_MEMORY; }
 };
@@ -55,12 +55,10 @@ template<int D> struct attn_bwd_combined_globals {
 template<int D> __launch_bounds__(NUM_THREADS, 1)
 __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr, bf16 *dO_ptr, bf16 *Q_ptr, bf16 *K_ptr, bf16 *V_ptr, float *L_vec_ptr, float *delta_vec_ptr) {

-  const int q_head_idx_fixed = blockIdx.x;  // This is the query head index [0, ATTN_H)
-  const int kv_head_idx = q_head_idx_fixed / GROUP_SIZE;
-  const int q_head_in_group = q_head_idx_fixed % GROUP_SIZE;
+  const int kv_head_idx = blockIdx.x;  // This is the KV head index
  const int seq_idx = blockIdx.y;
  const int batch_idx = blockIdx.z;
-  const int first_q_head = q_head_idx_fixed;
+  const int first_q_head = kv_head_idx * GROUP_SIZE;

  const int warpid = kittens::warpid();
  const int j = seq_idx * NUM_WARPS + warpid;
@@ -72,7 +70,7 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
  // first Q step that can overlap this K_span:
  const int first_step = max(0, k_start_min / STEP_QO);
  const int num_steps_per_head = total_steps_per_head - first_step;
-  const int num_steps = num_steps_per_head;
+  const int num_steps = num_steps_per_head * GROUP_SIZE;
  const int k_pos = j * WARP_SIZE_KV;

  constexpr float L_SCALE_FACTOR = 1.44269504089f;
@@ -3357,14 +3355,14 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
    }
  }

-  store<1>(g.dVg, dV_j, {batch_idx * GROUP_SIZE + q_head_in_group, 0, kv_head_idx, 0}, {0, j, 0, 0});
+  store<1>(g.dVg, dV_j, {batch_idx, 0, kv_head_idx, 0}, {0, j, 0, 0});
  __builtin_amdgcn_s_waitcnt(0);
  __builtin_amdgcn_s_barrier();

  // We first copy dV_j_T from accumulator GPRs to vector GPRs and then perform the store
  accvgpr_read(dV_j_T, dK_j_T);
  mul(dV_j_T, dV_j_T, dP_SCALE_FACTOR);
-  store<1>(g.dKg, dV_j, {batch_idx * GROUP_SIZE + q_head_in_group, 0, kv_head_idx, 0}, {0, j, 0, 0});
+  store<1>(g.dKg, dV_j, {batch_idx, 0, kv_head_idx, 0}, {0, j, 0, 0});

  // Write out final dQ_i slice
  mul(dQ_i_T, dQ_i_T, dP_SCALE_FACTOR);
@@ -66,7 +66,7 @@ template<int D, typename T=bf16, typename L=row_l, typename S=rt_32x16_s> using
 template<int D, typename T=bf16, typename L=col_l, typename S=rt_16x32_s> using qo_tile_transposed = rt<T, D, Q_BLOCK_SIZE, L, S>;
 template<int D, typename T=bf16, typename L=row_l, typename S=rt_32x16_s> using kv_tile = rt<T, KV_BLOCK_SIZE, D, L, S>;
 template<int D, typename T=bf16, typename L=col_l, typename S=rt_16x32_s> using kv_tile_transposed = rt<T, D, KV_BLOCK_SIZE, L, S>;
-template<typename T=float, typename L=col_l, typename S=rt_16x32_4_s> using attn_tile = rt<T, KV_BLOCK_SIZE, Q_BLOCK_SIZE, L, S>;
+template<int D, typename T=float, typename L=col_l, typename S=rt_16x32_4_s> using attn_tile = rt<T, KV_BLOCK_SIZE, Q_BLOCK_SIZE, L, S>;

 /**********************************************************/
 template<int THR_X, int THR_Y>
@@ -103,7 +103,7 @@ __device__ inline void mask_kv_tile(RT &dst, int q_abs, int k_abs, uint32_t neg_

    #pragma unroll
    for (int i = 0; i < dst.height; ++i) {
-        // Row base of the 32x* chunk produced by MFMA
+        // Row base of the 32x* chunk produced by MFMA 
        const int row_base = (i * 32) + ((lane >> 5) << 2); // multiplesof 4

        // Relative index of the FIRST element in this row-chunk w.r.t. q_pos
@@ -148,7 +148,7 @@ __device__ inline void mask_kv_tile(RT &dst, int q_abs, int k_abs, uint32_t neg_
 /**********************************************************/

 template<int D> struct attn_globals {
-    _gl_QKVO Qg, Kg, Vg, Og;
+    _gl_QKVO Qg, Kg, Vg, Og; 
    gl<float, -1, -1, -1, -1> L_vec;
    dim3 grid() { return dim3(ATTN_H, ((ATTN_N / Q_BLOCK_SIZE + NUM_WARPS - 1) / NUM_WARPS), ATTN_B); }
    dim3 block() { return dim3(NUM_THREADS); }
@@ -196,10 +196,10 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p

    kv_tile<D, bf16, col_l, rt_16x32_4_s> v_reg;
    qo_tile_transposed<D, float, col_l, rt_32x32_s> o_reg; // Output tile.
-    attn_tile<float, col_l, rt_32x32_s> att_block[2]; // attention tile, in float.
-    attn_tile<bf16, col_l, rt_32x32_s> att_block_bf16;
-    attn_tile<bf16, col_l, rt_16x32_4_s> att_block_bf16_in;
-    typename attn_tile<float, col_l, rt_32x32_s>::row_vec max_vec, norm_vec, max_vec_prev, scale_vec;
+    attn_tile<D, float, col_l, rt_32x32_s> att_block[2]; // attention tile, in float.
+    attn_tile<D, bf16, col_l, rt_32x32_s> att_block_bf16;
+    attn_tile<D, bf16, col_l, rt_16x32_4_s> att_block_bf16_in;
+    typename attn_tile<D, float, col_l, rt_32x32_s>::row_vec max_vec, norm_vec, max_vec_prev, scale_vec;

    zero(o_reg);
    zero(norm_vec);
@@ -241,8 +241,8 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    zero(att_block[0]);
    transpose(k_reg_transposed, k_reg);
    mma_AtB(att_block[0], k_reg_transposed, q_reg_transposed, att_block[0]);
-    __builtin_amdgcn_sched_barrier(0);
-    if constexpr (causal) {
+    __builtin_amdgcn_sched_barrier(0); 
+    if constexpr (causal) { 
        const int kv_end_pos = (1) * KV_BLOCK_SIZE;
        if (__builtin_expect(q_start_pos < kv_end_pos, 0)) {  // Only mask if needed
            mask_kv_tile(att_block[0], tile_idx, 0, neg_inf_v, lane);
@@ -269,7 +269,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    load(k_reg, k_smem[1]);
    // All warps then collaboratively load in the third slice of K (K2) into shared memory
    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, 2, head_idx_kv, 0}, swizzled_offsets_K);
-    // All warps then collaboratively load in the second slice of V (V1) into shared memory
+    // All warps then collaboratively load in the second slice of V (V1) into shared memory 
    G::load<1, false>(v_smem[1], g.Vg, {batch_idx, 1, head_idx_kv, 0}, swizzled_offsets_V);
    asm volatile("s_waitcnt lgkmcnt(0)");
    asm volatile("s_waitcnt vmcnt(4)");
@@ -288,7 +288,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
        mul(norm_vec, norm_vec, scale_vec);
        col_sum(norm_vec, att_block[0], norm_vec);
        copy(att_block_bf16, att_block[0]);
-        att_block_bf16_in = *reinterpret_cast<attn_tile< bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
+        att_block_bf16_in = *reinterpret_cast<attn_tile<D, bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
        sched_barrier_exp_pairs<6, 3, 1>();
        sched_barrier_pairs<10, 5, 1>();
        __builtin_amdgcn_sched_barrier(0);
@@ -296,7 +296,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
        __builtin_amdgcn_sched_barrier(0);

        // Cluster 1:
-        //      Load K3 into shared
+        //      Load K3 into shared 
        G::load<1, false>(k_smem[1], g.Kg, {batch_idx, j, head_idx_kv, 0}, swizzled_offsets_K);
        //      Load V0 into registers
        load(v_reg, v_smem[0]);
@@ -348,7 +348,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
        mul(norm_vec, norm_vec, scale_vec);
        col_sum(norm_vec, att_block[1], norm_vec);
        copy(att_block_bf16, att_block[1]);
-        att_block_bf16_in = *reinterpret_cast<attn_tile<bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
+        att_block_bf16_in = *reinterpret_cast<attn_tile<D, bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
        sched_barrier_exp_pairs<6, 3, 3>();
        sched_barrier_pairs<10, 5, 3>();
        __builtin_amdgcn_s_setprio(0);
@@ -417,7 +417,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p

    col_sum(norm_vec, att_block[0], norm_vec);
    copy(att_block_bf16, att_block[0]);
-    att_block_bf16_in = *reinterpret_cast<attn_tile<bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
+    att_block_bf16_in = *reinterpret_cast<attn_tile<D, bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
    sched_barrier_exp_pairs<6, 3, 5>();
    sched_barrier_pairs<10, 5, 5>();
    __builtin_amdgcn_sched_barrier(0);
@@ -482,7 +482,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    mul(norm_vec, norm_vec, scale_vec);
    col_sum(norm_vec, att_block[1], norm_vec);
    copy(att_block_bf16, att_block[1]);
-    att_block_bf16_in = *reinterpret_cast<attn_tile<bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
+    att_block_bf16_in = *reinterpret_cast<attn_tile<D, bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
    sched_barrier_exp_pairs<6, 3, 7>();
    sched_barrier_pairs<10, 5, 7>();
    __builtin_amdgcn_sched_barrier(0);
@@ -544,7 +544,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    mul(norm_vec, norm_vec, scale_vec);
    col_sum(norm_vec, att_block[0], norm_vec);
    copy(att_block_bf16, att_block[0]);
-    att_block_bf16_in = *reinterpret_cast<attn_tile<bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
+    att_block_bf16_in = *reinterpret_cast<attn_tile<D, bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
    sched_barrier_exp_pairs<6, 3, 9>();
    sched_barrier_pairs<10, 5, 9>();
    __builtin_amdgcn_sched_barrier(0);
@@ -586,7 +586,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p

    col_sum(norm_vec, att_block[1], norm_vec);
    copy(att_block_bf16, att_block[1]);
-    att_block_bf16_in = *reinterpret_cast<attn_tile<bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);
+    att_block_bf16_in = *reinterpret_cast<attn_tile<D, bf16, col_l, rt_16x32_4_s>*>(&att_block_bf16);

    __builtin_amdgcn_sched_barrier(0);
    mul_col(o_reg, o_reg, scale_vec);
@@ -505,9 +505,7 @@ tiny_backend_out = {**{f"aten.{x}.out":getattr(Tensor,x) for x in simple_tensor_
  "aten.lt.Tensor_out": Tensor.__lt__, "aten.lt.Scalar_out": Tensor.__lt__,
  "aten.le.Tensor_out": Tensor.__le__, "aten.le.Scalar_out": Tensor.__le__,
  "aten.clamp_max.Tensor_out": lambda input,max_: input.clamp(max_=max_),
-  "aten.clamp_max.out": lambda input,max_: input.clamp(max_=max_),
  "aten.clamp_min.Tensor_out": lambda input,min_: input.clamp(min_=min_),
-  "aten.clamp_min.out": lambda input,min_: input.clamp(min_=min_),
  "aten.fmod.Tensor_out": lambda input,other: input-input.div(other, rounding_mode="trunc")*other,
  # TODO: this might result in overflow issues
  "aten.round.decimals_out": lambda self,decimals: (self*10**decimals).round()/10**decimals,
@@ -590,7 +588,7 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
  "aten.repeat": lambda x,*repeats: Tensor.repeat(x,*repeats).contiguous(), # not a view
  "aten._softmax": lambda self,dim,half_to_float: self.softmax(dim),
  "aten._log_softmax": lambda self,dim,half_to_float: self.log_softmax(dim),
-  "aten.random_": lambda self: Tensor.randint(*self.shape, low=self.dtype.min, high=self.dtype.max, device=self.device, dtype=self.dtype),
+  "aten.random_": lambda self: Tensor.randint(*self.shape, low=dtypes.min(self.dtype), high=dtypes.max(self.dtype), device=self.device, dtype=self.dtype),
  "aten.random_.from": lambda self, from_, to: Tensor.randint(*self.shape, low=from_, high=to, device=self.device, dtype=self.dtype),
  "aten.uniform_": lambda self, low=0, high=1: Tensor.uniform(*self.shape, low=low, high=high, dtype=self.dtype),
  "aten.normal_": lambda self, mean=0, std=1: Tensor.normal(*self.shape, mean=mean, std=std, dtype=self.dtype),
@@ -23,8 +23,7 @@ if __name__ == "__main__":

  kernel_count = GlobalCounters.kernel_count
  assert kernel_count > 0, "No kernels, test failed"
-  # NOTE: this is 124 on torch 2.10.0
-  expected_kernels = 332
+  expected_kernels = 228
  expectation = f"ResNet18 kernels are {kernel_count} vs {expected_kernels} expected."
  if kernel_count < expected_kernels: warnings.warn(f"{expectation} Expectation can be lowered.", UserWarning)
  assert kernel_count <= expected_kernels, f"{expectation}"
@@ -1,6 +1,7 @@
 # simple tests
 import unittest
 import torch
+import warnings
 from tinygrad.helpers import getenv, GlobalCounters
 if getenv("TINY_BACKEND2"):
  import extra.torch_backend.backend2
@@ -17,13 +18,15 @@ class TestKernelFusionRegression(unittest.TestCase):
    torch.manual_seed(42)
    GlobalCounters.reset()
    fn().detach().cpu().numpy()
-    self.assertEqual(GlobalCounters.kernel_count, expected_kernels)
+    expectation = f"{GlobalCounters.kernel_count} vs {expected_kernels} expected."
+    if GlobalCounters.kernel_count < expected_kernels: warnings.warn(f"{expectation} Expectation can be lowered.", UserWarning)
+    self.assertLessEqual(GlobalCounters.kernel_count, expected_kernels, f"{expectation}")

  def test_elementwise_fusion(self):
    def fn():
      x = torch.randn(128, 128, device=device)
      return (x + 1.0) * 2.0 - 0.5
-    self._check_kernel_count(fn, 5)
+    self._check_kernel_count(fn, 6)

  def test_relu_fusion(self):
    def fn():
@@ -31,7 +34,7 @@ class TestKernelFusionRegression(unittest.TestCase):
      conv = torch.nn.Conv2d(3, 16, 3, padding=1).to(device)
      with torch.no_grad():
        return torch.nn.functional.relu(conv(x))
-    self._check_kernel_count(fn, 6)
+    self._check_kernel_count(fn, 8)

  def test_batchnorm_fusion(self):
    def fn():
@@ -41,20 +44,20 @@ class TestKernelFusionRegression(unittest.TestCase):
      bn.eval()
      with torch.no_grad():
        return torch.nn.functional.relu(bn(conv(x)))
-    self._check_kernel_count(fn, 10)
+    self._check_kernel_count(fn, 16)

  def test_reduce_fusion(self):
    def fn():
      x = torch.randn(64, 64, device=device)
      return (x * 2.0).sum()
-    self._check_kernel_count(fn, 5)
+    self._check_kernel_count(fn, 7)

  def test_matmul_elementwise_fusion(self):
    def fn():
      x = torch.randn(32, 32, device=device)
      w = torch.randn(32, 32, device=device)
      return torch.nn.functional.relu(x @ w + 1.0)
-    self._check_kernel_count(fn, 7)
+    self._check_kernel_count(fn, 6)

  def test_pooling_fusion(self):
    def fn():
@@ -68,7 +71,7 @@ class TestKernelFusionRegression(unittest.TestCase):
      identity = torch.randn(1, 8, 16, 16, device=device)
      out = x + identity
      return torch.nn.functional.relu(out)
-    self._check_kernel_count(fn, 7)
+    self._check_kernel_count(fn, 6)

  def test_inplace_add_relu_fusion(self):
    def fn():
@@ -76,7 +79,7 @@ class TestKernelFusionRegression(unittest.TestCase):
      y = torch.randn(1, 16, 32, 32, device=device)
      x += y
      return torch.nn.functional.relu(x)
-    self._check_kernel_count(fn, 7)
+    self._check_kernel_count(fn, 6)

  def test_conv_bn_add_relu_fusion(self):
    def fn():
@@ -89,7 +92,7 @@ class TestKernelFusionRegression(unittest.TestCase):
        out = bn(conv(x))
        out += identity
        return torch.nn.functional.relu(out)
-    self._check_kernel_count(fn, 12)
+    self._check_kernel_count(fn, 16)

  def test_multiple_inplace_ops_fusion(self):
    def fn():
@@ -114,7 +117,7 @@ class TestKernelFusionRegression(unittest.TestCase):
      bn.train()
      with torch.no_grad():
        return bn(x)
-    self._check_kernel_count(fn, 8)
+    self._check_kernel_count(fn, 10)

  # this is a minimal extra/other_mnist/beautiful_mnist_torch.py to cover fusion for training with optimizer
  def test_mnist_training_fusion(self):
@@ -135,7 +138,7 @@ class TestKernelFusionRegression(unittest.TestCase):
      loss.backward()
      optimizer.step()
      return loss
-    self._check_kernel_count(fn, 24)
+    self._check_kernel_count(fn, 33)

 if __name__ == "__main__":
  unittest.main()
@@ -1,17 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
-<plist version="1.0">
-<dict>
-	<key>com.apple.application-identifier</key>
-	<string>9YG3G8543N.org.tinygrad.tinygpu.edriver</string>
-	<key>com.apple.developer.driverkit</key>
-	<true/>
-	<key>com.apple.developer.driverkit.transport.pci</key>
-	<array>
-		<dict>
-			<key>IOPCIPrimaryMatch</key>
-			<string>0x000010de&amp;0x0000FFFF</string>
-		</dict>
-	</array>
-</dict>
-</plist>
@@ -1,33 +0,0 @@
-#!/bin/bash
-set -e
-
-xcodebuild clean build CODE_SIGN_IDENTITY="" CODE_SIGNING_REQUIRED=NO -alltargets -configuration Release build
-
-cp "../profiles/edriver_rel_2.provisionprofile" "./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext/embedded.provisionprofile"
-cp "../profiles/installer_provisioning.provisionprofile" "./build/Release/TinyGPU.app/Contents/embedded.provisionprofile"
-
-codesign \
-    --sign "Developer ID Application: tinygrad, Corp. (9YG3G8543N)" \
-    --entitlements ./TinyGPUDriverExtension/TinyGPUDriver.NV.Release.entitlements \
-    --verbose \
-    --options runtime \
-    --timestamp \
-    --force \
-    ./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext
-
-codesign \
-    --sign "Developer ID Application: tinygrad, Corp. (9YG3G8543N)" \
-    --entitlements ./macOS/macOS.entitlements \
-    --options runtime \
-    --verbose \
-    --timestamp \
-    --force \
-    ./build/Release/TinyGPU.app
-
-codesign --verify --deep --strict --verbose=4 ./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext
-
-codesign --verify --deep --strict --verbose=4 ./build/Release/TinyGPU.app
-
-spctl -a -vv ./build/Release/TinyGPU.app
-
-spctl -a -vv ./build/Release/TinyGPU.app/Contents/Library/SystemExtensions/org.tinygrad.tinygpu.edriver.dext
@@ -3,7 +3,3 @@ set -e

 ditto -c -k --keepParent ./build/Release/TinyGPU.app ./build/Release/TinyGPU.zip
 xcrun notarytool submit ./build/Release/TinyGPU.zip --keychain-profile "hgwJFhdheiIEy82nDN" --wait
-
-rm ./build/Release/TinyGPU.zip
-xcrun stapler staple ./build/Release/TinyGPU.app
-ditto -c -k --keepParent ./build/Release/TinyGPU.app ./build/Release/TinyGPU.zip
@@ -1,17 +1,17 @@
 A command line tool for exploring the VIZ trace.

-After running with VIZ=-1, use `extra/viz/cli.py` to explore the saved trace files.
+After running with VIZ=-1, use `PYTHONPATH=. extra/viz/cli.py` to explore the saved trace files.

 ## Inspect runtime profiling

-Use `extra/viz/cli.py --profile` to list all traced devices.
+Use `PYTHONPATH=. extra/viz/cli.py --profile` to list all traced devices.

 List top slowest kernels on a device: `--profile --device "AMD"`
 List samples of a kernel on a device: `--profile --device "AMD" --kernel E_3`

 ## Inspect codegen and PatternMatcher

-Use `extra/viz/cli.py --rewrites` to list all traced kernels.
+Use `PYTHONPATH=. extra/viz/cli.py --rewrites` to list all traced kernels.

 List all codegen steps for a kernel: `--rewrites --kernel E_3`
 Get source code: `--rewrites --kernel E_3 --select "View Source"`
@@ -1,78 +1,44 @@
 #!/usr/bin/env python3
 import os
 os.environ["VIZ"] = "0"
-import argparse, pathlib, sys, struct, json
+import argparse, pathlib
 from typing import Iterator
 from tinygrad.viz import serve as viz
 from tinygrad.uop.ops import RewriteTrace
 from tinygrad.helpers import temp, ansistrip, colored, time_to_str, ansilen
-
-# ** generic helpers
+from test.null.test_viz import load_profile

 def optional_eq(val:dict, arg:str|None) -> bool: return arg is None or ansistrip(val["name"]) == arg

 def print_data(data:dict) -> None:
  if isinstance(data.get("value"), Iterator):
    for m in data["value"]:
-      if m.get("uop"): print(f"Input UOp:\n{m['uop']}")
-      if m.get("diff"):
-        loc = pathlib.Path(m["upat"][0][0])
-        print(f"Rewrite at {loc.parent.name}/{loc.name}:{m['upat'][0][1]}\n{m['upat'][1]}")
-        for line in m["diff"]: print(colored(line, "red" if line.startswith("-") else "green" if line.startswith("+") else None))
+      if m.get("uop"):
+        print("Input UOp:")
+        print(m["uop"])
+      if not m["diff"]: continue
+      print("Rewrites:")
+      fp = pathlib.Path(m["upat"][0][0])
+      print(f"{fp.parent.name}/{fp.name}:{m['upat'][0][1]}")
+      print(m["upat"][1])
+      for line in m["diff"]:
+        color = "red" if line.startswith("-") else "green" if line.startswith("+") else None
+        print(colored(line, color))
  if data.get("src") is not None: print(data["src"])

-# ** Profiler trace decoder
-
-# 0 means None, otherwise it's an enum value
-def option(i:int) -> int|None: return None if i == 0 else i-1
-
-def decode_profile(data:bytes) -> dict:
-  ret, off = data, 0
-  def u(fmt:str) -> tuple:
-    nonlocal off
-    vals = struct.unpack_from(fmt, ret, off)
-    off += struct.calcsize(fmt)
-    return vals
-  total_dur, global_peak, index_len, layout_len = u("<IQII")
-  strings, dtypes, markers = json.loads(ret[off:off+index_len]).values()
-  off += index_len
-  layout:dict[str, dict] = {}
-  for _ in range(layout_len):
-    klen = u("<B")[0]
-    k = ret[off:off+klen].decode()
-    off += klen
-    layout[k] = v = {"events":[]}
-    event_type, event_count = u("<BI")
-    if event_type == 0:
-      for _ in range(event_count):
-        name, ref, key, st, dur, fmt = u("<IIIIfI")
-        v["events"].append({"name":strings[name], "ref":option(ref), "key":option(key), "st":st, "dur":dur, "fmt":strings[fmt]})
-    else:
-      v["linear"] = u("<B")[0]
-      v["peak"] = u("<Q")[0]
-      for _ in range(event_count):
-        if v["linear"]:
-          ts, value = u("<IQ")
-          v["events"].append({"event":"freq", "ts":ts, "value":value})
-        else:
-          alloc, ts, key = u("<BII")
-          if alloc: v["events"].append({"event":"alloc", "ts":ts, "key":key, "arg": {"dtype":strings[u("<I")[0]], "sz":u("<Q")[0]}})
-          else: v["events"].append({"event":"free", "ts":ts, "key":key, "arg": {"users":[u("<IIIB") for _ in range(u("<I")[0])]}})
-  return {"dur":total_dur, "peak":global_peak, "layout":layout, "markers":markers}
-
 if __name__ == "__main__":
  parser = argparse.ArgumentParser()
  g_mode = parser.add_argument_group("mode")
  g_mode.add_argument("--profile", action="store_true", help="View profile trace")
  g_mode.add_argument("--rewrites", action="store_true", help="View rewrites trace")
-  g_common = parser.add_argument_group("common options")
-  g_common.add_argument("--kernel", type=str, default=None, metavar="NAME", help="Select a kernel by name (optional name, default: only list names)")
  g_profile = parser.add_argument_group("profile options")
  g_profile.add_argument("--device", type=str, default=None, metavar="NAME", help="Select a device (optional name, default: only list names)")
  g_profile.add_argument("--top", type=int, default=10, metavar="N", help="Number of top kernels to show (-1 for all, default: 10)")
  g_rewrites = parser.add_argument_group("rewrites options")
  g_rewrites.add_argument("--select", type=str, default=None, metavar="NAME",
                          help="Select an item within the chosen kernel (optional name, default: only list names)")
+  g_common = parser.add_argument_group("common options")
+  g_common.add_argument("--kernel", type=str, default=None, metavar="NAME", help="Select a kernel by name (optional name, default: only list names)")
  parser.add_argument("--profile-path", type=pathlib.Path, metavar="PATH", help="Path to profile (optional file, default: latest profile)",
                        default=pathlib.Path(temp("profile.pkl", append_user=True)))
  parser.add_argument("--rewrites-path", type=pathlib.Path, metavar="PATH", help="Path to rewrites (optional file, default: latest rewrites)",
@@ -80,14 +46,14 @@ if __name__ == "__main__":
  args = parser.parse_args()
  if not args.profile and not args.rewrites:
    parser.print_help()
-    sys.exit(0)
+    exit(0)

  viz.trace = viz.load_pickle(args.rewrites_path, default=RewriteTrace([], [], {}))
  viz.ctxs = viz.get_rewrites(viz.trace)

  if args.profile:
    from tabulate import tabulate
-    profile = decode_profile(viz.get_profile(viz.load_pickle(args.profile_path, default=[])))
+    profile = load_profile(viz.load_pickle(args.profile_path, default=[]))
    agg, total, n = {}, 0, 0
    if args.device is None: print("Select a device:")
    for k,v in profile["layout"].items():
@@ -97,7 +63,7 @@ if __name__ == "__main__":
      for e in v.get("events", []):
        et = e["dur"]*1e-6
        if args.kernel is not None:
-          if optional_eq(e, args.kernel) and n < 10:
+          if ansistrip(e["name"]) == args.kernel and n < 10:
            ptm = colored(time_to_str(et, w=9), "yellow" if et > 0.01 else None) if et is not None else ""
            name = e["name"]+(" " * (46 - ansilen(e["name"])))
            print(f"{name} {ptm}/{(et or 0)*1e3:9.2f}ms  "+e['fmt'].replace('\n', ' | ')+"  ")
@@ -115,7 +81,7 @@ if __name__ == "__main__":
        other_t = total-sum(t for _, (t, _) in sel)
        table.append([f"Other ({len(other)} unique)", time_to_str(other_t, w=9), sum(c for _,(_,c) in other), f"{other_t/total*100.0:.2f}%"])
      print(tabulate(table, headers=["name", "total", "count", "pct"], tablefmt="github"))
-    sys.exit(0)
+    exit(0)

  for k in viz.ctxs:
    if not optional_eq(k, args.kernel): continue
@@ -74,7 +74,7 @@ testing_minimal = [
  "hypothesis>=6.148.9",
  "z3-solver<4.15.4",  # 4.15.4 has a segfault when creating many z3.Context()
 ]
-testing_unit = ["tinygrad[testing_minimal]", "tqdm", "safetensors", "tabulate", "openai", "gguf>=0.18"]
+testing_unit = ["tinygrad[testing_minimal]", "tqdm", "safetensors", "tabulate", "openai", "ggml-python"]
 testing = [
  "tinygrad[testing_unit]",
  "pillow",
@@ -324,12 +324,6 @@ def _disasm_smem(inst: SMEM) -> str:
  if name in ('s_memrealtime', 's_memtime'): return f"{name} {_fmt_sdst(inst.sdata, dst_n, cdna)}"
  return f"{name} {_fmt_sdst(inst.sdata, dst_n, cdna)}, {sbase_str}, {off_s}" + _mods((inst.glc, " glc"), (getattr(inst, 'dlc', 0), " dlc"))

-R4_TH_LOAD = {1: 'TH_LOAD_NT', 2: 'TH_LOAD_HT', 3: 'TH_LOAD_LU', 4: 'TH_LOAD_RT_WB', 5: 'TH_LOAD_NT_WB'}
-R4_TH_STORE = {1: 'TH_STORE_NT', 2: 'TH_STORE_HT', 3: 'TH_STORE_ST', 4: 'TH_STORE_RT_WB', 5: 'TH_STORE_NT_WB'}
-R4_TH_ATOMIC = {1: 'TH_ATOMIC_RETURN', 2: 'TH_ATOMIC_NT', 3: 'TH_ATOMIC_RETURN_NT',
-                4: 'TH_ATOMIC_CASCADE_RT', 5: 'TH_ATOMIC_CASCADE_RETURN', 6: 'TH_ATOMIC_CASCADE_NT', 7: 'TH_ATOMIC_CASCADE_RETURN_NT'}
-R4_SCOPE = {1: 'SCOPE_SE', 2: 'SCOPE_DEV', 3: 'SCOPE_SYS'}
-
 def _disasm_flat(inst: FLAT) -> str:
  name, cdna, r4 = inst.op_name.lower(), _is_cdna(inst), _is_r4(inst)
  acc = getattr(inst, 'acc', 0)
@@ -337,10 +331,9 @@ def _disasm_flat(inst: FLAT) -> str:
  if r4: seg = 'flat' if (cls_name:=inst.__class__.__name__) == 'VFLAT' else ('global' if cls_name == 'VGLOBAL' else 'scratch')
  else: seg = ['flat', 'scratch', 'global'][inst.seg] if inst.seg < 3 else 'flat'
  instr = f"{seg}_{name.split('_', 1)[1] if '_' in name else name}"
-  # Global/scratch uses 13-bit signed offset (RDNA3/CDNA), 24-bit signed offset (RDNA4)
+  # Global/scratch uses 13-bit signed offset
  offset = inst.ioffset if r4 else inst.offset  # type: ignore[attr-defined]
-  if r4: off_val = offset if offset < (1 << 23) else offset - (1 << 24)  # sign extend 24-bit
-  elif seg != 'flat':
+  if seg != 'flat':
    if cdna:
      # CDNA: bit 12 is sign bit but not in offset field
      raw = int.from_bytes(inst.to_bytes(), 'little')
@@ -355,9 +348,7 @@ def _disasm_flat(inst: FLAT) -> str:
  w = regs.get('data', regs.get('d', 1)) if 'store' in name or 'atomic' in name else regs.get('d', 1)
  off_s = f" offset:{off_val}" if off_val else ""
  if cdna: mods = f"{off_s}{' sc0' if inst.sc0 else ''}{' nt' if inst.nt else ''}{' sc1' if getattr(inst, 'sc1', 0) else ''}"  # type: ignore[attr-defined]
-  elif r4:
-    th_names = R4_TH_ATOMIC if 'atomic' in name else (R4_TH_STORE if 'store' in name else R4_TH_LOAD)
-    mods = off_s + (f" th:{th_names[inst.th]}" if inst.th in th_names else "") + (f" scope:{R4_SCOPE[inst.scope]}" if inst.scope in R4_SCOPE else "")
+  elif r4: mods = f"{off_s}{' scope' if inst.scope else ''}{' th' if inst.th else ''}"  # type: ignore[attr-defined]
  else: mods = f"{off_s}{' glc' if inst.glc else ''}{' slc' if inst.slc else ''}{' dlc' if inst.dlc else ''}"
  if seg == 'flat': saddr_s = ""
  elif _unwrap(inst.saddr) in (0x7F, 124): saddr_s = ", off"
@@ -366,7 +357,7 @@ def _disasm_flat(inst: FLAT) -> str:
    saddr_s = f", {(SPECIAL_PAIRS_CDNA if cdna else SPECIAL_PAIRS)[_unwrap(inst.saddr)]}"
  elif t := _ttmp(inst.saddr, 2): saddr_s = f", {t}"
  else: saddr_s = f", {_sreg(inst.saddr, 2) if _unwrap(inst.saddr) < 106 else decode_src(_unwrap(inst.saddr), cdna)}"
-  if 'addtid' in name: return f"{instr} {reg_fn((inst.vsrc if r4 else inst.data) if 'store' in name else inst.vdst)}{saddr_s}{mods}"
+  if 'addtid' in name: return f"{instr} {reg_fn(inst.data if 'store' in name else inst.vdst)}{saddr_s}{mods}"
   # RDNA4: vaddr instead of addr, vsrc instead of data
  addr = inst.vaddr if r4 else inst.addr  # type: ignore[attr-defined]
  data = inst.vsrc if r4 else inst.data  # type: ignore[attr-defined]
@@ -381,7 +372,7 @@ def _disasm_flat(inst: FLAT) -> str:
  addr_s = "off" if not inst.sve and seg == 'scratch' else _vreg(addr, addr_w)
  data_s, vdst_s = reg_fn(data, w), reg_fn(inst.vdst, w // 2 if 'cmpswap' in name else w)
  if 'atomic' in name:
-    glc_or_sc0 = inst.sc0 if cdna else (inst.th & 1 if r4 else inst.glc)  # type: ignore[attr-defined]
+    glc_or_sc0 = inst.sc0 if cdna else inst.glc  # type: ignore[attr-defined]
    sfx = f"{saddr_s if seg != 'flat' else ''}{mods}"
    return f"{instr} {vdst_s}, {addr_s}, {data_s}{sfx}" if glc_or_sc0 else f"{instr} {addr_s}, {data_s}{sfx}"
  if 'store' in name: return f"{instr} {addr_s}, {data_s}{saddr_s}{mods}"
@@ -1,12 +1,22 @@
 """Shared test helpers for AMD tests."""
 import ctypes
+from dataclasses import dataclass
 from tinygrad.helpers import unwrap
 from tinygrad.runtime.autogen import llvm
 from tinygrad.runtime.support.elf import elf_loader

+@dataclass
+class KernelInfo:
+  code: bytes
+  src: str
+  global_size: tuple[int, int, int]
+  local_size: tuple[int, int, int]
+  buf_idxs: list[int]  # indices into shared buffer pool
+  buf_sizes: list[int]  # sizes for each buffer index
+
 ARCH_TO_TARGET:dict[str, list[str]] = {
  "rdna3":["gfx1100"],
-  "rdna4":["gfx1200", "gfx1201"],
+  "rdna4":["gfx1200"],
  "cdna":["gfx950", "gfx942"],
 }

@@ -104,34 +104,6 @@ class TestCmpClass(unittest.TestCase):
    st = run_program(instructions, n_lanes=1)
    self.assertEqual(st.vcc & 1, 0, "Signaling NaN should not match quiet mask")

-  def test_v_cmp_lg_f32_nan(self):
-    """v_cmp_lg_f32 is ordered not-equal (<>): NaN <> x should be False per IEEE 754."""
-    quiet_nan = 0x7fc00000
-    one_f32 = 0x3f800000  # 1.0f
-    instructions = [
-      s_mov_b32(s[0], quiet_nan),
-      v_mov_b32_e32(v[0], s[0]),
-      s_mov_b32(s[1], one_f32),
-      v_mov_b32_e32(v[1], s[1]),
-      v_cmp_lg_f32_e32(v[0], v[1]),
-    ]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc & 1, 0, "v_cmp_lg_f32(NaN, 1.0) should be 0")
-
-  def test_v_cmp_neq_f32_nan(self):
-    """v_cmp_neq_f32 is unordered not-equal (!=): NaN != x should be True per IEEE 754."""
-    quiet_nan = 0x7fc00000
-    one_f32 = 0x3f800000  # 1.0f
-    instructions = [
-      s_mov_b32(s[0], quiet_nan),
-      v_mov_b32_e32(v[0], s[0]),
-      s_mov_b32(s[1], one_f32),
-      v_mov_b32_e32(v[1], s[1]),
-      v_cmp_neq_f32_e32(v[0], v[1]),
-    ]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc & 1, 1, "v_cmp_neq_f32(NaN, 1.0) should be 1")
-
  def test_v_cmp_sets_vcc_bits(self):
    """V_CMP_EQ sets VCC bits based on per-lane comparison."""
    instructions = [
@@ -6,6 +6,7 @@ from tinygrad import Device

 from test.mockgpu.amd.emu import WaveState, _decode_at, WAVE_SIZE, VCC_LO, EXEC_LO, SCC
 from tinygrad.renderer.amd import decode_inst
+from test.amd.helpers import KernelInfo
 import tinygrad
 REMU_PATH = Path(tinygrad.__file__).parent.parent / "extra/remu/target/release/libremu.so"
 if not REMU_PATH.exists(): REMU_PATH = Path(tinygrad.__file__).parent.parent / "extra/remu/target/release/libremu.dylib"
@@ -21,15 +22,6 @@ def _vals_equal(a: int, b: int) -> bool:
  if a == b: return True
  return _is_f32_nan(a) and _is_f32_nan(b)

-@dataclass
-class KernelSnapshot:
-  code: bytes
-  src: str
-  global_size: tuple[int, int, int]
-  local_size: tuple[int, int, int]
-  buf_idxs: list[int]  # indices into shared buffer pool
-  buf_sizes: list[int]  # sizes for each buffer index
-
@dataclass
 class StateSnapshot:
  pc: int
@@ -293,7 +285,7 @@ def run_single_kernel(kernel: bytes, n_lanes: int, args_ptr: int, global_size: t

  return True, f"Completed {gx*gy*gz} workgroups", total_steps

-def compare_emulators_multi_kernel(kernels: list[KernelSnapshot], buf_pool: dict[int, int], max_steps: int = 1000,
+def compare_emulators_multi_kernel(kernels: list[KernelInfo], buf_pool: dict[int, int], max_steps: int = 1000,
                                    debug: bool = False, trace_len: int = 10, buf_data: dict[int, bytes] | None = None) -> tuple[bool, str]:
  """Run all kernels through both emulators with shared buffer pool."""
  if buf_data is None: buf_data = {}
@@ -357,7 +349,7 @@ def compare_emulators_with_memory(kernel: bytes, n_lanes: int, buf_sizes: list,
  ok, msg, _ = run_single_kernel(kernel, n_lanes, args_ptr, global_size, (n_lanes, 1, 1), max_steps, debug, trace_len)
  return ok, msg

-def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelSnapshot], dict[int, int], dict[int, bytes]]:
+def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelInfo], dict[int, int], dict[int, bytes]]:
  """Compile a tinygrad operation and extract all kernels with their buffer mappings."""
  from tinygrad import Tensor
  from tinygrad.runtime.support.elf import elf_loader
@@ -395,7 +387,7 @@ def get_kernels_from_tinygrad(op_fn) -> tuple[list[KernelSnapshot], dict[int, in
                buf_pool[buf_id] = b.nbytes
              buf_idxs.append(buf_id)
              buf_sizes.append(b.nbytes)
-            kernels.append(KernelSnapshot(
+            kernels.append(KernelInfo(
              code=bytes(sec.content),
              src=lowered.prg.p.src,
              global_size=tuple(lowered.prg.p.global_size),
@@ -1,12 +1,10 @@
 import unittest
-import functools
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.renderer import Estimates
+
 from tinygrad.runtime.autogen.amd.rdna3.ins import *
-from tinygrad.runtime.autogen.amd.rdna4.ins import s_barrier_wait, s_barrier_signal
 from tinygrad.renderer.amd.dsl import s, v
-from test.amd.helpers import TARGET_TO_ARCH

 def custom_add_one(A:UOp) -> UOp:
  A = A.flatten()
@@ -45,26 +43,9 @@ def custom_add_var(A:UOp, B:UOp) -> UOp:
  sink = UOp.sink(A.base, B.base, var, threads, arg=KernelInfo(f"custom_add_var_{A.size}"))
  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))

-def custom_wave_sync(A:UOp, arch:str) -> UOp:
-  # 4 waves across 1024 WG — enough to saturate a SIMD with many concurrent WGs
-  # s_sleep yields the SIMD so waves from different WGs interleave, causing barrier packet reordering
-  threads = UOp.special(128, "lidx0")
-  wg = UOp.special(1024, "gidx0")
-  insts = []
-  for _ in range(4):
-    insts.append(s_sleep(4))
-    insts += [s_barrier()] if arch == "rdna3" else [s_barrier_signal(), s_barrier_wait()]
-    insts += [s_nop(0)]*4
-  insts.append(s_endpgm())
-  sink = UOp.sink(A.base, threads, wg, arg=KernelInfo("custom_wave_sync"))
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.DEVICE, arg="AMD"), UOp(Ops.LINEAR, src=tuple([UOp(Ops.INS, arg=x) for x in insts]))))
-
@unittest.skipUnless(Device.DEFAULT == "AMD", "requires AMD device")
 class TestCustomKernel(unittest.TestCase):
-  def setUp(self): self.arch = TARGET_TO_ARCH[Device["AMD"].arch]
-
  def test_simple(self):
-    if self.arch != "rdna3": self.skipTest("only rdna3")
    a = Tensor.full((16, 16), 1.).contiguous().realize()
    a = Tensor.custom_kernel(a, fxn=custom_add_one)[0]
    ei = a.schedule()[-1].lower()
@@ -74,7 +55,6 @@ class TestCustomKernel(unittest.TestCase):
    self.assertTrue((a.numpy() == 2.).all())

  def test_variable(self):
-    if self.arch != "rdna3": self.skipTest("only rdna3")
    b = Tensor.full((16, 16), 1, dtype=dtypes.uint32).contiguous().realize()
    a = Tensor.zeros_like(b).contiguous().realize()
    a = Tensor.custom_kernel(a, b, fxn=custom_add_var)[0]
@@ -83,9 +63,5 @@ class TestCustomKernel(unittest.TestCase):
      ei.run({"var":i})
      self.assertTrue((a.numpy() == 1+i).all())

-  def test_wave_sync(self):
-    if self.arch not in {"rdna3", "rdna4"}: self.skipTest("only rdna3 or rdna4")
-    Tensor.empty(1).custom_kernel(fxn=functools.partial(custom_wave_sync, arch=self.arch))[0].realize()
-
 if __name__ == "__main__":
  unittest.main()
@@ -40,7 +40,7 @@ RDNA4_FILES = ['gfx12_asm_sop1.s', 'gfx12_asm_sop2.s', 'gfx12_asm_sopp.s', 'gfx1
  'gfx12_asm_vop1.s', 'gfx12_asm_vop2.s', 'gfx12_asm_vopc.s', 'gfx12_asm_vopcx.s', 'gfx12_asm_vop3.s', 'gfx12_asm_vop3c.s',
  'gfx12_asm_vop3cx.s', 'gfx12_asm_vop3p.s', 'gfx12_asm_vop3_from_vop1.s', 'gfx12_asm_vop3_from_vop2.s',
  'gfx12_asm_vop3p_features.s', 'gfx12_asm_vopd.s', 'gfx12_asm_vopd_features.s',
-  'gfx12_asm_ds.s', 'gfx12_asm_smem.s', 'gfx12_asm_vflat.s',
+  'gfx12_asm_ds.s', 'gfx12_asm_smem.s',
  'gfx12_asm_wmma_w32.s']

 def _parse_llvm_tests(text: str, pattern: str) -> list[tuple[str, bytes]]:
@@ -9,8 +9,8 @@ from tinygrad.renderer.amd import decode_inst
 from tinygrad.runtime.autogen.amd.rdna3.ins import SOPP
 from tinygrad.runtime.autogen.amd.rdna3.enum import SOPPOp
 from tinygrad.renderer.amd.sqtt import (decode, LAYOUT_HEADER, WAVESTART, WAVESTART_RDNA4, WAVEEND, INST, INST_RDNA4, VALUINST,
-                                     IMMEDIATE, IMMEDIATE_MASK, PACKET_TYPES_RDNA3, PACKET_TYPES_RDNA4, PACKET_TYPES_CDNA, CDNA_WAVESTART,
-                                     InstOp, InstOpRDNA4, print_packets, CDNA_WAVEEND, CDNA_INST)
+                                     IMMEDIATE, IMMEDIATE_MASK, PACKET_TYPES_RDNA3, PACKET_TYPES_RDNA4,
+                                     InstOp, InstOpRDNA4, print_packets)
 from test.amd.helpers import TARGET_TO_ARCH

 import tinygrad
@@ -21,7 +21,7 @@ OTHER_SIMD_OPS = {InstOp.OTHER_LDS_LOAD, InstOp.OTHER_LDS_STORE, InstOp.OTHER_LD
                  InstOp.OTHER_FLAT_STORE_128, InstOp.OTHER_GLOBAL_LOAD, InstOp.OTHER_GLOBAL_LOAD_VADDR,
                  InstOp.OTHER_GLOBAL_STORE_64, InstOp.OTHER_GLOBAL_STORE_96, InstOp.OTHER_GLOBAL_STORE_128,
                  InstOp.OTHER_GLOBAL_STORE_VADDR_128}
-OTHER_SIMD_OPS_RDNA4 = {InstOpRDNA4.OTHER_VMEM, InstOpRDNA4.OTHER_VMEM_5, InstOpRDNA4.OTHER_LDS_1, InstOpRDNA4.OTHER_LDS_2}
+OTHER_SIMD_OPS_RDNA4 = {InstOpRDNA4.OTHER_VMEM, InstOpRDNA4.UNK_60}

 # ═══════════════════════════════════════════════════════════════════════════════
 # ROCPROF DECODER
@@ -125,7 +125,7 @@ class SQTTExamplesTestBase(unittest.TestCase):
          self.assertIsInstance(packets[0], LAYOUT_HEADER, f"first packet should be LAYOUT_HEADER in {name}")

  def test_packet_types_valid(self):
-    all_classes = set(PACKET_TYPES_RDNA3.values()) | set(PACKET_TYPES_RDNA4.values()) | set(PACKET_TYPES_CDNA.values())
+    all_classes = set(PACKET_TYPES_RDNA3.values()) | set(PACKET_TYPES_RDNA4.values())
    for name, (events, *_) in self.examples.items():
      for i, event in enumerate(events):
        with self.subTest(example=name, event=i):
@@ -138,8 +138,8 @@ class SQTTExamplesTestBase(unittest.TestCase):
      if "empty" in name: continue
      with self.subTest(example=name):
        all_packets = [p for e in events for p in decode(e.blob)]
-        self.assertGreater(len([p for p in all_packets if isinstance(p, (WAVESTART, WAVESTART_RDNA4, CDNA_WAVESTART))]), 0, f"no WAVESTART in {name}")
-        self.assertGreater(len([p for p in all_packets if isinstance(p, (WAVEEND, CDNA_WAVEEND))]), 0, f"no WAVEEND in {name}")
+        self.assertGreater(len([p for p in all_packets if isinstance(p, (WAVESTART, WAVESTART_RDNA4))]), 0, f"no WAVESTART in {name}")
+        self.assertGreater(len([p for p in all_packets if isinstance(p, WAVEEND)]), 0, f"no WAVEEND in {name}")

  def test_time_monotonic(self):
    for name, (events, *_) in self.examples.items():
@@ -153,10 +153,7 @@ class SQTTExamplesTestBase(unittest.TestCase):
      if "gemm" not in name: continue
      with self.subTest(example=name):
        all_packets = [p for e in events for p in decode(e.blob)]
-        inst_packets = [p for p in all_packets if isinstance(p, (INST, INST_RDNA4, CDNA_INST))]
-        self.assertGreater(len(inst_packets), 0, f"no INST packets in {name}")
-        if isinstance(inst_packets[0], (INST, INST_RDNA4)):
-          self.assertGreater(len([p for p in inst_packets if p.op.name.startswith("JUMP")]), 0, f"no JUMP packets in {name}")
+        self.assertGreater(len([p for p in all_packets if isinstance(p, (INST, INST_RDNA4))]), 0, f"no INST packets in {name}")

  expected: dict[str, list[int]] = {}  # override in subclasses
  def test_packet_counts(self):
@@ -184,8 +181,8 @@ class SQTTExamplesTestBase(unittest.TestCase):
        for event in events:
          wave_starts: dict[tuple[int, int, int], int] = {}
          for p in decode(event.blob):
-            if isinstance(p, (WAVESTART, CDNA_WAVESTART, WAVESTART_RDNA4)): wave_starts[(p.wave, p.simd, p.cu)] = p._time
-            elif isinstance(p, (WAVEEND, CDNA_WAVEEND)) and (key := (p.wave, p.simd, p.cu)) in wave_starts:
+            if isinstance(p, (WAVESTART, WAVESTART_RDNA4)): wave_starts[(p.wave, p.simd, p.cu)] = p._time
+            elif isinstance(p, WAVEEND) and (key := (p.wave, p.simd, p.cu)) in wave_starts:
              our_waves.append((wave_starts[key], p._time))
        self.assertEqual(sorted(our_waves), sorted(roc_waves), f"wave times mismatch in {name}")

@@ -211,22 +208,17 @@ class SQTTExamplesTestBase(unittest.TestCase):
 class TestSQTTExamplesRDNA3(SQTTExamplesTestBase):
  target = "gfx1100"
  expected = {
-    "profile_empty_run_0": [1880, 1867, 1920, 1971, 1998, 1904],
-    "profile_empty_run_1": [1880, 1867, 1920, 1971, 1998, 1904],
-    "profile_gemm_run_0": [3275, 3278, 2426, 2475, 2511, 2431],
-    "profile_gemm_run_1": [3264, 3268, 2420, 2469, 2504, 2401],
-    "profile_ops_run_0": [1944, 4903, 1984, 2035, 2062, 1968],
-    "profile_ops_run_1": [1944, 4918, 1984, 2035, 2062, 1968],
-    "profile_plus_run_0": [1938, 1932, 1978, 2029, 2056, 1962],
-    "profile_plus_run_1": [1891, 1874, 1931, 1982, 2009, 1915],
+    "profile_empty_run_0": [1844, 1885, 1905, 1956, 1983, 1889],
+    "profile_empty_run_1": [1780, 1885, 1905, 1956, 1983, 1889],
+    "profile_gemm_run_0": [2656, 2025, 2045, 2096, 2123, 2029, 3183, 2019, 2039, 2090, 2117, 2023, 19119, 2013, 2033, 2084, 2111, 2017],
+    "profile_gemm_run_1": [2662, 2025, 2045, 2096, 2123, 2029, 3179, 2019, 2039, 2090, 2117, 2023, 19113, 2071, 2091, 2142, 2169, 2075],
+    "profile_plus_run_0": [1886, 2013, 2033, 2084, 2111, 2017],
+    "profile_plus_run_1": [1988, 2071, 2091, 2142, 2169, 2075],
  }

 class TestSQTTExamplesRDNA4(SQTTExamplesTestBase): target = "gfx1200"
-
-class TestSQTTExamplesCDNA(SQTTExamplesTestBase):
-  target = "gfx950"
-  def test_rocprof_wave_times_match(self): self.skipTest("TODO: requires timestamp patching")
-  def test_rocprof_inst_times_match(self): self.skipTest("TODO: requires timestamp patching")
+@unittest.skip("TODO: fix CDNA")
+class TestSQTTExamplesCDNA(SQTTExamplesTestBase): target = "gfx950"

 if __name__ == "__main__":
  unittest.main()
@@ -1,94 +0,0 @@
-import unittest, contextlib
-from tinygrad import Device, Tensor, Context, TinyJit
-from tinygrad.device import Compiled, ProfileProgramEvent, ProfileDeviceEvent
-from tinygrad.viz.serve import load_amd_counters
-
-@contextlib.contextmanager
-def save_sqtt():
-  yield (ret:=[])
-  Device[Device.DEFAULT].synchronize()
-  Device[Device.DEFAULT]._at_profile_finalize()
-  load_amd_counters(ret, Compiled.profile_events)
-  ret[:] = [r for r in ret if r["name"].startswith("Exec")]
-
-@unittest.skipUnless(Device.DEFAULT == "AMD", "only runs on AMD")
-class TestSQTTProfiler(unittest.TestCase):
-  # TODO: can we enable SQTT profiling in context?
-  @classmethod
-  def setUpClass(cls):
-    if not Device[Device.DEFAULT].sqtt_enabled: raise unittest.SkipTest("device must be in SQTT profiling mode")
-
-  def setUp(self):
-    Device[Device.DEFAULT].synchronize()
-    Compiled.profile_events[:] = [e for e in Compiled.profile_events if isinstance(e, (ProfileProgramEvent, ProfileDeviceEvent))]
-
-  def test_simple(self):
-    t = Tensor.empty(1) + 1
-    with save_sqtt() as sqtt:
-      ei = t.schedule()[0].lower()
-      ei.run()
-    self.assertEqual(len(sqtt), 1)
-    self.assertEqual(sqtt[0]["name"], f"Exec {ei.prg.p.function_name}")
-
-  def test_multiple_runs(self):
-    t = Tensor.empty(1) + 1
-    with save_sqtt() as sqtt:
-      ei = t.schedule()[0].lower()
-      for _ in range(N:=3):
-        ei.run()
-    self.assertEqual(len(sqtt), N)
-    for i in range(1, N):
-      self.assertEqual(sqtt[i]["name"], f"Exec {ei.prg.p.function_name} n{i+1}")
-
-  def test_multiple_kernels(self):
-    t = ((Tensor.empty(1) + 1).contiguous() + 2)
-    sched = t.schedule()
-    with save_sqtt() as sqtt:
-      for si in sched: si.lower().run()
-    self.assertEqual(len(sqtt), len(sched))
-    for i,k in enumerate(sched):
-      self.assertEqual(sqtt[i]["name"], f"Exec {k.lower().prg.p.function_name}")
-
-  def test_multiple_kernels_lower(self):
-    t = ((Tensor.empty(1) + 1).contiguous() + 2)
-    sched = t.schedule()
-    with save_sqtt() as sqtt:
-      prgs = [si.lower() for si in sched]
-      for p in prgs: p.run()
-    self.assertEqual(len(sqtt), len(sched))
-    for i,ei in enumerate(prgs):
-      self.assertEqual(sqtt[i]["name"], f"Exec {ei.prg.p.function_name}")
-
-  def test_jit(self):
-    @TinyJit
-    def f(a): return a + 1
-    t = Tensor.empty(1)
-    with save_sqtt() as sqtt:
-      for _ in range(N:=5):
-        f(t).realize()
-    self.assertEqual(len(sqtt), N)
-    kernel_name = sqtt[0]["name"]
-    for i,s in enumerate(sqtt[1:], start=1): self.assertEqual(s["name"], f"{kernel_name} n{i+1}")
-
-  # TODO: can we trace SQTT for graphed kernels?
-  def test_jit_graph(self, kernel_count=3*2):
-    @TinyJit
-    def f(a): return ((a + 1).contiguous() + 2).contiguous().sum()
-    t = Tensor.empty(32)
-    with save_sqtt() as sqtt:
-      for _ in range(5):
-        f(t).realize()
-    names = [s["name"] for s in sqtt]
-    k0, k1, k2 = names[:3]
-    for i in range(3, len(sqtt), 3):
-      n = (i // 3)+1
-      self.assertEqual(names[i], f"{k0} n{n}")
-      self.assertEqual(names[i+1], f"{k1} n{n}")
-      self.assertEqual(names[i+2], f"{k2} n{n}")
-    self.assertEqual(len(sqtt), kernel_count)
-
-  @Context(JIT=2)
-  def test_jit_multiple_kernels(self): self.test_jit_graph(kernel_count=3*5)
-
-if __name__ == "__main__":
-  unittest.main()
@@ -0,0 +1,183 @@
+"""Tests comparing sqtt.py PACKET_TYPES_RDNA3/RDNA4 against AMD's rocprof-trace-decoder binary."""
+import unittest, struct, ctypes, pickle
+from pathlib import Path
+
+ROCPROF_LIB = Path("/usr/lib/librocprof-trace-decoder.so")
+import tinygrad
+EXAMPLES_DIR = Path(tinygrad.__file__).parent.parent / "extra/sqtt/examples"
+
+# CDNA pkt_fmt -> size in bytes (extracted from rocprof hash table)
+CDNA_PKT_SIZES = {0: 2, 1: 8, 2: 8, 3: 4, 4: 2, 5: 6, 6: 2, 7: 2, 8: 2, 9: 2, 10: 2, 11: 8, 12: 6, 13: 4, 14: 8, 15: 6}
+
+def _find_segment(perms: str):
+  """Find a segment of the loaded library with given permissions (e.g. 'rw-p', 'r--p')."""
+  with open('/proc/self/maps', 'r') as f:
+    for line in f:
+      if 'librocprof-trace-decoder.so' in line and f' {perms} ' in line:
+        parts = line.split()
+        return int(parts[0].split('-')[0], 16), int(parts[2], 16)
+  return None, None
+
+def _read_array(file_offset: int, count: int):
+  """Read an array of uint8 at file_offset from the loaded library."""
+  base, seg_offset = _find_segment('rw-p')
+  if base is None: return None
+  return list((ctypes.c_uint8 * count).from_address(base + (file_offset - seg_offset)))
+
+def _load_lib():
+  if not ROCPROF_LIB.exists(): return False
+  ctypes.CDLL(str(ROCPROF_LIB))
+  return True
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# RDNA EXTRACTION (nibble-based format)
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def extract_bit_tables():
+  """Extract bit budget tables. Returns (layout2, layout3, layout4) or None."""
+  if not _load_lib(): return None
+  return _read_array(0x2d220, 32), _read_array(0x2d280, 32), _read_array(0x2d2c0, 32)
+
+def extract_delta_fields():
+  """Extract delta bitfield tables. Returns (layout2, layout3, layout4) dicts mapping type_id -> (lo, hi)."""
+  if not _load_lib(): return None
+  ro_base, ro_offset = _find_segment('r--p')
+  if ro_base is None: return None
+
+  def read_table(file_offset, num_entries):
+    addr = ro_base + (file_offset - ro_offset)
+    data = bytes((ctypes.c_uint8 * (num_entries * 12)).from_address(addr))
+    return {type_id: (lo, hi) for j in range(0, len(data), 12)
+            for type_id, lo, hi in [struct.unpack('<III', data[j:j+12])] if type_id < 32}
+
+  return read_table(0x26800, 24), read_table(0x26dc0, 25), read_table(0x27300, 27)
+
+def extract_packet_encodings():
+  """Extract packet encodings. Returns (L2, L3, L4) dicts mapping type_id -> (mask, value)."""
+  if not _load_lib(): return None
+  rw_base, rw_offset = _find_segment('rw-p')
+  if rw_base is None: return None
+
+  # Read base encodings from registration vector at 0x2d340
+  vec_start = ctypes.c_void_p.from_address(rw_base + (0x2d340 - rw_offset)).value
+  vec_end = ctypes.c_void_p.from_address(rw_base + (0x2d348 - rw_offset)).value
+  base = {}
+  if vec_start and vec_end:
+    for i in range((vec_end - vec_start) // 32):
+      addr = vec_start + i * 32
+      type_id = ctypes.c_uint8.from_address(addr).value
+      pat_start = ctypes.c_void_p.from_address(addr + 8).value
+      pat_end = ctypes.c_void_p.from_address(addr + 16).value
+      if pat_start and pat_end and 0 < (n := pat_end - pat_start) <= 8:
+        pat = list((ctypes.c_uint8 * n).from_address(pat_start))
+        base[type_id] = (sum(1 << j for j in range(n)), sum(b << j for j, b in enumerate(pat)))
+
+  return {**base, 17: (0x7f, 0x51), 25: (0x7f, 0x31)}, base, {**base}  # L2 has overrides
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# CDNA EXTRACTION (16-bit header format)
+# ═══════════════════════════════════════════════════════════════════════════════
+
+def extract_cdna_packet_sizes():
+  """Extract CDNA pkt_fmt -> size mapping by running rocprof decoder to populate its hash table."""
+  if not _load_lib(): return None
+  from test.amd.test_sqtt_examples import run_rocprof_decoder
+
+  if not (pkl_path := next((EXAMPLES_DIR / "gfx950").glob("*.pkl"), None)): return None
+  with open(pkl_path, "rb") as f: data = pickle.load(f)
+  sqtt_events = [e for e in data if type(e).__name__ == "ProfileSQTTEvent"]
+  prg = next((e for e in data if type(e).__name__ == "ProfileProgramEvent"), None)
+  if not sqtt_events or not prg: return None
+
+  # Run decoder to trigger hash table initialization
+  run_rocprof_decoder([e.blob for e in sqtt_events], prg.lib, prg.base, "gfx950")
+
+  # Extract hash table: head at 0x2d4f0, nodes are 16 bytes (next[8], key[4], value[4])
+  rw_base, rw_offset = _find_segment('rw-p')
+  if not (head := ctypes.c_void_p.from_address(rw_base + (0x2d4f0 - rw_offset)).value if rw_base else None): return None
+
+  pkt_sizes: dict[int, int] = {}
+  node, seen = head, set()
+  while node and node not in seen and len(pkt_sizes) < 20:
+    seen.add(node)
+    key, val = ctypes.c_uint32.from_address(node + 8).value, ctypes.c_uint32.from_address(node + 12).value
+    if key < 16 and val in (0x10, 0x20, 0x30, 0x40): pkt_sizes[key] = {0x10: 2, 0x20: 4, 0x30: 6, 0x40: 8}[val]
+    node = ctypes.c_void_p.from_address(node).value  # type: ignore[assignment]
+  return pkt_sizes if len(pkt_sizes) == 16 else None
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# TESTS
+# ═══════════════════════════════════════════════════════════════════════════════
+
+class TestSQTTMatchesBinary(unittest.TestCase):
+  def test_bit_counts_match_layout3(self): self._test_bit_counts(3)
+  def test_bit_counts_match_layout4(self): self._test_bit_counts(4)
+  def test_encodings_match_layout3(self): self._test_encodings(3)
+  def test_encodings_match_layout4(self): self._test_encodings(4)
+  def test_delta_fields_match_layout3(self): self._test_delta_fields(3)
+  def test_delta_fields_match_layout4(self): self._test_delta_fields(4)
+
+  def test_cdna_packet_sizes(self):
+    """Extract and verify CDNA pkt_fmt -> size mapping from rocprof's hash table."""
+    if not (EXAMPLES_DIR / "gfx950").exists(): self.skipTest("no CDNA examples")
+    if not (pkt_sizes := extract_cdna_packet_sizes()): self.skipTest("rocprof-trace-decoder not installed")
+    for pkt_fmt, size in CDNA_PKT_SIZES.items():
+      with self.subTest(pkt_fmt=pkt_fmt): self.assertEqual(pkt_sizes.get(pkt_fmt), size)
+
+  def test_cdna_packet_definitions(self):
+    from tinygrad.renderer.amd.sqtt import PACKET_TYPES_CDNA
+    for pkt_fmt, pkt_cls in PACKET_TYPES_CDNA.items():
+      with self.subTest(packet=pkt_cls.__name__):
+        self.assertEqual(pkt_cls.encoding.default, pkt_fmt)
+        self.assertEqual(CDNA_PKT_SIZES[pkt_fmt] * 2, pkt_cls._size_nibbles)  # type: ignore[attr-defined]
+
+  def _test_bit_counts(self, layout: int):
+    if not (tables := extract_bit_tables()): self.skipTest("rocprof-trace-decoder not installed")
+    from tinygrad.renderer.amd.sqtt import PACKET_TYPES_RDNA3, PACKET_TYPES_RDNA4
+    for type_id, pkt_cls in {3: PACKET_TYPES_RDNA3, 4: PACKET_TYPES_RDNA4}[layout].items():
+      with self.subTest(packet=pkt_cls.__name__):
+        self.assertEqual(pkt_cls._size_nibbles * 4, tables[layout - 2][type_id])  # type: ignore[attr-defined]
+
+  def _test_encodings(self, layout: int):
+    if not (encodings := extract_packet_encodings()): self.skipTest("rocprof-trace-decoder not installed")
+    from tinygrad.renderer.amd.sqtt import PACKET_TYPES_RDNA3, PACKET_TYPES_RDNA4
+    for type_id, pkt_cls in {3: PACKET_TYPES_RDNA3, 4: PACKET_TYPES_RDNA4}[layout].items():
+      with self.subTest(packet=pkt_cls.__name__):
+        self.assertEqual((pkt_cls.encoding.mask, pkt_cls.encoding.default), encodings[layout - 2][type_id])
+
+  def _test_delta_fields(self, layout: int):
+    if not (deltas := extract_delta_fields()): self.skipTest("rocprof-trace-decoder not installed")
+    from tinygrad.renderer.amd.sqtt import PACKET_TYPES_RDNA3, PACKET_TYPES_RDNA4
+    for type_id, pkt_cls in {3: PACKET_TYPES_RDNA3, 4: PACKET_TYPES_RDNA4}[layout].items():
+      if type_id not in deltas[layout - 2]: continue
+      delta = getattr(pkt_cls, 'delta', None)
+      actual = (0, 0) if delta is None else (delta.lo, delta.hi + 1)
+      with self.subTest(packet=pkt_cls.__name__): self.assertEqual(actual, deltas[layout - 2][type_id])
+
+if __name__ == "__main__":
+  tables = extract_bit_tables()
+  encodings = extract_packet_encodings()
+  deltas = extract_delta_fields()
+
+  TYPE_NAMES = {1: 'VALUINST', 2: 'VMEMEXEC', 3: 'ALUEXEC', 4: 'IMMEDIATE', 5: 'IMMEDIATE_MASK', 6: 'WAVERDY',
+    7: 'TS_DELTA_S8_W3', 8: 'WAVEEND', 9: 'WAVESTART', 10: 'TS_DELTA_S5_W2', 11: 'WAVEALLOC', 12: 'TS_DELTA_S5_W3',
+    13: 'PERF', 14: 'UTILCTR', 15: 'TS_DELTA_SHORT', 16: 'NOP', 17: 'TS_WAVE_STATE', 18: 'EVENT', 19: 'EVENT_BIG',
+    20: 'REG', 21: 'SNAPSHOT', 22: 'TS_DELTA_OR_MARK', 23: 'LAYOUT_HEADER', 24: 'INST', 25: 'UNK_25'}
+
+  print("L2:", tables[0], "\nL3:", tables[1], "\nL4:", tables[2])
+  if encodings and tables:
+    print(f"\n{'TypeID':>6} {'Name':>18} {'L2 enc':>12} {'L3 enc':>12} {'L4 enc':>12}"
+          f" {'L2':>4} {'L3':>4} {'L4':>4} {'L2 delta':>12} {'L3 delta':>12} {'L4 delta':>12}")
+    print("-" * 140)
+    for type_id in sorted(set(encodings[0]) | set(encodings[1]) | set(encodings[2])):
+      name = TYPE_NAMES.get(type_id, f'UNK_{type_id}')
+      bits = [tables[i][type_id] if type_id < len(tables[i]) else 0 for i in range(3)]
+      enc_strs = [f"0x{encodings[i][type_id][0]:02x}/0x{encodings[i][type_id][1]:02x}" if type_id in encodings[i] else "-" for i in range(3)]
+      delta_strs = [f"[{d[1]-1}:{d[0]}]" if (d := deltas[i].get(type_id, (0, 0)))[1] > d[0] else "-" for i in range(3)]
+      print(f"{type_id:6d} {name:>18} {enc_strs[0]:>12} {enc_strs[1]:>12} {enc_strs[2]:>12}"
+            f" {bits[0]:4d} {bits[1]:4d} {bits[2]:4d} {delta_strs[0]:>12} {delta_strs[1]:>12} {delta_strs[2]:>12}")
+
+  cdna = extract_cdna_packet_sizes()
+  if cdna: print(f"\nCDNA packet sizes: {cdna}")
+
+  unittest.main()
@@ -2,10 +2,9 @@
 import unittest, pickle
 from typing import Iterator
 from pathlib import Path
-from tinygrad.helpers import DEBUG, getenv, temp
+from tinygrad.helpers import DEBUG
 from tinygrad.renderer.amd.sqtt import print_packets, map_insts
 from tinygrad.runtime.autogen.amd.rdna3.ins import s_endpgm
-from tinygrad.viz.serve import sqtt_timeline
 from test.amd.disasm import disasm

 import tinygrad
@@ -15,7 +14,7 @@ def rocprof_inst_traces_match(sqtt, prg, target):
  from tinygrad.viz.serve import amd_decode
  from extra.sqtt.roc import decode as roc_decode, InstExec
  addr_table = amd_decode(prg.lib, target)
-  disasm_map = {addr+prg.base:inst for addr,inst in addr_table.items()}
+  disasm_map = {addr+prg.base:(disasm(inst), inst.size()) for addr,inst in addr_table.items()}
  rctx = roc_decode([sqtt], {prg.tag:disasm_map})
  rwaves = rctx.inst_execs.get((sqtt.kern, sqtt.exec_tag), [])
  rwaves_iter:dict[int, list[Iterator[InstExec]]] = {} # wave unit (0-15) -> list of inst trace iterators for all executions on that unit
@@ -25,13 +24,13 @@ def rocprof_inst_traces_match(sqtt, prg, target):

  passed_insts = 0
  for pkt, info in map_insts(sqtt.blob, prg.lib, target):
-    if DEBUG >= 2: print_packets([(pkt, info)])
+    if DEBUG >= 2: print_packets([pkt])
    if info is None: continue
    if DEBUG >= 2: print(f"{' '*29}{disasm(info.inst)}")
    rocprof_inst = next(rwaves_iter[info.wave][0])
    ref_pc = rocprof_inst.pc-prg.base
    # always check pc matches
-    assert ref_pc == info.pc, f"pc mismatch {ref_pc}:{disasm_map[rocprof_inst.pc]} != {info.pc}:{disasm(info.inst)}"
+    assert ref_pc == info.pc, f"pc mismatch {ref_pc}:{disasm_map[rocprof_inst.pc][0]} != {info.pc}:{disasm(info.inst)}"
    # special handling for s_endpgm, it marks the wave completion.
    if info.inst == s_endpgm():
      completed_wave = list(rwaves_iter[info.wave].pop(0))
@@ -54,7 +53,7 @@ class TestSQTTMapBase(unittest.TestCase):
  def setUpClass(cls):
    if cls is TestSQTTMapBase: raise unittest.SkipTest("base class")
    cls.examples = {}
-    for pkl_path in ([Path(temp("profile.pkl", append_user=True))] if getenv("LOAD_PROFILE") else sorted((EXAMPLES_DIR/cls.target).glob("*.pkl"))):
+    for pkl_path in sorted((EXAMPLES_DIR/cls.target).glob("*.pkl")):
      with open(pkl_path, "rb") as f:
        data = pickle.load(f)
      sqtt_events = [e for e in data if type(e).__name__ == "ProfileSQTTEvent"]
@@ -64,8 +63,6 @@ class TestSQTTMapBase(unittest.TestCase):

  def test_rocprof_inst_traces_match(self):
    for name, (events, kern_events, target) in self.examples.items():
-      if "sync" in name and self.target.startswith("gfx12"):
-        self.skipTest("our timestamps are off by a few cycles because rocprof patches timestamps for rdna4 barriers")
      for event in events:
        if not event.itrace: continue
        if event.kern not in kern_events: continue
@@ -73,41 +70,9 @@ class TestSQTTMapBase(unittest.TestCase):
          passed_insts, n_waves, n_units = rocprof_inst_traces_match(event, kern_events[event.kern], target)
          if n_waves: print(f"{name}: passed for {passed_insts} instructions across {n_waves} waves scheduled on {n_units} wave units")

-  def test_sqtt_timeline(self):
-    for name, (events, kern_events, target) in self.examples.items():
-      for event in events:
-        if (p:=kern_events.get(event.kern)) is None: continue
-        with self.subTest(example=name, kern=event.kern):
-          if not (timeline:=sqtt_timeline(event.blob, p.lib, target)): continue
-          frequency = [e.key for e in timeline if type(e).__name__ == "ProfilePointEvent" and e.name == "freq_hz"]
-          mean = sum(frequency) / len(frequency)
-          variance = sum((v - mean) ** 2 for v in frequency) / len(frequency)
-          self.assertGreater(mean, 0)
-          if DEBUG >= 2: print(f"{name:20s} SE:{event.se} {mean/1e9:.2f} GHz mean, {variance/1e18:.2f} GHz^2 variance")
-          events = [e for e in timeline if type(e).__name__ == "ProfileRangeEvent"]
-          insts, execs = 0, 0
-          for e in events:
-            if "EXEC" in e.device:
-              if "ALT" not in e.name.display_name: execs += 1
-            elif "WAVE" in e.device:
-              # sopk/immediates don't get ALU/MEM EXEC
-              if e.name.display_name not in {"IMMEDIATE", "IMMEDIATE_MASK", "JUMP", "JUMP_NO", "MESSAGE", "BARRIER", "BARRIER_SIGNAL"}: insts += 1
-            else: raise Exception(f"timeline row must be INST or EXEC, got {e.device}")
-          self.assertEqual(execs, insts)
-
-  def test_wave_sync(self):
-    for name, (events, kern_events, target) in self.examples.items():
-      for event in events:
-        wave_barriers = {}
-        for e in sqtt_timeline(event.blob, kern_events[event.kern].lib, target):
-          if type(e).__name__ == "ProfileRangeEvent" and e.name.display_name == "BARRIER": wave_barriers.setdefault(e.device, []).append(e)
-        if not wave_barriers: continue
-        for row, events in wave_barriers.items():
-          for e in events:
-            assert e.en-e.st > 1, f"all barriers must have a duration greater than 1, got {e}"
-
 class TestSQTTMapRDNA3(TestSQTTMapBase): target = "gfx1100"

+@unittest.skip("this doesn't work")
 class TestSQTTMapRDNA4(TestSQTTMapBase): target = "gfx1200"

 if __name__ == "__main__":
@@ -2,7 +2,7 @@ import unittest
 from tinygrad import Tensor, Device, dtypes, Context
 from tinygrad.device import is_dtype_supported
 from tinygrad.helpers import getenv
-from extra.gemm.cdna_asm_gemm import asm_gemm
+from extra.gemm.asm.cdna.gemm import asm_gemm
 from test.helpers import needs_second_gpu

 # On non CDNA4 it will only validate the Tensor.custom_kernel integration
@@ -47,18 +47,6 @@ def verify_asm_gemm(batch:int, M:int, N:int, K:int, dtype=dtypes.float16, gpus:i
 def verify_asm_gemm_k_sharded(M:int, N:int, K:int, dtype=dtypes.float16, gpus:int=8) -> None:
  run_asm_gemm((M, K), (K, N), dtype=dtype, a_shard=1, b_shard=0, gpus=gpus)

-def verify_asm_gemm_n_sharded(batch:int, M:int, N:int, K:int, dtype=dtypes.float16, gpus:int=2) -> None:
-  run_asm_gemm((batch, M, K), (K, N), dtype=dtype, a_shard=None, b_shard=1, gpus=gpus)
-
-def verify_asm_gemm_m_sharded(M:int, N:int, K:int, dtype=dtypes.float16, gpus:int=2) -> None:
-  run_asm_gemm((M, K), (K, N), dtype=dtype, a_shard=0, b_shard=None, gpus=gpus)
-
-def verify_asm_gemm_n_sharded_2d(M:int, N:int, K:int, dtype=dtypes.float16, gpus:int=2) -> None:
-  run_asm_gemm((M, K), (K, N), dtype=dtype, a_shard=None, b_shard=1, gpus=gpus)
-
-def verify_asm_gemm_k_sharded_3d(batch:int, M:int, N:int, K:int, dtype=dtypes.float16, gpus:int=2) -> None:
-  run_asm_gemm((batch, M, K), (K, N), dtype=dtype, a_shard=2, b_shard=0, gpus=gpus)
-
 # 128x smaller than usual
 # uses the UOp GEMM, runs on non CDNA4 and CI
@unittest.skipUnless(is_dtype_supported(dtypes.half), "need half")
@@ -72,14 +60,6 @@ class TestGemm(unittest.TestCase):
  def test_gemm_multi(self): verify_asm_gemm(2, 64, 32, 32, gpus=2)
  @needs_second_gpu
  def test_gemm_k_sharded(self): verify_asm_gemm_k_sharded(64, 64, 2*64, gpus=2)
-  @needs_second_gpu
-  def test_gemm_m_sharded(self): verify_asm_gemm_m_sharded(2*64, 64, 32, gpus=2)
-  @needs_second_gpu
-  def test_gemm_n_sharded(self): verify_asm_gemm_n_sharded(1, 64, 64, 32, gpus=2)
-  @needs_second_gpu
-  def test_gemm_n_sharded_2d(self): verify_asm_gemm_n_sharded_2d(64, 2*64, 32, gpus=2)
-  @needs_second_gpu
-  def test_gemm_k_sharded_3d(self): verify_asm_gemm_k_sharded_3d(1, 64, 32, 2*64, gpus=2)

 # uses the Asm GEMM on CDNA4 only for speed reasons
 class TestGemmLarge(unittest.TestCase):
@@ -87,7 +67,6 @@ class TestGemmLarge(unittest.TestCase):
    if not is_cdna4():
      self.skipTest("very slow on non mi350x")

-  def test_tiny(self): verify_asm_gemm(1, 256, 256, 64)
  def test_simple(self): verify_asm_gemm(1, N:=getenv("N", 4096), N, N, dtype=dtypes.half)
  def test_gemm(self): verify_asm_gemm(1, 8192, 4096, 14336)
  def test_gemm_batched(self): verify_asm_gemm(2, 8192, 4096, 4096)
@@ -121,20 +100,6 @@ class TestGemmLarge(unittest.TestCase):
      verify_asm_gemm(3, 256, 256, 256)
  def test_gemm_previously_unsupported(self): verify_asm_gemm(8, 1024, 1024, 4096, gpus=8)

-  # M-sharded 2D
-  def test_m_sharded_1(self): verify_asm_gemm_m_sharded(8*8192, 4096, 4096, dtype=dtypes.bfloat16, gpus=8)
-  def test_m_sharded_2(self): verify_asm_gemm_m_sharded(8*4096, 14336, 4096, dtype=dtypes.bfloat16, gpus=8)
-
-  # N-sharded 2D
-  def test_n_sharded_2d_1(self): verify_asm_gemm_n_sharded_2d(8192, 8*4096, 4096, dtype=dtypes.bfloat16, gpus=8)
-  def test_n_sharded_2d_2(self): verify_asm_gemm_n_sharded_2d(4096, 8*14336, 4096, dtype=dtypes.bfloat16, gpus=8)
-
-  # tensor parallel shapes (Llama 8B, MP=8)
-  def test_tp_n_sharded_wq(self): verify_asm_gemm_n_sharded(1, 8192, 4096, 4096, dtype=dtypes.bfloat16, gpus=8)
-  def test_tp_n_sharded_w1(self): verify_asm_gemm_n_sharded(1, 8192, 14336, 4096, dtype=dtypes.bfloat16, gpus=8)
-  def test_tp_k_sharded_wo(self): verify_asm_gemm_k_sharded_3d(1, 8192, 4096, 4096, dtype=dtypes.bfloat16, gpus=8)
-  def test_tp_k_sharded_w2(self): verify_asm_gemm_k_sharded_3d(1, 8192, 4096, 14336, dtype=dtypes.bfloat16, gpus=8)
-
  # more shapes: vary M, N, K independently
  def test_shape_small_square(self): verify_asm_gemm(1, 256, 256, 256)
  def test_shape_small_rect_m(self): verify_asm_gemm(1, 512, 256, 256)
@@ -157,7 +122,7 @@ class TestGemmLarge(unittest.TestCase):

 class TestMagicGu(unittest.TestCase):
  def test_magicgu_matches_old(self):
-    from extra.gemm.cdna_asm_gemm import _magicgu_mulhi, TILE_M, TILE_N, TILE_K
+    from extra.gemm.asm.cdna.asm import _magicgu_mulhi, TILE_M, TILE_N, TILE_K
    old_iters_args = {64: (67108864, 0), 128: (33554432, 0), 224: (613566757, 2147483656)}
    old_gemm_shapes = [
      (8192, 4096, 4096), (8192, 14336, 4096), (8192, 4096, 14336),
@@ -10,9 +10,7 @@ def _check_ast_count(desired_count:int, t:Tensor):
  # NOTE: this has side effect because everything can be scheduled only once
  schedule = t.schedule()
  asts = [s for s in schedule if s.ast.op is Ops.SINK]
-  len(asts)
-  # NOT SUPPORTED ANYMORE
-  #assert len(asts) == desired_count, f"{len(asts)} != {desired_count}"
+  assert len(asts) == desired_count, f"{len(asts)} != {desired_count}"

 class TestMovedConstFolding(unittest.TestCase):
  def test_add_shrunk_zero(self):
@@ -27,11 +25,6 @@ class TestMovedConstFolding(unittest.TestCase):
  def test_add_padded_one(self):
    _check_ast_count(1, Tensor([1.0, 2, 3, 4]) * Tensor.ones(2).pad(((1, 1),)))

-  def test_copy_padded_const(self):
-    schedule = Tensor.ones(4, device="CPU:0").pad(((1, 1),)).to("CPU:1").schedule()
-    assert not any(si.ast.op is Ops.COPY for si in schedule), "const copy should be folded"
-    np.testing.assert_equal(Tensor.ones(4, device="CPU:0").pad(((1, 1),)).to("CPU:1").numpy(), [0, 1, 1, 1, 1, 0])
-
  def test_cast_padded(self):
    # NOTE: it's always 1 kernel when calling .numpy, limitation of _check_ast_count
    if is_dtype_supported(dtypes.int16):
@@ -265,6 +265,8 @@ class TestCustomKernel(unittest.TestCase):
    Expected schedule order: [A2, B2, E, custom_addmul, final_sum]
    The custom_addmul kernel should be at index 3.
    """
+    from tinygrad.engine.schedule import create_schedule
+    from tinygrad.schedule.rangeify import get_rangeify_map

    A, B = Tensor.empty(4, 4), Tensor.empty(4, 4)
    A2 = (A + 1).contiguous()                      # kernel 0: depends on A
@@ -273,7 +275,11 @@ class TestCustomKernel(unittest.TestCase):
    C, D, _, _ = Tensor.custom_kernel(C, D, A2, B2, fxn=custom_elementwise_addmul_kernel)  # depends on A2 AND B2
    E = (A2 * 3).contiguous()                      # kernel 2: depends only on A2
    result = (C + D + E).sum()                     # kernel 3: custom_addmul, then kernel 4: sum
-    schedule = result.schedule()
+
+    big_sink = result.uop.sink()
+    tensor_map = get_rangeify_map(big_sink)
+    sched_sink = big_sink.substitute(tensor_map)
+    schedule, _ = create_schedule(sched_sink)

    # Find the custom_addmul kernel position
    custom_idx = next((i for i, item in enumerate(schedule)
@@ -10,7 +10,7 @@ from tinygrad.renderer.nir import NIRRenderer
 from tinygrad import Context, Device, Tensor, dtypes
 from hypothesis import given, settings, strategies as strat
 from test.helpers import rand_for_dtype
-from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX, FP8E4M3FNUZ_MAX, FP8E5M2FNUZ_MAX
+from test.unit.test_dtype_spec import _assert_eq, core_dtypes, dtype_ints, dtype_floats, FP8E4M3_MAX, FP8E5M2_MAX
 import pytest
 pytestmark = pytest.mark.filterwarnings("ignore")

@@ -101,14 +101,14 @@ class TestDType(unittest.TestCase):
  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)), "skip for now")
  def test_uint_overflow(self):
    if not dtypes.is_unsigned(self.DTYPE): raise unittest.SkipTest("only for unsigned")
-    v = self.DTYPE.max
+    v = dtypes.max(self.DTYPE)
    _test_to_np(Tensor(v, dtype=self.DTYPE)+2, _to_np_dtype(self.DTYPE), np.array(v, dtype=_to_np_dtype(self.DTYPE))+2)
    _test_to_np(Tensor(v, dtype=self.DTYPE)*2, _to_np_dtype(self.DTYPE), np.array(v, dtype=_to_np_dtype(self.DTYPE))*2)

  def test_dtypes_DTYPES_DICT(self):
    self.assertIn("float", DTYPES_DICT)
    self.assertIn("float32", DTYPES_DICT)
-    self.assertEqual(len(DTYPES_DICT), 28)
+    self.assertEqual(len(DTYPES_DICT), 26)
    self.assertTrue(all(isinstance(value, DType) for value in DTYPES_DICT.values()))
    self.assertTrue(all(issubclass(_to_np_dtype(value), np.generic) for value in DTYPES_DICT.values() if _to_np_dtype(value) is not None))

@@ -143,8 +143,6 @@ def _test_ops(a_dtype:DType, b_dtype:DType, target_dtype=None):
 class TestFp8s(unittest.TestCase):
  def test_fp8e4m3_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e4m3).dtype == dtypes.fp8e4m3
  def test_fp8e5m2_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e5m2).dtype == dtypes.fp8e5m2
-  def test_fp8e4m3fnuz_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e4m3fnuz).dtype == dtypes.fp8e4m3fnuz
-  def test_fp8e5m2fnuz_creation(self): assert Tensor([-1, 1, 2], dtype=dtypes.fp8e5m2fnuz).dtype == dtypes.fp8e5m2fnuz

 class TestFp8sConversions(unittest.TestCase):
  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=False, allow_infinity=False, min_value=-FP8E4M3_MAX, max_value=FP8E4M3_MAX))
@@ -152,16 +150,28 @@ class TestFp8sConversions(unittest.TestCase):
    np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3), torch.tensor(x, dtype=torch.float8_e4m3fn).view(torch.uint8).item())

  def test_float_to_fp8e4m3_extreme_values(self):
-    for x in [FP8E4M3_MAX, FP8E4M3_MAX*1.01, -FP8E4M3_MAX, -FP8E4M3_MAX*1.01, math.inf, -math.inf, math.nan, -math.nan]:
-      np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3), torch.tensor(x, dtype=torch.float8_e4m3fn).view(torch.uint8).item())
+    np.testing.assert_equal(float_to_fp8(FP8E4M3_MAX, dtypes.fp8e4m3), 126)
+    np.testing.assert_equal(float_to_fp8(FP8E4M3_MAX*1.01, dtypes.fp8e4m3), 126)
+    np.testing.assert_equal(float_to_fp8(math.inf, dtypes.fp8e4m3), 127)
+    np.testing.assert_equal(float_to_fp8(-FP8E4M3_MAX, dtypes.fp8e4m3), 254)
+    np.testing.assert_equal(float_to_fp8(-FP8E4M3_MAX*1.01, dtypes.fp8e4m3), 254)
+    np.testing.assert_equal(float_to_fp8(-math.inf, dtypes.fp8e4m3), 255)
+    np.testing.assert_equal(float_to_fp8(math.nan, dtypes.fp8e4m3), 127)
+    np.testing.assert_equal(float_to_fp8(-math.nan, dtypes.fp8e4m3), 255)

  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=False, allow_infinity=False, min_value=-FP8E5M2_MAX, max_value=FP8E5M2_MAX))
  def test_float_to_fp8e5m2(self, x):
    np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e5m2), torch.tensor(x, dtype=torch.float8_e5m2).view(torch.uint8).item())

  def test_float_to_fp8e5m2_extreme_values(self):
-    for x in [FP8E5M2_MAX, FP8E5M2_MAX*1.01, -FP8E5M2_MAX, -FP8E5M2_MAX*1.01, math.inf, -math.inf, math.nan, -math.nan]:
-      np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e5m2), torch.tensor(x, dtype=torch.float8_e5m2).view(torch.uint8).item())
+    np.testing.assert_equal(float_to_fp8(FP8E5M2_MAX, dtypes.fp8e5m2), 123)
+    np.testing.assert_equal(float_to_fp8(FP8E5M2_MAX*1.01, dtypes.fp8e5m2), 123)
+    np.testing.assert_equal(float_to_fp8(math.inf, dtypes.fp8e5m2), 124)
+    np.testing.assert_equal(float_to_fp8(-FP8E5M2_MAX, dtypes.fp8e5m2), 251)
+    np.testing.assert_equal(float_to_fp8(-FP8E5M2_MAX*1.01, dtypes.fp8e5m2), 251)
+    np.testing.assert_equal(float_to_fp8(-math.inf, dtypes.fp8e5m2), 252)
+    np.testing.assert_equal(float_to_fp8(math.nan, dtypes.fp8e5m2), 126)
+    np.testing.assert_equal(float_to_fp8(-math.nan, dtypes.fp8e5m2), 254)

  @given(strat.integers(min_value=0, max_value=255))
  def test_fp8e4m3_to_float(self, x):
@@ -171,30 +181,6 @@ class TestFp8sConversions(unittest.TestCase):
  def test_fp8e5m2_to_float(self, x):
    np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e5m2), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e5m2).float().item())

-  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=False, allow_infinity=False, min_value=-FP8E4M3FNUZ_MAX, max_value=FP8E4M3FNUZ_MAX))
-  def test_float_to_fp8e4m3fnuz(self, x):
-    np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3fnuz), torch.tensor(x, dtype=torch.float8_e4m3fnuz).view(torch.uint8).item())
-
-  def test_float_to_fp8e4m3fnuz_extreme_values(self):
-    for x in [FP8E4M3FNUZ_MAX, FP8E4M3FNUZ_MAX*1.01, -FP8E4M3FNUZ_MAX, -FP8E4M3FNUZ_MAX*1.01, math.inf, -math.inf, math.nan, 0.0, -0.0]:
-      np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e4m3fnuz), torch.tensor(x, dtype=torch.float8_e4m3fnuz).view(torch.uint8).item())
-
-  @given(strat.floats(width=32, allow_subnormal=True, allow_nan=False, allow_infinity=False, min_value=-FP8E5M2FNUZ_MAX, max_value=FP8E5M2FNUZ_MAX))
-  def test_float_to_fp8e5m2fnuz(self, x):
-    np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.float8_e5m2fnuz).view(torch.uint8).item())
-
-  def test_float_to_fp8e5m2fnuz_extreme_values(self):
-    for x in [FP8E5M2FNUZ_MAX, FP8E5M2FNUZ_MAX*1.01, -FP8E5M2FNUZ_MAX, -FP8E5M2FNUZ_MAX*1.01, math.inf, -math.inf, math.nan, 0.0, -0.0]:
-      np.testing.assert_equal(float_to_fp8(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.float8_e5m2fnuz).view(torch.uint8).item())
-
-  @given(strat.integers(min_value=0, max_value=255))
-  def test_fp8e4m3fnuz_to_float(self, x):
-    np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e4m3fnuz), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e4m3fnuz).float().item())
-
-  @given(strat.integers(min_value=0, max_value=255))
-  def test_fp8e5m2fnuz_to_float(self, x):
-    np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e5m2fnuz).float().item())
-
@unittest.skipUnless(is_dtype_supported(dtypes.bfloat16), "bfloat16 not supported")
 class TestBFloat16(unittest.TestCase):
  def test_bf16_creation_numpy(self):
@@ -516,3 +502,4 @@ class TestOpsBFloat16(unittest.TestCase):

 if __name__ == '__main__':
  unittest.main()
+
@@ -52,8 +52,6 @@ class ht:
 ht.bfloat16 = ht.uint16.filter(lambda x: ((x >> 7) & 0xFF) != 0)  # filter subnormal bfloat16
 ht.fp8e4m3 = ht.uint8
 ht.fp8e5m2 = ht.uint8
-ht.fp8e4m3fnuz = ht.uint8
-ht.fp8e5m2fnuz = ht.uint8

 def universal_test(a, b, dtype, op):
  if not isinstance(op, tuple): op = (op, op)
@@ -69,8 +67,7 @@ def universal_test(a, b, dtype, op):
    if not is_dtype_supported(dtype) or dtype in EMULATED_DTYPES.tolist(dtypes): # denormals are zero
      fe, fm = dtypes.finfo(dtype)
      atol, rtol = 2 ** (2 - (1 << (fe - 1))), 2 ** (-fm)
-    else: atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1),
-                        dtypes.fp8e4m3fnuz:(1e-1, 1e-1), dtypes.fp8e5m2fnuz:(5e-1, 5e-1)}.get(dtype, (1e-10, 1e-7))
+    else: atol, rtol = {dtypes.bfloat16:(1e-3, 1e-2), dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2:(1.0, 5e-1)}.get(dtype, (1e-10, 1e-7))
    np.testing.assert_allclose(tensor_value, numpy_value, atol=atol, rtol=rtol)
  else: np.testing.assert_equal(tensor_value, numpy_value)

@@ -90,8 +87,7 @@ def universal_test_unary(a, dtype, op):
  else: tensor_value, numpy_value = op[0](ta).numpy(), op[1](ta.numpy())
  if dtype in dtypes.floats:
    atol, rtol = { dtypes.float16:(1e-3, 1e-2), dtypes.bfloat16:(1e-3, 2e-2),
-      dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2: (1.0, 5e-1),
-      dtypes.fp8e4m3fnuz:(1e-1, 1e-1), dtypes.fp8e5m2fnuz: (5e-1, 5e-1)}.get(dtype, (1e-6, 1e-5))
+      dtypes.fp8e4m3:(1e-1, 1e-1), dtypes.fp8e5m2: (1.0, 5e-1)}.get(dtype, (1e-6, 1e-5))
    np.testing.assert_allclose(tensor_value, numpy_value, atol=atol, rtol=rtol)
  else: np.testing.assert_equal(tensor_value, numpy_value)

@@ -159,26 +155,6 @@ class TestDTypeALU(unittest.TestCase):
  def test_emulated_fp8e5m2(self, a, b, op):
    universal_test(from_storage_scalar(a, dtypes.fp8e5m2), from_storage_scalar(b, dtypes.fp8e5m2), dtypes.fp8e5m2, op)

-  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3fnuz), f"no fp8e4m3fnuz on {Device.DEFAULT}")
-  @given(ht.fp8e4m3fnuz, ht.fp8e4m3fnuz, strat.sampled_from(binary_operations))
-  def test_fp8e4m3fnuz(self, a, b, op):
-    universal_test(from_storage_scalar(a, dtypes.fp8e4m3fnuz), from_storage_scalar(b, dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
-
-  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e5m2fnuz), f"no fp8e5m2fnuz on {Device.DEFAULT}")
-  @given(ht.fp8e5m2fnuz, ht.fp8e5m2fnuz, strat.sampled_from(binary_operations))
-  def test_fp8e5m2fnuz(self, a, b, op):
-    universal_test(from_storage_scalar(a, dtypes.fp8e5m2fnuz), from_storage_scalar(b, dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, op)
-
-  @given(ht.fp8e4m3fnuz, ht.fp8e4m3fnuz, strat.sampled_from(binary_operations))
-  @Context(EMULATED_DTYPES="fp8e4m3fnuz")
-  def test_emulated_fp8e4m3fnuz(self, a, b, op):
-    universal_test(from_storage_scalar(a, dtypes.fp8e4m3fnuz), from_storage_scalar(b, dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
-
-  @given(ht.fp8e5m2fnuz, ht.fp8e5m2fnuz, strat.sampled_from(binary_operations))
-  @Context(EMULATED_DTYPES="fp8e5m2fnuz")
-  def test_emulated_fp8e5m2fnuz(self, a, b, op):
-    universal_test(from_storage_scalar(a, dtypes.fp8e5m2fnuz), from_storage_scalar(b, dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, op)
-
  @given(ht.float32, strat.sampled_from(unary_operations))
  def test_float32_unary(self, a, op): universal_test_unary(a, dtypes.float32, op)

@@ -222,30 +198,6 @@ class TestDTypeALU(unittest.TestCase):
    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e5m2) != 0.0)
    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e5m2), dtypes.fp8e5m2, op)

-  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e4m3fnuz), f"no fp8e4m3fnuz on {Device.DEFAULT}")
-  @given(ht.fp8e4m3fnuz, strat.sampled_from(unary_operations))
-  def test_fp8e4m3fnuz_unary(self, a, op):
-    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz) != 0.0)
-    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
-
-  @unittest.skipUnless(is_dtype_supported(dtypes.fp8e5m2fnuz), f"no fp8e5m2fnuz on {Device.DEFAULT}")
-  @given(ht.fp8e5m2fnuz, strat.sampled_from(unary_operations))
-  def test_fp8e5m2fnuz_unary(self, a, op):
-    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz) != 0.0)
-    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, op)
-
-  @given(ht.fp8e4m3fnuz, strat.sampled_from(unary_operations))
-  @Context(EMULATED_DTYPES="fp8e4m3fnuz")
-  def test_emulated_fp8e4m3fnuz_unary(self, a, op):
-    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz) != 0.0)
-    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e4m3fnuz), dtypes.fp8e4m3fnuz, op)
-
-  @given(ht.fp8e5m2fnuz, strat.sampled_from(unary_operations))
-  @Context(EMULATED_DTYPES="fp8e5m2fnuz")
-  def test_emulated_fp8e5m2fnuz_unary(self, a, op):
-    if op[1] == np.reciprocal: assume(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz) != 0.0)
-    universal_test_unary(from_storage_scalar(a, dtype=dtypes.fp8e5m2fnuz), dtypes.fp8e5m2fnuz, op)
-
  @given(ht.uint8, ht.uint8, strat.sampled_from(integer_binary_operations))
  def test_uint8(self, a, b, op): universal_test(a, b, dtypes.uint8, op)

@@ -366,7 +318,7 @@ class TestDTypeALU(unittest.TestCase):

  @unittest.expectedFailure
  def test_unsafe_cast_float_to_int_failure(self):
-    val = float(dtypes.int32.max - 1)
+    val = float(dtypes.max(dtypes.int32) - 1)
    t1 = Tensor([val], dtype=dtypes.float32).cast(dtypes.int32)
    t2 = Tensor(val, dtype=dtypes.float32).cast(dtypes.int32)
    np.testing.assert_equal(t1.item(), t2.item())
@@ -15,7 +15,7 @@ from test.helpers import needs_second_gpu
 np.random.seed(1337)
 Tensor.manual_seed(1337)
 BUF_SIZE = 4096
-RUN_CNT = 5
+RUN_CNT = 4

 cached_prgs = {}
 def helper_exec_op(device, outbuf, inbufs):
@@ -47,17 +47,6 @@ def helper_create_offset_rawbuffer(base, offset=0):
  x = Buffer(base.device, base.size-offset, base.dtype, base=base, offset=offset)
  return x.ensure_allocated()

-def helper_alloc_rawbuffer_sized(device, size, fill=False):
-  rawbuf = Buffer(device, size, dtypes.int).ensure_allocated()
-  if fill:
-    with Context(DEBUG=0):
-      data = np.random.randint(-10000, 10000, size=rawbuf.size, dtype=_to_np_dtype(rawbuf.dtype))
-      rawbuf.copyin(Tensor(data).realize().uop.base.realized.as_memoryview())
-  return rawbuf
-
-def helper_make_view(base, offset_elems, size_elems):
-  return Buffer(base.device, size_elems, base.dtype, base=base, offset=offset_elems * base.dtype.itemsize).ensure_allocated()
-
 def helper_run_jit(jis, bufs, out_buffers):
  for rawbuf in out_buffers:
    mv = memoryview(bytearray(rawbuf.size * rawbuf.dtype.itemsize))
@@ -91,14 +80,6 @@ def helper_test_graphs(graph_impl, graphs, runs=RUN_CNT):

@unittest.skipUnless(Device[Device.DEFAULT].graph is not None, "graph support required")
 class TestGraph(unittest.TestCase):
-  def skip_if_no_offset(self):
-    if not hasattr(Device[Device.DEFAULT].allocator, "_offset"): self.skipTest("device does not support _offset")
-
-  def skip_if_not_multigraph(self):
-    graph = g.func if isinstance(g:=(d:=Device[Device.DEFAULT]).graph, functools.partial) else g
-    if not issubclass(graph, MultiGraphRunner): self.skipTest("graph is not supported (not MultiGraphRunner)")
-    if not hasattr(d.allocator, '_transfer') or not d.allocator.supports_transfer: self.skipTest("device is not supported (no transfers)")
-
  def test_order_2_writes_to_same_buf(self):
    d0 = Device.DEFAULT
    b0 = [helper_alloc_rawbuffer(d0, fill=True) for _ in range(5)]
@@ -129,6 +110,11 @@ class TestGraph(unittest.TestCase):

    helper_test_graphs(Device[d0].graph, graphs)

+  def skip_if_not_multigraph(self):
+    graph = g.func if isinstance(g:=(d:=Device[Device.DEFAULT]).graph, functools.partial) else g
+    if not issubclass(graph, MultiGraphRunner): self.skipTest("graph is not supported (not MultiGraphRunner)")
+    if not hasattr(d.allocator, '_transfer') or not d.allocator.supports_transfer: self.skipTest("device is not supported (no transfers)")
+
  def test_order_copy_writed(self):
    self.skip_if_not_multigraph()

@@ -279,58 +265,5 @@ class TestGraph(unittest.TestCase):

    helper_test_graphs(Device[d0].graph, graphs)

-  def test_partial_write_preserves_write_dep(self):
-    self.skip_if_not_multigraph()
-    self.skip_if_no_offset()
-    d0 = Device.DEFAULT
-
-    base = helper_alloc_rawbuffer_sized(d0, BUF_SIZE * 2, fill=True)
-    copy_src_full = helper_alloc_rawbuffer_sized(d0, BUF_SIZE * 2, fill=True)
-    copy_src_lo = helper_alloc_rawbuffer(d0, fill=True)
-    v_lo = helper_make_view(base, 0, BUF_SIZE)
-    v_hi = helper_make_view(base, BUF_SIZE, BUF_SIZE)
-    a, c = [helper_alloc_rawbuffer(d0, fill=True) for _ in range(2)]
-
-    graphs = [
-      [helper_copy_op(d0, base, copy_src_full), helper_copy_op(d0, v_lo, copy_src_lo), helper_exec_op(d0, c, [v_hi, a])]
-    ]
-    helper_test_graphs(Device[d0].graph, graphs)
-
-  def test_partial_write_preserves_read_dep(self):
-    self.skip_if_not_multigraph()
-    self.skip_if_no_offset()
-    d0 = Device.DEFAULT
-
-    base = helper_alloc_rawbuffer_sized(d0, BUF_SIZE * 2, fill=True)
-    copy_dst = helper_alloc_rawbuffer_sized(d0, BUF_SIZE * 2, fill=True)
-    copy_src_lo = helper_alloc_rawbuffer(d0, fill=True)
-    v_lo = helper_make_view(base, 0, BUF_SIZE)
-    v_hi = helper_make_view(base, BUF_SIZE, BUF_SIZE)
-    a, b = [helper_alloc_rawbuffer(d0, fill=True) for _ in range(2)]
-
-    graphs = [
-      [helper_copy_op(d0, copy_dst, base), helper_copy_op(d0, v_lo, copy_src_lo), helper_exec_op(d0, v_hi, [a, b])]
-    ]
-    helper_test_graphs(Device[d0].graph, graphs)
-
-  def test_middle_write_splits_write_dep(self):
-    self.skip_if_not_multigraph()
-    self.skip_if_no_offset()
-    d0 = Device.DEFAULT
-
-    base = helper_alloc_rawbuffer_sized(d0, BUF_SIZE * 3, fill=True)
-    copy_src_full = helper_alloc_rawbuffer_sized(d0, BUF_SIZE * 3, fill=True)
-    copy_src_mid = helper_alloc_rawbuffer(d0, fill=True)
-    v_lo = helper_make_view(base, 0, BUF_SIZE)
-    v_mid = helper_make_view(base, BUF_SIZE, BUF_SIZE)
-    v_hi = helper_make_view(base, BUF_SIZE * 2, BUF_SIZE)
-    a, c, e = [helper_alloc_rawbuffer(d0, fill=True) for _ in range(3)]
-
-    graphs = [
-      [helper_copy_op(d0, base, copy_src_full), helper_copy_op(d0, v_mid, copy_src_mid),
-       helper_exec_op(d0, c, [v_lo, a]), helper_exec_op(d0, e, [v_hi, a])]
-    ]
-    helper_test_graphs(Device[d0].graph, graphs)
-
 if __name__ == '__main__':
  unittest.main()
@@ -115,7 +115,7 @@ class TestImageDType(unittest.TestCase):
    tst = data.numpy()
    it = data.cast(dtypes.imagef((9,27,4))).realize()
    # the underlying UOp is identical
-    #self.assertIs(it.uop.base.realized, data.uop.base.realized)
+    self.assertIs(it.uop.base.realized, data.uop.base.realized)
    np.testing.assert_equal(tst, it.numpy())

  def test_image_and_back_wrong_shape(self):
@@ -332,6 +332,7 @@ class TestJit(unittest.TestCase):
    assert len(res3) == 10, "All values should be different, rand works in jit."
    assert res3 != res2, "Jit rand is diff with diff seeds"

+  #@unittest.expectedFailure # requires contiguous folding
  def test_jit_random_after_unrealized_random(self):
    @TinyJit
    def f(): return Tensor.rand()
@@ -475,7 +476,7 @@ class TestJit(unittest.TestCase):
    b = f(Tensor([2.0]))
    assert abs((a - b).item()) > 0.5

-  def test_jit_init_empty(self):
+  def test_jit_init_with_empty_different_size(self):
    @TinyJit
    def f(x:Tensor) -> Tensor: return (x + 1).realize()

@@ -484,16 +485,9 @@ class TestJit(unittest.TestCase):
    # scalar const input is not allowed
    with self.assertRaises(JitError):
      f(Tensor(2.0)).item()
-    # self.assertEqual(f(Tensor([2.0])).item(), 1.0) # TODO: wrong output, should be 3.0. currently depends on empty value
-
-  def test_jit_init_empty_alt(self):
-    @TinyJit
-    def f(a:Tensor, b:Tensor) -> Tensor: return b.assign(a+1)
-    for i in range(4):
-      a = Tensor([i])
-      b = Tensor.empty_like(a)
-      c = f(a, b)
-      self.assertEqual(c.item(), i+1)
+    # list input has different view structure than empty(1)
+    with self.assertRaises(JitError):
+      f(Tensor([2.0])).item()

@unittest.skip("Pending multioutput implementation #3607")
 class TestMultioutputJit(unittest.TestCase):
@@ -651,8 +645,8 @@ class TestJitFree(unittest.TestCase):
  def test_replan_buffers_memory_layout(self):
    if not hasattr(Device[Device.DEFAULT].allocator, '_offset'): raise unittest.SkipTest("replan_buffers_memory_layout useless")

-    ext_tensor = Tensor([1,24,23,45,1]).contiguous()
-    ext_tensor_2 = Tensor([2,2,2,2,2]).contiguous()
+    ext_tensor = Tensor([1,24,23,45,1])
+    ext_tensor_2 = Tensor([2,2,2,2,2])
    @TinyJit
    def fxn(x:Tensor):
      out = (x*ext_tensor_2+ext_tensor).reshape(5,1).expand(5, 100).contiguous()
@@ -660,9 +654,9 @@ class TestJitFree(unittest.TestCase):
    for i in range(5):
      out = fxn(Tensor([i,1,2,3,4]))
      self.assertEqual(out.item(), 11400+200*i)
-    self.assertEqual(len(set([b.base for item in fxn.captured.jit_cache for b in item.bufs if b is not None])), 4)
+    assert len(set([b.base for item in fxn.captured.jit_cache for b in item.bufs if b is not None])) == 4
    fxn.captured.replan_buffers_memory_layout()
-    self.assertEqual(len(set([b.base for item in fxn.captured.jit_cache for b in item.bufs if b is not None])), 2)
+    assert len(set([b.base for item in fxn.captured.jit_cache for b in item.bufs if b is not None])) == 2

    out = fxn(Tensor([11,1,2,3,4]))
    self.assertEqual(out.item(), 13600)
@@ -3,7 +3,8 @@ import unittest
 from dataclasses import replace

 from tinygrad.codegen.opt import Opt, OptOps
-from tinygrad.uop.ops import UOp, Ops, GroupOp, AxisType
+from tinygrad.codegen.gpudims import get_grouped_dims
+from tinygrad.uop.ops import UOp, Ops, GroupOp, AxisType, PatternMatcher, graph_rewrite, UPat
 from tinygrad.device import Device, Buffer, is_dtype_supported
 from tinygrad.tensor import Tensor, _to_np_dtype
 from tinygrad.engine.realize import run_schedule, CompiledRunner, get_program
@@ -252,6 +253,100 @@ class TestLinearizer(unittest.TestCase):
      if any(x.op is Ops.END and x.src[1].op in GroupOp.ALU for x in u.src):
        assert end_range < uops.index(u)

+  def test_grouped_dims(self):
+    def _assert_grouped_dims(prefix, dims, max_sizes, reverse_dims, expected_sizes, assert_same_length = True):
+      idxs = get_grouped_dims(prefix, dims, max_sizes, reverse_dims)
+      loop_idxs = dedup(flatten([[y for y in x.toposort() if y.op is Ops.SPECIAL] for x in idxs]))
+      loop_idxs = sorted(loop_idxs, key=lambda uop: uop.arg)
+      sizes = [x.src[0].arg for x in loop_idxs]
+      assert len(idxs) == len(dims), f"expected idxs to have same length as dims {len(dims)}, got {len(idxs)}"
+      if assert_same_length:
+        assert len(loop_idxs) == min(len(sizes), len(dims)), f"expected idxs to have length {min(len(sizes), len(dims))}, got {len(loop_idxs)}"
+      assert sizes == expected_sizes, f"expected sizes={expected_sizes}, got {sizes=}"
+      # TODO: add these back after uop symbolic
+      # for i in range(len(dims)):
+      #   assert idxs[i].max+1 == dims[i], f"idxs[{i}] should have max {dims[i]-1}"
+      # for i in range(len(loop_idxs)):
+      #   assert loop_idxs[i].expr.startswith(prefix), f"loop_idxs[{i}] must start with {prefix}"
+      #   assert loop_idxs[i].max+1 == sizes[i], f"loop_idxs[{i}] should have max {sizes[i]-1}"
+
+    # no-op
+    _assert_grouped_dims("gidx", (2,), (16,16,16), False, [2])
+    _assert_grouped_dims("gidx", (2,3), (16,16,16), False, [2,3])
+
+    # check reverse dims
+    _assert_grouped_dims("gidx", (2,3), (16,16,16), True, [3,2])
+    _assert_grouped_dims("gidx", (2,3,4), (16,16,16), False, [2,3,4])
+
+    # test splitting globals:    len(dims) == len(max)
+    _assert_grouped_dims("gidx", (64,3,4), (16,16,16), False, [16,12,4])
+    _assert_grouped_dims("gidx", (64,3,4), (16,4,16), False, [16,3,16])
+    _assert_grouped_dims("gidx", (64,3,4), (16,16,16), True, [16,3,16])
+    _assert_grouped_dims("gidx", (128,3,4), (16,4,256), False, [16,3,32])
+    _assert_grouped_dims("gidx", (4,4,512), (16,4,256), False, [8,4,256])
+
+    # prefer group_dim strategy when possible
+    _assert_grouped_dims("gidx", (512,4,2), (8192,2,2), False, [2048,2])
+
+    # test splitting globals:    len(dims) < len(max)
+    #                            len(dim)        ->          len(limited)
+    #                              1             ->             2
+    _assert_grouped_dims("gidx", (128,), (16,16,256), False, [16,8], False)
+    #                              1             ->             3
+    _assert_grouped_dims("gidx", (65536,), (16,16,256), False, [16,16,256], False)
+    #                              2             ->             3
+    _assert_grouped_dims("gidx", (128,128), (16,16,256), False, [16,16,64], False)
+    #                              2             ->             2
+    _assert_grouped_dims("gidx", (65536,2), (65535,65535,65535), False, [32768,4], False)
+    # test when the only divisor is the square root of dim
+    _assert_grouped_dims("gidx", (121,), (12,12,12), False, [11,11], False)
+
+    # collapse on onto the left most axis
+    _assert_grouped_dims("gidx", (2,3,4,5), (16,16,16), False, [6,4,5])
+    _assert_grouped_dims("gidx", (2,3,4,5), (32,16,16), True, [20,3,2])
+    # _assert_grouped_dims("gidx", (Variable("start_pos",1,2),3,4,5), (32,16,16), True, [20,3,Variable("start_pos",1,2)])
+
+    # collapse on left-most available axis (the left most is too small)
+    _assert_grouped_dims("gidx", (2,3,4,5), (4,16,16), False, [2,12,5])
+    _assert_grouped_dims("gidx", (2,3,4,5), (16,16,16), True, [5,12,2])
+
+    # _assert_grouped_dims("gidx", (Variable("start_pos",1,2),3,4,5), (16,16,16), False, [Variable("start_pos",1,2)*3,4,5])
+
+    # dim too large and not factorable
+    with self.assertRaises(RuntimeError):
+      get_grouped_dims("gidx", (23,), (16,16,16), False,)
+    with self.assertRaises(RuntimeError):
+      get_grouped_dims("gidx", (128,3,4), (16,2,2), False,)
+
+    # too large for sizes
+    with self.assertRaises(RuntimeError):
+      get_grouped_dims("gidx", (2,3,4,5,6), (16,16,16))
+
+    # TODO: In the above cases we only test if the shape after reshape is correct, never the indices.
+    # We should check if the returned indices are correct, for all cases.
+    # (65536, 2) -> (32768, 4)
+    dims, expected_limited_dims = (65536,2), (32768, 4)
+    idxs = get_grouped_dims("gidx", dims, (65535,65535,65535))
+    def match_div(): raise RuntimeError("match_div")
+    def match_mod(): raise RuntimeError("match_mod")
+    flat_idx_pattern = UPat(Ops.SPECIAL, arg='gidx0')*expected_limited_dims[1]+UPat(Ops.SPECIAL, arg='gidx1')
+    pm = PatternMatcher([
+      (flat_idx_pattern//dims[1], match_div),
+      (flat_idx_pattern%dims[1], match_mod)
+    ])
+
+    with self.assertRaises(RuntimeError) as error:
+      graph_rewrite(idxs[0], pm)
+    self.assertIn("match_div", str(error.exception))
+
+    with self.assertRaises(RuntimeError) as error:
+      graph_rewrite(idxs[1], pm)
+    self.assertIn("match_mod", str(error.exception))
+
+    # # variable too large
+    # with self.assertRaises(AssertionError):
+    #   get_grouped_dims("gidx", (Variable("start_pos",0,16),3,4), (16,16,16), False,)
+
  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
  def test_default_global_reversed(self):
    # shrink so that the dims do not collapse
@@ -94,7 +94,7 @@ class TestMultiTensor(unittest.TestCase):
  def _test_shard_op(self, op, out, n=4):
    t = Tensor.ones(n).contiguous().realize().shard(devices_2, 0)
    r = op(t).realize()
-    #assert t.uop.is_realized, "shard didn't realize"
+    assert t.uop.is_realized, "shard didn't realize"
    self.assertEqual(r.tolist(), out)
  def test_shard_reshape(self): self._test_shard_op(lambda t:t.reshape(2, 2), [[1.,1.],[1.,1.]])
  def test_shard_elementwise(self): self._test_shard_op(lambda t:(t+t).reshape(2, 2), [[2.,2.],[2.,2.]])
@@ -135,6 +135,34 @@ class TestMultiTensor(unittest.TestCase):
      si.run()
    self.assertEqual(len(set(names)), 1, "function was relinearized")

+  @unittest.skip("this doesn't fold because shard_ calls contiguous on all lbs")
+  def test_sharded_memory(self):
+    # Buffer may be stuck in track_cross_buffer
+    for x in (d0, d1, d2, d3, d4): Device[x].synchronize()
+    mem_base = GlobalCounters.mem_used
+
+    X = Tensor.ones(256).contiguous().realize()
+    assert GlobalCounters.mem_used-mem_base== X.dtype.itemsize * 256, GlobalCounters.mem_used-mem_base
+    X.shard_(devices_4).realize()
+    for x in (d0, d1, d2, d3, d4): Device[x].synchronize()
+    assert GlobalCounters.mem_used-mem_base == X.dtype.itemsize * 256 * 4, GlobalCounters.mem_used-mem_base
+
+    X = Tensor.ones(256).contiguous().realize()
+    assert GlobalCounters.mem_used-mem_base == X.dtype.itemsize * 256, GlobalCounters.mem_used-mem_base
+    X.shard_(devices_4, axis=0).realize()
+    for x in (d0, d1, d2, d3, d4): Device[x].synchronize()
+    assert GlobalCounters.mem_used-mem_base == X.dtype.itemsize * 256, GlobalCounters.mem_used-mem_base
+
+    X = Tensor.ones(256).realize()
+    assert GlobalCounters.mem_used-mem_base == 0
+    X.shard_(devices_4).realize()
+    assert GlobalCounters.mem_used-mem_base == 0
+
+    X = Tensor.ones(256).realize()
+    assert GlobalCounters.mem_used-mem_base == 0
+    X.shard_(devices_4, axis=0).realize()
+    assert GlobalCounters.mem_used-mem_base == 0
+
  def test_shard_same_device(self):
    X = Tensor.ones(256).contiguous().realize()
    X.shard_((d1, X.device), 0)
@@ -228,17 +256,17 @@ class TestMultiTensor(unittest.TestCase):
      a,b = _test_allreduce(Tensor.rand(256, 256))
      np.testing.assert_almost_equal(a.numpy(), b.numpy(), decimal=5)

-  def test_multiple_to_single_device(self):
-    kernel_counts = {}
-    for ring in (0, 2):
-      GlobalCounters.reset()
-      with Context(RING=ring, SCACHE=0):
-        t = Tensor.arange(32).contiguous().shard(devices_4, 0).to(Device.DEFAULT)
-        t.realize()
-      kernel_counts[ring] = GlobalCounters.kernel_count
-      self.assertEqual(t.device, Device.DEFAULT)
-      np.testing.assert_equal(t.numpy(), np.arange(32))
-    self.assertNotEqual(kernel_counts[0], kernel_counts[2])
+  def test_multiple_to_single_device_naive(self):
+    with Context(RING=0):
+      t = Tensor.arange(32).shard(devices_4, 0).to(Device.DEFAULT).realize()
+    self.assertEqual(t.device, Device.DEFAULT)
+    np.testing.assert_equal(t.numpy(), np.arange(32))
+
+  def test_multiple_to_single_device_ring(self):
+    with Context(RING=2):
+      t = Tensor.arange(32).shard(devices_4, 0).to(Device.DEFAULT).realize()
+    self.assertEqual(t.device, Device.DEFAULT)
+    np.testing.assert_equal(t.numpy(), np.arange(32))

  def test_allreduce_all2all(self):
    with Context(ALL2ALL=2):
@@ -626,6 +654,54 @@ class TestMultiTensor(unittest.TestCase):
    assert isinstance(jf.jit_cache[4].prg, BufferCopy)
    assert isinstance(jf.jit_cache[5].prg, graph_d1)

+  @unittest.skip("no longer supports uneven shard")
+  def test_uneven_shard(self):
+    for N in range(1, 6):
+      X = Tensor.rand(4, 1, 257).contiguous().realize()
+      n = X.numpy()
+      devices = tuple(f"{Device.DEFAULT}:{i}" for i in range(N))
+      X.shard_(devices, 2)
+      np.testing.assert_equal(X.numpy(), n)
+      np.testing.assert_equal(X.reshape(2, 2, 257).numpy(), n.reshape((2, 2, 257)))
+      np.testing.assert_equal(X.shrink(((0,2), (0, 1), (0,257))).numpy(), n[0:2, 0:1, 0:257])
+      np.testing.assert_equal(X.expand((4, 4, 257)).numpy(), np.tile(n, (1, 4, 1)))
+      np.testing.assert_equal(X.permute((0, 2, 1)).numpy(), np.transpose(n, (0, 2, 1)))
+
+  @unittest.skip("no longer supports uneven shard")
+  def test_uneven_multiple_zeros(self):
+    for data in ([1, 2, 3, 4], [1, 2, 3], [1, 2], [1], []):
+      for N in (1, 2, 3, 4):
+        devices = tuple(f"{Device.DEFAULT}:{i}" for i in range(N))
+        # make sure something is computed on each device
+        X = ((Tensor(data).shard(devices, axis=0) + 1).realize() - 1).realize()
+        np.testing.assert_equal(X.numpy(), data)
+
+  @unittest.skip("no longer supports uneven shard")
+  def test_uneven_shard_with_empty(self):
+    N = 4
+    X = Tensor.rand(16, 1, 3).contiguous().realize()
+    np_x = X.numpy()
+    devices = tuple(f"{Device.DEFAULT}:{i}" for i in range(N))
+
+    # test empty shard
+    np.testing.assert_equal(X.shard(devices, 0).numpy(), np_x)
+
+    # test reshape with empty shard
+    np.testing.assert_equal(X.shard(devices, 0).reshape(8, 1, 6).numpy(), np_x.reshape(8, 1, 6))
+
+  @unittest.skip("no longer supports uneven shard")
+  def test_multiple_uneven_shard(self):
+    N = 4
+    X = Tensor.rand(4, 1, 257).contiguous().realize()
+    Y = Tensor.rand(4, 1, 257).contiguous().realize()
+    np_x, np_y = X.numpy(), Y.numpy()
+    devices = tuple(f"{Device.DEFAULT}:{i}" for i in range(N))
+    X.shard_(devices, 2)
+    Y.shard_(devices, 2)
+    np.testing.assert_equal(X.numpy(), np_x)
+    np.testing.assert_equal(Y.numpy(), np_y)
+    np.testing.assert_equal((X + Y).numpy(), np_x + np_y)
+
  def test_bn_ast_on_devices(self):
    t = Tensor.empty((16, 64, 112, 112)).shard(devices_4, axis=0)
    bn = nn.BatchNorm2d(64)
@@ -676,7 +752,34 @@ class TestMultiTensor(unittest.TestCase):

    # test no left join
    with self.assertRaises((AssertionError, ValueError)):
-      t0.reshape((26*15,7)).contiguous().schedule()
+      t0.reshape((26*15,7)).schedule()
+
+  @unittest.skip("no longer supports uneven shard")
+  def test_reshape_on_axis_uneven(self):
+    def reshape_helper(t0, t, t_axis):
+      assert t.uop.axis == t_axis
+      np.testing.assert_allclose(t0.reshape(t.shape).numpy(), t.numpy())
+
+    t0 = Tensor.rand((4, 42, 15)).shard(devices_3, axis=1, splits=[14, 7, 21])
+
+    # ok to reshape as long as elements remain on same device
+    reshape_helper(t0, t0.reshape(2, 2, 42, 3, 5), 2)
+    # split to the right
+    reshape_helper(t0, t0.reshape(2, 2, 6, 7, 15), 2)
+    # split off and merge to the right
+    reshape_helper(t0, t0.reshape(4, 6, 105), 1)
+    # really blend the axes together
+    reshape_helper(t0, t0.reshape(4, 30, 21), 1)
+    # split off 1-shape
+    reshape_helper(t0, t0.reshape(4, 1, 42, 15), 2)
+    reshape_helper(t0, t0.reshape(4, 6, 1, 7, 15), 1)
+
+    # assert if cannot maintain shard axis without moving items between devices
+    with self.assertRaises(AssertionError): t0.reshape(4, 7, 6, 15)
+    # assert for degenerate reshape
+    with self.assertRaises(AssertionError): t0.reshape(4, 5, 7, 15)
+    # assert for cannot maintain axis
+    with self.assertRaises(AssertionError): t0.reshape(4, 3, 2, 7, 15)

  # it doesn't work like this anymore
  # NOTE: this never failed in assign_multi, it failed tensor spec because MULTI was never pushed in the graph
@@ -746,6 +849,16 @@ class TestMultiTensor(unittest.TestCase):
    self.assertEqual(rab.device, devices_4)
    self.assertEqual(rab.uop.axis, 0)

+  @unittest.skip("no longer supports uneven shard")
+  def test_rand_like_uneven_shard(self):
+    t = Tensor.empty((4, 42, 15)).shard(devices_3, axis=1)
+    t2 = Tensor.rand_like(t)
+    self.assertEqual(t.shape, t2.shape)
+    self.assertEqual(t.device, t2.device)
+    self.assertEqual(t.dtype, t2.dtype)
+    self.assertEqual(t.uop.axis, t2.uop.axis)
+    assert all(tlb.shape == t2lb.shape for tlb, t2lb in zip(t.uop.src, t2.uop.src))
+
  def test_rand_like_none_shard(self):
    t = Tensor.empty((16, 16)).shard(devices_2)
    t2 = Tensor.rand_like(t)
@@ -781,14 +894,6 @@ class TestMultiTensor(unittest.TestCase):
    t2.realize()
  def test_full_like_on_shard_axis(self): self.test_full_like_on_shard(0)

-  def test_full_like_shrink_on_shard_axis(self):
-    t = Tensor.ones(16, 16, dtype=dtypes.int).shard(devices_2, axis=0)
-    out = Tensor.full_like(t, 2)[:, :8]
-    sched = out.schedule()
-    self.assertEqual(len(sched), 2) # TODO: 0. fix mstack_early_shrink
-    run_schedule(sched)
-    self.assertEqual(out.tolist(), [[2]*8]*16)
-
  def test_dropout_on_shard(self):
    with Tensor.train():
      X = Tensor.ones(256).to(devices_2)
@@ -805,6 +910,15 @@ class TestMultiTensor(unittest.TestCase):
      assert set(unique) == {0, 2}, unique
      assert 200 < counts[0] < 312, counts[0]

+  @unittest.skip("no longer supports uneven shard")
+  def test_dropout_on_uneven_shard_axis(self):
+    with Tensor.train():
+      X = Tensor.ones(256).shard(devices_3, axis=0)
+      output = X.dropout(0.5).numpy()
+      unique, counts = np.unique(output, return_counts=True)
+      assert set(unique) == {0, 2}, unique
+      assert 100 < counts[0] < 156, counts[0]
+
  @unittest.skip("TODO: this requires forced_realize to be deleted.")
  def test_shard_memory(self):
    devices = (d0, d1, d2, d3)
@@ -812,15 +926,13 @@ class TestMultiTensor(unittest.TestCase):
    t.shard_(devices, axis=0).realize()
    assert all([lb is lb.base and lb.realized.base.size == 4 * 16 for lb in t.uop.src])

+  @unittest.skip("this is unreliable on OSX")
  def test_clone(self):
-    for axis in (None, 0):
-      t = Tensor.arange(16).reshape(4, 4).shard(devices_2, axis=axis).contiguous().realize()
-      t_clone = t.clone().realize()
-      self.assertEqual(t_clone.device, t.device)
-      self.assertEqual(t_clone.uop.axis, axis)
-      self.assertEqual(t_clone.tolist(), t.tolist())
-      t_clone += 1
-      self.assertNotEqual(t_clone.tolist(), t.tolist())
+    t = Tensor.rand(16, 16).shard(devices_2, axis=None)
+    np.testing.assert_allclose(t.numpy(), t.clone().numpy())
+
+    t = Tensor.rand(16, 16).shard(devices_2, axis=0)
+    np.testing.assert_allclose(t.numpy(), t.clone().numpy())

  @unittest.skip("RANGEIFY doesn't support multi const folding")
  def test_multi_const_folding(self):
@@ -869,18 +981,18 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):

    with self.assertRaises(AssertionError):
      # sharded axis shrink on non-device boundry is not allowed
-      a = t.shrink(((0, 3), (0, 8))).contiguous()
+      a = t.shrink(((0, 3), (0, 8)))
+      a.schedule()
+    with self.assertRaises(AssertionError):
+      # cannot shrink sharded and non-sharded axis at the same time
+      a = t.shrink(((0, 2), (2, 4)))
      a.schedule()
-    a = t.shrink(((0, 2), (2, 4)))
-    assert a.shape == (2, 2)
-    ref = Tensor.arange(64).reshape(8, 8).shrink(((0, 2), (2, 4)))
-    np.testing.assert_equal(a.numpy(), ref.numpy())

-    a = t.shrink(((0, 2), (0, 8))).contiguous()
+    a = t.shrink(((0, 2), (0, 8)))
    a.schedule()
    assert a.shape == (2, 8)

-    p = a.pad(((0, 6), (0, 0))).contiguous()
+    p = a.pad(((0, 6), (0, 0)))
    p.schedule()
    assert p.shape == (8, 8)

@@ -930,6 +1042,24 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):
      np.testing.assert_allclose(a.reshape((2, 1, 8)).expand((2, 5, 8)).numpy(), b.reshape((2, 1, 8)).expand((2, 5, 8)).numpy(), rtol=1e-7, atol=1e-3)
      np.testing.assert_allclose(a.flip(-1).numpy(), b.flip(-1).numpy(), rtol=1e-7, atol=1e-3)

+  @unittest.skip("no longer supports uneven shard")
+  def test_uneven(self):
+    t = Tensor.arange(24).reshape(3, 8).contiguous().realize()
+    t.shard_([f"{Device.DEFAULT}:{i}" for i in range(2)], axis=0)
+
+    a = t.shrink(((0, 2), None))
+    b = t.shrink(((2, 3), None))
+    na = t.numpy()[0:2]
+    nb = t.numpy()[2:3]
+    np.testing.assert_equal(a.numpy(), na)
+    np.testing.assert_equal(b.numpy(), nb)
+    np.testing.assert_equal((a+1).numpy(), na+1)
+    np.testing.assert_equal((b+1).numpy(), nb+1)
+    np.testing.assert_equal((1+a).numpy(), 1+na)
+    np.testing.assert_equal((1+b).numpy(), 1+nb)
+    np.testing.assert_equal((a+a).numpy(), na+na)
+    np.testing.assert_equal((b+b).numpy(), nb+nb)
+
  def test_add_two_partitions(self):
    t = Tensor.arange(64).reshape(8, 8).contiguous().realize()
    t.shard_([f"{Device.DEFAULT}:{i}" for i in range(4)], axis=0)
@@ -6,11 +6,9 @@ from tinygrad.helpers import getenv, IMAGE, DEBUG, CI, Context, CPU_LLVM, AMD_LL
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.device import is_dtype_supported
-from tinygrad.renderer.cstyle import QCOMCLRenderer
 from tinygrad.renderer.nir import NIRRenderer

-TINY_BACKEND = getenv("TINY_BACKEND")
-if TINY_BACKEND:
+if getenv("TINY_BACKEND"):
  import tinygrad.nn.torch # noqa: F401 # pylint: disable=unused-import
  torch.set_default_device("tiny")

@@ -420,6 +418,7 @@ class TestOps(unittest.TestCase):
    helper_test_op(None, lambda x: x.round(), vals=[[1.499, 1.5, 1.501, 1.0, 2.1, 0.0, -5.0, -2.499, -2.5, -2.501]], forward_only=True)
    helper_test_op(None, lambda x: x.round(), vals=[[2.5, -1.5]], forward_only=True)

+  @unittest.skipIf(Device.DEFAULT == "WEBGPU" and CI, "isinf check of 'nan' fails on CI software-based vulkan")
  def test_isinf(self):
    val = [float('-inf'), 0., float('inf'), float('nan'), 1.1]
    helper_test_op(None, torch.isinf, Tensor.isinf, vals=[val], forward_only=True)
@@ -437,7 +436,7 @@ class TestOps(unittest.TestCase):
    helper_test_op([(45,35), (45,35), (45,35)], lambda x,y,z: x.lerp(y,z))
    helper_test_op(None, lambda x,y,z: x.lerp(y,z), vals=[[1.,2.,3.], [4.,5.,6.], 0.5])

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
+  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
  def test_tril(self):
    helper_test_op([(3,3)], lambda x: x.tril())
    helper_test_op([(3,3)], lambda x: x.tril(1))
@@ -455,7 +454,7 @@ class TestOps(unittest.TestCase):
    helper_test_op([(5,3,3)], lambda x: x.tril(1))
    helper_test_op(None, lambda x: x.tril(), vals=[[[True] * 3] * 3], forward_only=True)

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
+  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
  def test_triu(self):
    helper_test_op([(3,3)], lambda x: x.triu())
    helper_test_op([(3,3)], lambda x: x.triu(1))
@@ -479,9 +478,9 @@ class TestOps(unittest.TestCase):
    helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[1., 0., 3., -4.], 3.])
    helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[1., 0., 3., -4.], [-1., -2., 3., 0.]])
    helper_test_op(None, torch.maximum, Tensor.maximum,
-                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.max], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.max(dtypes.int)], forward_only=True)
    helper_test_op(None, torch.maximum, Tensor.maximum,
-                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.min], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.min(dtypes.int)], forward_only=True)
    helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[True, False, False], True], forward_only=True)
    helper_test_op(None, torch.maximum, Tensor.maximum, vals=[[True, False, False], [True, True, False]], forward_only=True)

@@ -496,9 +495,9 @@ class TestOps(unittest.TestCase):
    helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[1., 0., 3., -4.], 3.])
    helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[1., 0., 3., -4.], [-1., -2., 3., 0.]])
    helper_test_op(None, torch.minimum, Tensor.minimum,
-                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.max], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.max(dtypes.int)], forward_only=True)
    helper_test_op(None, torch.minimum, Tensor.minimum,
-                   vals=[[-1234, 0, 1234, dtypes.int.max, dtypes.int.min], dtypes.int.min], forward_only=True)
+                   vals=[[-1234, 0, 1234, dtypes.max(dtypes.int), dtypes.min(dtypes.int)], dtypes.min(dtypes.int)], forward_only=True)
    helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[True, False, False], True], forward_only=True)
    helper_test_op(None, torch.minimum, Tensor.minimum, vals=[[True, False, False], [True, True, False]], forward_only=True)

@@ -641,6 +640,8 @@ class TestOps(unittest.TestCase):
    helper_test_op([(45,65), (45,65)], lambda x,y: x**y)
    helper_test_op([(45,65), (45,65)], lambda x,y: x.pow(y))

+  # TODO: WEBGPU NaN handling in pow operations
+  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "WEBGPU NaN handling differs")
  def test_pow(self):
    helper_test_op([(45,65)], lambda x: x**0)
    helper_test_op([(45,65)], lambda x: x**1)
@@ -759,14 +760,12 @@ class TestOps(unittest.TestCase):
    data = [[1,-8,1],[32,1,6]]
    tor = torch.tensor(data, dtype=torch.int)
    ten = Tensor(data, dtype=dtypes.int32)
-    # NOTE: this breaks assigns because it's folded to 0!
    helper_test_op([], lambda: tor^tor, lambda: ten^ten, forward_only=True)
    helper_test_op([], lambda: tor^0x1337, lambda: ten^0x1337, forward_only=True)
    helper_test_op([], lambda: 0x1337^tor, lambda: 0x1337^ten, forward_only=True)

    self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_xor(y), expected=RuntimeError)

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_and(self):
    data = [[1,-8,1],[32,1,6]]
    tor = torch.tensor(data, dtype=torch.int)
@@ -784,7 +783,6 @@ class TestOps(unittest.TestCase):

    self.helper_test_exception([(4), (4)], lambda x,y: x.bitwise_and(y), expected=RuntimeError)

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_or(self):
    data = [[1,-8,1],[32,1,6]]
    tor = torch.tensor(data, dtype=torch.int)
@@ -1173,7 +1171,6 @@ class TestOps(unittest.TestCase):
    helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[False, True]])
    helper_test_op(None, lambda x: x.type(torch.int32).argmax().type(torch.int32), lambda x: x.argmax(), forward_only=True, vals=[[True, False]])

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_argmin(self):
    # check if it returns the first index for multiple occurrences
    helper_test_op(None, lambda x: x.argmin().type(torch.int32), lambda x: x.argmin(), forward_only=True, vals=[[2, 2]])
@@ -1479,7 +1476,6 @@ class TestOps(unittest.TestCase):
  def test_prod_dtype_arg(self):
    with self.assertRaises(AttributeError): Tensor([1.0, 2.0]).prod(dtype="")

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_min(self):
    helper_test_op([(3,3)], lambda x: x.min())
    helper_test_op([(45,3)], lambda x: x.min())
@@ -1508,6 +1504,7 @@ class TestOps(unittest.TestCase):
    helper_test_op([(3,3)], lambda x: torch.full_like(x, 2).prod(), lambda x: (x.full_like(2)).prod(), forward_only=True)
    helper_test_op([(3,3)], lambda x: torch.full_like(x, 2).max(), lambda x: (x.full_like(2)).max(), forward_only=True)

+  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
  def test_any(self):
    helper_test_op([(3,4,5,6)], lambda x: x.any(), forward_only=True)
    helper_test_op(None, lambda x: x.any(), vals=[[True, True]], forward_only=True)
@@ -1519,7 +1516,7 @@ class TestOps(unittest.TestCase):
  def test_any_zero_axis(self):
    helper_test_op([(1,0,3,0,5)], lambda x: x.any(axis=(1,3)), forward_only=True)

-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
+  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
  def test_all(self):
    helper_test_op([(3,4,5,6)], lambda x: x.all(), forward_only=True)
    helper_test_op(None, lambda x: x.all(), vals=[[True, True]], forward_only=True)
@@ -1546,6 +1543,7 @@ class TestOps(unittest.TestCase):
    helper_test_op([(3, 4, 5, 6)], lambda x: x.isclose(x + 1e-9, rtol=0.01), forward_only=True)
    helper_test_op(None, lambda x,y: x.isclose(y), vals=[[1e-7, 1e-8, 1e-9], [0.0, 0.0, 0.0]], forward_only=True)

+  @unittest.skipIf(Device.DEFAULT == "WEBGPU" and CI, "isinf check of 'nan' fails on CI software-based vulkan")
  def test_isclose_edge_cases(self):
    for a in [math.inf, -math.inf, math.nan, 0.0]:
      for b in [math.inf, -math.inf, math.nan, 0.0]:
@@ -1669,15 +1667,6 @@ class TestOps(unittest.TestCase):
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(1), atol=1e-7, grad_atol=1e-7)
    helper_test_op([(10,10,10)], lambda x: x.log_softmax(2), atol=1e-7, grad_atol=1e-7)

-  def test_normalize(self):
-    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x), lambda x: x.normalize(), atol=1e-7, grad_atol=1e-7)
-    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x, dim=0), lambda x: x.normalize(dim=0), atol=1e-7, grad_atol=1e-7)
-    helper_test_op([(10,10,10)], lambda x: torch.nn.functional.normalize(x, dim=2), lambda x: x.normalize(dim=2), atol=1e-7, grad_atol=1e-7)
-    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x, p=1), lambda x: x.normalize(p=1), atol=1e-7, grad_atol=1e-7)
-    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x, p=3, dim=0), lambda x: x.normalize(p=3, dim=0), atol=1e-7, grad_atol=1e-7)
-    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x, p=0), lambda x: x.normalize(p=0), atol=1e-7, grad_atol=1e-7)
-    helper_test_op([(45,65)], lambda x: torch.nn.functional.normalize(x, p=-1), lambda x: x.normalize(p=-1), atol=1e-7, grad_atol=1e-7)
-
  def test_logsumexp(self):
    helper_test_op([(45,65)], lambda x: torch.logsumexp(x, dim=0), lambda x: x.logsumexp(0), atol=1e-7, grad_atol=1e-7)
    helper_test_op([(45,65)], lambda x: torch.logsumexp(x, dim=0, keepdim=True), lambda x: x.logsumexp(0, True), atol=1e-7, grad_atol=1e-7)
@@ -2893,7 +2882,6 @@ class TestOps(unittest.TestCase):
    helper_test_op([(2,5,6,5,3,4)], lambda x: x[...,c,:,e], lambda x: x[...,k,:,p])

  @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_slice_fancy_indexing_dim_collapse_int(self):
    a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
    # dim collapse from int
@@ -2904,7 +2892,6 @@ class TestOps(unittest.TestCase):
    helper_test_op([(2,5,6,5,3,4)], lambda x: x[1,:,3:11:2,d,0:2], lambda x: x[1,:,3:11:2,o,0:2])

  @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_slice_fancy_indexing_dim_inject_none(self):
    a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
    # dim injection from None
@@ -2939,7 +2926,6 @@ class TestOps(unittest.TestCase):
                            lambda x: x[Tensor([[0,1,-1],[-1,-2,0]]), Tensor([2,1,-1])])

  @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_slice_fancy_indexing_list_indices(self):
    a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
    helper_test_op([(2,5,6,5,3,4)], lambda x: x[((0,),)])
@@ -2951,7 +2937,6 @@ class TestOps(unittest.TestCase):
    helper_test_op([(2,5,6,5,3,4)], lambda x: x[a,(2,1,0),c,(-2,1,0),e], lambda x: x[i,(2,1,0),k,(-2,1,0),p])

  @slow_test
-  @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, QCOMCLRenderer), "QCOM CL vectorized bool bug")
  def test_slice_fancy_indexing_tuple_indices(self):
    a,b,c,d,e,i,j,k,o,p = self._get_index_randoms()
    helper_test_op([(2,5,6,5,3,4)], lambda x: x[(((0,),),)], lambda x: x[(((0,),),)])
@@ -3293,6 +3278,7 @@ class TestOps(unittest.TestCase):
    helper_test_op([(20,)], lambda x: (x>0.5).nonzero().int(), lambda x: (x>0.5).nonzero(), forward_only=True)
    helper_test_op([(10, 5, 3)], lambda x: (x>0.5).nonzero().int(), lambda x: (x>0.5).nonzero(), forward_only=True)

+  @unittest.skipIf(Device.DEFAULT == "QCOM", "OpenCL fails to compile this (both on GPU(qcom)/QCOM backends)")
  def test_cast(self):
    helper_test_op([(3, 3)], lambda x: x.float())
    helper_test_op(None, lambda x: x.float(), vals=[[0, 1, 2, 3]], forward_only=True)
@@ -0,0 +1,230 @@
+import unittest
+import numpy as np
+from tinygrad import Tensor, UOp, nn
+from tinygrad.uop.ops import AxisType, Ops
+
+class TestOuterworldReduce(unittest.TestCase):
+  def test_reduce(self):
+    x = Tensor.ones(5, 5).contiguous()
+    a = UOp.range(5, -1, AxisType.REDUCE)
+    out = x[a]
+    # TODO: syntax for this
+    t = Tensor(UOp(Ops.REDUCE, dtype=out.uop.dtype, src=(out.uop, a), arg=Ops.ADD))
+    self.assertListEqual(t.tolist(), [5.,5.,5.,5.,5.])
+
+# TODO: delete test_outerworld_range?
+class TestOuterRange(unittest.TestCase):
+  def test_simple_range(self):
+    a = Tensor.ones(10).contiguous()
+    acc = Tensor.zeros().contiguous()
+    Tensor.realize(a, acc)
+
+    # this is fold
+    i = UOp.range(10, -100, AxisType.OUTER)
+    acc_i = acc.uop.after(i)
+    vi = UOp.variable("i", i.vmin, i.vmax).bind(i)
+    out = Tensor(acc.uop.after(acc_i.store(acc_i + a[vi].uop).end(i)))
+    out.realize()
+    assert out.item() == 10.0
+
+  def test_inner_range(self):
+    a = Tensor.ones(10, 10).contiguous()
+    acc = Tensor.zeros(10).contiguous()
+    Tensor.realize(a, acc)
+
+    # this is fold
+    i = UOp.range(10, -100, AxisType.OUTER)
+    acc_i = acc.uop.after(i)
+    vi = UOp.variable("i", i.vmin, i.vmax).bind(i)
+    out = Tensor(acc.uop.after(acc_i.store(acc_i + a[:, vi].uop).end(i)))
+    out.realize()
+    self.assertEqual(out.tolist(), [10.0]*10)
+
+  def test_range_matmul(self):
+    vec = Tensor.randn(1, 10).realize()
+    mats = Tensor.randn(3, 10, 10).realize()
+
+    # 3 matmuls in "scan"
+    ref = ((vec @ mats[0]) @ mats[1]) @ mats[2]
+    ref.realize()
+
+    # 3 matmuls with outer world range
+    i = UOp.range(3, -100, AxisType.OUTER)
+    vec_i = Tensor(vec.uop.after(i))
+    comp = vec_i.contiguous() @ mats[i]
+    store = vec_i.uop.store(comp.uop).end(i)
+    out = Tensor(vec.uop.after(store))
+    out.realize()
+
+    # TODO: testing allclose
+    assert Tensor.allclose(ref, out, atol=1e-5), f"max diff {(ref-out).abs().max().item()}"
+
+class TestOuterScan(unittest.TestCase):
+  def _test_scan(self):
+    vec = Tensor.randn(1, 10).realize()
+    mats = Tensor.randn(3, 10, 10).realize()
+
+    # 3 matmuls in "scan"
+    vec1 = vec @ mats[0]
+    vec2 = vec1 @ mats[1]
+    vec3 = vec2 @ mats[2]
+    ref = Tensor.stack(vec1, vec2, vec3)
+    ref.realize()
+    return vec, mats, ref
+
+  def test_uop_scan_matmul(self):
+    vec, mats, ref = self._test_scan()
+
+    # 3 matmuls with SCAN
+    i = UOp.range(3, -100, AxisType.OUTER)
+    out = Tensor.empty(3, 1, 10)
+    phi = Tensor(i.eq(0).where(vec.uop, out[(i-1).maximum(0)].uop))
+    comp = phi @ mats[i]
+    store = out[i].uop.store(comp.uop).end(i)
+    out = Tensor(out.uop.after(store))
+    out.realize()
+
+    # TODO: testing allclose
+    assert Tensor.allclose(ref, out, atol=1e-5), f"max diff {(ref-out).abs().max().item()}"
+
+class TestOuterworld(unittest.TestCase):
+  def test_range_plus_1(self):
+    t = Tensor.arange(100).reshape(10,10).realize()
+
+    # passthrough ranges
+    a = UOp.range(10, -1)
+    sel = t[a] + 1
+    assert sel.shape == (10,)
+    cpy = sel.reshape(1, 10).expand(a, 10).contiguous().realize()
+
+    self.assertTrue((t+1==cpy).all().item())
+
+  def test_range_plus_1_transpose(self):
+    t = Tensor.arange(100).reshape(10,10).realize()
+
+    # passthrough ranges
+    a = UOp.range(10, -1)
+    sel = t[a] + 1
+    assert sel.shape == (10,)
+    cpy = sel.reshape(10, 1).expand(10, a).contiguous().realize()
+
+    self.assertTrue(((t+1).T==cpy).all().item())
+
+  def test_flip_range(self):
+    t = Tensor.rand(10, 10).realize()
+
+    # passthrough ranges
+    a = UOp.range(10, -1)
+    sel = t[9-a]
+    cpy = sel.reshape(1, 10).expand(a, 10).contiguous().realize()
+
+    self.assertTrue((t.flip(0)==cpy).all().item())
+
+  def test_vmap(self):
+    def f(x): return x.sum(axis=0)*2
+
+    x = Tensor.ones(3, 10, 2).contiguous()
+
+    # vmap across axis 0
+    a = UOp.range(3, -1)
+    out = f(x[a])
+    out = out.reshape(1, 2).expand(a, 2).contiguous()
+
+    # 3x2 grid of 20
+    out.realize()
+    self.assertTrue((out==20).all().item())
+
+  def test_fancy_vmap(self):
+    def f(x,y): return x+y
+
+    x = Tensor.arange(9).reshape(3,3).contiguous()
+    y = Tensor.arange(9).reshape(3,3).contiguous()
+
+    a = UOp.range(3, -1)
+    out = f(x[:,a], y[a,:])
+    # TODO: this should support flatten
+    out = out.reshape(1, 3).expand(a, 3).contiguous().realize()
+    self.assertListEqual([[0,4,8],[4,8,12],[8,12,16]], out.tolist())
+
+class TestVmap(unittest.TestCase):
+  def test_vmap_inner(self, axis_type=AxisType.LOOP, fuse=False, grad=False):
+    x = Tensor.ones(1, 10).contiguous().requires_grad_()
+    mats = Tensor.ones(3, 10, 10).contiguous().requires_grad_()
+
+    ref = x @ mats
+    if fuse: ref = ref * 2
+
+    # vmap across axis 0
+    a = UOp.range(3, -1, axis_type)
+    out = x @ mats[a]
+    out = out.reshape(1, 10).pad(((a,(3-a)-1), None))
+    out = Tensor(out.uop.reduce(a, arg=Ops.ADD))
+    if fuse: out = out * 2
+    if grad:
+      out.mean().backward()
+      np.testing.assert_allclose(mats.grad.numpy(), (2./30) if fuse else (1./30))
+    out.realize()
+
+    # TODO: testing allclose
+    assert Tensor.allclose(ref, out, atol=1e-6), f"max diff {(ref-out).abs().max().item()}"
+  def test_vmap_inner_fuse(self): self.test_vmap_inner(fuse=True)
+  def test_vmap_outer(self): self.test_vmap_inner(AxisType.OUTER)
+  def test_vmap_outer_fuse(self): self.test_vmap_inner(AxisType.OUTER, fuse=True)
+
+  def test_vmap_inner_grad(self): self.test_vmap_inner(grad=True)
+  def test_vmap_inner_fuse_grad(self): self.test_vmap_inner(fuse=True, grad=True)
+  def test_vmap_outer_grad(self): self.test_vmap_inner(AxisType.OUTER, grad=True)
+
+  def test_vmap_convs(self):
+    layers = [
+      nn.Conv2d(1, 8, 3), Tensor.relu,
+      nn.Conv2d(8, 8, 3), Tensor.relu]
+    img = Tensor.randn(4, 1, 16, 16).realize(*nn.state.get_parameters(layers))
+    a = UOp.range(4, -1, AxisType.OUTER)
+    out = img[a:a+1].sequential(layers)
+    out = out.pad(((a,(4-a)-1), None, None, None))
+    out = Tensor(out.uop.reduce(a, arg=Ops.ADD))
+    out.realize()
+    np.testing.assert_allclose(out.numpy(), img.sequential(layers).numpy(), atol=1e-6)
+
+  def test_vmap_gemm(self):
+    layers = [
+      nn.Linear(16, 16, bias=False), Tensor.relu,
+      nn.Linear(16, 16, bias=False), Tensor.relu]
+    img = Tensor.randn(4, 16).realize(*nn.state.get_parameters(layers))
+    a = UOp.range(4, -1, AxisType.OUTER)
+    out = img[a:a+1].sequential(layers)
+    out = out.pad(((a,(4-a)-1), None))
+    out = Tensor(out.uop.reduce(a, arg=Ops.ADD))
+    out.realize()
+    np.testing.assert_allclose(out.numpy(), img.sequential(layers).numpy(), atol=1e-6)
+
+  @unittest.skip("this is broken, we need to lower the outer reduce in the outer graph")
+  def test_vmap_gemm_grad(self):
+    layers = [
+      nn.Linear(16, 16, bias=False), Tensor.relu,
+      nn.Linear(16, 16, bias=False), Tensor.relu]
+    layer_tensors = nn.state.get_parameters(layers)
+    img = Tensor.randn(4, 16).realize(*layer_tensors)
+    for l in layer_tensors: l.requires_grad_()
+    a = UOp.range(4, -1, AxisType.OUTER)
+    out = img[a:a+1].sequential(layers)
+    out = out.pad(((a,(4-a)-1), None))
+    out = Tensor(out.uop.reduce(a, arg=Ops.ADD))
+    out.mean().backward()
+    grads = [l.grad for l in layer_tensors]
+    out.realize(*grads)
+    out_grads = [x.numpy() for x in grads]
+
+    # compute reference grads
+    for l in layer_tensors: l.grad = None
+    img.sequential(layers).mean().backward()
+    grads = [l.grad for l in layer_tensors]
+    out.realize(*grads)
+    ref_grads = [x.numpy() for x in grads]
+
+    # compare
+    for o,r in zip(out_grads, ref_grads): np.testing.assert_allclose(o, r, atol=1e-6)
+
+if __name__ == '__main__':
+  unittest.main()
@@ -0,0 +1,19 @@
+import unittest
+from tinygrad import Tensor
+
+class TestOuterCall(unittest.TestCase):
+  def test_outer_call_assign(self):
+    a = Tensor.zeros(10,10).contiguous()
+    b = Tensor.ones(10,10).contiguous()
+    Tensor.realize(a,b)
+
+    pa = a.as_param(0)
+    pb = b.as_param(1)
+    out = Tensor.call(a, b, fxn=pa.assign(pa+pb))
+    out.realize()
+
+    print(a.numpy())
+    assert (a == 1).all().item()
+
+if __name__ == '__main__':
+  unittest.main()
@@ -0,0 +1,148 @@
+import unittest
+from tinygrad import Tensor, nn, Variable, UOp
+
+# outerworld range should support three things
+# 1. full optimizer steps (test_model_bound_range)
+# 2. gradient accumulation (you want to end the range before running the optimizer)
+# 3. stacked linear layers
+
+class Model:
+  def __init__(self): self.w = nn.Linear(64, 8, bias=False)
+  def __call__(self, x:Tensor) -> Tensor: return self.w(x)
+
+def get_model_and_opt():
+  Tensor.manual_seed(1337)
+  m = Model()
+  opt = nn.optim.SGD(nn.state.get_parameters(m), lr=0.1, weight_decay=0)
+  return m, opt
+
+class TestOuterworldRange(unittest.TestCase):
+  STEPS = 5
+  BS = 20
+
+  @classmethod
+  def setUpClass(cls):
+    Tensor.manual_seed(1338)
+    # it learns to compute mean
+    cls.X = Tensor.randn(cls.STEPS, cls.BS, 64).contiguous().realize()
+    cls.Y = cls.X.reshape(cls.STEPS, cls.BS, 8, 8).mean(axis=-1).contiguous().realize()
+    cls.losses = cls._get_model_baseline()
+
+  def _compare(self, losses):
+    for i,(x,y) in enumerate(zip(self.losses, losses)):
+      self.assertAlmostEqual(x, y, places=5, msg=f"mismatch at {i} in {self.losses} vs {losses}")
+
+  @classmethod
+  @Tensor.train()
+  def _get_model_baseline(self):
+    m, opt = get_model_and_opt()
+    losses = []
+    for i in range(self.STEPS):
+      opt.zero_grad()
+      loss = (m(self.X[i]) - self.Y[i]).square().mean()
+      loss.backward()
+      loss.realize(*opt.schedule_step())
+      losses.append(loss.item())
+    return losses
+
+  @Tensor.train()
+  def test_model_grad_acc(self):
+    m, opt = get_model_and_opt()
+    losses = []
+    for i in range(self.STEPS):
+      opt.zero_grad()
+      sub_batch_size = self.BS//2
+      loss = 0
+      scaling_factor = self.BS//sub_batch_size
+      for j in range(0, self.BS, sub_batch_size):
+        sub_loss = (m(self.X[i][j:j+sub_batch_size]) - self.Y[i][j:j+sub_batch_size]).square().mean() / scaling_factor
+        sub_loss.backward()
+        loss += sub_loss
+      loss.realize(*opt.schedule_step())
+      losses.append(loss.item())
+    self._compare(losses)
+
+  @Tensor.train()
+  def test_model_variable(self):
+    m, opt = get_model_and_opt()
+    losses = []
+    vi = Variable('i', 0, self.STEPS-1)
+    for i in range(self.STEPS):
+      vib = vi.bind(i)
+      opt.zero_grad()
+      loss = (m(self.X[vib]) - self.Y[vib]).square().mean()
+      loss.backward()
+      loss.realize(*opt.schedule_step())
+      losses.append(loss.item())
+    self._compare(losses)
+
+  @Tensor.train()
+  def test_model_scheduled(self):
+    m, opt = get_model_and_opt()
+    losses = []
+    for i in range(self.STEPS):
+      opt.zero_grad()
+      loss = (m(self.X[i]) - self.Y[i]).square().mean()
+      loss.backward()
+      opt.schedule_step()
+      losses.append(loss)
+    self._compare(Tensor.stack(*losses).tolist())
+
+  @Tensor.train()
+  def test_model_scheduled_setitem(self):
+    m, opt = get_model_and_opt()
+    losses = Tensor.empty(self.STEPS)
+    for i in range(self.STEPS):
+      opt.zero_grad()
+      loss = (m(self.X[i]) - self.Y[i]).square().mean()
+      loss.backward()
+      opt.schedule_step()
+      # TODO: this shouldn't realize
+      losses[i] = loss.requires_grad_(False)
+    self._compare(losses.tolist())
+
+  @unittest.expectedFailure
+  @Tensor.train()
+  def test_model_scheduled_variable(self):
+    m, opt = get_model_and_opt()
+    losses = []
+    vi = Variable('i', 0, self.STEPS-1)
+    for i in range(self.STEPS):
+      vib = vi.bind(i)
+      opt.zero_grad()
+      loss = (m(self.X[vib]) - self.Y[vib]).square().mean()
+      loss.backward()
+      opt.schedule_step()
+      losses.append(loss)
+    self._compare(Tensor.stack(*losses).tolist())
+
+  @unittest.expectedFailure
+  @Tensor.train()
+  def test_model_scheduled_variable_setitem(self):
+    m, opt = get_model_and_opt()
+    losses = Tensor.empty(self.STEPS)
+    vi = Variable('i', 0, self.STEPS-1)
+    for i in range(self.STEPS):
+      vib = vi.bind(i)
+      opt.zero_grad()
+      loss = (m(self.X[vib]) - self.Y[vib]).square().mean()
+      loss.backward()
+      opt.schedule_step()
+      losses[vib] = loss.requires_grad_(False)
+    self._compare(losses.tolist())
+
+  @unittest.expectedFailure
+  @Tensor.train()
+  def test_model_bound_range(self):
+    m, opt = get_model_and_opt()
+    # TODO: should ranges be unique so you don't have to pass in the -1?
+    rng = UOp.range(self.STEPS, -1)
+    vib = Variable('i', 0, self.STEPS-1).bind(rng)
+    loss = (m(self.X[vib]) - self.Y[vib]).square().mean()
+    loss.backward()
+    losses = Tensor.empty(self.STEPS)
+    losses[vib] = loss
+    losses.realize(*opt.schedule_step())
+
+if __name__ == "__main__":
+  unittest.main()
@@ -1,4 +1,4 @@
-import unittest, struct, contextlib, statistics, gc
+import unittest, struct, contextlib, statistics, time, gc
 from tinygrad import Device, Tensor, dtypes, TinyJit
 from tinygrad.helpers import CI, getenv, Context, ProfileRangeEvent, cpu_profile, cpu_events, ProfilePointEvent, dedup
 from tinygrad.device import Buffer, BufferSpec, Compiled, ProfileDeviceEvent, ProfileGraphEvent
@@ -20,7 +20,7 @@ def helper_collect_profile(*devs):
  cpu_events.clear()

  profile_list = []
-  with Context(PROFILE=1):
+  with Context(VIZ=1, PROFILE=1):
    yield profile_list
    for dev in devs: dev.synchronize()
    for dev in devs: dev._at_profile_finalize()
@@ -170,19 +170,30 @@ class TestProfiler(unittest.TestCase):
    for (i1, d1), (i2, d2) in pairs:
      assert abs(jitter_matrix[i1][i2]) < 0.5, "jitter should be less than 0.5us"

+  @unittest.skip("this test is flaky")
  def test_cpu_profile(self):
    def test_fxn(err=False):
+      time.sleep(0.1)
      if err: raise Exception()
+      time.sleep(0.1)

    with helper_collect_profile(dev:=TestProfiler.d0) as profile:
-      with cpu_profile("test_1", dev):
+      with cpu_profile("test_1", dev.device):
        test_fxn(err=False)
      with self.assertRaises(Exception):
-        with cpu_profile("test_2", dev):
+        with cpu_profile("test_2", dev.device):
          test_fxn(err=True)

-    range_events = [p for p in profile if isinstance(p, ProfileRangeEvent) and p.device == dev]
+    range_events = [p for p in profile if isinstance(p, ProfileRangeEvent)]
    self.assertEqual(len(range_events), 2)
+    # record start/end time up to exit (error or success)
+    for e in range_events:
+      self.assertGreater(e.en, e.st)
+    e1, e2 = range_events
+    self.assertEqual([e1.name, e2.name], ["test_1", "test_2"])
+    # TODO: this is flaky
+    #self.assertLess(e1.st, e2.st)
+    #self.assertGreater(e1.en-e1.st, e2.en-e2.st)

  @unittest.skip("this test is flaky")
  @unittest.skipUnless(Device[Device.DEFAULT].graph is not None, "graph support required")
@@ -204,13 +204,13 @@ class TestQuantizeOnnx(unittest.TestCase):
    W = Tensor(m2:=(np.random.uniform(0, 255, size=(N,N)).astype(wi))).realize()
    tg_dtype = dtypes.int8 if xi == np.int8 else dtypes.uint8
    out = (X.int().matmul(W.int())//1000)
-    if clip: out = out.clip(tg_dtype.min, tg_dtype.max)
+    if clip: out = out.clip(dtypes.min(tg_dtype),dtypes.max(tg_dtype))
    out = out.cast(tg_dtype)
    opts = [Opt(op=OptOps.UPCAST, axis=1, arg=128), Opt(op=OptOps.UNROLL, axis=0, arg=4)] if opts is None else opts
    sexec(out, opts, replace_src, run_count=1)
    tout = out.numpy()
    mout = ((m1.astype(np.int32) @ m2.astype(np.int32)) // 1000)
-    if clip: mout = mout.clip(tg_dtype.min, tg_dtype.max)
+    if clip: mout = mout.clip(dtypes.min(tg_dtype),dtypes.max(tg_dtype))
    mout = mout.astype(xi)
    print(tout)
    print(mout)
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	ca906371fa	fix assign copy	2026-02-18 12:14:18 +08:00
geohot	fdb0b48ce9	assign after copy shouldn't contig	2026-02-18 12:03:20 +08:00