refactor to u.

refactor to use uop
python prg is pkl uops
2026-05-29 19:05:20 -07:00 · 2026-05-29 19:01:10 -07:00 · 2026-05-29 18:54:55 -07:00 · 2026-05-29 21:16:56 -04:00 · 2026-05-29 18:00:37 -07:00 · 2026-05-29 19:20:31 -04:00
133 changed files with 2114 additions and 1743 deletions
@@ -5,6 +5,7 @@ runs:
  steps:
    - name: Run process replay tests
      shell: bash
+      if: env.CAPTURE_PROCESS_REPLAY == '1'
      run: |
        export PR_TITLE=$(jq -r .pull_request.title "$GITHUB_EVENT_PATH")
        export CURRENT_SHA=${{ github.event.pull_request && github.event.pull_request.head.sha || github.sha }}
@@ -228,6 +228,11 @@ runs:

        sudo chown -R $USER:$USER /var/cache/apt/archives/

+    - name: Add clang to PATH (Linux)
+      if: inputs.llvm == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: echo "/usr/lib/llvm-20/bin" >> "$GITHUB_PATH"
+
    # **** AMD ****
    - name: Setup AMD (Linux)
      if: inputs.amd == 'true' && runner.os == 'Linux'
@@ -281,7 +286,7 @@ runs:
        cache-name: cache-gpuocelot-build-1
      with:
        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-gpuocelot-f463259669c69abce7b3a0567b6c284f348d0f32-rebuild-${{ env.CACHE_VERSION }}
    - name: Cache gpuocelot
      if: inputs.ocelot == 'true' && github.event_name != 'pull_request'
      id: cache-build
@@ -290,14 +295,14 @@ runs:
        cache-name: cache-gpuocelot-build-1
      with:
        path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-${{ env.CACHE_VERSION }}
+        key: ${{ runner.os }}-gpuocelot-f463259669c69abce7b3a0567b6c284f348d0f32-rebuild-${{ env.CACHE_VERSION }}
    - name: Clone/compile gpuocelot
      if: inputs.ocelot == 'true' && steps.cache-build-pr.outputs.cache-hit != 'true' && steps.cache-build.outputs.cache-hit != 'true'
      shell: bash
      run: |
-        git clone --recurse-submodules https://github.com/gpuocelot/gpuocelot.git ${{ github.workspace }}/gpuocelot
+        git clone --recurse-submodules https://github.com/tinygrad/gpuocelot.git ${{ github.workspace }}/gpuocelot
        cd ${{ github.workspace }}/gpuocelot/ocelot
-        git checkout b16039dc940dc6bc4ea0a98380495769ff35ed99
+        git checkout f463259669c69abce7b3a0567b6c284f348d0f32
        mkdir build
        cd build

@@ -306,10 +311,7 @@ runs:
          sudo xcode-select -s /Applications/Xcode_16.2.app/Contents/Developer
          CMAKE_ARGS="$CMAKE_ARGS -DBoost_INCLUDE_DIR=$(brew --prefix boost)/include -DBoost_LIBRARY_DIR=$(brew --prefix boost)/lib"
        else
-          curl -fL https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/linux-x86_64/cuda_nvcc-linux-x86_64-11.5.119-archive.tar.xz \
-            | sudo tar -xJ -C /usr/ --strip-components=1
-          curl -fL https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/linux-x86_64/cuda_cudart-linux-x86_64-11.5.117-archive.tar.xz \
-            | sudo tar -xJ -C /usr/ --strip-components=1
+          CMAKE_ARGS="$CMAKE_ARGS -DLLVM_DIR=$(llvm-config-15 --cmakedir)"
        fi

        cmake .. $CMAKE_ARGS
@@ -25,7 +25,7 @@ jobs:
      CI: ""
      CAPTURE_PROCESS_REPLAY: "0"
    runs-on: [self-hosted, macOS]
-    timeout-minutes: 3
+    timeout-minutes: 4
    defaults:
      run:
        shell: bash -e -o pipefail {0}
@@ -806,3 +806,16 @@ jobs:
        pkill -f 'extra/remote/serve.py' || true
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
+
+  llvmspeed:
+    name: LLVM Speed
+    runs-on: [self-hosted, Linux, tinyboxrandom]
+    timeout-minutes: 20
+    if: github.repository_owner == 'tinygrad'
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Speed Test
+      run: DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+    - name: Speed Test (BEAM=2)
+      run: BEAM=2 DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py
@@ -2,7 +2,7 @@ name: Unit Tests
 env:
  # increment this when downloads substantially change to avoid the internet
  CACHE_VERSION: '19'
-  CAPTURE_PROCESS_REPLAY: 1
+  CAPTURE_PROCESS_REPLAY: ${{ github.event_name == 'pull_request' && contains(github.event.pull_request.title, '[pr]') && '1' || '0' }}
  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  PYTHONPATH: ${{ github.workspace }}
  CHECK_OOB: 1
@@ -14,28 +14,14 @@ on:
  pull_request:
  workflow_dispatch:

-jobs:
-  llvmspeed:
-    name: LLVM Speed
-    runs-on: ubuntu-24.04
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: llvm-speed
-        deps: testing_unit
-        llvm: 'true'
-    - name: Speed Test
-      run: DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py
-    - name: Speed Test (BEAM=2)
-      run: BEAM=2 DEV=CPU:LLVM THREADS=0 python3 test/speed/external_test_speed_v_torch.py
+concurrency:
+  group: test-${{ github.event_name }}-${{ github.event_name == 'pull_request' && github.event.pull_request.number || github.run_id }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

+jobs:
  docs:
    name: Docs
-    runs-on: ubuntu-22.04
+    runs-on: &linux ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
    timeout-minutes: 10
    env:
      CHECK_OOB: 0
@@ -89,7 +75,7 @@ jobs:

  torchbackend:
    name: Torch Backend Tests
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -125,7 +111,7 @@ jobs:

  torchbackendmore:
    name: Torch Backend Tests More
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -147,7 +133,7 @@ jobs:

  bepython:
    name: Python Backend
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -215,7 +201,7 @@ jobs:

  linter:
    name: Linters
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 10

    steps:
@@ -246,7 +232,7 @@ jobs:

  nulltest:
    name: Null Tests
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 15

    steps:
@@ -277,7 +263,7 @@ jobs:

  unittest:
    name: Unit Tests
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 15

    steps:
@@ -290,7 +276,6 @@ jobs:
        pydeps: "pillow ftfy regex pre-commit"
        deps: testing_unit
        llvm: 'true'
-        amd: 'true'
    - name: Run pre-commit test hooks
      run: SKIP=ruff,mypy pre-commit run --all-files
    - name: Check Device.DEFAULT
@@ -321,7 +306,7 @@ jobs:
      matrix:
        group: [1, 2]
    name: SPEC=2 (${{ matrix.group }})
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -332,12 +317,13 @@ jobs:
        key: spec-unit
        deps: testing_unit
        python-version: '3.14'
+        llvm: 'true'
    - name: Test SPEC=2
      run: SPEC=2 pytest --maxfail=10 -n auto --durations=30 test/unit test/backend test/opt --ignore test/backend/test_custom_kernel.py --ignore test/unit/test_hashing.py --timeout 60 -k "not test_setitem_big" -k "not test_conv2d_ceildiv_edge_case" --splits 2 --group ${{ matrix.group }}

  fuzzing:
    name: Fuzzing
-    runs-on: ubuntu-latest
+    runs-on: *linux
    timeout-minutes: 10
    steps:
    - name: Checkout Code
@@ -358,7 +344,7 @@ jobs:

  testopenclimage:
    name: CL IMAGE Tests
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -378,7 +364,7 @@ jobs:

  testgpumisc:
    name: CL Misc tests
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 10
    steps:
      - name: Checkout Code
@@ -403,7 +389,7 @@ jobs:

  testopenpilot:
    name: openpilot Compile Tests
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -418,8 +404,6 @@ jobs:
      - name: Test openpilot model kernel count and gate usage
        run: |
          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1468 ALLOWED_GATED_READ_IMAGE=18 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
-      - name: Test openpilot CL compile fp16
-        run: FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
      - name: Test openpilot CL compile fp32 (test correctness)
        run: |
          DEV=CL IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
@@ -433,7 +417,7 @@ jobs:

  testonnxcpu:
    name: ONNX (CPU) Tests
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 20

    steps:
@@ -461,7 +445,7 @@ jobs:

  testopencl:
    name: ONNX (CL)+Optimization Tests
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 20
    steps:
      - name: Checkout Code
@@ -495,7 +479,7 @@ jobs:

  testllm:
    name: Test LLM
-    runs-on: ubuntu-24.04
+    runs-on: *linux
    timeout-minutes: 15
    env:
      CHECK_OOB: 0
@@ -520,7 +504,7 @@ jobs:

  testmodels:
    name: Models (llvm+cpu+gpu)
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -563,7 +547,7 @@ jobs:

  testdsp:
    name: Linux (DSP)
-    runs-on: ubuntu-24.04
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -572,8 +556,7 @@ jobs:
      uses: ./.github/actions/setup-tinygrad
      with:
        key: dsp-minimal
-        deps: testing_unit
-        pydeps: "onnx==1.18.0 onnxruntime ml_dtypes"
+        deps: testing
        llvm: "true"
        qemu: "true"
    - name: Set MOCKDSP env
@@ -581,13 +564,24 @@ jobs:
    - name: Run test_tiny on DSP
      run: DEBUG=2 DEV=DSP python test/test_tiny.py
    - name: Test transcendentals
-      run: CC=clang-20 DEBUG=2 DEV=DSP python test/backend/test_transcendental.py TestTranscendentalVectorized
+      run: DEBUG=2 DEV=DSP python test/backend/test_transcendental.py TestTranscendentalVectorized
    - name: Test quantize onnx
      run: DEBUG=2 DEV=DSP python3 test/backend/test_quantize_onnx.py

-  testwebgpu:
-    name: Linux (WebGPU)
-    runs-on: ubuntu-22.04
+  testlinux:
+    strategy:
+      fail-fast: false
+      matrix:
+        dev:
+          - 'CPU:CLANG'
+          - 'CPU:LLVM'
+          - 'CPU:LVP'
+          - 'CPU:X86'
+          - 'CL'
+          - 'WEBGPU'
+
+    name: Linux (DEV=${{ matrix.dev }})
+    runs-on: *linux
    timeout-minutes: 20
    steps:
    - name: Checkout Code
@@ -595,23 +589,27 @@ jobs:
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: webgpu-minimal
+        key: linux-${{ matrix.dev }}
        deps: testing_unit
        python-version: '3.12'
-        webgpu: 'true'
-    - name: Check Device.DEFAULT (WEBGPU) and print some source
+        llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') || contains(matrix.dev, 'CLANG') }}
+        mesa: ${{ contains(matrix.dev, 'LVP') && 'cpu' || 'false' }}
+        webgpu: ${{ matrix.dev == 'WEBGPU' }}
+        opencl: ${{ matrix.dev == 'CL' }}
+    - name: Set env
+      run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
+    - name: Check Device.DEFAULT and print some source
      run: |
-        DEV=WEBGPU python -c "from tinygrad import Device; assert Device.DEFAULT == 'WEBGPU', Device.DEFAULT"
-        DEV=WEBGPU DEBUG=4 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
-    - name: Run selected webgpu tests
-      run: |
-          DEV=WEBGPU WEBGPU_BACKEND="WGPUBackendType_Vulkan" python3 -m pytest -n=auto test/backend --durations=20
+        python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
+        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+    - name: Run backend tests
+      run: python -m pytest -n=auto test/backend --durations=20
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

  testamdasm:
    name: AMD ASM IDE
-    runs-on: ubuntu-24.04
+    runs-on: *linux
    timeout-minutes: 20
    env:
      DEV: MOCKKFD+AMD
@@ -658,7 +656,7 @@ jobs:

  testmockam:
    name: Linux (am)
-    runs-on: ubuntu-24.04
+    runs-on: *linux
    timeout-minutes: 15
    env:
      DEV: MOCKPCI+AMD
@@ -694,7 +692,7 @@ jobs:
        arch: [gfx1100, gfx1201, gfx950]

    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 15
    env:
      DEV: MOCKKFD+AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
@@ -729,7 +727,7 @@ jobs:
        backend: [ptx, nv]

    name: Linux (${{ matrix.backend }})
-    runs-on: ubuntu-22.04
+    runs-on: *linux
    timeout-minutes: 20
    env:
      FORWARD_ONLY: 1
@@ -757,39 +755,6 @@ jobs:
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

-  testcpuopencl:
-    strategy:
-      fail-fast: false
-      matrix:
-        backend: [llvm, cpu, opencl, lvp, x86]
-
-    name: Linux (${{ matrix.backend }})
-    runs-on: ubuntu-22.04
-    timeout-minutes: 20
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: ${{ matrix.backend }}-minimal
-          deps: testing_unit
-          opencl: ${{ matrix.backend == 'opencl' && 'true' }}
-          llvm: ${{ matrix.backend != 'opencl' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'cpu' }}
-      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'DEV=CPU:LLVM' || matrix.backend == 'cpu' && 'CC=clang-20\nDEV=CPU\nCPU_COUNT=2' || matrix.backend == 'opencl' && 'DEV=CL' || matrix.backend == 'lvp' && 'DEV=CPU:LVP' || matrix.backend == 'x86' && 'DEV=CPU:X86' }}" >> $GITHUB_ENV
-      - name: Check Device.DEFAULT and print some source
-        run: |
-          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['CPU','CL'], Device.DEFAULT"
-          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
-      - name: Run pytest (${{ matrix.backend }})
-        run: python -m pytest -n=auto test/backend --durations=20
-      - name: Run TRANSCENDENTAL math
-        run: TRANSCENDENTAL=2 python -m pytest -n=auto test/backend/test_ops.py::TestOps::test_sin test/backend/test_ops.py::TestOps::test_cos test/backend/test_ops.py::TestOps::test_tan test/backend/test_ops.py::TestOps::test_exp test/backend/test_ops.py::TestOps::test_log --durations=20
-      - name: Run process replay tests
-        uses: ./.github/actions/process-replay
-
 # ****** OSX Tests ******

  testmetal:
@@ -849,84 +814,56 @@ jobs:
    - name: Run process replay tests
      uses: ./.github/actions/process-replay

-  osxwebgpu:
-    name: MacOS (WebGPU)
-    runs-on: macos-14
-    timeout-minutes: 10
+  testmacos:
+    strategy:
+      fail-fast: false
+      matrix:
+        dev:
+          - 'CPU:CLANG'
+          - 'CPU:LLVM'
+          - 'CPU:LVP'
+          - 'METAL'
+          - 'WEBGPU'
+
+    name: MacOS (DEV=${{ matrix.dev }})
+    runs-on: macos-15
+    timeout-minutes: 20
    steps:
    - name: Checkout Code
      uses: actions/checkout@v6
    - name: Setup Environment
      uses: ./.github/actions/setup-tinygrad
      with:
-        key: osx-webgpu
-        deps: testing
-        webgpu: 'true'
-    - name: Build WEBGPU Efficientnet
-      run: DEV=WEBGPU WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m examples.compile_efficientnet
-    - name: Run selected webgpu tests
-      run: DEV=WEBGPU WEBGPU_BACKEND="WGPUBackendType_Metal" python3 -m pytest -n=auto test/backend --durations=20
-    #- name: Clean npm cache
-    #  run: npm cache clean --force
-    #- name: Install Puppeteer
-    #  run: npm install puppeteer
-    # this is also flaky
-    #- name: Run WEBGPU Efficientnet
-    #  run: node test/web/test_webgpu.js
-    # this is flaky
-    #- name: Run VIZ tests as external package
-    #  run: |
-    #    mkdir $GITHUB_WORKSPACE/test_dir
-    #    cd $GITHUB_WORKSPACE/test_dir
-    #    python -m venv venv
-    #    source venv/bin/activate
-    #    pip install $GITHUB_WORKSPACE
-    #    cp $GITHUB_WORKSPACE/test/web/test_viz.js .
-    #    node test_viz.js
-    - name: Test ONNX Runner (WEBGPU)
-      run: DEV=WEBGPU python3 test/external/external_test_onnx_runner.py
-
-  osxtests:
-    strategy:
-      fail-fast: false
-      matrix:
-        backend: [metal, llvm, cpu, lvp]
-    name: MacOS (${{ matrix.backend }})
-    runs-on: macos-15
-    timeout-minutes: 20
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: macos-${{ matrix.backend }}-minimal
-          deps: testing_unit
-          llvm: ${{ matrix.backend == 'llvm' || matrix.backend == 'lvp' }}
-          mesa: ${{ matrix.backend == 'lvp' && 'cpu' }}
-      - name: Set env
-        run: printf "${{ matrix.backend == 'llvm' && 'DEV=CPU:LLVM' || matrix.backend == 'cpu' && 'DEV=CPU\nCPU_COUNT=2' || matrix.backend == 'metal' && 'DEV=METAL' || matrix.backend == 'lvp' && 'DEV=CPU:LVP' }}" >> $GITHUB_ENV
-      - name: Check Device.DEFAULT and print some source
-        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU','LVP':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
-          DEBUG=4 python3 test/test_tiny.py TestTiny.test_plus
-      - name: Run pytest (${{ matrix.backend }})
-        run: python3 -m pytest -n=auto test/backend --durations=20
-      - name: Run process replay tests
-        uses: ./.github/actions/process-replay
-      - name: Run macOS-specific unit test
-        if: matrix.backend == 'llvm'
-        run: python3 -m pytest test/unit/test_disk_tensor.py::TestDiskTensor::test_copy_to_cpu_not_truncated test/unit/test_cpu.py
+        key: macos-${{ matrix.dev }}
+        deps: testing_unit
+        python-version: '3.12'
+        llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') }}
+        mesa: ${{ contains(matrix.dev, 'LVP') && 'cpu' || 'false' }}
+        webgpu: ${{ matrix.dev == 'WEBGPU' }}
+    - name: Set env
+      run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
+    - name: Check Device.DEFAULT and print some source
+      run: |
+        python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
+        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+    - name: Run backend tests
+      run: python -m pytest -n=auto test/backend --durations=20
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay

 # ****** Windows Tests ******

-  wintests:
+  testwindows:
    strategy:
      fail-fast: false
      matrix:
-        backend: [llvm, cpu, webgpu, x86]
+        dev:
+          - 'CPU:CLANG'
+          - 'CPU:LLVM'
+          - 'CPU:X86'
+          - 'WEBGPU'

-    name: Windows (${{ matrix.backend }})
+    name: Windows (DEV=${{ matrix.dev }})
    runs-on: windows-latest
    timeout-minutes: 15
    steps:
@@ -935,25 +872,20 @@ jobs:
      - name: Setup Environment
        uses: ./.github/actions/setup-tinygrad
        with:
-          key: windows-${{ matrix.backend }}-minimal
+          key: windows-${{ matrix.dev }}-minimal
          deps: testing_unit
-          pydeps: ${{ matrix.backend == 'webgpu' && 'dawn-python' || '' }}
+          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
      - name: Set env
        shell: bash
-        run:  printf "${{ matrix.backend == 'llvm' && 'DEV=CPU:LLVM' || matrix.backend == 'cpu' && 'DEV=CPU\nCPU_COUNT=2' || matrix.backend == 'webgpu' && 'DEV=WEBGPU' || matrix.backend == 'x86' && 'DEV=CPU:X86' }}" >> $GITHUB_ENV
-      - name: Run unit tests
-        if: matrix.backend=='llvm'
-        # test_newton_schulz hits RecursionError
-        run: python -m pytest -n=auto test/unit/ --ignore=test/unit/test_disk_tensor.py --ignore=test/unit/test_tar.py --ignore=test/unit/test_linalg.py --durations=20
-      - name: Run NULL backend tests
-        if: matrix.backend=='llvm'
-        shell: bash
-        run: DEV=NULL python -m pytest -n=auto test/null/ --ignore=test/null/test_elf.py --durations=20
-      - name: Run pytest (${{ matrix.backend }})
+        run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
+      - name: Check Device.DEFAULT and print some source
        shell: bash
        run: |
-          python -c "from tinygrad import Device; assert Device.DEFAULT == {'LLVM':'CPU', 'X86':'CPU'}.get(x:='${{ matrix.backend }}'.upper(), x), Device.DEFAULT"
-          python -m pytest -n=auto test/test_tiny.py test/backend/test_ops.py --durations=20
+          python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ matrix.dev }}').device"
+          DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+      - name: Run test_tiny
+        shell: bash
+        run: python -m pytest -n=auto test/test_tiny.py --durations=20

 # ****** Compile-only Tests ******

@@ -963,7 +895,7 @@ jobs:
      matrix:
        backend: [ir3, nak]
    name: Compile-only (${{ matrix.backend }})
-    runs-on: ubuntu-24.04
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -995,7 +927,7 @@ jobs:
          python -m pytest -n=auto test/backend/test_ops.py --durations=20
  qcomclcompiletests:
    name: Compile-only (QCOM CL)
-    runs-on: ubuntu-24.04-arm
+    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'MEMBER' && 'namespace-profile-tinygrad-arm64' || 'ubuntu-24.04-arm' }}
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -1419,10 +1419,7 @@ def train_llama3():

  for p in optim.params:
    grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
-    if isinstance(p.device, tuple) and p.uop.axis is not None:
-      p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device[0]).shard_(p.device, axis=p.uop.axis).contiguous()
-    else:
-      p.grad = Tensor.zeros(p.shape, dtype=grad_dtype, device=p.device).contiguous()
+    p.grad = p.zeros_like(dtype=grad_dtype).contiguous()
  grads = [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)
@@ -1438,16 +1435,19 @@ def train_llama3():

  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts] if hasattr(model, "_fp8_grad_amax") else []
-  fp8_inv_scales = list(model._fp8_inv_scale.values())
+  fp8_inv_scales = list(model._fp8_inv_scale.values()) + list(model._fp8_next_inv_scale.values())

  from tinygrad.nn.state import get_state_dict
  model_state = get_state_dict(model)
  for wname in model._fp8_inv_scale:
    w = model_state[wname]
    w._inv_scale = model._fp8_inv_scale[wname]
+    w._next_inv_scale = model._fp8_next_inv_scale[wname]
    if optim.master_params:
      idx = next(j for j, p in enumerate(optim.params) if p is w)
-      optim.master_params[idx].assign((optim.master_params[idx] * w._inv_scale.reshape(-1, *([1]*(w.ndim-1)))).contiguous())
+      master = optim.master_params[idx]
+      inv = w._inv_scale if w._inv_scale.device == master.device else w._inv_scale.to(master.device)
+      master.assign((master * inv.reshape(-1, *([1]*(w.ndim-1)))).contiguous())

  # realize everything here
  if optim.master_params: Tensor.realize(*optim.master_params)
@@ -1498,7 +1498,7 @@ def train_llama3():
  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)
+      fake_data_np = np.random.randint(0, real_vocab_size, size=(bs, SEQLEN + 1), dtype=np.int32)
      yield Tensor(fake_data_np, device="NPY")

  def get_train_iter():
@@ -136,6 +136,7 @@ class FlatTransformer:
    w_scales = [("wqkv", s_qkv), ("wo", s_o), ("w2", s_2)]
    w_scales += [("w1", s_1), ("w3", s_3)] if SPLIT_W13 else [("w13", s_13)]
    self._fp8_inv_scale = {name: s.float().contiguous().is_param_(False) for name, s in w_scales}
+    self._fp8_next_inv_scale = {name: s.float().contiguous().is_param_(False) for name, s in w_scales}

  def lin_per_layer(self, in_features:int, out_features:int, std:float=0.02):
    if getenv("ZEROS"): w = Tensor.zeros(self.n_layers, out_features, in_features)
@@ -221,14 +222,19 @@ class FlatTransformer:
      for v in get_parameters(self): v.shard_(device, axis=None)
    else:
      # flat per-layer weights: axis 0 is n_layers, so shard axes are +1 vs per-layer Transformer
-      self.wqkv.shard_(device, axis=1).realize()          # (n_layers, out, dim) shard out
-      self.wo.shard_(device, axis=2).realize()            # (n_layers, dim, in) shard in
+      def _shard_fp8(name:str, axis:int):
+        getattr(self, name).shard_(device, axis=axis)
+        self._fp8_inv_scale[name] = self._fp8_inv_scale[name].to(device).contiguous().is_param_(False)
+        self._fp8_next_inv_scale[name] = self._fp8_next_inv_scale[name].to(device).contiguous().is_param_(False)
+        Tensor.realize(getattr(self, name), self._fp8_inv_scale[name], self._fp8_next_inv_scale[name])
+      _shard_fp8("wqkv", 1)          # (n_layers, out, dim) shard out
+      _shard_fp8("wo", 2)            # (n_layers, dim, in) shard in
      if SPLIT_W13:
-        self.w1.shard_(device, axis=1).realize()
-        self.w3.shard_(device, axis=1).realize()
+        _shard_fp8("w1", 1)
+        _shard_fp8("w3", 1)
      else:
-        self.w13.shard_(device, axis=1).realize()           # (n_layers, hidden*2, dim) shard out
-      self.w2.shard_(device, axis=2).realize()            # (n_layers, dim, hidden) shard in
+        _shard_fp8("w13", 1)         # (n_layers, hidden*2, dim) shard out
+      _shard_fp8("w2", 2)            # (n_layers, dim, hidden) shard in
      self.attention_norm.shard_(device, axis=None).realize()
      self.ffn_norm.shard_(device, axis=None).realize()
      self.norm.weight.shard_(device, axis=None).realize()
@@ -239,8 +245,6 @@ class FlatTransformer:
        for name in amax_dict:
          for i in range(len(amax_dict[name])):
            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().is_param_(False)
-      for name in self._fp8_inv_scale:
-        self._fp8_inv_scale[name] = self._fp8_inv_scale[name].to(device).contiguous().is_param_(False)

  def __call__(self, tokens:Tensor, save:bool=True):
    h = self.tok_embeddings(tokens)
@@ -322,11 +326,7 @@ if __name__ == "__main__":

  # preallocate all the grad buffers and zero them out
  grad_dtype = lambda x: dtypes.bfloat16 if x.dtype in dtypes.fp8s else x.dtype
-  def _make_grad(x):
-    if isinstance(x.device, tuple) and x.uop.axis is not None:
-      return Tensor.zeros(x.shape, dtype=grad_dtype(x), device=x.device[0]).shard_(x.device, axis=x.uop.axis).contiguous()
-    return Tensor.zeros(x.shape, dtype=grad_dtype(x), device=x.device).contiguous()
-  grads = {x:_make_grad(x) for x in state.values() if x.is_param}
+  grads = {x:x.zeros_like(dtype=grad_dtype(x)).contiguous() for x in state.values() if x.is_param}

  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts]
@@ -6,6 +6,7 @@ from tinygrad.uop.ops import UOp, Ops

 STOCHASTIC_ROUND = getenv("STOCHASTIC_ROUND", 0)
 MASTER_WEIGHTS = getenv("MASTER_WEIGHTS", 0)
+FP8_AMAX_MARGIN = getenv("FP8_AMAX_MARGIN", 1.1)

 def stochastic_round_bf16(x:Tensor) -> Tensor:
  bits = x.bitcast(dtypes.uint32)
@@ -25,7 +26,10 @@ class GradAccClipAdamW(Optimizer):
    self.m = self._new_optim_param()
    self.v = self._new_optim_param()
    self.grad_acc, self.clip_norm = grad_acc, clip_norm
-    self.master_params:list[Tensor]|None = [p.float().contiguous() for p in self.params] if MASTER_WEIGHTS and self.params[0].dtype != dtypes.float32 else None
+    if MASTER_WEIGHTS and self.params[0].dtype != dtypes.float32:
+      self.master_params:list[Tensor]|None = [p.to(self.device).float().contiguous() for p in self.params]
+    else:
+      self.master_params = None

  def fstep(self, grads:list[Tensor]):
    if self.fused:
@@ -36,7 +40,8 @@ class GradAccClipAdamW(Optimizer):
    for i, tt in enumerate(self.params): tt.assign(self._apply_update(tt, updates[i], self.master_params[i] if self.master_params else None))
    # collect inv_scale tensors attached to fp8 params (set by _apply_update)
    fp8_inv_scales = [tt._inv_scale for tt in self.params if hasattr(tt, '_inv_scale')]
-    to_realize = extra+self.params+self.buffers+(self.master_params or [])+fp8_inv_scales
+    fp8_next_inv_scales = [tt._next_inv_scale for tt in self.params if hasattr(tt, '_next_inv_scale')]
+    to_realize = extra+self.params+self.buffers+(self.master_params or [])+fp8_inv_scales+fp8_next_inv_scales

    Tensor.realize(*to_realize)
    return extra[-1]
@@ -78,13 +83,23 @@ class GradAccClipAdamW(Optimizer):
    up = up.float().shard_like(w) + self.lr.to(w.device) * wd * w.detach()
    new_w = w.detach() - up
    if master is not None: master.assign(new_w)
-    if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16: return stochastic_round_bf16(new_w)
+    # when master is offloaded to a different device than the param, results are resharded back onto the param's (sharded) device
+    offloaded = master is not None and master.device != t.device
+    if STOCHASTIC_ROUND and t.dtype == dtypes.bfloat16:
+      out = stochastic_round_bf16(new_w)
+      return out.shard_like(t) if offloaded else out
    if t.dtype in dtypes.fp8s:
      from examples.mlperf.models.flat_llama import FP8_MAX
-      amax = new_w.float().abs().max(axis=tuple(range(1, new_w.ndim))).detach()  # per-layer amax for (n_layers, out, in)
-      scale = FP8_MAX / (amax + 1e-8)
-      fp8_w = (new_w * scale.reshape(-1, *([1]*(new_w.ndim-1)))).clamp(-FP8_MAX, FP8_MAX).cast(t.dtype)
-      if hasattr(t, '_inv_scale'):
-        t._inv_scale.assign(((amax + 1e-8) / FP8_MAX).cast(t._inv_scale.dtype))
-      return fp8_w
-    return new_w.cast(t.dtype)
+      # delayed scaling: reuse previous step's inv_scale
+      t._inv_scale.assign(t._next_inv_scale)
+      inv_scale = t._inv_scale.to(new_w.device) if offloaded else t._inv_scale
+      scale = inv_scale.reciprocal().reshape(-1, *([1]*(new_w.ndim-1)))
+      scaled = (new_w * scale).clamp(-FP8_MAX, FP8_MAX)
+      ret = scaled.cast(t.dtype)
+      # update inv_scale for next step from quantized result
+      new_amax = (ret.float().abs().max(axis=tuple(range(1, ret.ndim))) * inv_scale * FP8_AMAX_MARGIN).detach()
+      new_inv = ((new_amax + 1e-8) / FP8_MAX).cast(t._inv_scale.dtype)
+      t._next_inv_scale.assign(new_inv.shard_like(t._next_inv_scale) if offloaded else new_inv)
+      return ret.shard_like(t) if offloaded else ret
+    out = new_w.cast(t.dtype)
+    return out.shard_like(t) if offloaded else out
@@ -3,4 +3,4 @@ export BENCHMARK=5
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=0 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
 SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
-python -m tinygrad.viz.cli -s "$SRC" -t
+python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
@@ -23,7 +23,7 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li

  def name_of(bu:UOp, is_out:bool) -> str:
    nonlocal n
-    if bu.op is Ops.PARAM: key, name, size = ("in", bu.arg), f"input{bu.arg}", prod(bu.shape)*bu.dtype.itemsize
+    if bu.op is Ops.PARAM: key, name, size = ("in", bu.arg.slot), f"input{bu.arg.slot}", prod(bu.shape)*bu.dtype.itemsize
    else:
      b = bu.buffer
      key, size = (id(b.base), b.offset, b.size, b.dtype), b.size*b.dtype.itemsize
@@ -10,8 +10,8 @@ from tinygrad.runtime.support.memory import BumpAllocator
 from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, graph_rewrite
 from extra.hcq2.hcq2 import HCQ2Compiled, HCQ2DeviceCtx, HCQ2LowerCtx, pm_prep_runtime, pm_lower_ops
 from extra.hcq2.hcq2 import pm_split_into_queues, pm_add_barriers, pm_add_signals
-from extra.hcq2.hcq2 import pm_bufferize, pm_lift_after, pm_resolve_patches, pm_parametrize_host_buffers
-from extra.hcq2.hcq2 import pm_finalize_submit, pm_callify, pm_calc_kernargs_sizes
+from extra.hcq2.hcq2 import pm_bufferize, pm_lift_patches_to_cmdbuf, pm_resolve_patches, pm_parametrize_host_buffers
+from extra.hcq2.hcq2 import pm_add_timeline_inc, pm_callify, pm_calc_kernargs_sizes

 # **************** insert deps ****************

@@ -28,8 +28,8 @@ pm_insert_deps = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), insert_deps)]

 pm_replace_params = PatternMatcher([
  (UPat(Ops.PARAM, name="p"), lambda ctx, p: ctx.input_addrs_uop.index(UOp.const(dtypes.int, p.arg))),
-  (UPat(Ops.BUFFER_VIEW, src=(UPat(Ops.INDEX, name="addr"),), name="bv"),
-    lambda ctx, bv, addr: addr.cast(dtypes.uint64) + UOp.const(dtypes.uint64, bv.arg[1] * bv.dtype.itemsize)),
+  (UPat(Ops.SLICE, src=(UPat(Ops.INDEX, name="addr"), UPat(Ops.CONST, dtype=dtypes.weakint, name="off")), name="bv"),
+    lambda ctx, bv, addr, off: addr.cast(dtypes.uint64) + UOp.const(dtypes.uint64, off.arg * ctx.input_uops[addr.src[1].arg].dtype.itemsize)),
 ])

 # **************** graph-only passes ****************
@@ -75,14 +75,13 @@ def drop_dead_stores(ctx:HCQ2Graph, outer:UOp) -> UOp:
  return outer.replace(src=tuple(q.replace(src=tuple(x for x in q.src if x.op is not Ops.STORE or x in live)) for q in outer.src))
 pm_drop_dead_stores = PatternMatcher([(UPat(Ops.LINEAR, src=UPat(Ops.LINEAR), name="outer"), drop_dead_stores)])

-def add_queue_sig_resets(ctx:HCQ2Graph, cf:UOp) -> UOp|None:
-  if not ctx.queue_sig_bufs or cf.arg not in ("submit_compute", "submit_copy"): return None
+def add_queue_sig_resets(ctx:HCQ2Graph, x:UOp, cmdbuf:UOp) -> UOp|None:
+  if not ctx.queue_sig_bufs or cmdbuf.tag not in ("compute", "copy"): return None
  resets = tuple((b:=UOp.from_buffer(sig)).index(UOp.const(dtypes.int, 0), dtype=b.dtype.ptr())
                 .cast(dtypes.uint64.ptr()).store(UOp.const(dtypes.uint64, 0)) for sig in ctx.queue_sig_bufs)
-  patched = cf.src[0]
-  new_patched = patched.replace(src=patched.src + resets) if patched.op is Ops.AFTER else patched.after(*resets)
-  return cf.replace(src=(new_patched,))
-pm_add_queue_sig_resets = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, name="cf"), add_queue_sig_resets)])
+  return x.replace(src=x.src + resets)
+pm_add_queue_sig_resets = PatternMatcher([(UPat(Ops.AFTER, src=(UPat(Ops.BUFFER, name="cmdbuf"),), allow_any_len=True, name="x"),
+                                           add_queue_sig_resets)])

 # **************** Graph ****************

@@ -111,18 +110,18 @@ class HCQ2Graph(GraphRunner):
    self.linear = graph_rewrite(self.linear, pm_optimize_queue_deps, ctx=self, name="hcq: optimize queue deps", walk=True)
    self.linear = graph_rewrite(self.linear, pm_drop_dead_stores, ctx=self, name="hcq: drop dead stores")
    self.linear = graph_rewrite(self.linear, pm_add_signals, ctx=self.hcq_ctx, name="hcq: add signals", walk=True)
+    self.linear = graph_rewrite(self.linear, pm_add_timeline_inc, ctx=self.hcq_ctx, name="hcq: add submit", walk=True)
    self.linear = graph_rewrite(self.linear, self.dev.pm_lower, ctx=self.hcq_ctx, name=f"hcq: encode cmdbuf {self.dev.device}", walk=True)

    graph_rewrite(self.linear, pm_calc_kernargs_sizes, ctx=(sizes:={}), name=None)
    for dev_name, sz in sizes.items():
      buf = Buffer(dev_name, sz, dtypes.uint8, options=BufferSpec(cpu_access=True), preallocate=True)
-      self.hcq_ctx.devs[dev_name] = HCQ2DeviceCtx(dev_name, UOp.from_buffer(buf, dev_name), UOp.const(dtypes.uint64, buf._buf.va_addr))
+      self.hcq_ctx.dev_ctx[dev_name] = HCQ2DeviceCtx(dev_name, UOp.from_buffer(buf, dev_name), UOp.const(dtypes.uint64, buf._buf.va_addr))

    self.linear = graph_rewrite(self.linear, pm_bufferize, ctx=self.hcq_ctx, bottom_up=True, name="realize binaries")
-    self.linear = graph_rewrite(self.linear, pm_lift_after, ctx=self.hcq_ctx, bottom_up=False, name="lift patches to root")
+    self.linear = graph_rewrite(self.linear, pm_lift_patches_to_cmdbuf, ctx=self.hcq_ctx, bottom_up=False, name="lift patches to cmdbuf")
    self.linear = graph_rewrite(self.linear, pm_resolve_patches, ctx=self.hcq_ctx, bottom_up=False, name="simplify patches")
    self.linear = graph_rewrite(self.linear, pm_add_queue_sig_resets, ctx=self, name="hcq: add queue sig resets", walk=True)
-    self.linear = graph_rewrite(self.linear, pm_finalize_submit + self.dev.pm_lower, ctx=self.hcq_ctx, bottom_up=True, name="lower submits")
    self.linear = graph_rewrite(self.linear, pm_parametrize_host_buffers, ctx=self.hcq_ctx, bottom_up=True, name="parametrize host buffers")
    self.host_call = graph_rewrite(self.linear, pm_callify, ctx=self.hcq_ctx, name="hcq: callify")

@@ -1,40 +1,35 @@
 from __future__ import annotations
 from typing import cast, Callable, TypeVar, Generic, Any, TYPE_CHECKING
-import struct, functools, time, collections
+import struct, functools, time, collections, importlib, itertools
 from dataclasses import replace
 if TYPE_CHECKING: from tinygrad.engine.realize import ExecContext
-from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, mv_address, round_up, DEBUG, dedup, all_same
+from tinygrad.helpers import DEV, getenv, select_first_inited, select_by_name, suppress_finalizing, mv_address, round_up, DEBUG, dedup, pluralize
 from tinygrad.device import Device, Buffer, BufferSpec, Compiled, LRUAllocator, MultiBuffer
-from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, track_rewrites, buffers
-from tinygrad.uop.symbolic import symbolic, symbolic_simple
+from tinygrad.uop.ops import Ops, sint, UOp, UPat, PatternMatcher, KernelInfo, graph_rewrite, track_rewrites, GroupOp
+from tinygrad.uop.symbolic import symbolic_simple, symbolic
 from tinygrad.dtype import dtypes, DType
 from dataclasses import dataclass, field
 from tinygrad.runtime.support.memory import BumpAllocator
 from tinygrad.runtime.support.hcq import MMIOInterface
 from tinygrad.renderer import Renderer, Estimates
-from tinygrad.engine.realize import to_program, track_stats, get_call_arg_uops, resolve_params
+from tinygrad.engine.realize import to_program, track_stats, get_call_arg_uops, resolve_params, pm_flatten_linear

 HCQDeviceType = TypeVar('HCQDeviceType', bound='HCQ2Compiled')

 class HCQ2Compiled(Compiled):
-  """
-  A base class for devices compatible with the HCQ (Hardware Command Queue) API.
-  """
-  timestamp_divider: float = 1000.0  # GPU timestamp counter ticks per microsecond; override per device
+  timestamp_divider: float = 1000.0 # GPU timestamp counter ticks per microsecond; override per device

-  def __init__(self, device:str, allocator:'HCQAllocator', compilers:list[type[Renderer]], runtime,
-               kernargs_size=(16 << 20), can_recover:bool=False, arch=None):
+  def __init__(self, device:str, allocator:'HCQAllocator', compilers:list[type[Renderer]], runtime, can_recover:bool=False, arch=None):
    self.device_id:int = int(device.split(":")[1]) if ":" in device else 0

-    from extra.hcq2.graph.hcq import HCQ2Graph
-    super().__init__(device, allocator, compilers, lambda *a, **kw: None, HCQ2Graph, arch=arch)
+    # default pm bufferize
+    self.pm_bufferize = PatternMatcher([
+      (UPat(Ops.BUFFER, tag="timeline_signal"), lambda ctx: ctx.timeline_signal),
+      (UPat(Ops.BUFFER, tag="timeline_value"), lambda ctx: ctx.timeline_value),
+      (UPat(Ops.BUFFER, name="b"), lambda ctx, b: Buffer(ctx.device, b.arg, b.dtype, options=BufferSpec(host=True, uncached=True, cpu_access=True))),
+    ])

-    self.kernargs_size = kernargs_size
-    self.kernargs_offset_allocator:BumpAllocator = BumpAllocator(kernargs_size, wrap=True)
-
-  @functools.cached_property
-  def kernargs_buf(self) -> Buffer:
-    return Buffer(self.device, self.kernargs_size, dtypes.uint8, options=BufferSpec(cpu_access=True), preallocate=True)
+    super().__init__(device, allocator, compilers, lambda *a, **kw: None, None, arch=arch)

  @functools.cached_property
  def timeline_signal(self) -> Buffer:
@@ -60,14 +55,6 @@ class HCQ2Compiled(Compiled):

  def device_props(self) -> dict[str,Any]: return {} # to be overridden if needed. dict keys are backend dependent.

-  def _realloc(self, oldbuf:HCQ2Buffer|None, new_size:int, options:BufferSpec|None=None, force=False) -> tuple[HCQ2Buffer, bool]:
-    if oldbuf is not None: self.allocator.free(oldbuf, oldbuf.size, options=options)
-    try: buf, realloced = self.allocator.alloc(new_size, options=options), True
-    except MemoryError:
-      if force: raise
-      buf, realloced = self.allocator.alloc(oldbuf.size if oldbuf is not None else new_size, options=options), False
-    return buf, realloced
-
  def count(self) -> int: return self.iface.count if hasattr(self, 'iface') else 1

  def _select_iface(self):
@@ -126,7 +113,7 @@ class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
  def _copy(self, dst:Buffer, src:Buffer):
    from tinygrad.engine.realize import run_linear
    su = UOp.from_buffer(src)
-    run_linear(UOp(Ops.LINEAR, dtypes.void, (su.copy_to_device(dst.device).call(UOp.from_buffer(dst), su),)), jit=True, update_stats=False)
+    run_linear(UOp(Ops.LINEAR, dtypes.void, (su.copy_to_device(dst.device).call(UOp.from_buffer(dst), su),)), update_stats=False)

  def _copyin(self, dest:HCQ2Buffer, src:memoryview):
    s = Buffer(self.dev.device, len(src), dtypes.uint8, options=BufferSpec(host=True), preallocate=True)
@@ -139,33 +126,18 @@ class HCQAllocator(LRUAllocator[HCQDeviceType], Generic[HCQDeviceType]):
    self.dev.synchronize()
    dest[:] = d._buf.cpu_view()[:len(dest)]

-  def _as_buffer(self, buf): return buf.cpu_view().mv
-
-# **************** lower context ****************
+  # def _as_buffer(self, buf): return buf.cpu_view().mv

 def unwrap_after(uop):
  while uop.op is Ops.AFTER: uop = uop.src[0]
  return uop

-@dataclass
-class HCQ2DeviceCtx:
-  device:str                       # device name; resolve to instance via Device[device]
-  kernargs_host:UOp                # UOp whose .buffer is dev.kernargs_buf (BUFFER UOp in runtime, PARAM in graph)
-  kernargs_gpu:UOp                 # va_addr const of dev.kernargs_buf
-  kernargs_allocator:BumpAllocator = field(default_factory=lambda: BumpAllocator(2 << 20, wrap=False))
-
-@dataclass
-class HCQ2LowerCtx:
-  name:str
-  inputs:list[Buffer] = field(default_factory=list)
-  holds:list[UOp] = field(default_factory=list)
-  devs:dict[str, HCQ2DeviceCtx] = field(default_factory=dict)
-
 class HCQEncoder:
-  def __init__(self, device:str): self.device, self.blob, self.patches = device, b'', []
+  def __init__(self): self.blob, self.patches = b'', []

  def get_dev_addr(self, uop:UOp) -> UOp:
-    return UOp(Ops.GETADDR, dtypes.uint64, src=(uop,)) if unwrap_after(uop).op in (Ops.BUFFER, Ops.BUFFER_VIEW, Ops.BINARY) else uop
+    if unwrap_after(uop).op not in (Ops.BUFFER, Ops.SLICE, Ops.BINARY, Ops.MSTACK, Ops.MSELECT): return uop
+    return UOp(Ops.GETADDR, dtypes.uint64, src=(uop, UOp(Ops.DEVICE, arg=self.dev.device)))

  def append(self, *data, dtype=dtypes.uint32):
    for d in data:
@@ -176,267 +148,248 @@ class HCQEncoder:

  def q(self, *values): self.append(*values)

-  def uop(self, dev:str|None=None, dtype=dtypes.uint64, tag:str|None=None) -> UOp:
-    buf = UOp.new_buffer(dev or self.device, len(self.blob), dtypes.uint8)
+  def uop(self, dev:str|tuple[str, ...], tag:str|None=None) -> UOp:
+    buf = UOp.new_buffer(dev, len(self.blob), dtypes.uint8)
    if tag: buf = buf.rtag(tag)
    blob_uop = UOp(Ops.BINARY, dtypes.void, src=(), arg=self.blob)
-    stores = [buf.index(UOp.const(dtypes.int, off)).cast(dt.ptr()).store(val.cast(dt)) for off, val, dt in self.patches]
+    stores = [buf.index(UOp.const(dtypes.int, off), dtype=buf.dtype.ptr()).cast(dt.ptr()).store(val.cast(dt)) for off, val, dt in self.patches]
    return buf.after(buf.store(blob_uop), *stores)

-# **************** prepare runtime ****************
+# *****************
+# 0. helpers

-def lower_kernargs(call:UOp, prg:UOp) -> UOp:
+HCQ_DEVS = frozenset(("AMD",))
+HCQ_P2P_DEVS = HCQ_DEVS | frozenset(("CPU",))
+
+def to_tuple(d): return d if isinstance(d, tuple) else (d,)
+
+def all_devices_in(d:Any, c:frozenset[str]) -> bool: return {x.split(":")[0] for x in to_tuple(d)} <= c
+
+# *****************
+# 1.1. prep runtimes: staging copies
+
+def _need_staging(a, b): return all_devices_in(a.device, HCQ_DEVS) and not all_devices_in(b.device, HCQ_P2P_DEVS)
+
+def stage_copy(dst:UOp, src:UOp) -> UOp|None:
+  if not (_need_staging(src, dst) or _need_staging(dst, src)): return None
+
+  stage = UOp.new_buffer("CPU", src.buffer.nbytes, dtypes.uint8)
+  return UOp(Ops.LINEAR, dtypes.void, (src.copy_to_device("CPU").call(stage, src), stage.copy_to_device(dst.device).call(dst, stage)))
+pm_insert_copy_staging = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.COPY), UPat(name="dst"), UPat(name="src"))), stage_copy)])
+
+# *****************
+# 1.2. prep runtimes: programs/kernargs
+
+@functools.cache
+def get_pm_prep_program(name:str) -> PatternMatcher|None:
+  try:
+    importlib.import_module(f'tinygrad.runtime.ops_{name.lower()}') # TODO: remove that
+    return importlib.import_module(f'extra.hcq2.ops_{name.lower()}2').pm_prep_program
+  except ImportError: return None
+
+def prep_program(call:UOp, prg:UOp) -> UOp|None:
+  dev = call.src[1].device
+  if (pm:=get_pm_prep_program(to_tuple(dev)[0].split(":")[0])) is None or (lowered:=pm.rewrite(prg)) is None: return None
+
+  data, image_bytes = lowered
+  buf = UOp.new_buffer(dev, len(image_bytes), dtypes.uint8).rtag("program")
+  blob = UOp(Ops.BINARY, dtypes.void, src=(), arg=image_bytes)
+  return call.replace(src=(prg.replace(src=(buf.after(buf.store(blob)),), arg=(data, prg.arg)),) + call.src[1:])
+
+def prep_kernargs(call:UOp, prg:UOp) -> UOp:
  data, info = prg.arg
-  dev_name = unwrap_after(prg.src[0]).src[1].arg
+  patches = [(i*dtypes.uint64.itemsize, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], UOp(Ops.DEVICE, arg=call.src[1+gi].device))),
+              dtypes.uint64) for i,gi in enumerate(info.globals)] \
+          + [(len(info.globals)*dtypes.uint64.itemsize + i*dtypes.uint32.itemsize, v, dtypes.uint32) for i,v in enumerate(info.vars)]

-  enc = HCQEncoder(dev_name)
-  for gi in info.globals: enc.append(call.src[1+gi], dtype=dtypes.uint64)
-  for v in info.vars: enc.append(v, dtype=dtypes.uint32)
+  buf = UOp.new_buffer(call.src[1].device, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
+  kernargs = buf.after(*tuple(buf.index(UOp.const(dtypes.int, o), dtype=buf.dtype.ptr()).cast(dt.ptr()).store(val.cast(dt)) for o, val, dt in patches))

-  enc.blob += b'\x00' * (data.kernargs_alloc_size - len(enc.blob)) # pad blob
-  return call.replace(src=(prg.replace(src=prg.src + (enc.uop(tag="kernargs"),), arg=(data, info)),) + call.src[1:])
+  return call.replace(src=(prg.replace(src=prg.src + (kernargs,), arg=(data, info)),) + call.src[1:])

 pm_prep_runtime = PatternMatcher([
-  # device-specific lowering of the program
-  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(), UPat(Ops.DEVICE), UPat(), UPat(), UPat(Ops.BINARY)), name="p"),), name="c", allow_any_len=True),
-    lambda c, p: c.replace(src=(Device[p.src[1].arg].pm_lower.rewrite(p),) + c.src[1:])),
+  # bind generic PROGRAM device to the call's actual dev(s), then run device-specific lowering
+  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(), UPat(), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"),),
+    name="call", allow_any_len=True), prep_program),

  # lower kernargs (PROGRAM.src[0] is now AFTER(BUFFER, COPY) — the lowered program image)
-  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(Ops.AFTER),), name="prg"),), name="call", allow_any_len=True), lower_kernargs),
+  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(Ops.AFTER),), name="prg"),), name="call", allow_any_len=True), prep_kernargs),
 ])

-# **************** lower ops ****************
-
-def _devices(buf) -> tuple[str, ...]: return tuple(b.device for b in buf.bufs) if isinstance(buf, MultiBuffer) else (buf.device,)
+# *****************
+# 2.1. lowering to hcq ir

 def lower_program(call:UOp, prg:UOp) -> UOp:
-  q = UOp(Ops.LINEAR, dtypes.void, (prg,), arg=(_devices(call.src[1].buffer), "COMPUTE"))
-  return UOp(Ops.LINEAR, dtypes.void, (q,), tag=call.tag)
+  q = UOp(Ops.LINEAR, dtypes.void, (prg,), arg=(call.src[1].device, "COMPUTE"))
+  return call.replace(src=(q,) + call.src[1:]).rtag('hcq')

-def lower_copy(call:UOp, copy:UOp) -> UOp:
+def lower_copy(call:UOp, copy:UOp) -> UOp|None:
  dst, src = call.src[1], call.src[2]
-  q = UOp(Ops.LINEAR, dtypes.void, (UOp(Ops.COPY, dtypes.void, src=(dst, src), arg=src.buffer.nbytes),), arg=(_devices(dst.buffer), "COPY"))
-  return UOp(Ops.LINEAR, dtypes.void, (q,), tag=call.tag)
+  if (hcq_dev:=next((b.device for b in (dst, src) if b.device.split(":")[0] in HCQ_DEVS), None)) is None: return None
+
+  q = UOp(Ops.LINEAR, dtypes.void, (UOp(Ops.COPY, dtypes.void, src=(dst, src), arg=src.buffer.nbytes),), arg=(hcq_dev, "COPY"))
+  return call.replace(src=(q,) + call.src[1:]).rtag('hcq')

 pm_lower_ops = PatternMatcher([
-  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(Ops.AFTER), UPat()), name="prg"),), name="call", allow_any_len=True), lower_program),
+  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, src=(UPat(Ops.AFTER), UPat(Ops.AFTER)), name="prg"),), name="call", allow_any_len=True), lower_program),
  (UPat(Ops.CALL, src=(UPat(Ops.COPY, name="copy"),), name="call", allow_any_len=True), lower_copy),
 ])

-def split_into_queues(outer:UOp) -> UOp:
-  groups:dict[tuple, list[UOp]] = collections.defaultdict(list)
-  for child in outer.src:
-    wrapper = child.src[0] if child.op is Ops.AFTER else child
-    for q in wrapper.src: groups[q.arg].extend(q.src)
-  return outer.replace(src=tuple(UOp(Ops.LINEAR, dtypes.void, tuple(cmds), arg=k) for k, cmds in groups.items()))
-pm_split_into_queues = PatternMatcher([(UPat(Ops.LINEAR, src=UPat(Ops.LINEAR, src=UPat(Ops.LINEAR)).or_after(), name="outer"), split_into_queues)])
+# *****************
+# 2.2. queue split

-def add_signals(q:UOp) -> UOp:
+# def split_into_queues(linear:UOp) -> UOp:
+#   out = []
+#   for k, grp in itertools.groupby(linear.src, lambda c: c.src[0].arg if c.op is Ops.CALL and c.src[0].op is Ops.LINEAR else None):
+#     if k is None: out.extend(grp)
+#     else:
+#       calls = list(grp)
+#       items = tuple(x for c in calls for x in c.src[0].src)
+#       args = tuple(a for c in calls for a in c.src[1:])
+#       out.append(calls[0].replace(src=(UOp(Ops.LINEAR, dtypes.void, items, arg=k),) + args))
+#   return linear.replace(src=tuple(out))
+# pm_split_into_queues = PatternMatcher([(UPat(Ops.LINEAR, name="linear"), split_into_queues)])
+
+# *****************
+# 2.3. barriers / signals / timeline inc
+
+def add_barriers(call:UOp, q:UOp) -> UOp:
+  return call.replace(src=(q.replace(src=(UOp(Ops.BARRIER, dtypes.void), *q.src)),) + call.src[1:])
+pm_add_barriers = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.LINEAR, name="q"),), name="call", allow_any_len=True), add_barriers)])
+
+def add_signals(call:UOp, q:UOp) -> UOp:
  sig = UOp.new_buffer(q.arg[0], 0x100, dtypes.uint8).rtag("timeline_signal")
  tl = UOp.new_buffer(q.arg[0], 1, dtypes.uint64).rtag("timeline_value").index(UOp.const(dtypes.int, 0))
-  return q.replace(src=(sig.wait(tl-1), *q.src, sig.store(tl)), arg=q.arg)
-pm_add_signals = PatternMatcher([(UPat(Ops.LINEAR, src=UPat(Ops.LINEAR), name="outer"),
-  lambda outer: outer.replace(src=tuple(add_signals(q) for q in outer.src)))])
+  return call.replace(src=(q.replace(src=(sig.wait(tl-1), *q.src, sig.store(tl)), arg=q.arg),) + call.src[1:])
+pm_add_signals = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.LINEAR, name="q"),), name="call", allow_any_len=True), add_signals)])

-pm_add_barriers = PatternMatcher([(UPat(Ops.LINEAR, src=UPat(Ops.LINEAR), name="outer"),
-  lambda outer: outer.replace(src=tuple(q.replace(src=(UOp(Ops.BARRIER, dtypes.void), *q.src)) for q in outer.src)))])
+# *****************
+# 3.1. encode cmdbufs

-# **************** build host program ****************
+@functools.cache
+def get_pm_lower(name:str) -> PatternMatcher|None:
+  try:
+    importlib.import_module(f'tinygrad.runtime.ops_{name.lower()}') # TODO: remove that
+    return importlib.import_module(f'extra.hcq2.ops_{name.lower()}2').pm_lower
+  except ImportError: return None

-def calc_kernargs_sizes(ctx:dict[str,int], u:UOp) -> None:
-  if u.tag != "kernargs": return
-  dev_name = u.src[1].arg
-  ctx[dev_name] = ctx.get(dev_name, 0) + round_up(u.arg, 16)
-pm_calc_kernargs_sizes = PatternMatcher([(UPat(Ops.BUFFER, name="u"), calc_kernargs_sizes)])
+def encode_cmdbuf(call:UOp, q:UOp) -> UOp|None:
+  if (pm:=get_pm_lower(to_tuple(q.arg[0])[0].split(":")[0])) is None or (encoded:=pm.rewrite(q)) is None: return None
+  return call.replace(src=(encoded,) + call.src[1:])
+pm_encode_cmdbufs = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.LINEAR, name="q"),), name="call", allow_any_len=True), encode_cmdbuf)])

-def _lower_stores(host_buf:UOp, buf_node:UOp, stores:tuple[UOp, ...]) -> list[UOp]:
-  # blob stores substitute buf_node directly; indexed patches re-target the INDEX onto host_buf with byte→element offset conversion.
-  def lower(s:UOp) -> UOp:
-    if s.src[1].op is Ops.BINARY: return s.substitute({buf_node: host_buf})
-    idx = s.src[0].src[0]
-    return s.substitute({idx: host_buf.index(UOp.const(dtypes.int, idx.src[1].arg // host_buf.dtype.base.itemsize), dtype=host_buf.dtype.ptr())})
-  return [lower(s) for s in stores]
+# *****************
+# 3.2. add timeline inc

-_program_uop_cache:dict[bytes, tuple[UOp,UOp]] = {}
-def bufferize_binary(ctx:HCQ2LowerCtx, target:UOp, buf_node:UOp) -> UOp|None:
-  dev_name, stores = buf_node.src[1].arg, target.src[1:]
+def add_timeline_inc(call:UOp, s:UOp) -> UOp:
+  tl = UOp.new_buffer(s.device, 1, dtypes.uint64).rtag("timeline_value")
+  return call.replace(src=(tl.after(s).index(UOp.const(dtypes.int, 0), dtype=tl.dtype.ptr()).store(tl.index(UOp.const(dtypes.int, 0)) + 1),) + call.src[1:])
+pm_add_timeline_inc = PatternMatcher([(UPat(Ops.CALL, tag="hcq", src=(UPat(name="s"),), name="call", allow_any_len=True), add_timeline_inc)])

-  # program
-  if buf_node.tag == "program":
-    blob = target.src[1].src[1].arg
-    if (cached:=_program_uop_cache.get(blob)) is None:
-      lib_gpu = Buffer(dev_name, round_up(len(blob), 0x1000), dtypes.uint8, options=BufferSpec(nolru=True), preallocate=True)
-      Device[dev_name].allocator._copyin(lib_gpu._buf, memoryview(bytearray(blob)))
-      Device[dev_name].synchronize()
-      cached = _program_uop_cache[blob] = (UOp.from_buffer(lib_gpu, dev_name), UOp.const(dtypes.uint64, lib_gpu._buf.va_addr))
-    lib_uop, result = cached
-    if lib_uop not in ctx.holds: ctx.holds.append(lib_uop)
-    return result
+# *****************
+# 3.3. lift patches to the command buffer (root)

-  # kernargs
-  if buf_node.tag == "kernargs":
-    dctx = ctx.devs[dev_name]
-    isz = dctx.kernargs_host.dtype.base.itemsize
-    off = dctx.kernargs_allocator.alloc(buf_node.arg, 16)
-    host_buf = UOp(Ops.BUFFER_VIEW, dctx.kernargs_host.dtype, src=(dctx.kernargs_host,), arg=(buf_node.arg // isz, off // isz))
-    return (dctx.kernargs_gpu + off).after(*_lower_stores(host_buf, buf_node, stores))
-
-  # compute/copy cmdbufs
-  if buf_node.tag in ("compute", "copy"):
-    host_buf = UOp.from_buffer(Buffer(dev_name, buf_node.arg // dtypes.uint32.itemsize, dtypes.uint32,
-                                      options=BufferSpec(cpu_access=True, nolru=True), preallocate=True), dev_name)
-    return UOp(Ops.CUSTOM_FUNCTION, dtypes.void, src=(host_buf.after(*_lower_stores(host_buf, buf_node, stores)),), arg=f"submit_{buf_node.tag}")
-
-  return None
-
-# resolve timeline_signal/timeline_value placeholders to the real device buffers
-def resolve_timeline(b:UOp) -> UOp|None:  # TODO: multi device
-  if b.tag == "timeline_signal": return UOp.from_buffer(Device[b.src[1].arg[0]].timeline_signal)
-  if b.tag == "timeline_value": return UOp.from_buffer(Device[b.src[1].arg[0]].timeline_value)
-  return None
-
-pm_bufferize = PatternMatcher([
-  (UPat(Ops.AFTER, src=(UPat(Ops.BUFFER, name="buf_node"),), allow_any_len=True, name="target"), bufferize_binary),
-  (UPat(Ops.BUFFER, name="b"), resolve_timeline),
+def lift_patches_to_cmdbuf(cmdbuf:UOp) -> UOp|None:
+  if not (patches:=dedup(u for store in cmdbuf.src[1:] for u in store.toposort() if u.op is Ops.AFTER)): return None
+  deps = tuple(d for p in patches for d in p.src[1:])
+  return cmdbuf.replace(src=cmdbuf.src + deps).substitute({p: p.src[0] for p in patches})
+pm_lift_patches_to_cmdbuf = PatternMatcher([
+  (UPat(Ops.AFTER, src=(UPat(Ops.BUFFER, tag={"compute", "copy"}),), allow_any_len=True, name="cmdbuf"), lift_patches_to_cmdbuf),
 ])

-# afters keep patches linked to their binaries. lift nested patches to root afters so symbolic can resolve them all.
-def lift_after(ctx:HCQ2LowerCtx, after:UOp) -> UOp|None:
-  if not (inners:=[u for s in after.src[1:] for u in s.toposort() if u.op is Ops.AFTER]): return None
-  subs = {i: i.src[0] for i in inners}
-  return (s:=after.substitute(subs)).replace(src=s.src[:1] + tuple(d.substitute(subs) for i in inners for d in i.src[1:]) + s.src[1:])
-pm_lift_after = PatternMatcher([(UPat(Ops.AFTER, name="after", allow_any_len=True), lift_after)])
+# *****************
+# 4. bufferize placeholders: replace placeholders with real buffers.

-def resolve_getaddr(ctx:HCQ2LowerCtx, ga:UOp, buf:UOp) -> UOp:
-  if buf not in ctx.holds: ctx.holds.append(buf)
-  return UOp.const(dtypes.uint64, buf.buffer.get_buf(buf.device).va_addr)
+def bufferize_buf(buf:UOp) -> UOp|None:
+  if buf.tag is None: return None
+  uops = tuple(UOp.from_buffer((dv:=Device[dev]).pm_bufferize.rewrite(buf, ctx=dv), dev) for dev in to_tuple(buf.src[1].arg))
+  return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, uops)
+pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, name="buf"), bufferize_buf)])

-def fold_const_store(ctx:HCQ2LowerCtx, buf:UOp, off:UOp, val:UOp) -> UOp:
-  struct.pack_into(f'<{val.dtype.fmt}', buf.buffer.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * buf.dtype.base.itemsize, val.arg)
+# *****************
+# 5.1. capture buffers reachable from each hcq call as BIND, so we don't drop their refs
+
+def hold_call_buffers(call:UOp) -> UOp|None:
+  if not (bufs:=tuple(dedup(u for u in call.src[0].toposort() if u.op is Ops.BUFFER and u not in call.src))): return None
+  return call.replace(src=call.src + (UOp(Ops.BIND, dtypes.void, src=bufs),))
+pm_hold_call_buffers = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), hold_call_buffers)])
+
+# *****************
+# 5.2. resolve patches
+
+def push_stack(op, s): return UOp(Ops.STACK, op.dtype.scalar().vec(len(s.src)),
+  tuple(op.replace(dtype=op.dtype.scalar(), src=tuple(x if y is s else y for y in op.src)) for x in s.src))
+
+def fold_blob_store(buf:UOp, blob:UOp) -> UOp:
+  for b in (buf.src if buf.op is Ops.MSTACK else (buf,)): b.buffer.ensure_allocated()._buf.cpu_view().mv.cast('B')[:len(blob.arg)] = blob.arg
  return UOp(Ops.NOOP)

-def fold_blob_store(ctx:HCQ2LowerCtx, buf:UOp, blob:UOp) -> UOp:
-  buf.buffer.ensure_allocated()._buf.cpu_view().mv.cast('B')[:len(blob.arg)] = blob.arg
+def fold_const_store(buf:UOp, off:UOp, val:UOp) -> UOp:
+  for b, v in zip((buf.src if buf.op is Ops.MSTACK else (buf,)), (val.src if val.op is Ops.STACK else (val,))):
+    struct.pack_into(f'<{v.dtype.fmt}', b.buffer.ensure_allocated()._buf.cpu_view().mv.cast('B'), off.arg * b.dtype.base.itemsize, v.arg)
  return UOp(Ops.NOOP)

-pm_resolve_patches = symbolic_simple + PatternMatcher([
-  # resolve getaddrs
-  (UPat(Ops.GETADDR, src=(UPat(Ops.BUFFER_VIEW, name="bv"),)), # getaddr(buffer_view(x)) -> offset+getaddr(x)
-    lambda ctx, bv: UOp(Ops.GETADDR, dtypes.uint64, src=(bv.src[0],)) + UOp.const(dtypes.uint64, bv.arg[1] * bv.dtype.itemsize)),
-  (UPat(Ops.GETADDR, src=(UPat(Ops.BUFFER, name="buf"),), name="ga"), resolve_getaddr), # getaddr(buffer) -> const(va_addr)
-  (UPat(Ops.GETADDR, src=(UPat.cvar("const"),)), lambda ctx, const: const), # getaddr(const) -> const
+def resolve_getaddr(buf:UOp, g:UOp) -> UOp:
+  if isinstance(b:=buf.buffer, Buffer): return UOp.const(dtypes.uint64, b.get_buf(g.src[1].arg).va_addr)
+  return UOp(Ops.STACK, dtypes.uint64.vec(len(b.bufs)), tuple(UOp.const(dtypes.uint64, x.ensure_allocated()._buf.va_addr) for x in b.bufs))

-  # write consts and binaries directly into the buffer
-  (UPat((Ops.BUFFER, Ops.BUFFER_VIEW), name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
-  (UPat((Ops.BUFFER, Ops.BUFFER_VIEW), name="buf").index(UPat.cvar("off")).or_casted().store(UPat.cvar("val")), fold_const_store),
-])
+pm_resolve_patches = PatternMatcher([
+  # multi
+  (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
+  (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),

-def parametrize_host_buffer(ctx:HCQ2LowerCtx, buf:UOp) -> UOp:
-  # register a host buffer as a launcher input and return its placeholder
-  if (b:=buf.buffer) not in ctx.inputs: ctx.inputs.append(b)
-  return UOp.placeholder((b.size,), b.dtype, ctx.inputs.index(b))
+  # getaddr
+  (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"), UPat(Ops.DEVICE, name="dev"))), # getaddr(slice(x)) -> offset+getaddr(x)
+    lambda bv, dev: UOp(Ops.GETADDR, dtypes.uint64, src=(bv.src[0], dev)) + UOp.const(dtypes.uint64, bv.src[1].arg * bv.src[0].dtype.itemsize)),
+  (UPat(Ops.GETADDR, src=(UPat({Ops.BUFFER, Ops.MSTACK, Ops.MSELECT}, name="buf"), UPat(Ops.DEVICE)), name="g"), resolve_getaddr),

-pm_parametrize_host_buffers = PatternMatcher([
-  # resolve buffer views to parametrize only root buffers
-  (UPat(Ops.INDEX, src=(UPat(Ops.BUFFER_VIEW, name="bv"), UPat.var("idx")), name="bi"),
-    lambda bv, idx, bi: bi.replace(src=(bv.src[0], idx + bv.arg[1]))),
+  # folders
+  (UPat({Ops.BUFFER, Ops.MSTACK}, name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
+  (UPat({Ops.BUFFER, Ops.MSTACK}, name="buf").index(UPat.cvar("off")).or_casted().store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))),
+    fold_const_store),
+]) + symbolic_simple

-  # parametrize host buffers
-  (UPat((Ops.BUFFER, Ops.BUFFER_VIEW), name="buf"), parametrize_host_buffer),
+# *****************
+# 6. callify hcq programs

-  # remove UNIQUE/DEVICE to dedup CONST
+pm_fixup = PatternMatcher([ # TODO: this should gone?
  (UPat(Ops.CONST, name="c"), lambda c: c.replace(src=()) if len(c.src) else None),
 ])

-def finalize_submit(cf:UOp) -> UOp|None:
-  if not cf.arg.startswith("submit_") or cf.tag is not None: return None
-  tl = UOp.from_buffer(Device['AMD'].timeline_value, "CPU")
-  done = tl.after(UOp(Ops.BARRIER, dtypes.void, src=(cf.rtag("AMD"),)))
-  return done.index(UOp.const(dtypes.int, 0), dtype=tl.dtype.ptr()).store(tl.index(UOp.const(dtypes.int, 0)) + 1)
-pm_finalize_submit = PatternMatcher([(UPat(Ops.CUSTOM_FUNCTION, name="cf"), finalize_submit)])
+def to_param(bufs:list[UOp], ref:UOp) -> UOp:
+  bufs.append(ref)
+  return UOp.placeholder((ref.buffer.size,), ref.dtype, len(bufs)-1)
+pm_to_param = PatternMatcher([(UPat({Ops.MSELECT, Ops.MSTACK, Ops.BUFFER}, name="r"), lambda ctx, r: to_param(ctx, r))])

-def hcq_callify(ctx:HCQ2LowerCtx, l:UOp) -> UOp:
-  sink = UOp.sink(*l.src, arg=KernelInfo(name=ctx.name, estimates=Estimates()), tag=1)
-  call = to_program(sink, Device["CPU"].renderer).call(*[UOp.from_buffer(b, "CPU") if isinstance(b, Buffer) else b for b in ctx.inputs])
-  return call.replace(src=call.src + (UOp(Ops.BIND, dtypes.void, src=tuple(ctx.holds)),)) if ctx.holds else call
-pm_callify = PatternMatcher([(UPat(Ops.LINEAR, name="l", allow_any_len=True), hcq_callify)])
+def parametrize_host_buffers(call:UOp) -> UOp:
+  body = graph_rewrite(call.src[0], pm_to_param, ctx=(bufs:=[]), bottom_up=True, name="parametrize host buffers")
+  return call.replace(src=(body, *bufs) + call.src[1:], tag="hcq_param")
+pm_parametrize_host_buffers = PatternMatcher([(UPat(Ops.CALL, tag="hcq", name="call"), parametrize_host_buffers)])

-# **************** schedule ****************
+def callify_hcq(call:UOp) -> UOp:
+  sink = UOp.sink(call.src[0], arg=KernelInfo(name="hcq_submit", estimates=Estimates()), tag=1)
+  return to_program(sink, Device["CPU"].renderer).call(*call.src[1:])
+pm_callify_hcq = PatternMatcher([(UPat(Ops.CALL, tag="hcq_param", name="call"), callify_hcq)])

-@track_rewrites(name=lambda linear,ast,**kw: f"hcq schedule {getattr(ast.arg, 'name', ast.op.name.lower())}")
-def hcq_schedule(linear:UOp, ast:UOp) -> UOp:
-  # runtime preparation: device-specific program, kernargs for each program
-  linear = graph_rewrite(linear, pm_prep_runtime, name="hcq: prepare runtime")
+@track_rewrites(lambda _,ret: f"HCQ Schedule {pluralize('Kernel', len(ret.src))}")
+def hcq_schedule(linear:UOp) -> UOp:
+  linear = graph_rewrite(linear, pm_insert_copy_staging + pm_flatten_linear, name="insert copy staging")
+  linear = graph_rewrite(linear, pm_prep_runtime, name="prepare runtime")

-  # lower ops into hcq style per-device operations
-  linear = graph_rewrite(linear, pm_lower_ops, name="hcq: lower ops")
+  linear = graph_rewrite(linear, pm_lower_ops, name="lower ops into hcq ir")
+  # linear = graph_rewrite(linear, pm_split_into_queues, name="split into queues")
+  linear = graph_rewrite(linear, pm_add_barriers, walk=True, name="add barriers")
+  linear = graph_rewrite(linear, pm_add_signals, walk=True, name="add signals")
+  linear = graph_rewrite(linear, pm_encode_cmdbufs, walk=True, name="encode cmdbufs")
+  linear = graph_rewrite(linear, pm_add_timeline_inc, walk=True, name="add timeline inc")
+  linear = graph_rewrite(linear, pm_lift_patches_to_cmdbuf, name="lift patches to cmdbuf", enter_calls=True)

-  # split ops into logical queues
-  linear = graph_rewrite(linear, pm_split_into_queues, name="hcq: split into queues")
+  # realize starts from here
+  linear = graph_rewrite(linear, pm_bufferize, bottom_up=True, name="bufferize placeholders", enter_calls=True)
+  linear = graph_rewrite(linear, pm_hold_call_buffers, walk=True, name="hold call buffers")
+  linear = graph_rewrite(linear, pm_resolve_patches, bottom_up=False, name="simplify patches", enter_calls=True)
+  linear = graph_rewrite(linear, pm_fixup, bottom_up=False, name="fixup", enter_calls=True)
+  linear = graph_rewrite(linear, pm_parametrize_host_buffers, name="parametrize host buffers")
+  linear = graph_rewrite(linear, pm_callify_hcq, name="callify hcq")

-  # runtime-specific lowering
-  linear = graph_rewrite(linear, pm_add_barriers, walk=True, name="hcq: add barriers")
-  linear = graph_rewrite(linear, pm_add_signals, walk=True, name="hcq: add signals")
-
-  # encode cmdbuffers
-  # TODO: remove dev
-  dev = Device["AMD"]
-  return graph_rewrite(linear, dev.pm_lower, walk=True, name="hcq: encode cmdbuf")
-
-@track_rewrites(name=lambda ctx,linear,ast,**kw: f"hcq realize {getattr(ast.arg, 'name', ast.op.name.lower())}")
-def hcq_realize(ctx:HCQ2LowerCtx, linear:UOp, ast:UOp) -> UOp:
-  # allocate lowering structs
-  graph_rewrite(linear, pm_calc_kernargs_sizes, ctx=(sizes:={}), name=None)
-
-  for dev_name, sz in sizes.items():
-    dev = Device[dev_name]
-    off = dev.kernargs_offset_allocator.alloc(sz, 16)
-    ctx.devs[dev_name] = HCQ2DeviceCtx(dev_name, UOp.from_buffer(dev.kernargs_buf.view(sz, dtypes.uint8, off), dev_name),
-                                       UOp.const(dtypes.uint64, dev.kernargs_buf.get_buf(dev_name).va_addr + off))
-
-  dev = Device['AMD']
-  linear = graph_rewrite(linear, pm_bufferize, ctx=ctx, bottom_up=True, name="realize binaries")
-  linear = graph_rewrite(linear, pm_lift_after, ctx=ctx, bottom_up=False, name="lift patches to root")
-  linear = graph_rewrite(linear, pm_resolve_patches, ctx=ctx, bottom_up=False, name="simplify patches")
-  linear = graph_rewrite(linear, pm_finalize_submit + dev.pm_lower, ctx=ctx, bottom_up=True, name="lower submits")
-  linear = graph_rewrite(linear, pm_parametrize_host_buffers, ctx=ctx, bottom_up=True, name="parametrize host buffers")
-  return graph_rewrite(linear, pm_callify, ctx=ctx, name="hcq: callify")
-
-def ensure_accessible(ctx:HCQ2LowerCtx, call:UOp, copy:UOp) -> UOp|None:
-  src_buf = call.src[2].buffer # TODO: cleanup
-  dev = call.src[1].buffer.device
-  try: src_buf.get_buf(dev)
-  except Exception:
-    (cpubuf := Buffer("CPU", src_buf.nbytes, dtypes.uint8, preallocate=True)).copyin(src_buf.ensure_allocated().as_memoryview())
-    ctx.holds.append(buf_uop:=UOp.from_buffer(cpubuf, dev))
-    return call.replace(src=call.src[:2] + (buf_uop,) + call.src[3:])
-pm_ensure_bufs_accessible = PatternMatcher([(UPat(Ops.CALL, src=(UPat(Ops.COPY, name="copy"),), name="call", allow_any_len=True), ensure_accessible)])
-
-def hcq_exec(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
-  from tinygrad.engine.realize import run_linear
-
-  if ast.src[1].arg.split(":")[0] != "AMD": return None
-
-  # TODO: this mess should gone
-  resolved_call = call.replace(src=(ast,) + tuple(resolve_params(call, ctx.input_uops)) + tuple(s for s in call.src[1:] if s.op is Ops.BIND))
-  bufs = [cast(Buffer, resolved_call.src[1+gi].buffer) for gi in ast.arg.globals] if ast.op is Ops.PROGRAM \
-    else [cast(Buffer, resolved_call.src[i].buffer) for i in range(1, len(resolved_call.src))]
-  hcq_ctx = HCQ2LowerCtx(name="submit")
-  linear = graph_rewrite(UOp(Ops.LINEAR, dtypes.void, (resolved_call,)), pm_ensure_bufs_accessible, ctx=hcq_ctx)
-
-  linear = hcq_schedule(linear, ast)
-
-  dev = Device["AMD"]
-  host_call = hcq_realize(hcq_ctx, linear, ast)
-
-  with track_stats(ctx, call, dev.device, bufs, ctx.var_vals) as tm:
-    st = time.perf_counter() if ctx.wait else 0.0
-    run_linear(UOp(Ops.LINEAR, dtypes.void, (host_call,)), var_vals=ctx.var_vals, jit=True, update_stats=DEBUG>=3)
-    if ctx.wait:
-      dev.synchronize()
-      tm[0] = time.perf_counter() - st
-  return tm[0] if tm[0] is not None else 0.0
-
-pm_hcq_exec = PatternMatcher([
-  (UPat(Ops.CALL, src=(UPat({Ops.PROGRAM, Ops.COPY}, name="ast"),), name="call", allow_any_len=True), hcq_exec),
-])
+  return linear
@@ -27,9 +27,9 @@ from tinygrad.engine.realize import get_runtime
 from tinygrad.uop.ops import Ops, UPat, PatternMatcher, graph_rewrite

 class AMDComputeQueue(HCQEncoder):
-  def __init__(self, dev:AMDDevice):
-    super().__init__(dev.device)
-    self.dev = dev
+  def __init__(self, dev:AMDDevice, devs:tuple[str, ...]|None=None):
+    super().__init__()
+    self.dev, self.devs = dev, devs or (dev.device,)
    self.pm4, self.gc, self.nbio, self.soc = dev.pm4, dev.gc, dev.nbio, dev.soc

  def pkt3(self, cmd, *vals): self.q(self.pm4.PACKET3(cmd, len(vals) - 1), *vals)
@@ -105,10 +105,12 @@ class AMDComputeQueue(HCQEncoder):

    self.acquire_mem(gli=0, gl2=0)

+    scratch_addr = self.get_dev_addr(UOp.new_buffer(self.devs, data.private_segment_size, dtypes.uint8).rtag("scratch"))
    args_addr = self.get_dev_addr(args)
+
    user_regs = []
    if data.enable_private_segment_sgpr:
-      scratch_hilo = data64_le(self.dev.scratch.va_addr)
+      scratch_hilo = data64_le(scratch_addr)
      user_regs = [scratch_hilo[0], scratch_hilo[1] | 1 << 31, 0xffffffff, 0x20c14000]
    if data.enable_dispatch_ptr: user_regs += [*data64_le(args_addr + data.kernargs_segment_size)]
    user_regs += [*data64_le(args_addr)]
@@ -116,10 +118,10 @@ class AMDComputeQueue(HCQEncoder):
    self.wreg(self.gc.regCOMPUTE_PGM_LO, *data64_le(prog_addr >> 8))
    self.wreg(self.gc.regCOMPUTE_PGM_RSRC1, data.rsrc1, data.rsrc2)
    self.wreg(self.gc.regCOMPUTE_PGM_RSRC3, data.rsrc3)
-    self.wreg(self.gc.regCOMPUTE_TMPRING_SIZE, self.dev.tmpring_size)
+    self.wreg(self.gc.regCOMPUTE_TMPRING_SIZE, self.dev.tmpring_size(data.private_segment_size))

    for xcc_id in range(self.dev.xccs):
-      scratch_base = self.dev.scratch.va_addr + (self.dev.scratch.size // self.dev.xccs * xcc_id)
+      scratch_base = scratch_addr + (data.private_segment_size // self.dev.xccs * xcc_id)
      self.wreg(self.gc.regCOMPUTE_DISPATCH_SCRATCH_BASE_LO, *data64_le(scratch_base >> 8))

    self.wreg(self.gc.regCOMPUTE_RESTART_X, 0, 0, 0)
@@ -140,28 +142,28 @@ amd_inner_pm = PatternMatcher([
  (UPat(Ops.LINEAR, src=(UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM)), UPat()), name="x"),)), lambda ctx, x: ctx.store(x)),
 ])

-def amd_lower_pm4(linear):
-  enc = AMDComputeQueue(Device["AMD"])
+def amd_lower_pm4(linear, devs):
+  enc = AMDComputeQueue(Device[devs[0]], devs)
  graph_rewrite(linear.replace(src=tuple(UOp(Ops.LINEAR, dtypes.void, (cmd,)) for cmd in linear.src)), amd_inner_pm, ctx=enc, name="amd: encode")
-  return enc.uop(dev="CPU", dtype=dtypes.void, tag="compute")
+  return enc.uop(dev=devs if len(devs) > 1 else devs[0], tag="compute")

-def amd_submit_pm4(cf):
-  # the cmdbuf to submit + the patch writes that fill it
-  cmdbuf, stores = cf.src[0].src[0], cf.src[0].src[1:]
-  size, zero = UOp.const(dtypes.uint32, cmdbuf.arg), UOp.const(dtypes.int, 0)
+def amd_submit_pm4(cmdbuf, devs):
+  size, zero = UOp.const(dtypes.uint32, cmdbuf.src[0].arg // dtypes.uint32.itemsize), UOp.const(dtypes.int, 0)

-  # the compute queue's ring and its host-side ring/write/put pointers
+  # the compute queue's ring and its host-side ring/write/put pointers (placeholders, resolved in pm_bufferize)
  q = Device['AMD'].compute_queue
-  ring, wptr, doorbell, put_ptr = (UOp.from_buffer(b, "CPU") for b in (q.ring, q.write_ptr, q.doorbell, q.put_value))
+  ring, wptr, doorbell, put_ptr = (UOp.new_buffer(devs, b.size, b.dtype).rtag(("COMPUTE:0", name))
+    for name, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))

  # place the cmdbuf at the ring's write offset, wrapping the ring
  put = put_ptr.index(zero)
  next_put = put + size.cast(put.dtype)
-  i = UOp.range(size, 0, dtype=dtypes.int, src=stores)
+  i = UOp.range(size, 0, dtype=dtypes.int, src=(cmdbuf,))
  ring_idx = ((put + i.cast(put.dtype)) % q.ring.size).cast(dtypes.int)

  # copy the cmdbuf into the ring and advance the put/write pointers
-  copy_to_ring = ring.index(ring_idx, dtype=ring.dtype.ptr()).store(cmdbuf.index(i)).end(i)
+  copy_to_ring = ring.index(ring_idx, dtype=ring.dtype.ptr()).store(
+    cmdbuf.index(i*4, dtype=cmdbuf.dtype.ptr()).cast(dtypes.uint32.ptr()).load()).end(i)
  bump_put_ptr = put_ptr.index(zero, dtype=put_ptr.dtype.ptr()).store(next_put)
  bump_wptr = wptr.index(zero, dtype=wptr.dtype.ptr()).store(next_put)

@@ -171,7 +173,7 @@ def amd_submit_pm4(cf):

 class AMDCopyQueue(HCQEncoder):
  def __init__(self, dev:AMDDevice, queue_idx=0):
-    super().__init__(dev.device)
+    super().__init__()
    self.dev = dev
    self.sdma, self.queue_idx, self.max_copy_size = dev.sdma, queue_idx, dev.max_copy_size

@@ -198,12 +200,10 @@ class AMDCopyQueue(HCQEncoder):
    self.q(self.sdma.SDMA_OP_TIMESTAMP | self.sdma.SDMA_PKT_TIMESTAMP_GET_HEADER_SUB_OP(self.sdma.SDMA_SUBOP_TIMESTAMP_GET_GLOBAL),
           *data64_le(self.get_dev_addr(x.src[0])))

-def amd_lower_sdma(linear):
-  copy = next(s for s in linear.src if s.op is Ops.COPY)
-  dev = Device[dev_name:=copy.src[0].buffer.device]
-  enc = AMDCopyQueue(dev)
+def amd_lower_sdma(linear, devs):
+  enc = AMDCopyQueue(Device[devs[0]])
  graph_rewrite(linear.replace(src=tuple(UOp(Ops.LINEAR, dtypes.void, (cmd,)) for cmd in linear.src)), amd_inner_sdma_pm, ctx=enc, name="amd: encode sdma")
-  return enc.uop(dev="CPU", dtype=dtypes.void, tag="copy")
+  return enc.uop(dev=devs if len(devs) > 1 else devs[0], tag="copy")

 amd_inner_sdma_pm = PatternMatcher([
  (UPat(Ops.LINEAR, src=(UPat(Ops.WAIT, name="x"),)), lambda ctx, x: ctx.wait(x)),
@@ -213,14 +213,14 @@ amd_inner_sdma_pm = PatternMatcher([
  (UPat(Ops.LINEAR, src=(UPat(Ops.STORE, src=(UPat((Ops.BUFFER, Ops.PARAM)), UPat()), name="x"),)), lambda ctx, x: ctx.store(x)),
 ])

-def amd_submit_sdma(cf):
+def amd_submit_sdma(cmdbuf, devs):
  # the cmdbuf to submit + the patch writes that fill it
-  cmdbuf, stores = cf.src[0].src[0], cf.src[0].src[1:]
-  size_dw, zero = cmdbuf.arg, UOp.const(dtypes.int, 0)
+  size_dw, zero = cmdbuf.src[0].arg // dtypes.uint32.itemsize, UOp.const(dtypes.int, 0)

  # the sdma queue's ring and its host-side ring/write/put pointers
  q = Device['AMD'].sdma_queue(0)
-  ring, wptr, doorbell, put_ptr = (UOp.from_buffer(b, "CPU") for b in (q.ring, q.write_ptr, q.doorbell, q.put_value))
+  ring, wptr, doorbell, put_ptr = (UOp.new_buffer(devs, b.size, b.dtype).rtag(("SDMA:0", name))
+    for name, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))

  # sdma needs the cmdbuf contiguous: if it won't fit before the ring end, restart at 0 and zero the tail
  put_b = put_ptr.index(zero)
@@ -230,10 +230,11 @@ def amd_submit_sdma(cf):
  zero_amt_dw = (1 - fits) * (q.ring.size - tail_off_dw)

  # zero the wrapped tail, then copy the cmdbuf into the ring
-  zi = UOp.range(zero_amt_dw, 0, dtype=dtypes.int, src=stores)
+  zi = UOp.range(zero_amt_dw, 0, dtype=dtypes.int, src=(cmdbuf,))
  zero_tail = ring.index(tail_off_dw + zi, dtype=ring.dtype.ptr()).store(UOp.const(dtypes.uint32, 0)).end(zi)
-  i = UOp.range(UOp.const(dtypes.int, size_dw), 0, dtype=dtypes.int, src=stores)
-  copy_to_ring = ring.index(start_dw + i, dtype=ring.dtype.ptr()).store(cmdbuf.index(i)).end(i)
+  i = UOp.range(UOp.const(dtypes.int, size_dw), 0, dtype=dtypes.int, src=(cmdbuf,))
+  copy_to_ring = ring.index(start_dw + i, dtype=ring.dtype.ptr()).store(
+    cmdbuf.index(i*4, dtype=cmdbuf.dtype.ptr()).cast(dtypes.uint32.ptr()).load()).end(i)

  # advance the put/write pointers past the zeroed tail and the cmdbuf
  next_put_b = put_b + ((zero_amt_dw + size_dw) * 4).cast(put_b.dtype)
@@ -247,13 +248,13 @@ def amd_submit_sdma(cf):
@dataclass(frozen=True)
 class AMDProgramData:
  entry_point_offset:int; rsrc1:int; rsrc2:int; rsrc3:int; wave32:bool
-  kernargs_segment_size:int; kernargs_alloc_size:int
+  private_segment_size:int; kernargs_segment_size:int; kernargs_alloc_size:int
  enable_dispatch_ptr:int; enable_private_segment_sgpr:int

 _amd_program_cache:dict[tuple[bytes,str], tuple[AMDProgramData,bytes]] = {}

 def amd_build_program(prg:UOp) -> UOp:
-  dev = Device[prg.src[1].arg]
+  dev = Device[prg.src[1].arg] # TODO: rm this
  if (cached:=_amd_program_cache.get(key:=(lib:=prg.src[4].arg, dev.device))) is None:
    image, sections, relocs = elf_loader(lib)
    rodata = next(sh.header.sh_addr for sh in sections if sh.name == ".rodata")
@@ -263,29 +264,29 @@ def amd_build_program(prg:UOp) -> UOp:
    desc = amdgpu_kd.llvm_amdhsa_kernel_descriptor_t.from_buffer_copy(bytes(image[rodata:rodata+ctypes.sizeof(amdgpu_kd.llvm_amdhsa_kernel_descriptor_t)]))
    if (lds:=((desc.group_segment_fixed_size+511)//512)&0x1FF) > (dev.iface.props['lds_size_in_kb']*1024)//512:
      raise RuntimeError("Too many resources requested: group_segment_size")
-    dev._ensure_has_local_memory(desc.private_segment_fixed_size)
    edp = desc.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_DISPATCH_PTR
    cached = _amd_program_cache[key] = (AMDProgramData(
      entry_point_offset=rodata + desc.kernel_code_entry_byte_offset,
      rsrc1=desc.compute_pgm_rsrc1 | ((1<<20) if dev.target[0]==11 else 0),  # priv=1 on gfx11 for cwsr
      rsrc2=desc.compute_pgm_rsrc2 | (lds<<15), rsrc3=desc.compute_pgm_rsrc3,
      wave32=bool(desc.kernel_code_properties & 0x400),
+      private_segment_size=desc.private_segment_fixed_size,
      kernargs_segment_size=desc.kernarg_size,
      kernargs_alloc_size=desc.kernarg_size + (ctypes.sizeof(hsa.hsa_kernel_dispatch_packet_t) if edp else 0),
      enable_dispatch_ptr=edp,
-      enable_private_segment_sgpr=desc.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER,
-    ), bytes(image))
-  data, image_bytes = cached
-  buf_uop = UOp.new_buffer(dev.device, len(image_bytes), dtypes.uint8).rtag("program")
-  blob_uop = UOp(Ops.BINARY, dtypes.void, src=(), arg=image_bytes)
-  return prg.replace(src=(buf_uop.after(buf_uop.store(blob_uop)),), arg=(data, prg.arg))
+      enable_private_segment_sgpr=desc.kernel_code_properties & hsa.AMD_KERNEL_CODE_PROPERTIES_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER), bytes(image))
+  return cached
+
+pm_prep_program = PatternMatcher([
+  (UPat(Ops.PROGRAM, src=(UPat(), UPat(Ops.DEVICE, arg="AMD"), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"), amd_build_program),
+])

 class AMDAllocator(HCQAllocator['AMDDevice']):
  def __init__(self, dev:AMDDevice):
    super().__init__(dev, supports_copy_from_disk=dev.has_sdma_queue, supports_transfer=dev.has_sdma_queue and not dev.is_usb())

  def _alloc(self, size:int, options:BufferSpec) -> HCQ2Buffer:
-    return self.dev.iface.alloc(size, host=True, uncached=options.uncached, cpu_access=True)
+    return self.dev.iface.alloc(size, host=options.host, uncached=options.uncached, cpu_access=options.cpu_access or not self.dev.has_sdma_queue)

  def _do_free(self, opaque, options:BufferSpec): self.dev.iface.free(opaque)

@@ -372,19 +373,18 @@ class PCIIface(PCIIfaceBase):

 def _mock(iface, name=None): return type(name or f"MOCK{iface.__name__}", (iface,), {})

-def encode_queues(outer:UOp) -> UOp:
-  return outer.replace(src=tuple(amd_lower_pm4(q) if q.arg[1] == "COMPUTE" else amd_lower_sdma(q) for q in outer.src))
+def encode_queue(q:UOp) -> UOp|None:
+  if not (isinstance(q.arg, tuple) and len(q.arg) == 2 and q.arg[1] in ("COMPUTE", "COPY")): return None
+  devs = (q.arg[0],) if isinstance(q.arg[0], str) else q.arg[0]  # TODO: make this prettier
+  return amd_submit_pm4(amd_lower_pm4(q, devs), devs) if q.arg[1] == "COMPUTE" else amd_submit_sdma(amd_lower_sdma(q, devs), devs)
+
+pm_lower = PatternMatcher([
+  (UPat(Ops.LINEAR, name="q"), encode_queue),
+])

 class AMDDevice(HCQ2Compiled):
  timestamp_divider = 100.0  # AMD GPU clock: ticks/us

-  pm_lower = PatternMatcher([
-    (UPat(Ops.PROGRAM, src=(UPat(), UPat(), UPat(), UPat(), UPat(Ops.BINARY)), name="prg"), amd_build_program),
-    (UPat(Ops.LINEAR, src=UPat(Ops.LINEAR), name="outer"), encode_queues),
-    (UPat(Ops.CUSTOM_FUNCTION, arg="submit_compute", name="cf"), amd_submit_pm4),
-    (UPat(Ops.CUSTOM_FUNCTION, arg="submit_copy", name="cf"), amd_submit_sdma),
-  ])
-
  ifaces = [PCIIface]

  def is_am(self) -> bool: return isinstance(self.iface, (PCIIface,))
@@ -423,14 +423,13 @@ class AMDDevice(HCQ2Compiled):

    self.max_copy_size = 0x40000000 if self.iface.ip_versions[am.SDMA0_HWIP][0] >= 5 else 0x400000
    self.sdma_queues:dict = {}
-    self.has_sdma_queue = self.sdma_queue(0) is not None
+    self.has_sdma_queue = True # self.sdma_queue(0) is not None, TODO: think of this

-    super().__init__(device, AMDAllocator(self), [HIPRenderer, AMDLLVMRenderer, HIPCCRenderer], None,
-                     kernargs_size=16 << 20, can_recover=self.is_am(), arch=self.arch)
+    super().__init__(device, AMDAllocator(self), [HIPRenderer, AMDLLVMRenderer, HIPCCRenderer], None, can_recover=self.is_am(), arch=self.arch)

    # Scratch setup
    self.max_private_segment_size = 0
-    self._ensure_has_local_memory(128) # set default scratch size to 128 bytes per thread
+    self.pm_bufferize = PatternMatcher([(UPat(Ops.BUFFER, tag="scratch", name="b"), lambda ctx, b: ctx.scratch_buffer(b.arg))]) + self.pm_bufferize

    self.pmc_enabled:bool = PROFILE > 0 and PMC > 0
    if self.pmc_enabled:
@@ -457,19 +456,6 @@ class AMDDevice(HCQ2Compiled):
      self.sqtt_wptrs = self.allocator.alloc(round_up(self.se_cnt * self.xccs * 4, 0x1000), BufferSpec(cpu_access=True, nolru=True))
      self.sqtt_next_cmd_id = itertools.count(0)

-  @functools.cached_property
-  def compute_queue(self) -> AMDQueueDesc:
-    # https://gitlab.freedesktop.org/agd5f/linux/-/blob/a1fc9f584c4aaf8bc1ebfa459fc57a3f26a290d8/drivers/gpu/drm/amd/amdkfd/kfd_queue.c#L391
-    sgrp_size_per_cu, hwreg_size_per_cu = 0x4000, 0x1000
-    lds_size_per_cu = self.iface.props["lds_size_in_kb"] << 10 if self.target[:2] == (9,5) else 0x10000
-    vgpr_size_per_cu = 0x60000 if self.target in {(11,0,0), (11,0,1), (11,5,1), (12,0,0), (12,0,1)} else 0x80000 if self.target[0] == 9 else 0x40000
-    wg_data_size = round_up((vgpr_size_per_cu + sgrp_size_per_cu + lds_size_per_cu + hwreg_size_per_cu) * self.cu_cnt, mmap.PAGESIZE)
-    ctl_stack_size = round_up((12 if self.target[0] != 9 else 8) * self.wave_cnt + 8 + 40, mmap.PAGESIZE)
-    return self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL if self.is_aql else kfd.KFD_IOC_QUEUE_TYPE_COMPUTE,
-      0x2000 if self.is_usb() else (16 << 20), eop_buffer_size=0x1000,
-      ctx_save_restore_size=0 if self.is_am() else wg_data_size + ctl_stack_size, ctl_stack_size=ctl_stack_size,
-      debug_memory_size=round_up(self.wave_cnt * 32, 64))
-
  def create_queue(self, queue_type, ring_size, ctx_save_restore_size=0, eop_buffer_size=0, ctl_stack_size=0, debug_memory_size=0, idx=0):
    ring = self.iface.alloc(ring_size, uncached=True, cpu_access=True)
    gart = self.iface.alloc(0x100, uncached=True, cpu_access=True)
@@ -485,9 +471,29 @@ class AMDDevice(HCQ2Compiled):
    cwsr_buffer = self.iface.alloc(cwsr_buffer_size) if ctx_save_restore_size else None
    eop_buffer = self.iface.alloc(eop_buffer_size) if eop_buffer_size else None

-    return (self.iface.create_queue(queue_type, ring, gart, rptr=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
-            wptr=getattr(hsa.amd_queue_t, 'write_dispatch_id').offset, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer,
-            ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size, idx=idx))
+    queue = (self.iface.create_queue(queue_type, ring, gart, rptr=getattr(hsa.amd_queue_t, 'read_dispatch_id').offset,
+             wptr=getattr(hsa.amd_queue_t, 'write_dispatch_id').offset, eop_buffer=eop_buffer, cwsr_buffer=cwsr_buffer,
+             ctx_save_restore_size=ctx_save_restore_size, ctl_stack_size=ctl_stack_size, idx=idx))
+
+    qname = f"{'SDMA' if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA else 'COMPUTE'}:{idx}"
+    self.pm_bufferize = PatternMatcher([
+      (UPat(Ops.BUFFER, tag={(qname, name)}), lambda ctx, b=getattr(queue, name): b) for name in ["ring", "write_ptr", "doorbell", "put_value"]
+    ]) + self.pm_bufferize
+
+    return queue
+
+  @functools.cached_property
+  def compute_queue(self) -> AMDQueueDesc:
+    # https://gitlab.freedesktop.org/agd5f/linux/-/blob/a1fc9f584c4aaf8bc1ebfa459fc57a3f26a290d8/drivers/gpu/drm/amd/amdkfd/kfd_queue.c#L391
+    sgrp_size_per_cu, hwreg_size_per_cu = 0x4000, 0x1000
+    lds_size_per_cu = self.iface.props["lds_size_in_kb"] << 10 if self.target[:2] == (9,5) else 0x10000
+    vgpr_size_per_cu = 0x60000 if self.target in {(11,0,0), (11,0,1), (11,5,1), (12,0,0), (12,0,1)} else 0x80000 if self.target[0] == 9 else 0x40000
+    wg_data_size = round_up((vgpr_size_per_cu + sgrp_size_per_cu + lds_size_per_cu + hwreg_size_per_cu) * self.cu_cnt, mmap.PAGESIZE)
+    ctl_stack_size = round_up((12 if self.target[0] != 9 else 8) * self.wave_cnt + 8 + 40, mmap.PAGESIZE)
+    return self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE_AQL if self.is_aql else kfd.KFD_IOC_QUEUE_TYPE_COMPUTE,
+      0x2000 if self.is_usb() else (16 << 20), eop_buffer_size=0x1000,
+      ctx_save_restore_size=0 if self.is_am() else wg_data_size + ctl_stack_size, ctl_stack_size=ctl_stack_size,
+      debug_memory_size=round_up(self.wave_cnt * 32, 64))

  def sdma_queue(self, idx:int):
    if getenv("AMD_DISABLE_SDMA"): return None
@@ -496,38 +502,49 @@ class AMDDevice(HCQ2Compiled):
      self.sdma_queues[idx] = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_SDMA, 0x200 if self.is_usb() else (16 << 20), idx=idx)
    return self.sdma_queues.get(idx, None)

-  def _ensure_has_local_memory(self, private_segment_size):
-    if self.max_private_segment_size >= private_segment_size: return
+  def tmpring_size(self, private_segment_size):
+    private_segment_size = max(private_segment_size, 128)

    lanes_per_wave = 64 # wave64
    mem_alignment_size = 256 if self.target[0] != 9 else 1024
    size_per_thread = round_up(private_segment_size, mem_alignment_size // lanes_per_wave)
    size_per_xcc = size_per_thread * lanes_per_wave * self.iface.props['max_slots_scratch_cu'] * self.cu_cnt
-    self.scratch, ok = self._realloc(getattr(self, 'scratch', None), size_per_xcc * self.xccs)
-    if ok:
-      # NOTE: xcc logic is correct only for GFX9.
-      max_scratch_waves = self.cu_cnt * self.iface.props['max_slots_scratch_cu'] * self.xccs
-      wave_scratch = ceildiv(lanes_per_wave * size_per_thread, mem_alignment_size)
-      num_waves = (size_per_xcc // (wave_scratch * mem_alignment_size)) // (self.se_cnt if self.target[0] != 9 else 1)

-      tmpring_t = getattr(hsa, f'union_COMPUTE_TMPRING_SIZE{"_GFX"+str(self.target[0]) if self.target[0] != 9 else ""}_bitfields')
-      self.tmpring_size = int.from_bytes(tmpring_t(WAVES=min(num_waves, max_scratch_waves), WAVESIZE=wave_scratch), 'little')
+    # NOTE: xcc logic is correct only for GFX9.
+    max_scratch_waves = self.cu_cnt * self.iface.props['max_slots_scratch_cu'] * self.xccs
+    wave_scratch = ceildiv(lanes_per_wave * size_per_thread, mem_alignment_size)
+    num_waves = (size_per_xcc // (wave_scratch * mem_alignment_size)) // (self.se_cnt if self.target[0] != 9 else 1)
+
+    tmpring_t = getattr(hsa, f'union_COMPUTE_TMPRING_SIZE{"_GFX"+str(self.target[0]) if self.target[0] != 9 else ""}_bitfields')
+    tmpring = int.from_bytes(tmpring_t(WAVES=min(num_waves, max_scratch_waves), WAVESIZE=wave_scratch), 'little')
+
+    if hasattr(self, 'aql_desc'):
+      gfx9_rsrc = {'NUM_FORMAT':hsa.BUF_NUM_FORMAT_UINT, 'DATA_FORMAT':hsa.BUF_DATA_FORMAT_32, 'ELEMENT_SIZE':1, 'INDEX_STRIDE':3}
+      rsrc = {'DST_SEL_X':hsa.SQ_SEL_X, 'DST_SEL_Y':hsa.SQ_SEL_Y, 'DST_SEL_Z':hsa.SQ_SEL_Z, 'DST_SEL_W':hsa.SQ_SEL_W, 'ADD_TID_ENABLE':1,
+              'TYPE':hsa.SQ_RSRC_BUF, **(gfx9_rsrc if self.target[0] == 9 else {'FORMAT':hsa.BUF_FORMAT_32_UINT, 'OOB_SELECT':2})}
+      rsrc1_t = getattr(hsa, f'union_SQ_BUF_RSRC_WORD1{"_GFX11" if self.target[0] != 9 else ""}_bitfields')
+      rsrc3_t = getattr(hsa, f'union_SQ_BUF_RSRC_WORD3{"_GFX"+str(self.target[0]) if self.target[0] != 9 else ""}_bitfields')
+
+      self.aql_desc.scratch_backing_memory_location = int(self.scratch.get_buf().va_addr)
+      self.aql_desc.scratch_wave64_lane_byte_size = self.max_private_segment_size * lanes_per_wave // 64
+      self.aql_desc.scratch_resource_descriptor[:] = [lo32(self.scratch.get_buf().va_addr),
+        int.from_bytes(rsrc1_t(BASE_ADDRESS_HI=hi32(self.scratch.get_buf().va_addr), SWIZZLE_ENABLE=1), 'little'),
+        lo32(size_per_xcc), int.from_bytes(bytes(rsrc3_t(**rsrc)), 'little')]
+      self.aql_desc.compute_tmpring_size = tmpring
+      self.aql_gart.cpu_view()[:ctypes.sizeof(self.aql_desc)] = bytes(self.aql_desc)
+
+    return tmpring
+
+  def scratch_buffer(self, private_segment_size):
+    private_segment_size = max(private_segment_size, 128)
+    if self.max_private_segment_size < private_segment_size:
+      lanes_per_wave = 64 # wave64
+      mem_alignment_size = 256 if self.target[0] != 9 else 1024
+      size_per_thread = round_up(private_segment_size, mem_alignment_size // lanes_per_wave)
+      size_per_xcc = size_per_thread * lanes_per_wave * self.iface.props['max_slots_scratch_cu'] * self.cu_cnt
+      self.scratch = Buffer(self.device, size_per_xcc * self.xccs, dtypes.uint8, options=BufferSpec(nolru=True), preallocate=True)
      self.max_private_segment_size = private_segment_size
-
-      if hasattr(self, 'aql_desc'):
-        gfx9_rsrc = {'NUM_FORMAT':hsa.BUF_NUM_FORMAT_UINT, 'DATA_FORMAT':hsa.BUF_DATA_FORMAT_32, 'ELEMENT_SIZE':1, 'INDEX_STRIDE':3}
-        rsrc = {'DST_SEL_X':hsa.SQ_SEL_X, 'DST_SEL_Y':hsa.SQ_SEL_Y, 'DST_SEL_Z':hsa.SQ_SEL_Z, 'DST_SEL_W':hsa.SQ_SEL_W, 'ADD_TID_ENABLE':1,
-                'TYPE':hsa.SQ_RSRC_BUF, **(gfx9_rsrc if self.target[0] == 9 else {'FORMAT':hsa.BUF_FORMAT_32_UINT, 'OOB_SELECT':2})}
-        rsrc1_t = getattr(hsa, f'union_SQ_BUF_RSRC_WORD1{"_GFX11" if self.target[0] != 9 else ""}_bitfields')
-        rsrc3_t = getattr(hsa, f'union_SQ_BUF_RSRC_WORD3{"_GFX"+str(self.target[0]) if self.target[0] != 9 else ""}_bitfields')
-
-        self.aql_desc.scratch_backing_memory_location = int(self.scratch.va_addr)
-        self.aql_desc.scratch_wave64_lane_byte_size = self.max_private_segment_size * lanes_per_wave // 64
-        self.aql_desc.scratch_resource_descriptor[:] = [lo32(self.scratch.va_addr),
-          int.from_bytes(rsrc1_t(BASE_ADDRESS_HI=hi32(self.scratch.va_addr), SWIZZLE_ENABLE=1), 'little'),
-          lo32(size_per_xcc), int.from_bytes(bytes(rsrc3_t(**rsrc)), 'little')]
-        self.aql_desc.compute_tmpring_size = self.tmpring_size
-        self.aql_gart.cpu_view()[:ctypes.sizeof(self.aql_desc)] = bytes(self.aql_desc)
+    return self.scratch

  def on_device_hang(self): self.iface.on_device_hang()

@@ -5,19 +5,19 @@ from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.renderer import Estimates
 from extra.llama_kernels import FP8_MAX, NUM_WG, THREADS_PER_WG, compile_cpp, alloc_like, alloc_local, scalar_amax, dname_of

-# module-level mailbox: grad_xw13 UOp -> (grad_xw13_fp8 UOp, inv_scale UOp, new_amax UOp, store_effect)
+# module-level mailbox: grad_xw13 UOp -> (grad_xw13_fp8 UOp, inv_scale UOp)
 # lets cdna_asm_gemm's bwd reuse the fp8 companion produced by the fused silu_mul bwd kernel
 # instead of doing a redundant bf16 -> fp8 quantize.
-_grad_fp8_mailbox:dict = {}
+_grad_fp8_mailbox:dict[UOp, tuple[UOp, UOp]] = {}

@functools.cache
-def _custom_fused_bwd_w13(grad_xw13:UOp, grad_xw13_fp8:UOp, grad_amax_buf:UOp,
+def _custom_fused_bwd_w13(grad_xw13_fp8:UOp, grad_amax_buf:UOp,
                          xw13:UOp, grad_x2:UOp, amax_state:UOp, grad_amax_state:UOp, dname:str) -> UOp:
  hidden = xw13.shape[2] // 2
  n_elems = xw13.shape[0] * xw13.shape[1] * hidden
  threads, workgroups = UOp.special(THREADS_PER_WG, "lidx0"), UOp.special(NUM_WG, "gidx0")
-  mem = n_elems * 2 * 5 + n_elems * 2 + NUM_WG * 4 + 4
-  sink = UOp.sink(grad_xw13.base, grad_xw13_fp8.base, grad_amax_buf.base,
+  mem = n_elems * 2 * 3 + n_elems * 2 + NUM_WG * 4 + 4
+  sink = UOp.sink(grad_xw13_fp8.base, grad_amax_buf.base,
                  xw13.base, grad_x2.base, amax_state.base, grad_amax_state.base, threads, workgroups,
                  arg=KernelInfo(f"fused_silu_mul_bwd_w13_{n_elems}", estimates=Estimates(ops=10*n_elems, mem=mem)))
  src, lib = compile_cpp(pathlib.Path(__file__).parent, "cast_amax_bwd_w13.cpp", n_elems, hidden)
@@ -41,23 +41,23 @@ def _fused_quantize_bwd_w13(gradient:UOp, kernel:UOp):
  _, _, xw13, amax_state, grad_amax_state = kernel.src[1:]
  device = xw13.device
  axis = xw13.axis if isinstance(device, tuple) else None
-  grad_xw13     = alloc_like(xw13.shape, dtypes.bfloat16, device, axis)
  grad_xw13_fp8 = alloc_like(xw13.shape, dtypes.fp8e4m3,  device, axis)
  grad_amax_buf = alloc_local((NUM_WG,), dtypes.float32,  device, axis)
  grad_amax_state_t = Tensor(grad_amax_state, device=device)
  fxn = functools.partial(_custom_fused_bwd_w13, dname=dname_of(device))
-  grad_xw13, grad_xw13_fp8, grad_amax_buf, *_ = Tensor.custom_kernel(
-    grad_xw13, grad_xw13_fp8, grad_amax_buf,
+  grad_xw13_fp8, grad_amax_buf, *_ = Tensor.custom_kernel(
+    grad_xw13_fp8, grad_amax_buf,
    Tensor(xw13, device=device), Tensor(gradient, device=device).cast(dtypes.bfloat16),
    Tensor(amax_state, device=device), grad_amax_state_t, fxn=fxn)
+  grad_xw13_uop = grad_xw13_fp8.uop.cast(dtypes.bfloat16)
  inv_scale = (grad_amax_state_t.float() + 1e-8) / FP8_MAX
  new_grad_amax = scalar_amax(grad_amax_buf)
  store_effect = grad_amax_state_t.uop.store(new_grad_amax.uop)
  assert grad_xw13_fp8.uop.op is Ops.AFTER, f"expected AFTER, got {grad_xw13_fp8.uop.op}"
  grad_xw13_fp8_uop = grad_xw13_fp8.uop.replace(src=grad_xw13_fp8.uop.src + (store_effect,))
  # Stash fp8 companion for cdna_asm_gemm's bwd to attach to grad_a.
-  _grad_fp8_mailbox[grad_xw13.uop] = (grad_xw13_fp8_uop, inv_scale.uop)
-  return (None, None, grad_xw13.uop, None, None)
+  _grad_fp8_mailbox[grad_xw13_uop] = (grad_xw13_fp8_uop, inv_scale.uop)
+  return (None, None, grad_xw13_uop, None, None)

 def fused_quantize_fp8_w13(xw13:Tensor, amax_state:Tensor, fp8_dtype, grad_amax_state:Tensor) -> tuple[Tensor, Tensor, Tensor]:
  # NOTE: silu(xw1)*xw3 -> fp8 + amax over fused xw13 layout. Returns (fp8, inv_scale, new_amax)
@@ -21,15 +21,13 @@ constexpr float FP8_MAX = 448.0f;
 static_assert(N_ELEMS % VEC == 0, "N_ELEMS must be divisible by VEC");
 static_assert(HIDDEN % VEC == 0, "HIDDEN must be divisible by VEC");

-// fused silu*mul backward, three outputs in a single HBM pass:
-//   1) bf16 grad_xw13      — consumed by downstream bf16 autograd chain
-//   2) fp8  grad_xw13_fp8  — delayed-scale quantize using grad_amax_state (mailbox to matmul bwd)
-//   3) fp32 grad_amax_buf  — per-WG partial |grad_xw13|, reduced into next step's grad_amax_state
+// fused silu*mul backward, two outputs in a single HBM pass:
+//   1) fp8  grad_xw13_fp8  — delayed-scale quantize using grad_amax_state (mailbox to matmul bwd)
+//   2) fp32 grad_amax_buf  — per-WG partial |grad_xw13|, reduced into next step's grad_amax_state
 // grad_amax_state is read for the fp8 scale. The store of new_grad_amax into grad_amax_state's
 // buffer is built in Python as a separate effect and threaded into grad_a via .after(store).
 extern "C" __global__ __launch_bounds__(THREADS_PER_WG) void
 fused_silu_mul_bwd_w13(
-    __hip_bfloat16*       __restrict__ grad_xw13_out,        // bf16, 2*N_ELEMS
    __hip_fp8_storage_t*  __restrict__ grad_xw13_fp8_out,    // fp8,  2*N_ELEMS
    float*                __restrict__ grad_amax_buf,        // fp32, NUM_WG per-WG partials
    const __hip_bfloat16* __restrict__ xw13,                 // bf16, 2*N_ELEMS
@@ -62,7 +60,6 @@ fused_silu_mul_bwd_w13(
    const __hip_bfloat16 *x3 = reinterpret_cast<const __hip_bfloat16*>(&x3_raw);
    const __hip_bfloat16 *gv = reinterpret_cast<const __hip_bfloat16*>(&g_raw);

-    __hip_bfloat16 out1[VEC], out3[VEC];
    __hip_fp8_storage_t fp8_1[VEC], fp8_3[VEC];
    #pragma unroll
    for (int i = 0; i < VEC; i++) {
@@ -75,15 +72,11 @@ fused_silu_mul_bwd_w13(
      const float gs = fg * scale;
      const float g1 = gs * silu_prime * f3;
      const float g3 = gs * silu;
-      out1[i] = static_cast<__hip_bfloat16>(g1);
-      out3[i] = static_cast<__hip_bfloat16>(g3);
      local_max = fmaxf(local_max, fmaxf(fabsf(g1), fabsf(g3)));
      fp8_1[i] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, g1 * g_scale)), __HIP_SATFINITE, __HIP_E4M3);
      fp8_3[i] = __hip_cvt_float_to_fp8(fmaxf(-FP8_MAX, fminf(FP8_MAX, g3 * g_scale)), __HIP_SATFINITE, __HIP_E4M3);
    }

-    *reinterpret_cast<float4*>(&grad_xw13_out[xw1_off]) = *reinterpret_cast<float4*>(out1);
-    *reinterpret_cast<float4*>(&grad_xw13_out[xw3_off]) = *reinterpret_cast<float4*>(out3);
    *reinterpret_cast<uint64_t*>(&grad_xw13_fp8_out[xw1_off]) = *reinterpret_cast<uint64_t*>(fp8_1);
    *reinterpret_cast<uint64_t*>(&grad_xw13_fp8_out[xw3_off]) = *reinterpret_cast<uint64_t*>(fp8_3);
  }
@@ -1,7 +1,6 @@
 import unittest
 import numpy as np

-from test.helpers import CI
 from tinygrad.helpers import BEAM, Timing, prod
 from tinygrad import Variable, Device, Tensor
 from tinygrad.nn import Conv2d
@@ -65,7 +64,7 @@ class TestBeamSearch(unittest.TestCase):
    actual = a.numpy()
    np.testing.assert_allclose(actual, desired)

-  @unittest.skipIf(CI, "flaky. CL_OUT_OF_RESOURCES")
+  @unittest.skip("flaky. CL_OUT_OF_RESOURCES")
  def test_conv_beam(self):
    c = Conv2d(3, 16, (3,3))
    x = rand(1,3,32,32)
@@ -0,0 +1,294 @@
+#!/usr/bin/env python3
+"""
+Self-contained Qualcomm Hexagon **testsig** generator.
+
+Replicates: python2 elfsigner.py -t 0x67489311 -o .
+Dependencies: standard library + cryptography (pip install cryptography).
+Multiple serial numbers: use -t multiple times.
+"""
+
+import argparse, base64, hashlib, os, struct
+from datetime import datetime, timedelta, timezone
+from cryptography import x509
+from cryptography.hazmat.primitives import hashes, serialization
+from cryptography.hazmat.primitives.asymmetric import rsa
+from cryptography.x509.name import _ASN1Type
+from cryptography.x509.oid import ExtensionOID, NameOID, ObjectIdentifier
+
+# Embedded assets (raw base64) --- no external files needed
+# Compact test_elf_nop.so contents used for signing. Only bytes covered by
+# program headers are needed; section headers and alignment gaps are not.
+ORIG_PHDRS = [
+    {'t': 1, 'o': 0x0000, 'v': 0x0000, 'p': 0x0000, 'fs': 0x02fc, 'ms': 0x02fc, 'fl': 0x4, 'al': 0x1000},
+    {'t': 1, 'o': 0x1000, 'v': 0x1000, 'p': 0x1000, 'fs': 0x0104, 'ms': 0x0104, 'fl': 0x5, 'al': 0x1000},
+    {'t': 1, 'o': 0x2000, 'v': 0x2000, 'p': 0x2000, 'fs': 0x0004, 'ms': 0x0004, 'fl': 0x4, 'al': 0x1000},
+    {'t': 1, 'o': 0x3000, 'v': 0x4000, 'p': 0x4000, 'fs': 0x00d0, 'ms': 0x0100, 'fl': 0x6, 'al': 0x1000},
+    {'t': 2, 'o': 0x3010, 'v': 0x4010, 'p': 0x4010, 'fs': 0x00a8, 'ms': 0x00a8, 'fl': 0x6, 'al': 0x4},
+]
+ORIG_SEGS = {
+    0x0000: base64.b64decode("""
+f0VMRgEBAQAAAAAAAAAAAAMApAABAAAAsBAAADQAAACIMQAAAwAAADQAIAAFACgAFQASAAEAAAAAAAAAAAAAAAAAAAD8AgAA/AIA
+AAQAAAAAEAAAAQAAAAAQAAAAEAAAABAAAAQBAAAEAQAABQAAAAAQAAABAAAAACAAAAAgAAAAIAAABAAAAAQAAAAEAAAAABAAAAEA
+AAAAMAAAAEAAAABAAADQAAAAAAEAAAYAAAAAEAAAAgAAABAwAAAQQAAAEEAAAKgAAACoAAAABgAAAAQAAAADAAAAEwAAABIAAAAR
+AAAADgAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACgAAAAAAAAAJAAAACwAAAAwAAAANAAAA
+DwAAABAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAQAAAAAAAAAwAFAAAAAACwEAAAAAAAAAMABwAAAAAAwBAAAAAAAAADAAgAAAAA
+AAAgAAAAAAAAAwAJAAAAAAAAQAAAAAAAAAMACgAAAAAACEAAAAAAAAADAAsAAAAAAMxAAAAAAAAAAwAPAAAAAAAAQQAAAAAAAAMA
+EAAPAAAAABAAAGwAAAASAAUAFQAAAAAAAAAAAAAAEgAAAGkAAAAAQQAAAAAAABAA8f9KAAAAzEAAAAQAAAARAA8AMQAAAARAAAAA
+AAAAEAAKAAEAAAAIQAAAAAAAABAACwBcAAAAwBAAAEQAAAASAAgAYgAAANBAAAAAAAAAEADx/3UAAAAAQQAAAAAAABAA8f9GAAAA
+sBAAAAQAAAASAAcAAF9fRFRPUl9MSVNUX18AX2luaXQAX19yZWdpc3Rlcl9mcmFtZV9pbmZvX2Jhc2VzAF9fQ1RPUl9FTkRfXwBs
+aWJjLnNvAG5vcABub3BfdmFyAGxpYmdjYy5zbwBfZmluaQBfZWRhdGEAX19ic3Nfc3RhcnQAX2VuZAB0ZXN0X2VsZl9ub3Auc28A
+AADIQAAAIgoAAAAAAAA=
+"""),
+    0x1000: base64.b64decode("""
+AcCdoADbnaEB2J2hGMAJalTP6nH//+pyGNgq8///4HJI3+BxAMAY8wHAgJEIwAEQAkAAeAEoAyg0wABa///7ckz/+3Eb2xjzm//7
+vwDAm5EGwAAQAMCgUPj//1k4wJ2RG0CdkR7AHpAAwJ9SAAAAAMFAAAAcxElqDkKc4k9AnJE8wJyRDkIOjADAnFIAAAAAAAAAAAAA
+AAAAAAAAAAAAAMBAAAAO1ElqHMCOkQDAnFIAwJ9SAAAAAAAAAAAAAAAAAcCdoADbnaEPwAlqENDqcf//6nIPzyrz///7clD/+3Eb
+2w/zm8AbsADAm5EGwAAQAMCgUPj//1kbQJ2RHsAekADAn1I=
+"""),
+    0x2000: base64.b64decode("AAAAAA=="),
+    0x3000: base64.b64decode("""
+AAAAAAAAAAAAAAAAAAAAAAEAAAA+AAAAAQAAAFIAAAAOAAAAegAAAAwAAAAAEAAADQAAAMAQAAAEAAAA1AAAAAUAAABkAgAABgAA
+ADQBAAAKAAAAigAAAAsAAAAQAAAAAwAAALhAAAACAAAADAAAABQAAAAHAAAAFwAAAPACAAABAABwAwAAAAAAAAAAAAAAAAAAAAAA
+AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAABBAAAAAAAAAAAAAAAAAAABwEAAACgAAAA==
+"""),
+}
+ORIG_EHDR = ORIG_SEGS[0][:0x34]
+
+ATTESTCA_KEY = base64.b64decode("""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""")
+ATTESTCA_CERT = base64.b64decode("""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""")
+ROOTCA_CERT = base64.b64decode("""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""")
+
+PT_NULL = 0
+PF_OS_PHDR = 0x07000000
+PF_OS_HASH = 0x02200000
+MBN_V3 = 3
+SIG_SIZE = 256
+CC_SIZE = 0x1800
+PAD_FF = b'\xff'
+IPAD = 0x3636363636363636
+OPAD = 0x5C5C5C5C5C5C5C5C
+
+
+def _pad(data, size, pad=PAD_FF):
+  if len(data) > size:
+    raise ValueError("data too large: %d > %d" % (len(data), size))
+  return data + pad * (size - len(data))
+
+
+def _orig_segment_data(ph):
+  off, size = ph['o'], ph['fs']
+  for base_off, data in ORIG_SEGS.items():
+    rel = off - base_off
+    if 0 <= rel and rel + size <= len(data):
+      return data[rel:rel + size]
+  raise KeyError("missing original segment at 0x%x" % off)
+
+
+def _build_ehdr(base_ehdr, num_phdrs):
+  e = bytearray(base_ehdr)
+  struct.pack_into('<I', e, 0x20, 0)
+  struct.pack_into('<H', e, 0x2c, num_phdrs)
+  struct.pack_into('<H', e, 0x30, 0)
+  struct.pack_into('<H', e, 0x32, 0)
+  return bytes(e)
+
+
+def _build_elf(ehdr, phs, segs):
+  phoff = struct.unpack_from('<I', ehdr, 0x1c)[0]
+  d = bytes(ehdr)
+  if len(d) < phoff:
+    d += b'\x00' * (phoff - len(d))
+  for ph in phs:
+    d += struct.pack('<IIIIIIII', ph['t'], ph['o'], ph['v'],
+             ph['p'], ph['fs'], ph['ms'], ph['fl'], ph['al'])
+  for off, sdata in sorted(segs.items()):
+    if len(d) < off:
+      d += b'\x00' * (off - len(d))
+    d = d[:off] + sdata + d[off + len(sdata):]
+  return d
+
+
+def _qti_hmac(data, msm=0, sw=0):
+  def _u(v):
+    return bytes.fromhex(format(v, 'x').zfill(16))
+  Si, So = _u(sw ^ IPAD), _u(msm ^ OPAD)
+  a = hashlib.sha256(data).digest()
+  b = hashlib.sha256(Si + a).digest()
+  c = hashlib.sha256(So + b).digest()
+  return c
+
+
+def _raw_pkcs1_sign(private_key, data):
+  """Raw RSA-PKCS1-v1_5 signing WITHOUT DigestInfo wrapper.
+  This matches OpenSSL:  pkeyutl -sign -pkeyopt rsa_padding_mode:pkcs1"""
+  numbers = private_key.private_numbers()
+  d = numbers.d
+  n = numbers.public_numbers.n
+  key_len = (n.bit_length() + 7) // 8
+  pad_len = key_len - 3 - len(data)
+  if pad_len < 8:
+    raise ValueError("data too long for key size")
+  em = b'\x00\x01' + b'\xff' * pad_len + b'\x00' + data
+  m_int = int.from_bytes(em, 'big')
+  sig_int = pow(m_int, d, n)
+  return sig_int.to_bytes(key_len, 'big')
+
+
+def _new_cert(ca_key, ca_cert, attrs):
+  k = rsa.generate_private_key(public_exponent=3, key_size=2048)
+  def _na(oid, value, typ):
+    return x509.NameAttribute(oid, value, _type=typ)
+
+  # Match the SecTools/OpenSSL attestation cert profile accepted by DSP loaders.
+  n = [
+    _na(NameOID.COUNTRY_NAME, "US", _ASN1Type.PrintableString),
+    _na(NameOID.COMMON_NAME, "SecTools Test User", _ASN1Type.PrintableString),
+    _na(NameOID.LOCALITY_NAME, "San Diego", _ASN1Type.PrintableString),
+    _na(NameOID.ORGANIZATION_NAME, "SecTools", _ASN1Type.PrintableString),
+    _na(NameOID.STATE_OR_PROVINCE_NAME, "California", _ASN1Type.PrintableString),
+    _na(NameOID.ORGANIZATIONAL_UNIT_NAME, "01 %.16X SW_ID" % attrs['sw'], _ASN1Type.T61String),
+    _na(NameOID.ORGANIZATIONAL_UNIT_NAME, "02 %.16X HW_ID" % attrs['hw'], _ASN1Type.T61String),
+    _na(NameOID.ORGANIZATIONAL_UNIT_NAME, "04 %.4X OEM_ID" % attrs['oid'], _ASN1Type.T61String),
+    _na(NameOID.ORGANIZATIONAL_UNIT_NAME, "05 %.8X SW_SIZE" % attrs['sz'], _ASN1Type.T61String),
+    _na(NameOID.ORGANIZATIONAL_UNIT_NAME, "06 %.4X MODEL_ID" % attrs['mid'], _ASN1Type.T61String),
+    _na(NameOID.ORGANIZATIONAL_UNIT_NAME, "07 0001 %s" % attrs['ha'], _ASN1Type.PrintableString),
+    _na(NameOID.ORGANIZATIONAL_UNIT_NAME, "03 %.16X DEBUG" % attrs['dbg'], _ASN1Type.PrintableString),
+  ]
+
+  nvb = datetime.now(timezone.utc).replace(microsecond=0)
+  nva = nvb + timedelta(days=20 * 365)
+  b = (x509.CertificateBuilder()
+     .subject_name(x509.Name(n))
+     .issuer_name(ca_cert.subject)
+     .public_key(k.public_key())
+     .serial_number(1)
+     .not_valid_before(nvb)
+     .not_valid_after(nva))
+  b = b.add_extension(x509.AuthorityKeyIdentifier.from_issuer_public_key(ca_key.public_key()), False)
+  b = b.add_extension(x509.UnrecognizedExtension(ExtensionOID.BASIC_CONSTRAINTS, b'\x30\x03\x02\x01\x00'), False)
+  b = b.add_extension(x509.KeyUsage(
+    digital_signature=True,
+    content_commitment=False,
+    key_encipherment=False,
+    data_encipherment=False,
+    key_agreement=False,
+    key_cert_sign=False,
+    crl_sign=False,
+    encipher_only=False,
+    decipher_only=False,
+  ), False)
+  b = b.add_extension(x509.UnrecognizedExtension(ObjectIdentifier("1.3.6.1.4.1.1449.9.6.3"), b'\x00\x01\xe2\x40\x00\x01\xe2\x40'), False)
+  return k, b.sign(ca_key, hashes.SHA256())
+
+
+def _build_chain(attest, ca, root):
+  c = (attest.public_bytes(serialization.Encoding.DER) +
+     ca.public_bytes(serialization.Encoding.DER) +
+     root.public_bytes(serialization.Encoding.DER))
+  return _pad(c, CC_SIZE, b'\xff')
+
+
+def _sign(serial_num, out_dir):
+  orig = ORIG_PHDRS
+  base_ehdr = ORIG_EHDR
+
+  nph = len(orig) + 2
+  phsz = nph * 32
+  ehsz = len(base_ehdr)
+  phoff = struct.unpack_from('<I', base_ehdr, 0x1c)[0]
+  hv = 0x5000
+  ho = 0x1000
+  hfs = 0x1a08
+
+  prog = {'t': PT_NULL, 'o': 0, 'v': 0, 'p': 0,
+      'fs': ehsz + phsz, 'ms': 0, 'fl': PF_OS_PHDR, 'al': 0}
+  hph = {'t': PT_NULL, 'o': ho, 'v': hv, 'p': hv,
+       'fs': hfs, 'ms': 0x2000, 'fl': PF_OS_HASH, 'al': 0x1000}
+  shifted = [dict(ph, o=ph['o'] + 0x3000) for ph in orig]
+  allph = [prog, hph] + shifted
+
+  # Temporary ELF for hash 0
+  tmp_segs = {}
+  for i, ph in enumerate(shifted):
+    tmp_segs[ph['o']] = _orig_segment_data(orig[i])
+  ehdr = _build_ehdr(base_ehdr, nph)
+  tmp = _build_elf(ehdr, allph, tmp_segs)
+
+  phb = tmp[phoff:phoff + phsz]
+  hash0 = hashlib.sha256(ehdr + phb).digest()
+  hash1 = struct.pack('<I', serial_num) + b'\x00' * 28
+  hs = [hash0, hash1]
+  for ph in orig:
+    hs.append(hashlib.sha256(_orig_segment_data(ph)).digest())
+  ht = b''.join(hs)
+
+  cs, ss, ccs = len(ht), SIG_SIZE, CC_SIZE
+  dst = hv + 40
+  sp = dst + cs
+  cp = sp + ss
+  isz = cs + ss + ccs
+
+  mbn = struct.pack('<IIIIIIIIII',
+            0, MBN_V3, 0, dst, isz, cs, sp, ss, cp, ccs)
+  dts = mbn + ht
+  hmac = _qti_hmac(dts)
+
+  ca_key = serialization.load_pem_private_key(ATTESTCA_KEY, password=None)
+  ca_cert = x509.load_der_x509_certificate(ATTESTCA_CERT)
+  root_cert = x509.load_der_x509_certificate(ROOTCA_CERT)
+
+  attrs = {'sw': 0, 'hw': 0, 'oid': 0, 'mid': 0,
+       'sz': len(dts), 'ha': 'SHA256', 'dbg': 2}
+  nk, ac = _new_cert(ca_key, ca_cert, attrs)
+  sig = _raw_pkcs1_sign(nk, hmac)
+  sig = _pad(sig, SIG_SIZE, b'\x00')
+  cc = _build_chain(ac, ca_cert, root_cert)
+
+  hseg = mbn + ht + sig + cc
+  if len(hseg) != hfs:
+    raise RuntimeError("hash seg size mismatch %d vs %d" % (len(hseg), hfs))
+
+  segs = {ho: hseg}
+  for i, ph in enumerate(shifted):
+    segs[ph['o']] = _orig_segment_data(orig[i])
+
+  final = _build_elf(ehdr, allph, segs)
+
+  os.makedirs(out_dir, exist_ok=True)
+  out = os.path.join(out_dir, "testsig-0x%08X.so" % serial_num)
+  with open(out, 'wb') as f:
+    f.write(final)
+  print("Signing complete! Output saved at %s" % out)
+  return out
+
+
+def main():
+  p = argparse.ArgumentParser(description="Qualcomm testsig generator")
+  p.add_argument("-t", "--testsig", action="append", default=None,
+           help="serial number (e.g. 0x67489311); repeatable. Defaults to value from /sys/devices/soc0/serial_number")
+  p.add_argument("-o", "--output_dir", default=".")
+  args = p.parse_args()
+
+  serials = args.testsig if args.testsig else []
+
+  if not serials:
+    # Read default serial number from device
+    try:
+      with open('/sys/devices/soc0/serial_number', 'r') as f:
+        serial_str = f.read().strip()
+      serials = [serial_str]
+      print("Using serial number from /sys/devices/soc0/serial_number: %s" % serial_str)
+    except FileNotFoundError:
+      raise SystemExit("Error: No serial number provided (-t) and /sys/devices/soc0/serial_number not found.")
+    except PermissionError:
+      raise SystemExit("Error: Cannot read /sys/devices/soc0/serial_number (permission denied).")
+
+  for s in serials:
+    v = int(s.strip(), 0)
+    if not (0 <= v <= 0xFFFFFFFF):
+      raise ValueError("bad serial %r" % s)
+    _sign(v, args.output_dir)
+
+
+if __name__ == "__main__":
+  main()
@@ -564,8 +564,8 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
  "aten.__rshift__.Scalar": lambda x,y: x>>y,
  "aten.__irshift__.Scalar": lambda x,y: x>>y,
  # inplace ops using replace for fusion
-  "aten.zero_": lambda x: x.zeros_like(),
-  "aten.fill_.Scalar": lambda x, y: x.full_like(y),
+  "aten.zero_": lambda x: x.const_like(0),
+  "aten.fill_.Scalar": lambda x, y: x.const_like(y),
  "aten.add_.Tensor": lambda self, other, alpha=1.0: self + other * alpha,
  "aten.add_.Scalar": lambda self, other, alpha=1.0: self + other * alpha,
  "aten.mul_.Tensor": lambda self, other: self * other,
@@ -617,7 +617,7 @@ tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
  "aten.asinh": Tensor.asinh,
  "aten.mul": Tensor.mul,
  "aten.atanh": Tensor.atanh,
-  "aten.fill_.Tensor": lambda self, value: Tensor.full(self.shape, value.reshape(()).item(), device=self.device, dtype=self.dtype),
+  "aten.fill_.Tensor": lambda self, value: self.const_like(value.reshape(()).item()),
  "aten.flip": Tensor.flip,
  "aten.scatter_reduce.two": Tensor.scatter_reduce,
  "aten.squeeze_.dim": Tensor.squeeze,
@@ -240,9 +240,9 @@ class TestTorchBackend(unittest.TestCase):
    np.testing.assert_equal(result.cpu().numpy(), [3., 3., 2.])

  def test_mnist_index(self):
+    # from tinygrad.nn.datasets import mnist
+    X_train, Y_train = Tensor.randint(60000, 1, 28, 28, dtype='uchar').realize(), Tensor.randint(60000, dtype='uchar').realize()
    GlobalCounters.reset()
-    from tinygrad.nn.datasets import mnist
-    X_train, Y_train, _, _ = mnist()
    X_train = torch.tensor(X_train.float().numpy(), device=device)
    Y_train = torch.tensor(Y_train.cast('int64').numpy(), device=device)
    samples = torch.randint(0, X_train.shape[0], (32,))
@@ -53,8 +53,6 @@ All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op
  Placeholder with shape $\mathbf{s}$. Substituted in \op{Function}. \\[4pt]
 \op{Buffer} & () & size, dtype, device, addrspace &
  Shape $(n \cdot \textit{size},)$ if device is $n$-tuple, else $(\textit{size},)$. \\
-\op{BufferView} & (buf,) & size, dtype, offset &
-  Typed access into a buffer. Zero-copy $(\textit{size},)$ slice at offset; inherits addrspace. \\
 \op{Const} & () & value, dtype &
  A scalar constant with shape $(\ )$. \\
  & & & Form vector consts with \op{Stack} \\
@@ -66,7 +64,7 @@ All nodes in the tinygrad graph are \textbf{UOps}. A UOp is a tuple $(\mathrm{op
 A \op{Buffer}'s \textbf{addrspace} is \texttt{GLOBAL}, \texttt{LOCAL}, or \texttt{REG}.

 %% ============================================================
-\subsection*{{\color{movgreen}Movement Ops} \normalfont\small--- no arithmetic, shapes are $(k,)$-shaped UOps with dtype \texttt{index} in src}
+\subsection*{{\color{movgreen}Movement Ops} \normalfont\small--- no arithmetic; view, indexing, and reinterpretation only}

 \begin{tabular}{@{}l l l l@{}}
 \toprule
@@ -76,11 +74,13 @@ A \op{Buffer}'s \textbf{addrspace} is \texttt{GLOBAL}, \texttt{LOCAL}, or \textt
 \op{Flip} & $(T,)$ & bools $\mathbf{f}$ & Reverse along flagged axes. \\
 \op{Reshape} & $(T, \mathbf{s'})$ & --- & Reinterpret in row-major order. $\prod s_k = \prod s'_k$. \\
 \op{Expand} & $(T, \mathbf{s'})$ & --- & Broadcast size-1 axes. $s_k \in \{1, s'_k\}$. \\
-\op{Pad} & $(T, \mathbf{b}, \mathbf{e})$ & --- & Pad with $0$s: $b_k$ before, $e_k$ after each axis. \\
-\op{Shrink} & $(T, \mathbf{b}, \mathbf{e})$ & --- & Keep $[b_k, e_k)$ per axis. Inverse of \op{Pad}. \\
+\op{Pad} & $(T, \mathbf{o}, \mathbf{s'})$ & --- & Place $T$ at offset $o_k$ in a zero-filled output of shape $s'_k$. \\
+\op{Shrink} & $(T, \mathbf{o}, \mathbf{s'})$ & --- & Keep $s'_k$ elements starting at offset $o_k$ per axis. Inverse of \op{Pad}. \\
 \op{Index} & $(T, i_0, i_1, \ldots)$ & --- & Index from left. $()$-shaped $i$ removes dim; $(k,)$-shaped makes it $k$. \\
 \op{Stack} & $(T_0, T_1, \ldots)$ & --- & Join along a newly created leading axis. All shapes must match. \\
 \op{Replicated} & $(T,)$ & axes & Mark $T$ as replicated along axes. Collapse axes to $1$. \\
+\op{Slice} & $(T, \mathrm{offset})$ & size, dtype & Zero-copy \textit{size} elems of dtype; offset is elems of $T$ dtype. \\
+\op{Bitcast} & $(T,)$ & dtype & Reinterpret storage as target dtype; preserve total bytes. \\
 \bottomrule
 \end{tabular}

@@ -166,8 +166,7 @@ Unary & $(T,)$
  & $\mathrm{trunc}(x)$: round toward zero. \\
 & & \op{Cast}
  & Convert to target dtype (specified in arg). \\
-& & \op{Bitcast}
-  & Reinterpret bits as target dtype. Must be same size. \\[4pt]
+  \\[4pt]
 Binary & $(A, B)$
  & \op{Add}, \op{Mul}, \op{Max}, \op{Mod}, \op{Idiv}
  & $a+b$, $a \cdot b$, $\max(a,b)$, $a \bmod b$, $\lfloor a/b \rfloor$ \\
@@ -0,0 +1,20 @@
+#!/usr/bin/env python3
+import unittest
+import tinygrad.runtime.autogen.amd.cdna.ins as cdna
+from test.amd.hw.test_cdna_vop3 import run_cdna
+
+class TestCDNASDWA(unittest.TestCase):
+  def test_v_add_co_u32_e32_writes_vcc(self):
+    out = run_cdna([
+      cdna.s_mov_b32(cdna.s[0], 0xffffffff),
+      cdna.v_mov_b32_e32(cdna.v[0], cdna.s[0]),
+      cdna.v_mov_b32_e32(cdna.v[13], 1),
+      cdna.v_add_co_u32_e32(cdna.v[0], cdna.SDWA, cdna.v[13], vsrc0=cdna.v[0], dst_sel=6, src0_sel=6),
+      cdna.v_mov_b32_e32(cdna.v[2], cdna.VCC_LO),
+      cdna.v_lshlrev_b32_e32(cdna.v[2], 31, cdna.v[2]),
+      cdna.v_or_b32_e32(cdna.v[2], cdna.v[2], cdna.v[0]),
+    ])
+    self.assertEqual(out, 0x80000000)
+
+if __name__ == "__main__":
+  unittest.main()
@@ -167,7 +167,7 @@ class TestDSPcodePatterns(unittest.TestCase):

  def test_global_atomic_add_f32_parsing(self):
    """Test GLOBAL_ATOMIC_ADD_F32 keeps memory values in float dtype."""
-    vmem = UOp(Ops.PARAM, dtypes.uint32.ptr(1024), arg=2)
+    vmem = UOp.param(2, dtypes.uint32.ptr(1024))
    srcs = {
      'ADDR': UOp.const(dtypes.uint64, 0),
      'DATA': UOp.const(dtypes.uint32, 0x3f800000),
@@ -198,7 +198,7 @@ class TestDSPcodePatterns(unittest.TestCase):
  def test_mem_read_parsing(self):
    """Test MEM[addr].type read expression parsing."""
    # Create a mock LDS buffer
-    lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
+    lds = UOp.param(3, dtypes.uint32.ptr(16384))
    addr = UOp.const(dtypes.uint32, 0)
    vrs = {'_lds': lds, 'ADDR': addr, 'OFFSET': UOp.const(dtypes.uint32, 0)}

@@ -233,7 +233,7 @@ class TestDSPcodePatterns(unittest.TestCase):
    pcode = PCODE.get(DSOp.DS_LOAD_2ADDR_B32)
    self.assertIsNotNone(pcode)
    assert pcode is not None
-    lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
+    lds = UOp.param(3, dtypes.uint32.ptr(16384))
    srcs = {
      'ADDR': UOp.const(dtypes.uint32, 0),
      'OFFSET0': UOp.const(dtypes.uint32, 0),
@@ -314,7 +314,7 @@ class TestConcatWidthParsing(unittest.TestCase):
      self.assertEqual(parsed.simplify().arg, expected)

  def test_permlane64_wave64_pcode_indices(self):
-    vgpr = UOp(Ops.PARAM, dtypes.uint32.ptr(256), arg=0)
+    vgpr = UOp.param(0, dtypes.uint32.ptr(256))
    srcs = {
      'SRC0': UOp.const(dtypes.uint32, 0),
      'VDST': UOp.const(dtypes.uint32, 1),
@@ -347,7 +347,7 @@ class TestAllPcode(unittest.TestCase):
  def _make_srcs(self):
    """Create dummy source variables for pcode parsing."""
    u32, u64 = lambda v=0: UOp.const(dtypes.uint32, v), lambda v=0: UOp.const(dtypes.uint64, v)
-    lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
+    lds = UOp.param(3, dtypes.uint32.ptr(16384))
    return {'laneId': u32(), 'laneID': u32(), 'S0': u32(), 'S1': u32(), 'S2': u32(), 'S3': u32(), 'SRC0': u32(),
            'D0': u32(), 'D1': u32(), 'DST': u32(), 'VDST': u32(), 'SDST': u32(),
            'VCC': u64(), 'VCCZ': u32(), 'EXEC': u64(), 'EXEC_LO': u32(), 'EXECZ': u32(), 'SCC': u32(),
@@ -19,7 +19,7 @@ class TestArange(unittest.TestCase):
    self.assertEqual(self._get_flops(Tensor.arange(256), np.arange(256)), 0)
    self.assertEqual(self._get_flops(Tensor.arange(2560), np.arange(2560)), 0)

-  @unittest.skipIf(Device.DEFAULT == "CL", "TODO: fails on CI CL")
+  @unittest.skipIf(Device.DEFAULT == "CL", "flaky in CI")
  def test_arange_cumsum(self):
    np.testing.assert_equal(Tensor.arange(513).cumsum(0).numpy(), np.arange(513).cumsum())

@@ -125,8 +125,8 @@ class TestIndexing(unittest.TestCase):
  def test_index_mnist(self, noopt=1, op_limit=512*784*13, split_reduceop=0):
    # WEBGPU generates more ops due to bitpacking of < 4-byte dtypes
    if Device.DEFAULT == "WEBGPU": op_limit *= 15
-    from tinygrad.nn.datasets import mnist
-    X_train, Y_train, _, _ = mnist()
+    # from tinygrad.nn.datasets import mnist
+    X_train, Y_train = Tensor.randint(DSET, 1, 28, 28, dtype='uchar').realize(), Tensor.randint(DSET, dtype='uchar').realize()
    with Context(NOOPT=noopt, SPLIT_REDUCEOP=split_reduceop):
      samples = Tensor.randint(getenv("BS", 512), high=X_train.shape[0]).realize()
      GlobalCounters.reset()
@@ -155,7 +155,7 @@ class TestCustomKernel(unittest.TestCase):
    self.assertTrue((ref == tst).all().item())

  def test_eye(self):
-    ref = Tensor.eye(1024).contiguous().realize()
+    ref = Tensor.eye(1024).clone().realize()
    tst = Tensor.empty_like(ref)
    tst = tst.custom_kernel(fxn=custom_eye_kernel)[0]
    self.assertTrue((ref == tst).all().item())
@@ -335,7 +335,7 @@ class TestCustomKernel(unittest.TestCase):
    assert all(x == expected for x in result), f"expected all {expected}, got {result}"

  def test_custom_kernel_sched(self, use_custom=False):
-    x = Tensor.arange(32).reshape(8, 4).realize()
+    x = Tensor.arange(32).reshape(8, 4).clone().realize()
    y = Tensor.empty_like(x)
    y = Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
    if use_custom:
@@ -352,7 +352,7 @@ class TestCustomKernel(unittest.TestCase):

  @unittest.expectedFailure
  def test_sliced_buffer_function(self):
-    x = Tensor.arange(32).reshape(8, 4).realize()
+    x = Tensor.arange(32).reshape(8, 4).clone().realize()
    from tinygrad import function
    @function(precompile=True)
    def run(x:Tensor) -> Tensor:
@@ -2,13 +2,13 @@ import contextlib, unittest, math
 import numpy as np
 import torch
 from typing import Any, List
-from tinygrad.helpers import getenv, DEBUG, EMULATED_DTYPES
+from tinygrad.helpers import getenv, DEBUG, EMULATED_DTYPES, DEV
 from tinygrad.dtype import DType, DTYPES_DICT, least_upper_dtype, fp8_to_float, float_to_fp8, _to_np_dtype, _to_torch_dtype, truncate
 from tinygrad.renderer.ptx import PTXRenderer
 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, CI
+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
 import pytest
 pytestmark = pytest.mark.filterwarnings("ignore")
@@ -110,7 +110,9 @@ def _test_ops(a_dtype:DType, b_dtype:DType, target_dtype=None):
  _assert_eq((Tensor([1], dtype=a_dtype).cast(b_dtype)+Tensor([1], dtype=a_dtype).cast(b_dtype)).cast(a_dtype), a_dtype, [2])
  _assert_eq(Tensor([1,2,3,4], dtype=a_dtype)*Tensor([1,2,3,4], dtype=b_dtype), target_dtype, [1,4,9,16])
  _assert_eq(Tensor([[1,2],[3,4]], dtype=a_dtype)@Tensor.eye(2, dtype=b_dtype), target_dtype, [[1,2],[3,4]])
-  _assert_eq(Tensor([1,1,1,1], dtype=a_dtype)+Tensor.ones((4,4), dtype=b_dtype), target_dtype, 2*Tensor.ones(4,4).numpy())
+  _assert_eq(Tensor([1,1,1,1], dtype=a_dtype)+Tensor.ones((4,4), dtype=b_dtype), target_dtype, 2*np.ones((4,4)))
+  _assert_eq(Tensor([1,1,1,1], dtype=a_dtype)+Tensor.ones((4,4), dtype=b_dtype).clone(), target_dtype, 2*np.ones((4,4)))
+  _assert_eq(Tensor.ones((4,4), dtype=b_dtype).clone(), b_dtype, np.ones((4,4)))

 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))
@@ -225,7 +227,7 @@ class TestFloatDType(TestDType):
@unittest.skipUnless(dtypes.double in supported_dtypes, f"no double on {Device.DEFAULT}")
 class TestDoubleDType(TestDType):
  DTYPE = dtypes.double
-  @unittest.skipIf((CI and Device.DEFAULT in {"CUDA", "NV"}) or \
+  @unittest.skipIf((DEV.interface.startswith("MOCK") and Device.DEFAULT in {"CUDA", "NV"}) or \
   isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)), "conversion not supported on CI CUDA, PTX, and NIR")  # TODO: why not?
  def test_float64_increased_precision(self):
    for func in [
@@ -7,7 +7,6 @@ from tinygrad.runtime.ops_python import from_storage_scalar
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
 from tinygrad.uop import Ops
-from test.helpers import CI
 import numpy as np
 import pytest
 from hypothesis import assume, given, strategies as strat, settings
@@ -112,6 +111,7 @@ def universal_test_cast(a, in_dtype, dtype):
 def universal_test_midcast(a, b, c, op1, op2, d1:DType, d2:DType):
  if not isinstance(op1, tuple): op1 = (op1, op1)
  if not isinstance(op2, tuple): op2 = (op2, op2)
+  if op1[0] == operator.mod and b == 0: return
  # lt and max with nan is undefined in tinygrad
  if op1[0] in (operator.lt, Tensor.maximum) and (math.isnan(a) or math.isnan(b)): return
  if op2[0] in (operator.lt, Tensor.maximum) and math.isnan(c): return
@@ -331,12 +331,12 @@ class TestDTypeALU(unittest.TestCase):
  @given(ht.bool, ht.bool, strat.sampled_from(((operator.add, operator.add), (operator.mul, operator.mul))))
  def test_bool(self, a, b, op): universal_test(a, b, dtypes.bool, op)

-  @unittest.skipIf(not CI and Device.DEFAULT == "METAL", "broken on local M3")
  @given(ht.int32, ht.int32, ht.float32, strat.sampled_from(integer_binary_operations), strat.sampled_from(binary_operations))
  def test_int32_midcast_float(self, a, b, c, op1, op2): universal_test_midcast(a, b, c, op1, op2, dtypes.int32, dtypes.float32)

-  # Metal and CUDA and HIP and NIR behave differently than numpy in CI for overflows
-  skip_overflow = (CI and Device.DEFAULT in {"AMD", "NV", "CUDA"}) or isinstance(Device[Device.DEFAULT].renderer, NIRRenderer)
+  # Metal and (MOCK)CUDA and HIP and NIR behave differently than numpy for overflows
+  skip_overflow = ((DEV.interface.startswith("MOCK") and Device.DEFAULT in {"AMD", "NV", "CUDA"})
+                   or isinstance(Device[Device.DEFAULT].renderer, NIRRenderer))
  @given(strat.floats(width=32, min_value=0, max_value=10.0) if skip_overflow else ht.float32,
         strat.floats(width=32, min_value=0, max_value=10.0) if skip_overflow else ht.float32,
         ht.int32, strat.sampled_from(binary_operations), strat.sampled_from(integer_binary_operations))
@@ -1,5 +1,5 @@
 #!/usr/bin/env python
-import unittest
+import unittest, os
 import torch
 import numpy as np

@@ -7,7 +7,6 @@ from tinygrad.helpers import DEV
 from tinygrad.tensor import Tensor
 from tinygrad.device import Device
 from tinygrad.dtype import _from_torch_dtype, _to_torch_dtype
-from test.helpers import CI

 MOCKGPU = DEV.interface.startswith("MOCK")

@@ -28,7 +27,7 @@ class TestInterop(unittest.TestCase):
    tg_out = tg_data[:, :, 0] * 0.2989 + tg_data[:, :, 1] * 0.5870 + tg_data[:, :, 2] * 0.1140
    tg_res = tg_out.numpy()

-    if self.torch_device == "mps" and CI:
+    if self.torch_device == "mps" and os.getenv("CI", "") != "":
      # MPS backend out of memory: https://discuss.pytorch.org/t/mps-back-end-out-of-memory-on-github-action/189773
      # Calculate expected value on cpu.
      inp = inp.cpu()
@@ -57,7 +57,7 @@ class TestIselX86(unittest.TestCase):
    # need to move src from gpr to xmm before broadcasting
    self.assertTrue(n.arg is X86Ops.VPBROADCASTD and n.src[0].arg is X86Ops.VMOVD)
    # if we can fuse a load we can skip the move and access memory directly
-    load = UOp(Ops.PARAM, dtypes.int32.ptr(), arg=0).index(UOp.const(dtypes.int32, 0), ptr=True).load()
+    load = UOp.param(0, dtypes.int32.ptr()).index(UOp.const(dtypes.int32, 0), ptr=True).load()
    n = self.isel_rewrite(load.broadcast(4))
    self.assertTrue(n.arg is X86Ops.VPBROADCASTD and len(n.src) == 3)

@@ -122,20 +122,20 @@ class TestIselX86(unittest.TestCase):
  # complex address is [base + index*scale + displacement]
  def test_complex_address(self):
    a = UOp.variable("a", 0, 0, dtypes.int32)
-    load = UOp(Ops.PARAM, dtypes.int32.ptr(), arg=0).index(a + 1, ptr=True).load()
+    load = UOp.param(0, dtypes.int32.ptr()).index(a + 1, ptr=True).load()
    n = self.isel_rewrite(load)
    # displacement is the constant in "a" scaled to the buffer element size, dtype is int8 when the value fits otherwise int32
    self.assertTrue(n.src[2].op is Ops.CONST and n.src[2].dtype is dtypes.int8 and n.src[2].arg == 4)

  def test_fold_load(self):
-    load1 = UOp(Ops.PARAM, dtypes.int32.ptr(), arg=0).index(UOp.const(dtypes.int32, 0), ptr=True).load()
-    load2 = UOp(Ops.PARAM, dtypes.int32.ptr(), arg=0).index(UOp.const(dtypes.int32, 1), ptr=True).load()
+    load1 = UOp.param(0, dtypes.int32.ptr()).index(UOp.const(dtypes.int32, 0), ptr=True).load()
+    load2 = UOp.param(0, dtypes.int32.ptr()).index(UOp.const(dtypes.int32, 1), ptr=True).load()
    n = self.isel_rewrite(load1 + load2)
    self.assertTrue(len(n.src) == 4)

  # don't fold when used multiple times
  def test_dont_fold_load(self):
-    load = UOp(Ops.PARAM, dtypes.int32.ptr(), arg=0).index(UOp.const(dtypes.int32, 0), ptr=True).load()
+    load = UOp.param(0, dtypes.int32.ptr()).index(UOp.const(dtypes.int32, 0), ptr=True).load()
    # used by multiple users
    n = self.isel_rewrite(load + 1 + load)
    self.assertTrue(len(n.src) == 2)
@@ -144,4 +144,4 @@ class TestIselX86(unittest.TestCase):
    self.assertTrue(len(n.src) == 2)

 if __name__ == "__main__":
-  unittest.main()
+  unittest.main()
@@ -49,7 +49,7 @@ class TestJit(unittest.TestCase):
      y = (x + 1).contiguous().realize()
      z = x.shrink(((st, st + N),)).contiguous().realize()
      return y, z
-    x = Tensor.arange(2*N).contiguous().realize()
+    x = Tensor.arange(2*N).clone().realize()
    for _ in range(3): y, z = f(x, Variable("a", 0, N).bind(0))
    self.assertEqual(y.shape, (2*N,))
    self.assertEqual(z.shape, (N,))
@@ -92,7 +92,7 @@ class TestJit(unittest.TestCase):
    @TinyJit
    def f(x): return (x[2:5].contiguous() + 1).realize()
    for i in range(5):
-      x = (Tensor.arange(10).float() + i * 10).contiguous().realize()
+      x = (Tensor.arange(10).float() + i * 10).clone().realize()
      np.testing.assert_allclose(f(x).numpy(), x.numpy()[2:5] + 1)

  def test_jit_multiple_outputs(self):
@@ -8,10 +8,10 @@ class TestKernelCache(unittest.TestCase):
    if Device.DEFAULT not in ["CPU"]:
      self.skipTest("No custom kernel cache is implemented")

-    unique_const = 0.6765677269
+    const_value = 0.6765677269
    a = Tensor.rand(4,4).realize()
    b = Tensor.rand(4,4).realize()
-    x = a + b + unique_const
+    x = a + b + const_value
    x.realize()

    a1 = Tensor.rand(4,4).realize()
@@ -20,7 +20,7 @@ class TestKernelCache(unittest.TestCase):
    Device['CPU'].compiler.compile_cached = None # making it not callable

    try:
-      x1 = a1 + b1 + unique_const
+      x1 = a1 + b1 + const_value
      x1.realize() # Same kernel should be from cache.
    finally:
      Device['CPU'].compiler.compile_cached = orig_compile_func
@@ -11,16 +11,16 @@ from tinygrad.codegen import to_program
 class TestLinearizerFailure(unittest.TestCase):
  @unittest.skipUnless(Device.DEFAULT == "METAL", "only tested on METAL")
  def test_failure_beam_mnist(self):
-    c0 = UOp(Ops.PARAM, dtypes.uchar.ptr(4014080), arg=0, src=())
+    c0 = UOp.param(0, dtypes.uchar.ptr(4014080))
    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 0, AxisType.GLOBAL)
    c2 = UOp.range(UOp.const(dtypes.weakint, 784), 1, AxisType.GLOBAL)
    c3 = UOp.range(UOp.const(dtypes.weakint, 10), 3, AxisType.GLOBAL)
-    c4 = UOp(Ops.PARAM, dtypes.int.ptr(512), arg=1, src=())
+    c4 = UOp.param(1, dtypes.int.ptr(512))
    c5 = c4.index(c1.valid(UOp.const(dtypes.bool, True)))
    c6 = UOp.range(UOp.const(dtypes.weakint, 6000), 1004, AxisType.REDUCE)
    c7 = UOp.range(UOp.const(dtypes.weakint, 3750), 2006, AxisType.REDUCE)
    c8 = UOp.range(UOp.const(dtypes.weakint, 16), 2007, AxisType.GROUP_REDUCE)
-    c9 = UOp(Ops.PARAM, dtypes.uchar.ptr(47040000), arg=2, src=())
+    c9 = UOp.param(2, dtypes.uchar.ptr(47040000))
    c10 = c9.index((((c3*UOp.const(dtypes.weakint, 4704000))+c2)+(c6*UOp.const(dtypes.weakint, 784))).valid(UOp.const(dtypes.bool, True)))
    c11 = c5.alu(Ops.CMPNE, ((((c3*UOp.const(dtypes.weakint, 6000))+c6)+((c7*UOp.const(dtypes.weakint, 16))+c8)).alu(Ops.CMPLT, UOp.const(dtypes.weakint, 59999)).where(UOp.const(dtypes.int, 0), UOp.const(dtypes.int, 1)).reduce(c7, c8, arg=Ops.ADD)+UOp.const(dtypes.int, -1))).where(UOp.const(dtypes.uchar, 0), c10).reduce(c6, arg=Ops.ADD)
    c12 = c0.index((((c1*UOp.const(dtypes.weakint, 7840))+(c2*UOp.const(dtypes.weakint, 10)))+c3).valid(UOp.const(dtypes.bool, True))).store(c11).end(c1, c2, c3)
@@ -187,7 +187,7 @@ class TestMultiTensor(unittest.TestCase):
    np.testing.assert_allclose(O.numpy(), X.numpy()[0:2]*W.numpy()[0:2] < 2)

  def test_shrink_on_shard_axis(self):
-    X = Tensor.arange(4*4).reshape(4,4).realize()
+    X = Tensor.arange(4*4).reshape(4,4).clone().realize()
    X_np = X.numpy()
    X.shard_(devices_2, 0)
    # only shrink on the device that owns the shard, this is enabled by the mselect simplifier
@@ -239,7 +239,28 @@ class TestMultiTensor(unittest.TestCase):
      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])
+    self.assertEqual(kernel_counts[0], kernel_counts[2])
+
+  def test_to_single_device_gather_memory(self):
+    nrows, ncols = 64, 1024
+    nbytes = nrows*ncols*4
+    for devs in (devices_2, devices_4):
+      ndev = len(devs)
+      for axis in (0, 1):
+        sh = Tensor.arange(nrows*ncols).reshape(nrows, ncols).clone().shard(devs, axis).realize()
+        kernels, mem = {}, {}
+        for ring in (0, 2):
+          GlobalCounters.reset()
+          with Context(RING=ring, SCACHE=0):
+            t = sh.to(Device.DEFAULT)
+            t.realize()
+          kernels[ring], mem[ring] = GlobalCounters.kernel_count, GlobalCounters.global_mem
+          self.assertEqual(t.device, Device.DEFAULT)
+          np.testing.assert_equal(t.numpy(), np.arange(nrows*ncols).reshape(nrows, ncols))
+        self.assertEqual(kernels[0], kernels[2])
+        self.assertEqual(mem[0], mem[2])
+        self.assertLess(kernels[0], 2*ndev)
+        self.assertLessEqual(mem[0], 4*nbytes)

  def test_allreduce_all2all(self):
    with Context(ALL2ALL=2):
@@ -272,7 +293,7 @@ class TestMultiTensor(unittest.TestCase):
    @TinyJit
    def f(x): return (x+1).contiguous().sum()
    for _ in range(5):
-      tt = Tensor.arange(0, 4).contiguous().realize().shard((d1,d2), 0).realize()
+      tt = Tensor.arange(0, 4).clone().realize().shard((d1,d2), 0).realize()
      out = f(tt)
      assert out.item() == 1+2+3+4

@@ -288,7 +309,7 @@ class TestMultiTensor(unittest.TestCase):
    @TinyJit
    def f(x): return (x.shard((d1,d2), 0)+1).contiguous().sum()
    for _ in range(5):
-      tt = Tensor.arange(0, 4).contiguous().realize()
+      tt = Tensor.arange(0, 4).clone().realize()
      out = f(tt)
      assert out.item() == 1+2+3+4

@@ -800,9 +821,9 @@ 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):
+  def test_const_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]
+    out = t.const_like(2)[:, :8]
    linear, var_vals = out.linear_with_vars()
    self.assertEqual(len(linear.src), 0)
    run_linear(linear, var_vals)
@@ -844,7 +865,7 @@ class TestMultiTensor(unittest.TestCase):
  @unittest.skip("RANGEIFY doesn't support multi const folding")
  def test_multi_const_folding(self):
    with Context(TRACK_MATCH_STATS=0):
-      a = Tensor.arange(3).realize()
+      a = Tensor.arange(3).clone().realize()
      zeros = Tensor.zeros(3).realize()
    b = a.to(devices_2)*zeros.to(devices_2)
    sched = b.schedule_linear().src
@@ -883,7 +904,7 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):

  # shrink a multitensor on sharded axis
  def test_shrink_bad_args(self):
-    t = Tensor.arange(64).reshape(8, 8).contiguous().realize()
+    t = Tensor.arange(64).reshape(8, 8).clone().realize()
    t.shard_([f"{Device.DEFAULT}:{i}" for i in range(4)], axis=0)

    with self.assertRaises(AssertionError):
@@ -906,7 +927,7 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):
  @given(strat.sampled_from([dtypes.float, dtypes.int, dtypes.int64, dtypes.int16]))
  def test_ops(self, dtype):
    if dtype not in Device[Device.DEFAULT].renderer.supported_dtypes(): return
-    t = Tensor.arange(64).reshape(8, 8).contiguous().realize()
+    t = Tensor.arange(64).reshape(8, 8).clone().realize()
    t.shard_([f"{Device.DEFAULT}:{i}" for i in range(4)], axis=0)
    for i in range(4):
      print(f"{i=}")
@@ -950,7 +971,7 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):
      np.testing.assert_allclose(a.flip(-1).numpy(), b.flip(-1).numpy(), rtol=1e-7, atol=1e-3)

  def test_add_two_partitions(self):
-    t = Tensor.arange(64).reshape(8, 8).contiguous().realize()
+    t = Tensor.arange(64).reshape(8, 8).clone().realize()
    t.shard_([f"{Device.DEFAULT}:{i}" for i in range(4)], axis=0)

    a = t.shrink(((2, 4), None))
@@ -967,7 +988,7 @@ class TestShrinkMultiTensorShardedAxis(unittest.TestCase):

  def test_add_different_tensors(self):
    devices = [f"{Device.DEFAULT}:{i}" for i in range(4)]
-    x = Tensor.arange(64).reshape(8, 8).contiguous().realize().shard(devices, axis=0)
+    x = Tensor.arange(64).reshape(8, 8).clone().realize().shard(devices, axis=0)

    to_add = []
    for i in range(len(devices)):
@@ -1077,7 +1098,7 @@ class TestBatchNorm(unittest.TestCase):
  @given(strat.sampled_from((False, True)))
  def test_batchnorm(self, is_training):
    devices = [f"{Device.DEFAULT}:{i}" for i in range(4)]
-    x = Tensor.arange(4096).reshape(8, 8, 8, 8).contiguous().realize().shard(devices, axis=0)
+    x = Tensor.arange(4096).reshape(8, 8, 8, 8).clone().realize().shard(devices, axis=0)

    with Tensor.train(is_training):
      bns = []
@@ -1163,7 +1184,7 @@ class TestMultiBufferView(unittest.TestCase):

  @unittest.skip("flaky on LLVM")
  def test_shrink_non_shard_axis(self):
-    ref = Tensor.arange(8*4*10).reshape(8, 4, 10).contiguous().realize()
+    ref = Tensor.arange(8*4*10).reshape(8, 4, 10).clone().realize()
    a = Tensor.arange(8*4*10).reshape(8, 4, 10).clone().shard(devices_2, axis=1).realize()
    self._check(ref, a, lambda t: t[3])

@@ -1275,7 +1296,7 @@ class TestMultiSetitem(unittest.TestCase):
  @needs_second_gpu
  def setUp(self): pass

-  def _t(self, axis): return Tensor.arange(16).contiguous().realize().shard(self.device, axis=axis)
+  def _t(self, axis): return Tensor.arange(16).clone().realize().shard(self.device, axis=axis)

  def test_setitem_scalar_axis0(self):
    t = self._t(0)
@@ -1,4 +1,4 @@
-import time, math, unittest, functools, platform, warnings
+import time, math, unittest, functools, platform, warnings, sys
 import numpy as np
 from typing import List, Callable
 import torch
@@ -7,7 +7,6 @@ from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.renderer.cstyle import QCOMCLRenderer
 from tinygrad.renderer.nir import NIRRenderer
-from test.helpers import CI

 TINY_BACKEND = getenv("TINY_BACKEND")
 if TINY_BACKEND:
@@ -74,7 +73,7 @@ def helper_test_op(shps, torch_fxn, tinygrad_fxn=None, atol=1e-6, rtol=1e-3, gra
    for i, (t, torch_grad) in enumerate(zip(tiny_grads, torch_grads)):
      compare(f"backward pass tensor {i}", t.numpy(), torch_grad.detach().cpu().numpy(), atol=grad_atol, rtol=grad_rtol)

-  if not CI:
+  if sys.stdout.isatty():
    print("\ntesting %40r   torch/tinygrad fp: %.2f / %.2f ms  bp: %.2f / %.2f ms " % \
          (shps, torch_fp*1000, tinygrad_fp*1000, torch_fbp*1000, tinygrad_fbp*1000), end="")

@@ -103,7 +102,7 @@ class TestOps(unittest.TestCase):
    with self.assertRaises(expected) as tinygrad_cm:
      tinygrad_fxn(*tst)
    if exact: self.assertEqual(str(torch_cm.exception), str(tinygrad_cm.exception))
-    if not CI: print("\ntesting %40r   torch/tinygrad exception: %s / %s" % (shps, torch_cm.exception, tinygrad_cm.exception), end="")
+    if sys.stdout.isatty(): print("\ntesting %40r   torch/tinygrad exception: %s / %s" % (shps, torch_cm.exception, tinygrad_cm.exception), end="")

  def test_full_like(self):
    a = Tensor([[1,2,3],[4,5,6]], dtype=dtypes.float32)
@@ -1449,9 +1448,7 @@ class TestOps(unittest.TestCase):
                                                                         np.arange(64,128,dtype=np.float32).reshape(8,8)])
  def test_small_gemm_eye(self):
    helper_test_op(None, lambda x,y: x.matmul(y), lambda x,y: x@y, vals=[np.eye(8).astype(np.float32), np.eye(8).astype(np.float32)])
-  @unittest.skipIf(CI and Device.DEFAULT in ["NV", "CL", "CUDA"] or (Device.DEFAULT == "CPU" and DEV.renderer == "LLVM") or IMAGE
-  or (Device.DEFAULT == "WEBGPU" and platform.system() == "Windows"), "not supported on these in CI/IMAGE")
-  @unittest.skipIf(Device.DEFAULT == "QCOM", "not precise enough")
+  @unittest.skipUnless(dtypes.half in Device[Device.DEFAULT].renderer.supported_dtypes(), "not precise enough when emulating")
  def test_gemm_fp16(self):
    helper_test_op([(64,64), (64,64)], lambda x,y: x.half().matmul(y.half()), atol=5e-3, rtol=5e-3, grad_atol=5e-3, grad_rtol=5e-3)
  def test_gemm(self):
@@ -2848,6 +2845,7 @@ class TestOps(unittest.TestCase):
    for dim in range(-1, 2):
      helper_test_op([(45,65), (45,65), (45,65)], lambda x,y,z: torch.cat((x,y,z), dim), lambda x,y,z: x.cat(y, z, dim=dim))

+  @unittest.skipIf(COMPILE_ONLY, "test requires runtime")
  def test_stack(self):
    for dim in range(-1, 3):
      helper_test_op([(5,6,3), (5,6,3), (5,6,3)], lambda x, y, z: torch.stack((x, y, z), dim), lambda x, y, z: Tensor.stack(x, y, z, dim=dim))
@@ -2858,8 +2856,7 @@ class TestOps(unittest.TestCase):
    with self.assertRaises(ValueError):
      Tensor.stack((Tensor([1, 2]), Tensor([3, 4])), Tensor([5, 6]))

-    a = Tensor(3.14)
-    np.testing.assert_allclose(Tensor.stack(a, a).numpy(), Tensor([3.14, 3.14]).numpy())
+    np.testing.assert_allclose(Tensor.stack(Tensor(3.14), Tensor(3.14)).numpy(), np.array([3.14, 3.14]))

  def test_stack_max(self):
    helper_test_op(None, lambda x, y: torch.stack((x, y)).max(axis=0)[0], lambda x, y: Tensor.stack(x, y).max(axis=0), vals=[[1.], [2.]])
@@ -3349,6 +3346,7 @@ class TestOps(unittest.TestCase):
    # fill_value must not alter output dtype
    self.assertEqual(Tensor([1.0, 2.0]).masked_select(Tensor([True, False]), size=3, fill_value=-1).dtype, dtypes.default_float)

+  @slow_test
  def test_nonzero(self):
    helper_test_op([(32, 10)], lambda x: (x>0.5).nonzero().int(), lambda x: (x>0.5).nonzero(), forward_only=True)
    helper_test_op([(20,)], lambda x: (x>0.5).nonzero().int(), lambda x: (x>0.5).nonzero(), forward_only=True)
@@ -67,7 +67,7 @@ class TestPickle(unittest.TestCase):

  # NOTE: currently Buffer exists on the uop, not tensor
  def test_pickle_buffer_uop(self):
-    t = Tensor.arange(4).realize()
+    t = Tensor.arange(4).clone().realize()
    a = t.uop
    assert a.is_realized
    self.assertIsNotNone(buffer:=a.base.realized)
@@ -95,7 +95,7 @@ class TestPickle(unittest.TestCase):
    np.testing.assert_equal(vt2.numpy(), 20)

  def test_pickle_buffer_view(self):
-    t = Tensor.arange(10, device="CPU").contiguous().realize()
+    t = Tensor.arange(10).clone(device="CPU").realize()
    vt = t[3:5].contiguous().realize()
    assert hasattr(vt.uop.buffer, 'base')
    ref_value = vt.tolist()
@@ -5,7 +5,6 @@ from tinygrad.device import Buffer, BufferSpec, Compiled, ProfileDeviceEvent, Pr
 from tinygrad.runtime.support.hcq import HCQCompiled
 from tinygrad.engine.realize import get_runtime
 from tinygrad.codegen import to_program
-from test.helpers import CI

 MOCKGPU = DEV.interface.startswith("MOCK")
 def _dev_base(d):
@@ -145,7 +144,8 @@ class TestProfiler(unittest.TestCase):
    assert len(graph_evs) == 2, "2 graph events are expected"
    assert len(graph_evs[0].ents) == 2, "two entities are expected"

-  @unittest.skipIf(CI or not issubclass(type(Device[Device.DEFAULT]), HCQCompiled), "skip CI")
+  @unittest.skipIf(MOCKGPU, "skip MOCKGPU")
+  @unittest.skipUnless(issubclass(type(Device[Device.DEFAULT]), HCQCompiled), "must be HCQ")
  def test_dev_jitter_matrix(self):
    dev_cnt = 6
    try: devs = [Device[f"{Device.DEFAULT}:{i}"] for i in range(dev_cnt)]
@@ -1,15 +1,15 @@
 import unittest, math
-from functools import partial

-from tinygrad import nn, dtypes, Tensor, Device, TinyJit, Variable
-from tinygrad.helpers import getenv, OSX
+from tinygrad import dtypes, Tensor, Device
+from tinygrad.helpers import getenv, DEV
 from tinygrad.codegen import to_program

 from tinygrad.uop.ops import Ops
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.nir import NIRRenderer
 from tinygrad.renderer.isa.x86 import X86Renderer
-from test.helpers import not_support_multi_device, needs_second_gpu, CI
+from test.helpers import not_support_multi_device, needs_second_gpu
+from test.unit.test_randomness import equal_distribution, normal_test

 import numpy as np
 import torch
@@ -18,54 +18,6 @@ from hypothesis import given, settings, strategies as strat
 settings.register_profile("my_profile", max_examples=200, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
 settings.load_profile("my_profile")

-# https://gist.github.com/devries/11405101
-def ksprob(a):
-  fac, total, termbf = 2.0, 0.0, 0.0
-  a2 = -2.0 * a * a
-  for j in range(1, 101):
-    term = fac * math.exp(a2 * j * j)
-    total += term
-    if math.fabs(term) <= 0.001 * termbf or math.fabs(term) <= 1e-8 * total:
-      return total
-    fac = -fac
-    termbf = math.fabs(term)
-  return 1.0
-
-def kstest(l1, l2):
-  n1, n2 = len(l1), len(l2)
-  l1.sort()
-  l2.sort()
-  j1, j2, d, fn1, fn2 = 0, 0, 0.0, 0.0, 0.0
-  while j1 < n1 and j2 < n2:
-    d1, d2 = l1[j1], l2[j2]
-    if d1 <= d2:
-      fn1 = (float(j1) + 1.0) / float(n1)
-      j1 += 1
-    if d2 <= d1:
-      fn2 = (float(j2) + 1.0) / float(n2)
-      j2 += 1
-    dtemp = math.fabs(fn2 - fn1)
-    if dtemp > d:
-      d = dtemp
-  ne = float(n1 * n2) / float(n1 + n2)
-  nesq = math.sqrt(ne)
-  prob = ksprob((nesq + 0.12 + 0.11 / nesq) * d)
-  return prob
-
-def equal_distribution(tiny_func, torch_func=None, numpy_func=None, shape=(40, 43), alpha=0.04):
-  Tensor.manual_seed(1337)
-  torch.manual_seed(1337)
-  np.random.seed(1337)
-  assert not (torch_func is None and numpy_func is None), "no function to compare with"
-  x1 = tiny_func(*shape).numpy().flatten()
-  x2 = tiny_func(shape).numpy().flatten()
-  if numpy_func is not None: y = numpy_func(shape).flatten()
-  if torch_func is not None: z = torch_func(shape).numpy().flatten()
-  return (numpy_func is None or (kstest(x1, y) >= alpha and kstest(x2, y) >= alpha)) and \
-    (torch_func is None or (kstest(x1, z) >= alpha and kstest(x2, z) >= alpha))
-
-def normal_test(func, shape=(20, 45), alpha=0.05): return equal_distribution(func, numpy_func=lambda x: np.random.randn(*x), shape=shape, alpha=alpha)
-
 class TestRandomness(unittest.TestCase):
  def test_rand(self):
    self.assertFalse(normal_test(Tensor.rand))
@@ -96,7 +48,7 @@ class TestRandomness(unittest.TestCase):
    assert nx[nx == 0].size > 0
    equal_distribution(lambda *x: Tensor.rand(*x, dtype=dtypes.float16), torch.rand, lambda x: np.random.rand(*x), shape=(2, N, N))

-  @unittest.skipIf(CI and Device.DEFAULT in {"NV", "CUDA"}, "gpuocelot doesn't support certain ops needed for threefry")
+  @unittest.skipIf(DEV.interface.startswith("MOCK") and Device.DEFAULT in {"NV", "CUDA"}, "gpuocelot doesn't support certain ops needed for threefry")
  def test_threefry_against_reference(self):
    Tensor.manual_seed(1337)

@@ -274,11 +226,6 @@ class TestRandomness(unittest.TestCase):
    assert rand.dtype == dtypes.float16
    assert rand.device == empty.device

-  def test_randn(self):
-    self.assertEqual(Tensor.randn(3,3,dtype=dtypes.half).dtype, dtypes.half)
-    self.assertTrue(normal_test(Tensor.randn))
-    self.assertTrue(equal_distribution(Tensor.randn, torch.randn, lambda x: np.random.randn(*x)))
-
  def test_randn_device(self):
    self.assertEqual(Tensor.randn(3,3,device="CPU").device, "CPU")

@@ -296,128 +243,6 @@ class TestRandomness(unittest.TestCase):
    assert math.isfinite(mn), mn
    dtypes.default_float = old_default_float

-  def test_randint(self):
-    self.assertFalse(normal_test(Tensor.randint))
-    self.assertTrue(equal_distribution(partial(Tensor.randint, low=-2, high=5),
-                                       numpy_func=lambda x: np.random.randint(low=-2, high=5, size=x)))
-    self.assertTrue(equal_distribution(partial(Tensor.randint, low=-2, high=5, dtype="int32"),
-                                       numpy_func=lambda x: np.random.randint(low=-2, high=5, size=x)))
-    self.assertTrue(Tensor.randint(1, device="CPU").device=="CPU")
-    # check types of args
-    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0.1, high=3)
-    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0, high=3.5)
-    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=1, high=3, dtype="float")
-    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0, high=3, dtype=dtypes.float32)
-    # check low < high
-    with self.assertRaises(ValueError): Tensor.randint((3, 4), low=10, high=5)
-    with self.assertRaises(ValueError): Tensor.randint((3, 4), low=10, high=10)
-    np.testing.assert_array_equal(Tensor.randint(16, low=5, high=6).numpy(), 5)
-
-  def test_normal(self):
-    self.assertTrue(normal_test(Tensor.normal))
-    self.assertTrue(equal_distribution(Tensor.normal, lambda x: torch.nn.init.normal_(torch.empty(x), mean=0, std=1),
-                                                      lambda x: np.random.normal(loc=0, scale=1, size=x)))
-    # check std >= 0
-    with self.assertRaises(ValueError): Tensor.normal((3, 4), mean=0, std=-1)
-
-  def test_uniform(self):
-    self.assertFalse(normal_test(Tensor.uniform))
-    self.assertTrue(equal_distribution(Tensor.uniform, lambda x: torch.nn.init.uniform_(torch.empty(x)), lambda x: np.random.uniform(size=x)))
-    self.assertTrue(equal_distribution(partial(Tensor.uniform, low=-100, high=100, dtype=dtypes.int32),
-                                       numpy_func=lambda x: np.random.randint(low=-100, high=100, size=x)))
-    # check low < high
-    with self.assertRaises(ValueError): Tensor.uniform((3, 4), low=5.0, high=3.0)
-    with self.assertRaises(ValueError): Tensor.uniform((3, 4), low=1.0, high=1.0)
-
-  def test_scaled_uniform(self):
-    self.assertFalse(normal_test(Tensor.scaled_uniform))
-    self.assertTrue(equal_distribution(Tensor.scaled_uniform, lambda x: torch.nn.init.uniform_(torch.empty(x), a=-1, b=1) / math.sqrt(math.prod(x)),
-                                                              lambda x: np.random.uniform(-1, 1, size=x) / math.sqrt(math.prod(x))))
-
-  def test_glorot_uniform(self):
-    self.assertFalse(normal_test(Tensor.glorot_uniform))
-    self.assertTrue(equal_distribution(Tensor.glorot_uniform, lambda x: torch.nn.init.xavier_uniform_(torch.empty(x)),
-                                                              lambda x: np.random.uniform(-1, 1, size=x) * math.sqrt(6 / (x[0] + math.prod(x[1:])))))
-
-  def test_kaiming_uniform(self):
-    for shape in [(32, 16, 3, 3), (20, 44), (5, 15, 35)]:
-      self.assertTrue(equal_distribution(Tensor.kaiming_uniform, lambda x: torch.nn.init.kaiming_uniform_(torch.empty(x)), shape=shape))
-
-  def test_kaiming_normal(self):
-    for shape in [(32, 16, 3, 3), (20, 44), (3, 15, 35)]:
-      self.assertTrue(equal_distribution(Tensor.kaiming_normal, lambda x: torch.nn.init.kaiming_normal_(torch.empty(x)), shape=shape))
-
-  def test_multinomial(self):
-    self.assertRaises(AssertionError, lambda: Tensor(2).multinomial(1, replacement=False))
-    self.assertRaises(AssertionError, lambda: Tensor([1, 9]).multinomial(0, replacement=False))
-    def _check_with_torch(w, num_samples, replacement):
-      tiny_res = Tensor(w).multinomial(num_samples, replacement=replacement)
-      torch_res = torch.tensor(w).multinomial(num_samples, replacement=replacement)
-      self.assertEqual(tiny_res.shape, torch_res.shape)
-      if torch_res.ndim == 1:
-        tiny_res = tiny_res.unsqueeze(0)
-        torch_res = torch_res.unsqueeze(0)
-      for i in range(torch_res.shape[0]):
-        self.assertTrue(equal_distribution(lambda *_: tiny_res[i], lambda _: torch_res[i]))
-    _check_with_torch(w=[0.231, 0., 1., 0.5], num_samples=300, replacement=True)
-    _check_with_torch(w=[[0.2, 0.8]], num_samples=300, replacement=True)  # 2D but only 1 row
-    _check_with_torch(w=[[0.453, 0., 1., 0.81], [0.1, 0.8, 0., 0.1]], num_samples=300, replacement=True)
-    # no-replacement
-    w = [0.1, 0.9]
-    self.assertRaises(AssertionError, lambda: Tensor(w).multinomial(100, replacement=False))
-
-    @TinyJit
-    def sample_one(): return Tensor(w).multinomial(1, replacement=False).realize()
-
-    tiny_samples = [sample_one().item() for _ in range(400)]
-    torch_samples = [torch.tensor(w).multinomial(1, replacement=False).item() for _ in range(400)]
-    self.assertTrue(equal_distribution(lambda *_: Tensor(tiny_samples), lambda _: torch.tensor(torch_samples)))
-
-    w = list(range(32))
-    s1 = Tensor(w).multinomial(5, replacement=False).numpy()
-    self.assertEqual(len(set(s1.tolist())), 5)
-    s2 = Tensor(w).multinomial(5, replacement=False).numpy()
-    self.assertFalse(np.array_equal(s1, s2))
-    full = Tensor(w).multinomial(len(w), replacement=False).numpy()
-    self.assertEqual(sorted(full.tolist()), w)
-
-    w = [0.1, 0.2, 0.3, 0.4]
-    @TinyJit
-    def sample_three(): return Tensor(w).multinomial(3, replacement=False).realize()
-
-    tiny_draws = np.array([sample_three().numpy() for _ in range(400)])
-    torch_draws = np.array([torch.tensor(w).multinomial(3, replacement=False).numpy() for _ in range(400)])
-    for pos in range(3):
-      self.assertTrue(equal_distribution(lambda *_: Tensor(tiny_draws[:, pos]), lambda _: torch.tensor(torch_draws[:, pos])))
-
-  @unittest.skip("this test is flaky")
-  def test_multinomial_counterexample(self):
-    tiny_res = Tensor([0.3, 0.6, 0.1]).multinomial(4000, replacement=True)
-    torch_res = torch.tensor([0.3, 0.6, 0.1]).multinomial(4000, replacement=True)
-    self.assertTrue(equal_distribution(lambda *_: tiny_res, lambda _: torch_res))
-    torch_res = torch.tensor([0.2, 0.7, 0.1]).multinomial(4000, replacement=True)
-    self.assertFalse(equal_distribution(lambda *_: tiny_res, lambda _: torch_res))
-
-  def test_conv2d_init(self):
-    params = (128, 256, (3,3))
-    assert equal_distribution(lambda *_: nn.Conv2d(*params).weight, lambda _: torch.nn.Conv2d(*params).weight.detach())
-    assert equal_distribution(lambda *_: nn.Conv2d(*params).bias, lambda _: torch.nn.Conv2d(*params).bias.detach())
-
-  def test_linear_init(self):
-    params = (64, 256)
-    assert equal_distribution(lambda *_: nn.Linear(*params).weight, lambda _: torch.nn.Linear(*params).weight.detach())
-    assert equal_distribution(lambda *_: nn.Linear(*params).bias, lambda _: torch.nn.Linear(*params).bias.detach())
-
-  def test_bn_init(self):
-    params = (64,)
-    assert equal_distribution(lambda *_: nn.BatchNorm2d(*params).weight, lambda _: torch.nn.BatchNorm2d(*params).weight.detach())
-    assert equal_distribution(lambda *_: nn.BatchNorm2d(*params).bias, lambda _: torch.nn.BatchNorm2d(*params).bias.detach())
-
-  def test_rand_chain(self):
-    # NOTE: this fails if property propagates deeper than stack limit
-    for _ in range(833): Tensor.rand(1)
-    Tensor.rand(1).schedule_linear()
-
  def test_random_counter_overflow(self):
    device = Device.DEFAULT
    Tensor.manual_seed(1337)
@@ -433,20 +258,5 @@ class TestRandomness(unittest.TestCase):
    c = Tensor._device_rng_counters[device].numpy()
    np.testing.assert_allclose(c, [14, 1])

-# TODO: still fails with MAX_KERNEL_BUFFERS
-@unittest.skipIf(Device.DEFAULT == "WEBGPU" and not OSX, "WEBGPU Vulkan can only run kernels with up to 10 buffers")
-class TestSample(unittest.TestCase):
-  def test_sample(self):
-    X = Tensor.rand(1000, 50).realize()
-    BS = 16
-    idxs = np.random.randint(0, X.shape[0], size=(BS))
-    # this uncovered a bug with arg sort order
-    batch = [Variable(f'idx{i}', 0, X.shape[0]-1).bind(s) for i,s in enumerate(idxs.tolist())]
-    x = Tensor.cat(*[X.shrink(((batch[i], batch[i]+1), None)) for i in range(BS)])
-    print(idxs)
-    ret = x.numpy()
-    base = X.numpy()[idxs]
-    np.testing.assert_equal(ret, base)
-
 if __name__ == "__main__":
  unittest.main()
@@ -22,8 +22,8 @@ def _test_uop_result(inputs:list[Tensor], sink:UOp, local_size=None):

 def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
  dtype = alu_src_uops[0].dtype
-  a = UOp(Ops.PARAM, dtype.ptr(), (), 0)
-  b = UOp(Ops.PARAM, dtype.ptr(), (), 1)
+  a = UOp.param(0, dtype.ptr())
+  b = UOp.param(1, dtype.ptr())
  idx = UOp.const(dtypes.int, 0)
  ld = b.index(idx)
  alu = ld.alu(alu_op, *alu_src_uops)
@@ -33,7 +33,7 @@ def _setup_and_test_alu(alu_op:Ops, input_val:ConstType, *alu_src_uops:UOp):
 class TestRendererFailures(unittest.TestCase):
  @unittest.skipIf(not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, PythonRenderer)), "test is for ptx or python renderer")
  def test_gated_store_with_alu(self):
-    a = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    a = UOp.param(0, dtypes.int.ptr())
    gate_alu = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index(lidx0.valid(gate_alu)), UOp.const(dtypes.int, 1)))
    sink = UOp(Ops.SINK, dtypes.void, (gated_alu_store,), arg=KernelInfo())
@@ -42,7 +42,7 @@ class TestRendererFailures(unittest.TestCase):

  @unittest.skipIf(not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, PythonRenderer)), "test is for ptx or python renderer")
  def test_gated_store_with_alu_2d(self):
-    a = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    a = UOp.param(0, dtypes.int.ptr())
    gate_alu_0 = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
    gate_alu_1 = (lidx1:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 2),), 'lidx1')).ne(0)
    gated_alu_store = UOp(Ops.STORE, dtypes.void, (a.index((lidx0+lidx1*4).valid(gate_alu_0&gate_alu_1)), UOp.const(dtypes.int, 1)))
@@ -87,7 +87,7 @@ class TestWGSLFailures(unittest.TestCase):
 class TestPTXFailures(unittest.TestCase):
  @unittest.skip("INDEX can only have a gate ALU parent, not an IF")
  def test_gated_store_with_if(self):
-    a = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    a = UOp.param(0, dtypes.int.ptr())
    gate_alu = (lidx0:=UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'lidx0')).ne(0)
    val = UOp.const(dtypes.int, 1)
    if_uop = UOp(Ops.IF, dtypes.void, (gate_alu,))
@@ -10,9 +10,8 @@ from hypothesis import assume, given, strategies as strat
 from tinygrad import nn, dtypes, Device, Tensor, Variable
 from tinygrad.dtype import DType
 from tinygrad.uop.ops import UOp, Ops, UPat
-from tinygrad.helpers import DEBUG, OSX, GlobalCounters, Context, getenv, all_same, temp
+from tinygrad.helpers import DEBUG, DEV, OSX, GlobalCounters, Context, getenv, all_same, temp
 from tinygrad.engine.realize import compile_linear, run_linear
-from test.helpers import CI

 supported_dtypes = Device[Device.DEFAULT].renderer.supported_dtypes()

@@ -83,13 +82,6 @@ class TestSchedule(unittest.TestCase):
  def test_arange_avgpool2d_fused_noopt(self):
    with Context(NOOPT=1): self.test_arange_avgpool2d(kcount=1)

-  # linearizer error
-  @unittest.skip("recursion error no longer raised")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.supports_float4, "needs supports_float4 to fail")
-  def test_arange_avgpool2d_fused(self):
-    with self.assertRaises(RecursionError):
-      with Context(NOOPT=0): self.test_arange_avgpool2d(kcount=1)
-
  # when we're fusing a reduce, all ReduceOps must have the same N in the dimensions
  # all permutes, reshapes, expands and shrinks push through the reduce
  def test_arange_sum(self):
@@ -115,7 +107,6 @@ class TestSchedule(unittest.TestCase):
    run_linear(*check_schedule(b, 1))
    np.testing.assert_allclose(b.numpy(), np.broadcast_to(a.numpy().astype(np.float16), (2, 4, 4))+2, rtol=1e-3)

-  @unittest.skipIf(CI and Device.DEFAULT == "NV", "crashes on NV CI")
  def test_add_chain_buffers(self):
    N = 31
    with Context(TRACK_MATCH_STATS=0, DEBUG=0):
@@ -149,11 +140,6 @@ class TestSchedule(unittest.TestCase):
    run_linear(*check_schedule(z, 1, [x,y]))
    self.assertEqual(z.item(), 32)

-  def test_constants_can_store(self):
-    a = Tensor(2).contiguous()
-    run_linear(*check_schedule(a, 1))
-    np.testing.assert_equal(a.numpy(), 2)
-
  def test_allow_push_permutes(self):
    a = Tensor.randn(10,10,10).realize()
    b = Tensor.randn(10,10,1).realize()
@@ -235,8 +221,8 @@ class TestSchedule(unittest.TestCase):
    np.testing.assert_allclose(out.numpy(), (x.numpy() - x.numpy().max(keepdims=True)).max())

  def test_example_matmul_contig(self):
-    x = Tensor.eye(64).contiguous().realize()
-    y = Tensor.eye(64).contiguous().realize()
+    x = Tensor.eye(64).clone().realize()
+    y = Tensor.eye(64).clone().realize()
    z = y.matmul(x).sum()
    z.backward()
    out = x.grad.contiguous()
@@ -712,7 +698,7 @@ class TestSchedule(unittest.TestCase):
    x = Tensor.empty(3,3,3,3)
    y = x.pad((-1,2,2,-1), mode="replicate")
    dx = y.sum().gradient(x)[0]
-    sched = check_schedule(dx, 1)
+    sched = check_schedule(dx, 0)
    run_linear(*sched)
    np.testing.assert_allclose(dx.numpy(), [[[[0.,3.,9.],[0,1.,3.],[0.,0.,0.]]]*3]*3)

@@ -777,7 +763,6 @@ class TestSchedule(unittest.TestCase):
    gc.collect()
    self.assertEqual(GlobalCounters.mem_used-base, 1024)

-  @unittest.skipIf(Device.DEFAULT != "CL", "image only supported on CL")
  def test_image_dot_f16_fusion(self):
    with Context(FLOAT16=1, OPENPILOT_HACKS=1):
      def cnt():
@@ -789,7 +774,6 @@ class TestSchedule(unittest.TestCase):
      with Context(IMAGE=1):
        self.assertEqual(cnt(), 5)

-  @unittest.skipIf(Device.DEFAULT != "CL", "image only supported on CL")
  def test_image_f16_residual_fusion(self):
    with Context(FLOAT16=1, OPENPILOT_HACKS=1):
      def cnt():
@@ -805,7 +789,6 @@ class TestSchedule(unittest.TestCase):
      with Context(IMAGE=1):
        self.assertEqual(cnt(), 9)

-  @unittest.skipIf(Device.DEFAULT != "CL", "image only supported on CL")
  def test_image_conv_fusion(self):
    with Context(OPENPILOT_HACKS=1):
      def cnt():
@@ -843,7 +826,7 @@ class TestSchedule(unittest.TestCase):
    self._test_fusion([(32, 32)], lambda a:a-a.sum(1), 2)

  def test_cast_padded_view(self):
-    a = Tensor.arange(4).reshape(1, 4)
+    a = Tensor.arange(4).reshape(1, 4).clone().realize()
    casted_view = a.pad(((0, 1), (0, 0))).cast(dtypes.float)
    casted_view.realize()
    self.assertEqual(casted_view.uop.base.realized.size, 8)
@@ -853,7 +836,7 @@ class TestSchedule(unittest.TestCase):

  # NOTE: we only reorder CAST if it's an EXPAND
  def test_cast_after_shrink(self):
-    a = Tensor.arange(4).reshape(1, 4)
+    a = Tensor.arange(4).reshape(1, 4).clone().realize()
    casted_view = a.shrink(((0, 1), (0, 2))).cast(dtypes.float)
    casted_view.realize()
    self.assertEqual(casted_view.uop.base.realized.size, 2)
@@ -862,7 +845,7 @@ class TestSchedule(unittest.TestCase):
    self.assertListEqual(realized_view.tolist(), [[0, 1]])

  def test_cast_const_view(self):
-    a = Tensor.ones((4, 4), dtype=dtypes.float32)
+    a = Tensor.ones((4, 4), dtype=dtypes.float32, buffer=False)
    casted_view = a.cast(dtypes.int32)
    run_linear(*check_schedule(casted_view, 1))
    realized_const_view = casted_view.contiguous()
@@ -942,7 +925,7 @@ class TestSchedule(unittest.TestCase):
    np.testing.assert_equal(a.numpy(), (np.arange(4)*x.numpy()).T.sum())

  def test_div_padded_arange(self):
-    x = Tensor.full((2,2), 16)
+    x = Tensor.full((2,2), 16, buffer=False)
    y = x.div(Tensor.linspace(2, 8, steps=4, dtype=dtypes.int).reshape(2,2), rounding_mode="trunc").pad(((1,1), (1,1)))
    out = y.sum(axis=1)
    run_linear(*check_schedule(out, 1))
@@ -989,13 +972,6 @@ class TestSchedule(unittest.TestCase):
    run_linear(*check_schedule(out, 2))
    np.testing.assert_allclose(out.numpy(), (x.numpy()+(np.arange(10)+1)[2]).sum(), atol=1e-5, rtol=1e-6)

-  @unittest.skip("BUFFER_VIEW no longer supported on non-disk devices")
-  def test_arange_view_op(self):
-    a = Tensor.arange(12).reshape(4, 3).shrink(((1, 2), (1, 3))).contiguous()
-    sched = run_linear(*check_schedule(a, 1))
-    self.assertIs(sched[1].ast.op, Ops.BUFFER_VIEW)
-    np.testing.assert_equal(a.numpy(), [[4, 5]])
-
  @unittest.skipUnless(dtypes.half in supported_dtypes, "need half")
  def test_precompute_freqs_cis(self):
    from extra.models.llama import precompute_freqs_cis
@@ -1015,7 +991,7 @@ class TestSchedule(unittest.TestCase):

  def test_assign_non_contiguous_alt(self): self.test_assign_non_contiguous(alt=True)
  def test_assign_non_contiguous(self, alt=False):
-    x = (Tensor.arange(16)-100).reshape(4,4).contiguous().realize()
+    x = (Tensor.arange(16)-100).reshape(4,4).clone().realize()
    xref = x.numpy()
    if alt:
      y = Tensor.randint(2, 4).contiguous().realize()
@@ -1031,7 +1007,7 @@ class TestSchedule(unittest.TestCase):
    np.testing.assert_equal(tst.numpy(), a.numpy())

  def test_setitem_sched(self, mop=lambda x:x, expected_kcount=1):
-    a = Tensor.arange(16, device="CPU").reshape(4, 4).contiguous().realize()
+    a = Tensor.arange(16).reshape(4, 4).clone(device="CPU").realize()
    a2 = mop(a)
    expected = (a+a2).tolist()
    a.assign(a+a2)
@@ -1045,7 +1021,7 @@ class TestSchedule(unittest.TestCase):

  def test_setitem_const_fused(self):
    # https://github.com/tinygrad/tinygrad/issues/10690
-    a = Tensor.arange(16).contiguous().realize()
+    a = Tensor.arange(16).clone().realize()
    GlobalCounters.reset()
    a[4] = 3
    self.assertEqual(GlobalCounters.kernel_count, 0)
@@ -1082,11 +1058,10 @@ class TestSchedule(unittest.TestCase):
    self.assertEqual(a.tolist(), [0., 0.])
    self.assertEqual(b.tolist(), [False, False])

-  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "Validation error on WebGPU")
  def test_mnist_val(self):
-    from tinygrad.nn.datasets import mnist
+    # from tinygrad.nn.datasets import mnist
    import torch
-    _, Y_train, _, _ = mnist()
+    Y_train = Tensor.randint(60000, dtype='uchar').realize()
    samples = Tensor.randint(BS:=getenv("BS", 512), high=cast(int,Y_train.shape[-1])).realize()
    yt = Tensor.randn(BS, 10).realize()
    loss = yt.sparse_categorical_crossentropy(Y_train[samples])
@@ -1119,7 +1094,7 @@ class TestSchedule(unittest.TestCase):
      self.assertListEqual(a.tolist(), [[1.]*shape[1]]*shape[0])

 class TestLimitBufs(unittest.TestCase):
-  @unittest.skipIf(CI and Device.DEFAULT == "NV", "crashes on NV CI")
+  @unittest.skipIf(DEV.interface.startswith("MOCK") and Device.DEFAULT == "NV", "crashes in ocelot")
  def test_limit_bufs_with_var(self):
    N = 31
    with Context(TRACK_MATCH_STATS=0, DEBUG=0):
@@ -1289,7 +1264,7 @@ class TestView(unittest.TestCase):
 class TestCopyFolding(unittest.TestCase):
  def test_const_copy_is_free(self):
    b = Tensor(1).to("CPU") * 4
-    run_linear(*check_schedule(b, 1, filter_sink=False))
+    run_linear(*check_schedule(b, 0, filter_sink=False))
    assert b.item() == 4

  def test_one_hot_with_copy(self):
@@ -1298,13 +1273,13 @@ class TestCopyFolding(unittest.TestCase):
    check_schedule(x, 3, filter_sink=False)

  def test_const_copy_multi(self):
-    x = Tensor.ones(1, device="CPU").to_(["CPU", "CPU:1"]) * 2
+    x = Tensor.ones(1, device="CPU", buffer=False).to_(["CPU", "CPU:1"]) * 2
    run_linear(*check_schedule(x, 2, filter_sink=False))
    self.assertEqual(x.item(), 2.0)

  def test_late_const_copy_folding(self):
-    a = Tensor.arange(3).realize()
-    zeros = Tensor.zeros(3).realize()
+    a = Tensor.arange(3).clone().realize()
+    zeros = Tensor.zeros(3, buffer=False).realize()
    b = (a*zeros).to("CPU") + 1
    run_linear(*check_schedule(b, 1, filter_sink=False))
    self.assertListEqual(b.tolist(), [1, 1, 1])
@@ -1378,14 +1353,14 @@ class TestCopyFolding(unittest.TestCase):
    self.assertListEqual(b.tolist(), [[0, 2], [1, 3]])

  def test_permute_on_disk(self):
-    with open(temp('dt_arange_4_permute'), "wb") as f: f.write(Tensor.arange(4).realize().uop.base.buffer.as_memoryview())
+    with open(temp('dt_arange_4_permute'), "wb") as f: f.write(Tensor.arange(4).clone().realize().uop.base.buffer.as_memoryview())
    a = Tensor.empty(4, dtype=dtypes.int32, device=f"disk:{temp('dt_arange_4_permute')}")
    b = a.reshape(2, 2).permute(1, 0).to("CPU")
    b.realize()
    self.assertListEqual(b.tolist(), [[0, 2], [1, 3]])

  def test_permute_on_disk_contiguous(self):
-    with open(temp('dt_arange_4_permute_contig'), "wb") as f: f.write(Tensor.arange(4).realize().uop.base.buffer.as_memoryview())
+    with open(temp('dt_arange_4_permute_contig'), "wb") as f: f.write(Tensor.arange(4).clone().realize().uop.base.buffer.as_memoryview())
    a = Tensor.empty(4, dtype=dtypes.int32, device=f"disk:{temp('dt_arange_4_permute_contig')}")
    b = a.reshape(2, 2).permute(1, 0).contiguous().to("CPU")
    b.realize()
@@ -1399,7 +1374,7 @@ class TestCopyFolding(unittest.TestCase):

  # NOTE: disk permute must come after COPY
  def test_permute_after_shrink_on_disk(self):
-    with open(temp('dt_arange_5_permute'), "wb") as f: f.write(Tensor.arange(5).realize().uop.base.buffer.as_memoryview())
+    with open(temp('dt_arange_5_permute'), "wb") as f: f.write(Tensor.arange(5).clone().realize().uop.base.buffer.as_memoryview())
    a = Tensor.empty(5, dtype=dtypes.int32, device=f"disk:{temp('dt_arange_5_permute')}")
    b = a.shrink(((0, 4),)).reshape(2, 2).permute(1, 0).to("CPU")
    b.realize()
@@ -32,13 +32,14 @@ class TestSetitem(unittest.TestCase):
    self.assertListEqual(t.tolist(), [0, 1, 11, 3, 11, 5, 6, 7, 8, 9])

  def test_setitem_inplace_mul(self):
-    t = Tensor.arange(10).realize()
+    t = Tensor.arange(10).clone().realize()
    t[:3] *= 10
    self.assertListEqual(t.tolist(), [0, 10, 20, 3, 4, 5, 6, 7, 8, 9])

+  @unittest.skip("crashed in LLVM CI")
  def test_setitem_fancy_on_unrealized_view(self):
    # fancy indexing setitem on unrealized SHRINK view (triggered infinite loop in graph_rewrite)
-    base = Tensor.arange(20, dtype=dtypes.float).reshape(4, 5)
+    base = Tensor.arange(20, dtype=dtypes.float).reshape(4, 5).clone().realize()
    sub = base[1:3]
    flat = sub.reshape(sub.numel()).contiguous()
    idx = Tensor([0, 3, 7, 9])
@@ -229,7 +230,7 @@ class TestSetitem(unittest.TestCase):
    np.testing.assert_equal(t.numpy(), n)

  def test_setitem_swap_rows(self):
-    t = Tensor.arange(6, dtype=dtypes.float).reshape(3, 2).contiguous().realize()
+    t = Tensor.arange(6, dtype=dtypes.float).reshape(3, 2).clone().realize()
    tmp = t[0]
    t[0] = t[1]
    t[2] = tmp
@@ -237,7 +238,7 @@ class TestSetitem(unittest.TestCase):
    np.testing.assert_allclose(t.numpy(), [[2, 3], [2, 3], [2, 3]])

    # eager version
-    t = Tensor.arange(6, dtype=dtypes.float).reshape(3, 2).contiguous().realize()
+    t = Tensor.arange(6, dtype=dtypes.float).reshape(3, 2).clone().realize()
    tmp = t[0].realize()
    t[0] = t[1].realize()
    t[2] = tmp.realize()
@@ -269,8 +270,8 @@ class TestSetitem(unittest.TestCase):
  def test_cross_assign_independence(self):
    # when assigning to two tensors using computations from both,
    # both assigns should see the OLD values of both tensors
-    a = Tensor.arange(4, dtype=dtypes.float).contiguous().realize()
-    b = Tensor.arange(4, 8, dtype=dtypes.float).contiguous().realize()
+    a = Tensor.arange(4, dtype=dtypes.float).clone().realize()
+    b = Tensor.arange(4, 8, dtype=dtypes.float).clone().realize()
    new_a = a + b    # [4, 6, 8, 10]
    new_b = a * 2    # [0, 2, 4, 6] -- should use OLD a
    a.assign(new_a)
@@ -283,8 +284,8 @@ class TestSetitem(unittest.TestCase):
      np.testing.assert_allclose(b.numpy(), [8, 12, 16, 20])

    # eager version
-    a = Tensor.arange(4, dtype=dtypes.float).contiguous().realize()
-    b = Tensor.arange(4, 8, dtype=dtypes.float).contiguous().realize()
+    a = Tensor.arange(4, dtype=dtypes.float).clone().realize()
+    b = Tensor.arange(4, 8, dtype=dtypes.float).clone().realize()
    new_a = (a + b).realize()
    new_b = (a * 2).realize()
    a.assign(new_a).realize()
@@ -322,7 +323,7 @@ class TestWithGrad(unittest.TestCase):

  def test_set_overlapping_backward(self):
    z = Tensor.zeros(6)
-    x = Tensor.ones(4)
+    x = Tensor.ones(4).contiguous()
    y = Tensor.ones(4) * 2
    z[:4] = x
    z[2:] = y
@@ -36,7 +36,7 @@ class TestSubBuffer(unittest.TestCase):
    assert len(mv) == 5

  def test_subbuffer_used(self):
-    t = Tensor.arange(0, 10, dtype=dtypes.uint8).realize()
+    t = Tensor.arange(0, 10, dtype=dtypes.uint8).clone().realize()
    vt = t[2:4].realize()
    out = (vt + 100).tolist()
    assert out == [102, 103]
@@ -44,7 +44,7 @@ class TestSubBuffer(unittest.TestCase):
  @needs_second_gpu
  @unittest.skipIf(Device.DEFAULT not in {"CUDA", "NV", "AMD"} or DEV.interface.startswith("MOCK"), "only NV, AMD, CUDA")
  def test_subbuffer_transfer(self):
-    t = Tensor.arange(0, 10, dtype=dtypes.uint8).realize()
+    t = Tensor.arange(0, 10, dtype=dtypes.uint8).clone().realize()
    vt = t[2:5].contiguous().realize()
    out = vt.to(f"{Device.DEFAULT}:1").realize().tolist()
    assert out == [2, 3, 4]
@@ -551,7 +551,7 @@ class TestTinygrad(unittest.TestCase):
      Tensor.zeros(2, 2).realize()

  def test_shrink(self):
-    t = Tensor.arange(32).contiguous().realize()
+    t = Tensor.arange(32).clone().realize()
    self.assertListEqual(t[16:20].tolist(), [16,17,18,19])
    self.assertListEqual(t.shrink_to(16).tolist(), list(range(16)))
    t = t.reshape(4, 8).contiguous().realize()
@@ -10,7 +10,7 @@ class TestTensorVariable(unittest.TestCase):

  def test_inner_tvar_node(self):
    vv = Variable("w", 0, 10).bind(2)
-    ret = Tensor.from_uop(vv * 4).item()
+    ret = Tensor(vv * 4).item()
    assert ret == 8

  def test_inner_tvar_mul(self):
@@ -2,7 +2,6 @@ import unittest
 from tinygrad import Tensor, Device, dtypes
 from tinygrad.tensor import _to_np_dtype
 from tinygrad.helpers import Context, getenv, DEV, OSX
-from test.helpers import CI
 from test.backend.test_schedule import check_schedule
 from test.backend.test_dtype_alu import ht, dtypes_float
 import numpy as np
@@ -32,7 +31,7 @@ class TestTranscendentalMath(unittest.TestCase):
    ([(Tensor.sin, np.sin)] if dtypes.ulong in supported_dtypes else [])))
  def test_float32(self, x, op):
    # wrong nan behavior on Vulkan
-    if (math.isnan(x) or (x < 0 and op[0] == Tensor.log)) and CI and Device.DEFAULT == "WEBGPU" and not OSX: return
+    if (math.isnan(x) or (x < 0 and op[0] == Tensor.log)) and Device.DEFAULT == "WEBGPU" and not OSX: return
    with Context(TRANSCENDENTAL=2), np.errstate(all='ignore'):
      np.testing.assert_allclose(op[0](Tensor([x], dtype=dtypes.float32)).numpy(),
                                 op[1](np.array([x], dtype=_to_np_dtype(dtypes.float32))),
@@ -43,7 +42,7 @@ class TestTranscendentalMath(unittest.TestCase):
    ([(Tensor.sin, np.sin)] if dtypes.ulong in supported_dtypes else [])))
  def test_float16(self, x, op):
    # wrong nan behavior on Vulkan
-    if (math.isnan(x) or (x < 0 and op[0] == Tensor.log)) and CI and Device.DEFAULT == "WEBGPU" and not OSX: return
+    if (math.isnan(x) or (x < 0 and op[0] == Tensor.log)) and Device.DEFAULT == "WEBGPU" and not OSX: return
    with Context(TRANSCENDENTAL=2), np.errstate(all='ignore'):
      np.testing.assert_allclose(op[0](Tensor([x], dtype=dtypes.float16)).numpy(),
                                 op[1](np.array([x], dtype=_to_np_dtype(dtypes.float16))),
@@ -117,7 +116,7 @@ class TestFloat16Log2(unittest.TestCase):
        np.testing.assert_allclose(result, expected, rtol=1e-3, err_msg=f"log2({val})")

  @unittest.skipUnless(dtypes.float16 in supported_dtypes, f"no float16 on {Device.DEFAULT}")
-  @unittest.skipIf(Device.DEFAULT == "WEBGPU" and CI, "Nan handling differs on Vulkan")
+  @unittest.skipIf(Device.DEFAULT == "WEBGPU" and not OSX, "Nan handling differs on Vulkan")
  def test_float16_log2_special(self):
    # special values: inf, -inf, nan, 0, negative
    with Context(TRANSCENDENTAL=2), np.errstate(all='ignore'):
@@ -11,7 +11,7 @@ from tinygrad.engine.realize import run_linear
 from tinygrad.codegen import to_program
 from tinygrad.codegen.opt import Opt, OptOps
 from tinygrad.renderer.ptx import PTXRenderer
-from test.helpers import to_uops_list, CI
+from test.helpers import to_uops_list

 def run_uops(uops_list:list[UOp], bufs:list[Buffer]):
  buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in bufs]
@@ -20,6 +20,7 @@ def run_uops(uops_list:list[UOp], bufs:list[Buffer]):

 def uop(uops:list[UOp], op:Ops, dtype:Optional[DType], src:tuple[UOp, ...], arg:Any=None) -> UOp:
  if op is Ops.CONST: uops.append(UOp.const(dtype, arg))
+  elif op is Ops.PARAM: uops.append(UOp.param(arg, dtype).replace(src=()))
  else: uops.append(UOp(op, dtype, tuple(src), arg))
  return uops[-1]

@@ -173,8 +174,6 @@ class TestBoolUOps(TestUOps):
  def test_where_bool(self): self._test_top_bool_fxn(Ops.WHERE, lambda a,b,c: b if a else c)

 class TestLocalAccess(unittest.TestCase):
-  # NOTE: this is failing on METAL CI, no idea why. Works locally.
-  @unittest.skipIf(Device.DEFAULT == "METAL" and CI, "failing only in CI")
  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared memory")
  def test_local_basic(self):
    uops = []
@@ -222,8 +221,8 @@ class TestLocalAccess(unittest.TestCase):
@unittest.skipUnless(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "This only tests assembly backends")
 class TestAssembly(unittest.TestCase):
  def test_bitshift_left(self):
-    g1 = UOp(Ops.PARAM, dtypes.int32.ptr(), (), 0)
-    out = UOp(Ops.PARAM, dtypes.int32.ptr(), (), 1)
+    g1 = UOp.param(0, dtypes.int32.ptr())
+    out = UOp.param(1, dtypes.int32.ptr())
    c1 = UOp.const(dtypes.int, 2)
    c2 = UOp.const(dtypes.int, 3)
    l1 = g1.index(c1)
@@ -250,7 +249,7 @@ class TestAssembly(unittest.TestCase):
    self.assertGreaterEqual(len([x.op for x in uops if x.op is Ops.MULACC]), 4)

  def test_mulacc_shl(self):
-    g1 = UOp(Ops.PARAM, dtypes.int32.ptr(), (), 0)
+    g1 = UOp.param(0, dtypes.int32.ptr())
    c1 = UOp.const(dtypes.int, 0)
    c2 = UOp.const(dtypes.int, 1)
    expr = g1.index(c1) * UOp.const(dtypes.int, 4096) + g1.index(c2)
@@ -259,7 +258,7 @@ class TestAssembly(unittest.TestCase):
    self.assertIn(Ops.MULACC, [x.op for x in uops])

  def test_use_cmpeq(self):
-    g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)
+    g = UOp.param(0, dtypes.uint32.ptr())
    c = UOp.const(dtypes.uint, 7)
    comp = g.index(c).ne(c).ne(True)
    uops = to_uops_list([comp], ren=Device[Device.DEFAULT].renderer)
@@ -12,7 +12,7 @@ from tinygrad.dtype import ImageDType, Invalid
 # PYTHONPATH="." DEV=QCOM FLOAT16=1 IMAGE=2 NOLOCALS=1 taskset -c 4-7 python3 examples/openpilot/compile3.py https://github.com/commaai/openpilot/raw/720392c9a5b986981fdbed1bb8c47a6c5573a50e/selfdrive/modeld/models/driving_vision.onnx

 def vision_conv_143():
-  c0 = UOp(Ops.PARAM, dtypes.imageh((16, 1024, 4)), (), 0)
+  c0 = UOp.param(0, dtypes.imageh((16, 1024, 4)))
  c2 = UOp.range(32, 3, AxisType.LOOP)
  c5 = UOp.range(128, 4, AxisType.LOOP)
  c8 = UOp.range(16, 2, AxisType.LOOP)
@@ -22,13 +22,13 @@ def vision_conv_143():
  c26 = UOp.range(7, 1, AxisType.REDUCE)
  c27 = c2*2+c26
  c32 = ((c27<3)!=True)&(c27<67)
-  c34 = UOp(Ops.PARAM, dtypes.imageh((32, 1024, 4)), (), 1)
+  c34 = UOp.param(1, dtypes.imageh((32, 1024, 4)))
  c38 = c5//2
  c45 = (c32&c24).where((c27*64+c38+c17*4096+-12480), UOp.const(dtypes.weakint, Invalid))
  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp(Ops.PARAM, dtypes.imageh((64, 49, 4)), (), 2)
+  c49 = UOp.param(2, dtypes.imageh((64, 49, 4)))
  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
-  c63 = UOp(Ops.PARAM, dtypes.float.ptr(128), (), 3)
+  c63 = UOp.param(3, dtypes.float.ptr(128))
  c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
  c67 = c0.index((c2*128+c5+c8*4096), ptr=True).store(c65).end(c8, c2, c5)

@@ -38,7 +38,7 @@ def vision_conv_143():
  return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))

 def vision_conv_153():
-  c0 = UOp(Ops.PARAM, dtypes.imageh((8, 1024, 4)), (), 0)
+  c0 = UOp.param(0, dtypes.imageh((8, 1024, 4)))
  c2 = UOp.range(16, 3, AxisType.LOOP)
  c5 = UOp.range(256, 4, AxisType.LOOP)
  c8 = UOp.range(8, 2, AxisType.LOOP)
@@ -48,13 +48,13 @@ def vision_conv_153():
  c26 = UOp.range(7, 1, AxisType.REDUCE)
  c27 = c2*2+c26
  c32 = ((c27<3)!=True)&(c27<35)
-  c34 = UOp(Ops.PARAM, dtypes.imageh((16, 1024, 4)), (), 1)
+  c34 = UOp.param(1, dtypes.imageh((16, 1024, 4)))
  c38 = c5//2
  c45 = (c32&c24).where((c27*128+c38+c17*4096+-12672), UOp.const(dtypes.weakint, Invalid))
  c48 = (c24&c32).where(c34.index(c45), UOp.const(dtypes.float, 0.0))
-  c49 = UOp(Ops.PARAM, dtypes.imageh((128, 49, 4)), (), 2)
+  c49 = UOp.param(2, dtypes.imageh((128, 49, 4)))
  c61 = c48*c49.index((c26*4+c5%2+c16*28+c38*196))
-  c63 = UOp(Ops.PARAM, dtypes.float.ptr(256), (), 3)
+  c63 = UOp.param(3, dtypes.float.ptr(256))
  c65 = c61.reduce(c16, c26, arg=Ops.ADD)+c63.index(c5)
  c67 = c0.index((c2*256+c5+c8*4096), ptr=True).store(c65).end(c8, c2, c5)

@@ -64,16 +64,16 @@ def vision_conv_153():
  return c67.sink(arg=KernelInfo(name="conv", opts_to_apply=opts))

 def dm_conv_172():
-  c0 = UOp(Ops.PARAM, dtypes.imageh((1, 240, 4)), (), 0)
+  c0 = UOp.param(0, dtypes.imageh((1, 240, 4)))
  c2 = UOp.range(960, 4, AxisType.LOOP)
-  c5 = UOp(Ops.PARAM, dtypes.imageh((8, 384, 4)), (), 1)
+  c5 = UOp.param(1, dtypes.imageh((8, 384, 4)))
  c7 = UOp.range(32, 0, AxisType.REDUCE)
  c10 = UOp.range(4, 1, AxisType.REDUCE)
  c13 = UOp.range(12, 3, AxisType.REDUCE)
  c18 = UOp.range(8, 2, AxisType.REDUCE)
-  c23 = UOp(Ops.PARAM, dtypes.imageh((240, 128, 4)), (), 2)
+  c23 = UOp.param(2, dtypes.imageh((240, 128, 4)))
  c35 = c5.index((c7*4+c10+c13*128+c18*1536))*c23.index((c10*4+c2%4+c7*16+c2//4*512))
-  c37 = UOp(Ops.PARAM, dtypes.float.ptr(960), (), 3)
+  c37 = UOp.param(3, dtypes.float.ptr(960))
  c39 = c35.reduce(c7, c10, arg=Ops.ADD)+c37.index(c2)
  c50 = (1.0+((c39+0.044708251953125*(c39*(c39*c39)))*-2.3021129851685216).exp2()).reciprocal()*c39
  c53 = c50.reduce(c18, c13, arg=Ops.ADD)*0.010416666666666666
@@ -99,4 +99,4 @@ bufs = [Buffer(ps.arg.device, g.size, g.dtype if isinstance(g.dtype, ImageDType)

 gsize, lsize = ps.arg.launch_dims({})
 t = rt(*[b._buf for b in bufs], global_size=gsize, local_size=lsize, vals=ps.arg.vals({}), wait=True)
-print(f"{t*1e6:.2f} us")
+print(f"{t*1e6:.2f} us")
@@ -1,8 +1,7 @@
-import unittest
+import unittest, sys
 from tinygrad import Device
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import getenv, OSX
-from test.helpers import CI

 def multidevice_test(fxn):
  exclude_devices = getenv("EXCLUDE_DEVICES", "").split(",")
@@ -10,15 +9,15 @@ def multidevice_test(fxn):
    for device in Device._devices:
      # broken on OSX USB AMD, why?
      if device in ["DISK", "NPY", "FAKE", "DSP", "NULL"] or (OSX and device in ["AMD"]): continue
-      if not CI: print(device)
+      if sys.stdout.isatty(): print(device)
      if device in exclude_devices:
-        if not CI: print(f"WARNING: {device} test is excluded")
+        if sys.stdout.isatty(): print(f"WARNING: {device} test is excluded")
        continue
      with self.subTest(device=device):
        try:
          Device[device]
        except Exception:
-          if not CI: print(f"WARNING: {device} test isn't running")
+          if sys.stdout.isatty(): print(f"WARNING: {device} test isn't running")
          continue
        fxn(self, device)
  return ret
@@ -1,10 +1,9 @@
 import unittest, ctypes, struct, time, array
 from tinygrad import Device, Tensor, dtypes
-from tinygrad.helpers import to_mv
+from tinygrad.helpers import to_mv, DEV
 from tinygrad.device import Buffer, BufferSpec
 from tinygrad.engine.realize import get_runtime
 from tinygrad.codegen import to_program
-from test.helpers import CI

 def _time_queue(q, d):
  st = time.perf_counter()
@@ -149,7 +148,7 @@ class TestHCQ(unittest.TestCase):
    val = TestHCQ.b.uop.buffer.as_memoryview().cast("f")[1]
    assert val == 0.0, f"got val {val}, should not be updated"

-  @unittest.skipIf(CI, "Can't handle async update on CPU")
+  @unittest.skipIf(DEV.interface.startswith("MOCK"), "Can't handle async update on CPU")
  def test_wait_signal(self):
    temp_signal = TestHCQ.d0._alloc_signal(value=0)
    TestHCQ.compute_queue().wait(temp_signal, value=1).signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
@@ -160,7 +159,7 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0._wait_signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value, timeout=100)
    TestHCQ.d0.timeline_value += 1

-  @unittest.skipIf(CI, "Can't handle async update on CPU")
+  @unittest.skipIf(DEV.interface.startswith("MOCK"), "Can't handle async update on CPU")
  def test_wait_copy_signal(self):
    temp_signal = TestHCQ.d0._alloc_signal(value=0)
    TestHCQ.copy_queue().wait(temp_signal, value=1).signal(TestHCQ.d0.timeline_signal, TestHCQ.d0.timeline_value).submit(TestHCQ.d0)
@@ -3,7 +3,7 @@ import unittest
 import numpy as np
 from tinygrad import Tensor, dtypes
 from tinygrad.engine.jit import TinyJit
-from test.helpers import derandomize_model, CI
+from test.helpers import derandomize_model

 from examples.llama import Transformer

@@ -27,7 +27,6 @@ class TestJittedModels(unittest.TestCase):
    helper_test_jitted_correctness(lambda: (Tensor([[1,]]),), test, test_jit)
    dtypes.default_float = old_float

-  @unittest.skipUnless(not CI, "huge for CI")
  def test_jitted_stable_diffusion(self):
    from examples.stable_diffusion import UNetModel, unet_params
    model = UNetModel(**unet_params)
@@ -29,7 +29,7 @@ def gradient_test():
  z = y.matmul(x).sum()
  z.backward()
 def realized_eye():
-  Tensor.eye(3).realize()
+  Tensor.eye(3).clone().realize()
 def realized_list():
  Tensor([[2.0,0,-2.0]]).realize()
 def kernel_matmul():
@@ -4,7 +4,7 @@ import os, multiprocessing, logging, pickle, sqlite3, difflib, warnings, functoo
 from dataclasses import replace
 from typing import Callable, Any

-ASSERT_DIFF = int((flag:="[pr]") in os.getenv("COMMIT_MESSAGE", flag) or flag in os.getenv("PR_TITLE", flag))
+ASSERT_DIFF = int((flag:="[PR]") in os.getenv("COMMIT_MESSAGE", flag) or flag in os.getenv("PR_TITLE", flag))
 if not int(os.getenv("ASSERT_PROCESS_REPLAY", "1")): ASSERT_DIFF = 0

 try:
@@ -8,14 +8,11 @@ from tinygrad.tensor import _to_np_dtype
 from tinygrad.codegen import to_program
 from tinygrad.dtype import DType
 from tinygrad.nn.state import get_parameters
-from tinygrad.helpers import T, Target
+from tinygrad.helpers import T, Target, DEV
 from tinygrad.renderer import Renderer
 from tinygrad.codegen import full_rewrite_to_sink, line_rewrite, pm_linearize_cleanups
 from tinygrad.codegen.late.linearizer import linearize

-# TODO: remove this everywhere!
-CI = os.getenv("CI", "") != ""
-
 # decorator to skip slow tests by default, run with RUN_SLOW=1 to include them
 slow = unittest.skipUnless(os.getenv("RUN_SLOW"), "slow test, set RUN_SLOW=1 to run")
 from tinygrad.runtime.ops_python import PythonProgram, PythonRenderer, PythonCompiler
@@ -85,7 +82,7 @@ def eval_uop(uop:UOp, inputs:list[tuple[DType, list[Any]]]|None=None, vals:tuple
  for buf_dt, data in inputs or []:
    bufs.append(buf:=allocator.alloc(len(data) * buf_dt.itemsize))
    allocator._copyin(buf, memoryview(struct.pack(str(len(data)) + (buf_dt.fmt or ""), *data)))
-  g = UOp(Ops.PARAM, uop.dtype.ptr(), arg=0, src=())
+  g = UOp.param(0, uop.dtype.ptr())
  prg = to_program(UOp.store(g.index(UOp.const(dtypes.int, 0)), uop).sink(arg=KernelInfo()), PythonRenderer(Target("PYTHON")))
  prog = PythonProgram("run", PythonCompiler().compile(prg.src[3].arg))
  prog(out_buf:=allocator.alloc(uop.dtype.itemsize), *bufs, vals=vals)
@@ -100,7 +97,7 @@ def to_uops_list(u:list[UOp], ren=None) -> list[UOp]:

 def not_support_multi_device():
  # CL and CUDA don't support multi device if in CI
-  return CI and Device.DEFAULT in ("CL", "CUDA")
+  return (Device.DEFAULT == "CL" and Device[Device.DEFAULT].count() < 2) or (Device.DEFAULT == "CUDA" and DEV.interface.startswith("MOCK"))

 def needs_second_gpu(fn):
  @functools.wraps(fn)
@@ -423,10 +423,10 @@ def _collect_data_slices(assigns: list[tuple[str, UOp]], data_prefix: str, pcode
 class _Ctx:
  """Context for instruction compilation - holds buffers and helpers."""
  __slots__ = ('inst_size', 'dyn_fields', '_axis_id', 'wave_size', 'vgpr', 'accvgpr')
-  sgpr = UOp(Ops.PARAM, dtypes.uint32.ptr(SGPR_COUNT), arg=0)
-  vmem = UOp(Ops.PARAM, dtypes.uint32.ptr(1 << 46), arg=2)
-  lds = UOp(Ops.PARAM, dtypes.uint32.ptr(16384), arg=3)
-  scratch = UOp(Ops.PARAM, dtypes.uint8.ptr(1 << 30), arg=4)
+  sgpr = UOp.param(0, dtypes.uint32.ptr(SGPR_COUNT))
+  vmem = UOp.param(2, dtypes.uint32.ptr(1 << 46))
+  lds = UOp.param(3, dtypes.uint32.ptr(16384))
+  scratch = UOp.param(4, dtypes.uint8.ptr(1 << 30))
  # Cache PARAM UOps by wave_size so all _Ctx instances with same wave_size share identical UOp references
  _vgpr_cache: dict[int, UOp] = {}
  _accvgpr_cache: dict[int, UOp] = {}
@@ -434,10 +434,10 @@ class _Ctx:
  def __init__(self, inst_size: int, wave_size: int = 32):
    self.inst_size, self._axis_id, self.wave_size = inst_size, 0, wave_size
    self.dyn_fields: list[tuple[int, int]] = []  # (lo, hi) of fields read dynamically
-    if wave_size not in _Ctx._vgpr_cache: _Ctx._vgpr_cache[wave_size] = UOp(Ops.PARAM, dtypes.uint32.ptr(256 * wave_size), arg=1)
+    if wave_size not in _Ctx._vgpr_cache: _Ctx._vgpr_cache[wave_size] = UOp.param(1, dtypes.uint32.ptr(256 * wave_size))
    self.vgpr = _Ctx._vgpr_cache[wave_size]
    if wave_size == 64:
-      if wave_size not in _Ctx._accvgpr_cache: _Ctx._accvgpr_cache[wave_size] = UOp(Ops.PARAM, dtypes.uint32.ptr(256 * wave_size), arg=5)
+      if wave_size not in _Ctx._accvgpr_cache: _Ctx._accvgpr_cache[wave_size] = UOp.param(5, dtypes.uint32.ptr(256 * wave_size))
      self.accvgpr = _Ctx._accvgpr_cache[wave_size]
    else:
      self.accvgpr = self.vgpr
@@ -523,7 +523,7 @@ class _Ctx:
    """Write SGPR with dynamic register index. On RDNA, index 124 = NULL (writes discarded). On CDNA, index 124 = M0 (read/write)."""
    # RDNA: NULL (124) discards writes. CDNA: M0 (124) is writable.
    valid = None if self.wave_size == 64 else reg.ne(_c(124))
-    return self.sgpr.index(reg.cast(dtypes.int).valid(valid) if valid is not None else reg.cast(dtypes.int)).store(val.cast(dtypes.uint32))
+    return self.sgpr.index(reg.cast(dtypes.int).valid(valid) if valid is not None else reg.cast(dtypes.int), ptr=True).store(val.cast(dtypes.uint32))

  def wmask(self, reg: UOp, val: UOp) -> list[UOp]:
    """Write a lane mask (VCC/EXEC). Splits into lo/hi for wave64."""
@@ -1,19 +1,18 @@
 import torch
 from torch import nn
-import unittest
+import unittest, sys
 import numpy as np
 from tinygrad.nn.state import get_parameters, get_state_dict
 from tinygrad.nn import optim, Linear, Conv2d, BatchNorm2d
 from tinygrad.tensor import Tensor
 from extra.datasets import fetch_mnist
-from test.helpers import CI

 def compare_tiny_torch(model, model_torch, X, Y):
  with Tensor.train():
    model_torch.train()
    model_state_dict = get_state_dict(model)
    for k,v in model_torch.named_parameters():
-      if not CI: print(f"initting {k} from torch")
+      if sys.stdout.isatty(): print(f"initting {k} from torch")
      model_state_dict[k].assign(Tensor(v.detach().numpy())).realize()

    optimizer = optim.SGD(get_parameters(model), lr=0.001)
@@ -35,14 +34,14 @@ def compare_tiny_torch(model, model_torch, X, Y):
    loss_torch.backward()

    # assert losses match
-    if not CI: print(loss.realize().numpy())
-    if not CI: print(loss_torch.detach().numpy())
+    if sys.stdout.isatty(): print(loss.realize().numpy())
+    if sys.stdout.isatty(): print(loss_torch.detach().numpy())
    np.testing.assert_allclose(loss.realize().numpy(), loss_torch.detach().numpy(), atol=1e-4)

    for k,v in list(model_torch.named_parameters())[::-1]:
      g = model_state_dict[k].grad.numpy()
      gt = v.grad.detach().numpy()
-      if not CI: print("testing grads", k, model_state_dict[k].grad.dtype)
+      if sys.stdout.isatty(): print("testing grads", k, model_state_dict[k].grad.dtype)
      np.testing.assert_allclose(g, gt, atol=1e-3, err_msg=f'grad mismatch {k}')

    # take the steps
@@ -51,7 +50,7 @@ def compare_tiny_torch(model, model_torch, X, Y):

    # assert weights match
    for k,v in model_torch.named_parameters():
-      if not CI: print("testing weight", k, model_state_dict[k].dtype)
+      if sys.stdout.isatty(): print("testing weight", k, model_state_dict[k].dtype)
      np.testing.assert_allclose(model_state_dict[k].numpy(), v.detach().numpy(), atol=1e-3, err_msg=f'weight mismatch {k}')

 def get_mnist_data():
@@ -1,7 +1,7 @@
 import unittest, io
 from contextlib import redirect_stdout
 from tinygrad import Tensor, dtypes, Device
-from tinygrad.helpers import OSX, DEV
+from tinygrad.helpers import OSX
 from tinygrad.engine.realize import compile_linear
 from tinygrad.codegen import to_program

@@ -16,8 +16,7 @@ class TestCompileFailures(unittest.TestCase):
    self.compile((Tensor.empty(1024, dtype='uint8') + Tensor.empty(1024, dtype='uint8')).max())

 class TestDisassembly(unittest.TestCase):
-  # TODO: fails on llvm. llvm.LLVMGetHostCPUName() returns "generic"
-  @unittest.skipUnless(Device.DEFAULT in ("CPU",) and DEV.renderer not in ("LLVM", "LVP") and OSX, "m series cpus support fp16 arithmetic")
+  @unittest.skipUnless(Device.DEFAULT == "CPU" and OSX, "m series cpus support fp16 arithmetic")
  def test_float16_alu(self):
    c = Tensor([1], dtype=dtypes.float16) + Tensor([1], dtype=dtypes.float16)
    s = c.schedule_linear().src[-1]
@@ -134,14 +134,14 @@ class TestBitcastConstFolding(unittest.TestCase):
 # folds advance indexing into basic indexing
 class TestIndexingConstFolding(unittest.TestCase):
  def test_scalar_index(self):
-    t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
+    t = Tensor.arange(16).float().reshape(1,1,4,4).clone().realize()
    _check_ast_count(1, t[:,:,Tensor(1),:])
    _check_ast_count(1, t[:,:,Tensor(1)+2,:])
    _check_ast_count(1, t[:,:,Tensor(1),Tensor(0)])

  def test_const_tensor_index(self):
    # TODO: these can be 0, implement const tensor folded indexing
-    t = Tensor.arange(16).float().reshape(1,1,4,4).realize()
+    t = Tensor.arange(16).float().reshape(1,1,4,4).clone().realize()
    _check_ast_count(1, t[:,:,Tensor.ones(2,1,dtype=dtypes.int),:])
    _check_ast_count(1, t[:,:,Tensor.ones(1,2,dtype=dtypes.int)+2,:])
    _check_ast_count(1, t[:,:,Tensor.ones(1,1,dtype=dtypes.int),Tensor.zeros(2,1,2,dtype=dtypes.int)])
@@ -5,7 +5,6 @@ from tinygrad import Tensor
 from tinygrad.device import Device, Compiler, enumerate_devices_str
 from tinygrad.helpers import diskcache_get, diskcache_put, getenv, Context, Target, WIN, OSX, DEV
 from tinygrad.runtime.support.c import DLL
-from test.helpers import CI

 class TestDevice(unittest.TestCase):
  def test_canonicalize(self):
@@ -29,8 +28,8 @@ class TestDevice(unittest.TestCase):
  def test_nonexistent_renderer(self):
    with self.assertRaisesRegex(RuntimeError, "has no renderer"):
      with Context(DEV="CPU:TYPO"): Device[Device.DEFAULT].renderer
-    with self.assertRaisesRegex(RuntimeError, "did you mean: 'CLANGJIT'"):
-      with Context(DEV="CPU:CLANG"): Device[Device.DEFAULT].renderer
+    with self.assertRaisesRegex(RuntimeError, "did you mean: 'CLANG'"):
+      with Context(DEV="CPU:CLANGJIT"): Device[Device.DEFAULT].renderer

  @unittest.skipIf(Device.DEFAULT != "AMD", "only run on AMD")
  def test_nonexistent_iface(self):
@@ -67,20 +66,20 @@ class TestDevice(unittest.TestCase):
    self.assertNotEqual(result.returncode, 0)
    self.assertIn(b"deprecated", result.stderr)

-  @unittest.skipIf(WIN and CI, "skipping windows test") # TODO: subprocess causes memory violation?
+  @unittest.skipIf(WIN, "skipping windows test") # TODO: subprocess causes memory violation?
  def test_env_overwrite_default_compiler(self):
    if Device.DEFAULT == "CPU":
-      from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangJITCompiler
-      try: _, _ = CPULLVMCompiler(), ClangJITCompiler()
+      from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangCompiler
+      try: _, _ = CPULLVMCompiler(), ClangCompiler()
      except Exception as e: self.skipTest(f"skipping compiler test: not all compilers: {e}")

-      imports = "from tinygrad import Device; from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangJITCompiler"
+      imports = "from tinygrad import Device; from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangCompiler"
      subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, CPULLVMCompiler)"'],
                        shell=True, check=True, env={**os.environ, "DEV": "CPU:LLVM"})
-      subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, ClangJITCompiler)"'],
+      subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, ClangCompiler)"'],
                        shell=True, check=True, env={**os.environ, "DEV": "CPU"})
-      subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, ClangJITCompiler)"'],
-                        shell=True, check=True, env={**os.environ, "DEV": "CPU:CLANGJIT"})
+      subprocess.run([f'python3 -c "{imports}; assert isinstance(Device[Device.DEFAULT].compiler, ClangCompiler)"'],
+                        shell=True, check=True, env={**os.environ, "DEV": "CPU:CLANG"})
    elif Device.DEFAULT == "AMD":
      from tinygrad.runtime.support.compiler_amd import HIPCompiler, AMDLLVMCompiler
      try: _, _ = HIPCompiler(Device[Device.DEFAULT].arch), AMDLLVMCompiler(Device[Device.DEFAULT].arch)
@@ -95,17 +94,17 @@ class TestDevice(unittest.TestCase):
                        shell=True, check=True, env={**os.environ, "DEV": "AMD:HIP"})
    else: self.skipTest("only run on CPU/AMD")

-  @unittest.skipIf(WIN and CI, "skipping windows test")
+  @unittest.skipIf(WIN, "skipping windows test")
  def test_env_online(self):
-    from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangJITCompiler
-    try: _, _ = CPULLVMCompiler(), ClangJITCompiler()
+    from tinygrad.runtime.support.compiler_cpu import CPULLVMCompiler, ClangCompiler
+    try: _, _ = CPULLVMCompiler(), ClangCompiler()
    except Exception as e: self.skipTest(f"skipping compiler test: not all compilers: {e}")

    with Context(DEV="CPU:LLVM"):
      inst = Device["CPU"].compiler
      self.assertIsInstance(Device["CPU"].compiler, CPULLVMCompiler)
    with Context(DEV="CPU"):
-      self.assertIsInstance(Device["CPU"].compiler, ClangJITCompiler)
+      self.assertIsInstance(Device["CPU"].compiler, ClangCompiler)
    with Context(DEV="CPU:LLVM"):
      self.assertIsInstance(Device["CPU"].compiler, CPULLVMCompiler)
      assert inst is Device["CPU"].compiler  # cached
@@ -119,7 +118,7 @@ class TestDevice(unittest.TestCase):

    dev = Device["CPU"]
    dev.cached_renderer.clear()
-    with patch("tinygrad.renderer.cstyle.ClangJITRenderer.__init__", side_effect=RuntimeError("broken")):
+    with patch("tinygrad.renderer.cstyle.ClangRenderer.__init__", side_effect=RuntimeError("broken")):
      self.assertIsInstance(dev.renderer.compiler, CPULLVMCompiler)

  def test_dev_contextvar(self):
@@ -1,5 +1,5 @@
 import unittest, subprocess, platform
-from tinygrad.runtime.support.compiler_cpu import ClangJITCompiler
+from tinygrad.runtime.support.compiler_cpu import ClangCompiler
 from tinygrad.runtime.support.elf import elf_loader

 class TestElfLoader(unittest.TestCase):
@@ -23,7 +23,7 @@ class TestElfLoader(unittest.TestCase):
      }
    '''
    with self.assertRaisesRegex(RuntimeError, 'evil_external_function'):
-      ClangJITCompiler([{'AMD64':'x86_64', 'aarch64':'arm64'}.get(m:=platform.machine(), m), "native"]).compile(src)
+      ClangCompiler([{'AMD64':'x86_64', 'aarch64':'arm64'}.get(m:=platform.machine(), m), "native"]).compile(src)
  def test_link(self):
    src = '''
      float powf(float, float); // from libm
@@ -96,7 +96,7 @@ class TestGroupedDims(unittest.TestCase):

  def test_global_prod_max(self):
    g, l = UOp.range(256, 0, AxisType.GLOBAL), UOp.range(256, 1, AxisType.LOCAL)
-    sink = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0).index(g + l).store(UOp.const(dtypes.float, 1.0)).end(g, l).sink(arg=KernelInfo())
+    sink = UOp.param(0, dtypes.float.ptr()).index(g + l).store(UOp.const(dtypes.float, 1.0)).end(g, l).sink(arg=KernelInfo())
    class R(Renderer): global_max, local_max, global_prod_max = (256, 256, 256), (128, 128, 128), (128, 128, 128)
    specials = [u for u in add_gpudims(R(Target()), sink).toposort() if u.op is Ops.SPECIAL]
    self.assertGreater(len([s for s in specials if "lidx" in s.arg]), 1)
@@ -7,14 +7,14 @@ from tinygrad.codegen import to_program

 class TestLinearizerFailures(unittest.TestCase):
  def test_fail_1(self):
-    c0 = UOp(Ops.PARAM, dtypes.float.ptr(64), arg=0, src=())
+    c0 = UOp.param(0, dtypes.float.ptr(64))
    c1 = UOp.range(UOp.const(dtypes.weakint, 2), 1, AxisType.LOOP)
    c2 = UOp.range(UOp.const(dtypes.weakint, 32), 2, AxisType.LOOP)
    c3 = ((c1*UOp.const(dtypes.weakint, 32))+c2)
-    c4 = UOp(Ops.PARAM, dtypes.float.ptr(163840), arg=1, src=())
+    c4 = UOp.param(1, dtypes.float.ptr(163840))
    c5 = UOp.range(UOp.const(dtypes.weakint, 2560), 0, AxisType.REDUCE)
    c6 = c4.index(((((((c5//UOp.const(dtypes.weakint, 8))%UOp.const(dtypes.weakint, 8))*UOp.const(dtypes.weakint, 8))+(c5%UOp.const(dtypes.weakint, 8)))+(((c2*UOp.const(dtypes.weakint, 40))+(c5//UOp.const(dtypes.weakint, 64)))*UOp.const(dtypes.weakint, 64)))+(c1*UOp.const(dtypes.weakint, 81920))))
-    c7 = UOp(Ops.PARAM, dtypes.float.ptr(64), arg=2, src=())
+    c7 = UOp.param(2, dtypes.float.ptr(64))
    c8 = c7.index(c3)
    c9 = ((((c6+(c8*UOp.const(dtypes.float, -1.0)))*(c6+(c8*UOp.const(dtypes.float, -1.0)))).reduce(c5, arg=Ops.ADD)*UOp.const(dtypes.float, 0.000390625))+UOp.const(dtypes.float, 1e-05)).sqrt().reciprocal()
    c10 = c0.index(c3).store(c9).end(c1, c2)
@@ -26,7 +26,7 @@ def _make_linear(buffer_lists, copies=None):
 def _get_arena(buf, linear, result):
  for orig_si, new_si in zip(linear.src, result.src):
    for orig, new in zip(orig_si.src[1:], new_si.src[1:]):
-      if orig is buf and new.op is Ops.BUFFER_VIEW: return new.src[0]
+      if orig is buf and new.op is Ops.SLICE: return new.src[0]
  return None

 def check_assign(buffer_lists, copies=None):
@@ -37,8 +37,8 @@ def check_assign(buffer_lists, copies=None):
  replace_map: dict[int, tuple[UOp, int, int]] = {}
  for orig_si, new_si in zip(linear.src, result.src):
    for orig, new in zip(orig_si.src[1:], new_si.src[1:]):
-      if new.op is Ops.BUFFER_VIEW and id(orig) not in replace_map:
-        replace_map[id(orig)] = (new.src[0], new.arg[1] * new.dtype.itemsize, new.arg[0] * new.dtype.itemsize)
+      if new.op is Ops.SLICE and id(orig) not in replace_map:
+        replace_map[id(orig)] = (new.src[0], new.src[1].arg * new.src[0].dtype.itemsize, new.arg * new.dtype.itemsize)

  # verify pinned buffers are not planned
  for buf in held_bufs:
@@ -8,7 +8,7 @@ class TestDataset(unittest.TestCase):
    X_train[0].contiguous().realize()
    GlobalCounters.reset()
    X_train[0].contiguous().realize()
-    self.assertLessEqual(GlobalCounters.kernel_count, 1)  # 0 if BUFFER_VIEW (zero-copy), 1 otherwise
+    self.assertLessEqual(GlobalCounters.kernel_count, 1)  # 0 if SLICE (zero-copy), 1 otherwise

 if __name__ == '__main__':
  unittest.main()
@@ -118,7 +118,7 @@ class TestContiguous(unittest.TestCase):
  def test_size_change_buffer_view(self):
    a = Tensor.empty(4)
    b = a.reshape((1, 1, 4)).shrink(((0, 1), (0, 1), (0, 3))).contiguous()
-    check_schedule(b, 0)  # contiguous shrink of a realized buffer is a zero-copy BUFFER_VIEW
+    check_schedule(b, 0)  # contiguous shrink of a realized buffer is a zero-copy SLICE

  def test_double_contiguous_realizes_once(self):
    a = Tensor.empty(4, 1)
@@ -414,7 +414,7 @@ class TestSchedule(unittest.TestCase):
    check_schedule([a+b, a+b], 1)

  def test_const_realize(self):
-    t = Tensor.ones(2)
+    t = Tensor.ones(2, buffer=False)
    check_schedule(t[0], 0)
    check_schedule(t[1], 0)

@@ -429,7 +429,7 @@ class TestSchedule(unittest.TestCase):
      img = Tensor.empty(1,32,4,4)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
      out = bn(img)
-      check_schedule(out, 3)
+      check_schedule(out, 3, nn.state.get_parameters(bn))

  def test_fold_conv_batchnorm_notrain(self):
    with Tensor.train(False):
@@ -437,7 +437,7 @@ class TestSchedule(unittest.TestCase):
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=True)
      out = bn(c1(img)).relu()
-      check_schedule(out, 1, [c1.weight, c1.bias])
+      check_schedule(out, 1, [c1.weight, c1.bias, *nn.state.get_parameters(bn)])

  def test_fold_conv_batchnorm_notrain_no_running_stats(self):
    with Tensor.train(False):
@@ -445,7 +445,7 @@ class TestSchedule(unittest.TestCase):
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
      out = bn(c1(img)).relu()
-      check_schedule(out, 4, [c1.weight, c1.bias])
+      check_schedule(out, 4, [c1.weight, c1.bias, *nn.state.get_parameters(bn)])

  def test_fold_conv_batchnorm(self):
    with Tensor.train():
@@ -453,17 +453,17 @@ class TestSchedule(unittest.TestCase):
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
      out = bn(c1(img)).relu()
-      check_schedule(out, 4, [c1.weight, c1.bias])
+      check_schedule(out, 4, [c1.weight, c1.bias, *nn.state.get_parameters(bn)])

  def test_fold_conv_batchnorm_optim(self, adam=False):
-    # 2 is too low?
-    optim, cnt = (nn.optim.Adam, 16) if adam else (nn.optim.SGD, 2)
+    optim, cnt = (nn.optim.Adam, 29) if adam else (nn.optim.SGD, 15)
    with Tensor.train():
      img = Tensor.ones(1,3,4,4)
      c1 = nn.Conv2d(3,32,3)
      bn = nn.BatchNorm2d(32, track_running_stats=False)
      _realize_weights([c1, bn])
      opt = optim(nn.state.get_parameters([c1, bn]))
+      Tensor.realize(img, *nn.state.get_parameters(opt))
      img_bn = bn(c1(img)).elu().sum()
      opt.zero_grad()
      img_bn.backward()
@@ -477,14 +477,14 @@ class TestSchedule(unittest.TestCase):
      fw = bn(x).contiguous_backward().relu().contiguous()
      fw.sum().backward()
      # TODO: this is too many
-      check_schedule([x.grad, bn.weight.grad, bn.bias.grad, fw], 9)
+      check_schedule([x.grad, bn.weight.grad, bn.bias.grad, fw], 10, nn.state.get_parameters(bn))

  def test_fold_conv_relu(self):
    c1 = nn.Conv2d(3,16,3)
    # run
    img = Tensor.ones(2,3,64,64)
    out = c1(img).relu()
-    check_schedule(out, 1, [c1.weight, c1.bias])
+    check_schedule(out, 1, [c1.weight, c1.bias, img])

  def test_fold_conv_relu_alt(self):
    img = Tensor.ones(1,4,8,8)
@@ -766,12 +766,6 @@ class TestSchedule(unittest.TestCase):
    out = x + y
    check_schedule(out, 1)

-  def test_const_no_recompute(self):
-    x = Tensor(2) + Tensor(2)
-    y = Tensor(2) + Tensor(2)
-    out = x.contiguous() + y.contiguous()
-    check_schedule(out, 2, filter_sink=False)
-
  def test_reduce_shrink_child(self):
    a = Tensor.empty(100, 100)
    b = Tensor.empty(10,)
@@ -827,6 +821,7 @@ class TestSchedule(unittest.TestCase):
      layer = nn.Linear(32, 32*4)
      _realize_weights(layer)
      opt = nn.optim.Adam(nn.state.get_parameters(layer), lr=1e-4)
+      Tensor.realize(*nn.state.get_parameters(opt))
      layer(x).relu().sum().backward()
      check_schedule(opt.schedule_step(), 13)

@@ -836,6 +831,7 @@ class TestSchedule(unittest.TestCase):
      c1 = nn.Conv2d(3,32,3)
      _realize_weights(c1)
      opt = nn.optim.Adam(nn.state.get_parameters(c1), lr=1e-4)
+      Tensor.realize(*nn.state.get_parameters(opt))
      opt.zero_grad()
      c1(img).relu().sum().backward()
      check_schedule(opt.schedule_step(), 13)
@@ -847,6 +843,7 @@ class TestSchedule(unittest.TestCase):
      c2 = nn.Conv2d(16,32,2,bias=False)
      _realize_weights([c1, c2])
      opt = nn.optim.Adam(nn.state.get_parameters([c1, c2]), lr=1e-4)
+      Tensor.realize(*nn.state.get_parameters(opt))
      opt.zero_grad()
      c2(c1(img).relu()).relu().sum().backward()
      check_schedule(opt.schedule_step(), 15)
@@ -879,6 +876,7 @@ class TestSchedule(unittest.TestCase):
      c2 = nn.Conv2d(16,32,2,bias=False)
      _realize_weights([c1, c2])
      opt = nn.optim.SGD(nn.state.get_parameters([c1, c2]), nesterov=True, momentum=0.9, weight_decay=0.1)
+      Tensor.realize(*nn.state.get_parameters(opt))
      opt.zero_grad()
      c2(c1(img).relu()).relu().sum().backward()
      check_schedule(opt.schedule_step(), 11)
@@ -993,7 +991,7 @@ class TestSchedule(unittest.TestCase):
  def test_fuse_arange_pad_circular_mode_bw(self):
    x = Tensor.empty(1,1,5,5,5)
    out = x.pad((1,2,3,5,1,2), mode="circular")
-    g = out.sum().gradient(x)[0]
+    g = out.sum().gradient(x)[0].clone()
    linear, _ = check_schedule(g, 1)
    self.assertEqual(len([x for x in linear.src[0].src[0].backward_slice_with_self if x.op is Ops.REDUCE]), 0)

@@ -1008,7 +1006,7 @@ class TestSchedule(unittest.TestCase):
      out = bn1(conv1(x)).relu()
      out = bn2(conv2(out))
      out = (out + x).relu()
-      run_linear(*check_schedule(out, 2, [conv1.weight, conv2.weight]))
+      run_linear(*check_schedule(out, 2, [conv1.weight, conv2.weight, *nn.state.get_parameters(bn1), *nn.state.get_parameters(bn2)]))

 class TestSwizzle(unittest.TestCase):
  def test_softmax_one_kernel(self):
@@ -1212,10 +1210,10 @@ class TestFusionOp(unittest.TestCase):
    self.assertEqual(len(linear.src), 1)
    self.assertLess(time.perf_counter()-st, 2.0)

-# NOTE: the NULL backend supports BUFFER_VIEW
+# NOTE: the NULL backend supports SLICE
 class TestBufferView(unittest.TestCase):
  def test_shrink_contiguous_is_buffer_view(self):
-    # simple 1D shrink of a realized buffer should be BUFFER_VIEW, not a copy kernel
+    # simple 1D shrink of a realized buffer should be SLICE, not a copy kernel
    a = Tensor.arange(100).clone().realize()
    b = a.shrink(((10, 50),)).contiguous()
    run_linear(*check_schedule(b, 0))
@@ -1231,7 +1229,7 @@ class TestBufferView(unittest.TestCase):
    run_linear(*check_schedule(b, 0))

  def test_shrink_non_shard_axis_is_buffer_view_multi(self):
-    # indexing a non-shard axis of a realized sharded tensor should be BUFFER_VIEW on each device, not copy kernels
+    # indexing a non-shard axis of a realized sharded tensor should be SLICE on each device, not copy kernels
    # this is the flat_llama pattern: weight[layer_idx] where weight is (n_layers, out, dim) sharded on axis=1
    devices = ("NULL:1", "NULL:2")
    a = Tensor.arange(8*4*10).reshape(8, 4, 10).clone().shard(devices, axis=1).realize()
@@ -15,13 +15,13 @@ def simplify_image_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move

 def get_gated_load_uop(valid:UOp, idx:UOp):
  return UOp(Ops.LOAD, dtypes.float, (
-    UOp(Ops.PARAM, dtypes.float.ptr(), arg=0).index(idx.valid(valid), ptr=True),
+    UOp.param(0, dtypes.float.ptr()).index(idx.valid(valid), ptr=True),
    UOp.const(dtypes.float, 0.0)
  ))

 def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UOp]):
  return UOp(Ops.LOAD, dtypes.float.vec(4), (
-    UOp(Ops.PARAM, dtypes.imagef(image_shape), arg=0).index(idx[1].valid(valid), idx[0].valid(valid), ptr=True),
+    UOp.param(0, dtypes.imagef(image_shape)).index(idx[1].valid(valid), idx[0].valid(valid), ptr=True),
    UOp(Ops.STACK, dtypes.float.vec(4), src=(UOp.const(dtypes.float, 0.0),) * 4)
  ))

@@ -513,7 +513,7 @@ class TestDropTrueGate(unittest.TestCase):
    from tinygrad.codegen.late.devectorizer import load_store_indexing
    from tinygrad.uop.ops import graph_rewrite
    from tinygrad.uop.symbolic import sym
-    buf = UOp(Ops.PARAM, dtypes.int.ptr(), arg=0)
+    buf = UOp.param(0, dtypes.int.ptr())
    idx = UOp.const(dtypes.weakint, 0)
    true_gate = UOp.const(dtypes.bool, True)
    index_with_gate = UOp(Ops.INDEX, dtypes.int.ptr(), (buf, idx.valid(true_gate)))
@@ -557,7 +557,7 @@ class TestRangeShrink(unittest.TestCase):
    # one load guards r < 4, but another load uses r without a gate -> no shrink
    r = Range(0, 204)
    load1 = get_gated_load_uop(r < UOp.const(dtypes.weakint, 4), r)
-    load2 = UOp(Ops.LOAD, dtypes.float, (UOp(Ops.PARAM, dtypes.float.ptr(), arg=1).index(r, ptr=True),))
+    load2 = UOp(Ops.LOAD, dtypes.float, (UOp.param(1, dtypes.float.ptr()).index(r, ptr=True),))
    ranges = self.get_ranges(UOp.sink(load1, load2))
    self.assertEqual(len(ranges), 1)
    self.assertEqual(ranges[0].src[0].arg, 204)
@@ -583,7 +583,7 @@ class TestRangeShrink(unittest.TestCase):
    from tinygrad.dtype import Invalid
    r = Range(0, 204)
    x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
-    ranges = self.get_ranges(UOp(Ops.PARAM, dtypes.float.ptr(), arg=0).index(r).store((r < 4).where(x, 0)).sink())
+    ranges = self.get_ranges(UOp.param(0, dtypes.float.ptr()).index(r).store((r < 4).where(x, 0)).sink())
    self.assertEqual(len(ranges), 1)
    self.assertEqual(ranges[0].src[0].arg, 4)

@@ -592,7 +592,7 @@ class TestRangeShrink(unittest.TestCase):
    from tinygrad.dtype import Invalid
    r = Range(0, 204)
    x = (r < 4).where(UOp.const(dtypes.float, 1), Invalid)
-    ranges = self.get_ranges(UOp(Ops.PARAM, dtypes.float.ptr(), arg=0).index(r).store((r < 4).where(0, x)).sink())
+    ranges = self.get_ranges(UOp.param(0, dtypes.float.ptr()).index(r).store((r < 4).where(0, x)).sink())
    self.assertEqual(len(ranges), 1)
    self.assertEqual(ranges[0].src[0].arg, 4)

@@ -3,8 +3,7 @@ from tinygrad import Tensor, dtypes
 from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, graph_rewrite

 _strip_unique_pm = PatternMatcher([
-  (UPat(Ops.CONST, src=(UPat(Ops.UNIQUE), UPat(Ops.DEVICE, name="d")), name="b"), lambda b,d: b.replace(src=(d,))),
-  (UPat(Ops.BUFFER, src=(UPat(Ops.UNIQUE), UPat(Ops.DEVICE, name="d")), name="b"), lambda b,d: b.replace(src=(UOp.unique(0), d))),
+  (UPat((Ops.UNIQUE, Ops.LUNIQUE), name="u"), lambda u: u.replace(arg=0) if u.arg != 0 else None),
 ])
 def _strip_unique(u: UOp) -> UOp: return graph_rewrite(u, _strip_unique_pm)

@@ -395,10 +394,8 @@ class TestTensorUOpCreation(unittest.TestCase):
    self.assertIs(_strip_unique(Tensor.full((2, 3), 42, dtype=dtypes.int8, device="NULL").uop),
                  _strip_unique(UOp.full((2, 3), 42, dtype=dtypes.int8, device="NULL")))
  def test_full_symbolic_fill(self):
-    # bound symbolic variable — flows through Tensor.__init__'s UOp branch, no UNIQUE added
    t = Tensor.full((2, 3), UOp.variable("x", 1, 10).bind(5))
    self.assertEqual(t.shape, (2, 3))
-    self.assertFalse(t.uop.op_in_backward_slice_with_self(Ops.UNIQUE))
  def test_zeros(self):
    self.assertIs(_strip_unique(Tensor.zeros(2, 3).uop), _strip_unique(UOp.zeros(2, 3)))
  def test_ones(self):
@@ -1,7 +1,7 @@
 import unittest, math
 import numpy as np
 from tinygrad import dtypes
-from tinygrad.uop.ops import UOp, Ops
+from tinygrad.uop.ops import UOp
 from tinygrad.uop.decompositions import TRANSCENDENTAL_DTYPES, payne_hanek_reduction, cody_waite_reduction
 from tinygrad.uop.decompositions import frexp, rintk, xpow, xexp2, xlog2, trig_poly, pow2if
 from test.helpers import eval_uop
@@ -10,7 +10,7 @@ class TestTranscendentalFunctions(unittest.TestCase):
  def test_payne_hanek_reduction(self):
    # TODO: Test constant input when constant folding is fixed (or maybe test both variants)
    # Load input value from a buffer to prevent constant folding
-    input_buf = UOp(Ops.PARAM, dtypes.double.ptr(), arg=1, src=())
+    input_buf = UOp.param(1, dtypes.double.ptr())
    loaded_value = input_buf.index(UOp.const(dtypes.int, 0))
    def eval_payne_hanek_reduction(v:float) -> tuple[float, int]:
      return tuple(eval_uop(u, [(dtypes.float64, [v])]) for u in payne_hanek_reduction(loaded_value))
@@ -260,7 +260,7 @@ class TestUOpGraph(unittest.TestCase):

  @unittest.skip("this test isn't valid uops")
  def test_noop_vectorize_fold(self):
-    d0 = UOp(Ops.PARAM, dtypes.float.ptr(), arg=0)
+    d0 = UOp.param(0, dtypes.float.ptr())
    idx = UOp.const(dtypes.int, 0)
    ld = UOp(Ops.LOAD, dtypes.float.vec(2), (d0, idx))
    vec = UOp(Ops.STACK, dtypes.float.vec(2), (ld,))
@@ -272,9 +272,9 @@ class TestUOpGraph(unittest.TestCase):

  @unittest.skip("this test isn't valid uops")
  def test_gep_vec_fold(self):
-    d0 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
-    d1 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 1)
-    d2 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 2)
+    d0 = UOp.param(0, dtypes.float.ptr())
+    d1 = UOp.param(1, dtypes.float.ptr())
+    d2 = UOp.param(2, dtypes.float.ptr())
    idx = UOp.const(dtypes.int, 0)
    def _test_vec(geps, count=4):
      vec = UOp(Ops.STACK, dtypes.float.vec(count), geps)
@@ -380,8 +380,8 @@ class TestUOpGraph(unittest.TestCase):
      self.assertEqual(uops[-2], wmma)  # -2 to skip SINK

  def test_cast_alu_fold(self):
-    d0 = UOp(Ops.PARAM, dtypes.bool.ptr(), arg=0)
-    d1 = UOp(Ops.PARAM, dtypes.int.ptr(), arg=1)
+    d0 = UOp.param(0, dtypes.bool.ptr())
+    d1 = UOp.param(1, dtypes.int.ptr())
    idx = UOp.const(dtypes.int, 0)
    ld = d1.index(idx)
    alu = (ld<1).cast(dtypes.bool)
@@ -390,8 +390,8 @@ class TestUOpGraph(unittest.TestCase):
    self.assertEqual(len([x for x in uops if x.op is Ops.CAST]), 0)

  def test_double_cast_fold(self):
-    d0 = UOp(Ops.PARAM, dtypes.float.ptr(), arg=0)
-    d1 = UOp(Ops.PARAM, dtypes.int.ptr(), arg=1)
+    d0 = UOp.param(0, dtypes.float.ptr())
+    d1 = UOp.param(1, dtypes.int.ptr())
    idx = UOp.const(dtypes.int, 0)
    ld = d1.index(idx)
    alu = ld.cast(dtypes.float).cast(dtypes.float)
@@ -414,7 +414,7 @@ class TestUOpGraph(unittest.TestCase):

  def test_bitcast_to_same_dtype_fold(self):
    for dt in dtypes.ints + dtypes.floats + (dtypes.bool,):
-      d0 = UOp(Ops.PARAM, dt.ptr(), arg=0)
+      d0 = UOp.param(0, dt.ptr())
      v = d0.index(UOp.const(dtypes.int, 0))
      uops = to_uops_list([v.bitcast(dt)])
      self.assertEqual(len([x for x in uops if x.op is Ops.BITCAST and x.dtype is dt]), 0, f"dtype = {dt}")
@@ -427,10 +427,10 @@ class TestUOpGraph(unittest.TestCase):

  def test_where_on_gated_load_fold(self):
    ridx0 = UOp.range(100, 0)
-    d0 = UOp(Ops.PARAM, dtypes.long.ptr(), (), 0)
+    d0 = UOp.param(0, dtypes.long.ptr())
    ld = d0.index(ridx0.valid(ridx0<50))
    w = (ridx0<50).where(ld, 5)
-    out = UOp(Ops.PARAM, dtypes.long.ptr(), (), 1)
+    out = UOp.param(1, dtypes.long.ptr())
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
      assert u.op is not Ops.WHERE
@@ -438,7 +438,7 @@ class TestUOpGraph(unittest.TestCase):

  def test_where_on_gated_load_folds_swapped_branches(self):
    ridx0 = UOp.range(100, 0)
-    d0 = UOp(Ops.PARAM, dtypes.long.ptr(), (), 0)
+    d0 = UOp.param(0, dtypes.long.ptr())
    ld = d0.index(ridx0.valid((ridx0<50).logical_not()))
    w = (ridx0<50).where(5, ld)
    uops = to_uops_list([w])
@@ -448,11 +448,11 @@ class TestUOpGraph(unittest.TestCase):

  def test_where_on_gated_load_with_cast(self):
    ridx0 = UOp.range(100, 0)
-    d0 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    d0 = UOp.param(0, dtypes.int.ptr())
    gate_idx = ridx0.valid((ridx0<50))
    ld = d0.index(gate_idx).cast(dtypes.float)
    w = (ridx0<50).where(ld, 5.0)
-    out = UOp(Ops.PARAM, dtypes.float.ptr(), (), 1)
+    out = UOp.param(1, dtypes.float.ptr())
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
      assert u.op is not Ops.WHERE
@@ -460,27 +460,27 @@ class TestUOpGraph(unittest.TestCase):

  def test_where_on_casted_gated_load_extra_cond(self):
    ridx0 = UOp.range(100, 0)
-    d0 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
+    d0 = UOp.param(0, dtypes.float.ptr())
    ld = d0.index(ridx0.valid(ridx0<50))
    w = ((ridx0<50) & (ridx0>30)).where(ld, UOp.const(dtypes.float, 0)).cast(dtypes.half)
-    out = UOp(Ops.PARAM, dtypes.half.ptr(), (), 1)
+    out = UOp.param(1, dtypes.half.ptr())
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
      assert u.op is not Ops.WHERE

  def test_where_on_casted_gated_load_extra_cond_swapped(self):
    ridx0 = UOp.range(100, 0)
-    d0 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
+    d0 = UOp.param(0, dtypes.float.ptr())
    ld = d0.index(ridx0.valid(ridx0<50))
    w = ((ridx0<50) & (ridx0>30)).where(UOp.const(dtypes.float, 0), ld).cast(dtypes.half)
-    out = UOp(Ops.PARAM, dtypes.half.ptr(), (), 1)
+    out = UOp.param(1, dtypes.half.ptr())
    uops = to_uops_list([out.index(ridx0).store(w)])
    for u in uops:
      assert u.op is not Ops.WHERE

  def test_where_in_store_becomes_gate(self):
    ridx0 = UOp.range(100, 0)
-    d0 = UOp(Ops.PARAM, dtypes.long.ptr(), (), 0)
+    d0 = UOp.param(0, dtypes.long.ptr())
    idx = d0.index(ridx0)
    ld = idx.load()
    val = (ridx0<50).where(5, ld)
@@ -493,14 +493,14 @@ class TestUOpGraph(unittest.TestCase):
  def test_load_idx_becomes_int(self):
    # mnist indexing with split reduceop
    # Make sure we are not doign math on the loaded index, which would promote it to long
-    c0 = UOp(Ops.PARAM, dtypes.uchar.ptr(128000), arg=0, src=())
+    c0 = UOp.param(0, dtypes.uchar.ptr(128000))
    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
-    c3 = UOp(Ops.PARAM, dtypes.int.ptr(512), arg=1, src=())
+    c3 = UOp.param(1, dtypes.int.ptr(512))
    c4 = c3.index(c1)
    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
    c6 = ((c2*UOp.const(dtypes.weakint, 240))+c5)
-    c7 = UOp(Ops.PARAM, dtypes.uchar.ptr(60000), arg=2, src=())
+    c7 = UOp.param(2, dtypes.uchar.ptr(60000))
    c8 = c7.index(c6)
    c9 = ((c4<0).where((c4+60000), c4)!=c6.cast(dtypes.int)).where(0, c8.cast(dtypes.uint).cast(dtypes.uchar)).reduce(c5, arg=Ops.ADD)
    c10 = c0.index(((c1*UOp.const(dtypes.weakint, 250))+c2)).store(c9).end(c1, c2)
@@ -510,14 +510,14 @@ class TestUOpGraph(unittest.TestCase):

  def test_load_idx_no_math_on_loaded(self):
    # test the (x+y)<c pattern where x has loads - we shouldn't do math on loaded indices
-    c0 = UOp(Ops.PARAM, dtypes.uchar.ptr(128000), arg=0, src=())
+    c0 = UOp.param(0, dtypes.uchar.ptr(128000))
    c1 = UOp.range(UOp.const(dtypes.weakint, 512), 1, AxisType.LOOP)
    c2 = UOp.range(UOp.const(dtypes.weakint, 250), 2, AxisType.LOOP)
-    c3 = UOp(Ops.PARAM, dtypes.int.ptr(512), arg=1, src=())
+    c3 = UOp.param(1, dtypes.int.ptr(512))
    c4 = c3.index(c1)  # c4 is a load
    c5 = UOp.range(UOp.const(dtypes.weakint, 240), 0, AxisType.REDUCE)
    c6 = ((c2*UOp.const(dtypes.weakint, 240))+c5)
-    c7 = UOp(Ops.PARAM, dtypes.uchar.ptr(60000), arg=2, src=())
+    c7 = UOp.param(2, dtypes.uchar.ptr(60000))
    c8 = c7.index(c6)
    # (loaded + range) < const pattern - loaded value shouldn't be promoted to long
    loaded_idx = c4.cast(dtypes.weakint)
@@ -529,9 +529,9 @@ class TestUOpGraph(unittest.TestCase):
      self.assertNotEqual(u.dtype, dtypes.long)

  def test_fold_gated_load(self):
-    glbl0 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
-    glbl1 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 1)
-    glbl2 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 2)
+    glbl0 = UOp.param(0, dtypes.int.ptr())
+    glbl1 = UOp.param(1, dtypes.int.ptr())
+    glbl2 = UOp.param(2, dtypes.int.ptr())
    idx = UOp.const(dtypes.int, 0)
    ld0 = glbl1.index(UOp.invalid())
    ld1 = glbl2.index(idx.valid(UOp.const(dtypes.bool, True)))
@@ -541,7 +541,7 @@ class TestUOpGraph(unittest.TestCase):
    self.assertEqual(ld0, UOp.load(glbl2.index(idx, ptr=True), dtype=dtypes.int))

  def test_fold_gated_load_local(self):
-    glbl0 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    glbl0 = UOp.param(0, dtypes.int.ptr())
    smem = UOp(Ops.DEFINE_LOCAL, dtypes.int.ptr(size=18, addrspace=AddrSpace.LOCAL), (), "temp")
    lidx = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 16),), "lidx0")
    st = UOp(Ops.STORE, dtypes.void, (smem.index(lidx, ptr=True), glbl0.index(lidx, ptr=True).load()))
@@ -555,7 +555,7 @@ class TestUOpGraph(unittest.TestCase):
    self.assertEqual(ld0.src[0], smem.after(barrier).index(lidx+2, ptr=True))

  def test_fold_gated_store(self):
-    glbl = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    glbl = UOp.param(0, dtypes.int.ptr())
    idx0 = UOp.const(dtypes.int, 0)
    idx1 = UOp.const(dtypes.int, 0)
    val = UOp.const(dtypes.int, 42)
@@ -563,12 +563,12 @@ class TestUOpGraph(unittest.TestCase):
    st1 = glbl.index(idx0.valid(UOp.const(dtypes.bool, True)), ptr=True).store(val)
    uops = to_uops_list([st0, st1])
    # only the second store happens
-    self.assertEqual(len(uops), 6)  # +1 for SINK
+    self.assertEqual(len(uops), 7)  # +1 for SINK, +1 for PARAM shape sentinel
    self.assertEqual(uops[-2], glbl.index(idx1, ptr=True).store(val))  # -2 to skip SINK

  @unittest.skip("this is a uop type error")
  def test_asserts_bad_gate(self):
-    glbl0 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
+    glbl0 = UOp.param(0, dtypes.int.ptr())
    idx = UOp.const(dtypes.int, 0)
    bad_gate = UOp.const(dtypes.int, 1)
    with self.assertRaises(AssertionError): to_uops_list([UOp(Ops.STORE, dtypes.void, (glbl0, idx, UOp.const(dtypes.int, 42), bad_gate))])
@@ -779,7 +779,7 @@ class TestLoadStoreFolding(unittest.TestCase):
  def test_gated_load_gep_preserves_alt(self):
    """Test that LOAD(GEP, alt) preserves alt value after rewrite"""
    from tinygrad.codegen.late.devectorizer import load_store_folding
-    buf = UOp(Ops.PARAM, dtypes.float.vec(4).ptr(), (), 0)
+    buf = UOp.param(0, dtypes.float.vec(4).ptr())
    idx = UOp.const(dtypes.int, 0)
    gate = UOp.const(dtypes.bool, True)
    gated_index = buf.index(idx.valid(gate))
@@ -797,8 +797,8 @@ class TestLoadStoreFolding(unittest.TestCase):
  def test_gated_load_ptrcat_preserves_alt(self):
    """Test that LOAD(PTRCAT, alt) preserves alt value after rewrite"""
    from tinygrad.codegen.late.devectorizer import load_store_folding
-    buf1 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
-    buf2 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 1)
+    buf1 = UOp.param(0, dtypes.float.ptr())
+    buf2 = UOp.param(1, dtypes.float.ptr())
    idx = UOp.const(dtypes.int, 0)
    idx1 = buf1.index(idx)
    idx2 = buf2.index(idx)
@@ -951,7 +951,7 @@ class TestSymbolic(unittest.TestCase):
    expr = cond.where(a, b).cast(dtypes.half)

    # TODO: copied from render, render does not support cast
-    glbl = UOp(Ops.PARAM, dtypes.int.ptr(), arg=0)
+    glbl = UOp.param(0, dtypes.int.ptr())
    uops = get_uops(UOp(Ops.STORE, dtypes.void, (glbl.index(UOp.const(dtypes.int, 0)), expr)).sink())
    rewritten_uop = [uop for uop in uops if uop.op is Ops.STORE][0].src[1]

@@ -1270,7 +1270,7 @@ class TestStoreLoadFolding(unittest.TestCase):
  """Tests for store(index, load(index)) -> NOOP rule. This rule matches patterns that EMERGE during simplification."""
  def test_store_load_folding(self):
    # store(idx, load(idx)) -> NOOP, including emergent patterns like store(idx, load(idx) + 0)
-    buf = UOp(Ops.PARAM, dtypes.int.ptr(), arg=0)
+    buf = UOp.param(0, dtypes.int.ptr())
    index = buf.index(UOp.const(dtypes.weakint, 0))
    # Direct: store(idx, load(idx)) -> NOOP
    self.assertEqual(graph_rewrite(index.store(index.load()), sym).op, Ops.NOOP)
@@ -1340,7 +1340,7 @@ class TestRangeSplitting(unittest.TestCase):
    from tinygrad.codegen.simplify import pm_split_ranges, pm_flatten_range
    r0 = UOp.range(uconst(8), 0)
    # create a simple expression using the range with mod: store range%2 to a buffer
-    buf = UOp(Ops.PARAM, dtypes.int.ptr(), arg=0)
+    buf = UOp.param(0, dtypes.int.ptr())
    val = (r0 % uconst(2)).cast(dtypes.int)
    store = UOp(Ops.STORE, dtypes.void, (buf.index(uconst(0)), val))
    sink = UOp(Ops.SINK, dtypes.void, (UOp(Ops.END, dtypes.void, (store, r0)),))
@@ -82,7 +82,7 @@ class TestVminVmaxProperties(unittest.TestCase):
  def test_vmin_vmax_multiplication_0_inf(self):
    # vmin and vmax for multiplication with a variable
    x = UOp.const(dtypes.float, 0.0)
-    y = UOp.load(UOp(Ops.PARAM, dtypes.float.ptr(1), (), 0), UOp.const(dtypes.int, 0), dtype=dtypes.float)
+    y = UOp.load(UOp.param(0, dtypes.float.ptr(1)), UOp.const(dtypes.int, 0), dtype=dtypes.float)
    uop = x * y
    # TODO: these should be 0, but definitely should not be nan
    self.assertEqual(uop.vmin, -math.inf)
@@ -316,7 +316,7 @@ class TestVminVmaxVConst(unittest.TestCase):

  def test_vmin_vmax_vector_with_gep(self):
    # vmin and vmax for a vector constant of bool values
-    d1 = UOp(Ops.PARAM, dtypes.int.ptr(), (), 1)
+    d1 = UOp.param(1, dtypes.int.ptr())
    idx = UOp.const(dtypes.int, 0)
    val = UOp(Ops.LOAD, dtypes.int.vec(2), (d1.index(idx).cast(dtypes.int.vec(2).ptr()),))
    uop = (val // 32).gep(0)
@@ -110,7 +110,7 @@ class TestExecALU(unittest.TestCase):

 class TestGatedStoreRewrite(unittest.TestCase):
  def test_tiny_gate_store(self):
-    gmem = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
+    gmem = UOp.param(0, dtypes.float.ptr())
    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
    gate = gidx0<UOp.const(dtypes.int, 1)
    idx = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem, (gidx0 * UOp.const(dtypes.int, 2)).valid(gate)))
@@ -123,10 +123,11 @@ class TestGatedStoreRewrite(unittest.TestCase):
    gated_uops = tuple(uops[uops.index(if_uop)+1:uops.index(endif)])
    self.assertEqual(len(gated_uops), 1)
    self.assertIs(gated_uops[-1].op, Ops.STORE)
+    self.assertEqual(len(gated_uops[-1].src), 2)

  def test_gate_some_stores(self):
-    gmem0 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
-    gmem1 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 1)
+    gmem0 = UOp.param(0, dtypes.float.ptr())
+    gmem1 = UOp.param(1, dtypes.float.ptr())
    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
    idx = gidx0 * UOp.const(dtypes.int, 2)
    idx0 = UOp(Ops.INDEX, dtypes.float.ptr(), (gmem0, idx.valid(gidx0<UOp.const(dtypes.int, 1))))
@@ -140,12 +141,13 @@ class TestGatedStoreRewrite(unittest.TestCase):
    gated_uops = tuple(uops[uops.index(if_uop)+1:uops.index(endif)])
    self.assertEqual(len(gated_uops), 1)
    self.assertIs(gated_uops[-1].op, Ops.STORE)
+    self.assertEqual(len(gated_uops[-1].src), 2)

  # scaled down version of TestLinearizerDumb.test_unmerged_ifs
  @unittest.skip("we don't merge ifs anymore")
  def test_merge_ifs_alt(self):
-    gmem0 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 0)
-    gmem1 = UOp(Ops.PARAM, dtypes.float.ptr(), (), 1)
+    gmem0 = UOp.param(0, dtypes.float.ptr())
+    gmem1 = UOp.param(1, dtypes.float.ptr())
    gidx0 = UOp(Ops.SPECIAL, dtypes.int, (UOp.const(dtypes.int, 4),), 'gidx0')
    idx = gidx0*UOp.const(dtypes.int, 2)
    gate = gidx0<UOp.const(dtypes.int, 1)
@@ -161,13 +163,14 @@ class TestGatedStoreRewrite(unittest.TestCase):
    gated_uops = tuple(uops[uops.index(ifs[0])+1:uops.index(endifs[0])])
    self.assertEqual(len(gated_uops), 2)
    for x in gated_uops: self.assertIs(x.op, Ops.STORE)
+    for x in gated_uops: self.assertEqual(len(x.src), 2)

@unittest.skipIf(Device.DEFAULT == "METAL", "compiler bug")
@unittest.skipUnless(Ops.SHR in Device[Device.DEFAULT].renderer.code_for_op, "fast_idiv requires SHR")
 class TestFastIdiv(unittest.TestCase):
  def test_division_power_of_two(self):
    for dt in (dtypes.int32, dtypes.uint32):
-      g = UOp(Ops.PARAM, dt.ptr(), (), 0)
+      g = UOp.param(0, dt.ptr())
      c = UOp.const(dt, 2)
      l = g.index(c)
      a = UOp(Ops.CDIV, dt, (l, c))
@@ -180,7 +183,7 @@ class TestFastIdiv(unittest.TestCase):
  def test_floormod_power_of_two(self):
    # FLOORMOD by a power of two lowers to AND (correct floor mod for any sign in two's complement)
    for dt in (dtypes.int32, dtypes.uint32):
-      g = UOp(Ops.PARAM, dt.ptr(), (), 0)
+      g = UOp.param(0, dt.ptr())
      c = UOp.const(dt, 8)
      a = UOp(Ops.FLOORMOD, dt, (g.index(c), c))
      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
@@ -192,7 +195,7 @@ class TestFastIdiv(unittest.TestCase):
  def test_floordiv_power_of_two_uint(self):
    # uint FLOORDIV by a power of two lowers to a shift, leaving no IDIV/FLOORDIV in the kernel
    for dt in (dtypes.uint32, dtypes.uint64):
-      g = UOp(Ops.PARAM, dt.ptr(), (), 0)
+      g = UOp.param(0, dt.ptr())
      c = UOp.const(dt, 2)
      a = UOp(Ops.FLOORDIV, dt, (g.index(c), c))
      uops = to_uops_list([a], ren=Device[Device.DEFAULT].renderer)
@@ -204,7 +207,7 @@ class TestFastIdiv(unittest.TestCase):
  @Context(DISABLE_FAST_IDIV=0)
  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "WEBGPU doesn't support long")
  def test_fast_idiv_and_mod(self):
-    g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)
+    g = UOp.param(0, dtypes.uint32.ptr())
    c = UOp.const(dtypes.uint, 3)
    l = g.index(c)
    a = UOp(Ops.CDIV, dtypes.uint, (l, c))
@@ -239,7 +242,7 @@ class TestFastIdiv(unittest.TestCase):
  @unittest.expectedFailure
  def test_fast_idiv_overflow(self):
    # This will be possible with a slightly different method for fast_idiv
-    g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)
+    g = UOp.param(0, dtypes.uint32.ptr())
    c = UOp.const(dtypes.uint, 7)
    l = UOp(Ops.LOAD, dtypes.uint, (g.index(c),))
    a = UOp(Ops.CDIV, dtypes.uint, (l, c))
@@ -250,7 +253,7 @@ class TestFastIdiv(unittest.TestCase):
    self.assertNotIn(Ops.CDIV, ops)

  def test_disable_fast_idiv(self):
-    g = UOp(Ops.PARAM, dtypes.uint32.ptr(), (), 0)
+    g = UOp.param(0, dtypes.uint32.ptr())
    c = UOp.const(dtypes.uint, 3)
    l = g.index(c)
    a = UOp(Ops.CDIV, dtypes.uint, (l, c))
@@ -287,8 +290,8 @@ class TestUOpMethod(unittest.TestCase):
    self.assertEqual((gidx0*3+1).const_factor(), 1)

  def test_replace(self):
-    x = UOp(Ops.PARAM, dtypes.int.ptr(), (), 0)
-    self.assertIs(x.replace(arg=None).arg, None)
+    x = UOp.param(0, dtypes.int.ptr())
+    self.assertEqual(x.replace(arg=UOp.param(1, dtypes.int.ptr()).arg).arg.slot, 1)
    with self.assertRaises(AssertionError): x.replace(field="a")

  def test_const_zero_neg_zero_different(self):
@@ -139,7 +139,7 @@ class TestUOpsStats(unittest.TestCase):

  #MULACC should have the same stats as MUL + ADD
  def test_mulacc(self):
-    globl = UOp(Ops.PARAM, dtypes.int.ptr(), tuple())
+    globl = UOp.param(0, dtypes.int.ptr())
    o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
    o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
    u1 = globl.index(o1)
@@ -149,7 +149,7 @@ class TestUOpsStats(unittest.TestCase):
    u5 = UOp(Ops.ADD, dtypes.int, (u4,u3))
    uops = tuple(u5.toposort())

-    globl = UOp(Ops.PARAM, dtypes.int.ptr(), tuple())
+    globl = UOp.param(0, dtypes.int.ptr())
    o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
    o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
    u1 = globl.index(o1)
@@ -11,7 +11,7 @@ class TestValidateOOB(unittest.TestCase):
  # basic index patterns
  def test_const_index(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      to_uops_list([buf.index(UOp.const(dtypes.int, 0), ptr=True).load(dtype=dtypes.int)])  # valid
      to_uops_list([buf.index(UOp.const(dtypes.int, 15), ptr=True).load(dtype=dtypes.int)])  # valid (last element)
      with self.assertRaises(RuntimeError):
@@ -21,7 +21,7 @@ class TestValidateOOB(unittest.TestCase):

  def test_variable_index(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      to_uops_list([buf.index(Variable("i", 0, 15), ptr=True).load(dtype=dtypes.int)])  # valid
      with self.assertRaises(RuntimeError):
        to_uops_list([buf.index(Variable("i", 0, 20), ptr=True).load(dtype=dtypes.int)])  # oob
@@ -30,7 +30,7 @@ class TestValidateOOB(unittest.TestCase):

  def test_range_with_mask(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      r = UOp.range(42, 0, AxisType.GLOBAL)
      to_uops_list([buf.index(r.valid(r < 16), ptr=True).load(dtype=dtypes.int)])  # valid
      with self.assertRaises(RuntimeError):
@@ -38,7 +38,7 @@ class TestValidateOOB(unittest.TestCase):

  def test_variable_with_mask(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      v = Variable("v", -5, 80)
      to_uops_list([buf.index(v.valid((v >= 0) & (v < 16)), ptr=True).load(dtype=dtypes.int)])  # valid
      with self.assertRaises(RuntimeError):
@@ -46,7 +46,7 @@ class TestValidateOOB(unittest.TestCase):

  def test_gated_store(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      v = Variable("v", 0, 20)
      to_uops_list([buf.index(v.valid(v < 16), ptr=True).store(0)])  # valid
      with self.assertRaises(RuntimeError):
@@ -55,14 +55,14 @@ class TestValidateOOB(unittest.TestCase):
  # ALU ops in index
  def test_floordiv(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      to_uops_list([buf.index(UOp.range(32, 0, AxisType.GLOBAL) // 2, ptr=True).load(dtype=dtypes.int)])  # 0..15 valid
      with self.assertRaises(RuntimeError):
        to_uops_list([buf.index(UOp.range(34, 0, AxisType.GLOBAL) // 2, ptr=True).load(dtype=dtypes.int)])  # 0..16 oob

  def test_mod(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      r = UOp.range(100, 0, AxisType.GLOBAL)
      to_uops_list([buf.index(r % 16, ptr=True).load(dtype=dtypes.int)])  # 0..15 valid
      with self.assertRaises(RuntimeError):
@@ -70,14 +70,14 @@ class TestValidateOOB(unittest.TestCase):

  def test_shr(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      to_uops_list([buf.index(UOp.range(64, 0, AxisType.GLOBAL) >> 2, ptr=True).load(dtype=dtypes.int)])  # 0..15 valid
      with self.assertRaises(RuntimeError):
        to_uops_list([buf.index(UOp.range(128, 0, AxisType.GLOBAL) >> 2, ptr=True).load(dtype=dtypes.int)])  # 0..31 oob

  def test_shl(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(64), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(64))
      r = UOp.range(8, 0, AxisType.GLOBAL)
      to_uops_list([buf.index(r << 2, ptr=True).load(dtype=dtypes.int)])  # 0..28 valid
      with self.assertRaises(RuntimeError):
@@ -85,7 +85,7 @@ class TestValidateOOB(unittest.TestCase):

  def test_and(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      r = UOp.range(100, 0, AxisType.GLOBAL)
      to_uops_list([buf.index(r & 15, ptr=True).load(dtype=dtypes.int)])  # 0..15 valid
      with self.assertRaises(RuntimeError):
@@ -93,14 +93,14 @@ class TestValidateOOB(unittest.TestCase):

  def test_max(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      to_uops_list([buf.index(Variable("v", -10, 15).maximum(0), ptr=True).load(dtype=dtypes.int)])  # 0..15 valid
      with self.assertRaises(RuntimeError):
        to_uops_list([buf.index(Variable("v2", -10, 20).maximum(0), ptr=True).load(dtype=dtypes.int)])  # 0..20 oob

  def test_xor_in_mask(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      r = UOp.range(32, 0, AxisType.GLOBAL)
      to_uops_list([buf.index(r.valid((r < 8) ^ ((r >= 8) & (r < 16))), ptr=True).load(dtype=dtypes.int)])  # 0..15 valid
      with self.assertRaises(RuntimeError):
@@ -109,22 +109,22 @@ class TestValidateOOB(unittest.TestCase):
  # cast patterns
  def test_float_cast_in_index(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(16))
      r = UOp.range(20, 0)
      i = (r.cast(dtypes.float) * 0.68).trunc().cast(dtypes.int)
      to_uops_list([buf.index(i.valid((i >= 0) & (i < 16)), ptr=True).load(dtype=dtypes.int)])

  def test_bool_cast_in_mask(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf = UOp(Ops.PARAM, dtypes.int.ptr(1), (), 0)
+      buf = UOp.param(0, dtypes.int.ptr(1))
      r = UOp.range(20, 0)
      to_uops_list([buf.index(r.valid(r.cast(dtypes.bool).logical_not()), ptr=True).load(dtype=dtypes.int)])  # only r=0 valid

  # load result as index/mask
  def test_load_as_index(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf0 = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
-      buf1 = UOp(Ops.PARAM, dtypes.int.ptr(64), (), 1)
+      buf0 = UOp.param(0, dtypes.int.ptr(16))
+      buf1 = UOp.param(1, dtypes.int.ptr(64))
      r = UOp.range(42, 0, AxisType.GLOBAL)
      ld0 = buf0.index(r.valid(r < 8), ptr=True).load(dtype=dtypes.int).cast(dtypes.weakint)
      to_uops_list([buf1.index((ld0 * 2).valid((ld0 >= 0) & (ld0 < 32)), ptr=True).load(dtype=dtypes.int)])  # valid
@@ -133,8 +133,8 @@ class TestValidateOOB(unittest.TestCase):

  def test_load_bool_as_mask(self):
    with Context(CHECK_OOB=1, SPEC=2):
-      buf_bool = UOp(Ops.PARAM, dtypes.bool.ptr(16), (), 0)
-      buf_int = UOp(Ops.PARAM, dtypes.int.ptr(8), (), 1)
+      buf_bool = UOp.param(0, dtypes.bool.ptr(16))
+      buf_int = UOp.param(1, dtypes.int.ptr(8))
      gidx = UOp(Ops.SPECIAL, dtypes.weakint, (UOp.const(dtypes.weakint, 16),), "gidx0")
      ld_bool = buf_bool.index(gidx, ptr=True).load()
      with self.assertRaises(RuntimeError):
@@ -145,7 +145,7 @@ class TestValidateOOB(unittest.TestCase):
  def test_in_bounds_access_gated_local(self):
    with Context(CHECK_OOB=1):
      # Define buffers
-      gbuf = UOp(Ops.PARAM, dtypes.uint.ptr(400), (), 0)
+      gbuf = UOp.param(0, dtypes.uint.ptr(400))
      sbuf = UOp(Ops.DEFINE_LOCAL, dtypes.uint.ptr(8, addrspace=AddrSpace.LOCAL), (), "temp0")

      # Define indices, valids and barrier
@@ -169,8 +169,8 @@ class TestValidateOOB(unittest.TestCase):
  @unittest.skip("Bool load is not supported yet")
  def test_load_mask(self):
    with Context(CHECK_OOB=1):
-      glbl0 = UOp(Ops.PARAM, dtypes.int.ptr(16), (), 0)
-      mask = UOp(Ops.PARAM, dtypes.bool.ptr(16), (), 0)
+      glbl0 = UOp.param(0, dtypes.int.ptr(16))
+      mask = UOp.param(0, dtypes.bool.ptr(16))
      ridx = UOp.range(20, 0)
      ld0 = UOp(Ops.LOAD, dtypes.int, (glbl0.index(UOp.const(ridx, ridx<16&mask), ptr=True)))
      to_uops_list([ld0])
@@ -1,19 +1,18 @@
-import unittest, decimal, sys, json, contextlib, tempfile, pickle, io
+import unittest, decimal, sys, json, contextlib, tempfile, pickle, io, math
 from pathlib import Path
 from dataclasses import dataclass
 from typing import Generator

 from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, TrackedPatternMatcher, graph_rewrite, track_rewrites, profile_matches
 from tinygrad.uop.symbolic import sym
-from tinygrad.dtype import dtypes
+from tinygrad.dtype import dtypes, AddrSpace
 from tinygrad.helpers import colored, ansistrip, flatten, TracingKey, ProfileRangeEvent, ProfileEvent, Context, cpu_events, profile_marker
 from tinygrad.helpers import cpu_profile, ProfilePointEvent, unwrap
 from tinygrad.device import Buffer

 from tinygrad.uop.ops import tracked_keys, tracked_ctxs, uop_fields, active_rewrites, active_group, _name_cnt, RewriteTrace
-from tinygrad.viz.serve import load_rewrites, get_full_rewrite, uop_to_json, VizData, get_render
-from tinygrad.codegen import to_program_cache
-from tinygrad.codegen import to_program
+from tinygrad.viz.serve import load_rewrites, get_full_rewrite, uop_to_json, VizData, get_render, addrspace_colors
+from tinygrad.codegen import do_to_program

@track_rewrites(name=True)
 def exec_rewrite(sink:UOp, pm_lst:list[PatternMatcher], names:None|list[str]=None) -> UOp:
@@ -41,7 +40,6 @@ class VizTrace:
@contextlib.contextmanager
 def save_viz():
  for lst in [tracked_keys, tracked_ctxs, active_rewrites, active_group, _name_cnt]: lst.clear()
-  to_program_cache.clear()
  Buffer.profile_events.clear()
  cpu_events.clear()
  viz = VizTrace()
@@ -219,32 +217,47 @@ class TestViz(unittest.TestCase):
    with save_viz() as viz:
      a = UOp.variable("a", 0, 10, dtype=dtypes.int)
      z = UOp.const(a.dtype, 0)
+      y = UOp.const(dtypes.float, math.pi)
      alu = a*z
-      exec_rewrite(alu, [sym])
+      ret = exec_rewrite(sink:=UOp.sink(alu, y), [sym])
    lst = viz.list_items()
    self.assertEqual(len(lst), 1)
    graphs = [x["graph"] for x in viz.get_details(0, 0)]
-    # embed const in the parent node when possible
-    self.assertEqual(list(graphs[0]), [id(a), id(alu)])
-    self.assertEqual(list(graphs[1]), [id(z)])
+    # const is always in the graph, client side hides exclude=True nodes by default
+    self.assertEqual(list(graphs[0]), [id(a), id(z), id(alu), id(y), id(sink)])
+    self.assertTrue(graphs[0][id(z)]["exclude"])
+    self.assertTrue(graphs[0][id(y)]["exclude"])
+    self.assertFalse(graphs[0][id(alu)]["exclude"])
+    self.assertEqual(graphs[0][id(y)]["label"].split("\n")[:2], ["CONST", "3.14159"])
+    self.assertEqual(list(graphs[1]), [id(z), id(y), id(ret)])

-  # TODO: DEFINE_VAR (shape ()) now gets wrapped in RESHAPE+EXPAND when broadcast against a shaped operand
-  # (due to shared OpMixin._binop using _broadcasted). Either extend viz to fold RESHAPE/EXPAND around
-  # DEFINE_VAR/RANGE/SPECIAL the way it does for CONST, or redesign scalar-compiler-op broadcasting.
-  @unittest.expectedFailure
  def test_const_reshape_expand_folded(self):
    # CONST->RESHAPE->EXPAND should be folded into the ALU node, not shown as separate RESHAPE/EXPAND nodes
    c = UOp.const(dtypes.float, 1.0, device="CPU", shape=(3,4))  # creates CONST->RESHAPE->EXPAND chain
    a = UOp(Ops.DEFINE_VAR, dtypes.float, arg=("a", 0.0, 10.0))
    alu = a + c
-    graph = uop_to_json(VizData(), alu)
-    # the RESHAPE and EXPAND nodes from the const should not appear in the graph
-    labels = {v["label"].split("\n")[0] for v in graph.values()}
-    self.assertNotIn("RESHAPE", labels)
-    self.assertNotIn("EXPAND", labels)
-    # the CONST should be inlined into the ALU node's label
-    alu_label = graph[id(alu)]["label"]
-    self.assertIn("CONST", alu_label)
+    with save_viz() as viz:
+      graph_rewrite(alu, PatternMatcher([]))
+    graph = [x["graph"] for x in viz.get_details(0, 0)][0]
+    excluded_nodes = {v["label"].split("\n")[0] for v in graph.values() if v["exclude"]}
+    self.assertIn("CONST", excluded_nodes)
+    self.assertIn("STACK", excluded_nodes)
+    self.assertIn("RESHAPE", excluded_nodes)
+    self.assertIn("EXPAND", excluded_nodes)
+    self.assertIn("CONST1 1 Ops.DEVICE", graph[id(alu)]["label"])
+
+  def test_stack_movement_not_folded_unless_all_const(self):
+    a = UOp.variable("a", 0, 10, dtype=dtypes.int)
+    c = UOp.const(dtypes.int, 1)
+    stack = a.vectorize(c)
+    reshaped = stack.reshape((1, 2))
+    graph = uop_to_json(VizData(), reshaped)
+    self.assertFalse(graph[id(stack)]["exclude"])
+
+    const_stack = c.vectorize(UOp.const(dtypes.int, 2))
+    const_reshaped = const_stack.reshape((1, 2))
+    const_graph = uop_to_json(VizData(), const_reshaped)
+    self.assertTrue(const_graph[id(const_stack)]["exclude"])

 # VIZ displays nested graph_rewrites in a tree view

@@ -327,12 +340,15 @@ class TestVizIntegration(unittest.TestCase):
  def test_codegen_tracing(self):
    with save_viz() as viz:
      ast = (Tensor.empty(4)+Tensor.empty(4)).schedule_linear().src[0].src[0]
-      prg = to_program(ast, Device[Device.DEFAULT].renderer)
+      prg = do_to_program(ast, Device[Device.DEFAULT].renderer)
    lst = viz.list_items()
    self.assertEqual(len(lst), 3)
    self.assertEqual(lst[0]["name"], "Callify 1 Buffer n1")
    self.assertEqual(lst[1]["name"], "Schedule 1 Kernel n1")
    self.assertEqual(lst[2]["name"], prg.arg.name)
+    input_ast = next(viz.get_details(2, 0))["graph"].values()
+    for u in input_ast:
+      if u["label"].startswith("PARAM\n"): self.assertEqual(u["addrspace"], addrspace_colors[AddrSpace.GLOBAL])

  # schedule graph CALL nodes have a link to jump to codegen
  def test_link_sched_codegen(self):
@@ -344,7 +360,7 @@ class TestVizIntegration(unittest.TestCase):
      from tinygrad.engine.realize import compile_linear
      sched = compile_linear(sched)
      with Context(NO_COLOR=0):
-        prgs = [to_program(si.src[0], Device[c1.device].renderer).arg.name for si in sched.src]
+        prgs = [do_to_program(si.src[0], Device[c1.device].renderer).arg.name for si in sched.src]
    lst = viz.list_items()
    sched_idx = next(i for i,l in enumerate(lst) if l["name"].startswith("Schedule"))
    viz_kernel = next(i for i,s in enumerate(lst[sched_idx]["steps"]) if s["name"] == "View Kernel Graph")
@@ -406,9 +422,16 @@ class TestVizIntegration(unittest.TestCase):
      default_test(c+2)
    ls = viz.list_items()
    self.assertEqual(len(ls), 2)
-    self.assertEqual(list(next(viz.get_details(0, 0))["graph"]), [id(c+1)])
+    graph = next(viz.get_details(0, 0))["graph"]
+    self.assertEqual(list(graph), [id(c), id(c+1)])
+    self.assertTrue(graph[id(c)]["exclude"])
+    self.assertFalse(graph[id(c+1)]["exclude"])
    self.assertEqual(list(next(viz.get_details(1, 0))["graph"]), [id(c)])
-    self.assertEqual(list(next(viz.get_details(1, 1))["graph"]), [id(c+2)])
+    graph = next(viz.get_details(1, 1))["graph"]
+    self.assertEqual(list(graph), [id(c), id(c.const_like(2)), id(c+2)])
+    self.assertTrue(graph[id(c)]["exclude"])
+    self.assertTrue(graph[id(c.const_like(2))]["exclude"])
+    self.assertFalse(graph[id(c+2)]["exclude"])

  def test_recurse(self):
    with save_viz() as viz:
@@ -744,7 +767,7 @@ class TestCfg(unittest.TestCase):
    with save_viz() as viz:
      with Context(DEV=f"NULL::{self.arch}"):
        out = Tensor.custom_kernel(Tensor.empty(1), fxn=fxn)[0]
-        _ = to_program(out.schedule_linear().src[-1].src[0], Device[out.device].renderer)
+        _ = do_to_program(out.schedule_linear().src[-1].src[0], Device[out.device].renderer)
    codegen_rewrites = next(s for s in viz.list_items() if s["name"] == name)
    disasm = next(s for s in codegen_rewrites["steps"] if s["name"] == "View Disassembly")
    return get_render(viz.data, disasm["query"])
@@ -981,8 +1004,9 @@ class TestCLI(unittest.TestCase):
  def test_dedup(self):
    with save_viz() as viz:
      for _ in range(CNT:=4):
-        Tensor.empty(4, device="NULL").add(1).realize()
-        Tensor.empty(8, device="NULL").add(1).realize()
+        # use kernel names unique to this test
+        Tensor.custom_kernel(Tensor.empty(4, device="NULL"), fxn=lambda _: UOp.sink(arg=KernelInfo("k1_test_viz_dedup")))[0].realize()
+        Tensor.custom_kernel(Tensor.empty(8, device="NULL"), fxn=lambda _: UOp.sink(arg=KernelInfo("k2_test_viz_dedup")))[0].realize()
    with write_files(viz) as files, Context(NO_COLOR=1):
      name = run_cli(*files, "-s", "NULL")[0]["name"]
      with Context(DEBUG=3):
@@ -1008,5 +1032,22 @@ class TestCLI(unittest.TestCase):
    for i,n in enumerate(call_nodes):
      assert prgs[i] in n["label"], f"CALL must contain kernel name, got {n['label']}"

+  def test_interval(self):
+    def emit_kernel(name:str): Tensor.custom_kernel(Tensor.empty(1, device="NULL"), fxn=lambda _: UOp.sink(arg=KernelInfo(name=name)))[0].realize()
+    with save_viz() as viz:
+      emit_kernel("pre_1")
+      emit_kernel("pre_2")
+      profile_marker("interval_start")
+      emit_kernel("target_1")
+      emit_kernel("target_2")
+      profile_marker("interval_end")
+      emit_kernel("post_1")
+      emit_kernel("post_2")
+    with write_files(viz) as files, Context(NO_COLOR=1):
+      flat = run_cli(*files, "-s", "NULL", "--interval", "interval_start", "interval_end")
+      aggregate = run_cli(*files, "-s", "NULL", "--interval", "interval_start", "interval_end", "-t")
+    self.assertEqual([s["name"] for s in flat], ["interval_start", "target_1", "target_2", "interval_end"])
+    self.assertEqual(sorted(s["name"] for s in aggregate), ["target_1", "target_2"])
+
 if __name__ == "__main__":
  unittest.main()
@@ -1,7 +1,6 @@
 import unittest, sys
 from tinygrad import Tensor, GlobalCounters, dtypes, Context
-from tinygrad.helpers import Profiling, WINO
-from test.helpers import CI
+from tinygrad.helpers import WINO

@unittest.skipIf(sys.platform.startswith("win"), "flaky on Windows")
 class TestWinograd(unittest.TestCase):
@@ -11,11 +10,6 @@ class TestWinograd(unittest.TestCase):
  def tearDown(self):
    WINO.value = self.old

-  def test_profile(self):
-    x,w = Tensor.rand(1,4,9,9).realize(), Tensor.rand(4,4,3,3).realize()
-    with Profiling(enabled=not CI, sort='time'):
-      Tensor.conv2d(x,w).realize()
-
  def test_forward_kernels(self):
    x,w = Tensor.rand(1,4,9,9).realize(), Tensor.rand(4,4,3,3).realize()
    out = Tensor.conv2d(x,w)
@@ -1,7 +1,6 @@
-import unittest, numpy as np
+import unittest, numpy as np, os
 from tinygrad import Tensor, Device, TinyJit
-from tinygrad.helpers import Timing, OSX, getenv
-from test.helpers import CI
+from tinygrad.helpers import Timing, getenv
 import multiprocessing.shared_memory as shared_memory

 N = getenv("NSZ", 256)
@@ -12,7 +11,7 @@ class TestCopySpeed(unittest.TestCase):
  def testCopySHMtoDefault(self):
    s = shared_memory.SharedMemory(name="test_X", create=True, size=N*N*4)
    s.close()
-    if CI and not OSX:
+    if os.path.exists("/dev/shm"):
      t = Tensor.empty(N, N, device="disk:/dev/shm/test_X").realize()
    else:
      t = Tensor.empty(N, N, device="disk:shm:test_X").realize()
@@ -77,11 +76,8 @@ class TestCopySpeed(unittest.TestCase):
        Device[Device.DEFAULT].synchronize()
      np.testing.assert_equal(t.numpy(), x.numpy())

-  @unittest.skipIf(CI, "CI doesn't have 6 GPUs")
-  @unittest.skipIf(Device.DEFAULT != "CL", "only test this on CL")
+  @unittest.skipIf(Device.DEFAULT != "CL" or Device[Device.DEFAULT].count() != 6, "only test this on CL, with 6 gpus")
  def testCopyCPUto6GPUs(self):
-    from tinygrad.runtime.ops_cl import CLDevice
-    if len(CLDevice.device_ids) != 6: raise unittest.SkipTest("computer doesn't have 6 GPUs")
    t = Tensor.ones(N, N, device="CPU").contiguous().realize()
    print(f"buffer: {t.nbytes()*1e-9:.2f} GB")
    for _ in range(3):
@@ -1,9 +1,9 @@
 import unittest
 from tinygrad import Tensor, Device, dtypes
-from test.helpers import CI
+from tinygrad.helpers import DEV
 # similar to test/external/external_test_gpu_ast.py, but universal

-@unittest.skipIf(Device.DEFAULT in {"CUDA", "NV"} and CI, "slow on CUDA CI")
+@unittest.skipIf(Device.DEFAULT in {"CUDA", "NV"} and DEV.interface.startswith("MOCK"), "slow on ocelot")
 class TestSpecific(unittest.TestCase):
  # from openpilot

@@ -9,11 +9,11 @@ import torch
 torch.set_num_threads(1)
 import time
 import numpy as np
+import sys
 np.set_printoptions(linewidth=160)
 from tinygrad import Tensor, Device, GlobalCounters, TinyJit
 from tinygrad.nn import Conv2d
 from tinygrad.helpers import colorize_float, getenv, DEV
-from test.helpers import CI

 IN_CHANS = [int(x) for x in getenv("IN_CHANS", "4,16,64").split(",")]

@@ -96,7 +96,7 @@ def helper_test_generic(name, f1, f1_args, f2, f2_args):
  desc = "faster" if et_torch > et_tinygrad else "slower"
  flops = save_ops*1e-6
  mem = save_mem*1e-6
-  print(("\r" if not CI else "")+f"{name:42s} {et_torch:7.2f} ms ({flops/et_torch:9.2f} GFLOPS {mem/et_torch:7.2f} GB/s) in torch, {et_tinygrad:7.2f} ms ({flops/et_tinygrad:9.2f} GFLOPS {mem/et_tinygrad:7.2f} GB/s) in tinygrad, {colorize_float(et_tinygrad/et_torch)} {desc} {flops:10.2f} MOPS {mem:8.2f} MB")  # noqa: E501
+  print(("\r" if sys.stdout.isatty() else "")+f"{name:42s} {et_torch:7.2f} ms ({flops/et_torch:9.2f} GFLOPS {mem/et_torch:7.2f} GB/s) in torch, {et_tinygrad:7.2f} ms ({flops/et_tinygrad:9.2f} GFLOPS {mem/et_tinygrad:7.2f} GB/s) in tinygrad, {colorize_float(et_tinygrad/et_torch)} {desc} {flops:10.2f} MOPS {mem:8.2f} MB")  # noqa: E501
  atol, rtol = (1e-2, 1e-2) if torch_dt == torch.float16 else (1e-3, 1e-3)
  np.testing.assert_allclose(val_tinygrad, val_torch, atol=atol, rtol=rtol)

@@ -2,7 +2,6 @@
 import unittest, random
 from tinygrad import Tensor, Context, Variable, TinyJit, dtypes, Device, nn
 from tinygrad.helpers import getenv
-from test.helpers import CI

 class TestTiny(unittest.TestCase):

@@ -112,7 +111,7 @@ class TestTiny(unittest.TestCase):
  # *** a model ***

  # TODO: this is failing because of how swizzling rewrites the ShapeTracker of the final STORE
-  @unittest.skipIf(CI and Device.DEFAULT == "DSP", "failing because of make things that can't be images not images")
+  @unittest.skipIf(Device.DEFAULT == "DSP", "failing because of make things that can't be images not images")
  def test_mnist(self):
    layers = [
      nn.Conv2d(1, 32, 5), Tensor.relu,
@@ -131,7 +130,7 @@ class TestTiny(unittest.TestCase):
    self.assertEqual(len(probs[0]), 10)

  # TODO: this is failing because of how swizzling rewrites the ShapeTracker of the final STORE
-  @unittest.skipIf(CI and Device.DEFAULT == "DSP", "failing because of make things that can't be images not images")
+  @unittest.skipIf(Device.DEFAULT == "DSP", "failing because of make things that can't be images not images")
  def test_mnist_backward(self):
    # NOTE: we don't have the whole model here for speed
    layers = [
@@ -2,12 +2,12 @@ import unittest, time
 from unittest.case import skipIf

 from extra.bench_log import BenchEvent, InstantBenchEvent, WallTimeEvent, KernelTimeEvent, log_event_instant, _events, clear_events
-from tinygrad.helpers import Context
+from tinygrad.helpers import Context, DEV
 from tinygrad.tensor import Tensor
 from tinygrad.device import Device
-from test.helpers import CI

-_SKIP_KERNEL_TIMING = Device.DEFAULT == "WEBGPU"  # WEBGPU kernel timing not supported
+# WEBGPU kernel timing not supported, ocelot CUDA is inaccurate
+_SKIP_KERNEL_TIMING = Device.DEFAULT == "WEBGPU" or (Device.DEFAULT == "CUDA" and DEV.interface.startswith("MOCK"))

 class TestBenchLog(unittest.TestCase):
  def setUp(self):
@@ -38,7 +38,7 @@ class TestBenchLog(unittest.TestCase):
      self.assertGreater(_events[event]["wall"][0], 0)
      self.assertGreater(_events[event]["wall"][1], 0)

-  @skipIf(CI or _SKIP_KERNEL_TIMING, "ci timing is not accurate")
+  @skipIf(_SKIP_KERNEL_TIMING, "ci timing is not accurate")
  def test_log_single_kernel_time(self):
    wall_times = []

@@ -55,7 +55,7 @@ class TestBenchLog(unittest.TestCase):
      self.assertLess(_events[event]["kernel"][0], wall_times[0])
      self.assertGreater(_events[event]["kernel"][0], 0)

-  @skipIf((CI and Device.DEFAULT == "CUDA") or _SKIP_KERNEL_TIMING, "ci cuda timing is not accurate")
+  @skipIf(_SKIP_KERNEL_TIMING, "ci cuda timing is not accurate")
  def test_interleaved_wall_kernel_time(self):
    wall_times = []
    with Context(DEBUG=2):
@@ -77,7 +77,7 @@ class TestBenchLog(unittest.TestCase):
      self.assertLess(_events[event]["kernel"][0], wall_times[0])
      self.assertGreater(_events[event]["kernel"][0], 0)

-  @skipIf((CI and Device.DEFAULT == "CUDA") or _SKIP_KERNEL_TIMING, "ci cuda timing is not accurate")
+  @skipIf(_SKIP_KERNEL_TIMING, "ci cuda timing is not accurate")
  def test_stacked_wall_kernel_time(self):
    with Context(DEBUG=2):
      for event in BenchEvent:
@@ -4,7 +4,6 @@ import numpy as np
 from tinygrad import dtypes, Tensor, TinyJit, GlobalCounters, Variable
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.helpers import temp, DEV, Context
-from test.helpers import CI

 N = 200  # has to be bigger than the cache to fail

@@ -140,7 +139,7 @@ class TestAssign(unittest.TestCase):
  def test_assign_changes_realized_alt(self): return self.test_assign_changes_alt(realize=True)

  def test_assign_changes_buffer_alt(self):
-    a, b = [Tensor(Tensor(0).contiguous().realize().uop.buf_uop) for _ in range(2)]
+    a, b = [Tensor(Tensor([0]).realize().uop.buf_uop) for _ in range(2)]
    Tensor.realize(a.contiguous().assign(1), b.contiguous().assign(2))
    self.assertEqual((a + b).item(), 3)

@@ -189,7 +188,7 @@ class TestAssign(unittest.TestCase):
    new = a + times_a
    np.testing.assert_allclose(new.numpy(), 8)

-  @unittest.skipIf(CI and DEV.renderer == "LVP", "flaky in CI")
+  @unittest.skipIf(DEV.renderer == "LVP", "flaky in CI")
  def test_double_assign(self):
    a = Tensor.ones(4).contiguous().realize()
    a += 1
@@ -282,15 +281,15 @@ class TestAssign(unittest.TestCase):
    np.testing.assert_equal(t.numpy(), [[100, 104, 108, 112], [101, 105, 109, 113], [102, 106, 110, 114], [103, 107, 111, 115]])

  def test_assign_contiguous(self):
-    b = Tensor.arange(16).reshape(4,4).contiguous().realize()
-    a = (Tensor.arange(16).reshape(4,4).contiguous().realize() + 1)
+    b = Tensor.arange(16).reshape(4,4).clone().realize()
+    a = (Tensor.arange(16).reshape(4,4).clone().realize() + 1)
    GlobalCounters.reset()
    b.assign(a.contiguous()).realize()
    self.assertEqual(GlobalCounters.kernel_count, 2)

  def test_assign_contiguous_permute(self):
-    b = Tensor.arange(16).reshape(4,4).contiguous().realize()
-    a = (Tensor.arange(16).reshape(4,4).contiguous().realize() + 1).permute((1,0))
+    b = Tensor.arange(16).reshape(4,4).clone().realize()
+    a = (Tensor.arange(16).reshape(4,4).clone().realize() + 1).permute((1,0))
    GlobalCounters.reset()
    b.assign(a.contiguous()).realize()
    self.assertEqual(GlobalCounters.kernel_count, 2)
@@ -326,29 +325,29 @@ class TestAssign(unittest.TestCase):
    np.testing.assert_allclose(a.numpy(), np.arange(N*N).reshape((N,N)) + np.arange(N*N).reshape((N,N)).transpose(1,0))

  def test_post_permuted_assignment_alt(self):
-    a = Tensor.arange(N*N).reshape(N,N).contiguous().realize()
-    b = Tensor.arange(N*N).reshape(N,N).contiguous().realize()
+    a = Tensor.arange(N*N).reshape(N,N).clone().realize()
+    b = Tensor.arange(N*N).reshape(N,N).clone().realize()
    new_a = (a.T+b).numpy()
    a.assign(a.T+b)
    np.testing.assert_allclose(a.numpy(), new_a)

  def test_post_flipped_assignment(self):
-    a = Tensor.arange(N*N).reshape(N,N).contiguous().realize()
-    b = Tensor.arange(N*N).reshape(N,N).contiguous().realize()
+    a = Tensor.arange(N*N).reshape(N,N).clone().realize()
+    b = Tensor.arange(N*N).reshape(N,N).clone().realize()
    new_a = (a.flip(0)+b).numpy()
    a.assign(a.flip(0)+b)
    np.testing.assert_allclose(a.numpy(), new_a)

  def test_post_flipped_assignment_axis1(self):
-    a = Tensor.arange(N*N).reshape(N,N).contiguous().realize()
-    b = Tensor.arange(N*N).reshape(N,N).contiguous().realize()
+    a = Tensor.arange(N*N).reshape(N,N).clone().realize()
+    b = Tensor.arange(N*N).reshape(N,N).clone().realize()
    new_a = (a.flip(1)+b).numpy()
    a.assign(a.flip(1)+b)
    np.testing.assert_allclose(a.numpy(), new_a)

  def test_post_reshape_assignment_fine(self):
-    a = Tensor.arange(N*N).reshape(N, N).contiguous().realize()
-    b = Tensor.arange(N*N).reshape(N, N).contiguous().realize()
+    a = Tensor.arange(N*N).reshape(N, N).clone().realize()
+    b = Tensor.arange(N*N).reshape(N, N).clone().realize()
    rhs = a.reshape(-1).reshape(N, N)
    new_a = (rhs+b).numpy()
    a.assign(rhs+b)  # self-assign with reshape view is fine
@@ -356,7 +355,7 @@ class TestAssign(unittest.TestCase):

  @unittest.skip("multi output not supported anymore")
  def test_simple_assignment_multioutput(self):
-    a = Tensor.arange(32*32).reshape(32, 32).contiguous().realize()
+    a = Tensor.arange(32*32).reshape(32, 32).clone().realize()
    b = Tensor.full((32, ), 1.).contiguous().realize()
    c = Tensor.full((32, ), 2.).contiguous().realize()
    d = Tensor.full((32, ), 3.).contiguous().realize()
@@ -376,15 +375,15 @@ class TestAssign(unittest.TestCase):
  # NOTE: if the assign target is read/write in a single kernel, it should be contiguous

  def test_permuted_assignment_correct(self):
-    a = Tensor.arange(4 * 4).reshape(4, 4).contiguous().realize()
-    b = Tensor.arange(4 * 4).reshape(4, 4).contiguous().realize()
+    a = Tensor.arange(4 * 4).reshape(4, 4).clone().realize()
+    b = Tensor.arange(4 * 4).reshape(4, 4).clone().realize()
    a = a.permute(1, 0)
    new_val = a + b
    a.assign(new_val)
    np.testing.assert_equal(a.numpy(), np.arange(4 * 4).reshape(4, 4).transpose(1, 0) + np.arange(4 * 4).reshape(4, 4))

  def test_permuted_reduceop_child_dual_use(self):
-    a = Tensor.arange(32*32*32).reshape(32, 32, 32).contiguous().realize()
+    a = Tensor.arange(32*32*32).reshape(32, 32, 32).clone().realize()
    b = Tensor.ones(32, 32, dtype=dtypes.int).contiguous().realize()
    r = a.sum(axis=1)
    b.assign(r + b.permute(1, 0))
@@ -393,7 +392,7 @@ class TestAssign(unittest.TestCase):

  @unittest.skip("multi output not supported anymore")
  def test_permuted_reduceop_multioutput_dual_use(self):
-    a = Tensor.arange(32*32*32).reshape(32, 32, 32).contiguous().realize()
+    a = Tensor.arange(32*32*32).reshape(32, 32, 32).clone().realize()
    b = Tensor.full((32, 32), 1.).contiguous().realize()
    c = Tensor.full((32, 32), 2.).contiguous().realize()

@@ -406,9 +405,9 @@ class TestAssign(unittest.TestCase):

  @unittest.skip("multi output not supported anymore")
  def test_permuted_reduceop_multioutput_dual_use_possible(self):
-    a = Tensor.arange(32*32*32).reshape(32, 32, 32).contiguous().realize()
-    b = Tensor.arange(32 * 32).reshape(32, 32).realize()
-    c = Tensor.arange(32 * 32).reshape(32, 32).realize()
+    a = Tensor.arange(32*32*32).reshape(32, 32, 32).clone().realize()
+    b = Tensor.arange(32 * 32).reshape(32, 32).clone().realize()
+    c = Tensor.arange(32 * 32).reshape(32, 32).clone().realize()

    GlobalCounters.reset()
    r = a.sum(axis=1)
@@ -442,7 +441,7 @@ class TestAssign(unittest.TestCase):
    # Forward shift: read index > write index in overlap
    N = 100000
    shift = 1000
-    a = Tensor.arange(N).float().contiguous().realize()
+    a = Tensor.arange(N).float().clone().realize()
    expected = np.arange(N, dtype=np.float32)
    expected[:N-shift] = expected[shift:].copy()
    with Context(NOOPT=1): a[0:N-shift].assign(a[shift:N]).realize()
@@ -452,7 +451,7 @@ class TestAssign(unittest.TestCase):
    # Reverse shift: write index > read index in overlap
    N = 100000
    shift = 1000
-    a = Tensor.arange(N).float().contiguous().realize()
+    a = Tensor.arange(N).float().clone().realize()
    expected = np.arange(N, dtype=np.float32)
    expected[shift:] = expected[:N-shift].copy()
    with Context(NOOPT=1): a[shift:N].assign(a[0:N-shift]).realize()
@@ -460,7 +459,7 @@ class TestAssign(unittest.TestCase):

  def test_nonoverlapping_shrink_assignment(self):
    # TODO: non-overlapping shrinks don't actually need contiguous, could be 1 kernel with smarter range analysis
-    a = Tensor.arange(100).float().contiguous().realize()
+    a = Tensor.arange(100).float().clone().realize()
    expected = np.arange(100, dtype=np.float32)
    expected[0:10] = expected[50:60].copy()
    GlobalCounters.reset()
@@ -495,6 +495,25 @@ class TestFunctionTuple(unittest.TestCase):
    Tensor.realize(a.grad)
    np.testing.assert_allclose(a.grad.numpy(), [2., 2., 2., 2.])

+  def test_custom_kernel_precompile_multidevice(self):
+    # a custom_kernel output placeholder (invalids) under multi-device @function(precompile=True) must return the
+    # kernel's computed result. read it back through .numpy() so the cross-device gather reads the output buffer
+    devs = ("CPU:0", "CPU:1")
+    def double_kernel(C:UOp, A:UOp) -> UOp:
+      C, A = C.flatten(), A.flatten()
+      i = UOp.range(A.numel(), 0)
+      return C[i].store(A[i] * 2.0).end(i).sink(arg=KernelInfo(name="double_kernel"))
+    def double_grad(d_c:UOp, call:UOp): return (None, (Tensor(d_c) * 2.0).uop)
+
+    @function(precompile=True, precompile_backward=True)
+    def f(a:Tensor):
+      c = Tensor(Tensor.invalids(a.shape[0]//len(devs), a.shape[1], dtype=a.dtype, device=devs).uop.multi(0), device=devs)
+      return Tensor.custom_kernel(c, a, fxn=double_kernel, grad_fxn=double_grad)[0]
+
+    a = Tensor.full((4, 4), 7.0).contiguous().shard(devs, axis=0)
+    Tensor.realize(a)
+    np.testing.assert_allclose(f(a).numpy(), 14.0)
+
  def test_custom_kernel_precompile_further_compute(self):
    def my_kernel(C:UOp, A:UOp) -> UOp:
      i = UOp.range(A.shape[0], 0)
@@ -166,8 +166,8 @@ class TestMultiOutputGradient(unittest.TestCase):
    c, d, e, _, _ = Tensor.custom_kernel(Tensor.empty(4, 4), Tensor.empty(4, 4), Tensor.empty(4, 4), a, b,
                                          fxn=addmulsub_kernel, grad_fxn=backward_addmulsub)
    (c.sum() + d.sum() + e.sum()).backward()
-    np.testing.assert_allclose(a.grad.numpy(), a_ref.grad.numpy(), rtol=1e-5)
-    np.testing.assert_allclose(b.grad.numpy(), b_ref.grad.numpy(), rtol=1e-5)
+    np.testing.assert_allclose(a.grad.numpy(), a_ref.grad.numpy(), atol=1e-6, rtol=1e-5)
+    np.testing.assert_allclose(b.grad.numpy(), b_ref.grad.numpy(), atol=1e-6, rtol=1e-5)

 class TestViewGradient(unittest.TestCase):
  def test_expand(self):
@@ -1,7 +1,7 @@
 import unittest
 from unittest.mock import MagicMock
 from tinygrad import Device
-from tinygrad.uop.ops import Ops
+from tinygrad.uop.ops import Ops, UOp
 from tinygrad.dtype import dtypes

@unittest.skipUnless(Device.DEFAULT == "METAL", "Metal device required to run")
@@ -14,8 +14,10 @@ class TestMetalGraph(unittest.TestCase):
  def metal_buf(self, offset):
    buf = MagicMock()
    if offset > 0:
-      buf.op = Ops.BUFFER_VIEW
-      buf.arg = (None, offset)
+      buf.op = Ops.SLICE
+      src = MagicMock()
+      src.dtype = dtypes.uint8
+      buf.src = (src, UOp.const(dtypes.weakint, offset))
      buf.dtype = dtypes.uint8
    else:
      buf.op = Ops.BUFFER
@@ -34,7 +36,7 @@ class TestMetalGraph(unittest.TestCase):
    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(0x100000000))) is False

  def test_supports_uop_nonmetal_buf(self):
-    # non-BUFFER_VIEW ops should not be checked for offset
+    # non-SLICE ops should not be checked for offset
    buf = MagicMock()
    buf.op = Ops.BUFFER
    buf.device = Device.DEFAULT
@@ -0,0 +1,194 @@
+import unittest, math, torch
+import numpy as np
+from functools import partial
+from tinygrad import nn, dtypes, Tensor, Device, TinyJit, Variable
+from tinygrad.helpers import OSX
+
+# https://gist.github.com/devries/11405101
+def ksprob(a):
+  fac, total, termbf = 2.0, 0.0, 0.0
+  a2 = -2.0 * a * a
+  for j in range(1, 101):
+    term = fac * math.exp(a2 * j * j)
+    total += term
+    if math.fabs(term) <= 0.001 * termbf or math.fabs(term) <= 1e-8 * total:
+      return total
+    fac = -fac
+    termbf = math.fabs(term)
+  return 1.0
+
+def kstest(l1, l2):
+  n1, n2 = len(l1), len(l2)
+  l1.sort()
+  l2.sort()
+  j1, j2, d, fn1, fn2 = 0, 0, 0.0, 0.0, 0.0
+  while j1 < n1 and j2 < n2:
+    d1, d2 = l1[j1], l2[j2]
+    if d1 <= d2:
+      fn1 = (float(j1) + 1.0) / float(n1)
+      j1 += 1
+    if d2 <= d1:
+      fn2 = (float(j2) + 1.0) / float(n2)
+      j2 += 1
+    dtemp = math.fabs(fn2 - fn1)
+    if dtemp > d:
+      d = dtemp
+  ne = float(n1 * n2) / float(n1 + n2)
+  nesq = math.sqrt(ne)
+  prob = ksprob((nesq + 0.12 + 0.11 / nesq) * d)
+  return prob
+
+def equal_distribution(tiny_func, torch_func=None, numpy_func=None, shape=(40, 43), alpha=0.04):
+  Tensor.manual_seed(1337)
+  torch.manual_seed(1337)
+  np.random.seed(1337)
+  assert not (torch_func is None and numpy_func is None), "no function to compare with"
+  x1 = tiny_func(*shape).numpy().flatten()
+  x2 = tiny_func(shape).numpy().flatten()
+  if numpy_func is not None: y = numpy_func(shape).flatten()
+  if torch_func is not None: z = torch_func(shape).numpy().flatten()
+  return (numpy_func is None or (kstest(x1, y) >= alpha and kstest(x2, y) >= alpha)) and \
+    (torch_func is None or (kstest(x1, z) >= alpha and kstest(x2, z) >= alpha))
+
+def normal_test(func, shape=(20, 45), alpha=0.05): return equal_distribution(func, numpy_func=lambda x: np.random.randn(*x), shape=shape, alpha=alpha)
+
+class TestRandomness(unittest.TestCase):
+  def test_randn(self):
+    self.assertEqual(Tensor.randn(3,3,dtype=dtypes.half).dtype, dtypes.half)
+    self.assertTrue(normal_test(Tensor.randn))
+    self.assertTrue(equal_distribution(Tensor.randn, torch.randn, lambda x: np.random.randn(*x)))
+
+  def test_randint(self):
+    self.assertFalse(normal_test(Tensor.randint))
+    self.assertTrue(equal_distribution(partial(Tensor.randint, low=-2, high=5),
+                                       numpy_func=lambda x: np.random.randint(low=-2, high=5, size=x)))
+    self.assertTrue(equal_distribution(partial(Tensor.randint, low=-2, high=5, dtype="int32"),
+                                       numpy_func=lambda x: np.random.randint(low=-2, high=5, size=x)))
+    self.assertTrue(Tensor.randint(1, device="CPU").device=="CPU")
+    # check types of args
+    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0.1, high=3)
+    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0, high=3.5)
+    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=1, high=3, dtype="float")
+    with self.assertRaises(TypeError): Tensor.randint((3, 4), low=0, high=3, dtype=dtypes.float32)
+    # check low < high
+    with self.assertRaises(ValueError): Tensor.randint((3, 4), low=10, high=5)
+    with self.assertRaises(ValueError): Tensor.randint((3, 4), low=10, high=10)
+    np.testing.assert_array_equal(Tensor.randint(16, low=5, high=6).numpy(), 5)
+
+  def test_normal(self):
+    self.assertTrue(normal_test(Tensor.normal))
+    self.assertTrue(equal_distribution(Tensor.normal, lambda x: torch.nn.init.normal_(torch.empty(x), mean=0, std=1),
+                                                      lambda x: np.random.normal(loc=0, scale=1, size=x)))
+    # check std >= 0
+    with self.assertRaises(ValueError): Tensor.normal((3, 4), mean=0, std=-1)
+
+  def test_uniform(self):
+    self.assertFalse(normal_test(Tensor.uniform))
+    self.assertTrue(equal_distribution(Tensor.uniform, lambda x: torch.nn.init.uniform_(torch.empty(x)), lambda x: np.random.uniform(size=x)))
+    self.assertTrue(equal_distribution(partial(Tensor.uniform, low=-100, high=100, dtype=dtypes.int32),
+                                       numpy_func=lambda x: np.random.randint(low=-100, high=100, size=x)))
+    # check low < high
+    with self.assertRaises(ValueError): Tensor.uniform((3, 4), low=5.0, high=3.0)
+    with self.assertRaises(ValueError): Tensor.uniform((3, 4), low=1.0, high=1.0)
+
+  def test_scaled_uniform(self):
+    self.assertFalse(normal_test(Tensor.scaled_uniform))
+    self.assertTrue(equal_distribution(Tensor.scaled_uniform, lambda x: torch.nn.init.uniform_(torch.empty(x), a=-1, b=1) / math.sqrt(math.prod(x)),
+                                                              lambda x: np.random.uniform(-1, 1, size=x) / math.sqrt(math.prod(x))))
+
+  def test_glorot_uniform(self):
+    self.assertFalse(normal_test(Tensor.glorot_uniform))
+    self.assertTrue(equal_distribution(Tensor.glorot_uniform, lambda x: torch.nn.init.xavier_uniform_(torch.empty(x)),
+                                                              lambda x: np.random.uniform(-1, 1, size=x) * math.sqrt(6 / (x[0] + math.prod(x[1:])))))
+
+  def test_kaiming_uniform(self):
+    for shape in [(32, 16, 3, 3), (20, 44), (5, 15, 35)]:
+      self.assertTrue(equal_distribution(Tensor.kaiming_uniform, lambda x: torch.nn.init.kaiming_uniform_(torch.empty(x)), shape=shape))
+
+  def test_kaiming_normal(self):
+    for shape in [(32, 16, 3, 3), (20, 44), (3, 15, 35)]:
+      self.assertTrue(equal_distribution(Tensor.kaiming_normal, lambda x: torch.nn.init.kaiming_normal_(torch.empty(x)), shape=shape))
+
+  def test_multinomial(self):
+    self.assertRaises(AssertionError, lambda: Tensor(2).multinomial(1, replacement=False))
+    self.assertRaises(AssertionError, lambda: Tensor([1, 9]).multinomial(0, replacement=False))
+    def _check_with_torch(w, num_samples, replacement):
+      tiny_res = Tensor(w).multinomial(num_samples, replacement=replacement)
+      torch_res = torch.tensor(w).multinomial(num_samples, replacement=replacement)
+      self.assertEqual(tiny_res.shape, torch_res.shape)
+      if torch_res.ndim == 1:
+        tiny_res = tiny_res.unsqueeze(0)
+        torch_res = torch_res.unsqueeze(0)
+      for i in range(torch_res.shape[0]):
+        self.assertTrue(equal_distribution(lambda *_: tiny_res[i], lambda _: torch_res[i]))
+    _check_with_torch(w=[0.231, 0., 1., 0.5], num_samples=300, replacement=True)
+    _check_with_torch(w=[[0.2, 0.8]], num_samples=300, replacement=True)  # 2D but only 1 row
+    _check_with_torch(w=[[0.453, 0., 1., 0.81], [0.1, 0.8, 0., 0.1]], num_samples=300, replacement=True)
+    # no-replacement
+    w = [0.1, 0.9]
+    self.assertRaises(AssertionError, lambda: Tensor(w).multinomial(100, replacement=False))
+
+    @TinyJit
+    def sample_one(): return Tensor(w).multinomial(1, replacement=False).realize()
+
+    tiny_samples = [sample_one().item() for _ in range(400)]
+    torch_samples = [torch.tensor(w).multinomial(1, replacement=False).item() for _ in range(400)]
+    self.assertTrue(equal_distribution(lambda *_: Tensor(tiny_samples), lambda _: torch.tensor(torch_samples)))
+
+    w = list(range(32))
+    s1 = Tensor(w).multinomial(5, replacement=False).numpy()
+    self.assertEqual(len(set(s1.tolist())), 5)
+    s2 = Tensor(w).multinomial(5, replacement=False).numpy()
+    self.assertFalse(np.array_equal(s1, s2))
+    full = Tensor(w).multinomial(len(w), replacement=False).numpy()
+    self.assertEqual(sorted(full.tolist()), w)
+
+    w = [0.1, 0.2, 0.3, 0.4]
+    @TinyJit
+    def sample_three(): return Tensor(w).multinomial(3, replacement=False).realize()
+
+    tiny_draws = np.array([sample_three().numpy() for _ in range(400)])
+    torch_draws = np.array([torch.tensor(w).multinomial(3, replacement=False).numpy() for _ in range(400)])
+    for pos in range(3):
+      self.assertTrue(equal_distribution(lambda *_: Tensor(tiny_draws[:, pos]), lambda _: torch.tensor(torch_draws[:, pos])))
+
+  @unittest.skip("this test is flaky")
+  def test_multinomial_counterexample(self):
+    tiny_res = Tensor([0.3, 0.6, 0.1]).multinomial(4000, replacement=True)
+    torch_res = torch.tensor([0.3, 0.6, 0.1]).multinomial(4000, replacement=True)
+    self.assertTrue(equal_distribution(lambda *_: tiny_res, lambda _: torch_res))
+    torch_res = torch.tensor([0.2, 0.7, 0.1]).multinomial(4000, replacement=True)
+    self.assertFalse(equal_distribution(lambda *_: tiny_res, lambda _: torch_res))
+
+  def test_conv2d_init(self):
+    params = (128, 256, (3,3))
+    assert equal_distribution(lambda *_: nn.Conv2d(*params).weight, lambda _: torch.nn.Conv2d(*params).weight.detach())
+    assert equal_distribution(lambda *_: nn.Conv2d(*params).bias, lambda _: torch.nn.Conv2d(*params).bias.detach())
+
+  def test_linear_init(self):
+    params = (64, 256)
+    assert equal_distribution(lambda *_: nn.Linear(*params).weight, lambda _: torch.nn.Linear(*params).weight.detach())
+    assert equal_distribution(lambda *_: nn.Linear(*params).bias, lambda _: torch.nn.Linear(*params).bias.detach())
+
+  def test_bn_init(self):
+    params = (64,)
+    assert equal_distribution(lambda *_: nn.BatchNorm2d(*params).weight, lambda _: torch.nn.BatchNorm2d(*params).weight.detach())
+    assert equal_distribution(lambda *_: nn.BatchNorm2d(*params).bias, lambda _: torch.nn.BatchNorm2d(*params).bias.detach())
+
+# TODO: still fails with MAX_KERNEL_BUFFERS
+@unittest.skipIf(Device.DEFAULT == "WEBGPU" and not OSX, "WEBGPU Vulkan can only run kernels with up to 10 buffers")
+class TestSample(unittest.TestCase):
+  def test_sample(self):
+    X = Tensor.rand(1000, 50).realize()
+    BS = 16
+    idxs = np.random.randint(0, X.shape[0], size=(BS))
+    # this uncovered a bug with arg sort order
+    batch = [Variable(f'idx{i}', 0, X.shape[0]-1).bind(s) for i,s in enumerate(idxs.tolist())]
+    x = Tensor.cat(*[X.shrink(((batch[i], batch[i]+1), None)) for i in range(BS)])
+    print(idxs)
+    ret = x.numpy()
+    base = X.numpy()[idxs]
+    np.testing.assert_equal(ret, base)
+
+if __name__ == "__main__":
+  unittest.main()
@@ -15,6 +15,10 @@ class TestRealizeIsRealized(unittest.TestCase):
    t = Tensor.zeros(10).contiguous().realize()
    assert t.uop.is_realized

+  def test_ones(self):
+    t = Tensor.ones(4, 4).realize()
+    assert t.uop.is_realized
+
  def test_bytes(self):
    t = Tensor(b'\x01\x02\x03').realize()
    assert t.uop.is_realized
@@ -51,10 +55,6 @@ class TestRealizeIsRealized(unittest.TestCase):
    t = Tensor(3.14).realize()
    assert not t.uop.is_realized

-  def test_ones_not_realized(self):
-    t = Tensor.ones(4, 4).realize()
-    assert not t.uop.is_realized
-
  def test_none_not_realized(self):
    t = Tensor(None).realize()
    assert not t.uop.is_realized
@@ -2,11 +2,10 @@ import unittest
 import multiprocessing.shared_memory as shared_memory
 from tinygrad.helpers import WIN
 from tinygrad import Tensor, Device
-from test.helpers import CI
 import numpy as np

 class TestRawShmBuffer(unittest.TestCase):
-  @unittest.skipIf(WIN and CI, "only fails on CI windows instance")
+  @unittest.skipIf(WIN, "only fails on CI windows instance")
  def test_e2e(self):
    t = Tensor.randn(2, 2, 2).realize()

@@ -65,6 +65,10 @@ class TestTensorData(unittest.TestCase):
    assert dat.tolist() == 3
    assert dat.shape == ()

+  def test_const_dtype_for_uop(self):
+    self.assertEqual(Tensor.const(dtypes.int8, UOp.const(dtypes.float32, 1.0)).dtype, dtypes.int8)
+    self.assertEqual(Tensor.const(dtypes.int32, UOp.variable("x", 1, 10).bind(5)).item(), 5)
+
  def test_data_float32(self):
    a = Tensor([[1,2.5],[3,4]], dtype=dtypes.float32)
    dat = a.data()
@@ -1,4 +1,5 @@
 from dataclasses import dataclass, field
+from tinygrad.dtype import dtypes, AddrSpace, PtrDType, ImageDType
 from tinygrad.uop.ops import UOp, UPat, PatternMatcher, Ops, GroupOp, graph_rewrite, track_rewrites
 from tinygrad.helpers import VIZ, pluralize, all_int

@@ -56,17 +57,21 @@ def replace_store_after_with_contig(u:UOp, src:UOp):
  if assigned_to.op is not Ops.BUFFER: return src.contiguous(tag=u.tag)

 def _make_buffer_view(src:UOp) -> UOp|None:
-  """If movement ops on src collapse to a contiguous range, return BUFFER_VIEW.reshape(src.shape). Otherwise None."""
+  """If movement ops on src collapse to a contiguous range, return SLICE.reshape(src.shape). Otherwise None."""
  if (offset := src.contiguous_view_offset()) is None: return None
  buf = src.base
-  if buf.op is Ops.BUFFER_VIEW: offset, buf = offset + buf.arg[1], buf.src[0]
-  return UOp(Ops.BUFFER_VIEW, src.dtype, (buf,), (src.numel(), offset)).reshape(src.shape)
+  if buf.op is Ops.SLICE:
+    byte_offset = buf.src[1].arg * buf.src[0].dtype.itemsize + offset * src.dtype.itemsize
+    buf = buf.src[0]
+    if byte_offset % buf.dtype.itemsize != 0: return None
+    offset = byte_offset // buf.dtype.itemsize
+  return UOp(Ops.SLICE, src.dtype, (buf, UOp.const(dtypes.weakint, offset)), src.numel()).reshape(src.shape)

 def contiguous_mops_to_view(c:UOp, src:UOp):
-  """CONTIGUOUS(MOPS(BUFFER)) → CONTIGUOUS(BUFFER_VIEW) when movement ops collapse to a contiguous range."""
+  """CONTIGUOUS(MOPS(BUFFER)) → CONTIGUOUS(SLICE) when movement ops collapse to a contiguous range."""
  buf = src.base
-  if buf.op not in {Ops.BUFFER, Ops.BUFFER_VIEW}: return None
-  if src.op is Ops.RESHAPE and src.src[0].op in {Ops.BUFFER, Ops.BUFFER_VIEW}: return None
+  if buf.op not in {Ops.BUFFER, Ops.SLICE}: return None
+  if src.op is Ops.RESHAPE and src.src[0].op in {Ops.BUFFER, Ops.SLICE}: return None

  # no symbolic shape
  if not all_int(c.shape): return None
@@ -79,11 +84,11 @@ def contiguous_mops_to_view(c:UOp, src:UOp):

  x = src
  while x.op in GroupOp.Movement: x = x.src[0]
-  # NOTE: this contiguous is removed because this BUFFER_VIEW/RESHAPE has_buffer_identity
+  # NOTE: this contiguous is removed because this SLICE/RESHAPE has_buffer_identity
  if x.op is not Ops.MULTI and (view := _make_buffer_view(src)) is not None:
    return view.contiguous(tag=c.tag)

-  # for MULTI tensors, use multi_pm to resolve per-shard movement ops, then create BUFFER_VIEW on the resolved result
+  # for MULTI tensors, use multi_pm to resolve per-shard movement ops, then create SLICE on the resolved result
  if not isinstance(c.device, str):
    from tinygrad.schedule.multi import multi_pm
    resolved = graph_rewrite(src, multi_pm, name="multi_buffer_view")
@@ -137,7 +142,7 @@ pm_early_transform_tensor_graph = PatternMatcher([
  # resolve TUPLE+GETTUPLE (for precompiled calls)
  (UPat(Ops.GETTUPLE, src=(UPat(Ops.TUPLE, name="t"),), name="g"), lambda g,t: t.src[g.arg]),

-  # CONTIGUOUS(MOPS(BUFFER/BUFFER_VIEW)) → CONTIGUOUS(BUFFER_VIEW) when movement ops collapse to contiguous range
+  # CONTIGUOUS(MOPS(BUFFER/SLICE)) → CONTIGUOUS(SLICE) when movement ops collapse to contiguous range
  (UPat(Ops.CONTIGUOUS, src=(UPat(GroupOp.Movement, name="src"),), name="c"), contiguous_mops_to_view),

  # add CONTIGUOUS to tagged UOps
@@ -171,7 +176,8 @@ def finalize_after(ctx:AllocCtx, x:UOp):
 def replace_input_buffer(ctx:AllocCtx, b:UOp):
  ctx.replacements.append(b)
  return UOp.param(len(ctx.replacements)-1, b.dtype, b.shape, b.device,
-                   b._min_max if b.op is Ops.BIND else None, b.src[0].arg[0] if b.op is Ops.BIND else None)
+                   b._min_max if b.op is Ops.BIND else None, b.src[0].arg[0] if b.op is Ops.BIND else None,
+                   b.addrspace if isinstance(b.dtype, (PtrDType, ImageDType)) else AddrSpace.GLOBAL)

 pm_finalize_call = PatternMatcher([
  (UPat(Ops.AFTER, name="x"), finalize_after),
@@ -183,8 +189,8 @@ pm_finalize_call = PatternMatcher([
 pm_replace_buf = PatternMatcher([
  # replace BUFFER with PARAM for cache key normalization
  (UPat(Ops.BUFFER, src=(UPat(Ops.UNIQUE), UPat(Ops.DEVICE)), name="b"), replace_input_buffer),
-  # replace BUFFER_VIEW with PARAM. this rewrite is bottom up so BUFFERs we don't need won't be in the input
-  (UPat(Ops.BUFFER_VIEW, src=(UPat(Ops.BUFFER),), name="b"), replace_input_buffer),
+  # replace SLICE with PARAM. this rewrite is bottom up so BUFFERs we don't need won't be in the input
+  (UPat(Ops.SLICE, src=(UPat(Ops.BUFFER), UPat(Ops.CONST, dtype=dtypes.weakint)), name="b"), replace_input_buffer),
  # strip value from BIND for cache key normalization, so different values hit same cache
  (UPat(Ops.BIND, src=(UPat(Ops.DEFINE_VAR), UPat(Ops.CONST)), name="b"), replace_input_buffer),
 ])
@@ -105,6 +105,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
  sink = graph_rewrite(sink, pm_add_control_flow, ctx=CFGContext(sink), name="add control flow", bottom_up=True)

  if VIZ: graph_rewrite(sink, PatternMatcher([]), name="View Output AST")
+  if SPEC: type_verify(sink, spec_program)

  # return the rewritten sink
  return sink
@@ -115,7 +116,7 @@ pm_linearize_cleanups = PatternMatcher([
  (UPat((Ops.IF, Ops.ENDIF)), lambda: panic(RuntimeError, "if not allowed in graph")),
  # gated STORE becomes IF-STORE-ENDIF. this is the only use of IF-ENDIF
  (UPat(Ops.STORE, name="u", src=(UPat(Ops.INDEX).or_casted(), UPat(), UPat(name="gate", dtype=dtypes.bool))),
-   lambda u, gate: (u, [mif:=UOp(Ops.IF, src=(gate, u.src[0])), u, UOp(Ops.ENDIF, src=(mif,))]))
+   lambda u, gate: ((st:=u.replace(src=u.src[0:2])), [mif:=UOp(Ops.IF, src=(gate, u.src[0])), st, UOp(Ops.ENDIF, src=(mif,))]))
 ])

 # requires lst be toposorted. like graph rewrite, but for lines
@@ -132,7 +133,6 @@ def line_rewrite(lst:list[UOp], pm:PatternMatcher, ctx=None) -> list[UOp]:
 def do_linearize(ctx:Renderer, prg:UOp, sink:UOp) -> UOp:
  if DEBUG >= 3 and sink.arg.applied_opts: print(f"{sink.arg.function_name:<25} opts: {sink.arg.applied_opts}")
  lst = line_rewrite(linearize(sink), pm_linearize_cleanups)
-  if SPEC: type_verify(lst, spec_program)
  # isa renderers need to allocate registers
  if isinstance(ctx, ISARenderer):
    if ctx.pre_regalloc_matcher is not None: lst = line_rewrite(lst, ctx.pre_regalloc_matcher, PreRegAllocContext())
@@ -91,7 +91,7 @@ def add_gpudims(ctx:Renderer, s:UOp):
  subs = {}
  for r in s_topo:
    # look for local INDEXes that are not used in the GLOBAL store, then add them as an INVALID
-    if r.op is Ops.STORE and (idx := r.src[0]).src[0].ptrdtype.addrspace == AddrSpace.GLOBAL:
+    if r.op is Ops.STORE and (idx := r.src[0]).src[0].addrspace == AddrSpace.GLOBAL:
      missing_locals = [all_ranges[rng] for rng in local_dims if all_ranges[rng] not in idx.ranges]
      if len(missing_locals):
        assert len(idx.src) == 2, "index has 2 sources"
@@ -104,7 +104,7 @@ def fold_expanded_index(midx:UOp):
    for grp in grouped_offsets:
      # get the index offset for this element. using [0] is okay, because they are the same
      lidx = midx.src[offsets[grp[0]][0]]
-      if len(grp) > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(len(grp)).ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace))
+      if len(grp) > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(len(grp)).ptr(size=buf.max_numel(), addrspace=buf.addrspace))
      # set the idxs of the output
      for i,g in enumerate(grp):
        for oo in offsets[g]: idxs[oo] = global_offset+i
@@ -113,7 +113,7 @@ def fold_expanded_index(midx:UOp):
      global_offset += len(grp)
  assert None not in idxs, f"some idxs are missing {idxs}"
  # this base thing is for image, we want the CAT to be a normal pointer
-  post_cat = UOp(Ops.PTRCAT, buf.ptrdtype.base.ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace).vec(global_offset), tuple(ret))
+  post_cat = UOp(Ops.PTRCAT, buf.ptrdtype.base.ptr(size=buf.max_numel(), addrspace=buf.addrspace).vec(global_offset), tuple(ret))
  return post_cat.gep(tuple(cast(list[int], idxs)))

 def cat_after_store(cat:UOp, data:UOp):
@@ -165,7 +165,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
    must_divide = False
  elif buf.dtype.base not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not isinstance(buf.dtype, ImageDType):
    pass
-  elif buf.ptrdtype.addrspace == AddrSpace.REG:
+  elif buf.addrspace == AddrSpace.REG:
    pass
  elif isinstance(buf.dtype, ImageDType):
    lengths = [4]
@@ -186,7 +186,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
    for fold_length in lengths:
      if global_offset+fold_length > sz: continue
      lidx = buf.index((offset + global_offset).valid(mask), ptr=True)
-      if fold_length > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(fold_length).ptr(size=buf.ptrdtype.size, addrspace=buf.ptrdtype.addrspace))
+      if fold_length > 1: lidx = lidx.cast(buf.ptrdtype.base.vec(fold_length).ptr(size=buf.max_numel(), addrspace=buf.addrspace))
      if ls.op is Ops.STORE: ret.append(ls.replace(src=(lidx,ls.src[1].gep(tuple(range(global_offset, global_offset+fold_length))))))
      else: ret.append(ls.replace(src=(lidx,)+ls.src[1:], dtype=ls.dtype.scalar().vec(fold_length)))
      global_offset += fold_length
@@ -243,7 +243,9 @@ def no_vectorized_alu(alu:UOp):
  return UOp(Ops.STACK, alu.dtype, alus)

 def no_vectorized_buf(buf:UOp):
-  return buf.replace(dtype=buf.ptrdtype.base.scalar().ptr(buf.ptrdtype.size*buf.ptrdtype.count, buf.ptrdtype.addrspace)).cast(buf.dtype)
+  # TODO: this fails on regs
+  #assert buf.max_numel() == buf.ptrdtype.size
+  return buf.replace(dtype=buf.ptrdtype.base.scalar().ptr(buf.ptrdtype.size*buf.ptrdtype.count, buf.addrspace)).cast(buf.dtype)

 def no_vectorized_index(buf:UOp, cast:UOp, idx:UOp, bcast:UOp|None=None):
  cnt = cast.dtype.count
@@ -22,7 +22,7 @@ def linearize(sink:UOp) -> list[UOp]:
    extra = None
    match u.op:
      # the order and placement of these defines is important
-      case Ops.PARAM: priority, extra = -20, u.arg
+      case Ops.PARAM: priority, extra = -20, u.arg.slot
      case Ops.DEFINE_VAR: priority, extra = -19, u.arg
      case Ops.DEFINE_REG: priority = -18
      case Ops.DEFINE_LOCAL: priority = -17
@@ -93,4 +93,4 @@ def do_split_ends(e:UOp):
 pm_split_ends = PatternMatcher([
  # split the ends
  (UPat(Ops.END, name="e"), do_split_ends),
-])
+])
@@ -95,7 +95,7 @@ class Scheduler:
    if (old_sz:=rng.src[0].divides(amount)) is None:
      raise KernelOptError(f"{amount} can't divide {rng.src[0]} in {self.colored_shape()}")
    new_rng = UOp.range(amount, next(self.opt_range), new_type) if input_new_rng is None else input_new_rng
-    replaced_rng = rng.replace(src=(UOp.const(dtypes.int, old_sz),))
+    replaced_rng = rng.replace(src=(old_sz,))
    sub_axis = (new_rng * old_sz + replaced_rng) if top else (replaced_rng * amount + new_rng)
    self.ast = self.ast.substitute({rng:sub_axis}, name=f"shift {rng.arg[:-1]} {amount} {str(new_type).split('.')[1].lower()}")
    return replaced_rng, new_rng
@@ -329,8 +329,8 @@ class Scheduler:
  def group_for_reduces(self) -> int: return len(self.axes_of(AxisType.GROUP_REDUCE))

 def bufs_from_ast(ast:UOp, dname:str) -> list[Buffer]:
-  glbls = sorted([x for x in ast.backward_slice if x.op is Ops.PARAM], key=lambda x: x.arg)
-  return [Buffer(dname, x.ptrdtype.size, x.dtype.base) for x in glbls]
+  glbls = sorted([x for x in ast.backward_slice if x.op is Ops.PARAM], key=lambda x: x.arg.slot)
+  return [Buffer(dname, x.max_numel(), x.dtype.base) for x in glbls]

 def apply_opts(ast:UOp, ren:Renderer, beam:int=0) -> UOp:
  if ast.tag is not None: return ast
@@ -43,7 +43,7 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
    current_batch, current_batch_devs = [], []

  for si in linear.src:
-    if si.src[0].op is Ops.BUFFER_VIEW: continue
+    if si.src[0].op is Ops.SLICE: continue

    devs = dedup([Device[x] for b in si.src[1:] if b.op is not Ops.BIND for x in (b.device if isinstance(b.device, tuple) else (b.device,))])
    graph_t = graph_class(devs[0]) if devs[0].graph is not None else None
@@ -94,7 +94,7 @@ class GraphRunner:
    self.runtimes: list[Any|None] = []
    self.uop_replace: list[list[tuple[int, int]]] = []
    for call in self.linear.src:
-      replace = [(p, b.arg) for p, b in enumerate(get_call_arg_uops(call)) if b.op is Ops.PARAM]
+      replace = [(p, b.arg.slot) for p, b in enumerate(get_call_arg_uops(call)) if b.op is Ops.PARAM]
      for dev_idx, (bufs, device_vars) in enumerate(unwrap_multi(call, resolve_params(call, input_uops))):
        self.calls.append((dev_idx, call.src[0], [b.ensure_allocated() for b in bufs], device_vars))
        self.runtimes.append(get_runtime(bufs[0].device, call.src[0]) if call.src[0].op is Ops.PROGRAM else None)
@@ -193,7 +193,7 @@ class CapturedJit(Generic[ReturnType]):
      if call.op is not Ops.CALL: continue
      arg_uops = get_call_arg_uops(call)
      outs, ins = get_call_outs_ins(call)
-      out |= {arg_uops[k] for k in set(outs) - set(ins) if arg_uops[k].op in (Ops.BUFFER, Ops.BUFFER_VIEW)}
+      out |= {arg_uops[k] for k in set(outs) - set(ins) if arg_uops[k].op in (Ops.BUFFER, Ops.SLICE)}
    return out

  def __call__(self, input_uops:list[UOp], var_vals:dict[str, int]) -> ReturnType:
@@ -18,7 +18,7 @@ def get_call_arg_uops(call:UOp) -> tuple[UOp, ...]: return tuple(s for s in call
 def get_call_outs_ins(call:UOp) -> tuple[tuple[int, ...], tuple[int, ...]]:
  ast = call.src[0]
  if ast.op is Ops.PROGRAM: return tuple(ast.arg.outs), tuple(ast.arg.ins)
-  if ast.op in (Ops.COPY, Ops.BUFFER_VIEW): return (0,), (1,)
+  if ast.op in (Ops.COPY, Ops.SLICE): return (0,), (1,)
  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return (0,), tuple(range(1, len(get_call_arg_uops(call))))
  return (), ()

@@ -27,7 +27,9 @@ def get_call_name(call:UOp, bufs:list[Buffer], var_vals:dict[str, int]|None=None

  ast, arg_uops = call.src[0], get_call_arg_uops(call)
  if ast.op is Ops.PROGRAM: return ast.arg.name
-  if ast.op is Ops.BUFFER_VIEW: return colored(f"view {_uop_sz_to_str(arg_uops[0]):>10} @ {ast.arg[1] * arg_uops[1].dtype.itemsize:<10d}", "yellow")
+  if ast.op is Ops.SLICE:
+    offset = ast.src[1].arg * arg_uops[1].dtype.itemsize
+    return colored(f"view {_uop_sz_to_str(arg_uops[0]):>10} @ {offset:<10d}", "yellow")
  if ast.op is Ops.COPY: return colored(f"copy {_uop_sz_to_str(arg_uops[0]):>10}, {bufs[0].device[:7]:>7s} <- {bufs[1].device[:7]:7s}", "yellow")
  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return colored(f"enc/dec {_uop_sz_to_str(arg_uops[0])}", "yellow")
  if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph": return colored(f"batched {len(ast.src[0].src)}", "cyan")
@@ -135,8 +137,8 @@ class ExecContext:
  cache: bool = True

 def _resolve(b:UOp, inputs:tuple[UOp, ...]) -> UOp:
-  if b.op in (Ops.BUFFER_VIEW, Ops.MSELECT) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg], *b.src[1:]))
-  return inputs[b.arg] if b.op is Ops.PARAM else b
+  if b.op in (Ops.SLICE, Ops.MSELECT) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg.slot], *b.src[1:]))
+  return inputs[b.arg.slot] if b.op is Ops.PARAM else b
 def resolve_params(call:UOp, inputs:tuple[UOp, ...]) -> list[UOp]: return [_resolve(b, inputs) for b in get_call_arg_uops(call)]

 def unwrap_multi(call:UOp, resolved:list[UOp]) -> Iterator[tuple[list[Buffer], dict[str, int]]]:
@@ -149,7 +151,7 @@ def unwrap_multi(call:UOp, resolved:list[UOp]) -> Iterator[tuple[list[Buffer], d
 def exec_view(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
  resolved = resolve_params(call, ctx.input_uops)
  bufs = [cast(Buffer, b.buffer) for b in resolved]
-  bv = bufs[1].view(resolved[0].arg, ast.dtype, ast.arg[1]*bufs[1].dtype.itemsize)
+  bv = bufs[1].view(resolved[0].arg, ast.dtype, ast.src[1].arg*bufs[1].dtype.itemsize)
  with track_stats(ctx, call, bv.device, [bv, bufs[1]], ctx.var_vals): buffers[resolved[0]] = bv
  return None

@@ -175,7 +177,7 @@ def exec_kernel(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
    rt = get_runtime(device:=bufs[0].device, ast, cache=ctx.cache)
    global_size, local_size = ast.arg.launch_dims(var_vals)
    with track_stats(ctx, call, device, prg_bufs, var_vals) as tm:
-      et = tm[0] = rt(*[b._buf for b in prg_bufs], global_size=global_size, local_size=local_size, vals=ast.arg.vals(var_vals),
+      et = tm[0] = rt(*[b.get_buf(device) for b in prg_bufs], global_size=global_size, local_size=local_size, vals=ast.arg.vals(var_vals),
                       wait=ctx.wait, timeout=ctx.timeout)
  return et

@@ -231,7 +233,7 @@ pm_optimize_local_size = PatternMatcher([
 ])

 pm_exec = PatternMatcher([
-  (UPat(Ops.CALL, src=(UPat(Ops.BUFFER_VIEW, name="ast"),), name="call", allow_any_len=True), exec_view),
+  (UPat(Ops.CALL, src=(UPat(Ops.SLICE, name="ast"),), name="call", allow_any_len=True), exec_view),
  (UPat(Ops.CALL, src=(UPat(Ops.COPY, name="ast"),), name="call", allow_any_len=True), exec_copy),
  (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="ast"),), name="call", allow_any_len=True), exec_kernel),
  (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="encdec", name="ast"),), name="call", allow_any_len=True), exec_encdec),
@@ -239,14 +241,13 @@ pm_exec = PatternMatcher([
  (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="validate", name="ast"),), name="call", allow_any_len=True), exec_validate),
 ])

-if getenv("HCQ2"):
-  from extra.hcq2.hcq2 import pm_hcq_exec
-  pm_exec = pm_hcq_exec + pm_exec
-
 def compile_linear(linear:UOp, beam:int|None=None, validate=False) -> UOp:
  if validate: linear = graph_rewrite(linear, pm_validate, name="validate", walk=True)
  if (beam_val:=BEAM.value if beam is None else beam) >= 1: linear = graph_rewrite(linear, pm_beam, ctx=beam_val, walk=True)
  linear = graph_rewrite(linear, pm_compile, name="precompile kernels", walk=True)
+  if getenv("HCQ2"):
+    from extra.hcq2.hcq2 import hcq_schedule
+    linear = hcq_schedule(linear)
  return graph_rewrite(linear, pm_optimize_local_size, name="optimize local size", walk=True)

 def run_linear(linear:UOp, var_vals:dict[str, int]|None=None, input_uops:tuple[UOp, ...]=(), update_stats=True, jit=False, wait=False):
@@ -40,7 +40,7 @@ class _function(Generic[ReturnType]):
    params = get_state_dict((args, kwargs), tensor_type=(Tensor, UOp)).values()

    # deduplicate input_uops, keeping the first occurrence index for each unique uop
-    call_uops: list[UOp] = dedup([(t.uop if isinstance(t, Tensor) else t) for t in params])
+    call_uops: list[UOp] = dedup([u for t in params if not ((u:=(t.uop if isinstance(t, Tensor) else t)).base.op is Ops.CONST and u.device is None)])

    # disable realize/schedule while this is running
    # run it and do surgery later
@@ -16,20 +16,24 @@ def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):

 def _compact_params(body:UOp, all_args:tuple[UOp, ...]) -> tuple[UOp, tuple[UOp, ...]]:
  """Remove unused PARAMs from body and return compacted (body, args)."""
-  used = sorted({p.arg: p for p in body.toposort() if p.op is Ops.PARAM}.items())
-  return body.substitute({p: p.replace(arg=j) for j,(_, p) in enumerate(used)}, walk=True), tuple(all_args[i] for i,_ in used)
+  used = sorted({p.arg.slot: p for p in body.toposort() if p.op is Ops.PARAM}.items())
+  body = body.substitute({p: p.replace(arg=dataclasses.replace(p.arg, slot=j)) for j,(_, p) in enumerate(used)}, walk=True)
+  return body, tuple(all_args[i] for i,_ in used)

 def call_gradient(ctx:UOp, k:UOp, needed:set[int]) -> tuple[UOp|None, ...]:
  fxn, args = k.src[0], k.src[1:]
  if k.arg.grad_fxn is not None:
+    # put const on a device, also TODO why do we still have NOOP...
+    def on_dev(g, i): return g.clone(device=args[i].device if k.op is Ops.CALL else k.device) if g.device is None else g
    if ctx.op is Ops.TUPLE:
-      real = [g for g in ctx.src if g.op is not Ops.NOOP]
+      real = [on_dev(g, i) for i,g in enumerate(ctx.src) if g.op is not Ops.NOOP]
      return (None,) + (k.arg.grad_fxn(*real, call=k) if len(real) > 1 else k.arg.grad_fxn(real[0], k))
-    return (None,) + k.arg.grad_fxn(ctx, k)
+    return (None,) + k.arg.grad_fxn(on_dev(ctx, 0), k)
  assert fxn.op is Ops.TUPLE, f"expected TUPLE body for gradient, got {fxn.op}"
-  params = {x.arg:x for x in fxn.toposort(enter_calls=False) if x.op == Ops.PARAM}
+  params = {x.arg.slot:x for x in fxn.toposort(enter_calls=False) if x.op == Ops.PARAM}
  grad_args = ctx.src
-  root_grad = UOp(Ops.TUPLE, src=tuple(UOp(Ops.NOOP) if g.op is Ops.NOOP else g.param_like(len(args)+i) for i,g in enumerate(grad_args)))
+  root_grad = UOp(Ops.TUPLE, src=tuple(UOp(Ops.NOOP) if g.op is Ops.NOOP else
+    g if g.base.op is Ops.CONST and g.device is None else g.param_like(len(args)+i) for i,g in enumerate(grad_args)))
  grads = compute_gradient(fxn, root_grad, set(params.values()))
  # for precompiled calls, substitute forward outputs with params so intermediates aren't recomputed
  fwd_subs = {src: src.param_like(len(args)+len(grad_args)+i) for i, src in enumerate(fxn.src)} if k.arg.precompile else {}
@@ -70,7 +74,7 @@ pm_gradient = PatternMatcher([
    (ctx.cast(sum_acc_dtype(ctx.dtype))._rop(Ops.ADD, tuple(i for i,(s,n) in enumerate(zip(ret.src[0].shape, ret.shape)) if s!=n))
     .cast(ctx.dtype), None)),
  (UPat(Ops.PAD, name="ret"), lambda ctx, ret: (ctx.shrink(tuple([(p[0], s+p[0]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
-  (UPat(Ops.SHRINK, name="ret"), lambda ctx, ret: (ctx.pad(tuple([(p[0], s-p[1]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
+  (UPat(Ops.SHRINK, name="ret"), lambda ctx, ret: (ctx.pad(tuple([(p[0], s-p[0]-p[1]) for s,p in zip(ret.src[0].shape, ret.marg)])), None, None)),
  (UPat(Ops.PERMUTE, name="ret"), lambda ctx, ret: (ctx.permute(argsort(ret.marg)),)),
  (UPat(Ops.FLIP, name="ret"), lambda ctx, ret: (ctx.flip([i for i,x in enumerate(ret.marg) if x]),)),
  (UPat(Ops.COPY, name="ret"), lambda ctx, ret: (ctx.copy_to_device(ret.src[0].device), None)),
@@ -398,7 +398,7 @@ class Transformer:
    v_start_pos = UOp.variable("start_pos", 0, self.max_context-1)
    v_toks = UOp.variable("toks", 1, chunk_size)
    # TODO: use UOp.variable for temperature once float variables are supported
-    temp = Tensor(temperature).contiguous()
+    temp = Tensor([temperature])
    # assign all input tokens once, then slice from start_pos for the model call
    t = Tensor(tokens + [0] * (self.max_context - len(tokens)), dtype="int32").reshape(1, self.max_context)
    # recompute start_pos from what's currently valid in the caches
@@ -11,7 +11,7 @@ from tinygrad.dtype import ConstType, DType, DTypeLike, Invalid, InvalidType, Pt
 from tinygrad.helpers import all_int, argfix, ceildiv, flatten, flat_to_grouped, make_tuple, prod, resolve_pool_pads, round_up

 if TYPE_CHECKING:
-  from tinygrad.uop.ops import sint
+  from tinygrad.uop.ops import sint, UOp

 ReductionStr = Literal["mean", "sum", "none"]

@@ -23,12 +23,13 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
  def const(dtype, b, device=None): raise NotImplementedError("creation helpers are only supported on Tensor and UOp")

  @classmethod
-  def full(cls, shape:tuple[sint, ...], fill_value:ConstType, **kwargs) -> Self:
+  def full(cls, shape:tuple[sint, ...], fill_value:ConstType|UOp, dtype:DTypeLike|None=None,
+           device:str|tuple[str, ...]|None=None, buffer=True) -> Self:
    """
    Creates a tensor with the given shape, filled with the given value.

    You can pass in `dtype` and `device` keyword arguments to control the data type and device of the tensor.
-    Additionally, all other keyword arguments are passed to the constructor of the tensor.
+    Pass `buffer=False` to get a broadcast const value instead of a materialized buffer.

    ```python exec="true" source="above" session="tensor" result="python"
    print(Tensor.full((2, 3), 42).numpy())
@@ -37,11 +38,13 @@ class OpMixin(ElementwiseMixin, ReduceMixin):
    print(Tensor.full((2, 3), False).numpy())
    ```
    """
+    from tinygrad.uop.ops import UOp
    new_shape = argfix(shape)
-    if not kwargs.pop("buffer", True):
-      dt = to_dtype(kwargs.pop("dtype", None) or dtypes.from_py(fill_value))
-      return cls.const(dt, fill_value, canonicalize_device(kwargs.pop("device", None))).reshape((1,)*len(new_shape)).expand(new_shape)
-    return cls.unique_const(fill_value, **kwargs).reshape((1,)*len(new_shape)).expand(new_shape)
+    dt = to_dtype(dtype) if dtype is not None else None
+    if isinstance(fill_value, UOp): val = cls.const(dt or fill_value.dtype, fill_value)
+    else: val = cls.const(dt or dtypes.from_py(fill_value), fill_value, None if buffer else canonicalize_device(device))
+    val = val.reshape((1,)*len(new_shape)).expand(new_shape)
+    return val.clone(device=device) if buffer else val

  @classmethod
  def invalids(cls, *shape, **kwargs) -> Self:
@@ -52,7 +55,8 @@ class OpMixin(ElementwiseMixin, ReduceMixin):

    Eventually Tensor.empty will be replaced by this.
    """
-    return cls.full(argfix(*shape), Invalid, **kwargs)
+    new_shape = argfix(*shape)
+    return cls.unique_const(Invalid, **kwargs).reshape((1,)*len(new_shape)).expand(new_shape)

  @classmethod
  def zeros(cls, *shape, **kwargs) -> Self:
@@ -178,7 +178,7 @@ class MovementMixin:
  def pad(self, arg:tuple[tuple[sint, sint] | None, ...]) -> Self:
    if self.ndim != len(arg):
      raise ValueError(f"{self.ndim=} != {len(arg)=}")
-    ret = self._mop(Ops.PAD, tuple(x if x is not None else (0, 0) for x in arg))
+    ret = self._mop(Ops.PAD, tuple((x[0], s+x[0]+x[1]) if x is not None else (0, s) for x, s in zip(arg, self.shape)))
    return self if ret.shape == self.shape else ret

  def shrink(self, arg: tuple[tuple[sint, sint] | None, ...]) -> Self:
@@ -200,7 +200,7 @@ class MovementMixin:
    """
    if self.ndim != len(arg):
      raise ValueError(f"{self.ndim=} != {len(arg)=}")
-    ret = self._mop(Ops.SHRINK, arg=[x if x is not None else (0, s) for x, s in zip(arg, self.shape)])
+    ret = self._mop(Ops.SHRINK, arg=[(x[0], x[1]-x[0]) if x is not None else (0, s) for x, s in zip(arg, self.shape)])
    return self if ret.shape == self.shape else ret

  def permute(self, order, *args) -> Self:
@@ -251,7 +251,7 @@ class MovementMixin:
    return self.shrink(tuple([None if ns is None else (0, ns) for ns in argfix(shape, *args)]))

  def pad_to(self, shape, *args) -> Self:
-    return self._mop(Ops.PAD, tuple([(0, 0 if ns is None else ns-s) for s,ns in zip(self.shape, argfix(shape, *args), strict=True)]))
+    return self._mop(Ops.PAD, tuple((0, s if ns is None else ns) for s,ns in zip(self.shape, argfix(shape, *args), strict=True)))

  def view(self, shape, *args) -> Self:
    """`.view` is an alias for `.reshape`."""
@@ -1000,7 +1000,7 @@ def get_onnx_ops() -> dict[str, types.FunctionType|dict[OpSetId, types.FunctionT
      if align_corners: return Tensor.linspace(-1, 1, steps, device=theta.device)
      return Tensor.linspace(-1+1/steps, 1-1/steps, steps, device=theta.device)
    grids = Tensor.meshgrid(*(generate_grid(d) for d in spatial_dims))
-    base_grid = Tensor.stack(*reversed(grids), Tensor.ones_like(grids[0], device=theta.device), dim=-1)
+    base_grid = Tensor.stack(*reversed(grids), grids[0].const_like(1), dim=-1)
    base_grid = base_grid.reshape(1, prod(spatial_dims), len(grids)+1).expand(N, -1, -1)
    return (base_grid @ theta.transpose(1, 2)).reshape(N, *spatial_dims, -1)

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	5d13dd1123	refactor to u.	2026-05-29 19:05:20 -07:00
geohot	058259e0ea	refactor to use uop	2026-05-29 19:01:10 -07:00
geohot	da9318e745	python prg is pkl uops	2026-05-29 18:54:55 -07:00
sirhcmandGitHub	c377d01491	ci: run dsp on tinygrad[testing] (#16442 )	2026-05-29 21:16:56 -04:00
wozeparrotandGitHub	c23652e486	llama: minimize peak init mem (#16440 )	2026-05-29 18:00:37 -07:00
sirhcmandGitHub	d943493b79	ci: remove duplicate op compile test (#16441 )	2026-05-29 19:20:31 -04:00
chenyuandGitHub	8ac62b28e5	fix AffineGrid fusion (#16439 )	2026-05-29 17:59:47 -04:00
sirhcmandGitHub	ef50a49693	ci: macos dev matrix (#16436 )	2026-05-29 17:40:32 -04:00
sirhcmandGitHub	434cfa96a3	ci: no fetch in backend tests (#16438 ) should make for less actions cache thrashing	2026-05-29 17:11:16 -04:00
chenyuandGitHub	b7280705a7	limit CONST(UNIQUE) to invalids only (#16432 )	2026-05-29 16:02:06 -04:00
George HotzandGitHub	9506b78d73	fix viz addrspace (#16437 ) * fix viz addrspace * revert that	2026-05-29 12:58:05 -07:00
nimlgenandGitHub	d69aca41a9	hcq2: rework pm_bufferize (#16431 )	2026-05-29 22:09:52 +03:00
George HotzandGitHub	e2a0434403	full derivation of addrspace (#16433 ) * full derivation of addrspace * w/e, it fixes it	2026-05-29 11:39:31 -07:00
wozeparrotandGitHub	6787de9f52	llama: fix mp (#16434 )	2026-05-29 11:21:43 -07:00
chenyuandGitHub	2d7e5baab4	remove vec= from UPat.cvar [PR] (#16430 )	2026-05-29 10:52:30 -04:00
chenyuandGitHub	fa666cefe8	remove dead branch in UOp [PR] (#16429 )	2026-05-29 10:38:49 -04:00
qazalandGitHub	81bc00c006	do not require clearing method_cache in viz tests (#16428 ) * update * update test_dedup	2026-05-29 18:12:34 +09:00
qazalandGitHub	54cfb794b8	viz: addrspace little colored box (#16427 ) * return addrspace * layout * render * addrspace encodes color * update colors * in input_ast all are params are green * update stroke	2026-05-29 17:25:07 +09:00
qazalandGitHub	814d414f41	viz: set label offset for asm (#16426 )	2026-05-29 13:16:34 +09:00
wozeparrotandGitHub	f86966af56	llama: optim amax margin (#16425 )	2026-05-28 20:18:11 -07:00
sirhcmandGitHub	6e0d5262dc	ci: autocancel outdated pr jobs (#16424 )	2026-05-28 23:14:35 -04:00
sirhcmandGitHub	69aa2054f6	rename clangjit to clang (#16423 )	2026-05-28 22:41:58 -04:00
sirhcmandGitHub	a909acb882	move llvmspeed to benchmarks (#16422 )	2026-05-28 22:26:22 -04:00
George HotzandGitHub	1e7f1dcf49	add ParamArgs [pr] (#16421 ) * add ParamArgs * fix export * cleanups * fixes * simpler	2026-05-28 19:17:17 -07:00
sirhcmandGitHub	7d38edffdb	ci: dev matrix (#16420 ) windows just runs test_tiny	2026-05-28 22:04:04 -04:00
wozeparrotandGitHub	36c8ff70c1	llama: use old scale for dequant in optim (#16417 )	2026-05-28 15:21:19 -07:00
George Hotz GitHub sirhcm	c87f3433d1	use namespace runners (#16387 ) Co-authored-by: Christopher Milan <[email protected]>	2026-05-28 18:05:46 -04:00
George HotzandGitHub	c9adde72c1	addrspace property (#16418 ) * addrspace property * movement addrspace * regs	2026-05-28 14:39:25 -07:00
sirhcmandGitHub	c8af163d2b	disable process replay by default (#16419 ) enable process replay with [pr] and assert with [PR] process replay no longer captures on master	2026-05-28 17:36:28 -04:00
nimlgenandGitHub	b0e49afaf1	hcq2: new multi (#16413 ) * hcq2: new multi * op	2026-05-28 22:16:10 +03:00
George HotzandGitHub	edca5df25a	flip offset and shape in pad and shrink (#16414 ) * flip offset and shape in pad and shrink * dumb test	2026-05-28 11:58:19 -07:00
chenyuandGitHub	d72d8ee065	.const() should not ignore dtype (#16412 ) fixed a bug in postrange, also cleaner	2026-05-28 10:49:15 -04:00
sirhcmandGitHub	0ae957bb0a	refactor webgpu (#16406 )	2026-05-27 23:13:08 -04:00
qazalandGitHub	202adc644e	viz: make call toggle easier to click on (#16411 ) * call tag is a rect * details * colors * simplify, better comment	2026-05-28 11:53:36 +09:00
George HotzandGitHub	5ee6b6b79e	fix slice store to remove the index (#16410 ) * fix slice store to remove the index * fix spec	2026-05-27 19:17:53 -07:00
qazalandGitHub	88e88d63d6	viz: click on +- toggles sources (#16409 )	2026-05-28 09:12:43 +09:00
George HotzandGitHub	b21afb4883	marg line cleanup (#16408 ) * marg line cleanup * bitcast is a mop	2026-05-27 16:41:04 -07:00
wozeparrotandGitHub	dac3743d75	llama: delayed scaling in optim (#16407 )	2026-05-27 15:40:03 -07:00
George HotzandGitHub	8ee3a37524	shrink/pad use (new_shape, offset) (#16405 ) * shrink uses offset and shape * pad does too * fix	2026-05-27 15:13:08 -07:00
sirhcmandGitHub	171401e8df	skip modulo by zero in test_dtype_alu (#16404 )	2026-05-27 17:09:05 -04:00
qazalandGitHub	452c7d4230	llama: don't allocate grad_xw13 in bf16 (#16359 )	2026-05-28 04:33:07 +09:00
nimlgenandGitHub	0c385e31c6	hcq2 rewrite (#16375 ) * hcq2 rewrite * fi * x * simpler	2026-05-27 22:25:35 +03:00
chenyuandGitHub	c33b767407	bring back test and torch backend change for unique const (#16403 )	2026-05-27 15:16:08 -04:00
sirhcmandGitHub	bacabf0866	webgpu: fix enums (#16402 )	2026-05-27 13:09:50 -04:00
chenyuandGitHub	6da785562b	test_custom_kernel_precompile_multidevice (#16401 ) add a test to show what invalids need	2026-05-27 11:19:16 -04:00
chenyuandGitHub	3e80f375ee	skip test_setitem_fancy_on_unrealized_view (#16400 ) crashes in linux llvm ci	2026-05-27 09:50:26 -04:00
chenyuandGitHub	945ed4f689	revert const unique changes (#16395 )	2026-05-27 00:06:41 -04:00
sirhcmandGitHub	aacc8addf4	ci: use ubuntu 24.04 (#16393 )	2026-05-26 23:22:01 -04:00
chenyuandGitHub	fa14cde05c	test update for arange and eye (#16394 ) these will need explicit clone to make a buffer	2026-05-26 22:48:34 -04:00
wozeparrotandGitHub	3a7a6da7d5	llama: fakedata uses real vocab size (#16389 )	2026-05-26 18:58:55 -07:00
George HotzandGitHub	156a4438d9	rename BUFFER_VIEW to SLICE (#16391 ) * rename BUFFER_VIEW to SLICE * fix comments	2026-05-26 18:15:00 -07:00
sirhcmandGitHub	3adf7f5d95	disable flaky cl test (#16388 )	2026-05-26 19:56:57 -04:00
sirhcmandGitHub	d23659d38b	cleanup some old test skips (#16384 )	2026-05-26 19:07:22 -04:00
George HotzandGitHub	fd963038a0	remove allow_any_len from store (#16385 ) * remove allow_any_len from store * a few more * no bv there * more fixes * fixes * oh that	2026-05-26 15:26:53 -07:00
chenyuandGitHub	0b88827482	remove CONST(UNIQUE) (#16383 )	2026-05-26 14:45:22 -04:00
chenyuandGitHub	d861c50dce	remove unique_const (#16382 )	2026-05-26 13:53:31 -04:00
George HotzandGitHub	bac82d4949	fix emu bug in gfx950 (#16381 ) * fix emu bug in gfx950 * fix renderer	2026-05-26 10:32:03 -07:00
chenyuandGitHub	9b00defc8c	Revert "remove unique_const (#16372 )" (#16380 ) This reverts commit `09019d6761`.	2026-05-26 12:30:07 -04:00
chenyuandGitHub	09019d6761	remove unique_const (#16372 ) * remove unique_const * fix SDWA thing * that?	2026-05-26 12:18:03 -04:00
George HotzandGitHub	7f1b02854e	bufferview offset is units of input dtype (#16378 )	2026-05-26 08:49:31 -07:00
qazalandGitHub	846a809af7	viz: add +- toggle for hidden UOps (#16368 ) * first * remove * move src toggles to client side * line * update viz server tests * remove those * logic * cleanup * call matches * fix const arg * add labels * keep changes * the stack on movement ops hiding change * structure * rename to expandedNodes * work * test intention	2026-05-26 22:31:54 +09:00
nimlgenandGitHub	032905dec9	hcq2: simpler (#16361 )	2026-05-26 14:28:48 +03:00
geohot	322693dcd3	hotfix: bump Mac pytest timeout to 4 minutes (try 2)	2026-05-25 18:23:21 -07:00
George HotzandGitHub	41ee7dab1c	script to generate testsig for DSP (#16371 ) * script to generate testsig for DSP * cleanups	2026-05-25 17:54:58 -07:00
wozeparrotandGitHub	76fc39ccc0	gather to single device (#16354 )	2026-05-25 17:27:08 -07:00
geohot	942cb42b97	Revert "hotfix: bump Mac pytest timeout to 4 minutes" This reverts commit `695a0069ed`.	2026-05-25 17:25:11 -07:00
sirhcmandGitHub	8ddd1328df	remove getenv(CI) (#16365 ) gone everywhere except test_interop, because torch MPS does not work in actions	2026-05-25 20:23:33 -04:00
geohot	695a0069ed	hotfix: bump Mac pytest timeout to 4 minutes	2026-05-25 17:20:19 -07:00
George HotzandGitHub	689ab6a49f	move buffer view offset to src (#16364 ) * this work? * failed	2026-05-25 17:07:55 -07:00
sirhcmandGitHub	d8f86be613	webgpu: shader-f16 support in arch (#16370 )	2026-05-25 19:20:59 -04:00
qazalandGitHub	4bcc53eb26	viz: stable node position for +- toggle (#16367 )	2026-05-26 06:30:47 +09:00
qazalandGitHub	3506eb08ec	viz: sidebar toggles always recenter (#16366 ) * viz: sidebar toggles always recenters * python brain	2026-05-26 06:14:32 +09:00
chenyuandGitHub	cdeb861828	invalids is empty [pr] (#16353 )	2026-05-25 16:11:38 -04:00
qazalandGitHub	b73d2d17b9	viz/cli: add --interval (#16363 ) * interval support * add test_interval * llama uses interval	2026-05-26 03:35:06 +09:00
C TandGitHub	2ab90f31b1	use windows-specific alias nvcuda when loading cuda on windows (#16260 ) This also makes it possible to use cuda on windows by specifying 3 env vars with direct dll paths: NVCUDA_PATH, NVRTC_PATH and NVJITLINK_PATH without name collision with CUDA_PATH which is used for cuda headers include path in NVRTCCompiler.	2026-05-25 08:50:50 -07:00
wozeparrotandGitHub	68d2102fd2	llama: offload master weights (#16355 )	2026-05-25 08:48:13 -07:00
qazalandGitHub	eecd4706ff	fix mailbox comment, add types (#16360 )	2026-05-25 22:24:00 +09:00
nimlgenandGitHub	64095cf2e2	use get_buf in exec_kernel (#16356 )	2026-05-25 15:13:40 +03:00
chenyuandGitHub	5d5e02871f	remove Tensor.from_uop (#16344 ) and no device for const in Tensor init	2026-05-24 18:53:09 -04:00
nimlgenandGitHub	a891727c9f	hcq2: multi (#16347 ) * hcq2: multi * cleaner a bit	2026-05-24 19:28:33 +03:00
chenyuandGitHub	926d125a63	update test_stack (#16345 ) also skip COMPILE_ONLY, it was comparing 0==0	2026-05-23 10:42:35 -04:00
chenyuandGitHub	149a87dac2	deviceless const cleanups (#16341 )	2026-05-22 20:11:01 -04:00
sirhcmandGitHub	35461d4d8f	ci: cleanup some deps [pr] (#16340 )	2026-05-22 19:16:08 -04:00
sirhcmandGitHub	451f38155c	start cleanup of the slowest tests (#16339 )	2026-05-22 18:39:36 -04:00
nimlgenandGitHub	26b3b3f6a2	hcq2: move submit lowering to schedule (#16330 ) * hcq: move submit lowering to schedule * Dx	2026-05-22 23:15:19 +03:00