wip: test change, do not merge

Merge pull request 'szdiff: make workflow work on both github and gitea (supersedes #8 )' (#9 ) from szdiff_portable into master
szdiff: make comment steps work on both github and gitea
2026-08-14 13:58:27 +00:00 · 2026-08-01 10:50:45 -07:00 · 2026-08-01 10:49:44 -07:00 · 2026-08-01 10:44:41 -07:00 · 2026-08-01 10:34:37 -07:00
148 changed files with 2239 additions and 3562 deletions
@@ -41,6 +41,10 @@ inputs:
    description: "Install LLVM?"
    required: false
    default: 'false'
+  tinydreno:
+    description: "Install tinydreno"
+    required: false
+    default: 'false'
  qemu:
    description: "Install qemu"
    required: false
@@ -273,6 +277,12 @@ runs:
      shell: bash
      run: brew install llvm@20

+    # *** tinydreno ***
+    - name: Install tinydreno (linux)
+      if: inputs.tinydreno == 'true' && runner.os == 'Linux'
+      shell: bash
+      run: sudo curl -fL https://github.com/sirhcm/tinydreno/raw/refs/heads/master/libllvm-qcom.so -o /usr/lib/libllvm-qcom.so
+
    # *** OpenCL ***
    - name: Install rusticl
      if: inputs.opencl == 'true'
@@ -104,9 +104,6 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Setup (NV)
-      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p weights
@@ -158,9 +155,6 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Setup (NV)
-      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -210,9 +204,6 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Setup (NV)
-      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: Symlink models and datasets
      run: |
        mkdir -p extra/datasets
@@ -259,9 +250,6 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Setup (NV)
-      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -304,9 +292,6 @@ jobs:
        ./extra/amdpci/setup_python_cap.sh
        ./extra/hcq/hcq_smi.py amd rmmod
        ./extra/hcq/hcq_smi.py amd kill_pids
-    - name: Setup (NV)
-      if: ${{ matrix.dev == 'NV' }}
-      run: sudo lsof -tQ /dev/nvidia* | { xargs -r sudo kill -9 || true; }
    - name: setup staging db
      if: github.ref == 'refs/heads/update_benchmark_staging'
      run: |
@@ -1,181 +0,0 @@
-name: Platform Tests
-env:
-  # increment this when downloads substantially change to avoid the internet
-  CACHE_VERSION: '19'
-  CAPTURE_PROCESS_REPLAY: ${{ github.event_name == 'pull_request' && contains(github.event.pull_request.title, '[pr]') && '1' || '0' }}
-  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-  PYTHONPATH: ${{ github.workspace }}
-  CHECK_OOB: 1
-
-on:
-  push:
-    branches:
-      - master
-  pull_request:
-  workflow_dispatch:
-
-concurrency:
-  group: platform-${{ 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:
-
-# ****** OSX Tests ******
-
-  unittestmacos:
-    name: MacOS (unit)
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-macos
-        deps: testing_unit
-    - name: Run unit tests
-      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
-    - name: Test tensor core ops (fake)
-      run: DEV=METAL DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
-    - name: Test tensor core ops (real)
-      run: DEV=METAL DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
-    - name: Test Beam Search
-      run: DEV=METAL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
-    - name: Test Device Specific
-      run: DEV=METAL python3 -m pytest test/device/test_metal.py
-    #- name: Fuzz Test linearizer
-    #  run: DEV=METAL DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  unittestmacosmock:
-    name: MacOS (unit, mock)
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: unittest-macos-mock
-        deps: testing_unit
-        amd: 'true'
-        ocelot: 'true'
-    - name: Run NULL backend tests
-      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
-    - name: Run pytest (amd)
-      env:
-        DEV: MOCKKFD+AMD
-        FORWARD_ONLY: 1
-      run: |
-        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
-    - name: Run pytest (ptx)
-      env:
-        DEV: "MOCK+NV:PTX"
-        FORWARD_ONLY: 1
-        # TODO: failing due to library loading error
-        CAPTURE_PROCESS_REPLAY: 0
-      run: |
-        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py \
-          test/testextra/test_hevc.py::TestHevc::test_hevc_decode_compile --durations=20
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmetal:
-    strategy:
-      fail-fast: false
-      matrix:
-        group: [1, 2]
-    name: MacOS (DEV=METAL) (${{ matrix.group }})
-    runs-on: macos-26
-    timeout-minutes: 20
-    env:
-      DEV: METAL
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: macos-metal
-        deps: testing_unit
-    - name: Check Device.DEFAULT and print some source
-      run: |
-        python -c "from tinygrad import Device; assert Device.DEFAULT == 'METAL'"
-        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
-    - name: Run backend tests
-      run: python -m pytest -n=auto test/backend --durations=20 --splits 2 --group ${{ matrix.group }}
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-  testmacos:
-    strategy:
-      fail-fast: false
-      matrix:
-        dev:
-          - 'CPU:CLANG'
-          - 'CPU:LLVM'
-          - 'CPU:LVP'
-          - 'WEBGPU'
-
-    name: MacOS (DEV=${{ matrix.dev }})
-    runs-on: macos-26
-    timeout-minutes: 20
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: macos-${{ matrix.dev }}
-        deps: "testing_unit${{ contains(matrix.dev, 'LVP') && ' mesa' || '' }}"
-        llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') }}
-        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 test_tiny
-      run: python -m pytest -n=auto test/test_tiny.py --durations=20
-    - name: Run process replay tests
-      uses: ./.github/actions/process-replay
-
-# ****** Windows Tests ******
-
-  testwindows:
-    strategy:
-      fail-fast: false
-      matrix:
-        dev:
-          - 'CPU:CLANG'
-          - 'CPU:LLVM'
-          - 'CPU:X86'
-          - 'WEBGPU'
-
-    name: Windows (DEV=${{ matrix.dev }})
-    runs-on: windows-2025
-    timeout-minutes: 15
-    steps:
-      - name: Checkout Code
-        uses: actions/checkout@v6
-      - name: Setup Environment
-        uses: ./.github/actions/setup-tinygrad
-        with:
-          key: windows-${{ matrix.dev }}-minimal
-          deps: testing_unit
-          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
-      - name: Set env
-        shell: bash
-        run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
-      - name: Check Device.DEFAULT and print some source
-        shell: bash
-        run: |
-          python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ 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
@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+# Sticky PR comment via the REST API: find an existing comment containing MARKER and PATCH it, or POST a new one.
+# Works on GitHub and Gitea (stdlib only, replaces marocchino/sticky-pull-request-comment which needs GraphQL).
+# Env vars: GITHUB_TOKEN, GITHUB_API_URL, GITHUB_REPOSITORY (set by the runner), PR_NUMBER, MARKER, and BODY_FILE or MESSAGE.
+import json, os, sys, urllib.request
+
+api, repo = os.environ["GITHUB_API_URL"], os.environ["GITHUB_REPOSITORY"]
+pr, marker = os.environ["PR_NUMBER"], os.environ["MARKER"]
+body = open(os.environ["BODY_FILE"]).read() if os.environ.get("BODY_FILE") else os.environ["MESSAGE"]
+
+if not body.strip():
+  print("comment body is empty, not posting")
+  sys.exit(0)
+
+def req(url, method="GET", payload=None):
+  r = urllib.request.Request(url, data=None if payload is None else json.dumps(payload).encode(), method=method,
+    headers={"Authorization": f"token {os.environ['GITHUB_TOKEN']}", "Accept": "application/json", "Content-Type": "application/json"})
+  return json.load(urllib.request.urlopen(r))
+
+# find the latest sticky comment (paginate, 100 comments per page)
+existing, page = None, 1
+while True:
+  comments = req(f"{api}/repos/{repo}/issues/{pr}/comments?per_page=100&page={page}")
+  stickies = [c for c in comments if marker in (c.get("body") or "")]
+  if stickies: existing = stickies[-1]
+  if not comments or len(comments) < 100: break
+  page += 1
+
+if existing is not None and existing["body"] == body:
+  print("comment is already up to date")
+  sys.exit(0)
+url = f"{api}/repos/{repo}/issues/comments/{existing['id']}" if existing is not None else f"{api}/repos/{repo}/issues/{pr}/comments"
+resp = req(url, 'PATCH' if existing is not None else 'POST', {'body': body})
+print(f"{'updated' if existing is not None else 'created'} comment {resp['id']}")
@@ -26,14 +26,14 @@ jobs:
      - name: Check whether branch is up-to-date
        id: brstat
        run: |
-          git remote add tinygrad https://github.com/tinygrad/tinygrad
-          git fetch tinygrad master
+          # fetch master from the base repo (tinygrad/tinygrad on GitHub, the mirror on Gitea), not the PR head remote
+          git fetch "${{ github.event.pull_request.base.repo.clone_url }}" master
          echo "${{ github.event.pull_request.head.sha }}"
-          git rev-list --left-right --count  tinygrad/master...${{ github.event.pull_request.head.sha }} | awk '{print "Behind "$1" - Ahead "$2""}'
-          count=$(git rev-list --left-right --count  tinygrad/master...${{ github.event.pull_request.head.sha }} | awk '{print $1}')
+          git rev-list --left-right --count FETCH_HEAD...${{ github.event.pull_request.head.sha }} | awk '{print "Behind "$1" - Ahead "$2""}'
+          count=$(git rev-list --left-right --count FETCH_HEAD...${{ github.event.pull_request.head.sha }} | awk '{print $1}')
          if [ $count -gt 0 ]
          then
-            echo "Current branch is behind tinygrad master branch!"
+            echo "Current branch is behind ${{ github.event.pull_request.base.repo.full_name }} master branch!"
            echo "stat=true" >> "$GITHUB_OUTPUT"
          else
            echo "stat=false" >> "$GITHUB_OUTPUT"
@@ -75,13 +75,13 @@ jobs:
          python sz.py "$BASE" "$PR" > loc_content.txt
      - name: Comment Code Line Diff
        continue-on-error: false
-        uses: marocchino/sticky-pull-request-comment@v3
-        with:
+        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          ignore_empty: true
-          skip_unchanged: true
-          recreate: true
-          path: loc_content.txt
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          MARKER: "### Changes"
+          BODY_FILE: loc_content.txt
+        # note: run the script from the base checkout, never from the PR checkout
+        run: python3 "$GITHUB_WORKSPACE/base/.github/workflows/sticky_comment.py"

  rebase:
    name: Core Library Line Difference
@@ -91,12 +91,14 @@ jobs:
    needs: checkbranch
    if: needs.checkbranch.outputs.branchstat == 'true'
    steps:
+      # pull_request_target: a plain checkout gets the base repo, so no PR code is executed
+      - uses: actions/checkout@v6
      - name: Comment Rebase
        continue-on-error: false
-        uses: marocchino/sticky-pull-request-comment@v3
-        with:
+        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          skip_unchanged: true
-          recreate: true
-          message: |
-            This branch currently is behind tinygrad/master. The line count difference bot is disabled.
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          MARKER: "line count difference bot is disabled"
+          MESSAGE: |
+            This branch currently is behind ${{ github.event.pull_request.base.repo.full_name }} master. The line count difference bot is disabled.
+        run: python3 .github/workflows/sticky_comment.py
@@ -21,7 +21,7 @@ concurrency:
 jobs:
  docs:
    name: Docs
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    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
@@ -61,7 +61,7 @@ jobs:

  torchbackend:
    name: Torch Backend Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -88,7 +88,7 @@ jobs:

  bepython:
    name: Python Backend
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -126,7 +126,7 @@ jobs:

  linter:
    name: Linters
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 10

    steps:
@@ -157,7 +157,7 @@ jobs:

  nulltest:
    name: Null Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15

    steps:
@@ -191,7 +191,7 @@ jobs:

  unittest:
    name: Unit Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15

    steps:
@@ -219,8 +219,8 @@ jobs:
      run: python3 test/external/external_benchmark_schedule.py
    - name: Run process replay tests
      uses: ./.github/actions/process-replay
-    - name: Repo line count <= 26000 lines
-      run: MAX_LINE_COUNT=26000 python sz.py
+    - name: Repo line count < 25000 lines
+      run: MAX_LINE_COUNT=25000 python sz.py

  spec:
    strategy:
@@ -228,7 +228,7 @@ jobs:
      matrix:
        group: [1, 2]
    name: SPEC=2 (${{ matrix.group }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -244,7 +244,7 @@ jobs:

  fuzzing:
    name: Fuzzing
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 10
    steps:
    - name: Checkout Code
@@ -260,7 +260,7 @@ jobs:

  testopenclimage:
    name: CL IMAGE Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -280,7 +280,7 @@ jobs:

  testopenpilot:
    name: openpilot Compile Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -294,7 +294,7 @@ jobs:
          llvm: 'true'
      - name: Test openpilot model kernel count and gate usage
        run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=38 FLOAT16=1 DEV="CL::IMAGE_PITCH_ALIGNMENT=64" IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=54 FLOAT16=1 DEV="CL::IMAGE_PITCH_ALIGNMENT=64" IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
          # IMAGE_PITCH_ALIGNMENT=64 matches adreno 630
      - name: Test openpilot CL compile fp32 (test correctness)
        run: |
@@ -309,7 +309,7 @@ jobs:

  testonnxcpu:
    name: ONNX (CPU) Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20

    steps:
@@ -328,7 +328,7 @@ jobs:

  testoptim:
    name: Optimization Tests
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    steps:
      - name: Checkout Code
@@ -360,7 +360,7 @@ jobs:

  testllm:
    name: Test LLM
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    env:
      CHECK_OOB: 0
@@ -387,7 +387,7 @@ jobs:

  testmodels:
    name: Models
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
      - name: Checkout Code
@@ -407,7 +407,7 @@ jobs:

  testdsp:
    name: Linux (DSP)
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    steps:
    - name: Checkout Code
@@ -435,7 +435,7 @@ jobs:
          - 'WEBGPU'

    name: Linux (DEV=${{ matrix.dev }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    steps:
    - name: Checkout Code
@@ -461,7 +461,7 @@ jobs:

  testamdasm:
    name: AMD ASM IDE
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    env:
      DEV: MOCKKFD+AMD
@@ -507,7 +507,7 @@ jobs:

  hcq2:
    name: hcq2
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 5
    steps:
      - name: Checkout Code
@@ -521,15 +521,16 @@ jobs:
      - name: Run HCQ2 tests
        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/test_tiny.py
      - name: Run HCQ2 multi-device tests
-        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest -n=auto test/backend/test_multitensor.py
+        run: |
+          HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_multitensor.py \
+            TestMultiTensor.test_simple_add TestMultiTensor.test_shard_reduce \
+            TestMultiTensor.test_backward_sum TestMultiTensor.test_matmul_shard_0_0
      - name: Run HCQ2 JIT tests
        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python test/unit/test_jit.py
-      - name: Run HCQ2 unit tests
-        run: HCQ_RUNTIME_DEV=PYTHON HCQ2=1 DEV=MOCKKFD+AMD FORWARD_ONLY=1 PYTHONPATH=. python -m pytest test/device/test_hcq2.py

  testmockam:
    name: Linux (am)
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    env:
      DEV: MOCKPCI+AMD
@@ -565,7 +566,7 @@ jobs:
        arch: [gfx1100, gfx1201, gfx950]

    name: Linux (${{ matrix.backend }} ${{ matrix.arch }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 15
    env:
      DEV: MOCKKFD+AMD:${{ matrix.backend == 'amdllvm' && 'LLVM' || '' }}:${{ matrix.arch }}
@@ -584,9 +585,6 @@ jobs:
        run: |
          python3 -c "from tinygrad import Device; assert Device.DEFAULT in ['AMD'], Device.DEFAULT"
          DEBUG=5 FORWARD_ONLY=1 python3 test/test_tiny.py TestTiny.test_plus
-      - name: Run MXFP4 Llama training on NULL backend
-        if: ${{ matrix.backend == 'amd' && matrix.arch == 'gfx950' }}
-        run: PYTHONPATH=. DEV=NULL:HIP:gfx950 MXFP4=1 LLAMA_LAYERS=2 BENCHMARK=3 NULL_ALLOW_COPYOUT=1 NO_HIPCC=1 ROCM_PATH=/opt/rocm JITBEAM=0 examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/profile.sh
      - name: Run pytest (amd)
        run: python -m pytest -n=auto test/backend/test_ops.py test/backend/test_dtype.py test/backend/test_dtype_alu.py test/backend/test_linearizer.py test/backend/test_randomness.py test/backend/test_jit.py test/backend/test_graph.py test/backend/test_multitensor.py test/device/test_hcq.py test/external/external_test_am.py test/backend/test_asm_gemm.py::TestAsmGEMM --durations=20
      - name: Run disk copy tests
@@ -603,7 +601,7 @@ jobs:
        backend: [ptx, nv]

    name: Linux (${{ matrix.backend }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
    timeout-minutes: 20
    env:
      FORWARD_ONLY: 1
@@ -631,23 +629,175 @@ jobs:
      - name: Run process replay tests
        uses: ./.github/actions/process-replay

+# ****** OSX Tests ******
+
+  unittestmacos:
+    name: MacOS (unit)
+    runs-on: macos-26
+    timeout-minutes: 20
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: unittest-macos
+        deps: testing_unit
+    - name: Run unit tests
+      run: DEV=METAL python -m pytest -n=auto test/unit/ --durations=20
+    - name: Test tensor core ops (fake)
+      run: DEV=METAL DEBUG=3 TC=2 python test/backend/test_ops.py TestOps.test_gemm
+    - name: Test tensor core ops (real)
+      run: DEV=METAL DEBUG=3 python test/backend/test_ops.py TestOps.test_big_gemm
+    - name: Test Beam Search
+      run: DEV=METAL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
+    - name: Test Device Specific
+      run: DEV=METAL python3 -m pytest test/device/test_metal.py
+    #- name: Fuzz Test linearizer
+    #  run: DEV=METAL DEPTH=4 FUZZ_N=50 FUZZ_MAX_SIZE=1000000 python test/external/fuzz_linearizer.py
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  unittestmacosmock:
+    name: MacOS (unit, mock)
+    runs-on: macos-26
+    timeout-minutes: 20
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: unittest-macos-mock
+        deps: testing_unit
+        amd: 'true'
+        ocelot: 'true'
+    - name: Run NULL backend tests
+      run: SPEC=2 DEV=NULL python -m pytest -n=auto test/null/ --durations=20
+    - name: Run pytest (amd)
+      env:
+        DEV: MOCKKFD+AMD
+        FORWARD_ONLY: 1
+      run: |
+        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
+    - name: Run pytest (ptx)
+      env:
+        DEV: "MOCK+NV:PTX"
+        FORWARD_ONLY: 1
+        # TODO: failing due to library loading error
+        CAPTURE_PROCESS_REPLAY: 0
+      run: |
+        python3 -m pytest -n=auto test/device/test_hcq.py test/test_tiny.py --durations=20
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testmetal:
+    strategy:
+      fail-fast: false
+      matrix:
+        group: [1, 2]
+    name: MacOS (DEV=METAL) (${{ matrix.group }})
+    runs-on: macos-26
+    timeout-minutes: 20
+    env:
+      DEV: METAL
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: macos-metal
+        deps: testing_unit
+    - name: Check Device.DEFAULT and print some source
+      run: |
+        python -c "from tinygrad import Device; assert Device.DEFAULT == 'METAL'"
+        DEBUG=4 python test/test_tiny.py TestTiny.test_plus
+    - name: Run backend tests
+      run: python -m pytest -n=auto test/backend --durations=20 --splits 2 --group ${{ matrix.group }}
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+  testmacos:
+    strategy:
+      fail-fast: false
+      matrix:
+        dev:
+          - 'CPU:CLANG'
+          - 'CPU:LLVM'
+          - 'CPU:LVP'
+          - 'WEBGPU'
+
+    name: MacOS (DEV=${{ matrix.dev }})
+    runs-on: macos-26
+    timeout-minutes: 20
+    steps:
+    - name: Checkout Code
+      uses: actions/checkout@v6
+    - name: Setup Environment
+      uses: ./.github/actions/setup-tinygrad
+      with:
+        key: macos-${{ matrix.dev }}
+        deps: "testing_unit${{ contains(matrix.dev, 'LVP') && ' mesa' || '' }}"
+        llvm: ${{ contains(matrix.dev, 'LLVM') || contains(matrix.dev, 'LVP') }}
+        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 test_tiny
+      run: python -m pytest -n=auto test/test_tiny.py --durations=20
+    - name: Run process replay tests
+      uses: ./.github/actions/process-replay
+
+# ****** Windows Tests ******
+
+  testwindows:
+    strategy:
+      fail-fast: false
+      matrix:
+        dev:
+          - 'CPU:CLANG'
+          - 'CPU:LLVM'
+          - 'CPU:X86'
+          - 'WEBGPU'
+
+    name: Windows (DEV=${{ matrix.dev }})
+    runs-on: windows-2025
+    timeout-minutes: 15
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v6
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: windows-${{ matrix.dev }}-minimal
+          deps: testing_unit
+          pydeps: ${{ matrix.dev == 'WEBGPU' && 'dawn-python' || '' }}
+      - name: Set env
+        shell: bash
+        run: printf "DEV=${{ matrix.dev }}${{ matrix.dev == 'CPU:CLANG' && '\nCPU_COUNT=2' || '' }}" >> $GITHUB_ENV
+      - name: Check Device.DEFAULT and print some source
+        shell: bash
+        run: |
+          python -c "from tinygrad import Device; from tinygrad.helpers import Target; assert Device.DEFAULT == Target.parse('${{ 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 ******

  compiletests:
    strategy:
      fail-fast: false
      matrix:
-        dev:
-          - 'NULL:IR3:a630'
-          - 'NULL:QCOMCL:a630'
-          - 'NULL:NAK:sm_120'
-    name: Compile-only (DEV=${{ matrix.dev }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+        backend: [ir3, nak]
+    name: Compile-only (${{ matrix.backend }})
+    runs-on: *linux
    timeout-minutes: 15
-    env:
-      NULL_ALLOW_COPYOUT: 1
-      DEV: ${{ matrix.dev }}${{ contains(matrix.dev, 'a630') && ',IMAGE_PITCH_ALIGNMENT=64' || '' }}
-      IMAGE: ${{ contains(matrix.dev, 'a630') && '1' || '0' }}
    steps:
      - name: Checkout Code
        uses: actions/checkout@v6
@@ -656,14 +806,51 @@ jobs:
        with:
          key: compile-${{ matrix.backend }}
          deps: "testing_unit mesa"
-          qemu: ${{ contains(matrix.dev, 'QCOMCL') }}
-      - name: Test IMAGE
+      - name: Set env
        shell: bash
-        if: contains(matrix.dev, 'a630')
-        run: DEBUG=7 python3 test/backend/test_ops.py TestOps.test_gemm | grep isam
+        run: printf "NULL_ALLOW_COPYOUT=1\n${{ matrix.backend == 'ir3' && 'DEV=NULL:IR3:a630' || matrix.backend == 'nak' && 'DEV=NULL:NAK:sm_120' }}" >> $GITHUB_ENV
      - name: Run test_ops
        shell: bash
        run: |
          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+      - name: Run test_ops (IMAGE)
+        if: matrix.backend == 'ir3'
+        shell: bash
+        env:
+          IMAGE: 1
+          DEV: "NULL:IR3:a630,IMAGE_PITCH_ALIGNMENT=64"
+        run: |
+          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_gemm | grep image_load
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+  qcomclcompiletests:
+    name: Compile-only (QCOM CL)
+    runs-on: ubuntu-24.04-arm
+    timeout-minutes: 15
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@v6
+      - name: Setup Environment
+        uses: ./.github/actions/setup-tinygrad
+        with:
+          key: compile-qcomcl
+          deps: testing_unit
+          tinydreno: 'true'
+      - name: Set env
+        shell: bash
+        run: printf "DEV=NULL:QCOMCL:a630\nNULL_ALLOW_COPYOUT=1" >> $GITHUB_ENV
+      - name: Run test_ops
+        shell: bash
+        run: |
+          python -c "from tinygrad import Device; assert Device.DEFAULT == 'NULL'"
+          DEBUG=4 python3 test/backend/test_ops.py TestOps.test_add
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20
+      - name: Run test_ops (IMAGE)
+        shell: bash
+        env:
+          IMAGE: 1
+          DEV: "NULL:QCOMCL:a630,IMAGE_PITCH_ALIGNMENT=64"
+        run: |
+          DEBUG=4 python test/backend/test_ops.py TestOps.test_gemm | grep read_imagef
+          python -m pytest -n=auto test/backend/test_ops.py --durations=20
@@ -1282,7 +1282,7 @@ def train_bert():
        previous_step = i

 def train_llama3():
-  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8, MXFP4
+  from examples.mlperf.models.flat_llama import FlatTransformer, apply_grad, FP8_DTYPE, MXFP8
  from examples.llama3 import MODEL_PARAMS
  from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup
  from examples.mlperf.optim import GradAccClipAdamW, clip_grads
@@ -1434,9 +1434,9 @@ def train_llama3():
    load_state_dict(scheduler, safe_load(fn), realize=False)

  fp8_amax = [t for ts in model._fp8_amax.values() for t in ts]
-  fp8_next_amax = [t for ts in model._fp8_next_amax.values() for t in ts]
-  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts]
-  fp8_next_grad_amax = [t for ts in model._fp8_next_grad_amax.values() for t in ts]
+  fp8_next_amax = [t for ts in model._fp8_next_amax.values() for t in ts] if hasattr(model, "_fp8_next_amax") else []
+  fp8_grad_amax = [t for ts in model._fp8_grad_amax.values() for t in ts] if hasattr(model, "_fp8_grad_amax") else []
+  fp8_next_grad_amax = [t for ts in model._fp8_next_grad_amax.values() for t in ts] if hasattr(model, "_fp8_next_grad_amax") else []
  fp8_inv_scales = list(model._fp8_inv_scale.values()) + list(model._fp8_next_inv_scale.values())

  from tinygrad.nn.state import get_state_dict
@@ -1458,12 +1458,12 @@ def train_llama3():

  # realize everything here
  if optim.master_params: Tensor.realize(*optim.master_params)
-  loss_acc = Tensor.zeros(1, dtype=dtypes.float32, device=device)
-  Tensor.realize(loss_acc, *optim.params, *fp8_inv_scales, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
+  Tensor.realize(*optim.params, *fp8_inv_scales, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)

  @TinyJit
  def minibatch(tokens:Tensor):
-    model.reset_amax()
+    for nxt in fp8_next_amax: nxt.assign(0)
+    for nxt in fp8_next_grad_amax: nxt.assign(0)
    if is_dp: tokens = tokens.to(None).shard(device, 0)
    if is_mp: tokens = tokens.shard(device)
    if not is_sharding: tokens = tokens.to(None)
@@ -1477,8 +1477,8 @@ def train_llama3():
    for g, new_g in zip(grads, loss.gradient(*optim.params)):
      apply_grad(g, new_g.uop)

-    loss_acc.assign(loss_acc + loss.flatten().float())
-    return loss_acc.realize(*grads, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)
+    loss_cpu = loss.flatten().float().to("CPU")
+    return loss_cpu.realize(*grads, *fp8_amax, *fp8_next_amax, *fp8_grad_amax, *fp8_next_grad_amax)

  @TinyJit
  def optim_step():
@@ -1487,14 +1487,14 @@ def train_llama3():
    scheduler.step()

    for g in grads: g.assign(0)
-    model.update_amax()
+    for cur, nxt in zip(fp8_amax, fp8_next_amax): cur.assign(nxt)
+    for cur, nxt in zip(fp8_grad_amax, fp8_next_grad_amax): cur.assign(nxt)

    lr_cpu = optim.lr.float().to("CPU")
    grad_norm_cpu = grad_norm.float().to("CPU")
-    loss_cpu = loss_acc.to("CPU")
-    Tensor.realize(lr_cpu, grad_norm_cpu, loss_cpu, loss_acc.assign(0), *grads, *fp8_inv_scales, *fp8_amax, *fp8_grad_amax)
+    Tensor.realize(lr_cpu, grad_norm_cpu, *grads, *fp8_inv_scales, *fp8_amax, *fp8_grad_amax)

-    return lr_cpu, grad_norm_cpu, loss_cpu
+    return lr_cpu, grad_norm_cpu

  @TinyJit
  @Context(TRAINING=0)
@@ -1549,8 +1549,8 @@ def train_llama3():
      st = time.perf_counter()

      stopped = False
-      data_time, dev_time = 0, 0
-      for _ in range(accum_steps:=grad_acc if i >= 2 else 1):
+      losses, data_time, dev_time = [], 0, 0
+      for _ in range(grad_acc if i >= 2 else 1):
        ist = time.perf_counter()
        try: tokens = next(train_iter)
        except StopIteration:
@@ -1558,15 +1558,16 @@ def train_llama3():
          break
        mst = time.perf_counter()
        data_time += mst - ist
-        minibatch(tokens)
+        losses.append(minibatch(tokens).item())
        dev_time += time.perf_counter() - mst
      if stopped: break

      gt = time.perf_counter()
      ret = optim_step()
-      lr, grad_norm, loss = ret[0].item(), ret[1].item(), ret[2].item() / accum_steps
+      lr, grad_norm = ret[0].item(), ret[1].item()
      et = time.perf_counter()

+      loss = sum(losses) / len(losses)
      optim_time = et - gt
      dev_time += optim_time
      step_time = et - st
@@ -1578,7 +1579,7 @@ def train_llama3():

      mem_gb = GlobalCounters.mem_used / 1e9
      gflops = GlobalCounters.global_ops / 1e9 / dev_time
-      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * (9.2e15 if MXFP4 else 4.6e15))) * 100
+      mfu = ((6 * num_params * SEQLEN * GBS) / (dev_time * device_count * 4.6e15)) * 100
      tqdm.write(
          f"{i:5} {step_time:.3f} s step, {gbs_time:.3f} s gbs, {optim_time:.3f} s optim, {data_time:.3f} s data, {loss:.4f} loss, " \
          f"{lr:.12f} LR, {grad_norm:.6f} grad_norm, {mem_gb:.2f} GB used, {gflops:9.2f} GFLOPS, {mfu:5.2f}% MFU")
@@ -83,8 +83,8 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
      return out, x_fp8
  return (x_fp8.dot(w.T, dtype=dtypes.float) * ((amax_x.float() + 1e-8) / FP8_MAX) * w_inv_scale).cast(dtypes.bfloat16), x_fp8

-def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
-                         next_amax_x:Tensor|None, grad_amax_state:Tensor|None, next_grad_amax_state:Tensor|None):
+def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor,
+                         next_amax_x:Tensor, grad_amax_state:Tensor, next_grad_amax_state:Tensor):
  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_rmsnorm_mul_quantize_fp8
    x_fp8, x_normed, rrms = fused_rmsnorm_mul_quantize_fp8(x, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
@@ -96,8 +96,8 @@ def norm_quantize_matmul(x:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, ep
                     next_grad_amax_state=next_grad_amax_state, next_amax_x=next_amax_x)
  return out, x_normed, rrms, ret

-def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor|None,
-                             next_amax_x:Tensor|None, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
+def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w_inv_scale:Tensor, eps:float, amax_x:Tensor,
+                             next_amax_x:Tensor, grad_amax_state:Tensor|None=None, next_grad_amax_state:Tensor|None=None):
  if FUSED_ADD_NORM_MUL_QUANTIZE and not MXFP4:
    from extra.llama_kernels.fused_rmsnorm_mul_quantize_fp8 import fused_add_rmsnorm_mul_quantize_fp8
    x_fp8, h, x_normed, rrms = fused_add_rmsnorm_mul_quantize_fp8(x, residual, norm, amax_x, eps, FP8_DTYPE, next_amax_x)
@@ -111,14 +111,9 @@ def add_norm_quantize_matmul(x:Tensor, residual:Tensor, norm:Tensor, w:Tensor, w
  return out, h, x_normed, rrms, ret

 def silu_w13_quantize_matmul(x_w13:Tensor, w2:Tensor, s_2:Tensor,
-                             amax_x2:Tensor|None, next_amax_x2:Tensor|None,
-                             grad_amax_xw13:Tensor|None, next_grad_amax_xw13:Tensor|None,
-                             grad_amax_xout:Tensor|None, next_grad_amax_xout:Tensor|None):
-  if FUSED_SILU_W13 and MXFP4:
-    from extra.llama_kernels.swiglu import swiglu
-    out, *ret = matmul(swiglu(x_w13), w2, amax_x=amax_x2, w_inv_scale=s_2, grad_amax_state=grad_amax_xout,
-                       next_grad_amax_state=next_grad_amax_xout, next_amax_x=next_amax_x2)
-    return out, ret
+                             amax_x2:Tensor, next_amax_x2:Tensor,
+                             grad_amax_xw13:Tensor, next_grad_amax_xw13:Tensor,
+                             grad_amax_xout:Tensor, next_grad_amax_xout:Tensor):
  if FUSED_SILU_W13 and not MXFP4:
    from extra.llama_kernels.cast_amax import fused_quantize_fp8_w13
    x2_fp8 = fused_quantize_fp8_w13(x_w13, amax_x2, FP8_DTYPE, grad_amax_state=grad_amax_xw13,
@@ -169,15 +164,14 @@ class FlatTransformer:
    self.freqs_cis = precompute_freqs_cis(dim // n_heads, max_context * 2, rope_theta).clone().is_param_(False)

    def _amax(): return Tensor.full((), FP8_MAX, dtype=dtypes.float32).contiguous().is_param_(False)
-    n_amax = 0 if MXFP4 else n_layers
    names = ["xqkv", "xo", "x2"]
    names += ["x1", "x3"] if SPLIT_W13 else ["x13"]
-    self._fp8_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
-    self._fp8_next_amax = {name: [_amax() for _ in range(n_amax)] for name in names}
+    self._fp8_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
+    self._fp8_next_amax = {name: [_amax() for _ in range(n_layers)] for name in names}
    grad_names = ["xqkv", "xo", "xout"]
    grad_names += ["xw1", "xw3"] if SPLIT_W13 else ["xw13"]
-    self._fp8_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
-    self._fp8_next_grad_amax = {name: [_amax() for _ in range(n_amax)] for name in grad_names}
+    self._fp8_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
+    self._fp8_next_grad_amax = {name: [_amax() for _ in range(n_layers)] for name in grad_names}
    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 if MXFP8 else s.float()).contiguous().is_param_(False) for name, s in w_scales}
@@ -201,10 +195,9 @@ class FlatTransformer:
    return (w * scale_b).clamp(-FP8_MAX, FP8_MAX).cast(FP8_DTYPE), inv_scale

  def attention(self, x:Tensor, freqs_cis:Tensor, *, attention_norm:Tensor, wqkv:Tensor, wo:Tensor,
-                amax_xqkv:Tensor|None, amax_xo:Tensor|None, s_qkv:Tensor, s_o:Tensor,
-                next_amax_xqkv:Tensor|None, next_amax_xo:Tensor|None,
-                grad_amax_xqkv:Tensor|None, grad_amax_xo:Tensor|None,
-                next_grad_amax_xqkv:Tensor|None, next_grad_amax_xo:Tensor|None):
+                amax_xqkv:Tensor, amax_xo:Tensor, s_qkv:Tensor, s_o:Tensor,
+                next_amax_xqkv:Tensor, next_amax_xo:Tensor,
+                grad_amax_xqkv:Tensor, grad_amax_xo:Tensor, next_grad_amax_xqkv:Tensor, next_grad_amax_xo:Tensor):
    bsz, seqlen, _ = x.shape
    saves = []

@@ -326,33 +319,28 @@ class FlatTransformer:
          for i in range(len(amax_dict[name])):
            amax_dict[name][i] = amax_dict[name][i].to(device).contiguous().is_param_(False)

-  def reset_amax(self):
-    for st in (self._fp8_next_amax, self._fp8_next_grad_amax):
-      for ts in st.values():
-        for t in ts: t.assign(0)
-
-  def update_amax(self):
-    for cur, nxt in ((self._fp8_amax, self._fp8_next_amax), (self._fp8_grad_amax, self._fp8_next_grad_amax)):
-      for name in cur:
-        for c, n in zip(cur[name], nxt[name]): c.assign(n)
-
  def __call__(self, tokens:Tensor, save:bool=True):
    h = self.tok_embeddings(tokens)
    freqs_cis = self.freqs_cis.cast(h.dtype)
    if not getenv("HK_FLASH_ATTENTION"): freqs_cis = freqs_cis[:, :tokens.shape[1], :, :, :]
    a, na, ga, nga, s = self._fp8_amax, self._fp8_next_amax, self._fp8_grad_amax, self._fp8_next_grad_amax, self._fp8_inv_scale
-    def amax_kwargs(i:int, act_names:tuple[str, ...], grad_names:tuple[str, ...]) -> dict[str, Tensor|None]:
-      specs = (("amax_", a, act_names), ("next_amax_", na, act_names), ("grad_amax_", ga, grad_names), ("next_grad_amax_", nga, grad_names))
-      if MXFP4: return dict.fromkeys(f"{prefix}{name}" for prefix, _, names in specs for name in names)
-      return {f"{prefix}{name}":val[name][i] for prefix, val, names in specs for name in names}
    for i in range(self.n_layers):
-      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wo=self.wo[i], s_qkv=s["wqkv"][i], s_o=s["wo"][i],
-                         **amax_kwargs(i, ("xqkv", "xo"), ("xqkv", "xo")))
-      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], w2=self.w2[i], s_2=s["w2"][i], **amax_kwargs(i, ("x2",), ("xout",)))
+      attn_kwargs = dict(attention_norm=self.attention_norm[i], wqkv=self.wqkv[i], wo=self.wo[i],
+                         amax_xqkv=a["xqkv"][i], amax_xo=a["xo"][i], s_qkv=s["wqkv"][i], s_o=s["wo"][i],
+                         next_amax_xqkv=na["xqkv"][i], next_amax_xo=na["xo"][i],
+                         grad_amax_xqkv=ga["xqkv"][i], grad_amax_xo=ga["xo"][i],
+                         next_grad_amax_xqkv=nga["xqkv"][i], next_grad_amax_xo=nga["xo"][i])
+      ffn_kwargs = dict(ffn_norm=self.ffn_norm[i], w2=self.w2[i],
+                        amax_x2=a["x2"][i], s_2=s["w2"][i], grad_amax_xout=ga["xout"][i], next_grad_amax_xout=nga["xout"][i],
+                        next_amax_x2=na["x2"][i])
      if SPLIT_W13:
-        ffn_kwargs.update(w1=self.w1[i], w3=self.w3[i], s_1=s["w1"][i], s_3=s["w3"][i], **amax_kwargs(i, ("x1", "x3"), ("xw1", "xw3")))
+        ffn_kwargs.update(w1=self.w1[i], w3=self.w3[i], amax_x1=a["x1"][i], amax_x3=a["x3"][i],
+                          next_amax_x1=na["x1"][i], next_amax_x3=na["x3"][i],
+                          s_1=s["w1"][i], s_3=s["w3"][i], grad_amax_xw1=ga["xw1"][i], grad_amax_xw3=ga["xw3"][i],
+                          next_grad_amax_xw1=nga["xw1"][i], next_grad_amax_xw3=nga["xw3"][i])
      else:
-        ffn_kwargs.update(w13=self.w13[i], s_13=s["w13"][i], **amax_kwargs(i, ("x13",), ("xw13",)))
+        ffn_kwargs.update(w13=self.w13[i], amax_x13=a["x13"][i], s_13=s["w13"][i], grad_amax_xw13=ga["xw13"][i],
+                          next_grad_amax_xw13=nga["xw13"][i], next_amax_x13=na["x13"][i])
      h, *_ = self.run_layer(h, freqs_cis, attn_kwargs, ffn_kwargs, save=save)

    logits = matmul(self.norm(h), self.output[0], fp8=False)[0]
@@ -436,7 +424,9 @@ if __name__ == "__main__":
  @TinyJit
  def fwd_bwd(tokens:Tensor):
    with Timing("python forward: "):
-      model.reset_amax()
+      for amax_dict in (model._fp8_next_amax, model._fp8_next_grad_amax):
+        for ts in amax_dict.values():
+          for nxt in ts: nxt.assign(0)
      logits = model(tokens[:, :-1], save=llama_size=="8B")
      loss = vocab_mask.where(-1e9, logits).sparse_categorical_crossentropy(tokens[:, 1:])
    with Timing("python backward: "):
@@ -12,7 +12,7 @@ from tinygrad.helpers import Timing, colored, GlobalCounters, profile_marker
 from tinygrad.uop.ops import Ops, UOp
 from extra.models.llama import apply_rotary_emb
 from extra.llama_kernels.rmsnorm import rmsnorm
-from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8, asm_gemm, can_use_asm_gemm
+from extra.gemm.cdna_asm_gemm import _mx_block_scale, _mx_block_scale_3d, quantize_mxfp8
 from extra.gemm.moe_gemm import grouped_mx_gemm
 from extra.gemm.moe_routing import route, dispatch, combine

@@ -182,12 +182,12 @@ class GPTOSS:
    xq, xk = apply_rotary_emb(xq, xk, freqs_cis)
    xq, xk, xv = xq.cast(dtypes.bfloat16), xk.cast(dtypes.bfloat16), xv.cast(dtypes.bfloat16)  # (B,N,H,D)/(B,N,KV,D)

-    if getenv("HK_FLASH_ATTENTION"):
-      from extra.thunder.amd.fa import flash_attention
-      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks, window=self.sliding_window if sliding else 0)
-      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
-    elif sliding:
+    if sliding:
      attn = self._sliding_attention(xq, xk, xv, sinks)
+    elif getenv("HK_FLASH_ATTENTION"):
+      from extra.thunder.amd.fa import flash_attention
+      attn, *_ = flash_attention(xq, xk, xv, is_causal=True, write_flat=True, sinks=sinks)
+      attn = attn.reshape(bsz, seqlen, self.n_heads * self.head_dim)
    else:
      xqm = xq.reshape(bsz, seqlen, self.n_kv_heads, self.n_rep, self.head_dim).permute(0, 2, 3, 1, 4)
      xkm, xvm = xk.permute(0, 2, 1, 3).unsqueeze(2), xv.permute(0, 2, 1, 3).unsqueeze(2)
@@ -263,11 +263,7 @@ class GPTOSS:
                        w_down=self.w_down[i], w_down_scale=self.w_down_scale[i], w_down_bias=self.w_down_bias[i])
      h, *_ = self.run_layer(h, freqs_cis, mask_full, i % 2 == 0, attn_kwargs, ffn_kwargs, save=save)

-    h_normed = self.norm(h)
-    pad = (-self.dim) % 256
-    h_padded, w_padded = h_normed.pad((None, None, (0, pad))), self.output.pad(((0, 0), (0, pad)))
-    if ASM_GEMM and can_use_asm_gemm(h_padded, w_padded.T): logits = asm_gemm(h_padded, w_padded.T)
-    else: logits = h_normed @ self.output.T
+    logits = self.norm(h) @ self.output.T
    return logits

 def _get_pads(uop:UOp) -> list[UOp]:
@@ -1,8 +1,8 @@
 #!/usr/bin/env bash

 export PYTHONPATH="."
-export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
-export PATH="$ROCM_PATH/bin:$PATH"
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -16,7 +16,7 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export MXFP4=${MXFP4:-1}
+export FP8=${FP8:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
 export FAST_CE=${FAST_CE:-1}
 export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -46,7 +46,7 @@ export DATA_SEED=${DATA_SEED:-5760}
 export JITBEAM=${JITBEAM:-3}
 export BEAM_UOPS_MAX=6000 BEAM_UPCAST_MAX=256 BEAM_LOCAL_MAX=1024 BEAM_MIN_PROGRESS=5 BEAM_PADTO=1

-export FAKEDATA=${FAKEDATA:-$([[ "$DEV" == NULL:* ]] && echo 1 || echo 0)} BENCHMARK=${BENCHMARK:-10}
+export FAKEDATA=${FAKEDATA:-1} BENCHMARK=${BENCHMARK:-10}
 if [ -z "$FULL_LAYERS" ]; then
  export LLAMA_LAYERS=${LLAMA_LAYERS:-2}
 fi
@@ -1,8 +1,8 @@
 #!/usr/bin/env bash

 export PYTHONPATH="."
-export ROCM_PATH=${ROCM_PATH:-/opt/rocm-7.1.1}
-export PATH="$ROCM_PATH/bin:$PATH"
+export PATH="/opt/rocm-7.1.1/bin:$PATH"
+export ROCM_PATH="/opt/rocm-7.1.1"
 export DEV=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -16,7 +16,7 @@ export USE_ATOMICS=${USE_ATOMICS:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
 export WQKV=${WQKV:-1}
 export MASTER_WEIGHTS=${MASTER_WEIGHTS:-1}
-export MXFP4=${MXFP4:-1}
+export FP8=${FP8:-1}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
 export FAST_CE=${FAST_CE:-1}
 export FUSED_INPUT_QUANTIZE=${FUSED_INPUT_QUANTIZE:-1}
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${FUSED_SILU_W13:-1}
 export SPLIT_W13=${SPLIT_W13:-0}
 export OFFLOAD_OPTIM=${OFFLOAD_OPTIM:-0}

-export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="float32"
+export DEFAULT_FLOAT="bfloat16" OPTIM_DTYPE="bfloat16"
 export DP=${DP:-8} MP=${MP:-1} BS=${BS:-16} EVAL_BS=${EVAL_BS:-8} GRADIENT_ACC_STEPS=${GRADIENT_ACC_STEPS:-2}
 export GBS=$((BS * GRADIENT_ACC_STEPS))

@@ -1,6 +1,4 @@
 #!/bin/bash
-set -e
-
 export BENCHMARK=${BENCHMARK:-5}
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
@@ -10,7 +10,6 @@ export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -10,7 +10,6 @@ export DEVICE_IN_FUNCTION_BUG=1

 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
-export ASM_GEMM=${ASM_GEMM:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}
@@ -1,31 +0,0 @@
-import argparse, time
-from tinygrad.llm.model import Transformer
-
-if __name__ == "__main__":
-  parser = argparse.ArgumentParser()
-  parser.add_argument("--model", required=True, help="path to gguf model")
-  parser.add_argument("--max-context", type=int, default=8192, help="max context length (default: %(default)s)")
-  parser.add_argument("--prompt-tokens", type=int, default=1024, help="number of prompt tokens (default: %(default)s)")
-  parser.add_argument("--decode-tokens", type=int, default=16, help="number of tokens to decode (default: %(default)s)")
-  parser.add_argument("--chunk-size", type=int, default=32, help="chunk size for prefill (default: %(default)s)")
-  args = parser.parse_args()
-
-  st = time.perf_counter()
-  model, _ = Transformer.from_gguf(args.model, args.max_context)
-  print(f"load {time.perf_counter()-st:.3f}s", flush=True)
-
-  st = time.perf_counter()
-  model.warmup()
-  print(f"warm {time.perf_counter()-st:.3f}s", flush=True)
-
-  prompt = [257] + [1000+i%1000 for i in range(args.prompt_tokens-1)]
-  gen = model.generate(prompt, chunk_size=args.chunk_size)
-  st = time.perf_counter()
-  # first token is time-to-first-token; counted as part of prefill
-  output = [next(gen)]
-  pt = time.perf_counter()
-  print(f"prefill {args.prompt_tokens/(pt-st):.3f} tok/s", flush=True)
-
-  for _ in range(args.decode_tokens): output.append(next(gen))
-  et = time.perf_counter()
-  print(f"decode {args.decode_tokens/(et-pt):.3f} tok/s output {output}", flush=True)
@@ -35,7 +35,7 @@ def compile_net(linear:UOp, output_bufs:List[Buffer]) -> Tuple[Dict[str,str], Li
    return name

  for call in iter_kernel_calls(linear):
-    arg_uops = [b for b in call.src[1:] if not b.is_bound_var]
+    arg_uops = [b for b in call.src[1:] if b.op is not Ops.BIND]
    prg = to_program(call.src[0], Device[arg_uops[0].device].renderer)
    info = prg.arg
    functions[info.function_name] = prg.src[2].arg
@@ -0,0 +1,203 @@
+from tinygrad import Tensor, UOp, getenv
+from tinygrad.uop.ops import AxisType, KernelInfo, Ops
+from tinygrad.dtype import AddrSpace, dtypes
+from tinygrad.helpers import DEBUG, GlobalCounters, Context
+import math
+
+BLOCK_M, BLOCK_N = 64, 64
+WARP_SIZE = 32
+WMMA_M, WMMA_N, WMMA_K = 16, 16, 16
+WAVES_M, WAVES_N = 4, 1
+LANES_PER_WAVE_M, LANES_PER_WAVE_N = 2, 16
+WMMA_ACC = WMMA_M // LANES_PER_WAVE_M
+THREADS_PER_BLOCK = WARP_SIZE * WAVES_M * WAVES_N
+LDS_PAD = 4  # pad LDS rows to reduce bank conflicts
+
+WMMA_ARG = (WMMA_M, WMMA_N, WMMA_K), 'AMD', 32
+LOG2E = math.log2(math.e)
+
+def warp_shfl_xor(val, offset, lane):
+  """Read val from lane ^ offset using ds_bpermute."""
+  idx = ((lane ^ offset) * 4).cast(dtypes.int)
+  if val.op is Ops.INDEX and val.addrspace == AddrSpace.REG: val = val.load()
+  return UOp(Ops.CUSTOM, dtypes.float, (idx, val),
+             arg="__builtin_bit_cast(float, __builtin_amdgcn_ds_bpermute({0}, __builtin_bit_cast(int, {1})))")
+
+def warp_reduce_max(val, lane):
+  """Tree reduce MAX across LANES_PER_WAVE_N=16 lanes."""
+  for offset in [8, 4, 2, 1]:
+    val = UOp(Ops.MAX, dtypes.float, (val, warp_shfl_xor(val, offset, lane)))
+  return val
+
+def warp_reduce_sum(val, lane):
+  """Tree reduce SUM across LANES_PER_WAVE_N=16 lanes."""
+  for offset in [8, 4, 2, 1]:
+    val = val + warp_shfl_xor(val, offset, lane)
+  return val
+
+def amd_flash_attention(o:UOp, q:UOp, k:UOp, v:UOp) -> UOp:
+  # inputs are (B*H, N, D)
+  BH, N, D = q.shape
+  assert N % BLOCK_M == 0 and N % BLOCK_N == 0, f"N={N} must be divisible by BLOCK_M={BLOCK_M} and BLOCK_N={BLOCK_N}"
+  assert D % WMMA_K == 0 and D % LANES_PER_WAVE_N == 0, f"D={D} must be divisible by WMMA_K={WMMA_K} and LANES_PER_WAVE_N={LANES_PER_WAVE_N}"
+  assert BLOCK_M % (WAVES_M * WMMA_M) == 0 and BLOCK_N % LANES_PER_WAVE_N == 0
+  TM = BLOCK_M // (WAVES_M * LANES_PER_WAVE_M)
+  TN = BLOCK_N // (WAVES_N * LANES_PER_WAVE_N)
+  TD = D // (WAVES_N * LANES_PER_WAVE_N)
+  SCALE = 1.0 / math.sqrt(D)
+
+  block_bh = UOp.range(BH, 0, AxisType.GLOBAL)
+  block_m = UOp.range(N // BLOCK_M, 1, AxisType.GLOBAL)
+
+  q = q.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
+  k = k.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
+  v = v.reshape(BH, N//BLOCK_N, BLOCK_N, D)[block_bh]
+  o = o.reshape(BH, N//BLOCK_M, BLOCK_M, D)[block_bh, block_m]
+
+  wave_m = UOp.range(WAVES_M, 2, AxisType.LOCAL)
+  wave_n = UOp.range(WAVES_N, 3, AxisType.LOCAL)
+  lane = UOp.range(WARP_SIZE, -1, AxisType.WARP)
+  tid = (wave_m * WAVES_N + wave_n) * WARP_SIZE + lane
+  lane_m = lane // LANES_PER_WAVE_N
+  lane_n = lane % LANES_PER_WAVE_N
+
+  # LDS allocation: slot 0 = Q then P (shared), slot 1 = K then V
+  # TODO: the memory planner should be able to find this reuse
+  ELEMS_PER_THREAD = BLOCK_M * D // THREADS_PER_BLOCK
+  QP_lds = UOp.placeholder((BLOCK_M, D + LDS_PAD), dtypes.half, slot=0, addrspace=AddrSpace.LOCAL)
+  KV_lds = UOp.placeholder((BLOCK_N, D + LDS_PAD), dtypes.half, slot=1, addrspace=AddrSpace.LOCAL)[:, :D]
+
+  # register state
+  acc = UOp.placeholder((TM, TD), dtypes.float, slot=2, addrspace=AddrSpace.REG)
+  m_i = UOp.placeholder((TM,), dtypes.float, slot=3, addrspace=AddrSpace.REG)
+  l_i = UOp.placeholder((TM,), dtypes.float, slot=4, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0)))
+  m_i = m_i.after(m_i.store(m_i.const_like(-math.inf)))
+  l_i = l_i.after(l_i.store(l_i.const_like(0)))
+
+  # ====== KV tile loop ======
+  n_tile = UOp.range(N // BLOCK_N, 100, AxisType.REDUCE)
+
+  # load Q + K into LDS (Q reloaded each iteration since P overwrites slot 0)
+  Q_lds = QP_lds[:, :D]
+  Q_store = Q_lds.after(n_tile).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
+    q.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
+  K_store = KV_lds.reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
+    k[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
+  # NOTE: no explicit barrier needed, the AFTER on the LOCAL buffers implies it in late codegen
+  Q_lds = Q_lds.after(UOp.group(Q_store, K_store))
+  KV_lds_k = KV_lds.after(UOp.group(Q_store, K_store))
+
+  # -- S = Q @ K^T via WMMA (re-init each n_tile) --
+  S_reg = UOp.placeholder((TM, TN), dtypes.float, slot=6, addrspace=AddrSpace.REG)
+  S_reg = S_reg.after(S_reg.after(n_tile).store(S_reg.const_like(0)))
+  k_qk = UOp.range(D // WMMA_K, 101, AxisType.REDUCE)
+  tm1 = UOp.range(TM // WMMA_ACC, 200)
+  tn1 = UOp.range(TN, 201)
+  S_frag = S_reg.reshape(TM // WMMA_ACC, WMMA_ACC, TN).permute(0, 2, 1)[tm1, tn1]
+  q_frag = Q_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, D // WMMA_K, WMMA_K)[wave_m, tm1, lane_n, k_qk]
+  k_frag = KV_lds_k.reshape(WAVES_N, TN, WMMA_N, D // WMMA_K, WMMA_K)[wave_n, tn1, lane_n, k_qk]
+  qk = UOp.wmma(q_frag, k_frag, S_frag.after(k_qk), *WMMA_ARG)
+  qk_done = S_frag.store(qk).end(tm1, tn1).end(k_qk)
+  S_reg = S_reg.after(qk_done)
+
+  # -- softmax in registers with warp shuffles --
+  S_reg = S_reg.after(S_reg.store(S_reg * SCALE))
+
+  # per-thread local row max over TN=4 elements, then warp reduce across 16 lanes
+  m_ij = UOp.placeholder((TM,), dtypes.float, slot=7, addrspace=AddrSpace.REG)
+  m_ij = m_ij.after(m_ij.after(n_tile).store(m_ij.const_like(-math.inf)))
+  rm2 = UOp.range(TN, 261, AxisType.REDUCE)
+  m_ij = m_ij.after(m_ij.store(m_ij.after(rm2).maximum(S_reg[:, rm2])).end(rm2))
+  # warp reduce max (in-place)
+  ri_w = UOp.range(TM, 270)
+  m_ij = m_ij.after(m_ij[ri_w].store(warp_reduce_max(m_ij[ri_w], lane)).end(ri_w))
+
+  # compute P = exp(S - m_ij) in S_reg
+  S_reg = S_reg.after(S_reg.store(((S_reg - m_ij.reshape(TM, 1).expand(TM, TN)) * LOG2E).exp2()))
+
+  p_local = UOp.placeholder((TM,), dtypes.float, slot=8, addrspace=AddrSpace.REG)
+  p_local = p_local.after(p_local.after(n_tile).store(p_local.const_like(0)))
+  rp2 = UOp.range(TN, 291, AxisType.REDUCE)
+  p_local = p_local.after(p_local.store(p_local.after(rp2) + S_reg[:, rp2]).end(rp2))
+  ri_ws = UOp.range(TM, 295)
+  p_sum = p_local.after(p_local[ri_ws].store(warp_reduce_sum(p_local[ri_ws], lane)).end(ri_ws))
+
+  # write P = exp(S - m_ij) to P_lds (reuses slot 0, Q no longer needed)
+  P_lds = QP_lds[:, :BLOCK_N]
+  P_write = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TN, LANES_PER_WAVE_N)
+  P_write = P_write.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TN)
+  P_store = P_write[tid].store(S_reg.cast(dtypes.half))
+
+  # -- online softmax correction --
+  ri4 = UOp.range(TM, 330)
+  m_new_val = m_i[ri4].maximum(m_ij[ri4])
+  alpha_val = ((m_i[ri4] - m_new_val) * LOG2E).exp2()
+  beta_val = ((m_ij[ri4] - m_new_val) * LOG2E).exp2()
+  rj4 = UOp.range(TD, 331)
+  correction = UOp.group(
+    acc[ri4, rj4].store(alpha_val * acc[ri4, rj4]).end(rj4),
+    l_i[ri4].store(alpha_val * l_i[ri4] + beta_val * p_sum[ri4]),
+    m_i[ri4].store(m_new_val),
+  ).end(ri4)
+  acc = acc.after(correction)
+  l_i = l_i.after(correction)
+  m_i = m_i.after(correction)
+
+  # load V into KV_lds (must wait for QK WMMA to finish reading K from KV_lds)
+  V_store = KV_lds.after(qk_done).reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid].store(
+    v[n_tile].reshape(THREADS_PER_BLOCK, ELEMS_PER_THREAD)[tid])
+  # NOTE: no explicit barrier needed, the AFTER on the LOCAL buffers implies it in late codegen
+  P_lds = P_lds.after(UOp.group(P_store, V_store))
+  KV_lds_v = KV_lds.after(UOp.group(P_store, V_store))
+
+  # -- acc += P @ V via WMMA --
+  k_pv = UOp.range(BLOCK_N // WMMA_K, 400, AxisType.REDUCE)
+  tm2 = UOp.range(TM // WMMA_ACC, 401)
+  tn2 = UOp.range(TD, 402)
+  acc_frag = acc.reshape(TM // WMMA_ACC, WMMA_ACC, TD).permute(0, 2, 1)[tm2, tn2]
+  p_frag = P_lds.reshape(WAVES_M, TM // WMMA_ACC, WMMA_M, BLOCK_N // WMMA_K, WMMA_K)[wave_m, tm2, lane_n, k_pv]
+  v_frag = KV_lds_v.reshape(WAVES_N, TD, WMMA_N, BLOCK_N // WMMA_K, WMMA_K)[wave_n, tn2, lane_n, k_pv]
+  pv = UOp.wmma(p_frag, v_frag, acc_frag.after(k_pv), *WMMA_ARG)
+
+  # end KV tile loop
+  n_tile_end = acc_frag.store(pv).end(tm2, tn2).end(k_pv).end(n_tile)
+  acc = acc.after(n_tile_end)
+  l_i = l_i.after(n_tile_end)
+  m_i = m_i.after(n_tile_end)
+
+  # normalize: acc /= l_i
+  acc = acc.after(acc.store(acc * (1 / l_i).reshape(TM, 1).expand(TM, TD)))
+
+  # store output
+  o = o.reshape(WAVES_M, TM // WMMA_ACC, WMMA_ACC, LANES_PER_WAVE_M, WAVES_N, TD, LANES_PER_WAVE_N)
+  o = o.permute((0, 4, 3, 6, 1, 2, 5)).reshape(THREADS_PER_BLOCK, TM, TD)
+  return o[tid].store(acc).end(wave_m, wave_n, lane).end(block_m, block_bh).sink(arg=KernelInfo(opts_to_apply=()))
+
+if __name__ == "__main__":
+  B, H, N, D = getenv("B", 1), getenv("H", 32), getenv("N", 1024), getenv("D", 64)
+  q = Tensor.rand(B, H, N, D).cast(dtypes.half)
+  k = Tensor.rand(B, H, N, D).cast(dtypes.half)
+  v = Tensor.rand(B, H, N, D).cast(dtypes.half)
+  o = Tensor.empty(B, H, N, D, dtype=dtypes.float)
+  with Context(DEBUG=0): Tensor.realize(q, k, v)
+
+  q_flat, k_flat, v_flat, o_flat = q.reshape(B*H, N, D), k.reshape(B*H, N, D), v.reshape(B*H, N, D), o.reshape(B*H, N, D)
+  NUM_RUNS = getenv("CNT", 5)
+  ets = []
+  with Context(DEBUG=2):
+    for _ in range(NUM_RUNS):
+      GlobalCounters.reset()
+      tst = Tensor.custom_kernel(o_flat, q_flat, k_flat, v_flat, fxn=amd_flash_attention)[0].realize()
+      ets.append(GlobalCounters.time_sum_s)
+  print(f"best time: {min(ets)*1e3:.2f}ms")
+
+  if getenv("VERIFY", 1):
+    with Context(DEBUG=0):
+      ref = q.float().scaled_dot_product_attention(k.float(), v.float()).reshape(B*H, N, D).realize()
+      err = (ref - tst).square().mean().item()
+    print(f"mean squared error {err}")
+    if err > 1e-2:
+      raise RuntimeError("flash attention is wrong!")
+    else:
+      print("flash attention is correct!")
@@ -6,7 +6,6 @@ from tinygrad.renderer import Estimates
 from tinygrad.helpers import getenv, all_same, DEBUG, ceildiv
 from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8
-from extra.llama_kernels.quantize_mxfp4 import quantize_mxfp4

 TILE_M, TILE_N, TILE_K = 256, 256, 64

@@ -122,30 +121,10 @@ def custom_mxfp4_gemm(C:UOp, A:UOp, B:UOp, scale_a:UOp, scale_b:UOp, *extra:UOp,
  groups_x, groups_y = UOp.special(ceildiv(N, tile_n), "gidx0"), UOp.special(ceildiv(M, tile_m), "gidx1")
  lds = UOp.placeholder((163840,), dtypes.uint8, 0, AddrSpace.LOCAL)
  sink = UOp.sink(C.base, A.base, B.base, scale_a.base, scale_b.base, *(x.base for x in extra), lds, threads, groups_x, groups_y,
-                  arg=KernelInfo(f"mxfp4_gemm_{M}_{N}_{K}",
-                                 estimates=Estimates(ops=2*M*N*K, mem=(M*half_k+N*half_k)*A.dtype.itemsize+M*N*C.dtype.itemsize)))
+                  arg=KernelInfo(f"custom_mxfp4_gemm_{M}_{N}_{K}", estimates=Estimates(ops=2*M*N*K)))
  insts = build_kernel(M, N, K, tile_m, tile_n)
  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=x) for x in insts))))

-def _mxfp4_gemm_quantized(a_q:Tensor, b_q:Tensor, scale_a:Tensor, scale_b:Tensor) -> Tensor:
-  M, half_k = a_q.shape
-  N, half_k_b = b_q.shape
-  assert half_k == half_k_b
-  is_multi = isinstance(a_q.device, tuple)
-  reduce_out = is_multi and (a_q.uop.axis == 1 or b_q.uop.axis == 1)
-  if not is_multi: out = Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a_q.device)
-  elif reduce_out: out = Tensor(Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a_q.device).uop.unshard(0), device=a_q.device)
-  elif a_q.uop.axis == 0:
-    out = Tensor(Tensor.invalids(1, M//len(a_q.device), N, dtype=dtypes.bfloat16, device=a_q.device).uop.unshard(1), device=a_q.device)
-  elif b_q.uop.axis == 0:
-    out = Tensor(Tensor.invalids(1, M, N//len(a_q.device), dtype=dtypes.bfloat16, device=a_q.device).uop.unshard(2), device=a_q.device)
-  else: out = Tensor.invalids(1, M, N, dtype=dtypes.bfloat16, device=a_q.device)
-  tile_m, tile_n = next((tm, tn) for tm, tn in ((256, 256), (192, 256), (128, 512)) if M % tm == N % tn == 0)
-  out = Tensor.custom_kernel(out, a_q, b_q, scale_a, scale_b,
-                             fxn=functools.partial(custom_mxfp4_gemm, tile_m=tile_m, tile_n=tile_n))[0]
-  if reduce_out: out = out.sum(0)
-  return out.squeeze(0)
-
 def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
  # 1x32 block scaling along the last axis
  *batch, K = x.shape
@@ -158,6 +137,50 @@ def quantize_mxfp8(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
  packed = mx_pack(e8) if len(batch) == 1 and scale_K % 4 == 0 else None
  return x_clamped.cast(FP8_DTYPE), e8, packed

+def _mxfp4_shuffle_weight(x:Tensor) -> Tensor:
+  # shuffle_weight(x, layout=(16, 16)) on the packed uint8 buffer.
+  if x.ndim == 3:
+    ndev, rows, half_k = x.shape
+    return x.reshape(ndev, rows//16, 16, half_k//32, 2, 16).permute(0, 1, 3, 4, 2, 5).reshape(ndev, rows, half_k).contiguous()
+  rows, half_k = x.shape
+  return x.reshape(rows//16, 16, half_k//32, 2, 16).permute(0, 2, 3, 1, 4).reshape(rows, half_k).contiguous()
+
+def _mxfp4_shuffle_scales(x:Tensor) -> Tensor:
+  # e8m0_shuffle: each 256x8 scale tile is arranged for the raw MFMA scale loads.
+  if x.ndim == 3:
+    ndev, rows, scale_k = x.shape
+    return x.reshape(ndev, rows//32, 2, 16, scale_k//8, 2, 4).permute(0, 1, 4, 6, 3, 5, 2).reshape(ndev, rows, scale_k).contiguous()
+  rows, scale_k = x.shape
+  return x.reshape(rows//32, 2, 16, scale_k//8, 2, 4).permute(0, 3, 5, 2, 4, 1).reshape(rows, scale_k).contiguous()
+
+def quantize_mxfp4(x:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  # OCP MXFP4: 1x32 blocks, E2M1 values packed low-nibble first, and E8M0 scales.
+  *batch, K = x.shape
+  rows = math.prod(batch)
+  assert x.ndim >= 2 and K % 256 == 0 and rows % 32 == 0, \
+    f"mxfp4 quantization needs rows%32 and K%256, got {x.shape}"
+  xb = x.float().reshape(*batch, K//32, 32)
+  amax = xb.abs().max(axis=-1)
+
+  # even scale rounding: round the fp32 significand before choosing 2^(floor(log2)-2).
+  amax_rounded = ((amax.bitcast(dtypes.uint32) + 0x200000) & 0xFF800000).bitcast(dtypes.float32)
+  scale_exp = (amax_rounded.maximum(2**-126).log2().floor() - 2).clamp(-127, 127)
+  e8 = (scale_exp + 127).cast(dtypes.uint8)
+  scaled = xb * (-scale_exp).exp2().reshape(*batch, K//32, 1)
+
+  mag = scaled.abs()
+  code = sum(x.cast(dtypes.uint8) for x in
+             (mag > .25, mag >= .75, mag > 1.25, mag >= 1.75, mag > 2.5, mag >= 3.5, mag > 5.0))
+  code = code | ((scaled < 0).cast(dtypes.uint8) << 3)
+  code = code.reshape(*batch, K)
+  packed = code[..., 0::2] | (code[..., 1::2] << 4)
+  if isinstance(x.device, tuple) and x.uop.axis == x.ndim-2 and x.shape[x.uop.axis] == len(x.device):
+    axis = x.uop.axis
+    order = (axis, *range(axis), *range(axis+1, e8.ndim))
+    e8_local = e8.permute(order)
+    return packed, e8, _mxfp4_shuffle_scales(e8_local.reshape(e8_local.shape[0], -1, K//32))
+  return packed, e8, _mxfp4_shuffle_scales(e8.reshape(rows, K//32))
+
 def mx_pack(e8:Tensor) -> Tensor:
  rows, scale_K = e8.shape
  return e8.reshape(rows, scale_K // 4, 4).bitcast(dtypes.uint32).reshape(rows, scale_K // 4).permute(1, 0).contiguous()
@@ -382,16 +405,15 @@ def custom_mx_gemm_bw(gradient:UOp, kernel:UOp, has_w_post:bool, w_stored:bool=F
 # ** mxfp4 gemm backward

 def custom_mxfp4_gemm_bw(gradient:UOp, kernel:UOp):
-  inputs = kernel.src[1:]  # out, row operands/scales, BF16 operands, column operands/scales
-  assert len(inputs) == 11
+  # The raw kernel consumes quantized buffers, while the final two inputs retain the BF16 operands for STE gradients.
+  inputs = kernel.src[1:]  # (out, a_q, b_q, scale_a, scale_b, a, w)
+  assert len(inputs) == 7
  a, w = Tensor(inputs[5], device=inputs[5].device), Tensor(inputs[6], device=inputs[6].device)
-  a_col, scale_a_col = Tensor(inputs[7], device=a.device), Tensor(inputs[8], device=a.device)
-  w_col, scale_w_col = Tensor(inputs[9], device=a.device), Tensor(inputs[10], device=a.device)
  g = Tensor(gradient, device=a.device)[:a.shape[0]].cast(dtypes.bfloat16)
-  g_row, scale_g_row, g_col, scale_g_col = quantize_mxfp4(g, flatten_row=True)
-  grad_a = _mxfp4_gemm_quantized(g_row, w_col, scale_g_row, scale_w_col).reshape(*a.shape[:-1], w.shape[-1])
-  grad_w = _mxfp4_gemm_quantized(g_col, a_col, scale_g_col, scale_a_col).reshape(w.shape)
-  return (None, None, None, None, None, grad_a.uop, grad_w.uop, None, None, None, None)
+  grad_a = asm_gemm(g, w, mxfp4=True)
+  a_flat, g_flat = a.reshape(-1, a.shape[-1]), g.reshape(-1, g.shape[-1])
+  grad_w = asm_gemm(g_flat.T, a_flat, mxfp4=True)
+  return (None, None, None, None, None, grad_a.uop, grad_w.uop)

 # ** main gemm function

@@ -437,10 +459,16 @@ def asm_gemm(a:Tensor, b:Tensor, x_scale:Tensor|None=None, w_scale:Tensor|None=N
      tile_m, tile_n = next((tm, tn) for tm, tn in ((256, 256), (192, 256), (128, 512)) if (batch*M) % tm == N % tn == 0)
      fxn = functools.partial(custom_mxfp4_gemm, tile_m=tile_m, tile_n=tile_n)
      w = b.T
-      a_q, scale_a, a_col, scale_a_col = quantize_mxfp4(a, shuffle_col=True)
-      b_q, scale_b, b_col, scale_b_col = quantize_mxfp4(w, shuffle_row=True, shuffle_col=True)
-      out = Tensor.custom_kernel(out, a_q, b_q, scale_a, scale_b, a, w,
-                                 a_col, scale_a_col, b_col, scale_b_col, fxn=fxn, grad_fxn=custom_mxfp4_gemm_bw)[0]
+      if k_sharded:
+        ndev = len(a.device)
+        a_q, _, scale_a = quantize_mxfp4(a.reshape(batch, M, ndev, K))
+        b_q, _, scale_b = quantize_mxfp4(w.reshape(w.shape[0], ndev, K))
+        b_q = _mxfp4_shuffle_weight(b_q.permute(1, 0, 2))
+      else:
+        a_q, _, scale_a = quantize_mxfp4(a.reshape(batch*M, K))
+        b_q, _, scale_b = quantize_mxfp4(w)
+        a_q, b_q = a_q.reshape(batch, M, K//2).contiguous(), _mxfp4_shuffle_weight(b_q)
+      out = Tensor.custom_kernel(out, a_q, b_q, scale_a, scale_b, a, w, fxn=fxn, grad_fxn=custom_mxfp4_gemm_bw)[0]
    elif mx:
      # mxfp8 1x32 block scaling
      if mx_scales is not None:
@@ -5,8 +5,7 @@ BLOCK_ROW = 256

 def _sharded_invalids(shape:tuple[int, ...], dtype, device) -> Tensor:
  if isinstance(device, tuple):
-    per = Tensor.invalids(shape[0]//len(device), *shape[1:], dtype=dtype, device=device)
-    return Tensor(per.uop.unshard(0), device=device)
+    return Tensor(Tensor.invalids(shape[0] // len(device), *shape[1:], dtype=dtype, device=device).uop.multi(0), device=device)
  return Tensor.invalids(*shape, dtype=dtype, device=device)

 def _atomic_add(device:str) -> str:
@@ -79,7 +79,7 @@ if __name__ == "__main__":
  linear, var_vals = C.linear_with_vars()
  last_call = linear.src[-1]
  ast = last_call.src[0]
-  bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]
+  bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]

  src = compiled.asm["ptx"]
  # specify the shared memory here so we don't need to do it dynamically
@@ -3,8 +3,8 @@ from typing import cast, Any, Callable
 import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit
 assert sys.platform != 'win32'
 from dataclasses import dataclass
-from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, encode_kernargs_clike, make_cmdbuf
-from tinygrad.runtime.support.hcq2 import make_binary_patch
+from tinygrad.runtime.support.hcq2 import HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_cmdbuf
+from tinygrad.runtime.support.hcq2 import make_binary_patch, make_patches
 from tinygrad.uop.ops import sint, UOp
 from tinygrad.device import Compiled, BufferSpec, Buffer, Device
 from tinygrad.dtype import dtypes
@@ -152,12 +152,19 @@ def pm4_submit(ctx, lin):
  ring, wptr, doorbell, put_ptr = (UOp.placeholder((b.size,), b.dtype, 0, device=devs).rtag(f"COMPUTE:0_{name}")
    for name, b in (("ring", q.ring), ("write_ptr", q.write_ptr), ("doorbell", q.doorbell), ("put_value", q.put_value)))

-  # the host fence at the start of the batch guarantees the ib is free to reuse
-  size_dw = sum(len(ins.src) for ins in lin.src)
+  # two tail dwords coordinate safe IB reuse: GPU completions and host submits
+  size_dw = sum(len(ins.src) for ins in lin.src) + len(release_mem(ctx, 0, 0).src)
  assert size_dw < (1 << 20), f"indirect buffer of {size_dw} dwords doesn't fit one packet"

-  ib = UOp.placeholder((size_dw,), dtypes.uint32, next(UOp.unique_num), device=devs, volatile=True).rtag("cmdbuf")
-  cmdbuf = make_cmdbuf(lin, devs, buf=ib)
+  ib = UOp.placeholder((size_dw + 2,), dtypes.uint32, next(UOp.unique_num), device=devs, volatile=True).rtag("cmdbuf")
+  done_idx, submit_idx = UOp.const(size_dw + 0, dtypes.int), UOp.const(size_dw + 1, dtypes.int)
+  init_counters = make_patches(ib, [((size_dw + i) * 4, UOp.const(0, dtypes.uint32)) for i in range(2)]).rtag("link")
+  submitted = (counter:=ib.after(init_counters).index(submit_idx)).load()
+  completed = ib.after(loop:=UOp.loop(0)).index(done_idx).load()
+  ib_free = completed.end(loop, completed != submitted)
+
+  bump_fence = pm4_store(ctx, UOp(Ops.SLICE, dtypes.uint32, (ib, UOp.const(size_dw)), 2), (submitted + 1).cast(dtypes.uint64))
+  cmdbuf = make_cmdbuf(lin.replace(src=lin.src + (bump_fence,)), devs, buf=ib, dep=ib_free)

  # the ring itself only carries a packet pointing at the ib, wrapping the ring
  put = put_ptr.index(zero:=UOp.const(0, dtypes.int))
@@ -167,7 +174,7 @@ def pm4_submit(ctx, lin):
  # advance the put/write pointers past the packet
  bump_put_ptr = put_ptr.index(zero).store(put + len(pkt))
  bump_wptr = wptr.index(zero).store(put + len(pkt))
-  flush = UOp.barrier(write_pkt, bump_put_ptr, bump_wptr)
+  flush = UOp.barrier(write_pkt, bump_put_ptr, bump_wptr, counter.store(submitted + 1))
  return doorbell.after(flush).index(zero).store(put + len(pkt))

 pm_pm4_submit = PatternMatcher([(UPat(Ops.LINEAR, name="lin"), pm4_submit)])
@@ -182,8 +189,7 @@ def sdma_copy(ctx, call):
  src_addr, dst_addr = call.src[2].getaddr(ctx.devs), call.src[1].getaddr(ctx.devs)
  return call.ins(SDMAOps.COPY, src=tuple(UOp.const(x, dtypes.uint32) for off in range(0, sz, ctx.max_copy_size) for x in (
    ctx.sdma.SDMA_OP_COPY | ctx.sdma.SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_COPY_LINEAR),
-    ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz-off, ctx.max_copy_size)-1), 0,
-    *data64_le(src_addr+UOp.const(off, dtypes.uint64)), *data64_le(dst_addr+UOp.const(off, dtypes.uint64)))))
+    ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz-off, ctx.max_copy_size)-1), 0, *data64_le(src_addr+off), *data64_le(dst_addr+off))))

 def sdma_wait(ctx, ins, dst, val):
  op = ctx.sdma.SDMA_OP_POLL_REGMEM | ctx.sdma.SDMA_PKT_POLL_REGMEM_HEADER_FUNC(WAIT_REG_MEM_FUNCTION_GEQ) \
@@ -288,16 +294,14 @@ def amd_build_program(prg:UOp) -> UOp:

 class AMDAllocator(HCQAllocator['AMDDevice']):
  def __init__(self, dev:AMDDevice):
-    super().__init__(dev, supports_copy_from_disk=dev.has_copy_queue, supports_transfer=dev.has_copy_queue and not dev.is_usb())
+    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) -> HCQBuffer:
-    return self.dev.iface.alloc(size, host=options.host, uncached=options.uncached, cpu_access=options.cpu_access or not self.dev.has_copy_queue)
+  def _alloc(self, size:int, options:BufferSpec) -> HCQ2Buffer:
+    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)

-  def _do_map(self, buf:HCQBuffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)
-
-  def _do_unmap(self, buf:HCQBuffer): self.dev.iface.unmap(buf)
+  def _do_map(self, buf:HCQ2Buffer): return self.dev.iface.map(buf._base if buf._base is not None else buf)

@dataclass
 class AMDQueueDesc:
@@ -390,24 +394,15 @@ class KFDIface:
    return hcqbuf

  def free(self, mem):
-    self._unmap(mem)
-    if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
-    kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
-
-  def unmap(self, mem):
-    self._unmap(mem)
-    if getattr(mem, '_owns_kfd_handle', False): kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)
-
-  def _unmap(self, mem):
    gpus = (ctypes.c_int32 * 1)(self.gpu_id)
    stm = kfd.AMDKFD_IOC_UNMAP_MEMORY_FROM_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(gpus), n_devices=1)
    assert stm.n_success == 1
+    if mem.owner == self.dev:
+      if mem.va_addr: FileIOInterface.munmap(mem.va_addr, mem.size)
+      kfd.AMDKFD_IOC_FREE_MEMORY_OF_GPU(self.kfd, handle=mem.meta.handle)

  def map(self, mem):
-    if mem.owner is not None and mem.owner._is_cpu():
-      mapped = self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)
-      mapped._owns_kfd_handle = True
-      return mapped
+    if mem.owner is not None and mem.owner._is_cpu(): return self.alloc(mem.size, host=True, cpu_addr=mem.va_addr)

    c_gpus = (ctypes.c_int32 * 1)(self.gpu_id)
    stm = kfd.AMDKFD_IOC_MAP_MEMORY_TO_GPU(self.kfd, handle=mem.meta.handle, device_ids_array_ptr=ctypes.addressof(c_gpus), n_devices=1)
@@ -479,7 +474,6 @@ class PCIIface(PCIIfaceBase):

  def require_profile_mode(self): return True
  def is_wgp_active(self, xcc, se, sa, wgp) -> bool: return True # TODO: account for WGP disablement on some asics.
-  def unmap(self, mem): self.free(mem)

  def _compute_props(self):
    self.ip_versions = self.dev_impl.ip_ver
@@ -520,12 +514,12 @@ class PCIIface(PCIIfaceBase):
    if drain_only: d.iface.dev_impl.ih.drain()
    else: d.iface.dev_impl.ih.interrupt_handler()

-    if reset and d.iface.dev_impl.recover(force=True):
+    if reset and d.iface.dev_impl.recover():
      cq = d.compute_queue
      for b in (cq.put_value, cq.read_ptr, cq.write_ptr): b._buf.view.view(fmt='Q')[0] = 0
      d.iface.dev_impl.gfx.setup_ring(*cq.params)
-      d.signal('timeline')._buf.cpu_view().mv.cast('Q')[0] = \
-        d.signal('value', 1).as_memoryview(force_zero_copy=True, no_sync=True).cast('Q')[0] - 1
+      d.timeline_signal('COMPUTE:0')._buf.cpu_view().mv.cast('Q')[0] = \
+        d.timeline_value('COMPUTE:0').as_memoryview(force_zero_copy=True).cast('Q')[0] - 1

  def sleep(self, timeout):
    if hasattr(self.pci_dev, 'irq_poller') and self.pci_dev.irq_poller is not None and (events_cnt:=len(self.pci_dev.irq_poller.poll(timeout))):
@@ -551,12 +545,9 @@ class AMDDevice(HCQ2Compiled):
  ])

  timestamp_divider = 100.0  # AMD GPU clock: ticks/us
-  max_scratch_psize = 0

  ifaces = [KFDIface, PCIIface, _mock(KFDIface, "MOCKIface"), _mock(KFDIface), _mock(PCIIface)]

-  def device_props(self): return self.iface.props
-
  def is_am(self) -> bool: return isinstance(self.iface, (PCIIface,))
  def is_usb(self) -> bool: return False

@@ -593,7 +584,7 @@ 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_copy_queue = not getenv("AMD_DISABLE_SDMA")
+    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, can_recover=self.is_am(), arch=self.arch)

@@ -648,6 +639,9 @@ class AMDDevice(HCQ2Compiled):
    qname = f"{'COPY' if queue_type == kfd.KFD_IOC_QUEUE_TYPE_SDMA else 'COMPUTE'}:{idx}"
    self.pm_bufferize = PatternMatcher([
      (UPat(Ops.PARAM, tag=f"{qname}_{name}"), lambda ctx, b=getattr(queue, name): b) for name in ["ring", "write_ptr", "doorbell", "put_value"]
+    ] + [
+      (UPat(Ops.PARAM, tag=f"{qname}_timeline_signal"), lambda ctx, q=qname: ctx[0].timeline_signal(q)),
+      (UPat(Ops.PARAM, tag=f"{qname}_timeline_value"), lambda ctx, q=qname: ctx[0].timeline_value(q)),
    ]) + self.pm_bufferize

    return queue
@@ -706,7 +700,7 @@ class AMDDevice(HCQ2Compiled):
    return tmpring

  def scratch_buffer(self, private_segment_size):
-    AMDDevice.max_scratch_psize = private_segment_size = max(private_segment_size, 128, AMDDevice.max_scratch_psize)
+    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
@@ -1,36 +0,0 @@
-import functools, math, pathlib
-from tinygrad import Tensor, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo
-from tinygrad.renderer import Estimates
-from extra.llama_kernels import alloc_like, compile_hip
-
-@functools.cache
-def _custom_quantize_mxfp4(row_fp4:UOp, row_scale:UOp, col_fp4:UOp, col_scale:UOp, x:UOp, *, shuffle_row:bool, shuffle_col:bool) -> UOp:
-  M, N = math.prod(x.shape[:-1]), x.shape[-1]
-  assert M % 256 == 0 and N % 256 == 0, f"MXFP4 quantization requires multiples of 256, got {x.shape}"
-  name = f"quantize_mxfp4_{int(shuffle_row)}_{int(shuffle_col)}_{M}_{N}"
-  mem = M*N*2 + M*N + M*N//16  # read bf16, write row+col fp4 + e8m0
-  outputs = (row_fp4, row_scale, col_fp4, col_scale)
-  sink = UOp.sink(*(o.base for o in outputs), x.base,
-                  *(UOp(Ops.CUSTOM, dtypes.void, (o.base.index(0),), arg="") for o in outputs),
-                  UOp.special(256, "lidx0"), UOp.special(M//128, "gidx0"), UOp.special(N//64, "gidx1"),
-                  arg=KernelInfo(name, estimates=Estimates(ops=12*M*N, mem=mem)))
-  src = (pathlib.Path(__file__).parent/"quantize_mxfp4.cpp").read_text()
-  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src),
-    UOp(Ops.BINARY, arg=compile_hip(src, [f"-DKERNEL_NAME={name}", f"-DM_DIM={M}", f"-DN_DIM={N}",
-                                             f"-DSHUFFLE_ROWWISE_FP4_VALUE={int(shuffle_row)}",
-                                             f"-DSHUFFLE_COLWISE_FP4_VALUE={int(shuffle_col)}"]))))
-
-def quantize_mxfp4(x:Tensor, *, shuffle_row:bool=False, shuffle_col:bool=False, flatten_row:bool=False) -> tuple[Tensor, Tensor, Tensor, Tensor]:
-  assert x.dtype == dtypes.bfloat16 and x.ndim >= 2, f"expected BF16 matrix, got {x.dtype} {x.shape}"
-  M, N = math.prod(x.shape[:-1]), x.shape[-1]
-  assert M % 256 == 0 and N % 256 == 0, f"MXFP4 quantization requires multiples of 256, got {x.shape}"
-  axis = x.uop.axis if isinstance(x.device, tuple) else None
-  row_axis = 0 if flatten_row and axis is not None else axis
-  col_axis = None if axis is None else (0 if axis == x.ndim-1 else 1)
-  outputs = (alloc_like((M, N//2) if flatten_row else (*x.shape[:-1], N//2), dtypes.uint8, x.device, row_axis),
-             alloc_like((M, N//32) if flatten_row else (*x.shape[:-1], N//32), dtypes.uint8, x.device, row_axis),
-             alloc_like((N, M//2), dtypes.uint8, x.device, col_axis),
-             alloc_like((N, M//32), dtypes.uint8, x.device, col_axis))
-  fxn = functools.partial(_custom_quantize_mxfp4, shuffle_row=shuffle_row, shuffle_col=shuffle_col)
-  return tuple(Tensor.custom_kernel(*outputs, x, fxn=fxn)[:4])
@@ -1,226 +0,0 @@
-// Copyright (c) 2025-2026, Advanced Micro Devices, Inc. All rights reserved.
-// SPDX-License-Identifier: MIT
-
-#include <hip/hip_runtime.h>
-#include <cstdint>
-
-#if !defined(KERNEL_NAME) || !defined(M_DIM) || !defined(N_DIM) || !defined(SHUFFLE_ROWWISE_FP4_VALUE) || \
-    !defined(SHUFFLE_COLWISE_FP4_VALUE)
-#error kernel dimensions and layouts must be defined
-#endif
-
-namespace {
-
-constexpr int BLOCK = 32;
-constexpr int TILE_M = 128;
-constexpr int TILE_N = 64;
-constexpr int THREADS = 256;
-constexpr int THREADS_PER_ROW = 8;
-constexpr int VALUES_PER_THREAD = 4;
-constexpr int SMEM_STRIDE = BLOCK + 2;
-constexpr int M = M_DIM;
-constexpr int N = N_DIM;
-constexpr int M_PACKED = M / 2;
-constexpr int N_PACKED = N / 2;
-constexpr int M_SCALES = M / BLOCK;
-constexpr int N_SCALES = N / BLOCK;
-constexpr bool SHUFFLE_ROWWISE_FP4 = SHUFFLE_ROWWISE_FP4_VALUE;
-constexpr bool SHUFFLE_COLWISE_FP4 = SHUFFLE_COLWISE_FP4_VALUE;
-
-static_assert(M % 256 == 0 && N % 256 == 0);
-
-struct Quantized4 {
-  uint16_t fp4;
-  uint8_t scale;
-};
-
-__device__ __forceinline__ float swizzle_xor1(float value) {
-  float result;
-  asm volatile("ds_swizzle_b32 %0, %1 offset:0x041f\n\ts_waitcnt lgkmcnt(0)" : "=v"(result) : "v"(value));
-  return result;
-}
-
-__device__ __forceinline__ float swizzle_xor2(float value) {
-  float result;
-  asm volatile("ds_swizzle_b32 %0, %1 offset:0x081f\n\ts_waitcnt lgkmcnt(0)" : "=v"(result) : "v"(value));
-  return result;
-}
-
-__device__ __forceinline__ float swizzle_xor4(float value) {
-  float result;
-  asm volatile("ds_swizzle_b32 %0, %1 offset:0x101f\n\ts_waitcnt lgkmcnt(0)" : "=v"(result) : "v"(value));
-  return result;
-}
-
-__device__ __forceinline__ float max8(float value) {
-  value = fmaxf(value, swizzle_xor4(value));
-  value = fmaxf(value, swizzle_xor2(value));
-  return fmaxf(value, swizzle_xor1(value));
-}
-
-__device__ __forceinline__ float4 load_bf16x4(const uint16_t* values) {
-  const uint32_t lo = *reinterpret_cast<const uint32_t*>(values);
-  const uint32_t hi = *reinterpret_cast<const uint32_t*>(values + 2);
-  return make_float4(__uint_as_float(lo << 16), __uint_as_float(lo & 0xffff0000u),
-                     __uint_as_float(hi << 16), __uint_as_float(hi & 0xffff0000u));
-}
-
-__device__ __forceinline__ void hadamard16(float4& value, int lane) {
-  const float a0 = value.x + value.y, a1 = value.x - value.y;
-  const float a2 = value.z + value.w, a3 = value.z - value.w;
-  value = make_float4(a0 + a2, a1 + a3, a0 - a2, a1 - a3);
-
-  const float4 xor1 = make_float4(swizzle_xor1(value.x), swizzle_xor1(value.y), swizzle_xor1(value.z), swizzle_xor1(value.w));
-  value = lane & 1 ? make_float4(xor1.x - value.x, xor1.y - value.y, xor1.z - value.z, xor1.w - value.w)
-                   : make_float4(xor1.x + value.x, xor1.y + value.y, xor1.z + value.z, xor1.w + value.w);
-
-  const float4 xor2 = make_float4(swizzle_xor2(value.x), swizzle_xor2(value.y), swizzle_xor2(value.z), swizzle_xor2(value.w));
-  value = lane & 2 ? make_float4(xor2.x - value.x, xor2.y - value.y, xor2.z - value.z, xor2.w - value.w)
-                   : make_float4(xor2.x + value.x, xor2.y + value.y, xor2.z + value.z, xor2.w + value.w);
-  value.x *= 0.25f;
-  value.y *= 0.25f;
-  value.z *= 0.25f;
-  value.w *= 0.25f;
-}
-
-__device__ __forceinline__ uint8_t e8m0_scale(float amax, float& scale) {
-  if (amax == 0.0f) {
-    scale = 1.0f;
-    return 127;
-  }
-
-  const uint32_t rounded = (__float_as_uint(amax) + 0x200000u) & 0xff800000u;
-  int exponent = static_cast<int>((rounded >> 23) & 0xff) - 129;
-  exponent = exponent < -127 ? -127 : exponent > 127 ? 127 : exponent;
-  scale = exponent == -127 ? __uint_as_float(0x00400000u) : __uint_as_float(static_cast<uint32_t>(exponent + 127) << 23);
-  return static_cast<uint8_t>(exponent + 127);
-}
-
-__device__ __forceinline__ uint16_t pack_fp4(float4 value, float scale) {
-  uint32_t lo = 0, hi = 0;
-  asm volatile("v_cvt_scalef32_pk_fp4_f32 %0, %1, %2, %3" : "+v"(lo) : "v"(value.x), "v"(value.y), "v"(scale));
-  asm volatile("v_cvt_scalef32_pk_fp4_f32 %0, %1, %2, %3" : "+v"(hi) : "v"(value.z), "v"(value.w), "v"(scale));
-  return static_cast<uint16_t>(lo | (hi << 8));
-}
-
-__device__ __forceinline__ Quantized4 quantize(float4 value, int lane) {
-  hadamard16(value, lane);
-  const float local_max = fmaxf(fmaxf(fabsf(value.x), fabsf(value.y)), fmaxf(fabsf(value.z), fabsf(value.w)));
-  float scale;
-  const uint8_t e8m0 = e8m0_scale(max8(local_max), scale);
-  return {pack_fp4(value, scale), e8m0};
-}
-
-__device__ __forceinline__ void store_scale(uint8_t* output, int row, int col, int cols, uint8_t value) {
-  const int tile = ((row >> 5) * (cols >> 3) + (col >> 3)) << 8;
-  const int offset = ((col & 3) << 6) + ((row & 15) << 2) + (((col >> 2) & 1) << 1) + ((row >> 4) & 1);
-  output[tile + offset] = value;
-}
-
-template<bool Shuffled>
-__device__ __forceinline__ void store_fp4(uint8_t* output, int row, int col, int packed_cols, uint16_t value) {
-  int index = row * packed_cols + col;
-  if constexpr (Shuffled) {
-    const int tile = (row >> 4) * (packed_cols << 4) + (col >> 5) * 512;
-    const int offset = ((col >> 4) & 1) * 256 + (row & 15) * 16 + (col & 15);
-    index = tile + offset;
-  }
-  *reinterpret_cast<uint16_t*>(output + index) = value;
-}
-
-__device__ __forceinline__ void load_tile(uint16_t* tile, const uint16_t* input, int tile_m, int tile_n) {
-  const int row = threadIdx.x / THREADS_PER_ROW;
-  const int col = threadIdx.x % THREADS_PER_ROW * VALUES_PER_THREAD;
-  const uint64_t packed = *reinterpret_cast<const uint64_t*>(input + (tile_m + row) * N + tile_n + col);
-  *reinterpret_cast<uint32_t*>(tile + row * SMEM_STRIDE + col) = static_cast<uint32_t>(packed);
-  *reinterpret_cast<uint32_t*>(tile + row * SMEM_STRIDE + col + 2) = static_cast<uint32_t>(packed >> 32);
-}
-
-__device__ __forceinline__ void quantize_row(uint16_t* tile, uint8_t* fp4_output, uint8_t* scale_output,
-                                             int tile_m, int tile_n, int local_row, int lane) {
-  const int row = tile_m + local_row;
-  const int col = lane * VALUES_PER_THREAD;
-  const Quantized4 result = quantize(load_bf16x4(tile + local_row * SMEM_STRIDE + col), lane);
-  store_fp4<SHUFFLE_ROWWISE_FP4>(fp4_output, row, (tile_n + col) / 2, N_PACKED, result.fp4);
-  if (lane == 0) store_scale(scale_output, row, tile_n / BLOCK, N_SCALES, result.scale);
-}
-
-__device__ __forceinline__ Quantized4 quantize_col(uint16_t* tile, int col, int lane) {
-  const int row = lane * VALUES_PER_THREAD;
-  return quantize(make_float4(
-    __uint_as_float(static_cast<uint32_t>(tile[(row + 0) * SMEM_STRIDE + col]) << 16),
-    __uint_as_float(static_cast<uint32_t>(tile[(row + 1) * SMEM_STRIDE + col]) << 16),
-    __uint_as_float(static_cast<uint32_t>(tile[(row + 2) * SMEM_STRIDE + col]) << 16),
-    __uint_as_float(static_cast<uint32_t>(tile[(row + 3) * SMEM_STRIDE + col]) << 16)), lane);
-}
-
-} // namespace
-
-extern "C" __global__ __launch_bounds__(THREADS, 8)
-void KERNEL_NAME(uint8_t* __restrict__ rowwise_fp4, uint8_t* __restrict__ rowwise_scale,
-                 uint8_t* __restrict__ colwise_fp4, uint8_t* __restrict__ colwise_scale,
-                 const uint16_t* __restrict__ input) {
-  __shared__ uint16_t tile[BLOCK * SMEM_STRIDE];
-  const int tid = threadIdx.x;
-  const int line = tid / THREADS_PER_ROW;
-  const int lane = tid % THREADS_PER_ROW;
-  const int block_m = blockIdx.x * TILE_M;
-  const int block_n = blockIdx.y * TILE_N;
-
-  if constexpr (!SHUFFLE_COLWISE_FP4) {
-    uint16_t col_fp4[TILE_N / BLOCK][TILE_M / BLOCK];
-    uint8_t col_scale[TILE_N / BLOCK][TILE_M / BLOCK];
-
-    for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++) {
-      for (int chunk_n = 0; chunk_n < TILE_N / BLOCK; chunk_n++) {
-        const int tile_m = block_m + chunk_m * BLOCK;
-        const int tile_n = block_n + chunk_n * BLOCK;
-        load_tile(tile, input, tile_m, tile_n);
-        __syncthreads();
-
-        quantize_row(tile, rowwise_fp4, rowwise_scale, tile_m, tile_n, line, lane);
-        const Quantized4 result = quantize_col(tile, line, lane);
-        col_fp4[chunk_n][chunk_m] = result.fp4;
-        col_scale[chunk_n][chunk_m] = result.scale;
-        __syncthreads();
-      }
-    }
-
-    for (int chunk_n = 0; chunk_n < TILE_N / BLOCK; chunk_n++) {
-      for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++)
-        tile[line * BLOCK + chunk_m * THREADS_PER_ROW + lane] = col_fp4[chunk_n][chunk_m];
-      __syncthreads();
-
-      for (int round = 0; round < BLOCK / THREADS_PER_ROW; round++) {
-        const int col = round * THREADS_PER_ROW + tid / BLOCK;
-        const int row_pair = tid % BLOCK;
-        *reinterpret_cast<uint16_t*>(colwise_fp4 + (block_n + chunk_n * BLOCK + col) * M_PACKED + block_m / 2 + row_pair * 2) =
-          tile[col * BLOCK + row_pair];
-      }
-
-      if (lane == 0) {
-        const int col = block_n + chunk_n * BLOCK + line;
-        for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++)
-          store_scale(colwise_scale, col, block_m / BLOCK + chunk_m, M_SCALES, col_scale[chunk_n][chunk_m]);
-      }
-      __syncthreads();
-    }
-  } else {
-    for (int chunk_m = 0; chunk_m < TILE_M / BLOCK; chunk_m++) {
-      for (int chunk_n = 0; chunk_n < TILE_N / BLOCK; chunk_n++) {
-        const int tile_m = block_m + chunk_m * BLOCK;
-        const int tile_n = block_n + chunk_n * BLOCK;
-        load_tile(tile, input, tile_m, tile_n);
-        __syncthreads();
-
-        quantize_row(tile, rowwise_fp4, rowwise_scale, tile_m, tile_n, line, lane);
-        const int row = lane * VALUES_PER_THREAD;
-        const int col = tile_n + line;
-        const Quantized4 result = quantize_col(tile, line, lane);
-        store_fp4<true>(colwise_fp4, col, (tile_m + row) / 2, M_PACKED, result.fp4);
-        if (lane == 0) store_scale(colwise_scale, col, tile_m / BLOCK, M_SCALES, result.scale);
-        __syncthreads();
-      }
-    }
-  }
-}
@@ -1,49 +0,0 @@
-import functools, math
-from tinygrad import Tensor, dtypes
-from tinygrad.uop.ops import UOp, KernelInfo
-from tinygrad.renderer import Estimates
-from extra.llama_kernels import alloc_like
-
-LOG2E = 1.4426950408889634
-
-@functools.cache
-def _custom_swiglu(out:UOp, x_w13:UOp) -> UOp:
-  rows, hidden = math.prod(x_w13.shape[:-1]), x_w13.shape[-1]//2
-  n_elems = rows * hidden
-  out, x_w13 = out.reshape(n_elems), x_w13.reshape(rows, 2*hidden)
-  i = UOp.range(n_elems, 0)
-  row, col = i // hidden, i % hidden
-  act, gate = x_w13[row, col].cast(dtypes.float), x_w13[row, hidden+col].cast(dtypes.float)
-  sigmoid = (1.0 + (-LOG2E * act).exp2()).reciprocal()
-  store = out[i].store((act * sigmoid * gate).cast(out.dtype))
-  return store.end(i).sink(arg=KernelInfo(f"swiglu_fwd_{n_elems}", estimates=Estimates(ops=5*n_elems, mem=6*n_elems)))
-
-@functools.cache
-def _custom_swiglu_bwd(grad_out:UOp, x_w13:UOp, grad_act:UOp) -> UOp:
-  rows, hidden = math.prod(x_w13.shape[:-1]), x_w13.shape[-1]//2
-  n_elems = rows * hidden
-  grad_out, x_w13, grad_act = grad_out.reshape(rows, 2*hidden), x_w13.reshape(rows, 2*hidden), grad_act.reshape(n_elems)
-  i = UOp.range(n_elems, 0)
-  row, col = i // hidden, i % hidden
-  act, gate = x_w13[row, col].cast(dtypes.float), x_w13[row, hidden+col].cast(dtypes.float)
-  grad = grad_act[i].cast(dtypes.float)
-  sigmoid = (1.0 + (-LOG2E * act).exp2()).reciprocal()
-  silu = act * sigmoid
-  dact = grad_out[row, col].store((grad * (sigmoid + silu * (1.0 - sigmoid)) * gate).cast(grad_out.dtype))
-  dgate = grad_out.after(dact)[row, hidden+col].store((grad * silu).cast(grad_out.dtype))
-  return dgate.end(i).sink(arg=KernelInfo(f"swiglu_bwd_{n_elems}", estimates=Estimates(ops=10*n_elems, mem=10*n_elems)))
-
-def _swiglu_bwd(gradient:UOp, kernel:UOp):
-  _, x_w13 = kernel.src[1:]
-  axis = x_w13.axis if isinstance(x_w13.device, tuple) else None
-  grad_out = alloc_like(x_w13.shape, dtypes.bfloat16, x_w13.device, axis)
-  grad_out, *_ = Tensor.custom_kernel(grad_out, Tensor(x_w13, device=x_w13.device), Tensor(gradient, device=x_w13.device),
-                                      fxn=_custom_swiglu_bwd)
-  return (None, grad_out.uop)
-
-def swiglu(x_w13:Tensor) -> Tensor:
-  assert x_w13.dtype == dtypes.bfloat16 and x_w13.ndim >= 2 and x_w13.shape[-1] % 32 == 0
-  *prefix, two_k = x_w13.shape
-  axis = x_w13.uop.axis if isinstance(x_w13.device, tuple) else None
-  out = alloc_like((*prefix, two_k//2), dtypes.bfloat16, x_w13.device, axis)
-  return Tensor.custom_kernel(out, x_w13, fxn=_custom_swiglu, grad_fxn=_swiglu_bwd)[0]
@@ -1,217 +0,0 @@
-# Runbook: Llama 3 8B Training on DigitalOcean MI350X
-
-## Machine Specs
- 8x MI350X GPUs (gfx950, device ID 75b0), 288GB VRAM each
- 2TB RAM, 192 CPUs, 2TB disk
- ROCm 7.14 at `/opt/rocm` (NOT `/opt/rocm-7.1.1` like the submission scripts assume)
- Python 3.12
-
-## Phase 1: System Setup
-
-### 1.1 Install packages
-```bash
-apt-get update
-apt-get install -y python3-pip python3-venv git tmux rclone clang
-```
-
-### 1.2 Install Python deps
-```bash
-python3 -m pip install --break-system-packages --ignore-installed typing-extensions numpy tqdm wandb tiktoken sentencepiece
-```
-Note: `--ignore-installed typing-extensions` is needed because the base image ships typing-extensions 4.10.0 without a RECORD file, so pip cannot uninstall it.
-
-### 1.3 Install ROCm dev headers
-The base image has ROCm runtime but NOT the HIP dev headers. Need:
-```bash
-apt-get install -y amdrocm-core-dev
-```
-This installs `hip/hip_runtime.h` at `/opt/rocm/core-7.14/include/hip/hip_runtime.h`.
-The symlink `/opt/rocm/include` → `/opt/rocm/core-7.14/include` makes it available at `/opt/rocm/include/hip/hip_runtime.h`.
-
-### 1.4 Configure ROCm comgr
-ROCm 7.14 ships comgr 3.3 at `/opt/rocm/lib/libamd_comgr.so`. tinygrad's DLL loader needs explicit env vars to find it (it searches for `libcomgr.so*` by default, not `libamd_comgr.so*`). Set these in the run command:
-```bash
-export COMGR_PATH=/opt/rocm/lib/libamd_comgr.so
-export COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so
-```
-Also add ROCm libs to ldconfig so comgr's shared library dependencies resolve:
-```bash
-cat > /etc/ld.so.conf.d/rocm.conf << 'EOF'
-/opt/rocm/lib
-/opt/rocm/lib/llvm/lib
-/opt/rocm/lib/rocm_sysdeps/lib
-EOF
-ldconfig
-```
-
-### 1.5 Install geohot tmux config
-```bash
-curl -sL https://raw.githubusercontent.com/geohot/configuration/master/.tmux.conf -o ~/.tmux.conf
-```
-
-### 1.6 Reload amdgpu driver
-tinygrad's HCQ backend needs `/dev/kfd` which is created by the amdgpu kernel driver.
-If the driver was unloaded, reload it:
-```bash
-modprobe amdgpu
-ls /dev/kfd  # should exist
-```
-
-## Phase 2: Clone tinygrad
-```bash
-cd /root
-git clone https://github.com/tinygrad/tinygrad.git
-cd tinygrad
-python3 -m pip install --break-system-packages -e .
-```
-
-## Phase 3: Download C4 Dataset
-
-The C4 data is on the MLCommons Cloudflare R2 bucket in Megatron-LM indexed format.
-
-```bash
-rclone config create mlc-training s3 provider=Cloudflare \
-  access_key_id=76ea42eadb867e854061a1806220ee1e \
-  secret_access_key=a53625c4d45e3ca8ac0df8a353ea3a41ffc3292aa25259addd8b7dc5a6ce2936 \
-  endpoint=c2686074cb2caf5cbaf6d134bdba8b47.r2.cloudflarestorage.com
-
-mkdir -p /raid/datasets/c4-8b
-rclone copy mlc-training:mlcommons-training-wg-public/llama3_1/datasets/c4/llama3_1_8b/ /raid/datasets/c4-8b/ -P
-```
-
-Files downloaded (~85GB total, ~6 minutes):
- `c4-train.en_6_text_document.bin` (79 GB)
- `c4-train.en_6_text_document.idx` (870 MB)
- `c4-validation-91205-samples.en_text_document.bin` (159 MB)
- `c4-validation-91205-samples.en_text_document.idx` (1.8 MB)
- `LICENSE.txt`, `NOTICE.txt`
-
-**Wait for rclone to fully complete before starting training.** Starting training while the dataset is still downloading will read a truncated .bin file, causing `ValueError: all input arrays must have the same shape` in the dataloader. The stale `.index_cache` and `.blend_cache` files must also be deleted if this happens:
-```bash
-rm -f /raid/datasets/c4-8b/*.index_cache /raid/datasets/c4-8b/*.blend_cache
-```
-
-## Phase 4: wandb Login
-```bash
-wandb login
-```
-Enter API key from https://wandb.ai/authorize
-
-Alternatively, pass the key directly:
-```bash
-wandb login <API_KEY>
-```
-
-## Phase 5: Run Training
-
-Run training in tmux so it survives SSH disconnects:
-```bash
-tmux new-session -d -s train 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm WANDB=1 bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh 2>&1 | tee /root/train.log'
-```
-Attach with `tmux attach -t train`.
-
-### 5.1 Smoke test (beam search, 2 layers, real data)
-Always run beam first to validate the pipeline:
-```bash
-tmux new-session -d -s beam 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh 2>&1 | tee /root/beam.log'
-```
-
-The beam test runs 10 training steps with 2 layers. Expected results:
- ~0.29s per step after warmup
- ~700K GFLOPS, ~7% MFU (low because only 2 layers)
- ~380 GB VRAM used
- Loss stable at ~12.55 with random init
-
-### 5.2 Full training run
-```bash
-tmux new-session -d -s train 'cd /root/tinygrad && COMGR_PATH=/opt/rocm/lib/libamd_comgr.so COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so CC=/opt/rocm/core-7.14/lib/llvm/bin/clang DEV=AMD:HIP ROCM_PATH=/opt/rocm WANDB=1 bash examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh 2>&1 | tee /root/train.log'
-```
-
-## Environment Variable Reference
-
-| Variable | Value | Why |
-|---|---|---|
-| `COMGR_PATH` | `/opt/rocm/lib/libamd_comgr.so` | tinygrad's DLL loader needs explicit path to find comgr 3.3 |
-| `COMGR_3_PATH` | `/opt/rocm/lib/libamd_comgr.so` | comgr 3.x uses a separate `comgr_3` module with its own path var |
-| `CC` | `/opt/rocm/core-7.14/lib/llvm/bin/clang` | System clang doesn't know gfx950; must use ROCm's bundled clang |
-| `DEV` | `AMD:HIP` | Force HIPRenderer (comgr-based) over HIPCCRenderer (hipcc subprocess) |
-| `ROCM_PATH` | `/opt/rocm` | Script defaults to `/opt/rocm-7.1.1` which doesn't exist |
-| `WANDB` | `1` | Enable wandb logging (off by default) |
-
-## Architecture
-
-| Component | Source file |
-|---|---|
-| Model | `examples/mlperf/models/flat_llama.py` — FlatTransformer, FP8 MXFP4 weights, fused QKV, flash attention |
-| Trainer | `examples/mlperf/model_train.py` → `train_llama3()` |
-| Optimizer | `examples/mlperf/optim.py` — GradAccClipAdamW, master weights, FP8 re-quant |
-| LR schedule | `examples/mlperf/lr_schedulers.py` — CosineAnnealingLRWithWarmup |
-| Dataloader | `examples/mlperf/dataloader.py` — Megatron-LM indexed bin format |
-| ASM GEMM | `extra/gemm/cdna_asm_gemm.py` — gfx950 MFMA assembly, MXFP4 |
-| Flash attention | `extra/thunder/amd/fa.py` |
-| Fused kernels | `extra/llama_kernels/` — rmsnorm, silu, quantize, fused_ce |
-| GPU driver | `tinygrad/runtime/ops_amd.py` — HCQ, direct KFD ioctl |
-| Renderer | `tinygrad/renderer/cstyle.py` — HIPRenderer for gfx950 |
-| comgr compiler | `tinygrad/runtime/support/compiler_amd.py` — HIPCompiler using comgr 3.3 |
-
-## Troubleshooting
-
-### `'hip/hip_runtime.h' file not found`
-Install `amdrocm-core-dev`:
-```bash
-apt-get install -y amdrocm-core-dev
-```
-
-### `'gfx950' is not a recognized processor` + LLVM crash
-System clang doesn't know gfx950. Set `CC=/opt/rocm/core-7.14/lib/llvm/bin/clang`.
-
-### `comgr not available: try setting COMGR_PATH?`
-Add ROCm libs to ldconfig and set `COMGR_PATH` and `COMGR_3_PATH`:
-```bash
-# /etc/ld.so.conf.d/rocm.conf should contain /opt/rocm/lib paths
-ldconfig
-```
-
-### `comgr not available: try setting COMGR_3_PATH?`
-comgr 3.x uses a separate module. Set `COMGR_3_PATH=/opt/rocm/lib/libamd_comgr.so` too.
-
-### `No such file or directory: 'clang'`
-Install clang: `apt-get install -y clang` (for CPU compilation).
-For gfx950 HIP compilation, comgr (not clang) is used — ensure the ROCm 7.14 comgr 3.3 is properly loaded via `COMGR_PATH` and `COMGR_3_PATH`.
-
-## Appendix: KVM Virtualization Observations
-
-### Virtualization detection
-```
-$ systemd-detect-virt
-kvm
-$ lspci -nn | grep AMD
-83:00.0 ... Device [1002:75b0]
-```
-CPU flags include `hypervisor`. `dmesg` shows `Hypervisor detected: KVM`.
-
-### Working path: amdgpu driver (KFDIface)
-The amdgpu driver loads on boot and binds to all 8 GPUs, creating `/dev/kfd` and 64 renderD nodes (`/dev/dri/renderD128` through `/dev/dri/renderD191`). tinygrad's `KFDIface` enumerates GPUs through `/sys/devices/virtual/kfd/kfd/topology/nodes` and uses `/dev/kfd` for ioctl. No PCI device ID patching is needed — the KFD path does not use `PCIIface` or `AMDev._run_discovery()`.
-
-This is the working configuration. No code changes to tinygrad are required.
-
-### PCIIface path (does not work on this VM)
-For reference, the `PCIIface` path was also explored but does not work in this KVM guest:
-
- `PCIIface` in `ops_amd.py` does not list device ID `0x75b0`. Adding it allows PCI detection but `AMDev._run_discovery()` fails because the VRAM BAR reads all `0xFF`.
- This was observed with the GPU unbound from any driver, after PCI reset, and with VFIO bound.
- VFIO binding (`vfio-pci` with `enable_unsafe_noiommu_mode=1`) succeeded but VRAM BAR still reads all `0xFF`.
- No IOMMU in guest — `dmesg` has no `AMD-Vi` entries, PCI devices have no `iommu_group` symlink.
-
-### amdgpu driver behavior
-On first boot, amdgpu loaded and bound to all 8 GPUs. On one boot it failed to initialize:
-```
-[  799.780369] amdgpu 0000:83:00.0: Failed to alloc msi vectors
-[  799.781476] amdgpu 0000:83:00.0: sw_init of IP block <vega20_ih> failed -22
-[  799.782724] amdgpu 0000:83:00.0: amdgpu_device_ip_init failed
-[  799.793885] amdgpu 0000:83:00.0: Fatal error during GPU init
-```
-On a subsequent boot, amdgpu initialized successfully (SMU initialized, VRAM ready). After unbinding all 8 GPUs from amdgpu, `rmmod amdgpu` wedged the module (stuck in "Unloading" state in `/proc/modules`), requiring a full VM reboot.
-
-### No fan control
-No `fan*` or `pwm*` hwmon entries exist. Only `temp*`, `power*`, `freq*` are exposed. GPU temps read 56-63°C, power ~265W per GPU.
@@ -2,7 +2,7 @@ import math, pathlib, functools, struct

 from tinygrad import Device, Tensor
 from tinygrad.dtype import DTypeLike, dtypes
-from tinygrad.helpers import DEBUG, getenv
+from tinygrad.helpers import DEBUG
 from tinygrad.renderer import Estimates
 from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from tinygrad.runtime.support.elf import elf_loader
@@ -110,49 +110,7 @@ def _sharded_empty_like(ref:Tensor, axis:int|None=None) -> Tensor:
  return _sharded_empty(ref.shape, ref, axis)

@functools.cache
-def _windowed_lse(xq:Tensor, xk:Tensor, sinks, W:int) -> Tensor:
-  B, N, H, hd = xq.shape
-  H_KV = xk.shape[2]; R = H // H_KV; nb = N // W; sm = hd ** -0.5
-  q = xq.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
-  k = xk.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
-  k_prev = k.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
-  sc_d = (q @ k.transpose(-1, -2)) * sm
-  sc_p = (q @ k_prev.transpose(-1, -2)) * sm
-  li, lj = Tensor.arange(W).reshape(W, 1), Tensor.arange(W).reshape(1, W)
-  pv = (Tensor.arange(nb).reshape(nb, 1, 1) >= 1)
-  sc_d = (lj <= li).where(sc_d, -float("inf"))
-  sc_p = ((li < lj) & pv).where(sc_p, -float("inf"))
-  m = sc_d.max(-1, keepdim=True).maximum(sc_p.max(-1, keepdim=True))
-  if sinks is not None: m = m.maximum(sinks.reshape(1, H_KV, R, 1, 1, 1).float())
-  denom = (sc_d - m).exp().sum(-1, keepdim=True) + (sc_p - m).exp().sum(-1, keepdim=True)
-  if sinks is not None: denom = denom + (sinks.reshape(1, H_KV, R, 1, 1, 1).float() - m).exp()
-  return (m + denom.log()).reshape(B, H, N).unsqueeze(2)  # (B, H, 1, N), matches saved l_vec
-
-def _windowed_delta(xq:Tensor, xk:Tensor, xv:Tensor, do:Tensor, sinks, W:int) -> Tensor:
-  B, N, H, hd = xq.shape
-  H_KV = xk.shape[2]; R = H // H_KV; nb = N // W; sm = hd ** -0.5
-  q = xq.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
-  k = xk.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
-  v = xv.permute(0, 2, 1, 3).reshape(B, H_KV, 1, nb, W, hd).float()
-  dob = do.reshape(B, N, H_KV, R, hd).permute(0, 2, 3, 1, 4).reshape(B, H_KV, R, nb, W, hd).float()
-  k_prev = k.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
-  v_prev = v.pad((None, None, None, (1, 0), None, None))[:, :, :, :nb]
-  sc_d = (q @ k.transpose(-1, -2)) * sm
-  sc_p = (q @ k_prev.transpose(-1, -2)) * sm
-  li, lj = Tensor.arange(W).reshape(W, 1), Tensor.arange(W).reshape(1, W)
-  pv = (Tensor.arange(nb).reshape(nb, 1, 1) >= 1)
-  sc_d = (lj <= li).where(sc_d, -float("inf"))
-  sc_p = ((li < lj) & pv).where(sc_p, -float("inf"))
-  m = sc_d.max(-1, keepdim=True).maximum(sc_p.max(-1, keepdim=True))
-  if sinks is not None: m = m.maximum(sinks.reshape(1, H_KV, R, 1, 1, 1).float())
-  e_d, e_p = (sc_d - m).exp(), (sc_p - m).exp()
-  denom = e_d.sum(-1, keepdim=True) + e_p.sum(-1, keepdim=True)
-  if sinks is not None: denom = denom + (sinks.reshape(1, H_KV, R, 1, 1, 1).float() - m).exp()
-  o = ((e_d / denom) @ v) + ((e_p / denom) @ v_prev)
-  delta = (dob * o).sum(-1)
-  return delta.reshape(B, H, N).unsqueeze(2)
-
-def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=0):
+def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink):
  def grad(dou:UOp, ker:UOp) -> tuple:
    do = Tensor(dou, device=dou.device)
    attn = Tensor(ker.src[1].after(ker), device=ker.src[1].device)
@@ -160,8 +118,6 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
    xq = Tensor(ker.src[3], device=ker.src[3].device)
    xk = Tensor(ker.src[4], device=ker.src[4].device)
    xv = Tensor(ker.src[5], device=ker.src[5].device)
-    if window:
-      l_vec = _windowed_lse(xq, xk, Tensor(ker.src[6], device=ker.src[6].device) if has_sink else None, window)

    dq = _sharded_empty((B, H, N, D), xq, axis=shard_axis_t)
    GROUP_SIZE = H_local // H_KV_local
@@ -172,10 +128,8 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
    # delta_vec = (do * attn).sum(-1, dtype=dtypes.float32).transpose(1, 2).unsqueeze(-2).detach()
    delta_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)
    delta_vec, dq = Tensor.custom_kernel(delta_vec, dq, attn, do, fxn=functools.partial(custom_fa_backward_pre, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:2]
-    if window:
-      delta_vec = _windowed_delta(xq, xk, xv, do, Tensor(ker.src[6], device=ker.src[6].device) if has_sink else None, window)

-    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, window=window))[:3]
+    dq, dk_partial, dv_partial = Tensor.custom_kernel(dq, dk_partial, dv_partial, do, xq, xk, xv, l_vec, delta_vec, fxn=functools.partial(custom_fa_backward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D))[:3]

    if D == 64:
      dq = dq.reshape(B, H, N//16, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2).permute(0, 1, 2, 8, 9, 10, 11, 3, 4, 6, 7, 5, 12).reshape(B, H, N, D).transpose(1, 2)
@@ -195,7 +149,7 @@ def _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, sha
  return grad

 # TODO: remove write_flat once scheduler can remove reshapes between custom_kernel. TestCustomKernel.test_simple_reshape
-def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None, window:int=0):
+def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False, write_flat:bool=False, sinks:Tensor|None=None):
  assert attn_mask is None, "attn_mask not supported"
  assert is_causal, "only causal attention supported"

@@ -222,18 +176,18 @@ def flash_attention(xq, xk, xv, attn_mask:Tensor|None=None, is_causal:bool=False
  attn = _sharded_empty((B, N, H * D), xq, axis=shard_axis) if write_flat else _sharded_empty_like(xq, axis=shard_axis)
  l_vec = _sharded_empty((B, H, 1, N), xq, dtype=dtypes.float32, axis=shard_axis_t)

-  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink, window=window)
+  grad = _fa_grad_fxn(B, H, N, D, H_local, H_KV_local, H_KV, B_local, shard_axis, shard_axis_t, single_device, arch, has_sink)

  fwd_inputs = (attn, l_vec, xq, xk, xv) + ((sinks,) if has_sink else ())
-  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink, window=window), grad_fxn=grad)[:2]
+  attn, l_vec = Tensor.custom_kernel(*fwd_inputs, fxn=functools.partial(custom_fa_forward, device=single_device, arch=arch, B=B_local, N=N, H=H_local, H_KV=H_KV_local, D=D, has_sink=has_sink), grad_fxn=grad)[:2]

  return attn, attn, l_vec

@functools.cache
-def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True, window:int=0):
+def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None, *, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, has_sink:bool=True):
  code = (pathlib.Path(__file__).parent / "fa_fwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}", f"-DWINDOW={window}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DATTN_SINK={int(has_sink)}"]

  Q_BLOCK_SIZE = 32
  NUM_WARPS = 8
@@ -252,11 +206,10 @@ def custom_fa_forward(o:UOp, l_vec:UOp, q:UOp, k:UOp, v:UOp, sinks:UOp|None=None
                  arg=KernelInfo(name="custom_fa_forward", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  if not getenv("NO_HIPCC"):
-    lib = bytearray(lib)
-    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-    struct.pack_into('<I', lib, rodata_off, 160000)
-    lib = bytes(lib)
+  lib = bytearray(lib)
+  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+  struct.pack_into('<I', lib, rodata_off, 160000)
+  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
@@ -283,20 +236,19 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
                  arg=KernelInfo(name="custom_fa_backward_pre", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  if not getenv("NO_HIPCC"):
-    lib = bytearray(lib)
-    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-    struct.pack_into('<I', lib, rodata_off, 160000)
-    lib = bytes(lib)
+  lib = bytearray(lib)
+  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+  struct.pack_into('<I', lib, rodata_off, 160000)
+  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))

@functools.cache
-def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int, window:int=0):
+def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_vec:UOp, delta_vec:UOp, device:str, arch:str, B:int, N:int, H:int, H_KV:int, D:int):
  code = (pathlib.Path(__file__).parent / "fa_bwd_causal.cpp").read_text()
  compile_args = [f"-I{(pathlib.Path(__file__).parent / 'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-DHIP_ENABLE_WARP_SYNC_BUILTINS", "-ffast-math",
-                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}", f"-DWINDOW={window}"]
+                  f"-DATTN_B={B}", f"-DATTN_N={N}", f"-DATTN_H={H}", f"-DATTN_H_KV={H_KV}", f"-DATTN_D={D}"]

  BLOCK_SIZE_KV = 256
  GROUP_SIZE = H // H_KV
@@ -316,11 +268,10 @@ def custom_fa_backward(dq:UOp, dk:UOp, dv:UOp, do:UOp, q:UOp, k:UOp, v:UOp, l_ve
                  arg=KernelInfo(name="custom_fa_backward", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  if not getenv("NO_HIPCC"):
-    lib = bytearray(lib)
-    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-    struct.pack_into('<I', lib, rodata_off, 160000)
-    lib = bytes(lib)
+  lib = bytearray(lib)
+  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+  struct.pack_into('<I', lib, rodata_off, 160000)
+  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
@@ -347,11 +298,10 @@ def custom_fa_backward_post(dq_out:UOp, dq_in:UOp, device:str, arch:str, B:int,
                  arg=KernelInfo(name="custom_fa_backward_post", estimates=estimates))

  lib = HIPCCCompiler(arch, compile_args).compile_cached(code)
-  if not getenv("NO_HIPCC"):
-    lib = bytearray(lib)
-    rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
-    struct.pack_into('<I', lib, rodata_off, 160000)
-    lib = bytes(lib)
+  lib = bytearray(lib)
+  rodata_off = next(sh.header.sh_offset for sh in elf_loader(bytes(lib))[1] if sh.name == ".rodata")
+  struct.pack_into('<I', lib, rodata_off, 160000)
+  lib = bytes(lib)

  return UOp(Ops.PROGRAM,
             src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=code), UOp(Ops.BINARY, arg=lib)))
@@ -43,10 +43,6 @@ constexpr int SLICE_QO = 32;
 constexpr int DOT_SLICE_QO = 16;
 constexpr int WARP_SIZE_KV = 64; // warp size for KV
 constexpr bool causal = true;
-// WINDOW>0: sliding-window backward (query i sees keys in [i-WINDOW+1, i])
-#ifndef WINDOW
-#define WINDOW 0
-#endif

 #define NUM_WARPS 4
 #define NUM_THREADS (kittens::WARP_THREADS * NUM_WARPS)
@@ -92,12 +88,7 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
  const int k_start_min = j_min * WARP_SIZE_KV;
  // first Q step that can overlap this K_span:
  const int first_step = max(0, k_start_min / STEP_QO);
-#if WINDOW
-  // cap the Q loop, padded by 2 masked steps: the epilogue's deferred dq path miscomputes in-window tail queries
-  const int num_steps_per_head = min(total_steps_per_head - first_step, (BLOCK_SIZE_KV + WINDOW) / STEP_QO + 2);
-#else
  const int num_steps_per_head = total_steps_per_head - first_step;
-#endif
  const int num_steps = num_steps_per_head * HEADS_PER_WG;
  const int k_pos = j * WARP_SIZE_KV;

@@ -389,13 +380,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            mov<0, 1, neg_inf_v>(P_ij);
            mov<0, 2, neg_inf_v>(P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-          // window lower boundary, mirror of the causal edge
-          } else if (q_pos - k_pos == WINDOW) {
-            make_window<0, 0, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -549,8 +533,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -656,13 +638,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            make_causal<0, 1, neg_inf_v>(P_ij, P_ij);
            mov<0, 2, neg_inf_v>(P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-          } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            make_window<0, 1, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -816,8 +791,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -922,14 +895,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            // Apply the causal mask to [0, 2] and set [0, 3:4] to -inf
            make_causal<0, 2, neg_inf_v>(P_ij, P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-          } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            mov<0, 1, neg_inf_v>(P_ij);
-            make_window<0, 2, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1083,8 +1048,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1188,15 +1151,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
          } else if (q_pos == k_pos) {
            // Apply the causal mask to [0, 3]
            make_causal<0, 3, neg_inf_v>(P_ij, P_ij);
-#if WINDOW
-          } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            mov<0, 1, neg_inf_v>(P_ij);
-            mov<0, 2, neg_inf_v>(P_ij);
-            make_window<0, 3, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1349,8 +1303,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1476,13 +1428,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            mov<0, 1, neg_inf_v>(P_ij);
            mov<0, 2, neg_inf_v>(P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-          // window lower boundary, mirror of the causal edge
-          } else if (q_pos - k_pos == WINDOW) {
-            make_window<0, 0, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1637,8 +1582,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -1746,13 +1689,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            make_causal<0, 1, neg_inf_v>(P_ij, P_ij);
            mov<0, 2, neg_inf_v>(P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-          } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            make_window<0, 1, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -1906,8 +1842,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -2012,14 +1946,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            // Apply the causal mask to [0, 2] and set [0, 3:4] to -inf
            make_causal<0, 2, neg_inf_v>(P_ij, P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-          } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            mov<0, 1, neg_inf_v>(P_ij);
-            make_window<0, 2, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -2173,8 +2099,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[tic][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -2278,15 +2202,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
          } else if (q_pos == k_pos) {
            // Apply the causal mask to [0, 3]
            make_causal<0, 3, neg_inf_v>(P_ij, P_ij);
-#if WINDOW
-          } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            mov<0, 1, neg_inf_v>(P_ij);
-            mov<0, 2, neg_inf_v>(P_ij);
-            make_window<0, 3, neg_inf_v>(P_ij, P_ij);
-          } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
          }
        }
        mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -2439,8 +2354,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        if constexpr (D == 128) load<0, 2>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        if constexpr (D == 128) load<0, 3>(Q_i, subtile_inplace<DOT_SLICE_QO, D>(Q_i_smem[toc][0], {0, 0}), Q_i_addr);
        mma_AtB<0, 0, 7>(dQ_i_T, K_j_col, dP_ij_bf16_col_T, dQ_i_T);
-        // D=64: wait out MFMA->VALU accumulator hazard on dQ_i_T
-        if constexpr (D == 64) asm volatile("s_nop 15");
        if constexpr (D == 128) mma_AtB<1, 0, 0>(dQ_i_T, K_j_col, dP_ij_bf16_col_T);
        // Load K_j from shared memory to registers
        // load(K_j, subtile_inplace<WARP_SIZE_KV, D>(K_j_smem, {warpid, 0}));
@@ -2558,12 +2471,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            mov<0, 1, neg_inf_v>(P_ij);
            mov<0, 2, neg_inf_v>(P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-        } else if (q_pos - k_pos == WINDOW) {
-            make_window<0, 0, neg_inf_v>(P_ij, P_ij);
-        } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
        }
      }
      mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -2825,13 +2732,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            make_causal<0, 1, neg_inf_v>(P_ij, P_ij);
            mov<0, 2, neg_inf_v>(P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-        } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            make_window<0, 1, neg_inf_v>(P_ij, P_ij);
-        } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
        }
      }
      mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -3088,14 +2988,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
            // Apply the causal mask to [0, 2] and set [0, 3:4] to -inf
            make_causal<0, 2, neg_inf_v>(P_ij, P_ij);
            mov<0, 3, neg_inf_v>(P_ij);
-#if WINDOW
-        } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            mov<0, 1, neg_inf_v>(P_ij);
-            make_window<0, 2, neg_inf_v>(P_ij, P_ij);
-        } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
        }
      }
      mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -3352,15 +3244,6 @@ __global__ void attend_bwd_combined_ker(bf16 *dQ_ptr, bf16 *dK_ptr, bf16 *dV_ptr
        } else if (q_pos == k_pos) {
            // Apply the causal mask to [0, 3]
            make_causal<0, 3, neg_inf_v>(P_ij, P_ij);
-#if WINDOW
-        } else if (q_pos - k_pos == WINDOW) {
-            mov<0, 0, neg_inf_v>(P_ij);
-            mov<0, 1, neg_inf_v>(P_ij);
-            mov<0, 2, neg_inf_v>(P_ij);
-            make_window<0, 3, neg_inf_v>(P_ij, P_ij);
-        } else if (q_pos - k_pos > WINDOW) {
-            mov<neg_inf_v>(P_ij);
-#endif
        }
      }
      mul<0, 2>(P_ij, P_ij, P_SCALE_FACTOR);
@@ -34,10 +34,6 @@ constexpr int ATTN_D = 128; // dimension
 constexpr int Q_BLOCK_SIZE = 32; // q block size
 constexpr int KV_BLOCK_SIZE = 64; // kv block size
 constexpr bool causal = true;
-// WINDOW>0: sliding-window attention, query i attends keys in [i-WINDOW+1, i]
-#ifndef WINDOW
-#define WINDOW 0
-#endif

 #define NUM_WARPS 8
 #define NUM_THREADS (kittens::WARP_THREADS * NUM_WARPS)
@@ -86,26 +82,11 @@ template<typename T=float, typename L=col_l, typename S=rt_16x32_4_s> using attn

 /**********************************************************/
 template<int THR_X, int THR_Y>
-__device__ inline void mask_vec2_imm(uint32_t rel_vgpr, uint32_t rel_hi_vgpr, uint32_t neg_inf_vgpr,
+__device__ inline void mask_vec2_imm(uint32_t rel_vgpr, uint32_t neg_inf_vgpr,
                                     uint32_t& x_ref, uint32_t& y_ref) {

    uint64_t x_mask, y_mask;
-#if WINDOW
-    // causal+window in one asm block to not disturb register allocation
-    asm volatile(
-        "v_cmp_lt_i32_e64 %0, %4, %5\n\t"
-        "v_cmp_lt_i32_e64 %1, %4, %7\n\t"
-        "v_cndmask_b32_e64 %2, %2, %6, %0\n\t"
-        "v_cndmask_b32_e64 %3, %3, %6, %1\n\t"
-        "v_cmp_ge_i32_e64 %0, %8, %5\n\t"
-        "v_cmp_ge_i32_e64 %1, %8, %7\n\t"
-        "v_cndmask_b32_e64 %2, %2, %6, %0\n\t"
-        "v_cndmask_b32_e64 %3, %3, %6, %1\n\t"
-        : "=s"(x_mask), "=s"(y_mask), "+v"(x_ref), "+v"(y_ref)
-        : "v"(rel_vgpr), "n"(THR_X), "v"(neg_inf_vgpr), "n"(THR_Y), "v"(rel_hi_vgpr)
-        : "vcc"
-    );
-#else
+    // uint32_t ox, oy;
    asm volatile(
        // x: rel < THR_X ?
        "v_cmp_lt_i32_e64 %0, %6, %7\n\t"
@@ -118,7 +99,7 @@ __device__ inline void mask_vec2_imm(uint32_t rel_vgpr, uint32_t rel_hi_vgpr, ui
          "n"(THR_X), "v"(neg_inf_vgpr), "n"(THR_Y)
        : "vcc"
    );
-#endif
+    // x_ref = ox; y_ref = oy;
 }

 template<ducks::rt::col_layout RT>
@@ -141,8 +122,6 @@ __device__ inline void mask_kv_tile(RT &dst, int q_abs, int k_abs, uint32_t neg_
        // (smaller rel ⇒ more "future" keys that must be -inf)
        const int rel0 = q_pos - (k_base + row_base);
        const uint32_t rel = static_cast<uint32_t>(rel0);
-        // rel-WINDOW keeps THR within the inline-constant range
-        const uint32_t rel_hi = static_cast<uint32_t>(rel0 - WINDOW);

        #pragma unroll
        for (int j = 0; j < dst.width; ++j) {
@@ -166,14 +145,14 @@ __device__ inline void mask_kv_tile(RT &dst, int q_abs, int k_abs, uint32_t neg_
            //  - reuse a single neg_inf register
            //  - keep VCC live across the pair
            //  - avoid reloading -inf or recomputing rel
-            mask_vec2_imm< 0, 1 >(rel, rel_hi, neg_inf_v, d0x, d0y);
-            mask_vec2_imm< 2, 3 >(rel, rel_hi, neg_inf_v, d1x, d1y);
-            mask_vec2_imm< 8, 9 >(rel, rel_hi, neg_inf_v, d2x, d2y);
-            mask_vec2_imm<10,11 >(rel, rel_hi, neg_inf_v, d3x, d3y);
-            mask_vec2_imm<16,17 >(rel, rel_hi, neg_inf_v, d4x, d4y);
-            mask_vec2_imm<18,19 >(rel, rel_hi, neg_inf_v, d5x, d5y);
-            mask_vec2_imm<24,25 >(rel, rel_hi, neg_inf_v, d6x, d6y);
-            mask_vec2_imm<26,27 >(rel, rel_hi, neg_inf_v, d7x, d7y);
+            mask_vec2_imm< 0, 1 >(rel, neg_inf_v, d0x, d0y);
+            mask_vec2_imm< 2, 3 >(rel, neg_inf_v, d1x, d1y);
+            mask_vec2_imm< 8, 9 >(rel, neg_inf_v, d2x, d2y);
+            mask_vec2_imm<10,11 >(rel, neg_inf_v, d3x, d3y);
+            mask_vec2_imm<16,17 >(rel, neg_inf_v, d4x, d4y);
+            mask_vec2_imm<18,19 >(rel, neg_inf_v, d5x, d5y);
+            mask_vec2_imm<24,25 >(rel, neg_inf_v, d6x, d6y);
+            mask_vec2_imm<26,27 >(rel, neg_inf_v, d7x, d7y);
        }
    }
 }
@@ -222,16 +201,6 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    else max_num_tiles = num_tiles;
    const int q_start_pos = tile_idx * Q_BLOCK_SIZE;

-#if WINDOW
-    // start at the first in-window tile; clamp keeps >=4 tiles for the pipeline unroll
-    const int block_min_q = block_tile_idx * NUM_WARPS * Q_BLOCK_SIZE;
-    int min_tile = (block_min_q - WINDOW + 1) / KV_BLOCK_SIZE;
-    if (min_tile < 0) min_tile = 0;
-    if (min_tile > max_num_tiles - 4) min_tile = max(0, max_num_tiles - 4);
-#else
-    constexpr int min_tile = 0;
-#endif
-
    constexpr float TEMPERATURE_SCALE = (D == 128) ? 0.08838834764f*1.44269504089f : 0.125f*1.44269504089f;
    uint32_t neg_inf_v = 0xff800000;

@@ -262,7 +231,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    G::prefill_swizzled_offsets<1, false>(k_smem[0], g.Kg, swizzled_offsets_K);
    G::prefill_swizzled_offsets<1, false>(v_smem[0], g.Vg, swizzled_offsets_V);

-    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, min_tile, head_idx_kv, 0}, swizzled_offsets_K);
+    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, 0, head_idx_kv, 0}, swizzled_offsets_K);
    __builtin_amdgcn_s_waitcnt(0);
    __builtin_amdgcn_sched_barrier(0);
    __builtin_amdgcn_s_barrier();
@@ -274,9 +243,9 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    transpose(q_reg_transposed, q_reg);

    // All warps then collaboratively load in the first slice of V (V0) and the second slice of K (K1) into shared memory
-    G::load<1, false>(k_smem[1], g.Kg, {batch_idx, min_tile + 1, head_idx_kv, 0}, swizzled_offsets_K);
+    G::load<1, false>(k_smem[1], g.Kg, {batch_idx, 1, head_idx_kv, 0}, swizzled_offsets_K);
    // All warps then load in the first slice of K (K0)
-    G::load<1, false>(v_smem[0], g.Vg, {batch_idx, min_tile, head_idx_kv, 0}, swizzled_offsets_V);
+    G::load<1, false>(v_smem[0], g.Vg, {batch_idx, 0, head_idx_kv, 0}, swizzled_offsets_V);
    load(k_reg, k_smem[0]);
    __builtin_amdgcn_sched_barrier(0);
    asm volatile("s_waitcnt lgkmcnt(0)");
@@ -290,20 +259,13 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    mma_AtB(att_block[0], k_reg_transposed, q_reg_transposed, att_block[0]);
    __builtin_amdgcn_sched_barrier(0);
    if constexpr (causal) {
-        const int kv_end_pos = (min_tile + 1) * KV_BLOCK_SIZE;
-        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, WINDOW ? 1 : 0)) {
-            mask_kv_tile(att_block[0], tile_idx, min_tile, neg_inf_v, lane);
+        const int kv_end_pos = (1) * KV_BLOCK_SIZE;
+        if (__builtin_expect(q_start_pos < kv_end_pos, 0)) {  // Only mask if needed
+            mask_kv_tile(att_block[0], tile_idx, 0, neg_inf_v, lane);
        }
    }
    // Each warp performs a partial softmax of QK0 (i.e. some of the online softmax up until but not including the second exponential scaling of the attention block likely)
-#if WINDOW
-    // floor the max: min_tile can be fully masked, which would NaN via exp2(-inf - -inf)
-    zero(max_vec_prev);
-    add(max_vec_prev, max_vec_prev, -1e4f);
-    col_max(max_vec, att_block[0], max_vec_prev);
-#else
    col_max(max_vec, att_block[0]);
-#endif

    copy(max_vec_prev, max_vec);
    exp2(scale_vec, scale_vec);
@@ -322,25 +284,21 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    // All warps then load in the second slice of K (K1)
    load(k_reg, k_smem[1]);
    // All warps then collaboratively load in the third slice of K (K2) into shared memory
-    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, min_tile + 2, head_idx_kv, 0}, swizzled_offsets_K);
+    G::load<1, false>(k_smem[0], g.Kg, {batch_idx, 2, head_idx_kv, 0}, swizzled_offsets_K);
    // All warps then collaboratively load in the second slice of V (V1) into shared memory
-    G::load<1, false>(v_smem[1], g.Vg, {batch_idx, min_tile + 1, head_idx_kv, 0}, swizzled_offsets_V);
+    G::load<1, false>(v_smem[1], g.Vg, {batch_idx, 1, head_idx_kv, 0}, swizzled_offsets_V);
    asm volatile("s_waitcnt lgkmcnt(0)");
    asm volatile("s_waitcnt vmcnt(" FA_VM4 ")");
    __builtin_amdgcn_sched_barrier(0);
    __builtin_amdgcn_s_barrier();

    // hot loop
-    for (int j = min_tile + 3; j < max_num_tiles - 1; j += 2) {
+    for (int j = 3; j < max_num_tiles - 1; j += 2) {
        // Cluster 0:
        //      QK1
        zero(att_block[1]);
        transpose(k_reg_transposed, k_reg);
        mma_AtB(att_block[1], k_reg_transposed, q_reg_transposed, att_block[1]);
-#if WINDOW
-        // window masks interior tiles that causal skips
-        mask_kv_tile(att_block[1], tile_idx, j - 2, neg_inf_v, lane);
-#endif
        //      Finish softmax for QK0
        exp2(att_block[0].tiles[1][0], att_block[0].tiles[1][0]);
        mul(norm_vec, norm_vec, scale_vec);
@@ -421,7 +379,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
        load(v_reg, v_smem[1]);
        if constexpr (causal) {
            const int kv_end_pos = (j) * KV_BLOCK_SIZE;
-            if (WINDOW || q_start_pos < kv_end_pos) {
+            if (q_start_pos < kv_end_pos) {  // Only mask if needed
                mask_kv_tile(att_block[0], tile_idx, j - 1, neg_inf_v, lane);
            }
        }
@@ -489,7 +447,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    load(v_reg, v_smem[0]);
    if constexpr (causal) {
        const int kv_end_pos = (max_num_tiles - 2) * KV_BLOCK_SIZE;
-        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, WINDOW ? 1 : 0)) {
+        if (__builtin_expect(q_start_pos < kv_end_pos, 0)) {  // Only mask if needed
            mask_kv_tile(att_block[1], tile_idx, max_num_tiles - 3, neg_inf_v, lane);
        }
    }
@@ -552,7 +510,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    load(v_reg, v_smem[1]);
    if constexpr (causal) {
        const int kv_end_pos = (max_num_tiles - 1) * KV_BLOCK_SIZE;
-        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, 1)) {
+        if (__builtin_expect(q_start_pos < kv_end_pos, 1)) {  // Only mask if needed
            mask_kv_tile(att_block[0], tile_idx, max_num_tiles - 2, neg_inf_v, lane);
        }
    }
@@ -614,7 +572,7 @@ __global__ void attend_ker(bf16 *O_ptr, float *L_vec_ptr, bf16 *Q_ptr, bf16 *K_p
    load(v_reg, v_smem[0]);
    if constexpr (causal) {
        const int kv_end_pos = (max_num_tiles) * KV_BLOCK_SIZE;
-        if (__builtin_expect(WINDOW || q_start_pos < kv_end_pos, 1)) {
+        if (__builtin_expect(q_start_pos < kv_end_pos, 1)) {  // Only mask if needed
            mask_kv_tile(att_block[1], tile_idx, max_num_tiles - 1, neg_inf_v, lane);
        }
    }
@@ -97,33 +97,4 @@ __device__ inline static void atomic_pk_add_bf16_with_warpid(const GL &dst, cons
            }(std::make_index_sequence<RT::width>{});
        }.template operator()<Ns>(), ...);
    }(std::make_index_sequence<RT::height>{});
-}
-// make_window: complement of make_causal for the window lower boundary (q_pos-k_pos == WINDOW). masks = ~(causal masks)
-template<int N, int M, int GPR, ducks::art::all T0, ducks::art::all T1>
-__device__ static inline void make_window(T0 &dst, const T1 &src) {
-    static_assert(std::is_same_v<typename T0::T, float> && std::is_same_v<typename T1::T, float>, "Only float to float window mask is supported");
-    static_assert(std::is_same_v<typename T0::layout, typename T1::layout>, "Only same layout is supported");
-    static_assert(std::is_same_v<typename T0::shape, typename T1::shape>, "Only same shape is supported");
-
-    if constexpr (std::is_same_v<typename T0::layout, typename ducks::rt_layout::col> && std::is_same_v<typename T0::shape, typename ducks::rt_shape::rt_16x16>) {
-        using range_type_T0 = ducks::art::get_nth_range_t<typename T0::register_ranges, N * T0::width + M>;
-        using registers_T0 = ducks::art::split_many_t<ducks::art::type_list<range_type_T0>, 1>;
-        using range_type_T1 = ducks::art::get_nth_range_t<typename T1::register_ranges, N * T1::width + M>;
-        using registers_T1 = ducks::art::split_many_t<ducks::art::type_list<range_type_T1>, 1>;
-        static_assert(registers_T0::size == registers_T1::size);
-
-        uint64_t window_mask = 0x1FFF01FF001F0001;
-        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 0>::lo, ducks::art::get_nth_range_t<registers_T1, 0>::lo, GPR>(window_mask);
-
-        window_mask = 0x3FFF03FF003F0003;
-        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 1>::lo, ducks::art::get_nth_range_t<registers_T1, 1>::lo, GPR>(window_mask);
-
-        window_mask = 0x7FFF07FF007F0007;
-        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 2>::lo, ducks::art::get_nth_range_t<registers_T1, 2>::lo, GPR>(window_mask);
-
-        window_mask = 0xFFFF0FFF00FF000F;
-        macros::v_cndmask_b32_e64<ducks::art::get_nth_range_t<registers_T0, 3>::lo, ducks::art::get_nth_range_t<registers_T1, 3>::lo, GPR>(window_mask);
-    } else {
-        static_assert(false, "Unsupported window mask");
-    }
-}
+}
@@ -471,20 +471,6 @@ class TestCmpFloat(unittest.TestCase):
    st = run_program(instructions, n_lanes=1)
    self.assertEqual(st.vcc & 1, 1, "Expected vcc=1 (1.0 != 2.0)")

-  def test_v_cmp_eq_f16_src0_hi(self):
-    """v_cmp_eq_f16 with src0 from high half (true16 384+n encoding)."""
-    cmp = v_cmp_eq_f16_e32(v[0], v[1])
-    cmp._raw += 128  # src0 v[0] -> v[0].h, the dsl can't encode hi-half src0 yet
-    instructions = [
-      s_mov_b32(s[0], 0x42003c00),  # hi=3.0, lo=1.0
-      v_mov_b32_e32(v[0], s[0]),
-      s_mov_b32(s[0], 0x47004200),  # hi=7.0, lo=3.0
-      v_mov_b32_e32(v[1], s[0]),
-      cmp,
-    ]
-    st = run_program(instructions, n_lanes=1)
-    self.assertEqual(st.vcc & 1, 1, "Expected vcc=1 (v0.hi 3.0 == v1.lo 3.0)")
-
  def test_v_cmp_nge_f16_inf_self(self):
    """v_cmp_nge_f16 comparing -inf with itself (unordered less than).

@@ -4,13 +4,13 @@ from tinygrad import Tensor, GlobalCounters, dtypes, nn, Device, Variable
 from tinygrad.helpers import Context, getenv, DEV
 from tinygrad.engine.realize import run_linear, estimate_uop, compile_linear
 from tinygrad.renderer.ptx import PTXRenderer
-from test.helpers import needs_second_gpu, check_schedule, assert_kernel_count, KernelCountException
+from test.helpers import needs_second_gpu

 class TestArange(unittest.TestCase):
  def _get_flops(self, tensor, desired):
    GlobalCounters.reset()
    linear = compile_linear(tensor.schedule_linear())
-    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
+    self.assertEqual(len(linear.src), 1)
    run_linear(linear)
    np.testing.assert_equal(tensor.numpy(), desired)
    return estimate_uop(linear.src[-1]).ops
@@ -55,7 +55,8 @@ class TestIndexing(unittest.TestCase):
    with Context(NOOPT=1):
      GlobalCounters.reset()
      out = ((Tensor.arange(1,16385)-1)*needle).sum()
-      linear, var_vals = check_schedule(out, 1)
+      linear, var_vals = out.linear_with_vars()
+      self.assertEqual(len(linear.src), 1)
      run_linear(linear, var_vals)
    self.assertEqual(out.item(), 1337)

@@ -71,7 +72,8 @@ class TestIndexing(unittest.TestCase):
      reshape_dataset = dataset.T.reshape(1, DDIM, DSET, 1).expand(4, DDIM, DSET, 1)
      full = (rng==idxs).where(reshape_dataset, Tensor.zeros(4, DDIM, DSET, 1, buffer=False))
      X = full.sum(axis=(2,3))
-      linear, var_vals = check_schedule(X, 1)
+      linear, var_vals = X.linear_with_vars()
+      self.assertEqual(len(linear.src), 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -96,7 +98,8 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      X = dataset[idxs]
      assert X.shape == (4,DDIM)
-      linear, var_vals = check_schedule(X, 1)
+      linear, var_vals = X.linear_with_vars()
+      self.assertEqual(len(linear.src), 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -110,7 +113,8 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      X = dataset[idxs]
      assert X.shape == (4,DDIM)
-      linear, var_vals = check_schedule(X, 1)
+      linear, var_vals = X.linear_with_vars()
+      self.assertEqual(len(linear.src), 1)
      run_linear(linear, var_vals)
      assert GlobalCounters.global_ops < 4*DSET, f"too many ops {GlobalCounters.global_ops} != {4*DSET}"
    np.testing.assert_allclose(real_index, X.numpy())
@@ -153,7 +157,7 @@ class TestIndexing(unittest.TestCase):
      GlobalCounters.reset()
      z = emb(x).realize()
      self.assertLessEqual(GlobalCounters.global_ops, op_limit)
-      assert_kernel_count(2)
+      self.assertEqual(GlobalCounters.kernel_count, 2)
    if getenv("CHECK", 1):
      import torch
      with torch.no_grad():
@@ -253,7 +257,7 @@ class TestIndexing(unittest.TestCase):
    xq_rope, _ = apply_rotary_emb(xq, xq, freqs_cis)
    xq_rope.sum().backward()
    linear = compile_linear(wq.grad.schedule_linear())
-    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
+    assert len(linear.src) == 1, f"expected one kernel for backward, got: {len(linear.src)}"
    bwd_ops = estimate_uop(linear.src[0]).ops
    expected_ops = bs*seqlen*dim*dim*ops_scale
    print(f"rope matmul bwd ({dtype}): {GlobalCounters.kernel_count} kernels, {bwd_ops:,} ops")
@@ -1,7 +1,7 @@
 import unittest
 from tinygrad import Tensor, Device, dtypes, Context
 from tinygrad.helpers import getenv, system, DEV
-from extra.gemm.cdna_asm_gemm import asm_gemm, hk_bf16_atb_gemm
+from extra.gemm.cdna_asm_gemm import asm_gemm, hk_bf16_atb_gemm, quantize_mxfp4
 from test.helpers import needs_second_gpu
 from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8, FP8_MAX

@@ -157,20 +157,13 @@ class TestMXFP4(unittest.TestCase):

  def test_quantize(self):
    import numpy as np
-    from extra.llama_kernels.quantize_mxfp4 import quantize_mxfp4
-    rng = np.random.default_rng(0)
-    x = np.triu(rng.standard_normal((256, 256), dtype=np.float32))
-    x += np.triu(x, 1).T
-    x[:32, :32] = 0
-    row, row_scale, col, col_scale = quantize_mxfp4(Tensor(x, dtype=dtypes.bfloat16))
-    Tensor.realize(row, row_scale, col, col_scale)
-    row, row_scale = row.numpy(), row_scale.numpy()
-    col, col_scale = col.numpy(), col_scale.numpy()
-    np.testing.assert_array_equal(row, col)
-    np.testing.assert_array_equal(row_scale, col_scale)
-    self.assertTrue(row.any())
-    self.assertTrue((row_scale == 127).any())
-    self.assertTrue((row_scale != 127).any())
+    block = np.array([0, .26, .74, .75, 1.26, 1.75, 2.51, 3.5, 5.1, 6, -6] + [0] * 21, dtype=np.float32)
+    x = Tensor(np.tile(block, (32, 8)), dtype=dtypes.bfloat16)
+    packed, scale, _ = quantize_mxfp4(x)
+    p = packed.numpy()
+    codes = np.stack((p & 0xF, p >> 4), axis=-1).reshape(32, 256)
+    np.testing.assert_array_equal(codes[0, :11], [0, 1, 1, 2, 3, 4, 5, 6, 7, 7, 15])
+    np.testing.assert_array_equal(scale.numpy(), np.full((32, 8), 127, dtype=np.uint8))

  def test_correctness(self):
    import numpy as np
@@ -178,9 +171,17 @@ class TestMXFP4(unittest.TestCase):
    rng = np.random.default_rng(1)
    a = Tensor(rng.standard_normal((M, K), dtype=np.float32), dtype=dtypes.bfloat16)
    b = Tensor(rng.standard_normal((N, K), dtype=np.float32), dtype=dtypes.bfloat16)
-    out = asm_gemm(a, b.T, mxfp4=True).realize().numpy().astype(np.float32)
-    ref = a.numpy().astype(np.float32) @ b.numpy().astype(np.float32).T
-    self.assertLess(np.linalg.norm(out-ref) / np.linalg.norm(ref), 0.2)
+    out = asm_gemm(a, b.T, mxfp4=True).realize()
+    # reference gemm
+    a_packed, scale_a, _ = quantize_mxfp4(a)
+    b_packed, scale_b, _ = quantize_mxfp4(b)
+    def unpack(x): return np.stack((x & 0xF, x >> 4), axis=-1).reshape(x.shape[0], -1)
+    code_a, code_b = unpack(a_packed.numpy()), unpack(b_packed.numpy())
+    lut = np.array([0, .5, 1, 1.5, 2, 3, 4, 6, -0., -.5, -1, -1.5, -2, -3, -4, -6], dtype=np.float32)
+    a_dequant = lut[code_a] * np.repeat(np.exp2(scale_a.numpy().astype(np.int16)-127), 32, axis=1)
+    b_dequant = lut[code_b] * np.repeat(np.exp2(scale_b.numpy().astype(np.int16)-127), 32, axis=1)
+    ref = Tensor(a_dequant @ b_dequant.T, dtype=dtypes.bfloat16).realize().numpy()
+    np.testing.assert_array_equal(out.numpy(), ref)

  def test_empty(self):
    M, N, K = getenv("M", 16384), getenv("N", 4096), getenv("K", 14336)
@@ -4,7 +4,6 @@ import numpy as np
 from tinygrad.dtype import AddrSpace, dtypes, Invalid
 from tinygrad.uop.ops import KernelInfo, AxisType, Ops
 from tinygrad.renderer.ptx import PTXRenderer
-from test.helpers import assert_kernel_count

 # **** kernels ****

@@ -190,12 +189,6 @@ class TestCustomKernel(unittest.TestCase):
    b = Tensor.custom_kernel(tst, a, fxn=custom_sum)[0]
    self.assertEqual(b.item(), 15)

-  def test_sum_outside(self):
-    a = Tensor([1.0, 2, 3, 4, 5])+1
-    tst = Tensor.empty(1)
-    b = Tensor.custom_kernel(tst, a, fxn=custom_sum)[0]
-    self.assertEqual(b.item(), 20)
-
  def test_sum_int(self):
    a = Tensor([1, 2, 3, 4, 5])
    tst = Tensor.empty(1, dtype=a.dtype)
@@ -283,7 +276,7 @@ class TestCustomKernel(unittest.TestCase):

    GlobalCounters.reset()
    out.realize()
-    assert_kernel_count(5)
+    self.assertEqual(GlobalCounters.kernel_count, 5)

  def test_simple_reshape(self):
    a = Tensor.ones(2,3,4).realize()
@@ -293,7 +286,7 @@ class TestCustomKernel(unittest.TestCase):
    GlobalCounters.reset()
    c.realize()
    assert all(i == 3. for i in c.flatten().tolist()), f"all 3 {c.tolist()}"
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 3)

  def test_multi_after_schedule_order(self):
    """Test correct scheduling order when custom_kernel has multiple outputs.
@@ -343,7 +336,7 @@ class TestCustomKernel(unittest.TestCase):
    c = Tensor.custom_kernel(c, a, fxn=custom_add_one_kernel)[0]
    GlobalCounters.reset()
    c.realize()
-    assert_kernel_count(len(devs))
+    self.assertEqual(GlobalCounters.kernel_count, len(devs))
    self.assertTrue((c == 2).all().item())

  def test_partial_invalid_store_keeps_uncovered_reads(self):
@@ -408,11 +401,13 @@ class TestCustomKernel(unittest.TestCase):
    else: z = y.T.T+1
    GlobalCounters.reset()
    z.realize()
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 2)
    self.assertEqual(z.tolist(), x.add(2).tolist())

+  @unittest.expectedFailure
  def test_custom_kernel_sched_copy(self): self.test_custom_kernel_sched(use_custom=True)

+  @unittest.expectedFailure
  def test_sliced_buffer_function(self):
    x = Tensor.arange(32).reshape(8, 4).clone().realize()
    from tinygrad import function
@@ -422,8 +417,8 @@ class TestCustomKernel(unittest.TestCase):
      return Tensor.custom_kernel(y, x, fxn=custom_add_one_kernel)[0]
    GlobalCounters.reset()
    y = run(x[0]).realize()
-    # backends that support contiguous views don't launch extra kernels
-    assert_kernel_count(2 if x[0].uop.contiguous_view() is None else 1)
+    # it's copying the input and the output
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(y.tolist(), [1, 2, 3, 4])

  @Context(DEV="CPU")
@@ -433,37 +428,11 @@ class TestCustomKernel(unittest.TestCase):
    # TODO: it currently requires a compiler for Ops.BINARY
    from tinygrad.device import Device
    binary = Device[a.device].renderer.compiler.compile(src)
-    def custom_src_kernel(A:UOp, B:UOp) -> UOp:
+    def custom_src_kernel(A:UOp) -> UOp:
      sink = UOp.sink(A, arg=KernelInfo(name="test_src"))
      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(sink.toposort())), UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=binary)))
-    a = Tensor.custom_kernel(a.reshape(2, 2).clone(), a.reshape(2, 2).T, fxn=custom_src_kernel)[0]
-    self.assertEqual(a.tolist(), [[1, 1], [2, 3]])
-
-  @Context(DEV="CPU")
-  def test_simple_from_source_alt(self):
-    a = Tensor.arange(4).clone().realize()
-    src = "void copy(int* restrict out, int* restrict in) { for (int i = 0; i < 4; i++) out[i] = in[i]; }"
-    from tinygrad.device import Device
-    binary = Device[a.device].renderer.compiler.compile(src)
-    def custom_src_kernel(out:UOp, inp:UOp) -> UOp:
-      sink = UOp.sink(out, inp, arg=KernelInfo(name="copy"))
-      return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=tuple(sink.toposort())), UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=binary)))
-    out = Tensor.custom_kernel(Tensor.empty_like(a), a+1, fxn=custom_src_kernel)[0]
-    GlobalCounters.reset()
-    out.realize()
-    assert_kernel_count(2)
-    self.assertEqual(out.tolist(), [1, 2, 3, 4])
-
-  @unittest.skip("this shouldn't be expected to work")
-  def test_inplace_transpose(self):
-    def custom_assign_row_max_kernel(A:UOp) -> UOp:
-      row = UOp.range(A.shape[0], 0)
-      col = UOp.range(A.shape[1], 1)
-      return A[row, col].store(A[row].max(axis=0)).end(col).end(row).sink(arg=KernelInfo(name=f"assign_row_max_{A.numel()}"))
-    a = Tensor.arange(4).clone().realize()
-    a = Tensor.custom_kernel(a.reshape(2, 2).T, fxn=custom_assign_row_max_kernel)[0]
-    self.assertEqual(a.flatten().tolist(), [2, 2, 3, 3])
-    self.assertEqual(a.shape, (2, 2))
+    a = Tensor.custom_kernel(a.reshape(2, 2).T, fxn=custom_src_kernel)[0]
+    self.assertEqual(a.tolist(), [[1, 2], [1, 3]])

 class TestCustomKernelInput(unittest.TestCase):
  def _test_mop(self, mop_fxn, max_kernels):
@@ -491,8 +460,8 @@ class TestCustomKernelInput(unittest.TestCase):

  def test_reshape(self): self._test_mop(lambda x: x.reshape(16, 2), max_kernels=2)
  def test_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T, max_kernels=3)
-  def test_double_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T.T, max_kernels=2)
-  def test_shrink(self): self._test_mop(lambda x: x[:4], max_kernels=1)
+  def test_double_permute(self): self._test_mop(lambda x: x.reshape(4, 8).T.T, max_kernels=3)
+  def test_shrink(self): self._test_mop(lambda x: x[:4], max_kernels=2)
  def test_pad(self): self._test_mop(lambda x: x[:4].pad(((0, 4),)), max_kernels=2)
  def test_flip(self): self._test_mop(lambda x: x.flip(0), max_kernels=2)
  def test_offset_shrink(self): self._test_mop(lambda x: x[4:8], max_kernels=2)
@@ -169,13 +169,6 @@ class TestFp8sConversions(unittest.TestCase):
  def test_fp8e5m2fnuz_to_float(self, x):
    np.testing.assert_equal(fp8_to_float(x, dtypes.fp8e5m2fnuz), torch.tensor(x, dtype=torch.uint8).view(torch.float8_e5m2fnuz).float().item())

-  def test_fp8e5m2fnuz_to_float_smallest_normals(self):
-    # fnuz bias exceeds half's, so exp-1 normals land below half's normal range: they flush to zero like denormals
-    if dtypes.half not in supported_dtypes or dtypes.half in EMULATED_DTYPES.tolist(dtypes) or dtypes.fp8e5m2fnuz in supported_dtypes:
-      self.skipTest("needs the emulated fp8 with a native half intermediate")
-    vals = Tensor([0x04, 0x05, 0x06, 0x07], dtype=dtypes.uint8).bitcast(dtypes.fp8e5m2fnuz).float().numpy()
-    np.testing.assert_equal(vals, [0., 0., 0., 0.])
-
 class TestBFloat16DType(unittest.TestCase):
  def test_bf16_to_float(self):
    _test_cast(Tensor([100000], dtype=dtypes.bfloat16), dtypes.float32)
@@ -399,10 +399,9 @@ class TestDTypeALU(unittest.TestCase):
    if float_dtype not in supported_dtypes: float_dtype = dtypes.float32
    universal_test_cast(a, float_dtype, unsigned_dtype)

-  def test_unsafe_cast_float_to_int(self):
-    # the value is off the float32 grid but rounds in-range: the buffer and const-fold paths must agree
-    # (out-of-range float->int cast stays undefined: hardware may saturate where the fold wraps)
-    val = 2147483000.0
+  @unittest.expectedFailure
+  def test_unsafe_cast_float_to_int_failure(self):
+    val = float(dtypes.int32.max - 1)
    t1 = Tensor([val], dtype=dtypes.float32).cast(dtypes.int32)
    t2 = Tensor(val, dtype=dtypes.float32).cast(dtypes.int32)
    np.testing.assert_equal(t1.item(), t2.item())
@@ -2,7 +2,7 @@
 import unittest
 import numpy as np

-from test.helpers import assert_jit_cache_len, call_is_graph, not_support_multi_device, needs_second_gpu, KernelCountException
+from test.helpers import assert_jit_cache_len, call_is_graph, not_support_multi_device, needs_second_gpu
 from test.unit.test_jit import _simple_test
 from tinygrad import Tensor, Variable, TinyJit, Device, dtypes
 from tinygrad.engine.jit import graph_class
@@ -97,7 +97,7 @@ class TestJit(unittest.TestCase):
      prev = o

    # Checking that 2 graphs are inited.
-    if len(jf.captured.linear.src) != 2: raise KernelCountException(2, len(jf.captured.linear.src))
+    assert len(jf.captured.linear.src) == 2
    for si in jf.captured.linear.src:
      assert call_is_graph(si)

@@ -360,7 +360,7 @@ class TestJitGraphSplit(unittest.TestCase):
    self.expect(f, inp, inp_cpu,
      graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
      multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()],
-      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp()]) # cpu is hcq2 now, it does not join hcq graphs
+      hcqgraph=[self.ji_graph(4)])

  def test_jit_cpu_several(self):
    if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")
@@ -377,9 +377,9 @@ class TestJitGraphSplit(unittest.TestCase):
    inp = Tensor.randn(10, 10, device=Device.DEFAULT).realize()
    inp_cpu = Tensor.randn(10, 10, device="CPU").realize()
    self.expect(f, inp, inp_cpu,
-      graph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
-      multigraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()],
-      hcqgraph=[self.ji_graph(2), self.ji_comp(), self.ji_comp(), self.ji_comp()])
+      graph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
+      multigraph=[self.ji_graph(2), self.ji_graph(2), self.ji_comp()],
+      hcqgraph=[self.ji_graph(5)])

  def test_jit_multidev(self):
    if Device.DEFAULT == "CPU": raise unittest.SkipTest("CPU is not a valid default device for this test")
@@ -12,7 +12,7 @@ from tinygrad.dtype import DType, dtypes, AddrSpace
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.cstyle import CUDARenderer
 from tinygrad.renderer.isa import ISARenderer
-from test.helpers import replace_opts, check_schedule
+from test.helpers import replace_opts
 from test.backend.test_softmax_fusion import single_kernel_softmax
 MOCKGPU = DEV.interface.startswith("MOCK")

@@ -30,7 +30,7 @@ class TestLinearizer(unittest.TestCase):
    c = ((a.shrink(((0, 2),)) - a.shrink(((2, 4),))) - (b.shrink(((0, 2),)) - b.shrink(((2, 4),))))
    linear = c.schedule_linear()
    run_linear(linear)
-    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if not s.is_bound_var]
+    rawbufs = [s.buffer for s in linear.src[-1].src[1:] if s.op is not Ops.BIND]
    assert len(rawbufs) == 3 and set(rawbufs[1:]) == {a.uop.base.realized, b.uop.base.realized}
    np_c = (np_a[:2] - np_a[2:]) - (np_b[:2] - np_b[2:])
    np.testing.assert_allclose(np_c, c.numpy(), atol=1e-4, rtol=1e-4)
@@ -293,7 +293,8 @@ class TestLinearizer(unittest.TestCase):
    a = Tensor.ones(4, 4).contiguous().realize()
    b = a.shrink(((1, 2), None)).pad(((1, 2), None)).bool()
    a.assign(b.where(2, a))
-    linear, var_vals = check_schedule(a, 1)
+    linear, var_vals = a.linear_with_vars()
+    assert len(linear.src) == 1
    run_linear(linear, var_vals)
    np.testing.assert_equal(a.flatten().numpy(), [1.,1.,1.,1.,2.,2.,2.,2.,1.,1.,1.,1.,1.,1.,1.,1.])
    program = to_program(replace_opts(linear.src[-1].src[0], []), renderer=Device[Device.DEFAULT].renderer)
@@ -411,7 +412,7 @@ def helper_realized_ast(r:Tensor|list[Tensor]) -> tuple[UOp, list[Buffer]]:
  last_call = linear.src[-1]
  ast = last_call.src[0]
  assert ast.op is Ops.SINK, f"helper_realized_ast expects a SINK {last_call}"
-  last_bufs = [s.buffer for s in last_call.src[1:] if not s.is_bound_var]
+  last_bufs = [s.buffer for s in last_call.src[1:] if s.op is not Ops.BIND]
  # now all input buffers in last_call should be realized
  # create fresh buffers for the outputs
  bufs = [Buffer(x.device, x.size, x.dtype).allocate() if i < len(ast.src) else x for i,x in enumerate(last_bufs)]
@@ -437,7 +438,7 @@ def reset_bufs(bufs:list[Buffer]):
  for buf in bufs: buf.copy_from(Buffer("PYTHON", buf.size, buf.dtype, opaque=memoryview(bytearray(buf.nbytes))))

 def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[],
-                               apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[], check_default_opt=True):
+                               apply_tc=False, atol=1e-4, rtol=1e-4, color_sizes=[], wanna_output=[]):
  outbufs = real_bufs[:len(realized_ast.src)]
  wanna_output = [np.array(x).flatten() for x in wanna_output]
  buf_uops = [UOp.new_buffer(b.device, b.size, b.dtype) for b in real_bufs]
@@ -459,7 +460,9 @@ def _helper_linearizer_opt_ast(realized_ast:UOp, real_bufs:list[Buffer], opts=[]
    for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)

  # Check correctness of handcoded optimiztions.
-  if check_default_opt: check_opt(None)
+  reset_bufs(outbufs)
+  run_prg(opts=None)
+  for buf,want in zip(copyout_outputs(outbufs), wanna_output): np.testing.assert_allclose(buf, want, atol=atol, rtol=rtol)
  for x in opts: # Check custom transformations if any.
    check_opt(([Opt(OptOps.TC, 0, (TC_SELECT.value, TC_OPT.value, 1))] if apply_tc else [])+x)

@@ -5,10 +5,9 @@ from examples.mlperf.models.flat_llama import FP8_DTYPE, quantize_fp8
 from extra.llama_kernels.fused_ce import fused_ce_loss
 from extra.llama_kernels import local_abs_max
 from extra.llama_kernels.quantize_fp8_delayed import quantize_fp8_delayed, quantize_fp8_scalar
-from extra.llama_kernels.swiglu import swiglu
 from extra.models.llama import apply_rotary_emb, precompute_freqs_cis
 from extra.thunder.amd.fa import custom_fused_qkv_rope_backward, fused_qkv_rope
-from test.helpers import needs_second_gpu, assert_kernel_count
+from test.helpers import needs_second_gpu
 from test.backend.test_asm_gemm import has_hipcc

 def run_fused_ce(bs:int, seqlen:int, vocab:int, label_smoothing:float=0.0) -> None:
@@ -96,7 +95,7 @@ class TestLocalAmax(unittest.TestCase):
    x = Tensor.arange(16).reshape(4, 4).cast(dtypes.float).clone(devices[0]).realize().shard(devices, axis=0).realize()
    GlobalCounters.reset()
    out = (x * local_abs_max(x)).clone().realize()
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 2)
    self.assertEqual(out.tolist(), [[0., 7., 14., 21.], [28., 35., 42., 49.], [120., 135., 150., 165.], [180., 195., 210., 225.]])

@unittest.skipUnless(has_hipcc() and Device.DEFAULT == "AMD", "requires hipcc to compile and amd device to run")
@@ -162,31 +161,5 @@ class TestFusedQKVRoPE(unittest.TestCase):
    ref = Tensor.cat(dq_ref, dk_ref, dv_ref, dim=3).reshape(*dx.shape).realize()
    with Context(DEBUG=0): self.assertTrue(dx.allclose(ref, atol=2e-2, rtol=2e-2).item(), "backward mismatch")

-def run_swiglu(test:unittest.TestCase, shape:tuple[int, ...]) -> None:
-  Tensor.manual_seed(0)
-  x = (Tensor.randn(*shape) * 2).cast(dtypes.bfloat16).realize()
-  hidden = x.shape[-1] // 2
-  out, ref = swiglu(x), x[..., :hidden].silu() * x[..., hidden:]
-  Tensor.realize(out, ref)
-  with Context(DEBUG=0): test.assertTrue(out.allclose(ref, atol=2.5e-1, rtol=3e-2).item(), "SwiGLU forward mismatch")
-
-  grad = (Tensor.randn(*out.shape) * 2).cast(dtypes.bfloat16).realize()
-  grad_x, grad_ref = out.gradient(x, gradient=grad)[0], ref.gradient(x, gradient=grad)[0]
-  Tensor.realize(grad_x, grad_ref)
-  test.assertEqual(grad_x.shape, shape)
-  test.assertEqual(grad_x.dtype, dtypes.bfloat16)
-  with Context(DEBUG=0): test.assertTrue(grad_x.allclose(grad_ref, atol=2.5e-1, rtol=3e-2).item(), "SwiGLU backward mismatch")
-
-class TestSwiGLU(unittest.TestCase):
-  def setUp(self):
-    if dtypes.bfloat16 not in Device[Device.DEFAULT].renderer.supported_dtypes(): self.skipTest("need bfloat16")
-
-  def test_simple(self): run_swiglu(self, (2, 32, 64))
-
-  def test_llama_shape(self):
-    if Device.DEFAULT != "AMD" or not Device[Device.DEFAULT].renderer.target.arch.startswith("gfx950"):
-      self.skipTest("only run on real machine for speed")
-    run_swiglu(self, (2, 8192, 28672))
-
 if __name__ == '__main__':
  unittest.main()
@@ -1,12 +1,12 @@
 import unittest, random
 from tinygrad import Tensor, Device, nn, GlobalCounters, TinyJit, dtypes, Variable
-from tinygrad.uop.ops import Ops, UOp, AxisType, graph_rewrite
+from tinygrad.uop.ops import Ops, UOp, AxisType
 from tinygrad.helpers import getenv, prod, Context
 from tinygrad.nn.state import get_parameters
-from tinygrad.engine.realize import run_linear, compile_linear, pm_beam, pm_compile
+from tinygrad.engine.realize import run_linear, compile_linear
 import numpy as np
 from hypothesis import given, strategies as strat, settings
-from test.helpers import not_support_multi_device, needs_second_gpu, slow, call_is_graph, check_schedule, assert_kernel_count
+from test.helpers import not_support_multi_device, needs_second_gpu, slow, call_is_graph

 settings.register_profile("my_profile", max_examples=200, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
 settings.load_profile("my_profile")
@@ -62,7 +62,7 @@ class TestMultiTensor(unittest.TestCase):
  def test_shard_empty(self):
    GlobalCounters.reset()
    X = Tensor.empty(256).shard(devices_2, 0).realize()
-    assert_kernel_count(0)
+    assert GlobalCounters.kernel_count == 0
    (X + X).realize()

  # TODO: fix this to not copy on the src device
@@ -79,9 +79,9 @@ class TestMultiTensor(unittest.TestCase):
  def test_shard_beam(self):
    cpu_2 = ("CPU:1", "CPU:2")
    src = Tensor.ones(16).shard(cpu_2, 0).realize()
-    lin = UOp(Ops.LINEAR, src=(src.to(cpu_2[::-1]).schedule_linear().src[0],))
-    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): call = graph_rewrite(graph_rewrite(lin, pm_beam, ctx=1, walk=True), pm_compile, walk=True).src[0]
-    self.assertNotEqual(call.src[0].src[0].arg.applied_opts, ())
+    pad = src.to(cpu_2[::-1]).schedule_linear().src[0]
+    with Context(BEAM=1, IGNORE_BEAM_CACHE=1): prg = compile_linear(UOp(Ops.LINEAR, src=(pad,))).src[0].src[0]
+    self.assertNotEqual(prg.src[0].arg.applied_opts, ())

  def test_shard_same_device(self):
    X = Tensor.ones(256).contiguous().realize()
@@ -355,7 +355,8 @@ class TestMultiTensor(unittest.TestCase):
  def test_const_like_shrink_on_shard_axis(self):
    t = Tensor.ones(16, 16, dtype=dtypes.int).shard(devices_2, axis=0)
    out = t.const_like(2)[:, :8]
-    linear, var_vals = check_schedule(out, 0)
+    linear, var_vals = out.linear_with_vars()
+    self.assertEqual(len(linear.src), 0)
    run_linear(linear, var_vals)
    self.assertEqual(out.tolist(), [[2]*8]*16)

@@ -3,11 +3,11 @@ import unittest
 import numpy as np
 import torch
 from tinygrad import Tensor, Device, TinyJit, dtypes
+from tinygrad.uop.ops import Ops
 from tinygrad.helpers import GlobalCounters, Context
 from tinygrad.nn import Conv1d, ConvTranspose1d, Conv2d, ConvTranspose2d, Linear, Embedding
 from tinygrad.nn import BatchNorm, LayerNorm, LayerNorm2d, GroupNorm, InstanceNorm, RMSNorm, LSTMCell
 from tinygrad.nn.state import load_state_dict
-from test.helpers import check_schedule
 from tinygrad.engine.realize import run_linear
 from test.helpers import not_support_multi_device, needs_second_gpu, slow

@@ -428,14 +428,18 @@ class TestNN(unittest.TestCase):
    a = Tensor([[1, 5, 9, 11],
                [12, 19, 8, 1]])
    result = layer(a)
-    linear, var_vals = check_schedule(result, kcount)
+    linear, var_vals = result.linear_with_vars()
+    self.assertEqual(len([call for call in linear.src if call.src[0].op is Ops.SINK]), kcount,
+                     "first run realizes weight and embedding")
    run_linear(linear, var_vals)

    b = Tensor([[1, 2, 3],
                [4, 5, 6],
                [7, 8, 9]])
    result = layer(b)
-    linear, var_vals = check_schedule(result, 1)
+    linear, var_vals = result.linear_with_vars()
+    self.assertEqual(1, len([call for call in linear.src if call.src[0].op is Ops.SINK]),
+                     "second run realizes embedding only")
    run_linear(linear, var_vals)
    print(f"Embedding used {GlobalCounters.global_ops} ops")
    self.assertLessEqual(GlobalCounters.global_ops, ops)
@@ -728,17 +728,6 @@ class TestOps(unittest.TestCase):
        else:
          self.assertAlmostEqual(tiny_out, torch_out, msg=f"{x}, {c}")

-  def test_pow_neg_inf_frac_exponent(self):
-    # pow(-inf, 0.3) is +inf, so the gradient 0.3*pow(-inf, -0.7) is 0, never nan
-    helper_test_op(None, lambda x: x**0.3, vals=[[-math.inf]])
-    # is_odd truncates, so it calls 3.3 odd: only the non_int guard keeps pow(-inf, 3.3) from negating to -inf
-    helper_test_op(None, lambda x: x**3.3, vals=[[-math.inf]])
-
-  def test_pow_zero_exponent(self):
-    # x ** 0 is the constant 1 for every x, so the gradient with respect to the base is 0, never nan
-    # TODO: nan ** 0, failed on WEBGPU
-    helper_test_op(None, lambda x,y: x**y, vals=[[-math.inf, math.inf, 0.0], [0.0, 0.0, 0.0]])
-
  def test_pow_zero_tensor(self):
    helper_test_op(None, lambda x,y: x**y, vals=[[0.0], [0.0]])
    # TODO: fix WEBGPU
@@ -1535,8 +1524,6 @@ class TestOps(unittest.TestCase):

  def test_prod(self):
    helper_test_op(None, lambda x: x.prod(), vals=[[1.0, 2.0, 3.0]])
-    helper_test_op(None, lambda x: x.prod(), vals=[[0.0, 2.0, 3.0]])
-    helper_test_op(None, lambda x: x.prod(), vals=[[0.0, 0.0, 3.0]])
    with Context(NOOPT=1): helper_test_op(None, lambda x: x.prod(), vals=[[1.0, 2.0, 3.0]])
    helper_test_op([(3,4,5,6)], lambda x: x.prod(dim=3), lambda x: x.prod(axis=3))
    helper_test_op([(3,4,5,6)], lambda x: x.prod(dim=1), lambda x: x.prod(axis=1))
@@ -8,7 +8,6 @@ from tinygrad.helpers import prod
 from tinygrad.renderer.cstyle import CStyleLanguage
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.wgsl import WGSLRenderer
-from test.helpers import check_schedule
 from tinygrad.runtime.ops_python import PythonRenderer
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, python_alu
 from tinygrad.tensor import Tensor
@@ -62,7 +61,8 @@ class TestCStyleFailures(unittest.TestCase):
    dtype = "bool" if op in (Ops.OR, Ops.XOR, Ops.AND) else None
    ret = Tensor.empty(1, dtype=dtype)
    for _ in range(5): ret = python_alu[op](ret, Tensor.empty(1, dtype=dtype))
-    linear, _ = check_schedule(ret, 1)
+    linear = ret.schedule_linear()
+    assert len(linear.src) == 1
    src = to_program(linear.src[0].src[0], Device[Device.DEFAULT].renderer).src[2].arg
    self.assertEqual("("*5 not in src, should_strip_paren)

@@ -6,13 +6,34 @@ import unittest, time
 import numpy as np

 from tinygrad import nn, dtypes, Device, Tensor, Variable
-from tinygrad.uop.ops import Ops, UPat
-from tinygrad.helpers import DEV, GlobalCounters, Context, all_same, temp
-from tinygrad.engine.realize import run_linear
-from test.helpers import check_schedule, assert_kernel_count
+from tinygrad.uop.ops import UOp, Ops, UPat
+from tinygrad.helpers import DEBUG, DEV, GlobalCounters, Context, all_same, temp
+from tinygrad.engine.realize import compile_linear, run_linear

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

+class KernelCountException(Exception): pass
+def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Tensor]|None=None, filter_sink=True):
+  if to_prerealize:
+    with Context(DEBUG=0, TRACK_MATCH_STATS=0): Tensor.realize(*to_prerealize)
+  if isinstance(t, Tensor): linear, var_vals = t.linear_with_vars()
+  elif isinstance(t, list) and isinstance(t[0], Tensor): linear, var_vals = Tensor.linear_with_vars(*t)
+  else:
+    assert isinstance(t, UOp), f"can't schedule {t}"
+    linear, var_vals = Tensor(t).linear_with_vars()
+  kernel_cnt = sum((len(call.device) if isinstance(call.device, tuple) else 1)
+                   for call in linear.src if call.src[0].op is Ops.SINK or not filter_sink)
+  if kernel_cnt != allowed:
+    print(f"SCHEDULE ISSUE, expecting {allowed} got {kernel_cnt}")
+    if DEBUG >= 3:
+      for i,call in enumerate(linear.src):
+        print("kernel", i+1)
+        print(call.src[0])
+    raise KernelCountException(f"{kernel_cnt} != {allowed}")
+  # test compiling the linear
+  compile_linear(linear)
+  return linear, var_vals
+
 def _realize_weights(m):
  for p in nn.state.get_parameters(m): p.realize()

@@ -92,9 +113,11 @@ class TestSchedule(unittest.TestCase):
    a2 = mop(a)
    expected = (a+a2).tolist()
    a.assign(a+a2)
-    linear, var_vals = check_schedule(a, expected_kcount)
+    linear, var_vals = a.linear_with_vars()
+    kcount = len(linear.src)
    run_linear(linear, var_vals)
    self.assertListEqual(a.tolist(), expected)
+    self.assertEqual(kcount, expected_kcount)
  def test_setitem_permuted_sched(self): self.test_setitem_sched(lambda x: x.T, 2)
  def test_setitem_paddded_sched(self): self.test_setitem_sched(lambda x: x.shrink_to(4, 1).pad_to(4, 4), 1)

@@ -103,9 +126,9 @@ class TestSchedule(unittest.TestCase):
    a = Tensor.arange(16).clone().realize()
    GlobalCounters.reset()
    a[4] = 3
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    a.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertListEqual(a.tolist(), [0, 1, 2, 3, 3, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15])

  def test_no_extra_contiguous_on_setitem_assign_back(self):
@@ -4,7 +4,6 @@ from tinygrad import Tensor, GlobalCounters, Context, Device
 from tinygrad.dtype import DTypeLike, dtypes
 from tinygrad.engine.realize import run_linear
 from tinygrad.helpers import DEBUG, get_single_element
-from test.helpers import check_schedule

 def single_kernel_softmax(x_in:Tensor, axis=-1, dtype:DTypeLike|None=None) -> Tensor:
  # only support axis =-1
@@ -104,7 +103,8 @@ class TestFuse(unittest.TestCase):
    k = (x @ wk).contiguous()
    v = (x @ wv).contiguous()
    attn = q.scaled_dot_product_attention(k, v)
-    check_schedule(attn, 4) # 3 matmul and 1 attention
+    s = attn.schedule_linear()
+    self.assertEqual(len(s.src), 4) # 3 matmul and 1 attention

  @unittest.skip("needs RANGEIFY>1")
  def test_flash_attention(self):
@@ -6,15 +6,6 @@ from examples.gpt2 import Attention
 import numpy as np

 class TestSymbolicOps(unittest.TestCase):
-  def test_negative_slice(self):
-    a = Tensor.rand(3, 10, 4)
-    for i in range(3, 10):
-      vi = Variable("i", 1, 10).bind(i)
-      # negative int bounds against a symbolic dim must resolve against the size, like slice.indices
-      np.testing.assert_allclose(a[:, :vi][:, -3:-1].numpy(), a[:, :i][:, -3:-1].numpy(), atol=1e-6, rtol=1e-6)
-      np.testing.assert_allclose(a[:, :vi][:, -1:].numpy(), a[:, :i][:, -1:].numpy(), atol=1e-6, rtol=1e-6)
-      np.testing.assert_allclose(a[:, :vi][:, -1].numpy(), a[:, :i][:, -1].numpy(), atol=1e-6, rtol=1e-6)
-
  def test_plus1(self):
    def f(a): return (a+1).realize()
    a = Tensor.rand(3, 10)
@@ -2,7 +2,7 @@ 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 check_schedule
+from test.backend.test_schedule import check_schedule
 from test.backend.test_dtype_alu import ht, dtypes_float
 import numpy as np
 import math
@@ -330,7 +330,6 @@ class TestHCQ(unittest.TestCase):
  # Test profile api
  def test_speed_exec_time(self):
    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
-    st = time.perf_counter()
    TestHCQ.d0.hw_compute_queue_t().timestamp(sig_st) \
                                   .exec(TestHCQ.runtime, TestHCQ.kernargs_ba_ptr, TestHCQ.prg.arg.global_size, TestHCQ.prg.arg.local_size) \
                                   .timestamp(sig_en) \
@@ -338,13 +337,11 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    host_us = (time.perf_counter() - st) * 1e6

    et = float(sig_en.timestamp - sig_st.timestamp)

    print(f"exec kernel time: {et:.2f} us")
-    # emulated devices are only bounded by the host window around submit+wait
-    assert 0.1 <= et <= (host_us if MOCKGPU or Device.DEFAULT in {"CPU"} else 100)
+    assert 0.1 <= et <= (3000000 if MOCKGPU or Device.DEFAULT in {"CPU"} else 100)

  def test_speed_copy_bandwidth(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -355,7 +352,6 @@ class TestHCQ(unittest.TestCase):
    b = Buffer(Device.DEFAULT, SZ, dtypes.uint8, options=BufferSpec(nolru=True)).allocate()

    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
-    st = time.perf_counter()
    TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
                                .copy(a._buf, b._buf, SZ) \
                                .timestamp(sig_en) \
@@ -363,14 +359,13 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    host_ms = (time.perf_counter() - st) * 1e3

-    et_ms = float(sig_en.timestamp - sig_st.timestamp) / 1e3
-    assert 0 < et_ms <= host_ms # timestamps are in us and cover only the copy
+    et = float(sig_en.timestamp - sig_st.timestamp)
+    et_ms = et / 1e3

    gb_s = ((SZ / 1e9) / et_ms) * 1e3
    print(f"same device copy:  {et_ms:.2f} ms, {gb_s:.2f} GB/s")
-    assert (0 if MOCKGPU else 10) <= gb_s <= 1000
+    assert (0.2 if MOCKGPU else 10) <= gb_s <= 1000

  def test_speed_cross_device_copy_bandwidth(self):
    if TestHCQ.d0.hw_copy_queue_t is None: self.skipTest("device does not support copy queue")
@@ -384,7 +379,6 @@ class TestHCQ(unittest.TestCase):
    TestHCQ.d0.allocator._map(b._buf)

    sig_st, sig_en = TestHCQ.d0.new_signal(), TestHCQ.d0.new_signal()
-    st = time.perf_counter()
    TestHCQ.d0.hw_copy_queue_t().timestamp(sig_st) \
                                .copy(a._buf, b._buf, SZ) \
                                .timestamp(sig_en) \
@@ -392,14 +386,13 @@ class TestHCQ(unittest.TestCase):

    TestHCQ.d0.timeline_signal.wait(TestHCQ.d0.timeline_value)
    TestHCQ.d0.timeline_value += 1
-    host_ms = (time.perf_counter() - st) * 1e3

-    et_ms = float(sig_en.timestamp - sig_st.timestamp) / 1e3
-    assert 0 < et_ms <= host_ms # timestamps are in us and cover only the copy
+    et = float(sig_en.timestamp - sig_st.timestamp)
+    et_ms = et / 1e3

    gb_s = ((SZ / 1e9) / et_ms) * 1e3
    print(f"cross device copy: {et_ms:.2f} ms, {gb_s:.2f} GB/s")
-    assert (0 if MOCKGPU else 2) <= gb_s <= 100
+    assert (0.2 if MOCKGPU else 2) <= gb_s <= 100

  def test_timeline_signal_rollover(self):
    for queue_type in [TestHCQ.d0.hw_compute_queue_t, TestHCQ.d0.hw_copy_queue_t]:
@@ -1,14 +0,0 @@
-import unittest, numpy as np
-from unittest.mock import patch
-from tinygrad import Device, Tensor
-from tinygrad.helpers import getenv
-from tinygrad.runtime.support.hcq2 import HCQ_DEVS, all_devices_in
-
-@unittest.skipUnless(getenv("HCQ2") and all_devices_in(Device.DEFAULT, HCQ_DEVS), "hcq2 device required")
-class TestHCQ2(unittest.TestCase):
-  def test_copy_without_copy_queue(self):
-    with patch.object(Device[Device.DEFAULT], "has_copy_queue", False):
-      np.testing.assert_equal(Tensor(np.arange(61, dtype=np.float32)).to(Device.DEFAULT).contiguous().realize().numpy(), np.arange(61))
-
-if __name__ == "__main__":
-  unittest.main()
@@ -1,17 +0,0 @@
-from tinygrad import Device, Tensor, TinyJit, dtypes
-from tinygrad.helpers import Timing, Context
-
-GPUS, DEPTH, SZ = 8, 4, 128 * 2**20
-WARMUP, ITERS = 3, 5
-devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(GPUS))
-bufs = tuple(Tensor.empty(SZ, dtype=dtypes.uint8, device=dev).contiguous().realize() for _ in range(DEPTH) for dev in devs)
-
-@TinyJit
-def all_to_all(*srcs:Tensor): return Tensor.realize(*(src.to(dst) for i,src in enumerate(srcs) for j,dst in enumerate(devs) if i % GPUS != j))
-
-if __name__ == "__main__":
-  with Context(ALL2ALL=1, JIT_BATCH_SIZE=0):
-    for i in range(-WARMUP, ITERS):
-      with Timing("ALL2ALL ", lambda ns: f" {SZ*GPUS*(GPUS-1)*DEPTH/ns:.2f} GB/s", enabled=i>=0):
-        all_to_all(*bufs)
-        for dev in devs: Device[dev].synchronize()
@@ -1,5 +1,5 @@
-import unittest, time, itertools
-from tinygrad import Tensor, Context
+import unittest, time
+from tinygrad import Tensor

 class TestScheduleScaling(unittest.TestCase):
  """Test that .schedule() scales linearly with graph size (no O(n^2) behavior)."""
@@ -130,18 +130,5 @@ class TestScheduleScaling(unittest.TestCase):
      return parts[0].cat(*parts[1:])
    self._assert_linear(concat_chain)

-  @Context(DEV="NULL:HIP:gfx1100")
-  def test_custom_kernel_assign_scaling(self):
-    from tinygrad.uop.ops import UOp, Ops, KernelInfo
-    from tinygrad.runtime.autogen.amd.rdna3.ins import s_nop
-    count = itertools.count(0)
-    def custom_kernel_assign(n):
-      def custom_asm(out):
-        return UOp(Ops.PROGRAM, src=(UOp.sink(out, arg=KernelInfo(f"fxn_{next(count)}")),
-                                     UOp(Ops.LINEAR, src=tuple(UOp(Ops.INS, arg=s_nop(i)) for i in range(n*8)))))
-      call = Tensor.custom_kernel(Tensor.empty(1), fxn=custom_asm)[0]
-      return Tensor.cat(*[Tensor.empty(1).assign(call+i) for i in range(n)])
-    self._assert_linear(custom_kernel_assign, n_small=50, n_large=500)
-
 if __name__ == '__main__':
  unittest.main(verbosity=2)
@@ -8,11 +8,10 @@ from tinygrad.tensor import _to_np_dtype
 from tinygrad.codegen import to_program
 from tinygrad.dtype import DType, truncate
 from tinygrad.nn.state import get_parameters
-from tinygrad.helpers import T, Target, DEV, DEBUG, Context, GlobalCounters
+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
-from tinygrad.engine.realize import compile_linear

 # 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")
@@ -35,36 +34,6 @@ def derandomize_model(model):
    p.replace(Tensor.empty(p.shape, device=p.device, dtype=p.dtype))
    p.realize()

-class KernelCountException(Exception):
-  def __init__(self, expected:int, got:int):
-    self.expected, self.got = expected, got
-    super().__init__(f"expected {expected}, got {got}")
-
-def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Tensor]|None=None, filter_sink=True):
-  if to_prerealize:
-    with Context(DEBUG=0, TRACK_MATCH_STATS=0): Tensor.realize(*to_prerealize)
-  if isinstance(t, Tensor): linear, var_vals = t.linear_with_vars()
-  elif isinstance(t, list) and isinstance(t[0], Tensor): linear, var_vals = Tensor.linear_with_vars(*t)
-  else:
-    assert isinstance(t, UOp), f"can't schedule {t}"
-    linear, var_vals = Tensor(t).linear_with_vars()
-  # test compiling the linear
-  compile_linear(linear)
-  kernel_cnt = sum((len(call.device) if isinstance(call.device, tuple) else 1)
-                   for call in linear.src if call.src[0].op is Ops.SINK or not filter_sink)
-  if kernel_cnt != allowed:
-    print(f"SCHEDULE ISSUE, expecting {allowed} got {kernel_cnt}")
-    if DEBUG >= 3:
-      for i,call in enumerate(linear.src):
-        print("kernel", i+1)
-        print(call.src[0])
-    raise KernelCountException(allowed, kernel_cnt)
-  return linear, var_vals
-
-def assert_kernel_count(expected:int):
-  got = GlobalCounters.kernel_count
-  if got != expected: raise KernelCountException(expected, got)
-
 def call_is_graph(call:UOp) -> bool:
  ast = call.src[0]
  return ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph"
@@ -84,15 +53,15 @@ def jit_cache_count(linear:UOp) -> int:
 def assert_jit_cache_len(fxn, expected_len):
  linear = fxn.captured.linear if fxn.captured is not None else None
  if linear is None or not linear.src:
-    if expected_len != 0: raise KernelCountException(expected_len, 0)
+    assert expected_len == 0, expected_len
    return
-  if expected_len and all(call_is_hcq(call) for call in linear.src): expected_len = 4 # HCQ2: fence + reset + merged same-queue calls + finalizer
+  if expected_len and all(call_is_hcq(call) for call in linear.src): expected_len = 3 # HCQ2: merged same-queue calls + finalizer + bumps
  if call_is_graph(linear.src[0]):
-    if len(linear.src) != 1: raise KernelCountException(1, len(linear.src))
+    assert len(linear.src) == 1, len(linear.src)
    inner = linear.src[0].src[0].src[0]  # LINEAR UOp inside CUSTOM_FUNCTION
-    if len(inner.src) != expected_len: raise KernelCountException(expected_len, len(inner.src))
+    assert len(inner.src) == expected_len, f"expected {expected_len}, got {len(inner.src)}"
  else:
-    if len(linear.src) != expected_len: raise KernelCountException(expected_len, len(linear.src))
+    assert len(linear.src) == expected_len, f"expected {expected_len}, got {len(linear.src)}"

 def min_normal(dt:DType) -> float: return 2.0 ** (2 - (1 << (dtypes.finfo(dt)[0] - 1)))

@@ -109,7 +109,7 @@ def _init_sqtt_encoder():
  _SMEM = (ir3.SMEM, ir4.SMEM, irc.SMEM)
  _VALU = (ir3.VOP1, ir3.VOP2, ir3.VOP3, ir3.VOP3P, ir3.VOPC, ir3.VOPD, ir3.VOP3SD, ir3.VOP3_SDST, ir3.VOP1_SDST,
           ir4.VOP1, ir4.VOP2, ir4.VOP3, ir4.VOP3P, ir4.VOPC, ir4.VOPD, ir4.VOP3SD, ir4.VOP3_SDST, ir4.VOP1_SDST,
-           irc.VOP1, irc.VOP2, irc.VOP3, irc.VOP3P, irc.VOP3PX2, irc.VOPC, irc.VOP3SD, irc.VOP3_SDST)
+           irc.VOP1, irc.VOP2, irc.VOP3, irc.VOP3P, irc.VOPC, irc.VOP3SD, irc.VOP3_SDST)
  _DS = (ir3.DS, ir4.DS, irc.DS)
  _GLOBAL = (ir3.GLOBAL, ir4.VGLOBAL, irc.GLOBAL)
  _FLAT = (ir3.FLAT, ir4.VFLAT, irc.FLAT)
@@ -1150,9 +1150,6 @@ def _compile_vopc(inst: ir3.VOPC|ir3.VOPC_DPP16|ir3.VOP3|ir4.VOPC|ir4.VOPC_DPP16
  def get_cmp_bit(lane) -> UOp:
    lc = lane.cast(dtypes.int) if isinstance(lane, UOp) else _c(lane, dtypes.int)
    s0 = _load_dpp16_src0(ctx, inst, lc, _c(0)) if is_dpp16 else ctx.rsrc_dyn(src0_off, lc, bits['s0'], literal, is_f64)
-    if is_vopc and not isinstance(inst, irc.VOPC) and bits['s0'] == 16 and not is_dpp16:
-      src0_hi = src0_off >= _c(384)
-      s0 = src0_hi.where(_hi16(ctx.rvgpr_dyn(src0_hi.where(src0_off - _c(384), _c(0)), lc)), s0)
    s1 = _cond_hi16(vsrc1_hi, ctx.rsrc_dyn(src1_off, lc, bits['s1'], literal, is_f64)) if bits['s0'] == 16 \
      else ctx.rsrc_dyn(src1_off, lc, bits['s1'], literal, is_f64)
    if bits['s0'] == 16 and opsel: s0, s1 = _apply_opsel(s0, 0, opsel), _apply_opsel(s1, 1, opsel)
@@ -1326,7 +1323,7 @@ def _compile_vop3sd(inst: ir3.VOP3SD | ir4.VOP3SD | irc.VOP3SD, ctx: _Ctx) -> UO
  else:
    return ctx.compile_vop_pcode(inst.op, srcs, lane, vdst_reg, exec_mask, sdst_reg=inst.sdst.offset)

-def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
+def _compile_mfma(inst: irc.VOP3P, ctx: _Ctx) -> UOp:
  """CDNA MFMA matrix multiply-accumulate emulation.

  Uses local temp arrays to cache inputs, avoiding aliasing issues when vdst overlaps src0/src1.
@@ -1352,25 +1349,6 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
  src0_is_vgpr = src0_off >= _c(256)
  src1_is_vgpr = src1_off >= _c(256)

-  scaled = isinstance(inst, irc.VOP3PX2)
-  if scaled:
-    assert isinstance(inst, irc.VOP3PX2)
-    # F8F6F4 input formats: 0=FP8(E4M3), 1=BF8(E5M2). FP6/FP4 (2-4) not emulated.
-    src0_fmt, src1_fmt = int(inst.cbsz), int(inst.blgp)
-    if src0_fmt > 1 or src1_fmt > 1: raise RuntimeError(f"unsupported scaled MFMA formats cbsz={src0_fmt} blgp={src1_fmt}")
-    # scale_src0/scale_src1 are source operands pointing at 32-bit registers holding 4 packed E8M0 scale exponents.
-    # The 2-bit opsel/opsel_hi select which byte applies to A/B for this instruction.
-    scale0_off = ctx.inst_field(type(inst).scale_src0)
-    scale1_off = ctx.inst_field(type(inst).scale_src1)
-    sel0, sel1 = int(inst.opsel) & 3, int(inst.opsel_hi) & 3
-    def _scale_exp(off: UOp, sel: int, lane: UOp) -> UOp:
-      sv = ctx.rsrc_dyn(off, lane, 32)
-      byte = (sv >> UOp.const(sel * 8, dtypes.uint32)) & UOp.const(0xFF, dtypes.uint32)
-      return byte.cast(dtypes.int32) - UOp.const(127, dtypes.int32)
-    # combined A*B scale for this lane: 2^(ea-127) * 2^(eb-127)
-    def scale_factor(lane: UOp) -> UOp:
-      return UOp.exp2((_scale_exp(scale0_off, sel0, lane) + _scale_exp(scale1_off, sel1, lane)).cast(dtypes.float32))
-
  m = _re.search(r'(\d+)X(\d+)X(\d+)', op_name)
  if m is None: raise ValueError(f"could not parse MFMA dimensions from {op_name}")
  M, N, K = int(m.group(1)), int(m.group(2)), int(m.group(3))
@@ -1426,18 +1404,7 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
  # The optimizer folds bitcast(uint32→float32) stores to float32 arrays, losing the conversion.
  tmp = UOp.placeholder((n_a_elems + n_b_elems,), dtypes.uint32, slot=0, addrspace=AddrSpace.LOCAL)

-  # Per-operand fp8 format ("fp8"=E4M3, "bf8"=E5M2) for A and B
-  if 'F8F6F4' in op_name:
-    assert isinstance(inst, (irc.VOP3P_MFMA, irc.VOP3PX2))
-    _fmts = {0: "fp8", 1: "bf8"}
-    a_fmt, b_fmt = _fmts.get(int(inst.cbsz), "fp8"), _fmts.get(int(inst.blgp), "fp8")
-  elif is_fp8:
-    # A/B formats from name suffix, e.g. V_MFMA_F32_16X16X32_BF8_FP8
-    suffixes = op_name.rsplit('_', 2)[-2:]
-    a_fmt, b_fmt = ("bf8" if sfx == "BF8" else "fp8" for sfx in suffixes)
-  else: a_fmt = b_fmt = "fp8"
-
-  def cvt_elem(raw: UOp, sub_idx: int, fp8_fmt: str = "fp8") -> UOp:
+  def cvt_elem(raw: UOp, sub_idx: int) -> UOp:
    if is_i8:
      # Extract i8, sign-extend to i32
      byte_val = (raw >> UOp.const(sub_idx * 8, dtypes.uint32)) & UOp.const(0xFF, dtypes.uint32)
@@ -1445,7 +1412,7 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
    elif is_f32_src:
      return raw  # already uint32 (f32 bit pattern)
    elif is_fp8:
-      return _FUNCS[f"{fp8_fmt}_to_f32"](raw >> UOp.const(sub_idx * 8, dtypes.uint32)).bitcast(dtypes.uint32)
+      return ((raw >> UOp.const(sub_idx * 8, dtypes.uint32)) & UOp.const(0xFF, dtypes.uint32)).cast(dtypes.uint32)
    elif is_bf16:
      # bf16→f32 bits: just shift left by 16 (bf16 is upper 16 bits of f32)
      return ((raw >> UOp.const(sub_idx * 16, dtypes.uint32)) & UOp.const(0xFFFF, dtypes.uint32)) << UOp.const(16, dtypes.uint32)
@@ -1487,7 +1454,7 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
    # Read A/B sources. Use rsrc_dyn for inline constants/SGPRs (src_off < 256), rvgpr_dyn for VGPRs (src_off >= 256).
    a_raw = src0_is_vgpr.where(ctx.rvgpr_dyn(src0_r + _c(reg_idx), read_lane),
                                ctx.rsrc_dyn(src0_off, _c(0, dtypes.int), 32))
-    a_val = cvt_elem(a_raw, sub_idx, a_fmt)
+    a_val = cvt_elem(a_raw, sub_idx)
    if M == 4:
      a_idx = grp_idx * UOp.const(M * K, dtypes.int) + mn_idx * UOp.const(K, dtypes.int) + UOp.const(kl, dtypes.int)
    else:
@@ -1496,7 +1463,7 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:

    b_raw = src1_is_vgpr.where(ctx.rvgpr_dyn(src1_r + _c(reg_idx), read_lane),
                                ctx.rsrc_dyn(src1_off, _c(0, dtypes.int), 32))
-    b_val = cvt_elem(b_raw, sub_idx, b_fmt)
+    b_val = cvt_elem(b_raw, sub_idx)
    if M == 4:
      b_idx = b_off + grp_idx * UOp.const(N * K, dtypes.int) + mn_idx * UOp.const(K, dtypes.int) + UOp.const(kl, dtypes.int)
    else:
@@ -1513,17 +1480,6 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
  #   Actually: 16 ACCVGPRs per lane, organized as 4 groups (l//32 gives half, each half has 2 sub-groups) of 4 rows
  tmp2 = tmp.after(read_phase)

-  def _dot_accum(acc: UOp, a_row: UOp, b_row: UOp, lane: UOp) -> UOp:
-    """acc += sum_k A[a_row+k] * B[b_row+k]. For scaled MFMA, only the dot product is scaled: D = dot*scale + C."""
-    def prod(k: int) -> UOp:
-      return tmp2.index(a_row + UOp.const(k, dtypes.int)).bitcast(acc_dt) * tmp2.index(b_row + UOp.const(k, dtypes.int)).bitcast(acc_dt)
-    if not scaled:
-      for k in range(K): acc = acc + prod(k)
-      return acc
-    dot = prod(0)
-    for k in range(1, K): dot = dot + prod(k)
-    return acc + dot * scale_factor(lane)
-
  compute_lane = ctx.range()
  compute_stores = []

@@ -1554,7 +1510,10 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
      else: acc_v = acc_v.bitcast(dtypes.float32)
      acc = src2_is_vgpr.where(acc_v, acc_scalar)

-      acc = _dot_accum(acc, m_base * UOp.const(K, dtypes.int), b_off + n_idx * UOp.const(K, dtypes.int), compute_lane)
+      for k in range(K):
+        a_val = tmp2.index(m_base * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+        b_val = tmp2.index(b_off + n_idx * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+        acc = acc + a_val * b_val

      if is_int_out:
        compute_stores.append((ctx.waccvgpr_dyn if use_acc else ctx.wvgpr_dyn)(
@@ -1576,13 +1535,17 @@ def _compile_mfma(inst: irc.VOP3P|irc.VOP3PX2, ctx: _Ctx) -> UOp:
      if M == 4:
        # 4x4: each group is independent. A/B indexed per-group.
        m_base = c_grp * UOp.const(M * K, dtypes.int) + UOp.const(out_reg * K, dtypes.int)
-        b_base = b_off + c_grp * UOp.const(N * K, dtypes.int) + n_idx * UOp.const(K, dtypes.int)
+        for k in range(K):
+          a_val = tmp2.index(m_base + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+          b_val = tmp2.index(b_off + c_grp * UOp.const(N*K, dtypes.int) + n_idx * UOp.const(K, dtypes.int)+UOp.const(k, dtypes.int)).bitcast(acc_dt)
+          acc = acc + a_val * b_val
      else:
        # 16x16: K is split across groups. Shared MxK/NxK arrays.
        m_base = c_grp * UOp.const(out_per_lane, dtypes.int) + UOp.const(out_reg, dtypes.int)
-        b_base = b_off + n_idx * UOp.const(K, dtypes.int)
-
-      acc = _dot_accum(acc, m_base if M == 4 else m_base * UOp.const(K, dtypes.int), b_base, compute_lane)
+        for k in range(K):
+          a_val = tmp2.index(m_base * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+          b_val = tmp2.index(b_off + n_idx * UOp.const(K, dtypes.int) + UOp.const(k, dtypes.int)).bitcast(acc_dt)
+          acc = acc + a_val * b_val

      if is_int_out:
        compute_stores.append((ctx.waccvgpr_dyn if use_acc else ctx.wvgpr_dyn)(
@@ -1600,41 +1563,33 @@ def _compile_wmma(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P, ctx: _Ctx) -> UOp:
  vdst_reg = ctx.inst_field(type(inst).vdst)
  src0_r = ctx.inst_field(type(inst).src0) - _c(256)
  src1_r = ctx.inst_field(type(inst).src1) - _c(256)
-  src2_r = ctx.inst_field(type(inst).src2)
-  src2_r = (src2_r >= 256).where(src2_r - _c(256), src2_r)
-  output_type = op_name.split("WMMA_", 1)[1].split("_", 1)[0]
+  src2_r = ctx.inst_field(type(inst).src2) - _c(256)
+  is_f16_output = 'F16_16X16X16_F16' in op_name or 'BF16_16X16X16_BF16' in op_name  # F16/BF16 output vs F32 output
  is_bf16 = 'BF16' in op_name
  cvt = _FUNCS['bf16_to_f32'] if is_bf16 else _FUNCS['f16_to_f32']
  is_rdna4 = isinstance(inst, ir4.VOP3P)
-  sz = 8 if "8" in op_name else 16
-  # read matrix from VGPRs → flat f32/i32 array[row*16+k]
-  def gval(src, lane, vgpr, ridx):
+  # read 16x16 F16/BF16 matrix from VGPRs → flat f32 array[row*16+k]
+  def read_f16_val(src, lane, vgpr, half):
    v = ctx.rvgpr_dyn(src + _c(vgpr), UOp.const(lane, dtypes.int))
-    pkd = v >> UOp.const(ridx * sz, dtypes.uint32) if ridx > 0 else v
-    pkd = pkd & UOp.const((1 << sz) - 1, dtypes.uint32)
-    if "F" in output_type: return cvt(pkd)
-    return (pkd << _c(24, dtypes.uint)).bitcast(dtypes.int32) >> _c(24, dtypes.int32) # sign extend
+    return cvt((v >> UOp.const(16, dtypes.uint32)) if half else (v & UOp.const(0xFFFF, dtypes.uint32)))

-  # RDNA3 f16/bf16: 16 lanes × 8 VGPRs × 2 halves,    k maps linearly
-  # RDNA3 iu8:      16 lanes × 4 VGPRs × 4 quarters,  k maps linearly
-  # RDNA4:          32 lanes x 4 VGPRS x 2 halves, k bits are scrambled (k[2] goes to lane bit 4)
-  def read_mat(src):
-    n = 32 // sz # values per vgpr
-    # (row, k) → (lane, vgpr, row index)
+  # RDNA3: 16 lanes × 8 VGPRs × 2 halves, k maps linearly
+  # RDNA4: 32 lanes × 4 VGPRs × 2 halves, k bits are scrambled (k[2] goes to lane bit 4)
+  def read_f16_mat(src):
+  # (row, k) → (lane, vgpr, half)
    def ab_map(i, k):
      elem, lane = ((k & 3) | ((k >> 1) & 4), i + ((k >> 2) & 1) * 16) if is_rdna4 else (k, i)
-      return lane, elem // n, elem % n
-    return [gval(src, *ab_map(row, k)) for row in range(16) for k in range(16)]
-
-  mat_a, mat_b = read_mat(src0_r), read_mat(src1_r)
+      return lane, elem // 2, elem % 2
+    return [read_f16_val(src, *ab_map(row, k)) for row in range(16) for k in range(16)]
+  mat_a, mat_b = read_f16_mat(src0_r), read_f16_mat(src1_r)
  # (row, col) -> (lane, vgpr)
  def d_map(m, n):
    lane_bit, vgpr = (m >> 3, m & 7) if is_rdna4 else (m & 1, m >> 1)
    return n + lane_bit * 16, vgpr
-  if output_type in ["F16", "BF16"]:
+  if is_f16_output:
    # read accumulator C with f16 layout: for RDNA4, pairs of f32 vgprs pack into one f16 vgpr
    # for RDNA3, same layout as f32 but only lo 16 bits used
-    mat_c = [gval(src2_r, *((lane, vgpr // 2, vgpr % 2) if is_rdna4 else (lane, vgpr, 0)))
+    mat_c = [read_f16_val(src2_r, *((lane, vgpr // 2, vgpr % 2) if is_rdna4 else (lane, vgpr, 0)))
             for m in range(16) for n in range(16) for lane, vgpr in [d_map(m, n)]]
    mat_d = [sum(mat_a[r*16+k] * mat_b[c*16+k] for k in range(16)) + mat_c[r*16+c] for r in range(16) for c in range(16)]
    def f32_to_f16_bits(v: UOp) -> UOp: return v.cast(dtypes.half).bitcast(dtypes.uint16).cast(dtypes.uint32)
@@ -1647,22 +1602,18 @@ def _compile_wmma(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P, ctx: _Ctx) -> UOp:
    else:  # (rdna3) 1 f16 per VGPR (lo half only)
      stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1]), UOp.const(d_map(m, n)[0], dtypes.int), out_cvt(mat_d[m*16+n]), exec_mask)
                for m in range(16) for n in range(16)]
-  else: # f32/i32
-    out_dt = dtypes.float32 if output_type == "F32" else dtypes.int32
-    mat_c = [ctx.rvgpr_dyn(src2_r + _c(d_map(m, n)[1]), UOp.const(d_map(m, n)[0], dtypes.int)).bitcast(out_dt)
+  else: # f32
+    mat_c = [ctx.rvgpr_dyn(src2_r + _c(d_map(m, n)[1]), UOp.const(d_map(m, n)[0], dtypes.int)).bitcast(dtypes.float32)
             for m in range(16) for n in range(16)]
    mat_d = [sum(mat_a[r*16+k] * mat_b[c*16+k] for k in range(16)) + mat_c[r*16+c] for r in range(16) for c in range(16)]
    stores = [ctx.wvgpr_dyn(vdst_reg + _c(d_map(m, n)[1]), UOp.const(d_map(m, n)[0], dtypes.int), mat_d[m*16+n].bitcast(dtypes.uint32), exec_mask)
              for m in range(16) for n in range(16)]
  return UOp.sink(*stores, *ctx.inc_pc())

-def _compile_vop3p(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P | irc.VOP3PX2, ctx: _Ctx) -> UOp:
+def _compile_vop3p(inst: ir3.VOP3P | ir4.VOP3P | irc.VOP3P, ctx: _Ctx) -> UOp:
  op_name = _op_name(inst)
-  if 'WMMA' in op_name:
-    assert not isinstance(inst, irc.VOP3PX2)
-    return _compile_wmma(inst, ctx)
-  if 'MFMA' in op_name and any(f'{s}X{s}X' in op_name for s in ('4', '16', '32')) and isinstance(inst, (irc.VOP3P, irc.VOP3PX2)):
-    return _compile_mfma(inst, ctx)
+  if 'WMMA' in op_name and ('16X16X16_F16' in op_name or '16X16X16_BF16' in op_name): return _compile_wmma(inst, ctx)
+  if 'MFMA' in op_name and any(f'{s}X{s}X' in op_name for s in ('4', '16', '32')) and isinstance(inst, irc.VOP3P): return _compile_mfma(inst, ctx)

  # ACCVGPR_WRITE/READ/MOV: copies between VGPR and ACCVGPR register files
  # Detect by checking operand types for ACCVGPR involvement
@@ -2093,7 +2044,7 @@ _INST_HANDLERS: dict[type, Callable[..., UOp]] = {
  irc.SOPP: _compile_sopp, irc.SMEM: _compile_smem, irc.SOP1: _compile_sop, irc.SOP2: _compile_sop, irc.SOPC: _compile_sop, irc.SOPK: _compile_sop,
  irc.VOP1: _compile_vop12, irc.VOP1_DPP16: _compile_vop12, irc.VOP2: _compile_vop12, irc.VOP2_DPP16: _compile_vop12,
  irc.VOPC: _compile_vopc, irc.VOP3: _compile_vop3,
-  irc.VOP3_SDST: _compile_vop3, irc.VOP3SD: _compile_vop3sd, irc.VOP3P: _compile_vop3p, irc.VOP3PX2: _compile_vop3p,
+  irc.VOP3_SDST: _compile_vop3, irc.VOP3SD: _compile_vop3sd, irc.VOP3P: _compile_vop3p,
  irc.VOP1_SDWA: _compile_sdwa, irc.VOP2_SDWA: _compile_sdwa, irc.VOP2_SDWA_SDST: _compile_sdwa, irc.VOPC_SDWA_SDST: _compile_sdwa,
  irc.DS: _compile_mem_op, irc.FLAT: _compile_mem_op, irc.GLOBAL: _compile_mem_op, irc.SCRATCH: _compile_mem_op,
  irc.MUBUF: _compile_mubuf,
@@ -74,7 +74,6 @@ class TestWhisper(unittest.TestCase):
        err
      )

-  @slow
  def test_transcribe_file1(self):
    self.assertEqual(transcribe_file(self.model, self.enc, TEST_FILE_1),  TRANSCRIPTION_1)

@@ -90,7 +89,6 @@ class TestWhisper(unittest.TestCase):
    self.assertEqual(TRANSCRIPTION_1,  transcriptions[0])
    self.assertEqual(TRANSCRIPTION_2,  transcriptions[1])

-  @slow
  def test_transcribe_batch21(self):
    waveforms = [load_file_waveform(TEST_FILE_2), load_file_waveform(TEST_FILE_1)]
    transcriptions = transcribe_waveform(self.model, self.enc, waveforms)
@@ -1,7 +1,7 @@
 import unittest
 from tinygrad import Tensor, dtypes, TinyJit, UOp
 from tinygrad.llm.model import apply_rope as apply_rope_new, precompute_freqs_cis
-from test.helpers import assert_jit_cache_len, check_schedule
+from test.helpers import assert_jit_cache_len

 def apply_rope(x:Tensor, start_pos:int):
  B, H, T, Hd = x.shape
@@ -16,8 +16,9 @@ class TestAttention(unittest.TestCase):
    k = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
    v = Tensor.ones(BS, seqlen, dim, dtype=dtypes.half).contiguous().realize()
    attn = q.scaled_dot_product_attention(k, v)
+    sched = attn.schedule_linear()
    # attention has 4 kernels now
-    check_schedule(attn, 4)
+    self.assertEqual(len(sched.src), 4)

  def test_apply_rope_jit_prune(self):
    def rope_fn(x_in, pos): return apply_rope(x_in, pos)
@@ -1,6 +1,6 @@
 import unittest, itertools, math
 from tinygrad import Tensor, dtypes, Context
-from tinygrad.dtype import DType, ConstType, truncate
+from tinygrad.dtype import DType, ConstType
 from tinygrad.uop.ops import Ops, UOp
 from test.helpers import full_rewrite
 import numpy as np
@@ -51,17 +51,6 @@ class TestWeakConstFolding(unittest.TestCase):
  def test_invalid_poison(self):
    self.assertTrue(UOp.invalid().alu(Ops.CDIV, UOp.const(0)).simplify().is_invalid)

-  def test_cast_commits_to_dtype_grid(self):
-    # committing a weak const to a stated width puts the value on that width's grid, same as storage packing and native compilers
-    v = 1/123008  # not representable in float16
-    out = UOp.const(v).cast(dtypes.half).simplify()
-    self.assertEqual((out.op, out.dtype, out.val), (Ops.CONST, dtypes.half, truncate[dtypes.half](v)))
-    self.assertNotEqual(out.val, v)
-    # the grid commit preserves the sign of zero
-    self.assertEqual(math.copysign(1, UOp.const(-0.0).cast(dtypes.half).simplify().val), -1)
-    # observable at tensor level: the const-folded comparison agrees with the committed value
-    self.assertTrue((Tensor(-3.2).cast(dtypes.float32) <= truncate[dtypes.float32](-3.2)).item())
-
 class TestBinaryOpsConstFolding(unittest.TestCase):
  def test_add_literal_zero(self):
    _check_ast_count(0, Tensor([1.0, 2, 3, 4]) + 0)
@@ -143,13 +143,13 @@ class TestModuloAndDivisionFolding(unittest.TestCase):
    self.assertIs(apply_rewrite(x_var_uop.cast(dtypes.weakint) % 10).render(simplify=False), x_var_uop.render(simplify=False))

  def test_full_graph_rewrite_division_with_remainder(self):
-    x_var_uop = UOp.variable('x', 7, 9, param=True)
+    x_var_uop = UOp.variable('x', 7, 9)
    optimized_sink = apply_rewrite(x_var_uop // 2)
    for x_value in range(7, 10):
      self.assertEqual(x_value // 2, evaluate_uop(optimized_sink, {'x': x_value}))

  def test_full_graph_rewrite_complex_mod_div_expression(self):
-    x_var_uop = UOp.variable('x', 1, 10, param=True)
+    x_var_uop = UOp.variable('x', 1, 10)
    optimized_sink = apply_rewrite(((x_var_uop * 5) % 3) // 2)
    for x_value in range(1, 11):
      original_result = ((x_value * 5) % 3) // 2
@@ -208,7 +208,7 @@ class TestGEPAndVectorizeRewrite(unittest.TestCase):


 import inspect
-from tinygrad.uop.ops import graph_rewrite, _substitute, rewrite_group
+from tinygrad.uop.ops import graph_rewrite, _substitute, track_rewrites
 from tinygrad.uop.symbolic import symbolic_simple

 class TestBottomUpRewrite(unittest.TestCase):
@@ -220,7 +220,7 @@ class TestBottomUpRewrite(unittest.TestCase):
    self.assertIs(gt, ret)

 # normally .substitute would be fine, but it's not tracked
-@rewrite_group()
+@track_rewrites()
 def named_substitute(name:str, uop:UOp, rel:dict[UOp, UOp]): return graph_rewrite(uop, _substitute, rel, bottom_up=True)
 def substitute(uop:UOp, rel:dict[UOp, UOp]): return named_substitute(inspect.stack()[1].function, uop, rel)

@@ -307,8 +307,8 @@ class TestRecurse(unittest.TestCase):
  def test_inf_loop(self):
    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
    ])
    with self.assertRaises(RuntimeError):
      graph_rewrite(a, pm)
@@ -316,8 +316,8 @@ class TestRecurse(unittest.TestCase):
  def test_inf_loop_bottom_up(self):
    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
    ])
    with self.assertRaises(RuntimeError):
      graph_rewrite(a, pm, bottom_up=True)
@@ -378,8 +378,8 @@ class TestWalkRewrite(unittest.TestCase):
    """A bouncing pattern applies once and stops instead of looping."""
    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
    ])
    with self.assertRaises(RuntimeError):
      graph_rewrite(a, pm, bottom_up=True)
@@ -456,8 +456,8 @@ class TestWalkRewrite(unittest.TestCase):
    """Bottom-up walk also applies once per node, no fixed-point iteration."""
    a = UOp.const(3)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
    ])
    ret = graph_rewrite(a, pm, bottom_up=True, walk=True)
    self.assertIs(ret, UOp.const(4))
@@ -511,7 +511,7 @@ class TestWalkRewrite(unittest.TestCase):
    def bpm_match(ctx, x):
      ctx.append((x.val if x.op is Ops.CONST else x.op, "bpm"))
      # rewrite const(1) -> const(10), short-circuiting its subtree
-      if x.op is Ops.CONST and x.val == 1: return UOp.const(10, x.dtype)
+      if x.op is Ops.CONST and x.val == 1: return x.replace(arg=10)
      return None
    def pm_match(ctx, x):
      ctx.append((x.val if x.op is Ops.CONST else x.op, "pm"))
@@ -109,28 +109,6 @@ class TestLLMServer(unittest.TestCase):

    self.assertGreater(len(contents), 0)

-  def test_interrupted_stream_logs_tokens(self):
-    with patch.object(self.mock_model, "generate", side_effect=lambda ids, **kwargs: iter([300, 301, 999])), \
-         patch("tinygrad.llm.serve.stderr_log") as log, patch("tinygrad.llm.serve.colored", side_effect=lambda text, color: text) as color:
-      stream = self.server.RequestHandlerClass.run_model(Mock(server=self.server), [200, 201, 202], "test")
-      next(stream)
-      next(stream)
-      stream.close()
-    interrupt = log.call_args.args[0]
-    self.assertFalse(interrupt.startswith("\n"))
-    self.assertTrue(interrupt.endswith("\n"))
-    self.assertIn("gen:", interrupt)
-    self.assertIn("out:    1", interrupt)
-    self.assertTrue(any(args[0].startswith("total:") and args[1] == "red" for args, _ in color.call_args_list))
-
-  def test_stream_disconnect_closes_source(self):
-    from tinygrad.llm.serve import Handler
-    source, handler = Mock(), Mock()
-    source.__iter__ = Mock(return_value=iter([{}]))
-    handler.wfile.write.side_effect = BrokenPipeError
-    Handler.stream_json(handler, source)
-    source.close.assert_called_once()
-
  def test_non_streaming(self):
    resp = self.client.chat.completions.create(
      model="test-model",
@@ -27,15 +27,10 @@ def _make_linear(buffer_lists, copies=None):
    calls.append(UOp(Ops.CALL, src=(src0, *bufs)))
  return UOp(Ops.LINEAR, src=tuple(calls))

-def _get_planned_view(buf:UOp) -> tuple[UOp, int, int]|None:
-  view = buf.src[0] if buf.op is Ops.BITCAST else buf
-  if view.op is not Ops.SHRINK or view.src[0].op is not Ops.BUFFER: return None
-  return (arena:=view.src[0]), view.src[1].val * arena.dtype.itemsize, view.src[2].val * arena.dtype.itemsize
-
 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 (planned:=_get_planned_view(new)) is not None: return planned[0]
+      if orig is buf and new.op is Ops.SLICE: return new.src[0]
  return None

 def check_assign(buffer_lists, copies=None):
@@ -46,8 +41,8 @@ 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 (planned:=_get_planned_view(new)) is not None and id(orig) not in replace_map:
-        replace_map[id(orig)] = planned
+      if new.op is Ops.SLICE and id(orig) not in replace_map:
+        replace_map[id(orig)] = (new.src[0], new.src[1].val * new.src[0].dtype.itemsize, new.arg * new.dtype.itemsize)

  # verify pinned buffers are not planned
  for buf in held_bufs:
@@ -5,7 +5,7 @@ from tinygrad.nn.state import get_parameters
 from tinygrad.engine.jit import TinyJit
 from tinygrad import Tensor, Device, GlobalCounters, dtypes, Variable
 from tinygrad.helpers import Context
-from test.helpers import slow, jit_cache_count, KernelCountException
+from test.helpers import slow, jit_cache_count
 from extra.lr_scheduler import OneCycleLR
 from test.helpers import derandomize_model

@@ -35,9 +35,8 @@ def helper_test(nm, gen, model, max_memory_allowed, max_kernels_allowed, all_jit
    assert mem_used < max_memory_allowed, f"{nm} used more than {max_memory_allowed:.3f} GB - {mem_used:.3} GB used"
    assert (max_memory_allowed - mem_used) / max_memory_allowed < 0.2, f"{max_memory_allowed:.3f} GB is too far from {mem_used:.3} GB used"
    if kernels_used:
-      if kernels_used > max_kernels_allowed: raise KernelCountException(max_kernels_allowed, kernels_used)
-      if (max_kernels_allowed - kernels_used) / max_kernels_allowed >= 0.2:
-        raise KernelCountException(max_kernels_allowed, kernels_used)
+      assert kernels_used <= max_kernels_allowed, f"{nm} used more than {max_kernels_allowed} kernels, it used {kernels_used}"
+      assert (max_kernels_allowed - kernels_used) / max_kernels_allowed < 0.2, f"{max_kernels_allowed=} is too far from {kernels_used=} used"
    if all_jitted:
      assert kernels_used > 0 and kernels_used == GlobalCounters.kernel_count or (kernels_used <= GlobalCounters.kernel_count and getattr(Device[Device.DEFAULT], "graph", None)), f"only {kernels_used} out of {GlobalCounters.kernel_count} were jitted"  # noqa: E501

@@ -2,11 +2,32 @@
 import gc, unittest, time
 from typing import cast
 from tinygrad import nn, dtypes, Device, Tensor, getenv
-from tinygrad.uop.ops import UOp, Ops, GroupOp, UPat, KernelInfo, AxisType
-from tinygrad.helpers import GlobalCounters, Context
-from tinygrad.engine.realize import run_linear, compile_linear
-from tinygrad.codegen import to_program, full_rewrite_to_sink
-from test.helpers import check_schedule, assert_kernel_count, KernelCountException
+from tinygrad.uop.ops import UOp, Ops, GroupOp, UPat, KernelInfo
+from tinygrad.helpers import DEBUG, GlobalCounters, Context
+from tinygrad.engine.realize import compile_linear, run_linear
+from tinygrad.codegen import to_program
+
+class KernelCountException(Exception): pass
+def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Tensor]|None=None, filter_sink=True):
+  if to_prerealize:
+    with Context(DEBUG=0, TRACK_MATCH_STATS=0): Tensor.realize(*to_prerealize)
+  if isinstance(t, Tensor): linear, var_vals = t.linear_with_vars()
+  elif isinstance(t, list) and isinstance(t[0], Tensor): linear, var_vals = Tensor.linear_with_vars(*t)
+  else:
+    assert isinstance(t, UOp), f"can't schedule {t}"
+    linear, var_vals = Tensor(t).linear_with_vars()
+  kernel_cnt = sum((len(call.device) if isinstance(call.device, tuple) else 1)
+                   for call in linear.src if call.src[0].op is Ops.SINK or not filter_sink)
+  if kernel_cnt != allowed:
+    print(f"SCHEDULE ISSUE, expecting {allowed} got {kernel_cnt}")
+    if DEBUG >= 3:
+      for i,call in enumerate(linear.src):
+        print("kernel", i+1)
+        print(call.src[0])
+    raise KernelCountException(f"{kernel_cnt} != {allowed}")
+  # test compiling the linear
+  compile_linear(linear)
+  return linear, var_vals

 def _realize_weights(m):
  for p in nn.state.get_parameters(m): p.realize()
@@ -122,7 +143,7 @@ class TestSimpleSchedule(unittest.TestCase):
    a = Tensor.empty(16,16).sum(axis=1)
    a1 = a.reshape(4,4)
    a2 = a.reshape(16,1,1)
-    check_schedule([a1, a2], 1)
+    self.assertEqual(len(Tensor.schedule_linear(a1, a2).src), 1)

 class TestSchedule(unittest.TestCase):
  def setUp(self):
@@ -134,7 +155,8 @@ class TestSchedule(unittest.TestCase):
  def test_arange_avgpool2d(self, kcount=1):
    x = Tensor.arange(25).reshape(1,1,5,5).cast(dtypes.float32)
    t = x.avg_pool2d(padding=1).clone()
-    check_schedule(t, kcount)
+    linear, var_vals = t.linear_with_vars()
+    self.assertEqual(len(linear.src), kcount)

  def test_arange_avgpool2d_fused_noopt(self):
    with Context(NOOPT=1): self.test_arange_avgpool2d(kcount=1)
@@ -202,7 +224,7 @@ class TestSchedule(unittest.TestCase):
    GlobalCounters.reset()
    expr = (a/b)/c
    expr.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertLessEqual(GlobalCounters.global_ops, 4*3)

  # NOTE: this is causing "LAZYCACHE=1 incorrectly reuses contiguous const" #4562
@@ -335,11 +357,6 @@ class TestSchedule(unittest.TestCase):
    out1 = a.sum() + b
    check_schedule([out0, out1], 2)

-  def test_reduce_broadcast_not_recomputed(self):
-    a = Tensor.empty(32, 16).realize()
-    out = a-a.mean(axis=0, keepdim=True)
-    check_schedule(out, 2)
-
  def test_scaled_dot_product_attention_multireduce_fusion(self):
    q = Tensor.empty(32,8,16,8).realize()
    k = Tensor.empty(32,8,16,8).realize()
@@ -592,7 +609,9 @@ class TestSchedule(unittest.TestCase):
    img = Tensor.randn(BS, CIN, 64, 64).realize()
    w = Tensor.uniform(16, CIN, 3, 3).realize()
    ret = Tensor.conv2d(img, w).relu().mean().backward()
-    check_schedule([ret, img.grad, w.grad], allowed)
+    linear, var_vals = Tensor.linear_with_vars(ret, img.grad, w.grad)
+    cnt = len([call for call in linear.src if call.src[0].op is Ops.SINK])
+    assert cnt == allowed, f"expected {allowed} kernels, got {cnt}"

  def test_conv2d_half(self): self.test_conv2d(4, dtype=dtypes.half)

@@ -613,8 +632,7 @@ class TestSchedule(unittest.TestCase):
        return len([call for call in linear.src if call.src[0].op is Ops.PROGRAM])

      with Context(IMAGE=1):
-        got = cnt()
-        if got != 5: raise KernelCountException(5, got)
+        self.assertEqual(cnt(), 5)

  def test_image_f16_residual_fusion(self):
    with Context(FLOAT16=1, OPENPILOT_HACKS=1):
@@ -629,8 +647,7 @@ class TestSchedule(unittest.TestCase):
        return len([call for call in linear.src if call.src[0].op is Ops.PROGRAM])

      with Context(IMAGE=1):
-        got = cnt()
-        if got != 9: raise KernelCountException(9, got)
+        self.assertEqual(cnt(), 9)

  def _test_fusion(self, shapes, f, cnt):
    with Context(DEBUG=0, TRACK_MATCH_STATS=0):
@@ -697,19 +714,6 @@ class TestSchedule(unittest.TestCase):
    xt = X[[Tensor([2]), Tensor([1])]]
    check_schedule(xt, 1)

-  def test_split_advanced_indexing_not_recomputed(self):
-    with Context(SPLIT_REDUCEOP=1):
-      X = Tensor.empty(32768, 4).realize()
-      idx = Tensor.randint(4, high=X.shape[0])
-      linear, _ = check_schedule(X[idx], 3, [Tensor._device_rng_counters[idx.device]])
-      # The split's final reduction remains, but the one-hot gather should collapse into a direct indexed load.
-      reduce_kernels = 0
-      for call in linear.src:
-        if call.src[0].op is not Ops.SINK: continue
-        sink = full_rewrite_to_sink(call.src[0], Device[call.device].renderer)
-        reduce_kernels += any(u.op is Ops.RANGE and u.arg[-1] is AxisType.REDUCE for u in sink.toposort())
-      self.assertEqual(reduce_kernels, 1)
-
  def test_push_through_reshape(self):
    x = Tensor.empty(10, 20).realize()
    out = x.argmax(1)
@@ -858,65 +862,6 @@ class TestSchedule(unittest.TestCase):
    x = Tensor.rand(32)
    check_schedule(x, 1, [Tensor._device_rng_counters[x.device]])

-  # **** custom kernel realize tests
-
-  @staticmethod
-  def _copy_fxn(name:str="copy"):
-    def copy_kernel(out:UOp, inp:UOp) -> UOp:
-      i = UOp.range(inp.numel(), 0)
-      return UOp.group(out[i].store(inp[i])).end(i).sink(arg=KernelInfo(name=name))
-    return copy_kernel
-
-  def _copy_call(self, out:Tensor, expr:Tensor, name:str="copy") -> Tensor:
-    # forge a custom kernel call with params and call args, like llm/kernels does (no Tensor.custom_kernel contiguous)
-    params = tuple(UOp.placeholder_like(u, slot=i) for i,u in enumerate((out.uop, expr.uop)))
-    return Tensor(out.uop.after(self._copy_fxn(name)(*params).call(out.uop, expr.uop)))
-
-  def test_custom_kernel_buffer_src(self):
-    # custom kernels need buffers: a buffer input must never add a realize kernel
-    y = Tensor.ones(64).contiguous().realize()
-    out = Tensor.empty_like(y)
-    check_schedule(self._copy_call(out, y), 1)
-
-  def test_custom_kernel_view_src(self):
-    # a RESHAPE over a buffer resolves to the buffer state (RESHAPEs on call args are stripped), no realize kernel
-    y = Tensor.ones(64).contiguous().realize()
-    out = Tensor.empty_like(y)
-    check_schedule(self._copy_call(out, y.reshape(8, 8).reshape(64)), 1)
-
-  def test_custom_kernel_elementwise_src(self):
-    # a computed input is not a buffer state: the call args are unwrapped to their base buffer,
-    # so the compute would be silently dropped. this must raise instead of producing wrong results
-    y = Tensor.ones(64).contiguous().realize()
-    out = Tensor.empty_like(y)
-    check_schedule(self._copy_call(out, y + y), 2)
-
-  def test_custom_kernel_lazy_const_src(self):
-    # a lazy const expression above the call has no buffer at all. this used to crash rangeify with a KeyError
-    x = Tensor.linspace(-1.0, 1.0, 64)
-    out = Tensor.empty_like(x)
-    check_schedule(self._copy_call(out, x), 2)
-
-  def test_custom_kernel_offset_view_src(self):
-    # a SHRINK with an offset over a buffer is not a buffer state either, the offset would be silently dropped
-    y = Tensor.ones(128).contiguous().realize()
-    out = Tensor.empty(64)
-    check_schedule(self._copy_call(out, y[16:80]), 2)
-
-  def test_custom_kernel_computed_src_api(self):
-    # the supported way to pass computed inputs: Tensor.custom_kernel makes inputs contiguous (one realize kernel)
-    y = Tensor.ones(64).contiguous().realize()
-    out = Tensor.empty_like(y)
-    check_schedule(Tensor.custom_kernel(out, y + y, fxn=self._copy_fxn())[0], 2)
-
-  def test_custom_kernel_on_custom_kernel(self):
-    # the output of a custom kernel is a buffer state, chaining custom kernels must not add kernels
-    y = Tensor.ones(64).contiguous().realize()
-    k1 = self._copy_call(Tensor.empty_like(y), y, name="k1")
-    k2 = self._copy_call(Tensor.empty_like(y), k1, name="k2")
-    sched, _ = check_schedule(k2, 2)
-    self.assertEqual([call.src[0].arg.name for call in sched.src], ["k1", "k2"])
-
  def test_empty_is_not_realized(self):
    a = Tensor.empty(10)
    child = a+2
@@ -929,7 +874,8 @@ class TestSchedule(unittest.TestCase):
    t = Tensor.zeros((3, 3)).contiguous().realize()
    v = t[1]  # view - is_realized but not has_buffer_identity
    assert v.uop.is_realized
-    check_schedule(v, 0)
+    linear, _ = Tensor.linear_with_vars(v)
+    self.assertEqual(len(linear.src), 0)

  # NOTE: because empty does not have a lowered kernel if realize is called on a childless empty, it never gets allocated.
  def test_childless_empty_never_allocates(self):
@@ -1511,7 +1457,8 @@ class TestSchedule(unittest.TestCase):
    Tensor.manual_seed(0)
    x = Tensor.randn(4, 12, 64, 64, dtype=dtypes.half).realize()
    out = x.softmax(dtype=dtypes.float)
-    linear, _ = check_schedule(out, 3)
+    linear = out.schedule_linear()
+    self.assertEqual(len(linear.src), 3)
    # max reduction stays in input dtype (no numerical loss), upcast happens after subtracting max
    self.assertEqual(linear.src[0].src[1].dtype, dtypes.half)
    self.assertEqual(linear.src[1].src[1].dtype, dtypes.float)
@@ -1926,7 +1873,8 @@ class TestFusionOp(unittest.TestCase):
    val = 1.0
    a = Tensor(val)
    for _ in range(24): a = Tensor.stack(a, a)[0]
-    check_schedule(a, 0)
+    linear = a.schedule_linear()
+    self.assertLessEqual(len(linear.src), 1)
    self.assertLess(time.perf_counter()-st, 2.0)

  def test_recursive_reshape(self):
@@ -1935,7 +1883,8 @@ class TestFusionOp(unittest.TestCase):
    b = Tensor.empty(16, 2).realize()
    r = a.sum(1)
    for _ in range(24): r = r.reshape(16, 2) + b
-    check_schedule(r, 1)
+    linear = r.schedule_linear()
+    self.assertEqual(len(linear.src), 1)
    self.assertLess(time.perf_counter()-st, 2.0)

 # NOTE: the NULL backend supports SLICE
@@ -2,8 +2,7 @@ import unittest, itertools

 from tinygrad.codegen.late.coalesce import indexing_simplify
 from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import UOp, Ops, graph_rewrite
-from tinygrad.uop.weak import pm_lower_index_dtype
+from tinygrad.uop.ops import UOp, Ops, graph_rewrite, pm_lower_index_dtype
 from tinygrad.uop.symbolic import simplify_valid, sym, pm_move_where_on_load
 from tinygrad.helpers import Context
 from test.helpers import full_rewrite
@@ -25,7 +24,7 @@ def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UO
  ))

 def Special(expr, nmax): return UOp(Ops.SPECIAL, src=(UOp.const(nmax),), arg=expr)
-def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax, param=True)
+def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
 def Range(n, nmax): return UOp.range(nmax, n)

 class TestValidIdxSimplification(unittest.TestCase):
@@ -333,7 +332,7 @@ class TestImageSimplification(unittest.TestCase):
    load = get_load_image_uop(shape, valid, idx)

    self.check(load,
-               "(((idx2*2)+r0)<11)",
+               "((((idx2*2)+r0)<11)&((((idx1*8)+r1)<3)!=True))",
               "(idx0+(idx1*512+r1*64)+-192)",
               "((((idx2*2)+r0)+(((idx1+((r1+5)//8))+1)//2))+-4)")

@@ -461,7 +460,7 @@ class TestImageSimplification(unittest.TestCase):
      self.check(load, None, "(gidx0+lidx0*1024+r0*1024+lidx1*128+-3168)", "0")
    except AssertionError:
      # TODO: fold valid
-      self.check(load, "(((lidx1<1)!=True)&((lidx0+r0)<19))",
+      self.check(load, "(((lidx1<1)!=True)&(((lidx0+r0)<3)!=True)&((lidx0+r0)<19))",
                       "(gidx0+lidx1*128+(lidx0*1024+r0*1024)+-3168)", "0")

  def test_simplify10(self):
@@ -480,7 +479,7 @@ class TestImageSimplification(unittest.TestCase):
      self.check(load, None, "(lidx2+gidx0*4+lidx0*1024+r0*1024+lidx1*256+-3264)", "0")
    except AssertionError:
      # TODO: fold valid
-      self.check(load, "(((lidx1<1)!=True)&((lidx0+r0)<11))",
+      self.check(load, "(((lidx1<1)!=True)&(((lidx0+r0)<3)!=True)&((lidx0+r0)<11))",
                       "(lidx2+gidx0*4+lidx1*256+(lidx0*1024+r0*1024)+-3264)", "0")

  def test_drop_non_monotonic_window(self):
@@ -1,11 +1,10 @@
 import unittest, pytest
-from tinygrad import dtypes, Variable, Device
+from tinygrad import dtypes, Variable
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import DEBUG, Context
-from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, graph_rewrite, GroupOp, AxisType, broadcast_axes, KernelInfo
+from tinygrad.uop.ops import Ops, UOp, UPat, PatternMatcher, graph_rewrite, GroupOp, AxisType, broadcast_axes
 from tinygrad.uop.symbolic import sym
 from test.helpers import to_uops_list
-from tinygrad.codegen import full_rewrite_to_sink

 simple_pm = PatternMatcher([
  (UPat.cvar('x', dtypes.weakint), lambda x: UOp.const(1.0) + UOp.const(2.0)),
@@ -157,7 +156,7 @@ class TestGraphRewrite(unittest.TestCase):
    self.assertEqual(nout.val, 3.0)

  def test_depth_2_fold(self):
-    v = UOp.variable("v", 0, 1, dtypes.float, param=True)
+    v = UOp.variable("v", 0, 1, dtypes.float)
    c1 = UOp.const(1.0)
    c2 = UOp.const(2.0)
    nout = graph_rewrite(v+c1+c2, simple_pm)
@@ -339,7 +338,7 @@ class TestUOpGraph(unittest.TestCase):
    self.assertEqual(len([x for x in uops if x.op is Ops.CAST]), 1)

  def test_depth_2_const_fold(self):
-    v = UOp.variable("tmp", 0, 1, dtypes.int, param=True)
+    v = UOp.variable("tmp", 0, 1, dtypes.int)
    c2 = UOp.const(2, dtypes.int)
    c4 = UOp.const(4, dtypes.int)
    vc = v+c2
@@ -537,15 +536,6 @@ class TestReduceCollapse(unittest.TestCase):
    # Should become add of two separate reduces
    self.assertEqual(result.op, Ops.ADD)

-  def test_reduce_shapeless_const_unroll(self):
-    """a REDUCE over a shapeless CONST (e.g. x*0 folded late in codegen) must collapse before the expander"""
-    out = UOp.param(0, dtypes.float, (1,))
-    red = UOp.const(3.0).cast(dtypes.float).reduce(UOp.range(4, 0, AxisType.UNROLL), arg=(Ops.ADD, 0))
-    ast = UOp.sink(out.index(UOp.const(0)).store(red)).replace(arg=KernelInfo())
-    uops = full_rewrite_to_sink(ast, Device["CPU"].renderer, optimize=False).toposort()
-    self.assertNotIn(Ops.REDUCE, [u.op for u in uops])
-    self.assertIn(12.0, [u.val for u in uops if u.op is Ops.CONST])
-
 class TestMovementOps(unittest.TestCase):
  def test_pm_mops_partial_reshape_index_removes_reshape(self):
    from tinygrad.schedule.rangeify import pm_mops
@@ -603,7 +593,7 @@ class TestUOpTags(unittest.TestCase):
  def test_inc_by_one(self):
    g = UOp.const(1) + UOp.const(1)
    assert g.ssimplify() == 2
-    pm_plus_1 = PatternMatcher([(UPat(Ops.CONST, name="x"), lambda x: UOp.const(x.val+1, x.dtype).rtag(1) if x.tag is None else None)])
+    pm_plus_1 = PatternMatcher([(UPat(Ops.CONST, name="x"), lambda x: x.replace(arg=x.val+1, tag=1) if x.tag is None else None)])
    pm_strip_tags = PatternMatcher([(UPat(GroupOp.All, name="x"), lambda x: x.replace(tag=None) if x.tag is not None else None)])
    g = graph_rewrite(g, pm_plus_1)
    assert g.ssimplify() == 4
@@ -6,7 +6,7 @@ from tinygrad.dtype import dtypes, ConstType, DType, Invalid
 from test.helpers import get_uops
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, sym_infer
 from tinygrad.uop.spec import spec_shared, type_verify
-from tinygrad.uop.symbolic import sym, pm_fold_cast_const, commutative, pm_simplify_valid, pm_move_where_on_load
+from tinygrad.uop.symbolic import sym, commutative, pm_simplify_valid, pm_move_where_on_load
 from tinygrad.uop.validate import uops_to_z3

 def check_uop_against_string(self, v:UOp, s:str):
@@ -16,8 +16,7 @@ def check_uop_against_string(self, v:UOp, s:str):
  s_eval = graph_rewrite(s_eval, commutative, name="cannonicalize eval")
  self.assertIs(s_eval, v, f"eval did not match simplified: {s_eval} != {v.render()} for {s}")

-def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint):
-  return UOp.variable(name, min_val, max_val, dtype, param=True)
+def Variable(name: str, min_val: ConstType, max_val: ConstType, dtype: DType=dtypes.weakint): return UOp.variable(name,min_val,max_val,dtype)
 def uconst(val): return UOp.const(val)
 def usum(ops): return functools.reduce(lambda x,y: x+y, ops)
 def uand(ops): return functools.reduce(lambda x,y: x*y, ops)
@@ -36,7 +35,7 @@ class TestSymbolic(unittest.TestCase):
    self.assertEqual(solver.check(expr1 != expr2), z3.unsat, "simplified expression not equal to original")

  def helper_test_variable(self, v, n, m, s, test_z3:bool=True):
-    v_simplified = graph_rewrite(v, sym+pm_fold_cast_const, name="simplify symbolic uop")
+    v_simplified = graph_rewrite(v, sym, name="simplify symbolic uop")
    if test_z3: self.check_equal_z3(v, v_simplified)
    nmin, nmax = v_simplified.vmin, v_simplified.vmax
    check_uop_against_string(self, v_simplified, s)
@@ -443,7 +442,7 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable(uand([uconst(1), Variable("a", 0, 1)]), 0, 1, "a")

  def test_masked_shr_fold(self):
-    x = UOp.variable('x', 0, 255, dtype=dtypes.uint32, param=True)
+    x = UOp.variable('x', 0, 255, dtype=dtypes.uint32)
    self.helper_test_variable((x & -4) >> 2, 0, 63, "(x>>2)")

  def test_bool_or_not_tautology(self):
@@ -484,12 +483,12 @@ class TestSymbolic(unittest.TestCase):

  def test_div_drop_small_terms(self):
    # from openpilot, shouldnt simplify
-    gidx0 = UOp.variable("gidx0", 0, 10, param=True)
-    gidx1 = UOp.variable("gidx1", 0, 10, param=True)
-    lidx0 = UOp.variable("lidx0", 0, 1, param=True)
-    lidx1 = UOp.variable("lidx1", 0, 1, param=True)
-    ridx1005 = UOp.variable("ridx1005", 0, 2, param=True)
-    ridx1006 = UOp.variable("ridx1006", 0, 2, param=True)
+    gidx0 = UOp.variable("gidx0", 0, 10)
+    gidx1 = UOp.variable("gidx1", 0, 10)
+    lidx0 = UOp.variable("lidx0", 0, 1)
+    lidx1 = UOp.variable("lidx1", 0, 1)
+    ridx1005 = UOp.variable("ridx1005", 0, 2)
+    ridx1006 = UOp.variable("ridx1006", 0, 2)
    self.helper_test_variable((lidx1+((gidx1*18)+(ridx1005*18)+(lidx0*162))+(gidx0*2)+(ridx1006*2)+-40)//18, -3, 20,
      "(gidx1+ridx1005+lidx0*9+(gidx0+ridx1006+7)//9+-3)")

@@ -949,11 +948,6 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable(cond.where(u0, u1), 0, 1, "((a<2)!=True)")
    self.helper_test_variable(cond.where(u0, u1).where(u0, u1), 0, 1, "(a<2)")

-  def test_equivalent_const_max(self):
-    x = Variable("x", -10, 10)
-    self.helper_test_variable((x < 0).where(0, x), 0, 10, "x.maximum(0)")
-    self.helper_test_variable((0 < x).where(x, 0), 0, 10, "x.maximum(0)")
-
  def test_where_combine(self):
    cond = Variable("x", 0, 3) < 2
    a = Variable("a", 0, 3)
@@ -998,7 +992,7 @@ class TestSymbolic(unittest.TestCase):
    self.helper_test_variable(cond.ne(False), 0, 1, "(x<2)")

  def test_bitcast_chain(self):
-    a = UOp.variable("a", 0, 3, dtype=dtypes.int32, param=True)
+    a = UOp.variable("a", 0, 3, dtype=dtypes.int32)
    self.assertIs(graph_rewrite(a.bitcast(dtypes.float32).bitcast(a.dtype), sym), a)

  def test_negation_in_where(self):
@@ -1181,7 +1175,7 @@ class TestSymbolicVariables(unittest.TestCase):
    assert (a//4 + a//6).variables() == [a]

  def test_variable_min_eq_max_bind_folds(self):
-    b = UOp.variable("x", 1, 1).bind(1)
+    b = Variable("x", 1, 1).bind(1)
    s = b.simplify()
    self.assertEqual(s.op, Ops.CONST)
    self.assertEqual(s.val, 1)
@@ -1,6 +1,6 @@
 import unittest, math
 from tinygrad.uop.ops import UOp, Ops
-from tinygrad.dtype import dtypes, Invalid, truncate
+from tinygrad.dtype import dtypes, Invalid

 class TestVminVmaxProperties(unittest.TestCase):
  def test_vmin_vmax_constant(self):
@@ -162,16 +162,6 @@ class TestVminVmaxProperties(unittest.TestCase):
    self.assertEqual(x_uint.vmin, dtypes.uint.min)
    self.assertEqual(x_uint.vmax, dtypes.uint.max)

-  def test_vmin_vmax_cast_float_to_int(self):
-    self.assertEqual(UOp.variable('x', -4.5, 4.5, dtypes.float).cast(dtypes.int)._min_max, (-4, 4))
-    self.assertEqual(UOp.const(4.5).cast(dtypes.float).cast(dtypes.int)._min_max, (4, 4))
-    x = UOp.const(4.5).cast(dtypes.float)
-    self.assertIs(x.ne(x.cast(dtypes.int).cast(dtypes.float)).simplify().arg, True)
-
-  def test_vmin_vmax_cast_int_to_float_grid(self):
-    # a cast to float only takes values on the float grid, so its bounds are the source bounds rounded at the destination
-    self.assertEqual(UOp.variable('x', 0, 16777219, dtypes.int).cast(dtypes.float)._min_max, (0.0, 16777220.0))
-
  def test_vmin_vmax_invalid(self):
    i = UOp.invalid()
    self.assertNotEqual(i.vmin, i.vmax)
@@ -321,8 +311,8 @@ class TestVminVmaxVConst(unittest.TestCase):
  def test_vmin_vmax_vconst_with_floats(self):
    # vmin and vmax for a vector constant of float values
    uop = UOp.const((1.5, -3.2, 0.0))
-    self.assertEqual(uop.vmin, truncate[dtypes.default_float](-3.2))
-    self.assertEqual(uop.vmax, truncate[dtypes.default_float](1.5))
+    self.assertEqual(uop.vmin, -3.2)
+    self.assertEqual(uop.vmax, 1.5)

  def test_vmin_vmax_vconst_with_bools(self):
    # vmin and vmax for a vector constant of bool values
@@ -5,8 +5,7 @@ from tinygrad.tensor import Tensor
 from tinygrad.helpers import Timing, Context, cdiv
 from tinygrad.dtype import dtypes, AddrSpace, ConstFloat, Invalid  # noqa: F401
 from tinygrad.device import Device
-from tinygrad.uop.ops import Ops, ParamArg, PatternMatcher, UOp, UPat, dtype_from_uop, exec_alu, graph_rewrite  # noqa: F401  # ParamArg used by eval(str(uop)) roundtrip tests
-from tinygrad.uop.weak import pm_lower_index_dtype
+from tinygrad.uop.ops import Ops, ParamArg, PatternMatcher, UOp, UPat, dtype_from_uop, exec_alu, graph_rewrite, pm_lower_index_dtype  # noqa: F401  # ParamArg used by eval(str(uop)) roundtrip tests
 from tinygrad.uop.spec import spec_program, spec_shared, type_verify
 from tinygrad.uop.symbolic import sym, pm_remove_invalid
 from test.helpers import eval_uop, to_uops_list
@@ -56,11 +55,11 @@ class TestDTypeFromUOp(unittest.TestCase):
    invalid = UOp.invalid()
    self.assertIs(invalid.dtype, dtypes.bool)
    self.assertIs(UOp.const(Invalid, dtypes.float32), invalid)
+    self.assertIs((moved:=invalid.reshape((1,))).cast(dtypes.float32), moved)
    scratch = Tensor.invalids(4, dtype=dtypes.float32)
    self.assertEqual((scratch.dtype, next(u.dtype for u in scratch.uop.toposort() if u.op is Ops.BUFFER), next(u.dtype for u in scratch.uop.toposort()
      if u.is_invalid)), (dtypes.float32, dtypes.float32, dtypes.bool))
    invalid, value = UOp.invalid(), UOp.const(1, dtypes.float32)
-    for u in (UOp.param(0, dtypes.bool, ()).where(value, invalid), value+invalid, UOp.stack(value, invalid)): self.assertIs(u.src[-1], invalid)
    for u in (UOp(Ops.STACK, dtypes.float32, src=(value, invalid)), UOp(Ops.ADD, dtypes.float32, src=(value, invalid)),
              UOp.const(True).where(value, invalid), UOp(Ops.CMPLT, src=(invalid, value)), UOp(Ops.CMPLT, src=(value, invalid)),
              UOp.param(0, dtypes.float32, (4,)).index(invalid)): type_verify(u, spec_shared)
@@ -127,12 +126,6 @@ class TestConstFloatEq(unittest.TestCase):
    self.assertFalse(nan == Invalid)
    self.assertTrue(nan != Invalid)    # __ne__ must defer to the reflected eq, not swallow NotImplemented

-  def test_invalid_eq_defers_to_reflected(self):
-    class HoldsInvalid:  # a carrier that knows it holds Invalid. returning False for foreign types would silence its eq
-      def __eq__(self, other): return other is Invalid
-    self.assertTrue(Invalid == HoldsInvalid())
-    self.assertFalse(Invalid != HoldsInvalid())
-
  def test_matchers_agree_on_nan(self):
    n = UOp.const(math.nan, dtypes.float32)
    for compiled in (False, True):
@@ -454,7 +447,7 @@ class TestUPatHelpers(unittest.TestCase):

 class TestUopsObject(unittest.TestCase):
  def test_timing(self):
-    with Timing("create 10k uops:"): ret = [UOp.const(10000000+i, dtypes.int) for i in range(10000)]
+    with Timing("create 10k uops:"): ret = [UOp(Ops.CONST, dtypes.int, arg=10000000+i) for i in range(10000)]
    assert len(ret) == 10000

  def test_nested(self):
@@ -147,21 +147,21 @@ class TestUOpsStats(unittest.TestCase):
  #MULACC should have the same stats as MUL + ADD
  def test_mulacc(self):
    globl = UOp.param(0, dtypes.int, (3,))
-    o1 = UOp.const(1, dtypes.int)
-    o2 = UOp.const(2, dtypes.int)
+    o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
+    o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
    u1 = globl.index(o1)
    u2 = globl.index(o2)
-    u3 = UOp.const(3, dtypes.int)
+    u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
    u4 = UOp(Ops.MUL, src=(u1,u2))
    u5 = UOp(Ops.ADD, src=(u4,u3))
    uops = tuple(u5.toposort())

    globl = UOp.param(0, dtypes.int, (3,))
-    o1 = UOp.const(1, dtypes.int)
-    o2 = UOp.const(2, dtypes.int)
+    o1 = UOp(Ops.CONST, dtypes.int, tuple(), 1)
+    o2 = UOp(Ops.CONST, dtypes.int, tuple(), 2)
    u1 = globl.index(o1)
    u2 = globl.index(o2)
-    u3 = UOp.const(3, dtypes.int)
+    u3 = UOp(Ops.CONST, dtypes.int, tuple(), 3)
    u4 = UOp(Ops.MULACC, src=(u1,u2,u3))
    uops_fma = tuple(u4.toposort())

@@ -1,11 +1,11 @@
 import unittest
 from tinygrad.helpers import DEBUG, Context
 from tinygrad.dtype import dtypes
-from tinygrad.uop.ops import UPat, rewrite_group, GroupOp, Ops
+from tinygrad.uop.ops import UPat, track_rewrites, GroupOp, Ops
 from tinygrad.uop.upat import _get_code, upat_compile
 import dis

-@rewrite_group()
+@track_rewrites()
 def do_compile(up):
  print("\n***** COMPILE", up)
  match_code = _get_code(up, False)
@@ -1,12 +1,10 @@
 import unittest
-from tinygrad import dtypes
+from tinygrad import dtypes, Variable
 from tinygrad.dtype import AddrSpace
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import Ops, UOp, AxisType
 from test.helpers import to_uops_list

-def Variable(name, nmin, nmax): return UOp.variable(name, nmin, nmax, param=True)
-
 class TestValidateOOB(unittest.TestCase):
  """Test z3 validation of index bounds for different ALU ops and patterns."""

@@ -3,7 +3,7 @@ from pathlib import Path
 from dataclasses import dataclass
 from typing import Generator

-from tinygrad.uop.ops import UOp, UPat, Ops, PatternMatcher, TrackedPatternMatcher, graph_rewrite, rewrite_group
+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, AddrSpace
 from tinygrad.helpers import colored, ansistrip, flatten, TracingKey, ProfileRangeEvent, ProfileEvent, Context, cpu_events, profile_marker
@@ -14,7 +14,7 @@ from tinygrad.uop.ops import tracked_keys, tracked_ctxs, uop_fields, active_rewr
 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

-@rewrite_group(name=True)
+@track_rewrites(name=True)
 def exec_rewrite(sink:UOp, pm_lst:list[PatternMatcher], names:None|list[str]=None) -> UOp:
  for i,pm in enumerate(pm_lst):
    sink = graph_rewrite(sink, TrackedPatternMatcher(pm.patterns), name=names[i] if names else None)
@@ -97,7 +97,7 @@ class TestViz(unittest.TestCase):
    # VIZ tracks rewrites up to and including the error
    def count_3(x:UOp):
      assert x.val <= 3
-      return UOp.const(x.val+1, x.dtype)
+      return x.replace(arg=x.val+1)
    err_pm = PatternMatcher([(UPat.cvar("x"), count_3),])
    a = UOp.const(1)
    with save_viz() as viz:
@@ -109,7 +109,7 @@ class TestViz(unittest.TestCase):
  def test_default_name(self):
    with save_viz() as viz:
      a = UOp.variable("a", 1, 10)
-      @rewrite_group()
+      @track_rewrites()
      def name_default(): return graph_rewrite(a, PatternMatcher([]))
      name_default()
    lst = viz.list_items()
@@ -118,7 +118,7 @@ class TestViz(unittest.TestCase):
  # name can also come from a function that returns a string
  def test_dyn_name_fxn(self):
    with save_viz() as viz:
-      @rewrite_group(name=lambda *args,ret,**kwargs: ret.render())
+      @track_rewrites(name=lambda *args,ret,**kwargs: ret.render())
      def name_from_fxn(s:UOp, arg:list|None=None): return graph_rewrite(s, PatternMatcher([]))
      name_from_fxn(UOp.variable("a", 1, 10)+1, arg=["test"])
    lst = viz.list_items()
@@ -128,18 +128,18 @@ class TestViz(unittest.TestCase):
  # name can also come from a function that returns a TracingKey
  def test_tracing_key(self):
    with save_viz() as viz:
-      @rewrite_group(name=lambda inp,ret: TracingKey("custom_name", (inp,)))
+      @track_rewrites(name=lambda inp,ret: TracingKey("custom_name", (inp,)))
      def test(s:UOp): return graph_rewrite(s, PatternMatcher([]))
      test(UOp.variable("a", 1, 10)+1)
    lst = viz.list_items()
    # NOTE: names from TracingKey do not get deduped
    self.assertEqual(lst[0]["name"], "custom_name")

-  def test_nested_rewrite_group(self):
+  def test_nested_track_rewrites(self):
    with save_viz() as viz:
-      @rewrite_group(name=lambda x,ret: TracingKey(f"inner fxn for {x.render()}", (ret,)))
+      @track_rewrites(name=lambda x,ret: TracingKey(f"inner fxn for {x.render()}", (ret,)))
      def inner(x:UOp): return graph_rewrite(x, PatternMatcher([]), name="each")
-      @rewrite_group(name=lambda *args,ret: f"outer rewrite of {len(args)} inputs")
+      @track_rewrites(name=lambda *args,ret: f"outer rewrite of {len(args)} inputs")
      def outer(*xs:tuple[UOp, ...]): return graph_rewrite(UOp.sink(*[inner(x) for x in xs]), PatternMatcher([]), name="all")
      items = ["a", "b", "c"]
      outer(*[UOp.variable(x, 1, 10) for x in items])
@@ -156,13 +156,13 @@ class TestViz(unittest.TestCase):
      self.assertEqual(len(steps), 1)
      self.assertEqual(steps[0]["name"], "each")

-  def test_rewrite_group_nested(self):
+  def test_profile_matches(self):
    with save_viz() as viz:
-      @rewrite_group(new_ctx=False)
+      @profile_matches
      def nested_function(u:UOp):
        for i in range(2): graph_rewrite(u, PatternMatcher([]), name=f"step {i+1}")

-      @rewrite_group()
+      @track_rewrites()
      def main_rewrite(u:UOp):
        graph_rewrite(u, PatternMatcher([]), name="init")
        nested_function(u)
@@ -173,9 +173,9 @@ class TestViz(unittest.TestCase):
    self.assertEqual(steps[1]["name"], "nested_function")
    self.assertEqual(len(steps), 4)

-  def test_rewrite_group_invalid_arg(self):
+  def test_profile_matches_invalid_arg(self):
    with save_viz():
-      @rewrite_group(new_ctx=False)
+      @profile_matches
      def invalid_fxn(arg:str): return graph_rewrite(UOp(Ops.SINK), PatternMatcher([]))
      with self.assertRaisesRegex(AssertionError, "invalid match tracing input"):
        invalid_fxn("test")
@@ -202,8 +202,8 @@ class TestViz(unittest.TestCase):
    a = UOp.const(3)
    b = UOp.const(4)
    pm = PatternMatcher([
-      (UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype)),
-      (UPat(Ops.CONST, arg=4, name="x"), lambda x: UOp.const(3, x.dtype)),
+      (UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4)),
+      (UPat(Ops.CONST, arg=4, name="x"), lambda x: x.replace(arg=3)),
    ])
    with save_viz() as viz:
      # use smaller stack limit for faster test (default is 250000)
@@ -224,7 +224,7 @@ class TestViz(unittest.TestCase):
    list(viz.get_details(0, 0))

  def test_enter_calls_rewrite(self):
-    pm = PatternMatcher([(UPat(Ops.CONST, arg=3, name="x"), lambda x: UOp.const(4, x.dtype))])
+    pm = PatternMatcher([(UPat(Ops.CONST, arg=3, name="x"), lambda x: x.replace(arg=4))])
    with save_viz() as viz:
      inner = UOp.const(3)
      call = UOp(Ops.CALL, src=(UOp(Ops.SINK, src=(inner,)),))
@@ -305,10 +305,10 @@ class TestVizTree(unittest.TestCase):

  def test_tree_view(self):
    with save_viz() as viz:
-      a = UOp.variable("a",0,10,param=True)
-      b = UOp.variable("b",0,10,param=True)
-      c = UOp.variable("c",0,10,param=True)
-      d = UOp.variable("d",0,10,param=True)
+      a = UOp.variable("a",0,10)
+      b = UOp.variable("b",0,10)
+      c = UOp.variable("c",0,10)
+      d = UOp.variable("d",0,10)
      sink = UOp.sink(a+b, c+d)
      def tree_rewrite(): return graph_rewrite(sink, root, name="root")
      tree_rewrite()
@@ -395,7 +395,7 @@ class TestVizIntegration(unittest.TestCase):
    graph = next(viz.get_details(0, 0))["graph"]
    self.assertEqual(len([n for n in graph.values() if repr(metadata) in n["label"]]), 1)

-  # tracing also works without a rewrite_group context
+  # tracing also works without a track_rewrites context
  # all graph_rewrites get put into the default group
  def test_default_tracing(self):
    with save_viz() as viz:
@@ -407,11 +407,11 @@ class TestVizIntegration(unittest.TestCase):
    self.assertEqual(len(ls), 1)
    self.assertEqual(ls[0]["name"], "default graph_rewrite")

-  # using @rewrite_group organizes function calls into groups
+  # using @track_rewrites organizes function calls into groups
  # and nicely counts function calls.
  def test_group_traces(self):
    with save_viz() as viz:
-      @rewrite_group()
+      @track_rewrites()
      def test(root):
        return graph_rewrite(root, sym)
      test(c:=UOp.const(1))
@@ -420,11 +420,11 @@ class TestVizIntegration(unittest.TestCase):
    self.assertEqual(len(ls), 2)
    for i in range(2): self.assertEqual(ls[i]["name"], f"test n{i+1}")

-  # @rewrite_group always starts a new group.
+  # @track_rewrites always starts a new group.
  def test_group_combined(self):
    with save_viz() as viz:
      def default_test(root): return graph_rewrite(root, sym)
-      tracked_test = rewrite_group()(default_test)
+      tracked_test = track_rewrites()(default_test)
      c = UOp.const(1)
      default_test(c+1) # goes to the default group
      tracked_test(c)   # all rewrites after this go inside the second group.
@@ -1,7 +1,6 @@
 import unittest, sys
 from tinygrad import Tensor, GlobalCounters, dtypes, Context
 from tinygrad.helpers import WINO
-from test.helpers import check_schedule

@unittest.skipIf(sys.platform.startswith("win"), "flaky on Windows")
 class TestWinograd(unittest.TestCase):
@@ -14,7 +13,7 @@ class TestWinograd(unittest.TestCase):
  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)
-    check_schedule(out, 4)
+    self.assertEqual(len(out.schedule_linear().src), 4)

  def test_backward_counters(self):
    # contiguous_backward on the pooled input keeps the input-transform adjoint out of the overlap accumulation, so
@@ -239,15 +239,6 @@ class TestKernelOpts(unittest.TestCase):
    helper_linearizer_opt(a.sum().exp(), [[Opt(OptOps.PADTO, 0, 32)],])
    helper_linearizer_opt(a.sum(0).exp(), [[Opt(OptOps.PADTO, 1, 32)],])

-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
-  @unittest.expectedFailure
-  def test_padto_group_full_unroll_sum(self):
-    a = Tensor.ones(2, 28, 4096, dtype=dtypes.bfloat16).realize()
-    out = ((a * 0.5).float().square()).sum(axis=(0, 2))
-    opts_to_apply = [Opt(OptOps.GROUPTOP, 1, 256), Opt(OptOps.PADTO, 3, 32), Opt(OptOps.UNROLL, 2, 0), Opt(OptOps.UPCAST, 0, 7)]
-    helper_linearizer_opt(out, [opts_to_apply], check_default_opt=False)
-
  def test_padto_sum(self):
    N = 18
    # NOTE: this setup prevents 17 * 17 contiguous merged into one dimension
@@ -79,8 +79,7 @@ class TestTensorCores(unittest.TestCase):
  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
  def test_tensor_cores(self):
    for tc in Device[Device.DEFAULT].renderer.tensor_cores:
-      with self.subTest(tc=tc):
-        helper_tc_allclose(tc.dims[0], tc.dims[1], tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)
+      helper_tc_allclose(tc.dims[0], tc.dims[1], tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)

  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
  def test_tensor_cores_nested_reduce(self):
@@ -102,8 +101,7 @@ class TestTensorCores(unittest.TestCase):
      if Device.DEFAULT == "CPU" and DEV.renderer == "LLVM":
        assert "0x201000" in prg.src[2].arg
      elif Device.DEFAULT == "AMD" and DEV.renderer == "LLVM":
-        # RDNA emits wmma intrinsics, CDNA emits mfma intrinsics
-        assert ("@llvm.amdgcn.wmma" in prg.src[2].arg) or ("@llvm.amdgcn.mfma" in prg.src[2].arg)
+        assert "@llvm.amdgcn.wmma" in prg.src[2].arg
      elif Device[Device.DEFAULT].renderer.suffix == "PTX":
        assert "mma.sync.aligned" in prg.src[2].arg
      else:
@@ -183,13 +181,11 @@ class TestTensorCores(unittest.TestCase):
  @unittest.skipIf(Device.DEFAULT == "PYTHON", "slow on EMULATED device")
  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
  def test_tensor_cores_unroll_phi(self):
-    # skip fp8 tcs: the unoptimized ALU baseline quantizes products to fp8 (JAX promotion), which legitimately
-    # differs from the MFMA path (f32 accumulation), so the baseline-vs-TC numerical gate can't hold for fp8.
-    tc = next(tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in not in dtypes.fp8s)
-    x, y = Tensor.rand(16, 64, dtype=tc.dtype_in), Tensor.rand(64, 16, dtype=tc.dtype_in)
+    tc = Device[Device.DEFAULT].renderer.tensor_cores[0]
+    x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
    r = x.matmul(y, dtype=tc.dtype_out)
-    opts = [Opt(OptOps.UNROLL, 0, 2)]
-    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3, check_default_opt=False)
+    opts = [Opt(OptOps.UNROLL, 0, 4)]
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
    for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[1].src):
      if u.op is Ops.WMMA:
        assert u.src[-1].src[0].op != Ops.STORE
@@ -199,11 +195,11 @@ class TestTensorCores(unittest.TestCase):
  @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "CPU does not support using a different type for accumulation")
  def test_tensor_cores_unroll_casted_phi(self):
-    tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out and tc.dtype_in not in dtypes.fp8s][0]
-    x, y = Tensor.rand(16, 64, dtype=tc.dtype_in), Tensor.rand(64, 16, dtype=tc.dtype_in)
+    tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
+    x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
    r = x.matmul(y, dtype=tc.dtype_out)
-    opts = [Opt(OptOps.UNROLL, 0, 2)]
-    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3, check_default_opt=False)
+    opts = [Opt(OptOps.UNROLL, 0, 4)]
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
    for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[1].src):
      if u.op is Ops.WMMA:
        #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
@@ -215,11 +211,11 @@ class TestTensorCores(unittest.TestCase):
  @unittest.skipIf(Device.DEFAULT in {"CPU"}, "CPU does not support using a different type for accumulation")
  def test_tensor_cores_unroll_casted_phi_with_children(self):
    # all STORE children are outside the loop
-    tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out and tc.dtype_in not in dtypes.fp8s][0]
-    x, y = Tensor.rand(16, 64, dtype=tc.dtype_in), Tensor.rand(64, 16, dtype=tc.dtype_in)
+    tc = [tc for tc in Device[Device.DEFAULT].renderer.tensor_cores if tc.dtype_in != tc.dtype_out][0]
+    x, y = Tensor.rand(128, 128, dtype=tc.dtype_in), Tensor.rand(128, 128, dtype=tc.dtype_in)
    r = x.matmul(y, dtype=tc.dtype_out).relu()
-    opts = [Opt(OptOps.UNROLL, 0, 2)]
-    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3, check_default_opt=False)
+    opts = [Opt(OptOps.UNROLL, 0, 4)]
+    ast = helper_linearizer_opt(r, [opts], apply_tc=True, atol=3e-2, rtol=1e-3)
    for u in tuple(to_program(replace_opts(ast, opts), Device[Device.DEFAULT].renderer).src[1].src):
      if u.op is Ops.WMMA:
        #assert u.src[-1].dtype == dtypes.float.vec(prod(tc.thread_local_sizes[2]))
@@ -39,10 +39,6 @@ class TestTiny(unittest.TestCase):
    out = Tensor.ones(N).contiguous().sum()
    self.assertEqual(out.item(), N)

-  def test_eye(self):
-    out = Tensor.eye(3).flatten()
-    self.assertListEqual(out.tolist(), [1.0,0.0,0.0, 0.0,1.0,0.0, 0.0,0.0,1.0])
-
  def test_gemm(self, N=getenv("GEMM_N", 64), dtype=dtypes.float):
    a = Tensor.ones(N,N, dtype=dtype).contiguous()
    b = Tensor.eye(N, dtype=dtype).clone()
@@ -1,9 +1,7 @@
 import unittest

-from tinygrad import Tensor, Device, Variable, dtypes
-from tinygrad.helpers import DEV, fetch, round_up
-from tinygrad.engine.realize import compile_linear
-from tinygrad.uop.ops import Ops
+from tinygrad import Tensor, Device, dtypes
+from tinygrad.helpers import fetch, round_up
 from extra.hevc.hevc import parse_hevc_file_headers, nv_gpu
 from extra.hevc.decode import hevc_decode

@@ -65,7 +63,7 @@ class TestHevc(unittest.TestCase):
    self.assertEqual(list(frame3.initreflistidxl1), [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
    self.assertEqual(list(frame3.RefDiffPicOrderCnts), [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])

-  @unittest.skipUnless(Device.DEFAULT == "NV" and not DEV.interface.startswith("MOCK"), "real NV only")
+  @unittest.skipUnless(Device.DEFAULT == "NV", "NV only")
  def test_hevc_decode(self):
    url = "https://github.com/haraschax/filedump/raw/09a497959f7fa6fd8dba501a25f2cdb3a41ecb12/comma_video.hevc"
    dat = fetch(url, headers={"Range": f"bytes=0-{512<<10}"}).read_bytes()
@@ -85,22 +83,5 @@ class TestHevc(unittest.TestCase):
      self.assertEqual(f.dtype, dtypes.uint8)
      self.assertEqual(f.device, "NV")

-  @unittest.skipUnless(Device.DEFAULT == "NV", "NV only")
-  def test_hevc_decode_compile(self):
-    url = "https://github.com/haraschax/filedump/raw/09a497959f7fa6fd8dba501a25f2cdb3a41ecb12/comma_video.hevc"
-    dat = fetch(url, headers={"Range": f"bytes=0-{512<<10}"}).read_bytes()
-
-    opaque, frame_info, _, _, luma_w, luma_h, _ = parse_hevc_file_headers(dat)
-    offset, sz, frame_pos, max_hist, _ = frame_info[1]
-    out_image_size = luma_h + (luma_h + 1) // 2, round_up(luma_w, 64)
-    history = [Tensor.empty(*out_image_size, dtype=dtypes.uint8, device="NV") for _ in range(max_hist)]
-    decoded = Tensor(dat, device="NV")[offset:offset+sz].decode_hevc_frame(
-      Variable("pos", 0, max_hist + 1).bind(frame_pos), out_image_size, opaque[1], history)
-
-    compiled = compile_linear(decoded.linear_with_vars()[0])
-    self.assertTrue(any(call.src[0].op is Ops.PROGRAM for call in compiled.src))
-    encdec_calls = [call for call in compiled.src if call.src[0].op is Ops.CUSTOM_FUNCTION and call.src[0].arg == "encdec"]
-    self.assertEqual(len(encdec_calls), 1)
-
 if __name__ == "__main__":
  unittest.main()
@@ -1,5 +1,5 @@
 import unittest
-from tinygrad import Tensor, UOp, dtypes
+from tinygrad import Tensor, dtypes
 from tinygrad.helpers import Context
 from tinygrad.uop.ops import Ops

@@ -43,13 +43,6 @@ class TestRingAllReduce(unittest.TestCase):
    self.assertEqual(len(sinks), 2)
    self.assertTrue(all(dst != src for dst, src in pairs))

-  def test_symbolic_shape(self):
-    rows = UOp.variable("rows", 1, 4).bind(3)
-    t = Tensor.ones(4, 4).shard(("CPU:0", "CPU:1"), axis=1).realize()
-    out = t[:rows].sum(1).realize()
-    self.assertEqual(out.shape, (rows,))
-    self.assertTrue((out == 4).all().item())
-
  def test_correct_ring(self):
    with Context(RING=2):
      N = 4
@@ -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 assert_kernel_count

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

@@ -43,7 +42,7 @@ class TestAssign(unittest.TestCase):
    # it should copy into the empty buffer
    GlobalCounters.reset()
    c.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)

  def test_assign_slice(self):
    X = Tensor([1,2,3,4]).realize()
@@ -51,7 +50,7 @@ class TestAssign(unittest.TestCase):
    xs.assign(xs+1)
    GlobalCounters.reset()
    self.assertListEqual(X.tolist(), [1,2,4,5])
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)

  def test_assign_slice_alt(self):
    X = Tensor([1,2,3,4]).realize()
@@ -59,7 +58,7 @@ class TestAssign(unittest.TestCase):
    xs1.assign(xs2+1)
    GlobalCounters.reset()
    self.assertListEqual(X.tolist(), [1,4,5,4])
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 2)

  def test_assign_flip(self):
    ref = np.arange(16, dtype=np.float32)
@@ -69,7 +68,7 @@ class TestAssign(unittest.TestCase):
    xs.assign(xs + X)
    ref = ref + ref[::-1]
    np.testing.assert_allclose(X.numpy(), ref)
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 2)

  def test_assign_add(self):
    for T in (1, 2, 10):#, 100): # this crashes in CI, not sure why
@@ -332,14 +331,14 @@ class TestAssign(unittest.TestCase):
    a = (Tensor.arange(16).reshape(4,4).clone().realize() + 1)
    GlobalCounters.reset()
    b.assign(a.contiguous()).realize()
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 2)

  def test_assign_contiguous_permute(self):
    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()
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 2)

  def test_permuted_assignment(self):
    a = Tensor(np.arange(N*N, dtype=np.float32)).reshape(N,N)
@@ -414,7 +413,7 @@ class TestAssign(unittest.TestCase):

    GlobalCounters.reset()
    Tensor.realize(b, c, d)
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    np.testing.assert_allclose(b.numpy(), a.sum(1).numpy()+1)
    np.testing.assert_allclose(c.numpy(), a.sum(1).numpy()+2)
    np.testing.assert_allclose(d.numpy(), a.sum(1).numpy()+3)
@@ -462,7 +461,7 @@ class TestAssign(unittest.TestCase):
    b.assign(r + b)
    c.assign(r + b_perm.contiguous())
    Tensor.realize(b, c)
-    assert_kernel_count(2)
+    self.assertEqual(GlobalCounters.kernel_count, 2)
    np.testing.assert_equal(b.numpy(), a.numpy().sum(1) + np.arange(32 * 32).reshape(32, 32))
    np.testing.assert_equal(c.numpy(), a.numpy().sum(1) + np.arange(32 * 32).reshape(32, 32).transpose(1, 0))

@@ -472,7 +471,7 @@ class TestAssign(unittest.TestCase):
    a.assign(a + b)
    GlobalCounters.reset()
    a.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    np.testing.assert_equal(a.numpy(), np.ones((4, 4))+np.pad(np.ones((4, 4))[:, 0:2], ((0, 0), (0, 2)), constant_values=2))

  def test_permuted_assignment_masked_view_not_contiguous(self):
@@ -511,7 +510,7 @@ class TestAssign(unittest.TestCase):
    expected[0:10] = expected[50:60].copy()
    GlobalCounters.reset()
    a[0:10].assign(a[50:60]).realize()
-    assert_kernel_count(2)  # currently conservative, forces contiguous
+    self.assertEqual(GlobalCounters.kernel_count, 2)  # currently conservative, forces contiguous
    np.testing.assert_allclose(a.numpy(), expected)

  def test_setitem_half(self):
@@ -540,10 +539,6 @@ class TestAssign(unittest.TestCase):
    c = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
    c[0:2].bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000], dtype=dtypes.uint32)).realize()
    np.testing.assert_allclose(c.numpy(), [4.0, 3.0, 3.0, 4.0])
-    # without .realize()
-    a = Tensor([1.0, 2.0, 3.0, 4.0], dtype=dtypes.float32).realize()
-    a.bitcast(dtypes.uint32).assign(Tensor([0x40800000, 0x40400000, 0x40000000, 0x3f800000], dtype=dtypes.uint32))
-    np.testing.assert_allclose(a.numpy(), [4.0, 3.0, 2.0, 1.0])

  def test_assign_bitcast_different_size(self):
    # assign to a shape-changing bitcast view (only works on DISK currently)
@@ -635,7 +630,7 @@ class TestAssign(unittest.TestCase):
    GlobalCounters.reset()
    x.realize()
    # N assigns (1 kernel each) producing N kernels total
-    assert_kernel_count(N)
+    self.assertEqual(GlobalCounters.kernel_count, N)

  def test_shared_computation_assign_kernel_count(self):
    """When a .contiguous() is shared between an assign value and the next layer's input (like QKV projection in LLM),
@@ -653,7 +648,7 @@ class TestAssign(unittest.TestCase):
    GlobalCounters.reset()
    caches[-1][:1].contiguous().realize()
    # N matmuls + N assigns + 1 final read = 2*N+1 (AFTER embedding allows full graph scheduling with shared contiguous reuse)
-    assert_kernel_count(2*N+1)
+    self.assertEqual(GlobalCounters.kernel_count, 2*N+1)

  def test_double_assign_from_const(self):
    a = Tensor.empty(2)
@@ -661,7 +656,7 @@ class TestAssign(unittest.TestCase):
    a.assign(Tensor.ones(2, buffer=False))
    GlobalCounters.reset()
    a.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(a.tolist(), [1.,1.])

  def test_assign_deviceless_const(self):
@@ -677,7 +672,7 @@ class TestAssign(unittest.TestCase):
    contig.assign(Tensor([1, 4, 3], dtype=dtypes.int64))
    GlobalCounters.reset()
    base.assign(contig).realize()
-    assert_kernel_count(2)  # TODO: first copy is dead, could be 1
+    self.assertEqual(GlobalCounters.kernel_count, 2)  # TODO: first copy is dead, could be 1
    self.assertEqual(base.tolist(), [1,4,3])

  def test_nested_after_contiguous_store_no_init(self):
@@ -687,7 +682,7 @@ class TestAssign(unittest.TestCase):
    contig.assign(Tensor([1, 4, 3], dtype=dtypes.int64))
    GlobalCounters.reset()
    base.assign(contig).realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(base.tolist(), [1,4,3])

 class TestAssignOrdering(unittest.TestCase):
@@ -212,18 +212,6 @@ class TestCallSchedule(unittest.TestCase):
    out = f(a, v.bind(5))
    np.testing.assert_allclose(out.numpy(), [5., 10., 15.])

-  def test_precompile_scoped_bind_arg(self):
-    @function(precompile=True)
-    def f(x:Tensor, scale:UOp) -> Tensor: return x * scale
-    a = Tensor.ones(3)
-    x = f(a, UOp.variable("scale_a", 1, 100).bind(2))
-    y = f(a, UOp.variable("scale_b", 1, 100).bind(3))
-    fx = next(u for u in x.uop.toposort() if u.op is Ops.FUNCTION)
-    fy = next(u for u in y.uop.toposort() if u.op is Ops.FUNCTION)
-    self.assertEqual(fx.src[0].key, fy.src[0].key)
-    np.testing.assert_equal(x.numpy(), [2, 2, 2])
-    np.testing.assert_equal(y.numpy(), [3, 3, 3])
-
  def test_precompile_schedule_cache_hit(self):
    """two instances of the same @function should produce identical function body keys (schedule cache hit)"""
    @function(precompile=True)
@@ -359,15 +347,5 @@ class TestCallMultiSharded(unittest.TestCase):
    np.testing.assert_allclose(a.grad.numpy(), b.numpy(), rtol=1e-5)
    np.testing.assert_allclose(b.grad.numpy(), a.numpy(), rtol=1e-5)

-  def test_symbolic_reshape_shard_axis(self):
-    toks = UOp.variable("toks", 1, 2).bind(2)
-    devs = ("CPU:0", "CPU:1")
-    x = Tensor(np.arange(16, dtype=np.float32).reshape(1, 2, 8)).shard(devs, axis=2).realize()
-    @function
-    def f(x:Tensor) -> Tensor: return x.reshape(1, x.shape[1], 2, 4)
-    out = f(x[:, :toks]).realize()
-    self.assertEqual(out.uop.axis, 2)
-    np.testing.assert_equal(out[:1, :2].to(devs[0]).numpy(), np.arange(16, dtype=np.float32).reshape(1, 2, 2, 4))
-
 if __name__ == '__main__':
  unittest.main()
@@ -222,12 +222,6 @@ class TestAutoCastType(unittest.TestCase):
    t.square().mean().backward()
    np.testing.assert_allclose(t.grad.numpy().flatten(), [60000 * 2 / (N*N)] * N*N)

-  @unittest.skipUnless(dtypes.half in supported_dtypes, "need half")
-  def test_var_half_precision_large_n(self):
-    # the element count (70000) exceeds half max (65504): the denominator must not be materialized in half
-    t = Tensor([[0.0, 1.0]], dtype=dtypes.half).expand(35000, 2).contiguous()
-    np.testing.assert_allclose(t.var().numpy(), 0.25, rtol=1e-3)
-
  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "Precision error")
  @unittest.skipUnless(dtypes.half in supported_dtypes, "need half")
  def test_softmax_dtype(self):
@@ -3,8 +3,7 @@ import tempfile, unittest, math
 from tinygrad import Tensor, dtypes, TinyJit
 from tinygrad.helpers import Context
 from tinygrad.dtype import least_upper_float
-from tinygrad.uop.ops import UOp, Ops, dtype_from_uop, graph_rewrite
-from tinygrad.uop.weak import pm_lower_index_dtype, pm_commit_weak
+from tinygrad.uop.ops import UOp, Ops, dtype_from_uop, graph_rewrite, pm_lower_index_dtype, pm_commit_weak
 from tinygrad.uop.symbolic import symbolic_simple
 from tinygrad.uop.spec import spec_shared, type_verify
 from tinygrad.engine.jit import JitError
@@ -45,13 +44,9 @@ class TestWeakPromotion(unittest.TestCase):
    r = Tensor([2], dtype=dtypes.uint8, device="CPU").copysign(Tensor([1], dtype=dtypes.uint32, device="CPU"))
    self.assertEqual((r.dtype, r.tolist()), (dtypes.uint32, [2]))

-  def test_minimum_reflects_weak_operand(self):
+  def test_minimum_commits_both_operands(self):
    r = Tensor(1).minimum(Tensor([2], dtype=dtypes.uint8, device="CPU"))
    self.assertEqual((r.dtype, r.tolist()), (dtypes.uint8, [1]))
-    for dt in dtypes.uints:
-      r = Tensor([dt.max], dtype=dt, device="CPU").minimum(1)
-      self.assertEqual((r.dtype, r.tolist()), (dt, [1]))
-      self.assertNotIn(Ops.CAST, [u.op for u in r._uop.toposort()])

  def test_broadcasted_keeps_const_weak(self):
    # a python scalar stays a bare weak CONST through _broadcasted, lifted only to the KIND of the lub
@@ -63,31 +58,6 @@ class TestWeakPromotion(unittest.TestCase):
    self.assertEqual((x._uop.base.op, x._uop.base.val, x.dtype, x.shape, y.dtype),
                     (Ops.CONST, 1, dtypes.weakfloat, (1,), dtypes.float32))

-  def test_weak_expression_anchors_at_strong_lub(self):
-    # regression test for the HALF bert nan (#17408, reverted in #17409): lub(int32, weakfloat)==weakfloat makes
-    # `loss_mask.sum() + 1e-5` a weakfloat EXPRESSION. Meeting a strong float in a binop must pin it at the lub
-    denom = (Tensor.zeros(912, dtype=dtypes.int32) != Tensor.zeros(912, dtype=dtypes.float32)).sum() + 1e-5
-    self.assertIs(denom.dtype, dtypes.weakfloat)  # the setup: the denominator expression itself is weak
-    x, y = Tensor([2048.0], dtype=dtypes.float32)._broadcasted(denom)
-    self.assertIs(y.dtype, dtypes.float32)
-    recips = [u for u in (x / y)._uop.toposort() if u.op is Ops.RECIPROCAL]
-    self.assertEqual([(u.dtype, u.src[0].dtype) for u in recips], [(dtypes.float32, dtypes.float32)])
-    with Context(DEFAULT_FLOAT=dtypes.float16):
-      committed = graph_rewrite((UOp.const(1).cast(dtypes.int32) + UOp.const(1.0)).cast(dtypes.float32), pm_lower_index_dtype, ctx={})
-    self.assertEqual([u.dtype for u in committed.toposort() if u.op is Ops.ADD], [dtypes.float32])
-
-  def test_cast_weak_expression_commits_at_cast_floor(self):
-    # the floor never narrows: a cast BELOW the default does not pull the compute width down with it
-    with Context(DEFAULT_FLOAT=dtypes.float32):
-      narrowed = graph_rewrite((UOp.const(1.0) + UOp.const(2.0)).cast(dtypes.float16), pm_lower_index_dtype, ctx={})
-    self.assertEqual((narrowed.dtype, narrowed.src[0].dtype), (dtypes.float16, dtypes.float32))
-
-  def test_cast_weak_expression_value_uses_cast_floor(self):
-    with Context(DEFAULT_FLOAT=dtypes.float16):
-      denom = Tensor.ones(1, dtype=dtypes.int32, device="CPU").sum() * 70000 + 1e-5
-      out = Tensor(1.0, dtype=dtypes.float32, device="CPU") / denom
-      self.assertAlmostEqual(out.item(), 1 / (70000 + 1e-5), places=10)
-
  def test_uop_scalar_const_lifts_kind(self):
    for dtype, value, out_dtype, const_dtype in ((dtypes.weakint, 1, dtypes.weakint, dtypes.weakint),
                                                 (dtypes.int32, 1, dtypes.int32, dtypes.weakint),
@@ -4,7 +4,6 @@ from tinygrad.function import function
 from tinygrad import Tensor, GlobalCounters, Device
 from tinygrad.dtype import Invalid
 from tinygrad.uop.ops import UOp, Ops, KernelInfo, ProgramInfo
-from test.helpers import assert_kernel_count

 class TestFunction(unittest.TestCase):
  def test_simple(self):
@@ -619,7 +618,7 @@ class TestFunctionTuple(unittest.TestCase):
    out = f(a)
    GlobalCounters.reset()
    out.realize()
-    assert_kernel_count(kernel_count)
+    self.assertEqual(GlobalCounters.kernel_count, kernel_count)
    np.testing.assert_allclose(out.numpy(), [3., 5., 7., 9.])

  def test_custom_kernel_precompile_further_compute_multi(self): self.test_custom_kernel_precompile_further_compute(multi=True, kernel_count=4)
@@ -25,10 +25,10 @@ class TestHCQUnit(unittest.TestCase):
    cpu_call = UOp(Ops.PROGRAM, src=(UOp.sink(),)).call(UOp.new_buffer("CPU", 1, dtypes.float))
    gpu_devs = [d0]

-    # CPU uses HCQ2 and is no longer batched into legacy HCQ graphs.
+    # local MMIO: GPU works alone and with CPU in batch (cpu_support=True)
    assert HCQGraph.supports_uop(gpu_devs, gpu_call) is True
-    assert HCQGraph.supports_uop(gpu_devs, cpu_call) is False
-    assert HCQGraph.supports_uop(gpu_devs + [cpu_dev], gpu_call) is False
+    assert HCQGraph.supports_uop(gpu_devs, cpu_call) is True
+    assert HCQGraph.supports_uop(gpu_devs + [cpu_dev], gpu_call) is True

    # USB MMIO: GPU-only still works, but CPU batching must be rejected (cpu_support=False)
    orig_view = d0.timeline_signal.base_buf.view
@@ -3,7 +3,6 @@ from unittest.mock import patch
 from tinygrad import Tensor, UOp
 from tinygrad.schedule import schedule_cache
 from tinygrad.llm.model import Transformer, TransformerConfig
-from tinygrad.llm.serve import StreamRouter

 TEST_CONFIG = TransformerConfig(num_blocks=1, dim=64, hidden_dim=128, n_heads=2, n_kv_heads=2,
                           norm_eps=1e-5, vocab_size=100, head_dim=32, rope_theta=10000.0, rope_dim=32, v_head_dim=32, max_context=32)
@@ -11,59 +10,12 @@ V_START_POS = UOp.variable("start_pos", 0, TEST_CONFIG.max_context-1)
 V_TOKS = UOp.variable("toks", 1, 32)  # 32 is the default chunk_size in generate

 class TestTransformerGenerate(unittest.TestCase):
-  def test_warmup(self):
-    model, calls = Transformer(TEST_CONFIG), []
-    def generate(tokens, **kwargs):
-      calls.append(tokens)
-      yield from (1, 2)
-    with patch.object(model, "generate", generate): model.warmup()
-    self.assertEqual(calls, [[0], [0]])
-
-  def test_warmup_then_generate_with_default_chunk(self):
-    # warmup must not capture JIT graphs that generate()'s default chunk_size then rejects
-    model = Transformer(TEST_CONFIG)
-    model.warmup()
-    self.assertIsInstance(next(model.generate([5, 6, 7, 8])), int)
-
-  def test_first_recurrent_generate_before_state_init(self):
-    model = Transformer(TEST_CONFIG)
-    model.has_recurrent_block = True
-    with patch.object(Transformer, '__call__', return_value=Tensor([[42]])):
-      self.assertEqual(next(model.generate([0])), 42)
-
-  def test_recurrent_live_state_reuse(self):
-    model = Transformer(TEST_CONFIG)
-    model.has_recurrent_block = True
-    model._cached_tokens = [1, 2, 3, 4, 5]
-    self.assertEqual(model.get_start_pos([1, 2, 3, 4, 5, 42, 10]), 5)
-    calls = []
-    def mock_call(self, tokens, start_pos, temperature, **kwargs):
-      calls.append((tokens.shape, start_pos))
-      return Tensor([[42]])
-    with patch.object(Transformer, '__call__', mock_call):
-      next(model.generate([1, 2, 3, 4, 5, 42, 10]))
-    self.assertEqual(calls, [((1, 1), V_START_POS.bind(5)), ((1, 1), V_START_POS.bind(6))])
-
-  def test_recurrent_divergent_prompt_restarts(self):
-    model, calls = Transformer(TEST_CONFIG), []
-    model.has_recurrent_block, model._cached_tokens = True, [1, 2, 9]
-    def mock_call(self, tokens, start_pos, temperature):
-      calls.append(start_pos)
-      return Tensor([[42]])
-    with patch.object(Transformer, '__call__', mock_call): next(model.generate([1, 2, 10, 11]))
-    self.assertEqual(calls[0], V_START_POS.bind(0))
-
-  def test_template_starts_reasoning(self):
-    router = StreamRouter(reasoning=True)
-    self.assertEqual(list(router.route("reasoning</think>answer")),
-                     [("reasoning_content", "reasoning"), ("content", "answer")])
-
  def test_kv_cache_reuse(self):
    """Test that generate reuses the KV cache when tokens extend the cached prefix."""
    model = Transformer(TEST_CONFIG)

    captured_inputs = []
-    def mock_call(self, tokens, start_pos, temperature, **kwargs):
+    def mock_call(self, tokens, start_pos, temperature):
      captured_inputs.append((tokens.shape, start_pos))
      return Tensor([[42]])

@@ -88,7 +40,7 @@ class TestTransformerGenerate(unittest.TestCase):
    model = Transformer(TEST_CONFIG)

    captured_inputs = []
-    def mock_call(self, tokens, start_pos, temperature, **kwargs):
+    def mock_call(self, tokens, start_pos, temperature):
      captured_inputs.append((tokens.shape, start_pos))
      return Tensor([[42]])

@@ -136,7 +88,7 @@ class TestTransformerGenerate(unittest.TestCase):

    def get_prefill_flags(tokens, chunk_size):
      is_prefill = []
-      def mock_call(self, tokens, start_pos, temperature, **kwargs):
+      def mock_call(self, tokens, start_pos, temperature):
        is_prefill.append(resolve(tokens.shape[1] != 1))
        return Tensor([[42]])
      with patch.object(Transformer, '__call__', mock_call):
@@ -193,17 +145,11 @@ class TestTransformerGenerate(unittest.TestCase):
    # with temperature=2.0, we should see at least 2 distinct outputs across 5 runs
    self.assertGreater(len(runs), 1, "high temperature should produce varied outputs")

-  def test_recurrent_temperature_high_produces_variety(self):
-    model = Transformer(TEST_CONFIG)
-    model.has_recurrent_block = True
-    outputs = {model.forward(Tensor([[1]]), 0, Tensor([2.0])).item() for _ in range(5)}
-    self.assertGreater(len(outputs), 1)
-
  def test_temperature_passed_to_forward(self):
    """Temperature from generate should be passed through to __call__."""
    model = Transformer(TEST_CONFIG)
    captured_temps = []
-    def mock_call(self, tokens, start_pos, temperature, **kwargs):
+    def mock_call(self, tokens, start_pos, temperature):
      captured_temps.append(float(temperature.item()))
      return Tensor([[42]])
    with patch.object(Transformer, '__call__', mock_call):
@@ -1,4 +1,5 @@
 import unittest
+from unittest.mock import MagicMock
 from tinygrad import Device
 from tinygrad.uop.ops import Ops, UOp
 from tinygrad.dtype import dtypes
@@ -10,27 +11,36 @@ class TestMetalGraph(unittest.TestCase):
    self.MetalGraph = MetalGraph
    self.dev = Device[Device.DEFAULT]

-  def metal_buf(self, offset, bitcast=False):
-    size = 4 if bitcast else 1
-    buf = UOp.new_buffer(Device.DEFAULT, offset+size, dtypes.uint8)
-    if offset: buf = buf[offset:offset+size]
-    return buf.bitcast(dtypes.float32) if bitcast else buf
+  def metal_buf(self, offset):
+    buf = MagicMock()
+    if offset > 0:
+      buf.op = Ops.SLICE
+      src = MagicMock()
+      src.dtype = dtypes.uint8
+      buf.src = (src, UOp.const(offset))
+      buf.dtype = dtypes.uint8
+    else:
+      buf.op = Ops.BUFFER
+    buf.device = Device.DEFAULT
+    return buf

-  def supports_uop(self, *bufs):
-    return self.MetalGraph.supports_uop([self.dev], UOp(Ops.PROGRAM, src=(UOp.sink(),)).call(*bufs))
+  def call(self, *bufs):
+    c = MagicMock()
+    c.src = (MagicMock(op=Ops.PROGRAM),) + tuple(bufs)
+    return c

  def test_supports_uop_normal_offset(self):
-    assert self.supports_uop(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF)) is True
+    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(100), self.metal_buf(0xFFFFFFFF))) is True

  def test_supports_uop_overflow_offset(self):
-    assert self.supports_uop(self.metal_buf(0), self.metal_buf(0x100000000)) is False
+    assert self.MetalGraph.supports_uop([self.dev], self.call(self.metal_buf(0), self.metal_buf(0x100000000))) is False

-  def test_supports_uop_non_view_buf(self):
-    assert self.supports_uop(self.metal_buf(0)) is True
-
-  def test_supports_uop_bitcast(self):
-    assert self.supports_uop(self.metal_buf(0xFFFFFFFF, bitcast=True)) is True
-    assert self.supports_uop(self.metal_buf(0x100000000, bitcast=True)) is False
+  def test_supports_uop_nonmetal_buf(self):
+    # non-SLICE ops should not be checked for offset
+    buf = MagicMock()
+    buf.op = Ops.BUFFER
+    buf.device = Device.DEFAULT
+    self.MetalGraph.supports_uop([self.dev], self.call(buf))

 if __name__ == "__main__":
  unittest.main()
@@ -2,7 +2,7 @@ import unittest, numpy as np
 from tinygrad import Tensor, Variable, Context, Device, TinyJit, GlobalCounters, dtypes, UOp, nn, getenv
 from tinygrad.nn.state import get_parameters, get_state_dict
 from tinygrad.uop.ops import Ops
-from test.helpers import not_support_multi_device, needs_second_gpu, slow, assert_kernel_count, KernelCountException
+from test.helpers import not_support_multi_device, needs_second_gpu, slow
 from hypothesis import given, strategies as strat, settings

 settings.register_profile("my_profile", max_examples=200, deadline=None, derandomize=getenv("DERANDOMIZE_CI", False))
@@ -143,8 +143,9 @@ class TestMultiTensor(unittest.TestCase):
    GlobalCounters.reset()
    with Context(ALLREDUCE_CAST=1, RING=0, ALL2ALL=0):
      tst.realize()
-    assert_kernel_count(kernel_count)
+    kernel_count = GlobalCounters.kernel_count
    np.testing.assert_allclose(tst.numpy(), (a_src.numpy()+b_src.numpy()).sum(0))
+    self.assertEqual(kernel_count, kernel_count)

  def test_allreduce_cast_half_assign(self): self.test_allreduce_cast_half(assign=True, kernel_count=10)

@@ -384,12 +385,6 @@ class TestMultiTensor(unittest.TestCase):
      np.testing.assert_allclose(r.numpy(), np.ones(256)+np.ones(256), atol=1e-4, rtol=1e-5)
    assert jf.captured is not None

-  def test_symbolic_broadcast_copy(self):
-    rows = Variable("rows", 1, 4).bind(3)
-    out = Tensor.ones(rows, 8).to(devices_2).realize()
-    self.assertEqual(out.shape, (rows, 8))
-    np.testing.assert_equal(out[:3].to(Device.DEFAULT).numpy(), np.ones((3, 8)))
-
  def test_multitensor_jit_in_list(self):
    # test MULTI tensor inside a list container - exercises the container unpacking + MULTI unpacking
    @TinyJit
@@ -589,7 +584,7 @@ class TestMultiTensor(unittest.TestCase):
      zeros = Tensor.zeros(3).realize()
    b = a.to(devices_2)*zeros.to(devices_2)
    sched = b.schedule_linear().src
-    if len(sched) != 0: raise KernelCountException(0, len(sched))
+    self.assertEqual(len(sched), 0)
    self.assertListEqual(b.tolist(), [0, 0, 0])

@unittest.skipIf(not_support_multi_device(), "no multi")
@@ -1,16 +1,15 @@
 import unittest
 from tinygrad import Tensor, dtypes, GlobalCounters
-from test.helpers import assert_kernel_count

 class TestSetitemInto(unittest.TestCase):
  def test_setitem_into_unrealized(self):
    GlobalCounters.reset()
    t = Tensor.arange(4, dtype=dtypes.int32).reshape(2, 2)
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(t.tolist(), [[0, 1], [5, 5]])

@@ -19,11 +18,11 @@ class TestSetitemInto(unittest.TestCase):
    GlobalCounters.reset()
    a = Tensor.arange(8, dtype=dtypes.int32).reshape(2, 4)
    w = a[0] + a[1]  # unrealized ADD with SHRINK in graph: [4, 6, 8, 10]
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    w[1] = 99
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    w.realize()
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(w.tolist(), [4, 99, 8, 10])

@@ -31,61 +30,61 @@ class TestSetitemInto(unittest.TestCase):
    GlobalCounters.reset()
    t = Tensor.empty(4, dtype=dtypes.int32)
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(GlobalCounters.global_mem, 4)
    t[1].realize()
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(t[1].item(), 5)

  def test_setitem_into_empty_alu(self):
    GlobalCounters.reset()
    t = Tensor.empty(4, dtype=dtypes.int32) + 1
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertLessEqual(GlobalCounters.global_mem, 32)
    t[1].realize()
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(t[1].item(), 5)

  def test_setitem_into_tensor(self):
    t = Tensor([1, 2, 3, 4], dtype=dtypes.int32).realize()
    GlobalCounters.reset()
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t[1].realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(GlobalCounters.global_mem, 4)
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertListEqual(t.tolist(), [1, 5, 3, 4])

  def test_setitem_into_tensor_alu(self):
    t = Tensor([1, 2, 3, 4], dtype=dtypes.int32).realize() + 1
    GlobalCounters.reset()
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t[1].realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertLessEqual(GlobalCounters.global_mem, 32)
    t[1].realize()
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertListEqual(t.tolist(), [2, 5, 4, 5])

  def test_setitem_into_const(self):
    GlobalCounters.reset()
    t = Tensor.ones(4, dtype=dtypes.int32, buffer=False)
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(t.tolist(), [1, 5, 1, 1])

@@ -93,9 +92,9 @@ class TestSetitemInto(unittest.TestCase):
    GlobalCounters.reset()
    t = Tensor.ones(4, dtype=dtypes.int32, buffer=False) + 1
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    self.assertEqual(GlobalCounters.global_mem, 0)
    self.assertListEqual(t.tolist(), [2, 5, 2, 2])

@@ -106,18 +105,18 @@ class TestSetitemInto(unittest.TestCase):
    t = Tensor.arange(4, dtype=dtypes.int32)
    self.assertIs(other.uop, t.uop)
    t[1] = 5
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    self.assertListEqual(t.tolist(), [0, 5, 2, 3])

  def test_setitem_slice_const(self):
    t = Tensor.zeros(100, dtype=dtypes.int32).contiguous().realize()
    GlobalCounters.reset()
    t[20:50] = 3
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(GlobalCounters.global_mem, 30*4)  # 30 elements written

  def test_setitem_slice_tensor(self):
@@ -125,18 +124,18 @@ class TestSetitemInto(unittest.TestCase):
    v = Tensor.zeros(30, dtype=dtypes.int32).contiguous().realize()
    GlobalCounters.reset()
    t[20:50] = v
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(GlobalCounters.global_mem, 30*4*2)  # 30 read + 30 written

  def test_setitem_full(self):
    t = Tensor.zeros(100, dtype=dtypes.int32).contiguous().realize()
    GlobalCounters.reset()
    t[:] = 3
-    assert_kernel_count(0)
+    self.assertEqual(GlobalCounters.kernel_count, 0)
    t.realize()
-    assert_kernel_count(1)
+    self.assertEqual(GlobalCounters.kernel_count, 1)
    self.assertEqual(GlobalCounters.global_mem, 100*4)  # full buffer written

 if __name__ == '__main__':
@@ -0,0 +1,221 @@
+from dataclasses import dataclass, field
+from tinygrad.dtype import dtypes, AddrSpace
+from tinygrad.uop.ops import UOp, UPat, PatternMatcher, Ops, GroupOp, ParamArg, graph_rewrite, track_rewrites
+from tinygrad.helpers import VIZ, pluralize, all_int
+
+@dataclass
+class AllocCtx:
+  uop_list: list[UOp] = field(default_factory=list)
+  buffer_map: dict[UOp, UOp] = field(default_factory=dict)
+  bases: set[UOp] = field(default_factory=set)
+  assigns: list[UOp] = field(default_factory=list)
+  replacements: list[UOp] = field(default_factory=list)
+
+def tag_uop(ctx:AllocCtx, x:UOp):
+  if x.tag is not None: return None
+  ctx.uop_list.append(x)
+  return x.replace(tag=(len(ctx.uop_list)-1,))
+
+def disk_like(u:UOp): return isinstance(u.device, str) and u.device.startswith(("DISK", "TINYFS"))
+
+def disk_copy_is_buffer(ctx:AllocCtx, u:UOp):
+  # copies to disk are replaced with the disk buffer
+  if disk_like(u) and u.tag is None:
+    ctx.buffer_map[u] = u.empty_like()
+    return u.rtag(())
+  # all copies from disk/numpy are realized into a real buffer
+  from_creation = isinstance(u.src[0].device, str) and u.src[0].device.startswith(("NPY", "DISK", "PYTHON", "TINYFS"))
+  if from_creation: return tag_uop(ctx, u)
+
+# CONTIGUOUS and AFTER + parents are the only nodes that get updated
+add_tags = PatternMatcher([
+  (UPat(Ops.COPY, name="u"), disk_copy_is_buffer),
+  # no tag on copies that are assigned via STORE+AFTER — merge COPY tag into AFTER
+  (UPat(Ops.AFTER, src=(UPat(), UPat(Ops.STORE, src=(UPat(name="dest"), UPat(Ops.COPY, name="c")))), name="a"),
+   lambda a,c,dest: a.replace(src=(a.src[0], a.src[1].replace(src=(dest, c.rtag(())))), tag=a.tag+c.tag) if a.tag and c.tag else None),
+  (UPat((Ops.CONTIGUOUS, Ops.AFTER), name="x"), tag_uop),
+  (UPat(GroupOp.All, name="x"), lambda ctx,x: tag_uop(ctx,x) if x in ctx.bases else None),
+])
+
+def replace_contig_with_store_after(u:UOp):
+  # can't allocate a buffer for a virtual value
+  if u.is_virtual: return None
+  # if size is 0, remove the contig
+  if 0 in u.shape: return u.src[0]
+  # no real contig for DISK/TINYFS tensors, they are left alone
+  if disk_like(u): return u.rtag(None)
+  buf = u.empty_like()
+  return buf.after(buf.store(u.src[0])).rtag(u.tag)
+
+def replace_store_after_with_contig(u:UOp, src:UOp):
+  assigned_to = u
+  while assigned_to.op in {Ops.BITCAST, Ops.AFTER, Ops.UNSHARD}: assigned_to = assigned_to.src[0].base
+  if assigned_to.op not in {Ops.BUFFER, Ops.SLICE}: 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 SLICE. Otherwise None."""
+  if (offset := src.contiguous_view_offset()) is None: return None
+  buf = src.base
+  if buf.op is Ops.SLICE:
+    byte_offset = buf.src[1].val * 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(offset)), src.numel())
+
+def contiguous_mops_to_view(c:UOp, src:UOp):
+  """MOPS(BUFFER) → SLICE when movement ops collapse to a contiguous range."""
+  buf = src.base
+  if buf.op not in {Ops.BUFFER, Ops.SLICE, Ops.UNSHARD}: return None
+  if src.op is Ops.RESHAPE and src.src[0].op in {Ops.BUFFER, Ops.SLICE} and c.op is not Ops.BITCAST: return None
+  if c.op is not Ops.BITCAST and src.op is Ops.BUFFER: return None
+
+  # no symbolic shape
+  if not all_int(c.shape): return None
+
+  if buf.op is not Ops.UNSHARD and (view := _make_buffer_view(src)) is not None:
+    view = (view.replace(dtype=c.dtype, arg=c.numel()) if c.op is Ops.BITCAST else view).reshape(c.shape)
+    return c.replace(src=(view,)) if c.op is Ops.COPY else view
+
+  # for UNSHARD 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")
+    if resolved.op is not Ops.UNSHARD: return None
+    if (view := _make_buffer_view(resolved.src[0])) is None: return None
+    return view.reshape(resolved.src[0].shape).unshard(resolved.arg, resolved.src[1:]).contiguous(tag=c.tag)
+
+  return None
+
+def _precompiled_output_redirect(s:UOp, t:UOp) -> UOp|None:
+  # how output s lands in the caller's buffer t, or None if it must be copied into t
+  # materialize straight into t
+  if s.op is Ops.CONTIGUOUS: return t.after(t.store(s.src[0]))
+  # rebind output storage to t
+  if s.op in {Ops.BUFFER, Ops.UNSHARD} and s.has_buffer_identity(): return t
+  return None
+
+def transform_precompiled_call(c:UOp) -> UOp|None:
+  if not c.arg.precompile: return None
+  assert c.src[0].op is Ops.TUPLE, f"expected TUPLE body for precompiled FUNCTION, got {c.src[0].op}"
+  input_buffers = tuple(x.contiguous() if x.op not in {Ops.AFTER, Ops.BIND} else x for x in c.src[1:])
+
+  # add the outputs to the call
+  srcs = c.src[0].src
+  resolved = [c.gettuple(i) for i in range(len(srcs))]
+  outs = tuple(r.empty_like() for r in resolved)
+  targets = [o.param_like(len(c.src)-1+i).shrink_to(s.shape) for i,(o,s) in enumerate(zip(outs, srcs))]
+
+  subs:dict[UOp, UOp] = {}
+  items:list[UOp] = []
+  for s, t in zip(srcs, targets):
+    after_deps:list[UOp] = []
+    while s.op is Ops.AFTER:
+      after_deps.extend(s.src[1:])
+      s = s.src[0]
+    if (placed := _precompiled_output_redirect(s, t)) is not None and s not in subs:
+      subs[s] = placed
+      items.append(s.after(*after_deps) if after_deps else s)
+    else:
+      items.append(t.after(t.store(s.after(*after_deps))))
+  fxn = UOp.sink(*(x.substitute(subs) for x in items))
+
+  # body switches from TUPLE to SINK, so the node becomes an opaque CALL (not FUNCTION)
+  new_call = UOp(Ops.CALL, src=(fxn, *input_buffers, *outs), arg=c.arg)
+  rets = tuple(o.after(new_call) for o in outs)
+
+  # if the CALL has symbolic shapes, shrink the max-sized output to the actual symbolic shape
+  # NOTE: must use resolved shapes from the FUNCTION (which substitutes PARAMs with external args), not raw body shapes
+  rets = tuple(r.shrink_to(rs.shape) for r,rs in zip(rets, resolved))
+
+  return UOp.maketuple(*rets)
+
+# NOTE: adding rules to here is bad. these all need to run before the schedule cache
+pm_early_transform_tensor_graph = PatternMatcher([
+  # transform precompiled FUNCTIONs into CALLs (body becomes SINK with stores)
+  (UPat(Ops.FUNCTION, name="c"), transform_precompiled_call),
+
+  # resolve TUPLE+GETTUPLE (for precompiled calls)
+  (UPat(Ops.GETTUPLE, src=(UPat(Ops.TUPLE, name="t"),), name="g"), lambda g,t: t.src[g.arg]),
+
+  # fold MOPS+BITCAST over BUFFER/SLICE into SLICE when movement ops collapse to contiguous range
+  (UPat((Ops.BITCAST, Ops.COPY, Ops.CONTIGUOUS), src=(UPat(GroupOp.Movement|{Ops.BUFFER}, name="src"),), name="c"), contiguous_mops_to_view),
+
+  # remove contiguous on movement ops before a copy on disk
+  (UPat(GroupOp.Movement-{Ops.SHRINK, Ops.RESHAPE}, name="x").f(Ops.CONTIGUOUS).f(Ops.COPY, name="copy"), lambda x,copy:
+   copy.replace(src=(x,), tag=None) if isinstance(x.device, str) and x.device.startswith("DISK") else None),
+  # push copy past movement ops to disk
+  (UPat(GroupOp.Movement-{Ops.SHRINK, Ops.RESHAPE}, name="x").f(Ops.COPY, name="copy"), lambda x,copy:
+   x.replace(src=(copy.replace(src=(x.src[0],), tag=None),)+x.src[1:]) \
+   if isinstance(x.device, str) and x.device.startswith("DISK") else None),
+
+  # add CONTIGUOUS to tagged UOps
+  (UPat(GroupOp.All-{Ops.CONTIGUOUS, Ops.AFTER, Ops.STORE}, name="x"),
+   lambda x: None if x.tag is None else x.rtag(None).contiguous(tag=x.tag) if x.tag else x.replace(tag=None)),
+  # remove extra CONTIGUOUS on AFTER (only when target is contiguous)
+  (UPat(Ops.CONTIGUOUS, src=(UPat(Ops.AFTER, name="a"),), name="c"),
+   lambda a,c: a.replace(tag=(a.tag or ())+(c.tag or ())) if a.src[0].has_buffer_identity() else None),
+  # replace AFTER+STORE with CONTIGUOUS when target is not a buffer
+  (UPat(Ops.AFTER, src=(UPat(), UPat(Ops.STORE, src=(UPat(), UPat(name="src")))), name="u"), replace_store_after_with_contig),
+  # replace CONTIGUOUS with STORE+AFTER
+  (UPat(Ops.CONTIGUOUS, name="u"), replace_contig_with_store_after),
+  # remove DETACH/CONTIGUOUS_BACKWARD (allows more contiguous removal)
+  (UPat((Ops.DETACH, Ops.CONTIGUOUS_BACKWARD), name="x"), lambda x: x.src[0]),
+])
+
+def finalize_after(ctx:AllocCtx, x:UOp):
+  # untagged: record as an assign for the call body
+  if x.tag is None:
+    ctx.assigns.append(x)
+    return None
+  # tagged: untag and map each original pre-rewrite UOp to the stripped buffer; the untagged result is reprocessed as untagged
+  ret = x.replace(tag=None)
+  replace_uop = ret
+  while replace_uop.op is Ops.AFTER: replace_uop = replace_uop.src[0]
+  for t in x.tag:
+    original_uop: UOp = ctx.uop_list[t]
+    ctx.buffer_map[original_uop] = replace_uop.shrink_to(original_uop.shape)
+  return ret
+
+def replace_input_buffer(ctx:AllocCtx, b:UOp):
+  ctx.replacements.append(b)
+  if b.op is Ops.BIND: return b.param_like(len(ctx.replacements)-1)
+  return UOp.param(len(ctx.replacements)-1, b.dtype, b.shape, b.device,
+                   addrspace=b.addrspace if b.addrspace is not None else AddrSpace.GLOBAL)
+
+pm_finalize_call = PatternMatcher([
+  (UPat(Ops.AFTER, name="x"), finalize_after),
+  (UPat(Ops.COPY, name="x"), lambda ctx,x: ctx.assigns.append(x) if isinstance(x.device, str) and x.device.startswith(("DISK", "TINYFS")) else None),
+])
+
+pm_replace_buf = PatternMatcher([
+  # replace BUFFER with PARAM for cache key normalization
+  (UPat(Ops.BUFFER, src=(UPat(),), name="b"), lambda ctx,b:
+   replace_input_buffer(ctx, b) if isinstance(b.arg, ParamArg) and b.addrspace is AddrSpace.GLOBAL else None),
+  # 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.PARAM), UPat(Ops.CONST)), name="b"), replace_input_buffer),
+])
+
+@track_rewrites(lambda _,ret: f"Callify {pluralize('Buffer', len(ret[1]))}")
+def transform_to_call(big_sink:UOp) -> tuple[UOp, dict[UOp, UOp]]:
+  if VIZ: graph_rewrite(big_sink, PatternMatcher([]), name="View Tensor Graph")
+  # uop list is a list in the original_sink graph and we can map to the tags later
+  # same predicate as Tensor.realize
+  ctx = AllocCtx(bases={base for x in big_sink.src if not (base:=x.base).is_virtual and not base.has_buffer_identity()
+                        and base.op is not Ops.AFTER and base.addrspace is not AddrSpace.ALU})
+
+  # this rewrite is "read-only", it adds simple things to buffer_map and may sink things on big_sink, bottom_up
+  # this is the only one where we have to be careful to not break the tensor graph
+  big_sink = graph_rewrite(big_sink, add_tags, ctx=ctx, bottom_up=True, name="number the uops")
+
+  # here we can break the tensor graph. this is the only place you need to maintain numbered tags
+  big_sink = graph_rewrite(big_sink, pm_early_transform_tensor_graph, name="early transform tensor graph")
+
+  # here we construct the final buffer_map: as-built nodes -> their final storage. values are never keys
+  graph_rewrite(big_sink, pm_finalize_call, ctx=ctx, name="finalize call")
+  ret = graph_rewrite(UOp.sink(*ctx.assigns), pm_replace_buf, ctx=ctx, bottom_up=True, name="replace bufs").call(*ctx.replacements)
+  assert not any(x in ctx.buffer_map for x in ctx.buffer_map.values())
+  if VIZ: graph_rewrite(ret, PatternMatcher([]), name="View Call")
+  return ret, ctx.buffer_map
@@ -2,8 +2,8 @@ from dataclasses import replace, dataclass
 import itertools, functools
 from tinygrad.helpers import DISABLE_FAST_IDIV, TRANSCENDENTAL, SPEC, DEBUG, VIZ, IMAGE, NOOPT, EMULATED_DTYPES, NOLOCALS, USE_TC
 from tinygrad.helpers import ALLOW_TF32, DEFAULT_FLOAT, DEFAULT_INT, TracingKey, Context, panic
-from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, Ops, UPat, rewrite_group, KernelInfo, ProgramInfo, GroupOp, AxisType
-from tinygrad.uop.weak import pm_lower_index_dtype, pm_commit_weak, pm_cast_weak
+from tinygrad.uop.ops import PatternMatcher, graph_rewrite, UOp, pm_lower_index_dtype, Ops, UPat, track_rewrites, KernelInfo, ProgramInfo, GroupOp
+from tinygrad.uop.ops import AxisType, pm_commit_weak, pm_cast_weak
 from tinygrad.uop.render import pyrender
 from tinygrad.uop.spec import type_verify, spec_tensor, spec_program
 from tinygrad.renderer import Renderer, Estimates
@@ -12,7 +12,7 @@ from tinygrad.dtype import dtypes, AddrSpace

 # import all pattern matchers here
 from tinygrad.codegen.gpudims import pm_add_gpudims
-from tinygrad.uop.symbolic import sym, symbolic_simple, symbolic, pm_fold_cast_const, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
+from tinygrad.uop.symbolic import sym, symbolic_simple, symbolic, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
 from tinygrad.uop.movement import mop_cleanup
 from tinygrad.codegen.decomp.dtype import pm_dtype_decomps
 from tinygrad.codegen.decomp.op import get_late_rewrite_patterns, get_simplifying_rewrite_patterns
@@ -20,7 +20,7 @@ from tinygrad.codegen.decomp.transcendental import get_transcendental_patterns
 from tinygrad.codegen.late.coalesce import indexing_simplify
 from tinygrad.codegen.opt.postrange import apply_opts
 from tinygrad.codegen.late.gater import pm_move_gates_from_index
-from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse, pm_reduce_unparented
+from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse
 from tinygrad.schedule.multi import multi_pm
 from tinygrad.schedule.rangeify import pm_mops
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
@@ -125,7 +125,8 @@ def do_devectorize(b:UOp):
  # broadcasting needs to be already unpacked, Invalid matches any dtype and shape
  if not all(x.shape == b.shape or x.base.is_invalid for x in b.src): return None
  src = []
-  for idx_c in itertools.product(*[[UOp.const(i) for i in range(x)] for x in b.shape]):
+  for idx in itertools.product(*[range(x) for x in b.shape]):
+    idx_c = [UOp.const(i) for i in idx]
    src.append(b.replace(dtype=None, src=tuple(x.base if x.base.is_invalid else x.index(*idx_c) for x in b.src)))
  return UOp.stack(*src).reshape(b.shape) if b.op is not Ops.STORE else UOp.group(*src)

@@ -213,7 +214,7 @@ def reduce_ranges_to_acc(ctx:ReduceContext, r:UOp):
  topo = r.src[0].toposort()
  ended_ranges = flatten([x.ended_ranges for x in topo if x.op is Ops.END])
  input_ranges = tuple(x for x in topo if x.op is Ops.RANGE and x not in r.src[1:] and x not in ended_ranges)
-  acc_init = acc.after(*input_ranges).store(UOp.const(identity_element(r.arg[0], r.dtype)))
+  acc_init = acc.after(*input_ranges).store(identity_element(r.arg[0], r.dtype))
  acc_initted = acc.after(acc_init, *r.src[1:])
  inp = r.src[0].reduce(arg=r.arg) if r.arg[1] else r.src[0]
  acc_out = acc_initted.store(acc_initted.alu(r.arg[0], inp)).end(*r.src[1:]).rtag("mergeable")
@@ -301,7 +302,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
    sink = graph_rewrite(sink, pm_split_ranges+pm_flatten_range, ctx={}, name="split ranges")

    # symbolic (NOTE: this is a requirement for pm_simplify_ranges to be correct)
-    sink = graph_rewrite(sink, sym+pm_fold_cast_const+pm_flatten_range, name="initial symbolic")
+    sink = graph_rewrite(sink, sym+pm_flatten_range, name="initial symbolic")

    # optimize (schedule) the AST
    sink = graph_rewrite(sink, pm_flatten_range+pm_simplify_ranges, ctx={}, name="simplify ranges")
@@ -310,8 +311,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
    sink = apply_opts(sink, ren, beam=ast.arg.beam)

  # ** expander (expand_rewrite) **
-  # reduce_unparented: a REDUCE whose src folded to a CONST (e.g. x*0) has no parented ranges, collapse it before the expander
-  sink = graph_rewrite(sink, sym+pm_move_where_on_load+pm_flatten_range+pm_reduce_unparented, name="postopt symbolic")
+  sink = graph_rewrite(sink, sym+pm_move_where_on_load+pm_flatten_range, name="postopt symbolic")

  # expand
  sink = graph_rewrite(sink, expander2, ctx=build_range_map(sink), name="expander")
@@ -330,23 +330,24 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
  sink = graph_rewrite(sink, symbolic_simple+pm_expand_broadcast+pm_add_loads, name="*** expand broadcast / add loads")

  # devectorize
-  sink = graph_rewrite(sink, symbolic_simple+devectorizer2+indexing_simplify, ctx=ren, name="devectorize2")
+  sink = graph_rewrite(sink, symbolic_simple+devectorizer2, ctx=ren, name="devectorize2")
+
+  # simplify indexing
+  sink = graph_rewrite(sink, indexing_simplify, name="simplify load/store indexing")

  # some coalescing misses without this
  sink = graph_rewrite(sink, sym, name="early symbolic")

  # do memory coalescing (late)
  sink = memory_coalescing(sink, ren)
-  sink = graph_rewrite(sink, symbolic_simple+ew_devectorizer+pm_simplify_add_image,
-                       name="add images", ctx=({}, ren), bottom_up=True)
+  sink = graph_rewrite(sink, symbolic_simple+ew_devectorizer+pm_simplify_add_image, name="add images", ctx=({}, ren), bottom_up=True)

  # extra symbolic before decomp. crashes without this?
-  # NOTE: also run indexing_simplify here, while the index is still weakint and (x+y)*c -> x*c+y*c applies
-  sink = graph_rewrite(sink, sym+indexing_simplify, name="extra symbolic")
+  sink = graph_rewrite(sink, sym, name="extra symbolic")

  # lower index dtype
  # NOTE: we need indexing_simplify to remove the cast to long using the Invalid
-  sink = graph_rewrite(sink, symbolic_simple+pm_fold_cast_const+pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")
+  sink = graph_rewrite(sink, pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")

  # final symbolic before decomp
  sink = graph_rewrite(sink, symbolic, name="final symbolic")
@@ -357,7 +358,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:

  # floordiv+mod / dtype decomp (early)
  supported_ops = tuple(ren.code_for_op.keys())
-  pm_decomp = symbolic_simple+pm_fold_cast_const+get_simplifying_rewrite_patterns(supported_ops)
+  pm_decomp = symbolic_simple+get_simplifying_rewrite_patterns(supported_ops)
  sink = graph_rewrite(sink, pm_decomp, name="early decompositions")

  # late decomps + move gates from unrenderable INVALID where
@@ -451,7 +452,7 @@ pm_to_program = PatternMatcher([
  (UPat(Ops.PROGRAM, src=(UPat(), UPat(Ops.LINEAR), UPat(Ops.SOURCE, name="source")), name="prg"), do_compile),
 ])

-@rewrite_group(name=lambda ast,renderer,ret,**kwargs: TracingKey(ret.src[0].arg.name,(ret.src[0].arg.function_name, ast), ret=renderer), replay=True)
+@track_rewrites(name=lambda ast,renderer,ret,**kwargs: TracingKey(ret.src[0].arg.name,(ret.src[0].arg.function_name, ast), ret=renderer), replay=True)
@Context(ALLOW_DEVICE_USAGE=0)
 def do_to_program(ast:UOp, renderer:Renderer) -> UOp:
  """
@@ -78,11 +78,9 @@ def l2i(op: Ops, dt: DType, *uops:UOp):
    case Ops.MAX: return l2i(Ops.WHERE, dt, l2i(Ops.CMPLT, dt, *uops), b0, b1, a0, a1)
    case _: raise NotImplementedError(f"long decomposition of {op} unsupported")

-def split_l2i(ctx:dict, op: Ops, dt: DType, *uops:UOp):
-  # l2i does arithmetic on its inputs; rules enter here to split them to 32-bit words first, l2i recurses on itself.
-  # both word halves of a node ask for the same split, so ctx memos it for the pass
-  if (key:=(op, dt, uops)) not in ctx: ctx[key] = l2i(op, dt, *graph_rewrite(UOp.sink(*uops), pm_long_decomp, ctx=ctx, bottom_up=True).src)
-  return ctx[key]
+def split_l2i(op: Ops, dt: DType, *uops:UOp):
+  # l2i does arithmetic on its inputs; rules enter here to split them to 32-bit words first, l2i recurses on itself
+  return l2i(op, dt, *graph_rewrite(UOp.sink(*uops), pm_long_decomp, bottom_up=True).src)

 # ***** floats *****
 f2f_dt = { f:getattr(dtypes, f"uint{f.bitsize}") for f in dtypes.floats }
@@ -99,8 +97,7 @@ def f2f(v, fr:DType, to:DType, sat=True):
    if fr in dtypes.fp8_fnuz:
      fnuz_nan = sign.ne(0) & nosign.eq(0)
      qnan = shl(shl(1, te) - 1, tm) | shl(1, tm - 1)
-      # the fnuz bias can exceed the target's: exp in [1, fb-tb] is normal in fr but lands below to's normal range, so it flushes like a denormal
-      return fnuz_nan.where(qnan, sign | (exp < max(fb - tb, 0) + 1).where(0, norm)).bitcast(to)
+      return fnuz_nan.where(qnan, sign | exp.eq(0).where(0, norm)).bitcast(to)
    # fp8e4m3 has only one nan
    is_nan = (nosign.eq(shl(1, fm + fe) - 1) if fr == dtypes.fp8e4m3 else exp.eq(shl(1, fe) - 1))
    return (sign | exp.eq(0).where(0, is_nan.where(nan, norm))).bitcast(to)
@@ -142,21 +139,21 @@ pm_long_decomp = PatternMatcher([
  (UPat(Ops.STORE, src=(UPat.var('idx', tuple(l2i_dt.keys())), UPat.var('val')), name='st'), lambda st,idx,val:
   st.replace(src=(idx.rtag((0, dt:=l2i_dt[idx.dtype])), val.rtag((0, dt)))).group(
     st.replace(src=(idx.rtag((1, dt)), val.rtag((1, dt))))) if val.tag is None else None),
-  (UPat(GroupOp.Comparison, src=[UPat.var('a', tuple(l2i_dt.keys())), UPat()], name="x"), lambda ctx,a,x:
-   split_l2i(ctx, x.op, dt:=l2i_dt[a.dtype], *flatten((s.rtag((0, dt)), s.rtag((1, dt))) for s in x.src))),
-  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda ctx,a,x:
-   split_l2i(ctx, Ops.BITCAST, l2i_dt[x.dtype], a.rtag((0, dt:=l2i_dt[a.dtype])), a.rtag((1, dt)))[x.tag[0]]),
-  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a'),), name="x"), lambda ctx,a,x:
-   split_l2i(ctx, x.op, x.dtype, a)[x.tag[0]] if x.tag is not None else None),
-  (UPat(Ops.CAST, src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda ctx,a,x:
-   split_l2i(ctx, x.op, x.dtype, a.rtag((0, dt:=l2i_dt[a.dtype])), a.rtag((1, dt))) if x.dtype not in l2i_dt and a.tag is None else None),
-  (UPat((Ops.SHL, Ops.SHR), tuple(l2i_dt.keys()), src=(UPat.var('a'), UPat.var('b')), name="x"), lambda ctx,a,b,x:
-   split_l2i(ctx, x.op, dt:=l2i_dt[x.dtype], a.rtag((0, dt)), a.rtag((1, dt)), b.rtag((0, dt)))[x.tag[0]] if x.tag is not None else None),
-  (UPat(Ops.WHERE, tuple(l2i_dt.keys()), src=(UPat.var('c'), UPat.var('a'), UPat.var('b')), name="x"), lambda ctx,a,b,c,x:
-   split_l2i(ctx, x.op, dt:=l2i_dt[x.dtype], c, a.rtag((0, dt)), a.rtag((1, dt)), b.rtag((0, dt)), b.rtag((1, dt)))[x.tag[0]]
+  (UPat(GroupOp.Comparison, src=[UPat.var('a', tuple(l2i_dt.keys())), UPat()], name="x"), lambda a,x:
+   split_l2i(x.op, dt:=l2i_dt[a.dtype], *flatten((s.rtag((0, dt)), s.rtag((1, dt))) for s in x.src))),
+  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda a,x:
+   split_l2i(Ops.BITCAST, l2i_dt[x.dtype], a.rtag((0, dt:=l2i_dt[a.dtype])), a.rtag((1, dt)))[x.tag[0]]),
+  (UPat(Ops.CAST, tuple(l2i_dt.keys()), src=(UPat.var('a'),), name="x"), lambda a,x:
+   split_l2i(x.op, x.dtype, a)[x.tag[0]] if x.tag is not None else None),
+  (UPat(Ops.CAST, src=(UPat.var('a', tuple(l2i_dt.keys())),), name="x"), lambda a,x:
+   split_l2i(x.op, x.dtype, a.rtag((0, dt:=l2i_dt[a.dtype])), a.rtag((1, dt))) if x.dtype not in l2i_dt and a.tag is None else None),
+  (UPat((Ops.SHL, Ops.SHR), tuple(l2i_dt.keys()), src=(UPat.var('a'), UPat.var('b')), name="x"), lambda a,b,x:
+   split_l2i(x.op, dt:=l2i_dt[x.dtype], a.rtag((0, dt)), a.rtag((1, dt)), b.rtag((0, dt)))[x.tag[0]] if x.tag is not None else None),
+  (UPat(Ops.WHERE, tuple(l2i_dt.keys()), src=(UPat.var('c'), UPat.var('a'), UPat.var('b')), name="x"), lambda a,b,c,x:
+   split_l2i(x.op, dt:=l2i_dt[x.dtype], c, a.rtag((0, dt)), a.rtag((1, dt)), b.rtag((0, dt)), b.rtag((1, dt)))[x.tag[0]]
   if x.tag is not None else None),
-  (UPat((*(GroupOp.ALU - GroupOp.Comparison - {Ops.SHL, Ops.SHR, Ops.WHERE}), Ops.BITCAST), tuple(l2i_dt.keys()), name="x"), lambda ctx,x:
-   split_l2i(ctx, x.op, l2i_dt[x.dtype], *flatten((a.rtag((0, l2i_dt[x.dtype])), a.rtag((1, l2i_dt[x.dtype]))) for a in x.src))[x.tag[0]]
+  (UPat((*(GroupOp.ALU - GroupOp.Comparison - {Ops.SHL, Ops.SHR, Ops.WHERE}), Ops.BITCAST), tuple(l2i_dt.keys()), name="x"), lambda x:
+   split_l2i(x.op, l2i_dt[x.dtype], *flatten((a.rtag((0, l2i_dt[x.dtype])), a.rtag((1, l2i_dt[x.dtype]))) for a in x.src))[x.tag[0]]
   if x.tag is not None else None),
  (UPat(Ops.LOAD, tuple(l2i_dt.keys()), src=(UPat.var('idx'),), name='x'), lambda x,idx:
   x.replace(dtype=l2i_dt[x.dtype], src=(reindex(idx, x.tag[0]).replace(dtype=l2i_dt[x.dtype], tag=None),), tag=None) if x.tag is not None else None),
@@ -181,9 +178,7 @@ pm_float_decomp = PatternMatcher([
   f2f(x.bitcast(f2f_dt[ctx[0]]), ctx[0], ctx[1]) if bc.dtype == ctx[0] else None),
  (UPat(Ops.CAST, dtypes.floats, src=(UPat.var("val"),), name="x"), lambda ctx,x,val:
   f2f_clamp(val.cast(ctx[1]), ctx[0]) if x.dtype == ctx[0] else None),
-  # a CONST has no srcs to cast, it restates its value at the emulating dtype
-  (UPat(Ops.CONST, dtypes.floats, name="x"), lambda ctx,x: UOp.const(x.val, ctx[1]) if x.dtype == ctx[0] else None),
-  (UPat(GroupOp.All-GroupOp.Defines-{Ops.CAST, Ops.BITCAST, Ops.CONST}, dtypes.floats, name="x"), lambda ctx,x:
+  (UPat(GroupOp.All-{Ops.BITCAST}, dtypes.floats, name="x"), lambda ctx,x:
   x.replace(dtype=ctx[1], src=tuple(s.cast(ctx[1]) if s.dtype == ctx[0] else s for s in x.src))
   if x.dtype == ctx[0] else None),
  (UPat(Ops.STORE, src=(UPat.var("idx"), UPat(Ops.BITCAST, dtypes.floats, name="val")), name='st'), lambda ctx,st,idx,val:
@@ -200,7 +195,7 @@ def do_dtype_decomps(sink:UOp, ctx:tuple[set[DType], Renderer]) -> UOp:
      to = dtypes.int if fr == dtypes.long else dtypes.half if not _should_emulate(dtypes.half) and fr in dtypes.fp8s else dtypes.float
      if DEBUG >= 2: print(f"emulating {fr} as {to}")
      pm = pm_float_decomp if fr in dtypes.floats else pm_long_decomp
-      sink = graph_rewrite(sink, pm, name=f"decomp {fr} -> {to}", ctx={} if pm is pm_long_decomp else (fr, to), bottom_up=True)
+      sink = graph_rewrite(sink, pm, name=f"decomp {fr} -> {to}", ctx=(fr, to), bottom_up=True)
  ctx[0].clear()
  return sink

@@ -47,17 +47,17 @@ def fast_idiv(ren: Renderer, x: UOp, d: int, dont_cast=False) -> UOp|None:

 def threefry2x32(x: UOp, key: UOp):
  # split x and key from uint64 to two uint32
-  x0, x1 = x.cast(dtypes.uint32), (x >> 32).cast(dtypes.uint32)
-  key0, key1 = key.cast(dtypes.uint32), (key >> 32).cast(dtypes.uint32)
+  x0, x1 = (x & 0xffffffff).cast(dtypes.uint32), ((x // 2**32) & 0xffffffff).cast(dtypes.uint32)
+  key0, key1 = (key & 0xffffffff).cast(dtypes.uint32), ((key // 2**32) & 0xffffffff).cast(dtypes.uint32)

  rotations = [[13, 15, 26, 6], [17, 29, 16, 24]]
  ks = [key1, key0 ^ key1 ^ 0x1BD11BDA, key0]
  xr:list[UOp] = [x0 + ks[-1], x1 + ks[0]]
  for i in range(5):
-    for r in rotations[i % 2]: xr[0], xr[1] = (x0 := xr[0] + xr[1]), x0 ^ ((xr[1] << r) + (xr[1] >> (32 - r)))
+    for r in rotations[i % 2]: xr[0], xr[1] = (x0 := xr[0] + xr[1]), x0 ^ ((xr[1] * 2**r) + (xr[1] // 2**(32 - r)))
    xr = [(xr[0] + ks[i % 3]), (xr[1] + ks[(i + 1) % 3] + i + 1)]

-  return (xr[1].cast(dtypes.uint64) << 32) | xr[0].cast(dtypes.uint64)
+  return xr[1].cast(dtypes.uint64) * 2**32 | xr[0].cast(dtypes.uint64)

 # ***** decomposition patterns *****

@@ -257,12 +257,12 @@ def xlog2(d:UOp) -> UOp:
 def xpow(base:UOp, exponent:UOp) -> UOp:
  # start with b ** e = exp2(e * log2(b))
  ret = (base < 0).where(-base, base).log2().mul(exponent).exp2()
-  # negative base: nan for non-integer exponent, negate for odd integer exponent. -inf is never nan, it stays |base| ** exponent
+  # negative base: nan for non-integer exponent, negate for odd integer exponent
  non_int = exponent != exponent.cast(dtypes.int32).cast(exponent.dtype)
  is_odd = (exponent < 0).where(-exponent, exponent).cast(dtypes.int32).mod(2).cast(dtypes.bool)
-  neg_base = non_int.where(base.ne(-math.inf).where(ret.const_like(math.nan), ret), is_odd.where(-ret, ret))
-  # x ** 0 = 1, including 0 ** 0 and inf ** 0
-  return exponent.eq(0).where(ret.const_like(1), (base < 0).where(neg_base, ret))
+  neg_base = non_int.where(ret.const_like(math.nan), is_odd.where(-ret, ret))
+  # fix 0 ** 0 = 1
+  return (base.eq(0) & exponent.eq(0)).where(ret.const_like(1), (base < 0).where(neg_base, ret))

@functools.cache
 def get_transcendental_patterns(ops:tuple[Ops, ...], force_transcendental:bool) -> PatternMatcher:
@@ -57,7 +57,7 @@ def add_gpudims(ctx:Renderer, s:UOp):

  # get the idxs
  ki: KernelInfo = s.arg
-  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int, param=True).cast(dtypes.weakint)]
+  if ctx.has_threads: idxs = [UOp.variable("core_id", 0, int(global_shape[0])-1, dtypes.int).cast(dtypes.weakint)]
  elif ki.dont_use_locals:
    assert not local_dims, "can't use locals if there's no local dims"
    idxs = get_grouped_dims("idx", global_shape, ctx.global_max, reverse=True)
@@ -89,7 +89,7 @@ def add_gpudims(ctx:Renderer, s:UOp):

 pm_device_to_var = PatternMatcher([
  # the DEVICE axis is not a program axis, it's bound per device at launch. lower it to the _device_num variable (like SPECIAL for devices)
-  (UPat(Ops.RANGE, name="r"), lambda r: UOp.variable("_device_num", 0, r.vmax, dtype=r.dtype, param=True) if r.arg[-1] is AxisType.DEVICE else None),
+  (UPat(Ops.RANGE, name="r"), lambda r: UOp.variable("_device_num", 0, r.vmax, dtype=r.dtype) if r.arg[-1] is AxisType.DEVICE else None),
  # ENDs that closed a DEVICE range no longer close it
  (UPat(Ops.END, name="e"), lambda e: e.replace(src=(e.src[0],)+tuple(s for s in e.src[1:] if s.op is not Ops.PARAM))
   if any(s.op is Ops.PARAM and s.arg.name == '_device_num' for s in e.src[1:]) else None),
@@ -1,8 +1,8 @@
 import itertools, functools
 from collections import defaultdict
 from tinygrad.dtype import dtypes, AddrSpace, Invalid, DType
-from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp, shape_to_shape_arg, graph_rewrite
-from tinygrad.uop.symbolic import uop_given_valid, parse_valid, invalid_gate, sym
+from tinygrad.uop.ops import UOp, Ops, PatternMatcher, UPat, GroupOp, shape_to_shape_arg
+from tinygrad.uop.symbolic import uop_given_valid, parse_valid, invalid_gate
 from tinygrad.helpers import getenv, IMAGE, OSX, ceildiv, is_image_shape
 from tinygrad.renderer import Renderer

@@ -26,15 +26,12 @@ def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
    # check if idx is out of bound when X is on the wrong side of the bound: X in [c+1, vmax] or [vmin, c-1]
    lo, hi = (c + 1, X.vmax) if is_upper_bound else (X.vmin, c - 1)
    if lo <= hi:
-      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype, param=True)
-      subs = [{X: fake}]
-      # idx may not have X itself, so also substitute a term of X: v -> fake - (X - v)
-      terms = list(X.split_uop(Ops.ADD))
-      v = next((u for u in terms if u.op in GroupOp.Irreducible and u.op is not Ops.CONST), None)
-      if v is not None and (rest:=[u for u in terms if u is not v]): subs.append({v: fake - UOp.usum(*rest)})
-      if any((testidx:=graph_rewrite(coord.substitute(sub), sym)).vmin >= b or testidx.vmax < 0
-             for sub in subs for coord,b in zip(idx.src, (width, height))):
-        drop_stmt.append(stmt)
+      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype)
+      for coord,b in zip(idx.src, (width, height)):
+        rw = coord.substitute({X:fake}).simplify()
+        if rw.vmin >= b or rw.vmax < 0:
+          drop_stmt.append(stmt)
+          break
  return drop_stmt

 def simplify_valid_load(buf:UOp, start_idx:UOp, valid:UOp) -> UOp|None:
@@ -125,7 +125,7 @@ def regalloc_rewrite(ctx:LinearScanRegallocContext, x:UOp):
  # alloc/dealloc stack
  if ctx.stack_size > 0:
    sp = ctx.ren.stack_pointer()
-    offset = UOp.const(ctx.stack_size, sp.dtype)
+    offset = UOp(Ops.CONST, sp.dtype, arg=ctx.stack_size)
    if i == 0: before = [ctx.ren.isel_matcher.rewrite(UOp(Ops.SUB, src=(sp, offset), tag=sp.tag))] + before
    elif i == len(ctx.uops) - 2: before += [ctx.ren.isel_matcher.rewrite(UOp(Ops.ADD, src=(sp, offset), tag=sp.tag))]

@@ -332,9 +332,9 @@ class Scheduler:
  @property
  def group_for_reduces(self) -> int: return len(self.axes_of(AxisType.GROUP_REDUCE))

-def args_from_ast(ast:UOp, dname:str) -> tuple[list[Buffer], dict[str, int]]:
+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 and x.arg.slot >= 0], key=lambda x: x.arg.slot)
-  return [Buffer(dname, x.max_numel(), x.dtype) for x in glbls], {k.expr:int(k.vmax+k.vmin)//2 for k in ast.variables()}
+  return [Buffer(dname, x.max_numel(), x.dtype) for x in glbls]

 def apply_opts(ast:UOp, ren:Renderer, beam:int=0) -> UOp:
  if ast.tag is not None: return ast
@@ -344,10 +344,10 @@ def apply_opts(ast:UOp, ren:Renderer, beam:int=0) -> UOp:
    for opt in ast.arg.opts_to_apply: k.apply_opt(opt)
  elif beam >= 1:
    from tinygrad.codegen.opt.search import beam_search
-    rawbufs, var_vals = args_from_ast(ast, ren.target.device)
+    rawbufs = bufs_from_ast(ast, ren.target.device)
    # beam search may open devices
    with Context(ALLOW_DEVICE_USAGE=1):
-      k = beam_search(k, rawbufs, var_vals, beam, bool(getenv("BEAM_ESTIMATE", 1)))
+      k = beam_search(k, rawbufs, beam, bool(getenv("BEAM_ESTIMATE", 1)))
  elif not NOOPT and (ast.arg is None or ast.arg.applied_opts == ()):
    from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
    # NOTE: hand_coded_optimizations doesn't support multiblock opts yet
@@ -1,6 +1,6 @@
 import math, time, multiprocessing, traceback, signal, atexit
 from dataclasses import replace
-from tinygrad.uop.ops import sym_infer, AxisType, UOp, Ops
+from tinygrad.uop.ops import sym_infer, AxisType, UOp
 from tinygrad.uop.render import pyrender
 from tinygrad.device import Device, Buffer
 from tinygrad.helpers import prod, flatten, DEBUG, CACHELEVEL, diskcache_get, diskcache_put, getenv, Context, colored, time_to_str
@@ -62,8 +62,7 @@ def _try_compile(x:tuple[int,Scheduler]) -> tuple[int, tuple[UOp, float]|None]:
  ret = None
  try:
    st = time.perf_counter()
-    ast, dev = x[1].copy().get_optimized_ast(name_override="test"), x[1].ren.target.device
-    prg = to_program(ast.substitute({p: p.replace(arg=replace(p.arg, device=dev)) for p in ast.toposort() if p.op is Ops.PARAM}), x[1].ren)
+    prg = to_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].ren)
    et = time.perf_counter() - st
    uops = prg.src[1].src
    if len(uops) >= (uops_max:=getenv("BEAM_UOPS_MAX", 3000)) > 0:
@@ -112,7 +111,7 @@ def get_kernel_actions(s:Scheduler, include_0=True, max_up:int|None=None) -> dic
  return acted

 beam_pool, BEAM_DEBUG = None, getenv("BEAM_DEBUG")
-def beam_search(s:Scheduler, rawbufs:list[Buffer], var_vals:dict[str,int], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
+def beam_search(s:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
  global beam_pool
  key = {"ast": s.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": s.ren.target.device, "suffix": s.ren.suffix}
  if not disable_cache and CACHELEVEL >= 1 and (val:=diskcache_get("beam_search", key)) is not None:
@@ -137,6 +136,7 @@ def beam_search(s:Scheduler, rawbufs:list[Buffer], var_vals:dict[str,int], amt:i

  try:
    rawbufs = _ensure_buffer_alloc(rawbufs)
+    var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in s.ast.variables()}
    exiting, st = False, time.perf_counter()
    dev = Device[s.ren.target.device]
    while not exiting:
@@ -1,7 +1,6 @@
 import math, functools
 from dataclasses import dataclass
 from tinygrad.dtype import DType, dtypes
-from tinygrad.uop.ops import PatternMatcher, UOp, UPat, Ops

@dataclass(frozen=True)
 class TensorCore: # D = A * B + C, A is (M x K), B is (K x N), C and D are (M x N)
@@ -136,42 +135,6 @@ amd_cdna4 = amd_cdna_1616128 + amd_cdna_161632 + amd_cdna_161616

 def get_amd(arch): return {"gfx942": amd_cdna3, "gfx950": amd_cdna4, "gfx1200": amd_rdna4, "gfx1201": amd_rdna4}.get(arch, amd_rdna3)

-pm_validate_wmma_rdna3 = PatternMatcher([
-  (UPat(Ops.WMMA, name="x", dtype=dtypes.int32), lambda x: x.replace(
-    src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
-    if x.src[0].dtype == dtypes.int8 and x.src[0].max_numel() == 16 else None),
-  (UPat(Ops.WMMA, name="x", dtype=dtypes.half), lambda x: UOp(Ops.STACK, src=tuple(x.replace(
-      src=(x.src[0], x.src[1], UOp(Ops.STACK, src=tuple(x.src[2].index(UOp.const(j//2, dtypes.int16))
-      if j%2 == 0 else UOp.const(0.0, x.src[2].dtype)
-      for j in range(x.max_numel()*2)))),
-      arg=(*x.arg[:4], None)).index(UOp.const(i*2, dtypes.int16))
-      for i in range(x.max_numel()))) if x.max_numel() == 8 else None),
-  (UPat(Ops.WMMA, name="x"), lambda x: x.replace(
-    src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-    if x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 16 else None),
-])
-
-pm_validate_wmma_rdna4 = PatternMatcher([
-  (UPat(Ops.WMMA, name="x", dtype=dtypes.bfloat16), lambda x: x.replace(
-    dtype=dtypes.uint16,
-    src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2].bitcast(dtypes.uint16)))
-      .bitcast(dtypes.bfloat16) if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None),
-  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-    if x.max_numel() == 8 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 8 else None)
-])
-
-pm_validate_wmma_cdna = PatternMatcher([
-  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint32), x.src[1].bitcast(dtypes.uint32), x.src[2]))
-    if x.arg[0][2] == 128 and x.src[0].dtype.itemsize <= 8 else None),
-  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint16), x.src[1].bitcast(dtypes.uint16), x.src[2]))
-    if x.max_numel() == 4 and x.src[0].dtype == dtypes.bfloat16 and x.src[0].max_numel() == 4 else None),
-  (UPat(Ops.WMMA, name="x", dtype=dtypes.float),
-    lambda x: x.replace(src=(x.src[0].bitcast(dtypes.uint64), x.src[1].bitcast(dtypes.uint64), x.src[2]))
-    if x.max_numel() == 4 and x.src[0].dtype in dtypes.fp8_ocp and x.src[0].max_numel() == 8 else None),
-])
 # ***** Apple Metal *****

 metal = [TensorCore(dims=(8,8,8), threads=32, elements_per_thread=(2,2,2), dtype_in=di, dtype_out=do,
@@ -137,7 +137,7 @@ def reduce_collapse(red:UOp, u:UOp, pm:PatternMatcher=pm_reduce_collapse) -> UOp
    for u in included:
      for s in u.src:
        if s in included or s in replaces or s.op in {Ops.CONST, Ops.PARAM, Ops.BUFFER}: continue
-        replaces[s] = UOp.variable(f'in{len(replaces)}', s.vmin, s.vmax, s.dtype, param=True)
+        replaces[s] = UOp.variable(f'in{len(replaces)}', s.vmin, s.vmax, s.dtype)
    collapse_fxn = u.substitute(replaces).reduce(r, arg=Ops.ADD)
    sink = graph_rewrite(collapse_fxn, pm, name="reduce_collapse")
    if not no_range(sink): return None
@@ -336,8 +336,6 @@ class Compiled:
  pm_lower:Any = None
  pm_bufferize:Any = None

-  has_copy_queue:bool = True
-
  def __init__(self, device:str, allocator:Allocator, renderers:list[type[Renderer]], runtime:type[Program[Self]]|None, graph=None, arch=None):
    from tinygrad.renderer import Renderer
    self.device, self.allocator, self.runtime_t, self.graph, self.renderers = device, allocator, runtime, graph, renderers or [Renderer]
@@ -27,7 +27,7 @@ class InvalidType:
  def __new__(cls):
    if cls._instance is None: cls._instance = object.__new__(cls)
    return cls._instance
-  def __eq__(self, other): return self is other if isinstance(other, InvalidType) else NotImplemented  # foreign types get the reflected eq
+  def __eq__(self, other): return self is other
  def __hash__(self): return id(self)
  def __repr__(self): return "Invalid"
  def __reduce__(self): return (InvalidType, ())  # unpickle returns the singleton
@@ -80,7 +80,7 @@ class DType(metaclass=DTypeMetaClass):
    # NOTE: float('nan') != float('nan'), so we canonicalize here
    if isinstance(val, float) and math.isnan(val): val = math.nan
    # int is the default. wrap floats in ConstFloat to distinguish -0.0 from 0.0 in cache
-    return ConstFloat(truncate.get(self, float)(float(val))) if dtypes.is_float(self) else bool(val) if dtypes.is_bool(self) else int(val)
+    return ConstFloat(float(val)) if dtypes.is_float(self) else bool(val) if dtypes.is_bool(self) else int(val)


 class DTypes:
@@ -293,12 +293,6 @@ truncate: dict[DType, Callable] = {dtypes.bool: bool,
  **{getattr(dtypes, n): (lambda x, c=getattr(ctypes, f'c_{n}'): c(x).value)
     for n in ('float', 'double', 'int8', 'int16', 'int32', 'int64', 'uint8', 'uint16', 'uint32', 'uint64')}}

-def bitcast(x, in_dtype:DType, out_dtype:DType):
-  assert in_dtype.itemsize == out_dtype.itemsize, "bitcast itemsize mismatch"
-  packed = struct.pack(storage_fmt_for_dtype(in_dtype), to_storage_scalar(x, in_dtype))
-  out_val = struct.unpack(storage_fmt_for_dtype(out_dtype), packed)[0]
-  return from_storage_scalar(out_val, out_dtype)
-
 # numpy and torch dtype interop

 def _to_np_dtype(dtype:DType) -> type|None:
@@ -4,7 +4,7 @@ from tinygrad.tensor import Tensor, all_tensors
 from tinygrad.helpers import flatten, merge_dicts, DEBUG, Context, BEAM, getenv, JIT, JIT_BATCH_SIZE, dedup, pluralize, VIZ, disable_gc
 from tinygrad.device import Buffer, Compiled, Device, MultiBuffer, DepsTracker
 from tinygrad.dtype import DType
-from tinygrad.uop.ops import UOp, PatternMatcher, Variable, sym_infer, Ops, buffers, rewrite_group, graph_rewrite
+from tinygrad.uop.ops import UOp, PatternMatcher, Variable, sym_infer, Ops, buffers, track_rewrites, graph_rewrite
 from tinygrad.renderer import Estimates
 from tinygrad.engine.realize import capturing, compile_linear, link_linear, run_linear, graph_cache, estimate_uop, get_runtime
 from tinygrad.engine.realize import unwrap_multi, resolve_params, get_call_arg_uops, get_call_outs_ins
@@ -44,7 +44,9 @@ def graph_split_rewrite(linear:UOp, max_batch_size:int=0) -> UOp:
    current_batch, current_batch_devs = [], []

  for si in linear.src:
-    devs = dedup([Device[x] for b in si.src[1:] if not b.is_bound_var for x in (b.device if isinstance(b.device, tuple) else (b.device,))])
+    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

    can_graph = graph_t is not None and graph_t.supports_uop(devs, si)
@@ -62,14 +64,14 @@ def _copy_input(u:UOp) -> UOp:
  run_linear(UOp(Ops.LINEAR, src=(u.copy_to_device(u.device).call(new:=UOp.new_buffer(u.device, u.max_numel(), u.dtype), u),)))
  return new

-@rewrite_group(lambda linear,held_bufs,input_uops,ret=(): f"JIT {pluralize('call', len(linear.src))}")
+@track_rewrites(lambda linear,held_bufs,input_uops,ret=(): f"JIT {pluralize('call', len(linear.src))}")
 def jit_lower(linear:UOp, held_bufs:set[UOp], input_uops:list[UOp]) -> UOp:
  if VIZ: graph_rewrite(linear, PatternMatcher([]), name="View captured linear")

  # parametrize input buffers: map each input buffer UOp to a PARAM with the correct slot index
  linear = linear.substitute({u: UOp.param(i, u.dtype, u.shape, u.device) for i,u in enumerate(input_uops)}, walk=True)
  linear = memory_plan_rewrite(linear, held_bufs)
-  linear = compile_linear(linear, beam=getenv("JITBEAM", BEAM.value))
+  linear = compile_linear(linear, beam=getenv("JITBEAM", BEAM.value), jit=True)
  if JIT < 2: linear = graph_split_rewrite(linear, max_batch_size=JIT_BATCH_SIZE.value)
  if VIZ: graph_rewrite(linear, PatternMatcher([]), name="View graphed linear")
  return linear
@@ -167,7 +169,7 @@ class CapturedJit(Generic[ReturnType]):
  expected_input_info: list[tuple[UOp, tuple[Variable, ...], DType, str]]  # (view, variables, dtype, device) per input

  @functools.cached_property
-  def linear(self) -> UOp: return link_linear(self._linear)
+  def linear(self) -> UOp: return link_linear(self._linear, jit=True)

  def __reduce__(self): return self.__class__, (self.ret, self._linear, self.expected_names, self.expected_input_info)

@@ -178,7 +180,7 @@ 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 |= {b for k in set(outs) - set(ins) if (b:=u if (cv:=(u:=arg_uops[k]).contiguous_view()) is None else cv[0]).op is Ops.BUFFER}
+      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:
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
geohot	1ec33e2ef6	wip: test change, do not merge	2026-08-01 10:50:45 -07:00
geohot	fd62ad3034	Merge pull request 'szdiff: make workflow work on both github and gitea (supersedes #8 )' (#9 ) from szdiff_portable into master	2026-08-01 10:49:44 -07:00
geohot	60996454a3	szdiff: make comment steps work on both github and gitea - checkbranch: fetch master from the base repo clone_url instead of the PR head remote (origin), which on GitHub is the fork for fork PRs - replace the curl comment snippets with sticky_comment.py: stdlib-only, works with the REST API on both platforms, and keeps the skip_unchanged/ignore_empty behavior of the old action	2026-08-01 10:44:41 -07:00
geohot	a54bb3b795	fix szdiff workflow for Gitea: use local master and REST API comments Two bugs prevented the 'Core Library Line Count' bot from working: 1. checkbranch fetched master from github.com/tinygrad/tinygrad instead of the local Gitea instance. Since the Gitea mirror lags behind GitHub, every PR appeared 'behind' and the szdiff job was always skipped. 2. marocchino/sticky-pull-request-comment@v3 uses GraphQL to find existing comments, but Gitea has no GraphQL API (returns 404). Both the szdiff and rebase comment steps failed with '404 page not found'. Fix 1: fetch origin/master (the local Gitea repo) instead of adding a remote to github.com. Fix 2: replace the third-party action with curl calls to the Gitea REST API (GET/POST/PATCH /repos/.../issues/.../comments), which fully supports the sticky-comment pattern (find existing, update or create).	2026-08-01 10:34:37 -07:00