kimi linear working

* hevc tests

* x

.github/actions/setup-tinygrad/action.yml

@@ -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'

.github/workflows/platform.yml

@@ -179,3 +179,35 @@ jobs:
       - name: Run test_tiny
         shell: bash
         run: python -m pytest -n=auto test/test_tiny.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

.github/workflows/test.yml

@@ -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
@@ -86,30 +86,9 @@ jobs:
     - name: Custom tests
       run: DEV=CPU:LLVM GPUS=4 TINY_BACKEND=1 python3 -m pytest -nauto extra/torch_backend/test.py extra/torch_backend/test_inplace.py extra/torch_backend/test_multigpu.py extra/torch_backend/test_kernel_fusion.py --durations=20
 
-  torchbackendtrain:
-    name: Torch Backend Training
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
-    timeout-minutes: 15
-    steps:
-    - name: Checkout Code
-      uses: actions/checkout@v6
-    - name: Setup Environment
-      uses: ./.github/actions/setup-tinygrad
-      with:
-        key: torch-backend-pillow-torchvision-et-pt
-        deps: testing_unit
-        pydeps: "pillow torchvision expecttest"
-        llvm: 'true'
-    - name: Install ninja
-      run: |
-        sudo apt update || true
-        sudo apt install -y --no-install-recommends ninja-build
-    - name: Test beautiful_mnist in torch with TINY_BACKEND
-      run: STEPS=20 DEV=CPU TARGET_EVAL_ACC_PCT=90.0 MAX_BUFFER_SIZE=0 TINY_BACKEND=1 python3 examples/other_mnist/beautiful_mnist_torch.py
-
   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
@@ -147,7 +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:
@@ -178,7 +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:
@@ -212,7 +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:
@@ -249,7 +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
@@ -265,7 +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
@@ -281,7 +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
@@ -301,7 +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
@@ -330,7 +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:
@@ -349,7 +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
@@ -381,7 +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
@@ -408,7 +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
@@ -428,7 +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
@@ -456,7 +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
@@ -482,7 +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
@@ -528,7 +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
@@ -542,15 +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
@@ -586,7 +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 }}
@@ -609,7 +589,7 @@ jobs:
         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
+        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 test/opt/test_tensor_cores.py --durations=20
       - name: Run disk copy tests
         run: python -m pytest test/unit/test_disk_tensor.py -k test_copy_from_disk
       - name: Run TRANSCENDENTAL math
@@ -624,7 +604,7 @@ jobs:
         backend: [ptx, nv]
 
     name: Linux (${{ matrix.backend }})
-    runs-on: ${{ github.repository == 'tinygrad/tinygrad' && github.event_name == 'pull_request' && github.event.pull_request.author_association == 'COLLABORATOR' && 'namespace-profile-tinygrad' || 'ubuntu-24.04' }}
+    runs-on: *linux
     timeout-minutes: 20
     env:
       FORWARD_ONLY: 1
@@ -658,17 +638,10 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        dev:
-          - 'NULL:IR3:a630'
-          - 'NULL:QCOMCL:a630'
-          - 'NULL:NAK:sm_120'
-    name: Compile-only (DEV=${{ 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: ${{ contains(matrix.dev, 'a630') && '1' || '0' }}
     steps:
       - name: Checkout Code
         uses: actions/checkout@v6
@@ -677,14 +650,21 @@ jobs:
         with:
           key: compile-${{ 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

README.md

@@ -140,7 +140,7 @@ Documentation along with a quick start guide can be found on the [docs website](
 ```python
 from tinygrad import Tensor
 
-x = Tensor.eye(3).clone()  # clone to make it a buffer
+x = Tensor.eye(3)
 y = Tensor([[2.0,0,-2.0]])
 z = y.matmul(x).sum()
 z.backward()

docs/kimi_k3_mi350.md

@@ -0,0 +1,224 @@
+# Kimi K3 on 8× MI350X
+
+This branch targets text generation directly from the official `moonshotai/Kimi-K3` checkpoint at `/raid/weights/kimi-k3`. It intentionally ignores the vision tower and multimodal projector. The checkpoint remains in its official 96-shard format; the loader never converts, rewrites, or creates a second 1.56 TB copy.
+
+The checked TP8 layout consumes 196.78 GB (183.27 GiB) of text weights per GPU. The compressed MLA cache adds 28.99 GB (27 GiB) per GPU at the full 1,048,576-token context, leaving approximately 62.23 GB of each nominal 288 GB MI350X for execution buffers and allocator overhead. Start much smaller.
+
+## Resume the current optimization session
+
+Work on branch `kimi_slop`. It was cleanly rebased onto `origin/kimi_slop` commit `553bdf68e` on 2026-08-10. The retained K3 commits after that base are `1b3732a6e`, `c6ac4961d`, `1d8620471`, `224bac031`, `f53f0e7e7`, and `2b1b8c22a`; verify the current hashes with `git log` because a later rebase may rewrite them. Before starting any benchmark, check that the worktree is clean and that no model process remains:
+
+```sh
+git status --short --branch
+git log --oneline --decorate -10
+pgrep -af 'tinygrad.llm.cli|benchmark_kimi_k3' || true
+```
+
+The active acceptance target is **more than 100 tok/s decode, more than 200 tok/s prefill, and less than 180 seconds cold startup** on TP8/gfx950. None is currently met. The authoritative official-checkpoint baseline is 389.84 seconds startup, 38.65 tok/s prefill, and 6.25 tok/s decode. The 1.56 TB checkpoint has a measured 6.9 GB/s single-XFS-NVMe read ceiling, giving a roughly 227-second physical cold-read floor; meeting the startup target therefore also requires a faster storage path, not only loader code.
+
+Use the fake-weight, one-layer loop for development. Do not repeatedly load the official checkpoint while optimizing:
+
+```sh
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode attention --iterations 30
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode block --iterations 30
+PROFILE=1 DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode block --iterations 5
+```
+
+The clean retained baseline is about 0.630 ms per attention layer and 1.37 ms per complete block, with fake initialization taking about 0.9/2 seconds respectively after the rebase. Since K3 has 93 sequential blocks, a 100 tok/s projection requires at most approximately 0.108 ms per complete block. Only run another 96-shard official validation after a candidate produces a large whole-block gain, remains finite and deterministic, and passes a direct numerical comparison. Test one candidate at a time and remove failed experiments before moving on.
+
+The immediate bottleneck is launch and synchronization granularity: an official four-token decode profile contained 6,304 kernel events, while packed expert work was only a small fraction of total GPU time. Continue with whole-component or whole-block fusion/replay work, not isolated expert microkernels. The latest fake-loop A/B retested the previously rejected dual gate/up and weighted-down MFMA prototypes: 1.374 ms baseline versus 1.375 ms fused, so they were removed again. A fused whole-core KDA recurrence was also slower in the exact fake attention gate (0.665 versus 0.633 ms) and must not be restored unchanged.
+
+Preserve these invariants when official validation resumes: use `/raid/weights/kimi-k3` directly, keep all 96 shards byte-for-byte untouched, run only one model process, begin at context 128, verify all eight devices are `gfx950`, and preserve the first failure instead of retrying over it. The most recent preserved official failure from a rejected KDA experiment was the invalid sequence `[198, 163840, 163840, 163840]`; token 163840 is outside the valid vocabulary. The retained path before that experiment produced deterministic in-range replay.
+
+After a synthetic candidate passes, run correctness and performance in this order: NULL gfx950 compile coverage, focused tests with `-n12` where supported, TP8 fake numerical comparison, official context-128 deterministic tokens, load/prefill/decode timing, and then context admission at 4K, 32K, 131K, and 262K. Run `python -m mypy tinygrad/` and `python -m ruff check .` when those tools are installed. Read `tinygrad/viz/README.md` before inspecting rewrite or device profiles.
+
+## Before renting the machine
+
+- Keep the existing 96 shards in `/raid/weights/kimi-k3`; no additional model-sized free space is required. Leave ordinary headroom for logs and temporary files.
+- The host should have roughly 3 TB RAM, in line with AMD's MI350X platform guidance. The loader itself is streaming and must not need checkpoint-sized RAM.
+- Use a recent kernel/ROCm stack supported by the host vendor, although tinygrad uses its own AMD userspace driver when `DEV=AMD`.
+- Clone this exact commit/branch and keep the official checkpoint directory separate from the repository.
+
+Validate the existing directory without modifying it:
+
+```sh
+python examples/kimi_k3_prepare.py /raid/weights/kimi-k3 --context 4096
+```
+
+For a metadata-only preflight, place the official `config.json` and `model.safetensors.index.json` in a directory and run:
+
+```sh
+python examples/kimi_k3_prepare.py /raid/weights/kimi-k3 --metadata-only
+```
+
+## Hardware admission checks
+
+Do these before loading weights. Stop if any device is missing or reports a different architecture.
+
+```sh
+lspci -d 1002:75a0
+amd-smi list
+DEV=AMD DEBUG=2 python - <<'PY'
+from tinygrad import Device
+for i in range(8):
+  dev = Device[f"AMD:{i}"]
+  print(i, dev.arch)
+PY
+```
+
+Expected architecture: `gfx950` on all eight devices. Then run the small TP8 graph tests:
+
+```sh
+python -m pytest test/unit/test_llm_k3.py test/null/test_kimi_k3.py -q -n12
+DEV=NULL:HIP:gfx950 NULL_ALLOW_COPYOUT=1 python -m pytest \
+  test/unit/test_llm_k3.py::TestKimiK3::test_chunked_recurrent_generate -q -n1
+DEV=AMD python examples/kimi_k3_smoke.py --devices 8
+```
+
+The last two commands are deliberately small. They compile CDNA4 kernels and then exercise the complete TP8 topology without loading the checkpoint.
+
+For performance iteration, use the exact-width fake-weight harness before another official load:
+
+```sh
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode attention --iterations 20
+DEV=AMD python extra/benchmark_kimi_k3_fake.py --mode block --iterations 20
+```
+
+It retains K3's 7,168-wide residual stream, 12,288-wide KDA state, 96 heads, 128×128 recurrent matrices, TP8 layouts, top-k 16 routing, packed MXFP4 expert shapes, collectives, and decode JIT, but uses one layer and 16 fake experts. Fake attention weights initialize in about 0.9 seconds and the full block in about 3 seconds. The retained path measured 0.630 ms per fake attention layer and 1.367 ms per complete fake block, projecting about 7.87 tok/s across 93 identical blocks versus 6.25 tok/s for the official heterogeneous model. Treat this as a candidate admission benchmark, not a correctness substitute for official weights.
+
+## First official load
+
+Start at a short context so cache allocation and compilation are bounded. The loader reads disk-backed safetensors, TP-shards every destination before realizing it, and drops each source shard/projection immediately afterward.
+
+```sh
+/usr/bin/time -v env DEV=AMD DEBUG=1 python -m tinygrad.llm.cli \
+  --model /raid/weights/kimi-k3 --devices 8 --max_context 128 </dev/null 2>&1 | tee kimi-k3-load.log
+```
+
+Watch host RAM, swap, HBM, temperatures, and XGMI traffic from a second terminal. Do not start with a one-million-token cache. If loading fails, preserve the first exception and the last loader progress line; do not retry with a larger host-side cache.
+
+## Correctness and performance sequence
+
+1. Load with context 128 and generate one token.
+2. Repeat a fixed prompt twice and confirm token-for-token deterministic greedy output.
+3. Compare the first several greedy tokens against the official Transformers implementation at temperature zero.
+4. Benchmark decode only after two warm-up tokens.
+5. Benchmark prefill at 128, 512, 2K, and 8K tokens. Increase context only while HBM and compile time remain healthy.
+6. Use `VIZ=1` plus `python -m tinygrad.viz.cli` to inspect kernels; use `VIZ=2` only for short SQTT captures because it adds overhead.
+
+Example decode benchmark:
+
+```sh
+DEV=AMD DEBUG=1 python -m tinygrad.llm.cli --model /raid/weights/kimi-k3 \
+  --devices 8 --max_context 4096 --warmup --benchmark 20
+```
+
+## MI350X validation results (2026-08-10)
+
+The official directory was audited in place: 96 shards, 497,220 indexed tensors, 497,052 language tensors, and 1,560,860,324,864 total bytes. All eight devices reported `gfx950`. No checkpoint file was converted, copied, or modified, and every model run used a single process. The actual text tower is 1,559,965,606,912 bytes; its checked TP8 layout is 196,784,397,312 bytes per GPU.
+
+The preserved first full-checkpoint error was an `A_log` shape mismatch, `(128,) -> (96, 1)`. K3 stores one decay value per 128-wide KDA channel, not one per head. The loader now keeps this field replicated and applies the official channel-wise broadcast. A numerical unit test covers the distinction from the older head-wise Kimi Linear behavior.
+
+Load speed was fixed before generation. The original loader opened thousands of individual expert tensors and independently realized eight strided TP slices. The MI350 path now does the following without changing the checkpoint:
+
+- parses safetensor headers selectively, constructing disk-backed tensors only for the 2,460 non-expert entries consumed by that pass instead of materializing metadata objects for every expert entry twice;
+- copies contiguous axis-zero shards and replicas directly into their final device buffers;
+- reads a replicated tensor once and fans it out over XGMI instead of issuing eight identical direct reads (14.31 GB less RAID traffic);
+- stages an inner-axis tensor once and schedules all eight TP slices together;
+- reads each layer's contiguous 15.72 GB expert region once, reorders its lexicographically stored expert records on GPU 0, and realizes all six packed/scale destinations together;
+- retains only final MultiBuffer identities, drops the reorder graph, and flushes the 15.72 GB staging allocation before the next layer.
+
+One real expert layer leaves exactly 1,965,293,568 bytes resident on each GPU and zero bytes in the GPU-0 allocator cache. Complete context-128 loads measured 527.20 seconds before the final staging cleanup and 490.05/489.59 seconds afterward. Peak host RSS for the unprofiled correctness run was 2.11 GiB with zero swap. RAID variability produced later loads from 489.06 to 532.85 seconds.
+
+The selective-metadata and bounded-GC pass reduced non-expert loading from 125.77 to 57.77 seconds. A subsequent full official context-128 load completed in 411.49 seconds, 78.10 seconds (16.0%) faster than the 489.59-second baseline. It read the 96 shards in place with 1,049,688 KiB peak host RSS and zero swap; no weight payload was converted, copied, or modified. Direct-I/O probes measured approximately 6.9 GB/s aggregate for both one and eight concurrent 1 GiB reads. At that rate the 1.56 TB checkpoint has a roughly 227-second cold-read lower bound, so this RAID cannot meet a true cold sub-three-minute startup regardless of loader overhead.
+
+Expert staging graphs are acyclic and are released by reference counting after each layer, so the loader now suppresses unnecessary cyclic-collector scans only around that loop and restores its prior state on every exit. A quiet context-128 load then completed in 391.54 seconds, 30.14 seconds (7.1%) faster than the immediately preceding 421.68-second run, with 1.04 GiB peak RSS and zero swap, although storage variability contributes to run-to-run timing. The host used for these measurements actually mounts `/raid` from one 3.5 TB XFS NVMe, not a multi-drive RAID; shard 28 has 218 extents and live reads fell to roughly 160 MB/s there. This storage layout, plus the physical checkpoint size, remains the limiting cold-start constraint. The weights were not defragmented, copied, or modified.
+
+The fixed XTML prompt `Reply with exactly: OK` encodes to 93 tokens. After excluding the cold JIT capture from replay comparison, two greedy runs produced the identical eight-token sequence:
+
+```text
+[9545, 59991, 10580, 14404, 9545, 59991, 9545, 59991]
+```
+
+At context 128, steady prefill was 14.32 seconds (6.49 tok/s) and eight-token decode was 2.27 seconds (3.53 tok/s, 283.3 ms/token). The same first tokens remained stable at every admitted context. These rates are much lower than the planning estimates below and should be treated as the current measured baseline.
+
+The retained gfx950 serving pass enables the validated wave64 recurrent prefill kernel with 128-token chunks, uses exact BF16 decode projections, combines the routed/shared final TP partials into one collective, and tiles four adjacent packed-expert outputs during multi-token execution. On the same 93-token prompt, two replay trials produced the identical sequence `[198, 92652, 220, 80225]`. Prefill replay measured 2.418--2.482 seconds (37.47--38.46 tok/s), and eight-token decode measured 1.294 seconds (6.18 tok/s, 161.81 ms/token). Peak RSS was 2.77 GiB with zero swap. The packed prefill tile changes floating-point reduction order: direct official-layer comparison against the original kernel had maximum differences of 0.015625 for gate and 0.0078125 for down, and the end-to-end greedy sequence was stable across replay.
+
+A subsequent gfx950 decode pass split the 7,168-wide replicated BF16 projections across eight waves per 16 output channels and used CDNA4 BF16 MFMA, with one FP32 LDS reduction at the end. It is enabled only for batch-one/token-one replicated projections whose dimensions satisfy the hardware tile; prefill, the FP32 router, and the output-sharded 12,288-wide KDA gate remain unchanged. The official retained path uses it for MLA q-a/kv-a and KDA f-a. Isolated TP8 measurements improved replicated 128/576-output projections by about 16--18%; applying it to the already output-sharded KDA gate was slower and was rejected. Random-shape comparison against the generic graph had maximum/mean absolute BF16 differences of 2.0/0.1114 because the split changes reduction order. Against a serial FP32 accumulation rounded once to BF16, the 7,168-to-1,536 kernel was bit-exact in the tested sample.
+
+The final official context-128 validation loaded in 389.84 seconds with 2.71 GiB peak RSS and zero swap. Two replay trials produced the identical four-token sequence `[198, 59675, 9817, 12519]`; prefill remained 2.406 seconds (38.65 tok/s), while eight-token decode improved to 1.280 seconds (6.25 tok/s, 160.00 ms/token). A one-wave MFMA variant and a full-wave fused decode recurrence were both rejected: the former delivered 6.02 tok/s, and the latter 6.179 tok/s, while both changed the greedy sequence without a useful speed gain.
+
+A final load-first experiment increased the disk-to-HBM io_uring queue depth from one to the 32 existing bounded 2 MiB staging buffers. On a direct 1 GiB read from fragmented shard 28 it measured 6.834 GB/s versus 6.832 GB/s for the original path, so the change was rejected. The subsequent unmodified official 96-shard load completed in 389.48 seconds, confirming both the prior result and the single-NVMe lower bound. Peak RSS was 2.75 GiB with zero swap.
+
+Two direct packed-expert MFMA prototypes were also rejected after that load. A fused gate/up kernel was about 29% faster in isolation at the TP8-local shape, and a routed-down kernel which combined projection, probability weighting, and route reduction measured 1.45 ms versus 2.42 ms in isolation. End-to-end, however, stable replay produced `[198, 2338, 2127, 148297]`, prefill measured 38.87 tok/s, and decode measured 6.263 tok/s. That is indistinguishable from the retained 38.65/6.25 tok/s path while changing floating-point reduction order, so neither kernel was retained.
+
+A whole-core KDA decode experiment fused convolution, Q/K normalization, channel decay, recurrence, RMS normalization, output gating, and four persistent state updates. Its raw kernel replayed in about 109 microseconds per local KDA layer and matched a one-step synthetic reference within `9.77e-4` output and `8.13e-4` state maximum error. The exact-width fake-layer gate caught that it was slower than the retained attention path (0.665 versus 0.633 ms/layer). The already-running official validation was stopped after its first invalid greedy sequence, `[198, 163840, 163840, 163840]`, where 163840 is outside the checkpoint's vocabulary. The kernel was rejected and removed.
+
+| Maximum context | Load | Short-prompt replay | Result |
+|---:|---:|---:|---|
+| 128 | 489.59s | 14.32s | stable 8-token replay |
+| 4,096 | 489.06s | 14.32s | stable replay, zero swap |
+| 32,768 | 532.85s | 14.33s | stable replay, zero swap |
+| 131,072 | 520.91s | 14.37s | stable first token, zero swap |
+| 262,144 | 497.34s | 14.41s | stable first token, zero swap |
+
+These are maximum-context/cache admission tests with the same 93-token prompt, not full-length 32K/131K/262K prefills. The full cache allocation path was exercised, but filling those contexts remains a separate long-running throughput test.
+
+Runtime profiling bracketed four steady decode tokens. It recorded 6,304 kernel events and about 474--478 ms of summed GPU work across the eight devices inside a roughly 1.5-second profiled wall interval. The packed `mxfp4_expert_linear_wave64` kernels accounted for only about 22.5 ms summed; the largest families were small 1,792-wide reductions. This identifies launch/synchronization granularity as the immediate MI350 bottleneck rather than packed-weight bandwidth. `JIT_BATCH_SIZE=64` produced the same original 3.53 tok/s as 32. A gfx950 fused MXFP8 QDQ experiment was bit-exact but slower on the real device (about 95 microseconds versus 57--64 microseconds), so it was rejected. Combining the routed and shared final TP partials removed one collective per routed decode layer and helped raise unprofiled decode to 6.18 tok/s, but the remaining sequential launch boundaries still dominate.
+
+The checkpoint's bundled Transformers code was used as the architectural reference for channel decay and tensor mapping. A full independent Transformers/vLLM token comparison was not run on this host because the required `compressed_tensors`/serving backend is not installed; deterministic tinygrad replay and the numerical KDA, loader-layout, NULL gfx950 compile, and real TP8 smoke tests are the completed correctness gates.
+
+## Known hardware-only gate
+
+The correctness path now consumes packed MXFP4 expert weights directly on gfx950 with a wave64 software-decode kernel, so it does not create selected-expert BF16 weight expansions. MXFP8 activation quantization is still emulated. tinygrad has gfx950/CDNA4 BF16 and FP8 matrix-core support, but this branch does not yet have a hardware-validated native MXFP4×MXFP8 expert GEMM. Expect the first run to be a correctness bring-up, not production throughput. Capture profiles on MI350X before changing the representation: native FP4 work cannot be validated faithfully on the available gfx1100 cards.
+
+Recurrent prefill is fused. The gfx950 wave-parallel kernel was compared directly with the portable graph at the official per-GPU shape through 128 tokens: maximum core/state differences remained below `8e-6`/`1e-6`, outputs were finite, and replay was about 2.7 ms versus about 8 ms for the portable kernel in the isolated test. Full K3 therefore uses 128-token recurrent chunks on gfx950. Chunk size remains part of the numerical configuration because different reduction orders can select different final greedy tokens.
+
+The following serving changes apply to the official K3 path: recurrent-state reset graph capture, direct AMD scalar readback without rebuilding a scheduler graph, materialized gate/up boundaries, separate greedy decode JITs, K3's uncorrected routed probability semantics, gfx950 KDA Q/K/V and exact BF16 partial projections, one combined routed/shared final collective, a gfx950 greedy output-head kernel, the wave64 packed-expert path, and the multi-token four-output packed tile. Software MXFP8 remains in use.
+
+After hardware admission on MI350X, profile before porting those kernels. The likely implementation order is:
+
+1. A native packed MXFP4×MXFP8 grouped expert GEMM using CDNA4 matrix instructions.
+2. A wave64/MFMA KDA Q/K/V decode projection.
+3. Combined routed/shared down-projection TP partials so each layer performs one XGMI all-reduce.
+4. A CDNA4 output-head matvec and router matvec if they remain visible in the profile.
+
+Every port needs a direct numerical comparison with the generic graph and an end-to-end greedy-token comparison before performance measurements. The wave64 packed-expert kernel has compile coverage through `NULL:HIP:gfx950`; numerical and performance validation still require real MI350X hardware. None of the remaining gfx11-only kernels should be enabled on gfx950 by changing only the architecture guard.
+
+## MI350X performance expectation
+
+Treat the first rental as bring-up, not a guaranteed throughput run. The loader reads every official expert tensor once into a transient GPU-0 staging buffer (at most one packed projection), then redistributes TP8 slices over the GPU fabric; it does not generate files or require checkpoint-sized host RAM. A reasonable planning range for the full text model on eight MI350X cards is 3–8 minutes to stream and TP-shard the 1.56 TB checkpoint, 150–400 tok/s for initial short/medium prefill, and 25–60 tok/s decode with the software packed-expert path. After a native CDNA4 MXFP4×MXFP8 grouped expert kernel, wave64/MFMA recurrent projections, and XGMI collective tuning, 500+ tok/s prefill and roughly 80–150 tok/s decode are plausible targets. These ranges are engineering estimates, not measurements.
+
+The nominal HBM bandwidth is not the main uncertainty: eight MI350X devices have enough aggregate bandwidth for K3's active weights. Utilization is limited by 93 sequential layers, small routed projections, and synchronization after TP input-sharded projections. Record actual HBM and XGMI counters before deciding whether the next port should target matrix instructions or collective count.
+
+The official checkpoint also contains MoonViT-V2 and multimodal projector weights. They are skipped by the text loader. Image input remains a separate implementation and validation task.
+
+## Local TP4 performance baseline
+
+The pre-rental benchmark uses the converted `Kimi-Linear-48B-A3B-Instruct-MXFP4-v2` checkpoint on four gfx1100 GPUs. It is a useful regression test for the KDA/MLA/MoE text path, not a projection of K3 throughput on MI350X.
+
+```sh
+DEV=AMD JIT_BATCH_SIZE=64 python extra/benchmark_kimi.py \
+  /raid/models/Kimi-Linear-48B-A3B-Instruct-MXFP4-v2 \
+  --devices 4 --max-context 128 --prompt-tokens 32 --decode-tokens 32 --chunk-size 32
+```
+
+Results from 2026-08-10:
+
+- load from RAID: 44.28s for the 29.27 GB checkpoint
+- first 32-token prefill includes roughly 10s of compilation/capture
+- steady fresh-prompt prefill replay: 0.118s, 270.20 tok/s
+- steady context-32 decode replay: 101.82 tok/s, 9.82 ms/token
+- peak host RSS: 729.9 MiB; swap was not used
+
+The load, prefill, and decode targets are all met in the bounded prompt-32 run. Decode improved from 23.03 tok/s to 101.82 tok/s. The retained greedy output was checked across 32 decode steps; rejected half-wave and unrounded recurrent reductions were faster but diverged and eventually collapsed to a repeated token.
+
+Fully warmed HTTP serving was also measured with `--max_context 4096`. Startup, including weight load, capture, and replay of both serving shapes, took 113.73s. After a two-turn cache test, the first aligned 64-token request reported 271 tok/s prefill and 101 tok/s decode over 64 generated tokens. A 99-token prompt reported 254 tok/s prefill and 99 tok/s decode; decode falls slightly as MLA context grows.
+
+Recurrent serving uses only the captured 32-token prefill graph and captured single-token graph. Warmup uses two consecutive chunks so both initial and nonzero-position prefill execution are ready before the socket opens. A prompt tail shorter than 32 tokens runs through the single-token graph instead of compiling a new static shape, so no request-time JIT capture is required. Exact extensions reuse recurrent and KV state—the live second turn logged `in: 18 + 15`—while divergent prompts reset both safely. Very short prompts can report less than 200 aggregate prefill tok/s because fixed reset and single-token costs dominate; aligned and medium/long prompts exercise the 200+ tok/s prefill path.
+
+Four 7900 XTX cards provide 96 GB aggregate VRAM and about 3.84 TB/s aggregate physical memory bandwidth. Their nominal aggregate vector FP16 rate is about 245.6 TFLOP/s, or about 492 TFLOP/s through matrix instructions. Kimi Linear activates roughly 3.107B parameters per token; a simple active-weight accounting gives approximately 4.05 GB/token and an optimistic bandwidth-only ceiling near 948 tok/s. The measured decode rate is much lower because this MoE decode workload is a collection of small matrix-vector operations plus PCIe collectives, not one ideal streaming kernel.
+
+The generic loader currently rereads logical TP shards and accounts for roughly 227 GB of disk traffic for a TP4 load. RAID bandwidth hides that inefficiency locally, but a direct one-pass shard loader remains worthwhile before slow remote storage is used. It was not retained here because the attempted direct-shard graph exposed an unresolved scheduler/renderer edge; correctness and bounded memory take priority over avoiding the redundant reads.
+
+Different chunk sizes can choose a different final token because their matrix kernels use different floating-point reduction orders. Each measured shape was repeatable between cold and captured execution. For official K3 validation, compare logits/tokens against the reference at one fixed chunk size and greedy settings rather than requiring bitwise agreement between performance shapes.

examples/convert_kimi_linear.py

@@ -0,0 +1,9 @@
+import argparse
+from tinygrad.llm.kimi import convert_kimi
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser(description="Convert official Kimi-Linear-48B-A3B BF16 weights to tinygrad MXFP4/BF16")
+  parser.add_argument("source", help="downloaded moonshotai/Kimi-Linear-48B-A3B-Instruct directory")
+  parser.add_argument("output", help="output directory")
+  args = parser.parse_args()
+  convert_kimi(args.source, args.output)

examples/kimi_k3_prepare.py

@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""Cheap preflight for an official moonshotai/Kimi-K3 checkout. Does not load model weights."""
+import argparse, json, pathlib, shutil
+from tinygrad.llm.kimi_k3 import KIMI_K3_TP8_BYTES_PER_GPU, audit_kimi_k3_checkpoint
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("model_dir", type=pathlib.Path)
+  parser.add_argument("--metadata-only", action="store_true", help="permit absent weight shards")
+  parser.add_argument("--context", type=int, default=4096, help="context length used for the memory estimate")
+  args = parser.parse_args()
+  stats = audit_kimi_k3_checkpoint(args.model_dir, require_shards=not args.metadata_only)
+  if not 1 <= args.context <= 1_048_576: raise ValueError("--context must be between 1 and 1048576")
+
+  # K3 has 24 MLA layers. Each token stores the 512-value compressed latent plus 64 RoPE values in BF16.
+  per_gpu_weights = KIMI_K3_TP8_BYTES_PER_GPU
+  mla_cache = 24 * args.context * (512 + 64) * 2
+  hbm = 288_000_000_000
+  print(json.dumps(stats, indent=2))
+  print(f"exact text weights/GPU under this TP8 layout: {per_gpu_weights/1e9:.2f} GB ({per_gpu_weights/2**30:.2f} GiB)")
+  print(f"replicated MLA cache/GPU at {args.context:,} tokens: {mla_cache/1e9:.2f} GB ({mla_cache/2**30:.2f} GiB)")
+  print(f"nominal MI350X headroom before runtime buffers: {(hbm-per_gpu_weights-mla_cache)/1e9:.2f} GB")
+  if not args.metadata_only:
+    usage = shutil.disk_usage(args.model_dir)
+    print(f"filesystem free space: {usage.free/1e9:.2f} GB")
+
+if __name__ == "__main__": main()

examples/kimi_k3_smoke.py

@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""Run a reduced, architecture-complete K3 prefill/decode on tensor-parallel devices."""
+import argparse, time
+from tinygrad import Tensor, Device, dtypes, nn
+from tinygrad.llm.kimi_k3 import _shard_kimi_k3, kimi_k3_smoke_config
+from tinygrad.llm.model import Transformer
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--devices", type=int, default=8)
+  args = parser.parse_args()
+  if args.devices not in (1, 2, 4, 8): raise ValueError("the K3 admission smoke test supports 1, 2, 4, or 8 devices")
+  devices = tuple(f"AMD:{i}" for i in range(args.devices))
+  model = Transformer(kimi_k3_smoke_config())
+  for name,value in nn.state.get_state_dict(model).items():
+    fill = 127 if name.endswith("weight_scale") else 0
+    dtype = value.dtype if value.dtype is dtypes.uint8 else dtypes.bfloat16
+    value.replace(Tensor.full(value.shape, fill, dtype=dtype, device="CPU"))
+  _shard_kimi_k3(model, devices)
+  temperature = Tensor([0.0], device=devices)
+  for label,tokens,start in (("prefill", [[1, 2]], 0), ("decode", [[3]], 2), ("decode replay", [[4]], 3)):
+    begin = time.perf_counter()
+    out = model(Tensor(tokens, dtype=dtypes.int32, device=devices), start, temperature).realize()
+    for device in devices: Device[device].synchronize()
+    print(f"{label}: shape={out.shape}, {time.perf_counter()-begin:.3f}s")
+
+if __name__ == "__main__": main()

examples/mlperf/model_train.py

@@ -1742,8 +1742,8 @@ def train_gptoss():
   )
 
   for p in optim.params:
-    p.grad = p.zeros_like(dtype=dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype).contiguous()
-    if getattr(p, "_zero2", False): p.grad = optim.optimizers[0]._zero_shard(p.grad)
+    grad_dtype = dtypes.bfloat16 if p.dtype == FP8_DTYPE else p.dtype
+    p.grad = p.zeros_like(dtype=grad_dtype).contiguous()
   grads = [p.grad for p in optim.params]
 
   from extra.gemm.cdna_asm_gemm import _mx_block_scale

examples/mlperf/models/gpt_oss.py

@@ -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
 
@@ -146,7 +146,6 @@ class GPTOSS:
       return w_q, w_e8.is_param_(False)
     if moe:
       qs = [_one(*shape[1:]) for _ in range(shape[0])]
-      for q in qs: q[0]._zero2 = True  # grad arrives sharded on the expert axis under ZeRO-2 (moe_gemm)
       return [q[0] for q in qs], [q[1] for q in qs]
     return _one(*shape)
 
@@ -183,12 +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)
@@ -264,11 +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]:

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -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

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_run.sh

@@ -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=${DEV:-AMD}
 export CHECK_OOB=0
 export REWRITE_STACK_LIMIT=5000000 HCQDEV_WAIT_TIMEOUT_MS=240000
@@ -26,7 +26,7 @@ export FUSED_SILU_W13=${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))
 

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_beam.sh

@@ -11,7 +11,6 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-2}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}

examples/mlperf/training_submission_v6.1/tinycorp/benchmarks/gpt_oss/implementations/tinybox_8xMI350X/dev_run.sh

@@ -11,7 +11,6 @@ export DEVICE_IN_FUNCTION_BUG=1
 export DEBUG=${DEBUG:-0}
 export HK_FLASH_ATTENTION=${HK_FLASH_ATTENTION:-1}
 export ASM_GEMM=${ASM_GEMM:-1}
-export GROUPED_MOE=${GROUPED_MOE:-1}
 export ALL2ALL=${ALL2ALL:-1}
 export LATE_ALLREDUCE=${LATE_ALLREDUCE:-0}
 export ALLREDUCE_CAST=${ALLREDUCE_CAST:-1}

extra/benchmark_kimi.py

@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+"""Benchmark Kimi-Linear load, prefill, and decode on its TP4 checkpoint."""
+import argparse, resource, time
+from tinygrad import Device, TinyJit
+from tinygrad.helpers import profile_marker
+from tinygrad.llm.kimi import load_kimi
+
+def sync(devices:int) -> None:
+  for i in range(devices): Device[f"AMD:{i}"].synchronize()
+
+def timed_next(gen, devices:int) -> tuple[int, float]:
+  begin = time.perf_counter()
+  token = next(gen)
+  sync(devices)
+  return token, time.perf_counter()-begin
+
+def fresh_generate(model, prompt:list[int], chunk_size:int):
+  # Force recurrent/KV state reset so repeated runs and chunk sweeps measure the entire prompt,
+  # rather than silently reusing the prefix cached by the previous measurement.
+  model._cached_tokens = [-1] * len(prompt)
+  return model.generate(prompt.copy(), chunk_size=chunk_size)
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("model", help="converted Kimi-Linear-48B-A3B MXFP4-v2 directory")
+  parser.add_argument("--devices", type=int, default=4)
+  parser.add_argument("--max-context", type=int, default=128)
+  parser.add_argument("--prompt-tokens", type=int, default=32)
+  parser.add_argument("--decode-tokens", type=int, default=8)
+  parser.add_argument("--chunk-size", type=int, default=32)
+  parser.add_argument("--sweep-chunks", help="comma-separated prefill chunk sizes; uses the fastest for decode")
+  args = parser.parse_args()
+  if args.prompt_tokens < 1 or args.prompt_tokens + args.decode_tokens + 1 > args.max_context:
+    raise ValueError("prompt and decode tokens must fit within --max-context")
+
+  begin = time.perf_counter()
+  model = load_kimi(args.model, max_context=args.max_context, devices=args.devices)
+  sync(args.devices)
+  print(f"load: {time.perf_counter()-begin:.3f}s", flush=True)
+
+  prompt = [1] + [1000+i%1000 for i in range(args.prompt_tokens-1)]
+  chunks = [int(x) for x in args.sweep_chunks.split(",")] if args.sweep_chunks else [args.chunk_size]
+  if any(x < 1 or x > args.prompt_tokens for x in chunks): raise ValueError("prefill chunks must be between 1 and --prompt-tokens")
+  timings:list[tuple[float, int]] = []
+  prefill_jits:dict[int, TinyJit] = {}
+  for chunk in chunks:
+    # Recurrent prefill has a static token dimension. Give each swept shape its own capture;
+    # the rollout JIT remains shared and independently benchmarks chunk 1/decode.
+    if chunk != 1: model.prefill_jit = TinyJit(model.forward)
+    cold = fresh_generate(model, prompt, chunk)
+    first, cold_prefill = timed_next(cold, args.devices)
+    print(f"chunk {chunk}: cold prefill {cold_prefill:.3f}s, token={first}", flush=True)
+    warm = fresh_generate(model, prompt, chunk)
+    warm_first, prefill = timed_next(warm, args.devices)
+    if first != warm_first: raise RuntimeError(f"chunk {chunk} is not repeatable: cold={first}, warm={warm_first}")
+    timings.append((prefill, chunk))
+    if chunk != 1: prefill_jits[chunk] = model.prefill_jit
+    print(f"chunk {chunk}: prefill {prefill:.3f}s ({args.prompt_tokens/prefill:.3f} tok/s), token={first}", flush=True)
+
+  prefill, best_chunk = min(timings)
+  if best_chunk != 1: model.prefill_jit = prefill_jits[best_chunk]
+  warm = fresh_generate(model, prompt, best_chunk)
+  first, replay_prefill = timed_next(warm, args.devices)
+  _, cold_decode = timed_next(warm, args.devices)
+  _, capture_decode = timed_next(warm, args.devices)
+  print(f"selected chunk: {best_chunk}; prefill replay {replay_prefill:.3f}s "
+        f"({args.prompt_tokens/replay_prefill:.3f} tok/s), token={first}", flush=True)
+  print(f"cold decode: {cold_decode:.3f}s", flush=True)
+  print(f"capture decode: {capture_decode:.3f}s", flush=True)
+  profile_marker("kimi decode steady start")
+  begin = time.perf_counter()
+  output = [next(warm) for _ in range(args.decode_tokens)]
+  sync(args.devices)
+  decode = time.perf_counter()-begin
+  profile_marker("kimi decode steady end")
+  print(f"decode: {decode:.3f}s ({args.decode_tokens/decode:.3f} tok/s, {decode/args.decode_tokens*1e3:.3f} ms/tok), output={output}", flush=True)
+  print(f"peak RSS: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss/1024:.1f} MiB", flush=True)
+
+if __name__ == "__main__": main()

extra/benchmark_kimi_k3.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""Bounded correctness and load/prefill/decode benchmark for the official TP8 Kimi K3 checkpoint."""
+import argparse, resource, time
+
+from tinygrad import Device
+from tinygrad.helpers import profile_marker
+from tinygrad.llm.cli import KimiK3Template, SimpleTokenizer
+from tinygrad.llm.kimi_k3 import load_kimi_k3, load_kimi_tokenizer_data
+
+def sync(devices:int) -> None:
+  for i in range(devices): Device[f"AMD:{i}"].synchronize()
+
+def fresh_generate(model, prompt:list[int], chunk_size:int):
+  # Never reuse a prefix or recurrent state across correctness/benchmark trials.
+  model._cached_tokens = [-1] * len(prompt)
+  return model.generate(prompt.copy(), chunk_size=chunk_size, temperature=0.0)
+
+def timed_next(gen, devices:int) -> tuple[int, float]:
+  begin = time.perf_counter()
+  token = next(gen)
+  sync(devices)
+  return token, time.perf_counter()-begin
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("model", help="official unmodified Kimi K3 checkpoint directory")
+  parser.add_argument("--devices", type=int, default=8)
+  parser.add_argument("--max-context", type=int, default=128)
+  parser.add_argument("--prompt", default="Reply with exactly: OK")
+  parser.add_argument("--stable-tokens", type=int, default=8)
+  parser.add_argument("--decode-tokens", type=int, default=8)
+  parser.add_argument("--chunk-size", type=int, default=128)
+  args = parser.parse_args()
+
+  begin = time.perf_counter()
+  model = load_kimi_k3(args.model, max_context=args.max_context, devices=args.devices)
+  sync(args.devices)
+  load_time = time.perf_counter()-begin
+  print(f"load: {load_time:.3f}s", flush=True)
+
+  normal, special, bos, eos = load_kimi_tokenizer_data(args.model)
+  tok = SimpleTokenizer(normal, special, "kimi-k2", bos_id=bos, eos_id=eos, eot_id=eos)
+  rendered = KimiK3Template().render(messages=[{"role":"user", "content":args.prompt}], add_generation_prompt=True)
+  prompt = tok.encode(rendered)
+  needed = len(prompt) + max(args.stable_tokens, args.decode_tokens+3)
+  if needed > args.max_context: raise ValueError(f"prompt and output need {needed} tokens but max context is {args.max_context}")
+  print(f"prompt: {len(prompt)} tokens, chunk={args.chunk_size}", flush=True)
+
+  sequences:list[list[int]] = []
+  # TinyJit executes uncaptured once, captures the second call, and replays from the third call.
+  # Compare two replay paths rather than capture numerics/timing against replay.
+  for trial in range(4):
+    gen = fresh_generate(model, prompt, args.chunk_size)
+    sequence:list[int] = []
+    prefill = 0.0
+    for step in range(args.stable_tokens):
+      token, elapsed = timed_next(gen, args.devices)
+      sequence.append(token)
+      if step == 0: prefill = elapsed
+    if trial >= 2: sequences.append(sequence)
+    label = ("uncaptured warmup", "capture warmup", "stable trial 1", "stable trial 2")[trial]
+    print(f"{label}: prefill={prefill:.3f}s "
+          f"({len(prompt)/prefill:.3f} tok/s), tokens={sequence}", flush=True)
+  if sequences[0] != sequences[1]: raise RuntimeError(f"greedy output is not repeatable: {sequences}")
+  print(f"stable text: {tok.decode(sequences[0])!r}", flush=True)
+
+  gen = fresh_generate(model, prompt, args.chunk_size)
+  profile_marker("kimi k3 steady prefill start")
+  first, prefill = timed_next(gen, args.devices)
+  profile_marker("kimi k3 steady prefill end")
+  warmup = [timed_next(gen, args.devices)[0] for _ in range(2)]
+  profile_marker("kimi k3 steady decode start")
+  begin = time.perf_counter()
+  output = [next(gen) for _ in range(args.decode_tokens)]
+  sync(args.devices)
+  decode = time.perf_counter()-begin
+  profile_marker("kimi k3 steady decode end")
+  print(f"prefill replay: {prefill:.3f}s ({len(prompt)/prefill:.3f} tok/s), token={first}", flush=True)
+  print(f"decode after warmup {warmup}: {decode:.3f}s ({args.decode_tokens/decode:.3f} tok/s, "
+        f"{decode/args.decode_tokens*1e3:.3f} ms/tok), output={output}", flush=True)
+  print(f"peak RSS: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss/1024:.1f} MiB", flush=True)
+
+if __name__ == "__main__": main()

extra/benchmark_kimi_k3_fake.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+"""Fast exact-shape K3 KDA/layer benchmark using bounded fake weights instead of the 1.56 TB checkpoint."""
+from __future__ import annotations
+import argparse, statistics, time
+from dataclasses import replace
+
+from tinygrad import Device, Tensor, TinyJit, dtypes, nn
+from tinygrad.helpers import profile_marker
+from tinygrad.llm.kimi_k3 import kimi_k3_config
+from tinygrad.llm.model import GatedDeltaNetBlock
+
+def tp_axis(name:str) -> int|None:
+  if "ffn_gate_exps.weight" in name or "ffn_up_exps.weight" in name: return 1
+  if "ffn_gate_exps.weight_scale" in name or "ffn_up_exps.weight_scale" in name: return 1
+  if "ffn_down_exps.weight" in name or "ffn_down_exps.weight_scale" in name: return 2
+  if name.endswith(("ffn_gate_shexp.weight", "ffn_up_shexp.weight")): return 0
+  if name.endswith(("ffn_down_shexp.weight", "ffn_routed_down.weight", "ffn_routed_up.weight", "ssm_out.weight")): return 1
+  if name.endswith(("attn_q.weight", "attn_k.weight", "attn_v.weight", "ssm_g_full.weight", "ssm_f_b.weight", "ssm_beta.weight")): return 0
+  if name.endswith(("ssm_q_conv1d.weight", "ssm_k_conv1d.weight", "ssm_v_conv1d.weight", "ssm_dt.bias")): return 0
+  return None
+
+def fake_value(name:str) -> tuple[int|float, object]:
+  if name.endswith("weight_scale"): return 120, dtypes.uint8
+  if name.endswith("_exps.weight"): return 0x11, dtypes.uint8
+  if name.endswith("ssm_a"): return -0.1, dtypes.float32
+  if name.endswith("ssm_dt.bias"): return 0.1, dtypes.float32
+  if "conv1d.weight" in name: return 0.1, dtypes.float32
+  if name.endswith("exp_probs_b.bias"): return 0.0, dtypes.float32
+  if name.endswith("norm.weight"): return 1.0, dtypes.bfloat16
+  return 0.001, dtypes.bfloat16
+
+def fake_tp_tensor(shape:tuple[int, ...], value:int|float, dtype, devices:tuple[str, ...], axis:int|None) -> Tensor:
+  if axis is not None and shape[axis] % len(devices): raise ValueError(f"shape {shape} is not TP{len(devices)} divisible on axis {axis}")
+  source = Tensor.full(shape, value, dtype=dtype, device=devices[0]).clone().realize()
+  return source.shard(devices, axis=axis).realize()
+
+def sync(devices:tuple[str, ...]) -> None:
+  for device in devices: Device[device].synchronize()
+
+def main() -> None:
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--devices", type=int, default=8)
+  parser.add_argument("--mode", choices=("attention", "block"), default="attention")
+  parser.add_argument("--iterations", type=int, default=20)
+  args = parser.parse_args()
+  devices = tuple(f"AMD:{i}" for i in range(args.devices))
+  # One exact-width KDA layer, but only 16 fake routed experts. This retains top-k 16 and every
+  # official per-GPU matrix/state shape while keeping fake expert storage below 300 MB per layer.
+  config = replace(kimi_k3_config(4), num_blocks=1, num_experts=16, num_experts_per_tok=16, ssm_layers=(True,),
+                   attn_res_block_size=0)
+  block = GatedDeltaNetBlock(config, config.ssm)
+  begin = time.perf_counter()
+  for name,tensor in nn.state.get_state_dict(block).items():
+    if args.mode == "attention" and name.startswith(("ffn_", "exp_probs_")): continue
+    value, dtype = fake_value(name)
+    tensor.replace(fake_tp_tensor(tuple(int(x) for x in tensor.shape), value, dtype, devices, tp_axis(name)))
+  sync(devices)
+  print(f"fake weights: {time.perf_counter()-begin:.3f}s", flush=True)
+  x_source = (((Tensor.arange(config.dim, dtype=dtypes.float32).reshape(1, 1, config.dim) % 31) / 31) \
+    .cast(dtypes.bfloat16).to(devices[0])).clone().realize()
+  x = x_source.shard(devices, axis=None).realize()
+  block._init_state(x)
+  # Use direct buffer-backed state shards. The production path reaches this form after prefill;
+  # the fake harness begins immediately at decode and must not feed lazy clone graphs to TinyJit.
+  for state,axis in ((block.conv_state_q, 2), (block.conv_state_k, 2), (block.conv_state_v, 2), (block.recurrent_state, 1)):
+    state.replace(Tensor.zeros(*state.shape, dtype=state.dtype, device=devices[0]).shard(devices, axis=axis).realize())
+
+  @TinyJit
+  def run(inp:Tensor) -> Tensor:
+    if args.mode == "attention": return block._attention(block.attn_norm(inp), 0).realize()
+    return block(inp, 0).realize()
+
+  # uncaptured, capture, then replay only
+  run(x); sync(devices)
+  run(x); sync(devices)
+  samples:list[float] = []
+  profile_marker(f"fake K3 {args.mode} start")
+  for _ in range(args.iterations):
+    begin = time.perf_counter(); out = run(x); sync(devices); samples.append((time.perf_counter()-begin)*1e3)
+  profile_marker(f"fake K3 {args.mode} end")
+  print(f"{args.mode}: median={statistics.median(samples):.3f} ms/layer, min={min(samples):.3f} ms/layer, "
+        f"projected_93_layer_rate={1000/(statistics.median(samples)*93):.3f} tok/s, finite={out.float().isfinite().all().item()}")
+
+if __name__ == "__main__": main()

extra/export_model.py

@@ -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

extra/gemm/cdna_asm_gemm.py

@@ -122,8 +122,7 @@ 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))))
 

extra/gemm/moe_gemm.py

@@ -1,32 +1,10 @@
 import functools, pathlib
 from tinygrad import Tensor, dtypes
-from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
-from tinygrad.helpers import getenv
+from tinygrad.uop.ops import UOp, Ops, KernelInfo
 from tinygrad.renderer import Estimates
 from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
 from extra.gemm.cdna_asm_gemm import quantize_mxfp8, _mx_block_scale, _mx_block_scale_3d
 
-ZERO_OPTIM = getenv("ZERO_OPTIM", 0)
-
-def reduce_scatter_devaxis(out:Tensor, shard_axis:int=0) -> Tensor:
-  # out: sharded on the device axis, shape (ndev, *rest); return the device-axis sum left sharded on shard_axis.
-  u = out.uop
-  devs, rest = u.device, u.shape[1:]
-  assert rest[shard_axis] % len(devs) == 0, f"reduce_scatter needs even shards: {rest[shard_axis]} % {len(devs)}"
-  # reach the raw per-device buffer below the UNSHARD, keeping the AFTERs so reads stay ordered after the kernel writes
-  node, barriers = u, []
-  while node.op is not Ops.UNSHARD:
-    if node.op is Ops.AFTER: barriers += node.src[1:]
-    node = node.src[0]
-  mbuf = node.src[0].after(*barriers) if barriers else node.src[0]
-  sz = rest[shard_axis] // len(devs)
-  shards = []
-  for i in range(len(devs)):
-    bounds = tuple((0,s) if a != shard_axis else (i*sz,(i+1)*sz) for a,s in enumerate(rest))
-    contribs = [mbuf.mselect(j).reshape(rest).shrink(bounds).copy_to_device(devs[i]) for j in range(len(devs))]
-    shards.append(functools.reduce(lambda a,b: a.alu(Ops.ADD, b), contribs))
-  return Tensor(UOp.mstack(*shards).unshard(shard_axis, UOp.range(len(devs), -1, AxisType.DEVICE)), device=devs)
-
 @functools.cache
 def custom_hk_grouped_mxfp8_gemm(C:UOp, A:UOp, B:UOp, scale_A:UOp, scale_B:UOp, *extra:UOp, dname:str, n_experts:int) -> UOp:
   M, K = A.shape
@@ -80,8 +58,7 @@ def grouped_mx_wgrad(g:Tensor, xg:Tensor, expert_off:Tensor, n_experts:int) -> T
   out = Tensor(inv.uop.unshard(0), device=g.device) if is_multi else inv
   out = Tensor.custom_kernel(out, gT, xT, g_si, x_si, expert_off,
                              fxn=functools.partial(custom_hk_grouped_mxfp8_wgrad, dname=dname, n_experts=n_experts))[0]
-  if is_multi and ZERO_OPTIM: out = reduce_scatter_devaxis(out, 0)
-  else: out = out.sum(0) if is_multi else out.squeeze(0)
+  out = out.sum(0) if is_multi else out.squeeze(0)
   return out.reshape(n_experts, N, K)
 
 def mx_pack_3d(e8:Tensor) -> Tensor:

extra/gemm/triton_nv_matmul.py

@@ -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

extra/hcq2/ops_amd2.py

@@ -3,7 +3,7 @@ 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 HCQ2Compiled, HCQAllocator, HCQ2Buffer, encode_kernargs_clike, make_cmdbuf
 from tinygrad.runtime.support.hcq2 import make_binary_patch
 from tinygrad.uop.ops import sint, UOp
 from tinygrad.device import Compiled, BufferSpec, Buffer, Device
@@ -179,10 +179,11 @@ class SDMAOps(FastEnum): COPY = auto(); POLL_REGMEM = auto(); FENCE = auto(); TR
 
 def sdma_copy(ctx, call):
   sz = call.src[2].max_numel() * call.src[2].dtype.itemsize
-  hdr = ctx.sdma.SDMA_OP_COPY | ctx.sdma.SDMA_PKT_COPY_LINEAR_HEADER_SUB_OP(ctx.sdma.SDMA_SUBOP_COPY_LINEAR)
-  return call.ins(SDMAOps.COPY, src=tuple(x for off in range(0, sz, ctx.max_copy_size) for x in (
-    *(UOp.const(v, dtypes.uint32) for v in (hdr, ctx.sdma.SDMA_PKT_COPY_LINEAR_COUNT_COUNT(min(sz-off, ctx.max_copy_size)-1), 0)),
-    *(a + UOp.const(off, dtypes.uint64) if off else a for a in (call.src[2].getaddr(ctx.devs), call.src[1].getaddr(ctx.devs))))))
+  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)))))
 
 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) \
@@ -287,16 +288,14 @@ 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:
@@ -389,24 +388,15 @@ 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)
@@ -478,7 +468,6 @@ 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
@@ -560,7 +549,9 @@ class AMDDevice(HCQ2Compiled):
   def is_usb(self) -> bool: return False
 
   def __init__(self, device:str=""):
-    self.iface = self._select_iface(device)
+    self.device_id = int(device.split(":")[1]) if ":" in device else 0
+
+    self.iface = self._select_iface()
 
     self.target:tuple[int, ...] = ((trgt:=self.iface.props['gfx_target_version']) // 10000, (trgt // 100) % 100, trgt % 100)
     self.arch = "gfx%d%x%x" % self.target
@@ -590,7 +581,7 @@ 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)
 

extra/runbook_digitalocean_mi350x.md

@@ -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.

extra/thunder/amd/fa.py

@@ -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
@@ -293,10 +247,10 @@ def custom_fa_backward_pre(delta_vec:UOp, dq:UOp, o:UOp, do:UOp, device:str, arc
              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

extra/thunder/amd/matvec_bf16_splitk.cpp

@@ -0,0 +1,74 @@
+#include "kittens.cuh"
+
+using namespace kittens;
+
+#ifndef MATVEC_N
+#define MATVEC_N 1536
+#endif
+#ifndef MATVEC_K
+#define MATVEC_K 7168
+#endif
+
+constexpr int SPLIT_WAVES = 8;
+
+template<int W>
+__device__ __forceinline__ float run_split(const bf16 *A_ptr, const bf16 *B_ptr, int out_base,
+                                           st_bf<16, 32, st_16x32_s> &As,
+                                           st_bf<16, 32, st_16x32_s> &Bs) {
+  constexpr int K = MATVEC_K;
+  rt_bf<16, 32, row_l, rt_16x32_s> A;
+  rt_bf<16, 32, row_l, rt_16x32_s> B;
+  rt_fl<16, 16, col_l, rt_16x16_s> C;
+  zero(C);
+  const int lane = laneid();
+  constexpr int k_begin = W * (K / SPLIT_WAVES), k_end = k_begin + K / SPLIT_WAVES;
+  #pragma unroll 1
+  for (int k = k_begin; k < k_end; k += 32) {
+    #pragma unroll
+    for (int idx = lane; idx < 16 * 32; idx += 64) {
+      const int row = idx / 32, col = idx % 32;
+      *reinterpret_cast<bf16 *>(reinterpret_cast<char *>(&As.data[0]) + As.swizzle({row, col})) = A_ptr[k + col];
+      *reinterpret_cast<bf16 *>(reinterpret_cast<char *>(&Bs.data[0]) + Bs.swizzle({row, col})) =
+        B_ptr[(out_base + row) * K + k + col];
+    }
+    asm volatile("s_waitcnt lgkmcnt(0)");
+    load(A, As);
+    load(B, Bs);
+    asm volatile("s_waitcnt lgkmcnt(0)");
+    mma_ABt(C, A, B, C);
+  }
+  return C.tiles[0][0].data[0].x;
+}
+
+// Eight waves split K for one 16-channel output tile.  Each wave uses MFMA on
+// a repeated activation row, then wave zero reduces the eight FP32 partials.
+__global__ __launch_bounds__(64 * SPLIT_WAVES, 1)
+void hk_bf16_matvec_splitk(bf16 *C_ptr, const bf16 *A_ptr, const bf16 *B_ptr, bf16 *unused) {
+  constexpr int N = MATVEC_N, K = MATVEC_K;
+  static_assert(N % 16 == 0 && K % (32 * SPLIT_WAVES) == 0);
+  __shared__ st_bf<16, 32, st_16x32_s> As[SPLIT_WAVES];
+  __shared__ st_bf<16, 32, st_16x32_s> Bs[SPLIT_WAVES];
+  __shared__ float partial[SPLIT_WAVES][16];
+  const int tid = threadIdx.x, wave = tid / 64, lane = tid & 63;
+  const int out_base = blockIdx.x * 16;
+  float result = 0.0f;
+  switch (wave) {
+    case 0: result = run_split<0>(A_ptr, B_ptr, out_base, As[0], Bs[0]); break;
+    case 1: result = run_split<1>(A_ptr, B_ptr, out_base, As[1], Bs[1]); break;
+    case 2: result = run_split<2>(A_ptr, B_ptr, out_base, As[2], Bs[2]); break;
+    case 3: result = run_split<3>(A_ptr, B_ptr, out_base, As[3], Bs[3]); break;
+    case 4: result = run_split<4>(A_ptr, B_ptr, out_base, As[4], Bs[4]); break;
+    case 5: result = run_split<5>(A_ptr, B_ptr, out_base, As[5], Bs[5]); break;
+    case 6: result = run_split<6>(A_ptr, B_ptr, out_base, As[6], Bs[6]); break;
+    case 7: result = run_split<7>(A_ptr, B_ptr, out_base, As[7], Bs[7]); break;
+  }
+  if (lane < 16) partial[wave][lane] = result;
+  asm volatile("s_waitcnt lgkmcnt(0)");
+  __builtin_amdgcn_s_barrier();
+  if (wave == 0 && lane < 16) {
+    float total = 0.0f;
+    #pragma unroll
+    for (int i = 0; i < SPLIT_WAVES; i++) total += partial[i][lane];
+    C_ptr[out_base + lane] = static_cast<bf16>(total);
+  }
+}

extra/thunder/tiny/tk/tiles.py

@@ -209,7 +209,7 @@ class ST:
     return cls(uop, rows, cols, layout, base_shape, ker)
 
   def swizzle(self, row, col):
-    swizzled_offset = self.base_shape.swizzle(row, col, self._uop.dtype)
+    swizzled_offset = self.base_shape.swizzle(row, col, self._uop.dtype.scalar())
 
     row = swizzled_offset // self.base_shape.cols
     col = swizzled_offset % self.base_shape.cols

extra/torch_backend/backend.py

@@ -4,7 +4,7 @@
 # A006 Lambda argument `input` is shadowing a Python builtin
 from tinygrad import Tensor, dtypes, Device
 from tinygrad.uop.ops import Ops, GroupOp
-from tinygrad.helpers import getenv, prod, strides_for_shape
+from tinygrad.helpers import getenv, prod, strides_for_shape, argfix
 import torch.lib
 TORCH_DEBUG = getenv("TORCH_DEBUG")
 import torch, pathlib, operator, functools, weakref
@@ -73,12 +73,6 @@ def wrap_view_op(fn):
     return wrap(ret)
   return _wrap
 
-# NOTE: list assignment raises IndexError on an out of range dim, and the index must be a tuple: a list of all ints is one advanced index
-def _index_dim(self, dim, idx):
-  idxs = [slice(None)] * self.ndim
-  idxs[dim] = idx
-  return self[tuple(idxs)]
-
 view_ops = {
   "aten.view": Tensor.reshape,
   "aten._unsafe_view": Tensor.reshape,  # when are views unsafe, and do we care?
@@ -88,13 +82,15 @@ view_ops = {
   "aten.transpose.int": Tensor.transpose,
   "aten.squeeze.dim": Tensor.squeeze,
   "aten.unsqueeze": Tensor.unsqueeze,
-  "aten.select.int": _index_dim,
+  "aten.select.int": lambda self, dim, idx: self[(slice(None),) * (dim%self.ndim) + (idx,)],
   "aten.permute": Tensor.permute,
   "aten.alias": lambda self: self,
   "aten.diagonal": Tensor.diagonal,
-  "aten.slice.Tensor": lambda self, dim=0, start=None, end=None, step=1: _index_dim(self, dim, slice(start, end, step)),
   }
 
+# torch 2.10 handles this natively
+if tuple(map(int, torch.__version__.split('.')[:2])) < (2, 10): view_ops.update({"aten.detach": Tensor.detach})
+
 for k,v in view_ops.items(): torch.library.impl(k.replace("aten.", "aten::"), "privateuseone")(wrap_view_op(v))
 
 def _get_view_ops(view): return getattr(view, "_view_ops", [])
@@ -103,21 +99,46 @@ def _apply_view_ops(target, ops):
   for fn, args, kwargs in ops: target = fn(target, *args, **kwargs)
   return target
 
-# a chain of reshapes is undone by reshaping the value back to the base
+# similar to https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/InferSize.h
+def _reshape_target_shape(shape:tuple[int, ...], args) -> tuple[int, ...]|None:
+  if not (req := argfix(*args)): return None
+  new_shape, infer_idx = [], -1
+  for i, s in enumerate(req):
+    if s is None: s = shape[i] if i < len(shape) else None
+    if not isinstance(s, int): return None
+    if s == -1:
+      if infer_idx != -1: return None
+      infer_idx = len(new_shape)
+    new_shape.append(s)
+  total = prod(shape)
+  if infer_idx != -1:
+    known = prod(x for x in new_shape if x != -1)
+    if known == 0:
+      if total != 0: return None
+      new_shape[infer_idx] = 0
+    else: new_shape[infer_idx] = total // known
+  return tuple(new_shape) if prod(new_shape) == total else None
+
+# TODO: can we get rid of this? only for test_flatten_reshape_add
 def _try_simple_reshape_view_write(base: Tensor, view: Tensor, val: Tensor) -> bool:
   if not (ops := _get_view_ops(view)): return False
-  if any(fn is not Tensor.reshape for fn, _, _ in ops): return False
-  base.assign(val.reshape(base.shape))
+  shapes = [base.shape]
+  for fn, args, _ in ops:
+    if fn is Tensor.reshape:
+      if not (next_shape := _reshape_target_shape(shapes[-1], args)): return False
+      shapes.append(next_shape)
+  if shapes[-1] != view.shape: return False
+  for s in reversed(shapes[:-1]): val = val.reshape(s)
+  base.assign(val)
   return True
 
 def _view_write(base: Tensor, view: Tensor, value: Tensor) -> None:
   val = value if value.dtype == base.dtype else value.cast(base.dtype)
+  if view.shape == base.shape: return base.assign(val)
   if _try_simple_reshape_view_write(base, view, val): return
   idx_base = Tensor.arange(base.numel(), dtype=dtypes.int32).reshape(base.shape)
   idx_view = _apply_view_ops(idx_base, _get_view_ops(view)).reshape(-1)
-  # clone, not contiguous: contiguous() on a base that already owns its buffer returns the base itself, and scattering
-  # into that is an in-place write to a buffer other tensors still hold, which setitem refuses
-  flat_base = base.reshape(base.numel()).clone()
+  flat_base = base.reshape(base.numel()).contiguous()
   flat_base[idx_view] = val.reshape(-1)
   base.assign(flat_base.reshape(base.shape))
 
@@ -145,6 +166,11 @@ def _index_put_impl_(self, indices, values, accumulate=False, unsafe=False):
 def index_put(self, indices, values, accumulate=False):
   return aten.index_put(self.cpu(), [z.cpu() if isinstance(z, torch.Tensor) else None for z in indices], values.clone().cpu(), accumulate).tiny()
 
+@torch.library.impl("aten::isin.Tensor_Tensor_out", "privateuseone")
+def isin_tensor_tensor_out(x, y, *, assume_unique=False, invert=False, out=None):
+  result = (unwrap(x).unsqueeze(-1) == unwrap(y).flatten()).any(-1)
+  return out.copy_(wrap(~result if invert else result))
+
 @torch.library.impl("aten::randperm.generator_out", "privateuseone")
 def randperm_generator(n, generator=None, out=None):
   if generator is not None: raise NotImplementedError("tinygrad torch backend does not support torch.Generator for randperm")
@@ -205,6 +231,49 @@ def as_strided(tensor:torch.Tensor, size, stride, storage_offset=None):
 def _reshape_alias(tensor:torch.Tensor, size, stride):
   return _as_strided(tensor, size, stride)
 
+@torch.library.impl("aten::empty_strided", "privateuseone")
+def empty_strided(size, stride, dtype=None, layout=None, device=None, pin_memory=False):
+  if TORCH_DEBUG: print(f"empty_strided {size=} {stride=} {dtype=} {layout=} {device=} {pin_memory=}")
+  ret = Tensor.empty(*size, dtype=_from_torch_dtype(dtype or torch.get_default_dtype()), device=_from_torch_device(device))
+  # TODO: should return with requested strides
+  return wrap(ret)
+
+@torch.library.impl("aten::empty.memory_format", "privateuseone")
+def empty_memory_format(size, dtype=None, layout=None, device=None, pin_memory=False, memory_format=None):
+  if TORCH_DEBUG: print(f"empty.memory_format {size=} {dtype=} {layout=} {device=} {pin_memory=} {memory_format=}")
+  ret = Tensor.empty(*size, dtype=_from_torch_dtype(dtype or torch.get_default_dtype()), device=_from_torch_device(device))
+  return wrap(ret)
+
+@torch.library.impl("aten::max_pool2d_with_indices", "privateuseone")
+def max_pool2d_with_indices(self:torch.Tensor, kernel_size:tuple[int, ...], stride=None, padding=0, dilation=1, ceil_mode=False):
+  # TODO: supprt stride [] in tinygrad?
+  if stride is not None and len(stride) == 0: stride = None
+  ret, idx = unwrap(self).max_pool2d(kernel_size, stride, dilation, padding, ceil_mode, return_indices=True)
+  return (wrap(ret), wrap(idx.cast(dtypes.int64)))
+
+@torch.library.impl("aten::max_pool2d_with_indices_backward", "privateuseone")
+def max_pool2d_with_indices_backward(grad_out:torch.Tensor, self:torch.Tensor, kernel_size:tuple[int, ...], stride=None, padding=0, dilation=1, ceil_mode=False, indices=None):
+  return wrap(Tensor.max_unpool2d(unwrap(grad_out), unwrap(indices), output_size=unwrap(self).shape))
+
+@torch.library.impl("aten::max_unpool2d", "privateuseone")
+def max_unpool2d(self:torch.Tensor, indices:torch.Tensor, output_size):
+  return wrap(unwrap(self).max_unpool2d(unwrap(indices), output_size=output_size))
+
+@torch.library.impl("aten::arange", "privateuseone")
+def arange(end, dtype=None, device=None, pin_memory=None):
+  has_float = isinstance(end, float)
+  return wrap(Tensor.arange(0, end, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
+
+@torch.library.impl("aten::arange.start", "privateuseone")
+def arange_start(start, end, dtype=None, device=None, pin_memory=None):
+  has_float = any(isinstance(x, float) for x in (start, end))
+  return wrap(Tensor.arange(start, end, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
+
+@torch.library.impl("aten::arange.start_step", "privateuseone")
+def arange_start_step(start, end, step, dtype=None, device=None, pin_memory=None):
+  has_float = any(isinstance(x, float) for x in (start, end, step))
+  return wrap(Tensor.arange(start, end, step, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if has_float else torch.int64))))
+
 @torch.library.impl("aten::convolution_overrideable", "privateuseone")
 def convolution_overrideable(input, weight, bias, stride, padding, dilation, transposed, output_padding, groups):
   if TORCH_DEBUG >= 1:
@@ -225,27 +294,12 @@ def convolution_backward_overrideable(grad_out, input, weight, stride, padding,
   grads = out.gradient(*[t for t,m in zip([input, weight, bias], output_mask) if m], gradient=grad_out)
   return tuple([wrap(grads.pop(0)) if m else None for m in output_mask])
 
-# the functional scatters. without an impl aten falls back to a path that assumes a real storage: "self.has_storage() INTERNAL ASSERT FAILED"
-def _scatter_into(self, src, dim, index):
-  out = unwrap(self).clone()
-  slices = [slice(None)] * out.ndim
-  slices[dim] = index
-  out[slices] = unwrap(src).cast(out.dtype)  # torch casts src to self's dtype, tinygrad setitem demands they already match
-  return wrap(out)
-
-@torch.library.impl("aten::slice_scatter", "privateuseone")
-def slice_scatter(self, src, dim=0, start=None, end=None, step=1): return _scatter_into(self, src, dim, slice(start, end, step))
-
-@torch.library.impl("aten::select_scatter", "privateuseone")
-def select_scatter(self, src, dim, index): return _scatter_into(self, src, dim, index)
-
-@torch.library.impl("aten::diagonal_scatter", "privateuseone")
-def diagonal_scatter(self, src, offset=0, dim1=0, dim2=1):
-  # a diagonal is not one axis, so scatter through the flat indices it picks out
-  base, out = unwrap(self), unwrap(self).clone().reshape(-1)
-  idx = Tensor.arange(base.numel(), dtype=dtypes.int32).reshape(base.shape).diagonal(offset, dim1, dim2).reshape(-1)
-  out[idx] = unwrap(src).cast(base.dtype).reshape(-1)
-  return wrap(out.reshape(base.shape))
+@torch.library.impl("aten::slice.Tensor", "privateuseone")
+@wrap_view_op
+def slice_tensor(self, dim=0, start=None, end=None, step=1):
+  slices = [slice(None)] * self.ndim
+  slices[dim] = slice(start, end, step)
+  return self[slices]
 
 @torch.library.impl("aten::slice_backward", "privateuseone")
 def slice_backward(grad_out, input_sizes, dim, start, end, step):
@@ -287,14 +341,19 @@ for dim in [1, 2, 3]:
     torch.library.impl(f"aten::{pad_type}_pad{dim}d", "privateuseone")(functools.partial(pad_forward, mode=mode))
     torch.library.impl(f"aten::{pad_type}_pad{dim}d_backward", "privateuseone")(functools.partial(pad_backward, mode=mode))
 
-# the schemas are all positional: (self, output_size, align_corners, *scales) for linear, (self, output_size, *scales) for nearest.
-def upsample(self, size, *args, mode=None):
-  return wrap(Tensor.interpolate(unwrap(self), size, mode=mode, align_corners=args[0] if mode == "linear" else False))
+def upsample(self, size, align_corners=False, mode=None): return wrap(Tensor.interpolate(unwrap(self), size, mode=mode, align_corners=align_corners))
 for i,pre in enumerate(["", "bi", "tri"]):
   torch.library.impl(f"aten::upsample_{pre}linear{i+1}d", "privateuseone")(functools.partial(upsample, mode="linear"))
   torch.library.impl(f"aten::upsample_nearest{i+1}d", "privateuseone")(functools.partial(upsample, mode="nearest"))
   torch.library.impl(f"aten::_upsample_nearest_exact{i+1}d", "privateuseone")(functools.partial(upsample, mode="nearest-exact"))
 
+@torch.library.impl("aten::scatter_add.out", "privateuseone")
+def scatter_add(self, dim, index, src, out):
+  self, index, src, out_unwrapped = unwrap(self), unwrap(index), unwrap(src), unwrap(out)
+  if self.shape == (): _apply_inplace(out_unwrapped, src)
+  else: _apply_inplace(out_unwrapped, Tensor.scatter_reduce(self, dim, index, src, reduce='sum'))
+  return out
+
 def _copy_between_devices(src, dest, cast_dtype, to_device, non_blocking=False):
   if src.is_tiny and dest.is_tiny:
     src_t, dest_t = unwrap(src), unwrap(dest)
@@ -345,11 +404,15 @@ def sort_values(input, dim=-1, descending=False, stable=True, values=None, indic
   _apply_inplace(unwrap(indices), out_indices.cast(dtypes.int64))
   return values, indices
 
+@torch.library.impl("aten::_linalg_svd", "privateuseone")
+def _linalg_svd(self, full_matrices=False):
+  U, S, Vh = unwrap(self).svd(full_matrices)
+  return wrap(U), wrap(S), wrap(Vh)
+
 # register some decompositions
 from torch._decomp import get_decompositions
 decomps = [
   aten.native_layer_norm_backward,
-  aten.native_group_norm_backward,
   aten.linalg_cross,
   aten.addmm,
   aten.addcmul,
@@ -384,20 +447,12 @@ decomps = [
   aten._softmax_backward_data, aten.embedding_dense_backward,
   aten.linalg_vector_norm,
   aten.binary_cross_entropy, aten.binary_cross_entropy_backward,
-  # the C++ mse/smooth_l1 kernels resize their out tensor, and a tiny tensor has no storage to resize
-  aten.mse_loss, aten.mse_loss_backward,
-  aten.smooth_l1_loss, aten.smooth_l1_loss_backward,
   aten.upsample_nearest2d.out,
-  # NOTE: only the "out" overload, the "vec" one is CompositeImplicitAutograd and overriding it loses the autograd kernel
-  aten.upsample_bicubic2d.out,
-  aten._adaptive_avg_pool2d,
   # activations
   aten.hardswish, aten.hardswish_backward,
   aten.hardtanh, aten.hardtanh_backward,
   aten.gelu, aten.gelu_backward,
-  # NOTE: no aten.logical_or here, its decomposition reaches aten.bitwise_or through a path that checks aliasing by
-  # reading storage, which a tiny tensor has none of. it gets a direct impl below instead
-  aten.logical_and, aten.logical_xor,
+  aten.logical_and,
   aten.randint,
   aten.eye,
   aten.hardsigmoid_backward,
@@ -440,7 +495,7 @@ simple_tensor_methods = [
   # reduce
   "all", "any", "argmax", "argmin", "cumsum", "cumprod",
   # complex
-  "linspace"]
+  "avg_pool2d", "linspace"]
 
 tiny_backend_out = {**{f"aten.{x}.out":getattr(Tensor,x) for x in simple_tensor_methods}, **{
   "aten.add.out": lambda input,other,alpha=1: input+alpha*other,
@@ -485,8 +540,6 @@ tiny_backend_out = {**{f"aten.{x}.out":getattr(Tensor,x) for x in simple_tensor_
   "aten.where.self_out": Tensor.where,
   "aten.prod.int_out": Tensor.prod,
   "aten.scatter.src_out": Tensor.scatter,
-  "aten.scatter_add.out": lambda self,dim,index,src: src if self.shape == () else Tensor.scatter_reduce(self, dim, index, src, reduce="sum"),
-  "aten.isin.Tensor_Tensor_out": lambda x,y,assume_unique=False,invert=False: (x.unsqueeze(-1)==y.flatten()).any(-1) != invert,
   # NOTE: axis=[] in torch means all, change tinygrad?
   "aten.sum.IntList_out": lambda self,axis,keepdim=False,dtype=None:
     self.sum(axis if axis is None or len(axis) else None, keepdim,
@@ -502,9 +555,10 @@ def wrap_out(f):
     assert out.shape == assigned.shape, f"shape mismatch: {assigned.shape} -> {out.shape}"
     assert out.device == assigned.device or out.device is None or assigned.device is None, f"device mismatch: {assigned.device} -> {out.device}"
     assert out.dtype == assigned.dtype, f"dtype mismatch: {assigned.dtype} -> {out.dtype}"
-    # writing out= is an in-place write like any other: through the base if it is a view, refreshing any derived views
-    _apply_inplace(out, assigned)
-    return out
+    # an out= that is a view has to be written through its base, and _apply_inplace gives a deviceless base its buffer first
+    if canonical_base(out) is not out: return _apply_inplace(out, assigned) or out
+    if out.device is None and assigned.device is not None: out.replace(out.empty_like(device=assigned.device))
+    return out.assign(assigned)
   return _wrap_out
 
 def _inplace_op(t, new_value):
@@ -512,14 +566,7 @@ def _inplace_op(t, new_value):
   else: _apply_inplace(t, new_value)
   return t
 
-# the three arange overloads are one function at different arity, and dtype/layout/device/pin_memory are keyword only in all of them
-def _arange(*args, dtype=None, **_):
-  return Tensor.arange(*args, dtype=_from_torch_dtype(dtype or (torch.get_default_dtype() if any(isinstance(x, float) for x in args) else torch.int64)))
-
-def _empty(size, dtype=None, device=None, **_):
-  return Tensor.empty(*size, dtype=_from_torch_dtype(dtype or torch.get_default_dtype()), device=_from_torch_device(device))
-
-tiny_backend = {**tiny_backend_out, **{
+tiny_backend = {**{k:wrap_out(v) for k,v in tiny_backend_out.items()}, **{
   "aten.remainder.Scalar_Tensor": lambda x,y: x%y,
   "aten.floor_divide": lambda x,y: x//y,
   "aten.floor_divide_.Tensor": lambda x,y: x//y,
@@ -532,8 +579,8 @@ tiny_backend = {**tiny_backend_out, **{
   # inplace ops using replace for fusion
   "aten.zero_": lambda x: x.const_like(0),
   "aten.fill_.Scalar": lambda x, y: x.const_like(y),
-  "aten.add_.Tensor": lambda self, other, alpha=1: self + other * alpha,
-  "aten.add_.Scalar": lambda self, other, alpha=1: self + other * alpha,
+  "aten.add_.Tensor": lambda self, other, alpha=1.0: self + other * alpha,
+  "aten.add_.Scalar": lambda self, other, alpha=1.0: self + other * alpha,
   "aten.mul_.Tensor": lambda self, other: self * other,
   "aten.mul_.Scalar": lambda self, other: self * other,
   # relu doesn't have an out form?
@@ -566,9 +613,7 @@ tiny_backend = {**tiny_backend_out, **{
   # these don't work in out form, they have size 0
   "aten.abs": Tensor.abs,
   "aten.logical_not": Tensor.logical_not,
-  # compare against zero first: logical_* is bool-valued for any input dtype, while | is bitwise
-  "aten.logical_or": lambda x, y: (x != 0) | (y != 0),
-  "aten.logical_or_": lambda x, y: (x != 0) | (y != 0),
+  "aten.logical_or_": lambda x, y: x | y,
   "aten.multinomial": Tensor.multinomial,
   "aten.masked_fill_.Scalar": lambda self, mask, value: self.masked_fill(mask, value),
   "aten.masked_fill_.Tensor": lambda self, mask, value: self.masked_fill(mask, value),
@@ -577,7 +622,14 @@ tiny_backend = {**tiny_backend_out, **{
   "aten.masked_select": Tensor.masked_select,
   "aten.all": Tensor.all,
   "aten.sgn": Tensor.sign,
+  "aten.acos": Tensor.acos,
   "aten.any": Tensor.any,
+  "aten.bitwise_not": Tensor.bitwise_not,
+  "aten.argmax": Tensor.argmax,
+  "aten.argmin": Tensor.argmin,
+  "aten.asinh": Tensor.asinh,
+  "aten.mul": Tensor.mul,
+  "aten.atanh": Tensor.atanh,
   "aten.fill_.Tensor": lambda self, value: self.const_like(value.reshape(()).item()),
   "aten.flip": Tensor.flip,
   "aten.scatter_reduce.two": Tensor.scatter_reduce,
@@ -588,22 +640,10 @@ tiny_backend = {**tiny_backend_out, **{
   "aten.add.Tensor": lambda input,other,alpha=1: input+alpha*other,
   "aten.linspace": lambda start, stop, steps, dtype=None, **kwargs:
     Tensor.linspace(start, stop, steps, **({"dtype": _from_torch_dtype(dtype)} if dtype is not None else {})),
-  # the functional copy_. without an impl the fallback segfaults on a tensor with no storage
-  "aten.copy": lambda self,src,non_blocking=False: src.cast(self.dtype).to(self.device).expand(self.shape),
-  "aten.arange": lambda end, **kwargs: _arange(0, end, **kwargs),
-  "aten.arange.start": _arange,
-  "aten.arange.start_step": _arange,
-  # empty_strided takes the strides and drops them: we always allocate contiguous
-  "aten.empty_strided": lambda size, stride, **kwargs: _empty(size, **kwargs),
-  "aten.empty.memory_format": _empty,
-  # TODO: supprt stride [] in tinygrad?
-  "aten.max_pool2d_with_indices": lambda self,kernel_size,stride=None,padding=0,dilation=1,ceil_mode=False: ((r:=Tensor.max_pool2d(self, kernel_size, stride or None, dilation, padding, ceil_mode, return_indices=True))[0], r[1].cast(dtypes.int64)),
-  "aten.max_pool2d_with_indices_backward": lambda grad_out,self,kernel_size,stride=None,padding=0,dilation=1,ceil_mode=False,indices=None: Tensor.max_unpool2d(grad_out, indices, output_size=self.shape),
-  "aten.max_unpool2d": lambda self,indices,output_size: Tensor.max_unpool2d(self, indices, output_size=output_size),
-  "aten._linalg_svd": lambda self,full_matrices=False: Tensor.svd(self, full_matrices),
   "aten.topk": Tensor.topk,
   "aten.constant_pad_nd": lambda self, padding, value=0.0: self.pad(padding, mode="constant", value=value).contiguous(),
-  "aten.cumsum": lambda self, dim: self.cumsum(dim),
+  # TODO: input contiguous is needed to prevent CFGContext circular dependency assertion for shapes >512 (see test_cumsum_arange_large)
+  "aten.cumsum": lambda self, dim: self.contiguous().cumsum(dim),
   "aten.logsumexp": lambda self, axis, keepdim=False: self.logsumexp(axis[0], keepdim=keepdim),
   "aten.roll": Tensor.roll,
   "aten.logcumsumexp": Tensor.logcumsumexp,
@@ -612,7 +652,6 @@ tiny_backend = {**tiny_backend_out, **{
     self.ones_like(**{k: v for k, v in {"dtype": _from_torch_dtype(dtype) if dtype else None,
                                         "device": _from_torch_device(device) if device else None}.items() if v is not None}),
   "aten.max.dim": lambda self, dim, keepdim=False: (self.max(dim, keepdim), self.argmax(dim, keepdim).cast(dtype=dtypes.int64)),
-  "aten.min.dim": lambda self, dim, keepdim=False: (self.min(dim, keepdim), self.argmin(dim, keepdim).cast(dtype=dtypes.int64)),
   "aten.cummax": lambda self, dim: ((r := self.cummax(dim))[0], r[1].cast(dtypes.int64)),
   "aten.cummin": lambda self, dim: ((r := self.cummin(dim))[0], r[1].cast(dtypes.int64)),
   "aten.nonzero": Tensor.nonzero,
@@ -674,16 +713,15 @@ def wrap_inplace_view_op(f):
   return nf
 
 # the aten schema says how an op is called: an inplace view retargets the view, a writable first arg is inplace,
-# and a writable out arg gets wrap_out's dtype cast, shape assert, and view write-through
+# and a writable out arg must have come from tiny_backend_out so that wrap_out was applied
 for k,v in tiny_backend.items():
   name, _, overload = k.removeprefix("aten.").partition(".")
   op = getattr(getattr(aten, name), overload or "default")
   writes = [a.name for a in op._schema.arguments if a.alias_info is not None and a.alias_info.is_write]
   if torch.Tag.inplace_view in op.tags: fxn = wrap_inplace_view_op(v)
   elif writes == [op._schema.arguments[0].name] and op._schema.returns: fxn = wrap_inplace(v)
-  elif not writes: fxn = wrap_fxn(k, v)
-  elif writes == ["out"]: fxn = wrap_fxn(k, wrap_out(v))
-  else: raise RuntimeError(f"{k} writes {writes}: unhandled writable arg in schema")
+  elif not writes or (writes == ["out"] and k in tiny_backend_out): fxn = wrap_fxn(k, v)
+  else: raise RuntimeError(f"{k} writes {writes}: expected an inplace first arg, or an out arg with {k} in tiny_backend_out")
   torch.library.impl(k.replace("aten.", "aten::"), "privateuseone")(fxn)
 
 @torch.library.impl("aten::equal", "privateuseone")

extra/torch_backend/test.py

@@ -83,12 +83,6 @@ class TestTorchBackend(unittest.TestCase):
     torch.add(torch.ones(5, device=device), torch.ones(5, device=device), out=a)
     self.assertEqual(a.detach().storage_offset(), 3)
 
-  def test_out_refreshes_views_of_base(self):
-    a = torch.zeros(4, device=device)
-    v = a[2:]
-    torch.add(torch.ones(4, device=device), torch.ones(4, device=device), out=a)
-    np.testing.assert_equal(v.cpu().numpy(), [2., 2.])
-
   @unittest.expectedFailure  # TODO: storage offset assumes a contiguous source, use UOp.contiguous_view_offset
   def test_storage_offset_non_contiguous_source(self):
     a = torch.arange(12., device=device).reshape(3,4)
@@ -172,15 +166,6 @@ class TestTorchBackend(unittest.TestCase):
     expected = np.array([[1.5, 5.2, 9.0], [13.2, 17.1, 18.4]], dtype=np.float32)
     np.testing.assert_equal(y3.cpu().numpy(), expected)
 
-  def test_argmax_argmin(self):
-    a = torch.arange(12, dtype=torch.float32, device=device).reshape(3, 4)
-    c = a.cpu()
-    for got, want in [(a.argmax(), c.argmax()), (a.argmin(0), c.argmin(0)), (a.argmax(1, keepdim=True), c.argmax(1, keepdim=True)),
-                      (torch.min(a, 1).indices, torch.min(c, 1).indices), (torch.max(a, 1).indices, torch.max(c, 1).indices),
-                      (torch.min(a, 1).values, torch.min(c, 1).values), (torch.min(a, 1, keepdim=True).indices, torch.min(c, 1, keepdim=True).indices)]:
-      self.assertEqual(got.dtype, want.dtype)  # torch's arg reduces are int64, tinygrad's are int32
-      np.testing.assert_equal(got.cpu().numpy(), want.numpy())
-
   def test_isfinite(self):
     a = torch.ones(4, device=device)
     np.testing.assert_equal(torch.isfinite(a).cpu().numpy(), [True, True, True, True])
@@ -388,22 +373,6 @@ class TestTorchBackend(unittest.TestCase):
     for bwd_eps in [1e-5, 0.3]:
       for got, want in zip(run(device, bwd_eps), run("cpu", bwd_eps)): np.testing.assert_allclose(got, want, atol=1e-4, rtol=1e-3)
 
-  def test_groupnorm_backward(self):
-    def run(dev):
-      x = torch.arange(24., device=dev).reshape(2, 4, 3).requires_grad_()
-      w = torch.linspace(0.5, 2.0, 4).to(dev).requires_grad_()
-      torch.nn.functional.group_norm(x, 2, w, torch.zeros(4, device=dev)).square().sum().backward()
-      return x.grad.cpu().numpy(), w.grad.cpu().numpy()
-    for got, want in zip(run(device), run("cpu")): np.testing.assert_allclose(got, want, atol=1e-4, rtol=1e-3)
-
-  def test_mse_smooth_l1_loss_backward(self):
-    def run(dev, loss):
-      x = torch.arange(4., device=dev).requires_grad_()
-      loss(x, torch.ones(4, device=dev)).backward()
-      return x.grad.cpu().numpy()
-    for loss in [torch.nn.functional.mse_loss, torch.nn.functional.smooth_l1_loss]:
-      np.testing.assert_allclose(run(device, loss), run("cpu", loss), atol=1e-6)
-
   def test_batchnorm_unsqueeze(self):
     bn = torch.nn.BatchNorm2d(4).to(device)
     x = torch.randn(8, 4, 3, 3, device=device)
@@ -547,15 +516,6 @@ class TestTorchBackend(unittest.TestCase):
     cpu_res = torch.arange(20, dtype=torch.float32)[::2][1:4].numpy()
     np.testing.assert_equal(torch_res, cpu_res)
 
-  def test_select_out_of_range_dim(self):
-    a = torch.arange(12, dtype=torch.int32, device=device).reshape(3, 4)
-    with self.assertRaises(IndexError): a.select(5, 0)
-
-  def test_select_collapses_the_only_dim(self):
-    a = torch.arange(3, dtype=torch.int32, device=device)
-    self.assertEqual(a.select(0, 1).shape, ())
-    np.testing.assert_equal(a.select(0, 1).cpu().numpy(), 1)
-
   def test_slice_negative_dim(self):
     a = torch.arange(13, dtype=torch.int32, device=device).repeat(8, 1)
     torch_chunks = a.chunk(3, -1)
@@ -836,86 +796,6 @@ class TestTorchBackend(unittest.TestCase):
     np.testing.assert_allclose(w_tiny.grad.cpu().numpy(), w_cpu.grad.numpy(), atol=1e-4, rtol=1e-3)
     np.testing.assert_allclose(b_tiny.grad.cpu().numpy(), b_cpu.grad.numpy(), atol=1e-4, rtol=1e-3)
 
-  def test_write_through_detach_of_unrealized(self):
-    a = torch.empty(4, device=device)
-    a.detach().fill_(3)
-    np.testing.assert_equal(a.cpu().numpy(), [3, 3, 3, 3])
-
-  def test_square_transpose_inplace(self):
-    # a same-shape transpose is not a reshape: writing the transposed values straight back would scramble the base
-    a = torch.tensor([[0., 1., 2.], [3., 4., 5.], [6., 7., 8.]], device=device)
-    a.transpose(0, 1).add_(100)
-    np.testing.assert_equal(a.cpu().numpy(), [[100., 101., 102.], [103., 104., 105.], [106., 107., 108.]])
-
-  def test_interpolate(self):
-    a = torch.arange(4, dtype=torch.float32, device=device).reshape(1, 1, 2, 2)
-    nearest = torch.nn.functional.interpolate(a, scale_factor=2.0)
-    np.testing.assert_equal(nearest.cpu().numpy()[0, 0], [[0, 0, 1, 1], [0, 0, 1, 1], [2, 2, 3, 3], [2, 2, 3, 3]])
-    linear = torch.nn.functional.interpolate(a, size=(4, 4), mode="bilinear", align_corners=False)
-    ref = torch.nn.functional.interpolate(a.cpu(), size=(4, 4), mode="bilinear", align_corners=False)
-    np.testing.assert_allclose(linear.cpu().numpy(), ref.numpy(), rtol=1e-5)
-
-  def test_interpolate_bicubic_area(self):
-    a = torch.arange(32, dtype=torch.float32, device=device).reshape(1, 2, 4, 4)
-    for mode, scale in [("bicubic", 2.0), ("area", 0.5)]:
-      ref = torch.nn.functional.interpolate(a.cpu(), scale_factor=scale, mode=mode)
-      np.testing.assert_allclose(torch.nn.functional.interpolate(a, scale_factor=scale, mode=mode).cpu().numpy(), ref.numpy(), atol=1e-4)
-
-  @unittest.expectedFailure
-  def test_interpolate_bicubic_backward(self):
-    # the forward comes from a decomposition, but aten::upsample_bicubic2d_backward has none (nor does
-    # aten::_adaptive_avg_pool2d_backward, for area), so training through these modes needs a real kernel
-    x = torch.arange(32., dtype=torch.float32, device=device).reshape(1, 2, 4, 4).requires_grad_()
-    torch.nn.functional.interpolate(x, scale_factor=2.0, mode="bicubic").sum().backward()
-
-  @unittest.expectedFailure
-  def test_interpolate_inexact_scale(self):
-    # torch forwards the raw scale_factor, Tensor.interpolate recomputes it from output_size, and they disagree here
-    a = torch.arange(6, dtype=torch.float32, device=device).reshape(1, 1, 2, 3)
-    tiny = torch.nn.functional.interpolate(a, scale_factor=2.5, mode="bilinear")
-    ref = torch.nn.functional.interpolate(a.cpu(), scale_factor=2.5, mode="bilinear")
-    np.testing.assert_allclose(tiny.cpu().numpy(), ref.numpy(), rtol=1e-5)
-
-  def test_logical_or_xor(self):
-    a = torch.tensor([True, True, False, False], device=device)
-    b = torch.tensor([True, False, True, False], device=device)
-    np.testing.assert_equal(torch.logical_or(a, b).cpu().numpy(), [True, True, True, False])
-    np.testing.assert_equal(torch.logical_xor(a, b).cpu().numpy(), [False, True, True, False])
-    # bool-valued whatever the input dtype, so this is not | and ^
-    i, j = torch.tensor([2, 0, 5, 0], device=device), torch.tensor([0, 0, 1, 1], device=device)
-    np.testing.assert_equal(torch.logical_or(i, j).cpu().numpy(), [True, False, True, True])
-    np.testing.assert_equal(torch.logical_xor(i, j).cpu().numpy(), [True, False, False, True])
-
-  def test_slice_scatter(self):
-    # the scatters are functional: they return a new tensor and must leave the one they were given alone
-    a = torch.arange(12, dtype=torch.float32, device=device).reshape(3, 4)
-    out = torch.slice_scatter(a, torch.ones(1, 4, device=device), 0, 0, 1)
-    np.testing.assert_equal(out.cpu().numpy(), [[1, 1, 1, 1], [4, 5, 6, 7], [8, 9, 10, 11]])
-    np.testing.assert_equal(a.cpu().numpy(), np.arange(12, dtype=np.float32).reshape(3, 4))
-
-  def test_slice_scatter_casts_src(self):
-    a = torch.zeros(3, 4, device=device)
-    out = torch.slice_scatter(a, torch.ones(1, 4, dtype=torch.int32, device=device), 0, 0, 1)
-    self.assertEqual(out.dtype, torch.float32)
-    np.testing.assert_equal(out.cpu().numpy()[0], np.ones(4, dtype=np.float32))
-
-  def test_select_scatter(self):
-    a = torch.arange(12, dtype=torch.float32, device=device).reshape(3, 4)
-    out = torch.select_scatter(a, torch.ones(4, device=device), 0, 1)
-    np.testing.assert_equal(out.cpu().numpy(), [[0, 1, 2, 3], [1, 1, 1, 1], [8, 9, 10, 11]])
-
-  def test_diagonal_scatter(self):
-    a = torch.zeros(3, 3, device=device)
-    out = torch.diagonal_scatter(a, torch.arange(3, dtype=torch.float32, device=device))
-    np.testing.assert_equal(out.cpu().numpy(), np.diag([0., 1., 2.]))
-    np.testing.assert_equal(a.cpu().numpy(), np.zeros((3, 3), dtype=np.float32))
-
-  def test_copy_functional(self):
-    # without an impl this segfaults rather than fails: a regression here takes the whole run down
-    a = torch.arange(4, dtype=torch.float32, device=device)
-    out = torch.ops.aten.copy(a, torch.zeros(4, device=device))
-    np.testing.assert_equal(out.cpu().numpy(), [0., 0., 0., 0.])
-    np.testing.assert_equal(a.cpu().numpy(), [0., 1., 2., 3.])
 
 from tinygrad import Tensor
 class TestBackendHelpers(unittest.TestCase):

test/backend/test_const_folding.py

@@ -188,6 +188,7 @@ class TestTautologicalCompare(unittest.TestCase):
     np.testing.assert_equal((Tensor(True) < Tensor(False)).numpy(), False)
     np.testing.assert_equal((Tensor(True) < Tensor(True)).numpy(), False)
 
+  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "WEBGPU doesn't support NaN comparison correctly")
   def test_a_eq_a(self):
     # self eq is always true for int or bool
     a = Tensor([1, 2, 3])

test/backend/test_custom_kernel.py

@@ -422,8 +422,9 @@ 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
+    # TODO: subbuffer usage has runtime specific behavior, this will be fixed after the removal of SLICE.
+    assert_kernel_count(2 if y.device in ("CL", "WEBGPU") else 1)
     self.assertEqual(y.tolist(), [1, 2, 3, 4])
 
   @Context(DEV="CPU")

test/backend/test_dtype.py

@@ -340,9 +340,6 @@ class TestUint64DType(TestDType):
   DTYPE = dtypes.uint64
   def test_uint64_load(self):
     assert Tensor(2**64 - 1, dtype=dtypes.uint64).numpy() == 2**64 - 1
-  @unittest.skipIf(dtypes.double not in supported_dtypes, "needs float64")
-  def test_uint64_cast_double(self):
-    assert Tensor([2**32 + 1], dtype=dtypes.uint64).cast(dtypes.double).numpy() == 2**32 + 1
 
 @unittest.skipIf(isinstance(Device[Device.DEFAULT].renderer, PTXRenderer), "PTX does indexing math with longs")
 class TestEmulatedUInt64DType(TestUint64DType):

test/backend/test_isel.py

@@ -7,7 +7,7 @@ from tinygrad.renderer.isa.x86 import X86Renderer, X86Ops
 from tinygrad.renderer.isa import IselContext
 
 # INDEX on a register value with a constant index extracts a single element (the old GEP)
-def lane(y:UOp, i:int) -> UOp: return y.index(UOp.const(i, dtypes.int), dtype=y.dtype)
+def lane(y:UOp, i:int) -> UOp: return y.index(UOp.const(i, dtypes.int), dtype=y.dtype.scalar())
 
 @unittest.skipUnless(isinstance(Device[Device.DEFAULT].renderer, X86Renderer), "only x86")
 class TestIselX86(unittest.TestCase):

test/backend/test_jit.py

@@ -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")

test/backend/test_linearizer.py

@@ -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)
@@ -252,7 +252,7 @@ class TestLinearizer(unittest.TestCase):
     for u in uops:
       if u.op is Ops.STORE and u.src[0].addrspace is AddrSpace.REG:
         if uops.index(u) < begin_range:
-          assert u.src[1].op not in GroupOp.ALU
+          assert u.src[1].op is Ops.CONST
         else:
           assert u.src[1].op in GroupOp.ALU
           assert begin_range < uops.index(u) < end_range
@@ -261,7 +261,6 @@ class TestLinearizer(unittest.TestCase):
         assert end_range < uops.index(u)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
-  @unittest.skipIf(Device[Device.DEFAULT].renderer.casted_consts, "reads a literal, which is casted here. TODO: flip this")
   def test_default_global_reversed(self):
     # shrink so that the dims do not collapse
     t = Tensor.ones(5, 6, 7).contiguous().realize().shrink(((0, 4), (0, 5), (0, 6)))
@@ -412,7 +411,7 @@ 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)]
@@ -438,7 +437,7 @@ 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]
@@ -460,7 +459,9 @@ 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)
 

test/backend/test_multitensor.py

@@ -1,9 +1,9 @@
 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
@@ -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()

test/backend/test_ops.py

@@ -720,11 +720,10 @@ class TestOps(unittest.TestCase):
       return torch.autograd.grad(t ** c, t)[0].item()
     for x in [-math.inf, 0, 1, math.inf]:
       for c in [-1, 0, 0.3, 1, 2]:
-        torch_out = get_torch_gradient(x, c)
-        # the pow backward routes through exp2/log2, whose 0/inf behavior is undefined on WEBGPU
-        if Device.DEFAULT == "WEBGPU" and not math.isfinite(torch_out): continue
         tiny_out = get_tiny_gradient(x, c)
+        torch_out = get_torch_gradient(x, c)
         if math.isnan(tiny_out):
+          if Device.DEFAULT == "WEBGPU": continue # TODO: WEBGPU issue with nan
           assert math.isnan(torch_out)
         else:
           self.assertAlmostEqual(tiny_out, torch_out, msg=f"{x}, {c}")
@@ -750,7 +749,6 @@ class TestOps(unittest.TestCase):
   def test_exp2_log2_zero_times_negative(self):
     # gallivm's exp2/log2 have "undefined behavior with infs, 0s and nans", so exp2(log2(0)*y) returns 0 instead of inf
     helper_test_op(None, lambda x,y: (x.log2()*y).exp2(), lambda x,y: (x.log2()*y).exp2(), vals=[[0.0], [-0.7]], forward_only=True)
-  @unittest.skipIf(Device.DEFAULT == "WEBGPU", "pow at 0 routes through exp2/log2, whose 0/inf behavior is undefined on WEBGPU")
   def test_pow_zero_const(self):
     helper_test_op(None, lambda x: x**0.3, vals=[[0.0]])
     helper_test_op(None, lambda x: x**0.0, vals=[[0.0]])

test/backend/test_pickle.py

@@ -2,7 +2,7 @@ import unittest, pickle, types, tracemalloc
 import numpy as np
 from tinygrad import Tensor, Device, TinyJit, Variable, dtypes
 from tinygrad.helpers import GlobalCounters, ContextVar, Context, DEV
-from tinygrad.uop.ops import PatternMatcher, UPat, UOp, deconstruct_function
+from tinygrad.uop.ops import PatternMatcher, UPat, UOp
 
 class TestPickle(unittest.TestCase):
   def test_pickle_code_object(self):
@@ -11,11 +11,6 @@ class TestPickle(unittest.TestCase):
     fxn = types.FunctionType(pickle.loads(code_str), globals())
     self.assertEqual(fxn(2), 4)
 
-  def test_deconstruct_function_nested_comprehension(self):
-    # pre PEP 709, each comprehension is its own code object, so dtypes here is referenced two code objects deep
-    def fxn(): return [[dtypes.int for _ in range(2)] for _ in range(2)]
-    self.assertEqual(types.FunctionType(*deconstruct_function(fxn))(), fxn())
-
   def test_pickle_pattern_matcher(self):
     pm = PatternMatcher([(UPat.cvar('x'), lambda x: x*2)])
     sink = UOp.const(2)

test/backend/test_quantize_onnx.py

@@ -82,7 +82,7 @@ class TestQuantizeOnnxCPU(unittest.TestCase):
       linear = run_onnx({"input":inp})["output"].schedule_linear()
       prg = to_program(linear.src[-2].src[0], renderer=Device[Device.DEFAULT].renderer)
       daccs = [u for u in tuple(prg.src[1].src) if u.op is Ops.BUFFER and u.addrspace is AddrSpace.REG]
-      assert all(u.dtype is dtypes.int for u in daccs)
+      assert all(u.dtype.scalar() is dtypes.int for u in daccs)
 
 @unittest.skipIf(Device.DEFAULT != "DSP", "only tests for DSP")
 class TestQuantizeOnnx(unittest.TestCase):

test/backend/test_tensor.py

@@ -653,19 +653,9 @@ class TestZeroShapeTensor(unittest.TestCase):
     np.testing.assert_equal(Tensor([[1, 2]]).pad_to(2, 3).numpy(), [[1, 2, 0], [0, 0, 0]])
     np.testing.assert_equal(Tensor([[1, 2]]).pad_to(1, 3).numpy(), [[1, 2, 0]])
     np.testing.assert_equal(Tensor([[1, 2]]).pad_to(None, 3).numpy(), [[1, 2, 0]])
-    np.testing.assert_equal(Tensor([1, 2]).pad_to(4, value=2).numpy(), [1, 2, 2, 2])
-    np.testing.assert_equal(Tensor([[1, 2]]).pad_to(2, 3, value=-1).numpy(), [[1, 2, -1], [-1, -1, -1]])
-    np.testing.assert_equal(Tensor([1, 2]).pad_to(None, value=5).numpy(), [1, 2])  # no-op pad ignores the fill
     with self.assertRaises(ValueError): Tensor([1, 2]).pad_to(2, 3)
     with self.assertRaises(ValueError): Tensor([[1, 2]]).pad_to(3)
 
-  def test_max_shape(self):
-    from tinygrad import UOp
-    t = Tensor.empty(2, UOp.variable('v', 1, 32), 4)
-    self.assertEqual(t.max_shape, (2, 32, 4))
-    self.assertEqual(t.max_numel(), 2*32*4)
-    self.assertEqual(Tensor.empty(2, 3).max_shape, (2, 3))
-
   def test_shrink_into_zero(self):
     t = Tensor.rand(3, 4).realize()
     assert t.shrink((None, (2, 2))).realize().shape == (3, 0)

test/device/test_hcq2.py

@@ -1,22 +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))
-
-  def test_overlapping_device_tuples(self):
-    # an op on a wide device tuple followed by an op on an overlapping smaller tuple used to MMU-fault the smaller one
-    d4, d2 = tuple(f"{Device.DEFAULT}:{i}" for i in range(4)), tuple(f"{Device.DEFAULT}:{i}" for i in range(2))
-    ref = Tensor.arange(16).contiguous().realize()
-    Tensor(ref.uop.copy_to_device(d4)).realize()
-    out = Tensor.ones(8).shard(d2, axis=0).contiguous().realize()
-    np.testing.assert_equal(out.numpy(), np.ones(8))
-
-if __name__ == "__main__":
-  unittest.main()

test/external/external_benchmark_all2all.py

@@ -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()

test/external/external_test_schedule_scaling.py

@@ -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)

test/external/external_uop_gc.py

@@ -44,7 +44,7 @@ def realized_matmul():
   z = y.matmul(x)
   Tensor.realize(z)
 def realized_gradient():
-  x = Tensor.eye(3).clone()
+  x = Tensor.eye(3)
   y = Tensor([[2.0,0,-2.0]])
   z = y.matmul(x).sum()
   z.backward()

test/helpers.py

@@ -48,8 +48,6 @@ def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Te
   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:
@@ -59,6 +57,8 @@ def check_schedule(t:Tensor|list[Tensor]|UOp, allowed:int, to_prerealize:list[Te
         print("kernel", i+1)
         print(call.src[0])
     raise KernelCountException(allowed, kernel_cnt)
+  # test compiling the linear
+  compile_linear(linear)
   return linear, var_vals
 
 def assert_kernel_count(expected:int):
@@ -86,9 +86,7 @@ def assert_jit_cache_len(fxn, expected_len):
   if linear is None or not linear.src:
     if expected_len != 0: raise KernelCountException(expected_len, 0)
     return
-  if expected_len and all(call_is_hcq(call) for call in linear.src): # HCQ2: one batch submitter, or fence + reset + merged calls + finalizer
-    from tinygrad.runtime.support.hcq2 import HCQ_RUNTIME_DEV
-    expected_len = 1 if HCQ_RUNTIME_DEV.value == "CPU" else 4
+  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))
     inner = linear.src[0].src[0].src[0]  # LINEAR UOp inside CUSTOM_FUNCTION

test/mockgpu/amd/emu.py

@@ -260,6 +260,19 @@ def _cond(cond, if_true, if_false):
 def _cond_hi16(cond, val: UOp) -> UOp: return _cond(cond, _hi16(val), val)
 def _apply_opsel(val: UOp, sel_bit: int, opsel: int) -> UOp: return _hi16(val) if opsel & (1 << sel_bit) else val
 
+def _set_lane_bit(old: UOp, lane: UOp, val: UOp, exec_mask: UOp) -> UOp:
+  """Set/clear a single bit in a mask based on lane index, respecting exec mask."""
+  if old.dtype in (dtypes.uint64, dtypes.int64):
+    dt = dtypes.uint64
+    mask = UOp.const(1, dt) << lane.cast(dt)
+    new_bit = _to_u32(val).cast(dt) << lane.cast(dt)
+    cleared = old.cast(dt) & (mask ^ UOp.const(0xFFFFFFFFFFFFFFFF, dt))
+    return _lane_active(exec_mask, lane).where(cleared | new_bit, old.cast(dt))
+  mask = _c(1) << lane.cast(dtypes.uint32)
+  new_bit = _to_u32(val) << lane.cast(dtypes.uint32)
+  cleared = old & (mask ^ _c(MASK32))
+  return _lane_active(exec_mask, lane).where(cleared | new_bit, old)
+
 def _val_to_u32(val: UOp) -> UOp:
   """Convert any value to uint32 for storage (bitcast floats, cast ints)."""
   if val.dtype == dtypes.uint32: return val
@@ -519,19 +532,6 @@ class _Ctx:
       return [self.wsgpr_dyn(reg, lo), self.wsgpr_dyn(reg + _c(1), hi)]
     return [self.wsgpr_dyn(reg, val)]
 
-  def wmask_lane_bit(self, reg: UOp, lane: UOp, val: UOp, exec_mask: UOp) -> list[UOp]:
-    """Set/clear bit `lane` of the mask at `reg` from val for exec-active lanes, preserving memory for inactive lanes"""
-    active, bit = _lane_active(exec_mask, lane), _to_u32(val)
-    if self.wave_size <= 32:
-      old = self.rsgpr_dyn(reg)
-      mask = _c(1) << lane.cast(dtypes.uint32)
-      return [self.wsgpr_dyn(reg, active.where((old & (mask ^ _c(MASK32))) | (bit << lane.cast(dtypes.uint32)), old))]
-    off = (lane & _c(31, dtypes.int)).cast(dtypes.uint32)
-    mask = _c(1) << off
-    def half(old: UOp, sel: UOp) -> UOp: return sel.where(active.where((old & (mask ^ _c(MASK32))) | (bit << off), old), old)
-    return [self.wsgpr_dyn(reg, half(self.rsgpr_dyn(reg), lane < _c(32, dtypes.int))),
-            self.wsgpr_dyn(reg + _c(1), half(self.rsgpr_dyn(reg + _c(1)), _c(32, dtypes.int) <= lane))]
-
   def rmask(self, reg: UOp) -> UOp:
     """Read a lane mask (VCC/EXEC). Combines lo/hi for wave64."""
     if self.wave_size > 32: return _u64(self.rsgpr_dyn(reg), self.rsgpr_dyn(reg + _c(1)))
@@ -718,7 +718,9 @@ class _Ctx:
         raw_stores.append(('vgpr_direct', self.vgpr.index(val[0].valid(active)).store(new_val)))
         continue
       if 'D0' in dest and '[laneId]' in dest:
-        raw_stores.extend([('vcc', s) for s in self.wmask_lane_bit(_c(VCC_LO.offset), lane, val, exec_mask)])
+        old_vcc = self.rmask(_c(VCC_LO.offset))
+        new_vcc = _set_lane_bit(old_vcc, lane, val, exec_mask)
+        raw_stores.extend([('vcc', s) for s in self.wmask(_c(VCC_LO.offset), new_vcc)])
       elif dest.startswith('D0'):
         dest_suffix = re.match(r'D0\.(\w+)', dest)
         if dest_suffix is not None:
@@ -1037,11 +1039,13 @@ def _compile_sdwa(inst: irc.VOP1_SDWA | irc.VOP2_SDWA | irc.VOP2_SDWA_SDST | irc
         result = _sdwa_write(old, result, dst_sel, dst_unused)
       stores.append(ctx.wvgpr_dyn(vdst_reg, lane, result, exec_mask))
     elif dest.startswith('VCC'):
-      stores.extend(ctx.wmask_lane_bit(_c(VCC_LO.offset), lane, val, exec_mask))
+      old_vcc = ctx.rmask(_c(VCC_LO.offset))
+      stores.extend(ctx.wmask(_c(VCC_LO.offset), _set_lane_bit(old_vcc, lane, val, exec_mask)))
   if vcc_val is not None:
     # Initialize sdst to 0 before lane loop (old value may be unrelated data), then set lane bits in loop
     init_stores = [ctx.wsgpr_dyn(sdst_off, _c(0)), ctx.wsgpr_dyn(sdst_off + _c(1), _c(0))]
-    stores.extend(ctx.wmask_lane_bit(sdst_off, lane, vcc_val, exec_mask))
+    old_sdst = ctx.rmask(sdst_off)
+    stores.extend(ctx.wmask(sdst_off, _set_lane_bit(old_sdst, lane, vcc_val, exec_mask)))
     if stores:
       return UOp.sink(*init_stores, UOp.sink(*stores).end(lane), *ctx.inc_pc())
     return UOp.sink(*init_stores, *ctx.inc_pc())

test/models/test_whisper.py

@@ -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)

test/null/test_const_folding.py

@@ -1,6 +1,6 @@
 import unittest, itertools, math
 from tinygrad import Tensor, dtypes, Context
-from tinygrad.dtype import DType, ConstType
+from tinygrad.dtype import DType, ConstType, truncate
 from tinygrad.uop.ops import Ops, UOp
 from test.helpers import full_rewrite
 import numpy as np
@@ -51,13 +51,16 @@ class TestWeakConstFolding(unittest.TestCase):
   def test_invalid_poison(self):
     self.assertTrue(UOp.invalid().alu(Ops.CDIV, UOp.const(0)).simplify().is_invalid)
 
-  def test_single_rounding_log10_backward(self):
-    # log10 backward folds log10(2)/log(2) = 1/log(10) in one rounding, not the double-rounded 1/float32(log(10))
-    x = Tensor([1.0, 2.0, 3.0])
-    ast = next(s.src[0] for s in x.log10().sum().gradient(x)[0].schedule_linear().src if s.src[0].op is Ops.SINK)
-    const = next(u.arg for u in full_rewrite(ast).toposort() if u.op is Ops.CONST and u.dtype is dtypes.float32)
-    # correctly rounded: within half a float32 ulp of the exact value (folding at float32 lands 0.66 ulp off)
-    self.assertLess(abs(const - 1/math.log(10)), 2**-26)
+  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):

test/null/test_dtype_spec.py

@@ -51,6 +51,10 @@ class TestHelpers(unittest.TestCase):
     assert dtypes.is_float(dtypes.fp8e4m3)
     assert dtypes.is_float(dtypes.fp8e5m2)
 
+  @given(strat.sampled_from([d for d in DTYPES_DICT.values() if dtypes.is_float(d) or dtypes.is_int(d)]))
+  def test_scalar(self, dtype):
+    assert dtype.scalar() == dtype
+
   def test_from_py(self):
     assert dtypes.from_py(True) == dtypes.bool
     assert dtypes.from_py(Invalid) == dtypes.bool

test/null/test_graph_rewrite.py

@@ -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

test/null/test_kimi.py

@@ -0,0 +1,30 @@
+import unittest
+from tinygrad import Tensor, dtypes, nn
+from tinygrad.llm.kimi import _shard_kimi
+from tinygrad.llm.model import SSMConfig, Transformer, TransformerConfig
+
+class TestKimiTP4(unittest.TestCase):
+  def test_prefill_and_decode_graph(self):
+    devices = ("NULL:0", "NULL:1", "NULL:2", "NULL:3")
+    config = TransformerConfig(num_blocks=2, dim=32, hidden_dim=128, n_heads=4, n_kv_heads=1, norm_eps=1e-5,
+      vocab_size=64, head_dim=12, rope_theta=10000, rope_dim=4, v_head_dim=8, max_context=4, kv_lora_rank=16,
+      num_experts=8, num_experts_per_tok=2, norm_topk_prob=True, shared_expert_dim=32, ssm_layers=(True, False),
+      ssm=SSMConfig(4, 8, 4, 4, 32, True), shared_expert_gate=False, leading_dense_blocks=1, dense_hidden_dim=64,
+      routed_scaling_factor=2.446, expert_bias=True, expert_mxfp4=True, bf16_activations=True, kda_split_qkv=True)
+    model = Transformer(config)
+    for name, value in nn.state.get_state_dict(model).items():
+      fill = 127 if name.endswith("weight_scale") else 0
+      value.replace(Tensor.full(value.shape, fill, dtype=value.dtype if value.dtype is dtypes.uint8 else dtypes.bfloat16, device="NULL"))
+    _shard_kimi(model, devices)
+
+    temperature = Tensor([0.0], device=devices)
+    prefill = model(Tensor([[1, 2]], dtype=dtypes.int32, device=devices), 0, temperature).realize()
+    model(Tensor([[1, 2]], dtype=dtypes.int32, device=devices), 0, temperature).realize()  # replay prefill JIT
+    decode = model(Tensor([[3]], dtype=dtypes.int32, device=devices), 2, temperature).realize()
+    model(Tensor([[4]], dtype=dtypes.int32, device=devices), 3, temperature).realize()  # replay decode JIT
+    self.assertEqual(prefill.shape, (1, 1))
+    self.assertEqual(decode.shape, (1, 1))
+    self.assertEqual(model.blk[0].recurrent_state.uop.axis, 1)
+    self.assertEqual(model.blk[1].cache_k.dtype, dtypes.bfloat16)
+
+if __name__ == "__main__": unittest.main()

test/null/test_kimi_k3.py

@@ -0,0 +1,29 @@
+import unittest
+from tinygrad import Tensor, dtypes, nn
+from tinygrad.llm.kimi_k3 import _shard_kimi_k3
+from test.unit.test_llm_k3 import small_k3_config
+from tinygrad.llm.model import Transformer
+
+class TestKimiK3TP8(unittest.TestCase):
+  @staticmethod
+  def _model():
+    model = Transformer(small_k3_config())
+    for name,value in nn.state.get_state_dict(model).items():
+      fill = 127 if name.endswith("weight_scale") else 0
+      dtype = value.dtype if value.dtype is dtypes.uint8 else dtypes.bfloat16
+      value.replace(Tensor.full(value.shape, fill, dtype=dtype, device="NULL"))
+    _shard_kimi_k3(model, tuple(f"NULL:{i}" for i in range(8)))
+    return model
+
+  def test_prefill_decode_and_jit_replay(self):
+    devices = tuple(f"NULL:{i}" for i in range(8))
+    model = self._model()
+    temperature = Tensor([0.0], device=devices)
+    self.assertEqual(model(Tensor([[1, 2]], dtype=dtypes.int32, device=devices), 0, temperature).realize().shape, (1, 1))
+    model(Tensor([[1, 2]], dtype=dtypes.int32, device=devices), 0, temperature).realize()
+    self.assertEqual(model(Tensor([[3]], dtype=dtypes.int32, device=devices), 2, temperature).realize().shape, (1, 1))
+    model(Tensor([[4]], dtype=dtypes.int32, device=devices), 3, temperature).realize()
+    self.assertEqual(model.blk[0].recurrent_state.uop.axis, 1)
+    self.assertEqual(model.blk[1].cache_k.dtype, dtypes.bfloat16)
+
+if __name__ == "__main__": unittest.main()

test/null/test_memory_planner.py

@@ -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:

test/null/test_simplify_valid_idx.py

@@ -25,7 +25,7 @@ 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):

test/null/test_tensor.py

@@ -69,7 +69,7 @@ class TestIdxUpcast(unittest.TestCase):
     if not isinstance(Device[Device.DEFAULT].renderer, (PTXRenderer, NIRRenderer)):
       assert idx.op is Ops.INDEX
       idx_val = idx.src[1]
-      self.assertFalse(idx_val.overflows(idx_val.dtype))
+      self.assertFalse(idx_val.overflows(idx_val.dtype.scalar()))
 
   # use expand to generate kernel that uses large idx
   def do_op_then_assert(self, dtype: DType, dim1, dim2, dim3):

test/null/test_uop_graph.py

@@ -157,7 +157,7 @@ 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 +339,7 @@ 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

test/null/test_uop_symbolic.py

@@ -3,10 +3,10 @@ import unittest, pickle, functools, math
 import z3
 
 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, commutative, pm_simplify_valid, pm_move_where_on_load
-from tinygrad.uop.weak import pm_cast_weak
+from tinygrad.uop.symbolic import sym, pm_fold_cast_const, 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_cast_weak, name="simplify symbolic uop")
+    v_simplified = graph_rewrite(v, sym+pm_fold_cast_const, 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):
@@ -1014,11 +1008,20 @@ class TestSymbolic(unittest.TestCase):
     self.helper_test_variable(-a<-b, False, True, "(b<a)")
 
   def test_where_cast(self):
-    cond = Variable("s", 0, 3, dtypes.int) < 2
+    s = Variable("s", 0, 3, dtypes.int)
+    cond = s < 2
     a = Variable("a", 0, 3, dtypes.int)
-    self.assertIs(graph_rewrite(cond.where(a, a+1).cast(dtypes.half), sym), cond.where(a.cast(dtypes.half), (a+1).cast(dtypes.half)))
-    self.assertIs(graph_rewrite(cond.where(a, uconst(2)).cast(dtypes.half), sym), cond.where(a.cast(dtypes.half), UOp.const(2, dtypes.half)))
-    self.assertIs(graph_rewrite(cond.where(a, UOp.invalid()).cast(dtypes.half), sym), cond.where(a.cast(dtypes.half), UOp.invalid()))
+    b = Variable("b", 0, 3, dtypes.int)
+    expr = cond.where(a, b).cast(dtypes.half)
+
+    # TODO: copied from render, render does not support cast
+    glbl = UOp.param(0, dtypes.int, (1,))
+    uops = get_uops(UOp(Ops.STORE, src=(glbl.index(UOp.const(0, dtypes.int)), expr)).sink())
+    rewritten_uop = [uop for uop in uops if uop.op is Ops.STORE][0].src[1]
+
+    # the vars are now scalar PARAMs
+    pvar = {u.expr: u for u in rewritten_uop.toposort() if u.op is Ops.PARAM}
+    self.assertEqual(rewritten_uop, (pvar['s']<UOp.const(2, dtypes.int)).where(pvar['a'].cast(dtypes.half), pvar['b'].cast(dtypes.half)))
 
   def test_where_merge_branches(self):
     cond1 = Variable("s", 0, 10) < 6
@@ -1172,7 +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)
@@ -1366,16 +1369,6 @@ class TestInvalidIndex(unittest.TestCase):
     c2 = UOp.const((1, Invalid, 1, 1))
     self.assertIs((c1+c2).simplify(), UOp.const((2, Invalid, Invalid, Invalid)))
 
-  def test_gated_load_keeps_index_valid(self):
-    # the load executes even on gated-off iterations: gated_given_valid must not erase its mask (PADTO OOB shape)
-    buf = UOp.param(0, dtypes.bool, (17,))
-    ridx = Variable("ridx", 0, 31)
-    cond = ridx < 17
-    load = buf.index(ridx.valid(cond))
-    out = graph_rewrite(cond.where(load.where(uconst(2), uconst(0)), UOp.invalid()), sym)
-    idx = next(u for u in out.toposort() if u.op is Ops.INDEX)
-    self.assertIs(idx.src[1].get_valid(), cond.simplify())
-
 class TestStoreLoadFolding(unittest.TestCase):
   """Tests for store(index, load(index)) -> NOOP rule. This rule matches patterns that EMERGE during simplification."""
   def test_store_load_folding(self):

test/null/test_validate_oob.py

@@ -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."""
 

test/null/test_viz.py

@@ -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()

test/opt/test_gen_float4.py

@@ -10,12 +10,12 @@ from test.helpers import replace_opts
 class TestFloat4(unittest.TestCase):
   @staticmethod
   def count_float4(uops: list[UOp], n=4):
-    return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype == dtypes.float and uop.shape == (4,)]),
-            len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype == dtypes.float and uop.shape == (4,)]))
+    return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype.scalar() == dtypes.float and uop.shape == (4,)]),
+            len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype.scalar() == dtypes.float and uop.shape == (4,)]))
   @staticmethod
   def count_half4(uops: list[UOp]):
-    return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype == dtypes.half and uop.shape == (4,)]),
-            len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype == dtypes.half and uop.shape == (4,)]))
+    return (len([uop for uop in uops if uop.op is Ops.LOAD and uop.dtype.scalar() == dtypes.half and uop.shape == (4,)]),
+            len([uop for uop in uops if uop.op is Ops.STORE and uop.src[1].dtype.scalar() == dtypes.half and uop.shape == (4,)]))
 
   def test_float4_basic(self):
     a = Tensor.empty(2, 8).realize()

test/opt/test_kernel_opts.py

@@ -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

test/opt/test_tensor_cores.py

@@ -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):
@@ -186,10 +185,10 @@ class TestTensorCores(unittest.TestCase):
     # 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)
+    x, y = Tensor.rand(64, 64, dtype=tc.dtype_in), Tensor.rand(64, 64, 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)
+    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
@@ -200,10 +199,10 @@ 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(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)
+    x, y = Tensor.rand(64, 64, dtype=tc.dtype_in), Tensor.rand(64, 64, 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)
+    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]))
@@ -216,10 +215,10 @@ class TestTensorCores(unittest.TestCase):
   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)
+    x, y = Tensor.rand(64, 64, dtype=tc.dtype_in), Tensor.rand(64, 64, 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)
+    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]))

test/unit/test_allreduce.py

@@ -64,7 +64,7 @@ class TestAllreduceCast(unittest.TestCase):
     with Context(ALLREDUCE_CAST=allreduce_cast, RING=0, SCACHE=0):
       t = Tensor.empty(4, 4, dtype=dtype).shard(ds, axis=0)
       linear = t.sum(0).linear_with_vars()[0]
-      return {si.src[1].buffer.dtype for si in linear.src if si.src[0].op is Ops.COPY}
+      return {si.src[1].buffer.dtype.scalar() for si in linear.src if si.src[0].op is Ops.COPY}
 
   def test_allreduce_cast_bf16(self):
     # with ALLREDUCE_CAST, allreduce copies stay in bfloat16 instead of promoting to float32

test/unit/test_assign.py

@@ -540,10 +540,6 @@ 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)

test/unit/test_attention.py

@@ -1,9 +1,11 @@
 import unittest
+from types import SimpleNamespace
 import numpy as np
 from tinygrad import Tensor, dtypes, nn
+from tinygrad.llm.kimi import _shard_kimi
 from tinygrad.llm.model import (
   GatedDeltaNetBlock, SSMConfig, TransformerBlock, TransformerConfig,
-  apply_rope as apply_rope_new, precompute_freqs_cis, pairwise_topk,
+  apply_rope as apply_rope_new, iterative_topk, l2norm, precompute_freqs_cis, pairwise_topk,
 )
 
 def apply_rope(x:Tensor, start_pos:int):
@@ -41,14 +43,19 @@ class TestAttention(unittest.TestCase):
     np.testing.assert_allclose(block.cache_kv[0, :, :, :seqlen, :].numpy(), expected.numpy(), rtol=1e-5, atol=1e-5)
 
 class TestGatedDeltaNetBlock(unittest.TestCase):
+  def test_kda_l2norm_matches_fla(self):
+    x = np.array([[1e-4, -2e-4, 3e-4], [1.0, 2.0, -3.0]], dtype=np.float32)
+    expected = x / np.sqrt((x*x).sum(axis=-1, keepdims=True) + 1e-6)
+    np.testing.assert_allclose(l2norm(Tensor(x)).numpy(), expected, rtol=1e-6, atol=1e-6)
+
   def _tensor_linspace(self, start:float, stop:float, shape:tuple[int, ...]) -> Tensor:
     return Tensor.linspace(start, stop, int(np.prod(shape)), dtype=dtypes.float32).reshape(*shape)
 
   def _make_config(self, **kwargs):
-    return TransformerConfig(**({"num_blocks":1, "dim":32, "hidden_dim":64, "n_heads":1, "n_kv_heads":1,
-                                 "norm_eps":1e-5, "vocab_size":32, "head_dim":32, "rope_theta":10000.0,
-                                 "rope_dim":32, "v_head_dim":32, "max_context":4, "ssm_layers":(True,),
-                                 "ssm":SSMConfig(conv_kernel=2, state_size=32, group_count=1, time_step_rank=1, inner_size=32)} | kwargs))
+    return TransformerConfig(**({"num_blocks":1, "dim":4, "hidden_dim":8, "n_heads":1, "n_kv_heads":1,
+                                 "norm_eps":1e-5, "vocab_size":32, "head_dim":4, "rope_theta":10000.0,
+                                 "rope_dim":4, "v_head_dim":4, "max_context":4, "ssm_layers":(True,),
+                                 "ssm":SSMConfig(conv_kernel=2, state_size=2, group_count=1, time_step_rank=1, inner_size=2)} | kwargs))
 
   def _make_block(self, config:TransformerConfig) -> GatedDeltaNetBlock:
     block = GatedDeltaNetBlock(config, config.ssm)
@@ -79,10 +86,6 @@ class TestGatedDeltaNetBlock(unittest.TestCase):
       recurrent_state = cache[:, conv_flat:].reshape(cache.shape[0], block.num_v_heads, block.head_v_dim, block.head_v_dim)
       return conv_state, recurrent_state
 
-  def _reset_state(self, block:GatedDeltaNetBlock):
-    Tensor.realize(block.conv_state.assign(block.conv_state.const_like(0)),
-                   block.recurrent_state.assign(block.recurrent_state.const_like(0)))
-
   def _linear_np(self, x:np.ndarray, weight:np.ndarray) -> np.ndarray:
     return x.astype(np.float32) @ weight.T.astype(np.float32)
 
@@ -90,7 +93,7 @@ class TestGatedDeltaNetBlock(unittest.TestCase):
     x_float = x.astype(np.float32)
     return (x_float / np.sqrt((x_float * x_float).mean(axis=-1, keepdims=True) + eps)) * weight.astype(np.float32)
 
-  def _normalize_np(self, x:np.ndarray, eps:float=1e-6) -> np.ndarray:
+  def _normalize_np(self, x:np.ndarray, eps:float=1e-12) -> np.ndarray:
     return x / np.maximum(np.sqrt((x * x).sum(axis=-1, keepdims=True)), eps)
 
   def _softplus_np(self, x:np.ndarray) -> np.ndarray:
@@ -152,12 +155,6 @@ class TestGatedDeltaNetBlock(unittest.TestCase):
     x = Tensor.linspace(-1.0, 1.0, 3 * config.dim, dtype=dtypes.float32).reshape(1, 3, config.dim)
 
     expected_outs, expected_conv, expected_recurrent = self._naive_attention(block, x)
-    out = self._run_attention(block, x, 0)
-    conv_state, recurrent_state = self._cache_views(block)
-    np.testing.assert_allclose(out, np.concatenate(expected_outs, axis=1), rtol=1e-3, atol=1e-3)
-    np.testing.assert_allclose(conv_state, expected_conv[-1], rtol=1e-3, atol=1e-3)
-    np.testing.assert_allclose(recurrent_state, expected_recurrent[-1], rtol=1e-3, atol=1e-3)
-    self._reset_state(block)
 
     for step in range(x.shape[1]):
       out = self._run_attention(block, x[:, step:step+1], step)
@@ -173,7 +170,7 @@ class TestGatedDeltaNetBlock(unittest.TestCase):
     prompt = Tensor.linspace(0.75, -0.75, 2 * config.dim, dtype=dtypes.float32).reshape(1, 2, config.dim)
 
     for i in range(warmup.shape[1]): self._run_attention(block, warmup[:, i:i+1], i)
-    self._reset_state(block)
+    Tensor.realize(*block._state_reset_ops())
     expected_outs, expected_conv, expected_recurrent = self._naive_attention(block, prompt)
 
     for step in range(prompt.shape[1]):
@@ -187,64 +184,99 @@ class TestGatedDeltaNetBlock(unittest.TestCase):
                                  err_msg=f"GatedDeltaNet reset recurrent cache mismatch at step {step}")
 
   def test_kda_channel_decay(self):
-    config = self._make_config(dim=4, hidden_dim=8, n_heads=2, head_dim=4, rope_dim=4, v_head_dim=4,
-      ssm=SSMConfig(conv_kernel=2, state_size=2, group_count=2, time_step_rank=2, inner_size=4, kda=True))
-    block, x = GatedDeltaNetBlock(config, config.ssm), Tensor([[[1., 2., 0., 0.], [2., 1., 0., 0.]]])
+    config = self._make_config(n_heads=2, ssm=SSMConfig(conv_kernel=2, state_size=2, group_count=2, time_step_rank=2, inner_size=4, kda=True))
+    block, x = GatedDeltaNetBlock(config, config.ssm), Tensor([[[1., 2., 0., 0.]]])
+    # f_b(f_a(x)) = [1, 2, 3, 4]
     block.ssm_f_a.weight = Tensor([[1., 0., 0., 0.], [0., 1., 0., 0.]])
     block.ssm_f_b.weight = Tensor([[1., 0.], [0., 1.], [1., 1.], [2., 1.]])
     block._init_state(x)
     initial_state = Tensor.arange(8, dtype=dtypes.float32).reshape(1, 2, 2, 2)
     block.recurrent_state.assign(initial_state).realize()
     block.ssm_a = Tensor([[-1.], [-1.]])
-    block._attention(x, x.shape[1]).realize()
-    alpha = np.exp(-self._softplus_np(np.array([[1, 2, 3, 4], [2, 1, 3, 5]])).reshape(2, 2, 2)).prod(0)
-    np.testing.assert_allclose(block.recurrent_state.numpy(), initial_state.numpy() * alpha[..., None], rtol=1e-5, atol=1e-5)
+    block._attention(x, 0).realize()
+    alpha = np.exp(-self._softplus_np(np.arange(1, 5)).reshape(1, 2, 1, 2))
+    np.testing.assert_allclose(block.recurrent_state.numpy(), initial_state.numpy() * alpha, rtol=1e-5, atol=1e-5)
 
-  def test_kda_prefill_matches_decode(self):
-    config = self._make_config(ssm=SSMConfig(conv_kernel=2, state_size=32, group_count=1, time_step_rank=1, inner_size=32, kda=True))
-    block = GatedDeltaNetBlock(config, config.ssm)
-    for p in nn.state.get_parameters(block):
-      p.replace(self._tensor_linspace(-0.05, 0.05, p.shape) if len(p.shape) > 1 else self._tensor_linspace(0.05, 0.1, p.shape))
-    x = self._tensor_linspace(-0.5, 0.5, (1, 3, config.dim))
-    prefill = self._run_attention(block, x, 0)
-    prefill_conv, prefill_recurrent = self._cache_views(block)
-    self._reset_state(block)
-    decode = np.concatenate([self._run_attention(block, x[:, i:i+1], i) for i in range(3)], axis=1)
-    decode_conv, decode_recurrent = self._cache_views(block)
-    np.testing.assert_allclose(prefill, decode, rtol=1e-3, atol=1e-3)
-    np.testing.assert_allclose(prefill_conv, decode_conv, rtol=1e-3, atol=1e-3)
-    np.testing.assert_allclose(prefill_recurrent, decode_recurrent, rtol=1e-3, atol=1e-3)
+  def test_kda_safe_gate_decay(self):
+    config = self._make_config(n_heads=2, kda_full_rank_gate=True, kda_gate_lower_bound=-5.0,
+      ssm=SSMConfig(conv_kernel=2, state_size=2, group_count=2, time_step_rank=2, inner_size=4, kda=True))
+    block, x = GatedDeltaNetBlock(config, config.ssm), Tensor([[[1., 2., 0., 0.]]])
+    block.ssm_f_a.weight = Tensor([[1., 0., 0., 0.], [0., 1., 0., 0.]])
+    block.ssm_f_b.weight = Tensor([[1., 0.], [0., 1.], [1., 1.], [2., 1.]])
+    block.ssm_dt["bias"] = Tensor.zeros(4)
+    block.ssm_a = Tensor([[-2.], [-3.]])  # stores -exp(A_log)
+    block._init_state(x)
+    initial_state = Tensor.arange(8, dtype=dtypes.float32).reshape(1, 2, 2, 2)
+    block.recurrent_state.assign(initial_state).realize()
+    block._attention(x, 0).realize()
+    gate_logits = np.arange(1, 5, dtype=np.float32).reshape(1, 2, 2)
+    exp_a = np.array([2., 3.], dtype=np.float32).reshape(1, 2, 1)
+    alpha = np.exp(-5.0 / (1.0 + np.exp(-(exp_a * gate_logits)))).reshape(1, 2, 1, 2)
+    np.testing.assert_allclose(block.recurrent_state.numpy(), initial_state.numpy() * alpha, rtol=2e-5, atol=2e-5)
 
-  def test_varied_chunk_sizes_match_decode(self):
-    for kda in (False, True):
-      ssm = SSMConfig(conv_kernel=2, state_size=32, group_count=1, time_step_rank=1, inner_size=32, kda=kda)
-      config = self._make_config(ssm=ssm)
-      if kda:
-        block = GatedDeltaNetBlock(config, config.ssm)
-        for p in nn.state.get_parameters(block):
-          p.replace(self._tensor_linspace(-0.05, 0.05, p.shape) if len(p.shape) > 1 else self._tensor_linspace(0.05, 0.1, p.shape))
-      else: block = self._make_block(config)
-      x = self._tensor_linspace(-0.5, 0.5, (1, 4, config.dim))
-      decode = np.concatenate([self._run_attention(block, x[:, i:i+1], i) for i in range(4)], axis=1)
-      decode_conv, decode_recurrent = self._cache_views(block)
-      for chunking in ([4], [2, 2], [1, 3], [3, 1], [2, 1, 1]):
-        self._reset_state(block)
-        outs, start = [], 0
-        for size in chunking:
-          outs.append(self._run_attention(block, x[:, start:start+size], start))
-          start += size
-        chunked_conv, chunked_recurrent = self._cache_views(block)
-        np.testing.assert_allclose(np.concatenate(outs, axis=1), decode, rtol=1e-3, atol=1e-3, err_msg=f"{kda=} {chunking=}")
-        np.testing.assert_allclose(chunked_conv, decode_conv, rtol=1e-3, atol=1e-3, err_msg=f"{kda=} {chunking=}")
-        np.testing.assert_allclose(chunked_recurrent, decode_recurrent, rtol=1e-3, atol=1e-3, err_msg=f"{kda=} {chunking=}")
+  def test_kda_per_channel_a(self):
+    config = self._make_config(n_heads=2, kda_full_rank_gate=True, kda_gate_lower_bound=-5.0,
+      ssm=SSMConfig(conv_kernel=2, state_size=2, group_count=2, time_step_rank=2, inner_size=4, kda=True, channel_decay=True))
+    block, x = GatedDeltaNetBlock(config, config.ssm), Tensor([[[1., 2., 0., 0.]]])
+    block.ssm_f_a.weight = Tensor([[1., 0., 0., 0.], [0., 1., 0., 0.]])
+    block.ssm_f_b.weight = Tensor([[1., 0.], [0., 1.], [1., 1.], [2., 1.]])
+    block.ssm_dt["bias"] = Tensor.zeros(4)
+    block.ssm_a = Tensor([[-2.], [-3.]])
+    block._init_state(x)
+    initial_state = Tensor.arange(8, dtype=dtypes.float32).reshape(1, 2, 2, 2)
+    block.recurrent_state.assign(initial_state).realize()
+    block._attention(x, 0).realize()
+    gate_logits = np.arange(1, 5, dtype=np.float32).reshape(1, 2, 2)
+    exp_a = np.array([2., 3.], dtype=np.float32).reshape(1, 1, 2)
+    alpha = np.exp(-5.0 / (1.0 + np.exp(-(exp_a * gate_logits)))).reshape(1, 2, 1, 2)
+    np.testing.assert_allclose(block.recurrent_state.numpy(), initial_state.numpy() * alpha, rtol=2e-5, atol=2e-5)
 
-  def test_start_zero_resets_realized_state(self):
-    config, x = self._make_config(max_context=3), self._tensor_linspace(-1, 1, (1, 3, 32))
-    block = self._make_block(config)
-    self._run_attention(block, x, 0)
-    restarted = self._run_attention(block, x[:, :2], 0)
-    fresh = self._run_attention(self._make_block(config), x[:, :2], 0)
-    np.testing.assert_allclose(restarted, fresh, rtol=1e-3, atol=1e-3)
+  def test_kda_chunked_prefill_matches_decode(self):
+    config = self._make_config(max_context=4, n_heads=2,
+      ssm=SSMConfig(conv_kernel=2, state_size=2, group_count=2, time_step_rank=2, inner_size=4, kda=True), kda_split_qkv=True)
+    x = Tensor.linspace(-1, 1, 4*config.dim, dtype=dtypes.float32).reshape(1, 4, config.dim).cast(dtypes.bfloat16)
+
+    chunked = GatedDeltaNetBlock(config, config.ssm)
+    sequential = GatedDeltaNetBlock(config, config.ssm)
+    for value in nn.state.get_state_dict(chunked).values(): value.replace(value.cast(dtypes.bfloat16).realize())
+    sequential_state = nn.state.get_state_dict(sequential)
+    for name, value in nn.state.get_state_dict(chunked).items(): sequential_state[name].replace(value)
+
+    chunked._init_state(x)
+    chunk_out = chunked._attention(x, 0).realize()
+    sequential._init_state(x)
+    seq_out = Tensor.cat(*[sequential._attention(x[:, t:t+1], t).realize() for t in range(x.shape[1])], dim=1).realize()
+
+    np.testing.assert_allclose(chunk_out.numpy(), seq_out.numpy(), rtol=1e-5, atol=1e-5)
+    for name in ("conv_state_q", "conv_state_k", "conv_state_v"):
+      np.testing.assert_allclose(getattr(chunked, name).numpy(), getattr(sequential, name).numpy(), rtol=2e-2, atol=4e-3)
+    np.testing.assert_allclose(chunked.recurrent_state.numpy(), sequential.recurrent_state.numpy(), rtol=2e-3, atol=2e-3)
+
+  def test_kda_tp_final_token_matches_unsharded(self):
+    config = self._make_config(dim=8, hidden_dim=16, n_heads=4, n_kv_heads=4, head_dim=2, rope_dim=2, v_head_dim=2,
+      ssm=SSMConfig(conv_kernel=2, state_size=2, group_count=4, time_step_rank=4, inner_size=8, kda=True), kda_split_qkv=True)
+    single, tp = GatedDeltaNetBlock(config, config.ssm), GatedDeltaNetBlock(config, config.ssm)
+    for name, value in nn.state.get_state_dict(single).items():
+      data = np.full(value.shape, 1.0, np.float32) if "norm.weight" in name else \
+             np.linspace(-0.2, 0.2, value.numel(), dtype=np.float32).reshape(value.shape)
+      if name == "ssm_a": data.fill(-0.1)
+      value.replace(Tensor(data, device="CPU", dtype=dtypes.bfloat16).realize())
+    tp_state = nn.state.get_state_dict(tp)
+    for name, value in nn.state.get_state_dict(single).items(): tp_state[name].replace(value)
+    devices = ("CPU", "CPU:1")
+    _shard_kimi(SimpleNamespace(blk=[tp]), devices)
+
+    x = Tensor(np.linspace(-1, 1, 32, dtype=np.float32).reshape(1, 4, 8), device="CPU", dtype=dtypes.bfloat16)
+    single._init_state(x)
+    expected = single._attention(x, 0).realize()
+    x_tp = x.shard(devices, axis=None)
+    tp._init_state(x_tp)
+    actual = tp._attention(x_tp, 0).realize()
+
+    np.testing.assert_equal(actual.numpy(), expected.numpy())
+    np.testing.assert_equal(tp.recurrent_state.numpy(), single.recurrent_state.numpy())
+    for name in ("conv_state_q", "conv_state_k", "conv_state_v"):
+      np.testing.assert_equal(getattr(tp, name).numpy(), getattr(single, name).numpy())
 
 class TestPairwiseTopk(unittest.TestCase):
   def test_basic_topk(self):
@@ -269,5 +301,13 @@ class TestPairwiseTopk(unittest.TestCase):
         self.assertEqual(set(sel.numpy()[b, t].tolist()), expected)
         np.testing.assert_allclose(vals.numpy()[b, t], data[b, t][sel.numpy()[b, t]])
 
+  def test_iterative_matches_numpy(self):
+    rng = np.random.default_rng(42)
+    data = rng.standard_normal((2, 3, 896), dtype=np.float32)
+    vals, sel = iterative_topk(Tensor(data), 16)
+    expected = np.argsort(-data, axis=-1, stable=True)[..., :16]
+    np.testing.assert_equal(sel.numpy(), expected)
+    np.testing.assert_allclose(vals.numpy(), np.take_along_axis(data, expected, axis=-1))
+
 if __name__ == '__main__':
   unittest.main()

test/unit/test_dtype_weak.py

@@ -3,12 +3,11 @@ 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, GroupOp, dtype_from_uop, graph_rewrite
+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.symbolic import symbolic_simple
 from tinygrad.uop.spec import spec_shared, type_verify
 from tinygrad.engine.jit import JitError
-from test.helpers import full_rewrite
 
 
 class TestWeakPromotion(unittest.TestCase):
@@ -77,11 +76,6 @@ class TestWeakPromotion(unittest.TestCase):
       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_div_sub_operand_kept_weak(self):
-    a = Tensor.empty(4, dtype=dtypes.float32)
-    for t in (a / 1, a - 0):
-      self.assertEqual(t.uop.src[1].dtype, dtypes.weakfloat)
-
   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):
@@ -94,13 +88,6 @@ class TestWeakPromotion(unittest.TestCase):
       out = Tensor(1.0, dtype=dtypes.float32, device="CPU") / denom
       self.assertAlmostEqual(out.item(), 1 / (70000 + 1e-5), places=10)
 
-  def test_stacked_weak_casts_convert_each_kind(self):
-    # each weak cast is a kind conversion: weakint truncates before weakfloat re-lifts (neither is only a marker)
-    x = Tensor([2.5, -3.7], dtype=dtypes.float32, device="CPU")
-    stacked = x.cast(dtypes.weakint).cast(dtypes.weakfloat)
-    self.assertIs(stacked.dtype, dtypes.weakfloat)
-    self.assertEqual(stacked.tolist(), [2.0, -3.0])
-
   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),
@@ -289,18 +276,5 @@ class TestSignedUint64Weakfloat(unittest.TestCase):
     self.assertAlmostEqual((i64 + u64).sin().item(), math.sin(2), places=5)  # Unary lowers before transcendental
 
 
-class TestNoRedundantWide(unittest.TestCase):
-  def wide_alu(self, t:Tensor) -> int:
-    return sum(sum(1 for u in full_rewrite(call.src[0]).toposort() if u.op in GroupOp.ALU and u.dtype in {dtypes.long, dtypes.ulong})
-               for call in t.schedule_linear().src if call.src[0].op is Ops.SINK)
-
-  def test_unbounded_long_stays_long(self):
-    self.assertGreater(self.wide_alu(Tensor.empty(16, dtype=dtypes.long)*3 + 1), 0)
-
-  def test_fancy_index_has_no_wide_alu(self):
-    j, o = Tensor([0, 1, 2]).reshape(3, 1), Tensor([0, 1]).reshape(1, 2)
-    self.assertEqual(self.wide_alu(Tensor.empty(8, 9, 10, 11, 12)[1, j, 2, o, 2]), 0)
-
-
 if __name__ == "__main__":
   unittest.main()

test/unit/test_gradient.py

@@ -51,10 +51,6 @@ class TestTensorGradient(unittest.TestCase):
     with self.assertRaises(RuntimeError): x.sum().gradient(x)
     with self.assertRaises(RuntimeError): x.float().sum().gradient(x)
 
-  def test_const_target_raise(self):
-    t = Tensor(2.0)
-    with self.assertRaises(RuntimeError): (t * 2.0).gradient(t)
-
   def test_copy_to_device_gradient(self):
     t = Tensor([1.0, 2, 3]).realize()
     t.to("CPU:1").square().sum().backward()
@@ -104,7 +100,7 @@ class TestTensorGradient(unittest.TestCase):
 
   def test_implicit_broadcast_where_gradient(self):
     # WHERE with a bare ()-shape branch: the scalar's gradient counts the positions where it is selected
-    cond, x, w = Tensor([True, False, True]), Tensor([1.0, 2.0, 3.0]), Tensor(4.0, dtype=dtypes.float32)
+    cond, x, w = Tensor([True, False, True]), Tensor([1.0, 2.0, 3.0]), Tensor(4.0)
     dw = Tensor(cond.uop.alu(Ops.WHERE, x.uop, w.uop)).sum().gradient(w)[0]
     self.assertEqual(dw.shape, ())
     self.assertEqual(dw.item(), 1.0)
@@ -113,7 +109,7 @@ class TestTensorGradient(unittest.TestCase):
 
   def test_implicit_broadcast_alu_gradient(self):
     # MUL with a bare ()-shape src, no EXPAND in the graph
-    x, w = Tensor([1.0, 2.0, 3.0]), Tensor(2.0, dtype=dtypes.float32)
+    x, w = Tensor([1.0, 2.0, 3.0]), Tensor(2.0)
     m = x.uop.alu(Ops.MUL, w.uop)
     self.assertIs(m.src[1], w.uop)
     dw = Tensor(m).sum().gradient(w)[0]
@@ -122,7 +118,7 @@ class TestTensorGradient(unittest.TestCase):
 
   def test_implicit_broadcast_intermediate_accumulation(self):
     # s is used directly and through an implicit broadcast edge, each edge's gradient reduces to s's shape before they sum
-    x, p = Tensor([1.0, 2.0, 3.0]), Tensor(0.5, dtype=dtypes.float32)
+    x, p = Tensor([1.0, 2.0, 3.0]), Tensor(0.5)
     s = p.sin()
     z = Tensor(x.uop.alu(Ops.MUL, s.uop)).sum() + s
     dp = z.gradient(p)[0]

test/unit/test_hcq_graph.py

@@ -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

test/unit/test_llm_cli_k3.py

@@ -0,0 +1,30 @@
+import unittest
+from tinygrad.llm.cli import KimiK3Template
+from tinygrad.llm.serve import StreamRouter
+
+class TestKimiK3Template(unittest.TestCase):
+  def test_simple_text_chat(self):
+    template = KimiK3Template()
+    got = template.render([{"role":"system", "content":"Be concise."}, {"role":"user", "content":"Hello"}])
+    self.assertTrue(got.startswith('<|open|>message role="system" type="thinking-effort"<|sep|>'))
+    self.assertIn('<|open|>message role="user"<|sep|>Hello<|close|>message<|sep|><|end_of_msg|>', got)
+    self.assertTrue(got.endswith('<|open|>message role="assistant"<|sep|><|open|>think<|sep|>'))
+
+  def test_preserves_assistant_thinking(self):
+    got = KimiK3Template().render([{"role":"assistant", "reasoning_content":"why", "content":"answer"}], add_generation_prompt=False)
+    self.assertIn('<|open|>think<|sep|>why<|close|>think<|sep|>', got)
+    self.assertIn('<|open|>response<|sep|>answer<|close|>response<|sep|>', got)
+
+  def test_rejects_unimplemented_modalities(self):
+    with self.assertRaisesRegex(ValueError, "text-only"):
+      KimiK3Template().render([{"role":"user", "content":[{"type":"image", "url":"x"}]}])
+    with self.assertRaisesRegex(ValueError, "tool rendering"):
+      KimiK3Template().render([{"role":"user", "content":"x"}], tools=[{"type":"function"}])
+
+  def test_xtml_stream_router(self):
+    router, routed = StreamRouter(reasoning=True, xtml=True), []
+    for piece in ("rea", "son<|close|>thi", "nk<|sep|><|open|>response<|sep|>ans", "wer<|close|>response<|sep|>"):
+      routed.extend(router.route(piece))
+    self.assertEqual(routed, [("reasoning_content", "rea"), ("reasoning_content", "son"), ("content", "ans"), ("content", "wer")])
+
+if __name__ == "__main__": unittest.main()

test/unit/test_llm_k3.py

@@ -0,0 +1,139 @@
+import tempfile, unittest
+from pathlib import Path
+from dataclasses import replace
+import numpy as np
+from tinygrad import Tensor, dtypes, nn
+from tinygrad.helpers import getenv
+from tinygrad.llm.kernels import bf16_mfma_splitk
+from tinygrad.llm.kimi_k3 import KIMI_K3_FULL_ATTN_LAYERS, KIMI_K3_SSM_LAYERS, KIMI_K3_TEXT_SIZE, KIMI_K3_TP8_BYTES_PER_GPU, \
+  _layer_sources, _load_stacked_experts, _replace, _safe_load_selected, _shard_kimi_k3, _validate_config, kimi_k3_config, kimi_k3_smoke_config
+from tinygrad.llm.model import FFNBlock, Transformer
+
+def small_k3_config(max_context:int=4): return replace(kimi_k3_smoke_config(max_context), num_experts=8)
+
+class TestKimiK3(unittest.TestCase):
+  def test_selective_safetensor_load(self):
+    with tempfile.TemporaryDirectory() as tmp:
+      path = Path(tmp) / "weights.safetensors"
+      nn.state.safe_save({"keep":Tensor.arange(8), "skip":Tensor.arange(16)}, str(path))
+      selected = _safe_load_selected(path, ["keep"])
+      self.assertEqual(list(selected), ["keep"])
+      np.testing.assert_equal(selected["keep"].numpy(), np.arange(8))
+      with self.assertRaisesRegex(ValueError, "missing tensor absent"): _safe_load_selected(path, ["absent"])
+
+  def test_smoke_config_preserves_gfx950_expert_alignment(self):
+    c = kimi_k3_smoke_config()
+    self.assertEqual(c.routed_expert_dim % 64, 0)
+    self.assertEqual((c.hidden_dim // 8) % 64, 0)
+
+  @unittest.skipUnless(getenv("DEV", "") == "NULL:HIP:gfx950", "gfx950 compile coverage")
+  def test_gfx950_mfma_splitk_compile(self):
+    x = Tensor.zeros(1, 1, 256, dtype=dtypes.bfloat16, device="NULL:HIP:gfx950")
+    weight = Tensor.zeros(16, 256, dtype=dtypes.bfloat16, device="NULL:HIP:gfx950")
+    self.assertEqual(bf16_mfma_splitk(x, weight).realize().shape, (1, 1, 16))
+
+  def test_official_config(self):
+    c = kimi_k3_config(1_048_576)
+    self.assertEqual((c.num_blocks, c.dim, c.n_heads, c.num_experts, c.num_experts_per_tok), (93, 7168, 96, 896, 16))
+    self.assertEqual((sum(KIMI_K3_SSM_LAYERS), len(KIMI_K3_FULL_ATTN_LAYERS)), (69, 24))
+    self.assertEqual(KIMI_K3_FULL_ATTN_LAYERS, (*range(3, 93, 4), 92))
+    self.assertEqual((c.routed_expert_dim, c.hidden_dim, c.shared_expert_dim), (3584, 3072, 6144))
+    self.assertTrue(c.route_weights_uncorrected and c.kda_full_rank_gate and c.attn_output_gate)
+    self.assertTrue(c.ssm is not None and c.ssm.channel_decay)
+    self.assertEqual((c.activation_situ_beta, c.activation_situ_linear_beta, c.kda_gate_lower_bound), (4.0, 25.0, -5.0))
+
+  def test_config_rejects_wrong_checkpoint(self):
+    with self.assertRaisesRegex(ValueError, "not the supported official"):
+      _validate_config({"model_type":"kimi_linear", "hidden_size":2304})
+
+  def test_official_mapping_covers_model(self):
+    model = Transformer(kimi_k3_config(1))
+    state = nn.state.get_state_dict(model)
+    targets = {"token_embd.weight", "output_norm.weight", "output.weight", "output_attn_res_norm.weight", "output_attn_res_proj.weight"}
+    for i,is_kda in enumerate(KIMI_K3_SSM_LAYERS):
+      for target in _layer_sources(i, is_kda).values(): targets.update(target.split("|"))
+      if i:
+        for name in ("ffn_gate_exps.weight", "ffn_gate_exps.weight_scale", "ffn_up_exps.weight", "ffn_up_exps.weight_scale",
+                     "ffn_down_exps.weight", "ffn_down_exps.weight_scale"): targets.add(f"blk.{i}.{name}")
+    self.assertEqual(targets, set(state))
+    self.assertEqual(state["blk.1.ffn_gate_exps.weight"].shape, (896, 3072, 1792))
+    self.assertEqual(state["blk.1.ffn_gate_exps.weight_scale"].shape, (896, 3072, 112))
+    self.assertEqual(state["blk.0.ssm_a"].shape, (128, 1))
+    _shard_kimi_k3(model, tuple(f"NULL:{i}" for i in range(8)))
+    total, per_gpu = 0, 0
+    for name,value in state.items():
+      dtype = dtypes.uint8 if name.endswith(("weight_scale", "_exps.weight")) else dtypes.float32 if name.endswith(
+        ("exp_probs_b.bias", "ssm_q_conv1d.weight", "ssm_k_conv1d.weight", "ssm_v_conv1d.weight", "ssm_norm.weight", "ssm_a", "ssm_dt.bias")) \
+        else dtypes.bfloat16
+      size = value.numel() * dtype.itemsize
+      total += size
+      per_gpu += size if value.uop.axis is None else size//8
+    self.assertEqual((total, per_gpu), (KIMI_K3_TEXT_SIZE, KIMI_K3_TP8_BYTES_PER_GPU))
+
+  def test_situ_matches_reference(self):
+    block = FFNBlock(small_k3_config())
+    gate, up = Tensor([[-8., -1., 0., 3.]]), Tensor([[-30., -2., 5., 40.]])
+    got = block._activation(gate, up).numpy()
+    g, u = gate.numpy().astype(np.float32), up.numpy().astype(np.float32)
+    expected = (4*np.tanh(g/4)/(1+np.exp(-g))) * (25*np.tanh(u/25))
+    np.testing.assert_allclose(got, expected, rtol=1e-5, atol=1e-5)
+
+  def test_attention_residual_matches_reference(self):
+    block = FFNBlock(small_k3_config())
+    block.attn_res_norm.weight.assign([1.0+i/16 for i in range(32)])
+    block.attn_res_proj.weight.assign([[(-1.0)**i/8 for i in range(32)]])
+    prefix, residual = Tensor.arange(64).reshape(2, 32).float()/16, Tensor.arange(128).reshape(2, 2, 32).float()/32
+    got = block._apply_attn_res(prefix, residual, block.attn_res_proj, block.attn_res_norm).numpy()
+    v = np.concatenate((residual.numpy(), prefix.numpy()[:, None]), axis=1).astype(np.float32)
+    k = v / np.sqrt(np.mean(v*v, axis=-1, keepdims=True) + 1e-5)
+    scores = np.sum(k * block.attn_res_norm.weight.numpy() * block.attn_res_proj.weight.numpy()[0], axis=-1)
+    probs = np.exp(scores-scores.max(axis=-1, keepdims=True))
+    probs /= probs.sum(axis=-1, keepdims=True)
+    expected = np.matmul(probs[:, None], v).squeeze(1)
+    np.testing.assert_allclose(got, expected, rtol=1e-5, atol=1e-5)
+
+  def test_tp8_schema(self):
+    model = Transformer(small_k3_config())
+    _shard_kimi_k3(model, tuple(f"NULL:{i}" for i in range(8)))
+    state = nn.state.get_state_dict(model)
+    for name,axis in (("token_embd.weight",0), ("blk.1.ffn_gate_exps.weight",1), ("blk.1.ffn_down_exps.weight_scale",2),
+                      ("blk.1.ffn_routed_down.weight",1), ("blk.0.ssm_g_full.weight",0), ("blk.1.attn_q_b.weight",0)):
+      self.assertEqual(state[name].uop.axis, axis, name)
+    self.assertIsNone(state["blk.1.attn_res_norm.weight"].uop.axis)
+    self.assertIsNone(state["blk.1.ffn_routed_norm.weight"].uop.axis)
+    self.assertIsNone(state["blk.0.ssm_a"].uop.axis)
+
+  def test_direct_expert_staging(self):
+    devices = tuple(f"PYTHON:{i}" for i in range(4))
+    sources = [Tensor([[(e*40+r*4+c)&255 for c in range(4)] for r in range(8)], dtype=dtypes.uint8,
+                      device=devices[0]).realize() for e in range(8)]
+    expected = Tensor.stack(*sources).numpy()
+    for axis in (1, 2):
+      dst = Tensor.zeros(8, 8, 4, dtype=dtypes.uint8, device=devices[0]).shard(devices, axis=axis)
+      _load_stacked_experts(dst, sources)
+      np.testing.assert_equal(dst.numpy(), expected)
+
+  def test_direct_tp_replacement(self):
+    devices = tuple(f"PYTHON:{i}" for i in range(4))
+    source = Tensor.arange(64, dtype=dtypes.float32).reshape(8, 8).realize()
+    expected = source.numpy()
+    for axis in (None, 0, 1):
+      dst = Tensor.zeros(8, 8, device="PYTHON").shard(devices, axis=axis)
+      _replace(dst, source)
+      np.testing.assert_equal(dst.numpy(), expected)
+
+  def test_chunked_recurrent_generate(self):
+    model = Transformer(small_k3_config(max_context=8))
+    for name,value in nn.state.get_state_dict(model).items():
+      fill = 127 if name.endswith("weight_scale") else 0
+      value.replace(Tensor.full(value.shape, fill, dtype=value.dtype if value.dtype is dtypes.uint8 else dtypes.bfloat16, device="PYTHON"))
+    self.assertIsInstance(next(model.generate([1], chunk_size=2)), int)
+    prompt = [1, 2, 3, 4]
+    for _ in range(3): self.assertIsInstance(next(model.generate(prompt.copy(), chunk_size=2)), int)
+    self.assertEqual(model.get_start_pos(model._cached_tokens + [42]), len(prompt))
+    self.assertEqual(model.get_start_pos([9, 2, 3, 4, 42]), 0)
+    self.assertIsInstance(next(model.generate([1, 2, 3, 4, 5], chunk_size=3)), int)
+    self.assertEqual(set(model.recurrent_greedy_prefill_jits), {2})
+    self.assertEqual(model._cached_tokens[:4], [1, 2, 3, 4])
+
+if __name__ == "__main__": unittest.main()

test/unit/test_llm_kimi.py

@@ -0,0 +1,37 @@
+import unittest
+from tinygrad import dtypes, nn
+from tinygrad.llm.kimi import KIMI_LOGICAL_BYTES, KIMI_SSM_LAYERS, KIMI_TENSOR_COUNT, _shard_kimi, _validate_kimi_state, kimi_config
+from tinygrad.llm.model import Transformer
+
+class TestKimiLinear(unittest.TestCase):
+  def test_architecture_config(self):
+    config = kimi_config(4096)
+    self.assertEqual((config.num_blocks, config.dim, config.n_heads, config.vocab_size), (27, 2304, 32, 163840))
+    self.assertEqual(tuple(i for i, is_kda in enumerate(KIMI_SSM_LAYERS) if not is_kda), (3, 7, 11, 15, 19, 23, 26))
+    self.assertEqual((config.num_experts, config.num_experts_per_tok, config.shared_expert_dim), (256, 8, 1024))
+    self.assertTrue(config.expert_mxfp4 and config.bf16_activations and config.kda_split_qkv)
+    self.assertFalse(config.shared_expert_gate)
+
+  def test_tp4_schema_and_axes(self):
+    model = Transformer(kimi_config(32))
+    state = nn.state.get_state_dict(model)
+    self.assertEqual(len(state), KIMI_TENSOR_COUNT)
+    self.assertNotIn("blk.1.ffn_gate_inp_shexp.weight", state)
+    self.assertEqual(state["blk.1.ffn_gate_exps.weight"].dtype, dtypes.uint8)
+    self.assertEqual(state["blk.1.ffn_gate_exps.weight_scale"].dtype, dtypes.uint8)
+
+    _shard_kimi(model, ("NULL:0", "NULL:1", "NULL:2", "NULL:3"))
+    state = nn.state.get_state_dict(model)
+    for name, axis in (("token_embd.weight", 0), ("blk.1.ffn_gate_exps.weight", 1),
+                       ("blk.1.ffn_down_exps.weight_scale", 2), ("blk.3.attn_k_b.weight", 0)):
+      self.assertEqual(state[name].uop.axis, axis, name)
+    self.assertIsNone(state["blk.1.attn_norm.weight"].uop.axis)
+
+  def test_converted_schema_validation(self):
+    model = Transformer(kimi_config(1))
+    state = {name:value if value.dtype is dtypes.uint8 else value.cast(dtypes.bfloat16)
+             for name,value in nn.state.get_state_dict(model).items()}
+    _validate_kimi_state(model, state)
+    self.assertEqual(sum(value.nbytes() for value in state.values()), KIMI_LOGICAL_BYTES)
+
+if __name__ == "__main__": unittest.main()

test/unit/test_llm_moe.py

@@ -2,7 +2,7 @@ import unittest
 import numpy as np
 from dataclasses import replace
 from tinygrad import Tensor
-from tinygrad.llm.model import ExpertGating, TransformerBlock, TransformerConfig
+from tinygrad.llm.model import TransformerBlock, TransformerConfig
 
 def _moe_config(dim=8, hidden=16, n_heads=2, num_experts=4, num_experts_per_tok=2):
   return TransformerConfig(
@@ -72,6 +72,20 @@ class TestMoEFeedForward(unittest.TestCase):
     expected = (Tensor([1.0]).silu().item() + Tensor([3.0]).silu().item()) / 2
     np.testing.assert_allclose(out.numpy()[0, 0, 0], expected, rtol=1e-2)
 
+  def test_kimi_correction_bias_affects_route_weights(self):
+    dim, hidden, n_heads, num_experts, k = 8, 16, 2, 4, 2
+    config = replace(_moe_config(dim, hidden, n_heads, num_experts, k), norm_topk_prob=True, expert_bias=True)
+    block = TransformerBlock(config)
+    block.ffn_gate_exps.weight = Tensor.stack(*[Tensor.eye(hidden, dim) * (i + 1) for i in range(num_experts)])
+    block.ffn_up_exps.weight = Tensor.stack(*[Tensor.eye(hidden, dim) for _ in range(num_experts)])
+    block.ffn_down_exps.weight = Tensor.stack(*[Tensor.eye(dim, hidden) for _ in range(num_experts)])
+    block.ffn_gate_inp.weight = Tensor.zeros(num_experts, dim)
+    block.exp_probs_b["bias"] = Tensor([0.2, 0.1, 0.0, -0.1])
+
+    out = block._feed_forward(Tensor.ones(1, 1, dim))
+    expected = (Tensor([1.0]).silu().item() * 0.7 + Tensor([2.0]).silu().item() * 0.6) / 1.3
+    np.testing.assert_allclose(out.numpy()[0, 0, 0], expected, rtol=1e-2)
+
   def test_moe_feed_forward_shared_expert(self):
     dim, hidden, n_heads = 8, 16, 2
     num_experts, k = 4, 2
@@ -96,32 +110,5 @@ class TestMoEFeedForward(unittest.TestCase):
     expected = moe_expected + shared_expected
     np.testing.assert_allclose(out.numpy(), expected, rtol=1e-2)
 
-  def test_moe_feed_forward_gating_funcs(self):
-    dim, hidden, n_heads = 8, 16, 2
-    num_experts, k = 4, 2
-    logits = np.array([4.0, 3.0, 0.0, -1.0], dtype=np.float32)
-    def softmax(x):
-      probs = np.exp(x - x.max())
-      return probs / probs.sum()
-    for gating_func in ExpertGating:
-      for norm_topk_prob in (False, True):
-        block = TransformerBlock(replace(_moe_config(dim, hidden, n_heads, num_experts, k),
-                                         expert_gating_func=gating_func, norm_topk_prob=norm_topk_prob))
-        block.ffn_gate_exps.weight = Tensor.stack(*[Tensor.eye(hidden, dim) for _ in range(num_experts)])
-        block.ffn_up_exps.weight = Tensor.stack(*[Tensor.eye(hidden, dim) * (i + 1) for i in range(num_experts)])
-        block.ffn_down_exps.weight = Tensor.stack(*[Tensor.eye(dim, hidden) for _ in range(num_experts)])
-        block.ffn_gate_inp.weight = Tensor((logits / dim)[None, :].repeat(dim, 0).T)
-        out = block._feed_forward(Tensor.ones(1, 1, dim)).numpy()[0, 0, 0]
-
-        if gating_func == ExpertGating.SOFTMAX: selection_scores = softmax(logits)
-        elif gating_func == ExpertGating.SIGMOID: selection_scores = 1 / (1 + np.exp(-logits))
-        elif gating_func == ExpertGating.SOFTMAX_WEIGHT: selection_scores = logits
-        else: selection_scores = np.sqrt(np.logaddexp(0, logits))
-        sel = np.argsort(selection_scores)[-k:]
-        weights = softmax(logits[sel]) if gating_func == ExpertGating.SOFTMAX_WEIGHT else selection_scores[sel]
-        if norm_topk_prob: weights /= weights.sum()
-        expected = (weights * (sel + 1)).sum() / (1 + np.exp(-1))
-        np.testing.assert_allclose(out, expected, rtol=1e-3)
-
 if __name__ == '__main__':
   unittest.main()

test/unit/test_llm_quant.py

@@ -0,0 +1,68 @@
+import unittest
+import numpy as np
+from tinygrad import Tensor, dtypes
+from tinygrad.llm.quant import MXFP4_VALUES, dequantize_mxfp4, quantize_dequantize_mxfp8, quantize_mxfp4, quantize_mxfp4_cpu
+
+class TestMXFormats(unittest.TestCase):
+  def test_mxfp4_known_codes_and_scale(self):
+    values = np.array(MXFP4_VALUES * 2, dtype=np.float32)
+    packed, scale = quantize_mxfp4(Tensor(values))
+    # Positive and negative zero are numerically identical, so nearest-value encoding canonicalizes to +0.
+    np.testing.assert_array_equal(packed.numpy(), np.array([0x10, 0x32, 0x54, 0x76, 0x90, 0xba, 0xdc, 0xfe] * 2, dtype=np.uint8))
+    np.testing.assert_array_equal(scale.numpy(), np.array([127], dtype=np.uint8))
+    np.testing.assert_array_equal(dequantize_mxfp4(packed, scale, dtypes.float32).numpy(), values)
+
+  def test_mxfp4_block_scales_and_zero(self):
+    x = Tensor(np.array([0.0]*32 + [12.0, -12.0] + [0.0]*30, dtype=np.float32))
+    packed, scale = quantize_mxfp4(x)
+    np.testing.assert_array_equal(scale.numpy(), np.array([127, 128], dtype=np.uint8))
+    np.testing.assert_allclose(dequantize_mxfp4(packed, scale, dtypes.float32).numpy(), x.numpy())
+
+  def test_mxfp4_scale_rounds_amax_over_format_max(self):
+    # OCP E8M0 scale selection rounds log2(amax / 6), rather than flooring the
+    # input exponent. At this boundary the two rules differ by a factor of two.
+    x = Tensor(np.array([8.0] + [0.0]*31, dtype=np.float32))
+    packed, scale = quantize_mxfp4(x)
+    np.testing.assert_array_equal(scale.numpy(), np.array([127], dtype=np.uint8))
+    self.assertEqual(dequantize_mxfp4(packed, scale, dtypes.float32).numpy()[0], 6.0)
+
+  def test_mxfp4_cpu_converter_matches_tensor_path(self):
+    x = Tensor(np.linspace(-13, 13, 64*32, dtype=np.float32).reshape(64, 32))
+    packed, scale = quantize_mxfp4(x)
+    cpu_packed, cpu_scale = quantize_mxfp4_cpu(x)
+    np.testing.assert_array_equal(cpu_packed.numpy(), packed.numpy())
+    np.testing.assert_array_equal(cpu_scale.numpy(), scale.numpy())
+
+  def test_mxfp4_midpoints_round_to_even(self):
+    midpoints = np.array([0.25, 0.75, 1.25, 1.75, 2.5, 3.5, 5.0], dtype=np.float32)
+    x = Tensor(np.pad(np.concatenate((midpoints, -midpoints)), (0, 18)))
+    packed, scale = quantize_mxfp4(x)
+    expected = np.pad(np.array([0, 1, 1, 2, 2, 4, 4, 0, -1, -1, -2, -2, -4, -4], dtype=np.float32), (0, 18))
+    np.testing.assert_array_equal(dequantize_mxfp4(packed, scale, dtypes.float32).numpy(), expected)
+
+  def test_mxfp8_roundtrip_and_dtype(self):
+    # All E4M3-exact values remain exact after extracting a shared exponent.
+    x = Tensor(np.array(([0.0, 0.5, 1.0, 1.5, 2.0, -3.0, 4.0, -6.0] * 4), dtype=np.float32))
+    out = quantize_dequantize_mxfp8(x)
+    self.assertEqual(out.dtype, dtypes.bfloat16)
+    np.testing.assert_array_equal(out.float().numpy(), x.numpy())
+
+  def test_mxfp8_subnormal_and_rounding(self):
+    x = np.zeros(32, dtype=np.float32)
+    x[:5] = [1.0, 1.0625, 1.07, 2**-9, 2**-10]
+    out = quantize_dequantize_mxfp8(Tensor(x), dtype=dtypes.float32).numpy()
+    # amax / 448 rounds to an E8M0 scale of 2**-9, saturating the largest
+    # values while retaining the E4M3 subnormal quantum for this block.
+    np.testing.assert_array_equal(out[:5], [0.875, 0.875, 0.875, 2**-9, 2**-10])
+
+  def test_mxfp8_uses_full_e4m3_range(self):
+    x = np.zeros(32, dtype=np.float32)
+    x[:4] = [448.0, 416.0, 400.0, -448.0]
+    np.testing.assert_array_equal(quantize_dequantize_mxfp8(Tensor(x), dtype=dtypes.float32).numpy()[:4], [448.0, 416.0, 384.0, -448.0])
+
+  def test_mxfp8_scale_rounds_amax_over_format_max(self):
+    x = np.zeros(32, dtype=np.float32)
+    x[0] = 512.0
+    self.assertEqual(quantize_dequantize_mxfp8(Tensor(x), dtype=dtypes.float32).numpy()[0], 448.0)
+
+if __name__ == "__main__": unittest.main()

test/unit/test_llm_server.py

@@ -1,9 +1,10 @@
 import unittest
+from dataclasses import replace
 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
+from tinygrad.llm.serve import StreamRouter, parse_kimi_tool_call
 
 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)
@@ -13,17 +14,29 @@ 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)
+    def generate(tokens, temperature):
+      calls.append((tokens, temperature))
       yield from (1, 2)
     with patch.object(model, "generate", generate): model.warmup()
-    self.assertEqual(calls, [[0], [0]])
+    self.assertEqual(calls, [([0], 0.0), ([0], 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_recurrent_warmup_captures_reset_replay(self):
+    model, calls = Transformer(TEST_CONFIG), []
+    model.has_recurrent_block = True
+    state = Tensor.ones(4).realize()
+    model.blk[0]._state_reset_ops = lambda: [state.assign(state.const_like(0))]
+    def generate(tokens, temperature):
+      if calls: model.reset_jit()
+      calls.append((tokens.copy(), temperature))
+      tokens.append(42)
+      yield from (1, 2)
+    with patch.object(model, "generate", generate): model.warmup()
+    prompt = [0] * (TEST_CONFIG.max_context-2)
+    self.assertEqual(calls, [(prompt, 0.0)] * 3 + [(prompt, 1.0)] * 3 + [(prompt + list(range(1, i+1)), 0.0) for i in range(1, 4)])
+    self.assertEqual(model.reset_jit.cnt, 8)
+    cache_size = len(schedule_cache)
+    model.reset_jit()
+    self.assertEqual(len(schedule_cache), cache_size)
 
   def test_first_recurrent_generate_before_state_init(self):
     model = Transformer(TEST_CONFIG)
@@ -31,6 +44,17 @@ class TestTransformerGenerate(unittest.TestCase):
     with patch.object(Transformer, '__call__', return_value=Tensor([[42]])):
       self.assertEqual(next(model.generate([0])), 42)
 
+  def test_recurrent_prefill_tail_uses_rollout_shape(self):
+    model = Transformer(TEST_CONFIG)
+    model.has_recurrent_block = True
+    model.config = replace(model.config, recurrent_prefill_chunked=True)
+    calls = []
+    def mock_call(self, tokens, start_pos, temperature, **kwargs):
+      calls.append(tokens.shape)
+      return Tensor([[42]])
+    with patch.object(Transformer, '__call__', mock_call): next(model.generate([1, 2, 3, 4, 5, 6], chunk_size=4))
+    self.assertEqual(calls, [(1, 4), (1, 1), (1, 1)])
+
   def test_recurrent_live_state_reuse(self):
     model = Transformer(TEST_CONFIG)
     model.has_recurrent_block = True
@@ -44,20 +68,37 @@ class TestTransformerGenerate(unittest.TestCase):
       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):
+  def test_recurrent_prompt_snapshot_reuse(self):
+    model = Transformer(TEST_CONFIG)
+    model.has_recurrent_block = True
+    state, calls = Tensor.ones(4).realize(), []
+    def mock_call(self, tokens, start_pos, temperature, **kwargs):
       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))
+    with patch.object(model, "_state_tensors", return_value=[state]), patch.object(model.blk[0], "_reusable_prefix_len", return_value=0), \
+         patch.object(Transformer, '__call__', mock_call):
+      next(model.generate([1, 2, 3]))
+      state.assign(state.const_like(5)).realize()
+      model._cached_tokens = [1, 2, 3, 9, 9]
+      calls.clear()
+      self.assertEqual(model.get_start_pos([1, 2, 3, 7, 8]), 3)
+      next(model.generate([1, 2, 3, 7, 8]))
+    self.assertEqual(calls, [V_START_POS.bind(3), V_START_POS.bind(4)])
+    self.assertEqual(state.tolist(), [1.0] * 4)
 
   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_kimi_tool_call_stream(self):
+    router = StreamRouter()
+    self.assertEqual(list(router.route("before<|tool_calls_section_beg")), [("content", "before")])
+    self.assertEqual(list(router.route("in|><|tool_call_begin|>functions.read:0<|tool_call_argument_begin|>"
+                                      '{"path":"/tmp/x"}<|tool_call_end|><|tool_calls_section_end|>')), [])
+    self.assertEqual(parse_kimi_tool_call("functions.read:0<|tool_call_argument_begin|>{\"path\":\"/tmp/x\"}"),
+                     ("read", {"path":"/tmp/x"}))
+
   def test_kv_cache_reuse(self):
     """Test that generate reuses the KV cache when tokens extend the cached prefix."""
     model = Transformer(TEST_CONFIG)
@@ -193,12 +234,6 @@ 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)

test/unit/test_metal_graph.py

@@ -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()

test/unit/test_multitensor.py

@@ -390,12 +390,6 @@ class TestMultiTensor(unittest.TestCase):
     self.assertEqual(out.shape, (rows, 8))
     np.testing.assert_equal(out[:3].to(Device.DEFAULT).numpy(), np.ones((3, 8)))
 
-  def test_symbolic_broadcast_consumed(self):
-    rows = Variable("rows", 1, 4).bind(3)
-    out = (Tensor.ones(rows).to(devices_2) + 1).realize()
-    self.assertEqual(out.shape, (rows,))
-    np.testing.assert_equal(out[:3].to(Device.DEFAULT).numpy(), np.full(3, 2))
-
   def test_multitensor_jit_in_list(self):
     # test MULTI tensor inside a list container - exercises the container unpacking + MULTI unpacking
     @TinyJit

test/unit/test_schedule_cache.py

@@ -1,16 +1,11 @@
 import unittest
 import functools
-from tinygrad import Tensor, Variable, UOp, function
+from tinygrad import Tensor, Variable, UOp
 from tinygrad.uop.ops import KernelInfo
 from tinygrad.schedule import schedule_cache
 
-def custom_add_kernel(A:UOp, B:UOp, num:int=0) -> UOp:
-  return A[0].set(B[0] + num).sink(arg=KernelInfo(f"custom_add_{num}"))
-
-def custom_add_backward(grad_output:UOp, _) -> tuple[None, UOp]:
-  grad = Tensor.invalids(*grad_output.shape, dtype=grad_output.dtype, device=grad_output.device)
-  grad = Tensor.custom_kernel(grad, Tensor(grad_output, device=grad_output.device), fxn=functools.partial(custom_add_kernel, num=0))[0]
-  return None, grad.uop
+def custom_set0_kernel(A:UOp, num:int) -> UOp:
+  return A[0].set(num).sink(arg=KernelInfo(f"custom_set0_{num}"))
 
 class TestScheduleCache(unittest.TestCase):
   def test_bound_variable_reuses_cache(self):
@@ -30,27 +25,27 @@ class TestScheduleCache(unittest.TestCase):
 
   def test_custom_kernel(self):
     for i in range(4):
-      a, b = Tensor.empty(1), Tensor.ones(1)
-      a = Tensor.custom_kernel(a, b, fxn=functools.partial(custom_add_kernel, num=i))[0]
+      a = Tensor.empty(1)
+      a = Tensor.custom_kernel(a, fxn=functools.partial(custom_set0_kernel, num=i))[0]
       a.realize()
-      self.assertEqual(a.item(), i+1)
+      self.assertEqual(a.item(), i)
 
   def test_same_custom_function_reuses_cache(self):
     schedule_cache.clear()
-    fxn = functools.partial(custom_add_kernel, num=10)
+    fxn = functools.partial(custom_set0_kernel, num=10)
 
     # first run
-    a, x = Tensor.empty(1), Tensor.ones(1)
-    a = Tensor.custom_kernel(a, x, fxn=fxn)[0]
+    a = Tensor.empty(1)
+    a = Tensor.custom_kernel(a, fxn=fxn)[0]
     a.realize()
-    self.assertEqual(a.item(), 11)
+    self.assertEqual(a.item(), 10)
     cache_size_after_first = len(schedule_cache)
 
     # second run with same function should reuse cache
-    b, x = Tensor.empty(1), Tensor.ones(1)
-    b = Tensor.custom_kernel(b, x, fxn=fxn)[0]
+    b = Tensor.empty(1)
+    b = Tensor.custom_kernel(b, fxn=fxn)[0]
     b.realize()
-    self.assertEqual(b.item(), 11)
+    self.assertEqual(b.item(), 10)
     self.assertEqual(len(schedule_cache), cache_size_after_first)
 
   def test_simple(self):
@@ -70,29 +65,5 @@ class TestScheduleCache(unittest.TestCase):
       print(num)
     self.assertEqual(len(schedule_cache), start_len_schedule_cache)
 
-  def test_simple_precompile(self):
-    @function(precompile=True, precompile_backward=True)
-    def f(x:Tensor) -> Tensor:
-      out = Tensor.invalids(*x.shape, dtype=x.dtype, device=x.device)
-      out = Tensor.custom_kernel(out, x, fxn=functools.partial(custom_add_kernel, num=10), grad_fxn=custom_add_backward)[0]
-      return out + x
-
-    # warmup
-    x = Tensor.ones(1).realize()
-    out = f(x)
-    out.backward(x)
-    self.assertEqual(out.item(), 12)
-    self.assertEqual(x.grad.item(), 2)
-
-    # use the cache next time function is called
-    start_len_schedule_cache = len(schedule_cache)
-    for _ in range(3):
-      x = Tensor.ones(1).realize()
-      out = f(x)
-      out.backward(x)
-      self.assertEqual(out.item(), 12)
-      self.assertEqual(x.grad.item(), 2)
-    self.assertEqual(len(schedule_cache), start_len_schedule_cache)
-
 if __name__ == "__main__":
   unittest.main()

tinygrad/codegen/__init__.py

@@ -5,14 +5,14 @@ from tinygrad.helpers import ALLOW_TF32, DEFAULT_FLOAT, DEFAULT_INT, TracingKey,
 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.render import pyrender
-from tinygrad.uop.spec import type_verify, spec_tensor, spec_program, spec_program_casted_consts
+from tinygrad.uop.spec import type_verify, spec_tensor, spec_program
 from tinygrad.renderer import Renderer, Estimates
 from tinygrad.renderer.isa import ISARenderer, IselContext, PreRegAllocContext
 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_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
+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.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
@@ -153,8 +153,8 @@ devectorizer2 = mop_cleanup+pm_mops+PatternMatcher([
   # unpack WMMA
   (UPat(Ops.WMMA, name="u"), do_stack_wmma),
   # stacked INDEX is many INDEX
-  (UPat(Ops.INDEX, src=(UPat((Ops.PARAM, Ops.BUFFER), name="b"), UPat(Ops.STACK, name="s")), name="x"),
-   lambda b,s,x: UOp.stack(*[x.replace(src=(b,u)) for u in s.src])),
+  (UPat(Ops.INDEX, src=(UPat((Ops.PARAM, Ops.BUFFER), name="b"), UPat(Ops.STACK, name="s"))),
+   lambda b,s: UOp.stack(*[b.index(u) for u in s.src])),
   # INDEX into RESHAPE moves the RESHAPE
   (UPat(Ops.INDEX, src=(UPat((Ops.PARAM, Ops.BUFFER), name="b"), UPat(Ops.RESHAPE, name="s"))),
    lambda b,s: b.index(s.src[0]).reshape(s.shape)),
@@ -281,10 +281,6 @@ pm_implicit_barriers = PatternMatcher([
   (UPat(Ops.END, name="end"), add_war_barrier),
 ])
 
-pm_casted_consts = PatternMatcher([
-  (UPat(Ops.CONST, dtypes.all, name="c"), lambda c: UOp(Ops.CAST, c.dtype, src=(UOp.const(c.val),), arg=c.dtype)),
-])
-
 def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   if VIZ: graph_rewrite(ast, PatternMatcher([]), name="View Base AST")
   if DEBUG >= 5: print(pyrender(ast))
@@ -305,7 +301,7 @@ 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_flatten_range, name="initial symbolic")
+    sink = graph_rewrite(sink, sym+pm_fold_cast_const+pm_flatten_range, name="initial symbolic")
 
     # optimize (schedule) the AST
     sink = graph_rewrite(sink, pm_flatten_range+pm_simplify_ranges, ctx={}, name="simplify ranges")
@@ -350,7 +346,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
 
   # lower index dtype
   # NOTE: we need indexing_simplify to remove the cast to long using the Invalid
-  sink = graph_rewrite(sink, symbolic_simple+pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")
+  sink = graph_rewrite(sink, symbolic_simple+pm_fold_cast_const+pm_lower_index_dtype+indexing_simplify, ctx={}, name="lower all index dtypes")
 
   # final symbolic before decomp
   sink = graph_rewrite(sink, symbolic, name="final symbolic")
@@ -361,7 +357,7 @@ 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+get_simplifying_rewrite_patterns(supported_ops)
+  pm_decomp = symbolic_simple+pm_fold_cast_const+get_simplifying_rewrite_patterns(supported_ops)
   sink = graph_rewrite(sink, pm_decomp, name="early decompositions")
 
   # late decomps + move gates from unrenderable INVALID where
@@ -387,11 +383,8 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   num_params = len([x for x in sink.toposort() if x.op is Ops.PARAM and x.arg.slot != -1])
   sink = graph_rewrite(sink, pm_number_params, ctx=[num_params], name="number params with -1", walk=True)
 
-  # TODO: delete once migration are done
-  if ren.casted_consts: sink = graph_rewrite(sink, pm_casted_consts, name="casted consts", walk=True)
-
   if VIZ: graph_rewrite(sink, PatternMatcher([]), name="View Output AST")
-  if SPEC: type_verify(sink, spec_program_casted_consts if ren.casted_consts else spec_program)
+  if SPEC: type_verify(sink, spec_program)
 
   # return the rewritten sink
   return sink

tinygrad/codegen/decomp/dtype.py

@@ -33,8 +33,7 @@ def l2i(op: Ops, dt: DType, *uops:UOp):
       return (lo:=uops[0].cast(l2i_dt[dt])), (uops[0] / 2**32).cast(l2i_dt[dt]) - ((uops[0] < 0) & lo.ne(0))
     case Ops.CAST if dt in dtypes.floats:
       small = (a1.eq(0) & (a0 >= 0)) | (a1.eq(-1) & (a0 < 0))
-      cdt = dt if dt == dtypes.float64 else dtypes.float32
-      return small.where(a0.cast(dt), ((a1.cast(cdt) * (2**32)) + a0.bitcast(dtypes.uint).cast(cdt)).cast(dt))
+      return small.where(a0.cast(dt), ((a1.cast(dtypes.float32) * (2**32)) + a0.bitcast(dtypes.uint).cast(dtypes.float32)).cast(dt))
     case Ops.CAST: return a0.bitcast(dtypes.uint).cast(dt)
     case Ops.BITCAST: return a0.bitcast(dt), a1.bitcast(dt)
     case Ops.SHL:

tinygrad/codegen/gpudims.py

@@ -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),

tinygrad/codegen/late/coalesce.py

@@ -26,7 +26,7 @@ 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)
+      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype)
       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))
@@ -149,7 +149,7 @@ def memory_coalescing(sink:UOp, ctx:Renderer) -> UOp:
         grp = full_grp[:length]
         # NOTE: we apply the valid again after we determine the length
         offset = offset.valid(valid) if valid is not None else offset
-        idx = UOp(Ops.SHRINK, src=(buf, offset, UOp.const(len(grp)))) if len(grp) > 1 else buf.index(offset, dtype=offsets[grp[0]][0].src[0].dtype)
+        idx = UOp(Ops.SHRINK, src=(buf, offset, UOp.const(len(grp)))) if len(grp) > 1 else buf.index(offset)
         if op == Ops.STORE:
           datas = []
           for i,g in enumerate(grp):

tinygrad/codegen/late/gater.py

@@ -3,7 +3,7 @@ from tinygrad.uop.ops import PatternMatcher, UPat, Ops
 from tinygrad.dtype import Invalid, dtypes
 
 def move_where_load(gate, l, a, w):
-  return l.replace(src=(l.src[0], l.vconst_like(0) if a.is_invalid else l.const_like(a.val) if a.op is Ops.CONST else
+  return l.replace(src=(l.src[0], l.vconst_like(0) if a.is_invalid else
     a.src[0] if a.op is Ops.CAST and a.src[0].dtype == l.dtype else a.cast(l.dtype), l.src[2])).cast(w.dtype)
 
 pm_move_gates_from_index = PatternMatcher([

tinygrad/codegen/opt/tc.py

@@ -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,

tinygrad/codegen/simplify.py

@@ -1,7 +1,7 @@
 import itertools
 from typing import Callable
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, graph_rewrite, _substitute, range_start, AxisType
-from tinygrad.uop.symbolic import symbolic, invalid_gate
+from tinygrad.uop.symbolic import symbolic, pm_fold_cast_const, invalid_gate
 from tinygrad.helpers import partition
 from tinygrad.dtype import dtypes
 
@@ -32,7 +32,7 @@ def simplify_merge_adjacent(u:UOp) -> UOp|None:
         s0, s1 = r0.src[0], r1.src[0]
         # do the merge
         new_range = r0.replace(src=(s0*s1,))
-        nidx = graph_rewrite(u, _substitute+symbolic+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
+        nidx = graph_rewrite(u, _substitute+symbolic+pm_fold_cast_const+pm_flatten_range, ctx={r0:new_range//s1, r1:new_range%s1},
                              name=f"check_merge_{r0.arg[0]}_{r1.arg[0]}")
 
         # check if it simplifies
@@ -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

tinygrad/device.py

@@ -1,8 +1,8 @@
 from __future__ import annotations
 from dataclasses import dataclass, replace
 from collections import defaultdict
-from typing import Any, Callable, Generic, TypeVar, Iterator, Generator, Self, TYPE_CHECKING
-import importlib, inspect, functools, pathlib, os, contextlib, re, atexit, pickle, decimal, subprocess, struct
+from typing import Any, Generic, TypeVar, Iterator, Generator, Self, TYPE_CHECKING
+import importlib, inspect, functools, pathlib, os, contextlib, re, atexit, pickle, decimal
 from tinygrad.helpers import LRU, getenv, diskcache_get, diskcache_put, DEBUG, GlobalCounters, PROFILE, temp, colored
 from tinygrad.helpers import Context, CCACHE, ALLOW_DEVICE_USAGE, MAX_BUFFER_SIZE, cpu_events, ProfileEvent, ProfilePointEvent, suppress_finalizing
 from tinygrad.helpers import select_by_name, select_first_inited, DEV, TracingKey, size_to_str, pluralize, Target, unwrap, round_up
@@ -103,7 +103,7 @@ class Buffer:
   def __init__(self, device:str, size:int, dtype:DType, opaque:Any=None, options:BufferSpec|None=None,
                initial_value:bytes|pickle.PickleBuffer|None=None, uop_refcount=0, base:Buffer|None=None, offset:int=0, preallocate=False):
     assert isinstance(dtype, DType)
-    self.device, self.size, self.dtype, self.options, self.offset, self.allocated_views = Device.canonicalize(device), size, dtype, options, offset, 0
+    self.device, self.size, self.dtype, self.options, self.offset, self.allocated_views = device, size, dtype, options, offset, 0
     self._bufs: dict[str, Any] = {}
     if base is None:
       assert offset == 0, "base buffers can't have offset"
@@ -116,7 +116,7 @@ class Buffer:
         if isinstance(initial_value, pickle.PickleBuffer): initial_value.release()
     else:
       assert base._base is None, "base can't have a base"
-      assert self.device == base.device, "base must have the same device"
+      assert device == base.device, "base must have the same device"
       self._base = base
     if preallocate: self.allocate()
   @property
@@ -133,7 +133,7 @@ class Buffer:
   # check if the underlying buffer is allocated, possibly from the base object
   def is_allocated(self) -> bool: return self.base.is_allocated() if self._base is not None else self.device in self._bufs
   def get_buf(self, device: str) -> Any:
-    if device not in self._bufs and (device:=Device.canonicalize(device)) not in self._bufs:
+    if device not in self._bufs:
       allocator = Device[device].allocator
       if device == self.device: self.ensure_allocated()
       elif self._base is not None: self._bufs[device] = allocator._offset(self._base.get_buf(device), self.nbytes, self.offset)
@@ -310,14 +310,6 @@ class Compiler:
       if self.cachekey is not None: diskcache_put(self.cachekey, src, lib)
     return lib
   def disassemble(self, lib:bytes): pass
-  def server(self, cmd:str, arch:str, *args) -> subprocess.Popen:
-    argv = f"{cmd} {pathlib.Path(__file__).parent}/runtime/support/compileserver.py {type(self).__module__}:{type(self).__name__} {arch}"
-    return subprocess.Popen(argv.split() + [str(a) for a in args], stdout=subprocess.PIPE, stdin=subprocess.PIPE, bufsize=0)
-  def compile_server(self, src:str, proc:subprocess.Popen) -> bytes:
-    unwrap(proc.stdin).write(struct.pack("I", len(src.encode())) + src.encode())
-    if (lib:=unwrap(proc.stdout).read(struct.unpack("I", unwrap(proc.stdout).read(4))[0])): return lib
-    raise CompileError("Compilation Error")
-
 
 @dataclass
 class TinyELF:
@@ -339,18 +331,15 @@ class Program(Generic[DeviceType]):
                wait=False) -> float|None: pass
 
 class Compiled:
-  ifaces:list[Callable] = []
   profile_events:list[ProfileEvent] = [ProfileDeviceEvent("CPU")] # NOTE: CPU is the default device.
 
-  has_copy_queue:bool = True
-
   pm_lower:Any = None
   pm_bufferize:Any = None
 
   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]
-    self.device_id, self.arch = (int(idx) if ":" in device and (idx:=device.split(":")[1]).isdigit() else 0), arch
+    self.arch = arch
     self.cached_renderer:dict[Any, Renderer] = {}
 
   @property
@@ -373,21 +362,11 @@ class Compiled:
     return select_first_inited(select_by_name(self.renderers, self._renderer_name, t.renderer, f"{self.device} has no renderer {t.renderer!r}"),
                                f"No renderer for {self.device} is available", self.cached_renderer, t)
 
-  def _select_iface(self, device:str):
-    self.device_id = int(device.split(":")[1]) if ":" in device else 0
-    assert (v:=getenv(k:=f'{type(self).__name__[:-6].upper()}_IFACE', "")) == "",  \
-      f"{k}={v} is deprecated, use DEV={replace(DEV.target(type(self).__name__[:-6]), interface=v)} instead"
-    t = DEV.target(dev:=type(self).__name__[:-6])
-    filtered = select_by_name(self.ifaces, lambda i: i.__name__[:-5], t.interface, f"{dev} has no interface {t.interface!r}")
-    filtered = [i for i in filtered if t.interface.startswith("MOCK") or not i.__name__[:-5].startswith("MOCK")] # never fallback to mock ifaces
-    return select_first_inited([functools.partial(iface, self, self.device_id) for iface in filtered],
-                               f"No interface for {dev}:{self.device_id} is available")
-
   def count(self) -> int:
     """
     Returns the number of physical accelerators available to the runtime.
     """
-    return self.iface.count if hasattr(self, 'iface') else 1
+    return 1
 
   def synchronize(self):
     """
@@ -405,7 +384,7 @@ class Compiled:
     """
     Called at the end of process lifetime to allow the device to finalize.
     """
-    if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()
+    # override this in your device implementation
 
 if PROFILE:
   @atexit.register
@@ -427,7 +406,7 @@ def enumerate_devices_str() -> Generator[str, None, None]:
     ren_results, iface_results = [], []
     try:
       d = Device[device]
-      for iface in [i for i in d.ifaces if not i.__name__.startswith("MOCK")]:
+      for iface in [i for i in getattr(d, 'ifaces', []) if not i.__name__.startswith("MOCK")]:
         try:
           name = iface.__name__[:-5]
           default_text, count = ("(default)", d.count()) if type(d.iface) is iface else (f"(DEV={name}+{device} to make default)", iface(d, 0).count) # type: ignore

tinygrad/dtype.py

@@ -66,6 +66,7 @@ class DType(metaclass=DTypeMetaClass):
   def __reduce__(self): return type(self), tuple(getattr(self, f.name) for f in fields(self))
   def __repr__(self): return f"dtypes.{INVERSE_DTYPES_DICT[self.name]}"
   def __lt__(self, o:DType): return (self.priority, self.bitsize, self.name, self.fmt) < (o.priority, o.bitsize, o.name, o.fmt)
+  def scalar(self) -> DType: return self
   @functools.cached_property
   def min(self):
     if dtypes.is_int(self): return 0 if dtypes.is_unsigned(self) else -2**(self.bitsize-1)

tinygrad/engine/jit.py

@@ -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)
@@ -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:
@@ -269,14 +271,10 @@ class _TinyJit(Generic[ReturnType]):
         big_linear, onetime_linear = prune_linear(big_linear, set(input_buf_uops))
         if DEBUG >= 1: print(f"pruned from {len(big_linear.src) + len(onetime_linear.src)} -> {len(big_linear.src)} kernels")
         run_linear(onetime_linear, var_vals)
-        del onetime_linear
 
       # hold all buffers reachable from live Tensors (e.g. lazy .grad created during capture), the memory planner can't suballocate those
       held_bufs = set(buffers) | {u for tref in list(all_tensors) if (t:=tref()) is not None for u in t.uop.toposort() if u.op is Ops.BUFFER}
       linear = jit_lower(big_linear, held_bufs, input_buf_uops)
-      # drop the pre-planning graph: it keeps the whole capture-time working set allocated (big_linear) or referenced (held_bufs).
-      # the planned linear only uses the arena/held buffers, so the intermediates must be freed before linking and first exec
-      del big_linear, held_bufs
       self.captured = CapturedJit(ret, linear, names, expected_input_info)
       ret = self.captured(input_buf_uops, var_vals)
     elif self.cnt >= 2:

tinygrad/engine/realize.py

@@ -3,24 +3,21 @@ from typing import cast, Iterator, Any, Sequence
 import time, random, itertools, math, contextlib, weakref, array
 from dataclasses import dataclass, replace, field
 from tinygrad.helpers import colored, DEBUG, GlobalCounters, ansilen, all_int, prod, flatten, Context, getenv, to_tuple
-from tinygrad.helpers import BEAM, size_to_str, time_to_str, VALIDATE_WITH_CPU, PROFILE, ProfilePointEvent, cpu_events
-from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, AxisType, sym_infer, graph_rewrite
-from tinygrad.device import Device, Buffer, MultiBuffer, ProfileGraphEntry
-from tinygrad.dtype import dtypes
+from tinygrad.helpers import BEAM, size_to_str, time_to_str, VALIDATE_WITH_CPU, PROFILE, ProfilePointEvent, cpu_events, wait_cond
+from tinygrad.uop.ops import Ops, PatternMatcher, UOp, UPat, AxisType, sym_infer, buffers, graph_rewrite
+from tinygrad.device import Device, Buffer, MultiBuffer
 from tinygrad.renderer import Estimates
 from tinygrad.codegen import to_program
 from tinygrad.codegen.opt.postrange import args_from_ast
 
 # **************** Helpers ****************
 
-def get_call_arg_uops(call:UOp) -> tuple[UOp, ...]: return tuple(s for s in call.src[1:] if not s.is_bound_var)
-def get_call_var_uops(call:UOp, prg:UOp) -> list[UOp]:
-  bound = {s.src[0].expr: s.src[1].src[1] for s in call.src[1:] if s.is_bound_var}
-  return [bound.get(v.expr, v) for v in prg.arg.vars]
+def get_call_arg_uops(call:UOp) -> tuple[UOp, ...]: return tuple(s for s in call.src[1:] if s.op is not Ops.BIND)
+
 def get_call_outs_ins(call:UOp) -> tuple[tuple[int, ...], tuple[int, ...]]:
   ast = call.src[0]
   if ast.op is Ops.PROGRAM: return tuple(ast.arg.outs), tuple(ast.arg.ins)
-  if ast.op is Ops.COPY: return (0,), (1,)
+  if ast.op in (Ops.COPY, Ops.SLICE): return (0,), (1,)
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return (0,), tuple(range(1, len(get_call_arg_uops(call))))
   return (), ()
 
@@ -30,6 +27,9 @@ def get_call_name(call:UOp, bufs:Sequence[Buffer|UOp], var_vals:dict[str, int]|N
 
   ast, arg_uops = call.src[0], get_call_arg_uops(call)
   if ast.op is Ops.PROGRAM: return ast.arg.name
+  if ast.op is Ops.SLICE:
+    offset = ast.src[1].val * arg_uops[1].dtype.itemsize
+    return colored(f"view {_uop_sz_to_str(arg_uops[0]):>10} @ {offset:<10d}", "yellow")
   if ast.op is Ops.COPY: return colored(f"copy {_uop_sz_to_str(arg_uops[0]):>10}, {_dev_str(bufs[0]):>7s} <- {_dev_str(bufs[1]):7s}", "yellow")
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "encdec": return colored(f"enc/dec {_uop_sz_to_str(arg_uops[0])}", "yellow")
   if ast.op is Ops.CUSTOM_FUNCTION and ast.arg == "graph": return colored(f"batched {len(ast.src[0].src)}", "cyan")
@@ -140,7 +140,7 @@ class ExecContext:
   cache: bool = True
 
 def _resolve(b:UOp, inputs:tuple[UOp, ...]) -> UOp:
-  if b.op in (Ops.MSELECT, Ops.SHRINK) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg.slot], *b.src[1:]))
+  if b.op in (Ops.SLICE, Ops.MSELECT) and b.src[0].op is Ops.PARAM: return b.replace(src=(inputs[b.src[0].arg.slot], *b.src[1:]))
   if b.op is Ops.MSTACK: return b.replace(src=tuple(_resolve(x, inputs) for x in b.src))
   return inputs[b.arg.slot] if b.op is Ops.PARAM else b
 def resolve_params(call:UOp, inputs:tuple[UOp, ...]) -> list[UOp]: return [_resolve(b, inputs) for b in get_call_arg_uops(call)]
@@ -154,6 +154,13 @@ def unwrap_multi(call:UOp, resolved:list[UOp]) -> Iterator[tuple[list[Buffer], d
                    for x in call.src[0].toposort())
     for j, per_dev in enumerate(zip(*[cast(MultiBuffer, b).bufs for b in bufs])): yield list(per_dev), {"_device_num": j} if has_dnum else {}
 
+def exec_view(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
+  resolved = resolve_params(call, ctx.input_uops)
+  bufs = [cast(Buffer, b.buffer) for b in resolved]
+  bv = bufs[1].view(resolved[0].max_numel(), ast.dtype, ast.src[1].val*bufs[1].dtype.itemsize)
+  with track_stats(ctx, call, bv.device, [bv, bufs[1]], ctx.var_vals): buffers[resolved[0]] = bv
+  return None
+
 def exec_copy(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
   for bufs, device_vars in unwrap_multi(call, resolve_params(call, ctx.input_uops)):
     dest, src = bufs[0].ensure_allocated(), bufs[1].ensure_allocated()
@@ -169,10 +176,9 @@ def exec_copy(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
 
 def exec_kernel(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
   et = None
-  resolved = resolve_params(call, ctx.input_uops)
-  for device, (bufs, device_vars) in zip(to_tuple(call.src[1].device), unwrap_multi(call, [resolved[i] for i in ast.arg.globals])):
+  for device, (bufs, device_vars) in zip(to_tuple(call.src[1].device), unwrap_multi(call, resolve_params(call, ctx.input_uops))):
     var_vals = {**ctx.var_vals, **device_vars}
-    prg_bufs = [b.ensure_allocated() for b in bufs]
+    prg_bufs = [bufs[i].ensure_allocated() for i in ast.arg.globals]
     rt = get_runtime(device, ast, cache=ctx.cache)
     global_size, local_size = ast.arg.launch_dims(var_vals)
     with track_stats(ctx, call, device, prg_bufs, var_vals) as tm:
@@ -204,28 +210,27 @@ def exec_graph(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
   return t[0]
 
 def exec_hcq(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
-  dev = cast(Any, Device[(info:= call.arg.aux).device[0]])
-  addrs = [(b.bufs[j] if isinstance(b:=_resolve(ctx.input_uops[k], ctx.input_uops).buffer, MultiBuffer) else b).get_buf(dev_name).va_addr
-           for devs, idxs in info.input_idxs for j, dev_name in enumerate(devs) for k in idxs]
-  dev.rt_buffer._buf.cpu_view().view(offset=(base:=dev.rt_allocator.alloc(len(addrs) * 8)), fmt='Q')[:len(addrs)] = array.array('Q', addrs)
+  if (inputs:=call.arg.aux.inputs) is not None:
+    bufs = [_resolve(ctx.input_uops[i], ctx.input_uops).buffer for i in call.arg.aux.input_idxs]
+    table = call.src[1+inputs].buffer
+    for j,dev in enumerate(call.arg.aux.device):
+      addrs = array.array('Q', [(b.bufs[j] if isinstance(b, MultiBuffer) else b).get_buf(dev).va_addr for b in bufs])
+      mv = (table.bufs[j] if isinstance(table, MultiBuffer) else table).ensure_allocated()._buf.cpu_view().view(fmt='Q')
+      wait_cond(lambda: mv[0], value=0, timeout_ms=ctx.timeout or getenv("HCQDEV_WAIT_TIMEOUT_MS", 30000), msg=f"{dev} hang detected")
+      mv[:len(addrs)] = addrs
 
-  tables = [UOp.from_buffer(dev.rt_buffer.view(len(idxs), dtypes.uint64, base + j*len(idxs)*8), HCQ_RUNTIME_DEV.value)
-            for devs, idxs in info.input_idxs for j in range(len(devs))]
-  if info.inputs is not None: call = call.substitute({call.src[1+info.inputs]: UOp.mstack(*tables)})
-  exec_kernel(replace(ctx, update_stats=DEBUG>=3, var_vals={**ctx.var_vals, "hcq_inputs_ptr": dev.rt_buffer._buf.va_addr + base}), call, ast)
+  exec_kernel(replace(ctx, update_stats=False), call, ast)
 
-  tms = []
-  for devices, stat_call, prof in info.kernels:
-    for device in devices:
-      tm = None
-      if prof:
-        (d:=cast(Any, Device[device])).prof_ents[prof[0]] = ProfileGraphEntry(device, stat_call.arg.name, *prof)
-        if ctx.wait:
-          d.synchronize(timeout=ctx.timeout)
-          st, en = (d.signal(x)._buf.cpu_view().view(fmt='Q')[0] for x in prof)
-          tms.append(tm:=float(en-st)/d.timestamp_divider/1e6)
-      with track_stats(ctx, stat_call, device, [], ctx.var_vals) as et: et[0] = tm
-  return max(tms) if tms else None
+  tms:list[float|None] = []
+  for e in (aux:=call.arg.aux).prof: cast(Any, Device[e.device]).prof_ents[e.st_id] = e
+  for d in [cast(Any, Device[x]) for x in aux.device]:
+    with track_stats(ctx, call, d.device, [], ctx.var_vals) as et:
+      if ctx.wait:
+        d.synchronize(timeout=ctx.timeout)
+        ts = [d.signal(i)._buf.cpu_view().view(fmt='Q')[0] for e in aux.prof if e.device == d.device for i in (e.st_id, e.en_id)]
+        if ts: et[0] = float(max(ts)-min(ts))/d.timestamp_divider/1e6
+      tms += et
+  return tms[0]
 
 # flatten LINEAR-in-LINEAR: any nested LINEAR child gets inlined into its parent's src
 pm_flatten_linear = PatternMatcher([
@@ -256,6 +261,7 @@ pm_optimize_local_size = PatternMatcher([
 ])
 
 pm_exec = PatternMatcher([
+  (UPat(Ops.CALL, src=(UPat(Ops.SLICE, name="ast"),), name="call", allow_any_len=True), exec_view),
   (UPat(Ops.CALL, src=(UPat(Ops.COPY, name="ast"),), name="call", allow_any_len=True), exec_copy),
   (UPat(Ops.CALL, src=(UPat(Ops.PROGRAM, name="ast"),), name="call", allow_any_len=True), exec_kernel),
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="encdec", name="ast"),), name="call", allow_any_len=True), exec_encdec),
@@ -264,15 +270,14 @@ pm_exec = PatternMatcher([
   (UPat(Ops.CALL, src=(UPat(Ops.CUSTOM_FUNCTION, arg="validate", name="ast"),), name="call", allow_any_len=True), exec_validate),
 ])
 
-if getenv("HCQ2"): from tinygrad.runtime.support.hcq2 import hcq_compile, hcq_link, HCQ_RUNTIME_DEV # noqa: E402 # down here, hcq2 imports realize
+if getenv("HCQ2"): from tinygrad.runtime.support.hcq2 import hcq_compile, hcq_link # noqa: E402 # down here, hcq2 imports the helpers above
 
 def compile_linear(linear:UOp, beam:int|None=None, validate=False, input_uops:list[UOp]|None=None, profile:bool|None=None) -> UOp:
   if validate: linear = graph_rewrite(linear, pm_validate, name="validate", walk=True)
   if (beam_val:=BEAM.value if beam is None else beam) >= 1: linear = graph_rewrite(linear, pm_beam, ctx=beam_val, walk=True)
   linear = graph_rewrite(linear, pm_compile, name="precompile kernels", walk=True)
-  linear = graph_rewrite(linear, pm_optimize_local_size, name="optimize local size", walk=True)
-  if getenv("HCQ2"): linear = hcq_compile(linear, input_uops, bool(PROFILE or DEBUG >= 2) if profile is None else profile)
-  return linear
+  if getenv("HCQ2"): linear = hcq_compile(linear, input_uops, bool(PROFILE) if profile is None else profile)
+  return graph_rewrite(linear, pm_optimize_local_size, name="optimize local size", walk=True)
 
 def link_linear(linear:UOp, cache=True) -> UOp: return hcq_link(linear, cache=cache) if getenv("HCQ2") else linear
 

tinygrad/function.py

@@ -1,5 +1,4 @@
 import functools, time
-from dataclasses import replace
 from typing import Generic, TypeVar, Callable, cast, overload
 from tinygrad.helpers import Context, dedup, getenv, DEBUG
 from tinygrad.uop.ops import UOp, Ops, graph_rewrite, PatternMatcher, UPat
@@ -13,7 +12,7 @@ def add_to_ctx(ctx, x:UOp):
   return ret
 
 pm_ctx = PatternMatcher([
-  (UPat(Ops.BUFFER, name="x"), add_to_ctx),
+  (UPat((Ops.BUFFER, Ops.BIND), name="x"), add_to_ctx),
   (UPat((Ops.AFTER, Ops.CONTIGUOUS), name="x"),
    lambda ctx,x: add_to_ctx(ctx,x) if not x.op_in_backward_slice_with_self(Ops.PARAM) and x.op_in_backward_slice_with_self(Ops.BUFFER) else None),
 ])
@@ -24,10 +23,6 @@ def invalid_outputs(uret:UOp) -> set[UOp]:
   return {u.src[0].buf_uop for u in uret.backward_slice_with_self
           if u.op is Ops.STORE and u.src[1].base.is_invalid and not u.src[0].buf_uop.is_realized}
 
-def renumber_invalid_outputs(uret:UOp) -> UOp:
-  return uret.substitute({b:b.replace(arg=replace(b.arg, slot=i))
-                          for i,b in enumerate(x for x in uret.toposort(enter_calls=False) if x in invalid_outputs(uret))})
-
 ReturnType = TypeVar('ReturnType')
 class _function(Generic[ReturnType]):
   depth = 0
@@ -70,7 +65,6 @@ class _function(Generic[ReturnType]):
     # the BUFFERs that are left are the implicit inputs
     num_explicit = len(call_uops)
     uret = graph_rewrite(uret, pm_ctx, (call_uops, invalid_outputs(uret)), bottom_up=True, name="get_implicit_inputs")
-    uret = renumber_invalid_outputs(uret)
     name = getattr(self.fxn, '__qualname__', None) or type(self.fxn).__qualname__
     if not self.allow_implicit:
       implicit_buffers = [x for x in call_uops[num_explicit:] if x.op is Ops.BUFFER]

tinygrad/helpers.py

@@ -1,9 +1,9 @@
 from __future__ import annotations
 import time
 START_TIME = time.perf_counter()
-import os, functools, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass, gc
+import os, functools, platform, re, contextlib, operator, hashlib, pickle, sqlite3, tempfile, pathlib, string, ctypes, sys, gzip, getpass, gc
 from collections import defaultdict
-import shutil, math, types, copyreg, inspect, importlib, decimal, itertools, difflib
+import subprocess, shutil, math, types, copyreg, inspect, importlib, decimal, itertools, difflib
 from dataclasses import dataclass, field, replace
 from typing import ClassVar, Iterable, Any, TypeVar, Callable, Sequence, TypeGuard, Iterator, Generic, Generator, cast, overload
 
@@ -13,7 +13,8 @@ U = TypeVar("U")
 def prod(x:Iterable[T]) -> T|int: return functools.reduce(operator.mul, x, 1)
 
 # NOTE: helpers is not allowed to import from anything else in tinygrad
-OSX, WIN = sys.platform == "darwin", sys.platform == "win32"
+OSX, WIN = platform.system() == "Darwin", sys.platform == "win32"
+ARCH_X86 = any(x in platform.processor() for x in ("Intel", "i386", "x86_64"))
 BASEDIR = pathlib.Path(__file__).parent
 
 # fix colors on Windows, https://stackoverflow.com/questions/12492810/python-how-can-i-make-the-ansi-escape-codes-to-work-also-in-windows
@@ -230,7 +231,7 @@ class _DEV(ContextVar):
 
 DEV, DEBUG, BEAM, NOOPT = _DEV("DEV", ""), ContextVar("DEBUG", 0), ContextVar("BEAM", 0), ContextVar("NOOPT", 0)
 IMAGE, FLOAT16, OPENPILOT_HACKS = ContextVar("IMAGE", 0), ContextVar("FLOAT16", 0), ContextVar("OPENPILOT_HACKS", 0)
-JIT, JIT_BATCH_SIZE = ContextVar("JIT", 1), ContextVar("JIT_BATCH_SIZE", 32)
+JIT, JIT_BATCH_SIZE = ContextVar("JIT", 2 if OSX and ARCH_X86 else 1), ContextVar("JIT_BATCH_SIZE", 32)
 CHUNK_SIZE = 2**20  # TinyFS content-addressed store: blob chunk + hash-tree node granularity
 WINO, CAPTURING, TRACEMETA, NO_COLOR = ContextVar("WINO", 0), ContextVar("CAPTURING", 1), ContextVar("TRACEMETA", 1), ContextVar("NO_COLOR", 0)
 TRAINING = ContextVar("TRAINING", 0)
@@ -453,9 +454,9 @@ def _ensure_downloads_dir() -> pathlib.Path:
   if pathlib.Path("/etc/tinybox-release").is_file():
     # try creating dir with sudo
     if not (downloads_dir := pathlib.Path("/raid/downloads")).exists():
-      system(f"sudo mkdir -p {downloads_dir}")
-      system(f"sudo chown tiny:root {downloads_dir}")
-      system(f"sudo chmod 775 {downloads_dir}")
+      subprocess.run(["sudo", "mkdir", "-p", downloads_dir], check=True)
+      subprocess.run(["sudo", "chown", "tiny:root", downloads_dir], check=True)
+      subprocess.run(["sudo", "chmod", "775", downloads_dir], check=True)
     return downloads_dir
   return pathlib.Path(cache_dir) / "downloads"
 
@@ -496,7 +497,6 @@ def fetch_fw(path:str, name:str, sha256:str) -> bytes:
 # *** Exec helpers
 
 def system(cmd:str, **kwargs) -> str:
-  import subprocess
   st = time.perf_counter()
   try: ret = subprocess.check_output(cmd.split(), stderr=subprocess.STDOUT, **kwargs).decode().strip()
   except subprocess.CalledProcessError as e:

tinygrad/llm/cli.py

@@ -1,5 +1,5 @@
 from __future__ import annotations
-import sys, argparse, codecs, itertools, typing, re, unicodedata, json, time
+import sys, argparse, codecs, itertools, typing, re, unicodedata, json, time, pathlib
 from typing import TYPE_CHECKING
 from tinygrad import nn
 from tinygrad.uop.ops import UOp, Ops
@@ -127,30 +127,95 @@ class FallbackTemplate:
       out += self.end_turn()
     return out + self.role("assistant") if add_generation_prompt else out
 
-from tinygrad.llm.serve import LLMServer
+class KimiK3Template:
+  """Official K3 XTML envelope for text-only system/user/assistant conversations."""
+  OPEN, CLOSE, SEP, END = "<|open|>", "<|close|>", "<|sep|>", "<|end_of_msg|>"
+  def _open(self, tag:str, attrs:tuple[tuple[str, str], ...]=()) -> str:
+    escaped = ((k, str(v).replace("&", "&amp;").replace('"', "&quot;")) for k,v in attrs)
+    return self.OPEN + tag + "".join(f' {k}="{v}"' for k,v in escaped) + self.SEP
+  def _close(self, tag:str) -> str: return self.CLOSE + tag + self.SEP
+  def _message(self, role:str, content:str, name:str|None=None) -> str:
+    attrs = (("role", role),) + (() if name is None else (("name", name),))
+    return self._open("message", attrs) + content + self._close("message") + self.END
+  @staticmethod
+  def _content(message:dict) -> str:
+    content = message.get("content")
+    if content is None: return ""
+    if isinstance(content, str): return content
+    if isinstance(content, list):
+      if any(part.get("type") != "text" for part in content): raise ValueError("Kimi K3 native loader is text-only; image content is not implemented")
+      return "".join(part["text"] for part in content)
+    raise ValueError(f"unsupported Kimi K3 content type {type(content).__name__}")
+  def render(self, messages:list[dict], tools=None, add_generation_prompt:bool=True, preserve_thinking:bool=False, **kwargs) -> str:
+    if tools or any(m.get("role") == "tool" or m.get("tool_calls") for m in messages):
+      raise ValueError("Kimi K3 XTML tool rendering is not implemented in the native text loader")
+    effort = kwargs.get("thinking_effort", "max")
+    if effort not in ("low", "high", "max"): raise ValueError(f"invalid Kimi K3 thinking_effort {effort!r}")
+    body = "`thinking_effort` guides on how much to think in your thinking channel (not including the response channel), " \
+           "supported values include `low`, `medium`, `high`, and `max`.\n" \
+           f"Now the system is invoked with `thinking_effort={effort}`."
+    out = self._open("message", (("role", "system"), ("type", "thinking-effort"))) + body + self._close("message") + self.END
+    for message in messages:
+      role = message["role"]
+      if role in ("user", "system"):
+        out += self._message(role, self._content(message), message.get("name"))
+      elif role == "assistant":
+        reasoning = message.get("reasoning_content") or message.get("reasoning") or ""
+        content = self._open("think") + str(reasoning) + self._close("think")
+        content += self._open("response") + self._content(message) + self._close("response")
+        out += self._message(role, content, message.get("name"))
+      else: raise ValueError(f"unsupported Kimi K3 role {role!r}")
+    if add_generation_prompt: out += self._open("message", (("role", "assistant"),)) + self._open("think")
+    return out
+
+from tinygrad.llm.serve import LLMServer, StreamRouter
 
 def main():
   parser = argparse.ArgumentParser()
-  parser.add_argument("--model", "-m", default=list(models.keys())[0], help=f"Model choice ({', '.join(models.keys())}) or path to a local GGUF file")
+  parser.add_argument("--model", "-m", default=list(models.keys())[0],
+                      help=f"Model choice ({', '.join(models.keys())}), local GGUF file, converted Kimi directory, or official Kimi K3 directory")
   parser.add_argument("--max_context", type=int, default=4096, help="Max Context Length")
   parser.add_argument("--serve", nargs='?', type=int, const=8000, metavar="PORT", help="Run OpenAI compatible API (optional port, default 8000)")
   parser.add_argument("--warmup", action="store_true", help="warmup the JIT")
   parser.add_argument("--benchmark", nargs='?', type=int, const=20, metavar="COUNT", help="Benchmark tok/s (optional count, default 20)")
+  parser.add_argument("--devices", type=int, default=1, help="Tensor-parallel device count (Kimi-Linear requires 4, Kimi K3 requires 8)")
   args = parser.parse_args()
 
   # load the model
-  model, kv = Transformer.from_gguf(fetch(models.get(args.model, args.model)), args.max_context)
-  model_name = kv.get('general.name') or kv.get('general.basename') or args.model
+  model_path = pathlib.Path(args.model)
+  kv:dict[str, typing.Any]
+  is_k3 = False
+  if model_path.is_dir() and (model_path / "config.json").exists():
+    raw_config = json.loads((model_path / "config.json").read_text())
+    is_k3 = raw_config.get("model_type") == "kimi_k3"
+  if is_k3:
+    from tinygrad.llm.kimi_k3 import load_kimi_k3, load_kimi_tokenizer_data
+    model, kv = load_kimi_k3(model_path, args.max_context, args.devices), {}
+    normal, special, bos, eos = load_kimi_tokenizer_data(model_path)
+    tok = SimpleTokenizer(normal, special, "kimi-k2", bos_id=bos, eos_id=eos, eot_id=eos)
+    model_name = "Kimi-K3"
+    tok_cfg = json.loads((model_path / "tokenizer_config.json").read_text())
+    ct = tok_cfg.get("chat_template")
+  elif model_path.is_dir() and (model_path / "tinygrad-kimi.json").exists():
+    from tinygrad.llm.kimi import load_kimi, load_kimi_tokenizer_data
+    model, kv = load_kimi(model_path, args.max_context, args.devices), {}
+    normal, special, bos, eos = load_kimi_tokenizer_data(model_path)
+    tok = SimpleTokenizer(normal, special, "kimi-k2", bos_id=bos, eos_id=eos, eot_id=eos)
+    model_name = "Kimi-Linear-48B-A3B-Instruct-MXFP4"
+    ct = (model_path / "chat_template.jinja").read_text() if (model_path / "chat_template.jinja").exists() else None
+  else:
+    if args.devices != 1: raise ValueError("--devices is currently supported by the native Kimi MXFP4 loader only")
+    model, kv = Transformer.from_gguf(fetch(models.get(args.model, args.model)), args.max_context)
+    model_name = kv.get('general.name') or kv.get('general.basename') or args.model
+    tok = SimpleTokenizer.from_gguf_kv(kv)
+    ct = kv.get('tokenizer.chat_template')
   file_sizes = [y.nbytes() for y in UOp.sink(*[x.uop for x in nn.state.get_parameters(model)]).toposort() if y.op is Ops.BUFFER]
   print(f"using model \"{model_name}\" with {sum(file_sizes):,} bytes and {sum(x.numel() for x in nn.state.get_parameters(model)):,} params, "
         f"max context {args.max_context} on {nn.state.get_parameters(model)[0].device}")
 
-  # get tokenizer
-  tok = SimpleTokenizer.from_gguf_kv(kv)
-
   # use the model's chat template if jinja2 is available (enables model-specific formatting)
-  template: jinja2.Template|FallbackTemplate = FallbackTemplate(tok)
-  if (ct := kv.get('tokenizer.chat_template')) is not None:
+  template: jinja2.Template|FallbackTemplate|KimiK3Template = KimiK3Template() if is_k3 else FallbackTemplate(tok)
+  if ct is not None:
     try:
       import jinja2
       env = jinja2.Environment()
@@ -162,9 +227,10 @@ def main():
       template = env.from_string(ct)
     except ImportError: print("warning: jinja2 is not installed, the model's chat template is disabled")
 
-  # warmup the JIT
-  if args.warmup or args.serve:
+  # Capture the default greedy serving shapes before accepting requests.
+  if args.warmup:
     with Context(DEBUG=max(DEBUG.value, 1)): model.warmup()
+  elif args.serve: model.warmup()
 
   # start server
   if args.serve: LLMServer(('', args.serve), model, model_name, tok, template).serve_forever()
@@ -189,15 +255,26 @@ def main():
   while 1:
     try: messages.append({"role":"user", "content":input('>>> ')})
     except EOFError: break
-    ids = tok.encode(template.render(messages=messages, add_generation_prompt=True))
-    reply, dec = "", tok.stream_decoder()
+    rendered = template.render(messages=messages, add_generation_prompt=True)
+    ids = tok.encode(rendered)
+    reply, reasoning_reply, dec = "", "", tok.stream_decoder()
+    xtml = rendered.rstrip().endswith("<|open|>think<|sep|>")
+    router = StreamRouter(reasoning=xtml or rendered.rstrip().endswith("<think>"), xtml=xtml)
     for next_id in model.generate(ids):
       if tok.is_end(next_id):
-        sys.stdout.write(dec() + "\n\n")
+        for field,text in router.route(dec(), final=True):
+          if field == "content": reply += text
+          elif field == "reasoning_content": reasoning_reply += text
+          sys.stdout.write(text)
+        sys.stdout.write("\n\n")
         break
-      reply += (piece := dec(next_id))
-      sys.stdout.write(piece)
-      sys.stdout.flush()
-    messages.append({"role":"assistant", "content":reply})
+      for field,text in router.route(dec(next_id)):
+        if field == "content": reply += text
+        elif field == "reasoning_content": reasoning_reply += text
+        sys.stdout.write(text)
+        sys.stdout.flush()
+    assistant = {"role":"assistant", "content":reply}
+    if reasoning_reply: assistant["reasoning_content"] = reasoning_reply
+    messages.append(assistant)
 
 if __name__ == "__main__": main()

tinygrad/llm/kernels/__init__.py

@@ -0,0 +1,238 @@
+import functools
+from typing import cast
+from tinygrad import Tensor, UOp, Device, Context, dtypes
+from tinygrad.device import Buffer, MultiBuffer
+from tinygrad.dtype import AddrSpace
+from tinygrad.uop.ops import AxisType, KernelInfo
+
+def amd_custom_kernels_supported(device:str|tuple[str, ...]|None) -> bool:
+  """The hand-written wave32 kernel is intentionally limited to RDNA3/gfx11."""
+  if device is None: return False
+  device = device[0] if isinstance(device, tuple) else device
+  with Context(ALLOW_DEVICE_USAGE=1):
+    return (target:=getattr(Device[device], "target", None)) is not None and target[0] == 11
+
+def amd_wave64_custom_kernels_supported(device:str|tuple[str, ...]|None) -> bool:
+  """CDNA4 wave64 kernels used by the MI350X K3 path."""
+  if device is None: return False
+  device = device[0] if isinstance(device, tuple) else device
+  if device.startswith("NULL:HIP:gfx950"): return True  # compile-only CDNA4 coverage without MI350X hardware
+  with Context(ALLOW_DEVICE_USAGE=1):
+    return (target:=getattr(Device[device], "target", None)) is not None and target[:2] == (9, 5)
+
+def amd_packed_mxfp4_supported(device:str|tuple[str, ...]|None) -> bool:
+  return amd_custom_kernels_supported(device) or amd_wave64_custom_kernels_supported(device)
+
+def amd_exact_bf16_custom_kernels_supported(device:str|tuple[str, ...]|None) -> bool:
+  """The LDS-reduced exact BF16 pair kernels are portable across RDNA3 and CDNA4."""
+  return amd_custom_kernels_supported(device) or amd_wave64_custom_kernels_supported(device)
+
+def amd_int32_item(x:Tensor, host:memoryview) -> int:
+  """Copy a realized replicated AMD scalar without constructing a new scheduler graph."""
+  if x.numel() != 1 or x.dtype != dtypes.int32 or host.nbytes != 4: raise ValueError("expected one int32 and a four-byte host view")
+  buf = x.uop.buffer
+  if isinstance(buf, MultiBuffer): buf = buf.bufs[0]
+  if not isinstance(buf, Buffer) or not buf.device.startswith("AMD"): raise ValueError("expected a realized AMD buffer")
+  buf.allocator._copyout(host, buf._buf)
+  return int.from_bytes(host, byteorder="little", signed=True)
+
+def mxfp4_expert_linear(sel:Tensor, x:Tensor, weight:Tensor, scale:Tensor, partial:bool=False) -> Tensor:
+  """Run a TP routed projection without materializing selected BF16 weights."""
+  from tinygrad.llm.kernels.amd import (_mxfp4_expert_linear_kernel, _mxfp4_expert_linear_wave64_kernel,
+                                        _mxfp4_expert_linear_wave64_prefill_kernel)
+  batch, tokens, topk = sel.shape
+  out_features = weight.shape[1]
+  weight_axis = weight.uop.axis
+  if isinstance(weight.device, tuple):
+    devices = weight.device
+    # Gate/up shard their output dimension. Down shards its reduction dimension;
+    # represent each GPU's partial as a size-one device axis, then all-reduce it.
+    axis = 3 if weight_axis == 1 else 4
+    shard_shape: tuple[int|UOp, ...]
+    if weight_axis == 1:
+      if out_features % len(devices): raise ValueError(f"expert output {out_features} is not divisible by {len(devices)} devices")
+      shard_shape = (batch, tokens, topk, out_features//len(devices))
+    elif weight_axis == 2:
+      shard_shape = (batch, tokens, topk, out_features, 1)
+    else: raise ValueError(f"unsupported expert TP axis {weight_axis}")
+    partial_dtype = dtypes.float32 if weight_axis == 2 else dtypes.bfloat16
+    parts = [Tensor.empty(*shard_shape, dtype=partial_dtype, device=device).uop for device in devices]
+    out = Tensor(parts[0].mstack(*parts[1:]).unshard(axis))
+  else:
+    out = Tensor.empty(batch, tokens, topk, out_features, dtype=dtypes.bfloat16, device=weight.device)
+  if amd_wave64_custom_kernels_supported(weight.device):
+    kernel = _mxfp4_expert_linear_wave64_prefill_kernel if tokens > 1 else _mxfp4_expert_linear_wave64_kernel
+  else: kernel = _mxfp4_expert_linear_kernel
+  out = Tensor.custom_kernel(out, sel.contiguous(), x.contiguous(), weight, scale, fxn=kernel)[0]
+  return out if weight_axis == 2 and partial else out.sum(4).cast(dtypes.bfloat16) if weight_axis == 2 else out
+
+def bf16_partial_linear(x:Tensor, weight:Tensor) -> Tensor:
+  """Return output-shaped FP32 TP partials with a final device axis, without all-reduce."""
+  from tinygrad.llm.kernels.amd import _bf16_partial_linear_kernel
+  if not isinstance(weight.device, tuple) or weight.uop.axis != 1: raise ValueError("partial linear expects input-sharded TP weight")
+  batch, tokens, _ = x.shape
+  devices, out_features = weight.device, weight.shape[0]
+  shard_shape = (batch, tokens, out_features, 1)
+  parts = [Tensor.empty(*shard_shape, dtype=dtypes.float32, device=device).uop for device in devices]
+  out = Tensor(parts[0].mstack(*parts[1:]).unshard(3))
+  return Tensor.custom_kernel(out, x.contiguous(), weight, fxn=_bf16_partial_linear_kernel)[0]
+
+def bf16_matvec(x:Tensor, weight:Tensor) -> Tensor:
+  from tinygrad.llm.kernels.amd import _bf16_matvec_kernel
+  batch, tokens, _ = x.shape
+  out_features = weight.shape[0]
+  if isinstance(weight.device, tuple):
+    devices = weight.device
+    if weight.uop.axis == 0:
+      shard_shape = (batch, tokens, out_features//len(devices))
+      parts = [Tensor.empty(*shard_shape, dtype=dtypes.bfloat16, device=device).uop for device in devices]
+      out = Tensor(parts[0].mstack(*parts[1:]).unshard(2))
+    elif weight.uop.axis is None: out = Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=weight.device)
+    else: raise ValueError("bf16_matvec expects output-sharded or replicated TP weight")
+  else: out = Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=weight.device)
+  return Tensor.custom_kernel(out, x.contiguous(), weight, fxn=_bf16_matvec_kernel)[0]
+
+def bf16_mfma_splitk(x:Tensor, weight:Tensor) -> Tensor:
+  """gfx950 decode matvec for replicated or output-sharded BF16 weights."""
+  from tinygrad.llm.kernels.amd import _bf16_mfma_splitk_kernel
+  batch, tokens, in_features = x.shape
+  out_features = weight.shape[0]
+  if batch != 1 or tokens != 1 or weight.shape[1] != in_features or in_features % 256:
+    raise ValueError(f"unsupported MFMA split-K shapes {x.shape} {weight.shape}")
+  if not amd_wave64_custom_kernels_supported(weight.device): raise ValueError("MFMA split-K requires gfx950")
+  if isinstance(weight.device, tuple):
+    devices = weight.device
+    if weight.uop.axis == 0:
+      if out_features % (16*len(devices)): raise ValueError("local MFMA output must be divisible by 16")
+      shape = (batch, tokens, out_features//len(devices))
+      parts = [Tensor.empty(*shape, dtype=dtypes.bfloat16, device=device).uop for device in devices]
+      out = Tensor(parts[0].mstack(*parts[1:]).unshard(2))
+    elif weight.uop.axis is None:
+      if out_features % 16: raise ValueError("MFMA output must be divisible by 16")
+      out = Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=devices)
+    else: raise ValueError("MFMA split-K requires replicated or output-sharded weight")
+  else:
+    if out_features % 16: raise ValueError("MFMA output must be divisible by 16")
+    out = Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=weight.device)
+  return Tensor.custom_kernel(out, x.contiguous(), weight, fxn=_bf16_mfma_splitk_kernel)[0]
+
+def mxfp8_quantize_dequantize(x:Tensor) -> Tensor:
+  """gfx11 software MXFP8 round trip without a multi-kernel reduction graph."""
+  from tinygrad.llm.kernels.amd import _mxfp8_qdq_kernel
+  out = Tensor.empty_like(x, dtype=dtypes.bfloat16)
+  return Tensor.custom_kernel(out, x.contiguous(), fxn=_mxfp8_qdq_kernel)[0]
+
+def kda_qkv_linear(x:Tensor, qw:Tensor, kw:Tensor, vw:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  """Fuse equal-sized output-sharded KDA Q/K/V decode projections."""
+  from tinygrad.llm.kernels.amd import _kda_qkv_kernel
+  batch, tokens, _ = x.shape
+  out_features = qw.shape[0]
+  if not (qw.shape == kw.shape == vw.shape): raise ValueError("fused KDA Q/K/V weights must have equal shapes")
+  if isinstance(qw.device, tuple):
+    devices = qw.device
+    if qw.uop.axis != 0 or out_features % len(devices): raise ValueError("fused KDA Q/K/V expects output-sharded weights")
+    shard_shape = (batch, tokens, out_features//len(devices))
+    def make_out() -> Tensor:
+      parts = [Tensor.empty(*shard_shape, dtype=dtypes.bfloat16, device=device).uop for device in devices]
+      return Tensor(parts[0].mstack(*parts[1:]).unshard(2))
+    outs: tuple[Tensor, Tensor, Tensor] = (make_out(), make_out(), make_out())
+  else:
+    outs = (Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=qw.device),
+            Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=qw.device),
+            Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=qw.device))
+  ret = Tensor.custom_kernel(*outs, x.contiguous(), qw, kw, vw, fxn=_kda_qkv_kernel)
+  return ret[0], ret[1], ret[2]
+
+def dual_bf16_matvec(x:Tensor, aw:Tensor, bw:Tensor, fast:bool=False) -> tuple[Tensor, Tensor]:
+  """Fuse two equal-shaped BF16 decode projections that consume the same input."""
+  from tinygrad.llm.kernels.amd import _dual_bf16_matvec_fast_kernel, _dual_bf16_matvec_kernel
+  batch, tokens, _ = x.shape
+  out_features = aw.shape[0]
+  if aw.shape != bw.shape: raise ValueError("fused BF16 weights must have equal shapes")
+  if isinstance(aw.device, tuple) and aw.uop.axis == 0:
+    devices = aw.device
+    if out_features % len(devices): raise ValueError("fused BF16 output is not divisible by the device count")
+    shard_shape = (batch, tokens, out_features//len(devices))
+    def make_out() -> Tensor:
+      parts = [Tensor.empty(*shard_shape, dtype=dtypes.bfloat16, device=device).uop for device in devices]
+      return Tensor(parts[0].mstack(*parts[1:]).unshard(2))
+    outs = (make_out(), make_out())
+  else:
+    outs = (Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=aw.device),
+            Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=aw.device))
+  ret = Tensor.custom_kernel(*outs, x.contiguous(), aw, bw, fxn=_dual_bf16_matvec_fast_kernel if fast else _dual_bf16_matvec_kernel)
+  return ret[0], ret[1]
+
+def dual_input_bf16_matvec(ax:Tensor, bx:Tensor, aw:Tensor, bw:Tensor) -> tuple[Tensor, Tensor]:
+  """Fuse equal-shaped BF16 projections with separate inputs and identical TP layouts."""
+  from tinygrad.llm.kernels.amd import _dual_input_bf16_matvec_kernel
+  if ax.shape != bx.shape or aw.shape != bw.shape: raise ValueError("fused BF16 inputs and weights must have equal shapes")
+  batch, tokens, _ = ax.shape
+  out_features = aw.shape[0]
+  if isinstance(aw.device, tuple) and aw.uop.axis == 0:
+    devices = aw.device
+    if out_features % len(devices): raise ValueError("fused BF16 output is not divisible by the device count")
+    shard_shape = (batch, tokens, out_features//len(devices))
+    def make_out() -> Tensor:
+      parts = [Tensor.empty(*shard_shape, dtype=dtypes.bfloat16, device=device).uop for device in devices]
+      return Tensor(parts[0].mstack(*parts[1:]).unshard(2))
+    outs = (make_out(), make_out())
+  else:
+    outs = (Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=aw.device),
+            Tensor.empty(batch, tokens, out_features, dtype=dtypes.bfloat16, device=aw.device))
+  ret = Tensor.custom_kernel(*outs, ax.contiguous(), bx.contiguous(), aw, bw, fxn=_dual_input_bf16_matvec_kernel)
+  return ret[0], ret[1]
+
+def kda_fgb_linear(x:Tensor, gw:Tensor, fw:Tensor, bw:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  """Fuse replicated KDA g/f low-rank projections with its output-sharded beta projection."""
+  from tinygrad.llm.kernels.amd import _kda_fgb_kernel
+  if gw.shape != fw.shape or not isinstance(gw.device, tuple) or gw.uop.axis is not None or \
+     bw.device != gw.device or bw.uop.axis != 0: raise ValueError("unsupported KDA f/g/beta TP layout")
+  batch, tokens, _ = x.shape
+  devices, rank, beta_features = gw.device, gw.shape[0], bw.shape[0]
+  gout = Tensor.empty(batch, tokens, rank, dtype=dtypes.bfloat16, device=devices)
+  fout = Tensor.empty(batch, tokens, rank, dtype=dtypes.bfloat16, device=devices)
+  beta_shape = (batch, tokens, beta_features//len(devices))
+  parts = [Tensor.empty(*beta_shape, dtype=dtypes.bfloat16, device=device).uop for device in devices]
+  bout = Tensor(parts[0].mstack(*parts[1:]).unshard(2))
+  ret = Tensor.custom_kernel(gout, fout, bout, x.contiguous(), gw, fw, bw, fxn=_kda_fgb_kernel)
+  return ret[0], ret[1], ret[2]
+
+@functools.cache
+def _gated_delta_prefill_kernel(core:UOp, next_state:UOp, q:UOp, k:UOp, v:UOp, beta:UOp, alpha:UOp, state:UOp, kq:UOp) -> UOp:
+  batch, heads, tokens, value_dim = cast(tuple[int, int, int, int], core.shape)
+  key_dim, alpha_dim = cast(int, q.shape[-1]), cast(int, alpha.shape[-1]) if len(alpha.shape) == 4 else 1
+  core, v = (x.reshape(batch*heads, tokens, value_dim) for x in (core, v))
+  q, k = (x.reshape(batch*heads, tokens, key_dim) for x in (q, k))
+  beta, kq = (x.reshape(batch*heads, tokens) for x in (beta, kq))
+  alpha = alpha.reshape(batch*heads, tokens, alpha_dim)
+  state, next_state = (x.reshape(batch*heads, value_dim, key_dim) for x in (state, next_state))
+  bh, row, cols = UOp.range(batch*heads, 0, AxisType.GLOBAL), UOp.range(value_dim, 2), tuple(range(key_dim))
+  current = UOp.placeholder((key_dim,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  current = current.after(UOp.group(*(current[col].store(state[bh, row, col].float()) for col in cols)))
+  token = UOp.range(tokens, 1, AxisType.REDUCE)
+  previous = tuple(current.after(token)[col].load() for col in cols)
+  keys, queries = (tuple(x[bh, token, col].load() for col in cols) for x in (k, q))
+  av = tuple(alpha[bh, token, col if alpha_dim > 1 else 0].load() for col in cols)
+  bv = beta[bh, token].load()
+  state_k = sum((x*a*y for x,a,y in zip(previous, av, keys)), UOp.const(0, dtypes.float32))
+  state_q = sum((x*a*y for x,a,y in zip(previous, av, queries)), UOp.const(0, dtypes.float32))
+  delta = (v[bh, token, row].load() - state_k) * bv
+  step = UOp.group(core[bh, token, row].store(state_q + delta*kq[bh, token]),
+                   *(current[col].store(x*a + delta*y) for col,x,a,y in zip(cols, previous, av, keys))).end(token)
+  stores = (next_state[bh, row, col].store(current.after(step)[col].load().cast(next_state.dtype)) for col in cols)
+  return UOp.group(*stores).end(row, bh).sink(arg=KernelInfo(name="gated_delta_prefill", opts_to_apply=()))
+
+def gated_delta_prefill(q:Tensor, k:Tensor, v:Tensor, beta:Tensor, alpha:Tensor, state:Tensor) -> tuple[Tensor, Tensor]:
+  batch, heads, tokens, key_dim = q.shape
+  value_dim = v.shape[-1]
+  assert q.shape == k.shape and v.shape[:3] == q.shape[:3] and beta.shape == (batch, heads, tokens)
+  assert alpha.shape in ((batch, heads, tokens), (batch, heads, tokens, key_dim))
+  assert state.shape == (batch, heads, value_dim, key_dim)
+  kernel = _gated_delta_prefill_kernel
+  if amd_exact_bf16_custom_kernels_supported(q.device) and key_dim % 32 == 0 and value_dim % 4 == 0:
+    from tinygrad.llm.kernels.amd import _gated_delta_prefill_kernel as kernel
+  core, next_state, kq = Tensor.empty_like(v), Tensor.empty_like(state), (q*k).sum(-1).contiguous()
+  result = Tensor.custom_kernel(core, next_state, q.contiguous(), k.contiguous(), v.contiguous(), beta.contiguous(), alpha.contiguous(), state, kq,
+                                fxn=kernel)
+  return result[0], result[1]

tinygrad/llm/kernels/amd.py

@@ -0,0 +1,353 @@
+from __future__ import annotations
+import functools
+import pathlib
+from typing import cast
+from tinygrad import UOp
+from tinygrad.uop.ops import AxisType, KernelInfo, Ops
+from tinygrad.dtype import AddrSpace, dtypes
+
+@functools.cache
+def _bf16_mfma_splitk_kernel(out:UOp, x:UOp, weight:UOp) -> UOp:
+  """CDNA4 BF16 matvec with eight waves splitting K per 16 output channels."""
+  from tinygrad.renderer import Estimates
+  from tinygrad.runtime.support.compiler_amd import HIPCCCompiler
+  out_features, in_features = cast(tuple[int, int], weight.shape)
+  assert out.numel() == out_features and x.numel() == in_features and out_features % 16 == 0 and in_features % 256 == 0
+  threads, workgroups = UOp.special(512, "lidx0"), UOp.special(out_features//16, "gidx0")
+  sink = UOp.sink(out.base, x.base, weight.base, threads, workgroups,
+                  arg=KernelInfo(name=f"bf16_mfma_splitk_{out_features}_{in_features}",
+                                 estimates=Estimates(ops=2*out_features*in_features,
+                                                     mem=(out_features*in_features+in_features+out_features)*2)))
+  root = pathlib.Path(__file__).parents[3]/"extra"/"thunder"/"amd"
+  src = (root/"matvec_bf16_splitk.cpp").read_text()
+  lib = HIPCCCompiler("gfx950", [f"-I{(root/'include').as_posix()}", "-std=c++20", "-DKITTENS_CDNA4", "-ffast-math",
+                                   "-DHIP_ENABLE_WARP_SYNC_BUILTINS", f"-DMATVEC_N={out_features}",
+                                   f"-DMATVEC_K={in_features}"]).compile_cached(src)
+  return UOp(Ops.PROGRAM, src=(sink, UOp(Ops.LINEAR, src=(*sink.src, sink)), UOp(Ops.SOURCE, arg=src), UOp(Ops.BINARY, arg=lib)))
+
+def warp_reduce(val:UOp, full_wave:bool=False, maximum:bool=False) -> UOp:
+  for offset in ((16, 8, 4, 2, 1) if full_wave else (8, 4, 2, 1)):
+    if val.op is Ops.INDEX and val.addrspace == AddrSpace.REG: val = val.load()
+    other = UOp(Ops.CUSTOM, dtypes.float, (val,), arg=
+      f"__builtin_bit_cast(float, __builtin_amdgcn_ds_swizzle(__builtin_bit_cast(int, {{0}}), {0x1f | offset<<10}))")
+    val = val.maximum(other) if maximum else val + other
+  return val
+
+@functools.cache
+def _mxfp8_qdq_kernel(out:UOp, x:UOp) -> UOp:
+  """Software OCP E4M3/E8M0 round trip, one wave per 32-value MX block."""
+  groups = cast(int, x.shape[-1])//32
+  outer = x.numel()//cast(int, x.shape[-1])
+  block, lane = UOp.range(outer*groups, 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  value = x.reshape(outer, groups, 32)[block//groups, block%groups, lane].float()
+  amax = warp_reduce(value.abs(), full_wave=True, maximum=True)
+  exponent = (amax.maximum(1e-38)/448.0).log2().round().maximum(-127.0).minimum(127.0)
+  block_scale = amax.eq(0).where(1.0, exponent.exp2())
+  normalized = value/block_scale
+  magnitude = normalized.abs().minimum(448.0)
+  elem_exp = magnitude.maximum(2**-9).log2().floor().maximum(-6.0).minimum(8.0)
+  quantum = (elem_exp-3.0).exp2()
+  quantized = (magnitude/quantum).round()*quantum
+  quantized = (normalized < 0).where(-quantized, quantized).maximum(-448.0).minimum(448.0)
+  store = out.reshape(outer, groups, 32)[block//groups, block%groups, lane].store((quantized*block_scale).cast(out.dtype))
+  return store.end(lane, block).sink(arg=KernelInfo(name="mxfp8_qdq", opts_to_apply=()))
+
+def _mxfp4_value(code:UOp) -> UOp:
+  """Decode one OCP E2M1 nibble without a lookup-table memory access."""
+  magnitude = code & 7
+  value = magnitude.eq(7).where(6.0, magnitude.eq(6).where(4.0, magnitude.eq(5).where(3.0, magnitude.float()*0.5)))
+  return (code & 8).ne(0).where(-value, value)
+
+def _e8m0_value(scale:UOp) -> UOp:
+  """Decode an E8M0 byte with IEEE exponent bits instead of a transcendental exp2."""
+  bits = scale.cast(dtypes.uint32) << 23
+  # E8M0 byte zero denotes 2**-127, halfway through IEEE's subnormal exponent bin.
+  return scale.eq(0).where(UOp.const(0x00400000, dtypes.uint32).bitcast(dtypes.float32), bits.bitcast(dtypes.float32))
+
+@functools.cache
+def _kda_qkv_kernel(qout:UOp, kout:UOp, vout:UOp, x:UOp, qw:UOp, kw:UOp, vw:UOp) -> UOp:
+  """Fused BF16 decode projection for equal-sized KDA Q/K/V tensors."""
+  batch, tokens, out_features = cast(tuple[int, int, int], qout.shape)
+  in_features, output_tile = cast(int, x.shape[-1]), 1
+  assert qout.shape == kout.shape == vout.shape and out_features % output_tile == 0 and in_features % 32 == 0
+  row, lane = UOp.range(batch*tokens*(out_features//output_tile), 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  token, output_block = row // (out_features//output_tile), row % (out_features//output_tile)
+  outputs = tuple(output_block*output_tile+i for i in range(output_tile))
+  acc = UOp.placeholder((3, output_tile), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0.0)))
+  group = UOp.range(in_features//32, 2, AxisType.REDUCE)
+  activation = x.reshape(batch*tokens, in_features)[token, group*32+lane].float()
+  updates = [acc.after(group)[p, i].load()+activation*w[output, group*32+lane].float()
+             for p,w in enumerate((qw, kw, vw)) for i,output in enumerate(outputs)]
+  update = acc.store(UOp.stack(*updates).reshape(3, output_tile)).end(group)
+  outs = (qout, kout, vout)
+  stores = (outs[p].reshape(batch*tokens, out_features)[token, output.valid(lane.eq(0))].store(
+    warp_reduce(acc.after(update)[p, i], full_wave=True).cast(outs[p].dtype))
+    for p in range(3) for i,output in enumerate(outputs))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="kda_qkv", opts_to_apply=()))
+
+@functools.cache
+def _dual_bf16_matvec_kernel(aout:UOp, bout:UOp, x:UOp, aw:UOp, bw:UOp) -> UOp:
+  """Exact pair of BF16 decode projections with one activation read."""
+  batch, tokens, out_features = cast(tuple[int, int, int], aout.shape)
+  in_features = cast(int, x.shape[-1])
+  assert aout.shape == bout.shape and out_features == aw.shape[0] == bw.shape[0] and in_features % 16 == 0
+  row, lane = UOp.range(batch*tokens*out_features, 0), UOp.range(16, 1, axis_type=AxisType.LOCAL)
+  token, output = row//out_features, row%out_features
+  acc = UOp.placeholder((2,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0.0)))
+  chunk, group = in_features//16, UOp.range(in_features//16, 2, AxisType.REDUCE)
+  input_idx = lane*chunk+group
+  activation = x.reshape(batch*tokens, in_features)[token, input_idx].float()
+  update = acc.store(UOp.stack(*(acc.after(group)[i].load()+(activation*w[output, input_idx].float()).cast(dtypes.bfloat16).float()
+                                for i,w in enumerate((aw, bw))))).end(group)
+  local = UOp.placeholder((2, 16), dtypes.float32, slot=1, addrspace=AddrSpace.LOCAL)
+  barrier = UOp.group(*(local[i, lane].store(acc.after(update)[i]) for i in range(2))).barrier()
+  stores = (out.reshape(batch*tokens, out_features)[token, output.valid(lane.eq(0))].store(
+    sum((local.after(barrier)[i, j] for j in range(16)), UOp.const(0, dtypes.float32)).cast(out.dtype))
+            for i,out in enumerate((aout, bout)))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="dual_bf16_matvec", opts_to_apply=()))
+
+@functools.cache
+def _dual_bf16_matvec_fast_kernel(aout:UOp, bout:UOp, x:UOp, aw:UOp, bw:UOp) -> UOp:
+  """Coalesced pair used where one changed reduction boundary does not feed recurrent state."""
+  batch, tokens, out_features = cast(tuple[int, int, int], aout.shape)
+  in_features = cast(int, x.shape[-1])
+  assert aout.shape == bout.shape and out_features == aw.shape[0] == bw.shape[0] and in_features % 32 == 0
+  row, lane = UOp.range(batch*tokens*out_features, 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  token, output = row//out_features, row%out_features
+  acc = UOp.placeholder((2,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0.0)))
+  group = UOp.range(in_features//32, 2, AxisType.REDUCE)
+  input_idx = group*32+lane
+  activation = x.reshape(batch*tokens, in_features)[token, input_idx].float()
+  update = acc.store(UOp.stack(*(acc.after(group)[i].load()+(activation*w[output, input_idx].float()).cast(dtypes.bfloat16).float()
+                                for i,w in enumerate((aw, bw))))).end(group)
+  stores = (out.reshape(batch*tokens, out_features)[token, output.valid(lane.eq(0))].store(
+    warp_reduce(acc.after(update)[i], full_wave=True).cast(out.dtype)) for i,out in enumerate((aout, bout)))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="dual_bf16_matvec_fast", opts_to_apply=()))
+
+@functools.cache
+def _dual_input_bf16_matvec_kernel(aout:UOp, bout:UOp, ax:UOp, bx:UOp, aw:UOp, bw:UOp) -> UOp:
+  """Exact pair of equal-shaped BF16 projections with distinct inputs."""
+  batch, tokens, out_features = cast(tuple[int, int, int], aout.shape)
+  in_features = cast(int, ax.shape[-1])
+  assert aout.shape == bout.shape and ax.shape == bx.shape and out_features == aw.shape[0] == bw.shape[0] and in_features % 16 == 0
+  row, lane = UOp.range(batch*tokens*out_features, 0), UOp.range(16, 1, axis_type=AxisType.LOCAL)
+  token, output = row//out_features, row%out_features
+  acc = UOp.placeholder((2,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0.0)))
+  chunk, group = in_features//16, UOp.range(in_features//16, 2, AxisType.REDUCE)
+  input_idx = lane*chunk+group
+  update = acc.store(UOp.stack(*(acc.after(group)[i].load()+(inp.reshape(batch*tokens, in_features)[token, input_idx].float()*
+    weight[output, input_idx].float()).cast(dtypes.bfloat16).float() for i,(inp,weight) in enumerate(((ax,aw), (bx,bw)))))).end(group)
+  local = UOp.placeholder((2, 16), dtypes.float32, slot=1, addrspace=AddrSpace.LOCAL)
+  barrier = UOp.group(*(local[i, lane].store(acc.after(update)[i]) for i in range(2))).barrier()
+  stores = (out.reshape(batch*tokens, out_features)[token, output.valid(lane.eq(0))].store(
+    sum((local.after(barrier)[i, j] for j in range(16)), UOp.const(0, dtypes.float32)).cast(out.dtype))
+            for i,out in enumerate((aout, bout)))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="dual_input_bf16_matvec", opts_to_apply=()))
+
+@functools.cache
+def _kda_fgb_kernel(gout:UOp, fout:UOp, bout:UOp, x:UOp, gw:UOp, fw:UOp, bw:UOp) -> UOp:
+  """Mixed-output wave32 KDA g/f/beta projection."""
+  batch, tokens, rank = cast(tuple[int, int, int], gout.shape)
+  beta_features, in_features = cast(int, bout.shape[-1]), cast(int, x.shape[-1])
+  assert gout.shape == fout.shape and rank == gw.shape[0] == fw.shape[0] and beta_features == bw.shape[0] and in_features % 32 == 0
+  rows = batch*tokens*(2*rank+beta_features)
+  row, lane = UOp.range(rows, 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  token, projection_row = row//(2*rank+beta_features), row%(2*rank+beta_features)
+  is_g, is_f = projection_row < rank, (projection_row >= rank) & (projection_row < 2*rank)
+  g_row = projection_row.valid(is_g)
+  f_row = (projection_row-rank).valid(is_f)
+  b_row = (projection_row-2*rank).valid(~is_g & ~is_f)
+  acc = UOp.placeholder((), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(0.0))
+  group = UOp.range(in_features//32, 2, AxisType.REDUCE)
+  input_idx = group*32+lane
+  weight = is_g.where(gw[g_row, input_idx], is_f.where(fw[f_row, input_idx], bw[b_row, input_idx])).float()
+  product = (x.reshape(batch*tokens, in_features)[token, input_idx].float()*weight).cast(dtypes.bfloat16).float()
+  update = acc.store(acc.after(group)+product).end(group)
+  total = warp_reduce(acc.after(update)[0], full_wave=True).cast(dtypes.bfloat16)
+  stores = (gout.reshape(batch*tokens, rank)[token, g_row.valid(lane.eq(0))].store(total),
+            fout.reshape(batch*tokens, rank)[token, f_row.valid(lane.eq(0))].store(total),
+            bout.reshape(batch*tokens, beta_features)[token, b_row.valid(lane.eq(0))].store(total))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="kda_fgb", opts_to_apply=()))
+
+def _mxfp4_expert_linear_impl(out:UOp, sel:UOp, x:UOp, weight:UOp, scale:UOp) -> UOp:
+  """Wave32 decode GEMM which consumes selected experts directly from packed MXFP4 storage."""
+  batch, tokens, topk, out_features = cast(tuple[int, int, int, int], out.shape[:4])
+  partials = cast(int, out.shape[4]) if len(out.shape) == 5 else 1
+  output_tile = 1
+  assert out_features % output_tile == 0
+  in_features = cast(int, weight.shape[-1])*2
+  assert in_features % 32 == 0 and x.shape[-1] == in_features and sel.shape == (batch, tokens, topk)
+  xchoices = cast(int, x.shape[-2])
+  assert xchoices in (1, topk)
+  total_rows = batch*tokens*topk*(out_features//output_tile)*partials
+  row, lane = UOp.range(total_rows, 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  partial, output_block, route = row % partials, (row//partials) % (out_features//output_tile), \
+    row // ((out_features//output_tile)*partials)
+  outputs = tuple(output_block*output_tile+i for i in range(output_tile))
+  token, choice = route // topk, route % topk
+  expert = sel.reshape(batch*tokens, topk)[token, choice]
+  xv = x.reshape(batch*tokens, xchoices, in_features)
+  acc = UOp.placeholder((output_tile,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0.0)))
+  group = UOp.range(in_features//32, 2, AxisType.REDUCE)
+  activation = xv[token, 0 if xchoices == 1 else choice, group*32+lane].float()
+  updates = []
+  for i,output in enumerate(outputs):
+    packed = weight[expert, output, group*16+lane//2]
+    code = (packed >> ((lane&1)*4).cast(dtypes.uint8)) & 15
+    w = _mxfp4_value(code) * _e8m0_value(scale[expert, output, group])
+    updates.append(acc.after(group)[i].load()+activation*w)
+  update = acc.store(UOp.stack(*updates)).end(group)
+  out = out.reshape(batch*tokens, topk, out_features, partials)
+  stores = (out[token, choice, output, partial.valid(lane.eq(0))].store(warp_reduce(acc.after(update)[i], full_wave=True).cast(out.dtype))
+            for i,output in enumerate(outputs))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="mxfp4_expert_linear", opts_to_apply=()))
+
+@functools.cache
+def _mxfp4_expert_linear_kernel(out:UOp, sel:UOp, x:UOp, weight:UOp, scale:UOp) -> UOp:
+  return _mxfp4_expert_linear_impl(out, sel, x, weight, scale)
+
+@functools.cache
+def _mxfp4_expert_linear_wave64_kernel(out:UOp, sel:UOp, x:UOp, weight:UOp, scale:UOp) -> UOp:
+  """Wave64 decode GEMM for CDNA, with a workgroup-wide reduction independent of local-id decomposition."""
+  batch, tokens, topk, out_features = cast(tuple[int, int, int, int], out.shape[:4])
+  partials = cast(int, out.shape[4]) if len(out.shape) == 5 else 1
+  in_features = cast(int, weight.shape[-1])*2
+  assert in_features % 64 == 0 and x.shape[-1] == in_features and sel.shape == (batch, tokens, topk)
+  xchoices = cast(int, x.shape[-2])
+  assert xchoices in (1, topk)
+  total_rows = batch*tokens*topk*out_features*partials
+  row, lane = UOp.range(total_rows, 0), UOp.range(64, 1, axis_type=AxisType.LOCAL)
+  partial, output, route = row%partials, (row//partials)%out_features, row//(out_features*partials)
+  token, choice = route//topk, route%topk
+  expert = sel.reshape(batch*tokens, topk)[token, choice]
+  xv = x.reshape(batch*tokens, xchoices, in_features)
+  acc = UOp.placeholder((), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(0.0))
+  group = UOp.range(in_features//64, 2, AxisType.REDUCE)
+  activation = xv[token, 0 if xchoices == 1 else choice, group*64+lane].float()
+  packed = weight[expert, output, group*32+lane//2]
+  code = (packed >> ((lane&1)*4).cast(dtypes.uint8)) & 15
+  weight_value = _mxfp4_value(code) * _e8m0_value(scale[expert, output, group*2+lane//32])
+  update = acc.store(acc.after(group)+activation*weight_value).end(group)
+  local = UOp.placeholder((64,), dtypes.float32, slot=1, addrspace=AddrSpace.LOCAL)
+  barrier = local[lane].store(acc.after(update)[0]).barrier()
+  total = sum((local.after(barrier)[i] for i in range(64)), UOp.const(0, dtypes.float32))
+  out = out.reshape(batch*tokens, topk, out_features, partials)
+  store = out[token, choice, output, partial.valid(lane.eq(0))].store(total.cast(out.dtype))
+  return store.end(lane, row).sink(arg=KernelInfo(name="mxfp4_expert_linear_wave64", opts_to_apply=()))
+
+@functools.cache
+def _mxfp4_expert_linear_wave64_prefill_kernel(out:UOp, sel:UOp, x:UOp, weight:UOp, scale:UOp) -> UOp:
+  """Tiled CDNA4 prefill GEMM. Four adjacent outputs share activation loads and each wave half reduces in parallel."""
+  batch, tokens, topk, out_features = cast(tuple[int, int, int, int], out.shape[:4])
+  partials = cast(int, out.shape[4]) if len(out.shape) == 5 else 1
+  in_features, output_tile = cast(int, weight.shape[-1])*2, 4
+  assert tokens > 1 and in_features % 64 == 0 and out_features % output_tile == 0
+  assert x.shape[-1] == in_features and sel.shape == (batch, tokens, topk)
+  xchoices = cast(int, x.shape[-2])
+  assert xchoices in (1, topk)
+  total_rows = batch*tokens*topk*(out_features//output_tile)*partials
+  row, lane = UOp.range(total_rows, 0), UOp.range(64, 1, axis_type=AxisType.LOCAL)
+  partial, output_block, route = row%partials, (row//partials)%(out_features//output_tile), row//((out_features//output_tile)*partials)
+  outputs = tuple(output_block*output_tile+i for i in range(output_tile))
+  token, choice = route//topk, route%topk
+  expert = sel.reshape(batch*tokens, topk)[token, choice]
+  xv = x.reshape(batch*tokens, xchoices, in_features)
+  acc = UOp.placeholder((output_tile,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0.0)))
+  group = UOp.range(in_features//64, 2, AxisType.REDUCE)
+  activation = xv[token, 0 if xchoices == 1 else choice, group*64+lane].float()
+  updates = []
+  for i,output in enumerate(outputs):
+    packed = weight[expert, output, group*32+lane//2]
+    code = (packed >> ((lane&1)*4).cast(dtypes.uint8)) & 15
+    weight_value = _mxfp4_value(code) * _e8m0_value(scale[expert, output, group*2+lane//32])
+    updates.append(acc.after(group)[i].load()+activation*weight_value)
+  update = acc.store(UOp.stack(*updates)).end(group)
+  half_totals = tuple(warp_reduce(acc.after(update)[i], full_wave=True) for i in range(output_tile))
+  local = UOp.placeholder((output_tile, 2), dtypes.float32, slot=1, addrspace=AddrSpace.LOCAL)
+  half = (lane//32).valid((lane&31).eq(0))
+  barrier = UOp.group(*(local[i, half].store(total) for i,total in enumerate(half_totals))).barrier()
+  out = out.reshape(batch*tokens, topk, out_features, partials)
+  stores = (out[token, choice, output, partial.valid(lane.eq(0))].store(
+    (local.after(barrier)[i, 0]+local.after(barrier)[i, 1]).cast(out.dtype)) for i,output in enumerate(outputs))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="mxfp4_expert_linear_wave64_prefill", opts_to_apply=()))
+
+@functools.cache
+def _bf16_partial_linear_kernel(out:UOp, x:UOp, weight:UOp) -> UOp:
+  """Per-device BF16 down projection; its dummy final axis is reduced after combining TP partials."""
+  batch, tokens, out_features, partials = cast(tuple[int, int, int, int], out.shape)
+  in_features, output_tile = cast(int, x.shape[-1]), 1
+  assert out_features % output_tile == 0 and in_features % 32 == 0
+  row, lane = UOp.range(batch*tokens*(out_features//output_tile)*partials, 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  partial, output_block, token = row%partials, (row//partials)%(out_features//output_tile), row//(partials*(out_features//output_tile))
+  outputs = tuple(output_block*output_tile+i for i in range(output_tile))
+  acc = UOp.placeholder((output_tile,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(acc.const_like(0.0)))
+  group = UOp.range(in_features//32, 2, AxisType.REDUCE)
+  input_idx = group*32+lane
+  activation = x.reshape(batch*tokens, in_features)[token, input_idx].float()
+  update = acc.store(UOp.stack(*(acc.after(group)[i].load()+(activation*weight[output, input_idx].float()).cast(dtypes.bfloat16).float()
+    for i,output in enumerate(outputs)))).end(group)
+  stores = (out.reshape(batch*tokens, out_features, partials)[token, output, partial.valid(lane.eq(0))].store(
+    warp_reduce(acc.after(update)[i], full_wave=True)) for i,output in enumerate(outputs))
+  return UOp.group(*stores).end(lane, row).sink(arg=KernelInfo(name="bf16_partial_linear", opts_to_apply=()))
+
+@functools.cache
+def _bf16_matvec_kernel(out:UOp, x:UOp, weight:UOp) -> UOp:
+  batch, tokens, out_features = cast(tuple[int, int, int], out.shape)
+  in_features = cast(int, x.shape[-1])
+  assert in_features % 32 == 0
+  row, lane = UOp.range(batch*tokens*out_features, 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  token, output = row//out_features, row%out_features
+  acc = UOp.placeholder((), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  acc = acc.after(acc.store(0.0))
+  group = UOp.range(in_features//32, 2, AxisType.REDUCE)
+  input_idx = group*32+lane
+  product = (x.reshape(batch*tokens, in_features)[token, input_idx].float()*weight[output, input_idx].float()).cast(dtypes.bfloat16).float()
+  update = acc.store(acc.after(group)+product).end(group)
+  total = warp_reduce(acc.after(update)[0], full_wave=True)
+  return out.reshape(batch*tokens, out_features)[token, output.valid(lane.eq(0))].store(total.cast(out.dtype)).end(lane, row).sink(
+    arg=KernelInfo(name="bf16_matvec", opts_to_apply=()))
+
+@functools.cache
+def _gated_delta_prefill_kernel(core:UOp, next_state:UOp, q:UOp, k:UOp, v:UOp, beta:UOp, alpha:UOp, state:UOp, kq:UOp) -> UOp:
+  batch, heads, tokens, value_dim, row_tile = *core.shape, 4
+  key_dim, alpha_dim = q.shape[-1], alpha.shape[-1] if len(alpha.shape) == 4 else 1
+  assert all(isinstance(x, int) for x in (batch, heads, tokens, value_dim, key_dim)) and key_dim % 32 == 0 and value_dim % row_tile == 0
+  batch, heads, tokens, value_dim, key_dim = cast(tuple[int, int, int, int, int], (batch, heads, tokens, value_dim, key_dim))
+  core, v = (x.reshape(batch*heads, tokens, value_dim) for x in (core, v))
+  q, k = (x.reshape(batch*heads, tokens, key_dim) for x in (q, k))
+  beta, kq = (x.reshape(batch*heads, tokens) for x in (beta, kq))
+  alpha = alpha.reshape(batch*heads, tokens, alpha_dim)
+  state, next_state = (x.reshape(batch*heads, value_dim, key_dim) for x in (state, next_state))
+  bh_row, lane = UOp.range(batch*heads*value_dim//row_tile, 0), UOp.range(32, 1, axis_type=AxisType.LOCAL)
+  bh, row_base = bh_row // (value_dim//row_tile), (bh_row % (value_dim//row_tile))*row_tile
+  rows = tuple(row_base+i for i in range(row_tile))
+  cols = tuple(lane + i*32 for i in range(key_dim//32))
+  current = UOp.placeholder((row_tile*key_dim//32,), dtypes.float32, slot=0, addrspace=AddrSpace.REG)
+  current = current.after(current.store(UOp.stack(*(state[bh, row, col].float() for row in rows for col in cols))))
+  token = UOp.range(tokens, 2, AxisType.REDUCE)
+  keys = tuple(k[bh, token, col].load() for col in cols)
+  queries = tuple(q[bh, token, col].load() for col in cols)
+  updates:list[UOp] = []
+  stores:list[UOp] = []
+  for row_idx,row in enumerate(rows):
+    previous = tuple(current.after(token)[row_idx*key_dim//32+i].load() for i in range(key_dim//32))
+    av = tuple(alpha[bh, token, col if alpha_dim > 1 else 0].load() for col in cols)
+    bv = beta[bh, token].load()
+    state_k = warp_reduce(sum((x*a*y for x,a,y in zip(previous, av, keys)), UOp.const(0, dtypes.float32)), full_wave=True)
+    state_q = warp_reduce(sum((x*a*y for x,a,y in zip(previous, av, queries)), UOp.const(0, dtypes.float32)), full_wave=True)
+    delta = (v[bh, token, row].load() - state_k) * bv
+    updates += [x*a + delta*y for x,a,y in zip(previous, av, keys)]
+    stores.append(core[bh, token, row.valid(lane.eq(0))].store(state_q + delta*kq[bh, token]))
+  step = UOp.group(*stores, current.store(UOp.stack(*updates))).end(token)
+  state_stores = (next_state[bh, row, col].store(current.after(step)[row_idx*key_dim//32+i].load().cast(next_state.dtype))
+                  for row_idx,row in enumerate(rows) for i,col in enumerate(cols))
+  return UOp.group(*state_stores).end(lane, bh_row).sink(arg=KernelInfo(name="gated_delta_prefill", opts_to_apply=()))

tinygrad/llm/kimi.py

@@ -0,0 +1,181 @@
+from __future__ import annotations
+import base64, gc, json, pathlib, shutil
+from tinygrad import Tensor, Device, dtypes, nn
+from tinygrad.nn.state import safe_load, safe_save
+from tinygrad.llm.model import Transformer, TransformerConfig, SSMConfig
+from tinygrad.llm.quant import quantize_mxfp4_cpu
+
+KIMI_SSM_LAYERS = tuple(i not in (3, 7, 11, 15, 19, 23, 26) for i in range(27))
+KIMI_TENSOR_COUNT, KIMI_LOGICAL_BYTES = 688, 29_051_930_368
+KIMI_CHECKPOINT_FORMAT = "tinygrad-kimi-mxfp4-v2"
+
+def kimi_config(max_context:int, expert_mxfp4:bool=True) -> TransformerConfig:
+  return TransformerConfig(num_blocks=27, dim=2304, hidden_dim=1024, n_heads=32, n_kv_heads=32, norm_eps=1e-5,
+    vocab_size=163840, head_dim=192, rope_theta=10000.0, rope_dim=64, v_head_dim=128, q_lora_rank=0, kv_lora_rank=512,
+    num_experts=256, num_experts_per_tok=8, norm_topk_prob=True, shared_expert_dim=1024, leading_dense_blocks=1,
+    dense_hidden_dim=9216, routed_scaling_factor=2.446, expert_bias=True, max_context=max_context, expert_mxfp4=expert_mxfp4,
+    shared_expert_gate=False, bf16_activations=True, kda_split_qkv=True,
+    recurrent_prefill_chunked=True, recurrent_prefill_chunk_size=32,
+    ssm=SSMConfig(conv_kernel=4, state_size=128, group_count=32, time_step_rank=32, inner_size=4096, kda=True),
+    ssm_layers=KIMI_SSM_LAYERS)
+
+def _shard_kimi(model:Transformer, devices:tuple[str, ...]) -> None:
+  """Tensor-parallel layout. Every GPU owns a slice of every expert (not a replicated expert set)."""
+  for name, value in nn.state.get_state_dict(model).items():
+    axis = None
+    if name in ("token_embd.weight", "output.weight"): axis = 0
+    elif ".ffn_gate_exps.weight" in name or ".ffn_up_exps.weight" in name: axis = 1
+    elif ".ffn_gate_exps.weight_scale" in name or ".ffn_up_exps.weight_scale" in name: axis = 1
+    elif ".ffn_down_exps.weight" in name or ".ffn_down_exps.weight_scale" in name: axis = 2
+    elif name.endswith((".ffn_gate.weight", ".ffn_up.weight", ".ffn_gate_shexp.weight", ".ffn_up_shexp.weight")): axis = 0
+    elif name.endswith((".ffn_down.weight", ".ffn_down_shexp.weight", ".attn_output.weight", ".ssm_out.weight")): axis = 1
+    elif name.endswith((".attn_q.weight", ".attn_k.weight", ".attn_v.weight", ".attn_qkv.weight",
+                        ".ssm_f_b.weight", ".ssm_g_b.weight", ".ssm_beta.weight")): axis = 0
+    elif name.endswith((".ssm_conv1d.weight", ".ssm_q_conv1d.weight", ".ssm_k_conv1d.weight", ".ssm_v_conv1d.weight")): axis = 0
+    elif name.endswith((".ssm_a", ".ssm_dt.bias")): axis = 0
+    elif name.endswith((".attn_k_b.weight", ".attn_v_b.weight")): axis = 0
+    value.shard_(devices, axis=axis)
+
+def _validate_kimi_state(model:Transformer, state:dict[str, Tensor]) -> None:
+  model_state = nn.state.get_state_dict(model)
+  missing, unexpected = set(model_state)-set(state), set(state)-set(model_state)
+  if missing or unexpected: raise ValueError(f"invalid Kimi tensor names: missing={sorted(missing)}, unexpected={sorted(unexpected)}")
+  for name, value in state.items():
+    if value.shape != model_state[name].shape: raise ValueError(f"invalid shape for {name}: expected {model_state[name].shape}, got {value.shape}")
+    expected_dtype = dtypes.uint8 if name.endswith((".weight_scale", "_exps.weight")) else dtypes.bfloat16
+    if value.dtype != expected_dtype: raise ValueError(f"invalid dtype for {name}: expected {expected_dtype}, got {value.dtype}")
+  if len(state) != KIMI_TENSOR_COUNT or (nbytes := sum(x.nbytes() for x in state.values())) != KIMI_LOGICAL_BYTES:
+    raise ValueError(f"invalid Kimi checkpoint size: {len(state)} tensors, {nbytes} bytes")
+
+def _load_converted_state(model_dir:pathlib.Path, files:list[str]) -> dict[str, Tensor]:
+  state:dict[str, Tensor] = {}
+  for filename in files:
+    part = safe_load(model_dir / filename)
+    if duplicates := set(state) & set(part): raise ValueError(f"duplicate Kimi tensors in {filename}: {sorted(duplicates)}")
+    state.update(part)
+  return state
+
+def load_kimi(model_dir:str|pathlib.Path, max_context:int=4096, devices:int=4) -> Transformer:
+  model_dir = pathlib.Path(model_dir)
+  manifest = json.loads((model_dir / "tinygrad-kimi.json").read_text())
+  if manifest.get("format") != KIMI_CHECKPOINT_FORMAT: raise ValueError("unsupported Kimi checkpoint format")
+  if devices != 4: raise ValueError("Kimi-Linear MXFP4 checkpoint currently requires TP4 (--devices 4)")
+  devs = tuple(f"{Device.DEFAULT}:{i}" for i in range(devices))
+  model = Transformer(kimi_config(max_context, expert_mxfp4=True))
+  state = _load_converted_state(model_dir, manifest["files"])
+  _validate_kimi_state(model, state)
+  _shard_kimi(model, devs)
+  nn.state.load_state_dict(model, state, strict=True, consume=True, realize=True)
+  if state: raise ValueError(f"unexpected Kimi tensors: {sorted(state)}")
+  return model
+
+def _load_hf_state(src:pathlib.Path) -> dict[str, Tensor]:
+  index = json.loads((src / "model.safetensors.index.json").read_text())
+  state:dict[str, Tensor] = {}
+  for filename in sorted(set(index["weight_map"].values())): state.update(safe_load(src / filename))
+  return state
+
+def _layer_key(i:int, suffix:str) -> str: return f"model.layers.{i}.{suffix}"
+
+def _convert_attention(sd:dict[str, Tensor], i:int, is_kda:bool, consume:bool=False) -> dict[str, Tensor]:
+  p, out = f"blk.{i}.", {}
+  def get(suffix:str) -> Tensor:
+    key = _layer_key(i, suffix)
+    return (sd.pop(key) if consume else sd[key]).to("CPU")
+  if is_kda:
+    for src_name, dst_name in (("q_proj", "attn_q"), ("k_proj", "attn_k"), ("v_proj", "attn_v")):
+      out[p+dst_name+".weight"] = get(f"self_attn.{src_name}.weight")
+    for src_name, dst_name in (("q_conv1d", "ssm_q_conv1d"), ("k_conv1d", "ssm_k_conv1d"), ("v_conv1d", "ssm_v_conv1d")):
+      out[p+dst_name+".weight"] = get(f"self_attn.{src_name}.weight").squeeze(1)
+    for src_name, dst_name in (("f_a_proj", "ssm_f_a"), ("f_b_proj", "ssm_f_b"), ("g_a_proj", "ssm_g_a"),
+                               ("g_b_proj", "ssm_g_b"), ("b_proj", "ssm_beta"), ("o_proj", "ssm_out")):
+      out[p+dst_name+".weight"] = get(f"self_attn.{src_name}.weight")
+    out[p+"ssm_norm.weight"] = get("self_attn.o_norm.weight")
+    out[p+"ssm_dt.bias"] = get("self_attn.dt_bias")
+    out[p+"ssm_a"] = (-get("self_attn.A_log").float().exp()).reshape(32, 1)
+  else:
+    out[p+"attn_q.weight"] = get("self_attn.q_proj.weight")
+    out[p+"attn_kv_a_mqa.weight"] = get("self_attn.kv_a_proj_with_mqa.weight")
+    out[p+"attn_kv_a_norm.weight"] = get("self_attn.kv_a_layernorm.weight")
+    kv_b = get("self_attn.kv_b_proj.weight").reshape(32, 256, 512)
+    k_b, v_b = kv_b[:, :128], kv_b[:, 128:]
+    out[p+"attn_k_b.weight"], out[p+"attn_v_b.weight"] = k_b.transpose(1, 2), v_b
+    out[p+"attn_output.weight"] = get("self_attn.o_proj.weight")
+  return out
+
+def convert_kimi(src_dir:str|pathlib.Path, dst_dir:str|pathlib.Path) -> None:
+  """Stream the official BF16 checkpoint into the tinygrad TP4 MXFP4/BF16 representation."""
+  src, dst = pathlib.Path(src_dir), pathlib.Path(dst_dir)
+  dst.mkdir(parents=True, exist_ok=True)
+  config = json.loads((src / "config.json").read_text())
+  expected = {"hidden_size":2304, "num_hidden_layers":27, "num_attention_heads":32, "num_key_value_heads":32,
+              "vocab_size":163840, "intermediate_size":9216, "num_experts":256, "num_experts_per_token":8,
+              "moe_intermediate_size":1024, "num_shared_experts":1, "qk_nope_head_dim":128, "qk_rope_head_dim":64,
+              "v_head_dim":128, "kv_lora_rank":512, "first_k_dense_replace":1, "mla_use_nope":True}
+  if any(config.get(k) != v for k,v in expected.items()): raise ValueError(f"not Kimi-Linear-48B-A3B: expected {expected}")
+  sd, files = _load_hf_state(src), []
+
+  common = {"token_embd.weight":sd.pop("model.embed_tokens.weight").to("CPU"), "output_norm.weight":sd.pop("model.norm.weight").to("CPU"),
+            "output.weight":sd.pop("lm_head.weight").to("CPU")}
+  safe_save(common, str(dst / "model-common.safetensors"))
+  del common
+  gc.collect()
+  files.append("model-common.safetensors")
+  for i in range(27):
+    p = f"blk.{i}."
+    layer = _convert_attention(sd, i, KIMI_SSM_LAYERS[i], consume=True)
+    layer[p+"attn_norm.weight"] = sd.pop(_layer_key(i, "input_layernorm.weight")).to("CPU")
+    layer[p+"ffn_norm.weight"] = sd.pop(_layer_key(i, "post_attention_layernorm.weight")).to("CPU")
+    if i == 0:
+      for src_name, dst_name in (("gate_proj", "ffn_gate"), ("up_proj", "ffn_up"), ("down_proj", "ffn_down")):
+        layer[p+dst_name+".weight"] = sd.pop(_layer_key(i, f"mlp.{src_name}.weight")).to("CPU")
+    else:
+      base = _layer_key(i, "block_sparse_moe")
+      layer[p+"ffn_gate_inp.weight"] = sd.pop(base+".gate.weight").to("CPU")
+      # The official name is e_score_correction_bias; tolerate the early checkpoint spelling.
+      bias_name = next(k for k in (base+".gate.e_score_correction_bias", base+".gate.e_score_correction") if k in sd)
+      layer[p+"exp_probs_b.bias"] = sd.pop(bias_name).to("CPU")
+      for src_name, dst_name in (("gate_proj", "ffn_gate_shexp"), ("up_proj", "ffn_up_shexp"), ("down_proj", "ffn_down_shexp")):
+        layer[p+dst_name+".weight"] = sd.pop(base+f".shared_experts.{src_name}.weight").to("CPU")
+    layer_file = f"model-layer-{i:02d}.safetensors"
+    safe_save({k:v.cast(dtypes.bfloat16).contiguous() for k,v in layer.items()}, str(dst/layer_file))
+    del layer
+    gc.collect()
+    files.append(layer_file)
+
+    if i:
+      base = _layer_key(i, "block_sparse_moe.experts")
+      for wid, dst_name in (("w1", "ffn_gate_exps"), ("w3", "ffn_up_exps"), ("w2", "ffn_down_exps")):
+        packed, scales = [], []
+        for e in range(256):
+          q, s = quantize_mxfp4_cpu(sd.pop(f"{base}.{e}.{wid}.weight").to("CPU"))
+          packed.append(q)
+          scales.append(s)
+        expert_file = f"model-layer-{i:02d}-{wid}-mxfp4.safetensors"
+        safe_save({p+dst_name+".weight":Tensor.stack(*packed), p+dst_name+".weight_scale":Tensor.stack(*scales)}, str(dst/expert_file))
+        del packed, scales, q, s
+        gc.collect()
+        files.append(expert_file)
+
+  if sd: raise ValueError(f"unconverted Kimi source tensors: {sorted(sd)}")
+  for name in ("config.json", "tokenizer_config.json", "special_tokens_map.json", "tiktoken.model", "chat_template.jinja"):
+    if (src/name).exists(): shutil.copy2(src/name, dst/name)
+  converted = _load_converted_state(dst, files)
+  _validate_kimi_state(Transformer(kimi_config(max_context=1)), converted)
+  manifest = {"format":KIMI_CHECKPOINT_FORMAT, "tensor_count":KIMI_TENSOR_COUNT,
+              "logical_bytes":KIMI_LOGICAL_BYTES, "files":files}
+  (dst / "tinygrad-kimi.json").write_text(json.dumps(manifest, indent=2)+"\n")
+
+def load_kimi_tokenizer_data(model_dir:str|pathlib.Path) -> tuple[dict[str, int], dict[str, int], int, int]:
+  """Return byte-encoded normal tokens and specials for SimpleTokenizer without transformers/tiktoken."""
+  model_dir = pathlib.Path(model_dir)
+  normal:dict[str, int] = {}
+  bs = [*range(33, 127), *range(161, 173), *range(174, 256)]
+  byte_encoder = {b:chr(b) for b in bs} | {b:chr(256+i) for i,b in enumerate(b for b in range(256) if b not in bs)}
+  for line in (model_dir / "tiktoken.model").read_bytes().splitlines():
+    token, rank = line.split()
+    normal["".join(byte_encoder[b] for b in base64.b64decode(token))] = int(rank)
+  tc = json.loads((model_dir / "tokenizer_config.json").read_text())
+  specials = {v["content"]:int(k) for k,v in tc.get("added_tokens_decoder", {}).items()}
+  cfg = json.loads((model_dir / "config.json").read_text())
+  return normal, specials, cfg["bos_token_id"], cfg["eos_token_id"]

tinygrad/llm/kimi_k3.py

@@ -0,0 +1,337 @@
+from __future__ import annotations
+import gc, json, math, pathlib
+from dataclasses import replace
+from collections import defaultdict
+from typing import Callable, cast
+from tinygrad import Tensor, Device, dtypes, nn
+from tinygrad.device import Buffer
+from tinygrad.uop.ops import UOp
+from tinygrad.nn.state import safe_dtypes, safe_load_metadata
+from tinygrad.llm.kimi import load_kimi_tokenizer_data
+from tinygrad.llm.model import SSMConfig, Transformer, TransformerConfig
+
+KIMI_K3_TOTAL_SIZE = 1_560_860_324_864
+KIMI_K3_TEXT_SIZE = 1_559_965_606_912
+KIMI_K3_TP8_BYTES_PER_GPU = 196_784_397_312
+KIMI_K3_SHARDS = 96
+KIMI_K3_EXPERTS = 896
+KIMI_K3_LAYERS = 93
+KIMI_K3_FULL_ATTN_LAYERS = (*range(3, KIMI_K3_LAYERS, 4), 92)
+KIMI_K3_SSM_LAYERS = tuple(i not in KIMI_K3_FULL_ATTN_LAYERS for i in range(KIMI_K3_LAYERS))
+
+def kimi_k3_config(max_context:int) -> TransformerConfig:
+  """Official Kimi K3 text-tower configuration (zero-based full-attention layers)."""
+  return TransformerConfig(num_blocks=93, dim=7168, hidden_dim=3072, n_heads=96, n_kv_heads=96, norm_eps=1e-5,
+    vocab_size=163840, head_dim=192, rope_theta=10000.0, rope_dim=64, v_head_dim=128, max_context=max_context,
+    q_lora_rank=1536, kv_lora_rank=512, num_experts=896, num_experts_per_tok=16, norm_topk_prob=True,
+    shared_expert_dim=6144, leading_dense_blocks=1, dense_hidden_dim=33792, routed_scaling_factor=1.0,
+    expert_bias=True, expert_mxfp4=True, bf16_activations=True, kda_split_qkv=True,
+    ssm=SSMConfig(conv_kernel=4, state_size=128, group_count=96, time_step_rank=96, inner_size=12288, kda=True, channel_decay=True),
+    ssm_layers=KIMI_K3_SSM_LAYERS, shared_expert_gate=False, attn_output_gate=True,
+    activation_situ_beta=4.0, activation_situ_linear_beta=25.0, routed_expert_dim=3584, latent_moe_norm=True,
+    route_weights_uncorrected=True, attn_res_block_size=12, kda_full_rank_gate=True, kda_gate_lower_bound=-5.0,
+    recurrent_prefill_chunked=True, recurrent_prefill_chunk_size=128)
+
+def kimi_k3_smoke_config(max_context:int=4) -> TransformerConfig:
+  """Reduced K3 with every architectural feature retained for cheap compile/hardware admission tests."""
+  # Keep both the routed latent and the TP8-local expert hidden dimension wave64 aligned. The real
+  # gfx950 packed-expert kernel requires this, so the hardware smoke test must preserve the constraint.
+  return replace(kimi_k3_config(max_context), num_blocks=2, dim=32, hidden_dim=512, n_heads=8, n_kv_heads=8,
+    vocab_size=64, head_dim=8, rope_dim=4, v_head_dim=4, q_lora_rank=16, kv_lora_rank=8, num_experts=512,
+    num_experts_per_tok=2, shared_expert_dim=32, dense_hidden_dim=64, routed_expert_dim=64,
+    ssm=SSMConfig(4, 4, 8, 8, 32, True, True), ssm_layers=(True, False), attn_res_block_size=1)
+
+def _shard_kimi_k3(model:Transformer, devices:tuple[str, ...]) -> None:
+  """Tensor parallel layout for K3. The official dimensions are divisible by TP8."""
+  if len(devices) not in (1, 2, 4, 8): raise ValueError(f"Kimi K3 tensor parallelism requires 1, 2, 4, or 8 devices, got {len(devices)}")
+  for name, value in nn.state.get_state_dict(model).items():
+    axis = None
+    if name in ("token_embd.weight", "output.weight"): axis = 0
+    elif ".ffn_gate_exps.weight" in name or ".ffn_up_exps.weight" in name: axis = 1
+    elif ".ffn_gate_exps.weight_scale" in name or ".ffn_up_exps.weight_scale" in name: axis = 1
+    elif ".ffn_down_exps.weight" in name or ".ffn_down_exps.weight_scale" in name: axis = 2
+    elif name.endswith((".ffn_gate.weight", ".ffn_up.weight", ".ffn_gate_shexp.weight", ".ffn_up_shexp.weight")): axis = 0
+    elif name.endswith((".ffn_down.weight", ".ffn_down_shexp.weight", ".ffn_routed_down.weight", ".ffn_routed_up.weight",
+                        ".attn_output.weight", ".ssm_out.weight")): axis = 1
+    elif name.endswith((".attn_q_b.weight", ".attn_k_b.weight", ".attn_v_b.weight", ".attn_gate.weight",
+                        ".attn_q.weight", ".attn_k.weight", ".attn_v.weight", ".ssm_f_b.weight", ".ssm_g_full.weight", ".ssm_beta.weight")): axis = 0
+    elif name.endswith((".ssm_q_conv1d.weight", ".ssm_k_conv1d.weight", ".ssm_v_conv1d.weight", ".ssm_dt.bias")): axis = 0
+    value.shard_(devices, axis=axis)
+
+def _validate_config(config:dict) -> None:
+  text = config.get("text_config", config)
+  expected = {"model_type":"kimi_linear", "hidden_size":7168, "num_hidden_layers":93, "num_attention_heads":96,
+              "vocab_size":163840, "intermediate_size":33792, "num_experts":896, "num_experts_per_token":16,
+              "moe_intermediate_size":3072, "num_shared_experts":2, "q_lora_rank":1536, "kv_lora_rank":512,
+              "qk_nope_head_dim":128, "qk_rope_head_dim":64, "v_head_dim":128, "routed_expert_hidden_size":3584,
+              "attn_res_block_size":12, "hidden_act":"situ", "mla_use_nope":True, "mla_use_output_gate":True,
+              "activation_situ_beta":4.0, "activation_situ_linear_beta":25.0, "latent_moe_use_norm":True,
+              "moe_renormalize":True, "first_k_dense_replace":1, "num_expert_group":1, "topk_group":1}
+  bad = {k:(text.get(k), v) for k,v in expected.items() if text.get(k) != v}
+  linear = text.get("linear_attn_config", {})
+  linear_expected = {"head_dim":128, "num_heads":96, "short_conv_kernel_size":4, "use_full_rank_gate":True,
+                     "gate_lower_bound":-5.0, "full_attn_layers":[i+1 for i in KIMI_K3_FULL_ATTN_LAYERS],
+                     "kda_layers":[i+1 for i,x in enumerate(KIMI_K3_SSM_LAYERS) if x]}
+  bad.update({f"linear_attn_config.{k}":(linear.get(k), v) for k,v in linear_expected.items() if linear.get(k) != v})
+  quant = text.get("quantization_config", {})
+  if quant.get("format") != "mxfp4-pack-quantized": bad["quantization_config.format"] = (quant.get("format"), "mxfp4-pack-quantized")
+  if bad: raise ValueError(f"not the supported official Kimi K3 checkpoint: {bad}")
+
+def audit_kimi_k3_checkpoint(model_dir:str|pathlib.Path, require_shards:bool=True) -> dict[str, int]:
+  """Validate checkpoint metadata only. This never opens weight data and is safe on small hosts."""
+  root = pathlib.Path(model_dir)
+  _validate_config(json.loads((root / "config.json").read_text()))
+  index = json.loads((root / "model.safetensors.index.json").read_text())
+  weight_map, total = index.get("weight_map", {}), index.get("metadata", {}).get("total_size")
+  language = [k for k in weight_map if k.startswith("language_model.")]
+  experts = [k for k in language if ".block_sparse_moe.experts." in k]
+  missing_files = {fn for fn in weight_map.values() if not (root / fn).is_file()}
+  if total != KIMI_K3_TOTAL_SIZE: raise ValueError(f"unexpected checkpoint size {total}, expected {KIMI_K3_TOTAL_SIZE}")
+  if len(set(weight_map.values())) != KIMI_K3_SHARDS: raise ValueError("official Kimi K3 must contain 96 safetensor shards")
+  if len(experts) != 92 * KIMI_K3_EXPERTS * 3 * 2: raise ValueError(f"unexpected routed-expert tensor count {len(experts)}")
+  if require_shards and missing_files: raise FileNotFoundError(f"missing {len(missing_files)} checkpoint shards, first: {sorted(missing_files)[0]}")
+  return {"tensors":len(weight_map), "language_tensors":len(language), "expert_tensors":len(experts),
+          "shards":len(set(weight_map.values())), "missing_shards":len(missing_files), "total_size":total}
+
+def _layer_sources(i:int, is_kda:bool) -> dict[str, str]:
+  src, dst = f"language_model.model.layers.{i}.", f"blk.{i}."
+  out = {
+    src+"input_layernorm.weight":dst+"attn_norm.weight", src+"post_attention_layernorm.weight":dst+"ffn_norm.weight",
+    src+"self_attention_res_norm.weight":dst+"attn_res_norm.weight", src+"self_attention_res_proj.weight":dst+"attn_res_proj.weight",
+    src+"mlp_res_norm.weight":dst+"mlp_res_norm.weight", src+"mlp_res_proj.weight":dst+"mlp_res_proj.weight",
+  }
+  if is_kda:
+    for a,b in (("q_proj","attn_q"),("k_proj","attn_k"),("v_proj","attn_v"),("g_proj","ssm_g_full"),
+                ("f_a_proj","ssm_f_a"),("f_b_proj","ssm_f_b"),("b_proj","ssm_beta"),("o_proj","ssm_out")):
+      out[src+f"self_attn.{a}.weight"] = dst+b+".weight"
+    for a,b in (("q_conv1d","ssm_q_conv1d"),("k_conv1d","ssm_k_conv1d"),("v_conv1d","ssm_v_conv1d")):
+      out[src+f"self_attn.{a}.weight"] = dst+b+".weight"
+    out[src+"self_attn.o_norm.weight"], out[src+"self_attn.dt_bias"], out[src+"self_attn.A_log"] = \
+      dst+"ssm_norm.weight", dst+"ssm_dt.bias", dst+"ssm_a"
+  else:
+    for a,b in (("q_a_proj","attn_q_a"),("q_a_layernorm","attn_q_a_norm"),("q_b_proj","attn_q_b"),
+                ("kv_a_proj_with_mqa","attn_kv_a_mqa"),("kv_a_layernorm","attn_kv_a_norm"),
+                ("g_proj","attn_gate"),("o_proj","attn_output")):
+      out[src+f"self_attn.{a}.weight"] = dst+b+".weight"
+    # kv_b_proj is split into head-wise K and V tensors while loading.
+    out[src+"self_attn.kv_b_proj.weight"] = dst+"attn_k_b.weight|"+dst+"attn_v_b.weight"
+  if i == 0:
+    for a,b in (("gate_proj","ffn_gate"),("up_proj","ffn_up"),("down_proj","ffn_down")): out[src+f"mlp.{a}.weight"] = dst+b+".weight"
+  else:
+    base = src+"block_sparse_moe."
+    out[base+"gate.weight"], out[base+"gate.e_score_correction_bias"] = dst+"ffn_gate_inp.weight", dst+"exp_probs_b.bias"
+    for a,b in (("gate_proj","ffn_gate_shexp"),("up_proj","ffn_up_shexp"),("down_proj","ffn_down_shexp"),
+                ("routed_expert_down_proj","ffn_routed_down"),("routed_expert_up_proj","ffn_routed_up"),
+                ("routed_expert_norm","ffn_routed_norm")):
+      out[base+(f"shared_experts.{a}.weight" if a.endswith("_proj") and not a.startswith("routed_") else a+".weight")] = dst+b+".weight"
+  return out
+
+def _replace(dst:Tensor, src:Tensor) -> None:
+  if dst.shape != src.shape: raise ValueError(f"shape mismatch: expected {dst.shape}, got {src.shape}")
+  if not isinstance(dst.device, tuple):
+    dst.replace(src.to(dst.device)).realize()
+    return
+  if isinstance(src.device, tuple):
+    dst.replace(src.shard_like(dst)).realize()
+    return
+
+  # Build the final MultiBuffer directly. The generic shard().realize() path schedules several
+  # kernels per tensor and recompiles them for every DISK:<filename> device. Axis-0 shards and
+  # replicas are contiguous, so copy those bytes straight into their final device buffers.
+  devices, axis, shape = dst.device, dst.uop.axis, tuple(int(x) for x in dst.shape)
+  try: src_buffer = cast(Buffer, src.uop.buffer)
+  except (AssertionError, RuntimeError):
+    src = src.clone().realize()
+    src_buffer = cast(Buffer, src.uop.buffer)
+
+  if axis is None:
+    # Replicas are identical on every device. Read the disk tensor once, retain that allocation on
+    # GPU 0, and fan it out over XGMI instead of issuing eight identical direct reads.
+    staging = Tensor.empty(*shape, dtype=src.dtype, device=devices[0]).realize()
+    cast(Buffer, staging.uop.buffer).ensure_allocated().copy_from(src_buffer.ensure_allocated())
+    parts = [staging]
+    for device in devices[1:]:
+      part = Tensor.empty(*shape, dtype=src.dtype, device=device).realize()
+      cast(Buffer, part.uop.buffer).ensure_allocated().copy_from(cast(Buffer, staging.uop.buffer).ensure_allocated())
+      parts.append(part)
+    dst.replace(Tensor(parts[0].uop.mstack(*(x.uop for x in parts[1:]))))
+    return
+  if axis == 0:
+    part_shape = (shape[0]//len(devices), *shape[1:])
+    part_numel = math.prod(part_shape)
+    parts:list[Tensor] = []
+    for i,device in enumerate(devices):
+      part = Tensor.empty(*part_shape, dtype=src.dtype, device=device).realize()
+      source = src_buffer.view(part_numel, src.dtype, i*part_numel*src.dtype.itemsize)
+      cast(Buffer, part.uop.buffer).ensure_allocated().copy_from(source.ensure_allocated())
+      parts.append(part)
+  else:
+    # Inner-axis TP slices are strided in row-major safetensors. Stage one complete tensor on
+    # GPU 0, then schedule all slice kernels and peer copies as one multi-device graph.
+    staging = Tensor.empty(*shape, dtype=src.dtype, device=devices[0]).realize()
+    cast(Buffer, staging.uop.buffer).ensure_allocated().copy_from(src_buffer.ensure_allocated())
+    dst.replace(staging.shard(devices, axis=axis)).realize()
+    return
+  dst.replace(Tensor(parts[0].uop.mstack(*(x.uop for x in parts[1:])).unshard(axis)))
+
+def _safe_load_selected(fn:pathlib.Path, keys:tuple[str, ...]|list[str]) -> dict[str, Tensor]:
+  """Create disk-backed tensors only for selected safetensor entries, without touching payload data."""
+  source, data_start, metadata = safe_load_metadata(fn)
+  data = source[data_start:]
+  missing = [key for key in keys if key not in metadata]
+  if missing: raise ValueError(f"missing tensor {missing[0]} from {fn.name}")
+  out:dict[str, Tensor] = {}
+  for key in keys:
+    entry = metadata[key]
+    out[key] = data[entry["data_offsets"][0]:entry["data_offsets"][1]].bitcast(safe_dtypes[entry["dtype"]]).reshape(entry["shape"])
+  return out
+
+def _load_stacked_experts(dst:Tensor, sources:list[Tensor]) -> None:
+  """Read expert tensors once into a transient GPU staging buffer, then redistribute TP slices over the GPU fabric."""
+  if not sources or not isinstance(dst.device, tuple) or dst.uop.axis is None: raise ValueError("expected a TP-sharded expert destination")
+  devices, axis = dst.device, dst.uop.axis
+  shape = (len(sources), *sources[0].shape)
+  if dst.shape != shape or any(x.shape != sources[0].shape or x.dtype != dst.dtype for x in sources):
+    raise ValueError(f"expert source shape/dtype does not match destination {dst.shape} {dst.dtype}")
+
+  # Each official expert tensor is contiguous in its safetensor file. Assemble it once on GPU 0;
+  # Buffer.copy_from uses the AMD driver's bounded DISK->GPU staging path and never allocates host-sized storage.
+  staging = Tensor.empty(*shape, dtype=dst.dtype, device=devices[0]).realize()
+  staging_buffer = cast(Buffer, staging.uop.buffer)
+  def free_staging_cache() -> None:
+    if (free_cache:=getattr(Device[devices[0]].allocator, "free_cache", None)) is not None: free_cache()
+  offset = 0
+  for source in sources:
+    source_buffer = cast(Buffer, source.uop.buffer)
+    staging_buffer.view(cast(int, source.numel()), source.dtype, offset).ensure_allocated().copy_from(source_buffer.ensure_allocated())
+    offset += source.nbytes()
+
+  # Schedule all TP slices and peer copies together. This avoids eight independent realization
+  # passes and lets the runtime overlap the multi-device transfer graph.
+  dst.replace(staging.shard(devices, axis=axis)).realize()
+  del staging
+  gc.collect()
+  free_staging_cache()
+
+def _load_nonexperts(root:pathlib.Path, weight_map:dict[str, str], model:Transformer, progress:Callable[[str], None]) -> set[str]:
+  model_state, mappings = nn.state.get_state_dict(model), {
+    "language_model.model.embed_tokens.weight":"token_embd.weight", "language_model.model.norm.weight":"output_norm.weight",
+    "language_model.lm_head.weight":"output.weight", "language_model.model.output_attn_res_norm.weight":"output_attn_res_norm.weight",
+    "language_model.model.output_attn_res_proj.weight":"output_attn_res_proj.weight"}
+  for i,is_kda in enumerate(KIMI_K3_SSM_LAYERS): mappings.update(_layer_sources(i, is_kda))
+  by_file:dict[str, list[str]] = defaultdict(list)
+  for source in mappings:
+    if source not in weight_map: raise ValueError(f"missing Kimi K3 tensor {source}")
+    by_file[weight_map[source]].append(source)
+  consumed:set[str] = set()
+  for filename, sources in sorted(by_file.items()):
+    progress(f"loading non-expert tensors from {filename}")
+    shard = _safe_load_selected(root / filename, sources)
+    for source in sources:
+      value, targets = shard[source], mappings[source].split("|")
+      # A_log is the only checkpoint tensor requiring arithmetic during load. Realize its 128
+      # channel values on CPU so replicating it does not try to render the disk/PYTHON graph.
+      if source.endswith("A_log"): value = (-value.to("CPU").float().exp()).reshape(model_state[targets[0]].shape).realize()
+      if source.endswith("conv1d.weight"): value = value.squeeze(1)
+      if source.endswith("kv_b_proj.weight"):
+        # Splitting K/V includes a transpose, which cannot be rendered against a disk buffer.
+        # Materialize only this one 25 MiB projection on CPU, then release it with the shard.
+        value = value.to("CPU").realize().reshape(96, 256, 512)
+        values:tuple[Tensor, ...] = (value[:, :128].transpose(1, 2), value[:, 128:])
+      else: values = (value,)
+      for target,tensor in zip(targets, values): _replace(model_state[target], tensor)
+      consumed.add(source)
+    del shard
+  gc.collect()
+  return consumed
+
+def _load_experts(root:pathlib.Path, weight_map:dict[str, str], model:Transformer, progress:Callable[[str], None]) -> set[str]:
+  model_state, consumed = nn.state.get_state_dict(model), set[str]()
+  for i in range(1, KIMI_K3_LAYERS):
+    base = f"language_model.model.layers.{i}.block_sparse_moe.experts"
+    fields = tuple((wid, suffix, dst_name) for wid,dst_name in (("w1","ffn_gate_exps"),("w2","ffn_down_exps"),("w3","ffn_up_exps"))
+                   for suffix in ("weight_packed", "weight_scale"))
+    keys = {(e,wid,suffix):f"{base}.{e}.{wid}.{suffix}" for e in range(KIMI_K3_EXPERTS) for wid,suffix,_ in fields}
+    files = sorted({weight_map[k] for k in keys.values()})
+    progress(f"loading layer {i}/92 routed experts from {', '.join(files)}")
+    shards = {fn:_safe_load_selected(root / fn, [key for key in keys.values() if weight_map[key] == fn]) for fn in files}
+
+    # Official K3 stores all six tensors for an expert contiguously and all 896 experts for a
+    # layer in one contiguous shard region, ordered lexicographically by expert name. Read that
+    # region once, then reorder/split on GPU 0 directly into the six TP8 destinations.
+    blocks:list[tuple[int, int, int, list[Buffer]]] = []
+    for e in range(KIMI_K3_EXPERTS):
+      bufs = [cast(Buffer, shards[weight_map[keys[e,wid,suffix]]][keys[e,wid,suffix]].uop.buffer) for wid,suffix,_ in fields]
+      if not all(bufs[j].device == bufs[0].device and bufs[j].offset+bufs[j].nbytes == bufs[j+1].offset for j in range(len(bufs)-1)):
+        raise ValueError(f"layer {i} expert {e} tensors are not contiguous in the official shard")
+      blocks.append((bufs[0].offset, bufs[-1].offset+bufs[-1].nbytes, e, bufs))
+    blocks.sort()
+    if len(files) != 1 or not all(blocks[j][1] == blocks[j+1][0] for j in range(len(blocks)-1)):
+      raise ValueError(f"layer {i} routed experts are not one contiguous official-shard region")
+    row_bytes = blocks[0][1]-blocks[0][0]
+    if any(end-start != row_bytes for start,end,_,_ in blocks): raise ValueError(f"layer {i} expert records have inconsistent sizes")
+
+    devices = cast(tuple[str, ...], model_state[f"blk.{i}.ffn_gate_exps.weight"].device)
+    raw = Tensor.empty(KIMI_K3_EXPERTS, row_bytes, dtype=dtypes.uint8, device=devices[0]).realize()
+    raw_buffer, first_buffer = cast(Buffer, raw.uop.buffer), blocks[0][3][0]
+    raw_buffer.ensure_allocated().copy_from(first_buffer.base.view(KIMI_K3_EXPERTS*row_bytes, dtypes.uint8, blocks[0][0]).ensure_allocated())
+    lexpos = {expert:pos for pos,(_,_,expert,_) in enumerate(blocks)}
+    permutation = Tensor([lexpos[e] for e in range(KIMI_K3_EXPERTS)], device=devices[0])
+    field_offset, outputs = 0, []
+    for field_idx,(wid,suffix,dst_name) in enumerate(fields):
+      field_bytes = blocks[0][3][field_idx].nbytes
+      dst = model_state[f"blk.{i}.{dst_name}.weight" + ("_scale" if suffix == "weight_scale" else "")]
+      axis = dst.uop.axis
+      if axis is None: raise ValueError(f"layer {i} expert destination {dst_name} is not TP-sharded")
+      value = raw[:, field_offset:field_offset+field_bytes][permutation].reshape(dst.shape).shard(devices, axis=axis)
+      # Realize into a buffer-identity tensor, then retain only that identity in the model. Keeping
+      # value's arithmetic UOp would also keep the 15.7 GB raw staging tensor and its reorder graph
+      # alive for every loaded weight, wasting about 44 GB on GPU 0 after the load completes.
+      shard_shape = tuple(int(x) for x in value.uop.shard_shape)
+      storage = UOp.new_buffer(devices, math.prod(shard_shape), dst.dtype).reshape(shard_shape).unshard(axis)
+      final = Tensor(storage)
+      final.assign(value)
+      outputs.append((dst, final, storage))
+      field_offset += field_bytes
+    if field_offset != row_bytes: raise ValueError(f"layer {i} expert field sizes do not cover the contiguous record")
+    outputs[0][1].realize(*(value for _,value,_ in outputs[1:]))
+    for dst,_,storage in outputs: dst.replace(Tensor(storage))
+    consumed.update(keys.values())
+    # Drop the realized assignment graphs before flushing the allocator cache. Their final storage
+    # UOps remain in model_state, while the graphs themselves still reference raw and permutation.
+    del shards, raw, raw_buffer, permutation, outputs, value, final, storage, dst
+    if (free_cache:=getattr(Device[devices[0]].allocator, "free_cache", None)) is not None: free_cache()
+  return consumed
+
+def load_kimi_k3(model_dir:str|pathlib.Path, max_context:int=4096, devices:int=8,
+                 progress:Callable[[str], None]=print) -> Transformer:
+  """Load the official native K3 checkpoint without ever materializing it in host RAM.
+
+  Safetensor shards remain disk-backed. Expert tensors are read once into a bounded GPU staging
+  buffer, redistributed as TP slices, and discarded after every projection. Vision tensors are intentionally ignored.
+  """
+  root = pathlib.Path(model_dir)
+  _validate_config(json.loads((root / "config.json").read_text()))
+  index = json.loads((root / "model.safetensors.index.json").read_text())
+  weight_map = index["weight_map"]
+  if devices != 8: raise ValueError("official Kimi K3 currently requires --devices 8")
+  if index.get("metadata", {}).get("total_size") != KIMI_K3_TOTAL_SIZE or len(set(weight_map.values())) != KIMI_K3_SHARDS:
+    raise ValueError("checkpoint index does not match the official 96-shard Kimi K3 release")
+  missing_files = {fn for fn in weight_map.values() if not (root / fn).is_file()}
+  if missing_files: raise FileNotFoundError(f"missing {len(missing_files)} checkpoint shards, first: {sorted(missing_files)[0]}")
+  model = Transformer(kimi_k3_config(max_context))
+  _shard_kimi_k3(model, tuple(f"{Device.DEFAULT}:{i}" for i in range(devices)))
+  consumed = _load_nonexperts(root, weight_map, model, progress)
+  # Expert staging graphs are acyclic and released by reference counting after each layer. Avoid
+  # unnecessary cyclic-collector scans across the complete persistent model graph during this loop.
+  gc_was_enabled = gc.isenabled()
+  gc.disable()
+  try: consumed.update(_load_experts(root, weight_map, model, progress))
+  finally:
+    if gc_was_enabled: gc.enable()
+  unused_language = {k for k in weight_map if k.startswith("language_model.")} - consumed
+  if unused_language: raise ValueError(f"unmapped language tensors: {sorted(unused_language)[:20]}")
+  return model
+
+__all__ = ["KIMI_K3_FULL_ATTN_LAYERS", "KIMI_K3_SSM_LAYERS", "KIMI_K3_TEXT_SIZE", "KIMI_K3_TP8_BYTES_PER_GPU",
+           "audit_kimi_k3_checkpoint", "kimi_k3_config", "kimi_k3_smoke_config", "load_kimi_k3", "load_kimi_tokenizer_data"]

tinygrad/llm/model.py

@@ -1,17 +1,18 @@
 from __future__ import annotations
-import enum, functools, itertools, pathlib
+import array, functools, itertools, pathlib
 from dataclasses import dataclass, replace
+from typing import Callable, cast
 from tinygrad import Tensor, nn, UOp, TinyJit, getenv, function, dtypes
+from tinygrad.device import MultiBuffer
 from tinygrad.nn import Linear
 from tinygrad.llm.gguf import gguf_load
+from tinygrad.llm.quant import dequantize_mxfp4, quantize_dequantize_mxfp8
+from tinygrad.llm.kernels import amd_custom_kernels_supported, amd_exact_bf16_custom_kernels_supported, amd_int32_item, \
+  amd_packed_mxfp4_supported, amd_wave64_custom_kernels_supported, bf16_matvec, bf16_mfma_splitk, bf16_partial_linear, \
+  dual_bf16_matvec, dual_input_bf16_matvec, \
+  gated_delta_prefill, kda_fgb_linear, kda_qkv_linear, mxfp4_expert_linear, mxfp8_quantize_dequantize
 from tinygrad.uop.ops import resolve
 
-class ExpertGating(enum.IntEnum):
-  SOFTMAX = 1
-  SIGMOID = 2
-  SOFTMAX_WEIGHT = 3  # softmax over the top-k selected logits
-  SQRT_SOFTPLUS = 4
-
 @functools.cache
 def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0, device:str|None=None) -> Tensor:
   freqs = 1.0 / (theta ** (Tensor.arange(0, dim, 2)[:(dim // 2)] / dim))
@@ -26,12 +27,37 @@ class ExpertWeights:
     # sel: (B, T, k), x: (B, T, 1, in) or (B, T, k, in) -> output: (B, T, k, out)
     return (x.unsqueeze(-2) @ self.weight[sel].transpose(-1, -2)).contiguous().squeeze(-2)
 
+class MXFP4ExpertWeights:
+  """Routed-expert weights stored as packed OCP MXFP4 with one E8M0 scale per 32 values."""
+  def __init__(self, num_experts:int, in_features:int, out_features:int):
+    if in_features % 32: raise ValueError(f"MXFP4 expert input size must be divisible by 32, got {in_features}")
+    self.in_features, self.out_features = in_features, out_features
+    self.weight = Tensor.zeros(num_experts, out_features, in_features//2, dtype=dtypes.uint8)
+    self.weight_scale = Tensor.full((num_experts, out_features, in_features//32), 127, dtype=dtypes.uint8)
+  def __call__(self, sel:Tensor, x:Tensor, quantized:bool=False, partial:bool=False) -> Tensor:
+    # Only selected weights are expanded, so packed storage remains resident during generation.
+    if isinstance(self.weight.device, tuple) and not isinstance(sel.device, tuple): sel = sel.shard(self.weight.device, axis=None)
+    if not quantized:
+      x = mxfp8_quantize_dequantize(x.cast(dtypes.bfloat16)) if amd_custom_kernels_supported(x.device) else \
+        quantize_dequantize_mxfp8(x.cast(dtypes.bfloat16))
+    # gfx11 has no native FP4 instructions, but decoding nibbles inside the dot product still avoids
+    # the much larger selected-expert BF16 temporary. Gate/up weights are output-sharded in TP.
+    if amd_packed_mxfp4_supported(self.weight.device):
+      return mxfp4_expert_linear(sel, x, self.weight, self.weight_scale, partial=partial)
+    weight = dequantize_mxfp4(self.weight[sel], self.weight_scale[sel], dtype=dtypes.bfloat16)
+    return (x.unsqueeze(-2) @ weight.transpose(-1, -2)).contiguous().squeeze(-2)
+
 def apply_rope(x:Tensor, freqs_cis:Tensor) -> Tensor:
   assert x.shape[-1] % 2 == 0
   cos, sin = freqs_cis.reshape(1, 1, x.shape[2], -1).chunk(2, dim=-1)
   x1, x2 = x.chunk(2, dim=-1)
   return (x1 * cos - x2 * sin).cat(x2 * cos + x1 * sin, dim=-1)
 
+def l2norm(x:Tensor, eps:float=1e-6) -> Tensor:
+  """FLA-compatible L2 normalization: FP32 reduction and epsilon inside the square root."""
+  dtype, x = x.dtype, x.float()
+  return (x * (x.square().sum(axis=-1, keepdim=True, dtype=dtypes.float32) + eps).rsqrt()).cast(dtype)
+
 def pairwise_topk(x: Tensor, k: int) -> tuple[Tensor, Tensor]:
   n = x.shape[-1]
   vals = Tensor.arange(n).reshape(1,1,n).cast(x.dtype).expand(x.shape)
@@ -40,6 +66,16 @@ def pairwise_topk(x: Tensor, k: int) -> tuple[Tensor, Tensor]:
   sel = x.const_like(0).scatter(-1, cmp.sum(axis=-1).cast('int32'), vals)[:,:,n-k:].cast('int32')
   return x.gather(-1, sel), sel
 
+def iterative_topk(x:Tensor, k:int) -> tuple[Tensor, Tensor]:
+  """O(k*N) top-k for very wide MoE routers, with stable first-index tie breaking."""
+  work, values, indices = x, [], []
+  for _ in range(k):
+    sel = work.argmax(-1, keepdim=True)
+    values.append(x.gather(-1, sel))
+    indices.append(sel)
+    work = work.scatter(-1, sel, x.dtype.min)
+  return values[0].cat(*values[1:], dim=-1), indices[0].cat(*indices[1:], dim=-1)
+
 @dataclass(frozen=True)
 class SSMConfig:
   conv_kernel: int
@@ -48,6 +84,7 @@ class SSMConfig:
   time_step_rank: int
   inner_size: int
   kda: bool = False
+  channel_decay: bool = False
 
 @dataclass(frozen=True)
 class TransformerConfig:
@@ -67,7 +104,6 @@ class TransformerConfig:
   num_experts: int = 0
   num_experts_per_tok: int = 0
   norm_topk_prob: bool = False
-  expert_gating_func: ExpertGating = ExpertGating.SOFTMAX
   q_lora_rank: int = 0
   kv_lora_rank: int = 0
   shared_expert_dim: int = 0
@@ -80,6 +116,20 @@ class TransformerConfig:
   routed_scaling_factor: float = 1.0
   qkv_bias: bool = False
   expert_bias: bool = False
+  expert_mxfp4: bool = False
+  bf16_activations: bool = False
+  kda_split_qkv: bool = False
+  # Kimi K3 extensions. Defaults preserve all existing model behavior.
+  activation_situ_beta: float = 0.0
+  activation_situ_linear_beta: float = 0.0
+  routed_expert_dim: int = 0
+  latent_moe_norm: bool = False
+  route_weights_uncorrected: bool = False
+  attn_res_block_size: int = 0
+  kda_full_rank_gate: bool = False
+  kda_gate_lower_bound: float = 0.0
+  recurrent_prefill_chunked: bool = False
+  recurrent_prefill_chunk_size: int = 0
 
 class FFNBlock:
   def __init__(self, config:TransformerConfig):
@@ -93,9 +143,15 @@ class FFNBlock:
     if config.num_experts > 0:
       self.ffn_gate_inp = Linear(config.dim, config.num_experts, bias=False)  # router
       if config.expert_bias: self.exp_probs_b = {"bias": Tensor.zeros(config.num_experts)}
-      self.ffn_gate_exps = ExpertWeights(config.num_experts, config.dim, config.hidden_dim)
-      self.ffn_up_exps = ExpertWeights(config.num_experts, config.dim, config.hidden_dim)
-      self.ffn_down_exps = ExpertWeights(config.num_experts, config.hidden_dim, config.dim)
+      expert_cls = MXFP4ExpertWeights if config.expert_mxfp4 else ExpertWeights
+      expert_dim = config.routed_expert_dim or config.dim
+      self.ffn_gate_exps = expert_cls(config.num_experts, expert_dim, config.hidden_dim)
+      self.ffn_up_exps = expert_cls(config.num_experts, expert_dim, config.hidden_dim)
+      self.ffn_down_exps = expert_cls(config.num_experts, config.hidden_dim, expert_dim)
+      if config.routed_expert_dim:
+        self.ffn_routed_down = Linear(config.dim, expert_dim, bias=False)
+        self.ffn_routed_up = Linear(expert_dim, config.dim, bias=False)
+        if config.latent_moe_norm: self.ffn_routed_norm = nn.RMSNorm(expert_dim, config.norm_eps)
       if config.shared_expert_dim > 0:
         self.ffn_gate_shexp = Linear(config.dim, config.shared_expert_dim, bias=False)
         self.ffn_up_shexp = Linear(config.dim, config.shared_expert_dim, bias=False)
@@ -106,43 +162,100 @@ class FFNBlock:
       self.ffn_up      = Linear(config.dim, config.hidden_dim, bias=False)
       self.ffn_down    = Linear(config.hidden_dim, config.dim, bias=False)
 
+    if config.attn_res_block_size:
+      self.attn_res_norm, self.mlp_res_norm = nn.RMSNorm(config.dim, config.norm_eps), nn.RMSNorm(config.dim, config.norm_eps)
+      self.attn_res_proj, self.mlp_res_proj = Linear(config.dim, 1, bias=False), Linear(config.dim, 1, bias=False)
+
+  def _activation(self, gate:Tensor, up:Tensor) -> Tensor:
+    if not self.config.activation_situ_beta: return gate.silu() * up
+    gate32, up32, beta = gate.float(), up.float(), self.config.activation_situ_beta
+    gate32 = beta * (gate32 / beta).tanh() * gate32.sigmoid()
+    if (linear_beta := self.config.activation_situ_linear_beta): up32 = linear_beta * (up32 / linear_beta).tanh()
+    return (gate32 * up32).cast(gate.dtype)
+
   def _feed_forward(self, x:Tensor) -> Tensor:
     if hasattr(self, 'ffn_gate_exps'):
       h = x.unsqueeze(2)  # (B, T, 1, D) - add expert dim for broadcasting
-      logits = self.ffn_gate_inp(x)
-      bias = self.exp_probs_b["bias"] if hasattr(self, 'exp_probs_b') else None
-      gating, normalize_topk = self.config.expert_gating_func, self.config.norm_topk_prob
-      # fast path: without selection bias, normalized SOFTMAX is equivalent to SOFTMAX_WEIGHT
-      if gating == ExpertGating.SOFTMAX and bias is None and normalize_topk:
-        gating, normalize_topk = ExpertGating.SOFTMAX_WEIGHT, False
-      if   gating == ExpertGating.SOFTMAX_WEIGHT: scores = logits
-      elif gating == ExpertGating.SOFTMAX:        scores = logits.softmax(-1)
-      elif gating == ExpertGating.SIGMOID:        scores = logits.sigmoid()
-      elif gating == ExpertGating.SQRT_SOFTPLUS:  scores = logits.softplus().sqrt()
-
-      _, sel = pairwise_topk(scores if bias is None else scores + bias, self.config.num_experts_per_tok)
-      probs = scores.gather(-1, sel)
-      # SOFTMAX_WEIGHT applies softmax after top-k selection
-      if gating == ExpertGating.SOFTMAX_WEIGHT: probs = probs.softmax(-1)
-      if normalize_topk: probs = probs / probs.sum(axis=-1, keepdim=True)
+      # Kimi computes router logits in FP32 even though the residual stream and weights are BF16.
+      logits = x.float().linear(self.ffn_gate_inp.weight.float().transpose()) if self.config.bf16_activations else self.ffn_gate_inp(x)
+      if hasattr(self, 'exp_probs_b'):
+        scores = logits.sigmoid()
+        adjusted_scores = scores + self.exp_probs_b["bias"]
+        topk = iterative_topk if self.config.num_experts >= 512 else pairwise_topk
+        _, sel = topk(adjusted_scores, self.config.num_experts_per_tok)
+        probs = (scores if self.config.route_weights_uncorrected else adjusted_scores).gather(-1, sel)
+        # Kimi-Linear-48B's older reference weights corrected scores. K3 selects with the correction
+        # but gathers the uncorrected sigmoid scores, so keep this an explicit compatibility switch.
+        if self.config.norm_topk_prob: probs = probs / (probs.sum(axis=-1, keepdim=True) + 1e-20)
+      else:
+        vals, sel = pairwise_topk(logits, self.config.num_experts_per_tok)
+        probs = vals.softmax(-1) if self.config.norm_topk_prob else logits.softmax(-1).gather(-1, sel)
       probs = probs * self.config.routed_scaling_factor
-      x_down = self.ffn_down_exps(sel, (self.ffn_gate_exps(sel, h).silu() * self.ffn_up_exps(sel, h)).contiguous())  # (B, T, k, D)
-      out = (x_down * probs.unsqueeze(-1)).sum(axis=2)  # (B, T, D)
+      if hasattr(self, 'ffn_routed_down'): h = self.ffn_routed_down(x).unsqueeze(2)
+      if isinstance(self.ffn_gate_exps, MXFP4ExpertWeights) and amd_packed_mxfp4_supported(h.device):
+        hq = mxfp8_quantize_dequantize(h.cast(dtypes.bfloat16)) if amd_custom_kernels_supported(h.device) else \
+          quantize_dequantize_mxfp8(h.cast(dtypes.bfloat16))
+        gate = self.ffn_gate_exps(sel, hq, quantized=True)
+        up = cast(MXFP4ExpertWeights, self.ffn_up_exps)(sel, hq, quantized=True)
+      else: gate, up = self.ffn_gate_exps(sel, h), self.ffn_up_exps(sel, h)
+      routed_activation = self._activation(gate, up).contiguous()
+      combine_down = resolve(x.shape[1] == 1) and isinstance(self.ffn_down_exps, MXFP4ExpertWeights) and \
+        hasattr(self, 'ffn_gate_shexp') and not hasattr(self, 'ffn_routed_up') and amd_custom_kernels_supported(x.device)
+      x_down = cast(MXFP4ExpertWeights, self.ffn_down_exps)(sel, routed_activation, partial=True) if combine_down else \
+        self.ffn_down_exps(sel, routed_activation)
+      out = (x_down * probs.unsqueeze(-1).unsqueeze(-1)).sum(axis=2) if combine_down else \
+        (x_down * probs.unsqueeze(-1)).sum(axis=2).cast(x_down.dtype)  # (B, T, D[, devices])
+      combine_final = resolve(x.shape[1] == 1) and hasattr(self, 'ffn_routed_up') and hasattr(self, 'ffn_gate_shexp') and \
+        not hasattr(self, 'ffn_gate_inp_shexp') and isinstance(self.ffn_routed_up.weight.device, tuple) and \
+        isinstance(self.ffn_down_shexp.weight.device, tuple) and self.ffn_routed_up.weight.uop.axis == self.ffn_down_shexp.weight.uop.axis == 1 and \
+        self.ffn_routed_up.weight.shape[1] % (32*len(self.ffn_routed_up.weight.device)) == 0 and \
+        self.ffn_down_shexp.weight.shape[1] % (32*len(self.ffn_down_shexp.weight.device)) == 0 and \
+        amd_exact_bf16_custom_kernels_supported(x.device)
+      if hasattr(self, 'ffn_routed_up'):
+        if hasattr(self, 'ffn_routed_norm'): out = self.ffn_routed_norm(out)
+        out = bf16_partial_linear(out, self.ffn_routed_up.weight) if combine_final else self.ffn_routed_up(out)
       if hasattr(self, 'ffn_gate_shexp'):
-        shexp = self.ffn_down_shexp(self.ffn_gate_shexp(x).silu().contiguous() * self.ffn_up_shexp(x))
-        if hasattr(self, 'ffn_gate_inp_shexp'): shexp = shexp * (x * self.ffn_gate_inp_shexp["weight"]).sum(axis=-1, keepdim=True).sigmoid()
-        out = out + shexp
+        if resolve(x.shape[1] == 1) and amd_exact_bf16_custom_kernels_supported(x.device) and \
+           self.ffn_gate_shexp.weight.shape == self.ffn_up_shexp.weight.shape:
+          shared_gate, shared_up = dual_bf16_matvec(x, self.ffn_gate_shexp.weight, self.ffn_up_shexp.weight,
+                                                    fast=amd_custom_kernels_supported(x.device))
+        else: shared_gate, shared_up = self.ffn_gate_shexp(x).contiguous(), self.ffn_up_shexp(x).contiguous()
+        shared_activation = self._activation(shared_gate, shared_up).contiguous()
+        if combine_down:
+          out = (out + bf16_partial_linear(shared_activation, self.ffn_down_shexp.weight)).sum(3).cast(dtypes.bfloat16)
+        elif combine_final:
+          out = (out + bf16_partial_linear(shared_activation, self.ffn_down_shexp.weight)).sum(3).cast(dtypes.bfloat16)
+        else:
+          shexp = self.ffn_down_shexp(shared_activation)
+          if hasattr(self, 'ffn_gate_inp_shexp'):
+            shexp = shexp * (x * self.ffn_gate_inp_shexp["weight"]).sum(axis=-1, keepdim=True).sigmoid()
+          out = out + shexp
       return out
     # TODO: remove the need for this contiguous
-    return self.ffn_down(self.ffn_gate(x).silu().contiguous() * self.ffn_up(x))
+    if resolve(x.shape[1] == 1) and amd_exact_bf16_custom_kernels_supported(x.device) and \
+       self.ffn_gate.weight.shape == self.ffn_up.weight.shape:
+      dense_gate, dense_up = dual_bf16_matvec(x, self.ffn_gate.weight, self.ffn_up.weight, fast=amd_custom_kernels_supported(x.device))
+    else: dense_gate, dense_up = self.ffn_gate(x).contiguous(), self.ffn_up(x).contiguous()
+    dense_activation = self._activation(dense_gate, dense_up).contiguous()
+    if resolve(x.shape[1] == 1) and isinstance(self.ffn_down.weight.device, tuple) and self.ffn_down.weight.uop.axis == 1 and \
+       self.ffn_down.weight.shape[1] % (32*len(self.ffn_down.weight.device)) == 0 and amd_exact_bf16_custom_kernels_supported(x.device):
+      return bf16_partial_linear(dense_activation, self.ffn_down.weight).sum(3).cast(dtypes.bfloat16)
+    return self.ffn_down(dense_activation)
 
   # given the token-prefix match, return how much cached state this block can still reuse
   def _reusable_prefix_len(self, prefix_len:int, cached_len:int) -> int: return prefix_len
+  # return writes that reset this block's state after a cache mismatch
+  def _state_reset_ops(self) -> list[Tensor]: return []
   def _init_state(self, x:Tensor): raise NotImplementedError
   def _attention(self, x:Tensor, start_pos:int|UOp) -> Tensor: raise NotImplementedError
 
   def __call__(self, x: Tensor, start_pos: int|UOp):
     self._init_state(x)
+    # Kimi's heterogeneous TP shards are captured by the outer TinyJit; per-block precompilation
+    # cannot represent their differently shaped local buffers as one implicit parameter bundle.
+    if self.config.bf16_activations:
+      h = x + self._attention(self.attn_norm(x), start_pos)
+      return (h + self._feed_forward(self.ffn_norm(h))).contiguous()
     # we pass in the weights implicitly so we unpack the GGUF on the fly
     @function(precompile=True, allow_implicit=True)
     def _run(x:Tensor, start_pos:int|UOp):
@@ -150,6 +263,32 @@ class FFNBlock:
       return (h + self._feed_forward(self.ffn_norm(h))).contiguous()
     return _run(x, start_pos)
 
+  @staticmethod
+  def _apply_attn_res(prefix_sum:Tensor, block_residual:Tensor, proj:Linear, norm:nn.RMSNorm) -> Tensor:
+    # Both inputs are flattened over B*T. Scoring is intentionally FP32, matching K3 eager inference.
+    v = block_residual.cat(prefix_sum.unsqueeze(1), dim=1)
+    vf = v.float()
+    k = vf * (vf.square().mean(axis=-1, keepdim=True) + norm.eps).rsqrt()
+    assert norm.weight is not None
+    scores = (k * (norm.weight.float() * proj.weight.squeeze(0).float())).sum(axis=-1)
+    return (scores.softmax(-1).unsqueeze(1) @ vf).squeeze(1).cast(v.dtype)
+
+  def attn_residual(self, x:Tensor, start_pos:int|UOp, block_residual:Tensor, layer_idx:int) -> tuple[Tensor, Tensor]:
+    self._init_state(x)
+    shape, prefix_sum = x.shape, x
+    prefix:Tensor|None = prefix_sum
+    if block_residual.shape[1]: x = self._apply_attn_res(x.reshape(-1, shape[-1]), block_residual,
+                                                        self.attn_res_proj, self.attn_res_norm).reshape(shape)
+    if layer_idx % self.config.attn_res_block_size == 0:
+      block_residual = block_residual.cat(prefix_sum.reshape(-1, shape[-1]).unsqueeze(1), dim=1)
+      prefix = None
+    attn = self._attention(self.attn_norm(x), start_pos)
+    prefix = attn if prefix is None else prefix + attn
+    x = self._apply_attn_res(prefix.reshape(-1, shape[-1]), block_residual,
+                             self.mlp_res_proj, self.mlp_res_norm).reshape(shape)
+    mlp = self._feed_forward(self.ffn_norm(x))
+    return (prefix + mlp).contiguous(), block_residual
+
 class TransformerBlock(FFNBlock):
   def __init__(self, config:TransformerConfig):
     super().__init__(config)
@@ -191,7 +330,9 @@ class TransformerBlock(FFNBlock):
 
     # NOTE: this mask is causal_lower_right, not the causal_upper_left generated by is_casual = True
     # TODO: this if statement should be removed and it shouldn't generate extra kernels
-    mask = Tensor.full((1, 1, T, start_pos+T), float("-inf"), dtype=x.dtype, buffer=False).triu(start_pos+1) \
+    # Build the static T×T causal corner on-device, then prepend the unmasked cached prefix.
+    # A broadcast const with symbolic width otherwise defaults to CPU in multi-device graphs.
+    mask = Tensor.full((1, 1, T, T), float("-inf"), dtype=x.dtype, device=x.device).triu(1).pad(((0, 0),)*3+((start_pos, 0),)) \
       if resolve(T != 1) else None
     attn = q.scaled_dot_product_attention(k, v, attn_mask=mask, enable_gqa=True)     # (B,H,T,Hd)
     attn = attn.transpose(1, 2).reshape(B, T, -1)                                    # back to (B,T,D)
@@ -199,8 +340,9 @@ class TransformerBlock(FFNBlock):
 
   def _init_state(self, x:Tensor):
     if not hasattr(self, "cache_kv"):
+      # TODO: how is the dtype of this determined?
       self.cache_kv = Tensor.empty(2, x.shape[0], self.config.n_kv_heads, self.config.max_context, self.config.head_dim,
-                                   dtype=dtypes.default_float, device=x.device)
+                                   device=x.device, dtype=x.dtype)
       self.freqs_cis = precompute_freqs_cis(self.config.rope_dim, self.config.max_context, self.config.rope_theta, device=x.device)
 
 class MLATransformerBlock(FFNBlock):
@@ -218,17 +360,23 @@ class MLATransformerBlock(FFNBlock):
     self.attn_k_b = {"weight": Tensor.zeros(config.n_heads, config.kv_lora_rank, qk_nope_head_dim)}
     self.attn_v_b = {"weight": Tensor.zeros(config.n_heads, config.v_head_dim, config.kv_lora_rank)}
     self.attn_output = Linear(config.n_heads * config.v_head_dim, config.dim, bias=False)
+    if config.attn_output_gate: self.attn_gate = Linear(config.dim, config.n_heads * config.v_head_dim, bias=False)
 
   def _attention(self, x:Tensor, start_pos:int|UOp) -> Tensor:
     B, T, _ = x.shape
     q_nope_head_dim = self.config.head_dim - self.config.rope_dim
-    q_proj = self.attn_q_b(self.attn_q_a_norm(self.attn_q_a(x))) if self.config.q_lora_rank > 0 else self.attn_q(x)
+    mfma_decode = resolve(T == 1) and x.shape[-1] % 256 == 0 and amd_wave64_custom_kernels_supported(x.device)
+    q_a_mfma = self.config.q_lora_rank > 0 and mfma_decode and self.attn_q_a.weight.shape[0] % 16 == 0
+    q_a = bf16_mfma_splitk(x, self.attn_q_a.weight) if q_a_mfma else \
+      self.attn_q_a(x) if self.config.q_lora_rank > 0 else None
+    q_proj = self.attn_q_b(self.attn_q_a_norm(q_a)) if q_a is not None else self.attn_q(x)
     q = q_proj.reshape(B, T, self.config.n_heads, self.config.head_dim).transpose(1, 2)
     q_nope, q_rope = q[..., :q_nope_head_dim], q[..., q_nope_head_dim:]
     if not self.config.ssm or not self.config.ssm.kda: q_rope = apply_rope(q_rope, self.freqs_cis[start_pos:start_pos+T])
     q = (q_nope @ self.attn_k_b["weight"].transpose(-1, -2)).cat(q_rope, dim=-1)
 
-    kv_a = self.attn_kv_a_mqa(x)
+    kv_a_mfma = mfma_decode and self.attn_kv_a_mqa.weight.shape[0] % 16 == 0
+    kv_a = bf16_mfma_splitk(x, self.attn_kv_a_mqa.weight) if kv_a_mfma else self.attn_kv_a_mqa(x)
     c_kv = self.attn_kv_a_norm(kv_a[..., :self.config.kv_lora_rank])
     k_rope = kv_a[..., self.config.kv_lora_rank:].reshape(B, T, 1, self.config.rope_dim).transpose(1, 2)
     if not self.config.ssm or not self.config.ssm.kda: k_rope = apply_rope(k_rope, self.freqs_cis[start_pos:start_pos+T])
@@ -237,17 +385,23 @@ class MLATransformerBlock(FFNBlock):
     k = Tensor(self.cache_k.uop.after(self.cache_k[:, :, start_pos:start_pos+T, :].uop.store(k_store.uop)))[:, :, 0:start_pos+T, :]
     v = k[..., :self.config.kv_lora_rank]
 
-    mask = Tensor.full((1, 1, T, start_pos+T), float("-inf"), dtype=x.dtype, buffer=False).triu(start_pos+1) \
+    mask = Tensor.full((1, 1, T, T), float("-inf"), dtype=x.dtype, device=x.device).triu(1).pad(((0, 0),)*3+((start_pos, 0),)) \
       if resolve(T != 1) else None
     attn = q @ k.transpose(-1, -2) * (1.0 / self.config.head_dim ** 0.5)
     if mask is not None: attn = attn + mask
-    attn = attn.softmax(-1)
+    # Match eager Kimi MLA: normalize attention scores in FP32, then return to the query dtype.
+    attn = attn.softmax(-1, dtype=dtypes.float32).cast(q.dtype)
     attn = ((attn @ v) @ self.attn_v_b["weight"].transpose(-1, -2)).transpose(1, 2).reshape(B, T, -1)
+    if hasattr(self, "attn_gate"): attn = attn * self.attn_gate(x).sigmoid()
+    if resolve(T == 1) and isinstance(self.attn_output.weight.device, tuple) and \
+       self.attn_output.weight.shape[1] % (32*len(self.attn_output.weight.device)) == 0 and amd_exact_bf16_custom_kernels_supported(attn.device):
+      return bf16_partial_linear(attn, self.attn_output.weight).sum(3).cast(dtypes.bfloat16)
     return self.attn_output(attn)
 
   def _init_state(self, x:Tensor):
     if not hasattr(self, "cache_k"):
-      self.cache_k = Tensor.empty(x.shape[0], 1, self.config.max_context, self.config.kv_lora_rank + self.config.rope_dim, device=x.device)
+      self.cache_k = Tensor.empty(x.shape[0], 1, self.config.max_context, self.config.kv_lora_rank + self.config.rope_dim,
+                                  device=x.device, dtype=x.dtype)
       self.freqs_cis = precompute_freqs_cis(self.config.rope_dim, self.config.max_context, self.config.rope_theta, device=x.device)
 
 class GatedDeltaNetBlock(FFNBlock):
@@ -257,88 +411,144 @@ class GatedDeltaNetBlock(FFNBlock):
     assert self.num_v_heads % self.num_k_heads == 0
     self.head_v_dim, self.ssm_conv_kernel = ssm.inner_size // ssm.time_step_rank, ssm.conv_kernel
     self.conv_channels, self.q_dim = ssm.inner_size + 2*ssm.group_count*ssm.state_size, ssm.state_size*ssm.group_count
-    self.attn_qkv = Linear(config.dim, self.conv_channels, bias=False)
+    if ssm.kda and config.kda_split_qkv:
+      self.attn_q, self.attn_k = Linear(config.dim, self.q_dim, bias=False), Linear(config.dim, self.q_dim, bias=False)
+      self.attn_v = Linear(config.dim, ssm.inner_size, bias=False)
+      self.ssm_q_conv1d = {"weight": Tensor.zeros(self.q_dim, self.ssm_conv_kernel)}
+      self.ssm_k_conv1d = {"weight": Tensor.zeros(self.q_dim, self.ssm_conv_kernel)}
+      self.ssm_v_conv1d = {"weight": Tensor.zeros(ssm.inner_size, self.ssm_conv_kernel)}
+    else:
+      self.attn_qkv = Linear(config.dim, self.conv_channels, bias=False)
+      self.ssm_conv1d = {"weight": Tensor.zeros(self.conv_channels, self.ssm_conv_kernel)}
     if ssm.kda:
-      self.ssm_g_a, self.ssm_g_b = Linear(config.dim, self.head_v_dim, bias=False), Linear(self.head_v_dim, ssm.inner_size, bias=False)
+      if config.kda_full_rank_gate: self.ssm_g_full = Linear(config.dim, ssm.inner_size, bias=False)
+      else: self.ssm_g_a, self.ssm_g_b = Linear(config.dim, self.head_v_dim, bias=False), Linear(self.head_v_dim, ssm.inner_size, bias=False)
       self.ssm_f_a, self.ssm_f_b = Linear(config.dim, self.head_k_dim, bias=False), Linear(self.head_k_dim, ssm.inner_size, bias=False)
     else:
       self.attn_gate = Linear(config.dim, ssm.inner_size, bias=False)
       self.ssm_alpha = Linear(config.dim, self.num_v_heads, bias=False)
     self.ssm_beta = Linear(config.dim, self.num_v_heads, bias=False)
-    self.ssm_conv1d = {"weight": Tensor.zeros(self.conv_channels, self.ssm_conv_kernel)}
     self.ssm_dt = {"bias": Tensor.zeros(ssm.inner_size if ssm.kda else self.num_v_heads)}
-    self.ssm_a = Tensor.zeros(self.num_v_heads, 1) if ssm.kda else Tensor.zeros(self.num_v_heads)
+    self.ssm_a = Tensor.zeros(self.head_v_dim if ssm.channel_decay else self.num_v_heads, 1) if ssm.kda else Tensor.zeros(self.num_v_heads)
+    self.kda_channel_decay = ssm.channel_decay
     self.ssm_norm, self.ssm_out = nn.RMSNorm(self.head_v_dim, config.norm_eps), Linear(ssm.inner_size, config.dim, bias=False)
 
   def _attention(self, x:Tensor, start_pos:int|UOp) -> Tensor:
     B, T, _ = x.shape
-    # bind ints to a variable so the reset flag stays a runtime value (it toggles when generation restarts at position 0)
-    start_pos = start_pos if isinstance(start_pos, UOp) else UOp.variable("start_pos", 0, self.config.max_context-1).bind(start_pos)
-    initial = Tensor(start_pos).eq(0)
-    is_kda = hasattr(self, "ssm_g_a")
-    symbolic = isinstance(T, UOp)
-    T_pad = x.max_shape[1]  # symbolic chunks are padded to their max size: one graph serves every size
 
     # input processing
-    x = x.half()
-    out_gate = self.ssm_g_b(self.ssm_g_a(x)) if is_kda else self.attn_gate(x)
-    out_gate = out_gate.reshape(B, T, self.num_v_heads, self.head_v_dim)
-    beta = self.ssm_beta(x).sigmoid().reshape(B, T, self.num_v_heads)
-    alpha = self.ssm_f_b(self.ssm_f_a(x)) if is_kda else self.ssm_alpha(x)
-    log_alpha = ((alpha.float() + self.ssm_dt["bias"]).softplus().reshape(B, T, self.num_v_heads, -1) *
-                 self.ssm_a.reshape(self.num_v_heads, -1))
+    # Kimi-Linear is a BF16 model. Qwen 3.5 GGDN checkpoints historically use FP16 here.
+    x = x.cast(dtypes.bfloat16) if self.config.ssm and self.config.ssm.kda else x.half()
+    fused_fg = hasattr(self, "ssm_g_a") and resolve(T == 1) and amd_custom_kernels_supported(x.device) and \
+      self.ssm_g_a.weight.shape == self.ssm_f_a.weight.shape
+    mfma_decode = resolve(T == 1) and x.shape[-1] % 256 == 0 and amd_wave64_custom_kernels_supported(x.device)
+    if hasattr(self, "ssm_g_full"): out_gate = self.ssm_g_full(x)
+    elif hasattr(self, "ssm_g_a"):
+      if fused_fg:
+        gate_a, alpha_a, beta_logits = kda_fgb_linear(x, self.ssm_g_a.weight, self.ssm_f_a.weight, self.ssm_beta.weight)
+        out_gate, alpha_logits = dual_input_bf16_matvec(gate_a, alpha_a, self.ssm_g_b.weight, self.ssm_f_b.weight)
+      else: out_gate = self.ssm_g_b(self.ssm_g_a(x))
+    else: out_gate = self.attn_gate(x)
+    if not fused_fg: beta_logits = self.ssm_beta(x)
+    if not fused_fg:
+      if hasattr(self, "ssm_f_a"):
+        f_a = bf16_mfma_splitk(x, self.ssm_f_a.weight) if mfma_decode and self.ssm_f_a.weight.shape[0] % 16 == 0 else self.ssm_f_a(x)
+        alpha_logits = self.ssm_f_b(f_a)
+      else: alpha_logits = self.ssm_alpha(x)
 
-    # qkv conv, conv_state is reset when starting from position 0
-    conv_state = initial.where(0, self.conv_state)
-    # assemble the conv window in a static-size buffer: [conv_state | qkv rows | zero-pad].
-    # padded steps are exact no-ops: beta=0 (delta rule off), log_alpha=0 (decay 1 after exp)
-    win = Tensor.zeros(B, self.ssm_conv_kernel-1 + T_pad, self.conv_channels).uop
-    win = win.after(win[:, :self.ssm_conv_kernel-1].store(conv_state.cast(win.dtype).uop))
-    win = win.after(win[:, self.ssm_conv_kernel-1:self.ssm_conv_kernel-1+T].store(self.attn_qkv(x).cast(win.dtype).uop))
-    conv_window = Tensor(win)
-    # the last conv_kernel-1 columns of the window become the next conv state
-    conv_state_store = self.conv_state.uop.store(conv_window[:, T:T+self.ssm_conv_kernel-1].cast(self.conv_state.dtype).uop)
+    # Causal depthwise Q/K/V convolution. All tokens are projected together, then the recurrent
+    # update is fused into one kernel so prefill doesn't build a Python-unrolled graph.
+    split_qkv = hasattr(self, "attn_q")
+    if split_qkv:
+      if resolve(T == 1) and amd_packed_mxfp4_supported(x.device) and \
+         self.attn_q.weight.shape == self.attn_k.weight.shape == self.attn_v.weight.shape:
+        projected_q, projected_k, projected_v = kda_qkv_linear(x, self.attn_q.weight, self.attn_k.weight, self.attn_v.weight)
+      else: projected_q, projected_k, projected_v = self.attn_q(x), self.attn_k(x), self.attn_v(x)
+      # Snapshot mutable caches before constructing the recurrence. Otherwise the final store can
+      # overwrite their buffers before earlier outputs in a multi-token lazy graph consume them.
+      conv_state_q, conv_state_k, conv_state_v = self.conv_state_q.clone(), self.conv_state_k.clone(), self.conv_state_v.clone()
+    else: projected, conv_state = self.attn_qkv(x), self.conv_state
+    def causal_conv(projected:Tensor, state:Tensor, weight:Tensor) -> tuple[Tensor, Tensor]:
+      window = state.cat(projected, dim=1)
+      out = functools.reduce(lambda a,b: a+b, (window[:, i:i+T] * weight[:, i] for i in range(self.ssm_conv_kernel))).silu()
+      return out, window[:, T:T+self.ssm_conv_kernel-1]
+    if split_qkv:
+      q, conv_state_q = causal_conv(projected_q, conv_state_q, self.ssm_q_conv1d["weight"])
+      k, conv_state_k = causal_conv(projected_k, conv_state_k, self.ssm_k_conv1d["weight"])
+      v, conv_state_v = causal_conv(projected_v, conv_state_v, self.ssm_v_conv1d["weight"])
+    else:
+      conv_out, conv_state = causal_conv(projected, conv_state, self.ssm_conv1d["weight"])
+      q, k, v = conv_out.split([self.q_dim, self.q_dim, self.conv_channels - 2*self.q_dim], dim=-1)
 
-    conv_out = functools.reduce(lambda a,b: a+b,
-      (conv_window[:, i:i+T_pad] * self.ssm_conv1d["weight"][:, i] for i in range(self.ssm_conv_kernel))).silu()
-    if symbolic:
-      out_gate = out_gate.pad_to((B, T_pad, self.num_v_heads, self.head_v_dim))
-      beta, log_alpha = beta.pad_to((B, T_pad, self.num_v_heads)), log_alpha.pad_to((B, T_pad, *log_alpha.shape[2:]))
-    q, k, v = conv_out.split([self.q_dim, self.q_dim, self.conv_channels - 2*self.q_dim], dim=-1)
-    qk_eps = 1e-12 if is_kda else 1e-6
-    q, k = (z.reshape(B, T_pad, self.num_k_heads, self.head_k_dim).normalize(dim=-1, eps=qk_eps)
-            .repeat(1, 1, self.num_v_heads//self.num_k_heads, 1) for z in (q, k))
-    v = v.reshape(B, T_pad, self.num_v_heads, self.head_v_dim)
-    # layout the per-step operands to broadcast against the (B, H, V, K) state
-    q, k, v, beta = (z.transpose(1, 2).float() for z in (q, k, v, beta))
-    q, k, v, beta = q.unsqueeze(-2) * self.head_k_dim**-0.5, k.unsqueeze(-2), v.unsqueeze(-1), beta.unsqueeze(-1).unsqueeze(-1)
-    alpha = log_alpha.transpose(1, 2).exp().unsqueeze(-1)  # per-channel decay for kda, per-head otherwise (B, H, T, V|1, 1)
+    q, k = q.reshape(B, T, self.num_k_heads, self.head_k_dim), k.reshape(B, T, self.num_k_heads, self.head_k_dim)
+    q, k = (l2norm(q), l2norm(k)) if self.config.ssm and self.config.ssm.kda else (q.normalize(dim=-1), k.normalize(dim=-1))
+    q = q.repeat(1, 1, self.num_v_heads//self.num_k_heads, 1).transpose(1, 2).float() * self.head_k_dim**-0.5
+    k = k.repeat(1, 1, self.num_v_heads//self.num_k_heads, 1).transpose(1, 2).float()
+    v = v.reshape(B, T, self.num_v_heads, self.head_v_dim).transpose(1, 2).float()
+    beta = (beta_logits.float() if self.config.ssm and self.config.ssm.kda else beta_logits).sigmoid().transpose(1, 2)
+    gate_logits = (alpha_logits.float() + self.ssm_dt["bias"]).reshape(B, T, self.num_v_heads, -1)
+    a_shape = (1, 1, 1, self.head_v_dim) if self.kda_channel_decay else (1, 1, self.num_v_heads, 1)
+    if self.config.kda_gate_lower_bound:
+      log_alpha = self.config.kda_gate_lower_bound * ((-self.ssm_a).reshape(a_shape) * gate_logits).sigmoid()
+    else: log_alpha = gate_logits.softplus() * self.ssm_a.reshape(a_shape)
+    alpha = log_alpha.squeeze(-1).transpose(1, 2).exp() if log_alpha.shape[-1] == 1 else log_alpha.permute(0, 2, 1, 3).exp()
+    if T == 1:
+      # Keep decode on the small elementwise graph. The fused prefill kernel writes a temporary
+      # recurrent matrix, which is worthwhile for multiple tokens but needlessly copies state at T=1.
+      decay = alpha if len(alpha.shape) == 4 else alpha.unsqueeze(-1)
+      recurrent_state = self.recurrent_state * decay
+      k1, q1 = k[:, :, 0].unsqueeze(-1), q[:, :, 0].unsqueeze(-1)
+      recurrent_state = recurrent_state + ((v[:, :, 0].unsqueeze(-1) - recurrent_state@k1) * beta[:, :, 0].reshape(B, self.num_v_heads, 1, 1)) @ \
+                        k1.transpose(-1, -2)
+      core = (recurrent_state @ q1).squeeze(-1).unsqueeze(2)
+    else: core, recurrent_state = gated_delta_prefill(q, k, v, beta, alpha, self.recurrent_state)
+    core = core.transpose(1, 2)
 
-    # recurrent: scan over the (padded) tokens, updating the recurrent state. collect the per-step outputs
-    state = Tensor(self.recurrent_state.uop.after(conv_state_store)).float()  # carry the conv write into this graph
-    state = initial.where(0, state)
-    outs = []
-    for t in range(T_pad):
-      s1 = state * alpha[:, :, t]  # decay the state
-      delta = (v[:, :, t] - (s1*k[:, :, t]).sum(-1, keepdim=True)) * beta[:, :, t]  # the delta rule update
-      state = s1 + delta * k[:, :, t]
-      outs.append((state * q[:, :, t]).sum(-1))
+    # Store each cache with its own AFTER. Multi-device lowering handles one sharded STORE per
+    # AFTER; grouping these effects under one cache silently drops stores on the other shards.
+    state_updates:list[Tensor]
+    if split_qkv:
+      state_updates = [self.conv_state_q.assign(conv_state_q.cast(self.conv_state_q.dtype)),
+                       self.conv_state_k.assign(conv_state_k.cast(self.conv_state_k.dtype)),
+                       self.conv_state_v.assign(conv_state_v.cast(self.conv_state_v.dtype))]
+    else: state_updates = [self.conv_state.assign(conv_state.cast(self.conv_state.dtype))]
+    state_updates.append(self.recurrent_state.assign(recurrent_state.cast(self.recurrent_state.dtype)))
+    core_attn_out = self.ssm_norm(core.cast(x.dtype) if self.config.ssm and self.config.ssm.kda else core)
+    gate = out_gate.reshape(B, T, self.num_v_heads, self.head_v_dim)
+    gate = gate.float().sigmoid().cast(core_attn_out.dtype) if hasattr(self, "ssm_g_a") else gate.silu()
+    out = (core_attn_out * gate).reshape(B, T, -1)
+    out = out.cast(x.dtype)
+    ret = bf16_partial_linear(out, self.ssm_out.weight).sum(3).cast(dtypes.bfloat16) if resolve(T == 1) and \
+      isinstance(self.ssm_out.weight.device, tuple) and self.ssm_out.weight.shape[1] % (32*len(self.ssm_out.weight.device)) == 0 and \
+      amd_exact_bf16_custom_kernels_supported(out.device) else self.ssm_out(out)
+    return ret.realize(*state_updates)
 
-    # store the updated recurrent state in place, then read the stacked outputs after the write
-    core = Tensor(outs[0].stack(*outs[1:], dim=1).contiguous().uop.after(self.recurrent_state.uop.store(state.cast(self.recurrent_state.dtype).uop)))
-
-    # output; undo the padding before the output projection
-    z = (self.ssm_norm(core) * (out_gate.sigmoid() if is_kda else out_gate.silu())).cast(x.dtype).contiguous()
-    if symbolic: z = z[:, :T]
-    return self.ssm_out(z.reshape(B, T, -1))
+  # Recurrent state can be reused only when the new prompt exactly extends all currently valid state.
+  def _state_tensors(self) -> tuple[Tensor, ...]:
+    if hasattr(self, "conv_state_q"):
+      return self.conv_state_q, self.conv_state_k, self.conv_state_v, self.recurrent_state
+    return (self.conv_state, self.recurrent_state) if hasattr(self, "conv_state") else ()
+  def _state_reset_ops(self): return [s.assign(s.const_like(0)) for s in self._state_tensors()]
+  def _reusable_prefix_len(self, prefix_len:int, cached_len:int) -> int: return prefix_len if prefix_len == cached_len else 0
 
   def _init_state(self, x):
-    if not hasattr(self, "conv_state"):
-      self.conv_state = Tensor.zeros(x.shape[0], self.ssm_conv_kernel-1, self.conv_channels, device=x.device).clone()
-      self.recurrent_state = Tensor.zeros(x.shape[0], self.num_v_heads, self.head_v_dim, self.head_k_dim, device=x.device).clone()
+    if not hasattr(self, "conv_state") and not hasattr(self, "conv_state_q"):
+      if hasattr(self, "attn_q"):
+        device = x.device[0] if isinstance(x.device, tuple) else x.device
+        self.conv_state_q = Tensor.zeros(x.shape[0], self.ssm_conv_kernel-1, self.q_dim, device=device, dtype=x.dtype).clone()
+        self.conv_state_k = Tensor.zeros(x.shape[0], self.ssm_conv_kernel-1, self.q_dim, device=device, dtype=x.dtype).clone()
+        self.conv_state_v = Tensor.zeros(x.shape[0], self.ssm_conv_kernel-1, self.num_v_heads*self.head_v_dim, device=device, dtype=x.dtype).clone()
+        self.recurrent_state = Tensor.zeros(x.shape[0], self.num_v_heads, self.head_v_dim, self.head_k_dim, device=device).clone()
+        if isinstance(x.device, tuple):
+          for state in (self.conv_state_q, self.conv_state_k, self.conv_state_v): state.shard_(x.device, axis=2)
+          self.recurrent_state.shard_(x.device, axis=1)
+      else:
+        self.conv_state = Tensor.zeros(x.shape[0], self.ssm_conv_kernel-1, self.conv_channels, device=x.device, dtype=x.dtype).clone()
+        self.recurrent_state = Tensor.zeros(x.shape[0], self.num_v_heads, self.head_v_dim, self.head_k_dim, device=x.device).clone()
 
 class Transformer:
   def __init__(self, config:TransformerConfig):
+    self.config = config
     dense_config = replace(config, num_experts=0, num_experts_per_tok=0, shared_expert_dim=0, hidden_dim=config.dense_hidden_dim or config.hidden_dim)
     if config.ssm: config = replace(config, qk_norm=config.head_dim)
     block_cls = MLATransformerBlock if config.kv_lora_rank > 0 else TransformerBlock
@@ -348,23 +558,76 @@ class Transformer:
     self.token_embd  = nn.Embedding(config.vocab_size, config.dim)
     self.output_norm = nn.RMSNorm(config.dim, config.norm_eps)
     self.output = Linear(config.dim, config.vocab_size, bias=False)
+    if config.attn_res_block_size:
+      self.output_attn_res_norm = nn.RMSNorm(config.dim, config.norm_eps)
+      self.output_attn_res_proj = Linear(config.dim, 1, bias=False)
     self.max_context = config.max_context
     self.has_recurrent_block = any(isinstance(b, GatedDeltaNetBlock) for b in self.blk)
     self._cached_tokens: list[int] = []
+    self._snapshot_tokens: list[int] = []
+    self._state_snapshots:list[Tensor] = []
+    self._token_buffer:Tensor|None = None
+    self._temperature_buffer:Tensor|None = None
     # we specialize the JIT for prefill and rollout
     self.prefill_jit = TinyJit(self.forward)
     self.rollout_jit = TinyJit(self.forward)
+    self.greedy_prefill_jit = TinyJit(self.forward)
+    self.greedy_rollout_jit = TinyJit(self.forward)
+    self.recurrent_prefill_jits:dict[int, Callable[..., Tensor]] = {}
+    self.recurrent_greedy_prefill_jits:dict[int, Callable[..., Tensor]] = {}
+    self.reset_jit = TinyJit(self._reset_state)
+    self.save_state_jit = TinyJit(self._save_state)
+    self.restore_state_jit = TinyJit(self._restore_state)
 
-  def forward(self, tokens:Tensor, start_pos:int|UOp, temperature:Tensor) -> Tensor:
-    x = self.token_embd(tokens).float()                   # (B, T, D)
-    for block in self.blk: x = block(x, start_pos)
-    # only run the output projection on the last token
-    logits = self.output(self.output_norm(x[:, -1:]))[:, -1, :]
+  def _reset_state(self) -> None:
+    if resets := [r for b in self.blk for r in b._state_reset_ops()]: Tensor.realize(*resets)
+
+  def _state_tensors(self) -> list[Tensor]:
+    return [s for block in self.blk if isinstance(block, GatedDeltaNetBlock) for s in block._state_tensors()]
+
+  def _init_state_snapshots(self) -> None:
+    if not self._state_snapshots: self._state_snapshots = [s.clone().realize() for s in self._state_tensors()]
+
+  def _save_state(self) -> None:
+    if writes := [dst.assign(src) for dst,src in zip(self._state_snapshots, self._state_tensors())]: Tensor.realize(*writes)
+
+  def _restore_state(self) -> None:
+    if writes := [dst.assign(src) for dst,src in zip(self._state_tensors(), self._state_snapshots)]: Tensor.realize(*writes)
+
+  def forward(self, tokens:Tensor, start_pos:int|UOp, temperature:Tensor|None) -> Tensor:
+    if len(tokens.shape) == 1: tokens = tokens.reshape(1, -1)
+    x = self.token_embd(tokens).cast(dtypes.bfloat16) if self.config.bf16_activations else self.token_embd(tokens).float()
+    block_residual = Tensor.zeros(x.shape[0]*x.shape[1], 0, x.shape[2], device=x.device, dtype=x.dtype) \
+      if self.config.attn_res_block_size else None
+    for i, block in enumerate(self.blk):
+      if block_residual is not None: x, block_residual = block.attn_residual(x, start_pos, block_residual, i)
+      else: x = block(x, start_pos)
+      # Tensor indexing lowers selected experts through a fused one-hot reduction. Keeping all 26
+      # of those high-level graphs alive until the final output is scheduled exhausts host memory.
+      # A realization boundary lowers one block at a time; TinyJit still captures and memory-plans
+      # the resulting schedules for rollout replay.
+      if self.config.expert_mxfp4: x.realize()
+    if block_residual is not None:
+      x = FFNBlock._apply_attn_res(x.reshape(-1, x.shape[-1]), block_residual,
+                                   self.output_attn_res_proj, self.output_attn_res_norm).reshape(x.shape)
+    final_x = self.output_norm(x)
+    if temperature is None and resolve(tokens.numel() == 1) and amd_exact_bf16_custom_kernels_supported(x.device):
+      return bf16_matvec(final_x, self.output.weight).argmax(-1, keepdim=True)
+    logits = self.output(final_x)[:, -1, :]
+    if temperature is None: return logits.argmax(-1, keepdim=True)
     # Gumbel-max trick: argmax(logits/temp - log(-log(uniform))) is equivalent to sampling from softmax(logits/temp)
     return (logits / temperature.maximum(1e-12) - (Tensor.rand_like(logits).maximum(1e-12).log().neg()).log()).argmax(-1, keepdim=True)
 
-  def __call__(self, tokens:Tensor, start_pos:int|UOp, temperature:Tensor) -> Tensor:
-    return (self.prefill_jit if resolve(tokens.shape[1] != 1) else self.rollout_jit)(tokens.contiguous(), start_pos, temperature)
+  def __call__(self, tokens:Tensor, start_pos:int|UOp, temperature:Tensor|None) -> Tensor:
+    token_count = tokens.numel()
+    if self.has_recurrent_block and resolve(token_count != 1):
+      assert isinstance(token_count, int)
+      cache = self.recurrent_greedy_prefill_jits if temperature is None else self.recurrent_prefill_jits
+      jit = cache.setdefault(token_count, TinyJit(self.forward))
+      return jit(tokens.flatten().contiguous(), start_pos, temperature)
+    if temperature is None:
+      return (self.greedy_prefill_jit if resolve(token_count != 1) else self.greedy_rollout_jit)(tokens.flatten().contiguous(), start_pos, None)
+    return (self.prefill_jit if resolve(token_count != 1) else self.rollout_jit)(tokens.flatten().contiguous(), start_pos, temperature)
 
   @staticmethod
   def from_gguf(gguf:Tensor|str|pathlib.Path, max_context:int|None=None,
@@ -430,7 +693,6 @@ class Transformer:
       qk_norm=int(state_dict['blk.0.attn_q_norm.weight'].shape[0]) if 'blk.0.attn_q_norm.weight' in state_dict else 0,
       num_experts=kv.get(f'{arch}.expert_count', 0), num_experts_per_tok=kv.get(f'{arch}.expert_used_count', 0),
       norm_topk_prob=kv.get(f'{arch}.expert_weights_norm', arch in ('qwen3moe', 'qwen35moe', 'kimi-linear')),
-      expert_gating_func=ExpertGating(kv.get(f'{arch}.expert_gating_func', ExpertGating.SOFTMAX)),
       kv_lora_rank=kv_lora_rank, q_lora_rank=kv.get(f'{arch}.attention.q_lora_rank', 0),
       leading_dense_blocks=kv.get(f'{arch}.leading_dense_block_count', 0),
       shared_expert_dim=kv.get(
@@ -451,34 +713,91 @@ class Transformer:
     return model, kv
 
   def warmup(self):
-    for _ in range(2): list(zip(range(2), self.generate([0])))
-
-  def get_start_pos(self, tokens:list[int]) -> int:
-    # recurrent state can't be partially reused after divergence: reuse it only when tokens extend the cached prefix
+    # Capture the only two shapes used by recurrent serving: a full prefill chunk and one-token rollout.
+    # Two chunks exercise both the initial and nonzero-position prefill paths before the server opens.
+    recurrent_chunk = self.config.recurrent_prefill_chunk_size or 32
+    prompt = [0] * max(1, min(recurrent_chunk*2, self.max_context-2)) if self.has_recurrent_block else [0]
+    # Recurrent serving captures both greedy and sampled graphs, then executes one replay so graph
+    # creation/lowering cannot leak into request latency for either HTTP temperature path.
+    # generate mutates its token list, so each pass needs a fresh prompt to exercise cache reset.
+    for temperature in ((0.0, 1.0) if self.has_recurrent_block else (0.0,)):
+      for _ in range(3 if self.has_recurrent_block else 2): list(zip(range(2), self.generate(prompt.copy(), temperature=temperature)))
+    # Capture prompt-boundary restore using successively extended prompts so every restore starts
+    # from the checkpoint made by the previous pass.
     if self.has_recurrent_block:
-      return len(self._cached_tokens) if self._cached_tokens and len(self._cached_tokens) < len(tokens) \
-        and tokens[:len(self._cached_tokens)] == self._cached_tokens else 0
+      for i in range(1, 4): list(zip(range(2), self.generate(prompt + list(range(1, i+1)), temperature=0.0)))
+
+  def _cache_start(self, tokens:list[int]) -> tuple[int, bool]:
     prefix_len = sum(1 for _ in itertools.takewhile(lambda ab: ab[0] == ab[1], zip(tokens[:-1], self._cached_tokens)))
-    return min(block._reusable_prefix_len(prefix_len, len(self._cached_tokens)) for block in self.blk)
+    live_start = min(block._reusable_prefix_len(prefix_len, len(self._cached_tokens)) for block in self.blk)
+    snapshot_prefix = sum(1 for _ in itertools.takewhile(lambda ab: ab[0] == ab[1], zip(tokens[:-1], self._snapshot_tokens)))
+    snapshot_start = len(self._snapshot_tokens) if snapshot_prefix == len(self._snapshot_tokens) else 0
+    return (snapshot_start, True) if snapshot_start > live_start else (live_start, False)
+
+  def get_start_pos(self, tokens:list[int]) -> int: return self._cache_start(tokens)[0]
 
   def generate(self, tokens:list[int], chunk_size:int=32, temperature:float=0.0):
-    if self.has_recurrent_block: chunk_size = 1
+    chunked_recurrent = self.has_recurrent_block and self.config.recurrent_prefill_chunked
+    if chunked_recurrent and self.config.recurrent_prefill_chunk_size:
+      chunk_size = min(chunk_size, self.config.recurrent_prefill_chunk_size)
+    if self.has_recurrent_block and not chunked_recurrent: chunk_size = 1
     v_start_pos = UOp.variable("start_pos", 0, self.max_context-1)
     v_toks = UOp.variable("toks", 1, chunk_size)
     # TODO: use UOp.variable for temperature once float variables are supported
-    temp = Tensor([temperature])
-    # assign all input tokens once, then slice from start_pos for the model call
-    t = Tensor(tokens + [0] * (self.max_context - len(tokens)), dtype="int32").reshape(1, self.max_context)
+    model_device = self.token_embd.weight.device
+    amd_tp = isinstance(model_device, tuple) and all(d.startswith("AMD") for d in model_device)
+    if temperature == 0.0: temp = None
+    elif amd_tp:
+      if self._temperature_buffer is None: self._temperature_buffer = Tensor.empty(1, device=model_device).realize()
+      temp_storage = self._temperature_buffer.uop.buf_uop.buffer
+      temp_buffers = temp_storage.bufs if isinstance(temp_storage, MultiBuffer) else [temp_storage]
+      temp_host = memoryview(array.array('f', [temperature])).cast('B')
+      for buf in temp_buffers: buf.ensure_allocated().allocator._copyin(buf._buf, temp_host)
+      temp = self._temperature_buffer
+    else: temp = Tensor([temperature], device=model_device)
+    # Keep the replicated AMD token buffer identity stable across HTTP requests so captured graphs
+    # see the same input topology. Updating this small int32 buffer is cheaper than rebuilding JITs.
+    if amd_tp:
+      if self._token_buffer is None:
+        self._token_buffer = Tensor.empty(1, self.max_context, dtype=dtypes.int32, device=model_device).realize()
+      token_storage = self._token_buffer.uop.buf_uop.buffer
+      token_buffers = token_storage.bufs if isinstance(token_storage, MultiBuffer) else [token_storage]
+      input_host = memoryview(array.array('i', tokens + [0] * (self.max_context-len(tokens)))).cast('B')
+      for buf in token_buffers: buf.ensure_allocated().allocator._copyin(buf._buf, input_host)
+      t = self._token_buffer
+    else: t = Tensor(tokens + [0] * (self.max_context - len(tokens)), dtype="int32", device=model_device).reshape(1, self.max_context)
     # recompute start_pos from what's currently valid in the caches
-    start_pos = self.get_start_pos(tokens)
+    start_pos, restore_snapshot = self._cache_start(tokens)
+    # This graph is captured by warmup. Resetting on-device avoids hundreds of synchronous copies
+    # to sharded AMD state buffers and guarantees unrelated requests don't schedule new kernels.
+    if restore_snapshot: self.restore_state_jit()
+    elif start_pos < len(self._cached_tokens) and self.has_recurrent_block: self.reset_jit()
     out, prompt_len = None, len(tokens)
+    token_host = memoryview(bytearray(4)) if amd_tp else None
     while len(tokens) < self.max_context:
-      n_toks = min(chunk_size, len(tokens) - start_pos)
-      sp, nt = v_start_pos.bind(start_pos), v_toks.bind(n_toks)
-      out = self(t[:, sp:sp+nt] if start_pos < prompt_len or out is None else out, sp, temp).realize()
+      remaining = len(tokens) - start_pos
+      # Full recurrent chunks use the high-throughput prefill graph. Process the tail through the
+      # rollout graph so every request uses only the two shapes captured during server warmup.
+      n_toks = chunk_size if chunked_recurrent and remaining >= chunk_size else 1 if chunked_recurrent else min(chunk_size, remaining)
+      # Recurrent blocks execute an explicit recurrence over T. Give them a static chunk length so
+      # Python constructs the recurrence once per encountered size; decode remains the T=1 JIT.
+      if chunked_recurrent:
+        # Token count is static for the recurrent kernel, but cache position must remain a runtime
+        # variable so repeated chunks do not replay MLA stores at the capture position.
+        sp = v_start_pos.bind(start_pos)
+        model_input = t[:, sp:sp+n_toks] if start_pos < prompt_len or out is None else out
+      else:
+        sp = v_start_pos.bind(start_pos)
+        nt = v_toks.bind(n_toks)
+        model_input = t[:, sp:sp+nt] if start_pos < prompt_len or out is None else out
+      out = self(model_input, sp, temp).realize()
       start_pos += n_toks
       # chunked prefill: keep processing until all prompt tokens are consumed
       if start_pos < len(tokens): continue
-      tokens.append(int(out.item()))
+      if self.has_recurrent_block and len(tokens) == prompt_len and self._state_tensors():
+        self._init_state_snapshots()
+        self.save_state_jit()
+        self._snapshot_tokens = tokens.copy()
+      tokens.append(amd_int32_item(out, token_host) if token_host is not None else int(out.item()))
       self._cached_tokens = tokens[:-1]
       yield tokens[-1]

tinygrad/llm/quant.py

@@ -0,0 +1,81 @@
+from tinygrad import Tensor, dtypes
+
+MX_BLOCK_SIZE = 32
+MXFP4_VALUES = (0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0,
+                -0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0)
+
+def _e8m0_scale(scale:Tensor) -> Tensor:
+  """Decode an OCP E8M0 scale byte. 127 encodes 2**0."""
+  return (scale.cast(dtypes.float32) - 127.0).exp2()
+
+def quantize_mxfp4(x:Tensor) -> tuple[Tensor, Tensor]:
+  """Quantize the last dimension to OCP MXFP4 (E2M1 values, E8M0 scale, block size 32)."""
+  if x.shape[-1] % MX_BLOCK_SIZE: raise ValueError(f"MXFP4 requires a multiple-of-32 last dimension, got {x.shape}")
+  *outer, k = x.shape
+  blocks = x.float().reshape(*outer, k//MX_BLOCK_SIZE, MX_BLOCK_SIZE)
+  amax = blocks.abs().max(axis=-1)
+  # Match the OCP/MLX reference: quantize amax / E2M1_MAX to the nearest E8M0
+  # power of two. This deliberately differs from extracting exponent bits: blocks
+  # whose maximum is near a power-of-two boundary can select a scale 2x smaller.
+  exponent = (amax.maximum(1e-38).div(6.0).log2().round()).clamp(-127, 127)
+  scale = (amax == 0).where(127, exponent + 127).cast(dtypes.uint8)
+  normalized = blocks / _e8m0_scale(scale).unsqueeze(-1)
+  # Midpoint bins avoid materializing a 16x larger distance tensor while preserving nearest-value encoding.
+  magnitude = normalized.abs()
+  # OCP formats use round-to-nearest-even: at alternating midpoints, the upper code has an even mantissa LSB.
+  code = sum(((magnitude >= midpoint) if upper_even else (magnitude > midpoint)).cast(dtypes.uint8)
+             for midpoint, upper_even in ((0.25, False), (0.75, True), (1.25, False), (1.75, True),
+                                           (2.5, False), (3.5, True), (5.0, False)))
+  code = ((normalized < 0) & (code != 0)).where(code + 8, code).reshape(*outer, k)
+  # Safetensors has no nibble dtype. Store the earlier element in the low nibble.
+  packed = code[..., ::2] + code[..., 1::2] * 16
+  return packed.contiguous(), scale.contiguous()
+
+def quantize_mxfp4_cpu(x:Tensor) -> tuple[Tensor, Tensor]:
+  """CPU converter fast path for large checkpoints. Inference itself does not depend on numpy."""
+  import numpy as np
+  if x.shape[-1] % MX_BLOCK_SIZE: raise ValueError(f"MXFP4 requires a multiple-of-32 last dimension, got {x.shape}")
+  array = x.float().numpy()
+  blocks = array.reshape(*array.shape[:-1], array.shape[-1]//MX_BLOCK_SIZE, MX_BLOCK_SIZE)
+  amax = np.max(np.abs(blocks), axis=-1)
+  exponent = np.clip(np.rint(np.log2(np.maximum(amax, 1e-38) / 6.0)), -127, 127)
+  scale = np.where(amax == 0, 127, exponent + 127).astype(np.uint8)
+  normalized = blocks / np.exp2(scale.astype(np.float32) - 127)[..., None]
+  magnitude = np.abs(normalized)
+  code = sum(((magnitude >= midpoint) if upper_even else (magnitude > midpoint)).astype(np.uint8)
+             for midpoint, upper_even in ((0.25, False), (0.75, True), (1.25, False), (1.75, True),
+                                           (2.5, False), (3.5, True), (5.0, False)))
+  code = np.where((normalized < 0) & (code != 0), code + 8, code).astype(np.uint8).reshape(array.shape)
+  packed = code[..., ::2] + code[..., 1::2] * 16
+  return Tensor(packed), Tensor(scale)
+
+def dequantize_mxfp4(packed:Tensor, scale:Tensor, dtype=dtypes.bfloat16) -> Tensor:
+  """Decode the packed representation emitted by quantize_mxfp4."""
+  if packed.shape[-1] != scale.shape[-1] * 16:
+    raise ValueError(f"incompatible MXFP4 values/scales: {packed.shape} and {scale.shape}")
+  lo = packed - packed.div(16, rounding_mode="trunc") * 16
+  hi = packed.div(16, rounding_mode="trunc")
+  code = Tensor.stack(lo, hi, dim=-1).reshape(*packed.shape[:-1], packed.shape[-1]*2)
+  values = Tensor(MXFP4_VALUES, dtype=dtypes.float32, device=packed.device)[code]
+  scales = _e8m0_scale(scale).unsqueeze(-1).expand(*scale.shape, MX_BLOCK_SIZE).reshape(*scale.shape[:-1], scale.shape[-1]*MX_BLOCK_SIZE)
+  return (values * scales).cast(dtype)
+
+def quantize_dequantize_mxfp8(x:Tensor, dtype=dtypes.bfloat16) -> Tensor:
+  """Apply the Kimi expert-activation MXFP8 E4M3/E8M0 round trip in 32-value blocks."""
+  if x.shape[-1] % MX_BLOCK_SIZE: raise ValueError(f"MXFP8 requires a multiple-of-32 last dimension, got {x.shape}")
+  *outer, k = x.shape
+  blocks = x.float().reshape(*outer, k//MX_BLOCK_SIZE, MX_BLOCK_SIZE)
+  amax = blocks.abs().max(axis=-1)
+  # As for MXFP4, the E8M0 scale is nearest-power-of-two(amax / E4M3_MAX).
+  exponent = (amax.maximum(1e-38).div(448.0).log2().round()).clamp(-127, 127)
+  scale = (amax == 0).where(127, exponent + 127).cast(dtypes.uint8)
+  normalized = blocks / _e8m0_scale(scale).unsqueeze(-1)
+  # Software OCP E4M3 rounding is required on gfx1100 (RDNA3 has no native FP8 dtype).
+  # E4M3 has three explicit mantissa bits and a minimum normal exponent of -6;
+  # using e=-6 also gives the 2**-9 subnormal quantum.
+  magnitude = normalized.abs().clamp(max_=448.0)
+  elem_exp = magnitude.maximum(2**-9).log2().floor().clamp(-6, 8)
+  quantum = (elem_exp - 3).exp2()
+  quantized = (magnitude / quantum).round() * quantum
+  quantized = (normalized < 0).where(-quantized, quantized).clamp(-448.0, 448.0)
+  return (quantized * _e8m0_scale(scale).unsqueeze(-1)).reshape(*outer, k).cast(dtype)

tinygrad/llm/serve.py

@@ -24,6 +24,11 @@ def parse_tool_call(s:str) -> tuple[str, typing.Any]|None:
     return fm.group(1), args
   return None
 
+def parse_kimi_tool_call(s:str) -> tuple[str, typing.Any]|None:
+  if (m := re.match(r"\s*(?:functions\.)?([^:\s]+)(?::[^\s]+)?\s*<\|tool_call_argument_begin\|>\s*(.*?)\s*\Z", s, re.DOTALL)) is None: return None
+  try: return m.group(1), json.loads(m.group(2))
+  except json.JSONDecodeError: return None
+
 def normalize_messages(messages:list[dict]) -> None:
   # chat templates expect tool_call arguments as dicts (OpenAI clients send JSON strings)
   for m in messages:
@@ -34,9 +39,10 @@ def normalize_messages(messages:list[dict]) -> None:
 
 class StreamRouter:
   # routes streamed output text to (field, text) deltas, keeping tool_call regions in .buf for the final parse
-  def __init__(self, reasoning:bool=False):
+  def __init__(self, reasoning:bool=False, xtml:bool=False):
     self.buf = ""
     self.mode = "reasoning" if reasoning else "undecided"  # output inside a think block is sent as reasoning_content
+    self.xtml = xtml
   def split(self, tag:str, final:bool) -> tuple[str, bool]:
     # split buf on the first full tag, holding back a partial tag at the end unless final
     if tag in self.buf:
@@ -45,20 +51,39 @@ class StreamRouter:
     hold = max((i for i in range(1, min(len(self.buf), len(tag))+1) if tag.startswith(self.buf[-i:])), default=0) if not final else 0
     emit, self.buf = self.buf[:len(self.buf)-hold], self.buf[len(self.buf)-hold:]
     return emit, False
+  def split_any(self, tags:tuple[str, ...], final:bool) -> tuple[str, str|None]:
+    found = [(self.buf.index(tag), tag) for tag in tags if tag in self.buf]
+    if found:
+      pos, tag = min(found)
+      before, self.buf = self.buf[:pos], self.buf[pos+len(tag):]
+      return before, tag
+    hold = max((i for tag in tags for i in range(1, min(len(self.buf), len(tag))+1) if tag.startswith(self.buf[-i:])), default=0) if not final else 0
+    emit, self.buf = self.buf[:len(self.buf)-hold], self.buf[len(self.buf)-hold:]
+    return emit, None
   def route(self, piece:str, final:bool=False) -> typing.Iterator[tuple[str, str]]:
     self.buf += piece
     if self.mode == "undecided":  # decide whether the output starts with a think block
       if not final and len(self.buf) < len("<think>") and "<think>".startswith(self.buf): return
       self.mode, self.buf = ("reasoning", self.buf[len("<think>"):]) if self.buf.startswith("<think>") else ("content", self.buf)
     if self.mode == "reasoning":
-      emit, done = self.split("</think>", final)
+      emit, done = self.split("<|close|>think<|sep|>" if self.xtml else "</think>", final)
       if emit: yield "reasoning_content", emit
       if not done: return
+      self.mode = "content_open" if self.xtml else "content"
+    if self.mode == "content_open":
+      _, found = self.split("<|open|>response<|sep|>", final)
+      if not found: return
       self.mode = "content"
+    if self.mode == "done": return
+    if self.xtml and self.mode == "content":
+      emit, found = self.split("<|close|>response<|sep|>", final)
+      if emit: yield "content", emit
+      if found: self.mode = "done"
+      return
     if self.mode == "tool": return
-    emit, found = self.split("<tool_call>", final)
+    emit, tool_tag = self.split_any(("<tool_call>", "<|tool_calls_section_begin|>"), final)
     if emit: yield "content", emit
-    if found: self.mode, self.buf = "tool", "<tool_call>" + self.buf
+    if tool_tag: self.mode, self.buf = "tool", tool_tag + self.buf
 
 class Handler(HTTPRequestHandler):
   server: LLMServer
@@ -67,7 +92,7 @@ class Handler(HTTPRequestHandler):
     if self.path == "/v1/models": self.send_data(json.dumps({"object":"list","data":[{"id":self.server.model_name,"object":"model"}]}).encode())
     else: self.send_data((pathlib.Path(__file__).parent / "chat.html").read_bytes(), content_type="text/html")
   def run_model(self, ids:list[int], model_name:str, include_usage=False, max_tokens:int|None=None, temperature:float=0.0,
-                reasoning:bool=False):
+                reasoning:bool=False, xtml:bool=False):
     model, tok = self.server.model, self.server.tok
     prompt_tokens = len(ids)
     cache_start_pos = model.get_start_pos(ids)
@@ -78,7 +103,7 @@ class Handler(HTTPRequestHandler):
     finish_reason = "stop"
     st = pt = time.perf_counter()
     dec = tok.stream_decoder()
-    router = StreamRouter(reasoning)
+    router = StreamRouter(reasoning, xtml)
     def log_stats(interrupted:bool=False):
       et = time.perf_counter()
       total = f"total:{et-st:6.2f}s"
@@ -106,6 +131,14 @@ class Handler(HTTPRequestHandler):
           name, args = parsed
           tool_calls.append({"index":len(tool_calls), "id":f"call_{uuid.uuid4().hex[:24]}", "type":"function",
                              "function":{"name":name, "arguments":args if isinstance(args, str) else json.dumps(args)}})
+      for m in re.finditer(r"<\|tool_call_begin\|>(.*?)<\|tool_call_end\|>", router.buf, re.DOTALL):
+        if (parsed := parse_kimi_tool_call(m.group(1))) is None:
+          stderr_log(f"failed to parse Kimi tool call: {m.group(1)[:200]}")
+          yield chunk({"content":m.group(0)})
+        else:
+          name, args = parsed
+          tool_calls.append({"index":len(tool_calls), "id":f"call_{uuid.uuid4().hex[:24]}", "type":"function",
+                             "function":{"name":name, "arguments":json.dumps(args)}})
       if tool_calls:
         yield chunk({"tool_calls":tool_calls})
         if finish_reason == "stop": finish_reason = "tool_calls"
@@ -139,9 +172,10 @@ class Handler(HTTPRequestHandler):
 
       # reply
       max_tokens = body.get("max_completion_tokens") or body.get("max_tokens")
+      xtml = rendered.rstrip().endswith("<|open|>think<|sep|>")
       chunks = self.run_model(ids, body["model"], not body.get("stream") or body.get("stream_options",{}).get("include_usage", False),
                               max_tokens=max_tokens, temperature=float(body.get("temperature", 0.0)),
-                              reasoning=rendered.rstrip().endswith("<think>"))
+                              reasoning=xtml or rendered.rstrip().endswith("<think>"), xtml=xtml)
       if body.get("stream"): self.stream_json(chunks)
       else:
         out, reasoning, tool_calls, finish_reason = [], [], [], "stop"

tinygrad/mixin/elementwise.py

@@ -115,8 +115,7 @@ class ElementwiseMixin(CreationMixin):
     ```
     """
     a, b = self._broadcasted(x, reverse)
-    # alu, not +: _broadcasted already promoted these, and a second promote would cast -b (only a bare weak CONST is kept weak)
-    return a.alu(Ops.ADD, -b)
+    return a + (-b)
 
   def mul(self, x: Self | ConstType, reverse: bool = False) -> Self:
     """
@@ -246,9 +245,8 @@ class ElementwiseMixin(CreationMixin):
     if dtypes.is_int(a.dtype) and dtypes.is_int(b.dtype):
       if rounding_mode == "trunc": return a.alu(Ops.CDIV, b)
       if rounding_mode == "floor": return a.alu(Ops.FLOORDIV, b)
-    if dtypes.is_int(a.dtype) or a.dtype == dtypes.bool: a = a.cast(dtypes.default_float)
-    # alu, not *: _broadcasted already promoted these, and a second promote would cast 1/b (only a bare weak CONST is kept weak)
-    d = a.alu(Ops.MUL, b.reciprocal())
+      a = a.cast(dtypes.default_float)
+    d = a * b.reciprocal()
     if rounding_mode is None: return d
     if rounding_mode == "trunc": return d.trunc()
     if rounding_mode == "floor": return d.floor()

tinygrad/mixin/gradient.py

@@ -3,7 +3,6 @@ import math, dataclasses
 from tinygrad.uop.ops import UOp, PatternMatcher, UPat, Ops, all_metadata, broadcast_axes
 from tinygrad.helpers import argsort
 from tinygrad.dtype import sum_acc_dtype
-from tinygrad.function import renumber_invalid_outputs
 
 def reduce_gradient(ctx:UOp, ret:UOp, op:Ops):
   if op == Ops.ADD: return (ctx._broadcast_to(ret.src[0].shape),)
@@ -33,7 +32,7 @@ def call_gradient(ctx:UOp, k:UOp, needed:set[int]) -> tuple[UOp|None, ...]:
   params = {x.arg.slot:x for x in fxn.toposort(enter_calls=False) if x.op == Ops.PARAM}
   grad_args = ctx.src
   root_grad = UOp(Ops.TUPLE, src=tuple(UOp(Ops.NOOP) if g.op is Ops.NOOP else
-    g if g.device is None else g.param_like(len(args)+i) for i,g in enumerate(grad_args)))
+    g if g.base.op is Ops.CONST else g.param_like(len(args)+i) for i,g in enumerate(grad_args)))
   grads = compute_gradient(fxn, root_grad, set(params.values()))
   # for precompiled calls, substitute forward outputs with params so intermediates aren't recomputed
   fwd_subs = {src: src.param_like(len(args)+len(grad_args)+i) for i, src in enumerate(fxn.src)} if k.arg.precompile else {}
@@ -41,7 +40,6 @@ def call_gradient(ctx:UOp, k:UOp, needed:set[int]) -> tuple[UOp|None, ...]:
   # collect needed gradient bodies, compact unused params, create a single backward CALL
   grad_bodies = [(i, grads[p]) for i in needed if (p:=params.get(i)) is not None and p in grads]
   bwd_body = UOp.maketuple(*(gb for _, gb in grad_bodies)).substitute(fwd_subs, walk=True)
-  bwd_body = renumber_invalid_outputs(bwd_body)
   bwd_body, compact_args = _compact_params(bwd_body, (*args, *grad_args, *fwd_outs))
   bwd_call = bwd_body.call(*compact_args, name=(k.arg.name or "")+"_backward", precompile=k.arg.precompile_backward)
   gb_map = {i: idx for idx, (i, _) in enumerate(grad_bodies)}

tinygrad/mixin/movement.py

@@ -46,16 +46,6 @@ class MovementMixin:
     """
     return prod(self.shape)
 
-  @property
-  def max_shape(self) -> tuple[int, ...]:
-    """The shape with every symbolic dimension replaced by its maximum."""
-    from tinygrad.uop.ops import to_max_shape  # deferred: ops.py imports the mixins
-    return to_max_shape(self.shape)
-
-  def max_numel(self) -> int:
-    """The number of elements in `max_shape`."""
-    return prod(self.max_shape)
-
   def size(self, dim:int|None=None) -> sint|tuple[sint, ...]:
     """
     Returns the size of the tensor. If `dim` is specified, return the length along dimension `dim`. Otherwise return the shape of the tensor.