Merge pull request '[CI] Build GB10 ARM64 image with Gitea Actions' (#2) from karylab_agents/vllm:ci/build-gb10-arm64-image into karylab/gb10

Reviewed-on: #2

.gitea/workflows/build-gb10.yml

@@ -0,0 +1,105 @@
+name: Build GB10 ARM64 image
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - karylab/gb10
+    paths:
+      - ".gitea/workflows/build-gb10.yml"
+      - "CMakeLists.txt"
+      - "cmake/**"
+      - "csrc/**"
+      - "docker/**"
+      - "pyproject.toml"
+      - "requirements/**"
+      - "setup.py"
+      - "vllm/**"
+
+env:
+  REGISTRY: git.karylab.com
+  IMAGE_NAME: karylab_agents/vllm-gb10
+
+jobs:
+  build-arm64:
+    # This label is registered by gitea-act-runner.yml.
+    runs-on: vllm-x86-builder
+    timeout-minutes: 720
+
+    # Do not add pull_request as a trigger. This runner controls the build
+    # host's Docker socket and must only execute reviewed code on main.
+    permissions:
+      contents: read
+
+    steps:
+      - name: Check out source
+        uses: actions/checkout@v4
+
+      - name: Enable ARM64 emulation
+        uses: docker/setup-qemu-action@v3
+        with:
+          platforms: arm64
+
+      - name: Create Buildx builder
+        uses: docker/setup-buildx-action@v3
+
+      - name: Validate registry credential
+        env:
+          REGISTRY_TOKEN: ${{ secrets.REGISTRY_TOKEN }}
+        run: |
+          if [ -z "${REGISTRY_TOKEN}" ]; then
+            echo "Missing repository Actions secret: REGISTRY_TOKEN" >&2
+            exit 1
+          fi
+
+      - name: Log in to Gitea Container Registry
+        env:
+          REGISTRY_TOKEN: ${{ secrets.REGISTRY_TOKEN }}
+        run: |
+          printf '%s' "${REGISTRY_TOKEN}" |
+            docker login "${REGISTRY}" \
+              --username karylab_agents \
+              --password-stdin
+
+      - name: Derive immutable image tag
+        id: image
+        shell: bash
+        run: |
+          short_sha="$(git rev-parse --short=12 HEAD)"
+          echo "tag=gb10-${short_sha}" >> "${GITHUB_OUTPUT}"
+          echo "revision=$(git rev-parse HEAD)" >> "${GITHUB_OUTPUT}"
+
+      - name: Build and push ARM64 image
+        env:
+          IMAGE_TAG: ${{ steps.image.outputs.tag }}
+          REVISION: ${{ steps.image.outputs.revision }}
+        run: |
+          image="${REGISTRY}/${IMAGE_NAME}"
+          run_url="${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}"
+
+          docker buildx build \
+            --platform linux/arm64 \
+            --file docker/Dockerfile \
+            --target vllm-openai \
+            --build-arg "torch_cuda_arch_list=12.0" \
+            --build-arg "max_jobs=4" \
+            --build-arg "nvcc_threads=8" \
+            --build-arg "VLLM_BUILD_COMMIT=${REVISION}" \
+            --build-arg "VLLM_BUILD_PIPELINE=gitea-actions" \
+            --build-arg "VLLM_BUILD_URL=${run_url}" \
+            --build-arg "VLLM_IMAGE_TAG=${image}:${IMAGE_TAG}" \
+            --cache-from "type=registry,ref=${image}:buildcache" \
+            --cache-to "type=registry,ref=${image}:buildcache,mode=max,compression=zstd,oci-mediatypes=true" \
+            --label "org.opencontainers.image.revision=${REVISION}" \
+            --label "org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}" \
+            --tag "${image}:${IMAGE_TAG}" \
+            --tag "${image}:latest" \
+            --push \
+            .
+
+      - name: Verify published image
+        env:
+          IMAGE_TAG: ${{ steps.image.outputs.tag }}
+        run: |
+          docker buildx imagetools inspect \
+            "${REGISTRY}/${IMAGE_NAME}:${IMAGE_TAG}"

tests/model_executor/test_deepseek_v4_mhc_warmup.py

@@ -0,0 +1,668 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""CPU-pure tests for DeepSeek V4 MHC warmup selection and gating logic.
+
+The TileLang JIT kernels require CUDA; these tests verify CPU-side selection,
+layer-finding, and gating without a GPU.  ``_select_mhc_split_key_token_sizes``
+tests monkeypatch ``compute_num_split`` (CUDA-backed) with a deterministic fake
+so the real function runs under CPU-pure coverage.
+"""
+
+import importlib
+import sys
+import types
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+    _find_deepseek_v4_model,
+    _find_first_mhc_layer,
+    _normalize_token_sizes,
+    _select_mhc_split_key_token_sizes,
+    _select_mhc_warmup_token_sizes,
+)
+from vllm.utils.math_utils import cdiv
+
+# ── Helpers ───────────────────────────────────────────────────────────────
+
+
+def _fake_compute_num_split(n_sms: int):
+    """Deterministic ``compute_num_split`` with explicit SM count instead of
+    ``torch.cuda``."""
+
+    def _inner(block_k: int, k: int | None, grid_size: int) -> int:
+        split_k = n_sms // grid_size
+        if k is not None:
+            split_k = min(split_k, cdiv(k, block_k) // 4)
+        return max(split_k, 1)
+
+    return _inner
+
+
+def _patch_compute_split(*, n_sms: int, monkeypatch: pytest.MonkeyPatch) -> None:
+    """Stub ``tilelang_kernels`` in ``sys.modules`` via ``monkeypatch.setitem``
+    so the local import inside ``_select_mhc_split_key_token_sizes`` resolves
+    without triggering the real module (which requires TileLang + CUDA).  The
+    stub is automatically restored after each test."""
+
+    MODULE_PATH = "vllm.model_executor.kernels.mhc.tilelang_kernels"
+
+    # Ensure parent packages exist in sys.modules so the dotted-path import
+    # resolves through parent lookups.
+    for parent_path in (
+        "vllm",
+        "vllm.model_executor",
+        "vllm.model_executor.kernels",
+        "vllm.model_executor.kernels.mhc",
+    ):
+        if parent_path not in sys.modules:
+            monkeypatch.setitem(sys.modules, parent_path, types.ModuleType(parent_path))
+
+    stub = types.ModuleType(MODULE_PATH)
+    stub.compute_num_split = _fake_compute_num_split(n_sms)
+    monkeypatch.setitem(sys.modules, MODULE_PATH, stub)
+
+
+# ── _normalize_token_sizes ───────────────────────────────────────────────
+
+
+class TestNormalizeTokenSizes:
+    def test_empty_when_no_sizes(self) -> None:
+        assert _normalize_token_sizes((), max_tokens=100) == []
+
+    def test_removes_out_of_range(self) -> None:
+        assert _normalize_token_sizes([0, 1, 50, 100, 200], max_tokens=100) == [
+            1,
+            50,
+            100,
+        ]
+
+    def test_deduplicates_and_sorts(self) -> None:
+        assert _normalize_token_sizes([4, 1, 4, 8, 2], max_tokens=100) == [1, 2, 4, 8]
+
+    def test_accepts_iterator(self) -> None:
+        assert _normalize_token_sizes(iter({1, 2, 3}), max_tokens=10) == [1, 2, 3]
+
+
+# ── _select_mhc_warmup_token_sizes ────────────────────────────────────────
+
+
+class TestSelectMhcWarmupTokenSizes:
+    def test_empty_on_zero_max_tokens(self) -> None:
+        assert (
+            _select_mhc_warmup_token_sizes(max_tokens=0, cudagraph_capture_sizes=[])
+            == []
+        )
+
+    def test_contains_1(self) -> None:
+        assert 1 in _select_mhc_warmup_token_sizes(
+            max_tokens=10, cudagraph_capture_sizes=[]
+        )
+
+    def test_bounded_by_max_tokens(self) -> None:
+        sizes = _select_mhc_warmup_token_sizes(max_tokens=5, cudagraph_capture_sizes=[])
+        assert all(1 <= s <= 5 for s in sizes)
+
+    def test_includes_cudagraph_capture_sizes(self) -> None:
+        sizes = _select_mhc_warmup_token_sizes(
+            max_tokens=100, cudagraph_capture_sizes=[7, 33]
+        )
+        assert 7 in sizes and 33 in sizes
+
+    def test_includes_max_auto_tokens(self) -> None:
+        sizes = _select_mhc_warmup_token_sizes(
+            max_tokens=100, cudagraph_capture_sizes=[]
+        )
+        assert 100 in sizes
+
+    def test_respects_auto_warmup_cap(self) -> None:
+        sizes = _select_mhc_warmup_token_sizes(
+            max_tokens=20000, cudagraph_capture_sizes=[]
+        )
+        assert max(sizes) == 16384
+
+    def test_does_not_include_zero(self) -> None:
+        assert _select_mhc_warmup_token_sizes(
+            max_tokens=1, cudagraph_capture_sizes=[]
+        ) == [1]
+
+
+# ── _select_mhc_split_key_token_sizes ─────────────────────────────────────
+
+
+class TestSelectMhcSplitKeyTokenSizes:
+    """Real ``_select_mhc_split_key_token_sizes`` invoked under a deterministic
+    ``compute_num_split`` monkeypatch.  No CUDA required outside the skip-guard
+    cross-check below."""
+
+    # ── Broadcast-variant key counts (k_size = hidden_size) ──────────────
+
+    @pytest.mark.parametrize(
+        "max_tokens,k_size,n_sms,expected_keys",
+        [
+            (8192, 4096, 188, 16),  # RTX PRO 6000 Blackwell, broadcast K
+            (8192, 4096, 132, 15),  # H100 SXM
+            (8192, 4096, 80, 12),
+            (8192, 2048, 188, 8),
+            (8192, 1024, 188, 4),
+            (1024, 4096, 188, 6),
+            (1, 4096, 188, 1),
+        ],
+    )
+    def test_broadcast_key_count(
+        self, monkeypatch, max_tokens, k_size, n_sms, expected_keys
+    ) -> None:
+        _patch_compute_split(n_sms=n_sms, monkeypatch=monkeypatch)
+        reps = _select_mhc_split_key_token_sizes(max_tokens=max_tokens, k_size=k_size)
+        assert len(reps) == expected_keys, (
+            f"Expected {expected_keys} keys for max_tokens={max_tokens}, "
+            f"k_size={k_size}, n_sms={n_sms}, got {len(reps)}"
+        )
+
+    # ── Non-broadcast-variant key counts (k_size = hc_mult * hidden_size) ──
+
+    @pytest.mark.parametrize(
+        "max_tokens,k_size,n_sms,expected_keys",
+        [
+            (8192, 65536, 188, 26),  # DSv4 default: hc_mult=16, hidden=4096
+            (8192, 65536, 132, 22),  # H100 SXM
+            (8192, 65536, 80, 16),
+            (8192, 32768, 188, 26),  # hc_mult=8, hidden=4096; same as 188 SM full
+            (8192, 8192, 188, 22),
+            (1024, 65536, 188, 16),
+            (1, 65536, 188, 1),
+        ],
+    )
+    def test_non_broadcast_key_count(
+        self, monkeypatch, max_tokens, k_size, n_sms, expected_keys
+    ) -> None:
+        _patch_compute_split(n_sms=n_sms, monkeypatch=monkeypatch)
+        reps = _select_mhc_split_key_token_sizes(max_tokens=max_tokens, k_size=k_size)
+        assert len(reps) == expected_keys, (
+            f"Expected {expected_keys} keys for max_tokens={max_tokens}, "
+            f"k_size={k_size}, n_sms={n_sms}, got {len(reps)}"
+        )
+
+    # ── Semantic invariants ─────────────────────────────────────────────
+
+    def test_all_keys_distinct(self, monkeypatch) -> None:
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+        reps = _select_mhc_split_key_token_sizes(max_tokens=8192, k_size=4096)
+        fake_cns = _fake_compute_num_split(188)
+        ns_values = [fake_cns(64, 4096, cdiv(t, 64)) for t in reps]
+        assert len(set(ns_values)) == len(ns_values), (
+            f"Duplicate n_splits: {dict(zip(reps, ns_values))}"
+        )
+
+    def test_last_key_n_splits_is_one(self, monkeypatch) -> None:
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+        reps = _select_mhc_split_key_token_sizes(max_tokens=8192, k_size=4096)
+        last_ns = _fake_compute_num_split(188)(64, 4096, cdiv(reps[-1], 64))
+        assert last_ns == 1, f"Last key n_splits={last_ns}, expected 1"
+
+    # ── Exact broadcast sequence (k_size=hidden_size=4096, 188 SMs) ──────
+
+    def test_exact_broadcast_sequence_188_sms(self, monkeypatch) -> None:
+        # fmt: off
+        expected = [1, 705, 769, 833, 897, 961, 1089, 1153,
+                    1281, 1473, 1665, 1985, 2369, 3009, 3969, 6017]
+        # fmt: on
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+        reps = _select_mhc_split_key_token_sizes(max_tokens=8192, k_size=4096)
+        assert reps == expected, (
+            f"188-SM broadcast (K=4096) sequence mismatch\n"
+            f"  Expected ({len(expected)}): {expected}\n"
+            f"  Got      ({len(reps)}): {reps}"
+        )
+
+    # ── Exact non-broadcast sequence (k_size=65536, 188 SMs) ─────────────
+
+    def test_exact_non_broadcast_sequence_188_sms(self, monkeypatch) -> None:
+        # fmt: off
+        expected = [1, 65, 129, 193, 257, 321, 385, 449, 513,
+                    577, 641, 705, 769, 833, 897, 961, 1089,
+                    1153, 1281, 1473, 1665, 1985, 2369, 3009,
+                    3969, 6017]
+        # fmt: on
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+        reps = _select_mhc_split_key_token_sizes(max_tokens=8192, k_size=65536)
+        assert reps == expected, (
+            f"188-SM non-broadcast (K=65536) sequence mismatch\n"
+            f"  Expected ({len(expected)}): {expected}\n"
+            f"  Got      ({len(reps)}): {reps}"
+        )
+
+    # ── CUDA real cross-check (skipped when GPU or TileLang unavailable) ─
+
+    def test_selector_matches_compute_num_split(self, monkeypatch) -> None:
+        if not torch.cuda.is_available():
+            pytest.skip("CUDA required for production cross-check")
+
+        MODULE_PATH = "vllm.model_executor.kernels.mhc.tilelang_kernels"
+        monkeypatch.delitem(sys.modules, MODULE_PATH, raising=False)
+        importlib.invalidate_caches()
+
+        try:
+            from vllm.model_executor.kernels.mhc.tilelang_kernels import (
+                compute_num_split,
+            )
+        except Exception as e:
+            pytest.skip(f"TileLang real module unavailable: {e}")
+
+        real_module = sys.modules[MODULE_PATH]
+        assert real_module.__file__ is not None
+        assert "vllm" in real_module.__file__
+
+        reps = _select_mhc_split_key_token_sizes(max_tokens=8192, k_size=4096)
+        assert len(reps) >= 1
+        ns_values = [compute_num_split(64, 4096, cdiv(t, 64)) for t in reps]
+        assert len(set(ns_values)) == len(ns_values), (
+            f"Duplicate n_splits: {dict(zip(reps, ns_values))}"
+        )
+        assert all(ns >= 1 for ns in ns_values), "Some n_splits are zero"
+
+
+# ── _find_first_mhc_layer ───────────────────────────────────────────────
+
+
+class TestFindFirstMhcLayer:
+    def test_finds_layer_with_all_required_attrs(self) -> None:
+        class MockLayer:
+            hc_pre = hc_post = hc_attn_fn = hc_attn_scale = None
+            hc_attn_base = hc_ffn_fn = hc_ffn_scale = hc_ffn_base = None
+
+        MockLayer.__name__ = MockLayer.__qualname__ = "DeepseekV4DecoderLayer"
+
+        class MockModel:
+            def modules(self):
+                yield self
+                yield MockLayer()
+
+        result = _find_first_mhc_layer(MockModel())
+        assert result is not None
+        assert result.__class__.__name__ == "DeepseekV4DecoderLayer"
+
+    def test_skips_layer_missing_required_attr(self) -> None:
+        class IncompleteLayer:
+            hc_post = hc_attn_fn = hc_attn_scale = hc_attn_base = None
+            hc_ffn_fn = hc_ffn_scale = hc_ffn_base = None
+
+        IncompleteLayer.__name__ = IncompleteLayer.__qualname__ = (
+            "DeepseekV4DecoderLayer"
+        )
+
+        class MockModel:
+            def modules(self):
+                yield self
+                yield IncompleteLayer()
+
+        assert _find_first_mhc_layer(MockModel()) is None
+
+
+# ── _find_deepseek_v4_model ──────────────────────────────────────────────
+
+
+class TestFindDeepseekV4Model:
+    def test_finds_model_with_all_required_attrs(self) -> None:
+        class MockDsModel:
+            hc_head_fn = hc_head_scale = hc_head_base = None
+
+        MockDsModel.__name__ = MockDsModel.__qualname__ = "DeepseekV4Model"
+
+        class MockModel:
+            def modules(self):
+                yield self
+                yield MockDsModel()
+
+        result = _find_deepseek_v4_model(MockModel())
+        assert result is not None
+        assert result.__class__.__name__ == "DeepseekV4Model"
+
+    def test_skips_model_missing_required_attr(self) -> None:
+        class IncompleteModel:
+            hc_head_fn = hc_head_scale = None
+
+        IncompleteModel.__name__ = IncompleteModel.__qualname__ = "DeepseekV4Model"
+
+        class MockModel:
+            def modules(self):
+                yield self
+                yield IncompleteModel()
+
+        assert _find_deepseek_v4_model(MockModel()) is None
+
+    def test_skips_wrong_class_name(self) -> None:
+        class OtherModel:
+            hc_head_fn = hc_head_scale = hc_head_base = None
+
+        OtherModel.__name__ = OtherModel.__qualname__ = "OtherModel"
+
+        class MockModel:
+            def modules(self):
+                yield self
+                yield OtherModel()
+
+        assert _find_deepseek_v4_model(MockModel()) is None
+
+
+# ── Broadcast no-op gates ─────────────────────────────────────────────────
+
+
+class TestWarmupBroadcastNoOpConditions:
+    @staticmethod
+    def _install_broadcast_spy(monkeypatch) -> list[str]:
+        calls: list[str] = []
+        module_path = "vllm.model_executor.kernels.mhc.tilelang"
+        stub = types.ModuleType(module_path)
+        stub.mhc_pre_broadcast_tilelang = lambda *args, **kwargs: calls.append(
+            "mhc_pre_broadcast_tilelang"
+        )
+        monkeypatch.setitem(sys.modules, module_path, stub)
+        return calls
+
+    @staticmethod
+    def _model_with_layer(layer):
+        layer.__name__ = layer.__qualname__ = "DeepseekV4DecoderLayer"
+
+        class MockModel:
+            def modules(self):
+                yield self
+                yield layer()
+
+        return MockModel()
+
+    def test_noop_when_no_broadcast_layer(self) -> None:
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            _warmup_broadcast_mhc,
+        )
+
+        class NoBroadcastLayer:
+            hc_attn_fn_broadcast = None
+
+        _warmup_broadcast_mhc(
+            self._model_with_layer(NoBroadcastLayer), token_sizes=[1, 2, 4]
+        )
+
+    def test_noop_when_device_not_cuda(self, monkeypatch) -> None:
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            _warmup_broadcast_mhc,
+        )
+
+        calls = self._install_broadcast_spy(monkeypatch)
+
+        class CpuLayer:
+            hc_attn_fn_broadcast = torch.empty(0)
+            hc_attn_fn = torch.empty(0, device="cpu")
+
+        _warmup_broadcast_mhc(self._model_with_layer(CpuLayer), token_sizes=[1, 2, 4])
+        assert not calls
+
+    def test_noop_when_broadcast_not_a_tensor(self, monkeypatch) -> None:
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            _warmup_broadcast_mhc,
+        )
+
+        calls = self._install_broadcast_spy(monkeypatch)
+
+        class BoolBroadcastLayer:
+            hc_attn_fn_broadcast = True
+            hc_attn_fn = SimpleNamespace(device=torch.device("cuda"))
+
+        _warmup_broadcast_mhc(
+            self._model_with_layer(BoolBroadcastLayer), token_sizes=[1, 2, 4]
+        )
+        assert not calls
+
+    def test_noop_when_broadcast_device_mismatch(self, monkeypatch) -> None:
+        from unittest import mock
+
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            _warmup_broadcast_mhc,
+        )
+
+        calls = self._install_broadcast_spy(monkeypatch)
+        broadcast = mock.MagicMock(spec=torch.Tensor)
+        broadcast.device = torch.device("cuda:1")
+
+        class MismatchedLayer:
+            hc_attn_fn_broadcast = broadcast
+            hc_attn_fn = SimpleNamespace(device=torch.device("cuda:0"))
+
+        _warmup_broadcast_mhc(
+            self._model_with_layer(MismatchedLayer), token_sizes=[1, 2, 4]
+        )
+        assert not calls
+
+
+# ── _warmup_layer_mhc union behavior ─────────────────────────────────────
+
+
+class TestWarmupLayerMhcUnion:
+    """Verify ``_warmup_layer_mhc`` unions general ``token_sizes`` with
+    split-key reps computed from ``hc_mult * hidden_size``, and calls
+    ``hc_pre``/``hc_post`` for each token size in the union."""
+
+    def test_unions_general_sizes_with_split_key_reps(self, monkeypatch) -> None:
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            _warmup_layer_mhc,
+        )
+
+        # Patch compute_num_split so _select_mhc_split_key_token_sizes works
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+
+        called_sizes: list[int] = []
+
+        class MockFn:
+            device = torch.device("cpu")
+
+        class MockLayer:
+            hidden_size = 4096
+            hc_mult = 16
+            hc_attn_fn = hc_attn_scale = hc_attn_base = MockFn()
+            hc_ffn_fn = hc_ffn_scale = hc_ffn_base = MockFn()
+
+            def hc_pre(self, residual_slice, fn, scale, base):
+                called_sizes.append(residual_slice.shape[0])
+                return (None, None, None)
+
+            def hc_post(self, layer_input, residual_slice, post_mix, comb_mix):
+                pass
+
+        MockLayer.__name__ = MockLayer.__qualname__ = "DeepseekV4DecoderLayer"
+
+        general_sizes = [1, 2, 4, 8, 16, 32, 64, 128]
+        _warmup_layer_mhc(MockLayer(), general_sizes)
+
+        # Compute expected union: general | split-key (K=65536, 188 SMs)
+        split_key_sizes = _select_mhc_split_key_token_sizes(
+            max_tokens=max(general_sizes), k_size=65536
+        )
+        expected_union = sorted(set(general_sizes) | set(split_key_sizes))
+
+        # hc_pre is called once per size per (attn + ffn) = 2x per size.
+        # Deduplicate to check unique sizes covered.
+        unique_called = sorted(set(called_sizes))
+        assert unique_called == expected_union, (
+            f"_warmup_layer_mhc called sizes: {unique_called}\n"
+            f"Expected union: {expected_union}\n"
+            f"General: {general_sizes}\n"
+            f"Split-key: {split_key_sizes}"
+        )
+
+    def test_calls_both_attn_and_ffn_for_each_size(self, monkeypatch) -> None:
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            _warmup_layer_mhc,
+        )
+
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+
+        call_log: list[str] = []
+
+        class MockFn:
+            device = torch.device("cpu")
+
+        class MockLayer:
+            hidden_size = 4096
+            hc_mult = 16
+            hc_attn_fn = hc_attn_scale = hc_attn_base = MockFn()
+            hc_ffn_fn = hc_ffn_scale = hc_ffn_base = MockFn()
+
+            def hc_pre(self, residual_slice, fn, scale, base):
+                size = residual_slice.shape[0]
+                call_log.append(f"hc_pre(size={size})")
+                return (None, None, None)
+
+            def hc_post(self, layer_input, residual_slice, post_mix, comb_mix):
+                call_log.append("hc_post")
+
+        MockLayer.__name__ = MockLayer.__qualname__ = "DeepseekV4DecoderLayer"
+
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+        _warmup_layer_mhc(MockLayer(), [1, 2])
+
+        # For each size: 2 calls (attn + ffn) * hc_pre+post = 4 log entries
+        # For 2 general sizes + any split-key reps bounded by max_tokens=2
+        # (which should be zero split-key reps since max_tokens=2 < 65)
+        # So 2 sizes * (hc_pre_attn, hc_post, hc_pre_ffn, hc_post) = 8 entries
+        assert len(call_log) == 8, (
+            f"Expected 8 log entries for 2 general sizes, got {len(call_log)}: "
+            f"{call_log}"
+        )
+
+
+# ── Model-type gate ───────────────────────────────────────────────────────
+
+
+class TestDeepseekV4ModelGate:
+    def test_returns_early_for_non_dsv4_model_type(self) -> None:
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            deepseek_v4_mhc_warmup,
+        )
+
+        class OtherModel:
+            config = SimpleNamespace(model_type="llama")
+
+            def modules(self):
+                return iter([])
+
+        deepseek_v4_mhc_warmup(OtherModel(), max_tokens=1024)
+
+
+# ── Orchestration ─────────────────────────────────────────────────────────
+
+
+class TestDeepseekV4MhcWarmupOrchestration:
+    """All three internal stages called in order.  TileLang dependencies
+    monkeypatched; mock modules use CUDA-like device attributes."""
+
+    def test_all_three_stages_called_in_order(self, monkeypatch) -> None:
+        from vllm.model_executor.warmup.deepseek_v4_mhc_warmup import (
+            deepseek_v4_mhc_warmup,
+        )
+
+        calls: list[tuple] = []
+        _fake_cuda = SimpleNamespace(device=torch.device("cuda"))
+
+        def _record_layer(layer, token_sizes):
+            calls.append(("_warmup_layer_mhc", token_sizes))
+
+        def _record_broadcast(model_arg, token_sizes):
+            calls.append(("_warmup_broadcast_mhc", model_arg, token_sizes))
+
+        def _record_head(model_arg, token_sizes):
+            calls.append(("_warmup_hc_head", token_sizes))
+
+        class MockLayer:
+            hc_pre = hc_post = lambda *a: None
+            hc_attn_fn = hc_attn_scale = hc_attn_base = _fake_cuda
+            hc_ffn_fn = hc_ffn_scale = hc_ffn_base = _fake_cuda
+            hidden_size = 4096
+            hc_mult = 16
+
+        MockLayer.__name__ = MockLayer.__qualname__ = "DeepseekV4DecoderLayer"
+
+        class MockDsModel:
+            hc_head_fn = hc_head_scale = hc_head_base = _fake_cuda
+            config = SimpleNamespace(hidden_size=4096)
+            hc_mult = 16
+            hc_eps = rms_norm_eps = 1e-6
+
+        MockDsModel.__name__ = MockDsModel.__qualname__ = "DeepseekV4Model"
+
+        class MockModel:
+            config = SimpleNamespace(model_type="deepseek_v4")
+
+            def modules(self):
+                yield self
+                yield MockDsModel()
+                yield MockLayer()
+
+        monkeypatch.setattr(
+            "vllm.model_executor.warmup.deepseek_v4_mhc_warmup._warmup_layer_mhc",
+            _record_layer,
+        )
+        monkeypatch.setattr(
+            "vllm.model_executor.warmup.deepseek_v4_mhc_warmup._warmup_broadcast_mhc",
+            _record_broadcast,
+        )
+        monkeypatch.setattr(
+            "vllm.model_executor.warmup.deepseek_v4_mhc_warmup._warmup_hc_head",
+            _record_head,
+        )
+        monkeypatch.setattr(
+            "vllm.model_executor.warmup.deepseek_v4_mhc_warmup.torch.accelerator.synchronize",
+            lambda: calls.append(("synchronize",)),
+        )
+        monkeypatch.setattr(
+            "vllm.model_executor.warmup.deepseek_v4_mhc_warmup.logger.info",
+            lambda *a, **kw: None,
+        )
+
+        deepseek_v4_mhc_warmup(MockModel(), max_tokens=1024)
+
+        assert len(calls) == 4, f"Expected 4 calls, got {len(calls)}: {calls}"
+
+        s1_name, s1_sizes = calls[0]
+        assert s1_name == "_warmup_layer_mhc"
+        assert isinstance(s1_sizes, list) and len(s1_sizes) > 0
+        assert 1 in s1_sizes
+
+        s2_name, s2_model, s2_sizes = calls[1]
+        assert s2_name == "_warmup_broadcast_mhc"
+        assert s2_sizes == s1_sizes
+        assert s2_model.__class__.__name__ == "MockModel"
+
+        s3_name, s3_sizes = calls[2]
+        assert s3_name == "_warmup_hc_head"
+        assert s3_sizes == s1_sizes
+
+        assert calls[3] == ("synchronize",)
+
+
+# ── No sys.modules leakage ───────────────────────────────────────────────
+
+
+class TestNoSysModulesLeakage:
+    """Verify that ``_select_mhc_split_key_token_sizes`` does not leave
+    ``tilelang_kernels`` permanently cached in ``sys.modules`` after the
+    stub is restored by monkeypatch cleanup."""
+
+    MODULE_PATH = "vllm.model_executor.kernels.mhc.tilelang_kernels"
+
+    def test_clean_modules_after_monkeypatch_cleanup(self, monkeypatch) -> None:
+        # Ensure module is not already in sys.modules
+        monkeypatch.delitem(sys.modules, self.MODULE_PATH, raising=False)
+
+        # Patch and invoke
+        _patch_compute_split(n_sms=188, monkeypatch=monkeypatch)
+        reps = _select_mhc_split_key_token_sizes(max_tokens=8192, k_size=4096)
+        assert len(reps) == 16
+
+        # Monkeypatch cleanup restores original state: module removed if absent
+        monkeypatch.undo()
+        # After undo, the module should not be present (it wasn't before)
+        assert self.MODULE_PATH not in sys.modules, (
+            f"{self.MODULE_PATH} leaked into sys.modules"
+        )

vllm/model_executor/warmup/deepseek_v4_mhc_warmup.py

@@ -15,6 +15,7 @@ import torch
 
 from vllm.logger import init_logger
 from vllm.tracing import instrument
+from vllm.utils.math_utils import cdiv
 
 logger = init_logger(__name__)
 
@@ -61,6 +62,39 @@ def _select_mhc_warmup_token_sizes(
     return _normalize_token_sizes(candidates, max_tokens=max_auto_tokens)
 
 
+def _select_mhc_split_key_token_sizes(
+    *,
+    max_tokens: int,
+    k_size: int,
+) -> list[int]:
+    """Select one representative token count per distinct n_splits compile key.
+
+    The MHC TileLang kernels compute n_splits at runtime via
+    ``compute_num_split(block_k, k_size, cdiv(tokens, block_k))``.  Because
+    ``n_splits`` is a TileLang compile-time parameter, every distinct value
+    produces a separate JIT compilation artifact.  This function returns
+    exactly one token per reachable ``n_splits`` value, stopping early when
+    ``n_splits`` drops to 1 (all remaining grid sizes map to n_splits=1).
+    """
+    from vllm.model_executor.kernels.mhc.tilelang_kernels import compute_num_split
+
+    block_k = 64
+    max_grid = cdiv(max_tokens, block_k)
+    reps: list[int] = []
+    seen: set[int] = set()
+    for g in range(1, max_grid + 1):
+        t = (g - 1) * block_k + 1
+        if t > max_tokens:
+            break
+        ns = compute_num_split(block_k, k_size, g)
+        if ns not in seen:
+            reps.append(t)
+            seen.add(ns)
+            if ns == 1:
+                break
+    return reps
+
+
 def _find_first_mhc_layer(model: torch.nn.Module) -> torch.nn.Module | None:
     for module in model.modules():
         if module.__class__.__name__ != "DeepseekV4DecoderLayer":
@@ -102,6 +136,16 @@ def _warmup_layer_mhc(
     hidden_size = int(layer.hidden_size)
     hc_mult = int(layer.hc_mult)
     device = layer.hc_attn_fn.device
+
+    # Union general token sizes with split-key reps for the non-broadcast
+    # MHC kernel (k_size = hc_mult * hidden_size).
+    k_size = hc_mult * hidden_size
+    split_key_sizes = _select_mhc_split_key_token_sizes(
+        max_tokens=max_tokens, k_size=k_size
+    )
+    all_sizes = sorted(set(token_sizes) | set(split_key_sizes))
+    max_tokens = max(all_sizes)
+
     residual = torch.zeros(
         max_tokens,
         hc_mult,
@@ -110,19 +154,19 @@ def _warmup_layer_mhc(
         device=device,
     )
 
-    for size in token_sizes:
+    for size in all_sizes:
         residual_slice = residual[:size]
         for fn, scale, base in (
             (layer.hc_attn_fn, layer.hc_attn_scale, layer.hc_attn_base),
             (layer.hc_ffn_fn, layer.hc_ffn_scale, layer.hc_ffn_base),
         ):
-            layer_input, post_mix, comb_mix = layer.hc_pre(
+            layer_input, post_mix, res_mix = layer.hc_pre(
                 residual_slice,
                 fn,
                 scale,
                 base,
             )
-            layer.hc_post(layer_input, residual_slice, post_mix, comb_mix)
+            layer.hc_post(layer_input, residual_slice, post_mix, res_mix)
 
 
 def _warmup_hc_head(
@@ -161,6 +205,86 @@ def _warmup_hc_head(
         )
 
 
+def _warmup_broadcast_mhc(
+    model: torch.nn.Module,
+    token_sizes: list[int],
+) -> None:
+    """Warm up the first-layer broadcast MHC TileLang kernel.
+
+    The first ``DeepseekV4DecoderLayer`` uses
+    ``mhc_pre_broadcast_tilelang`` (2-D input, ``fn_broadcast`` weight)
+    instead of the 3-D ``mhc_pre_tilelang`` used by all subsequent layers.
+    ``fn_broadcast`` is set during ``finalize_mhc_broadcast_weights()`` and
+    only exists on the very first decoder layer.  No-op for models without a
+    broadcast-capable layer.
+
+    Unlike the generic per-layer MHC warmup (which covers power-of-two token
+    sizes), the broadcast kernel uses ``n_splits`` as a TileLang compile-time
+    parameter.  Different token counts can map to the same ``n_splits`` value;
+    this function selects one representative token per distinct compile key to
+    avoid redundant JIT compilations while covering every reachable key.
+    """
+    first_broadcast_layer = None
+    for module in model.modules():
+        if module.__class__.__name__ != "DeepseekV4DecoderLayer":
+            continue
+        fn_broadcast = getattr(module, "hc_attn_fn_broadcast", None)
+        if fn_broadcast is not None:
+            first_broadcast_layer = module
+            break
+    if first_broadcast_layer is None:
+        return
+
+    from vllm.model_executor.kernels.mhc.tilelang import mhc_pre_broadcast_tilelang
+
+    device = first_broadcast_layer.hc_attn_fn.device
+    if device.type != "cuda":
+        return
+
+    # Fail closed unless hc_attn_fn_broadcast is a tensor on the same CUDA
+    # device as hc_attn_fn.  A non-tensor, a CPU tensor, or a tensor on a
+    # different device index indicates a partially initialized or mismatched
+    # broadcast-weight setup that must not be warmed up.
+    fn_broadcast = first_broadcast_layer.hc_attn_fn_broadcast
+    if not isinstance(fn_broadcast, torch.Tensor):
+        return
+    if fn_broadcast.device.type != "cuda":
+        return
+    if fn_broadcast.device != device:
+        return
+
+    hidden_size = first_broadcast_layer.hidden_size
+    broadcast_token_sizes = _select_mhc_split_key_token_sizes(
+        max_tokens=max(token_sizes),
+        k_size=hidden_size,
+    )
+
+    x_2d = torch.zeros(
+        max(broadcast_token_sizes),
+        hidden_size,
+        dtype=torch.bfloat16,
+        device=device,
+    )
+    norm_weight = first_broadcast_layer.attn_norm.weight.data
+    norm_eps = first_broadcast_layer.attn_norm.variance_epsilon
+
+    for size in broadcast_token_sizes:
+        mhc_pre_broadcast_tilelang(
+            x_2d[:size],
+            first_broadcast_layer.hc_attn_fn,
+            first_broadcast_layer.hc_attn_scale,
+            first_broadcast_layer.hc_attn_base,
+            first_broadcast_layer.rms_norm_eps,
+            first_broadcast_layer.hc_eps,
+            first_broadcast_layer.hc_eps,
+            first_broadcast_layer.hc_post_alpha,
+            first_broadcast_layer.hc_sinkhorn_iters,
+            norm_weight=norm_weight,
+            norm_eps=norm_eps,
+            fn_broadcast=first_broadcast_layer.hc_attn_fn_broadcast,
+        )
+
+
 @instrument(span_name="DeepSeek V4 mHC warmup")
 def deepseek_v4_mhc_warmup(
     model: torch.nn.Module,
@@ -199,6 +323,7 @@ def deepseek_v4_mhc_warmup(
     )
     with torch.inference_mode():
         _warmup_layer_mhc(layer, token_sizes)
+        _warmup_broadcast_mhc(model, token_sizes)
         if deepseek_model is not None:
             _warmup_hc_head(deepseek_model, token_sizes)
         torch.accelerator.synchronize()

vllm/models/deepseek_v4/nvidia/model.py

@@ -865,6 +865,43 @@ class DeepseekV4DecoderLayer(nn.Module):
             requires_grad=False,
         )
 
+    def hc_pre(
+        self,
+        x: torch.Tensor,
+        hc_fn: torch.Tensor,
+        hc_scale: torch.Tensor,
+        hc_base: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Run standalone MHC pre (non-broadcast 3D path).
+
+        Used by kernel warmup and model-level forward.  Passes fused RMSNorm
+        weight so the TileLang JIT compiles the production kernel path.
+        """
+        post_mix, res_mix, layer_input = mhc_pre_tilelang(
+            residual=x,
+            fn=hc_fn,
+            hc_scale=hc_scale,
+            hc_base=hc_base,
+            rms_eps=self.rms_norm_eps,
+            hc_pre_eps=self.hc_eps,
+            hc_sinkhorn_eps=self.hc_eps,
+            hc_post_mult_value=self.hc_post_alpha,
+            sinkhorn_repeat=self.hc_sinkhorn_iters,
+            norm_weight=self.attn_norm.weight.data,
+            norm_eps=self.attn_norm.variance_epsilon,
+        )
+        return layer_input, post_mix, res_mix
+
+    def hc_post(
+        self,
+        x: torch.Tensor,
+        residual: torch.Tensor,
+        post: torch.Tensor,
+        comb: torch.Tensor,
+    ) -> torch.Tensor:
+        """Run standalone MHC post.  Used by kernel warmup."""
+        return mhc_post_tilelang(x, residual, post, comb)
+
     def forward(
         self,
         x: torch.Tensor,