[Bugfix] Align Hub kernel dependency guidance

Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>

docs/user_guide/diffusion/attention_backends/huggingface_hub.md

@@ -12,7 +12,7 @@ implementations to reduce numerical drift and sampling divergence.
 ## Installation
 
 ```bash
-pip install kernels==0.14.1
+pip install kernels==0.15.2
 ```
 
 If `kernels` is unavailable, an explicit `FLASH_ATTN_HUB` or

requirements/common.txt

@@ -13,8 +13,9 @@ transformers >= 5.10.1, < 5.15
 av>=14.0.0
 omegaconf>=2.3.0
 diffusers==0.40.0
-# Hub kernels, including LTX-2.5 NATTEN
+# Hub kernels, including LTX-2.5 NATTEN. Keep this pin in sync with the
-kernels==0.14.1
+# supported transformers range above.
+kernels==0.15.2
 safetensors>=0.8.0
 accelerate==1.12.0
 soundfile>=0.13.1

tests/diffusion/models/ltx2/test_ltx25_pipeline.py

@@ -386,7 +386,7 @@ def test_ltx25_natten_failure_has_actionable_omni_remedy(monkeypatch):
         MissingNattenProcessor,
     )
 
-    with pytest.raises(RuntimeError, match=r"kernels==0\.14\.1.*supported GPU.*allow Hub access"):
+    with pytest.raises(RuntimeError, match=r"kernels==0\.15\.2.*supported GPU.*allow Hub access"):
         ltx2_components._create_ltx25_natten_processor()
 
 

tests/model_executor/models/test_kv_cache_scale_mapper.py

@@ -67,11 +67,8 @@ class _FakeHunyuanModel(_FakeModel):
             use_cla=False,
         )
 
-    def get_expert_mapping(self) -> list[object]:
+    def get_expert_mapping(self) -> tuple[list[object], dict[str, object]]:
-        return []
+        return [], {}
-
-    def _get_expert_weights_remapping(self) -> dict[str, object]:
-        return {}
 
     def _split_qkv_weight(self, weight: torch.Tensor) -> torch.Tensor:
         return weight
@@ -84,49 +81,6 @@ def test_model_loaders_do_not_call_removed_get_cache_scale(rel_path: str) -> Non
     assert ".get_cache_scale(" not in source
 
 
-@pytest.mark.parametrize(
-    ("module_name", "class_name"),
-    [
-        (
-            "vllm_omni.model_executor.models.hunyuan_image3.hunyuan_image3",
-            "HunyuanModel",
-        ),
-        (
-            "vllm_omni.diffusion.models.hunyuan_image3.hunyuan_image3_transformer",
-            "HunyuanImage3Model",
-        ),
-    ],
-)
-def test_hunyuan_expert_mapping_separates_public_and_local_contracts(
-    module_name: str,
-    class_name: str,
-    monkeypatch: pytest.MonkeyPatch,
-) -> None:
-    """The public expert mapping must be a flat list, not a loader-side pair."""
-    module = __import__(module_name, fromlist=[class_name])
-    expected_mapping = [("param", "weight", 0, "gate_proj")]
-    fake_model = SimpleNamespace(
-        config=SimpleNamespace(num_experts=1),
-        num_redundant_experts=0,
-    )
-    monkeypatch.setattr(module, "_is_moe", lambda config: True)
-    if hasattr(module, "fused_moe_make_expert_params_mapping"):
-        monkeypatch.setattr(module, "fused_moe_make_expert_params_mapping", lambda **kwargs: expected_mapping)
-    else:
-        monkeypatch.setattr(
-            module,
-            "FusedMoE",
-            SimpleNamespace(make_expert_params_mapping=lambda *args, **kwargs: expected_mapping),
-        )
-
-    model_cls = getattr(module, class_name)
-    assert model_cls.get_expert_mapping(fake_model) == expected_mapping
-    assert model_cls._get_expert_weights_remapping(fake_model) == {
-        "gate_proj": ("gate_and_up_proj", 1, 2),
-        "up_proj": ("gate_and_up_proj", 0, 2),
-    }
-
-
 @pytest.mark.parametrize(
     ("module_name", "class_name", "fake_model_cls"),
     [

vllm_omni/diffusion/models/hunyuan_image3/hunyuan_image3_transformer.py

@@ -2190,28 +2190,24 @@ class HunyuanImage3Model(nn.Module):
         return torch.concat((q, k, v))
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
-        if not _is_moe(self.config):
+        if _is_moe(self.config):
-            return []
+            # Params for weights, fp8 weight scales, fp8 activation scales
-
+            # (param_name, weight_name, expert_id, shard_id)
-        # Params for weights, fp8 weight scales, fp8 activation scales
+            fused_moe_expert_mapping = FusedMoE.make_expert_params_mapping(
-        # (param_name, weight_name, expert_id, shard_id)
+                self,
-        return FusedMoE.make_expert_params_mapping(
+                ckpt_gate_proj_name="gate_proj",
-            self,
+                ckpt_down_proj_name="down_proj",
-            ckpt_gate_proj_name="gate_proj",
+                ckpt_up_proj_name="up_proj",
-            ckpt_down_proj_name="down_proj",
+                num_experts=self.config.num_experts,
-            ckpt_up_proj_name="up_proj",
+                num_redundant_experts=self.num_redundant_experts,
-            num_experts=self.config.num_experts,
+            )
-            num_redundant_experts=self.num_redundant_experts,
+            expert_weights_remapping = {
-        )
+                "gate_proj": ("gate_and_up_proj", 1, 2),
-
+                "up_proj": ("gate_and_up_proj", 0, 2),
-    def _get_expert_weights_remapping(self) -> dict[str, tuple[str, int, int]]:
+            }
-        """Return the vLLM-Omni-local remapping for Hunyuan checkpoint weights."""
+            return fused_moe_expert_mapping, expert_weights_remapping
-        if not _is_moe(self.config):
+        else:
-            return {}
+            return [], {}
-        return {
-            "gate_proj": ("gate_and_up_proj", 1, 2),
-            "up_proj": ("gate_and_up_proj", 0, 2),
-        }
 
     # rename for delay load
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
@@ -2247,8 +2243,7 @@ class HunyuanImage3Model(nn.Module):
 
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
-        expert_params_mapping = self.get_expert_mapping()
+        expert_params_mapping, expert_weights_remapping = self.get_expert_mapping()
-        expert_weights_remapping = self._get_expert_weights_remapping()
 
         # List of unexpected keywords in weight names
         unexpected_keywords = [

vllm_omni/diffusion/models/ltx2/ltx2_components.py

@@ -298,7 +298,7 @@ def _create_ltx25_natten_processor() -> LTX2VideoVaeNeighborhoodNattenProcessor:
     except (ImportError, OSError, RuntimeError, ValueError) as exc:
         raise RuntimeError(
             "LTX-2.5 DiffVAE requires the shi-labs/natten Hub kernel. "
-            "Install kernels==0.14.1, use a supported GPU, leave "
+            "Install kernels==0.15.2, use a supported GPU, leave "
             "DIFFUSERS_DISABLE_REMOTE_CODE unset, and allow Hub access during kernel initialization."
         ) from exc
 

vllm_omni/model_executor/models/hunyuan_image3/hunyuan_image3.py

@@ -128,29 +128,25 @@ class HunyuanModel(HunYuanModel):
         v = v.reshape(-1, hidden_size)
         return torch.concat((q, k, v))
 
-    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
+    def get_expert_mapping(self) -> tuple[list[tuple[str, str, int, str]], dict[str, tuple[str, int, int]]]:
-        if not _is_moe(self.config):
+        if _is_moe(self.config):
-            return []
+            # Params for weights, fp8 weight scales, fp8 activation scales
-
+            # (param_name, weight_name, expert_id, shard_id)
-        # Params for weights, fp8 weight scales, fp8 activation scales
+            fused_moe_expert_mapping = fused_moe_make_expert_params_mapping(
-        # (param_name, weight_name, expert_id, shard_id)
+                model=self,
-        return fused_moe_make_expert_params_mapping(
+                ckpt_gate_proj_name="gate_proj",
-            model=self,
+                ckpt_down_proj_name="down_proj",
-            ckpt_gate_proj_name="gate_proj",
+                ckpt_up_proj_name="up_proj",
-            ckpt_down_proj_name="down_proj",
+                num_experts=self.config.num_experts,
-            ckpt_up_proj_name="up_proj",
+                num_redundant_experts=self.num_redundant_experts,
-            num_experts=self.config.num_experts,
+            )
-            num_redundant_experts=self.num_redundant_experts,
+            expert_weights_remapping = {
-        )
+                "gate_proj": ("gate_and_up_proj", 1, 2),
-
+                "up_proj": ("gate_and_up_proj", 0, 2),
-    def _get_expert_weights_remapping(self) -> dict[str, tuple[str, int, int]]:
+            }
-        """Return the vLLM-Omni-local remapping for Hunyuan checkpoint weights."""
+            return fused_moe_expert_mapping, expert_weights_remapping
-        if not _is_moe(self.config):
+        else:
-            return {}
+            return [], {}
-        return {
-            "gate_proj": ("gate_and_up_proj", 1, 2),
-            "up_proj": ("gate_and_up_proj", 0, 2),
-        }
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
         cla_factor = _get_cla_factor(self.config)
@@ -178,8 +174,7 @@ class HunyuanModel(HunYuanModel):
 
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
-        expert_params_mapping = self.get_expert_mapping()
+        expert_params_mapping, expert_weights_remapping = self.get_expert_mapping()
-        expert_weights_remapping = self._get_expert_weights_remapping()
 
         # List of unexpected keywords in weight names
         unexpected_keywords = [