[Model] Extend NPU support for HunyuanImage3 Diffusion Model (#1689)

Signed-off-by: ElleElleWu <1608928702@qq.com>
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
Co-authored-by: skf1999 <13234016272@163.com>
Co-authored-by: Just-it <1161406585@qq.com>
Co-authored-by: Semmer2 <semmer@live.cn>
Co-authored-by: gcanlin <canlinguosdu@gmail.com>

tests/diffusion/models/hunyuan_image_3/test_hunyuan_fused_moe.py

@@ -0,0 +1,112 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for HunyuanFusedMoE (Support HunyuanImage3 Diffusion Model, 5a779b4)."""
+
+import pytest
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+class TestHunyuanFusedMoEPlatformDispatch:
+    """Test platform dispatch via platform qualname hooks."""
+
+    def test_default_platform_uses_default_impl_qualname(self, mocker):
+        """HunyuanFusedMoE should resolve the impl class from the platform hook."""
+        import vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe as hunyuan_moe
+
+        mock_platform = mocker.MagicMock()
+        mock_platform.get_diffusion_model_impl_qualname.return_value = (
+            "vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe.HunyuanFusedMoEDefault"
+        )
+
+        mocker.patch.object(
+            hunyuan_moe,
+            "current_omni_platform",
+            mock_platform,
+        )
+        mock_resolve = mocker.patch.object(hunyuan_moe, "resolve_obj_by_qualname")
+        mock_impl = mocker.MagicMock()
+        mock_resolve.return_value = mock_impl
+
+        from vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe import (
+            HunyuanFusedMoE,
+        )
+
+        HunyuanFusedMoE(prefix="")
+
+        mock_platform.prepare_diffusion_op_runtime.assert_called_once_with("hunyuan_fused_moe")
+        mock_platform.get_diffusion_model_impl_qualname.assert_called_once_with("hunyuan_fused_moe")
+        mock_resolve.assert_called_once_with(
+            "vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe.HunyuanFusedMoEDefault"
+        )
+        mock_impl.assert_called_once_with(prefix="")
+
+
+class TestHunyuanFusedMoEFactory:
+    """Test HunyuanFusedMoE factory __new__ and make_expert_params_mapping delegation."""
+
+    def test_new_delegates_to_impl_class(self, mocker):
+        """HunyuanFusedMoE(prefix=..., **kwargs) should instantiate and return impl instance."""
+        import vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe as hunyuan_moe
+
+        class MockImpl:
+            def __init__(self, *, prefix: str = "", **kwargs):
+                self.prefix = prefix
+                self.kwargs = kwargs
+
+        mock_platform = mocker.MagicMock()
+        mock_platform.get_diffusion_model_impl_qualname.return_value = "mock.impl.Qualname"
+        mocker.patch.object(hunyuan_moe, "current_omni_platform", mock_platform)
+
+        mock_impl_class = mocker.MagicMock(return_value=MockImpl(prefix="test", a=1))
+        mocker.patch.object(hunyuan_moe, "resolve_obj_by_qualname", return_value=mock_impl_class)
+
+        from vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe import (
+            HunyuanFusedMoE,
+        )
+
+        result = HunyuanFusedMoE(prefix="test", a=1)
+
+        assert isinstance(result, MockImpl)
+        assert result.prefix == "test"
+        assert result.kwargs == {"a": 1}
+        mock_platform.prepare_diffusion_op_runtime.assert_called_once_with("hunyuan_fused_moe")
+        mock_platform.get_diffusion_model_impl_qualname.assert_called_once_with("hunyuan_fused_moe")
+        mock_impl_class.assert_called_once_with(prefix="test", a=1)
+
+    def test_make_expert_params_mapping_delegates_to_impl(self, mocker):
+        """make_expert_params_mapping should delegate to impl class method."""
+        import vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe as hunyuan_moe
+
+        expected_mapping = [("a", "b", 0, "c")]
+        mock_platform = mocker.MagicMock()
+        mock_platform.get_diffusion_model_impl_qualname.return_value = "mock.impl.Qualname"
+        mocker.patch.object(hunyuan_moe, "current_omni_platform", mock_platform)
+
+        mock_impl_class = mocker.MagicMock()
+        mock_impl_class.make_expert_params_mapping = mocker.MagicMock(return_value=expected_mapping)
+        mocker.patch.object(hunyuan_moe, "resolve_obj_by_qualname", return_value=mock_impl_class)
+
+        from vllm_omni.diffusion.models.hunyuan_image_3.hunyuan_fused_moe import (
+            HunyuanFusedMoE,
+        )
+
+        result = HunyuanFusedMoE.make_expert_params_mapping(
+            model=None,
+            ckpt_gate_proj_name="gate",
+            ckpt_down_proj_name="down",
+            ckpt_up_proj_name="up",
+            num_experts=4,
+            num_redundant_experts=0,
+        )
+
+        assert result == expected_mapping
+        mock_platform.get_diffusion_model_impl_qualname.assert_called_once_with("hunyuan_fused_moe")
+        mock_impl_class.make_expert_params_mapping.assert_called_once_with(
+            None,
+            ckpt_gate_proj_name="gate",
+            ckpt_down_proj_name="down",
+            ckpt_up_proj_name="up",
+            num_experts=4,
+            num_redundant_experts=0,
+        )

tests/diffusion/test_data_is_moe.py

@@ -0,0 +1,78 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Unit tests for OmniDiffusionConfig.is_moe (fix is_moe type and threshold, 6663c0b)."""
+
+import pytest
+
+from vllm_omni.diffusion.data import OmniDiffusionConfig, TransformerConfig
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+class TestOmniDiffusionConfigIsMoE:
+    """Tests for OmniDiffusionConfig.is_moe property.
+
+    Covers commit 6663c0b: fix is_moe type and threshold
+    - num_experts must be (list, tuple, int); otherwise return False.
+    - Threshold: is_moe is True when num_experts > 0 (not > 1).
+    """
+
+    def test_is_moe_missing_num_experts_returns_false(self):
+        """When num_experts is absent, is_moe should be False."""
+        tf_config = TransformerConfig.from_dict({})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is False
+
+    def test_is_moe_none_num_experts_returns_false(self):
+        """When num_experts is explicitly None (e.g. in params), is_moe should be False."""
+        tf_config = TransformerConfig.from_dict({"num_experts": None})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is False
+
+    def test_is_moe_non_allowed_type_returns_false(self):
+        """When num_experts is not int/list/tuple (e.g. str), is_moe should be False."""
+        tf_config = TransformerConfig.from_dict({"num_experts": "2"})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is False
+
+    def test_is_moe_int_zero_returns_false(self):
+        """num_experts int 0 should yield is_moe False (threshold > 0)."""
+        tf_config = TransformerConfig.from_dict({"num_experts": 0})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is False
+
+    def test_is_moe_int_one_returns_true(self):
+        """num_experts int 1 should yield is_moe True (threshold > 0, not > 1)."""
+        tf_config = TransformerConfig.from_dict({"num_experts": 1})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is True
+
+    def test_is_moe_int_gt_one_returns_true(self):
+        """num_experts int > 1 should yield is_moe True."""
+        tf_config = TransformerConfig.from_dict({"num_experts": 2})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is True
+
+    def test_is_moe_list_all_zero_returns_false(self):
+        """num_experts list with all <= 0 should yield is_moe False."""
+        tf_config = TransformerConfig.from_dict({"num_experts": [0]})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is False
+
+    def test_is_moe_list_has_positive_returns_true(self):
+        """num_experts list with any int > 0 should yield is_moe True."""
+        tf_config = TransformerConfig.from_dict({"num_experts": [0, 1]})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is True
+
+    def test_is_moe_tuple_has_positive_returns_true(self):
+        """num_experts tuple with any int > 0 should yield is_moe True."""
+        tf_config = TransformerConfig.from_dict({"num_experts": (0, 2)})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is True
+
+    def test_is_moe_list_non_int_ignored(self):
+        """num_experts list with only non-int entries should yield is_moe False."""
+        tf_config = TransformerConfig.from_dict({"num_experts": ["a", 0.0]})
+        config = OmniDiffusionConfig(model="test", tf_model_config=tf_config)
+        assert config.is_moe is False