[NPU] Initialize Ascend config for diffusion workers for v0.22.0 (#4654)

Signed-off-by: betta18 <jiangmengyu1@huawei.com>
Co-authored-by: betta18 <jiangmengyu1@huawei.com>

requirements/npu.txt

@@ -1,3 +1,3 @@
 -r common.txt
 onnxruntime-cann>=1.23.2
-torchaudio==2.9.0
+torchaudio==2.10.0

vllm_omni/diffusion/worker/diffusion_worker.py

@@ -239,6 +239,7 @@ class DiffusionWorker:
             "Final IR op priority after setting vLLM-Omni overrides: %s", vllm_config.kernel_config.ir_op_priority
         )
         self.vllm_config = vllm_config
+        current_omni_platform.init_diffusion_worker_vllm_config(vllm_config)
 
         # Initialize distributed environment
         with (

vllm_omni/platforms/interface.py

@@ -136,6 +136,11 @@ class OmniPlatform(Platform):
         """
         return "vllm_omni.diffusion.worker.diffusion_model_runner.DiffusionModelRunner"
 
+    @classmethod
+    def init_diffusion_worker_vllm_config(cls, vllm_config: Any) -> None:
+        """Initialize platform-specific state for diffusion worker VllmConfig."""
+        return None
+
     @classmethod
     def get_torch_device(cls, local_rank: int | None = None) -> torch.device:
         raise NotImplementedError

vllm_omni/platforms/npu/platform.py

@@ -55,6 +55,12 @@ class NPUOmniPlatform(OmniPlatform, NPUPlatform):
     def get_omni_generation_worker_cls(cls) -> str:
         return "vllm_omni.platforms.npu.worker.npu_generation_worker.NPUGenerationWorker"
 
+    @classmethod
+    def init_diffusion_worker_vllm_config(cls, vllm_config: Any) -> None:
+        from vllm_ascend.ascend_config import init_ascend_config
+
+        init_ascend_config(vllm_config)
+
     @classmethod
     def get_default_stage_config_path(cls) -> str:
         return "vllm_omni/platforms/npu/stage_configs"