[Bugfix]: Fix online serving failure when using deploy config (#3537)

Signed-off-by: KexiongYu <yukexiong1@huawei.com>
Signed-off-by: Y. Fisher <yukexiong1@huawei.com>

vllm_omni/config/pipeline_registry.py

@@ -69,7 +69,7 @@ _OMNI_PIPELINES: dict[str, tuple[str, str]] = {
         "vllm_omni.model_executor.models.glm_image.pipeline",
         "GLM_IMAGE_PIPELINE",
     ),
-    "hunyuan_image3": (
+    "hunyuan_image_3_moe": (
         "vllm_omni.model_executor.models.hunyuan_image3.pipeline",
         "HUNYUAN_IMAGE3_PIPELINE",
     ),

vllm_omni/deploy/hunyuan_image3.yaml

@@ -1,7 +1,7 @@
 # HunyuanImage-3.0-Instruct deploy: AR (stage 0) + DiT (stage 1)
 # with AR-to-DiT KV reuse. The base CUDA layout uses 2 GPUs for AR
 # and 2 GPUs for DiT.
-pipeline: hunyuan_image3
+pipeline: hunyuan_image_3_moe
 async_chunk: false
 trust_remote_code: true
 
@@ -22,7 +22,6 @@ connectors:
 
 stages:
   - stage_id: 0
-    is_comprehension: false
     final_output: true
     final_output_type: text
     max_num_seqs: 1
@@ -38,7 +37,7 @@ stages:
     omni_kv_config:
       need_send_cache: true
     output_connectors:
-      to_stage_1: shared_memory_connector
+      to_stage_1: rdma_connector
     default_sampling_params:
       temperature: 0.0
       top_p: 1
@@ -58,7 +57,7 @@ stages:
       tensor_parallel_size: 2
       enable_expert_parallel: true
     input_connectors:
-      from_stage_0: shared_memory_connector
+      from_stage_0: rdma_connector
     default_sampling_params:
       num_inference_steps: 50
       guidance_scale: 0
@@ -83,7 +82,9 @@ platforms:
         max_num_batched_tokens: 8192
         parallel_config:
           tensor_parallel_size: 4
-          enable_expert_parallel: true
+          enable_expert_parallel: false
+          sequence_parallel_size: 1
+          ulysses_degree: 1
 
   xpu:
     stages:

vllm_omni/model_executor/models/hunyuan_image3/hunyuan_image3.py

@@ -1128,6 +1128,7 @@ def _hunyuan_image3_unpack_packed_topk(
     gating_output: torch.Tensor,
     topk: int,
     renormalize: bool,
+    num_experts: int,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     """Unpack pre-computed ``(topk_weights, topk_indices)`` packed by
     :class:`HunyuanImage3SparseMoeBlock` into ``gating_output``.

vllm_omni/model_executor/models/hunyuan_image3/pipeline.py

@@ -17,7 +17,7 @@ _HUNYUAN_IMAGE3_INPUT_PROCESSOR = "vllm_omni.model_executor.stage_input_processo
 
 
 HUNYUAN_IMAGE3_PIPELINE = PipelineConfig(
-    model_type="hunyuan_image3",
+    model_type="hunyuan_image_3_moe",
     model_arch=_HUNYUAN_IMAGE3_MODEL_ARCH,
     hf_architectures=_HUNYUAN_IMAGE3_HF_ARCHS,
     stages=(