[Misc]feat: adapt to vLLM main (54503ece)

.github/vllm-main-verified.commit

@@ -1 +1 @@
-85c09e9885e346ea1612da30ebff5a75f67d2350
+54503ecec0f3ac31e5ecfc5f28652e4cc42307b5

tests/e2e/conftest.py

@@ -1924,7 +1924,7 @@ def qwen_prompt(questions: list[str]) -> list[str]:
 
 
 def hunyuan_prompt(questions: list[str]) -> list[str]:
-    placeholder = "<｜hy_place▁holder▁no▁100｜><｜hy_place▁holder▁no▁102｜><｜hy_place▁holder▁no▁101｜>"  # noqa: E501
+    placeholder = "<｜hy_place▁holder▁no▁102｜>"  # noqa: E501
     return [f"<｜hy_begin▁of▁sentence｜>{placeholder}{question}<｜hy_User｜>" for question in questions]
 
 

vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py

@@ -1137,7 +1137,8 @@ def _get_group_kv_caches(self, group_idx: int, layer_indices: list[int] | None =
         if layer_indices is None:
             _, layer_indices = self.kv_group2layeridx[group_idx]
         layer_index_set = set(layer_indices)
-        num_attn_module = 2 if self.vllm_config.model_config.hf_text_config.model_type == "longcat_flash" else 1
+        model_type = self.vllm_config.model_config.hf_text_config.model_type
+        num_attn_module = 2 if model_type in ("longcat_flash", "longcat_flash_ngram") else 1
         from vllm.v1.worker.utils import extract_layer_index
 
         def layer_in_group(layer_name: str) -> bool:
@@ -2125,7 +2126,8 @@ def _build_kv_group2layeridx(self) -> dict[int, tuple[dict[str, Any], list[int]]
         from vllm.v1.worker.utils import extract_layer_index
 
         kv_group2layeridx: dict[int, tuple[dict[str, Any], list[int]]] = {}
-        num_attn_module = 2 if self.vllm_config.model_config.hf_text_config.model_type == "longcat_flash" else 1
+        model_type = self.vllm_config.model_config.hf_text_config.model_type
+        num_attn_module = 2 if model_type in ("longcat_flash", "longcat_flash_ngram") else 1
         next_mtp_layer_idx = self.total_layers
         transfer_group_id = 0
         for kv_cache_group_id, group_spec in enumerate(self.kv_cache_config.kv_cache_groups):

vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_layerwise_connector.py

@@ -1336,7 +1336,8 @@ def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
         if use_kv_buffer:
             self.create_kv_buffer(kv_buffer)
 
-        num_attn_module = 2 if self.vllm_config.model_config.hf_text_config.model_type == "longcat_flash" else 1
+        model_type = self.vllm_config.model_config.hf_text_config.model_type
+        num_attn_module = 2 if model_type in ("longcat_flash", "longcat_flash_ngram") else 1
         mtp_layer_name = ""
         for layer_name in kv_caches:
             if "mtp" in layer_name:

vllm_ascend/patch/hunyuan_vl_processor_compat.py

@@ -148,8 +148,7 @@ def call_hf_processor(
 
 def install_hunyuan_vl_processor_compat() -> None:
     """Align both supported vLLM refs with Transformers 5.13 Hunyuan APIs."""
-    if not _remove_stale_registry_entries():
-        return
+    _remove_stale_registry_entries()
     from vllm.model_executor.models import hunyuan_vision as main_hunyuan_vision
 
     _patch_hunyuan_processor_loader(main_hunyuan_vision)

vllm_ascend/patch/platform/patch_speculative_config.py

@@ -111,7 +111,7 @@ def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
                 "architectures": ["Qwen3_5MoeMTP" if is_moe else "Qwen3_5MTP"],
             }
         )
-    if hf_config.model_type == "longcat_flash":
+    if hf_config.model_type in ("longcat_flash", "longcat_flash_ngram"):
         hf_config.model_type = "longcat_flash_mtp"
         n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
         hf_config.update({"n_predict": n_predict, "architectures": ["LongCatFlashMTPModel"]})

vllm_ascend/worker/model_runner_v1.py

@@ -3971,7 +3971,8 @@ def initialize_kv_cache_tensors(self, kv_cache_config: KVCacheConfig) -> dict[st
         else:
             from vllm.v1.worker.utils import bind_kv_cache
 
-            num_attn_module = 2 if self.model_config.hf_text_config.model_type == "longcat_flash" else 1
+            model_type = self.model_config.hf_text_config.model_type
+            num_attn_module = 2 if model_type in ("longcat_flash", "longcat_flash_ngram") else 1
             bind_kv_cache(
                 kv_caches,
                 self.compilation_config.static_forward_context,