[BUGFIX] 修复 Qwen3.5 在新版 transformers 下的配置兼容问题并统一 ROCm unified attention 路由

目的：
  修复 Qwen3.5 / Qwen3.5-MoE 在升级 transformers 后的配置解析兼容问题，并优化 ROCm 下 unified attention 的路由策略，避免prefill 和 decode落到不同实现路径上，降低后续排查和行为不一致的成本

vllm/transformers_utils/config.py

@@ -5,6 +5,7 @@ import os
 from collections.abc import Callable
 from dataclasses import asdict
 from functools import cache, partial
+from importlib.util import find_spec
 from importlib.metadata import version
 from pathlib import Path
 from typing import Any, Literal, TypeAlias
@@ -44,10 +45,14 @@ try:
     # Transformers v5
     from transformers.configuration_utils import ALLOWED_ATTENTION_LAYER_TYPES
 except ImportError:
-    # Transformers v4
-    from transformers.configuration_utils import (
-        ALLOWED_LAYER_TYPES as ALLOWED_ATTENTION_LAYER_TYPES,
-    )
+    try:
+        # Transformers v4.52+
+        from transformers.configuration_utils import (
+            ALLOWED_LAYER_TYPES as ALLOWED_ATTENTION_LAYER_TYPES,
+        )
+    except ImportError:
+        # Transformers v4.51 and earlier: neither symbol exists, use empty set
+        ALLOWED_ATTENTION_LAYER_TYPES: set = set()
 
 
 if envs.VLLM_USE_MODELSCOPE:
@@ -60,6 +65,18 @@ MISTRAL_CONFIG_NAME = "params.json"
 logger = init_logger(__name__)
 
 
+def _hf_has_native_qwen3_5_config(model_type: str | None) -> bool:
+    if model_type == "qwen3_5":
+        return find_spec(
+            "transformers.models.qwen3_5.configuration_qwen3_5"
+        ) is not None
+    if model_type == "qwen3_5_moe":
+        return find_spec(
+            "transformers.models.qwen3_5_moe.configuration_qwen3_5_moe"
+        ) is not None
+    return False
+
+
 class LazyConfigDict(dict):
     def __getitem__(self, key):
         if isinstance(value := super().__getitem__(key), type):
@@ -99,6 +116,8 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
     step3p5="Step3p5Config",
     qwen3_asr="Qwen3ASRConfig",
     qwen3_next="Qwen3NextConfig",
+    qwen3_5="Qwen3_5Config",
+    qwen3_5_moe="Qwen3_5MoeConfig",
     lfm2_moe="Lfm2MoeConfig",
     tarsier2="Tarsier2Config",
 )
@@ -151,7 +170,9 @@ class HFConfigParser(ConfigParserBase):
         if (hf_overrides := kwargs.pop("hf_overrides", None)) is not None:
             model_type = hf_overrides.get("model_type", model_type)
 
-        if model_type in _CONFIG_REGISTRY:
+        if model_type in _CONFIG_REGISTRY and not _hf_has_native_qwen3_5_config(
+            model_type
+        ):
             config_class = _CONFIG_REGISTRY[model_type]
             config = config_class.from_pretrained(
                 model,

vllm/transformers_utils/configs/qwen3_5.py

@@ -68,10 +68,10 @@ class Qwen3_5TextConfig(PretrainedConfig):
         eos_token_id=None,
         **kwargs,
     ):
-        kwargs["ignore_keys_at_rope_validation"] = [
+        kwargs["ignore_keys_at_rope_validation"] = {
             "mrope_section",
             "mrope_interleaved",
-        ]
+        }
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size

vllm/transformers_utils/configs/qwen3_5_moe.py

@@ -75,10 +75,10 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):
         eos_token_id=None,
         **kwargs,
     ):
-        kwargs["ignore_keys_at_rope_validation"] = [
+        kwargs["ignore_keys_at_rope_validation"] = {
             "mrope_section",
             "mrope_interleaved",
-        ]
+        }
         self.vocab_size = vocab_size
         self.max_position_embeddings = max_position_embeddings
         self.hidden_size = hidden_size

vllm/v1/attention/ops/triton_unified_attention.py

@@ -973,11 +973,24 @@ def unified_attention(
         is_prefill=False,
     )
 
+    # On ROCm, prefer the unified FA kernel whenever the cache layout and
+    # headdim match its supported path. This keeps both prefill and decode on
+    # the same implementation instead of routing decode-only batches to the 3D
+    # Triton kernel.
+    use_fa_unified_2d = (
+        current_platform.is_rocm()
+        and varlen_fwd_unified is not None
+        and block_size % 64 == 0
+        and head_size == 256
+    )
+
     # Launch the 2D kernel if
     # 1. No intermediate tiled softmax buffers for the 3D kernel have been allocated, or
     # 2. The batch includes at least one prefill request, or
     # 3. The number of sequences exceeds the configured threshold
     if (
+        use_fa_unified_2d
+        or
         seq_threshold_3D is None
         or num_par_softmax_segments is None
         or softmax_segm_output is None
@@ -987,12 +1000,6 @@ def unified_attention(
         or num_seqs > seq_threshold_3D
     ):
         # print(f"[2D Triton] k shape: {k.shape}, v shape: {v.shape}")
-        use_fa_unified_2d = (
-            current_platform.is_rocm()
-            and varlen_fwd_unified is not None
-            and block_size % 64 == 0
-            and head_size == 256
-        )
         if not use_fa_unified_2d:
             # print("Running Triton kernel")
             kernel_unified_attention_2d[