Merge branch 'v0.15.1-dev_yql_3.9' into 'v0.15.1-dev'

dpsk_v32的mtp层的dense加载适配

See merge request dcutoolkit/deeplearing/vllm!479

vllm/model_executor/models/deepseek_mtp.py

@@ -78,7 +78,9 @@ class DeepSeekMultiTokenPredictorLayer(nn.Module):
 
         self.device = current_platform.device_type
 
-        self.is_v32 = hasattr(config, "index_topk")
+        #添加判断，默认开启DSA
+        force_disable_dsa = os.environ.get("VLLM_DISABLE_DSA", "0") == "1"
+        self.is_v32 = hasattr(config, "index_topk") and not force_disable_dsa
         if self.is_v32:
             topk_tokens = config.index_topk
             topk_indices_buffer = torch.empty(
@@ -282,9 +284,18 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
 
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
+
+        # 判断是否加载"indexer"权重
+        model_has_indexer = any("indexer" in param_name for param_name in params_dict.keys())
+
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
+
+            #  跳过加载"indexer"权重
+            if "indexer" in name and not model_has_indexer:
+                logger.info(f"Skipping indexer weight (DSA disabled): {name}")
+                continue
             spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
             if spec_layer is None:
                 continue

vllm/v1/attention/ops/flashmla.py

@@ -136,7 +136,6 @@ def get_mla_metadata_dense_fp8(
             cache_seqlens,
             num_q_tokens_per_head_k,
             num_heads_k, 
-            # 16,
         )
     else:
         return torch.ops._flashmla_extension_C.get_mla_decoding_metadata_dense_fp8(