Merge branch 'v0.9.2-dev-wm' into 'v0.9.2-dev'

[fix]修复v1 mtp接受率低的问题

See merge request dcutoolkit/deeplearing/vllm!174

vllm/config.py

@@ -4106,7 +4106,7 @@ class CompilationConfig:
     are always used, it can set this to False. Otherwise, it should
     set this to True, and the compiler will copy the input to an
     internally managed buffer. Default is False."""
-    full_cuda_graph: bool = True
+    full_cuda_graph: bool = False
     """whether to use a full cuda graph for the entire forward pass rather than
     splitting certain operations such as attention into subgraphs. Thus this
     flag cannot be used together with splitting_ops. This may provide
@@ -4948,4 +4948,4 @@ def get_layers_from_vllm_config(vllm_config: VllmConfig,
         for layer_name, layer in
         vllm_config.compilation_config.static_forward_context.items()
         if isinstance(layer, layer_type)
-    }
\ No newline at end of file
+    }

vllm/v1/attention/backends/mla/common.py

@@ -546,11 +546,14 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
         device = self.runner.device
         block_table = self.block_table
         block_table_tensor = block_table.get_device_tensor()[:num_reqs]
-        block_table.slot_mapping[:num_actual_tokens].copy_(
-            block_table.slot_mapping_cpu[:num_actual_tokens],
-            non_blocking=True)
-        block_table.slot_mapping[num_actual_tokens:].fill_(-1)
-        slot_mapping = block_table.slot_mapping[:num_actual_tokens]
+
+        slot_mapping = common_attn_metadata.slot_mapping
+        if slot_mapping is None:
+            block_table.slot_mapping[:num_actual_tokens].copy_(
+                block_table.slot_mapping_cpu[:num_actual_tokens],
+                non_blocking=True)
+            block_table.slot_mapping[num_actual_tokens:].fill_(-1)
+            slot_mapping = block_table.slot_mapping[:num_actual_tokens]
 
         query_start_loc = common_attn_metadata.query_start_loc
         seq_lens = common_attn_metadata.seq_lens

vllm/v1/attention/backends/utils.py

@@ -45,6 +45,8 @@ class CommonAttentionMetadata:
     """(batch_size,), record the rejected tokens number in cpu and gpu"""
     num_speculative_tokens: int = 0
     """Number of speculative tokens"""
+    slot_mapping: torch.Tensor = None
+    """(batch_size, seq_len), slot mapping"""
 
 
 M = TypeVar("M")

vllm/v1/spec_decode/eagle.py

@@ -159,6 +159,7 @@ class EagleProposer:
                 num_actual_tokens=num_tokens,
                 max_query_len=max_query_len,
                 num_rejected_tokens=num_rejected_tokens,
+                slot_mapping=target_slot_mapping
             )
 
             assert self.runner is not None