rebase and change helper function name

Signed-off-by: qizixi <[email protected]>

Author: zixi-qi

vllm/v1/spec_decode/eagle.py

@@ -151,6 +151,11 @@ def propose(
         else:
             raise ValueError(f"Unsupported method: {self.method}")
 
+        # At this moment, we assume all eagle layers belong to the same KV
+        # cache group, thus using the same attention metadata.
+        per_layer_attn_metadata = {}
+        for layer_name in self.attn_layer_names:
+            per_layer_attn_metadata[layer_name] = attn_metadata
         if self.use_cuda_graph and \
             num_tokens <= self.cudagraph_batch_sizes[-1]:
             num_input_tokens = self.vllm_config.pad_for_cudagraph(num_tokens)
@@ -356,7 +361,8 @@ def dummy_run(
                 self.hidden_states[:num_tokens],
             )
 
-    def validate_kv_cache_group(self, kv_cache_config: KVCacheConfig) -> None:
+    def validate_same_kv_cache_group(self,
+                                     kv_cache_config: KVCacheConfig) -> None:
         """
         Validate that all eagle layers belong to the same KVCacheGroup.
         Need this assumption to ensure all eagle layers can use the

vllm/v1/worker/gpu_model_runner.py

@@ -2024,7 +2024,7 @@ def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
             assert isinstance(self.drafter, EagleProposer)
             # validate all draft model layers belong to the same kv cache
             # group
-            self.drafter.validate_kv_cache_group(kv_cache_config)
+            self.drafter.validate_same_kv_cache_group(kv_cache_config)
 
         bind_kv_cache(
             kv_caches,