[Spec] fix EAGLE v2 verify metadata init order on non-cuda-graph path (#26244)

python/sglang/srt/speculative/eagle_info_v2.py

@@ -306,11 +306,9 @@ class EagleVerifyInputV2Mixin:
         )
         if can_run_cuda_graph:
             target_worker.model_runner.graph_runner.replay_prepare(verify_forward_batch)
-        else:
+        # Non-cuda-graph: defer init to forward_extend, which runs after
-            if not batch.forward_mode.is_idle():
+        # `_forward_raw -> prepare_mlp_sync_batch` pads the batch. Initing
-                target_worker.model_runner.attn_backend.init_forward_metadata(
+        # here would use pre-pad shapes and trip DSv4 indexer shape match.
-                    verify_forward_batch
-                )
 
         return verify_forward_batch, can_run_cuda_graph
 

python/sglang/srt/speculative/eagle_worker_v2.py

@@ -1039,12 +1039,14 @@ class EAGLEWorkerV2(BaseSpecWorker):
                 verify_input.retrieve_next_token.shape
             ).cpu()
 
-        # Run target verify batch in the main compute stream (GPU compute)
+        # Run target verify batch in the main compute stream (GPU compute).
+        # Only skip metadata init when cuda-graph already ran replay_prepare;
+        # the non-cuda-graph path needs forward_extend's init (post-pad).
         forward_batch_output = self.target_worker.forward_batch_generation(
             batch=None,
             forward_batch=verify_forward_batch,
             is_verify=True,
-            skip_attn_backend_init=True,
+            skip_attn_backend_init=can_run_cuda_graph,
         )
         logits_output = forward_batch_output.logits_output