Merge tag 'v0.15.1' into v0.15.1-dev

vllm/v1/attention/backends/flash_attn.py

@@ -315,18 +315,6 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
         vllm_config: "VllmConfig",
         kv_cache_spec: "AttentionSpec",
     ) -> AttentionCGSupport:
-        # FA2 does not support CUDA graphs with encoder-decoder models due to
-        # accuracy issues reported in https://github.com/vllm-project/vllm/issues/33091
-        if (
-            vllm_config.model_config.is_encoder_decoder
-            and get_flash_attn_version() == 2
-        ):
-            logger.warning_once(
-                "FlashAttention2 does not support CUDA graphs with "
-                "encoder-decoder models due to accuracy issues reported in #33091. "
-                "Disabling CUDA graph."
-            )
-            return AttentionCGSupport.NEVER
         return cls._cudagraph_support
 
     def __init__(

vllm/v1/core/kv_cache_coordinator.py

@@ -479,6 +479,16 @@ class HybridKVCacheCoordinator(KVCacheCoordinator):
         hit_length = max_cache_hit_length
         hit_blocks_by_group: list[list[KVCacheBlock] | None] = [None] * num_groups
 
+        # Simple hybrid (1 full attn + 1 other): one iteration suffices.
+        # Full attn is always first if it exists. This avoids EAGLE drops
+        # being applied multiple times to non-full-attn groups.
+        # FIXME (yifan): However, for complex hybrid models with multiple attn
+        # groups, we still have the EAGLE spiral block dropping problem. See
+        # discussion in issue https://github.com/vllm-project/vllm/issues/32802.
+        is_simple_hybrid = len(self.attention_groups) == 2 and isinstance(
+            self.attention_groups[0][0], FullAttentionSpec
+        )
+
         while True:
             curr_hit_length = hit_length
 
@@ -495,10 +505,6 @@ class HybridKVCacheCoordinator(KVCacheCoordinator):
                     # the last iteration.
                     num_blocks = curr_hit_length // spec.block_size
                     curr_hit_length = num_blocks * spec.block_size
-                    for group_id in group_ids:
-                        blocks = hit_blocks_by_group[group_id]
-                        assert blocks is not None
-                        del blocks[num_blocks:]
                 else:
                     hit_blocks = manager_cls.find_longest_cache_hit(
                         block_hashes=_get_block_hashes(spec),
@@ -513,11 +519,21 @@ class HybridKVCacheCoordinator(KVCacheCoordinator):
                     for group_id, blocks in zip(group_ids, hit_blocks):
                         hit_blocks_by_group[group_id] = blocks
 
-            if curr_hit_length < hit_length:
+            if curr_hit_length >= hit_length:
+                break
             hit_length = curr_hit_length
-            else:
+            # Simple hybrid: exit after one iteration
+            if is_simple_hybrid:
                 break
 
+        # Truncate full attention blocks to final hit_length (if present)
+        spec, group_ids, _ = self.attention_groups[0]
+        if isinstance(spec, FullAttentionSpec):
+            num_blocks = hit_length // spec.block_size
+            for group_id in group_ids:
+                if (blks := hit_blocks_by_group[group_id]) is not None:
+                    del blks[num_blocks:]
+
         return tuple(
             blocks if blocks is not None else [] for blocks in hit_blocks_by_group
         ), hit_length

vllm/v1/core/sched/scheduler.py

@@ -1284,7 +1284,7 @@ class Scheduler(SchedulerInterface):
             scheduled_spec_token_ids = (
                 scheduler_output.scheduled_spec_decode_tokens.get(req_id)
             )
-            if scheduled_spec_token_ids:
+            if scheduled_spec_token_ids and generated_token_ids:
                 num_draft_tokens = len(scheduled_spec_token_ids)
                 num_accepted = len(generated_token_ids) - 1
                 num_rejected = num_draft_tokens - num_accepted

vllm/v1/worker/gpu_model_runner.py

@@ -1382,12 +1382,14 @@ class GPUModelRunner(
         num_scheduled_tokens: dict[str, int],
         kv_cache_spec: KVCacheSpec,
         num_reqs: int,
+        for_cudagraph_capture: bool = False,
     ) -> tuple[torch.Tensor | None, np.ndarray | None]:
         if not isinstance(kv_cache_spec, CrossAttentionSpec):
             return None, None
 
         # Zero out buffer for padding requests that are not actually scheduled (CGs)
         self.encoder_seq_lens.np[:num_reqs] = 0
+
         # Build encoder_seq_lens array mapping request indices to
         # encoder lengths for inputs scheduled in this batch
         for req_id in num_scheduled_tokens:
@@ -1404,6 +1406,15 @@ class GPUModelRunner(
                 feature.mm_position.length for feature in req_state.mm_features
             )
             self.encoder_seq_lens.np[req_index] = encoder_input_tokens
+        if for_cudagraph_capture:
+            # During CUDA graph capture, we need to use realistic encoder lengths
+            # so that max_seqlen_k is captured with the correct value.
+            max_encoder_len = getattr(
+                self.model_config.hf_config,
+                "max_source_positions",
+                self.max_encoder_len,
+            )
+            self.encoder_seq_lens.np[:num_reqs] = max_encoder_len
 
         self.encoder_seq_lens.copy_to_gpu(num_reqs)
         encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs]
@@ -1821,6 +1832,7 @@ class GPUModelRunner(
                 num_scheduled_tokens or {},
                 kv_cache_group.kv_cache_spec,
                 num_reqs_padded,
+                for_cudagraph_capture=for_cudagraph_capture,
             )
             if kv_cache_gid > 0:
                 cm.block_table_tensor = _get_block_table(kv_cache_gid)