[Bugfix] Fix Llama4 gibberish output with long context and CUDA graph (#6162)

python/sglang/srt/layers/attention/flashattention_backend.py

@@ -913,8 +913,10 @@ class FlashAttentionBackend(AttentionBackend):
         # Use precomputed metadata across all layers
         metadata = self.forward_metadata
         local_attn_metadata = getattr(metadata, "local_attn_metadata", None)
-        use_local_attention = (
+        use_local_attn = (
-            self.attention_chunk_size is not None and local_attn_metadata is not None
+            self.attention_chunk_size is not None
+            and local_attn_metadata is not None
+            and (hasattr(layer, "use_irope") and layer.use_irope)
         )
         # We do cascade attention for Draft Decode with topk > 1
         use_cascade_attn = self.topk > 1
@@ -970,7 +972,7 @@ class FlashAttentionBackend(AttentionBackend):
                     k_descale=k_descale,
                     v_descale=v_descale,
                 )
-            elif use_local_attention:
+            elif use_local_attn:
                 # Use chunked (local) attention batching for self-attention
                 o = flash_attn_with_kvcache(
                     q=q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
@@ -979,7 +981,7 @@ class FlashAttentionBackend(AttentionBackend):
                     page_table=local_attn_metadata.local_block_table,
                     cache_seqlens=local_attn_metadata.local_seqused_k,
                     cu_seqlens_q=local_attn_metadata.local_query_start_loc,
-                    cu_seqlens_k_new=metadata.cu_seqlens_k,
+                    cu_seqlens_k_new=None,
                     max_seqlen_q=local_attn_metadata.local_max_query_len,
                     softmax_scale=layer.scaling,
                     causal=True,
@@ -1127,7 +1129,6 @@ class FlashAttentionBackend(AttentionBackend):
         This creates fixed-size tensors that will be reused during CUDA graph replay
         to avoid memory allocations.
         """
-
         # This is being used by normal decode and draft decode when topk == 1
         self.decode_cuda_graph_metadata = {
             "cache_seqlens": torch.zeros(max_bs, dtype=torch.int32, device=self.device),
@@ -1154,6 +1155,34 @@ class FlashAttentionBackend(AttentionBackend):
             ),
         }
 
+        # Only allocate local attention buffers if local attention is enabled
+        # This prevents OOM errors when local attention is not being used
+        if self.attention_chunk_size is not None:
+            # Estimate maximum sizes for local attention metadata
+            max_seq_len = self.max_context_len
+            page_size = self.page_size or 1
+            attn_chunk_size = self.attention_chunk_size
+            max_virtual_batches = max_bs * (
+                (max_seq_len + attn_chunk_size - 1) // attn_chunk_size
+            )
+            max_blocks_per_seq = (max_seq_len + attn_chunk_size - 1) // attn_chunk_size
+            max_pages_per_block = (attn_chunk_size + page_size - 1) // page_size
+
+            self.decode_cuda_graph_local_attn_metadata = {
+                "local_query_start_loc": torch.zeros(
+                    max_virtual_batches + 1, dtype=torch.int32, device=self.device
+                ),
+                "local_seqused_k": torch.zeros(
+                    max_virtual_batches, dtype=torch.int32, device=self.device
+                ),
+                "local_block_table": torch.zeros(
+                    max_virtual_batches,
+                    max_blocks_per_seq * max_pages_per_block,
+                    dtype=torch.int32,
+                    device=self.device,
+                ),
+            }
+
         # This is used by draft decode's first half of metadata when topk > 1
         if self.topk > 1:
             self.draft_decode_metadata_topk_normal = {
@@ -1405,6 +1434,21 @@ class FlashAttentionBackend(AttentionBackend):
                 )
                 self.decode_cuda_graph_metadata[bs] = metadata
 
+                if self.attention_chunk_size is not None:
+                    metadata.local_attn_metadata = FlashAttentionMetadata.LocalAttentionMetadata(
+                        local_query_start_loc=self.decode_cuda_graph_local_attn_metadata[
+                            "local_query_start_loc"
+                        ],
+                        local_seqused_k=self.decode_cuda_graph_local_attn_metadata[
+                            "local_seqused_k"
+                        ],
+                        local_block_table=self.decode_cuda_graph_local_attn_metadata[
+                            "local_block_table"
+                        ],
+                        local_max_query_len=1,
+                        local_max_seq_len=1,
+                    )
+
         elif forward_mode.is_target_verify():
             if self.topk <= 1:
                 metadata.cache_seqlens_int32 = self.target_verify_metadata[
@@ -1572,8 +1616,7 @@ class FlashAttentionBackend(AttentionBackend):
                     metadata_expand.page_table[: cache_loc.shape[0]].copy_(
                         cache_loc[:, :decode_length].contiguous().to(torch.int32)
                     )
-                # TODO: we need to test this part for llama 4 eagle case
+                # TODO: Handle local attention metadata for draft decode when llama4 eagle is supported
-                self._init_local_attn_metadata(metadata, device)
             else:
                 metadata = self.decode_cuda_graph_metadata[bs]
                 # Normal Decode
@@ -1599,7 +1642,7 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata.page_table[:, :max_seq_pages].copy_(page_indices)
                 metadata.page_table[:, max_seq_pages:].fill_(0)
 
-                self._init_local_attn_metadata(metadata, device)
+                self._update_local_attn_metadata_for_replay(metadata, bs)
         elif forward_mode.is_target_verify():
             if self.topk <= 1:
                 metadata = self.target_verify_metadata[bs]
@@ -1755,6 +1798,7 @@ class FlashAttentionBackend(AttentionBackend):
             page_table,
             self.page_size,
         )
+
         local_metadata = FlashAttentionMetadata.LocalAttentionMetadata(
             local_query_start_loc=torch.from_numpy(cu_seqlens_q_local_np).to(device),
             local_seqused_k=torch.from_numpy(seqlens_k_local_np).to(device),
@@ -1764,6 +1808,79 @@ class FlashAttentionBackend(AttentionBackend):
         )
         metadata.local_attn_metadata = local_metadata
 
+    def _update_local_attn_metadata_for_replay(
+        self, metadata: FlashAttentionMetadata, bs: int
+    ):
+        """Update preallocated local attention metadata in-place before CUDA graph replay."""
+        if self.attention_chunk_size is None:
+            return
+
+        # Access preallocated buffers
+        local_q_buf = self.decode_cuda_graph_local_attn_metadata[
+            "local_query_start_loc"
+        ]
+        local_k_buf = self.decode_cuda_graph_local_attn_metadata["local_seqused_k"]
+        local_block_buf = self.decode_cuda_graph_local_attn_metadata[
+            "local_block_table"
+        ]
+        cu_seqlens_q = self.decode_cuda_graph_metadata["cu_seqlens_q"]
+
+        # Create a modified version for local attention that only processes the last token
+        # This mimics the normal decode pattern
+        cu_seqlens_q = torch.arange(
+            bs + 1, device=cu_seqlens_q.device, dtype=cu_seqlens_q.dtype
+        )
+        seqlens = metadata.cache_seqlens_int32[:bs]
+        # Slice the page_table to match the batch size and actual sequence length
+        # This serves three important purposes:
+        # 1. Ensures we only process the actual batch size (bs) and not the maximum batch size
+        # 2. Limits the sequence length to prevent processing padding tokens or garbage values
+        # 3. Prevents zeros in the block table which can cause garbage output during replay
+        #
+        # Without this slicing, the pre-allocated page_table may contain zeros or invalid indices
+        # beyond the actual sequence length, leading to incorrect attention calculations
+        max_seq_len = int(seqlens.max().item())
+        sliced_page_table = metadata.page_table[:bs, :max_seq_len]
+
+        cu_seqlens_q_np = cu_seqlens_q.cpu().numpy()
+        seqlens_np = seqlens.cpu().numpy()
+        (
+            seqlens_q_local_np,
+            cu_seqlens_q_local_np,
+            seqlens_k_local_np,
+            block_table_local,
+        ) = make_local_attention_virtual_batches(
+            self.attention_chunk_size,
+            cu_seqlens_q_np,
+            seqlens_np,
+            sliced_page_table,
+            self.page_size,
+        )
+
+        # Convert back to tensors
+        device = local_q_buf.device
+        cu_seqlens_q_local = torch.from_numpy(cu_seqlens_q_local_np).to(device)
+        seqlens_k_local = torch.from_numpy(seqlens_k_local_np).to(device)
+        block_table_local = block_table_local.to(device)
+        # Get sizes
+        q_len = cu_seqlens_q_local.shape[0]
+        k_len = seqlens_k_local.shape[0]
+        b0, b1 = block_table_local.shape
+
+        # In-place updates into preallocated tensors and zero out the unused space
+        local_q_buf[:q_len].copy_(cu_seqlens_q_local)
+        local_q_buf[q_len:].fill_(0)
+        local_k_buf[:k_len].copy_(seqlens_k_local)
+        local_k_buf[k_len:].fill_(0)
+        local_block_buf[:b0, :b1].copy_(block_table_local)
+        local_block_buf[b0:, :].fill_(0)
+        local_block_buf[:b0, b1:].fill_(0)
+
+        if metadata.local_attn_metadata is not None:
+            lam = metadata.local_attn_metadata
+            lam.local_max_query_len = int(seqlens_q_local_np.max())
+            lam.local_max_seq_len = int(seqlens_k_local_np.max())
+
 
 class FlashAttentionMultiStepBackend: