sglang 0.4.6.post2__py3-none-any.whl → 0.4.6.post4__py3-none-any.whl

sglang/srt/layers/attention/flashattention_backend.py

@@ -308,7 +308,7 @@ class FlashAttentionBackend(AttentionBackend):
         ), "Sliding window and cross attention are not supported together"
 
         self.forward_metadata: FlashAttentionMetadata = None
-        # extra metdata for handling speculative decoding topk > 1, extended draft decode and verify
+        # extra metadata for handling speculative decoding topk > 1, extended draft decode and verify
         self.forward_metadata_spec_decode_expand: FlashAttentionMetadata = None
         self.max_context_len = model_runner.model_config.context_len
         self.device = model_runner.device
@@ -338,7 +338,7 @@ class FlashAttentionBackend(AttentionBackend):
         """Initialize forward metadata hence all layers in the forward pass can reuse it."""
         metadata = FlashAttentionMetadata()
         seqlens_in_batch = forward_batch.seq_lens
-        batch_size = len(seqlens_in_batch)
+        batch_size = forward_batch.batch_size
         device = seqlens_in_batch.device
 
         if forward_batch.forward_mode.is_decode_or_idle():
@@ -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 = (
-            self.attention_chunk_size is not None and local_attn_metadata is not None
+        use_local_attn = (
+            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[
@@ -1525,12 +1569,9 @@ class FlashAttentionBackend(AttentionBackend):
                     metadata.max_seq_len_k = seq_lens_cpu.max().item() + (
                         self.speculative_step_id + 1
                     )
-                    metadata.cu_seqlens_k.copy_(
-                        torch.nn.functional.pad(
-                            torch.cumsum(
-                                metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                            ),
-                            (1, 0),
+                    metadata.cu_seqlens_k[1:].copy_(
+                        torch.cumsum(
+                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
                         )
                     )
 
@@ -1554,12 +1595,9 @@ class FlashAttentionBackend(AttentionBackend):
                     # metadata.max_seq_len_q = self.topk, already set in capture
                     metadata.max_seq_len_k = seq_lens_cpu.max().item()
                     # metadata.cu_seqlens_q already set in capture
-                    metadata.cu_seqlens_k.copy_(
-                        torch.nn.functional.pad(
-                            torch.cumsum(
-                                metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                            ),
-                            (1, 0),
+                    metadata.cu_seqlens_k[1:].copy_(
+                        torch.cumsum(
+                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
                         )
                     )
 
@@ -1578,8 +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
-                self._init_local_attn_metadata(metadata, device)
+                # TODO: Handle local attention metadata for draft decode when llama4 eagle is supported
             else:
                 metadata = self.decode_cuda_graph_metadata[bs]
                 # Normal Decode
@@ -1587,8 +1624,9 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata.max_seq_len_k = max_len
 
                 metadata.cache_seqlens_int32 = seq_lens.to(torch.int32)
-                metadata.cu_seqlens_k = torch.nn.functional.pad(
-                    torch.cumsum(seq_lens, dim=0, dtype=torch.int32), (1, 0)
+                # Optimize cumulative sequence length calculation
+                metadata.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(seq_lens, dim=0, dtype=torch.int32)
                 )
 
                 max_seq_pages = (
@@ -1604,7 +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]
@@ -1615,13 +1653,8 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata.max_seq_len_k = (
                     seq_lens_cpu.max().item() + self.speculative_num_draft_tokens
                 )
-                metadata.cu_seqlens_k.copy_(
-                    torch.nn.functional.pad(
-                        torch.cumsum(
-                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                        ),
-                        (1, 0),
-                    )
+                metadata.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(metadata.cache_seqlens_int32, dim=0, dtype=torch.int32)
                 )
                 max_seq_pages = (
                     metadata.max_seq_len_k + self.page_size - 1
@@ -1640,13 +1673,8 @@ class FlashAttentionBackend(AttentionBackend):
                 # metadata.max_seq_len_q = self.speculative_num_draft_tokens, already set in capture
                 metadata.max_seq_len_k = seq_lens_cpu.max().item()
                 # metadata.cu_seqlens_q already set in capture
-                metadata.cu_seqlens_k.copy_(
-                    torch.nn.functional.pad(
-                        torch.cumsum(
-                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                        ),
-                        (1, 0),
-                    )
+                metadata.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(metadata.cache_seqlens_int32, dim=0, dtype=torch.int32)
                 )
                 page_table = self.req_to_token[
                     req_pool_indices, : metadata.max_seq_len_k
@@ -1704,14 +1732,11 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata_expand.cache_seqlens_int32.copy_(
                     mask.sum(dim=1).to(torch.int32)
                 )
-                metadata_expand.cu_seqlens_k.copy_(
-                    torch.nn.functional.pad(
-                        torch.cumsum(
-                            metadata_expand.cache_seqlens_int32,
-                            dim=0,
-                            dtype=torch.int32,
-                        ),
-                        (1, 0),
+                metadata_expand.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(
+                        metadata_expand.cache_seqlens_int32,
+                        dim=0,
+                        dtype=torch.int32,
                     )
                 )
                 metadata_expand.max_seq_len_k = (
@@ -1722,11 +1747,8 @@ class FlashAttentionBackend(AttentionBackend):
             # Only support encoder size 1 for now
             metadata.encoder_max_seq_len_k = encoder_lens[0]
             metadata.encoder_lens_int32.copy_(encoder_lens[:1])
-            metadata.encoder_cu_seqlens_k.copy_(
-                torch.nn.functional.pad(
-                    torch.cumsum(metadata.encoder_lens_int32, dim=0, dtype=torch.int32),
-                    (1, 0),
-                )
+            metadata.encoder_cu_seqlens_k[1:].copy_(
+                torch.cumsum(metadata.encoder_lens_int32, dim=0, dtype=torch.int32)
             )
 
             metadata.encoder_page_table[:, : metadata.encoder_max_seq_len_k].copy_(
@@ -1776,6 +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),
@@ -1785,6 +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:
 

sglang/srt/layers/attention/flashinfer_backend.py

@@ -16,8 +16,9 @@ from typing import TYPE_CHECKING, Callable, List, Optional, Union
 import torch
 
 if os.environ["SGLANG_ENABLE_TORCH_COMPILE"] == "1":
-    import torch._dynamo
+    import logging
 
+    torch._logging.set_logs(dynamo=logging.ERROR)
     torch._dynamo.config.suppress_errors = True
 
 from sglang.global_config import global_config
@@ -107,6 +108,7 @@ class FlashInferAttnBackend(AttentionBackend):
         if (
             "Qwen2ForCausalLM" in model_runner.model_config.hf_config.architectures
             or "Qwen3ForCausalLM" in model_runner.model_config.hf_config.architectures
+            or "MiMoForCausalLM" in model_runner.model_config.hf_config.architectures
         ):
             global_config.flashinfer_workspace_size = 512 * 1024 * 1024
 
@@ -416,6 +418,7 @@ class FlashInferAttnBackend(AttentionBackend):
 
         logits_soft_cap = layer.logit_cap
 
+        q = q.contiguous()
         if not self.forward_metadata.use_ragged:
             if k is not None:
                 assert v is not None
@@ -425,7 +428,7 @@ class FlashInferAttnBackend(AttentionBackend):
                     )
 
             o = prefill_wrapper_paged.forward(
-                q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
+                q.view(-1, layer.tp_q_head_num, layer.head_dim),
                 forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
                 causal=not layer.is_cross_attention,
                 sm_scale=layer.scaling,
@@ -435,20 +438,27 @@ class FlashInferAttnBackend(AttentionBackend):
                 v_scale=layer.v_scale,
             )
         else:
-            o1, s1 = self.prefill_wrapper_ragged.forward_return_lse(
-                q.view(-1, layer.tp_q_head_num, layer.head_dim),
-                k.view(-1, layer.tp_k_head_num, layer.head_dim),
-                v.view(-1, layer.tp_v_head_num, layer.head_dim),
-                causal=True,
-                sm_scale=layer.scaling,
-                logits_soft_cap=logits_soft_cap,
-            )
-
             if self.forward_metadata.extend_no_prefix:
-                o = o1
+                o = self.prefill_wrapper_ragged.forward(
+                    q.view(-1, layer.tp_q_head_num, layer.head_dim),
+                    k.view(-1, layer.tp_k_head_num, layer.head_dim),
+                    v.view(-1, layer.tp_v_head_num, layer.head_dim),
+                    causal=True,
+                    sm_scale=layer.scaling,
+                    logits_soft_cap=logits_soft_cap,
+                )
+
             else:
+                o1, s1 = self.prefill_wrapper_ragged.forward_return_lse(
+                    q.view(-1, layer.tp_q_head_num, layer.head_dim),
+                    k.view(-1, layer.tp_k_head_num, layer.head_dim),
+                    v.view(-1, layer.tp_v_head_num, layer.head_dim),
+                    causal=True,
+                    sm_scale=layer.scaling,
+                    logits_soft_cap=logits_soft_cap,
+                )
                 o2, s2 = prefill_wrapper_paged.forward_return_lse(
-                    q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
+                    q.view(-1, layer.tp_q_head_num, layer.head_dim),
                     forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
                     causal=False,
                     sm_scale=layer.scaling,

sglang/srt/layers/attention/flashinfer_mla_backend.py

@@ -18,8 +18,9 @@ import torch
 import triton
 
 if os.environ["SGLANG_ENABLE_TORCH_COMPILE"] == "1":
-    import torch._dynamo
+    import logging
 
+    torch._logging.set_logs(dynamo=logging.ERROR)
     torch._dynamo.config.suppress_errors = True
 
 from sglang.global_config import global_config
@@ -338,23 +339,39 @@ class FlashInferMLAAttnBackend(AttentionBackend):
         layer: RadixAttention,
         forward_batch: ForwardBatch,
         save_kv_cache: bool = True,
+        q_rope: Optional[torch.Tensor] = None,
+        k_rope: Optional[torch.Tensor] = None,
     ):
 
         cache_loc = forward_batch.out_cache_loc
         logits_soft_cap = layer.logit_cap
         prefill_wrapper_paged = self.forward_metadata.prefill_wrapper
-        qall = q.view(-1, layer.tp_q_head_num, layer.head_dim)
         k_buf = forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id)
 
         # Save kv cache
         if save_kv_cache and k is not None:
             assert v is not None
             if save_kv_cache:
-                forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+                if k_rope is not None:
+                    forward_batch.token_to_kv_pool.set_mla_kv_buffer(
+                        layer, cache_loc, k, k_rope
+                    )
+                else:
+                    forward_batch.token_to_kv_pool.set_kv_buffer(layer, cache_loc, k, v)
+        if q_rope is not None:
+            q = q.view(-1, layer.tp_q_head_num, layer.v_head_dim)
+            q_rope = q_rope.view(
+                -1, layer.tp_q_head_num, layer.head_dim - layer.v_head_dim
+            )
 
         if self.forward_metadata.use_ragged:
             # ragged prefill
-            o, _ = self.prefill_wrapper_ragged.forward_return_lse(
+            if q_rope is not None:
+                q = torch.cat([q, q_rope], dim=-1)
+            qall = q.view(-1, layer.tp_q_head_num, layer.head_dim)
+            if k_rope is not None:
+                k = torch.cat([k, k_rope], dim=-1)
+            o = self.prefill_wrapper_ragged.forward(
                 qall,
                 k.view(-1, layer.tp_k_head_num, layer.head_dim),
                 v.view(-1, layer.tp_k_head_num, layer.v_head_dim),
@@ -364,11 +381,19 @@ class FlashInferMLAAttnBackend(AttentionBackend):
             )
         else:
             # mla paged prefill
+            if q_rope is None:
+                qall = q.view(-1, layer.tp_q_head_num, layer.head_dim)
+                q, q_rope = (
+                    qall[:, :, : layer.v_head_dim],
+                    qall[:, :, layer.v_head_dim :],
+                )
+            o = q.new_empty(q.shape)
             o = prefill_wrapper_paged.run(
-                qall[:, :, : layer.v_head_dim],
-                qall[:, :, layer.v_head_dim :],
+                q,
+                q_rope,
                 k_buf[:, :, : layer.v_head_dim],
                 k_buf[:, :, layer.v_head_dim :],
+                out=o,
             )
 
         return o.view(-1, layer.tp_q_head_num * layer.v_head_dim)
@@ -381,6 +406,9 @@ class FlashInferMLAAttnBackend(AttentionBackend):
         layer: RadixAttention,
         forward_batch: ForwardBatch,
         save_kv_cache: bool = True,
+        # For multi-head latent attention
+        q_rope: Optional[torch.Tensor] = None,
+        k_rope: Optional[torch.Tensor] = None,
     ):
         decode_wrapper = self.forward_metadata.decode_wrapper
         cache_loc = forward_batch.out_cache_loc
@@ -388,23 +416,42 @@ class FlashInferMLAAttnBackend(AttentionBackend):
         if k is not None:
             assert v is not None
             if save_kv_cache:
-                forward_batch.token_to_kv_pool.set_kv_buffer(
-                    layer,
-                    cache_loc,
-                    k,
-                    v,
-                )
+                if k_rope is not None:
+                    forward_batch.token_to_kv_pool.set_mla_kv_buffer(
+                        layer,
+                        cache_loc,
+                        k,
+                        k_rope,
+                    )
+                else:
+                    forward_batch.token_to_kv_pool.set_kv_buffer(
+                        layer,
+                        cache_loc,
+                        k,
+                        v,
+                    )
 
         # Reshape inputs
-        reshaped_q = q.view(-1, layer.tp_q_head_num, layer.head_dim)
+        if q_rope is not None:
+            q_nope = q.view(-1, layer.tp_q_head_num, layer.v_head_dim)
+            q_rope = q_rope.view(
+                -1, layer.tp_q_head_num, layer.head_dim - layer.v_head_dim
+            )
+        else:
+            reshaped_q = q.view(-1, layer.tp_q_head_num, layer.head_dim)
+            q_nope = reshaped_q[:, :, : layer.v_head_dim]
+            q_rope = reshaped_q[:, :, layer.v_head_dim :]
+
         k_buffer = forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id)
 
+        o = q_nope.new_empty(q_nope.shape)
         # Direct call to run without the wrapper
         o = decode_wrapper.run(
-            reshaped_q[:, :, : layer.v_head_dim],
-            reshaped_q[:, :, layer.v_head_dim :],
+            q_nope,
+            q_rope,
             k_buffer[:, :, : layer.v_head_dim],
             k_buffer[:, :, layer.v_head_dim :],
+            out=o,
         )
 
         return o.view(-1, layer.tp_q_head_num * layer.v_head_dim)

sglang/srt/layers/attention/merge_state.py

@@ -0,0 +1,46 @@
+from typing import Optional, Tuple
+
+import torch
+from sgl_kernel import merge_state_v2
+
+from sglang.srt.layers.attention.triton_ops.merge_state import merge_state_triton
+from sglang.srt.utils import is_cuda
+
+_is_cuda = is_cuda()
+
+
+# Automatically fallback to the Triton kernel in some cases
+# (e.g., for AMD GPUs, when the head dimension is not a multiple
+# of 4 or 8, and in FP8 precision)
+def _supported_dtypes(o: torch.Tensor) -> bool:
+    return o.dtype in [torch.float32, torch.half, torch.bfloat16]
+
+
+def _supported_headdim(o: torch.Tensor) -> bool:
+    headdim = o.shape[2]  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+    if o.dtype == torch.float32:
+        return headdim % 4 == 0
+    return headdim % 8 == 0
+
+
+def merge_state(
+    prefix_output: torch.Tensor,
+    prefix_lse: torch.Tensor,
+    suffix_output: torch.Tensor,
+    suffix_lse: torch.Tensor,
+    output: Optional[torch.Tensor] = None,
+    output_lse: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+    if (
+        _is_cuda
+        and _supported_dtypes(prefix_output)
+        and _supported_headdim(prefix_output)
+    ):
+        return merge_state_v2(
+            prefix_output, prefix_lse, suffix_output, suffix_lse, output, output_lse
+        )
+    else:
+        # Fallback to Triton kernel
+        return merge_state_triton(
+            prefix_output, prefix_lse, suffix_output, suffix_lse, output, output_lse
+        )

sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py

@@ -919,7 +919,7 @@ def _fwd_kernel(
 
         e_max = n_e_max
 
-    # stage 2: compute the trianlge part
+    # stage 2: compute the triangle part
 
     cur_block_m_end = tl.minimum(cur_seq_len_extend, (cur_block_m + 1) * BLOCK_M)
     for start_n in range(0, cur_block_m_end, BLOCK_N):