sglang 0.4.10.post1__py3-none-any.whl → 0.5.0rc0__py3-none-any.whl

sglang/srt/layers/attention/flashattention_backend.py

@@ -1406,7 +1406,7 @@ class FlashAttentionBackend(AttentionBackend):
                     )
                     metadata.page_table = self.decode_cuda_graph_metadata[
                         "page_table_draft_decode"
-                    ][req_pool_indices, :]
+                    ][:bs, :]
                     self.decode_cuda_graph_metadata[bs] = metadata
                 else:
                     # When top k > 1, we need two specific draft decode metadata, and then merge states
@@ -1424,7 +1424,7 @@ class FlashAttentionBackend(AttentionBackend):
                     ][: bs + 1]
                     metadata.page_table = self.draft_decode_metadata_topk_normal[
                         "page_table"
-                    ][req_pool_indices, :]
+                    ][:bs, :]
 
                     # 2. The second half of metadata for draft tokens (per_batch_num_tokens = topk)
                     metadata_expand.cache_seqlens_int32 = (
@@ -1461,7 +1461,7 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata.max_seq_len_k = seq_lens.max().item()
                 # Precompute page table
                 metadata.page_table = self.decode_cuda_graph_metadata["page_table"][
-                    req_pool_indices, :
+                    :bs, :
                 ]
                 # Precompute cumulative sequence lengths
                 metadata.cu_seqlens_q = torch.arange(
@@ -1498,9 +1498,7 @@ class FlashAttentionBackend(AttentionBackend):
                     : (bs + 1)
                 ]
 
-                metadata.page_table = self.target_verify_metadata["page_table"][
-                    req_pool_indices, :
-                ]
+                metadata.page_table = self.target_verify_metadata["page_table"][:bs, :]
 
                 self.target_verify_metadata[bs] = metadata
             else:
@@ -1519,7 +1517,7 @@ class FlashAttentionBackend(AttentionBackend):
                 ][: bs + 1]
                 metadata.page_table = self.target_verify_metadata_topk_normal[
                     "page_table"
-                ][req_pool_indices, :]
+                ][:bs, :]
 
                 # 2. The second half of metadata for draft tokens (per_batch_num_tokens = topk)
                 metadata_expand.cache_seqlens_int32 = (
@@ -1562,9 +1560,7 @@ class FlashAttentionBackend(AttentionBackend):
             metadata.cu_seqlens_k = self.draft_extend_metadata["cu_seqlens_k"][
                 : (bs + 1)
             ]
-            metadata.page_table = self.draft_extend_metadata["page_table"][
-                req_pool_indices, :
-            ]
+            metadata.page_table = self.draft_extend_metadata["page_table"][:bs, :]
 
             self.draft_extend_metadata[bs] = metadata
 
@@ -1578,7 +1574,7 @@ class FlashAttentionBackend(AttentionBackend):
             ][: (encoder_bs + 1)]
 
             metadata.encoder_page_table = self.encoder_metadata["encoder_page_table"][
-                req_pool_indices, :
+                :bs, :
             ]
 
         self.forward_metadata = metadata

sglang/srt/layers/attention/triton_backend.py

@@ -88,6 +88,7 @@ class TritonAttnBackend(AttentionBackend):
                 self.window_kv_indptr = torch.zeros_like(kv_indptr_buf)
 
         self.req_to_token = model_runner.req_to_token_pool.req_to_token
+        self.token_to_kv_pool_allocator = model_runner.token_to_kv_pool_allocator
 
         if not self.skip_prefill:
             self.qo_indptr = torch.zeros(
@@ -197,6 +198,7 @@ class TritonAttnBackend(AttentionBackend):
                             forward_batch.req_pool_indices,
                             bs,
                             self.device,
+                            self.token_to_kv_pool_allocator,
                         )
                     )
                     window_num_kv_splits = torch.empty(
@@ -225,7 +227,6 @@ class TritonAttnBackend(AttentionBackend):
             mask_indptr = None
             max_extend_len = None
         elif forward_batch.forward_mode.is_target_verify():
-            # TODO: Support sliding window in spec inference
             bs = len(forward_batch.req_pool_indices)
             qo_indptr = torch.arange(
                 0,
@@ -250,6 +251,20 @@ class TritonAttnBackend(AttentionBackend):
                 self.req_to_token.stride(0),
             )
 
+            if self.sliding_window_size is not None and self.sliding_window_size > 0:
+                window_kv_indptr, window_kv_indices, window_kv_lens = (
+                    update_sliding_window_buffer(
+                        self.window_kv_indptr,
+                        self.req_to_token,
+                        self.sliding_window_size,
+                        forward_batch.seq_lens,
+                        forward_batch.req_pool_indices,
+                        bs,
+                        self.device,
+                        self.token_to_kv_pool_allocator,
+                    )
+                )
+
             custom_mask = spec_info.custom_mask
             seq_mask_len = self.num_draft_tokens * (
                 forward_batch.seq_lens + self.num_draft_tokens
@@ -308,6 +323,7 @@ class TritonAttnBackend(AttentionBackend):
                     forward_batch.req_pool_indices,
                     bs,
                     self.device,
+                    self.token_to_kv_pool_allocator,
                 )
 
             qo_indptr = self.qo_indptr
@@ -423,14 +439,17 @@ class TritonAttnBackend(AttentionBackend):
                 ):
                     window_kv_indices = self.cuda_graph_window_kv_indices
                     window_num_kv_splits = self.cuda_graph_window_num_kv_splits
-                    window_kv_indptr, _ = update_sliding_window_buffer_cuda_graph(
-                        self.window_kv_indptr,
-                        window_kv_indices,
-                        self.req_to_token,
-                        self.sliding_window_size,
-                        seq_lens[:bs],
-                        req_pool_indices,
-                        bs,
+                    window_kv_indptr, window_kv_indices, _ = (
+                        update_sliding_window_buffer_cuda_graph(
+                            self.window_kv_indptr,
+                            window_kv_indices,
+                            self.req_to_token,
+                            self.sliding_window_size,
+                            seq_lens[:bs],
+                            req_pool_indices,
+                            bs,
+                            self.token_to_kv_pool_allocator,
+                        )
                     )
             else:
                 kv_indptr, kv_indices = spec_info.kv_indptr, spec_info.kv_indices
@@ -464,6 +483,22 @@ class TritonAttnBackend(AttentionBackend):
                 self.req_to_token.stride(0),
             )
 
+            if self.sliding_window_size is not None and self.sliding_window_size > 0:
+                window_kv_indices = self.cuda_graph_window_kv_indices
+                window_num_kv_splits = self.cuda_graph_window_num_kv_splits
+                window_kv_indptr, window_kv_indices, _ = (
+                    update_sliding_window_buffer_cuda_graph(
+                        self.window_kv_indptr,
+                        window_kv_indices,
+                        self.req_to_token,
+                        self.sliding_window_size,
+                        seq_lens,
+                        req_pool_indices,
+                        bs,
+                        self.token_to_kv_pool_allocator,
+                    )
+                )
+
             custom_mask = self.cuda_graph_custom_mask
             custom_mask[: spec_info.custom_mask.shape[0]] = spec_info.custom_mask
             seq_mask_len = self.num_draft_tokens * (seq_lens + self.num_draft_tokens)
@@ -557,7 +592,7 @@ class TritonAttnBackend(AttentionBackend):
                 ):
                     window_num_kv_splits = self.cuda_graph_window_num_kv_splits
                     window_kv_indices = self.cuda_graph_window_kv_indices
-                    _, window_kv_lens = update_sliding_window_buffer_cuda_graph(
+                    _, _, window_kv_lens = update_sliding_window_buffer_cuda_graph(
                         self.window_kv_indptr,
                         window_kv_indices,
                         self.req_to_token,
@@ -565,6 +600,7 @@ class TritonAttnBackend(AttentionBackend):
                         seq_lens[:bs],
                         req_pool_indices[:bs],
                         bs,
+                        self.token_to_kv_pool_allocator,
                     )
                     self.get_num_kv_splits(
                         window_num_kv_splits[:num_token], window_kv_lens[:bs]
@@ -599,6 +635,19 @@ class TritonAttnBackend(AttentionBackend):
                 kv_indices,
                 self.req_to_token.stride(0),
             )
+            if self.sliding_window_size is not None and self.sliding_window_size > 0:
+                window_num_kv_splits = self.cuda_graph_window_num_kv_splits
+                window_kv_indices = self.cuda_graph_window_kv_indices
+                _, _, window_kv_lens = update_sliding_window_buffer_cuda_graph(
+                    self.window_kv_indptr,
+                    window_kv_indices,
+                    self.req_to_token,
+                    self.sliding_window_size,
+                    seq_lens,
+                    req_pool_indices,
+                    bs,
+                    self.token_to_kv_pool_allocator,
+                )
             custom_mask = self.cuda_graph_custom_mask
             custom_mask[: spec_info.custom_mask.shape[0]] = spec_info.custom_mask
             seq_mask_len = self.num_draft_tokens * (seq_lens + self.num_draft_tokens)
@@ -637,6 +686,7 @@ class TritonAttnBackend(AttentionBackend):
         layer: RadixAttention,
         forward_batch: ForwardBatch,
         save_kv_cache=True,
+        sinks=None,
     ):
         # TODO: reuse the buffer across layers
         if layer.qk_head_dim != layer.v_head_dim:
@@ -680,7 +730,8 @@ class TritonAttnBackend(AttentionBackend):
             self.forward_metadata.max_extend_len,
             layer.scaling,
             layer.logit_cap,
-            sliding_window_size,
+            sliding_window_size=sliding_window_size,
+            sinks=sinks,
         )
         return o
 
@@ -692,6 +743,7 @@ class TritonAttnBackend(AttentionBackend):
         layer: RadixAttention,
         forward_batch: ForwardBatch,
         save_kv_cache=True,
+        sinks=None,
     ):
         # During torch.compile, there is a bug in rotary_emb that causes the
         # output value to have a 3D tensor shape. This reshapes the output correctly.
@@ -728,6 +780,7 @@ class TritonAttnBackend(AttentionBackend):
             self.max_kv_splits,
             layer.scaling,
             layer.logit_cap,
+            sinks=sinks,
         )
         return o
 
@@ -932,10 +985,11 @@ def update_sliding_window_buffer(
     req_pool_indices,
     bs,
     device,
+    token_to_kv_pool_allocator=None,
 ):
     window_kv_lens = torch.minimum(
         seq_lens,
-        torch.tensor(sliding_window_size + 1),
+        torch.tensor(sliding_window_size),
     )
     window_kv_indptr[1 : bs + 1] = torch.cumsum(window_kv_lens, dim=0)
     window_kv_indptr = window_kv_indptr[: bs + 1]
@@ -952,6 +1006,14 @@ def update_sliding_window_buffer(
         window_kv_indices,
         req_to_token.stride(0),
     )
+    # full to swa index mapping
+    if hasattr(token_to_kv_pool_allocator, "translate_loc_from_full_to_swa"):
+        kv_last_index = window_kv_indptr[-1]
+        window_kv_indices[:kv_last_index] = (
+            token_to_kv_pool_allocator.translate_loc_from_full_to_swa(
+                window_kv_indices[:kv_last_index]
+            )
+        )
     return window_kv_indptr, window_kv_indices, window_kv_lens
 
 
@@ -963,10 +1025,11 @@ def update_sliding_window_buffer_cuda_graph(
     seq_lens,
     req_pool_indices,
     bs,
+    token_to_kv_pool_allocator=None,
 ):
     window_kv_lens = torch.minimum(
         seq_lens,
-        torch.tensor(sliding_window_size + 1),
+        torch.tensor(sliding_window_size),
     )
     window_kv_indptr[1 : bs + 1] = torch.cumsum(window_kv_lens, dim=0)
     window_kv_indptr = window_kv_indptr[: bs + 1]
@@ -980,4 +1043,12 @@ def update_sliding_window_buffer_cuda_graph(
         window_kv_indices,
         req_to_token.stride(0),
     )
-    return window_kv_indptr, window_kv_lens
+    # full to swa index mapping
+    if hasattr(token_to_kv_pool_allocator, "translate_loc_from_full_to_swa"):
+        kv_last_index = window_kv_indptr[-1]
+        window_kv_indices[:kv_last_index] = (
+            token_to_kv_pool_allocator.translate_loc_from_full_to_swa(
+                window_kv_indices[:kv_last_index]
+            )
+        )
+    return window_kv_indptr, window_kv_indices, window_kv_lens

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

@@ -495,6 +495,7 @@ def _fwd_kernel_stage2(
     O,
     kv_indptr,
     num_kv_splits,
+    sink_ptr,
     stride_mid_ob,
     stride_mid_oh,
     stride_mid_os,
@@ -504,6 +505,7 @@ def _fwd_kernel_stage2(
     MIN_BLOCK_KV: tl.constexpr,
     BLOCK_DV: tl.constexpr,
     Lv: tl.constexpr,
+    HAS_SINK: tl.constexpr,
 ):
     cur_batch = tl.program_id(0)
     cur_head = tl.program_id(1)
@@ -545,6 +547,10 @@ def _fwd_kernel_stage2(
             e_sum = e_sum * old_scale + exp_logic
             e_max = n_e_max
 
+    if HAS_SINK:
+        cur_sink = tl.load(sink_ptr + cur_head)
+        e_sum += tl.exp(cur_sink - e_max)
+
     tl.store(
         O + cur_batch * stride_obs + cur_head * stride_oh + offs_d,
         acc / e_sum,
@@ -561,12 +567,14 @@ def _decode_softmax_reducev_fwd(
     kv_indptr,
     num_kv_splits,
     max_kv_splits,
+    sinks=None,
 ):
     batch, head_num = q.shape[0], q.shape[1]
     Lv = v_buffer.shape[-1]
     BLOCK_DV = triton.next_power_of_2(Lv)
 
     MAX_KV_SPLITS = max_kv_splits
+    HAS_SINK = sinks is not None
 
     extra_kargs = {}
     if _is_hip:
@@ -581,6 +589,7 @@ def _decode_softmax_reducev_fwd(
         o,
         kv_indptr,
         num_kv_splits,
+        sinks,
         logits.stride(0),
         logits.stride(1),
         logits.stride(2),
@@ -590,6 +599,7 @@ def _decode_softmax_reducev_fwd(
         MIN_BLOCK_KV=_MIN_BLOCK_KV,
         BLOCK_DV=BLOCK_DV,
         Lv=Lv,
+        HAS_SINK=HAS_SINK,
         num_warps=4,
         num_stages=2,
         **extra_kargs,
@@ -609,6 +619,7 @@ def decode_attention_fwd_normal(
     max_kv_splits,
     sm_scale,
     logit_cap=0.0,
+    sinks=None,
 ):
     _decode_att_m_fwd(
         q,
@@ -632,6 +643,7 @@ def decode_attention_fwd_normal(
         kv_indptr,
         num_kv_splits,
         max_kv_splits,
+        sinks,
     )
 
 
@@ -648,6 +660,7 @@ def decode_attention_fwd_grouped(
     max_kv_splits,
     sm_scale,
     logit_cap=0.0,
+    sinks=None,
 ):
     _decode_grouped_att_m_fwd(
         q,
@@ -671,6 +684,7 @@ def decode_attention_fwd_grouped(
         kv_indptr,
         num_kv_splits,
         max_kv_splits,
+        sinks,
     )
 
 
@@ -687,6 +701,7 @@ def decode_attention_fwd(
     max_kv_splits,
     sm_scale,
     logit_cap=0.0,
+    sinks=None,
 ):
     assert max_kv_splits == attn_logits.shape[2]
     assert q.shape[0] <= kv_indptr.shape[0] - 1
@@ -709,6 +724,7 @@ def decode_attention_fwd(
             max_kv_splits,
             sm_scale,
             logit_cap=logit_cap,
+            sinks=sinks,
         )
     else:
         # GQA/MQA/MLA
@@ -725,4 +741,5 @@ def decode_attention_fwd(
             max_kv_splits,
             sm_scale,
             logit_cap=logit_cap,
+            sinks=sinks,
         )