Add attention sink in attention backends (#22320)

WoosukKwon · LiuXiaoxuanPKU · simon-mo · web-flow · commit 6e20924350e3 · 2025-08-05T22:37:21.000-07:00
Signed-off-by: Woosuk Kwon &lt;woosuk.kwon@berkeley.edu&gt;

Co-authored-by: LiuXiaoxuanPKU &lt;lilyliupku@gmail.com&gt;
Co-authored-by: simon-mo &lt;xmo@berkeley.edu&gt;
Co-authored-by: Chen Zhang &lt;zhangch99@outlook.com&gt;
Co-authored-by: Hongxia Yang &lt;62075498+hongxiayang@users.noreply.github.com&gt;
Co-authored-by: Minseok Lee &lt;47620120+minseokl@users.noreply.github.com&gt;
Co-authored-by: Yongye Zhu &lt;zyy1102000@gmail.com&gt;
diff --git a/vllm/attention/ops/chunked_prefill_paged_decode.py b/vllm/attention/ops/chunked_prefill_paged_decode.py
@@ -28,6 +28,7 @@ def kernel_paged_attention_2d(
         query_ptr,  # [num_tokens, num_query_heads, head_size]
         key_cache_ptr,  # [num_blks, num_kv_heads, head_size // x, blk_size, x]
         value_cache_ptr,  # [num_blks, num_kv_heads, head_size, blk_size]
+        sink_ptr,  # [num_query_heads]
         block_tables_ptr,  # [num_seqs, max_num_blocks_per_seq]
         seq_lens_ptr,  # [num_seqs]
         alibi_slopes_ptr,  # [num_query_heads]
@@ -95,7 +96,17 @@ def kernel_paged_attention_2d(
 
     block_table_offset = seq_idx * block_table_stride
 
-    M = tl.full([num_queries_per_kv_padded], float("-inf"), dtype=tl.float32)
+    if sink_ptr is None:
+        M = tl.full([num_queries_per_kv_padded],
+                    float("-inf"),
+                    dtype=tl.float32)
+    else:
+        M = tl.load(
+            sink_ptr + query_head_idx,
+            mask=head_mask,
+            other=float("-inf"),
+        ).to(dtype=tl.float32)
+
     L = tl.full([num_queries_per_kv_padded], 1.0, dtype=tl.float32)
     acc = tl.zeros([num_queries_per_kv_padded, HEAD_SIZE_PADDED],
                    dtype=tl.float32)
@@ -223,6 +234,8 @@ def chunked_prefill_paged_decode(
     alibi_slopes=None,
     sliding_window=None,
     sm_scale=None,
+    # Optional tensor for sinks
+    sinks=None,
 ):
 
     if sm_scale is None:
@@ -253,6 +266,7 @@ def chunked_prefill_paged_decode(
             sliding_window=sliding_window,
             sm_scale=sm_scale,
             skip_decode=True,
+            sinks=sinks,
         )
 
     block_size = value_cache.shape[3]
@@ -281,11 +295,17 @@ def chunked_prefill_paged_decode(
     num_queries_per_kv_padded = max(triton.next_power_of_2(num_queries_per_kv),
                                     16)
 
-    use_custom = use_rocm_custom_paged_attention(query.dtype, head_size,
-                                                 block_size,
-                                                 num_queries_per_kv,
-                                                 max_seq_len, sliding_window,
-                                                 kv_cache_dtype, alibi_slopes)
+    use_custom = use_rocm_custom_paged_attention(
+        query.dtype,
+        head_size,
+        block_size,
+        num_queries_per_kv,
+        max_seq_len,
+        sliding_window,
+        kv_cache_dtype,
+        alibi_slopes,
+        sinks,
+    )
     if use_custom:
         _PARTITION_SIZE_ROCM = 256
         max_num_partitions = ((max_seq_len + _PARTITION_SIZE_ROCM - 1) //
@@ -334,6 +354,7 @@ def chunked_prefill_paged_decode(
             query_ptr=query,
             key_cache_ptr=key_cache,
             value_cache_ptr=value_cache,
+            sink_ptr=sinks,
             block_tables_ptr=block_table,
             seq_lens_ptr=seq_lens,
             alibi_slopes_ptr=alibi_slopes,
diff --git a/vllm/attention/ops/prefix_prefill.py b/vllm/attention/ops/prefix_prefill.py
@@ -38,6 +38,7 @@ def _fwd_kernel(Q,
                 V,
                 K_cache,
                 V_cache,
+                sink_ptr,
                 B_Loc,
                 sm_scale,
                 k_scale,
@@ -126,7 +127,15 @@ def _fwd_kernel(Q,
                 other=0.0)  # [M,D]
 
     # initialize pointer to m and l
-    m_i = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    if sink_ptr is None:
+        m_i = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    else:
+        m_i = tl.load(
+            sink_ptr + tl.full([BLOCK_M], cur_head, dtype=tl.int64),
+            mask=(offs_m < cur_batch_query_len),
+            other=float("-inf"),
+        ).to(dtype=tl.float32)
+
     l_i = tl.full([BLOCK_M], 1.0, dtype=tl.float32)
     acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED], dtype=tl.float32)  # [M,D]
 
@@ -732,7 +741,8 @@ def context_attention_fwd(q,
                           alibi_slopes=None,
                           sliding_window=None,
                           sm_scale=None,
-                          skip_decode=False):
+                          skip_decode=False,
+                          sinks=None):
 
     q_dtype_is_f32 = q.dtype is torch.float32
 
@@ -781,6 +791,7 @@ def context_attention_fwd(q,
         sliding_window = 0
 
     if alibi_slopes is not None:
+        assert sinks is None, "Sinks arg is not supported with alibi"
         # need to reduce num. blocks when using fp32
         # due to increased use of GPU shared memory
         # if q.dtype is torch.float32:
@@ -843,7 +854,7 @@ def context_attention_fwd(q,
     max_seq_len = 0 if max_seq_len is None else max_seq_len
     extra_kargs = {}
     if current_platform.is_rocm():
-        extra_kargs = {"kpack": 2, "waves_per_eu": 2}
+        extra_kargs = {"kpack": 1, "waves_per_eu": 2}
 
     grid = lambda META: (batch, head,
                          triton.cdiv(max_input_len, META["BLOCK_M"]))
@@ -853,6 +864,7 @@ def context_attention_fwd(q,
         v,
         k_cache,
         v_cache,
+        sinks,
         b_loc,
         sm_scale,
         k_scale,
diff --git a/vllm/attention/ops/triton_unified_attention.py b/vllm/attention/ops/triton_unified_attention.py
@@ -52,6 +52,7 @@ def kernel_unified_attention_2d(
         query_ptr,  # [num_tokens, num_query_heads, head_size]
         key_cache_ptr,  # [num_blks, blk_size, num_kv_heads, head_size]
         value_cache_ptr,  # [num_blks, blk_size, num_kv_heads, head_size]
+        sink_ptr,  # [num_query_heads]
         block_tables_ptr,  # [num_seqs, max_num_blocks_per_seq]
         seq_lens_ptr,  # [num_seqs]
         alibi_slopes_ptr,  # [num_query_heads]
@@ -131,7 +132,15 @@ def kernel_unified_attention_2d(
 
     block_table_offset = seq_idx * block_table_stride
 
-    M = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    if sink_ptr is None:
+        M = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    else:
+        M = tl.load(
+            sink_ptr + query_offset_1,
+            mask=query_mask_1,
+            other=float("-inf"),
+        ).to(dtype=tl.float32)
+
     L = tl.full([BLOCK_M], 1.0, dtype=tl.float32)
     acc = tl.zeros([BLOCK_M, HEAD_SIZE_PADDED], dtype=tl.float32)
 
@@ -292,6 +301,7 @@ def kernel_unified_attention_3d(
         query_ptr,  # [num_tokens, num_query_heads, head_size]
         key_cache_ptr,  # [num_blks, num_kv_heads, head_size // x, blk_size, x]
         value_cache_ptr,  # [num_blks, num_kv_heads, head_size, blk_size]
+        sink_ptr,  # [num_query_heads]
         block_tables_ptr,  # [num_seqs, max_num_blocks_per_seq]
         seq_lens_ptr,  # [num_seqs]
         alibi_slopes_ptr,  # [num_query_heads]
@@ -383,7 +393,15 @@ def kernel_unified_attention_3d(
 
     block_table_offset = seq_idx * block_table_stride
 
-    M = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    if sink_ptr is None or segm_idx != 0:
+        M = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    else:
+        M = tl.load(
+            sink_ptr + query_offset_1,
+            mask=query_mask_1,
+            other=float("-inf"),
+        ).to(dtype=tl.float32)
+
     L = tl.full([BLOCK_M], 1.0, dtype=tl.float32)
     acc = tl.zeros([BLOCK_M, HEAD_SIZE_PADDED], dtype=tl.float32)
 
@@ -627,6 +645,8 @@ def unified_attention(
     v_descale,
     alibi_slopes=None,
     qq_bias=None,
+    # Optional tensor for sinks
+    sinks=None,
 ):
     assert causal, "Only causal attention is supported"
     assert q_descale is None, "Q scales not supported"
@@ -635,6 +655,10 @@ def unified_attention(
     assert q.element_size() >= 2 or block_size >= 32, \
         "Block size must be at least 32 for fp8"
 
+    if sinks is not None:
+        assert sinks.shape[0] == q.shape[1], \
+        "Sinks must be num_query_heads size"
+
     use_alibi_slopes = alibi_slopes is not None
     use_qq_bias = qq_bias is not None
 
@@ -669,6 +693,7 @@ def unified_attention(
             query_ptr=q,
             key_cache_ptr=k,
             value_cache_ptr=v,
+            sink_ptr=sinks,
             block_tables_ptr=block_table,
             seq_lens_ptr=seqused_k,
             alibi_slopes_ptr=alibi_slopes,
@@ -741,6 +766,7 @@ def unified_attention(
                 query_ptr=q,
                 key_cache_ptr=k,
                 value_cache_ptr=v,
+                sink_ptr=sinks,
                 block_tables_ptr=block_table,
                 seq_lens_ptr=seqused_k,
                 alibi_slopes_ptr=alibi_slopes,
diff --git a/vllm/envs.py b/vllm/envs.py
@@ -17,6 +17,7 @@
     LD_LIBRARY_PATH: Optional[str] = None
     VLLM_USE_TRITON_FLASH_ATTN: bool = True
     VLLM_V1_USE_PREFILL_DECODE_ATTENTION: bool = False
+    VLLM_USE_AITER_UNIFIED_ATTENTION: bool = False
     VLLM_FLASH_ATTN_VERSION: Optional[int] = None
     LOCAL_RANK: int = 0
     CUDA_VISIBLE_DEVICES: Optional[str] = None
@@ -151,6 +152,8 @@
     VLLM_LOOPBACK_IP: str = ""
     VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
     VLLM_ENABLE_RESPONSES_API_STORE: bool = False
+    VLLM_USE_TRTLLM_CONTEXT_ATTENTION: bool = False
+    VLLM_USE_TRTLLM_DECODE_ATTENTION: bool = False
 
 
 def get_default_cache_root():
@@ -326,6 +329,12 @@ def get_vllm_port() -> Optional[int]:
     (os.getenv("VLLM_V1_USE_PREFILL_DECODE_ATTENTION", "False").lower() in
      ("true", "1")),
 
+    # Use AITER triton unified attention for V1 attention
+    "VLLM_USE_AITER_UNIFIED_ATTENTION":
+    lambda:
+    (os.getenv("VLLM_USE_AITER_UNIFIED_ATTENTION", "False").lower() in
+     ("true", "1")),
+
     # Force vllm to use a specific flash-attention version (2 or 3), only valid
     # when using the flash-attention backend.
     "VLLM_FLASH_ATTN_VERSION":
@@ -1022,9 +1031,13 @@ def get_vllm_port() -> Optional[int]:
     "VLLM_USE_CUDNN_PREFILL":
     lambda: bool(int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))),
 
-    # If set to 1, use the TRTLLM Attention backend in flashinfer.
-    "VLLM_USE_TRTLLM_ATTENTION":
-    lambda: os.getenv("VLLM_USE_TRTLLM_ATTENTION", None),
+    # If set to 1, use the TRTLLM Context Attention backend in flashinfer.
+    "VLLM_USE_TRTLLM_CONTEXT_ATTENTION":
+    lambda: bool(int(os.getenv("VLLM_USE_TRTLLM_CONTEXT_ATTENTION", "0"))),
+
+    # If set to 1, use the TRTLLM Decode Attention backend in flashinfer.
+    "VLLM_USE_TRTLLM_DECODE_ATTENTION":
+    lambda: bool(int(os.getenv("VLLM_USE_TRTLLM_DECODE_ATTENTION", "0"))),
 
     # Controls garbage collection during CUDA graph capture.
     # If set to 0 (default), enables GC freezing to speed up capture time.
diff --git a/vllm/v1/attention/backends/flash_attn.py b/vllm/v1/attention/backends/flash_attn.py
@@ -373,6 +373,7 @@ def __init__(
         logits_soft_cap: Optional[float] = None,
         attn_type: AttentionType = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[str] = None,
+        sinks: Optional[torch.Tensor] = None,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
@@ -410,6 +411,14 @@ def __init__(
             raise NotImplementedError(
                 "FlashAttention does not support fp8 kv-cache on this device.")
 
+        self.sinks = sinks
+        if self.sinks is not None:
+            assert self.vllm_flash_attn_version == 3, (
+                "Sinks are only supported in FlashAttention 3")
+            assert self.sinks.shape[0] == num_heads, (
+                "Sinks must have the same number of heads as the number of "
+                "heads in the layer")
+
     def forward(
         self,
         layer: torch.nn.Module,
@@ -534,6 +543,7 @@ def forward(
                 k_descale=layer._k_scale.expand(descale_shape),
                 v_descale=layer._v_scale.expand(descale_shape),
                 num_splits=attn_metadata.max_num_splits,
+                s_aux=self.sinks,
             )
             return output
 
diff --git a/vllm/v1/attention/backends/triton_attn.py b/vllm/v1/attention/backends/triton_attn.py
diff --git a/vllm/v1/attention/backends/utils.py b/vllm/v1/attention/backends/utils.py
