diff --git a/docs/design/attention_backends.md b/docs/design/attention_backends.md
index ed9f0d3016..329a4aacfb 100644
--- a/docs/design/attention_backends.md
+++ b/docs/design/attention_backends.md
@@ -177,7 +177,7 @@ Priority is **1 = highest** (tried first).
 | `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |
 | `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
 | `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder Only | Any |
-| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ✅ | ✅ | ❌ | ❌ | Decoder | N/A |
 | `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto` | %16 | Any | ✅ | ❌ | ✅ | ❌ | All | N/A |
 | `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder, Encoder Only | N/A |
 | `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ❌ | ✅ | ❌ | All | Any |
diff --git a/vllm/_aiter_ops.py b/vllm/_aiter_ops.py
index 2c9b939fa2..5a8b690433 100644
--- a/vllm/_aiter_ops.py
+++ b/vllm/_aiter_ops.py
@@ -2394,6 +2394,7 @@ class rocm_aiter_ops:
         alibi_slopes: torch.Tensor | None = None,
         return_lse: bool = False,
         out: torch.Tensor | None = None,
+        sink_ptr: torch.Tensor | None = None,
     ):
         """
         Flash attention with variable length sequences.
@@ -2422,6 +2423,7 @@ class rocm_aiter_ops:
             alibi_slopes=alibi_slopes,
             return_lse=return_lse,
             out=out,
+            sink_ptr=sink_ptr,
         )
 
     @staticmethod
diff --git a/vllm/v1/attention/backends/rocm_aiter_fa.py b/vllm/v1/attention/backends/rocm_aiter_fa.py
index d0cc011fe4..a9fa45debc 100644
--- a/vllm/v1/attention/backends/rocm_aiter_fa.py
+++ b/vllm/v1/attention/backends/rocm_aiter_fa.py
@@ -709,6 +709,11 @@ class AiterFlashAttentionMetadataBuilder(
 
 class AiterFlashAttentionBackend(AttentionBackend):
     supported_dtypes: ClassVar[list[torch.dtype]] = [torch.float16, torch.bfloat16]
+
+    @classmethod
+    def supports_sink(cls) -> bool:
+        return True
+
     supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = [
         "auto",
         "float16",
@@ -788,6 +793,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
         logits_soft_cap: float | None = None,
         attn_type: AttentionType = AttentionType.DECODER,
         kv_sharing_target_layer_name: int | None = None,
+        sinks: torch.Tensor | None = None,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
@@ -806,6 +812,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
             logits_soft_cap = 0.0
         self.logits_soft_cap = logits_soft_cap
         self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
+        self.sinks = sinks
 
         assert self.num_heads % self.num_kv_heads == 0
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
@@ -878,6 +885,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
             alibi_slopes=self.alibi_slopes,
             return_lse=False,
             out=output,
+            sink_ptr=self.sinks,
         )
 
     def extend_forward(
@@ -927,6 +935,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
             window_size=self.sliding_window,
             alibi_slopes=self.alibi_slopes,
             return_lse=True,
+            sink_ptr=self.sinks,
         )
         assert attn_metadata.extend_metadata is not None
         chunk_context_metadata = attn_metadata.extend_metadata.chunk_context_metadata
@@ -974,6 +983,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
                 window_size=self.sliding_window,
                 alibi_slopes=self.alibi_slopes,
                 return_lse=True,
+                sink_ptr=self.sinks,
             )
             if chunked_output is None:
                 chunked_output = suf_out
@@ -1092,6 +1102,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
                     window_size=self.sliding_window,
                     alibi_slopes=self.alibi_slopes,
                     out=output_actual_tokens[num_decode_tokens + num_extend_tokens :],
+                    sink_ptr=self.sinks,
                 )
 
             # calculate for extends
@@ -1136,11 +1147,17 @@ class AiterFlashAttentionImpl(AttentionImpl):
                 assert attn_metadata.decode_metadata is not None
                 decode_max_query_len = attn_metadata.decode_metadata.max_query_len
 
-                # Multi-token speculative decode path.
-                if decode_max_query_len > 1:
+                # Use unified_attention for speculative decoding (multi-token),
+                # sliding window, or sinks
+                # (pa_fwd_asm and paged_attention_v1 don't support sinks)
+                if (
+                    self.sliding_window[0] != -1
+                    or decode_max_query_len > 1
+                    or self.sinks is not None
+                ):
                     assert not rocm_aiter_ops.is_shuffle_kv_cache_enabled(), (
-                        "Shuffle KV cache layout is not supported with "
-                        "speculative decoding (multi-token decode)."
+                        "Shuffle KV cache layout is not supported with sliding "
+                        "window, sinks, or speculative decoding (multi-token decode)."
                     )
                     if not attn_metadata.causal:
                         from aiter.ops.triton.attention.mha_v3 import (
@@ -1207,6 +1224,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
                             q_descale=None,
                             k_descale=layer._k_scale.expand(descale_shape),
                             v_descale=layer._v_scale.expand(descale_shape),
+                            sinks=self.sinks,
                         )
                     return