From 1a9718085c7980443558db1ff4160c58096a3f0e Mon Sep 17 00:00:00 2001 From: "Mengtao (Martin) Yuan" Date: Fri, 6 Mar 2026 18:12:07 -0800 Subject: [PATCH] Fix CUDA graph decode capture crash in AITER FlashAttention (#36042) Signed-off-by: Martin Yuan Co-authored-by: Martin Yuan --- vllm/v1/attention/backends/rocm_aiter_fa.py | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/vllm/v1/attention/backends/rocm_aiter_fa.py b/vllm/v1/attention/backends/rocm_aiter_fa.py index 9c9da3dfde..b1adaa7240 100644 --- a/vllm/v1/attention/backends/rocm_aiter_fa.py +++ b/vllm/v1/attention/backends/rocm_aiter_fa.py @@ -1152,11 +1152,10 @@ class AiterFlashAttentionImpl(AttentionImpl): decode_max_query_len = attn_metadata.decode_metadata.max_query_len # Use unified_attention for speculative decoding (multi-token) - # or when sliding window is enabled - if self.sliding_window[0] != -1 or decode_max_query_len > 1: + if decode_max_query_len > 1: assert not rocm_aiter_ops.is_shuffle_kv_cache_enabled(), ( - "Shuffle KV cache layout is not supported with sliding " - "window or speculative decoding (multi-token decode)." + "Shuffle KV cache layout is not supported with " + "speculative decoding (multi-token decode)." ) from aiter.ops.triton.unified_attention import ( unified_attention,