A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 02:00 UTC
[ROCm][Qwen3-32B] Fix AITER MHA accuracy issue cause by #25763 (#28670)
Signed-off-by: Xiake Sun <xiake.sun@amd.com>
Xiake Sun committed
Nov 17, 2025 at 12:52 UTC
60e089f0b90b1fe9b65224b069c953927d1f3b44
1 file changed
+2
-2
vllm/v1/attention/backends/rocm_aiter_fa.py
+2
-2
@@ -729,7 +729,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
729
cu_seqlens_k=attn_metadata.prefill_metadata.query_start_loc,
730
max_seqlen_q=attn_metadata.prefill_metadata.max_query_len,
731
max_seqlen_k=attn_metadata.prefill_metadata.max_seq_len,
732
- min_seqlen_q=attn_metadata.prefill_metadata.min_query_len,
732
+ min_seqlen_q=1,
733
dropout_p=0.0,
734
softmax_scale=self.scale,
735
causal=True,
@@ -759,7 +759,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
759
cu_seqlens_q=attn_metadata.extend_metadata.query_start_loc,
760
max_seqlen_q=attn_metadata.extend_metadata.max_query_len,
761
max_seqlen_k=attn_metadata.extend_metadata.max_seq_len,
762
- min_seqlen_q=attn_metadata.extend_metadata.min_query_len,
762
+ min_seqlen_q=1,
763
block_table=attn_metadata.block_table[
764
num_decodes : num_decodes + num_extends
765
],