biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm][Qwen3-32B] Fix AITER MHA accuracy issue cause by #25763 (#28670)

Browse Source 
Signed-off-by: Xiake Sun <xiake.sun@amd.com>
This commit is contained in:
Xiake Sun
2025-11-17 12:52:11 +08:00
committed by GitHub
parent d64429bb36
commit 60e089f0b9
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/v1/attention/backends/rocm_aiter_fa.py
View File

@@ -729,7 +729,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
cu_seqlens_k=attn_metadata.prefill_metadata.query_start_loc,
max_seqlen_q=attn_metadata.prefill_metadata.max_query_len,
max_seqlen_k=attn_metadata.prefill_metadata.max_seq_len,
min_seqlen_q=attn_metadata.prefill_metadata.min_query_len,
min_seqlen_q=1,
dropout_p=0.0,
softmax_scale=self.scale,
causal=True,
@@ -759,7 +759,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
cu_seqlens_q=attn_metadata.extend_metadata.query_start_loc,
max_seqlen_q=attn_metadata.extend_metadata.max_query_len,
max_seqlen_k=attn_metadata.extend_metadata.max_seq_len,
min_seqlen_q=attn_metadata.extend_metadata.min_query_len,
min_seqlen_q=1,
block_table=attn_metadata.block_table[
num_decodes : num_decodes + num_extends
],