Skip to content

[BugFix] Fix FI accuracy issue when used for MLA prefill #26063

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Oct 2, 2025
+9 −2
Merged

[BugFix] Fix FI accuracy issue when used for MLA prefill #26063

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
af93f80
fix FI accuracy issue
LucasWilkinson Oct 2, 2025
6b35efb
Update vllm/v1/attention/backends/mla/common.py
LucasWilkinson Oct 2, 2025
8d58e0a
Merge branch 'main' into lwilkinson/fix-fi-accuracy-issue
mgoin Oct 2, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 9 additions & 2 deletions vllm/v1/attention/backends/mla/common.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1211,13 +1211,18 @@ def _run_prefill_new_tokens_fi(self, prefill: MLACommonPrefillMetadata, q,
k, v, return_softmax_lse):
assert isinstance(prefill, FlashInferPrefillMetadata)
assert prefill.prefill_main is not None
return prefill.prefill_main.run(
ret = prefill.prefill_main.run(
q=q,
k=k,
v=v,
return_lse=return_softmax_lse,
)

if isinstance(ret, tuple):
# Convert from (q_len, num_heads) to (num_heads, q_len)
return ret[0], ret[1].transpose(0, 1).contiguous()
return ret

def _run_prefill_new_tokens_cudnn(self, prefill: MLACommonPrefillMetadata,
q, k, v, return_softmax_lse):
assert isinstance(prefill, CudnnPrefillMetadata)
Expand Down Expand Up @@ -1260,12 +1265,14 @@ def _run_prefill_context_chunk_fa(self, prefill: MLACommonPrefillMetadata,
def _run_prefill_context_chunk_fi(self, prefill: MLACommonPrefillMetadata,
chunk_idx: int, q, k, v):
assert isinstance(prefill, FlashInferPrefillMetadata)
return prefill.prefill_chunks[chunk_idx].run(
attn_out, lse = prefill.prefill_chunks[chunk_idx].run(
q=q,
k=k,
v=v,
return_lse=True,
)
# Convert from (q_len, num_heads) to (num_heads, q_len)
return attn_out, lse.transpose(0, 1).contiguous()

def _run_prefill_context_chunk_cudnn(self,
prefill: MLACommonPrefillMetadata,
Expand Down