Skip to content

Fix MLA FP8 padding simulate MTP2 accuracy #645

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
valarLip merged 1 commit into from
Aug 27, 2025
Merged

Fix MLA FP8 padding simulate MTP2 accuracy #645

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
4 changes: 2 additions & 2 deletions vllm/attention/ops/rocm_aiter_mla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,7 +63,7 @@ def aiter_mla_decode_fwd(
kv_scale = torch.ones([1], dtype=torch.float, device=kv_buffer.device)
batch_size = q_fp8.shape[0]
q_fp8_padded = torch.ones(batch_size * 2, q_fp8.shape[1], q_fp8.shape[2], dtype=q_fp8.dtype, device=q_fp8.device)
q_fp8_padded[::2] = q_fp8
q_fp8_padded[1::2] = q_fp8
qo_indptr_padded = torch.arange(0, (batch_size + 1) * 2, 2, dtype=qo_indptr.dtype, device=qo_indptr.device)
o_padded = torch.empty((o.shape[0] * 2, o.shape[1], o.shape[2]), dtype=o.dtype, device=o.device).fill_(-1)
max_seqlen_q_new = 2
Expand All @@ -88,7 +88,7 @@ def aiter_mla_decode_fwd(
q_scale=q_scale,
kv_scale=kv_scale,
)
o[:] = o_padded[::2] # Extract every second element
o[:] = o_padded[1::2] # Extract every second element
else:
mla_decode_fwd_dispatch(q,
kv_buffer.view(-1, 1, 1, q.shape[-1]),
Expand Down
Loading