[ROCm][Bugfix][MLA] Fix FP8 MLA prefill unpad for interleave-padded heads (#60907)

Signed-off-by: Lucas Santos <Lucas.Santos@amd.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
This commit is contained in:
lucas-santos-amd
2026-10-14 14:34:28 -04:00
committed by GitHub
co-authored by Douglas Lehr
parent d1d8e6c096
commit e534024e49
2 changed files with 19 additions and 10 deletions
@@ -2317,10 +2317,12 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
#
# Exact, not approximate: after kv_b_proj gqa_ratio is 1, so q, k and v
# all carry num_heads heads and attention is independent per head.
# Padding all three identically makes padded head j a duplicate of real
# head j % num_heads, so the real heads [0:num_heads] are bit-identical
# to the unpadded result. Same argument as the decode path; only the
# target width differs.
# Padding all three identically makes every padded head a duplicate of
# a real head, so the real heads are bit-identical to the unpadded
# result. get_mla_padded_q repeat_interleaves when the target is a
# multiple of num_heads (real heads at stride target // num_heads) and
# tiles otherwise (real heads [0:num_heads]). Same argument as the
# decode path; only the target width differs.
_real_nhead = self.num_heads
nhead = AiterMLAHelper.get_fp8_prefill_num_heads(_real_nhead)
_pad = nhead != _real_nhead
@@ -2408,7 +2410,12 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
)
if _pad:
out.view(total_q, _real_nhead, v_head_dim).copy_(out_3d[:, :_real_nhead, :])
real_out = (
out_3d[:, :: nhead // _real_nhead, :]
if nhead % _real_nhead == 0
else out_3d[:, :_real_nhead, :]
)
out.view(total_q, _real_nhead, v_head_dim).copy_(real_out)
def forward_mha(
self,