[Bugfix] Fix quant RMS norm fusion for quantization with TMA-aligned scales (#33255)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
ElizaWszola
2026-02-17 23:35:04 -08:00
committed by GitHub
co-authored by Luka Govedič
parent a49ea5a58f
commit a88b3be7c4
12 changed files with 234 additions and 75 deletions
@@ -924,7 +924,16 @@ def per_token_group_quant_fp8(
# TODO(bnell): this causes some fp8 moe test to fail.
if current_platform.is_cuda() and x.is_contiguous():
torch.ops._C.per_token_group_fp8_quant(
x, x_q, x_s, group_size, eps, fp8_min, fp8_max, use_ue8m0
x,
x_q,
x_s,
group_size,
eps,
fp8_min,
fp8_max,
use_ue8m0,
column_major_scales,
tma_aligned_scales,
)
return x_q, x_s