[ROCm][Model][Bugfix] Enable GLM-5.2-MXFP4 on the deepseek_v32 path and fix sparse attention correctness (#51915)

Signed-off-by: Jack Hu <Jack.Hu@amd.com>
Signed-off-by: Jack Hu <jack.hu@amd.com>
Signed-off-by: Douglas Lehr <Doug.Lehr@amd.com>
Co-authored-by: James E T Smith <jamesETsmith@users.noreply.github.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
Co-authored-by: Douglas Lehr <Doug.Lehr@amd.com>
This commit is contained in:
Jack Hu
2026-09-22 18:08:51 -05:00
committed by GitHub
co-authored by James E T Smith Douglas Lehr Douglas Lehr
parent 9646f53064
commit 1c0eee919d
8 changed files with 318 additions and 211 deletions
@@ -1025,9 +1025,7 @@ class MLAAttention(nn.Module, AttentionLayerBase):
mqa_q_pe = mqa_pe_padded
if self.is_aiter_triton_fp4_bmm_enabled:
from aiter.ops.triton.batched_gemm_a16wfp4 import batched_gemm_a16wfp4
mqa_ql_nope = batched_gemm_a16wfp4(
mqa_ql_nope = rocm_aiter_ops.batched_gemm_a16wfp4(
mqa_q_nope,
self.W_K,
self.W_K_scale,