[ROCm][Perf] Add opt-in a4w4 (FP4 activation) MoE for DeepSeek V4.1 on AITER (#58819)

Signed-off-by: fai <fangzhouai@gmail.com>
Signed-off-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Co-authored-by: Cursor Agent (Claude Sonnet 5) <cursoragent@cursor.com>
This commit is contained in:
Fangzhou Ai
2026-09-30 15:07:44 +08:00
committed by GitHub
co-authored by Cursor Agent
parent e5f3d08d72
commit 77e8b02c8d
8 changed files with 204 additions and 11 deletions
@@ -220,6 +220,7 @@ class Mxfp4OnlineMoEMethod(OnlineMoEMethodBase):
w2_weight_scale=w2_scale,
w13_bias=w13_bias,
w2_bias=w2_bias,
use_separated_a4w4=self.moe.use_mxfp4_w4a4_dsv4,
)
)