EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[releases/v0.18.0][bugfix][eplb] remove unnecessary weight_scale wrap behaviour (#7732)

Browse Source 
### What this PR does / why we need it?
This PR simplifies the apply method in w8a8_dynamic.py by removing the
conditional logic that used fused_w1_scale and fused_w2_scale based on
the fused_scale_flag. This redundant wrap behavior leads to EPLB break
in int8 quantization scenarios.

Cherry-picked from #7188. Note that only bugfix lines in that PR are
picked.

Signed-off-by: linfeng-yuan <1102311262@qq.com>
This commit is contained in:
linfeng-yuan
2026-03-30 16:16:03 +08:00
committed by GitHub
parent fdd0726ae4
commit deceefd305
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/quantization/methods/w8a8_dynamic.py
View File

@@ -267,8 +267,8 @@ class AscendW8A8DynamicFusedMoEMethod(AscendMoEScheme):
log2phy=log2phy, log2phy=log2phy,
pertoken_scale=pertoken_scale, pertoken_scale=pertoken_scale,
activation=activation, activation=activation,
w1_scale=[layer.fused_w1_scale] if fused_scale_flag else w1_scale, w1_scale=w1_scale,
w2_scale=[layer.fused_w2_scale] if fused_scale_flag else w2_scale, w2_scale=w2_scale,
) )
) )
if zero_expert_num > 0 and zero_expert_type is not None: if zero_expert_num > 0 and zero_expert_type is not None: