[AMD][WA] force to use gate_mode interleaved to fix tp2/tp4/tp8 acc issue (#27201)

Co-authored-by: wunhuang <wunhuang@amd.com>
This commit is contained in:
kk
2026-06-05 20:18:11 -07:00
committed by GitHub
co-authored by wunhuang
parent bf66b7b6da
commit aa55657e9e
4 changed files with 73 additions and 55 deletions
@@ -76,7 +76,7 @@ MI35X_GPT_OSS_MODELS = [
# to INTERLEAVE, so opt out explicitly here.
env_vars={
"SGLANG_USE_AITER": "1",
"SGLANG_USE_AITER_MOE_GU_ITLV": "0",
"SGLANG_USE_AITER_MOE_GU_ITLV": "1",
},
),
ModelConfig(
@@ -97,7 +97,7 @@ MI35X_GPT_OSS_MODELS = [
],
env_vars={
"SGLANG_USE_AITER": "1",
"SGLANG_USE_AITER_MOE_GU_ITLV": "0",
"SGLANG_USE_AITER_MOE_GU_ITLV": "1",
},
),
]