minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Fix] Disable FlashInfer allreduce fusion in Nemotron-3-Nano lm-eval test (#30945)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-07-12 15:39:59 -05:00
committed by GitHub
parent 5ba3c5147e
commit 7a82178277
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/models_e2e/test_nvidia_nemotron_3_nano.py
+5
View File
@@ -27,6 +27,11 @@ class TestNvidiaNemotron3Nano30BFP8(LMEvalMixin, DefaultServerBase):
other_args = [ other_args = [
"--tp-size", "--tp-size",
"2", "2",
# FlashInfer trtllm allreduce fusion on flashinfer 0.6.14 degrades
# gsm8k for this model (strict-match 0.84 -> 0.78); the ground truth
# was calibrated with fusion disabled. Keep it off until the fusion
# numerics regression is resolved.
"--enforce-disable-flashinfer-allreduce-fusion",
] + NEMOTRON_3_NANO_THINKING_ARGS ] + NEMOTRON_3_NANO_THINKING_ARGS