From 7a82178277ee2a6afc894dad59291c973445bd1c Mon Sep 17 00:00:00 2001
From: Liangsheng Yin <hnyls2002@gmail.com>
Date: Sun, 12 Jul 2026 15:39:59 -0500
Subject: [PATCH] [Fix] Disable FlashInfer allreduce fusion in Nemotron-3-Nano
 lm-eval test (#30945)

---
 test/registered/models_e2e/test_nvidia_nemotron_3_nano.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/test/registered/models_e2e/test_nvidia_nemotron_3_nano.py b/test/registered/models_e2e/test_nvidia_nemotron_3_nano.py
index 87e2f708f..0813eaf9f 100644
--- a/test/registered/models_e2e/test_nvidia_nemotron_3_nano.py
+++ b/test/registered/models_e2e/test_nvidia_nemotron_3_nano.py
@@ -27,6 +27,11 @@ class TestNvidiaNemotron3Nano30BFP8(LMEvalMixin, DefaultServerBase):
     other_args = [
         "--tp-size",
         "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