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

2026-07-12 15:39:59 -05:00
parent 5ba3c5147e
commit 7a82178277
1 changed files with 5 additions and 0 deletions
@@ -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