From d60d6ff6fd8a2c348204b59aa6b48c58ff764710 Mon Sep 17 00:00:00 2001
From: Yukun He <23156053+hyukn@users.noreply.github.com>
Date: Wed, 21 Jan 2026 13:53:04 +0800
Subject: [PATCH] [None][fix] Cherry-pick #10715: Disable short profile for
 tunable ops with MERGE strategy (#10844)

Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
---
 tensorrt_llm/_torch/autotuner.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/tensorrt_llm/_torch/autotuner.py b/tensorrt_llm/_torch/autotuner.py
index c78b1a0319..f68d9c7f6c 100644
--- a/tensorrt_llm/_torch/autotuner.py
+++ b/tensorrt_llm/_torch/autotuner.py
@@ -1105,7 +1105,10 @@ class AutoTuner:
 
         disable_short_profile = os.environ.get(
             "TLLM_AUTOTUNER_DISABLE_SHORT_PROFILE", "0") == "1"
-        if fewer_repeat_avg_time > short_profile_threshold_ms and not disable_short_profile:
+
+        # Disable this feature for merged tuning strategy to avoid potential hang due to asymmetric tuning.
+        if fewer_repeat_avg_time > short_profile_threshold_ms and not disable_short_profile \
+            and tuning_config.distributed_tuning_strategy != DistributedTuningStrategy.MERGE:
             # directly use the few repeat estimated time to avoid redundant profiling
             avg_time = fewer_repeat_avg_time
         else: