diff --git a/src/post_training/slurm/job_llamafactory.sh.jinja b/src/post_training/slurm/job_llamafactory.sh.jinja
index 088ba26..b241b8e 100644
--- a/src/post_training/slurm/job_llamafactory.sh.jinja
+++ b/src/post_training/slurm/job_llamafactory.sh.jinja
@@ -43,7 +43,7 @@ export WORLD_SIZE=$(( SLURM_NNODES * GPUS_PER_NODE ))
 
 # NCCL tuning for multi-node stability
 export NCCL_IB_TIMEOUT=120
-export NCCL_DEBUG=INFO
+export NCCL_DEBUG=WARN
 export CUDA_DEVICE_MAX_CONNECTIONS=1
 export OMP_NUM_THREADS=1
 
diff --git a/src/post_training/slurm/job_trl_container.sh.jinja b/src/post_training/slurm/job_trl_container.sh.jinja
index 6269035..c963f3f 100644
--- a/src/post_training/slurm/job_trl_container.sh.jinja
+++ b/src/post_training/slurm/job_trl_container.sh.jinja
@@ -43,7 +43,7 @@ export WORLD_SIZE=$(( SLURM_NNODES * GPUS_PER_NODE ))
 
 # NCCL tuning for multi-node stability
 export NCCL_IB_TIMEOUT=120
-export NCCL_DEBUG=INFO
+export NCCL_DEBUG=WARN
 # CUDA_DEVICE_MAX_CONNECTIONS=1 is a Megatron-LM tensor-parallel flag that
 # serializes CUDA streams and hurts ZeRO-2 overlap_comm — do not set it.
 export OMP_NUM_THREADS=1