diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index e9af1ce19..e08778cc5 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -198,7 +198,7 @@ gptoss-fp4-b200-trt:
     - { tp: 8, conc-start: 4, conc-end: 8 }
 
 gptoss-fp4-b200-vllm:
-  image: vllm/vllm-openai:v0.10.2
+  image: vllm/vllm-openai:v0.11.0
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: b200
@@ -228,7 +228,7 @@ gptoss-fp4-b200-vllm:
     - { tp: 8, conc-start: 4, conc-end: 4 }
 
 gptoss-fp4-h100-vllm:
-  image: vllm/vllm-openai:v0.10.2
+  image: vllm/vllm-openai:v0.11.0
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: h100
@@ -286,7 +286,7 @@ gptoss-fp4-h200-trt:
     - { tp: 8, ep: 8, dp-attn: false, conc-start: 4, conc-end: 8 }
 
 gptoss-fp4-h200-vllm:
-  image: vllm/vllm-openai:v0.10.2
+  image: vllm/vllm-openai:v0.11.0
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: h200
diff --git a/benchmarks/gptoss_fp4_h100_docker.sh b/benchmarks/gptoss_fp4_h100_docker.sh
index eceff904f..a8bb57c16 100644
--- a/benchmarks/gptoss_fp4_h100_docker.sh
+++ b/benchmarks/gptoss_fp4_h100_docker.sh
@@ -9,6 +9,7 @@
 # CONC
 
 cat > config.yaml << EOF
+compilation-config: '{"cudagraph_mode":"PIECEWISE"}'
 async-scheduling: true
 no-enable-prefix-caching: true
 cuda-graph-sizes: 2048
@@ -20,7 +21,7 @@ export PYTHONNOUSERSITE=1
 
 set -x
 vllm serve $MODEL --host=0.0.0.0 --port=$PORT \
---config=config.yaml \
+--config config.yaml \
 --gpu-memory-utilization=0.9 \
 --tensor-parallel-size=$TP \
 --max-num-seqs=$CONC  \
diff --git a/benchmarks/gptoss_fp4_h100_slurm.sh b/benchmarks/gptoss_fp4_h100_slurm.sh
index f2b5e3a05..d2819b5b3 100644
--- a/benchmarks/gptoss_fp4_h100_slurm.sh
+++ b/benchmarks/gptoss_fp4_h100_slurm.sh
@@ -17,6 +17,7 @@
 echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
 
 cat > config.yaml << EOF
+compilation-config: '{"cudagraph_mode":"PIECEWISE"}'
 async-scheduling: true
 no-enable-prefix-caching: true
 cuda-graph-sizes: 2048
@@ -29,7 +30,7 @@ export TORCH_CUDA_ARCH_LIST="9.0"
 
 set -x
 PYTHONNOUSERSITE=1 vllm serve $MODEL --host=0.0.0.0 --port=$PORT \
---config=config.yaml \
+--config config.yaml \
 --gpu-memory-utilization=0.9 \
 --tensor-parallel-size=$TP \
 --max-num-seqs=$CONC  \
diff --git a/benchmarks/gptoss_fp4_h200_slurm.sh b/benchmarks/gptoss_fp4_h200_slurm.sh
index 23ac0bfa1..f92c60425 100644
--- a/benchmarks/gptoss_fp4_h200_slurm.sh
+++ b/benchmarks/gptoss_fp4_h200_slurm.sh
@@ -31,6 +31,7 @@ fi
 
 # Create config.yaml
 cat > config.yaml << EOF
+compilation-config: '{"cudagraph_mode":"PIECEWISE"}'
 async-scheduling: true
 no-enable-prefix-caching: true
 cuda-graph-sizes: 2048