Upgrade vLLM to v0.11.2 (#273)

* Upgrade vLLM to v0.11.2

Updated configs:

- Use FP8 kv-cache for GPT-OSS B200.
- Remove "custom_ops" from compilation-config for GPT-OSS.
- Remove "cudagraph_mode" from compilation-config for GPT-OSS.
- Remove VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB env var for
  GPT-OSS.
- Remove deprecated "--disable-log-requests" flag.
- Rename "cuda-graph-sizes" flag.

Signed-off-by: Po-Han Huang <pohanh@nvidia.com>

* make cw runners container writable

* undo make cw runners container writable

* coreweave cleanup

* coreweave cleanup pt 2

---------

Signed-off-by: Po-Han Huang <pohanh@nvidia.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
Co-authored-by: Cameron Quilici <cjquilici@gmail.com>

Author: 3 people

.github/configs/nvidia-master.yaml

@@ -209,7 +209,7 @@ gptoss-fp4-b200-trt:
     - { tp: 8, conc-start: 4, conc-end: 8 }
 
 gptoss-fp4-b200-vllm:
-  image: vllm/vllm-openai:v0.11.0
+  image: vllm/vllm-openai:v0.11.2
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: b200
@@ -240,7 +240,7 @@ gptoss-fp4-b200-vllm:
     - { tp: 8, conc-start: 4, conc-end: 4 }
 
 gptoss-fp4-h100-vllm:
-  image: vllm/vllm-openai:v0.11.0
+  image: vllm/vllm-openai:v0.11.2
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: h100
@@ -300,7 +300,7 @@ 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:v0.11.0
+  image: vllm/vllm-openai:v0.11.2
   model: openai/gpt-oss-120b
   model-prefix: gptoss
   runner: h200

benchmarks/gptoss_fp4_b200_docker.sh

@@ -29,16 +29,16 @@ else
 fi
 
 cat > config.yaml << EOF
-compilation-config: '{"pass_config":{"enable_fi_allreduce_fusion":true,"enable_attn_fusion":true,"enable_noop":true},"custom_ops":["+rms_norm"],"cudagraph_mode":"FULL_AND_PIECEWISE"}'
+kv-cache-dtype: fp8
+compilation-config: '{"pass_config":{"enable_fi_allreduce_fusion":true,"enable_noop":true}}'
 async-scheduling: true
 no-enable-prefix-caching: true
-cuda-graph-sizes: 2048
+max-cudagraph-capture-size: 2048
 max-num-batched-tokens: 8192
 max-model-len: $CALCULATED_MAX_MODEL_LEN
 EOF
 
 export TORCH_CUDA_ARCH_LIST="10.0"
-export VLLM_FLASHINFER_ALLREDUCE_FUSION_THRESHOLDS_MB='{"2":32,"4":32,"8":8}'
 export PYTHONNOUSERSITE=1
 export VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8=1
 
@@ -47,7 +47,7 @@ SERVER_LOG=$(mktemp /tmp/server-XXXXXX.log)
 set -x
 vllm serve $MODEL --host 0.0.0.0 --port $PORT --config config.yaml \
 --gpu-memory-utilization 0.9 --tensor-parallel-size $TP --max-num-seqs 512 \
---disable-log-requests > $SERVER_LOG 2>&1 &
+> $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 
@@ -69,4 +69,4 @@ run_benchmark_serving \
     --num-prompts "$NUM_PROMPTS" \
     --max-concurrency "$CONC" \
     --result-filename "$RESULT_FILENAME" \
-    --result-dir /workspace/
+    --result-dir /workspace/

benchmarks/gptoss_fp4_h100_docker.sh

@@ -12,10 +12,9 @@
 
 
 cat > config.yaml << EOF
-compilation-config: '{"cudagraph_mode":"PIECEWISE"}'
 async-scheduling: true
 no-enable-prefix-caching: true
-cuda-graph-sizes: 2048
+max-cudagraph-capture-size: 2048
 max-num-batched-tokens: 8192
 max-model-len: 10240
 EOF
@@ -29,7 +28,7 @@ vllm serve $MODEL --host=0.0.0.0 --port=$PORT \
 --gpu-memory-utilization=0.9 \
 --tensor-parallel-size=$TP \
 --max-num-seqs=$CONC  \
---disable-log-requests > $SERVER_LOG 2>&1 &
+> $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 
@@ -51,4 +50,4 @@ run_benchmark_serving \
     --num-prompts $(( $CONC * 10 )) \
     --max-concurrency 512 \
     --result-filename "$RESULT_FILENAME" \
-    --result-dir /workspace/
+    --result-dir /workspace/

benchmarks/gptoss_fp4_h100_slurm.sh

@@ -13,10 +13,9 @@
 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
+max-cudagraph-capture-size: 2048
 max-num-batched-tokens: 8192
 max-model-len: 10240
 EOF
@@ -30,7 +29,7 @@ PYTHONNOUSERSITE=1 vllm serve $MODEL --host=0.0.0.0 --port=$PORT \
 --gpu-memory-utilization=0.9 \
 --tensor-parallel-size=$TP \
 --max-num-seqs=$CONC  \
---disable-log-requests > $SERVER_LOG 2>&1 &
+ > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 

benchmarks/gptoss_fp4_h200_slurm.sh

@@ -27,10 +27,9 @@ 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
+max-cudagraph-capture-size: 2048
 max-num-batched-tokens: 8192
 max-model-len: $CALCULATED_MAX_MODEL_LEN
 EOF
@@ -42,7 +41,7 @@ export TORCH_CUDA_ARCH_LIST="9.0"
 
 PYTHONNOUSERSITE=1 vllm serve $MODEL --host 0.0.0.0 --port $PORT --config config.yaml \
  --gpu-memory-utilization 0.9 --tensor-parallel-size $TP --max-num-seqs $CONC  \
- --disable-log-requests > $SERVER_LOG 2>&1 &
+ > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
 

runners/launch_h200-cw.sh

@@ -27,6 +27,6 @@ srun --jobid=$JOB_ID \
 --container-mount-home \
 --container-workdir=/workspace/ \
 --no-container-entrypoint --export=ALL \
-bash benchmarks/${MODEL_CODE}_${PRECISION}_h200${FRAMEWORK_SUFFIX}_slurm.sh
+bash -c 'bash benchmarks/'"${EXP_NAME%%_*}_${PRECISION}"'_h200_slurm.sh; rm -rf /dev/shm/sagemaker_sessions'
 
 scancel $JOB_ID