[NVIDIA] Update Qwen3.5 STP recipe on B200

benchmarks/qwen3.5_bf16_b200.sh

@@ -19,20 +19,53 @@ nvidia-smi
 
 hf download "$MODEL"
 
+export NCCL_NVLS_ENABLE=1
+export SGL_ENABLE_JIT_DEEPGEMM=false
+export SGLANG_ENABLE_FLASHINFER_GEMM=true
+export PYTHONUNBUFFERED=1
+
 SERVER_LOG=/workspace/server.log
 PORT=${PORT:-8888}
 
-MEM_FRAC_STATIC=0.8
+# Low latency (conc 4,8): recv interval 10; max throughput (conc 16+): recv interval 30
+if [[ $CONC -ge 16 ]]; then
+  SCHEDULER_RECV_INTERVAL=30
+else
+  SCHEDULER_RECV_INTERVAL=10
+fi
+
+MEM_FRAC_STATIC=0.82
+CHUNKED_PREFILL_SIZE=32768
+MAX_PREFILL_TOKENS=32768
+CUDA_GRAPH_MAX_BATCH_SIZE=$CONC
+MAX_RUNNING_REQUESTS=128
+CONTEXT_LENGTH=$((ISL + OSL + 20))
+
+echo "SCHEDULER_RECV_INTERVAL: $SCHEDULER_RECV_INTERVAL, CONC: $CONC, ISL: $ISL, OSL: $OSL"
 
 ps aux
 
 set -x
 PYTHONNOUSERSITE=1 python3 -m sglang.launch_server \
     --model-path=$MODEL \
+    --served-model-name "Qwen/Qwen3.5-397B-A17B" \
     --host=0.0.0.0 \
     --port=$PORT \
+    --trust-remote-code \
     --tensor-parallel-size=$TP \
+    --disable-radix-cache \
     --mem-fraction-static $MEM_FRAC_STATIC \
+    --chunked-prefill-size $CHUNKED_PREFILL_SIZE \
+    --max-prefill-tokens $MAX_PREFILL_TOKENS \
+    --cuda-graph-max-bs $CUDA_GRAPH_MAX_BATCH_SIZE \
+    --max-running-requests $MAX_RUNNING_REQUESTS \
+    --context-length $CONTEXT_LENGTH \
+    --attention-backend trtllm_mha \
+    --moe-runner-backend flashinfer_trtllm \
+    --tokenizer-worker-num 6 \
+    --stream-interval 30 \
+    --scheduler-recv-interval $SCHEDULER_RECV_INTERVAL \
+    --enable-flashinfer-allreduce-fusion \
     > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!

perf-changelog.yaml

@@ -671,3 +671,13 @@
     - "Environment: VLLM_ROCM_USE_AITER=1"
     - "TP=2 and TP=4, concurrency 4-64 for 1k1k, 1k8k, and 8k1k sequence lengths"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/755
+
+- config-keys:
+    - qwen3.5-bf16-b200-sglang
+  description:
+    - "Update Qwen3.5-397B-A17B BF16 SGLang B200 benchmark launch config"
+    - "Image: lmsysorg/sglang:nightly-dev-20260216-d3bae71e"
+    - "Add trtllm_mha attention backend, flashinfer_trtllm MOE runner"
+    - "Add context-length, tokenizer-worker-num, env tuning (NCCL_NVLS_ENABLE, SGLANG_ENABLE_FLASHINFER_GEMM)"
+    - "Set cuda-graph-max-bs to match concurrency, scheduler-recv-interval based on concurrency"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/758