Multimodal large language models introduce a class of inference problem that standard vLLM profiling playbooks were not designed for. While LLM inference is well-understood, Multimodal MoE models introduce a chaotic new variable: image resolution. In a real-world product catalog, 20x variance in image size disrupts standard batch formation, drives preemption spikes.
This talk presents the profiling methodology and optimization path we followed to achieve the top result among all B200 submissions on the MLPerf Inference benchmark for Qwen3-VL-235B-A22B-Instruct, a 235-billion-parameter mixture-of-experts vision-language model with 22 billion active parameters. The workload is the MLPerf Shopify product catalog: 48,289 real e-commerce images paired with text, classified into structured JSON output. We submitted results on both 8xB200 and 8xH200 nodes running RHEL with CentML-optimized vLLM.

Results
Configuration
Throughput
Context
8×B200 — Server
67.86 samples/sec
#1 among all B200 submissions; 50% faster than top B300 submission
8×B200 — Offline
79.04 samples/sec
#1 B200 result; on par with top B300 submission
8×H200 — Offline
18.22 samples/sec
Only H200 submission for this model
8×H200 — Server
11.05 samples/sec
Only H200 submission for this model

We will present the profiling work that drove these results and how we identified the ViT encoder as the primary throughput constraint, why Shortest Job First scheduling outperformed continuous batching defaults on variable-image workloads, how FP8 multimodal attention was validated without compromising benchmark accuracy, and what the FlashInfer MoE kernel traces revealed about active-parameter utilization under the MLPerf traffic pattern.