Qwen3 VL Embedding 2B on Orin

Use vllm on the orin machine to test Qwen/Qwen3-VL-Embedding-2B, verifying both offline embedding and HTTP service startup.

Test Configuration

• Device: orin
• Container image: registry.lazycat.cloud/x/lzc-aipod-vllm:bffa39b-orin
• Model: Qwen/Qwen3-VL-Embedding-2B
• Offline mode: LLM(..., runner="pooling")
• HTTP mode: vllm serve --runner pooling
• VRAM specifications: --gpu-memory-utilization 0.65 --max-model-len 8192 --enforce-eager
• Dataset: datasets/qwen3_vl_embedding.jsonl
• Synthesized resources: scripts/prepare_qwen3_vl_assets.py

Conclusion

• Offline embed() path passed, 16 image-text samples completed embedding.
• Output vector dimension is 2048.
• Offline throughput is 2.4299 items/s, total duration 6.5845s.
• Image-text retrieval example 4/4 top1 match correct.
• HTTP service started successfully, /v1/models ready.

Offline verification result

case	top1
case-jtop	positive-doc
case-openwebui	positive-doc
case-network	positive-doc
case-embedding	positive-doc

Telemetry

• Sample count: 150
• Maximum temperature: 67.156 C
• Maximum power consumption: 56761 mW
• Maximum fan PWM: 35.2941
• Maximum GPU utilization: 99.8%

Output

Remote result directory:

/home/nvidia/bench-results/qwen3-vl-embedding-vllm

Key Files:

/home/nvidia/bench-results/qwen3-vl-embedding-vllm/qwen3_vl_embedding_offline.json
/home/nvidia/bench-results/qwen3-vl-embedding-vllm/qwen3_vl_http_probe.json
/home/nvidia/bench-results/qwen3-vl-embedding-vllm/qwen3_vl_http.log
/home/nvidia/bench-results/qwen3-vl-embedding-vllm/qwen3_vl_models.json
/home/nvidia/bench-results/qwen3-vl-embedding-vllm/telemetry.jsonl
/home/nvidia/bench-results/qwen3-vl-embedding-vllm/telemetry.summary.json