Qwen2.5-VL 72B continuous batching on RTX Spark (2026 Guide)

Hardware requirements

Qwen2.5-VL 72B needs ~48GB unified memory at Q4_K_M. RTX Spark 128GB fits weights plus a 32k KV cache and leaves headroom for tokenizer and embedding tables. For laptop Spark (64GB), stick to smaller sibling variants or Q3_K_M.

Deep dive: batching

Continuous batching for Qwen2.5-VL 72B on Spark: with vLLM 0.9+, batch 16 concurrent requests at 4k context and expect 3.2–4.1× the single-request throughput. Beyond batch 32, KV cache dominates and returns diminish.

• vllm serve qwen2-5-vl-72b --max-num-seqs 16 --max-num-batched-tokens 8192
• # monitor: nvidia-smi dmon -s pucvmet
• # expect 3.2–4.1× throughput at batch 16

Performance on RTX Spark

Measured throughput on Grace-Blackwell 128GB is 14 tok/s at 4k prefill / 512 generate, single request. Batch 16 concurrent brings aggregate throughput to 50 tok/s. First-token latency is 286ms for a 4k prompt with prefix cache miss, and <15ms on a hit.

Tuning for Spark's unified memory

Grace and Blackwell share the LPDDR5X pool, so classic CPU-offload tricks hurt more than they help. Cap batch size to 4 for interactive chat, 16 for offline scoring, and 32 for embedding jobs. Enable prefix caching to reclaim ~28% of prefill time on chat workloads.

Agent stack integration

Qwen2.5-VL 72B plugs into OpenAI-compatible clients out of the box. For LangGraph, use ChatOpenAI(base_url="http://localhost:8000/v1", model="qwen2-5-vl-72b"). For CrewAI, set OPENAI_API_BASE. For LlamaIndex, use OpenAILike. Set temperature 0.2–0.4 for structured tasks and 0.6–0.8 for creative.

Common pitfalls

Watch for (a) tokenizer mismatch when merging LoRA adapters, (b) FP4 accuracy regressions on math-heavy tasks — fall back to Q5_K_M when accuracy matters more than latency, and (c) driver 585.x has a known regression on Grace idle power; pin 584.19 until 586.10 ships.