Skip to main content

Stack target

Vision-language optimization sweeps three serving variants (baseline, balanced, throughput) tuned for image-bearing traffic. Every variant runs the multimodal prefix cache and a capped concurrency limit so image-heavy requests do not OOM.
TechniqueTriggerWinPlan
Multimodal prefix cacheRepeated image or document inputsFaster repeated-image promptsAll plans
Multimodal concurrency capAll vision-language deploymentsOOM protection for image-heavy requestsAll plans
Per-family sequence ceilingKnown vision-language familySafer defaults by model sizeAll plans
Throughput tuning (batch + chunked prefill)Bulk Q&A or OCR pipelinesHigher batched throughput within OOM-safe limitsAll plans
Image-bearing benchmark requestsBenchmark sweepMeasures real image-encoder costAll plans

Public contract

Vision-language optimization uses vision-specific serving techniques. It does not present generic LLM serving work as runnable vision-language optimization unless the selected runtime can actually apply it to image-bearing requests. RunInfra validates model family, runtime, image support, and benchmark evidence before promoting a vision-language variant.

Detection, routing, and application

DecisionSelection ruleRuntime behavior
Vision-language modalityPipeline model and request shapeUses the vision-language optimization recipe
Family detectionModel architecture and configChooses family-specific safety caps for Qwen-VL, LLaVA, Mllama, InternVL, Phi-Vision, Idefics, PaliGemma, Pixtral, Molmo, MiniCPM-V, and related families
Prefix cacheEvery vision-language variantEnables multimodal prefix caching on all variants
Concurrency capModel size and image loadCaps concurrent multimodal requests to prevent OOM
Throughput tuningThroughput variantRaises batch size and chunked prefill within OOM-safe sequence limits
Image-bearing benchmarkBenchmark profileMeasures prompt latency with real image inputs

Verification

  1. The optimization plan should name the vision-language technique and the selected image workload.
  2. Benchmarks should include image-bearing prompts rather than text-only substitutions.
  3. Runtime evidence should preserve multimodal cache, concurrency, and sequence-limit settings.
  4. Quality checks should preserve vision-specific signals rather than collapsing them into generic text quality.

Rollback

constraints.customerOverrides = {
  skipTechniques: ["multimodal_prefix_cache"],
};
Skipping multimodal prefix cache is rarely the right choice because it is normally quality-neutral. Use this override only when a deployment-specific issue requires it.

Availability notes

  • Per-family sequence ceilings are available for common vision-language families.
  • The multimodal prefix cache, concurrency cap, and sequence ceiling apply on every plan. There is no separate paid-tier gate for vision-language serving knobs today.
  • OCR-heavy workloads should keep the model at its native precision and rely on the throughput variant for batched gains, since aggressive batching can shift quality on document tasks.