Skip to main content

Why load test

A model can look healthy at low concurrency and still fail under production traffic. Load testing validates scheduler behavior, queueing, memory pressure, and tail latency before a deployment becomes critical. High-concurrency tests are useful for three failure modes:
  1. Scheduler saturation: requests queue once the runtime reaches its in-flight limit.
  2. Connection pressure: client or gateway connection pools become the bottleneck.
  3. VRAM pressure: KV cache or activation memory grows until the runtime must reject or restart work.

Test plan

Start below the expected production peak, then increase concurrency until one metric clearly bends.
StepPurpose
WarmupLoad weights, compile kernels where applicable, and avoid cold-start noise
BaselineMeasure healthy latency and throughput at expected traffic
RampIncrease concurrency in stages
StressHold the highest safe stage long enough to observe p95 and p99 behavior
RecoveryReturn to baseline and confirm the deployment stabilizes

What to report

MetricWhy it matters
p50, p95, p99 latencyShows scheduler health and tail behavior
Successful RPSShows useful throughput under load
Token or vector throughputNormalizes work across request shapes
Peak VRAMCatches memory cliffs before production
OOM thresholdIdentifies the unsafe concurrency region
Success rateReveals dropped requests or backend errors
Runtime settingsConfirms the optimized launch settings were preserved

Failure-mode triage

SymptomLikely causeFix
p99 latency spikes at high concurrencyRuntime queue depth is too lowRaise the in-flight request cap until VRAM approaches the ceiling
RPS plateaus mid-sweepGPU compute is saturatedAdd replicas, add GPUs, or choose a lower-cost latency target
OOM at higher concurrencyKV cache or activations exceed VRAMLower max context, lower concurrency, or move to a larger GPU
Runtime settings do not match the optimized variantDeployment driftRedeploy the selected optimized version
All requests fail immediatelyStartup or routing failureInspect deployment health and request IDs

Modality-specific notes

ModalityLoad-test focus
LLMFirst-token latency, output-token throughput, and KV-cache pressure
EmbeddingsBatch size, vector throughput, and input length distribution
ASRReal-time factor, audio duration mix, and chunk count
TTSTime to first audio chunk and stream stability
Image generationPer-image latency, image count, size, and scheduler behavior
Vision-languageImage-heavy prompts, cache hits, and multimodal concurrency
Use these results to set deployment SLOs and to decide whether to adjust replicas, GPU tier, context length, batch size, or optimization priority.