groq-guide
Ultra-low-latency inference on Groq's LPU hardware — when tokens-per-second is the product requirement.
- License 许可
- MIT
- 最近更新
- 2026-11-07
- Stars
- ★ 0
- Forks
- 1
01 前言 Front Matter
- name
- groq-guide
- category
- ai-research
- description
- Ultra-low-latency inference on Groq's LPU hardware — when tokens-per-second is the product requirement.
03 SKILL.md 正文
Overview
Groq runs language models on its own custom silicon — Language Processing Units (LPUs) — delivering exceptionally high tokens-per-second and low latency for supported open models. The value proposition is simple and sharp: when speed is the binding constraint, Groq's hardware-tuned inference is among the fastest available via API.
The tradeoff is model selection: Groq supports a curated set of models compiled for its hardware, not the full open-model universe. You trade breadth for velocity. For applications where latency defines the experience — voice agents, real-time copilots, interactive tools, agents making many sequential calls — that trade is often worth it.
Use Groq when you've measured your latency requirements and they demand hardware-level speed. It's a specialist tool: unmatched at its specialty, narrow outside it.
When to use
- Real-time applications: voice agents, live copilots, interactive chat where latency is the UX.
- Agent loops with many sequential model calls — per-call latency compounds.
- High-throughput batch processing where tokens/sec per dollar wins.
- Demos and prototypes where responsiveness sells the experience.
- Any workload where you can use a Groq-supported model and speed is the top requirement.
- Latency-sensitive evals and benchmarking.
Core concepts
- LPU hardware: purpose-built chips for LLM inference — deterministic, high-throughput execution. The speed comes from hardware/software co-design, not just optimization tricks.
- Curated model set: models compiled and supported for LPU execution. Check current availability — the set grows over time but remains curated, not comprehensive.
- Tokens-per-second leadership: the headline metric — Groq typically leads on raw generation speed for supported models. Validate on your prompt shapes.
- Low time-to-first-token: fast first tokens matter for interactive UX; Groq's architecture delivers here too. Measure TTFT for your use case.
- API compatibility: OpenAI-compatible API surface — swapping in Groq is usually a base-URL and key change, making evaluation cheap.
- Rate limits and tiers: generous free tiers for development; paid tiers for production throughput. Understand the rate limits for your expected concurrency.
- Deterministic performance: LPU execution is highly deterministic — less latency variance than GPU-based serving. Useful for SLO planning.
- Model versioning: supported models update over time; pin what's pinnable and re-validate on changes.
Practical workflow
- Confirm model availability. Check that a model meeting your quality bar is supported on Groq. If none is, Groq's speed doesn't help — quality bar first.
- Benchmark your workload. Measure TTFT, tokens/sec, and end-to-end latency on your actual prompts. Confirm the speed advantage holds for your shapes (long contexts, structured outputs).
- Evaluate quality parity. Run your eval set — ensure the Groq-served model meets your quality bar. Hardware-optimized serving should be behavior-preserving; verify, don't assume.
- Design for the speed. Ultra-low latency enables new UX: streaming voice, real-time collaboration, rapid agent loops. Don't just port a slow design — exploit the speed.
- Plan rate limits. Map your production concurrency to Groq's rate limits and tiers. Load-test at expected peak; arrange capacity before launch.
- Build fallbacks. For production, configure fallback providers — speed specialists are still single points of failure.
- Monitor latency in production. Track TTFT and tokens/sec on live traffic; alert on regressions. Deterministic hardware should give stable numbers — instability signals a problem.
Checklist for Groq in production:
- Supported model clears your quality bar on your evals.
- Latency benchmarked on your prompts (TTFT + throughput).
- Rate limits mapped to production concurrency.
- Fallback providers configured.
- Production latency monitored with alerts.
Common pitfalls
- Speed without a quality bar. Choosing Groq's fastest model instead of the slowest model that meets your quality needs. Quality gates the choice; speed ranks among qualifiers.
- Model set mismatch. Designing around a model Groq doesn't support, then compromising quality to get the speed. Check availability first.
- Ignoring rate limits. Free-tier limits in development masking production needs. Plan capacity for launch traffic, not dev traffic.
- No fallback. Groq as a single point of failure for a latency-critical product. Incidents happen to everyone.
- Over-engineering for speed you don't need. Batch/offline workloads paying a premium (in model choice constraints) for latency they can't use. Match the tool to the requirement.
- Assuming quality parity. Hardware-optimized inference is usually behavior-preserving, but "usually" isn't a guarantee — run your evals.
- Long-context surprises. Speed advantages can narrow at very long contexts. Benchmark at your actual context lengths.
- Latency SLOs without monitoring. Assuming deterministic hardware means no monitoring needed. Monitor anyway — networks, queues, and rate limits still vary.
04 同仓库更多 Skills
web-scrape
Turn web pages into LLM-ready markdown via the Firecrawl API (renders JS, anti-bot handling). Use when browser_open/fetch fails on JS-heavy pages, or when you need clean markdown/structured extraction from a site. Requires FIRECRAWL_API_KEY.
design-taste-frontend
tasteskill v2, the anti-slop frontend skill. Art direction for websites, distilled from a study of 247 curated gallery sites (One Page Love, Awwwards Sites of the Day). Deals researched directions (idea, composition, type, palette, photographs, section blueprints) from a library, enforces a strict no-slop house style with a real-browser inspector, and runs a look-and-fix loop on screenshots. Use this whenever you build, design, redesign, restyle, polish or review a landing page, marketing site, portfolio, product page, one-pager, hero or any public-facing web page, in any stack (HTML/CSS, React, Next.js, Vue, Svelte, Astro, Tailwind) — also when the user says a page looks generic, templated, "AI-made", sloppy, or asks to make it look designed, premium, award-level or less like every other site. Not for dashboards or app screens.
test-driven-development
Use when implementing any feature or bugfix, before writing implementation code
systematic-debugging
Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes
requesting-code-review
Use when completing tasks, implementing major features, or before merging to verify work meets requirements
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).