Groq Performance Tuning

Overview

Maximize Groq's ultra-low-latency LPU inference. Groq delivers sub-100ms token generation; tuning focuses on streaming efficiency, prompt caching, model selection for speed vs quality, and parallel request orchestration.

Prerequisites

• Groq API key with rate limit awareness
• groq-sdk npm package installed
• Understanding of LLM token economics
• Monitoring for TTFT (time to first token)

Instructions

Step 1: Select Optimal Model for Speed

import Groq from 'groq-sdk';

const groq = new Groq({ apiKey: process.env.GROQ_API_KEY });

// Model speed tiers (approximate TTFT):
// llama-3.3-70b-versatile: ~200ms TTFT, best quality
// llama-3.1-8b-instant:    ~80ms TTFT, fastest
// mixtral-8x7b-32768:      ~150ms TTFT, long context  # 32768 = configured value

async function fastCompletion(prompt: string) {
  return groq.chat.completions.create({
    model: 'llama-3.1-8b-instant', // Fastest model
    messages: [{ role: 'user', content: prompt }],
    temperature: 0,       // Deterministic = cacheable
    max_tokens: 256,      // Limit output for speed  # 256 bytes
  });
}

Step 2: Streaming for Perceived Performance

async function streamCompletion(
  messages: any[],
  onToken: (token: string) => void
) {
  const stream = await groq.chat.completions.create({
    model: 'llama-3.3-70b-versatile',
    messages,
    stream: true,
    max_tokens: 1024,  # 1024: 1 KB
  });

  let fullResponse = '';
  for await (const chunk of stream) {
    const token = chunk.choices[0]?.delta?.content || '';
    fullResponse += token;
    onToken(token);
  }
  return fullResponse;
}

Step 3: Semantic Prompt Cache

import { LRUCache } from 'lru-cache';
import { createHash } from 'crypto';

const promptCache = new LRUCache<string, string>({
  max: 500,  # HTTP 500 Internal Server Error
  ttl: 1000 * 60 * 10, // 10 min for deterministic prompts  # 1000: 1 second in ms
});

function hashPrompt(messages: any[], model: string): string {
  return createHash('sha256')
    .update(JSON.stringify({ messages, model }))
    .digest('hex');
}

async function cachedCompletion(messages: any[], model: string) {
  const key = hashPrompt(messages, model);
  const cached = promptCache.get(key);
  if (cached) return cached;

  const response = await groq.chat.completions.create({
    model,
    messages,
    temperature: 0,
  });

  const result = response.choices[0].message.content!;
  promptCache.set(key, result);
  return result;
}

Step 4: Parallel Request Orchestration

async function parallelCompletions(
  prompts: string[],
  concurrency = 5
) {
  const results: string[] = [];

  for (let i = 0; i < prompts.length; i += concurrency) {
    const batch = prompts.slice(i, i + concurrency);
    const batchResults = await Promise.all(
      batch.map(prompt =>
        cachedCompletion(
          [{ role: 'user', content: prompt }],
          'llama-3.1-8b-instant'
        )
      )
    );
    results.push(...batchResults);
  }
  return results;
}

Error Handling

Issue	Cause	Solution
Rate limit 429	Over RPM/TPM quota	Use exponential backoff, batch requests
High TTFT	Using 70b model	Switch to 8b-instant for latency-sensitive tasks
Stream disconnect	Network timeout	Implement reconnection with partial response recovery
Token overflow	max_tokens too high	Set conservative limits, truncate prompts

Examples

Latency Benchmark

async function benchmarkModels(prompt: string) {
  const models = ['llama-3.1-8b-instant', 'llama-3.3-70b-versatile'];

  for (const model of models) {
    const start = performance.now();
    await groq.chat.completions.create({
      model,
      messages: [{ role: 'user', content: prompt }],
      max_tokens: 100,
    });
    console.log(`${model}: ${(performance.now() - start).toFixed(0)}ms`);
  }
}

Resources

• Groq API Documentation
• Groq Rate Limits
• Groq Model Cards

Output

• Configuration files or code changes applied to the project
• Validation report confirming correct implementation
• Summary of changes made and their rationale