groq-performance-tuning

What this skill does

# Groq Performance Tuning

## Overview

Maximize Groq's LPU inference speed advantage. Groq already delivers extreme throughput (280-560 tok/s) and low latency (<200ms TTFT), but client-side optimization -- model selection, prompt size, streaming, caching, and parallelism -- determines whether your application fully exploits that speed.

## Groq Speed Benchmarks

| Model | TTFT | Throughput | Context |
|-------|------|-----------|---------|
| `llama-3.1-8b-instant` | ~50ms | ~560 tok/s | 128K |
| `llama-3.3-70b-versatile` | ~150ms | ~280 tok/s | 128K |
| `llama-3.3-70b-specdec` | ~100ms | ~400 tok/s | 128K |
| `meta-llama/llama-4-scout-17b-16e-instruct` | ~80ms | ~460 tok/s | 128K |

TTFT = Time to First Token. Actual values depend on prompt size and server load.

## Instructions

### Step 1: Choose the Right Model for Speed

```typescript
import Groq from "groq-sdk";

const groq = new Groq();

// Speed tiers for different use cases
const SPEED_MAP = {
  // Under 100ms TTFT -- use for latency-critical paths
  instant: "llama-3.1-8b-instant",
  // Under 200ms TTFT -- use for quality-sensitive paths
  balanced: "llama-3.3-70b-versatile",
  // Speculative decoding -- same quality as 70b, faster throughput
  fast70b: "llama-3.3-70b-specdec",
} as const;

type SpeedTier = keyof typeof SPEED_MAP;

async function tieredCompletion(prompt: string, tier: SpeedTier = "instant") {
  return groq.chat.completions.create({
    model: SPEED_MAP[tier],
    messages: [{ role: "user", content: prompt }],
    temperature: 0,        // Deterministic = cacheable
    max_tokens: 256,       // Only request what you need
  });
}
```

### Step 2: Minimize Token Count

```typescript
// Groq charges per token AND rate limits on TPM
// Smaller prompts = faster responses + less quota usage

// BAD: verbose system prompt (200+ tokens)
const verbosePrompt = "You are an AI assistant that classifies text. Given a piece of text, analyze it carefully and determine whether the sentiment is positive, negative, or neutral. Consider the tone, word choice, and overall message...";

// GOOD: concise system prompt (15 tokens)
const concisePrompt = "Classify as positive/negative/neutral. One word only.";

// BAD: high max_tokens for short expected output
const wasteful = { max_tokens: 4096 }; // for a one-word response

// GOOD: match max_tokens to expected output
const efficient = { max_tokens: 5 };   // "positive" is 1 token
```

### Step 3: Streaming for Perceived Performance

```typescript
async function streamWithMetrics(
  messages: any[],
  onToken: (token: string) => void
): Promise<{ content: string; ttftMs: number; totalMs: number; tokPerSec: number }> {
  const start = performance.now();
  let ttft = 0;
  let content = "";
  let tokenCount = 0;

  const stream = await groq.chat.completions.create({
    model: "llama-3.3-70b-versatile",
    messages,
    stream: true,
    max_tokens: 1024,
  });

  for await (const chunk of stream) {
    const token = chunk.choices[0]?.delta?.content || "";
    if (token) {
      if (!ttft) ttft = performance.now() - start;
      content += token;
      tokenCount++;
      onToken(token);
    }
  }

  const totalMs = performance.now() - start;
  return {
    content,
    ttftMs: Math.round(ttft),
    totalMs: Math.round(totalMs),
    tokPerSec: Math.round(tokenCount / (totalMs / 1000)),
  };
}
```

### Step 4: Semantic Prompt Cache

```typescript
import { LRUCache } from "lru-cache";
import { createHash } from "crypto";

const promptCache = new LRUCache<string, string>({
  max: 1000,
  ttl: 10 * 60_000,  // 10 min TTL for deterministic responses
});

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

async function cachedCompletion(
  messages: any[],
  model = "llama-3.1-8b-instant"
): Promise<string> {
  const key = hashRequest(messages, model);
  const cached = promptCache.get(key);
  if (cached) return cached;

  const response = await groq.chat.completions.create({
    model,
    messages,
    temperature: 0,  // Cache only works with deterministic output
  });

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

### Step 5: Parallel Request Orchestration

```typescript
import PQueue from "p-queue";

// Respect RPM limits while maximizing throughput
const queue = new PQueue({
  concurrency: 10,
  intervalCap: 25,
  interval: 60_000,
});

async function parallelCompletions(
  prompts: string[],
  model = "llama-3.1-8b-instant"
): Promise<string[]> {
  const results = await Promise.all(
    prompts.map((prompt) =>
      queue.add(() =>
        cachedCompletion(
          [{ role: "user", content: prompt }],
          model
        )
      )
    )
  );
  return results as string[];
}
```

### Step 6: Latency Benchmarking

```typescript
async function benchmarkModels(prompt: string, iterations = 3) {
  const models = [
    "llama-3.1-8b-instant",
    "llama-3.3-70b-versatile",
    "llama-3.3-70b-specdec",
  ];

  for (const model of models) {
    const latencies: number[] = [];
    const speeds: number[] = [];

    for (let i = 0; i < iterations; i++) {
      const start = performance.now();
      const result = await groq.chat.completions.create({
        model,
        messages: [{ role: "user", content: prompt }],
        max_tokens: 100,
      });
      const elapsed = performance.now() - start;
      latencies.push(elapsed);
      const tps = result.usage!.completion_tokens /
        ((result.usage as any).completion_time || elapsed / 1000);
      speeds.push(tps);
    }

    const avgLatency = latencies.reduce((a, b) => a + b) / latencies.length;
    const avgSpeed = speeds.reduce((a, b) => a + b) / speeds.length;
    console.log(
      `${model.padEnd(45)} | ${avgLatency.toFixed(0)}ms avg | ${avgSpeed.toFixed(0)} tok/s avg`
    );
  }
}
```

## Performance Decision Matrix

| Scenario | Model | max_tokens | stream | cache |
|----------|-------|-----------|--------|-------|
| Classification | 8b-instant | 5 | No | Yes |
| Chat response | 70b-versatile | 1024 | Yes | No |
| Data extraction | 8b-instant | 200 | No | Yes |
| Code generation | 70b-versatile | 2048 | Yes | No |
| Bulk processing | 8b-instant | 256 | No | Yes |

## Error Handling

| Issue | Cause | Solution |
|-------|-------|----------|
| High TTFT | Using 70b for simple tasks | Switch to `llama-3.1-8b-instant` |
| Rate limit (429) | Over RPM or TPM | Use queue with interval limiting |
| Stream disconnect | Network timeout | Implement reconnection with partial content |
| Token overflow | max_tokens too high | Set to expected output size |
| Cache miss rate high | Unique prompts | Normalize prompts, use template patterns |

## Resources

- [Groq Models & Speed](https://console.groq.com/docs/models)
- [Groq Rate Limits](https://console.groq.com/docs/rate-limits)
- [lru-cache on npm](https://www.npmjs.com/package/lru-cache)

## Next Steps

For cost optimization, see `groq-cost-tuning`.