File size: 9,606 Bytes
e706de2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 |
# Code Explanation: batch.js
This file demonstrates **parallel execution** of multiple LLM prompts using separate context sequences, enabling concurrent processing for better performance.
## Step-by-Step Code Breakdown
### 1. Import and Setup (Lines 1-10)
```javascript
import {getLlama, LlamaChatSession} from "node-llama-cpp";
import path from "path";
import {fileURLToPath} from "url";
/**
* Asynchronous execution improves performance in GAIA benchmarks,
* multi-agent applications, and other high-throughput scenarios.
*/
const __dirname = path.dirname(fileURLToPath(import.meta.url));
```
- Standard imports for LLM interaction
- Comment explains the performance benefit
- **GAIA benchmark**: A standard for testing AI agent performance
- Useful for multi-agent systems that need to handle many requests
### 2. Model Path Configuration (Lines 11-16)
```javascript
const modelPath = path.join(
__dirname,
"../",
"models",
"DeepSeek-R1-0528-Qwen3-8B-Q6_K.gguf"
)
```
- Uses **DeepSeek-R1**: An 8B parameter model optimized for reasoning
- **Q6_K quantization**: Balance between quality and size
- Model is loaded once and shared between sequences
### 3. Initialize Llama and Load Model (Lines 18-19)
```javascript
const llama = await getLlama();
const model = await llama.loadModel({modelPath});
```
- Standard initialization
- Model is loaded into memory once
- Will be used by multiple sequences simultaneously
### 4. Create Context with Multiple Sequences (Lines 20-23)
```javascript
const context = await model.createContext({
sequences: 2,
batchSize: 1024 // The number of tokens that can be processed at once by the GPU.
});
```
**Key parameters:**
- **sequences: 2**: Creates 2 independent conversation sequences
- Each sequence has its own conversation history
- Both share the same model and context memory pool
- Can be processed in parallel
- **batchSize: 1024**: Maximum tokens processed per GPU batch
- Larger = better GPU utilization
- Smaller = lower memory usage
- 1024 is a good balance for most GPUs
### Why Multiple Sequences?
```
Single Sequence (Sequential) Multiple Sequences (Parallel)
───────────────────────── ──────────────────────────────
Process Prompt 1 → Response 1 Process Prompt 1 ──┐
Wait... ├→ Both responses
Process Prompt 2 → Response 2 Process Prompt 2 ──┘ in parallel!
Total Time: T1 + T2 Total Time: max(T1, T2)
```
### 5. Get Individual Sequences (Lines 25-26)
```javascript
const sequence1 = context.getSequence();
const sequence2 = context.getSequence();
```
- Retrieves two separate sequence objects from the context
- Each sequence maintains its own state
- They can be used independently for different conversations
### 6. Create Separate Sessions (Lines 28-33)
```javascript
const session1 = new LlamaChatSession({
contextSequence: sequence1
});
const session2 = new LlamaChatSession({
contextSequence: sequence2
});
```
- Creates a chat session for each sequence
- Each session has its own conversation history
- Sessions are completely independent
- No system prompts in this example (could be added)
### 7. Define Questions (Lines 35-36)
```javascript
const q1 = "Hi there, how are you?";
const q2 = "How much is 6+6?";
```
- Two completely different questions
- Will be processed simultaneously
- Different types: conversational vs. computational
### 8. Parallel Execution with Promise.all (Lines 38-44)
```javascript
const [
a1,
a2
] = await Promise.all([
session1.prompt(q1),
session2.prompt(q2)
]);
```
**How this works:**
1. `session1.prompt(q1)` starts asynchronously
2. `session2.prompt(q2)` starts asynchronously (doesn't wait for #1)
3. `Promise.all()` waits for BOTH to complete
4. Returns results in array: [response1, response2]
5. Destructures into `a1` and `a2`
**Key benefit**: Both prompts are processed at the same time, not one after another!
### 9. Display Results (Lines 46-50)
```javascript
console.log("User: " + q1);
console.log("AI: " + a1);
console.log("User: " + q2);
console.log("AI: " + a2);
```
- Outputs both question-answer pairs
- Results appear in order despite parallel processing
## Key Concepts Demonstrated
### 1. Parallel Processing
Instead of:
```javascript
// Sequential (slow)
const a1 = await session1.prompt(q1); // Wait
const a2 = await session2.prompt(q2); // Wait again
```
We use:
```javascript
// Parallel (fast)
const [a1, a2] = await Promise.all([
session1.prompt(q1),
session2.prompt(q2)
]);
```
### 2. Context Sequences
A context can hold multiple independent sequences:
```
┌─────────────────────────────────────┐
│ Context (Shared) │
│ ┌───────────────────────────────┐ │
│ │ Model Weights (8B params) │ │
│ └───────────────────────────────┘ │
│ │
│ ┌─────────────┐ ┌─────────────┐ │
│ │ Sequence 1 │ │ Sequence 2 │ │
│ │ "Hi there" │ │ "6+6?" │ │
│ │ History... │ │ History... │ │
│ └─────────────┘ └─────────────┘ │
└─────────────────────────────────────┘
```
## Performance Comparison
### Sequential Execution
```
Request 1: 2 seconds
Request 2: 2 seconds
Total: 4 seconds
```
### Parallel Execution (This Example)
```
Request 1: 2 seconds ──┐
Request 2: 2 seconds ──┤ Both running
Total: ~2 seconds └─ simultaneously
```
**Speedup**: ~2x for 2 sequences, scales with more sequences
## Use Cases
### 1. Multi-User Applications
```javascript
// Handle multiple users simultaneously
const [user1Response, user2Response, user3Response] = await Promise.all([
session1.prompt(user1Query),
session2.prompt(user2Query),
session3.prompt(user3Query)
]);
```
### 2. Multi-Agent Systems
```javascript
// Multiple agents working on different tasks
const [
plannerResponse,
analyzerResponse,
executorResponse
] = await Promise.all([
plannerSession.prompt("Plan the task"),
analyzerSession.prompt("Analyze the data"),
executorSession.prompt("Execute step 1")
]);
```
### 3. Benchmarking
```javascript
// Test multiple prompts for evaluation
const results = await Promise.all(
testPrompts.map(prompt => session.prompt(prompt))
);
```
### 4. A/B Testing
```javascript
// Test different system prompts
const [responseA, responseB] = await Promise.all([
sessionWithPromptA.prompt(query),
sessionWithPromptB.prompt(query)
]);
```
## Resource Considerations
### Memory Usage
Each sequence needs memory for:
- Conversation history
- Intermediate computations
- KV cache (key-value cache for transformer attention)
**Rule of thumb**: More sequences = more memory needed
### GPU Utilization
- **Single sequence**: May underutilize GPU
- **Multiple sequences**: Better GPU utilization
- **Too many sequences**: May exceed VRAM, causing slowdown
### Optimal Number of Sequences
Depends on:
- Available VRAM
- Model size
- Context length
- Batch size
**Typical**: 2-8 sequences for consumer GPUs
## Limitations & Considerations
### 1. Shared Context Limit
All sequences share the same context memory pool:
```
Total context size: 8192 tokens
Sequence 1: 4096 tokens
Sequence 2: 4096 tokens
Maximum distribution!
```
### 2. Not True Parallelism for CPU
On CPU-only systems, sequences are interleaved, not truly parallel. Still provides better overall throughput.
### 3. Model Loading Overhead
The model is loaded once and shared, which is efficient. But initial loading still takes time.
## Why This Matters for AI Agents
### Efficiency in Production
Real-world agent systems need to:
- Handle multiple requests concurrently
- Respond quickly to users
- Make efficient use of hardware
### Multi-Agent Architectures
Complex agent systems often have:
- **Planner agent**: Thinks about strategy
- **Executor agent**: Takes actions
- **Critic agent**: Evaluates results
These can run in parallel using separate sequences.
### Scalability
This pattern is the foundation for:
- Web services with multiple users
- Batch processing of data
- Distributed agent systems
## Best Practices
1. **Match sequences to workload**: Don't create more than you need
2. **Monitor memory usage**: Each sequence consumes VRAM
3. **Use appropriate batch size**: Balance speed vs. memory
4. **Clean up resources**: Always dispose when done
5. **Handle errors**: Wrap Promise.all in try-catch
## Expected Output
Running this script should output something like:
```
User: Hi there, how are you?
AI: Hello! I'm doing well, thank you for asking...
User: How much is 6+6?
AI: 12
```
Both responses appear quickly because they were processed simultaneously!
|