273 lines
7.0 KiB
TypeScript
273 lines
7.0 KiB
TypeScript
/**
|
|
* vLLM Container
|
|
*
|
|
* Manages vLLM containers for high-performance LLM inference.
|
|
*/
|
|
|
|
import type {
|
|
IContainerConfig,
|
|
ILoadedModel,
|
|
TContainerType,
|
|
} from '../interfaces/container.ts';
|
|
import type {
|
|
IChatCompletionRequest,
|
|
IChatCompletionResponse,
|
|
IChatMessage,
|
|
} from '../interfaces/api.ts';
|
|
import { CONTAINER_IMAGES, CONTAINER_PORTS } from '../constants.ts';
|
|
import { logger } from '../logger.ts';
|
|
import { BaseContainer, type TModelPullProgress } from './base-container.ts';
|
|
|
|
/**
|
|
* vLLM model info response
|
|
*/
|
|
interface IVllmModelsResponse {
|
|
object: 'list';
|
|
data: Array<{
|
|
id: string;
|
|
object: 'model';
|
|
created: number;
|
|
owned_by: string;
|
|
}>;
|
|
}
|
|
|
|
/**
|
|
* vLLM container implementation
|
|
*
|
|
* vLLM serves a single model per instance and is optimized for:
|
|
* - High throughput with PagedAttention
|
|
* - Continuous batching
|
|
* - OpenAI-compatible API
|
|
*/
|
|
export class VllmContainer extends BaseContainer {
|
|
public readonly type: TContainerType = 'vllm';
|
|
public readonly displayName = 'vLLM';
|
|
public readonly defaultImage = CONTAINER_IMAGES.VLLM;
|
|
public readonly defaultPort = CONTAINER_PORTS.VLLM;
|
|
|
|
constructor(config: IContainerConfig) {
|
|
super(config);
|
|
|
|
// Set defaults if not provided
|
|
if (!config.image) {
|
|
config.image = this.defaultImage;
|
|
}
|
|
if (!config.port) {
|
|
config.port = this.defaultPort;
|
|
}
|
|
|
|
// Add default volume for model cache
|
|
if (!config.volumes || config.volumes.length === 0) {
|
|
config.volumes = [`modelgrid-vllm-${config.id}:/root/.cache/huggingface`];
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Create vLLM container configuration
|
|
*/
|
|
public static createConfig(
|
|
id: string,
|
|
name: string,
|
|
modelName: string,
|
|
gpuIds: string[],
|
|
options: Partial<IContainerConfig> = {},
|
|
): IContainerConfig {
|
|
// vLLM requires model to be specified at startup
|
|
const command = [
|
|
'--model', modelName,
|
|
'--host', '0.0.0.0',
|
|
'--port', String(options.port || CONTAINER_PORTS.VLLM),
|
|
];
|
|
|
|
// Add tensor parallelism if multiple GPUs
|
|
if (gpuIds.length > 1) {
|
|
command.push('--tensor-parallel-size', String(gpuIds.length));
|
|
}
|
|
|
|
// Add additional options
|
|
if (options.env?.VLLM_MAX_MODEL_LEN) {
|
|
command.push('--max-model-len', options.env.VLLM_MAX_MODEL_LEN);
|
|
}
|
|
|
|
return {
|
|
id,
|
|
name,
|
|
type: 'vllm',
|
|
image: options.image || CONTAINER_IMAGES.VLLM,
|
|
gpuIds,
|
|
port: options.port || CONTAINER_PORTS.VLLM,
|
|
externalPort: options.externalPort,
|
|
models: [modelName],
|
|
env: {
|
|
HF_TOKEN: options.env?.HF_TOKEN || '',
|
|
...options.env,
|
|
},
|
|
volumes: options.volumes || [`modelgrid-vllm-${id}:/root/.cache/huggingface`],
|
|
autoStart: options.autoStart ?? true,
|
|
restartPolicy: options.restartPolicy || 'unless-stopped',
|
|
memoryLimit: options.memoryLimit,
|
|
cpuLimit: options.cpuLimit,
|
|
command,
|
|
};
|
|
}
|
|
|
|
/**
|
|
* Check if vLLM is healthy
|
|
*/
|
|
public async isHealthy(): Promise<boolean> {
|
|
try {
|
|
const response = await this.fetch('/health', { timeout: 5000 });
|
|
return response.ok;
|
|
} catch {
|
|
return false;
|
|
}
|
|
}
|
|
|
|
/**
|
|
* List available models
|
|
* vLLM serves a single model per instance
|
|
*/
|
|
public async listModels(): Promise<string[]> {
|
|
try {
|
|
const data = await this.fetchJson<IVllmModelsResponse>('/v1/models');
|
|
return (data.data || []).map((m) => m.id);
|
|
} catch (error) {
|
|
logger.warn(`Failed to list vLLM models: ${error instanceof Error ? error.message : String(error)}`);
|
|
return this.config.models || [];
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Get loaded models with details
|
|
*/
|
|
public async getLoadedModels(): Promise<ILoadedModel[]> {
|
|
try {
|
|
const data = await this.fetchJson<IVllmModelsResponse>('/v1/models');
|
|
return (data.data || []).map((m) => ({
|
|
name: m.id,
|
|
size: 0, // vLLM doesn't expose size
|
|
loaded: true,
|
|
requestCount: 0,
|
|
}));
|
|
} catch {
|
|
// Return configured model as fallback
|
|
return this.config.models.map((name) => ({
|
|
name,
|
|
size: 0,
|
|
loaded: true,
|
|
requestCount: 0,
|
|
}));
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Pull a model
|
|
* vLLM downloads models automatically at startup
|
|
* This method is a no-op - models are configured at container creation
|
|
*/
|
|
public async pullModel(modelName: string, onProgress?: TModelPullProgress): Promise<boolean> {
|
|
logger.info(`vLLM downloads models at startup. Model: ${modelName}`);
|
|
logger.info('To use a different model, create a new vLLM container.');
|
|
|
|
if (onProgress) {
|
|
onProgress({
|
|
model: modelName,
|
|
status: 'vLLM models are loaded at container startup',
|
|
percent: 100,
|
|
});
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
/**
|
|
* Remove a model
|
|
* vLLM serves a single model per instance
|
|
*/
|
|
public async removeModel(modelName: string): Promise<boolean> {
|
|
logger.info(`vLLM serves a single model per instance.`);
|
|
logger.info(`To remove model ${modelName}, stop and remove this container.`);
|
|
return true;
|
|
}
|
|
|
|
/**
|
|
* Send a chat completion request
|
|
* vLLM is OpenAI-compatible
|
|
*/
|
|
public async chatCompletion(request: IChatCompletionRequest): Promise<IChatCompletionResponse> {
|
|
return this.fetchJson<IChatCompletionResponse>('/v1/chat/completions', {
|
|
method: 'POST',
|
|
body: {
|
|
...request,
|
|
stream: false,
|
|
},
|
|
timeout: 300000, // 5 minutes
|
|
});
|
|
}
|
|
|
|
/**
|
|
* Stream a chat completion request
|
|
* vLLM is OpenAI-compatible
|
|
*/
|
|
public async chatCompletionStream(
|
|
request: IChatCompletionRequest,
|
|
onChunk: (chunk: string) => void,
|
|
): Promise<void> {
|
|
const response = await this.fetch('/v1/chat/completions', {
|
|
method: 'POST',
|
|
body: {
|
|
...request,
|
|
stream: true,
|
|
},
|
|
timeout: 300000,
|
|
});
|
|
|
|
if (!response.ok) {
|
|
const error = await response.text();
|
|
throw new Error(`HTTP ${response.status}: ${error}`);
|
|
}
|
|
|
|
const reader = response.body?.getReader();
|
|
if (!reader) {
|
|
throw new Error('No response body');
|
|
}
|
|
|
|
const decoder = new TextDecoder();
|
|
|
|
while (true) {
|
|
const { done, value } = await reader.read();
|
|
if (done) break;
|
|
|
|
const text = decoder.decode(value);
|
|
// vLLM already sends data in SSE format
|
|
onChunk(text);
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Get vLLM-specific metrics
|
|
*/
|
|
public async getMetrics(): Promise<Record<string, unknown>> {
|
|
try {
|
|
const response = await this.fetch('/metrics', { timeout: 5000 });
|
|
if (response.ok) {
|
|
const text = await response.text();
|
|
// Parse Prometheus metrics
|
|
const metrics: Record<string, unknown> = {};
|
|
const lines = text.split('\n');
|
|
for (const line of lines) {
|
|
if (line.startsWith('#') || !line.trim()) continue;
|
|
const match = line.match(/^(\w+)(?:\{[^}]*\})?\s+([\d.e+-]+)/);
|
|
if (match) {
|
|
metrics[match[1]] = parseFloat(match[2]);
|
|
}
|
|
}
|
|
return metrics;
|
|
}
|
|
} catch {
|
|
// Metrics endpoint may not be enabled
|
|
}
|
|
return {};
|
|
}
|
|
}
|