feat(cluster,api,models,cli): add cluster-aware model catalog deployments and request routing

This commit is contained in:
2026-04-20 23:00:50 +00:00
parent 83cacd0cf1
commit 4f2266e1b7
55 changed files with 3970 additions and 1630 deletions
+149 -65
View File
@@ -1,58 +1,89 @@
/**
* Chat Completions Handler
*
* Handles /v1/chat/completions and /v1/completions endpoints.
* Chat completions handler.
*/
import * as http from 'node:http';
import type {
IChatCompletionRequest,
IChatCompletionResponse,
IApiError,
} from '../../interfaces/api.ts';
import { logger } from '../../logger.ts';
import type { IApiError, IChatCompletionRequest } from '../../interfaces/api.ts';
import { ClusterCoordinator } from '../../cluster/coordinator.ts';
import { ContainerManager } from '../../containers/container-manager.ts';
import { logger } from '../../logger.ts';
import { ModelRegistry } from '../../models/registry.ts';
import { ModelLoader } from '../../models/loader.ts';
/**
* Handler for chat completion requests
*/
export class ChatHandler {
private containerManager: ContainerManager;
private modelRegistry: ModelRegistry;
private modelLoader: ModelLoader;
private clusterCoordinator: ClusterCoordinator;
constructor(containerManager: ContainerManager, modelLoader: ModelLoader) {
constructor(
containerManager: ContainerManager,
modelRegistry: ModelRegistry,
modelLoader: ModelLoader,
clusterCoordinator: ClusterCoordinator,
) {
this.containerManager = containerManager;
this.modelRegistry = modelRegistry;
this.modelLoader = modelLoader;
this.clusterCoordinator = clusterCoordinator;
}
/**
* Handle POST /v1/chat/completions
*/
public async handleChatCompletion(
req: http.IncomingMessage,
res: http.ServerResponse,
body: IChatCompletionRequest,
): Promise<void> {
const modelName = body.model;
const isStream = body.stream === true;
const canonicalModel = await this.resolveCanonicalModel(body.model);
const requestBody: IChatCompletionRequest = {
...body,
model: canonicalModel,
};
logger.dim(`Chat completion request for model: ${modelName}`);
logger.dim(`Chat completion request for model: ${canonicalModel}`);
try {
// Find or load the model
const container = await this.findOrLoadModel(modelName);
if (!container) {
this.sendError(res, 404, `Model "${modelName}" not found or could not be loaded`, 'model_not_found');
const container = await this.findOrLoadLocalModel(canonicalModel);
if (container) {
if (requestBody.stream) {
await this.handleStreamingCompletion(res, container, requestBody);
} else {
await this.handleNonStreamingCompletion(res, container, requestBody);
}
return;
}
// Route to streaming or non-streaming handler
if (isStream) {
await this.handleStreamingCompletion(res, container, body);
} else {
await this.handleNonStreamingCompletion(res, container, body);
const ensured = await this.clusterCoordinator.ensureModelViaControlPlane(canonicalModel);
if (!ensured) {
this.sendError(
res,
404,
`Model "${canonicalModel}" not found or could not be deployed`,
'model_not_found',
);
return;
}
if (ensured.location.nodeName === this.clusterCoordinator.getLocalNodeName()) {
const localContainer = await this.findLocalContainer(canonicalModel);
if (!localContainer) {
this.sendError(
res,
503,
`Model "${canonicalModel}" was scheduled locally but is not ready`,
'server_error',
);
return;
}
if (requestBody.stream) {
await this.handleStreamingCompletion(res, localContainer, requestBody);
} else {
await this.handleNonStreamingCompletion(res, localContainer, requestBody);
}
return;
}
await this.proxyChatRequest(req, res, ensured.location.endpoint, requestBody);
} catch (error) {
const message = error instanceof Error ? error.message : String(error);
logger.error(`Chat completion error: ${message}`);
@@ -60,34 +91,38 @@ export class ChatHandler {
}
}
/**
* Find container with model or attempt to load it
*/
private async findOrLoadModel(
private async resolveCanonicalModel(modelName: string): Promise<string> {
const model = await this.modelRegistry.getModel(modelName);
return model?.id || modelName;
}
private async findLocalContainer(
modelName: string,
): Promise<import('../../containers/base-container.ts').BaseContainer | null> {
// First, check if model is already loaded
const container = await this.containerManager.findContainerForModel(modelName);
if (container) {
return container;
}
// Try to load the model
logger.info(`Model ${modelName} not loaded, attempting to load...`);
const loadResult = await this.modelLoader.loadModel(modelName);
if (!loadResult.success) {
logger.error(`Failed to load model: ${loadResult.error}`);
return null;
}
// Find the container again after loading
return this.containerManager.findContainerForModel(modelName);
}
/**
* Handle non-streaming chat completion
*/
private async findOrLoadLocalModel(
modelName: string,
): Promise<import('../../containers/base-container.ts').BaseContainer | null> {
const existing = await this.findLocalContainer(modelName);
if (existing) {
return existing;
}
if (!this.clusterCoordinator.shouldDeployLocallyFirst()) {
return null;
}
logger.info(`Model ${modelName} not loaded, attempting local deploy...`);
const loadResult = await this.modelLoader.loadModel(modelName);
if (!loadResult.success) {
return null;
}
return this.findLocalContainer(loadResult.model);
}
private async handleNonStreamingCompletion(
res: http.ServerResponse,
container: import('../../containers/base-container.ts').BaseContainer,
@@ -99,35 +134,85 @@ export class ChatHandler {
res.end(JSON.stringify(response));
}
/**
* Handle streaming chat completion
*/
private async handleStreamingCompletion(
res: http.ServerResponse,
container: import('../../containers/base-container.ts').BaseContainer,
body: IChatCompletionRequest,
): Promise<void> {
// Set SSE headers
res.writeHead(200, {
'Content-Type': 'text/event-stream',
'Cache-Control': 'no-cache',
'Connection': 'keep-alive',
Connection: 'keep-alive',
'X-Accel-Buffering': 'no',
});
// Stream chunks to client
await container.chatCompletionStream(body, (chunk) => {
res.write(`data: ${chunk}\n\n`);
res.write(chunk);
});
// Send final done message
res.write('data: [DONE]\n\n');
res.end();
}
/**
* Send error response
*/
private async proxyChatRequest(
req: http.IncomingMessage,
res: http.ServerResponse,
targetEndpoint: string,
body: IChatCompletionRequest,
): Promise<void> {
const response = await fetch(`${targetEndpoint}/v1/chat/completions`, {
method: 'POST',
headers: this.buildForwardHeaders(req),
body: JSON.stringify(body),
});
if (body.stream) {
res.writeHead(response.status, {
'Content-Type': response.headers.get('content-type') || 'text/event-stream',
'Cache-Control': 'no-cache',
Connection: 'keep-alive',
});
const reader = response.body?.getReader();
if (!reader) {
res.end();
return;
}
while (true) {
const { done, value } = await reader.read();
if (done) {
break;
}
res.write(value);
}
res.end();
return;
}
const text = await response.text();
res.writeHead(response.status, {
'Content-Type': response.headers.get('content-type') || 'application/json',
});
res.end(text);
}
private buildForwardHeaders(req: http.IncomingMessage): Record<string, string> {
const headers: Record<string, string> = {
'Content-Type': 'application/json',
};
if (typeof req.headers.authorization === 'string') {
headers.Authorization = req.headers.authorization;
}
if (typeof req.headers['x-request-id'] === 'string') {
headers['X-Request-Id'] = req.headers['x-request-id'];
}
return headers;
}
private sendError(
res: http.ServerResponse,
statusCode: number,
@@ -140,7 +225,6 @@ export class ChatHandler {
message,
type,
param,
code: null,
},
};