// CRITICALimport { spawn } from "node:child_process";import type { ChildProcess } from "node:child_process";import { createWriteStream, existsSync, readFileSync } from "node:fs";import type { WriteStream } from "node:fs";import { createInterface } from "node:readline";import { resolve } from "node:path";import { setTimeout as delayTimeout } from "node:timers/promises";import { parse as parseYaml } from "yaml";import type { Config } from "../config/env";import { delay } from "../core/async";import type { Logger } from "../core/logger";import type { LaunchResult, ProcessInfo, Recipe } from "../types/models";import type { EventManager } from "./event-manager";import { buildSglangCommand, buildVllmCommand } from "./backends";import { buildEnvironment, collectChildren, detectBackend, extractFlag, fetchTabbyModel, listProcesses, pidExists, buildProcessTree,} from "./process-utilities"; /** * Controller process manager. */export interface ProcessManager { findInferenceProcess: (port: number) => Promise<ProcessInfo | null>; launchModel: (recipe: Recipe) => Promise<LaunchResult>; evictModel: (force: boolean) => Promise<number | null>; killProcess: (pid: number, force: boolean) => Promise<boolean>;} /** * Create a process manager. * @param config - Runtime config. * @param logger - Logger instance. * @param eventManager - Event manager for log forwarding. * @returns Process manager. */export const createProcessManager = ( config: Config, logger: Logger, eventManager?: EventManager,): ProcessManager => { /** * Locate the inference process by port. * @param port - Port to match. * @returns Process info or null. */ const findInferenceProcess = async ( port: number, ): Promise<ProcessInfo | null> => { const processes = listProcesses(); for (const proc of processes) { const backend = detectBackend(proc.args); if (!backend) { continue; } const flagPort = extractFlag(proc.args, "--port"); if (backend === "tabbyapi") { if (port !== 8000) { continue; } } else if (!flagPort || Number(flagPort) !== port) { continue; } let modelPath = extractFlag(proc.args, "--model") || extractFlag(proc.args, "--model-path"); let servedModelName = extractFlag(proc.args, "--served-model-name"); if (!modelPath) { const serveIndex = proc.args.indexOf("serve"); if (serveIndex >= 0 && serveIndex + 1 < proc.args.length) { const candidate = proc.args[serveIndex + 1]; if (candidate && !candidate.startsWith("-")) { modelPath = candidate; } } } if (backend === "tabbyapi" && !modelPath) { const tabbyDirectory = config.tabby_api_dir || "/opt/tabbyAPI"; const configFlag = extractFlag(proc.args, "--config"); if (configFlag) { const configPath = resolve(tabbyDirectory, configFlag); if (existsSync(configPath)) { try { const content = readFileSync(configPath, "utf-8"); const parsed = parseYaml(content) as Record< string, unknown >; const model = parsed["model"] as | Record<string, unknown> | undefined; const modelName = model?.["model_name"]; if (typeof modelName === "string") { modelPath = resolve( config.models_dir, modelName, ); servedModelName = modelName; } } catch { return { pid: proc.pid, backend, model_path: "tabbyapi:unknown", port, served_model_name: servedModelName ?? "GLM-4.7", }; } } } if (!modelPath) { const tabbyResult = await fetchTabbyModel( port, tabbyDirectory, config.models_dir, ); modelPath = tabbyResult.modelPath ?? modelPath; servedModelName = tabbyResult.servedModelName ?? servedModelName; } } if (!modelPath && backend === "tabbyapi") { return { pid: proc.pid, backend, model_path: "tabbyapi:unknown", port, served_model_name: servedModelName ?? "GLM-4.7", }; } return { pid: proc.pid, backend, model_path: modelPath ?? null, port, served_model_name: servedModelName ?? null, }; } return null; }; /** * Kill a process and its children. * @param pid - Process id. * @param force - Force kill if true. * @returns True on success. */ const killProcess = async ( pid: number, force: boolean, ): Promise<boolean> => { if (!pidExists(pid)) { return true; } const tree = buildProcessTree(); const children = new Set<number>(); collectChildren(tree, pid, children); const allPids = [...children, pid]; const signal = force ? "SIGKILL" : "SIGTERM"; for (const childPid of allPids) { try { process.kill(childPid, signal); } catch { continue; } } if (!force) { const deadline = Date.now() + 10_000; while (Date.now() < deadline) { if (!pidExists(pid)) { break; } await delayTimeout(250); } if (pidExists(pid)) { try { process.kill(pid, "SIGKILL"); } catch { return false; } } } await delay(force ? 500 : 1000); return true; }; /** * Launch an inference backend for a recipe. * @param recipe - Recipe data. * @returns Launch result. */ const launchModel = async (recipe: Recipe): Promise<LaunchResult> => { const updatedRecipe: Recipe = { ...recipe, port: config.inference_port, }; const command = updatedRecipe.backend === "sglang" ? buildSglangCommand(updatedRecipe, config) : buildVllmCommand(updatedRecipe); const logFile = resolve("/tmp", `vllm_${updatedRecipe.id}.log`); const env = buildEnvironment(updatedRecipe); try { const entry = command[0]; if (!entry) { return { success: false, pid: null, message: "Invalid launch command", log_file: logFile, }; } let spawnError: string | null = null; // Use pipes to capture stdout/stderr for forwarding const child = spawn(entry, command.slice(1), { stdio: ["ignore", "pipe", "pipe"], env, detached: true, }) as ChildProcess; child.on("error", (error) => { spawnError = String(error); }); // Create log file stream let logStream: WriteStream | null = null; try { logStream = createWriteStream(logFile, { flags: "a" }); } catch (logError) { logger.warn("Failed to open log file", { error: String(logError), }); } // Forward stdout to log file and event manager if (child.stdout) { const rl = createInterface({ input: child.stdout, crlfDelay: Infinity, }); rl.on("line", (line) => { if (logStream) { logStream.write(line + "\n"); } if (eventManager) { eventManager .publishLogLine(updatedRecipe.id, line) .catch(() => {}); } }); } // Forward stderr to log file and event manager if (child.stderr) { const rl = createInterface({ input: child.stderr, crlfDelay: Infinity, }); rl.on("line", (line) => { if (logStream) { logStream.write(line + "\n"); } if (eventManager) { eventManager .publishLogLine(updatedRecipe.id, line) .catch(() => {}); } }); } // Close log stream when process exits child.on("exit", () => { if (logStream) { logStream.end(); } }); child.unref(); await delay(3000); if (spawnError) { if (logStream) { logStream.end(); } return { success: false, pid: null, message: spawnError, log_file: logFile, }; } if (child.exitCode !== null) { if (logStream) { logStream.end(); } return { success: false, pid: null, message: "Process exited early", log_file: logFile, }; } return { success: true, pid: child.pid ?? null, message: "Process started", log_file: logFile, }; } catch (error) { logger.error("Launch failed", { error: String(error) }); return { success: false, pid: null, message: String(error), log_file: logFile, }; } }; /** * Evict the running inference process. * @param force - Force kill if true. * @returns Evicted pid or null. */ const evictModel = async (force: boolean): Promise<number | null> => { const current = await findInferenceProcess(config.inference_port); if (!current) { return null; } await killProcess(current.pid, force); return current.pid; }; return { findInferenceProcess, launchModel, evictModel, killProcess, };};
// CRITICAL
import { spawn } from "node:child_process";
import type { ChildProcess } from "node:child_process";
import { createWriteStream, existsSync, readFileSync } from "node:fs";
import type { WriteStream } from "node:fs";
import { createInterface } from "node:readline";
import { resolve } from "node:path";
import { setTimeout as delayTimeout } from "node:timers/promises";
import { parse as parseYaml } from "yaml";
import type { Config } from "../config/env";
import { delay } from "../core/async";
import type { Logger } from "../core/logger";
import type { LaunchResult, ProcessInfo, Recipe } from "../types/models";
import type { EventManager } from "./event-manager";
import { buildSglangCommand, buildVllmCommand } from "./backends";
import {
buildEnvironment,
collectChildren,
detectBackend,
extractFlag,
fetchTabbyModel,
listProcesses,
pidExists,
buildProcessTree,
} from "./process-utilities";
/**
* Controller process manager.
*/
export interface ProcessManager {
findInferenceProcess: (port: number) => Promise<ProcessInfo | null>;
launchModel: (recipe: Recipe) => Promise<LaunchResult>;
evictModel: (force: boolean) => Promise<number | null>;
killProcess: (pid: number, force: boolean) => Promise<boolean>;
}
/**
* Create a process manager.
* @param config - Runtime config.
* @param logger - Logger instance.
* @param eventManager - Event manager for log forwarding.
* @returns Process manager.
*/
export const createProcessManager = (
config: Config,
logger: Logger,
eventManager?: EventManager,
): ProcessManager => {
/**
* Locate the inference process by port.
* @param port - Port to match.
* @returns Process info or null.
*/
const findInferenceProcess = async (
port: number,
): Promise<ProcessInfo | null> => {
const processes = listProcesses();
for (const proc of processes) {
const backend = detectBackend(proc.args);
if (!backend) {
continue;
}
const flagPort = extractFlag(proc.args, "--port");
if (backend === "tabbyapi") {
if (port !== 8000) {
continue;
}
} else if (!flagPort || Number(flagPort) !== port) {
continue;
}
let modelPath =
extractFlag(proc.args, "--model") ||
extractFlag(proc.args, "--model-path");
let servedModelName = extractFlag(proc.args, "--served-model-name");
if (!modelPath) {
const serveIndex = proc.args.indexOf("serve");
if (serveIndex >= 0 && serveIndex + 1 < proc.args.length) {
const candidate = proc.args[serveIndex + 1];
if (candidate && !candidate.startsWith("-")) {
modelPath = candidate;
}
}
}
if (backend === "tabbyapi" && !modelPath) {
const tabbyDirectory = config.tabby_api_dir || "/opt/tabbyAPI";
const configFlag = extractFlag(proc.args, "--config");
if (configFlag) {
const configPath = resolve(tabbyDirectory, configFlag);
if (existsSync(configPath)) {
try {
const content = readFileSync(configPath, "utf-8");
const parsed = parseYaml(content) as Record<
string,
unknown
>;
const model = parsed["model"] as
| Record<string, unknown>
| undefined;
const modelName = model?.["model_name"];
if (typeof modelName === "string") {
modelPath = resolve(
config.models_dir,
modelName,
);
servedModelName = modelName;
}
} catch {
return {
pid: proc.pid,
backend,
model_path: "tabbyapi:unknown",
port,
served_model_name: servedModelName ?? "GLM-4.7",
};
}
}
}
if (!modelPath) {
const tabbyResult = await fetchTabbyModel(
port,
tabbyDirectory,
config.models_dir,
);
modelPath = tabbyResult.modelPath ?? modelPath;
servedModelName =
tabbyResult.servedModelName ?? servedModelName;
}
}
if (!modelPath && backend === "tabbyapi") {
return {
pid: proc.pid,
backend,
model_path: "tabbyapi:unknown",
port,
served_model_name: servedModelName ?? "GLM-4.7",
};
}
return {
pid: proc.pid,
backend,
model_path: modelPath ?? null,
port,
served_model_name: servedModelName ?? null,
};
}
return null;
};
/**
* Kill a process and its children.
* @param pid - Process id.
* @param force - Force kill if true.
* @returns True on success.
*/
const killProcess = async (
pid: number,
force: boolean,
): Promise<boolean> => {
if (!pidExists(pid)) {
return true;
}
const tree = buildProcessTree();
const children = new Set<number>();
collectChildren(tree, pid, children);
const allPids = [...children, pid];
const signal = force ? "SIGKILL" : "SIGTERM";
for (const childPid of allPids) {
try {
process.kill(childPid, signal);
} catch {
continue;
}
}
if (!force) {
const deadline = Date.now() + 10_000;
while (Date.now() < deadline) {
if (!pidExists(pid)) {
break;
}
await delayTimeout(250);
}
if (pidExists(pid)) {
try {
process.kill(pid, "SIGKILL");
} catch {
return false;
}
}
}
await delay(force ? 500 : 1000);
return true;
};
/**
* Launch an inference backend for a recipe.
* @param recipe - Recipe data.
* @returns Launch result.
*/
const launchModel = async (recipe: Recipe): Promise<LaunchResult> => {
const updatedRecipe: Recipe = {
...recipe,
port: config.inference_port,
};
const command =
updatedRecipe.backend === "sglang"
? buildSglangCommand(updatedRecipe, config)
: buildVllmCommand(updatedRecipe);
const logFile = resolve("/tmp", `vllm_${updatedRecipe.id}.log`);
const env = buildEnvironment(updatedRecipe);
try {
const entry = command[0];
if (!entry) {
return {
success: false,
pid: null,
message: "Invalid launch command",
log_file: logFile,
};
}
let spawnError: string | null = null;
// Use pipes to capture stdout/stderr for forwarding
const child = spawn(entry, command.slice(1), {
stdio: ["ignore", "pipe", "pipe"],
env,
detached: true,
}) as ChildProcess;
child.on("error", (error) => {
spawnError = String(error);
});
// Create log file stream
let logStream: WriteStream | null = null;
try {
logStream = createWriteStream(logFile, { flags: "a" });
} catch (logError) {
logger.warn("Failed to open log file", {
error: String(logError),
});
}
// Forward stdout to log file and event manager
if (child.stdout) {
const rl = createInterface({
input: child.stdout,
crlfDelay: Infinity,
});
rl.on("line", (line) => {
if (logStream) {
logStream.write(line + "\n");
}
if (eventManager) {
eventManager
.publishLogLine(updatedRecipe.id, line)
.catch(() => {});
}
});
}
// Forward stderr to log file and event manager
if (child.stderr) {
const rl = createInterface({
input: child.stderr,
crlfDelay: Infinity,
});
rl.on("line", (line) => {
if (logStream) {
logStream.write(line + "\n");
}
if (eventManager) {
eventManager
.publishLogLine(updatedRecipe.id, line)
.catch(() => {});
}
});
}
// Close log stream when process exits
child.on("exit", () => {
if (logStream) {
logStream.end();
}
});
child.unref();
await delay(3000);
if (spawnError) {
if (logStream) {
logStream.end();
}
return {
success: false,
pid: null,
message: spawnError,
log_file: logFile,
};
}
if (child.exitCode !== null) {
if (logStream) {
logStream.end();
}
return {
success: false,
pid: null,
message: "Process exited early",
log_file: logFile,
};
}
return {
success: true,
pid: child.pid ?? null,
message: "Process started",
log_file: logFile,
};
} catch (error) {
logger.error("Launch failed", { error: String(error) });
return {
success: false,
pid: null,
message: String(error),
log_file: logFile,
};
}
};
/**
* Evict the running inference process.
* @param force - Force kill if true.
* @returns Evicted pid or null.
*/
const evictModel = async (force: boolean): Promise<number | null> => {
const current = await findInferenceProcess(config.inference_port);
if (!current) {
return null;
}
await killProcess(current.pid, force);
return current.pid;
};
return {
findInferenceProcess,
launchModel,
evictModel,
killProcess,
};
};