import type { EngineId, ComputeEngineSpec, EngineSupport, HostProfile, LaunchPlan, LaunchRequest, EngineRuntimeKind,} from "../contracts";import { applyDevices } from "./devices";import { exllamav3 } from "./exllamav3";import { llamacpp } from "./llamacpp";import { mlx } from "./mlx";import { sglang } from "./sglang";import { vllm } from "./vllm";const SPECS: Readonly<Record<EngineId, ComputeEngineSpec>> = { vllm, sglang, llamacpp, mlx, exllamav3,};export const engineSpec = (id: EngineId): ComputeEngineSpec => SPECS[id];export const allEngineSpecs: readonly ComputeEngineSpec[] = Object.values(SPECS);/** Engines this host can actually run, with the runtimes available for each. */export const availableEngines = ( host: HostProfile,): readonly { readonly id: EngineId; readonly support: EngineSupport }[] => allEngineSpecs.map((spec) => ({ id: spec.id, support: spec.supports(host) }));export const supportsRuntime = ( id: EngineId, host: HostProfile, runtime: EngineRuntimeKind,): boolean => { const support = SPECS[id].supports(host); return support.ok && support.runtimes.includes(runtime);};/** * The one entry point that turns a request into a runnable plan. Device selection is * folded in here so no engine has to know how its accelerator is addressed, and no * launcher has to re-derive it. */export const planLaunch = (request: LaunchRequest): LaunchPlan => applyDevices(SPECS[request.engine].plan(request), request.host.accelerator);
import type {
EngineId,
ComputeEngineSpec,
EngineSupport,
HostProfile,
LaunchPlan,
LaunchRequest,
EngineRuntimeKind,
} from "../contracts";
import { applyDevices } from "./devices";
import { exllamav3 } from "./exllamav3";
import { llamacpp } from "./llamacpp";
import { mlx } from "./mlx";
import { sglang } from "./sglang";
import { vllm } from "./vllm";
const SPECS: Readonly<Record<EngineId, ComputeEngineSpec>> = {
vllm,
sglang,
llamacpp,
mlx,
exllamav3,
};
export const engineSpec = (id: EngineId): ComputeEngineSpec => SPECS[id];
export const allEngineSpecs: readonly ComputeEngineSpec[] = Object.values(SPECS);
/** Engines this host can actually run, with the runtimes available for each. */
export const availableEngines = (
host: HostProfile,
): readonly { readonly id: EngineId; readonly support: EngineSupport }[] =>
allEngineSpecs.map((spec) => ({ id: spec.id, support: spec.supports(host) }));
export const supportsRuntime = (
id: EngineId,
host: HostProfile,
runtime: EngineRuntimeKind,
): boolean => {
const support = SPECS[id].supports(host);
return support.ok && support.runtimes.includes(runtime);
};
/**
* The one entry point that turns a request into a runnable plan. Device selection is
* folded in here so no engine has to know how its accelerator is addressed, and no
* launcher has to re-derive it.
*/
export const planLaunch = (request: LaunchRequest): LaunchPlan =>
applyDevices(SPECS[request.engine].plan(request), request.host.accelerator);