atelier/sidecar/providers/qwen-local.ts

import type { ImageAdapter, ImageRequest, ImageUsage, ProviderResult } from './types';

/**
 * Qwen-Image 2.1 on our own GPU, via `sd-server` from stable-diffusion.cpp with the Vulkan backend.
 *
 * It already speaks OpenAI's image API, which is the vocabulary `ImageRequest` is written in, so there
 * is no translation layer — only the things below that are true of a local GPU and not of a hosted API.
 *
 * ── It is slow, and the number matters ──
 *
 * Measured on this machine from the server's own log, not estimated: 512x512 completes in ~185s, and
 * 1024x1024 samples at ~41.4s per step over 20 steps, so about 14 minutes. The catalogue therefore
 * defaults to 512 and offers 1024 labelled for what it costs — a default nobody chose should not spend a
 * quarter of an hour of the only GPU there is.
 *
 * ── `timeout: false`, and it is not caution ──
 *
 * The server serialises strictly: a second caller waits for the first to finish ENTIRELY before its own
 * first step. So a queued 512 request can wait its own 3 minutes plus a 14-minute job ahead of it — 17
 * minutes before a single byte comes back. Bun's fetch has a default timeout, and every default anyone
 * would reach for is far under that. The caller's `signal` is still honoured, because a user cancelling
 * is a different thing from a client giving up.
 *
 * ── No `edit()` in this version ──
 *
 * Deliberate. It is the same GPU and the same serialisation, so it doubles the ways the queue can be
 * blocked while the generate path is still unproven in anger. `ImageAdapter` makes absence the way to
 * say a provider cannot do something, and the route answers 400 rather than the queue finding out late.
 */

/**
 * Where the server is. Overridable because the same service answers on the tailnet
 * (`http://100.64.0.1:8188`), and a second machine should not need a code change to reach it.
 */
const BASE = process.env.ATELIER_QWEN_LOCAL_URL ?? 'http://127.0.0.1:8188';

/**
 * What `GET /v1/models` actually returns, which is not what the model IS.
 *
 * sd-server hardcodes this id with no override, so the catalogue is the only side that can move: the id
 * stays `sd-cpp-local` so the two agree, and the catalogue carries "Qwen-Image 2.1 (local)" as the
 * label so the UI still says something meaningful. A catalogue that disagreed with `/v1/models` would
 * make every request name a model the server denies knowing.
 */
const MODEL_ID = 'sd-cpp-local';

/** Sent when the caller names no size. See the timings above for why it is not 1024. */
const DEFAULT_SIZE = '512x512';

type OpenAiImagesResponse = {
  data?: { b64_json?: string; revised_prompt?: string }[];
  usage?: ImageUsage;
  size?: string;
  quality?: string;
  background?: string;
};

/**
 * Only what this server understands.
 *
 * `quality`, `style`, `moderation` and `background` are OpenAI concepts with no Qwen equivalent — they
 * are dropped rather than passed, because sd-server ignores unknown fields silently and a parameter
 * that appears to have been honoured is worse than one that was visibly not offered. What was actually
 * used comes back in `echoed`.
 */
function body(request: ImageRequest): Record<string, unknown> {
  return {
    prompt: request.prompt,
    model: MODEL_ID,
    n: 1,
    size: request.size && request.size !== 'auto' ? request.size : DEFAULT_SIZE,
    response_format: 'b64_json',
    ...(request.output_format ? { output_format: request.output_format } : {}),
  };
}

export const qwenLocalAdapter: ImageAdapter = {
  name: 'qwen-local',
  label: 'Qwen-Image (local)',
  models: [MODEL_ID],
  defaultModel: MODEL_ID,
  keyless: true,

  async generate(request, _apiKey, signal): Promise<ProviderResult> {
    const sent = body(request);

    const response = await fetch(`${BASE}/v1/images/generations`, {
      method: 'POST',
      headers: { 'content-type': 'application/json' },
      body: JSON.stringify(sent),
      signal,
      // Bun extension. See the block comment: the honest worst case is 17 minutes.
      timeout: false,
    } as RequestInit);

    if (!response.ok) {
      const detail = await response.text().catch(() => '');
      throw new Error(`qwen-local ${response.status}: ${detail.slice(0, 400) || response.statusText}`);
    }

    const payload = (await response.json()) as OpenAiImagesResponse;
    const entries = payload.data ?? [];
    if (entries.length === 0) throw new Error('qwen-local returned no images');

    const format = request.output_format ?? 'png';
    const images = entries
      .filter((entry) => typeof entry.b64_json === 'string')
      .map((entry) => ({
        bytes: Uint8Array.from(atob(entry.b64_json!), (char) => char.charCodeAt(0)),
        format,
        ...(entry.revised_prompt ? { revisedPrompt: entry.revised_prompt } : {}),
      }));

    if (images.length === 0) throw new Error('qwen-local returned images with no b64_json');

    return {
      images,
      usage: payload.usage,
      // What was actually used, including the size we substituted — the caller asked for `auto` or for
      // nothing and got 512, and the row should record that rather than the absence.
      echoed: { model: MODEL_ID, size: sent.size, provider: 'qwen-local', ...(payload.quality ? { quality: payload.quality } : {}) },
    };
  },

  /**
   * Zero, and that is a fact rather than a gap.
   *
   * `null` here means "this model's price is not known", which is what the OpenAI adapter returns for
   * models it has no row for. Our own GPU costs no dollars per image, so zero is the true answer and
   * saying "unknown" would put this provider in the same bucket as one we simply failed to price.
   */
  costUsd(_model: string, _usage: ImageUsage | undefined): string {
    return '0.00';
  },
};