mlx-bun 0.0.13 → 0.3.0

package/src/serve/model-host.ts

@@ -0,0 +1,512 @@
+// Model-host lifecycle: load a model directory into the ServerContext that
+// createServer serves (model + tokenizer + template + lazy vision/audio
+// towers + adapters + optional draft), and the on-demand tower getters.
+// Extracted from src/server.ts (repo-taming Phase 4).
+import { existsSync, readFileSync } from "node:fs";
+import { bindLegacyDraftTarget } from "../backends/mlx/draft-target";
+import { loadModelConfig, type KvQuantSpec, type ModelConfig } from "../config";
+import { Weights } from "../weights";
+import { Gemma4Model } from "../model/gemma4";
+import {
+  createModel,
+  openGlm52RuntimeModel,
+  type Glm52RuntimeOpenOptions,
+  type RuntimeModel,
+} from "../model/factory";
+import { Glm52Model } from "../model/glm52";
+import { type Glm52MemoryPlan } from "../model/glm52-memory";
+import { resolveModelProfile, type ResolvedModelProfile } from "../model/profile";
+import { Glm52NativeMtpProvider } from "../spec/glm52-mtp-source";
+import { ChatTemplate } from "../chat-template";
+import { loadTokenizer, type LoadedTokenizer } from "../tokenizer";
+import { AdapterManager } from "../lora";
+import { VisionTower } from "../vision/embedder";
+import { SiglipVisionTower, parseSiglipConfig } from "../vision/siglip";
+import { Qwen3VLVisionTower } from "../vision/qwen3vl-tower";
+import { Qwen35Model } from "../model/qwen3_5";
+import type { AudioTokenIds, VisionTokenIds, VisionEncoder } from "../vision/prompt";
+import { AudioTower, parseAudioConfig } from "../audio/conformer";
+import { sidecarShipsAudioTower } from "../registry";
+import { fit } from "../fit";
+import type { ModelImplementationProvider } from "../model/implementation";
+import type { ServedModelInfo } from "./model-binding";
+
+export interface AdapterService extends Pick<AdapterManager, keyof AdapterManager> {}
+
+export interface ServerContext<Model = RuntimeModel> {
+  serving?: import("./model-binding").ModelServingBinding;
+  /** State serialization supplied by this model/backend implementation. */
+  stateCodecs?: import("../kv-store").CacheCodecProvider;
+  model: Model;
+  /** Declared external artifact/family profile that selected model
+   * construction. Request-level methods are resolved separately. */
+  profile: ResolvedModelProfile;
+  tokenizer: LoadedTokenizer;
+  template: ChatTemplate;
+  modelId: string;
+  /** Lazily-loaded vision tower cache — null until the first image request
+   *  (see `getVisionTower`). The tower (SigLIP ~hundreds of MB, encoder-free
+   *  smaller) is not loaded for text-only sessions. */
+  vision: VisionEncoder | null;
+  /** Loads + selects the vision tower on demand; null when the model has no
+   *  (supported) vision sidecar. Invoked at most once, then cached in
+   *  `vision`. */
+  loadVision: (() => VisionEncoder) | null;
+  visionTokenIds: VisionTokenIds;
+  /** Lazily-loaded Conformer audio tower — null until the first audio
+   *  request (see `getAudioTower`). Same sidecar file as vision
+   *  (optiq_vision.safetensors), separate tower; text-only sessions never
+   *  pay for it. */
+  audio: AudioTower | null;
+  /** Loads the audio tower on demand; null when the model can't do audio
+   *  (no `audio_config` in config.json or no sidecar). No flags — audio
+   *  auto-enables exactly like vision. */
+  loadAudio: (() => AudioTower) | null;
+  /** null when the model has no `audio_config` (audio-incapable). */
+  audioTokenIds: AudioTokenIds | null;
+  adapters: AdapterService;
+  /** Per-layer KV quantization from the repo's kv_config.json (null if
+   *  absent). Applied by default — optiq serve's headline behavior;
+   *  ServerOptions.kvQuant overrides ("off" | uniform bits). */
+  kvConfig: KvQuantSpec[] | null;
+  /** Model-author recommended sampling from generation_config.json —
+   *  optiq serve injects these as server defaults (gen_config.py);
+   *  explicit request fields always win. */
+  genDefaults: GenSamplingDefaults;
+  /** Speculative decoding (`serve --draft-model`, mlx_lm.server parity).
+   *  Server-level: when set, EVERY request routes to the serial lane
+   *  (upstream: is_batchable = draft is None) and spec-eligible ones decode
+   *  through src/spec/serve-loop.ts. null = no draft configured. */
+  draft?: {
+    provider: import("../spec/source").DraftProvider;
+    numDraftTokens: number;
+  } | null;
+  /** Present only for the direct Colibri GLM runtime. This is the exact
+   * header-derived process equation used before opening resident state. */
+  glmMemoryPlan?: Glm52MemoryPlan | null;
+}
+
+export type ServingContext = ServerContext<import("./model-binding").ServedModelInfo>;
+
+export interface ModelHostSource {
+  readonly modelDir: string;
+  readonly modelId: string | undefined;
+  readonly options: Omit<LoadContextOptions, "implementations">;
+}
+
+export interface LoadContextOptions<Model extends ServedModelInfo = RuntimeModel> {
+  /** Engine-owned complete implementations may own their loader and methods.
+   * Selection happens before opening the default resident or streamed weights. */
+  implementations?: ModelImplementationProvider<ModelHostSource, Promise<ServerContext<Model>>>;
+  profiles?: import("../model/profile").ResolveModelProfileOptions;
+  memoryBudgetBytes?: number;
+  /** Direct Colibri runtime/resource overrides. Ignored by other models. */
+  glm?: Glm52RuntimeOpenOptions;
+  /** Snapshot dir of a draft model for speculative decoding
+   * (`--draft-model`). Loaded alongside the target; the pair must share
+   * a tokenizer family. */
+  draftModelDir?: string;
+  /** Drafts per round (`--num-draft-tokens`, mlx_lm.server default 3). */
+  numDraftTokens?: number;
+  /** Draft-provider kind override (`--draft-kind`). */
+  draftKind?: DraftKind;
+  ngramMax?: number;
+  ngramMin?: number;
+}
+
+export interface GenSamplingDefaults {
+  temperature?: number;
+  topP?: number;
+  topK?: number;
+  repetitionPenalty?: number;
+}
+
+export type DraftKind = "dspark" | "deepspec" | "assistant" | "two-model" | "ngram" | "mtp";
+
+/** Detect the draft artifact's kind so the right provider is loaded. All
+ *  providers share ONE serve loop (src/spec/serve-loop.ts). Exported for the
+ *  bench harness (scripts/bench-matrix.ts features) — one detection, no drift.
+ *  "ngram" is never detected — it has no artifact (model-free prompt lookup,
+ *  src/spec/ngram-source.ts) and mounts via an explicit `--draft-kind ngram`. */
+export async function detectDraftKind(dir: string): Promise<DraftKind> {
+  if (await Bun.file(`${dir}/dspark.json`).exists()) return "dspark"; // our trained module
+  try {
+    const cfg = (await Bun.file(`${dir}/config.json`).json()) as {
+      model_type?: string;
+      architectures?: string[];
+    };
+    // DeepSeek's released DSpark drafters (DeepSpec reference): no
+    // dspark.json, plain HF config stamped Gemma4DSparkModel.
+    if (cfg.architectures?.[0] === "Gemma4DSparkModel") return "deepspec";
+    if (String(cfg.model_type ?? "").includes("assistant")) return "assistant";
+    // Native MTP heads split from a qwen3_5-family release
+    // (mlx-community/Qwen3.8-27B-MTP-*): model_type "qwen3_5_mtp". The
+    // target's recurrent DeltaNet caches roll back via the serve loop's
+    // spec-round snapshot/replay contract (SSMCache.specRound*).
+    if (String(cfg.model_type ?? "").endsWith("_mtp")) return "mtp";
+  } catch {
+    // no/unreadable config → fall through to a full second model
+  }
+  return "two-model";
+}
+
+async function loadGenSamplingDefaults(modelDir: string): Promise<GenSamplingDefaults> {
+  const file = Bun.file(`${modelDir}/generation_config.json`);
+  if (!(await file.exists())) return {};
+  try {
+    const raw = (await file.json()) as Record<string, unknown>;
+    const num = (v: unknown) => (typeof v === "number" ? v : undefined);
+    return {
+      temperature: num(raw.temperature),
+      topP: num(raw.top_p),
+      topK: num(raw.top_k),
+      repetitionPenalty: num(raw.repetition_penalty),
+    };
+  } catch {
+    return {};
+  }
+}
+
+export function loadContext<Model extends ServedModelInfo>(
+  modelDir: string, modelId: string | undefined,
+  opts: LoadContextOptions<Model> & {
+    implementations: ModelImplementationProvider<ModelHostSource, Promise<ServerContext<Model>>>;
+  },
+): Promise<ServerContext<Model>>;
+export function loadContext(modelDir: string, modelId?: string, opts?: LoadContextOptions): Promise<ServerContext>;
+export function loadContext<Model extends ServedModelInfo>(
+  modelDir: string, modelId: string | undefined, opts: LoadContextOptions<Model>,
+): Promise<ServingContext>;
+export async function loadContext(
+  modelDir: string, modelId?: string,
+  opts: LoadContextOptions<ServedModelInfo> = {},
+): Promise<ServingContext> {
+  const config = await loadModelConfig(modelDir);
+  const profile = resolveModelProfile(config, opts.profiles);
+  if (opts.implementations) {
+    const implementation = opts.implementations.select(config, profile);
+    const { implementations: _provider, ...options } = opts;
+    return implementation.create({ modelDir, modelId, options }, config, profile);
+  }
+  const glm = profile.profile.execution.loader === "colibri";
+  // Bundled MTP companion: `--draft-kind mtp` with no --draft-model resolves
+  // to the artifact's own mtp/ subfolder (single-repo packaging — the
+  // companion is a complete model dir the provider already loads). Explicit
+  // --draft-model still wins; a missing bundle is a clear refusal below.
+  if (opts.draftKind === "mtp" && !opts.draftModelDir) {
+    const bundled = `${modelDir}/mtp`;
+    if (await Bun.file(`${bundled}/config.json`).exists()) {
+      opts = { ...opts, draftModelDir: bundled };
+    } else {
+      throw new Error(
+        `--draft-kind mtp needs a companion: pass --draft-model <dir> or use ` +
+        `an artifact that bundles one at <model>/mtp/ (none at ${bundled})`,
+      );
+    }
+  }
+  const externalDraft = opts.draftModelDir !== undefined || opts.draftKind !== undefined;
+  const resolvedDraftKind = opts.draftModelDir
+    ? opts.draftKind ?? await detectDraftKind(opts.draftModelDir)
+    : opts.draftKind;
+  if (glm && externalDraft && opts.glm?.enableMtp === true)
+    throw new Error("native GLM MTP and --draft-model/--draft-kind are mutually exclusive");
+  let weights: Weights | null = null;
+  let model!: RuntimeModel;
+  let glmMemoryPlan: Glm52MemoryPlan | null = null;
+  if (glm) {
+    const enableNativeMtp = !externalDraft && opts.glm?.enableMtp !== false;
+    const opened = await openGlm52RuntimeModel(modelDir, {
+      ...opts.glm,
+      memoryBudgetBytes: opts.memoryBudgetBytes ?? opts.glm?.memoryBudgetBytes,
+      // Every drafter owns the serial speculative lane. Do not reserve a
+      // simultaneous ordinary batch that cannot be admitted while it is live.
+      batchSize: enableNativeMtp || externalDraft ? 1 : opts.glm?.batchSize,
+      // An explicit alternate drafter owns the one speculative lane.
+      enableMtp: enableNativeMtp,
+    });
+    model = opened.model;
+    glmMemoryPlan = opened.plan;
+  } else {
+    weights = await Weights.open(modelDir);
+  }
+  // memoryBudget enforcement at load (Phase 5): Weights.open only mmaps
+  // (no GPU allocation yet), so a model whose weights can never serve
+  // within the budget is refused HERE — before any unified-memory
+  // commitment — with an actionable error instead of a Metal OOM later.
+  if (!glm && opts.memoryBudgetBytes) {
+    const weightsBytes = [...weights!.shards.files.values()]
+      .reduce((a, f) => a + f.mmap.size, 0);
+    const report = fit(config, weightsBytes, 1, undefined, undefined, 0, opts.memoryBudgetBytes);
+    if (report.maxSafeContext < 1)
+      throw new Error(
+        `model does not fit the memory budget: weights ${(weightsBytes / 1e9).toFixed(2)} GB ` +
+        `+ prefill transient leave no room for any context within ` +
+        `${(opts.memoryBudgetBytes / 1e9).toFixed(2)} GB`,
+      );
+  }
+  if (!glm) model = createModel(weights!, config, profile);
+  const tokenizer = await loadTokenizer(modelDir);
+  // Generation must stop on the tokenizer's eos_token — the chat turn
+  // terminator (e.g. Qwen <|im_end|> = 248046). Some configs (Qwen3.5-4B)
+  // declare a different eos_token_id in config.json than the chat format
+  // emits, so without this a turn never ends and generation runs away,
+  // hallucinating both sides of the dialogue until max_tokens. mlx-lm stops on
+  // the tokenizer eos; union it in. No-op when already present (Gemma, 27B).
+  if (tokenizer.eosTokenId != null && !config.eosTokenIds.includes(tokenizer.eosTokenId))
+    config.eosTokenIds = [...config.eosTokenIds, tokenizer.eosTokenId];
+
+  // Speculative decoding: load the draft (mlx_lm.server --draft-model). The
+  // draft artifact's KIND selects the provider — all three share ONE serve
+  // loop (src/spec/serve-loop.ts): dspark.json → DSpark (KV-injected), a
+  // *_assistant config → the optiq KV-borrowing Gemma drafter, otherwise a
+  // full second model (mlx-lm parity). `--draft-kind` overrides the detect.
+  let draft: ServerContext["draft"] = null;
+  if (opts.draftKind === "ngram") {
+    // Model-free prompt lookup: no artifact, no dir, no probe/budget concerns
+    // (weightsBytes 0, open() never throws). Default γ=10 per the reference
+    // implementation — drafting is free, so wide blocks cost only verify-window
+    // width when wrong.
+    if (opts.draftModelDir)
+      throw new Error(
+        "--draft-kind ngram is model-free — drop --draft-model (it would be ignored)",
+      );
+    const { NgramProvider } = await import("../spec/ngram-source");
+    draft = {
+      provider: new NgramProvider({ max: opts.ngramMax, min: opts.ngramMin }),
+      numDraftTokens: Math.max(1, opts.numDraftTokens ?? 10),
+    };
+  } else if (opts.draftModelDir) {
+    const dir = opts.draftModelDir;
+    const kind = resolvedDraftKind!;
+    let provider: import("../spec/source").DraftProvider;
+    let numDraftTokens = Math.max(1, opts.numDraftTokens ?? 3);
+    if (kind === "dspark") {
+      const { DflashProvider } = await import("../spec/dflash-source");
+      const p = await DflashProvider.load(dir);
+      provider = p;
+      // Pin to the trained block width — the serve loop must never ask for
+      // more positions than the DSpark block was trained for (n ≤ cfg.gamma).
+      numDraftTokens = Math.max(1, Math.min(opts.numDraftTokens ?? p.gamma, p.gamma));
+    } else if (kind === "deepspec") {
+      const { DeepspecProvider } = await import("../spec/deepspec-source");
+      const p = await DeepspecProvider.load(dir);
+      provider = p;
+      // Same pin, from their config's block_size (e.g. 7 for the released
+      // dspark_gemma4_12b_block7).
+      numDraftTokens = Math.max(1, Math.min(opts.numDraftTokens ?? p.gamma, p.gamma));
+    } else if (kind === "assistant") {
+      const { AssistantProvider } = await import("../spec/assistant-source");
+      provider = await AssistantProvider.load(dir);
+    } else if (kind === "mtp") {
+      const { QwenMtpProvider } = await import("../spec/qwen-mtp-source");
+      const p = await QwenMtpProvider.load(dir);
+      provider = p;
+      // Default the round width to the head's trained block (block_size 3 →
+      // 2 recursive drafts + the pending row per round; the head was trained
+      // multi-step, so an explicit larger --num-draft-tokens is allowed but
+      // acceptance decides whether it pays).
+      const block = (await Bun.file(`${dir}/config.json`).json() as { block_size?: number }).block_size;
+      if (opts.numDraftTokens === undefined && typeof block === "number")
+        numDraftTokens = Math.max(1, block - 1);
+    } else {
+      const { TwoModelProvider } = await import("../spec/two-model");
+      provider = await TwoModelProvider.load(dir, config.text.vocabSize);
+      // Tokenizer-family hard check — two-model ONLY (it ships its own
+      // tokenizer). Exact-token-match acceptance is meaningful only when both
+      // models tokenize identically; a probe that ENCODES differently means
+      // different families — refuse instead of silently accepting ~0% of
+      // drafts. The KV-borrowing drafters share the target's tokenization by
+      // construction, so this check does not apply to them.
+      const draftTok = await loadTokenizer(dir);
+      const probe = "The 3 quick brown foxes jumped över the lazy dog?! 🦊";
+      if (JSON.stringify(tokenizer.encode(probe)) !== JSON.stringify(draftTok.encode(probe))) {
+        provider.dispose();
+        throw new Error(
+          `--draft-model tokenizer differs from the target's (probe string encodes ` +
+            `differently) — speculation needs the same tokenizer family`,
+        );
+      }
+    }
+    // Fail-fast pairing validation (2026-07-07 review): the KV-borrowing
+    // sources validate the (target, drafter) pairing in open() — non-Gemma4
+    // target, DeepSpec target-layer-count mismatch — which used to surface
+    // as a 500 from inside specServeRun on EVERY text request. Probe-open
+    // once with throwaway caches here so a mismatch refuses at load; open()
+    // allocates no per-request tensors before prefill/draft, so this is
+    // free. The probe sampler is never called during open().
+    {
+      const probeCaches = model.makeCache();
+      try {
+        provider
+          .open({
+            sampler: () => { throw new Error("probe sampler never samples"); },
+            target: bindLegacyDraftTarget(model, probeCaches),
+          })
+          .dispose();
+      } catch (err) {
+        provider.dispose();
+        throw new Error(
+          `--draft-model is incompatible with this target: ${err instanceof Error ? err.message : String(err)}`,
+        );
+      } finally {
+        for (const c of probeCaches) c.dispose();
+      }
+    }
+    if (opts.memoryBudgetBytes) {
+      const targetBytes = weights
+        ? [...weights.shards.files.values()].reduce((a, f) => a + f.mmap.size, 0)
+        : model.weightsBytes;
+      // Draft weights shrink the target's envelope. Draft KV is not modeled
+      // (small relative to its weights at serve contexts); admission stays
+      // approximately conservative via the combined-weights term.
+      const report = fit(config, targetBytes + provider.weightsBytes, 1, undefined, undefined, 0, opts.memoryBudgetBytes);
+      if (report.maxSafeContext < 1) {
+        provider.dispose();
+        throw new Error(
+          `target + draft do not fit the memory budget (draft adds ` +
+            `${(provider.weightsBytes / 1e9).toFixed(2)} GB)`,
+        );
+      }
+    }
+    draft = { provider, numDraftTokens };
+  } else if (opts.draftKind) {
+    // Every other kind names an artifact to load — refuse instead of silently
+    // serving without speculation.
+    throw new Error(`--draft-kind ${opts.draftKind} requires --draft-model`);
+  }
+
+  // GLM's checkpoint-native MTP row is the production default. It uses the
+  // already-planned bounded auxiliary expert tier and the same tokenizer, so
+  // there is no second artifact or compatibility probe to load.
+  if (!draft && model instanceof Glm52Model && glmMemoryPlan?.enableMtp) {
+    draft = {
+      provider: new Glm52NativeMtpProvider(model),
+      numDraftTokens: glmMemoryPlan.mtpDraftTokens,
+    };
+  }
+
+  return {
+    draft,
+    model,
+    profile,
+    glmMemoryPlan,
+    adapters: new AdapterManager(model),
+    kvConfig: config.kvQuant,
+    genDefaults: await loadGenSamplingDefaults(modelDir),
+    tokenizer,
+    template: await ChatTemplate.load(modelDir),
+    modelId: modelId ?? modelDir.split("/").filter(Boolean).at(-1)!,
+    // Vision is loaded lazily (getVisionTower) — text-only sessions never
+    // pay for the tower. The loader picks the encoder-free gemma4_unified
+    // (12B) tower vs the SigLIP encoder (e2b/e4b/26B/31B) by the sidecar's
+    // vision_config.model_type. Vision sidecars are a Gemma4 feature;
+    // MiniCPM5 never ships one.
+    vision: null,
+    loadVision: makeVisionLoader(modelDir, model, config),
+    visionTokenIds: {
+      imageTokenId: (config.raw.image_token_id as number) ?? 258880,
+      boiTokenId: (config.raw.boi_token_id as number) ?? 255999,
+      eoiTokenId: (config.raw.eoi_token_id as number) ?? 258882,
+    },
+    // Audio mirrors vision: lazy tower from the same sidecar, loaded on the
+    // first audio request only (docs/design/generic-model-support.md A4).
+    audio: null,
+    loadAudio: makeAudioLoader(modelDir, model, config),
+    audioTokenIds: config.raw.audio_config
+      ? {
+          audioTokenId: (config.raw.audio_token_id as number) ?? 258881,
+          boaTokenId: (config.raw.boa_token_id as number) ?? 256000,
+          eoaTokenId: (config.raw.eoa_token_id as number) ?? 258883,
+        }
+      : null,
+  };
+}
+
+/** Build the on-demand vision-tower loader, selecting the encoder-free
+ *  (gemma4_unified, 12B) tower vs the SigLIP encoder (gemma4_vision:
+ *  e2b/e4b/26B/31B) by the sidecar's vision_config.model_type. Returns null
+ *  when the model has no usable vision sidecar. */
+export function makeVisionLoader(
+  modelDir: string, model: RuntimeModel, config: ModelConfig,
+): (() => VisionEncoder) | null {
+  // Qwen3.5/3.8: the tower ships as optiq/optiq_vision.safetensors (the
+  // artifact's bf16 sidecar). The returned tower is a Qwen3VLVisionTower —
+  // it rides the same lazy slot/capability flags; the qwen chat branch is
+  // the only consumer and casts it back (the gemma branches are gated on
+  // `instanceof Gemma4Model`, so the union never crosses).
+  if (model instanceof Qwen35Model) {
+    // Vision weights arrive either as the OptiQ-convention sidecar OR in-main
+    // (mlx-vlm convention; our artifacts ship one copy in-main since
+    // 2026-08-18 — the tower loader handles both).
+    const hasSidecar = existsSync(`${modelDir}/optiq/optiq_vision.safetensors`);
+    const hasInMain = config.raw.vision_config !== undefined;
+    if (!hasSidecar && !hasInMain) return null;
+    return () =>
+      Qwen3VLVisionTower.load(modelDir) as unknown as VisionEncoder;
+  }
+  if (!(config.hasVisionSidecar && model instanceof Gemma4Model)) return null;
+  const vc = config.raw.vision_config as Record<string, any> | undefined;
+  if (vc?.model_type === "gemma4_vision") {
+    const sigCfg = parseSiglipConfig(vc);
+    return () => SiglipVisionTower.load(modelDir, sigCfg, model.embedScale);
+  }
+  // gemma4_unified_vision (or unlabelled): the encoder-free patch embedder.
+  return () => VisionTower.load(modelDir, model.embedScale, config.text.rmsNormEps);
+}
+
+/** Lazily load + cache the vision tower on first use. A sidecar that fails
+ *  to load is a capability gap, not a fatal error: returns null and the
+ *  request is answered with a 400 (the loader is cleared so we don't retry
+ *  a known-bad load every request). */
+export function getVisionTower(ctx: ServerContext): VisionEncoder | null {
+  if (ctx.vision) return ctx.vision;
+  if (!ctx.loadVision) return null;
+  try {
+    ctx.vision = ctx.loadVision();
+    return ctx.vision;
+  } catch (e) {
+    console.warn(`vision sidecar not loadable (${(e as Error).message}) — serving text-only`);
+    ctx.loadVision = null;
+    return null;
+  }
+}
+
+/** Build the on-demand audio-tower loader (gemma-4 Conformer, A4 of
+ *  docs/design/generic-model-support.md). Auto-enables — no flags — when
+ *  config.json carries an `audio_config` AND the optiq_vision.safetensors
+ *  sidecar exists (the audio tensors ship in the same sidecar as vision).
+ *  Returns null when the model can't do audio (no audio_config: 26B-A4B,
+ *  DiffusionGemma, bf16 assistants — architectural, not a porting gap). */
+export function makeAudioLoader(
+  modelDir: string, model: RuntimeModel, config: ModelConfig,
+): (() => AudioTower) | null {
+  if (!(config.hasVisionSidecar && config.raw.audio_config && model instanceof Gemma4Model))
+    return null;
+  // The sidecar header must actually name the Conformer tensors: the local
+  // 12B snapshot pairs audio_config with a STUB sidecar (embed_audio only),
+  // and a non-null loader here is what advertises `audio: true` on every
+  // capability surface (ws handshake, /v1/models). Header-only read.
+  if (!sidecarShipsAudioTower(`${modelDir}/optiq_vision.safetensors`)) return null;
+  const audioCfg = parseAudioConfig(config.raw.audio_config as Record<string, any>);
+  return () => AudioTower.load(modelDir, audioCfg, model.embedScale);
+}
+
+/** Lazily load + cache the audio tower on first use. Unlike vision's
+ *  warn-and-continue (text-only degrade is correct for requests WITHOUT
+ *  images), this is only ever consulted for requests WITH audio — the
+ *  caller turns null into an explicit 400, never a silent text-only
+ *  degrade. A failed load is not retried every request. (The stub-sidecar
+ *  case — the local 12B state — never gets here: makeAudioLoader checks the
+ *  sidecar header and returns a null loader.) */
+export function getAudioTower(ctx: ServerContext): AudioTower | null {
+  if (ctx.audio) return ctx.audio;
+  if (!ctx.loadAudio) return null;
+  try {
+    ctx.audio = ctx.loadAudio();
+    return ctx.audio;
+  } catch (e) {
+    console.warn(`audio sidecar not loadable (${(e as Error).message})`);
+    ctx.loadAudio = null;
+    return null;
+  }
+}

package/src/serve/openai-wire.ts

@@ -0,0 +1,103 @@
+// OpenAI wire formats: InferenceResult → chat.completion / text_completion
+// JSON, and the SSE chunk protocols for both. Anthropic and Responses have
+// their own (src/anthropic.ts, src/responses.ts); all three implement the
+// same CompletionStreamProtocol so the HTTP writer (http.ts) is shared.
+import type { CompletionSummary, CompletionUsage } from "./completion-executor";
+import type { CompletionEvent, CompletionStreamProtocol } from "./completion-sink";
+import type { InferenceResult } from "./inference-request";
+
+export interface OpenAiResponseMeta {
+  id: string;
+  created: number;
+  model: string;
+}
+
+/** OpenAI usage block from the executor's live/terminal usage view. */
+export const openAiUsage = (usage: CompletionUsage) => ({
+  prompt_tokens: usage.promptTokens,
+  completion_tokens: usage.completionTokens,
+  total_tokens: usage.totalTokens,
+  prompt_tokens_details: { cached_tokens: usage.cachedTokens },
+  ...(usage.speculation ? { speculation: usage.speculation } : {}),
+  ...(usage.fill ? { fill: usage.fill } : {}),
+});
+
+/** Terminal usage: the wire block plus mlx-bun's `lane` (batch|serial|…). */
+export const openAiSummaryUsage = (summary: CompletionSummary) => ({
+  ...openAiUsage(summary.usage),
+  lane: summary.lane,
+});
+
+export function chatCompletionJson(result: InferenceResult, meta: OpenAiResponseMeta) {
+  return {
+    ...meta, object: "chat.completion",
+    choices: [{
+      index: 0,
+      message: {
+        role: "assistant",
+        content: result.content || (result.toolCalls.length ? null : ""),
+        ...(result.reasoning ? { reasoning: result.reasoning } : {}),
+        ...(result.toolCalls.length ? { tool_calls: result.toolCalls } : {}),
+      },
+      ...(result.logprobs ? { logprobs: result.logprobs } : {}),
+      finish_reason: result.finishReason,
+    }],
+    usage: openAiSummaryUsage(result),
+  };
+}
+
+export function textCompletionJson(result: InferenceResult, meta: OpenAiResponseMeta) {
+  return {
+    ...meta, object: "text_completion",
+    choices: [{
+      index: 0, text: result.content,
+      ...(result.logprobs ? { logprobs: result.logprobs } : {}),
+      finish_reason: result.finishReason,
+    }],
+    usage: openAiSummaryUsage(result),
+  };
+}
+
+const sse = (obj: unknown) => `data: ${JSON.stringify(obj)}\n\n`;
+const DONE = "data: [DONE]\n\n";
+
+/** OpenAI chat SSE: a role primer chunk, one chunk per event (content /
+ *  reasoning / tool_calls delta), a terminal chunk carrying finish_reason +
+ *  usage, then the bare `[DONE]` sentinel strict SDK clients require. An
+ *  error mid-stream is a single `{error}` frame and nothing after it. */
+export function chatCompletionStream(meta: OpenAiResponseMeta): CompletionStreamProtocol {
+  const chunk = (delta: Record<string, unknown>, finish: string | null) => ({
+    ...meta, object: "chat.completion.chunk",
+    choices: [{ index: 0, delta, finish_reason: finish }],
+  });
+  let errored = false;
+  return {
+    start: () => [sse(chunk({ role: "assistant", content: "" }, null))],
+    addEvents: (events: CompletionEvent[]) => events.map((event) =>
+      event.type === "reasoning" ? sse(chunk({ reasoning: event.text }, null))
+      : event.type === "content" ? sse(chunk({ content: event.text }, null))
+      : sse(chunk({ tool_calls: event.calls.map((call, index) => ({ index, ...call })) }, null))),
+    finish: (reason, usage) => errored ? [] : [sse({ ...chunk({}, reason), usage }), DONE],
+    error: (message) => { errored = true; return [sse({ error: { message } })]; },
+  };
+}
+
+/** OpenAI text-completion SSE: content deltas as `text`, then a final chunk
+ *  with finish_reason + usage (mlx-lm gates usage behind
+ *  stream_options.include_usage; we always attach it — an additive superset
+ *  OpenAI clients ignore), then `[DONE]`. */
+export function textCompletionStream(meta: OpenAiResponseMeta): CompletionStreamProtocol {
+  const chunk = (text: string, finish: string | null) => ({
+    ...meta, object: "text_completion",
+    choices: [{ index: 0, text, finish_reason: finish }],
+  });
+  let errored = false;
+  return {
+    start: () => [],
+    addEvents: (events: CompletionEvent[]) => events
+      .filter((event) => event.type === "content")
+      .map((event) => sse(chunk((event as { text: string }).text, null))),
+    finish: (reason, usage) => errored ? [] : [sse({ ...chunk("", reason), usage }), DONE],
+    error: (message) => { errored = true; return [sse({ error: { message } })]; },
+  };
+}

package/src/serve/parent-application.ts

@@ -0,0 +1,56 @@
+import { handleStaticRoute } from "./static-routes";
+import { STATIC_ROUTE_ASSETS } from "../web-assets";
+import { createResponsesClient } from "./responses-client";
+import type { DisposableResource } from "../contracts/resources";
+import type { JobStore } from "../jobs/db";
+import { handleLabRoute, matchLabRoute } from "./lab-routes";
+import { handleAdminRoute, matchAdminRoute } from "./admin-routes";
+
+/** CPU application state stays in the parent when model workers restart/evict.
+ * Native inspection/merge/export routes still execute inside a model worker. */
+export function createParentApplication(options: {
+  acquireGpu(signal: AbortSignal): Promise<DisposableResource>;
+  serverPort(): number | undefined;
+  invalidateLibrary(): void;
+  createJobStore?(): Promise<JobStore>;
+}) {
+  const responses = createResponsesClient();
+  let jobs: Promise<JobStore> | undefined;
+  let closed = false;
+  const ensureJobs = () => {
+    if (closed) return Promise.reject(new Error("application host is closed"));
+    return jobs ??= (async () => {
+      const { JobStore, registerRunner } = await import("../jobs");
+      const store = options.createJobStore ? await options.createJobStore() : new JobStore();
+      store.markZombies();
+      const { datasetRunner } = await import("../dataset/job");
+      registerRunner("dataset", datasetRunner);
+      return store;
+    })();
+  };
+  return {
+    get responseStats() { return responses.stats; },
+    async handle(request: Request, forward?: (request: Request) => Promise<Response>): Promise<Response | null> {
+      const url = new URL(request.url);
+      const staticResponse = handleStaticRoute(url, request, STATIC_ROUTE_ASSETS);
+      if (staticResponse) return staticResponse;
+      if (forward && request.method === "POST" && url.pathname === "/v1/responses")
+        return responses.forward(request, forward);
+      const admin = matchAdminRoute(request.method, url.pathname);
+      if (admin) return handleAdminRoute(url, request, { ensureJobs, invalidateLibrary: options.invalidateLibrary });
+      const route = matchLabRoute(request.method, url.pathname);
+      if (route !== "quantize-submit" && route !== "finetune-submit" &&
+          route !== "dataset-submit" && route !== "dataset-templates") return null;
+      return handleLabRoute(url, request, { ensureJobs, acquireGpu: options.acquireGpu,
+        serverPort: options.serverPort, invalidateLibrary: options.invalidateLibrary });
+    },
+    async close() {
+      closed = true;
+      if (!jobs) return;
+      const store = await jobs;
+      const { closeSubprocessJobs, closeInProcessJobs } = await import("../jobs");
+      await Promise.all([closeSubprocessJobs(store), closeInProcessJobs(store)]);
+      store.close();
+    },
+  };
+}