mlx-bun 0.4.0 → 1.0.0

package/src/spec/source.ts

@@ -1,215 +0,0 @@
-// DraftSource — the seam between the serve-time speculative verify loop
-// (src/spec/serve-loop.ts) and whatever produces draft tokens. Designed in
-// docs/reference/cli.md §7 so every drafter shares ONE
-// verify/accept executor:
-//   - TwoModelSource (src/spec/two-model.ts) — mlx-lm parity (L1 oracle:
-//     mlx_lm.server --draft-model), a full second model. Ignores `target`.
-//   - AssistantSource (src/spec/assistant-source.ts) — the optiq KV-borrowing
-//     Gemma drafter (src/spec/drafter.ts; L2 oracle: optiq spec_generate).
-//     Reads the target's donor K/V + anchor hidden each step.
-//   - DflashSource (src/spec/dflash-source.ts) — DSpark (L3, KL/quality-gated).
-//     Taps the target's multi-layer hiddens (prefill + verify) into a growing
-//     H_ctx (docs/design/speculative-decoding.md).
-//   - NgramSource (src/spec/ngram-source.ts) — model-free prompt lookup.
-//   - Glm52NativeMtpSource (src/spec/glm52-mtp-source.ts) — the target
-//     artifact's native Colibri MTP row.
-//
-// The sources differ ONLY in what fills the draft; the serve loop, admission
-// accounting, and stats never change. The KV-borrowing sources need target
-// state the two-model source doesn't — carried by the `target` view at open()
-// and the optional prefill/draft/commit arguments below (all no-ops for
-// two-model). See [[dspark-seam-kv-borrowing]].
-
-import type { MlxArray } from "../mlx/array";
-import type { Dtype } from "../mlx/ffi";
-import type { Cache } from "../model/gemma4";
-import type { CheckpointAttachment } from "../backends/mlx/checkpoint-state";
-import type { MlxDraftRows } from "../backends/mlx/speculative-round";
-import type { PreparedStateChange } from "../contracts/resources";
-
-/** Numerical ports are backend-specific; returned arrays are caller-owned. */
-export interface DraftProjection {
-  readonly embed: {
-    encode(ids: MlxArray): MlxArray;
-    readonly scales: { readonly dtype: Dtype };
-  };
-  logitsFromHidden(hidden: MlxArray): MlxArray;
-}
-
-/** Read-only target state and embedding, independent of draft scheduling. */
-export interface AssistantRowsTarget {
-  readonly hiddenSize: number;
-  embed(ids: MlxArray): MlxArray;
-  readDonors(): import("./drafter").AssistantDonors & {
-    readonly positions: readonly number[];
-    dispose(): void;
-  };
-}
-
-export interface QwenMtpTarget {
-  readonly hiddenSize: number;
-  readonly layerCount: number;
-  embed(ids: MlxArray): MlxArray;
-  logitsFromHidden(hidden: MlxArray): MlxArray;
-}
-
-/** Graph-declared extensions over the target's live state. Sources request
- * only the ports they consume; they never inspect a concrete model or cache.
- * Absence refuses an unsupported pairing before draft-side allocation. */
-export interface TargetView {
-  /** Opaque identity for providers borrowing weights from one exact target. */
-  readonly identity: object;
-  readonly assistantRows?: AssistantRowsTarget;
-  readonly gemmaTaps?: { readonly layerCount: number; readonly projection: DraftProjection };
-  readonly qwenMtp?: QwenMtpTarget;
-}
-
-/** A per-request draft-token producer. Created per generation (owns its own
- *  draft-side state), disposed by the serve loop's finally. */
-export interface DraftSource {
-  /** Method-owned companion state, independent of retention and storage.
-   * Capture returns owned immutable views. Restore borrows its attachment;
-   * the source retains the views it needs before returning. */
-  readonly checkpoint?: {
-    readonly namespace: string;
-    capture(processedTokens: number): import("../backends/mlx/checkpoint-state").CheckpointAttachment;
-    restore(processedTokens: number, attachment: import("../backends/mlx/checkpoint-state").CheckpointAttachment): void;
-  };
-  /** Target prefill shape required by this source's oracle. Most mlx-lm
-   *  sources leave the final prompt token pending; native Colibri MTP starts
-   *  from a full-prompt target forward. */
-  readonly prefillMode?: "tail-split" | "full";
-  /** Request one fixed target kernel family across the speculative verify
-   * batch. Native GLM MTP uses the direct-Colibri SPEC_PIN contract. */
-  readonly pinTargetKernelFamily?: boolean;
-
-  /** Multi-layer target tap the source needs captured on the target's prefill
-   *  AND every verify forward (DSpark's H_ctx; e4b {20,31,41,42}). When set,
-   *  the serve loop sets model.hiddenTap around those forwards and passes the
-   *  captured context [1,L,m*H] into prefill()/commit(). Undefined for sources
-   *  that don't tap (two-model, assistant). */
-  readonly tapLayers?: number[];
-
-  /** Process the prompt (two-model: prefill the draft model's cache;
-   *  assistant/dflash: read the target's state — mostly a no-op, but DSpark
-   *  seeds H_ctx from `ctxML`, the tapped prefill context [1,Lp,m*H], present
-   *  iff tapLayers is set). A composed prefill may supply its exact processed
-   *  token count, so a history source need not infer the target's boundary. */
-  prefill(promptIds: number[], ctxML?: MlxArray, processedTokens?: number): void | Promise<void>;
-
-  /** Propose 0..n tokens (RETURN LENGTH IS AUTHORITATIVE — a source may
-   *  return fewer than n, e.g. DSpark's confidence-scheduled draft-length
-   *  pruning; ZERO means "skip drafting this round" — DeepSpec ℓ=0 semantics,
-   *  the serve loop degenerates to one plain target step, still tapped +
-   *  committed for context-growing sources). The serve loop verifies over
-   *  exactly the returned length. `feed` is the token(s) the draft has not
-   *  yet consumed: [pending]
-   *  after a rejection/first round, or [lastDraft, bonus] after an all-accept
-   *  round (mlx-lm's re-feed rule, generate.py:645-648). `stepBase` is the
-   *  emitted-token index, threaded to the sampler for per-step RNG streams.
-   *  `anchorHidden` is the target's final hidden [1,1,H] at the pending/anchor
-   *  position — the assistant source borrows it for its first draft step;
-   *  two-model and dflash ignore it. */
-  draft(
-    feed: number[],
-    n: number,
-    stepBase: number,
-    anchorHidden?: MlxArray,
-  ): number[] | Promise<number[]>;
-
-  /** Verify outcome for the last round: kAccept of d accepted, where d is the
-   *  length draft() actually RETURNED (≤ n). Two-model: trim the draft cache
-   *  by max(d - kAccept - 1, 0) — mlx-lm's rewind rule (generate.py:589-591).
-   *  DSpark: grow H_ctx by the accepted window from `vCtxML`, the verified
-   *  window's tapped context [1,d+1,m*H] (present iff tapLayers is set), and
-   *  drop the rejected tips. Native GLM MTP receives the target's verified
-   *  hidden window [1,d+1,H] in `verifiedHidden` and the kAccept accepted token
-   *  ids in `acceptedTokens`; it uses them to rebuild accepted MTP KV rows
-   *  from target state rather than recursively drafted hidden state. */
-  commit(
-    d: number,
-    kAccept: number,
-    vCtxML?: MlxArray,
-    verifiedHidden?: MlxArray,
-    acceptedTokens?: readonly number[],
-  ): void | Promise<void>;
-
-  /** Resident per-request draft weights, for admission accounting (0 when the
-   *  provider owns the weights). */
-  readonly weightsBytes: number;
-  dispose(): void;
-}
-
-export interface DraftRowSampling {
-  /** Borrow [B,V] log-probabilities; return owned [B] IDs on device. */
-  sample(logprobs: MlxArray, steps: readonly number[]): MlxArray;
-}
-
-export interface DraftRowCheckpoint {
-  readonly processedTokens: number;
-  readonly attachment: CheckpointAttachment;
-}
-
-/** Method-owned state membership, independent of scheduling and cache tiers.
- * Open/append borrow checkpoints; capture returns owned immutable state.
- * Membership changes and capture occur only between committed rounds.
- * The caller retains the provider's residency lease for the group's lifetime. */
-export interface DraftRowGroup extends MlxDraftRows {
-  readonly namespace: string;
-  readonly tapLayers: readonly number[];
-  readonly rowCount: number;
-  append(checkpoints: readonly DraftRowCheckpoint[]): void;
-  prepareAppend(checkpoints: readonly DraftRowCheckpoint[]): PreparedStateChange;
-  filterRows(keep: readonly number[]): void;
-  capture(row: number): DraftRowCheckpoint;
-  dispose(): void;
-}
-
-/** Draft preparation consumes each target chunk at the same batch geometry.
- * Inputs are borrowed and contain only new tokens/context, with equal chunk
- * lengths across rows. A null checkpoint appends a cold request. The provider
- * owns alignment and retained hidden state; scheduling and persistence do not.
- * Membership and capture occur between prefill calls. Capture returns owned
- * state after at least one target token has been consumed for that row. */
-export interface DraftPrefillGroup {
-  readonly namespace: string;
-  readonly prefillMode: NonNullable<DraftSource["prefillMode"]>;
-  readonly tapLayers: readonly number[];
-  readonly rowCount: number;
-  append(checkpoints: readonly (DraftRowCheckpoint | null)[]): void;
-  prepareAppend(checkpoints: readonly (DraftRowCheckpoint | null)[]): PreparedStateChange;
-  prefill(tokens: MlxArray, context?: MlxArray): void | Promise<void>;
-  /** Resolve borrowed backing before its external restore lease is released. */
-  materialize(): void;
-  filterRows(keep: readonly number[]): void;
-  capture(row: number): DraftRowCheckpoint;
-  dispose(): void;
-}
-
-export interface GroupedDraftProvider {
-  /** Draft state remains valid when target forwards run under a mounted
-   * adapter context. Unqualified learned draft graphs leave this absent. */
-  readonly supportsTargetAdapters?: boolean;
-  openPrefill(options: { target: TargetView;
-    checkpoints: readonly (DraftRowCheckpoint | null)[] }): DraftPrefillGroup;
-  open(options: { target: TargetView; sampling: DraftRowSampling;
-    checkpoints: readonly DraftRowCheckpoint[] }): DraftRowGroup;
-}
-
-/** Server-lifetime owner of the draft machinery (the loaded draft model);
- * open() mints a per-request DraftSource; grouped opens shared row state. */
-export interface DraftProvider {
-  /** Human-readable id (registry id / path tail) for logs + cache namespacing. */
-  readonly id: string;
-  readonly weightsBytes: number;
-  readonly grouped?: GroupedDraftProvider;
-  open(opts: {
-    /** The request's sampler over logprobs [1,V] → token array [1] (the SAME
-     *  sampler as the target — mlx-lm parity; greedy drafting under a
-     *  temperature>0 request is NOT parity). */
-    sampler: (logprobs: MlxArray, step: number) => MlxArray;
-    /** Bound target capabilities, with explicit donor-view ownership. */
-    target: TargetView;
-  }): DraftSource;
-  dispose(): void;
-}

package/src/spec/two-model-rows.ts

@@ -1,61 +0,0 @@
-import type { MlxAutoregressiveBinding } from "../backends/mlx/autoregressive";
-import { MlxAutoregressiveDraftRows, type MlxDraftLayout } from "../backends/mlx/autoregressive-draft-rows";
-import { prefillCacheLayout } from "../backends/mlx/cache-layout";
-import { MlxStateRows } from "../backends/mlx/state-rows";
-import { captureCacheAttachment, restoreCacheAttachment } from "../backends/mlx/cache-attachment";
-import { applyStateChanges, cleanupFailure, disposeResources } from "../engine/resources";
-import type { Cache } from "../model/gemma4-base";
-import type { DraftPrefillGroup, DraftRowCheckpoint, DraftRowGroup, DraftRowSampling, GroupedDraftProvider } from "./source";
-import type { PreparedStateChange } from "../contracts/resources";
-import type { MlxArray } from "../mlx/array";
-
-export interface StandaloneDraftRows extends DraftPrefillGroup {
-  draft: DraftRowGroup["draft"];
-  commit(accepted: readonly number[], context?: MlxArray): Promise<void>;
-}
-
-/** The standalone model supplies graph/state capabilities once. Target
- * verification, scheduling and cache tiers use the existing provider ports. */
-export function openStandaloneDraftRows(binding: MlxAutoregressiveBinding, namespace: string,
-  sampling: DraftRowSampling | null, checkpoints: readonly (DraftRowCheckpoint | null)[]): StandaloneDraftRows {
-    const empty = binding.makeCache(), layouts: MlxDraftLayout[] = [];
-    try { for (const cache of empty) layouts.push(prefillCacheLayout(cache) as MlxDraftLayout); }
-    catch (error) { disposeResources(layouts); throw error; }
-    finally { disposeResources(empty); }
-    const state = new MlxStateRows(layouts), rows = new MlxAutoregressiveDraftRows(binding.graph, state, sampling);
-    const prepareAppend = (checkpoints: readonly (DraftRowCheckpoint | null)[]): PreparedStateChange => {
-      const restored: Cache[][] = [], pending: Array<number | null> = [];
-      try {
-        for (const checkpoint of checkpoints) {
-          restored.push(checkpoint ? restoreCacheAttachment(checkpoint.attachment, binding.makeCache) : binding.makeCache());
-          pending.push(checkpoint && checkpoint.attachment.metadata.pendingLast !== false ? Number(checkpoint.attachment.metadata.pendingLast) : null);
-        }
-        return rows.prepareAppend(restored, pending);
-      } finally { disposeResources(restored.flat()); }
-    };
-    const append = (checkpoints: readonly (DraftRowCheckpoint | null)[]) => applyStateChanges([() => prepareAppend(checkpoints)]);
-    try { append(checkpoints); }
-    catch (error) { return cleanupFailure(error, () => rows.dispose()); }
-    return {
-      namespace, prefillMode: "tail-split", tapLayers: [],
-      get rowCount() { return state.rowCount; },
-      append, prepareAppend, prefill: tokens => rows.prefill(tokens), materialize: () => rows.materialize(),
-      filterRows: keep => rows.filterRows(keep), draft: rows.draft.bind(rows), commit: accepted => rows.commit(accepted),
-      capture(row): DraftRowCheckpoint {
-        const caches = state.extractRow(row), pendingLast = rows.pendingLast[row] ?? null;
-        try {
-          const processedTokens = caches[0]!.offset + Number(pendingLast !== null);
-          return { processedTokens, attachment: captureCacheAttachment("standalone-draft-v1", caches,
-            { processedTokens, pendingLast: pendingLast ?? false }) };
-        } finally { disposeResources(caches); }
-      },
-      dispose: () => rows.dispose(),
-    };
-}
-
-export function standaloneDraftGroups(binding: MlxAutoregressiveBinding, namespace: string): GroupedDraftProvider {
-  return {
-    open: options => openStandaloneDraftRows(binding, namespace, options.sampling, options.checkpoints),
-    openPrefill: options => openStandaloneDraftRows(binding, namespace, null, options.checkpoints),
-  };
-}

package/src/spec/two-model.ts

@@ -1,113 +0,0 @@
-// Two-model speculative drafting — mlx-lm parity (`mlx_lm.server
-// --draft-model`). A full second model drafts autoregressively; the serve
-// loop (src/spec/serve-loop.ts) verifies with the target. L1 oracle:
-// mlx_lm.server with the same target/draft pair, greedy, token-for-token
-// (spec-vs-spec — both batch the verify lm-head, so neither is bit-exact to
-// stock decode at knife-edges; see src/spec/generate.ts header).
-//
-// Faithfulness notes (read from the oracle venv's generate.py):
-//  - drafts are sampled with the REQUEST sampler (generate.py:593-601) —
-//    greedy drafting under a temperature>0 request is not parity;
-//  - draft-cache rewind is max(n - kAccept - 1, 0) (generate.py:589-591),
-//    with the all-accept re-feed of the last draft handled by the serve
-//    loop's `feed` (generate.py:645-648);
-//  - logits processors do NOT run on draft steps (they're target-side).
-
-import { MlxArray } from "../mlx/array";
-import * as ops from "../mlx/ops";
-import { clearCache } from "../mlx/ffi";
-import { flagOn } from "../runtime-config";
-import { loadModelConfig } from "../config";
-import { Weights } from "../weights";
-import { createModel, type RuntimeModel } from "../model/factory";
-import type { DraftProvider, DraftSource, GroupedDraftProvider } from "./source";
-import { standaloneDraftGroups, openStandaloneDraftRows } from "./two-model-rows";
-import { bindLegacyAutoregressiveModel } from "../backends/mlx/autoregressive";
-import { artifactIdentity } from "../model/artifact-identity";
-import { configFingerprint } from "../model/fingerprint";
-
-/** mlx-lm's prefill_step_size — the draft drain chunks at the same stride
- *  as the target's (serve-loop.ts); chunking is numerically exact for
- *  causal attention. */
-const PREFILL_CHUNK = 2048;
-
-export class TwoModelProvider implements DraftProvider {
-  readonly id: string;
-  readonly weightsBytes: number;
-  readonly grouped: GroupedDraftProvider;
-
-  private constructor(
-    readonly model: RuntimeModel,
-    private readonly weights: Weights,
-    id: string,
-    private readonly namespace: string,
-  ) {
-    this.id = id;
-    this.grouped = standaloneDraftGroups(bindLegacyAutoregressiveModel(model), namespace);
-    this.weightsBytes = [...weights.shards.files.values()].reduce(
-      (a, f) => a + f.mmap.size,
-      0,
-    );
-  }
-
-  /** Load the draft model from a snapshot dir. Vocab-size mismatch with the
-   *  target is a WARNING (mlx-lm parity, server.py:363-368); the tokenizer-
-   *  family hard check lives at the server (it owns both tokenizers). */
-  static async load(modelDir: string, targetVocabSize?: number): Promise<TwoModelProvider> {
-    const config = await loadModelConfig(modelDir);
-    const weights = await Weights.open(modelDir);
-    try {
-      const model = createModel(weights, config);
-      if (
-        targetVocabSize !== undefined &&
-        config.text.vocabSize !== targetVocabSize
-      )
-        console.warn(
-          `draft model vocab size ${config.text.vocabSize} != target ${targetVocabSize} — ` +
-            `speculation will accept nothing useful (mlx-lm warns identically)`,
-        );
-      const identity = await artifactIdentity(configFingerprint(config),
-        [...weights.shards.files].map(([name, shard]) => ({ name, path: shard.path })));
-      return new TwoModelProvider(model, weights, modelDir.split("/").filter(Boolean).at(-1)!,
-        `standalone-draft-v1:${identity}`);
-    } catch (error) { weights.dispose(); throw error; }
-  }
-
-  open(opts: Parameters<DraftProvider["open"]>[0]): DraftSource {
-    return new TwoModelSource(this.model, opts.sampler, this.namespace);
-  }
-
-  dispose(): void {
-    this.weights.dispose();
-  }
-}
-
-/** Compatibility adapter: the existing request loop also uses the shared
- * draft graph at B=1. Proposal generation and state recovery have one owner. */
-class TwoModelSource implements DraftSource {
-  readonly weightsBytes = 0;
-  private readonly rows: ReturnType<typeof openStandaloneDraftRows>;
-  private processed = 0;
-  constructor(model: RuntimeModel, sampler: (lp: MlxArray, step: number) => MlxArray, namespace: string) {
-    this.rows = openStandaloneDraftRows(bindLegacyAutoregressiveModel(model), namespace,
-      { sample: (logprobs, steps) => sampler(logprobs, steps[0]!) }, [null]);
-  }
-  async prefill(promptIds: number[]): Promise<void> {
-    const tailSplit = flagOn("MLX_BUN_PREFILL_TAIL_SPLIT", true);
-    const upTo = tailSplit && promptIds.length > 1 ? promptIds.length - 1 : promptIds.length;
-    for (let at = this.processed; at < upTo; at += PREFILL_CHUNK) {
-      const end = Math.min(at + PREFILL_CHUNK, upTo);
-      using ids = ops.fromInt32(promptIds.slice(at, end), [1, end - at]);
-      await this.rows.prefill(ids); clearCache();
-    }
-    this.processed = upTo;
-  }
-  async draft(feed: number[], count: number, step: number): Promise<number[]> {
-    return (await this.rows.draft([feed.at(-1)!], count, [step]))[0]!;
-  }
-  async commit(_drafted: number, accepted: number): Promise<void> {
-    await this.rows.commit([accepted]);
-    this.processed += accepted + 1;
-  }
-  dispose(): void { this.rows.dispose(); }
-}