@mlx-node/lm 0.0.10 → 0.0.12

package/dist/models/paged-config-override.js

@@ -12,12 +12,12 @@
 import { mkdir, mkdtemp, readFile, readdir, rm, stat, symlink, writeFile } from 'node:fs/promises';
 import { tmpdir } from 'node:os';
 import { dirname, isAbsolute, join, relative, resolve, sep } from 'node:path';
-/** Every chat-capable family currently discovered by `@mlx-node/agent`. */
-export const AGENT_PAGED_MODEL_TYPES = ['qwen3', 'qwen3_5', 'qwen3_5_moe', 'gemma4', 'lfm2', 'lfm2_moe'];
+import { CHAT_FAMILY_IDS } from '../family-data.js';
 /** Families historically forced paged by `mlx launch claude`. */
 export const QWEN35_PAGED_MODEL_TYPES = ['qwen3_5', 'qwen3_5_moe'];
 const QWEN35_CACHE_FLOOR_MODEL_TYPES = new Set(QWEN35_PAGED_MODEL_TYPES);
 const DEFAULT_QWEN35_PAGED_CACHE_MB = 16_384;
+const DEFAULT_QWEN35_PAGED_CACHE_INITIAL_MB = 2_048;
 const QWEN35_MTP_DRAFTER_DIR = 'mtp-drafter';
 const QWEN35_DENSE_NESTED_MTP_SIDECAR = 'mtp/weights.safetensors';
 /**
@@ -38,7 +38,7 @@ export class PagedConfigOverrideManager {
     cleanupPromise;
     disposed = false;
     constructor(options = {}) {
-        this.modelTypes = new Set(options.modelTypes ?? AGENT_PAGED_MODEL_TYPES);
+        this.modelTypes = new Set(options.modelTypes ?? CHAT_FAMILY_IDS);
         this.tempDirPrefix = options.tempDirPrefix ?? 'mlx-paged-overrides-';
         this.preserveEmbeddedGemmaDraft = options.preserveEmbeddedGemmaDraft ?? false;
     }
@@ -108,12 +108,20 @@ export class PagedConfigOverrideManager {
         const pagedEnabled = config.use_block_paged_cache === true;
         const configuredMemoryMb = positiveNumber(config.paged_cache_memory_mb);
         const memorySatisfied = cacheFloorMb === undefined || (configuredMemoryMb ?? 0) >= cacheFloorMb;
+        // The clone writes `paged_cache_initial_memory_mb = min(initialMb, maxMb)`.
+        // A source whose field already equals that value needs no clone; an absent
+        // or different field does, or the new start-small default would silently
+        // skip every already-paged Qwen3.5 checkpoint (they lack the field).
+        const initialMb = QWEN35_CACHE_FLOOR_MODEL_TYPES.has(modelType) ? resolveQwen35InitialMb() : undefined;
+        const configuredInitialMb = positiveNumber(config.paged_cache_initial_memory_mb);
+        const initialSatisfied = initialMb === undefined ||
+            configuredInitialMb === Math.min(initialMb, Math.max(configuredMemoryMb ?? 0, cacheFloorMb ?? 0));
         // Even an already-paged Gemma config must be cloned when `draft/` exists:
         // returning the source would expose the draft to native auto-discovery and
         // trigger the flat-speculation/paged-cache conflict. An authoritative
         // persist directive that disagrees with the source likewise blocks the
         // pass-through so the resolved value reaches the loader.
-        if (pagedEnabled && memorySatisfied && !hasEmbeddedGemmaDraft && !persistOverrideNeeded) {
+        if (pagedEnabled && memorySatisfied && initialSatisfied && !hasEmbeddedGemmaDraft && !persistOverrideNeeded) {
             return modelPath;
         }
         // Memoize per (source, resolved family, persist directive): the same
@@ -125,7 +133,7 @@ export class PagedConfigOverrideManager {
         const existing = this.overrides.get(cacheKey);
         if (existing !== undefined)
             return existing;
-        const pending = this.createOverride(sourcePath, config, cacheFloorMb, modelType, persistPagedCache);
+        const pending = this.createOverride(sourcePath, config, cacheFloorMb, initialMb, modelType, persistPagedCache);
         this.overrides.set(cacheKey, pending);
         try {
             return await pending;
@@ -155,7 +163,7 @@ export class PagedConfigOverrideManager {
             await rm(root, { recursive: true, force: true }).catch(() => undefined);
         }
     }
-    async createOverride(sourcePath, sourceConfig, cacheFloorMb, modelType, persistPagedCache) {
+    async createOverride(sourcePath, sourceConfig, cacheFloorMb, initialMb, modelType, persistPagedCache) {
         const root = await this.getRoot();
         const overrideDir = await mkdtemp(join(root, 'model-'));
         const config = {
@@ -173,6 +181,14 @@ export class PagedConfigOverrideManager {
         if (cacheFloorMb !== undefined) {
             config.paged_cache_memory_mb = Math.max(positiveNumber(sourceConfig.paged_cache_memory_mb) ?? 0, cacheFloorMb);
         }
+        if (initialMb !== undefined) {
+            // Authoritative initial (grow-on-demand) budget, clamped to the max just
+            // written above so it can never exceed the pool ceiling. `initialMb` is
+            // always >= 1 and the max is floored by `cacheFloorMb` (>= 1) for the
+            // same families, so this never writes 0 — which the native loader rejects.
+            const maxMb = positiveNumber(config.paged_cache_memory_mb) ?? 0;
+            config.paged_cache_initial_memory_mb = Math.min(initialMb, maxMb);
+        }
         await writeFile(join(overrideDir, 'config.json'), JSON.stringify(config, null, 2), 'utf-8');
         const sourceEntries = await readdir(sourcePath);
         for (const name of sourceEntries) {
@@ -289,3 +305,10 @@ function resolveQwen35CacheFloorMb() {
     const parsed = Number.parseInt(raw, 10);
     return Number.isFinite(parsed) && parsed > 0 ? parsed : DEFAULT_QWEN35_PAGED_CACHE_MB;
 }
+function resolveQwen35InitialMb() {
+    const raw = process.env.MLX_PAGED_CACHE_INITIAL_MB;
+    if (raw == null || raw === '')
+        return DEFAULT_QWEN35_PAGED_CACHE_INITIAL_MB;
+    const parsed = Number.parseInt(raw, 10);
+    return Number.isFinite(parsed) && parsed > 0 ? parsed : DEFAULT_QWEN35_PAGED_CACHE_INITIAL_MB;
+}

package/dist/models/qwen3_5-configs.d.ts

@@ -4,10 +4,8 @@
  * Supports both dense and MoE variants. MoE fields are optional -
  * when `numExperts` is undefined, the model uses dense MLP layers.
  */
-import type { Qwen35Config as RustQwen35Config, Qwen35GenerationConfig as RustQwen35GenerationConfig, Qwen35GenerationResult as RustQwen35GenerationResult } from '@mlx-node/core';
+import type { Qwen35Config as RustQwen35Config } from '@mlx-node/core';
 export type Qwen35Config = RustQwen35Config;
-export type Qwen35GenerationConfig = RustQwen35GenerationConfig;
-export type Qwen35GenerationResult = RustQwen35GenerationResult;
 /**
  * Default configurations for common Qwen3.5 models
  */

package/dist/models/qwen3_5-configs.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"qwen3_5-configs.d.ts","sourceRoot":"","sources":["../../src/models/qwen3_5-configs.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EACV,YAAY,IAAI,gBAAgB,EAChC,sBAAsB,IAAI,0BAA0B,EACpD,sBAAsB,IAAI,0BAA0B,EACrD,MAAM,gBAAgB,CAAC;AAExB,MAAM,MAAM,YAAY,GAAG,gBAAgB,CAAC;AAC5C,MAAM,MAAM,sBAAsB,GAAG,0BAA0B,CAAC;AAChE,MAAM,MAAM,sBAAsB,GAAG,0BAA0B,CAAC;AAEhE;;GAEG;AACH,eAAO,MAAM,cAAc,EAAE;IAAE,CAAC,GAAG,EAAE,MAAM,GAAG,YAAY,CAAA;CA8BzD,CAAC;AAEF;;;;;;GAMG;AACH,wBAAgB,eAAe,CAAC,IAAI,EAAE,MAAM,GAAG,YAAY,CAM1D"}
+{"version":3,"file":"qwen3_5-configs.d.ts","sourceRoot":"","sources":["../../src/models/qwen3_5-configs.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,IAAI,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAEvE,MAAM,MAAM,YAAY,GAAG,gBAAgB,CAAC;AAE5C;;GAEG;AACH,eAAO,MAAM,cAAc,EAAE;IAAE,CAAC,GAAG,EAAE,MAAM,GAAG,YAAY,CAAA;CA8BzD,CAAC;AAEF;;;;;;GAMG;AACH,wBAAgB,eAAe,CAAC,IAAI,EAAE,MAAM,GAAG,YAAY,CAM1D"}

package/dist/stream.d.ts

@@ -1,4 +1,4 @@
-import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative } from "@mlx-node/core";
+import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, MuseGlimmerModel as MuseGlimmerModelNative, NemotronHModel as NemotronHModelNative, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative } from "@mlx-node/core";
 import type { ChatStreamChunk, ChatStreamHandle, PerformanceMetrics, ToolCallResult } from "@mlx-node/core";
 import type { SessionCapableModel } from "./chat-session.js";
 export interface ChatStreamDelta {
@@ -184,12 +184,14 @@ type ResolvedApplyTemplate<O extends StreamingModelOptions> = O extends {
 } ? O["applyTemplate"] : O["recordModelPath"];
 /** @internal Method names whose callback ABI is replaced by the generator wrapper. */
 export type NativeStreamingMethod = keyof NativeStreamingInstance;
-type StreamingReplacementMethod<O extends StreamingModelOptions> = NativeStreamingMethod | (ResolvedApplyTemplate<O> extends true ? "applyChatTemplate" : never);
+type NativeSessionReplacementMethod = "chatSessionStart" | "chatSessionContinue" | "chatSessionContinueTool" | "beginChatSessionStart" | "beginChatSessionContinue" | "beginChatSessionContinueTool";
+type StreamingReplacementMethod<O extends StreamingModelOptions> = NativeStreamingMethod | NativeSessionReplacementMethod | (ResolvedApplyTemplate<O> extends true ? "applyChatTemplate" : never);
 /**
  * Instance surface of a generated streaming wrapper. Only methods replaced at
  * runtime are removed from the native instance: the three callback streaming
- * methods, plus `applyChatTemplate` when the wrapper installs its path-backed
- * implementation. Intersecting the remaining native surface with
+ * methods, the internal operation methods, plus `applyChatTemplate` when the
+ * wrapper installs its path-backed implementation. Intersecting the remaining
+ * native surface with
  * `SessionCapableModel` preserves required native capabilities such as
  * `hasBlockPagedCache()` while exposing the generator streaming signatures.
  *
@@ -225,7 +227,7 @@ declare const Qwen35Model_base: {
     new (): StreamingInstance<typeof Qwen35ModelNative, {
         readonly recordModelPath: true;
     }>;
-    load(path: string): Promise<StreamingInstance<typeof Qwen35ModelNative, {
+    load(path: string, options?: import("@mlx-node/core").Qwen35LoadOptions | null | undefined): Promise<StreamingInstance<typeof Qwen35ModelNative, {
         readonly recordModelPath: true;
     }>>;
 };
@@ -263,6 +265,17 @@ declare const Lfm2Model_base: {
 /** LFM2 model (text-only) — see {@link Qwen35Model} for the wrapper shape. */
 export declare class Lfm2Model extends Lfm2Model_base {
 }
+declare const NemotronHModel_base: {
+    new (): StreamingInstance<typeof NemotronHModelNative, {
+        readonly recordModelPath: true;
+    }>;
+    load(modelPath: string): Promise<StreamingInstance<typeof NemotronHModelNative, {
+        readonly recordModelPath: true;
+    }>>;
+};
+/** Nemotron 3.5 Lightning (text-only) — see {@link Qwen35Model} for the wrapper shape. */
+export declare class NemotronHModel extends NemotronHModel_base {
+}
 declare const Gemma4Model_base: {
     new (config: import("@mlx-node/core").Gemma4Config): StreamingInstance<typeof Gemma4ModelNative, {
         readonly recordModelPath: true;
@@ -274,6 +287,17 @@ declare const Gemma4Model_base: {
 /** Gemma4 model (text-only) — see {@link Qwen35Model} for the wrapper shape. */
 export declare class Gemma4Model extends Gemma4Model_base {
 }
+declare const MuseGlimmerModel_base: {
+    new (): StreamingInstance<typeof MuseGlimmerModelNative, {
+        readonly recordModelPath: true;
+    }>;
+    load(modelPath: string): Promise<StreamingInstance<typeof MuseGlimmerModelNative, {
+        readonly recordModelPath: true;
+    }>>;
+};
+/** Muse-Glimmer text model with embedded DFlash speculative decoding. */
+export declare class MuseGlimmerModel extends MuseGlimmerModel_base {
+}
 declare const Qwen3Model_base: {
     new (): StreamingInstance<typeof Qwen3ModelNative, {
         readonly recordModelPath: true;

package/dist/stream.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,SAAS,IAAI,eAAe,EAE5B,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AACxB,OAAO,KAAK,EAGV,eAAe,EACf,gBAAgB,EAChB,kBAAkB,EAElB,cAAc,EACf,MAAM,gBAAgB,CAAC;AAExB,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAE7D,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,KAAK,CAAC;IACZ,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,IAAI,CAAC;IACX,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,cAAc,EAAE,CAAC;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,2EAA2E;IAC3E,eAAe,EAAE,OAAO,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB;;;;OAIG;IACH,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B;;;;;;;;;;;;;;;;;;;;;OAqBG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,kBAAkB,CAAC;CAClC;AAED,MAAM,MAAM,eAAe,GAAG,eAAe,GAAG,eAAe,CAAC;AAKhE,KAAK,oBAAoB,GAAG,eAAe,GAAG,gBAAgB,CAAC;AAE/D,UAAU,qBAAqB;IAC7B,KAAK,EAAE,oBAAoB,CAAC;IAC5B;;;OAGG;IACH,wBAAwB,CAAC,EAAE,MAAM,CAAC;CACnC;AA6DD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AACH,wBAAuB,cAAc,CACnC,SAAS,EAAE,CACT,QAAQ,EAAE,CAAC,GAAG,EAAE,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,eAAe,KAAK,IAAI,KAC1D,OAAO,CAAC,gBAAgB,CAAC,EAC9B,MAAM,CAAC,EAAE,WAAW,GACnB,cAAc,CAAC,eAAe,CAAC,CA+JjC;AAcD;;;;;;;;;;GAUG;AACH,MAAM,WAAW,uBAAuB;IACtC,sBAAsB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACxE,yBAAyB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAC3E,6BAA6B,EAAE,CAC7B,GAAG,IAAI,EAAE,KAAK,EAAE,KACb,OAAO,CAAC,gBAAgB,CAAC,CAAC;CAChC;AAED;;;;;;;;;;GAUG;AACH,UAAU,mBAAmB;IAK3B,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,GAAG,uBAAuB,CAAC;IAKhD,IAAI,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IACzC,SAAS,EAAE,uBAAuB,CAAC;CACpC;AAED,mDAAmD;AACnD,UAAU,qBAAqB;IAC7B;;;;;;OAMG;IACH,eAAe,EAAE,OAAO,CAAC;IACzB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB;;;;OAIG;IACH,qBAAqB,CAAC,EAAE,qBAAqB,CAAC;IAC9C;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAC;CAClC;AAED;;;;;GAKG;AACH,MAAM,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAEjD;;;;;;GAMG;AACH,KAAK,qBAAqB,CAAC,CAAC,SAAS,qBAAqB,IAAI,CAAC,SAAS;IACtE,aAAa,EAAE,OAAO,CAAC;CACxB,GACG,CAAC,CAAC,eAAe,CAAC,GAClB,CAAC,CAAC,iBAAiB,CAAC,CAAC;AAEzB,sFAAsF;AACtF,MAAM,MAAM,qBAAqB,GAAG,MAAM,uBAAuB,CAAC;AAElE,KAAK,0BAA0B,CAAC,CAAC,SAAS,qBAAqB,IAC3D,qBAAqB,GACrB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAAG,mBAAmB,GAAG,KAAK,CAAC,CAAC;AAE1E;;;;;;;;;;;;;GAaG;AACH,MAAM,MAAM,iBAAiB,CAC3B,CAAC,SAAS,mBAAmB,EAC7B,CAAC,SAAS,qBAAqB,IAC7B,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE,0BAA0B,CAAC,CAAC,CAAC,CAAC,GACtD,mBAAmB,GACnB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAClC,QAAQ,CAAC,IAAI,CAAC,mBAAmB,EAAE,mBAAmB,CAAC,CAAC,GACxD,MAAM,CAAC,CAAC;AAEd;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,kBAAkB,CAChC,CAAC,SAAS,mBAAmB,EAC7B,KAAK,CAAC,CAAC,SAAS,qBAAqB,EAErC,WAAW,EAAE,CAAC,EACd,IAAI,EAAE,CAAC,GACN;IAcD,KAAK,GAAG,IAAI,EAAE,qBAAqB,CAAC,CAAC,CAAC,GAAG,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACjE,IAAI,CAAC,GAAG,IAAI,EAAE,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;CACxE,CAqIA;;;;;;;;;AAED;;;;;;;GAOG;AACH,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;;;AAEL,8EAA8E;AAC9E,qBAAa,SAAU,SAAQ,cAG7B;CAAG;;;;;;;;;AAEL,gFAAgF;AAChF,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;;;AAEL;;;;;;GAMG;AACH,qBAAa,UAAW,SAAQ,eAG9B;CAAG"}
+{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAEA,OAAO,EACL,WAAW,IAAI,iBAAiB,EAChC,SAAS,IAAI,eAAe,EAC5B,gBAAgB,IAAI,sBAAsB,EAC1C,cAAc,IAAI,oBAAoB,EAEtC,UAAU,IAAI,gBAAgB,EAC9B,WAAW,IAAI,iBAAiB,EAChC,cAAc,IAAI,oBAAoB,EACvC,MAAM,gBAAgB,CAAC;AACxB,OAAO,KAAK,EAIV,eAAe,EACf,gBAAgB,EAChB,kBAAkB,EAElB,cAAc,EACf,MAAM,gBAAgB,CAAC;AAExB,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mBAAmB,CAAC;AAkC7D,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,KAAK,CAAC;IACZ,WAAW,CAAC,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,IAAI,CAAC;IACX,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,cAAc,EAAE,CAAC;IAC5B,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAC;IACxB,2EAA2E;IAC3E,eAAe,EAAE,OAAO,CAAC;IACzB,SAAS,EAAE,MAAM,CAAC;IAClB,YAAY,EAAE,MAAM,CAAC;IACrB,eAAe,EAAE,MAAM,CAAC;IACxB,OAAO,EAAE,MAAM,CAAC;IAChB;;;;OAIG;IACH,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,iBAAiB,CAAC,EAAE,OAAO,CAAC;IAC5B;;;;;;;;;;;;;;;;;;;;;OAqBG;IACH,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,WAAW,CAAC,EAAE,kBAAkB,CAAC;CAClC;AAED,MAAM,MAAM,eAAe,GAAG,eAAe,GAAG,eAAe,CAAC;AAKhE,KAAK,oBAAoB,GAAG,eAAe,GAAG,gBAAgB,CAAC;AAE/D,UAAU,qBAAqB;IAC7B,KAAK,EAAE,oBAAoB,CAAC;IAC5B;;;OAGG;IACH,wBAAwB,CAAC,EAAE,MAAM,CAAC;CACnC;AAyDD;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AACH,wBAAuB,cAAc,CACnC,SAAS,EAAE,CACT,QAAQ,EAAE,CAAC,GAAG,EAAE,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,eAAe,KAAK,IAAI,KAC1D,OAAO,CAAC,gBAAgB,CAAC,EAC9B,MAAM,CAAC,EAAE,WAAW,GACnB,cAAc,CAAC,eAAe,CAAC,CAsLjC;AAwCD;;;;;;;;;;GAUG;AACH,MAAM,WAAW,uBAAuB;IACtC,sBAAsB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACxE,yBAAyB,EAAE,CAAC,GAAG,IAAI,EAAE,KAAK,EAAE,KAAK,OAAO,CAAC,gBAAgB,CAAC,CAAC;IAC3E,6BAA6B,EAAE,CAC7B,GAAG,IAAI,EAAE,KAAK,EAAE,KACb,OAAO,CAAC,gBAAgB,CAAC,CAAC;CAChC;AAED;;;;;;;;;;GAUG;AACH,UAAU,mBAAmB;IAK3B,KAAK,GAAG,IAAI,EAAE,KAAK,EAAE,GAAG,uBAAuB,CAAC;IAKhD,IAAI,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC;IACzC,SAAS,EAAE,uBAAuB,CAAC;CACpC;AAED,mDAAmD;AACnD,UAAU,qBAAqB;IAC7B;;;;;;OAMG;IACH,eAAe,EAAE,OAAO,CAAC;IACzB;;;;;;OAMG;IACH,aAAa,CAAC,EAAE,OAAO,CAAC;IACxB;;;;OAIG;IACH,qBAAqB,CAAC,EAAE,qBAAqB,CAAC;IAC9C;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAC;CAClC;AAED;;;;;GAKG;AACH,MAAM,MAAM,cAAc,GAAG,mBAAmB,CAAC;AAEjD;;;;;;GAMG;AACH,KAAK,qBAAqB,CAAC,CAAC,SAAS,qBAAqB,IAAI,CAAC,SAAS;IACtE,aAAa,EAAE,OAAO,CAAC;CACxB,GACG,CAAC,CAAC,eAAe,CAAC,GAClB,CAAC,CAAC,iBAAiB,CAAC,CAAC;AAEzB,sFAAsF;AACtF,MAAM,MAAM,qBAAqB,GAAG,MAAM,uBAAuB,CAAC;AAElE,KAAK,8BAA8B,GAC/B,kBAAkB,GAClB,qBAAqB,GACrB,yBAAyB,GACzB,uBAAuB,GACvB,0BAA0B,GAC1B,8BAA8B,CAAC;AAEnC,KAAK,0BAA0B,CAAC,CAAC,SAAS,qBAAqB,IAC3D,qBAAqB,GACrB,8BAA8B,GAC9B,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAAG,mBAAmB,GAAG,KAAK,CAAC,CAAC;AAE1E;;;;;;;;;;;;;;GAcG;AACH,MAAM,MAAM,iBAAiB,CAC3B,CAAC,SAAS,mBAAmB,EAC7B,CAAC,SAAS,qBAAqB,IAC7B,IAAI,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE,0BAA0B,CAAC,CAAC,CAAC,CAAC,GACtD,mBAAmB,GACnB,CAAC,qBAAqB,CAAC,CAAC,CAAC,SAAS,IAAI,GAClC,QAAQ,CAAC,IAAI,CAAC,mBAAmB,EAAE,mBAAmB,CAAC,CAAC,GACxD,MAAM,CAAC,CAAC;AAEd;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,kBAAkB,CAChC,CAAC,SAAS,mBAAmB,EAC7B,KAAK,CAAC,CAAC,SAAS,qBAAqB,EAErC,WAAW,EAAE,CAAC,EACd,IAAI,EAAE,CAAC,GACN;IAaD,KAAK,GAAG,IAAI,EAAE,qBAAqB,CAAC,CAAC,CAAC,GAAG,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACjE,IAAI,CAAC,GAAG,IAAI,EAAE,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,GAAG,OAAO,CAAC,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;CACxE,CAwMA;;;;;;;;;AAED;;;;;;;GAOG;AACH,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;;;AAEL,8EAA8E;AAC9E,qBAAa,SAAU,SAAQ,cAG7B;CAAG;;;;;;;;;AAEL,0FAA0F;AAC1F,qBAAa,cAAe,SAAQ,mBAElC;CAAG;;;;;;;;;AAEL,gFAAgF;AAChF,qBAAa,WAAY,SAAQ,gBAE/B;CAAG;;;;;;;;;AAEL,yEAAyE;AACzE,qBAAa,gBAAiB,SAAQ,qBAEpC;CAAG;;;;;;;;;;;AAEL;;;;;;GAMG;AACH,qBAAa,UAAW,SAAQ,eAG9B;CAAG"}

package/dist/stream.js

@@ -1,13 +1,10 @@
 import { join } from "node:path";
-import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, Qwen3Tokenizer, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative, } from "@mlx-node/core";
+import { Gemma4Model as Gemma4ModelNative, Lfm2Model as Lfm2ModelNative, MuseGlimmerModel as MuseGlimmerModelNative, NemotronHModel as NemotronHModelNative, Qwen3Tokenizer, Qwen3Model as Qwen3ModelNative, Qwen35Model as Qwen35ModelNative, Qwen35MoeModel as Qwen35MoeModelNative, } from "@mlx-node/core";
 const modelPathsForTokenizers = new WeakMap();
 const tokenizerPromises = new WeakMap();
 function getNativeIsReasoning(chunk) {
     return typeof chunk.isReasoning === "boolean" ? chunk.isReasoning : undefined;
 }
-function rememberModelPath(model, modelPath) {
-    modelPathsForTokenizers.set(model, modelPath);
-}
 async function applyChatTemplateFromModelPath(model, messages, addGenerationPrompt, tools, enableThinking, contentPolicy) {
     const modelPath = modelPathsForTokenizers.get(model);
     if (modelPath == null) {
@@ -62,8 +59,31 @@ async function applyChatTemplateFromModelPath(model, messages, addGenerationProm
  * public API — may change without notice.
  */
 export async function* _runChatStream(startCall, signal) {
+    // The native ThreadsafeFunction uses the same fixed ceiling. This JS-side
+    // guard also protects handwritten/test adapters that bypass the Rust glue.
+    const maxBufferedEvents = 64;
     const queue = [];
     let resolve = null;
+    let handle = null;
+    let cancelRequested = false;
+    let cancelled = false;
+    let overflowed = false;
+    const cancelOnce = () => {
+        if (cancelled)
+            return;
+        if (handle === null) {
+            cancelRequested = true;
+            return;
+        }
+        cancelled = true;
+        try {
+            handle.cancel();
+        }
+        catch {
+            // Cancellation is best-effort. The distinguished backlog error below
+            // remains authoritative for the consumer even if a backend throws.
+        }
+    };
     const waitForItem = () => queue.length > 0
         ? Promise.resolve()
         : new Promise((r) => {
@@ -77,31 +97,31 @@ export async function* _runChatStream(startCall, signal) {
         }
     };
     const callback = (err, chunk) => {
+        if (overflowed)
+            return;
+        if (queue.length >= maxBufferedEvents) {
+            overflowed = true;
+            cancelOnce();
+            // Keep the queue at its fixed ceiling while guaranteeing the consumer
+            // eventually observes why the stream was cancelled.
+            queue[queue.length - 1] = {
+                error: new Error(`Native chat stream backlog exceeded ${maxBufferedEvents} buffered events`),
+            };
+            notify();
+            return;
+        }
         queue.push(err ? { error: err } : { chunk });
         notify();
     };
-    const handle = await startCall(callback);
+    handle = await startCall(callback);
+    if (cancelRequested)
+        cancelOnce();
     // Guard against double-cancel. Some native backends throw on a
     // second `cancel()`; we route every cancel site through this
     // helper so the abort path (via `triggerAbort`) and the unwind
     // path (via the `finally` block) don't cancel twice and so any
     // backend that does throw is swallowed rather than escaping as
     // an error out of an otherwise-clean early termination.
-    let cancelled = false;
-    const cancelOnce = () => {
-        if (cancelled)
-            return;
-        cancelled = true;
-        try {
-            handle.cancel();
-        }
-        catch {
-            // Native backend threw on cancel — nothing actionable here.
-            // Swallow so aborted streams still surface as a clean early
-            // termination via the synthetic `aborted` marker rather than
-            // as an unexpected error out of the generator.
-        }
-    };
     // Signal-driven fast-abort. If the signal is already aborted at
     // attach time we still arm the listener so the synchronous abort
     // dispatch path runs below (calling `handle.cancel()` after the
@@ -211,6 +231,32 @@ export async function* _runChatStream(startCall, signal) {
         cancelOnce();
     }
 }
+/**
+ * Translate the public AbortSignal API to the native two-phase cancellation
+ * operation. Callers receive one ordinary Promise and use the same
+ * AbortController as fetch and the streaming APIs.
+ */
+async function runChatSessionCall(startCall, signal) {
+    if (signal.aborted)
+        throw new Error("chat session cancelled");
+    const call = await startCall();
+    let cancelled = false;
+    const cancelOnce = () => {
+        if (cancelled)
+            return;
+        cancelled = true;
+        call.cancel();
+    };
+    signal.addEventListener("abort", cancelOnce, { once: true });
+    if (signal.aborted)
+        cancelOnce();
+    try {
+        return await call.result();
+    }
+    finally {
+        signal.removeEventListener("abort", cancelOnce);
+    }
+}
 /**
  * Build the streaming-model subclass for a native chat model class.
  *
@@ -238,6 +284,7 @@ export function makeStreamingModel(NativeClass, opts) {
     const nativeStart = NativeClass.prototype.chatStreamSessionStart;
     const nativeContinue = NativeClass.prototype.chatStreamSessionContinue;
     const nativeContinueTool = NativeClass.prototype.chatStreamSessionContinueTool;
+    const nativeChat = NativeClass.prototype;
     // `NativeClass` is structurally a constructor; cast to a concrete
     // constructor type so `class extends` accepts it. Runtime behavior is
     // unchanged — we extend the real native class.
@@ -251,7 +298,7 @@ export function makeStreamingModel(NativeClass, opts) {
         }
         static async load(modelPath, ...rest) {
             // Forward any trailing family-specific load options verbatim (e.g.
-            // Gemma4's `Gemma4LoadOptions` with `draftModelPath`); families whose
+            // Gemma4/Qwen3.5 `draftModelPath`); families whose
             // native `load` takes only the path receive no extras. The public
             // signature is re-narrowed per family via `Parameters<C['load']>` in
             // the factory return type below.
@@ -260,10 +307,48 @@ export function makeStreamingModel(NativeClass, opts) {
             // concrete subclass declared per family supplies the prototype and
             // `instanceof ConcreteSubclass` holds.
             Object.setPrototypeOf(instance, this.prototype);
-            if (recordPath)
-                rememberModelPath(instance, modelPath);
+            if (recordPath) {
+                const resolvedAssetsPath = instance.modelAssetsPath?.();
+                modelPathsForTokenizers.set(instance, resolvedAssetsPath ?? modelPath);
+            }
             return instance;
         }
+        async chatSessionStart(messages, config, signal) {
+            if (signal == null) {
+                if (nativeChat.chatSessionStart == null) {
+                    throw new Error("Native model does not implement chatSessionStart");
+                }
+                return await nativeChat.chatSessionStart.call(this, messages, config);
+            }
+            if (nativeChat.beginChatSessionStart == null) {
+                throw new Error("Native model does not implement beginChatSessionStart");
+            }
+            return await runChatSessionCall(() => nativeChat.beginChatSessionStart.call(this, messages, config), signal);
+        }
+        async chatSessionContinue(messages, config, signal) {
+            if (signal == null) {
+                if (nativeChat.chatSessionContinue == null) {
+                    throw new Error("Native model does not implement chatSessionContinue");
+                }
+                return await nativeChat.chatSessionContinue.call(this, messages, config);
+            }
+            if (nativeChat.beginChatSessionContinue == null) {
+                throw new Error("Native model does not implement beginChatSessionContinue");
+            }
+            return await runChatSessionCall(() => nativeChat.beginChatSessionContinue.call(this, messages, config), signal);
+        }
+        async chatSessionContinueTool(messages, config, signal) {
+            if (signal == null) {
+                if (nativeChat.chatSessionContinueTool == null) {
+                    throw new Error("Native model does not implement chatSessionContinueTool");
+                }
+                return await nativeChat.chatSessionContinueTool.call(this, messages, config);
+            }
+            if (nativeChat.beginChatSessionContinueTool == null) {
+                throw new Error("Native model does not implement beginChatSessionContinueTool");
+            }
+            return await runChatSessionCall(() => nativeChat.beginChatSessionContinueTool.call(this, messages, config), signal);
+        }
         // The native methods are callback-based, but `Base` is typed as a
         // `SessionCapableModel` constructor (whose streaming methods already
         // return `AsyncGenerator<ChatStreamEvent>`), so these overrides are
@@ -313,11 +398,21 @@ export class Lfm2Model extends makeStreamingModel(Lfm2ModelNative, {
     replayAssistantRawText: true,
 }) {
 }
+/** Nemotron 3.5 Lightning (text-only) — see {@link Qwen35Model} for the wrapper shape. */
+export class NemotronHModel extends makeStreamingModel(NemotronHModelNative, {
+    recordModelPath: true,
+}) {
+}
 /** Gemma4 model (text-only) — see {@link Qwen35Model} for the wrapper shape. */
 export class Gemma4Model extends makeStreamingModel(Gemma4ModelNative, {
     recordModelPath: true,
 }) {
 }
+/** Muse-Glimmer text model with embedded DFlash speculative decoding. */
+export class MuseGlimmerModel extends makeStreamingModel(MuseGlimmerModelNative, {
+    recordModelPath: true,
+}) {
+}
 /**
  * Qwen3 (first-gen, text-only) model.
  *
@@ -344,12 +439,16 @@ function _assertSessionCapable() {
     const _moe = null;
     const _lfm2 = null;
     const _gemma4 = null;
+    const _museGlimmer = null;
     const _qwen3 = null;
+    const _nemotronH = null;
     void _qwen35;
     void _moe;
     void _lfm2;
     void _gemma4;
+    void _museGlimmer;
     void _qwen3;
+    void _nemotronH;
 }
 void _assertSessionCapable;
 /** Compile-time guard that both Qwen3.5 native classes and wrappers retain the exact media planner. */
@@ -364,17 +463,21 @@ function _assertExpandedPromptPlannerSurfaces() {
     void _wrappedMoe;
 }
 void _assertExpandedPromptPlannerSurfaces;
-/** Compile-time guard that the factory preserves every non-streaming native member. */
+/** Compile-time guard for every native member the factory does not replace. */
 function _assertPreservedNativeSurfaces() {
     const _qwen3 = null;
     const _qwen35 = null;
     const _moe = null;
     const _lfm2 = null;
     const _gemma4 = null;
+    const _museGlimmer = null;
+    const _nemotronH = null;
     void _qwen3;
     void _qwen35;
     void _moe;
     void _lfm2;
     void _gemma4;
+    void _museGlimmer;
+    void _nemotronH;
 }
 void _assertPreservedNativeSurfaces;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mlx-node/lm",
-  "version": "0.0.10",
+  "version": "0.0.12",
   "homepage": "https://github.com/mlx-node/mlx-node",
   "bugs": {
     "url": "https://github.com/mlx-node/mlx-node/issues"
@@ -21,6 +21,10 @@
     ".": {
       "types": "./dist/index.d.ts",
       "import": "./dist/index.js"
+    },
+    "./family-data": {
+      "types": "./dist/family-data.d.ts",
+      "import": "./dist/family-data.js"
     }
   },
   "scripts": {
@@ -28,9 +32,9 @@
     "test": "vite test run"
   },
   "dependencies": {
-    "@mlx-node/core": "0.0.10"
+    "@mlx-node/core": "0.0.12"
   },
   "devDependencies": {
-    "@types/node": "^26.0.0"
+    "@types/node": "^26.4.0"
   }
 }

package/dist/interfaces.d.ts

@@ -1,3 +0,0 @@
-/** Compatibility type barrel; model unions are derived from the loader registry. */
-export type { LoadableModel, TrainableModel } from './models/model-loader.js';
-//# sourceMappingURL=interfaces.d.ts.map

package/dist/interfaces.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"interfaces.d.ts","sourceRoot":"","sources":["../src/interfaces.ts"],"names":[],"mappings":"AAAA,oFAAoF;AACpF,YAAY,EAAE,aAAa,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAC"}

package/dist/interfaces.js

@@ -1 +0,0 @@
-export {};