Add eval, async_eval, and IO ops (load/save/safetensors/gguf)

node/scripts/mlx_sources.gypi

@@ -99,8 +99,11 @@
       'vendor/mlx/fft.cpp',
       'vendor/mlx/graph_utils.cpp',
       'vendor/mlx/io/load.cpp',
-      'vendor/mlx/io/no_gguf.cpp',
-      'vendor/mlx/io/no_safetensors.cpp',
+      'vendor/mlx/io/safetensors.cpp',
+      'vendor/mlx/io/gguf.cpp',
+      'vendor/mlx/io/gguf_quants.cpp',
+      'vendor/fp16.c',
+      'vendor/gguflib.c',
       'vendor/mlx/linalg.cpp',
       'vendor/mlx/ops.cpp',
       'vendor/mlx/primitives.cpp',

node/src/core/index.ts

@@ -283,6 +283,15 @@ import {
   broadcast_shapes,
   convolve,
   einsum_path,
+  // Batch 5: eval + IO
+  eval_op,
+  async_eval,
+  load,
+  save,
+  save_safetensors,
+  save_gguf,
+  type SaveSafetensorsOptions,
+  type LoadResult,
   type NanToNumOptions,
   type CloseOptions,
   type ContiguousOptions,
@@ -522,6 +531,12 @@ export {
   broadcast_shapes,
   convolve,
   einsum_path,
+  eval_op,
+  async_eval,
+  load,
+  save,
+  save_safetensors,
+  save_gguf,
 };
 export type {
   SplitOptions, PadOptions, SliceOptions, AsStridedOptions, NumberOfElementsOptions,
@@ -533,6 +548,7 @@ export type {
   ConvGeneralOptions, ConvTranspose1dOptions, ConvTranspose2dOptions, ConvTranspose3dOptions,
   BlockMaskedMMOptions, GatherMMOptions, QuantizeOptions, DequantizeOptions,
   QuantizedMatmulOptions, GatherQMMOptions,
+  SaveSafetensorsOptions, LoadResult,
 };
 export {
   default_device,
@@ -785,6 +801,12 @@ const core = {
   broadcast_shapes,
   convolve,
   einsum_path,
+  eval: eval_op,
+  async_eval,
+  load,
+  save,
+  save_safetensors,
+  save_gguf,
   device: deviceModule,
   Dtype,
   dtype: dtypeModule,

node/src/core/ops.ts

@@ -2993,3 +2993,100 @@ export function einsum_path(subscripts: string, ...operands: MLXArray[]): [strin
   return [path, subscripts];
 }
 
+// ============================================================
+// Eval ops
+// ============================================================
+
+/** Evaluate one or more arrays, forcing lazy computation */
+export function eval_op(...args: (MLXArray | MLXArray[])[]): void {
+  const natives: any[] = [];
+  for (const arg of args) {
+    if (Array.isArray(arg)) {
+      for (const a of arg) natives.push(toNativeHandle(a));
+    } else {
+      natives.push(toNativeHandle(arg));
+    }
+  }
+  addon.eval(...natives);
+}
+
+/** Asynchronously evaluate one or more arrays */
+export function async_eval(...args: (MLXArray | MLXArray[])[]): void {
+  const natives: any[] = [];
+  for (const arg of args) {
+    if (Array.isArray(arg)) {
+      for (const a of arg) natives.push(toNativeHandle(a));
+    } else {
+      natives.push(toNativeHandle(arg));
+    }
+  }
+  addon.async_eval(...natives);
+}
+
+// ============================================================
+// IO ops
+// ============================================================
+
+export interface SaveSafetensorsOptions {
+  metadata?: Record<string, string>;
+}
+
+export interface LoadResult {
+  arrays: Record<string, MLXArray>;
+  metadata: Record<string, any>;
+}
+
+/** Load an array or dict of arrays from a file (.npy, .npz, .safetensors, .gguf) */
+export function load(file: string, options?: StreamOptions): MLXArray | LoadResult {
+  const args: any[] = [file];
+  if (options?.stream) args.push(normalizeStream(options.stream));
+  const result = addon.load(...args);
+  // Single array (npy format)
+  if (result && typeof result === 'object' && result.constructor && result.constructor.name !== 'Object') {
+    return MLXArray.fromHandle(result);
+  }
+  // Dict format (safetensors, gguf) — result is {arrays: {}, metadata: {}}
+  if (result && result.arrays) {
+    const arrays: Record<string, MLXArray> = {};
+    for (const [key, val] of Object.entries(result.arrays)) {
+      arrays[key] = MLXArray.fromHandle(val as any);
+    }
+    return { arrays, metadata: result.metadata || {} };
+  }
+  return MLXArray.fromHandle(result);
+}
+
+/** Save a single array to a .npy file */
+export function save(file: string, arr: MLXArray): void {
+  addon.save(file, toNativeHandle(arr));
+}
+
+/** Save arrays to a .safetensors file */
+export function save_safetensors(
+  file: string,
+  arrays: Record<string, MLXArray>,
+  options?: SaveSafetensorsOptions,
+): void {
+  const nativeArrays: Record<string, any> = {};
+  for (const [key, val] of Object.entries(arrays)) {
+    nativeArrays[key] = toNativeHandle(val);
+  }
+  const args: any[] = [file, nativeArrays];
+  if (options?.metadata) {
+    args.push(options.metadata);
+  }
+  addon.save_safetensors(...args);
+}
+
+/** Save arrays to a .gguf file */
+export function save_gguf(
+  file: string,
+  arrays: Record<string, MLXArray>,
+): void {
+  const nativeArrays: Record<string, any> = {};
+  for (const [key, val] of Object.entries(arrays)) {
+    nativeArrays[key] = toNativeHandle(val);
+  }
+  addon.save_gguf(file, nativeArrays);
+}
+

node/src/index.ts

@@ -226,6 +226,13 @@ export const {
   broadcast_shapes,
   convolve,
   einsum_path,
+  // Batch 5: eval + IO
+  eval: eval_op,
+  async_eval,
+  load,
+  save,
+  save_safetensors,
+  save_gguf,
   // DType constants
   bool,
   int8,