# 接口：QuantizationConfig

## 层级

- **`QuantizationConfig`**

  ↳ [`MlxQuantizationConfig`](MlxQuantizationConfig)

## 属性

### 位

• `Optional` **bits**：`number`

#### 定义于[[bits.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:930](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L930)

___

### config\_groups

• `Optional` **config\_groups**: `Record`\<`string`, \{ `format?`: `string` ; `targets?`: `string`[] ; `weights?`: \{ `num_bits?`: `number`  }  }\>

#### 定义于[[configgroups.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:935](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L935)

___

### expert\_dtype

• `Optional` **expert\_dtype**：`string`

当路由专家精度与主量化器不同时（例如，FP4 专家搭配 FP8 注意力）。

#### 定义于[[expertdtype.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:920](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L920)

___

### 格式

• `Optional` **格式**：`string`

#### 定义于[[format.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:934](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L934)

___

### group\_size

• `Optional` **group\_size**：`number`

#### 定义于[[groupsize.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:928](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L928)

___

### ignore

• `Optional` **ignore**：`string`[]

compressed-tensors 将其排除列表命名为 `ignore`，而非 `modules_to_not_convert`，并使用与 `config_groups[].targets` 相同的 `re:` 前缀目标语法。

#### 定义于[[ignore.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:940](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L940)

___

### load\_in\_4bit

• `Optional` **load\_in\_4bit**：`boolean`

#### 定义于[[loadin4bit.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:931](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L931)

___

### load\_in\_8bit

• `Optional` **load\_in\_8bit**：`boolean`

#### 定义于[[loadin8bit.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:932](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L932)

___

### mode

• `Optional` **mode**：`string`

MLX 量化模式（例如 `affine`）；MLX 配置不声明 `quant_method`。

#### 定义于[[mode.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:927](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L927)

___

### modules\_to\_not\_convert

• `Optional` **modules\_to\_not\_convert**： `string`[]

#### 定义于[[modulestonotconvert.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:929](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L929)

___

### quant\_algo

• `Optional` **量化\_算法**: `string`

NVIDIA ModelOpt（`quant_method: "modelopt"`）算法，例如 `"NVFP4"`、`"W4A16_NVFP4"`、`"FP8"` 或 `"MIXED_PRECISION"`——此时每个模块的算法位于 `quantized_layers` 中。

#### 定义于[[quantalgo.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:945](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L945)

___

### quant\_method

• `Optional` **量化\_方法**: `string`

#### 定义于[[quantmethod.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:918](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L918)

___

### quantized\_layers

• `Optional` **量化\_层**：`Record`\<`string`, \{ `quant_algo?`: `string`  }\>

ModelOpt 混合精度：模块名称（或父模块，例如 `…mlp.experts`）-> 其算法。

#### 定义于[[quantizedlayers.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:947](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L947)

___

### store\_dtype

• `Optional` **store\_dtype**：`string`

与`expert_dtype`的作用相同，只是名称不同：MiMo-V2.6 的稠密层使用`quant_method: "fp8"`，但路由专家以`store_dtype: "mxfp4"`存储，每个`U8`打包两个。

#### 定义于[[storedtype.defined-in]]

[packages/hub/src/lib/parse-safetensors-metadata.ts:925](https://github.com/huggingface/huggingface.js/blob/main/packages/hub/src/lib/parse-safetensors-metadata.ts#L925)