Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / transformers /main /ko /main_classes /quantization.md
HuggingFaceDocBuilder
32 minutes ago
preview code
download
raw
39.2 kB
# 양자화[[quantization]]

양자화 기법은 가중치와 활성화를 8비트 정수(int8)와 같은 더 낮은 정밀도의 데이터 타입으로 표현함으로써 메모리와 계산 비용을 줄입니다. 이를 통해 일반적으로는 메모리에 올릴 수 없는 더 큰 모델을 로드할 수 있고, 추론 속도를 높일 수 있습니다. Transformers는 AWQ와 GPTQ 양자화 알고리즘을 지원하며, bitsandbytes를 통해 8비트와 4비트 양자화를 지원합니다.
Transformers에서 지원되지 않는 양자화 기법들은 `HfQuantizer` 클래스를 통해 추가될 수 있습니다.

모델을 양자화하는 방법은 이 [양자화](../quantization) 가이드를 통해 배울 수 있습니다.

## QuantoConfig[[transformers.QuantoConfig]][[transformers.QuantoConfig]]

#### transformers.QuantoConfig[[transformers.QuantoConfig]]

```python
transformers.QuantoConfig(weights = 'int8', activations = None, modules_to_not_convert: list | None = None, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1021)

**Parameters:**

weights (`str`, *optional*, defaults to `"int8"`) : The target dtype for the weights after quantization. Supported values are ("float8","int8","int4","int2")

activations (`str`, *optional*) : The target dtype for the activations after quantization. Supported values are (None,"int8","float8")

modules_to_not_convert (`list`, *optional*, default to `None`) : The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision (e.g. Whisper encoder, Llava encoder, Mixtral gate layers).
This is a wrapper class about all possible attributes and features that you can play with a model that has been
loaded using `quanto`.
#### post_init[[transformers.QuantoConfig.post_init]]
```python
post_init()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1049)
Safety checker that arguments are correct
## AqlmConfig[[transformers.AqlmConfig]][[transformers.AqlmConfig]]
#### transformers.AqlmConfig[[transformers.AqlmConfig]]
```python
transformers.AqlmConfig(in_group_size: int = 8, out_group_size: int = 1, num_codebooks: int = 1, nbits_per_codebook: int = 16, linear_weights_not_to_quantize: list[str] | None = None, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L865)
**Parameters:**
in_group_size (`int`, *optional*, defaults to 8) : The group size along the input dimension.
out_group_size (`int`, *optional*, defaults to 1) : The group size along the output dimension. It's recommended to always use 1.
num_codebooks (`int`, *optional*, defaults to 1) : Number of codebooks for the Additive Quantization procedure.

nbits_per_codebook (`int`, *optional*, defaults to 16) : Number of bits encoding a single codebook vector. Codebooks size is 2**nbits_per_codebook.
linear_weights_not_to_quantize (`Optional[list[str]]`, *optional*) : List of full paths of `nn.Linear` weight parameters that shall not be quantized.
kwargs (`dict[str, Any]`, *optional*) : Additional parameters from which to initialize the configuration object.
This is a wrapper class about `aqlm` parameters.
#### post_init[[transformers.AqlmConfig.post_init]]
```python
post_init()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L902)
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
## VptqConfig[[transformers.VptqConfig]][[transformers.VptqConfig]]
#### transformers.VptqConfig[[transformers.VptqConfig]]
```python
transformers.VptqConfig(enable_proxy_error: bool = False, config_for_layers: dict = {}, shared_layer_config: dict = {}, modules_to_not_convert: list | None = None, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L980)
**Parameters:**
enable_proxy_error (`bool`, *optional*, defaults to `False`) : calculate proxy error for each layer
config_for_layers (`Dict`, *optional*, defaults to `{}`) : quantization params for each layer
shared_layer_config (`Dict`, *optional*, defaults to `{}`) : shared quantization params among layers
modules_to_not_convert (`list`, *optional*, default to `None`) : The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision (e.g. Whisper encoder, Llava encoder, Mixtral gate layers).
kwargs (`dict[str, Any]`, *optional*) : Additional parameters from which to initialize the configuration object.
This is a wrapper class about `vptq` parameters.
#### post_init[[transformers.VptqConfig.post_init]]
```python
post_init()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1010)
Safety checker that arguments are correct
## AwqConfig[[transformers.AwqConfig]][[transformers.AwqConfig]]
#### transformers.AwqConfig[[transformers.AwqConfig]]
```python
transformers.AwqConfig(bits: int = 4, group_size: int = 128, zero_point: bool = True, backend: AwqBackend = <AwqBackend.AUTO: 'auto'>, modules_to_not_convert: list | None = None, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L799)
**Parameters:**

bits (`int`, *optional*, defaults to 4) : The number of bits to quantize to.

group_size (`int`, *optional*, defaults to 128) : The group size to use for quantization. Recommended value is 128 and -1 uses per-column quantization.
zero_point (`bool`, *optional*, defaults to `True`) : Whether to use zero point quantization.

backend (`AwqBackend`, *optional*, defaults to `AwqBackend.AUTO`) : The quantization backend.

modules_to_not_convert (`list`, *optional*, default to `None`) : The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision (e.g. Whisper encoder, Llava encoder, Mixtral gate layers). Note you cannot quantize directly with transformers, please refer to `AutoAWQ` documentation for quantizing HF models.
This is a wrapper class about all possible attributes and features that you can play with a model that has been
loaded using `auto-awq` library awq quantization relying on auto_awq backend.

## EetqConfig[[transformers.EetqConfig]][[transformers.EetqConfig]]

#### transformers.EetqConfig[[transformers.EetqConfig]]

```python
transformers.EetqConfig(weights: str = 'int8', modules_to_not_convert: list | None = None, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1062)

**Parameters:**

weights (`str`, *optional*, defaults to `"int8"`) : The target dtype for the weights. Supported value is only "int8"

modules_to_not_convert (`list`, *optional*, default to `None`) : The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision.
This is a wrapper class about all possible attributes and features that you can play with a model that has been
loaded using `eetq`.
#### post_init[[transformers.EetqConfig.post_init]]
```python
post_init()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1086)
Safety checker that arguments are correct
## GPTQConfig[[transformers.GPTQConfig]][[transformers.GPTQConfig]]
#### transformers.GPTQConfig[[transformers.GPTQConfig]]
```python
transformers.GPTQConfig(bits: int, tokenizer: Any = None, dataset: list[str] | str | None = None, group_size: int = 128, damp_percent: float = 0.1, desc_act: bool = False, act_group_aware: bool = True, sym: bool = True, true_sequential: bool = True, format: str = 'gptq', meta: dict[str, typing.Any] | None = None, backend: str | None = None, model_seqlen: int | None = None, block_name_to_quantize: str | None = None, module_name_preceding_first_block: list[str] | None = None, batch_size: int = 1, pad_token_id: int | None = None, max_input_length: int | None = None, cache_block_outputs: bool = True, modules_in_block_to_quantize: list[list[str]] | None = None, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L617)
**Parameters:**
bits (`int`) : The number of bits to quantize to, supported numbers are (2, 3, 4, 8).
tokenizer (`str` or `PreTrainedTokenizerBase`, *optional*) : The tokenizer used to process the dataset. You can pass either: - A custom tokenizer object. - A string, the *model id* of a predefined tokenizer hosted inside a model repo on huggingface.co. - A path to a *directory* containing vocabulary files required by the tokenizer, for instance saved using the [save_pretrained()](/docs/transformers/main/ko/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.save_pretrained) method, e.g., `./my_model_directory/`.
dataset (`Union[list[str]]`, *optional*) : The dataset used for quantization. You can provide your own dataset in a list of string or just use the original datasets used in GPTQ paper ['wikitext2','c4','c4-new']
group_size (`int`, *optional*, defaults to 128) : The group size to use for quantization. Recommended value is 128 and -1 uses per-column quantization.

damp_percent (`float`, *optional*, defaults to 0.1) : The percent of the average Hessian diagonal to use for dampening. Recommended value is 0.1.
desc_act (`bool`, *optional*, defaults to `False`) : Whether to quantize columns in order of decreasing activation size. Setting it to False can significantly speed up inference but the perplexity may become slightly worse. Also known as act-order.

act_group_aware (`bool`, *optional*, defaults to `True`) : Use GAR (group aware activation order) during quantization. Has measurable positive impact on quantization quality. Only applicable when `desc_act = False`. Will forced to be `False` when `desc_act = True`.

sym (`bool`, *optional*, defaults to `True`) : Whether to use symmetric quantization.

true_sequential (`bool`, *optional*, defaults to `True`) : Whether to perform sequential quantization even within a single Transformer block. Instead of quantizing the entire block at once, we perform layer-wise quantization. As a result, each layer undergoes quantization using inputs that have passed through the previously quantized layers.
format (`str`, *optional*, defaults to `"gptq"`) : GPTQ weight format. `gptq` (v1) is supported by gptqmodel. `gptq_v2` is gptqmodel only.

meta (`dict[str, any]`, *optional*) : Properties, such as tooling:version, that do not directly contributes to quantization or quant inference are stored in meta. i.e. `meta.quantizer`: ["optimum:_version_", "gptqmodel:_version_"]

backend (`str`, *optional*) : Controls which kernel to use. Valid values for gptqmodel are `auto`, `auto_trainable` and more. Ref gptqmodel backends: https://github.com/ModelCloud/GPTQModel/blob/main/gptqmodel/utils/backend.py

model_seqlen (`int`, *optional*) : The maximum sequence length that the model can take.
block_name_to_quantize (`str`, *optional*) : The transformers block name to quantize. If None, we will infer the block name using common patterns (e.g. model.layers)

module_name_preceding_first_block (`list[str]`, *optional*) : The layers that are preceding the first Transformer block.

batch_size (`int`, *optional*, defaults to 1) : The batch size used when processing the dataset
pad_token_id (`int`, *optional*) : The pad token id. Needed to prepare the dataset when `batch_size` > 1.

max_input_length (`int`, *optional*) : The maximum input length. This is needed to initialize a buffer that depends on the maximum expected input length. It is specific to the exllama backend with act-order.

cache_block_outputs (`bool`, *optional*, defaults to `True`) : Whether to cache block outputs to reuse as inputs for the succeeding block.

modules_in_block_to_quantize (`list[list[str]]`, *optional*) : List of list of module names to quantize in the specified block. This argument is useful to exclude certain linear modules from being quantized. The block to quantize can be specified by setting `block_name_to_quantize`. We will quantize each list sequentially. If not set, we will quantize all linear layers. Example: `modules_in_block_to_quantize =[["self_attn.k_proj", "self_attn.v_proj", "self_attn.q_proj"], ["self_attn.o_proj"]]`. In this example, we will first quantize the q,k,v layers simultaneously since they are independent. Then, we will quantize `self_attn.o_proj` layer with the q,k,v layers quantized. This way, we will get better results since it reflects the real input `self_attn.o_proj` will get when the model is quantized.

This is a wrapper class about all possible attributes and features that you can play with a model that has been
loaded using `optimum` api for GPTQ quantization relying on the gptqmodel backend.

#### from_dict_optimum[[transformers.GPTQConfig.from_dict_optimum]]

```python
from_dict_optimum(config_dict)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L788)

Get compatible class with optimum gptq config dict

#### post_init[[transformers.GPTQConfig.post_init]]

```python
post_init()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L738)

Safety checker that arguments are correct

#### to_dict_optimum[[transformers.GPTQConfig.to_dict_optimum]]

```python
to_dict_optimum()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L782)

Get compatible dict for optimum gptq config

## BitsAndBytesConfig[[#transformers.BitsAndBytesConfig]][[transformers.BitsAndBytesConfig]]

#### transformers.BitsAndBytesConfig[[transformers.BitsAndBytesConfig]]

```python
transformers.BitsAndBytesConfig(load_in_8bit = False, load_in_4bit = False, llm_int8_threshold = 6.0, llm_int8_skip_modules = None, llm_int8_enable_fp32_cpu_offload = False, llm_int8_has_fp16_weight = False, bnb_4bit_compute_dtype = None, bnb_4bit_quant_type = 'fp4', bnb_4bit_use_double_quant = False, bnb_4bit_quant_storage = None, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L390)

**Parameters:**

load_in_8bit (`bool`, *optional*, defaults to `False`) : This flag is used to enable 8-bit quantization with LLM.int8().

load_in_4bit (`bool`, *optional*, defaults to `False`) : This flag is used to enable 4-bit quantization by replacing the Linear layers with FP4/NF4 layers from `bitsandbytes`.

llm_int8_threshold (`float`, *optional*, defaults to 6.0) : This corresponds to the outlier threshold for outlier detection as described in `LLM.int8() : 8-bit Matrix Multiplication for Transformers at Scale` paper: https://huggingface.co/papers/2208.07339 Any hidden states value that is above this threshold will be considered an outlier and the operation on those values will be done in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], but there are some exceptional systematic outliers that are very differently distributed for large models. These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6, but a lower threshold might be needed for more unstable models (small models, fine-tuning).

llm_int8_skip_modules (`list[str]`, *optional*) : An explicit list of the modules that we do not want to convert in 8-bit. This is useful for models such as Jukebox that has several heads in different places and not necessarily at the last position. For example for `CausalLM` models, the last `lm_head` is kept in its original `dtype`.

llm_int8_enable_fp32_cpu_offload (`bool`, *optional*, defaults to `False`) : This flag is used for advanced use cases and users that are aware of this feature. If you want to split your model in different parts and run some parts in int8 on GPU and some parts in fp32 on CPU, you can use this flag. This is useful for offloading large models such as `google/flan-t5-xxl`. Note that the int8 operations will not be run on CPU.
llm_int8_has_fp16_weight (`bool`, *optional*, defaults to `False`) : This flag runs LLM.int8() with 16-bit main weights. This is useful for fine-tuning as the weights do not have to be converted back and forth for the backward pass.
bnb_4bit_compute_dtype (`torch.dtype` or str, *optional*, defaults to `torch.float32`) : This sets the computational type which might be different than the input type. For example, inputs might be fp32, but computation can be set to bf16 for speedups.

bnb_4bit_quant_type (`str`,  *optional*, defaults to `"fp4"`) : This sets the quantization data type in the bnb.nn.Linear4Bit layers. Options are FP4 and NF4 data types which are specified by `fp4` or `nf4`.
bnb_4bit_use_double_quant (`bool`, *optional*, defaults to `False`) : This flag is used for nested quantization where the quantization constants from the first quantization are quantized again.
bnb_4bit_quant_storage (`torch.dtype` or str, *optional*, defaults to `torch.uint8`) : This sets the storage type to pack the quantized 4-bit params.

kwargs (`dict[str, Any]`, *optional*) : Additional parameters from which to initialize the configuration object.

This is a wrapper class about all possible attributes and features that you can play with a model that has been
loaded using `bitsandbytes`.

Currently only supports `LLM.int8()`, `FP4`, and `NF4` quantization. If more methods are added to `bitsandbytes`,
then more arguments will be added to this class.

#### is_quantizable[[transformers.BitsAndBytesConfig.is_quantizable]]

```python
is_quantizable()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L551)

Returns `True` if the model is quantizable, `False` otherwise.

#### post_init[[transformers.BitsAndBytesConfig.post_init]]

```python
post_init()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L521)

Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.

#### quantization_method[[transformers.BitsAndBytesConfig.quantization_method]]

```python
quantization_method()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L557)

This method returns the quantization method used for the model. If the model is not quantizable, it returns
`None`.

#### to_diff_dict[[transformers.BitsAndBytesConfig.to_diff_dict]]

```python
to_diff_dict()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L588)

**Returns:** `dict[str, Any]`

Dictionary of all the attributes that make up this configuration instance,

Removes all attributes from config which correspond to the default config attributes for better readability and
serializes to a Python dictionary.

## HfQuantizer[[transformers.quantizers.HfQuantizer]][[transformers.quantizers.HfQuantizer]]

#### transformers.quantizers.HfQuantizer[[transformers.quantizers.HfQuantizer]]

```python
transformers.quantizers.HfQuantizer(quantization_config: QuantizationConfigMixin, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L74)

Abstract class of the HuggingFace quantizer. Supports for now quantizing HF transformers models for inference and/or quantization.
This class is used only for transformers.PreTrainedModel.from_pretrained and cannot be easily used outside the scope of that method
yet.
Attributes
quantization_config (`transformers.utils.quantization_config.QuantizationConfigMixin`):
The quantization config that defines the quantization parameters of your model that you want to quantize.
requires_calibration (`bool`):
Whether the quantization method requires to calibrate the model before using it.
#### adjust_max_memory[[transformers.quantizers.HfQuantizer.adjust_max_memory]]
```python
adjust_max_memory(max_memory: dict)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L148)
adjust max_memory argument for infer_auto_device_map() if extra memory is needed for quantization
#### dequantize[[transformers.quantizers.HfQuantizer.dequantize]]
```python
dequantize(model, dtype = None)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L239)
Potentially dequantize the model to retrieve the original model, with some loss in accuracy / performance.
Note not all quantization schemes support this.
#### get_load_operations[[transformers.quantizers.HfQuantizer.get_load_operations]]
```python
get_load_operations()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L323)
Op classes loading the serialized tensors, e.g. `{"deserialize": Mxfp4Deserialize, "dequantize": Mxfp4Dequantize}`.
A deserialization builds the params of one module, so it only runs on the general converters. A dequantization
also runs before the ops of the model's converters, e.g. before a per-expert merge.
#### get_new_dtype[[transformers.quantizers.HfQuantizer.get_new_dtype]]
```python
get_new_dtype(converter, param, dtype: torch.dtype | None)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L134)
This happens at materialization time, when the tensor is sent from disk to device.
`dtype` to cast the pre-quantized checkpoint tensor `param` in.
If that param is not quantized (float type) we use the input `dtype` that comes from the initialized `nn.Module`.
If the quantization method has a `Deserializer` or `Dequantizer` (load function) we return None so it's handled by them.
#### get_state_dict_and_metadata[[transformers.quantizers.HfQuantizer.get_state_dict_and_metadata]]
```python
get_state_dict_and_metadata(model)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L290)
Get state dict and metadata. Useful when we need to modify a bit the state dict due to quantization
#### get_weight_conversions[[transformers.quantizers.HfQuantizer.get_weight_conversions]]
```python
get_weight_conversions()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L331)
Return the loading to perform on any parameter that is not matched by the model's converters
#### param_needs_quantization[[transformers.quantizers.HfQuantizer.param_needs_quantization]]
```python
param_needs_quantization(model: PreTrainedModel, param_name: str, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L152)
Check whether a given param needs to be quantized.
#### postprocess_model[[transformers.quantizers.HfQuantizer.postprocess_model]]
```python
postprocess_model(model: PreTrainedModel, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L207)
**Parameters:**
model (`~transformers.PreTrainedModel`) : The model to quantize
kwargs (`dict`, *optional*) : The keyword arguments that are passed along `_process_model_after_weight_loading`.

Post-process the model post weights loading.
Make sure to override the abstract method `_process_model_after_weight_loading`.

#### preprocess_model[[transformers.quantizers.HfQuantizer.preprocess_model]]

```python
preprocess_model(model: PreTrainedModel, dtype = None, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L184)

**Parameters:**

model (`~transformers.PreTrainedModel`) : The model to quantize

kwargs (`dict`, *optional*) : The keyword arguments that are passed along `_process_model_before_weight_loading`.

Setting model attributes and/or converting model before weights loading. At this point
the model should be initialized on the meta device so you can freely manipulate the skeleton
of the model in order to replace modules in-place. Make sure to override the abstract method `_process_model_before_weight_loading`.

#### remove_quantization_config[[transformers.quantizers.HfQuantizer.remove_quantization_config]]

```python
remove_quantization_config(model)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L227)

Remove the quantization config from the model.

#### update_attn_implementation[[transformers.quantizers.HfQuantizer.update_attn_implementation]]

```python
update_attn_implementation(config)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L177)

Sees the requested `attn_implementation` before the model is built, so a quantizer can default it.

#### update_device_map[[transformers.quantizers.HfQuantizer.update_device_map]]

```python
update_device_map(device_map: dict[str, typing.Any] | None)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L112)

**Parameters:**

device_map (`Union[dict, str]`, *optional*) : The device_map that is passed through the `from_pretrained` method.

Override this method if you want to pass a override the existing device map with a new
one. E.g. for bitsandbytes, since `accelerate` is a hard requirement, if no device_map is
passed, the device_map is set to `"auto"``

#### update_dtype[[transformers.quantizers.HfQuantizer.update_dtype]]

```python
update_dtype(dtype: torch.dtype)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L100)

**Parameters:**

dtype (`torch.dtype`) : The input dtype that is passed in `from_pretrained`

Some quantization methods require to explicitly set the dtype of the model to a
target dtype. You need to override this method in case you want to make sure that behavior is
preserved

#### update_parallel_plan[[transformers.quantizers.HfQuantizer.update_parallel_plan]]

```python
update_parallel_plan(model, plan: dict)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L166)

Updates in place `model.tp_plan` or `model.ep_plan` (the full plans, merged from the config and all
sub-configs) to take into account the quantizer's custom keys like 'inv_scale'
#### update_weight_conversions[[transformers.quantizers.HfQuantizer.update_weight_conversions]]
```python
update_weight_conversions(weight_conversions, model)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L341)
Makes the model's converters (from the conversion mapping) load the serialized names. The model's params
decide, e.g. for a merge of `experts.*.gate_proj.weight` into `experts.gate_up_proj`:
- `experts.gate_up_proj_scale_inv` is a param: a copy merges `experts.*.gate_proj.weight_scale_inv` into it
- otherwise the converter also collects `experts.*.gate_proj.weight_scale_inv`, the dequantize op runs first
#### validate_environment[[transformers.quantizers.HfQuantizer.validate_environment]]
```python
validate_environment(*args, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/quantizers/base.py#L158)
This method is used to potentially check for potential conflicts with arguments that are
passed in `from_pretrained`. You need to define it for all future quantizers that are integrated with transformers.
If no explicit check are needed, simply return nothing.
## HqqConfig[[transformers.HqqConfig]][[transformers.HqqConfig]]
#### transformers.HqqConfig[[transformers.HqqConfig]]
```python
transformers.HqqConfig(nbits: int = 4, group_size: int = 64, view_as_float: bool = False, axis: int | None = None, dynamic_config: dict | None = None, skip_modules: list = ['lm_head'], **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L277)
**Parameters:**
nbits (`int`, *optional*, defaults to 4) : Number of bits. Supported values are (8, 4, 3, 2, 1).
group_size (`int`, *optional*, defaults to 64) : Group-size value. Supported values are any value that is divisible by weight.shape[axis]).

view_as_float (`bool`, *optional*, defaults to `False`) : View the quantized weight as float (used in distributed training) if set to `True`.

axis (`Optional[int]`, *optional*) : Axis along which grouping is performed. Supported values are 0 or 1.

dynamic_config (dict, *optional*) : Parameters for dynamic configuration. The key is the name tag of the layer and the value is a quantization config. If set, each layer specified by its id will use its dedicated quantization configuration.
skip_modules (`list[str]`, *optional*, defaults to `['lm_head']`) : List of `nn.Linear` layers to skip.

kwargs (`dict[str, Any]`, *optional*) : Additional parameters from which to initialize the configuration object.

This is wrapper around hqq's BaseQuantizeConfig.

#### from_dict[[transformers.HqqConfig.from_dict]]

```python
from_dict(config: dict)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L342)

Override from_dict, used in AutoQuantizationConfig.from_dict in quantizers/auto.py

#### post_init[[transformers.HqqConfig.post_init]]

```python
post_init()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L337)

Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.

#### to_diff_dict[[transformers.HqqConfig.to_diff_dict]]

```python
to_diff_dict()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L367)

**Returns:** `dict[str, Any]`

Dictionary of all the attributes that make up this configuration instance,

Removes all attributes from config which correspond to the default config attributes for better readability and
serializes to a Python dictionary.

## FbgemmFp8Config[[transformers.FbgemmFp8Config]][[transformers.FbgemmFp8Config]]

#### transformers.FbgemmFp8Config[[transformers.FbgemmFp8Config]]

```python
transformers.FbgemmFp8Config(activation_scale_ub: float = 1200.0, modules_to_not_convert: list | None = None, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1292)

**Parameters:**

activation_scale_ub (`float`, *optional*, defaults to 1200.0) : The activation scale upper bound. This is used when quantizing the input activation.

modules_to_not_convert (`list`, *optional*, default to `None`) : The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision.
This is a wrapper class about all possible attributes and features that you can play with a model that has been
loaded using fbgemm fp8 quantization.
## CompressedTensorsConfig[[transformers.CompressedTensorsConfig]][[transformers.CompressedTensorsConfig]]
#### transformers.CompressedTensorsConfig[[transformers.CompressedTensorsConfig]]
```python
transformers.CompressedTensorsConfig(config_groups: dict[str, typing.Union[ForwardRef('QuantizationScheme'), list[str]]] | None = None, format: str = 'dense', quantization_status: QuantizationStatus = 'initialized', kv_cache_scheme: typing.Optional[ForwardRef('QuantizationArgs')] = None, global_compression_ratio: float | None = None, ignore: list[str] | None = None, quant_method: str = 'compressed-tensors', run_compressed: bool | None = None, dequantize: bool = False, use_optimized_inference: bool = False, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1095)
**Parameters:**
config_groups (`typing.dict[str, typing.Union[ForwardRef('QuantizationScheme'), typing.list[str]]]`, *optional*) : dictionary mapping group name to a quantization scheme definition

format (`str`, *optional*, defaults to `"dense"`) : format the weights are stored in, e.g. `dense`, `float-quantized` or `pack-quantized`

quantization_status (`QuantizationStatus`, *optional*, defaults to `"initialized"`) : status of model in the quantization lifecycle, ie 'initialized', 'calibration', 'frozen'
kv_cache_scheme (`typing.Union[QuantizationArgs, NoneType]`, *optional*) : specifies quantization of the kv cache. If None, kv cache is not quantized.
global_compression_ratio (`typing.Union[float, NoneType]`, *optional*) : 0-1 float percentage of model compression
ignore (`typing.Union[typing.list[str], NoneType]`, *optional*) : layer names or types to not quantize, supports regex prefixed by 're:'
quant_method (`str`, *optional*, defaults to `"compressed-tensors"`) : do not override, should be compressed-tensors

run_compressed (`bool`, *optional*) : deprecated, use `dequantize` instead. Its only effect is to set `dequantize` to its opposite, `run_compressed=False` meaning `dequantize=True`.

dequantize (`bool`, *optional*, defaults to `False`) : when `True`, the quantized weights are dequantized back to the model dtype (e.g. BF16) at load time, which is what fine-tuning or saving the model in its original dtype needs. When `False`, the weights are left in their compressed form and compressed-tensors decompresses them on the first forward pass.

use_optimized_inference (`bool`, *optional*, defaults to `False`) : when `True`, the layers whose scheme we have kernels for keep their quantized weights and run through those kernels (currently W8A8 FP8, on CUDA SM89+, XPU, or MPS with torch >= 2.15). Inference only, and opt-in: without it the model goes through the regular compressed-tensors route. Ignored when `dequantize=True`.

This is a wrapper class that handles compressed-tensors quantization config options.
It is a wrapper around `compressed_tensors.QuantizationConfig`

#### from_dict[[transformers.CompressedTensorsConfig.from_dict]]

```python
from_dict(config_dict, return_unused_kwargs = False, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1200)

**Parameters:**

config_dict (`dict[str, Any]`) : Dictionary that will be used to instantiate the configuration object.
return_unused_kwargs (`bool`,*optional*, defaults to `False`) : Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in `PreTrainedModel`.

kwargs (`dict[str, Any]`) : Additional parameters from which to initialize the configuration object.

**Returns:** `QuantizationConfigMixin`

The configuration object instantiated from those parameters.

Instantiates a [CompressedTensorsConfig](/docs/transformers/main/ko/main_classes/quantization#transformers.CompressedTensorsConfig) from a Python dictionary of parameters.
Optionally unwraps any args from the nested quantization_config
#### to_dict[[transformers.CompressedTensorsConfig.to_dict]]
```python
to_dict()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1225)
Quantization config to be added to config.json
Serializes this instance to a Python dictionary. Returns:
`dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.
#### to_diff_dict[[transformers.CompressedTensorsConfig.to_diff_dict]]
```python
to_diff_dict()
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1240)
**Returns:** `dict[str, Any]`
Dictionary of all the attributes that make up this configuration instance,
Removes all attributes from config which correspond to the default config attributes for better readability and
serializes to a Python dictionary.
## TorchAoConfig[[transformers.TorchAoConfig]][[transformers.TorchAoConfig]]
#### transformers.TorchAoConfig[[transformers.TorchAoConfig]]
```python
transformers.TorchAoConfig(quant_type: AOBaseConfig, modules_to_not_convert: list | None = None, include_input_output_embeddings: bool = False, untie_embedding_weights: bool = False, **kwargs)
```
[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1468)
**Parameters:**
quant_type (`AOBaseConfig`) : A torchao `AOBaseConfig` instance specifying the quantization type, e.g. `Int4WeightOnlyConfig(group_size=32)`, `Int8WeightOnlyConfig()`, `Int8DynamicActivationInt8WeightConfig()`, `Float8WeightOnlyConfig()`, etc.

modules_to_not_convert (`list`, *optional*, default to `None`) : The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision.
include_input_output_embeddings (`bool`, *optional*, defaults to `False`) : Whether to include embedding in quantization or not, input embedding will be removed from the module_not_to_convert list as well if this flag is set.
untie_embedding_weights (`bool`, *optional*, defaults to `False`) : Whether to untie the weights when we are quantizing input embedding weights that is tied to other weights.
Config class for torchao quantization/sparsity techniques.
Example:
```python
from torchao.quantization import Int4WeightOnlyConfig
quantization_config = TorchAoConfig(Int4WeightOnlyConfig(group_size=32))
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="cuda", torch_dtype=torch.bfloat16, quantization_config=quantization_config
)
```
#### from_dict[[transformers.TorchAoConfig.from_dict]]

```python
from_dict(config_dict, return_unused_kwargs = False, **kwargs)
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1551)

Create configuration from a dictionary.

#### get_apply_tensor_subclass[[transformers.TorchAoConfig.get_apply_tensor_subclass]]

```python
get_apply_tensor_subclass()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1537)

Return the quantization config to apply.

#### post_init[[transformers.TorchAoConfig.post_init]]

```python
post_init()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1519)

Validate configuration and set defaults.

#### to_dict[[transformers.TorchAoConfig.to_dict]]

```python
to_dict()
```

[Source](https://github.com/huggingface/transformers/blob/main/src/transformers/utils/quantization_config.py#L1541)

Convert configuration to a dictionary.
Xet Storage Details

Size:: 39.2 kB
Xet hash:: 91b57b6fd1683da172c8a15df9ee9141b2e4e221d46e53af244a6fb595143482
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.