Quantization techniques reduce memory and computational costs by representing weights and activations with lower-precision data types like 8-bit integers (int8). This enables loading larger models you normally wouldn’t be able to fit into memory, and speeding up inference. Transformers supports the AWQ and GPTQ quantization algorithms and it supports 8-bit and 4-bit quantization with bitsandbytes.
Quantization techniques that aren’t supported in Transformers can be added with the HfQuantizer class.
Learn how to quantize models in the Quantization guide.
class transformers.QuantoConfig
< source >
( weights = 'int8' activations = None modules_to_not_convert: list | None = None **kwargs )
Parameters
str, optional, defaults to "int8") —
The target dtype for the weights after quantization. Supported values are (“float8”,“int8”,“int4”,“int2”) str, optional) —
The target dtype for the activations after quantization. Supported values are (None,“int8”,“float8”) 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.
Safety checker that arguments are correct
class transformers.AqlmConfig
< source >
( 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 )
Parameters
int, optional, defaults to 8) —
The group size along the input dimension. int, optional, defaults to 1) —
The group size along the output dimension. It’s recommended to always use 1. int, optional, defaults to 1) —
Number of codebooks for the Additive Quantization procedure. int, optional, defaults to 16) —
Number of bits encoding a single codebook vector. Codebooks size is 2**nbits_per_codebook. Optional[list[str]], optional) —
List of full paths of nn.Linear weight parameters that shall not be quantized. dict[str, Any], optional) —
Additional parameters from which to initialize the configuration object. This is a wrapper class about aqlm parameters.
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
class transformers.VptqConfig
< source >
( enable_proxy_error: bool = False config_for_layers: dict = {} shared_layer_config: dict = {} modules_to_not_convert: list | None = None **kwargs )
Parameters
bool, optional, defaults to False) — calculate proxy error for each layer Dict, optional, defaults to {}) — quantization params for each layer Dict, optional, defaults to {}) — shared quantization params among layers 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). dict[str, Any], optional) —
Additional parameters from which to initialize the configuration object. This is a wrapper class about vptq parameters.
Safety checker that arguments are correct
class transformers.AwqConfig
< source >
( bits: int = 4 group_size: int = 128 zero_point: bool = True backend: AwqBackend = <AwqBackend.AUTO: 'auto'> modules_to_not_convert: list | None = None **kwargs )
Parameters
int, optional, defaults to 4) —
The number of bits to quantize to. int, optional, defaults to 128) —
The group size to use for quantization. Recommended value is 128 and -1 uses per-column quantization. bool, optional, defaults to True) —
Whether to use zero point quantization. AwqBackend, optional, defaults to AwqBackend.AUTO) —
The quantization backend. 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.
class transformers.EetqConfig
< source >
( weights: str = 'int8' modules_to_not_convert: list | None = None **kwargs )
Parameters
str, optional, defaults to "int8") —
The target dtype for the weights. Supported value is only “int8” 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.
Safety checker that arguments are correct
class transformers.GPTQConfig
< source >
( bits: int tokenizer: typing.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 )
Parameters
int) —
The number of bits to quantize to, supported numbers are (2, 3, 4, 8). str or PreTrainedTokenizerBase, optional) —
The tokenizer used to process the dataset. You can pass either:
./my_model_directory/.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’] int, optional, defaults to 128) —
The group size to use for quantization. Recommended value is 128 and -1 uses per-column quantization. float, optional, defaults to 0.1) —
The percent of the average Hessian diagonal to use for dampening. Recommended value is 0.1. 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. 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. bool, optional, defaults to True) —
Whether to use symmetric quantization. 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. str, optional, defaults to "gptq") —
GPTQ weight format. gptq (v1) is supported by gptqmodel. gptq_v2 is gptqmodel only. 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”] 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 int, optional) —
The maximum sequence length that the model can take. str, optional) —
The transformers block name to quantize. If None, we will infer the block name using common patterns (e.g. model.layers) list[str], optional) —
The layers that are preceding the first Transformer block. int, optional, defaults to 1) —
The batch size used when processing the dataset int, optional) —
The pad token id. Needed to prepare the dataset when batch_size > 1. 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. bool, optional, defaults to True) —
Whether to cache block outputs to reuse as inputs for the succeeding block. 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.
Get compatible class with optimum gptq config dict
Safety checker that arguments are correct
Get compatible dict for optimum gptq config
class transformers.BitsAndBytesConfig
< source >
( 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 )
Parameters
bool, optional, defaults to False) —
This flag is used to enable 8-bit quantization with LLM.int8(). 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. 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). 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. 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. 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. 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. 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. bool, optional, defaults to False) —
This flag is used for nested quantization where the quantization constants from the first quantization are
quantized again. torch.dtype or str, optional, defaults to torch.uint8) —
This sets the storage type to pack the quantized 4-bit params. 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.
Returns True if the model is quantizable, False otherwise.
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
This method returns the quantization method used for the model. If the model is not quantizable, it returns
None.
to_diff_dict
< source >
( ) → dict[str, Any]
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.
class transformers.quantizers.HfQuantizer
< source >
( quantization_config: QuantizationConfigMixin **kwargs )
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 argument for infer_auto_device_map() if extra memory is needed for quantization
Potentially dequantize the model to retrieve the original model, with some loss in accuracy / performance. Note not all quantization schemes support this.
Override this method if you want to adjust the param_name.
Get state dict and metadata. Useful when we need to modify a bit the state dict due to quantization
param_needs_quantization
< source >
( model: PreTrainedModel param_name: str **kwargs )
Check whether a given param needs to be quantized.
postprocess_model
< source >
( model: PreTrainedModel **kwargs )
Post-process the model post weights loading.
Make sure to override the abstract method _process_model_after_weight_loading.
preprocess_model
< source >
( model: PreTrainedModel dtype = None **kwargs )
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 the quantization config from the model.
update_device_map
< source >
( device_map: dict[str, typing.Any] | None )
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
< source >
( dtype: torch.dtype )
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
updates the tp plan for the scales
updates the tp plan for the scales
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.
class transformers.HiggsConfig
< source >
( bits: int = 4 p: int = 2 modules_to_not_convert: list[str] | None = None hadamard_size: int = 512 group_size: int = 256 tune_metadata: dict[str, typing.Any] | None = None **kwargs )
Parameters
list, optional, default to [“lm_head”]) —
List of linear layers that should not be quantized. HiggsConfig is a configuration class for quantization using the HIGGS method.
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
class transformers.HqqConfig
< source >
( 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 )
Parameters
int, optional, defaults to 4) —
Number of bits. Supported values are (8, 4, 3, 2, 1). int, optional, defaults to 64) —
Group-size value. Supported values are any value that is divisible by weight.shape[axis]). bool, optional, defaults to False) —
View the quantized weight as float (used in distributed training) if set to True. Optional[int], optional) —
Axis along which grouping is performed. Supported values are 0 or 1. list[str], optional, defaults to ['lm_head']) —
List of nn.Linear layers to skip. dict[str, Any], optional) —
Additional parameters from which to initialize the configuration object. This is wrapper around hqq’s BaseQuantizeConfig.
Override from_dict, used in AutoQuantizationConfig.from_dict in quantizers/auto.py
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
to_diff_dict
< source >
( ) → dict[str, Any]
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.
class transformers.MetalConfig
< source >
( bits: int = 4 group_size: int = 64 modules_to_not_convert: list | None = None dequantize: bool = False **kwargs )
Configuration class for Metal affine quantization targeting Apple Silicon (MPS) devices.
This quantization method uses the mlx-quantization-metal-kernels Metal kernels from the Hugging Face Hub
to perform affine quantization (scales + qbiases) with configurable bit-width and group size.
The quantized weights are packed into uint32 tensors and the forward pass uses fused
dequantization + matmul Metal kernels.
class transformers.Mxfp4Config
< source >
( modules_to_not_convert: list | None = None dequantize: bool = False **kwargs )
Parameters
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. bool, optional, default to False) —
Whether we dequantize the model to bf16 precision or not This is a wrapper class about all possible attributes and features that you can play with a model that has been loaded using mxfp4 quantization.
class transformers.FbgemmFp8Config
< source >
( activation_scale_ub: float = 1200.0 modules_to_not_convert: list | None = None **kwargs )
Parameters
float, optional, defaults to 1200.0) —
The activation scale upper bound. This is used when quantizing the input activation. 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.
class transformers.CompressedTensorsConfig
< source >
( 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 sparsity_config: dict[str, typing.Any] | None = None quant_method: str = 'compressed-tensors' run_compressed: bool = True **kwargs )
Parameters
typing.dict[str, typing.Union[ForwardRef('QuantizationScheme'), typing.list[str]]], optional) —
dictionary mapping group name to a quantization scheme definition str, optional, defaults to "dense") —
format the model is represented as. Set run_compressed True to execute model as the
compressed format if not dense QuantizationStatus, optional, defaults to "initialized") —
status of model in the quantization lifecycle, ie ‘initialized’, ‘calibration’, ‘frozen’ typing.Union[QuantizationArgs, NoneType], optional) —
specifies quantization of the kv cache. If None, kv cache is not quantized. typing.Union[float, NoneType], optional) —
0-1 float percentage of model compression typing.Union[typing.list[str], NoneType], optional) —
layer names or types to not quantize, supports regex prefixed by ‘re:’ typing.dict[str, typing.Any], optional) —
configuration for sparsity compression str, optional, defaults to "compressed-tensors") —
do not override, should be compressed-tensors bool, optional, defaults to True) — alter submodules (usually linear) in order to
emulate compressed model execution if True, otherwise use default submodule This is a wrapper class that handles compressed-tensors quantization config options.
It is a wrapper around compressed_tensors.QuantizationConfig
from_dict
< source >
( config_dict return_unused_kwargs = False **kwargs ) → QuantizationConfigMixin
Parameters
dict[str, Any]) —
Dictionary that will be used to instantiate the configuration object. bool,optional, defaults to False) —
Whether or not to return a list of unused keyword arguments. Used for from_pretrained method in
PreTrainedModel. dict[str, Any]) —
Additional parameters from which to initialize the configuration object. Returns
QuantizationConfigMixin
The configuration object instantiated from those parameters.
Instantiates a CompressedTensorsConfig from a Python dictionary of parameters. Optionally unwraps any args from the nested quantization_config
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
< source >
( ) → dict[str, Any]
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.
class transformers.TorchAoConfig
< source >
( quant_type: AOBaseConfig modules_to_not_convert: list | None = None include_input_output_embeddings: bool = False untie_embedding_weights: bool = False **kwargs )
Parameters
AOBaseConfig) —
A torchao AOBaseConfig instance specifying the quantization type, e.g.
Int4WeightOnlyConfig(group_size=32), Int8WeightOnlyConfig(),
Int8DynamicActivationInt8WeightConfig(), Float8WeightOnlyConfig(), etc. 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. 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. 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:
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
< source >
( config_dict return_unused_kwargs = False **kwargs )
Create configuration from a dictionary.
Return the quantization config to apply.
Validate configuration and set defaults.
Convert configuration to a dictionary.
class transformers.BitNetQuantConfig
< source >
( modules_to_not_convert: list | None = None linear_class: str = 'bitlinear' quantization_mode: str = 'offline' use_rms_norm: bool = False rms_norm_eps: float | None = 1e-06 **kwargs )
Parameters
Optional[List], optional) —
Optionally, provides a list of full paths of nn.Linear weight parameters
that shall not be quantized. Defaults to None. str, optional, defaults to "bitlinear") —
The type of linear class to use. Can be either bitlinear or autobitlinear. str, optional, defaults to "offline") —
The quantization mode to use. Can be either online or offline.
In online mode, the weight quantization parameters are calculated dynamically
during each forward pass (e.g., based on the current weight values). This can
adapt to weight changes during training (Quantization-Aware Training - QAT).
In offline mode, quantization parameters are pre-calculated before inference.
These parameters are then fixed and loaded into the quantized model. This
generally results in lower runtime overhead compared to online quantization. bool, optional, defaults to False) —
Whether to apply RMSNorm on the activations before quantization. This matches the original BitNet paper’s approach
of normalizing activations before quantization/packing. float, optional, defaults to 1e-06) —
The epsilon value used in the RMSNorm layer for numerical stability. dict[str, Any], optional) —
Additional keyword arguments that may be used by specific quantization
backends or future versions. Configuration class for applying BitNet quantization.
Safety checker that arguments are correct
class transformers.SpQRConfig
< source >
( bits: int = 3 beta1: int = 16 beta2: int = 16 shapes: dict[str, int] | None = None modules_to_not_convert: list[str] | None = None **kwargs )
Parameters
int, optional, defaults to 3) —
Specifies the bit count for the weights and first order zero-points and scales.
Currently only bits = 3 is supported. int, optional, defaults to 16) —
SpQR tile width. Currently only beta1 = 16 is supported. int, optional, defaults to 16) —
SpQR tile height. Currently only beta2 = 16 is supported. Optional, optional) —
A dictionary holding the shape of each object. We need this because it’s impossible
to deduce the exact size of the parameters just from bits, beta1, beta2. Optional[list[str]], optional) —
Optionally, provides a list of full paths of nn.Linear weight parameters that shall not be quantized.
Defaults to None. dict[str, Any], optional) —
Additional parameters from which to initialize the configuration object. This is a wrapper class about spqr parameters. Refer to the original publication for more details.
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
class transformers.FineGrainedFP8Config
< source >
( activation_scheme: str = 'dynamic' weight_block_size: tuple = (128, 128) dequantize: bool = False modules_to_not_convert: list | None = None **kwargs )
Parameters
str, optional, defaults to "dynamic") —
The scheme used for activation, the defaults and only support scheme for now is “dynamic”. typing.tuple[int, int], optional, defaults to (128, 128)) —
The size of the weight blocks for quantization, default is (128, 128). bool, optional, defaults to False) —
Whether to dequantize the model during loading. list, optional) —
A list of module names that should not be converted during quantization. FineGrainedFP8Config is a configuration class for fine-grained FP8 quantization used mainly for deepseek models.
Safety checker that arguments are correct
class transformers.QuarkConfig
< source >
( **kwargs )
class transformers.FourOverSixConfig
< source >
( activation_dtype: str | None = None activation_scale_rule: str | None = None dtype: str = 'nvfp4' gradient_dtype: str | None = None gradient_scale_rule: str | None = None keep_master_weights: bool = False matmul_backend: str | None = None output_dtype: str | None = 'bfloat16' quantize_backend: str | None = None scale_rule: str = 'mse' weight_dtype: str | None = None weight_scale_2d: bool = False weight_scale_rule: str | None = None module_config_overrides: dict[str, dict[str, typing.Any]] | None = None modules_to_not_convert: list[str] | None = ['lm_head'] **kwargs )
Parameters
str, optional) —
Data type to use when quantizing activation tensors. If not provided, dtype is used. str, optional) —
Scaling rule to use when selecting a scale for blocks in activation tensors. If not
provided, scale_rule is used. str, default “nvfp4”, optional, defaults to "nvfp4") —
The data type to use for the layer’s weights, activations, and tensors. Can be
"nvfp4" or "mxfp4". str, optional) —
Data type to use when quantizing gradient tensors. If not provided, dtype is used. str, optional) —
Scaling rule to use when selecting a scale for blocks in gradient tensors. If not
provided, scale_rule is used. bool, default False, optional, defaults to False) —
Whether to keep the master weights. If True, high-precision weights are kept at all
times and weights are quantized online in each forward pass. This is useful for
quantized training. str, optional) —
The backend to use for matrix multiplications. Can be "cutlass" or "pytorch". If
not provided, CUTLASS will be used if available and PyTorch will be used otherwise. str, optional, defaults to "bfloat16") —
The data type to use for the output of the layer. Can be "bfloat16" or "float16". str, optional) —
The backend to use for quantization. Can be "cuda", "triton", or "pytorch". If
not provided, the fastest backend will be selected based on your environment, and based
on the options supported by each backend. Typically, "cuda" will be used for
inference, "triton" will be used for training, and "pytorch" will be used on
non-CUDA devices. str, default “mse”, optional, defaults to "mse") —
Rule to use when selecting block scales. Can be "mse", "mae", or "abs_max" for
Four Over Six, "static_6" for default NVFP4 quantization, or "static_4" to scale
all blocks to a maximum value of 4. str, optional) —
Data type to use when quantizing weight tensors. If not provided, dtype is used. bool, default False, optional, defaults to False) —
Whether to compute scale factors on weight tensors in 2D blocks. This should be done
during training. str, optional) —
Scaling rule to use when selecting a scale for blocks in weight tensors. If not
provided, scale_rule is used. dict[str, dict[str, Any]], optional) —
A dictionary of module-specific configuration overrides. Keys should be module names, and
values should be dictionaries containing the quantization configuration for that module.
This can be used to override the default configuration for specific modules. list[str], optional, defaults to ['lm_head']) —
The list of modules to exclude from quantization. By default, the lm_head is excluded. This is a wrapper class containing all options for quantization with fouroversix. In brief,
Four Over Six is a modification to NVFP4 quantization which adaptively scales the largest value
in each block of 16 FP4 values to either 4 or 6. Selecting a scale of 6 uses the full range of
FP4 values, but selecting a scale of 4 allows for a more uniform distribution of quantization
error. Refer to the original publication for more details: https://arxiv.org/abs/2512.02010.
class transformers.FPQuantConfig
< source >
( forward_dtype: str = 'nvfp4' forward_method: str = 'abs_max' backward_dtype: str = 'bf16' store_master_weights: bool = False hadamard_group_size: int | None = None pseudoquantization: bool = False transform_init: str = 'hadamard' modules_to_not_convert: list[str] | None = None **kwargs )
Parameters
str, optional, defaults to "nvfp4") —
The dtype to use for the forward pass. str, optional, defaults to "abs_max") —
The scaling to use for the forward pass. Can be "abs_max" or "quest". "abs_max" is better for PTQ, "quest" is better for QAT. str, optional, defaults to "bf16") —
The dtype to use for the backward pass. bool, optional, defaults to False) —
Whether to store the master weights. Needed for QAT over layer weights. int, optional) —
The group size for the hadamard transform before quantization for "quest" it matches the MXFP4 group size (32). If None, it will be set to 16 for "nvfp4" and 32 for "mxfp4". bool, optional, defaults to False) —
Whether to use Triton-based pseudo-quantization. Is mandatory for non-Blackwell GPUs. Doesn’t provide any speedup. For debugging purposes. str, optional, defaults to "hadamard") — a method to initialize the pre-processing matrix with. Can be "hadamard", "identity" or "gsr". list, optional) —
The list of modules to not quantize, useful for quantizing models that explicitly require to have
some modules left in their original precision. FPQuantConfig is a configuration class for quantization using the FPQuant method.
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
class transformers.AutoRoundConfig
< source >
( bits: int = 4 group_size: int = 128 sym: bool = True backend: str = 'auto' **kwargs )
Parameters
int, optional, defaults to 4) —
The number of bits to quantize to, supported numbers are (2, 3, 4, 8). int, optional, defaults to 128) — Group-size value bool, optional, defaults to True) — Symmetric quantization or not str, optional, defaults to "auto") — The kernel to use, e.g., ipex,marlin, exllamav2, triton, etc. Ref. https://github.com/intel/auto-round?tab=readme-ov-file#specify-backend This is a wrapper class about all possible attributes and features that you can play with a model that has been loaded AutoRound quantization.
Safety checker that arguments are correct.
class transformers.SinqConfig
< source >
( nbits: int = 4 group_size: int = 64 tiling_mode: str = '1D' method: str = 'sinq' modules_to_not_convert: list[str] | None = None **kwargs: typing.Any )
Parameters
int, default 4) —
Quantization bits for weights. int, default 64) —
Group size used in SINQ weight quantization (must be multiple of 8). str, default “1D”) —
Tiling mode for SINQ (typically “1D”; “2D” if supported in your backend). str, default “sinq”) —
“sinq” – calibration-free weight-only SINQ
“asinq” – A-SINQ (activation-aware), not supported in Hugging Face. Please refer to the official SINQ repository. list[str], optional) —
List of module names/prefixes to keep in full precision. _extra_kwargs for round-tripping). Quantization config for SINQ / A-SINQ.
Pass this to:
AutoModel.from_pretrained(…, quantization_config=SinqConfig(…))
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。