Hugging Face's logo

Bitsandbytes documentation

Overview

You are viewing main version, which requires installation from source. If you'd like regular pip install, checkout the latest stable version (v0.43.0).
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Overview

8-bit optimizers reduce the memory footprint of 32-bit optimizers without any performance degradation which means you can train large models with many parameters faster. At the core of 8-bit optimizers is block-wise quantization which enables quantization accuracy, computational efficiency, and stability.

bitsandbytes provides 8-bit optimizers through the base Optimizer8bit class, and additionally provides Optimizer2State and Optimizer1State for 2-state (for example, Adam) and 1-state (for example, Adagrad) optimizers respectively. To provide custom optimizer hyperparameters, use the GlobalOptimManager class to configure the optimizer.

Optimizer8bit

class bitsandbytes.optim.optimizer.Optimizer8bit

< >

( params defaults optim_bits = 32 is_paged = False )

__init__

< >

( params defaults optim_bits = 32 is_paged = False )

Parameters

  • params (torch.tensor) — The input parameters to optimize.
  • optim_bits (int, defaults to 32) — The number of bits of the optimizer state.
  • is_paged (bool, defaults to False) — Whether the optimizer is a paged optimizer or not.

Base 8-bit optimizer class.

Optimizer2State

class bitsandbytes.optim.optimizer.Optimizer2State

< >

( optimizer_name params lr = 0.001 betas = (0.9, 0.999) eps = 1e-08 weight_decay = 0.0 optim_bits = 32 args = None min_8bit_size = 4096 percentile_clipping = 100 block_wise = True max_unorm = 0.0 skip_zeros = False is_paged = False )

__init__

< >

( optimizer_name params lr = 0.001 betas = (0.9, 0.999) eps = 1e-08 weight_decay = 0.0 optim_bits = 32 args = None min_8bit_size = 4096 percentile_clipping = 100 block_wise = True max_unorm = 0.0 skip_zeros = False is_paged = False )

Parameters

  • optimizer_name (str) — The name of the optimizer.
  • params (torch.tensor) — The input parameters to optimize.
  • lr (float, defaults to 1e-3) — The learning rate.
  • betas (tuple, defaults to (0.9, 0.999)) — The beta values for the optimizer.
  • eps (float, defaults to 1e-8) — The epsilon value for the optimizer.
  • weight_decay (float, defaults to 0.0) — The weight decay value for the optimizer.
  • optim_bits (int, defaults to 32) — The number of bits of the optimizer state.
  • args (object, defaults to None) — An object with additional arguments.
  • min_8bit_size (int, defaults to 4096) — The minimum number of elements of the parameter tensors for 8-bit optimization.
  • percentile_clipping (int, defaults to 100) — Adapts clipping threshold automatically by tracking the last 100 gradient norms and clipping the gradient at a certain percentile to improve stability.
  • block_wise (bool, defaults to True) — Whether to independently quantize each block of tensors to reduce outlier effects and improve stability.
  • max_unorm (float, defaults to 0.0) — The maximum value to normalize each block with.
  • skip_zeros (bool, defaults to False) — Whether to skip zero values for sparse gradients and models to ensure correct updates.
  • is_paged (bool, defaults to False) — Whether the optimizer is a paged optimizer or not.

Base 2-state update optimizer class.

Optimizer1State

class bitsandbytes.optim.optimizer.Optimizer1State

< >

( optimizer_name params lr = 0.001 betas = (0.9, 0.0) eps = 1e-08 weight_decay = 0.0 optim_bits = 32 args = None min_8bit_size = 4096 percentile_clipping = 100 block_wise = True max_unorm = 0.0 skip_zeros = False is_paged = False )

__init__

< >

( optimizer_name params lr = 0.001 betas = (0.9, 0.0) eps = 1e-08 weight_decay = 0.0 optim_bits = 32 args = None min_8bit_size = 4096 percentile_clipping = 100 block_wise = True max_unorm = 0.0 skip_zeros = False is_paged = False )

Parameters

  • optimizer_name (str) — The name of the optimizer.
  • params (torch.tensor) — The input parameters to optimize.
  • lr (float, defaults to 1e-3) — The learning rate.
  • betas (tuple, defaults to (0.9, 0.0)) — The beta values for the optimizer.
  • eps (float, defaults to 1e-8) — The epsilon value for the optimizer.
  • weight_decay (float, defaults to 0.0) — The weight decay value for the optimizer.
  • optim_bits (int, defaults to 32) — The number of bits of the optimizer state.
  • args (object, defaults to None) — An object with additional arguments.
  • min_8bit_size (int, defaults to 4096) — The minimum number of elements of the parameter tensors for 8-bit optimization.
  • percentile_clipping (int, defaults to 100) — Adapts clipping threshold automatically by tracking the last 100 gradient norms and clipping the gradient at a certain percentile to improve stability.
  • block_wise (bool, defaults to True) — Whether to independently quantize each block of tensors to reduce outlier effects and improve stability.
  • max_unorm (float, defaults to 0.0) — The maximum value to normalize each block with.
  • skip_zeros (bool, defaults to False) — Whether to skip zero values for sparse gradients and models to ensure correct updates.
  • is_paged (bool, defaults to False) — Whether the optimizer is a paged optimizer or not.

Base 1-state update optimizer class.

Utilities

class bitsandbytes.optim.GlobalOptimManager

< >

( )

A global optimizer manager for enabling custom optimizer configs.

override_config

< >

( parameters key = None value = None key_value_dict = None )

Parameters

  • parameters (torch.Tensor or list(torch.Tensors)) — The input parameters.
  • key (str) — The hyperparamter to override. value — The hyperparameter values.
  • key_value_dict (dict) — A dictionary with multiple key-values to override.

Override initial optimizer config with specific hyperparameters.

The key-values of the optimizer config for the input parameters are overridden This can be both, optimizer parameters like betas or lr, or it can be 8-bit specific parameters like optim_bits or percentile_clipping.

Example:

import torch
import bitsandbytes as bnb

mng = bnb.optim.GlobalOptimManager.get_instance()

model = MyModel()
mng.register_parameters(model.parameters()) # 1. register parameters while still on CPU

model = model.cuda()
# use 8-bit optimizer states for all parameters
adam = bnb.optim.Adam(model.parameters(), lr=0.001, optim_bits=8)

# 2. override: the parameter model.fc1.weight now uses 32-bit Adam
mng.override_config(model.fc1.weight, 'optim_bits', 32)
< > Update on GitHub

←Papers, resources & how to cite AdaGrad→