solow 0.1.2

SOLO

Paper | PyTorch >= 2.3 | torchao >= 0.7.0

Installation

pip install solow

or

pip install git+https://github.com/MTandHJ/SOLO.git

Usage

from solo.adamw import AdamWQ

optimizer = AdamWQ(
    model.parameters(),
    lr = 0.001,
    weight_decay = 0.,
    betas = (0.8, 0.999),
    bits = (4, 2), # (4 bits for signed, 2 bits for unsigned)
    quantile = 0.1,
    block_sizes = (128, 128),
    quantizers = ('de', 'qema'),
    # A tensor whose size is less than `min_quantizable_tensor_size`
    # will be excluded from quantization.
    # For rigorous probing, this value is set to 0 in paper.
    # Assigning a larger value (such as the default of 4096 in torchao) 
    # may yield more stable results.
    min_quantizable_tensor_size = 0
)

• quantizers:
  • none: The orginal 32-bit state.
  • bf16: The BF16 format.
  • de: The dynamic exponent mapping without a stochastic rounding.
  • de-sr: The dynamic exponent mapping with a stochastic rounding.
  • linear: The linear mapping without a stochastic rounding.
  • linear-sr: The linear mapping with a stochastic rounding.
  • qema: The proposed logarithmic quantization.

Reference Code

• pytorch-optimizer: We implemented the low-bit Adafactor and AdaBelief optimiers based on this code.

Citation

@article{xu2025solo,
  title={Pushing the Limits of Low-Bit Optimizers: A Focus on EMA Dynamics},
  author={Xu, Cong and Liang, Wenbin and Yu, Mo and Liu, Anan and Zhang, Ke-Yue and Ma, Lizhuang and Wang, Jianyong and Wang, Jun and Zhang, Wei},
  journal={arXiv preprint arXiv:2505.00347},
  year={2025}
}