Expand description
Quantization configuration, algebraic modes, and bit-budget calculation.
Structs§
- BitBudget
- Resolved bit-budget and rank geometry for a single quantized linear layer.
- Layer
Shape - Linear layer dimension specification
(in_features, out_features, bias). - Little
BitConfig - Quantization configuration matching upstream
littlebit_config.json. - Little
BitConfig Overrides - Optional CLI or API overrides applied with highest priority over config files.
Enums§
- Bias
Mode - Whether a linear layer includes a bias vector.
- Cast
Policy - Serialization policy for checkpoint float tensors when saving.
- Inference
Mode - Runtime weight representation when loading quantized checkpoints for inference.
- Init
Mode - Weight initialization mode when constructing
crate::LittleBitLinear. - ItqMode
- LittleBit-2 Joint Iterative Quantization (Joint-ITQ) rotation configuration.
- Quant
Func - Binarization function and straight-through gradient estimator.
- Residual
Mode - Residual factorization mode for a
crate::LittleBitLinearlayer. - Train
Mode - Whether parameters inside a quantized layer hold autograd variables or frozen tensors.
- Weight
State - Whether factor weights
U,Vin a loadedLittleBitLinearare already binarized.
Constants§
- SMOOTH_
SIGN_ DEFAULT_ ALPHA - Default sharpness hyperparameter
alphaforQuantFunc::SmoothSign.
Functions§
- compute_
eff_ bits - Compute actual effective bits per weight for a layer geometry.
- estimate_
split_ dim - Estimate the floating-point latent rank
split_dimneeded to matcheff_bit_target. - finalize_
split_ dim - Round
split_float(orsplit_default) down to a multiple of 8 with a minimum floor. - parse_
eff_ bit_ layers - Parse a per-block effective-bit specification such as
"0-7,27:2.0;12:1.5".