bitorch_engine.utils.quant_operators.gptq_style_zeros_packing

bitorch_engine.utils.quant_operators.gptq_style_zeros_packing(zeros: Tensor, w_bit: int, out_features: int, group_size: int) → Tensor[source]

Packs the zeros tensor in GPTQ style for efficient storage and computation.

Parameters:

  • zeros (torch.Tensor) – Input tensor containing zeros.

  • w_bit (int) – Number of bits for weight quantization.

  • out_features (int) – Number of output features.

  • group_size (int) – Size of the group for packing.

Returns:

Packed tensor with reduced storage.

Return type:

torch.Tensor