Function dequantize_tensor

dequant.py:15–28

(tensor, dtype=None, dequant_dtype=None)

    return not is_torch_compatible(tensor)

def dequantize_tensor(tensor, dtype=None, dequant_dtype=None):
    qtype = getattr(tensor, "tensor_type", None)
    oshape = getattr(tensor, "tensor_shape", tensor.shape)

    if qtype in TORCH_COMPATIBLE_QTYPES:
        return tensor.to(dtype)
    elif qtype in dequantize_functions:
        dequant_dtype = dtype if dequant_dtype == "target" else dequant_dtype
        return dequantize(tensor.data, qtype, oshape, dtype=dequant_dtype).to(dtype)
    else:
        # this is incredibly slow
        tqdm.write(f"Falling back to numpy dequant for qtype: {getattr(qtype, 'name', repr(qtype))}")
        new = gguf.quants.dequantize(tensor.cpu().numpy(), qtype)
        return torch.from_numpy(new).to(tensor.device, dtype=dtype)

def dequantize(data, qtype, oshape, dtype=None):
    """