mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-21 14:52:20 +00:00
b1f9044d6c
fix incorrect output of Qwen2-7B-Instruct-GPTQ-Int4 and Qwen2-7B-Instruct-AWQ ipex kernel provide func like add_bias, so no need add it outside Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
48 lines
1.5 KiB
Python
48 lines
1.5 KiB
Python
from typing import Optional
|
|
import torch
|
|
import torch.nn as nn
|
|
import intel_extension_for_pytorch as ipex
|
|
|
|
|
|
class WQLinear(nn.Module):
|
|
def __init__(
|
|
self, w_bit, group_size, qweight, qzeros, scales, bias: Optional[torch.Tensor]
|
|
):
|
|
super().__init__()
|
|
|
|
if w_bit not in [4]:
|
|
raise NotImplementedError("Only 4-bit are supported for now.")
|
|
|
|
self.in_features = qweight.shape[0]
|
|
self.out_features = qweight.shape[1] * 32 // w_bit
|
|
|
|
self.w_bit = w_bit
|
|
self.group_size = group_size if group_size != -1 else self.in_features
|
|
# quick sanity check (make sure aligment)
|
|
assert self.in_features % self.group_size == 0
|
|
assert self.out_features % (32 // self.w_bit) == 0
|
|
|
|
self.qweight = qweight
|
|
self.qzeros = qzeros
|
|
self.scales = scales
|
|
self.bias = bias
|
|
self.woq_linear = (
|
|
ipex.llm.quantization.IPEXWeightOnlyQuantizedLinear.from_weight(
|
|
self.qweight,
|
|
self.scales,
|
|
self.qzeros,
|
|
self.in_features,
|
|
self.out_features,
|
|
bias=self.bias,
|
|
group_size=self.group_size,
|
|
quant_method=ipex.llm.quantization.QuantMethod.AWQ_GEMM,
|
|
dtype=ipex.llm.quantization.QuantDtype.INT4,
|
|
)
|
|
)
|
|
|
|
@torch.no_grad()
|
|
def forward(self, x):
|
|
out_shape = x.shape[:-1] + (self.out_features,)
|
|
out = self.woq_linear(x.reshape(-1, x.shape[-1]))
|
|
return out.reshape(out_shape)
|