ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

LLM: fix abnormal output of fp16 deepspeed autotp (#10558)

Browse source
This commit is contained in:
binbin Deng 2024-03-28 09:35:48 +08:00 committed by GitHub
parent e619142a16
commit 92dfed77be
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 6 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
python/llm/src/ipex_llm/transformers/low_bit_linear.py
View file

@ -702,12 +702,16 @@ class FP16Linear(nn.Linear):
if self.weight_type == 2:
self.weight = self.weight.transpose(0, 1).contiguous()
self.weight_type = 1
return F.linear(x, self.weight, self.bias)
result = F.linear(x, self.weight, self.bias)
else:
if self.weight_type == 1:
self.weight = self.weight.transpose(0, 1).contiguous()
self.weight_type = 2
return torch.ops.torch_ipex.matmul_bias_out(x, self.weight, self.bias)
result = torch.ops.torch_ipex.matmul_bias_out(x, self.weight, self.bias)
if self.mp_group is not None:
from deepspeed import comm as dist
dist.inference_all_reduce(result, group=self.mp_group)
return result
else:
if self.in_len == 4096 and self.weight_type != 3 or \
self.in_len == 11008 and self.weight_type != 1: