vLLM: set convert_to_half to False by default (#13172)
* init * remove * fix
This commit is contained in:
parent
1576347892
commit
154af7d7f7
2 changed files with 3 additions and 1 deletions
|
|
@ -293,6 +293,7 @@ def convert_vllm(module, qtype, in_features, out_features, mp_group, cur_qtype,
|
|||
mp_group=mp_group,
|
||||
optimize_lm_head=optimize_lm_head,
|
||||
enable_scale_search=enable_scale_search,
|
||||
conver_to_half=False,
|
||||
)
|
||||
return new_linear
|
||||
|
||||
|
|
@ -589,6 +590,7 @@ def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
|
|||
optimize_lm_head=False,
|
||||
act_order=act_order,
|
||||
enable_scale_search=enable_scale_search,
|
||||
conver_to_half=False,
|
||||
)
|
||||
device = module.qweight.data.device
|
||||
invalidInputError(device.type != "meta",
|
||||
|
|
|
|||
|
|
@ -654,7 +654,7 @@ class LowBitLinear(nn.Linear):
|
|||
else:
|
||||
w = self.weight.data
|
||||
|
||||
if use_batch_forward(x_2d, self.weight.qtype, self.out_len):
|
||||
if use_batch_forward(x_2d, self.weight.qtype, self.out_len) and self.conver_to_half:
|
||||
import xe_batch
|
||||
result = xe_batch.batch_forward(x_2d, w, self.qtype)
|
||||
elif not is_training and self.conver_to_half \
|
||||
|
|
|
|||
Loading…
Reference in a new issue