github/SHARK-Studio
1
1
Fork 0
mirror of https://github.com/nod-ai/SHARK-Studio.git synced 2026-01-10 06:17:55 -05:00
Code Issues Packages Projects Releases Wiki Activity

Fix brevitas quantize argument (#1633)

Browse Source
This commit is contained in:
jinchen62
2023-07-07 11:30:31 -07:00
committed by GitHub
parent 3a24cff901
commit 47ec7275e6
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
apps/language_models/scripts/vicuna.py
View File

@@ -695,7 +695,7 @@ class ShardedVicuna(SharkLLMBase):
weight_quant_type="asym",
weight_bit_width=weight_bit_width,
weight_param_method="stats",
weight_scale_type="float",
weight_scale_precision="float",
weight_quant_granularity="per_group",
weight_group_size=self.weight_group_size,
quantize_weight_zero_point=False,

4
apps/language_models/src/model_wrappers/vicuna_model.py
View File

@@ -21,7 +21,7 @@ class FirstVicuna(torch.nn.Module):
weight_quant_type="asym",
weight_bit_width=weight_bit_width,
weight_param_method="stats",
weight_scale_type="float",
weight_scale_precision="float",
weight_quant_granularity="per_group",
weight_group_size=weight_group_size,
quantize_weight_zero_point=False,
@@ -62,7 +62,7 @@ class SecondVicuna(torch.nn.Module):
weight_quant_type="asym",
weight_bit_width=weight_bit_width,
weight_param_method="stats",
weight_scale_type="float",
weight_scale_precision="float",
weight_quant_granularity="per_group",
weight_group_size=weight_group_size,
quantize_weight_zero_point=False,