Skip to content

Support fp16 model to weight-only quantization for PyTorch framework #1387

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Nov 16, 2023
Merged

Support fp16 model to weight-only quantization for PyTorch framework #1387

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ae1881b
Support fp16 model to weight-only quantization for PyTorch framework
PenghuiCheng Nov 15, 2023
91d60c9
Add UT
PenghuiCheng Nov 15, 2023
8399d64
update code
PenghuiCheng Nov 15, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions neural_compressor/adaptor/torch_utils/weight_only.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -399,6 +399,9 @@ def rtn_quantize(
for name, m in model.named_modules():
if m.__class__.__name__ not in supported_layers:
continue
orig_dtype = next(m.parameters()).dtype
if orig_dtype != torch.float:
m = m.float()
if name in weight_config: # pragma: no cover
num_bits = weight_config[name]["bits"]
group_size = weight_config[name]["group_size"]
Expand Down Expand Up @@ -466,6 +469,8 @@ def rtn_quantize(
)
q_weight = q_weight.T if group_dim == 0 else q_weight
m.weight.data.copy_(q_weight)
if orig_dtype != torch.float:
m = m.to(orig_dtype)
return model


Expand Down
3 changes: 2 additions & 1 deletion test/quantization/test_weight_only_quantization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@ def test_trace(self):

def test_rtn(self):
fp32_model = copy.deepcopy(self.model)
fp16_model = copy.deepcopy(self.model).to(torch.float16)
model1 = rtn_quantize(fp32_model, num_bits=3, group_size=-1)
self.assertTrue(isinstance(model1.fc1, torch.nn.Linear))
weight_config = {
Expand All @@ -67,7 +68,7 @@ def test_rtn(self):
},
}
model2 = rtn_quantize(fp32_model, weight_config=weight_config)
model2 = rtn_quantize(fp32_model, weight_config=weight_config, return_int=True)
model2 = rtn_quantize(fp16_model, weight_config=weight_config, return_int=True)
self.assertTrue(isinstance(model2.fc1, WeightOnlyLinear))

def test_awq(self):
Expand Down