Skip to content
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

Improve WOQ autoround export #1711

Merged
merged 1 commit into from
Apr 3, 2024
Merged

Improve WOQ autoround export #1711

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
62 changes: 48 additions & 14 deletions neural_compressor/model/torch_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -498,8 +498,7 @@ def export_compressed_model(
gptq_config = self.gptq_config if hasattr(self, "gptq_config") else {}

autoround_config = self.autoround_config if hasattr(self, "autoround_config") else {}

if gptq_config or (autoround_config and device == "xpu"):
if gptq_config:
for k, v in weight_config.items():
logger.debug(f"Compressing {k} on device {device}")
if v["dtype"] == "fp32":
Expand Down Expand Up @@ -558,19 +557,54 @@ def export_compressed_model(
)
new_module.pack(int_weight, gptq_scale, gptq_zp, m.bias, gptq_perm)
set_module(self.model, k, new_module)
elif autoround_config and (device == "cpu" or device == "auto"):
from auto_round.export.export_to_itrex.export import pack_model # pylint: disable=E0401
elif autoround_config:
if device == "xpu":
for k, v in weight_config.items():
logger.debug(f"Compressing {k} on device {device}")
if v["dtype"] == "fp32":
continue
else:
dtype = v["dtype"]
num_bits = v["bits"]
group_size = v["group_size"]
scheme = v["scheme"]
m = fetch_module(self.model, k)
autoround_conf = autoround_config[k]
fp32_weight = m.weight.data
autoround_scale = torch.tensor(autoround_conf["scale"], dtype=torch.float32)
autoround_zp = None if scheme == "sym" else torch.tensor(autoround_conf["zero"], dtype=torch.int32)
int_weight = quant_weight_w_scale(fp32_weight, autoround_scale, autoround_zp, group_size)
int_weight = int_weight.type(torch.int32)
new_module = WeightOnlyLinear(
m.in_features,
m.out_features,
num_bits,
group_size,
dtype=dtype,
zp=autoround_zp is not None,
bias=m.bias is not None,
g_idx=None,
compression_dtype=compression_dtype,
compression_dim=compression_dim,
scale_dtype=scale_dtype,
device=device,
use_optimum_format=use_optimum_format,
)
new_module.pack(int_weight, autoround_scale, autoround_zp, m.bias, None)
set_module(self.model, k, new_module)
else:
from auto_round.export.export_to_itrex.export import pack_model # pylint: disable=E0401

self.model = pack_model(
self.model,
weight_config=autoround_config,
enable_full_range=enable_full_range,
compression_dtype=compression_dtype,
compression_dim=compression_dim,
device=device,
use_optimum_format=use_optimum_format,
inplace=True,
)
self.model = pack_model(
self.model,
weight_config=autoround_config,
enable_full_range=enable_full_range,
compression_dtype=compression_dtype,
compression_dim=compression_dim,
device=device,
use_optimum_format=use_optimum_format,
inplace=True,
)
else:
for k, v in weight_config.items():
logger.debug(f"Compressing {k} on device {device}")
Expand Down
Loading