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

fix load_in_int8 behaviour for large models #495

Merged
merged 2 commits into from
Dec 26, 2023
Merged

fix load_in_int8 behaviour for large models #495

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3bc3d01
fix load_in_int8 behaviour for large models
eaidova Dec 25, 2023
eecf181
do not compress model twise when compress during export
eaidova Dec 25, 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
8 changes: 6 additions & 2 deletions optimum/intel/openvino/modeling_base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -259,7 +259,7 @@ def _from_transformers(
local_files_only: bool = False,
task: Optional[str] = None,
trust_remote_code: bool = False,
load_in_8bit: bool = False,
load_in_8bit: Optional[bool] = None,
**kwargs,
):
"""
Expand All @@ -283,6 +283,10 @@ def _from_transformers(
save_dir = TemporaryDirectory()
save_dir_path = Path(save_dir.name)

compression_option = None
if load_in_8bit is not None:
compression_option = "int8" if load_in_8bit else "fp32"

main_export(
model_name_or_path=model_id,
output=save_dir_path,
Expand All @@ -294,7 +298,7 @@ def _from_transformers(
local_files_only=local_files_only,
force_download=force_download,
trust_remote_code=trust_remote_code,
int8=load_in_8bit,
compression_option=compression_option,
)

config.save_pretrained(save_dir_path)
Expand Down
7 changes: 5 additions & 2 deletions optimum/intel/openvino/modeling_base_seq2seq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,7 +220,7 @@ def _from_transformers(
task: Optional[str] = None,
use_cache: bool = True,
trust_remote_code: bool = False,
load_in_8bit: bool = False,
load_in_8bit: Optional[bool] = None,
**kwargs,
):
"""
Expand Down Expand Up @@ -251,6 +251,9 @@ def _from_transformers(
if use_cache:
task = task + "-with-past"

compression_option = None
if load_in_8bit is not None:
compression_option = "int8" if load_in_8bit else "fp32"
main_export(
model_name_or_path=model_id,
output=save_dir_path,
Expand All @@ -262,7 +265,7 @@ def _from_transformers(
local_files_only=local_files_only,
force_download=force_download,
trust_remote_code=trust_remote_code,
compression_option="int8" if load_in_8bit else None,
compression_option=compression_option,
)

config.save_pretrained(save_dir_path)
Expand Down
7 changes: 5 additions & 2 deletions optimum/intel/openvino/modeling_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -211,7 +211,7 @@ def _from_transformers(
task: Optional[str] = None,
use_cache: bool = True,
trust_remote_code: bool = False,
load_in_8bit: bool = False,
load_in_8bit: Optional[bool] = None,
**kwargs,
):
if config.model_type.replace("_", "-") not in _SUPPORTED_ARCHITECTURES:
Expand All @@ -228,6 +228,9 @@ def _from_transformers(
if use_cache:
task = task + "-with-past"

compression_option = None
if load_in_8bit is not None:
compression_option = "int8" if load_in_8bit else "fp32"
main_export(
model_name_or_path=model_id,
output=save_dir_path,
Expand All @@ -239,7 +242,7 @@ def _from_transformers(
local_files_only=local_files_only,
force_download=force_download,
trust_remote_code=trust_remote_code,
compression_option="int8" if load_in_8bit else None,
compression_option=compression_option,
)

config.is_decoder = True
Expand Down
8 changes: 6 additions & 2 deletions optimum/intel/openvino/modeling_diffusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -286,13 +286,17 @@ def _from_transformers(
tokenizer: Optional["CLIPTokenizer"] = None,
scheduler: Union["DDIMScheduler", "PNDMScheduler", "LMSDiscreteScheduler"] = None,
feature_extractor: Optional["CLIPFeatureExtractor"] = None,
load_in_8bit: bool = False,
load_in_8bit: Optional[bool] = None,
tokenizer_2: Optional["CLIPTokenizer"] = None,
**kwargs,
):
save_dir = TemporaryDirectory()
save_dir_path = Path(save_dir.name)

compression_option = None
if load_in_8bit is not None:
compression_option = "int8" if load_in_8bit else "fp32"

main_export(
model_name_or_path=model_id,
output=save_dir_path,
Expand All @@ -304,7 +308,7 @@ def _from_transformers(
use_auth_token=use_auth_token,
local_files_only=local_files_only,
force_download=force_download,
int8=load_in_8bit,
compression_option=compression_option,
)

return cls._from_pretrained(
Expand Down
32 changes: 32 additions & 0 deletions tests/openvino/test_quantization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -272,6 +272,38 @@ def test_ovmodel_load_with_uncompressed_weights(self, model_cls, model_type):
_, num_int8, _ = get_num_quantized_nodes(model)
self.assertEqual(0, num_int8)

def test_ovmodel_load_large_model_with_default_compressed_weights(self):
with unittest.mock.patch("transformers.modeling_utils.ModuleUtilsMixin") as model_mixin_patch:
model_mixin_patch.num_parameters.return_value = 2e9
with unittest.mock.patch("openvino.runtime.ie_api.Core.read_model") as core_patch:
with unittest.mock.patch("optimum.exporters.openvino.convert._save_model") as save_model_patch:
_ = OVModelForCausalLM.from_pretrained(
MODEL_NAMES["llama"], export=True, compile=False, use_cache=False
)
saving_params = {
"model": unittest.mock.ANY,
"path": unittest.mock.ANY,
"compression_option": "int8",
"compression_ratio": None,
}
save_model_patch.aasert_called_with(saving_params)

def test_ovmodel_load_large_model_with_uncompressed_weights(self):
with unittest.mock.patch("transformers.modeling_utils.ModuleUtilsMixin") as model_mixin_patch:
model_mixin_patch.num_parameters.return_value = 2e9
with unittest.mock.patch("openvino.runtime.ie_api.Core.read_model") as core_patch:
with unittest.mock.patch("optimum.exporters.openvino.convert._save_model") as save_model_patch:
_ = OVModelForCausalLM.from_pretrained(
MODEL_NAMES["llama"], export=True, load_in_8bit=False, compile=False, use_cache=False
)
saving_params = {
"model": unittest.mock.ANY,
"path": unittest.mock.ANY,
"compression_option": "fp32",
"compression_ratio": None,
}
save_model_patch.aasert_called_with(saving_params)


class OVQuantizerQATest(unittest.TestCase):
SUPPORTED_ARCHITECTURES = (("hf-internal-testing/tiny-random-BertForQuestionAnswering",),)
Expand Down
Loading