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

Jump to bottom

FEAT / Bitsandbytes: Add dequantize API for bitsandbytes quantized models #30806

Merged
merged 11 commits into from
May 15, 2024
Merged

FEAT / Bitsandbytes: Add dequantize API for bitsandbytes quantized models #30806

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
e0c39a9
add method
younesbelkada May 14, 2024
c748425
change method name
younesbelkada May 14, 2024
14d51c2
more comments
younesbelkada May 14, 2024
be7af7c
Apply suggestions from code review
younesbelkada May 15, 2024
1cff84d
fixup
younesbelkada May 15, 2024
ba01b82
add docstrings and fix comment
younesbelkada May 15, 2024
1a4a906
warn users on the de-quantized dtype
younesbelkada May 15, 2024
309581b
Update src/transformers/quantizers/base.py
younesbelkada May 15, 2024
2813fb8
Update src/transformers/integrations/bitsandbytes.py
younesbelkada May 15, 2024
8b904f7
final suggestion - use private method
younesbelkada May 15, 2024
7f17efa
Merge remote-tracking branch 'origin/main' into add-dequant
younesbelkada May 15, 2024
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
23 changes: 21 additions & 2 deletions src/transformers/integrations/bitsandbytes.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -377,7 +377,7 @@ def _create_accelerate_new_hook(old_hook):
return new_hook


def dequantize_and_replace(
def _dequantize_and_replace(
model,
modules_to_not_convert=None,
current_key_name=None,
Expand Down Expand Up @@ -434,7 +434,7 @@ def dequantize_and_replace(
new_module.to(device)
model._modules[name] = new_module
if len(list(module.children())) > 0:
_, has_been_replaced = dequantize_and_replace(
_, has_been_replaced = _dequantize_and_replace(
module,
modules_to_not_convert,
current_key_name,
Expand All @@ -444,3 +444,22 @@ def dequantize_and_replace(
# Remove the last key for recursion
current_key_name.pop(-1)
return model, has_been_replaced
Copy link
Collaborator

@amyeroberts amyeroberts May 15, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

One general comment, if instead you could have a private method _dequantize_and_replace, which handles the recursion, you don't need to return has_been_replaced here. When someone calls dequantize_and_replace, I don't think has_been_replaced is ever used and could be confusing e.g.:

# This is just dequantize_and_replace from before
def _dequantize_and_replace(
    model,
    modules_to_not_convert=None,
    current_key_name=None,
    quantization_config=None,
    has_been_replaced=False,
):
    ...
    return model, has_been_replaced

def dequantize_and_replace(
    model,
    modules_to_not_convert=None,
    current_key_name=None,
    quantization_config=None,
    has_been_replaced=False,
):
    model, has_been_replaced = _dequantize_and_replace(...)
    return model

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

makes sense ! Will do !

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done in 8b904f7 !



def dequantize_and_replace(
model,
modules_to_not_convert=None,
quantization_config=None,
):
model, has_been_replaced = _dequantize_and_replace(
model,
modules_to_not_convert=modules_to_not_convert,
quantization_config=quantization_config,
)

if not has_been_replaced:
logger.warning(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice :)

"For some reason the model has not been properly dequantized. You might see unexpected behavior."
)

return model
2 changes: 1 addition & 1 deletion src/transformers/quantizers/quantizer_bnb_4bit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -316,7 +316,7 @@ def is_trainable(self) -> bool:
def _dequantize(self, model):
from ..integrations import dequantize_and_replace

model, has_been_replaced = dequantize_and_replace(
model = dequantize_and_replace(
model, self.modules_to_not_convert, quantization_config=self.quantization_config
)
return model
2 changes: 1 addition & 1 deletion src/transformers/quantizers/quantizer_bnb_8bit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -285,7 +285,7 @@ def is_trainable(self) -> bool:
def _dequantize(self, model):
from ..integrations import dequantize_and_replace

model, has_been_replaced = dequantize_and_replace(
model = dequantize_and_replace(
model, self.modules_to_not_convert, quantization_config=self.quantization_config
)
return model
Loading