Skip to content

fix: support torchao >= 0.16.0 by importing renamed CamelCase Config classes#13304

Open
s-zx wants to merge 1 commit intohuggingface:mainfrom
s-zx:fix/torchao-0.16-api-rename
Open

fix: support torchao >= 0.16.0 by importing renamed CamelCase Config classes#13304
s-zx wants to merge 1 commit intohuggingface:mainfrom
s-zx:fix/torchao-0.16-api-rename

Conversation

@s-zx
Copy link

@s-zx s-zx commented Mar 22, 2026

What does this PR do?

Fixes ImportError when using TorchAoConfig with torchao >= 0.16.0.

torchao 0.15.0 deprecated the snake_case quantization functions (int4_weight_only, float8_weight_only, etc.) with deprecation warnings, and 0.16.0 removed them entirely, replacing them with CamelCase Config classes (Int4WeightOnlyConfig, Float8WeightOnlyConfig, etc.).

The TorchAoConfig._get_torchao_quant_type_to_method method was unconditionally importing the old snake_case names, causing:

ImportError: cannot import name 'float8_dynamic_activation_float8_weight' from 'torchao.quantization'

Fix: Add a version guard in _get_torchao_quant_type_to_method:

  • torchao >= 0.16.0: import new CamelCase Config classes, aliased to the old snake_case names so the rest of the method remains unchanged
  • torchao < 0.16.0: keep importing the old snake_case functions as before

The aliasing approach keeps the change minimal and self-contained — no other code in the method needs to change.

Fixes #13286

…classes

torchao 0.15.0 deprecated the snake_case quantization functions
(int4_weight_only, float8_weight_only, etc.) with a deprecation warning,
and 0.16.0 removed them entirely, replacing them with CamelCase Config
classes (Int4WeightOnlyConfig, Float8WeightOnlyConfig, etc.).

Add a version guard in TorchAoConfig._get_torchao_quant_type_to_method:
- torchao >= 0.16.0: import new CamelCase Config classes, aliased to the
  old snake_case names so the rest of the method remains unchanged
- torchao < 0.16.0: keep importing the old snake_case functions as before

Fixes huggingface#13286
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

torchao >= 0.16.0 quantization not supported

2 participants