Steven Liu 23b0004330 torch.dtype alias (#14162) ...

* alias

* add tests

* feedback

* Fix failing test_dtype_alias pipeline tests

Cosmos pipelines keep the safety_checker in float32 and require it to be
passed explicitly on reload, mirroring their test_torch_dtype_dict
overrides. PRX skips dtype-dict/save-load tests because the custom
T5GemmaEncoder is not compatible with transformers v5; do the same for
test_dtype_alias.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

2026-07-14 10:08:01 -07:00

..

__init__.py

Add Nunchaku Lite single-file quantization (#14100)

2026-07-14 10:09:08 +05:30

attention.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

cache.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

common.py

torch.dtype alias (#14162)

2026-07-14 10:08:01 -07:00

compile.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

ip_adapter.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

lora.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

memory.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

parallelism.py

add SP support for _flash_3_varlen_hub backend (#13809)

2026-07-08 09:12:29 +05:30

quantization.py

Add Nunchaku Lite single-file quantization (#14100)

2026-07-14 10:09:08 +05:30

single_file.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

training.py

update licensing for others to 2026. (#14128)

2026-07-08 07:40:20 +05:30

utils.py

fix _flash_3_varlen_hub mask handling (#14115)

2026-07-06 10:33:11 +05:30