Our unit tests fail with the new transformers 4.44.0 (currently latest).
Platform
CI and reproducible on Mac M3 with PYTORCH_ENABLE_MPS_FALLBACK=1 tox -e py
Sample Code
Output shows errors like this:
================================================================================================ short test summary info ================================================================================================
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_model - RuntimeError: The expanded size of the tensor (14) must match the existing size (22) at non-singleton dimension 2. Target sizes: [4...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_stream_out_model - RuntimeError: The expanded size of the tensor (14) must match the existing size (22) at non-singleton dimension 2. Targe...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_train_model - RuntimeError: The expanded size of the tensor (23) must match the existing size (31) at non-singleton dimension 2. Target sizes: ...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_gen_trained_mpt - RuntimeError: The expanded size of the tensor (23) must match the existing size (31) at non-singleton dimension 2. Target siz...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_train_model_classification_record - RuntimeError: The expanded size of the tensor (23) must match the existing size (31) at non-singleton dimens...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_repetition_penalty_0_works - RuntimeError: The expanded size of the tensor (16) must match the existing size (24) at non-singleton dimension...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_truncate_tokens_0 - RuntimeError: The expanded size of the tensor (16) must match the existing size (24) at non-singleton dimension 2. Targ...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_with_preserve_input_text - RuntimeError: The expanded size of the tensor (16) must match the existing size (24) at non-singleton dimension 2...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_sampling_param_ignored_greedy_decoding - RuntimeError: The expanded size of the tensor (16) must match the existing size (24) at non-singlet...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_with_custom_stop_criteria - RuntimeError: The expanded size of the tensor (16) must match the existing size (24) at non-singleton dimension ...
FAILED tests/modules/text_generation/test_peft_prompt_tuning.py::test_run_exponential_decay_len_penatly_object - RuntimeError: The expanded size of the tensor (16) must match the existing size (24) at non-singleton...
FAILED tests/toolkit/text_generation/test_model_run_utils.py::test_generate_text_func_serialization_json[to_dict-dict-causal_lm_dummy_model] - RuntimeError: The expanded size of the tensor (22) must match the exist...
FAILED tests/toolkit/text_generation/test_model_run_utils.py::test_generate_text_func_serialization_json[to_json-str-causal_lm_dummy_model] - RuntimeError: The expanded size of the tensor (22) must match the existi...
FAILED tests/toolkit/text_generation/test_model_run_utils.py::test_generate_text_func_serialization_json[to_proto-GeneratedTextResult-causal_lm_dummy_model] - RuntimeError: The expanded size of the tensor (22) must...
FAILED tests/toolkit/text_generation/test_model_run_utils.py::test_generate_text_func_preserve_input_causal_lm[causal_lm_dummy_model] - RuntimeError: The expanded size of the tensor (22) must match the existing siz...
=========================================================================== 15 failed, 316 passed, 5 skipped, 23 warnings in 63.65s (0:01:03) ===========================================================================
Expected behavior
We did not expect 4.44 to be a breaking change for this. It's a 6 day old release, so TBD issues/fixes/workarounds.
Observed behavior
See above.
Additional context
Happened to show up on a seemingly unrelated PR and then was reproduced on main.
Describe the bug
Our unit tests fail with the new transformers 4.44.0 (currently latest).
Platform
CI and reproducible on Mac M3 with
PYTORCH_ENABLE_MPS_FALLBACK=1 tox -e pySample Code
Output shows errors like this:
Expected behavior
We did not expect 4.44 to be a breaking change for this. It's a 6 day old release, so TBD issues/fixes/workarounds.
Observed behavior
See above.
Additional context
Happened to show up on a seemingly unrelated PR and then was reproduced on main.