FX symbolic_trace: do not test decoder_inputs_embeds

[fxmarty]

As per title, fixes #31574 (comment)

#31574 was meant to support inputs_embeds, not decoder_inputs_embeds. As some encoder-decoder models, we hit some errors (ValueError: You have to specify either decoder_input_ids or decoder_inputs_embeds).

[amyeroberts]

Thanks for fixing!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ydshieh]

Isn't the error happening for input_embed?

/usr/local/lib/python3.10/site-packages/transformers/models/blenderbot_small/modeling_blenderbot_small.py:711: in forward
raise ValueError("You have to specify either input_ids or inputs_embeds")
E ValueError: You have to specify either input_ids or inputs_embeds

(here is BlenderbotSmallEncoder).

[fxmarty]

@ydshieh yes, that's because inputs_embeds was not in the input_names list here

transformers/tests/test_modeling_common.py

Lines 1141 to 1151 in ae9dd02

	if model.config.is_encoder_decoder:
	model.config.use_cache = False # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
	labels = inputs.get("labels", None)
	input_names = [
	"attention_mask",
	"decoder_attention_mask",
	"decoder_input_ids",
	"input_features",
	"input_ids",
	"input_values",
	]

for encoder-decoder models, and is then filtered out

transformers/tests/test_modeling_common.py

Line 1228 in ae9dd02

filtered_inputs = {k: v for (k, v) in inps.items() if k in input_names}

.

[ydshieh]

Thanks!

[amyeroberts]

@fxmarty There's still errors with fx tests happening on main: https://app.circleci.com/pipelines/github/huggingface/transformers/97262/workflows/8822144c-e8ed-4722-a982-72ee876c3eda/jobs/1287254

FAILED tests/models/mobilebert/test_modeling_mobilebert.py::MobileBertModelTest::test_torch_fx - RuntimeError: mat1 and mat2 shapes cannot be multiplied (4x192 and 96x64)
FAILED tests/models/mobilebert/test_modeling_mobilebert.py::MobileBertModelTest::test_torch_fx_output_loss - RuntimeError: mat1 and mat2 shapes cannot be multiplied (4x192 and 96x64)

Could you take a look?