fix Parameter dtype in audio models

[ylacombe]

What does this PR do?

tests/pipelines/test_pipelines_automatic_speech_recognition.py::AutomaticSpeechRecognitionPipelineTests::test_wav2vec2_conformer_float16 was failing because masked_spec_embed is initialized as a fp32 tensor.

transformers/src/transformers/models/wav2vec2_conformer/modeling_wav2vec2_conformer.py

Line 1238 in 28a2283

self.masked_spec_embed = nn.Parameter(torch.FloatTensor(config.hidden_size).uniform_())

I thus changed the tensor from FloatTensor to Tensor. Ideally, I would have moved the .uniform_, to the _init_weights, but turns that model.apply doesn't iterate on Parameter.

Slow tests on every modeling tests of the files I've changed all passed, but I might have missed some other tests to run, any ideas @ydshieh ?

cc @ydshieh because of the failing test and @ArthurZucker because we've talked about it

[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.

[ArthurZucker]

LGTM

[ydshieh]

Let me check CI

[ydshieh]

The original failed pipeline test also passes. I will run a whole pipeline test and let's see. For model testing, since you checked every model modified, I also think it's fine and don't see what else should be checked manually.

[sanchit-gandhi]

LGTM - thanks @ylacombe! Fine for me to keep the .uniform() in the init for simplicity as well

[ydshieh]

all good! merge now, thanks a lot.