Make EncodecModel.decode ONNX exportable

[fxmarty]

As per title. This is needed for the ONNX export of Musicgen e.g. for transformers.js

This removes an important warning:

/home/fxmarty/hf_internship/transformers/src/transformers/models/encodec/modeling_encodec.py:121: TracerWarning: Converting a tensor to a Python float might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  ideal_length = (math.ceil(n_frames) - 1) * stride + (kernel_size - padding_total)

where a padding length was previously hard-coded and applied here

transformers/src/transformers/models/encodec/modeling_encodec.py

Lines 149 to 160 in f01e160

	extra_padding = self._get_extra_padding_for_conv1d(hidden_states, kernel_size, stride, padding_total)
	if self.causal:
	# Left padding for causal
	hidden_states = self._pad1d(hidden_states, (padding_total, extra_padding), mode=self.pad_mode)
	else:
	# Asymmetric padding required for odd strides
	padding_right = padding_total // 2
	padding_left = padding_total - padding_right
	hidden_states = self._pad1d(
	hidden_states, (padding_left, padding_right + extra_padding), mode=self.pad_mode
	)

[fxmarty]

Essentially, we need these ops to be on tensors and not on python types (hence the registration of buffers).

The .to(torch.int64) is added because the produced ONNX model is wrong otherwise (try to concat the float ideal_length - length with padding_total, which is illegal

[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.

[xenova]

Can confirm this fixes the issue with the ONNX export (tested with transformers.js)!

[ylacombe]

LGTM!

[ArthurZucker]

LGTM