Use case where n_audio > 1, n_group > 1

[mindest]

When n_audio > 1, n_group > 1,

whisper/whisper/decoding.py

Lines 733 to 737 in 5979f03

	# repeat text tensors by the group size, for beam search or best-of-n sampling
	tokens = tokens.repeat_interleave(self.n_group, dim=0).to(audio_features.device)
	# call the main sampling loop
	tokens, sum_logprobs, no_speech_probs = self._main_loop(audio_features, tokens)

is it true that we also need the audio_features to be repeated:

audio_features = audio_features.repeat_interleave(self.n_group, dim=0)

otherwise later q, k will have mismatched shapes (q [B * N, T, D], k [B, T, D])?