[Eng] Fix issues with generation on MAC with mps enabled

[tongbaojia]

Currently, if export SUNO_ENABLE_MPS=True, the audio generation doesn't work.

To test this:
export SUNO_ENABLE_MPS=True
then do:
python -m bark --text "tony is in town"

This is eventually traced down to generate_fine. And the cause is:
codebook_preds is type int64 from torch.multinomial.
in_buffer is type int32.

On CPU machines, the column slicing works.
On MPS, for some unknown reason, the column slicing doesn't and it just keeps on overwriting the first column!
Converting the type fixes this issue.

+                    codebook_preds = torch.multinomial(probs[rel_start_fill_idx:1024], num_samples=1).reshape(-1)
+                print("check", rel_start_fill_idx, nn)
+                print(in_buffer, in_buffer.shape)
+                codebook_preds = codebook_preds #.to(torch.int32)
                 in_buffer[0, rel_start_fill_idx:, nn] = codebook_preds
+                print("after", in_buffer)


----------


check 0 2
tensor([[[ 409,  993, 1024,  ..., 1024, 1024, 1024],
         [ 764,  956, 1024,  ..., 1024, 1024, 1024],
         [ 499,  942, 1024,  ..., 1024, 1024, 1024],
         ...,
         [1024, 1024, 1024,  ..., 1024, 1024, 1024],
         [1024, 1024, 1024,  ..., 1024, 1024, 1024],
         [1024, 1024, 1024,  ..., 1024, 1024, 1024]]], device='mps:0',
       dtype=torch.int32) torch.Size([1, 1024, 8])
after tensor([[[ 819,  993, 1024,  ..., 1024, 1024, 1024],
         [1000,  956, 1024,  ..., 1024, 1024, 1024],
         [ 345,  942, 1024,  ..., 1024, 1024, 1024],
         ...,
         [ 786, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 786, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 786, 1024, 1024,  ..., 1024, 1024, 1024]]], device='mps:0',
       dtype=torch.int32)
check 0 3
tensor([[[ 819,  993, 1024,  ..., 1024, 1024, 1024],
         [1000,  956, 1024,  ..., 1024, 1024, 1024],
         [ 345,  942, 1024,  ..., 1024, 1024, 1024],
         ...,
         [ 786, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 786, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 786, 1024, 1024,  ..., 1024, 1024, 1024]]], device='mps:0',
       dtype=torch.int32) torch.Size([1, 1024, 8])
after tensor([[[ 961,  993, 1024,  ..., 1024, 1024, 1024],
         [ 945,  956, 1024,  ..., 1024, 1024, 1024],
         [ 399,  942, 1024,  ..., 1024, 1024, 1024],
         ...,
         [1005, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 506, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 302, 1024, 1024,  ..., 1024, 1024, 1024]]], device='mps:0',
       dtype=torch.int32)
check 0 4
tensor([[[ 961,  993, 1024,  ..., 1024, 1024, 1024],
         [ 945,  956, 1024,  ..., 1024, 1024, 1024],
         [ 399,  942, 1024,  ..., 1024, 1024, 1024],
         ...,
         [1005, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 506, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 302, 1024, 1024,  ..., 1024, 1024, 1024]]], device='mps:0',
       dtype=torch.int32) torch.Size([1, 1024, 8])
after tensor([[[ 742,  993, 1024,  ..., 1024, 1024, 1024],
         [ 720,  956, 1024,  ..., 1024, 1024, 1024],
         [1013,  942, 1024,  ..., 1024, 1024, 1024],
         ...,
         [ 435, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 408, 1024, 1024,  ..., 1024, 1024, 1024],
         [ 816, 1024, 1024,  ..., 1024, 1024, 1024]]], device='mps:0',
       dtype=torch.int32)

[tongbaojia]

I get this from current main FYI
On Python 3.10.10, torch==2.0.1, torchaudio==2.0.2

export SUNO_ENABLE_MPS=True
python -m bark --text "tony is in town [laughs] lol"
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:42<00:00,  2.33it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 37/37 [01:19<00:00,  2.15s/it]
Oops, an error occurred: index out of range in self

[gkucsko]

amazing tyty!