MusicGen (#1020)

* Add MusicGen model

* add benchmarks

* change to from_pretrained

* symlinks

* add readme and requirements

* fix readme

* readme

encodec/encodec.py

@@ -1,7 +1,10 @@
 # Copyright © 2024 Apple Inc.
 
+import functools
+import json
 import math
-from dataclasses import dataclass
+from pathlib import Path
+from types import SimpleNamespace
 from typing import List, Optional, Tuple, Union
 
 import mlx.core as mx
@@ -669,3 +672,70 @@ class EncodecModel(nn.Module):
         if padding_mask is not None and padding_mask.shape[1] < audio_values.shape[1]:
             audio_values = audio_values[:, : padding_mask.shape[1]]
         return audio_values
+
+    @classmethod
+    def from_pretrained(cls, path_or_repo: str):
+        from huggingface_hub import snapshot_download
+
+        path = Path(path_or_repo)
+        if not path.exists():
+            path = Path(
+                snapshot_download(
+                    repo_id=path_or_repo,
+                    allow_patterns=["*.json", "*.safetensors", "*.model"],
+                )
+            )
+
+        with open(path / "config.json", "r") as f:
+            config = SimpleNamespace(**json.load(f))
+
+        model = EncodecModel(config)
+        model.load_weights(str(path / "model.safetensors"))
+        processor = functools.partial(
+            preprocess_audio,
+            sampling_rate=config.sampling_rate,
+            chunk_length=model.chunk_length,
+            chunk_stride=model.chunk_stride,
+        )
+        mx.eval(model)
+        return model, processor
+
+
+def preprocess_audio(
+    raw_audio: Union[mx.array, List[mx.array]],
+    sampling_rate: int = 24000,
+    chunk_length: Optional[int] = None,
+    chunk_stride: Optional[int] = None,
+):
+    r"""
+    Prepare inputs for the EnCodec model.
+
+    Args:
+        raw_audio (mx.array or List[mx.array]): The sequence or batch of
+            sequences to be processed.
+        sampling_rate (int): The sampling rate at which the audio waveform
+            should be digitalized.
+        chunk_length (int, optional): The model's chunk length.
+        chunk_stride (int, optional): The model's chunk stride.
+    """
+    if not isinstance(raw_audio, list):
+        raw_audio = [raw_audio]
+
+    raw_audio = [x[..., None] if x.ndim == 1 else x for x in raw_audio]
+
+    max_length = max(array.shape[0] for array in raw_audio)
+    if chunk_length is not None:
+        max_length += chunk_length - (max_length % chunk_stride)
+
+    inputs = []
+    masks = []
+    for x in raw_audio:
+        length = x.shape[0]
+        mask = mx.ones((length,), dtype=mx.bool_)
+        difference = max_length - length
+        if difference > 0:
+            mask = mx.pad(mask, (0, difference))
+            x = mx.pad(x, ((0, difference), (0, 0)))
+        inputs.append(x)
+        masks.append(mask)
+    return mx.stack(inputs), mx.stack(masks)