MusicForge commited on
Commit
b3fdaed
·
verified ·
1 Parent(s): 4711ada

Mirror from aufklarer/MAGNeT-Small-30secs-MLX-8bit

Browse files
Files changed (4) hide show
  1. README.md +67 -0
  2. compression_state_dict.bin +3 -0
  3. config.json +34 -0
  4. model.safetensors +3 -0
README.md ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ language:
4
+ - en
5
+ tags:
6
+ - mlx
7
+ - audio
8
+ - music
9
+ - text-to-music
10
+ - magnet
11
+ - quantized
12
+ - int8
13
+ base_model: facebook/magnet-small-30secs
14
+ library_name: mlx
15
+ pipeline_tag: text-to-audio
16
+ ---
17
+
18
+ # MAGNeT-Small-30secs-MLX-8bit
19
+
20
+ - [speech-swift](https://github.com/soniqo/speech-swift) — Apple SDK
21
+ - [soniqo.audio](https://soniqo.audio) — website
22
+ - [blog](https://soniqo.audio/blog) — blog
23
+
24
+ MLX port of Meta's [MAGNeT](https://arxiv.org/abs/2401.04577) — a masked
25
+ parallel text-to-music model — quantized to **INT8** weight-only for
26
+ on-device generation on Apple Silicon. EnCodec 32 kHz audio decoder; T5-base
27
+ text encoder; per-codebook iterative decoding with restricted-context
28
+ self-attention.
29
+
30
+ ## Model
31
+
32
+ | | |
33
+ |---|---|
34
+ | Parameters (LM) | 300M |
35
+ | Quantization | INT8 weight-only, group size 64 |
36
+ | Format | MLX safetensors |
37
+ | Sample rate | 32 kHz mono |
38
+ | Output length | 30 s per generation (fixed) |
39
+ | Decoding steps | 50 total (`[20, 10, 10, 10]` across 4 RVQ codebooks) |
40
+ | Bundle size | 691 MB on disk |
41
+ | Source | [facebook/magnet-small-30secs](https://huggingface.co/facebook/magnet-small-30secs) |
42
+
43
+ ## Performance (Apple Silicon, 30 s audio)
44
+
45
+ | Metric | Value |
46
+ |---|---|
47
+ | Real-time factor (wall / audio) | **0.29** |
48
+ | Peak RSS | 1543 MB |
49
+ | CLAP score (laion/clap-htsat-unfused, 5 prompts) | 0.438 |
50
+
51
+ ## Usage
52
+
53
+ ```python
54
+ from huggingface_hub import snapshot_download
55
+ bundle = snapshot_download("aufklarer/MAGNeT-Small-30secs-MLX-8bit")
56
+ # See https://github.com/soniqo/speech-swift for production usage.
57
+ ```
58
+
59
+ ## Source
60
+
61
+ - Upstream: [facebook/magnet-small-30secs](https://huggingface.co/facebook/magnet-small-30secs)
62
+ - Paper: [MAGNeT (arxiv 2401.04577)](https://arxiv.org/abs/2401.04577)
63
+ - Audiocraft: [facebookresearch/audiocraft](https://github.com/facebookresearch/audiocraft)
64
+
65
+ ## License
66
+
67
+ **CC-BY-NC 4.0** — inherited from upstream MAGNeT weights. Non-commercial use only.
compression_state_dict.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91598c7da3d183eb8e0cc19cbbdc4f64f2d0c53069f9c8aa84185d0e33873c67
3
+ size 236003715
config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_q": 4,
3
+ "card": 2048,
4
+ "dim": 1024,
5
+ "num_heads": 16,
6
+ "num_layers": 24,
7
+ "ffn_dim": 4096,
8
+ "segment_duration": 30,
9
+ "frame_rate": 50,
10
+ "subcodes_context": 5,
11
+ "span_len": 3,
12
+ "sample_rate": 32000,
13
+ "t5_name": "t5-base",
14
+ "t5_dim": 768,
15
+ "encodec_name": "encodec-32khz",
16
+ "quantization": {
17
+ "mode": "mlx_affine",
18
+ "bits": 8,
19
+ "group_size": 64,
20
+ "targets": [
21
+ "self_attn.q_proj.weight",
22
+ "self_attn.k_proj.weight",
23
+ "self_attn.v_proj.weight",
24
+ "self_attn.out_proj.weight",
25
+ "cross_attn.q_proj.weight",
26
+ "cross_attn.k_proj.weight",
27
+ "cross_attn.v_proj.weight",
28
+ "cross_attn.out_proj.weight",
29
+ "linear1.weight",
30
+ "linear2.weight"
31
+ ]
32
+ },
33
+ "format": "int8"
34
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bad102a8df600b9221b891986e8e15255ec2a40143bc92d5a3f265b0bff45052
3
+ size 488815540