Merge branch 'longgen' into our_hf2
Browse files- README.md +7 -6
- app.py +215 -87
- audiocraft/models/loaders.py +0 -2
- audiocraft/models/musicgen.py +82 -11
- audiocraft/modules/transformer.py +67 -27
- tests/models/test_musicgen.py +9 -1
- tests/modules/test_rope.py +9 -1
- tests/modules/test_transformer.py +40 -34
README.md
CHANGED
|
@@ -5,7 +5,7 @@ tags:
|
|
| 5 |
- "music generation"
|
| 6 |
- "language models"
|
| 7 |
- "LLMs"
|
| 8 |
-
app_file: "
|
| 9 |
emoji: 🎵
|
| 10 |
colorFrom: white
|
| 11 |
colorTo: blue
|
|
@@ -54,11 +54,12 @@ pip install -e . # or if you cloned the repo locally
|
|
| 54 |
|
| 55 |
## Usage
|
| 56 |
We offer a number of way to interact with MusicGen:
|
| 57 |
-
1.
|
| 58 |
-
2. You can
|
| 59 |
-
3.
|
| 60 |
-
4.
|
| 61 |
-
|
|
|
|
| 62 |
|
| 63 |
## API
|
| 64 |
|
|
|
|
| 5 |
- "music generation"
|
| 6 |
- "language models"
|
| 7 |
- "LLMs"
|
| 8 |
+
app_file: "app.py"
|
| 9 |
emoji: 🎵
|
| 10 |
colorFrom: white
|
| 11 |
colorTo: blue
|
|
|
|
| 54 |
|
| 55 |
## Usage
|
| 56 |
We offer a number of way to interact with MusicGen:
|
| 57 |
+
1. A demo is also available on the [`facebook/MusicGen` HuggingFace Space](https://huggingface.co/spaces/facebook/MusicGen) (huge thanks to all the HF team for their support).
|
| 58 |
+
2. You can run the extended demo on a Colab: [colab notebook](https://colab.research.google.com/drive/1fxGqfg96RBUvGxZ1XXN07s3DthrKUl4-?usp=sharing).
|
| 59 |
+
3. You can use the gradio demo locally by running `python app.py`.
|
| 60 |
+
4. You can play with MusicGen by running the jupyter notebook at [`demo.ipynb`](./demo.ipynb) locally (if you have a GPU).
|
| 61 |
+
5. Finally, checkout [@camenduru Colab page](https://github.com/camenduru/MusicGen-colab) which is regularly
|
| 62 |
+
updated with contributions from @camenduru and the community.
|
| 63 |
|
| 64 |
## API
|
| 65 |
|
app.py
CHANGED
|
@@ -1,70 +1,139 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
All rights reserved.
|
| 4 |
|
| 5 |
-
This source code is licensed under the license found in the
|
| 6 |
-
LICENSE file in the root directory of this source tree.
|
| 7 |
-
|
|
|
|
|
|
|
| 8 |
|
| 9 |
-
from tempfile import NamedTemporaryFile
|
| 10 |
import argparse
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
import torch
|
| 12 |
import gradio as gr
|
| 13 |
-
|
| 14 |
-
from audiocraft.
|
| 15 |
from audiocraft.data.audio import audio_write
|
|
|
|
| 16 |
|
| 17 |
-
MODEL = None
|
| 18 |
-
IS_SHARED_SPACE = "musicgen/MusicGen" in os.environ.get('SPACE_ID', '')
|
| 19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
-
def load_model(version):
|
| 22 |
-
print("Loading model", version)
|
| 23 |
-
return MusicGen.get_pretrained(version)
|
| 24 |
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
global MODEL
|
| 28 |
-
|
| 29 |
-
if MODEL is None or MODEL.name !=
|
| 30 |
-
MODEL =
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
MODEL.set_generation_params(
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
-
if
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
output = MODEL.generate_with_chroma(
|
| 50 |
-
descriptions=[text],
|
| 51 |
-
melody_wavs=melody,
|
| 52 |
-
melody_sample_rate=sr,
|
| 53 |
-
progress=False
|
| 54 |
)
|
| 55 |
else:
|
| 56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
waveform_video = gr.make_waveform(file.name)
|
| 64 |
-
return waveform_video
|
| 65 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
|
| 67 |
-
|
|
|
|
| 68 |
with gr.Blocks() as interface:
|
| 69 |
gr.Markdown(
|
| 70 |
"""
|
|
@@ -73,14 +142,6 @@ def ui(**kwargs):
|
|
| 73 |
presented at: ["Simple and Controllable Music Generation"](https://huggingface.co/papers/2306.05284)
|
| 74 |
"""
|
| 75 |
)
|
| 76 |
-
if IS_SHARED_SPACE:
|
| 77 |
-
gr.Markdown("""
|
| 78 |
-
⚠ This Space doesn't work in this shared UI ⚠
|
| 79 |
-
|
| 80 |
-
<a href="https://huggingface.co/spaces/musicgen/MusicGen?duplicate=true" style="display: inline-block;margin-top: .5em;margin-right: .25em;" target="_blank">
|
| 81 |
-
<img style="margin-bottom: 0em;display: inline;margin-top: -.25em;" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
|
| 82 |
-
to use it privately, or use the <a href="https://huggingface.co/spaces/facebook/MusicGen">public demo</a>
|
| 83 |
-
""")
|
| 84 |
with gr.Row():
|
| 85 |
with gr.Column():
|
| 86 |
with gr.Row():
|
|
@@ -88,10 +149,12 @@ def ui(**kwargs):
|
|
| 88 |
melody = gr.Audio(source="upload", type="numpy", label="Melody Condition (optional)", interactive=True)
|
| 89 |
with gr.Row():
|
| 90 |
submit = gr.Button("Submit")
|
|
|
|
|
|
|
| 91 |
with gr.Row():
|
| 92 |
model = gr.Radio(["melody", "medium", "small", "large"], label="Model", value="melody", interactive=True)
|
| 93 |
with gr.Row():
|
| 94 |
-
duration = gr.Slider(minimum=1, maximum=
|
| 95 |
with gr.Row():
|
| 96 |
topk = gr.Number(label="Top-k", value=250, interactive=True)
|
| 97 |
topp = gr.Number(label="Top-p", value=0, interactive=True)
|
|
@@ -99,9 +162,9 @@ def ui(**kwargs):
|
|
| 99 |
cfg_coef = gr.Number(label="Classifier Free Guidance", value=3.0, interactive=True)
|
| 100 |
with gr.Column():
|
| 101 |
output = gr.Video(label="Generated Music")
|
| 102 |
-
submit.click(
|
| 103 |
gr.Examples(
|
| 104 |
-
fn=
|
| 105 |
examples=[
|
| 106 |
[
|
| 107 |
"An 80s driving pop song with heavy drums and synth pads in the background",
|
|
@@ -137,7 +200,13 @@ def ui(**kwargs):
|
|
| 137 |
### More details
|
| 138 |
|
| 139 |
The model will generate a short music extract based on the description you provided.
|
| 140 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
|
| 142 |
We present 4 model variations:
|
| 143 |
1. Melody -- a music generation model capable of generating music condition on text and melody inputs. **Note**, you can also use text only.
|
|
@@ -154,27 +223,75 @@ def ui(**kwargs):
|
|
| 154 |
"""
|
| 155 |
)
|
| 156 |
|
| 157 |
-
|
| 158 |
-
launch_kwargs = {}
|
| 159 |
-
username = kwargs.get('username')
|
| 160 |
-
password = kwargs.get('password')
|
| 161 |
-
server_port = kwargs.get('server_port', 0)
|
| 162 |
-
inbrowser = kwargs.get('inbrowser', False)
|
| 163 |
-
share = kwargs.get('share', False)
|
| 164 |
-
server_name = kwargs.get('listen')
|
| 165 |
|
| 166 |
-
launch_kwargs['server_name'] = server_name
|
| 167 |
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
launch_kwargs['inbrowser'] = inbrowser
|
| 174 |
-
if share:
|
| 175 |
-
launch_kwargs['share'] = share
|
| 176 |
|
| 177 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
|
| 179 |
|
| 180 |
if __name__ == "__main__":
|
|
@@ -182,7 +299,11 @@ if __name__ == "__main__":
|
|
| 182 |
parser.add_argument(
|
| 183 |
'--listen',
|
| 184 |
type=str,
|
|
|
|
| 185 |
default='0.0.0.0',
|
|
|
|
|
|
|
|
|
|
| 186 |
help='IP to listen on for connections to Gradio',
|
| 187 |
)
|
| 188 |
parser.add_argument(
|
|
@@ -206,11 +327,18 @@ if __name__ == "__main__":
|
|
| 206 |
|
| 207 |
args = parser.parse_args()
|
| 208 |
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
server_port=args.server_port
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
| 2 |
+
# All rights reserved.
|
|
|
|
| 3 |
|
| 4 |
+
# This source code is licensed under the license found in the
|
| 5 |
+
# LICENSE file in the root directory of this source tree.
|
| 6 |
+
|
| 7 |
+
# Updated to account for UI changes from https://github.com/rkfg/audiocraft/blob/long/app.py
|
| 8 |
+
# also released under the MIT license.
|
| 9 |
|
|
|
|
| 10 |
import argparse
|
| 11 |
+
from concurrent.futures import ProcessPoolExecutor
|
| 12 |
+
import os
|
| 13 |
+
import subprocess as sp
|
| 14 |
+
from tempfile import NamedTemporaryFile
|
| 15 |
+
import time
|
| 16 |
+
import warnings
|
| 17 |
+
|
| 18 |
import torch
|
| 19 |
import gradio as gr
|
| 20 |
+
|
| 21 |
+
from audiocraft.data.audio_utils import convert_audio
|
| 22 |
from audiocraft.data.audio import audio_write
|
| 23 |
+
from audiocraft.models import MusicGen
|
| 24 |
|
|
|
|
|
|
|
| 25 |
|
| 26 |
+
MODEL = None # Last used model
|
| 27 |
+
IS_BATCHED = "facebook/MusicGen" in os.environ.get('SPACE_ID', '')
|
| 28 |
+
MAX_BATCH_SIZE = 12
|
| 29 |
+
BATCHED_DURATION = 15
|
| 30 |
+
INTERRUPTING = False
|
| 31 |
+
# We have to wrap subprocess call to clean a bit the log when using gr.make_waveform
|
| 32 |
+
_old_call = sp.call
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
def _call_nostderr(*args, **kwargs):
|
| 36 |
+
# Avoid ffmpeg vomitting on the logs.
|
| 37 |
+
kwargs['stderr'] = sp.DEVNULL
|
| 38 |
+
kwargs['stdout'] = sp.DEVNULL
|
| 39 |
+
_old_call(*args, **kwargs)
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
sp.call = _call_nostderr
|
| 43 |
+
# Preallocating the pool of processes.
|
| 44 |
+
pool = ProcessPoolExecutor(4)
|
| 45 |
+
pool.__enter__()
|
| 46 |
|
|
|
|
|
|
|
|
|
|
| 47 |
|
| 48 |
+
def interrupt():
|
| 49 |
+
global INTERRUPTING
|
| 50 |
+
INTERRUPTING = True
|
| 51 |
|
| 52 |
+
def make_waveform(*args, **kwargs):
|
| 53 |
+
# Further remove some warnings.
|
| 54 |
+
be = time.time()
|
| 55 |
+
with warnings.catch_warnings():
|
| 56 |
+
warnings.simplefilter('ignore')
|
| 57 |
+
out = gr.make_waveform(*args, **kwargs)
|
| 58 |
+
print("Make a video took", time.time() - be)
|
| 59 |
+
return out
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def load_model(version='melody'):
|
| 63 |
global MODEL
|
| 64 |
+
print("Loading model", version)
|
| 65 |
+
if MODEL is None or MODEL.name != version:
|
| 66 |
+
MODEL = MusicGen.get_pretrained(version)
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def _do_predictions(texts, melodies, duration, **gen_kwargs):
|
| 70 |
+
MODEL.set_generation_params(duration=duration, **gen_kwargs)
|
| 71 |
+
print("new batch", len(texts), texts, [None if m is None else (m[0], m[1].shape) for m in melodies])
|
| 72 |
+
be = time.time()
|
| 73 |
+
processed_melodies = []
|
| 74 |
+
target_sr = 32000
|
| 75 |
+
target_ac = 1
|
| 76 |
+
for melody in melodies:
|
| 77 |
+
if melody is None:
|
| 78 |
+
processed_melodies.append(None)
|
| 79 |
+
else:
|
| 80 |
+
sr, melody = melody[0], torch.from_numpy(melody[1]).to(MODEL.device).float().t()
|
| 81 |
+
if melody.dim() == 1:
|
| 82 |
+
melody = melody[None]
|
| 83 |
+
melody = melody[..., :int(sr * duration)]
|
| 84 |
+
melody = convert_audio(melody, sr, target_sr, target_ac)
|
| 85 |
+
processed_melodies.append(melody)
|
| 86 |
|
| 87 |
+
if any(m is not None for m in processed_melodies):
|
| 88 |
+
outputs = MODEL.generate_with_chroma(
|
| 89 |
+
descriptions=texts,
|
| 90 |
+
melody_wavs=processed_melodies,
|
| 91 |
+
melody_sample_rate=target_sr,
|
| 92 |
+
progress=True
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 93 |
)
|
| 94 |
else:
|
| 95 |
+
outputs = MODEL.generate(texts, progress=True)
|
| 96 |
+
|
| 97 |
+
outputs = outputs.detach().cpu().float()
|
| 98 |
+
out_files = []
|
| 99 |
+
for output in outputs:
|
| 100 |
+
with NamedTemporaryFile("wb", suffix=".wav", delete=False) as file:
|
| 101 |
+
audio_write(
|
| 102 |
+
file.name, output, MODEL.sample_rate, strategy="loudness",
|
| 103 |
+
loudness_headroom_db=16, loudness_compressor=True, add_suffix=False)
|
| 104 |
+
out_files.append(pool.submit(make_waveform, file.name))
|
| 105 |
+
res = [out_file.result() for out_file in out_files]
|
| 106 |
+
print("batch finished", len(texts), time.time() - be)
|
| 107 |
+
return res
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
def predict_batched(texts, melodies):
|
| 111 |
+
max_text_length = 512
|
| 112 |
+
texts = [text[:max_text_length] for text in texts]
|
| 113 |
+
load_model('melody')
|
| 114 |
+
res = _do_predictions(texts, melodies, BATCHED_DURATION)
|
| 115 |
+
return [res]
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
def predict_full(model, text, melody, duration, topk, topp, temperature, cfg_coef, progress=gr.Progress()):
|
| 119 |
+
global INTERRUPTING
|
| 120 |
+
INTERRUPTING = False
|
| 121 |
+
topk = int(topk)
|
| 122 |
+
load_model(model)
|
| 123 |
|
| 124 |
+
def _progress(generated, to_generate):
|
| 125 |
+
progress((generated, to_generate))
|
| 126 |
+
if INTERRUPTING:
|
| 127 |
+
raise gr.Error("Interrupted.")
|
| 128 |
+
MODEL.set_custom_progress_callback(_progress)
|
|
|
|
|
|
|
| 129 |
|
| 130 |
+
outs = _do_predictions(
|
| 131 |
+
[text], [melody], duration,
|
| 132 |
+
top_k=topk, top_p=topp, temperature=temperature, cfg_coef=cfg_coef)
|
| 133 |
+
return outs[0]
|
| 134 |
|
| 135 |
+
|
| 136 |
+
def ui_full(launch_kwargs):
|
| 137 |
with gr.Blocks() as interface:
|
| 138 |
gr.Markdown(
|
| 139 |
"""
|
|
|
|
| 142 |
presented at: ["Simple and Controllable Music Generation"](https://huggingface.co/papers/2306.05284)
|
| 143 |
"""
|
| 144 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
with gr.Row():
|
| 146 |
with gr.Column():
|
| 147 |
with gr.Row():
|
|
|
|
| 149 |
melody = gr.Audio(source="upload", type="numpy", label="Melody Condition (optional)", interactive=True)
|
| 150 |
with gr.Row():
|
| 151 |
submit = gr.Button("Submit")
|
| 152 |
+
# Adapted from https://github.com/rkfg/audiocraft/blob/long/app.py, MIT license.
|
| 153 |
+
_ = gr.Button("Interrupt").click(fn=interrupt, queue=False)
|
| 154 |
with gr.Row():
|
| 155 |
model = gr.Radio(["melody", "medium", "small", "large"], label="Model", value="melody", interactive=True)
|
| 156 |
with gr.Row():
|
| 157 |
+
duration = gr.Slider(minimum=1, maximum=120, value=10, label="Duration", interactive=True)
|
| 158 |
with gr.Row():
|
| 159 |
topk = gr.Number(label="Top-k", value=250, interactive=True)
|
| 160 |
topp = gr.Number(label="Top-p", value=0, interactive=True)
|
|
|
|
| 162 |
cfg_coef = gr.Number(label="Classifier Free Guidance", value=3.0, interactive=True)
|
| 163 |
with gr.Column():
|
| 164 |
output = gr.Video(label="Generated Music")
|
| 165 |
+
submit.click(predict_full, inputs=[model, text, melody, duration, topk, topp, temperature, cfg_coef], outputs=[output])
|
| 166 |
gr.Examples(
|
| 167 |
+
fn=predict_full,
|
| 168 |
examples=[
|
| 169 |
[
|
| 170 |
"An 80s driving pop song with heavy drums and synth pads in the background",
|
|
|
|
| 200 |
### More details
|
| 201 |
|
| 202 |
The model will generate a short music extract based on the description you provided.
|
| 203 |
+
The model can generate up to 30 seconds of audio in one pass. It is now possible
|
| 204 |
+
to extend the generation by feeding back the end of the previous chunk of audio.
|
| 205 |
+
This can take a long time, and the model might lose consistency. The model might also
|
| 206 |
+
decide at arbitrary positions that the song ends.
|
| 207 |
+
|
| 208 |
+
**WARNING:** Choosing long durations will take a long time to generate (2min might take ~10min). An overlap of 12 seconds
|
| 209 |
+
is kept with the previously generated chunk, and 18 "new" seconds are generated each time.
|
| 210 |
|
| 211 |
We present 4 model variations:
|
| 212 |
1. Melody -- a music generation model capable of generating music condition on text and melody inputs. **Note**, you can also use text only.
|
|
|
|
| 223 |
"""
|
| 224 |
)
|
| 225 |
|
| 226 |
+
interface.queue().launch(**launch_kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 227 |
|
|
|
|
| 228 |
|
| 229 |
+
def ui_batched(launch_kwargs):
|
| 230 |
+
with gr.Blocks() as demo:
|
| 231 |
+
gr.Markdown(
|
| 232 |
+
"""
|
| 233 |
+
# MusicGen
|
|
|
|
|
|
|
|
|
|
| 234 |
|
| 235 |
+
This is the demo for [MusicGen](https://github.com/facebookresearch/audiocraft), a simple and controllable model for music generation
|
| 236 |
+
presented at: ["Simple and Controllable Music Generation"](https://huggingface.co/papers/2306.05284).
|
| 237 |
+
<br/>
|
| 238 |
+
<a href="https://huggingface.co/spaces/facebook/MusicGen?duplicate=true" style="display: inline-block;margin-top: .5em;margin-right: .25em;" target="_blank">
|
| 239 |
+
<img style="margin-bottom: 0em;display: inline;margin-top: -.25em;" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
|
| 240 |
+
for longer sequences, more control and no queue.</p>
|
| 241 |
+
"""
|
| 242 |
+
)
|
| 243 |
+
with gr.Row():
|
| 244 |
+
with gr.Column():
|
| 245 |
+
with gr.Row():
|
| 246 |
+
text = gr.Text(label="Describe your music", lines=2, interactive=True)
|
| 247 |
+
melody = gr.Audio(source="upload", type="numpy", label="Condition on a melody (optional)", interactive=True)
|
| 248 |
+
with gr.Row():
|
| 249 |
+
submit = gr.Button("Generate")
|
| 250 |
+
with gr.Column():
|
| 251 |
+
output = gr.Video(label="Generated Music")
|
| 252 |
+
submit.click(predict_batched, inputs=[text, melody], outputs=[output], batch=True, max_batch_size=MAX_BATCH_SIZE)
|
| 253 |
+
gr.Examples(
|
| 254 |
+
fn=predict_batched,
|
| 255 |
+
examples=[
|
| 256 |
+
[
|
| 257 |
+
"An 80s driving pop song with heavy drums and synth pads in the background",
|
| 258 |
+
"./assets/bach.mp3",
|
| 259 |
+
],
|
| 260 |
+
[
|
| 261 |
+
"A cheerful country song with acoustic guitars",
|
| 262 |
+
"./assets/bolero_ravel.mp3",
|
| 263 |
+
],
|
| 264 |
+
[
|
| 265 |
+
"90s rock song with electric guitar and heavy drums",
|
| 266 |
+
None,
|
| 267 |
+
],
|
| 268 |
+
[
|
| 269 |
+
"a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions bpm: 130",
|
| 270 |
+
"./assets/bach.mp3",
|
| 271 |
+
],
|
| 272 |
+
[
|
| 273 |
+
"lofi slow bpm electro chill with organic samples",
|
| 274 |
+
None,
|
| 275 |
+
],
|
| 276 |
+
],
|
| 277 |
+
inputs=[text, melody],
|
| 278 |
+
outputs=[output]
|
| 279 |
+
)
|
| 280 |
+
gr.Markdown("""
|
| 281 |
+
### More details
|
| 282 |
+
|
| 283 |
+
The model will generate 12 seconds of audio based on the description you provided.
|
| 284 |
+
You can optionaly provide a reference audio from which a broad melody will be extracted.
|
| 285 |
+
The model will then try to follow both the description and melody provided.
|
| 286 |
+
All samples are generated with the `melody` model.
|
| 287 |
+
|
| 288 |
+
You can also use your own GPU or a Google Colab by following the instructions on our repo.
|
| 289 |
+
|
| 290 |
+
See [github.com/facebookresearch/audiocraft](https://github.com/facebookresearch/audiocraft)
|
| 291 |
+
for more details.
|
| 292 |
+
""")
|
| 293 |
+
|
| 294 |
+
demo.queue(max_size=8 * 4).launch(**launch_kwargs)
|
| 295 |
|
| 296 |
|
| 297 |
if __name__ == "__main__":
|
|
|
|
| 299 |
parser.add_argument(
|
| 300 |
'--listen',
|
| 301 |
type=str,
|
| 302 |
+
<<<<<<< HEAD
|
| 303 |
default='0.0.0.0',
|
| 304 |
+
=======
|
| 305 |
+
default='0.0.0.0' if 'SPACE_ID' in os.environ else '127.0.0.1',
|
| 306 |
+
>>>>>>> longgen
|
| 307 |
help='IP to listen on for connections to Gradio',
|
| 308 |
)
|
| 309 |
parser.add_argument(
|
|
|
|
| 327 |
|
| 328 |
args = parser.parse_args()
|
| 329 |
|
| 330 |
+
launch_kwargs = {}
|
| 331 |
+
if args.username and args.password:
|
| 332 |
+
launch_kwargs['auth'] = (args.username, args.password)
|
| 333 |
+
if args.server_port:
|
| 334 |
+
launch_kwargs['server_port'] = args.server_port
|
| 335 |
+
if args.inbrowser:
|
| 336 |
+
launch_kwargs['inbrowser'] = args.inbrowser
|
| 337 |
+
if args.share:
|
| 338 |
+
launch_kwargs['share'] = args.share
|
| 339 |
+
|
| 340 |
+
# Show the interface
|
| 341 |
+
if IS_BATCHED:
|
| 342 |
+
ui_batched(launch_kwargs)
|
| 343 |
+
else:
|
| 344 |
+
ui_full(launch_kwargs)
|
audiocraft/models/loaders.py
CHANGED
|
@@ -80,8 +80,6 @@ def load_lm_model(file_or_url_or_id: tp.Union[Path, str], device='cpu', cache_di
|
|
| 80 |
cfg = OmegaConf.create(pkg['xp.cfg'])
|
| 81 |
cfg.device = str(device)
|
| 82 |
if cfg.device == 'cpu':
|
| 83 |
-
cfg.transformer_lm.memory_efficient = False
|
| 84 |
-
cfg.transformer_lm.custom = True
|
| 85 |
cfg.dtype = 'float32'
|
| 86 |
else:
|
| 87 |
cfg.dtype = 'float16'
|
|
|
|
| 80 |
cfg = OmegaConf.create(pkg['xp.cfg'])
|
| 81 |
cfg.device = str(device)
|
| 82 |
if cfg.device == 'cpu':
|
|
|
|
|
|
|
| 83 |
cfg.dtype = 'float32'
|
| 84 |
else:
|
| 85 |
cfg.dtype = 'float16'
|
audiocraft/models/musicgen.py
CHANGED
|
@@ -36,13 +36,16 @@ class MusicGen:
|
|
| 36 |
used to map audio to invertible discrete representations.
|
| 37 |
lm (LMModel): Language model over discrete representations.
|
| 38 |
"""
|
| 39 |
-
def __init__(self, name: str, compression_model: CompressionModel, lm: LMModel
|
|
|
|
| 40 |
self.name = name
|
| 41 |
self.compression_model = compression_model
|
| 42 |
self.lm = lm
|
|
|
|
| 43 |
self.device = next(iter(lm.parameters())).device
|
| 44 |
self.generation_params: dict = {}
|
| 45 |
self.set_generation_params(duration=15) # 15 seconds by default
|
|
|
|
| 46 |
if self.device.type == 'cpu':
|
| 47 |
self.autocast = TorchAutocast(enabled=False)
|
| 48 |
else:
|
|
@@ -65,7 +68,7 @@ class MusicGen:
|
|
| 65 |
return self.compression_model.channels
|
| 66 |
|
| 67 |
@staticmethod
|
| 68 |
-
def get_pretrained(name: str = 'melody', device=
|
| 69 |
"""Return pretrained model, we provide four models:
|
| 70 |
- small (300M), text to music, # see: https://huggingface.co/facebook/musicgen-small
|
| 71 |
- medium (1.5B), text to music, # see: https://huggingface.co/facebook/musicgen-medium
|
|
@@ -73,6 +76,12 @@ class MusicGen:
|
|
| 73 |
- large (3.3B), text to music, # see: https://huggingface.co/facebook/musicgen-large
|
| 74 |
"""
|
| 75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
if name == 'debug':
|
| 77 |
# used only for unit tests
|
| 78 |
compression_model = get_debug_compression_model(device)
|
|
@@ -96,7 +105,7 @@ class MusicGen:
|
|
| 96 |
def set_generation_params(self, use_sampling: bool = True, top_k: int = 250,
|
| 97 |
top_p: float = 0.0, temperature: float = 1.0,
|
| 98 |
duration: float = 30.0, cfg_coef: float = 3.0,
|
| 99 |
-
two_step_cfg: bool = False, extend_stride: float =
|
| 100 |
"""Set the generation parameters for MusicGen.
|
| 101 |
|
| 102 |
Args:
|
|
@@ -113,11 +122,10 @@ class MusicGen:
|
|
| 113 |
should we extend the audio each time. Larger values will mean less context is
|
| 114 |
preserved, and shorter value will require extra computations.
|
| 115 |
"""
|
| 116 |
-
|
| 117 |
-
assert extend_stride <= 25, "Keep at least 5 seconds of overlap!"
|
| 118 |
self.extend_stride = extend_stride
|
|
|
|
| 119 |
self.generation_params = {
|
| 120 |
-
'max_gen_len': int(duration * self.frame_rate),
|
| 121 |
'use_sampling': use_sampling,
|
| 122 |
'temp': temperature,
|
| 123 |
'top_k': top_k,
|
|
@@ -126,6 +134,10 @@ class MusicGen:
|
|
| 126 |
'two_step_cfg': two_step_cfg,
|
| 127 |
}
|
| 128 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 129 |
def generate_unconditional(self, num_samples: int, progress: bool = False) -> torch.Tensor:
|
| 130 |
"""Generate samples in an unconditional manner.
|
| 131 |
|
|
@@ -268,20 +280,79 @@ class MusicGen:
|
|
| 268 |
Returns:
|
| 269 |
torch.Tensor: Generated audio, of shape [B, C, T], T is defined by the generation params.
|
| 270 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
| 271 |
def _progress_callback(generated_tokens: int, tokens_to_generate: int):
|
| 272 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 273 |
|
| 274 |
if prompt_tokens is not None:
|
| 275 |
-
assert
|
| 276 |
"Prompt is longer than audio to generate"
|
| 277 |
|
| 278 |
callback = None
|
| 279 |
if progress:
|
| 280 |
callback = _progress_callback
|
| 281 |
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 285 |
|
| 286 |
# generate audio
|
| 287 |
assert gen_tokens.dim() == 3
|
|
|
|
| 36 |
used to map audio to invertible discrete representations.
|
| 37 |
lm (LMModel): Language model over discrete representations.
|
| 38 |
"""
|
| 39 |
+
def __init__(self, name: str, compression_model: CompressionModel, lm: LMModel,
|
| 40 |
+
max_duration: float = 30):
|
| 41 |
self.name = name
|
| 42 |
self.compression_model = compression_model
|
| 43 |
self.lm = lm
|
| 44 |
+
self.max_duration = max_duration
|
| 45 |
self.device = next(iter(lm.parameters())).device
|
| 46 |
self.generation_params: dict = {}
|
| 47 |
self.set_generation_params(duration=15) # 15 seconds by default
|
| 48 |
+
self._progress_callback: tp.Optional[tp.Callable[[int, int], None]] = None
|
| 49 |
if self.device.type == 'cpu':
|
| 50 |
self.autocast = TorchAutocast(enabled=False)
|
| 51 |
else:
|
|
|
|
| 68 |
return self.compression_model.channels
|
| 69 |
|
| 70 |
@staticmethod
|
| 71 |
+
def get_pretrained(name: str = 'melody', device=None):
|
| 72 |
"""Return pretrained model, we provide four models:
|
| 73 |
- small (300M), text to music, # see: https://huggingface.co/facebook/musicgen-small
|
| 74 |
- medium (1.5B), text to music, # see: https://huggingface.co/facebook/musicgen-medium
|
|
|
|
| 76 |
- large (3.3B), text to music, # see: https://huggingface.co/facebook/musicgen-large
|
| 77 |
"""
|
| 78 |
|
| 79 |
+
if device is None:
|
| 80 |
+
if torch.cuda.device_count():
|
| 81 |
+
device = 'cuda'
|
| 82 |
+
else:
|
| 83 |
+
device = 'cpu'
|
| 84 |
+
|
| 85 |
if name == 'debug':
|
| 86 |
# used only for unit tests
|
| 87 |
compression_model = get_debug_compression_model(device)
|
|
|
|
| 105 |
def set_generation_params(self, use_sampling: bool = True, top_k: int = 250,
|
| 106 |
top_p: float = 0.0, temperature: float = 1.0,
|
| 107 |
duration: float = 30.0, cfg_coef: float = 3.0,
|
| 108 |
+
two_step_cfg: bool = False, extend_stride: float = 18):
|
| 109 |
"""Set the generation parameters for MusicGen.
|
| 110 |
|
| 111 |
Args:
|
|
|
|
| 122 |
should we extend the audio each time. Larger values will mean less context is
|
| 123 |
preserved, and shorter value will require extra computations.
|
| 124 |
"""
|
| 125 |
+
assert extend_stride < self.max_duration, "Cannot stride by more than max generation duration."
|
|
|
|
| 126 |
self.extend_stride = extend_stride
|
| 127 |
+
self.duration = duration
|
| 128 |
self.generation_params = {
|
|
|
|
| 129 |
'use_sampling': use_sampling,
|
| 130 |
'temp': temperature,
|
| 131 |
'top_k': top_k,
|
|
|
|
| 134 |
'two_step_cfg': two_step_cfg,
|
| 135 |
}
|
| 136 |
|
| 137 |
+
def set_custom_progress_callback(self, progress_callback: tp.Optional[tp.Callable[[int, int], None]] = None):
|
| 138 |
+
"""Override the default progress callback."""
|
| 139 |
+
self._progress_callback = progress_callback
|
| 140 |
+
|
| 141 |
def generate_unconditional(self, num_samples: int, progress: bool = False) -> torch.Tensor:
|
| 142 |
"""Generate samples in an unconditional manner.
|
| 143 |
|
|
|
|
| 280 |
Returns:
|
| 281 |
torch.Tensor: Generated audio, of shape [B, C, T], T is defined by the generation params.
|
| 282 |
"""
|
| 283 |
+
total_gen_len = int(self.duration * self.frame_rate)
|
| 284 |
+
max_prompt_len = int(min(self.duration, self.max_duration) * self.frame_rate)
|
| 285 |
+
current_gen_offset: int = 0
|
| 286 |
+
|
| 287 |
def _progress_callback(generated_tokens: int, tokens_to_generate: int):
|
| 288 |
+
generated_tokens += current_gen_offset
|
| 289 |
+
if self._progress_callback is not None:
|
| 290 |
+
# Note that total_gen_len might be quite wrong depending on the
|
| 291 |
+
# codebook pattern used, but with delay it is almost accurate.
|
| 292 |
+
self._progress_callback(generated_tokens, total_gen_len)
|
| 293 |
+
else:
|
| 294 |
+
print(f'{generated_tokens: 6d} / {total_gen_len: 6d}', end='\r')
|
| 295 |
|
| 296 |
if prompt_tokens is not None:
|
| 297 |
+
assert max_prompt_len >= prompt_tokens.shape[-1], \
|
| 298 |
"Prompt is longer than audio to generate"
|
| 299 |
|
| 300 |
callback = None
|
| 301 |
if progress:
|
| 302 |
callback = _progress_callback
|
| 303 |
|
| 304 |
+
if self.duration <= self.max_duration:
|
| 305 |
+
# generate by sampling from LM, simple case.
|
| 306 |
+
with self.autocast:
|
| 307 |
+
gen_tokens = self.lm.generate(
|
| 308 |
+
prompt_tokens, attributes,
|
| 309 |
+
callback=callback, max_gen_len=total_gen_len, **self.generation_params)
|
| 310 |
+
|
| 311 |
+
else:
|
| 312 |
+
# now this gets a bit messier, we need to handle prompts,
|
| 313 |
+
# melody conditioning etc.
|
| 314 |
+
ref_wavs = [attr.wav['self_wav'] for attr in attributes]
|
| 315 |
+
all_tokens = []
|
| 316 |
+
if prompt_tokens is None:
|
| 317 |
+
prompt_length = 0
|
| 318 |
+
else:
|
| 319 |
+
all_tokens.append(prompt_tokens)
|
| 320 |
+
prompt_length = prompt_tokens.shape[-1]
|
| 321 |
+
|
| 322 |
+
stride_tokens = int(self.frame_rate * self.extend_stride)
|
| 323 |
+
|
| 324 |
+
while current_gen_offset + prompt_length < total_gen_len:
|
| 325 |
+
time_offset = current_gen_offset / self.frame_rate
|
| 326 |
+
chunk_duration = min(self.duration - time_offset, self.max_duration)
|
| 327 |
+
max_gen_len = int(chunk_duration * self.frame_rate)
|
| 328 |
+
for attr, ref_wav in zip(attributes, ref_wavs):
|
| 329 |
+
wav_length = ref_wav.length.item()
|
| 330 |
+
if wav_length == 0:
|
| 331 |
+
continue
|
| 332 |
+
# We will extend the wav periodically if it not long enough.
|
| 333 |
+
# we have to do it here rather than in conditioners.py as otherwise
|
| 334 |
+
# we wouldn't have the full wav.
|
| 335 |
+
initial_position = int(time_offset * self.sample_rate)
|
| 336 |
+
wav_target_length = int(self.max_duration * self.sample_rate)
|
| 337 |
+
print(initial_position / self.sample_rate, wav_target_length / self.sample_rate)
|
| 338 |
+
positions = torch.arange(initial_position,
|
| 339 |
+
initial_position + wav_target_length, device=self.device)
|
| 340 |
+
attr.wav['self_wav'] = WavCondition(
|
| 341 |
+
ref_wav[0][:, positions % wav_length],
|
| 342 |
+
torch.full_like(ref_wav[1], wav_target_length))
|
| 343 |
+
with self.autocast:
|
| 344 |
+
gen_tokens = self.lm.generate(
|
| 345 |
+
prompt_tokens, attributes,
|
| 346 |
+
callback=callback, max_gen_len=max_gen_len, **self.generation_params)
|
| 347 |
+
if prompt_tokens is None:
|
| 348 |
+
all_tokens.append(gen_tokens)
|
| 349 |
+
else:
|
| 350 |
+
all_tokens.append(gen_tokens[:, :, prompt_tokens.shape[-1]:])
|
| 351 |
+
prompt_tokens = gen_tokens[:, :, stride_tokens:]
|
| 352 |
+
prompt_length = prompt_tokens.shape[-1]
|
| 353 |
+
current_gen_offset += stride_tokens
|
| 354 |
+
|
| 355 |
+
gen_tokens = torch.cat(all_tokens, dim=-1)
|
| 356 |
|
| 357 |
# generate audio
|
| 358 |
assert gen_tokens.dim() == 3
|
audiocraft/modules/transformer.py
CHANGED
|
@@ -25,6 +25,22 @@ from xformers import ops
|
|
| 25 |
from .rope import RotaryEmbedding
|
| 26 |
from .streaming import StreamingModule
|
| 27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
def _is_profiled() -> bool:
|
| 30 |
# Return true if we are currently running with a xformers profiler activated.
|
|
@@ -75,14 +91,22 @@ def create_sin_embedding(positions: torch.Tensor, dim: int, max_period: float =
|
|
| 75 |
|
| 76 |
def expand_repeated_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 77 |
"""torch.repeat_interleave(x, dim=2, repeats=n_rep) from xlformers"""
|
| 78 |
-
bs, slen, n_kv_heads, head_dim = x.shape
|
| 79 |
if n_rep == 1:
|
| 80 |
return x
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
|
| 87 |
|
| 88 |
class LayerScale(nn.Module):
|
|
@@ -210,6 +234,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 210 |
# Return a causal mask, accounting for potentially stored past keys/values
|
| 211 |
# We actually return a bias for the attention score, as this has the same
|
| 212 |
# convention both in the builtin MHA in Pytorch, and Xformers functions.
|
|
|
|
| 213 |
if self.memory_efficient:
|
| 214 |
from xformers.ops import LowerTriangularMask
|
| 215 |
if current_steps == 1:
|
|
@@ -222,7 +247,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 222 |
return LowerTriangularMask()
|
| 223 |
if self._streaming_state:
|
| 224 |
past_keys = self._streaming_state['past_keys']
|
| 225 |
-
past_steps = past_keys.shape[
|
| 226 |
else:
|
| 227 |
past_steps = 0
|
| 228 |
|
|
@@ -239,6 +264,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 239 |
torch.full([], float('-inf'), device=device, dtype=dtype))
|
| 240 |
|
| 241 |
def _complete_kv(self, k, v):
|
|
|
|
| 242 |
if self.cross_attention:
|
| 243 |
# With cross attention we assume all keys and values
|
| 244 |
# are already available, and streaming is with respect
|
|
@@ -247,20 +273,20 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 247 |
# Complete the key/value pair using the streaming state.
|
| 248 |
if self._streaming_state:
|
| 249 |
pk = self._streaming_state['past_keys']
|
| 250 |
-
nk = torch.cat([pk, k], dim=
|
| 251 |
if v is k:
|
| 252 |
nv = nk
|
| 253 |
else:
|
| 254 |
pv = self._streaming_state['past_values']
|
| 255 |
-
nv = torch.cat([pv, v], dim=
|
| 256 |
else:
|
| 257 |
nk = k
|
| 258 |
nv = v
|
| 259 |
|
| 260 |
-
assert nk.shape[
|
| 261 |
offset = 0
|
| 262 |
if self.past_context is not None:
|
| 263 |
-
offset = max(0, nk.shape[
|
| 264 |
if self._is_streaming:
|
| 265 |
self._streaming_state['past_keys'] = nk[:, offset:]
|
| 266 |
if v is not k:
|
|
@@ -271,8 +297,9 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 271 |
self._streaming_state['offset'] = torch.tensor(0)
|
| 272 |
return nk, nv
|
| 273 |
|
| 274 |
-
|
| 275 |
def _apply_rope(self, query: torch.Tensor, key: torch.Tensor):
|
|
|
|
|
|
|
| 276 |
# Apply rope embeddings to query and key tensors.
|
| 277 |
assert self.rope is not None
|
| 278 |
if 'past_keys' in self._streaming_state:
|
|
@@ -293,6 +320,11 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 293 |
assert not is_causal, ("new param added in torch 2.0.1 not supported, "
|
| 294 |
"use the causal args in the constructor.")
|
| 295 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 296 |
dtype = query.dtype
|
| 297 |
if self._is_streaming:
|
| 298 |
assert self.causal or self.cross_attention, \
|
|
@@ -325,8 +357,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 325 |
if self.qk_layer_norm is True:
|
| 326 |
q = self.q_layer_norm(q)
|
| 327 |
k = self.k_layer_norm(k)
|
| 328 |
-
|
| 329 |
-
q, k, v = [rearrange(x, "b t (h d) -> b h t d", h=self.num_heads) for x in [q, k, v]]
|
| 330 |
else:
|
| 331 |
if not _is_profiled():
|
| 332 |
# profiling breaks that propertysomehow.
|
|
@@ -334,7 +365,11 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 334 |
assert value is key, "specialized implementation"
|
| 335 |
projected = nn.functional.linear(query, self.in_proj_weight, self.in_proj_bias)
|
| 336 |
if self.kv_repeat == 1:
|
| 337 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 338 |
q, k, v = ops.unbind(packed, dim=2)
|
| 339 |
else:
|
| 340 |
embed_dim = self.embed_dim
|
|
@@ -345,18 +380,17 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 345 |
end = start + per_head_dim * kv_heads
|
| 346 |
k = projected[:, :, start: end]
|
| 347 |
v = projected[:, :, end:]
|
| 348 |
-
q = rearrange(q, "b t (h d) ->
|
| 349 |
-
k = rearrange(k, "b t (h d) ->
|
| 350 |
-
v = rearrange(v, "b t (h d) ->
|
| 351 |
|
| 352 |
if self.qk_layer_norm is True:
|
| 353 |
assert self.kv_repeat == 1
|
| 354 |
-
q, k = [rearrange(x, "
|
| 355 |
q = self.q_layer_norm(q)
|
| 356 |
k = self.k_layer_norm(k)
|
| 357 |
-
q, k = [rearrange(x, "b t (h d) ->
|
| 358 |
if self.rope:
|
| 359 |
-
assert False, "Not supported for now"
|
| 360 |
q, k = self._apply_rope(q, k)
|
| 361 |
k, v = self._complete_kv(k, v)
|
| 362 |
if self.kv_repeat > 1:
|
|
@@ -366,8 +400,11 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 366 |
q, k, v = [x.float() for x in [q, k, v]]
|
| 367 |
if self.memory_efficient:
|
| 368 |
p = self.dropout if self.training else 0
|
| 369 |
-
|
| 370 |
-
|
|
|
|
|
|
|
|
|
|
| 371 |
else:
|
| 372 |
# We include the dot product as float32, for consistency
|
| 373 |
# with the other implementations that include that step
|
|
@@ -377,18 +414,21 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
| 377 |
# extend a bit the range of operations done in float32,
|
| 378 |
# although this should make no difference.
|
| 379 |
q = q / q.shape[-1] ** 0.5
|
|
|
|
|
|
|
| 380 |
if self._is_streaming and self.safe_streaming and q.device.type == 'cuda':
|
| 381 |
with torch.autocast(device_type=q.device.type, dtype=torch.float32):
|
| 382 |
-
pre_w = torch.einsum("
|
| 383 |
else:
|
| 384 |
-
pre_w = torch.einsum("
|
| 385 |
if attn_mask is not None:
|
| 386 |
pre_w = pre_w + attn_mask
|
| 387 |
w = torch.softmax(pre_w, dim=-1)
|
| 388 |
w = F.dropout(w, self.dropout, training=self.training).to(v)
|
| 389 |
-
|
|
|
|
| 390 |
x = x.to(dtype)
|
| 391 |
-
x = rearrange(x, "
|
| 392 |
x = self.out_proj(x)
|
| 393 |
else:
|
| 394 |
key, value = self._complete_kv(key, value)
|
|
|
|
| 25 |
from .rope import RotaryEmbedding
|
| 26 |
from .streaming import StreamingModule
|
| 27 |
|
| 28 |
+
_efficient_attention_backend: str = 'torch'
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def set_efficient_attention_backend(backend: str = 'torch'):
|
| 32 |
+
# Using torch by default, it seems a bit faster on older P100 GPUs (~20% faster).
|
| 33 |
+
global _efficient_attention_backend
|
| 34 |
+
assert _efficient_attention_backend in ['xformers', 'torch']
|
| 35 |
+
_efficient_attention_backend = backend
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def _get_attention_time_dimension() -> int:
|
| 39 |
+
if _efficient_attention_backend == 'torch':
|
| 40 |
+
return 2
|
| 41 |
+
else:
|
| 42 |
+
return 1
|
| 43 |
+
|
| 44 |
|
| 45 |
def _is_profiled() -> bool:
|
| 46 |
# Return true if we are currently running with a xformers profiler activated.
|
|
|
|
| 91 |
|
| 92 |
def expand_repeated_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 93 |
"""torch.repeat_interleave(x, dim=2, repeats=n_rep) from xlformers"""
|
|
|
|
| 94 |
if n_rep == 1:
|
| 95 |
return x
|
| 96 |
+
if _efficient_attention_backend == 'torch':
|
| 97 |
+
bs, n_kv_heads, slen, head_dim = x.shape
|
| 98 |
+
return (
|
| 99 |
+
x[:, :, None, :, :]
|
| 100 |
+
.expand(bs, n_kv_heads, n_rep, slen, head_dim)
|
| 101 |
+
.reshape(bs, n_kv_heads * n_rep, slen, head_dim)
|
| 102 |
+
)
|
| 103 |
+
else:
|
| 104 |
+
bs, slen, n_kv_heads, head_dim = x.shape
|
| 105 |
+
return (
|
| 106 |
+
x[:, :, :, None, :]
|
| 107 |
+
.expand(bs, slen, n_kv_heads, n_rep, head_dim)
|
| 108 |
+
.reshape(bs, slen, n_kv_heads * n_rep, head_dim)
|
| 109 |
+
)
|
| 110 |
|
| 111 |
|
| 112 |
class LayerScale(nn.Module):
|
|
|
|
| 234 |
# Return a causal mask, accounting for potentially stored past keys/values
|
| 235 |
# We actually return a bias for the attention score, as this has the same
|
| 236 |
# convention both in the builtin MHA in Pytorch, and Xformers functions.
|
| 237 |
+
time_dim = _get_attention_time_dimension()
|
| 238 |
if self.memory_efficient:
|
| 239 |
from xformers.ops import LowerTriangularMask
|
| 240 |
if current_steps == 1:
|
|
|
|
| 247 |
return LowerTriangularMask()
|
| 248 |
if self._streaming_state:
|
| 249 |
past_keys = self._streaming_state['past_keys']
|
| 250 |
+
past_steps = past_keys.shape[time_dim]
|
| 251 |
else:
|
| 252 |
past_steps = 0
|
| 253 |
|
|
|
|
| 264 |
torch.full([], float('-inf'), device=device, dtype=dtype))
|
| 265 |
|
| 266 |
def _complete_kv(self, k, v):
|
| 267 |
+
time_dim = _get_attention_time_dimension()
|
| 268 |
if self.cross_attention:
|
| 269 |
# With cross attention we assume all keys and values
|
| 270 |
# are already available, and streaming is with respect
|
|
|
|
| 273 |
# Complete the key/value pair using the streaming state.
|
| 274 |
if self._streaming_state:
|
| 275 |
pk = self._streaming_state['past_keys']
|
| 276 |
+
nk = torch.cat([pk, k], dim=time_dim)
|
| 277 |
if v is k:
|
| 278 |
nv = nk
|
| 279 |
else:
|
| 280 |
pv = self._streaming_state['past_values']
|
| 281 |
+
nv = torch.cat([pv, v], dim=time_dim)
|
| 282 |
else:
|
| 283 |
nk = k
|
| 284 |
nv = v
|
| 285 |
|
| 286 |
+
assert nk.shape[time_dim] == nv.shape[time_dim]
|
| 287 |
offset = 0
|
| 288 |
if self.past_context is not None:
|
| 289 |
+
offset = max(0, nk.shape[time_dim] - self.past_context)
|
| 290 |
if self._is_streaming:
|
| 291 |
self._streaming_state['past_keys'] = nk[:, offset:]
|
| 292 |
if v is not k:
|
|
|
|
| 297 |
self._streaming_state['offset'] = torch.tensor(0)
|
| 298 |
return nk, nv
|
| 299 |
|
|
|
|
| 300 |
def _apply_rope(self, query: torch.Tensor, key: torch.Tensor):
|
| 301 |
+
# TODO: fix and verify layout.
|
| 302 |
+
assert _efficient_attention_backend == 'xformers', 'Rope not supported with torch attn.'
|
| 303 |
# Apply rope embeddings to query and key tensors.
|
| 304 |
assert self.rope is not None
|
| 305 |
if 'past_keys' in self._streaming_state:
|
|
|
|
| 320 |
assert not is_causal, ("new param added in torch 2.0.1 not supported, "
|
| 321 |
"use the causal args in the constructor.")
|
| 322 |
|
| 323 |
+
time_dim = _get_attention_time_dimension()
|
| 324 |
+
if time_dim == 2:
|
| 325 |
+
layout = "b h t d"
|
| 326 |
+
else:
|
| 327 |
+
layout = "b t h d"
|
| 328 |
dtype = query.dtype
|
| 329 |
if self._is_streaming:
|
| 330 |
assert self.causal or self.cross_attention, \
|
|
|
|
| 357 |
if self.qk_layer_norm is True:
|
| 358 |
q = self.q_layer_norm(q)
|
| 359 |
k = self.k_layer_norm(k)
|
| 360 |
+
q, k, v = [rearrange(x, f"b t (h d) -> {layout}", h=self.num_heads) for x in [q, k, v]]
|
|
|
|
| 361 |
else:
|
| 362 |
if not _is_profiled():
|
| 363 |
# profiling breaks that propertysomehow.
|
|
|
|
| 365 |
assert value is key, "specialized implementation"
|
| 366 |
projected = nn.functional.linear(query, self.in_proj_weight, self.in_proj_bias)
|
| 367 |
if self.kv_repeat == 1:
|
| 368 |
+
if time_dim == 2:
|
| 369 |
+
bound_layout = "b h p t d"
|
| 370 |
+
else:
|
| 371 |
+
bound_layout = "b t p h d"
|
| 372 |
+
packed = rearrange(projected, f"b t (p h d) -> {bound_layout}", p=3, h=self.num_heads)
|
| 373 |
q, k, v = ops.unbind(packed, dim=2)
|
| 374 |
else:
|
| 375 |
embed_dim = self.embed_dim
|
|
|
|
| 380 |
end = start + per_head_dim * kv_heads
|
| 381 |
k = projected[:, :, start: end]
|
| 382 |
v = projected[:, :, end:]
|
| 383 |
+
q = rearrange(q, f"b t (h d) -> {layout}", h=self.num_heads)
|
| 384 |
+
k = rearrange(k, f"b t (h d) -> {layout}", h=kv_heads)
|
| 385 |
+
v = rearrange(v, f"b t (h d) -> {layout}", h=kv_heads)
|
| 386 |
|
| 387 |
if self.qk_layer_norm is True:
|
| 388 |
assert self.kv_repeat == 1
|
| 389 |
+
q, k = [rearrange(x, f"{layout} -> b t (h d)") for x in [q, k]]
|
| 390 |
q = self.q_layer_norm(q)
|
| 391 |
k = self.k_layer_norm(k)
|
| 392 |
+
q, k = [rearrange(x, f"b t (h d) -> {layout}", h=self.num_heads) for x in [q, k]]
|
| 393 |
if self.rope:
|
|
|
|
| 394 |
q, k = self._apply_rope(q, k)
|
| 395 |
k, v = self._complete_kv(k, v)
|
| 396 |
if self.kv_repeat > 1:
|
|
|
|
| 400 |
q, k, v = [x.float() for x in [q, k, v]]
|
| 401 |
if self.memory_efficient:
|
| 402 |
p = self.dropout if self.training else 0
|
| 403 |
+
if _efficient_attention_backend == 'torch':
|
| 404 |
+
x = torch.nn.functional.scaled_dot_product_attention(
|
| 405 |
+
q, k, v, is_causal=attn_mask is not None, dropout_p=p)
|
| 406 |
+
else:
|
| 407 |
+
x = ops.memory_efficient_attention(q, k, v, attn_mask, p=p)
|
| 408 |
else:
|
| 409 |
# We include the dot product as float32, for consistency
|
| 410 |
# with the other implementations that include that step
|
|
|
|
| 414 |
# extend a bit the range of operations done in float32,
|
| 415 |
# although this should make no difference.
|
| 416 |
q = q / q.shape[-1] ** 0.5
|
| 417 |
+
key_layout = layout.replace('t', 'k')
|
| 418 |
+
query_layout = layout
|
| 419 |
if self._is_streaming and self.safe_streaming and q.device.type == 'cuda':
|
| 420 |
with torch.autocast(device_type=q.device.type, dtype=torch.float32):
|
| 421 |
+
pre_w = torch.einsum(f"{query_layout},{key_layout}-> b h t k", q, k)
|
| 422 |
else:
|
| 423 |
+
pre_w = torch.einsum(f"{query_layout},{key_layout}-> b h t k", q, k)
|
| 424 |
if attn_mask is not None:
|
| 425 |
pre_w = pre_w + attn_mask
|
| 426 |
w = torch.softmax(pre_w, dim=-1)
|
| 427 |
w = F.dropout(w, self.dropout, training=self.training).to(v)
|
| 428 |
+
# Key and value have the same format.
|
| 429 |
+
x = torch.einsum(f"b h t k, {key_layout} -> {layout}", w, v)
|
| 430 |
x = x.to(dtype)
|
| 431 |
+
x = rearrange(x, f"{layout} -> b t (h d)", h=self.num_heads)
|
| 432 |
x = self.out_proj(x)
|
| 433 |
else:
|
| 434 |
key, value = self._complete_kv(key, value)
|
tests/models/test_musicgen.py
CHANGED
|
@@ -13,7 +13,7 @@ from audiocraft.models import MusicGen
|
|
| 13 |
class TestSEANetModel:
|
| 14 |
def get_musicgen(self):
|
| 15 |
mg = MusicGen.get_pretrained(name='debug', device='cpu')
|
| 16 |
-
mg.set_generation_params(duration=2.0)
|
| 17 |
return mg
|
| 18 |
|
| 19 |
def test_base(self):
|
|
@@ -48,3 +48,11 @@ class TestSEANetModel:
|
|
| 48 |
wav = mg.generate(
|
| 49 |
['youpi', 'lapin dort'])
|
| 50 |
assert list(wav.shape) == [2, 1, 64000]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
class TestSEANetModel:
|
| 14 |
def get_musicgen(self):
|
| 15 |
mg = MusicGen.get_pretrained(name='debug', device='cpu')
|
| 16 |
+
mg.set_generation_params(duration=2.0, extend_stride=2.)
|
| 17 |
return mg
|
| 18 |
|
| 19 |
def test_base(self):
|
|
|
|
| 48 |
wav = mg.generate(
|
| 49 |
['youpi', 'lapin dort'])
|
| 50 |
assert list(wav.shape) == [2, 1, 64000]
|
| 51 |
+
|
| 52 |
+
def test_generate_long(self):
|
| 53 |
+
mg = self.get_musicgen()
|
| 54 |
+
mg.max_duration = 3.
|
| 55 |
+
mg.set_generation_params(duration=4., extend_stride=2.)
|
| 56 |
+
wav = mg.generate(
|
| 57 |
+
['youpi', 'lapin dort'])
|
| 58 |
+
assert list(wav.shape) == [2, 1, 32000 * 4]
|
tests/modules/test_rope.py
CHANGED
|
@@ -7,10 +7,11 @@
|
|
| 7 |
import torch
|
| 8 |
|
| 9 |
from audiocraft.modules.rope import RotaryEmbedding
|
| 10 |
-
from audiocraft.modules.transformer import StreamingTransformer
|
| 11 |
|
| 12 |
|
| 13 |
def test_rope():
|
|
|
|
| 14 |
B, T, H, C = 8, 75, 16, 128
|
| 15 |
|
| 16 |
rope = RotaryEmbedding(dim=C)
|
|
@@ -23,6 +24,7 @@ def test_rope():
|
|
| 23 |
|
| 24 |
|
| 25 |
def test_rope_io_dtypes():
|
|
|
|
| 26 |
B, T, H, C = 8, 75, 16, 128
|
| 27 |
|
| 28 |
rope_32 = RotaryEmbedding(dim=C, dtype=torch.float32)
|
|
@@ -46,6 +48,7 @@ def test_rope_io_dtypes():
|
|
| 46 |
|
| 47 |
|
| 48 |
def test_transformer_with_rope():
|
|
|
|
| 49 |
torch.manual_seed(1234)
|
| 50 |
for pos in ['rope', 'sin_rope']:
|
| 51 |
tr = StreamingTransformer(
|
|
@@ -61,6 +64,7 @@ def test_transformer_with_rope():
|
|
| 61 |
|
| 62 |
@torch.no_grad()
|
| 63 |
def test_rope_streaming():
|
|
|
|
| 64 |
torch.manual_seed(1234)
|
| 65 |
tr = StreamingTransformer(
|
| 66 |
16, 4, 2, causal=True, dropout=0.,
|
|
@@ -88,6 +92,7 @@ def test_rope_streaming():
|
|
| 88 |
|
| 89 |
@torch.no_grad()
|
| 90 |
def test_rope_streaming_past_context():
|
|
|
|
| 91 |
torch.manual_seed(1234)
|
| 92 |
|
| 93 |
for context in [None, 10]:
|
|
@@ -117,6 +122,7 @@ def test_rope_streaming_past_context():
|
|
| 117 |
|
| 118 |
|
| 119 |
def test_rope_memory_efficient():
|
|
|
|
| 120 |
torch.manual_seed(1234)
|
| 121 |
tr = StreamingTransformer(
|
| 122 |
16, 4, 2, custom=True, dropout=0., layer_scale=0.1,
|
|
@@ -137,6 +143,7 @@ def test_rope_memory_efficient():
|
|
| 137 |
|
| 138 |
|
| 139 |
def test_rope_with_xpos():
|
|
|
|
| 140 |
B, T, H, C = 8, 75, 16, 128
|
| 141 |
|
| 142 |
rope = RotaryEmbedding(dim=C, xpos=True)
|
|
@@ -149,6 +156,7 @@ def test_rope_with_xpos():
|
|
| 149 |
|
| 150 |
|
| 151 |
def test_positional_scale():
|
|
|
|
| 152 |
B, T, H, C = 8, 75, 16, 128
|
| 153 |
|
| 154 |
rope = RotaryEmbedding(dim=C, xpos=True, scale=0.0)
|
|
|
|
| 7 |
import torch
|
| 8 |
|
| 9 |
from audiocraft.modules.rope import RotaryEmbedding
|
| 10 |
+
from audiocraft.modules.transformer import StreamingTransformer, set_efficient_attention_backend
|
| 11 |
|
| 12 |
|
| 13 |
def test_rope():
|
| 14 |
+
set_efficient_attention_backend('xformers')
|
| 15 |
B, T, H, C = 8, 75, 16, 128
|
| 16 |
|
| 17 |
rope = RotaryEmbedding(dim=C)
|
|
|
|
| 24 |
|
| 25 |
|
| 26 |
def test_rope_io_dtypes():
|
| 27 |
+
set_efficient_attention_backend('xformers')
|
| 28 |
B, T, H, C = 8, 75, 16, 128
|
| 29 |
|
| 30 |
rope_32 = RotaryEmbedding(dim=C, dtype=torch.float32)
|
|
|
|
| 48 |
|
| 49 |
|
| 50 |
def test_transformer_with_rope():
|
| 51 |
+
set_efficient_attention_backend('xformers')
|
| 52 |
torch.manual_seed(1234)
|
| 53 |
for pos in ['rope', 'sin_rope']:
|
| 54 |
tr = StreamingTransformer(
|
|
|
|
| 64 |
|
| 65 |
@torch.no_grad()
|
| 66 |
def test_rope_streaming():
|
| 67 |
+
set_efficient_attention_backend('xformers')
|
| 68 |
torch.manual_seed(1234)
|
| 69 |
tr = StreamingTransformer(
|
| 70 |
16, 4, 2, causal=True, dropout=0.,
|
|
|
|
| 92 |
|
| 93 |
@torch.no_grad()
|
| 94 |
def test_rope_streaming_past_context():
|
| 95 |
+
set_efficient_attention_backend('xformers')
|
| 96 |
torch.manual_seed(1234)
|
| 97 |
|
| 98 |
for context in [None, 10]:
|
|
|
|
| 122 |
|
| 123 |
|
| 124 |
def test_rope_memory_efficient():
|
| 125 |
+
set_efficient_attention_backend('xformers')
|
| 126 |
torch.manual_seed(1234)
|
| 127 |
tr = StreamingTransformer(
|
| 128 |
16, 4, 2, custom=True, dropout=0., layer_scale=0.1,
|
|
|
|
| 143 |
|
| 144 |
|
| 145 |
def test_rope_with_xpos():
|
| 146 |
+
set_efficient_attention_backend('xformers')
|
| 147 |
B, T, H, C = 8, 75, 16, 128
|
| 148 |
|
| 149 |
rope = RotaryEmbedding(dim=C, xpos=True)
|
|
|
|
| 156 |
|
| 157 |
|
| 158 |
def test_positional_scale():
|
| 159 |
+
set_efficient_attention_backend('xformers')
|
| 160 |
B, T, H, C = 8, 75, 16, 128
|
| 161 |
|
| 162 |
rope = RotaryEmbedding(dim=C, xpos=True, scale=0.0)
|
tests/modules/test_transformer.py
CHANGED
|
@@ -9,7 +9,8 @@ from itertools import product
|
|
| 9 |
import pytest
|
| 10 |
import torch
|
| 11 |
|
| 12 |
-
from audiocraft.modules.transformer import
|
|
|
|
| 13 |
|
| 14 |
|
| 15 |
def test_transformer_causal_streaming():
|
|
@@ -86,19 +87,22 @@ def test_streaming_api():
|
|
| 86 |
|
| 87 |
def test_memory_efficient():
|
| 88 |
torch.manual_seed(1234)
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
tr_mem_efficient = StreamingTransformer(
|
| 92 |
-
16, 4, 2, dropout=0., memory_efficient=True, layer_scale=0.1)
|
| 93 |
-
tr_mem_efficient.load_state_dict(tr.state_dict())
|
| 94 |
-
tr.eval()
|
| 95 |
-
steps = 12
|
| 96 |
-
x = torch.randn(3, steps, 16)
|
| 97 |
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
|
| 103 |
|
| 104 |
def test_attention_as_float32():
|
|
@@ -129,30 +133,32 @@ def test_attention_as_float32():
|
|
| 129 |
@torch.no_grad()
|
| 130 |
def test_streaming_memory_efficient():
|
| 131 |
torch.manual_seed(1234)
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
16, 4, 2, dropout=0.,
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
|
|
|
|
|
|
| 140 |
|
| 141 |
-
|
| 142 |
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
|
| 157 |
|
| 158 |
def test_cross_attention():
|
|
@@ -204,7 +210,7 @@ def test_cross_attention_compat():
|
|
| 204 |
|
| 205 |
y = cross_attn(queries, keys, values)[0]
|
| 206 |
y_ref = ref_attn(queries, keys, values)[0]
|
| 207 |
-
assert torch.allclose(y, y_ref, atol=1e-7)
|
| 208 |
|
| 209 |
# Now let's check that streaming is working properly.
|
| 210 |
with cross_attn.streaming():
|
|
|
|
| 9 |
import pytest
|
| 10 |
import torch
|
| 11 |
|
| 12 |
+
from audiocraft.modules.transformer import (
|
| 13 |
+
StreamingMultiheadAttention, StreamingTransformer, set_efficient_attention_backend)
|
| 14 |
|
| 15 |
|
| 16 |
def test_transformer_causal_streaming():
|
|
|
|
| 87 |
|
| 88 |
def test_memory_efficient():
|
| 89 |
torch.manual_seed(1234)
|
| 90 |
+
for backend in ['torch', 'xformers']:
|
| 91 |
+
set_efficient_attention_backend(backend)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
|
| 93 |
+
tr = StreamingTransformer(
|
| 94 |
+
16, 4, 2, custom=True, dropout=0., layer_scale=0.1)
|
| 95 |
+
tr_mem_efficient = StreamingTransformer(
|
| 96 |
+
16, 4, 2, dropout=0., memory_efficient=True, layer_scale=0.1)
|
| 97 |
+
tr_mem_efficient.load_state_dict(tr.state_dict())
|
| 98 |
+
tr.eval()
|
| 99 |
+
steps = 12
|
| 100 |
+
x = torch.randn(3, steps, 16)
|
| 101 |
+
|
| 102 |
+
with torch.no_grad():
|
| 103 |
+
y = tr(x)
|
| 104 |
+
y2 = tr_mem_efficient(x)
|
| 105 |
+
assert torch.allclose(y, y2), ((y - y2).norm(), backend)
|
| 106 |
|
| 107 |
|
| 108 |
def test_attention_as_float32():
|
|
|
|
| 133 |
@torch.no_grad()
|
| 134 |
def test_streaming_memory_efficient():
|
| 135 |
torch.manual_seed(1234)
|
| 136 |
+
for backend in ['torch', 'xformers']:
|
| 137 |
+
set_efficient_attention_backend(backend)
|
| 138 |
+
tr = StreamingTransformer(16, 4, 2, causal=True, dropout=0., custom=True)
|
| 139 |
+
tr_mem_efficient = StreamingTransformer(
|
| 140 |
+
16, 4, 2, dropout=0., memory_efficient=True, causal=True)
|
| 141 |
+
tr.load_state_dict(tr_mem_efficient.state_dict())
|
| 142 |
+
tr.eval()
|
| 143 |
+
tr_mem_efficient.eval()
|
| 144 |
+
steps = 12
|
| 145 |
+
x = torch.randn(3, steps, 16)
|
| 146 |
|
| 147 |
+
ref = tr(x)
|
| 148 |
|
| 149 |
+
with tr_mem_efficient.streaming():
|
| 150 |
+
outs = []
|
| 151 |
+
# frame_sizes = [2] + [1] * (steps - 2)
|
| 152 |
+
frame_sizes = [1] * steps
|
| 153 |
|
| 154 |
+
for frame_size in frame_sizes:
|
| 155 |
+
frame = x[:, :frame_size]
|
| 156 |
+
x = x[:, frame_size:]
|
| 157 |
+
outs.append(tr_mem_efficient(frame))
|
| 158 |
|
| 159 |
+
out = torch.cat(outs, dim=1)
|
| 160 |
+
delta = torch.norm(out - ref) / torch.norm(out)
|
| 161 |
+
assert delta < 1e-6, delta
|
| 162 |
|
| 163 |
|
| 164 |
def test_cross_attention():
|
|
|
|
| 210 |
|
| 211 |
y = cross_attn(queries, keys, values)[0]
|
| 212 |
y_ref = ref_attn(queries, keys, values)[0]
|
| 213 |
+
assert torch.allclose(y, y_ref, atol=1e-7), (y - y_ref).norm() / y_ref.norm()
|
| 214 |
|
| 215 |
# Now let's check that streaming is working properly.
|
| 216 |
with cross_attn.streaming():
|