Spaces:
Running
on
L40S
Running
on
L40S
updated
Browse files
app.py
CHANGED
@@ -3,9 +3,10 @@ import torch
|
|
3 |
import gc
|
4 |
import numpy as np
|
5 |
import random
|
6 |
-
|
7 |
import os
|
8 |
os.environ['ELASTIC_LOG_LEVEL'] = 'DEBUG'
|
|
|
9 |
from elastic_models.transformers import MusicgenForConditionalGeneration
|
10 |
|
11 |
def set_seed(seed: int = 42):
|
@@ -24,6 +25,7 @@ def cleanup_gpu():
|
|
24 |
torch.cuda.synchronize()
|
25 |
gc.collect()
|
26 |
|
|
|
27 |
_generator = None
|
28 |
_processor = None
|
29 |
|
@@ -40,8 +42,7 @@ def load_model():
|
|
40 |
|
41 |
print("[MODEL] Loading processor...")
|
42 |
_processor = AutoProcessor.from_pretrained(
|
43 |
-
"facebook/musicgen-large"
|
44 |
-
cache_dir="/mnt/fs/huggingface_cache/"
|
45 |
)
|
46 |
|
47 |
print("[MODEL] Loading model...")
|
@@ -64,7 +65,7 @@ def load_model():
|
|
64 |
)
|
65 |
|
66 |
print("[MODEL] Model initialization completed successfully")
|
67 |
-
|
68 |
return _generator, _processor
|
69 |
|
70 |
def calculate_max_tokens(duration_seconds):
|
@@ -74,7 +75,6 @@ def calculate_max_tokens(duration_seconds):
|
|
74 |
return max_new_tokens
|
75 |
|
76 |
def generate_music(text_prompt, duration=10, guidance_scale=3.0):
|
77 |
-
"""Generate music based on text prompt using pipeline"""
|
78 |
try:
|
79 |
generator, processor = load_model()
|
80 |
|
@@ -84,7 +84,10 @@ def generate_music(text_prompt, duration=10, guidance_scale=3.0):
|
|
84 |
print(f"[GENERATION] Guidance scale: {guidance_scale}")
|
85 |
|
86 |
cleanup_gpu()
|
|
|
|
|
87 |
set_seed(42)
|
|
|
88 |
|
89 |
max_new_tokens = calculate_max_tokens(duration)
|
90 |
|
@@ -112,8 +115,25 @@ def generate_music(text_prompt, duration=10, guidance_scale=3.0):
|
|
112 |
print(f"[GENERATION] Audio shape: {audio_data.shape}")
|
113 |
print(f"[GENERATION] Sample rate: {sample_rate}")
|
114 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
audio_data = audio_data.astype(np.float32)
|
116 |
|
|
|
|
|
|
|
117 |
return sample_rate, audio_data
|
118 |
|
119 |
except Exception as e:
|
@@ -121,7 +141,8 @@ def generate_music(text_prompt, duration=10, guidance_scale=3.0):
|
|
121 |
cleanup_gpu()
|
122 |
return None, None
|
123 |
|
124 |
-
|
|
|
125 |
gr.Markdown("# 🎵 MusicGen Large Music Generator")
|
126 |
gr.Markdown("Generate music from text descriptions using Facebook's MusicGen Large model with elastic compression.")
|
127 |
|
@@ -156,7 +177,9 @@ with gr.Blocks(title="MusicGen Large - Music Generation", theme=gr.themes.Soft()
|
|
156 |
with gr.Column():
|
157 |
audio_output = gr.Audio(
|
158 |
label="Generated Music",
|
159 |
-
type="numpy"
|
|
|
|
|
160 |
)
|
161 |
|
162 |
with gr.Accordion("Tips", open=False):
|
@@ -170,21 +193,23 @@ with gr.Blocks(title="MusicGen Large - Music Generation", theme=gr.themes.Soft()
|
|
170 |
generate_btn.click(
|
171 |
fn=generate_music,
|
172 |
inputs=[text_input, duration, guidance_scale],
|
173 |
-
outputs=audio_output
|
|
|
174 |
)
|
175 |
|
|
|
176 |
gr.Examples(
|
177 |
examples=[
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
],
|
187 |
-
inputs=
|
188 |
label="Example Prompts"
|
189 |
)
|
190 |
|
|
|
3 |
import gc
|
4 |
import numpy as np
|
5 |
import random
|
6 |
+
import tempfile
|
7 |
import os
|
8 |
os.environ['ELASTIC_LOG_LEVEL'] = 'DEBUG'
|
9 |
+
from transformers import AutoProcessor, pipeline
|
10 |
from elastic_models.transformers import MusicgenForConditionalGeneration
|
11 |
|
12 |
def set_seed(seed: int = 42):
|
|
|
25 |
torch.cuda.synchronize()
|
26 |
gc.collect()
|
27 |
|
28 |
+
# Global variables for model caching with thread lock
|
29 |
_generator = None
|
30 |
_processor = None
|
31 |
|
|
|
42 |
|
43 |
print("[MODEL] Loading processor...")
|
44 |
_processor = AutoProcessor.from_pretrained(
|
45 |
+
"facebook/musicgen-large"
|
|
|
46 |
)
|
47 |
|
48 |
print("[MODEL] Loading model...")
|
|
|
65 |
)
|
66 |
|
67 |
print("[MODEL] Model initialization completed successfully")
|
68 |
+
|
69 |
return _generator, _processor
|
70 |
|
71 |
def calculate_max_tokens(duration_seconds):
|
|
|
75 |
return max_new_tokens
|
76 |
|
77 |
def generate_music(text_prompt, duration=10, guidance_scale=3.0):
|
|
|
78 |
try:
|
79 |
generator, processor = load_model()
|
80 |
|
|
|
84 |
print(f"[GENERATION] Guidance scale: {guidance_scale}")
|
85 |
|
86 |
cleanup_gpu()
|
87 |
+
|
88 |
+
import time
|
89 |
set_seed(42)
|
90 |
+
print(f"[GENERATION] Using seed: {42}")
|
91 |
|
92 |
max_new_tokens = calculate_max_tokens(duration)
|
93 |
|
|
|
115 |
print(f"[GENERATION] Audio shape: {audio_data.shape}")
|
116 |
print(f"[GENERATION] Sample rate: {sample_rate}")
|
117 |
|
118 |
+
# Fix audio format for Gradio display
|
119 |
+
if len(audio_data.shape) > 1:
|
120 |
+
# If stereo or multi-channel, take first channel
|
121 |
+
audio_data = audio_data[0] if audio_data.shape[0] < audio_data.shape[1] else audio_data[:, 0]
|
122 |
+
|
123 |
+
# Ensure it's 1D
|
124 |
+
audio_data = audio_data.flatten()
|
125 |
+
|
126 |
+
# Normalize audio to prevent clipping
|
127 |
+
max_val = np.max(np.abs(audio_data))
|
128 |
+
if max_val > 0:
|
129 |
+
audio_data = audio_data / max_val * 0.95 # Scale to 95% to avoid clipping
|
130 |
+
|
131 |
+
# Convert to float32 for Gradio
|
132 |
audio_data = audio_data.astype(np.float32)
|
133 |
|
134 |
+
print(f"[GENERATION] Final audio shape: {audio_data.shape}")
|
135 |
+
print(f"[GENERATION] Audio range: [{np.min(audio_data):.3f}, {np.max(audio_data):.3f}]")
|
136 |
+
|
137 |
return sample_rate, audio_data
|
138 |
|
139 |
except Exception as e:
|
|
|
141 |
cleanup_gpu()
|
142 |
return None, None
|
143 |
|
144 |
+
|
145 |
+
with gr.Blocks(title="MusicGen Large - Music Generation") as demo:
|
146 |
gr.Markdown("# 🎵 MusicGen Large Music Generator")
|
147 |
gr.Markdown("Generate music from text descriptions using Facebook's MusicGen Large model with elastic compression.")
|
148 |
|
|
|
177 |
with gr.Column():
|
178 |
audio_output = gr.Audio(
|
179 |
label="Generated Music",
|
180 |
+
type="numpy",
|
181 |
+
format="wav",
|
182 |
+
interactive=False
|
183 |
)
|
184 |
|
185 |
with gr.Accordion("Tips", open=False):
|
|
|
193 |
generate_btn.click(
|
194 |
fn=generate_music,
|
195 |
inputs=[text_input, duration, guidance_scale],
|
196 |
+
outputs=audio_output,
|
197 |
+
show_progress=True
|
198 |
)
|
199 |
|
200 |
+
# Example prompts - only text prompts now
|
201 |
gr.Examples(
|
202 |
examples=[
|
203 |
+
"A groovy funk bassline with a tight drum beat",
|
204 |
+
"Relaxing acoustic guitar melody",
|
205 |
+
"Electronic dance music with heavy bass",
|
206 |
+
"Classical violin concerto",
|
207 |
+
"Reggae with steel drums and bass",
|
208 |
+
"Rock ballad with electric guitar solo",
|
209 |
+
"Jazz piano improvisation with brushed drums",
|
210 |
+
"Ambient synthwave with retro vibes",
|
211 |
],
|
212 |
+
inputs=text_input,
|
213 |
label="Example Prompts"
|
214 |
)
|
215 |
|