Spaces:
Running
on
Zero
Running
on
Zero
add dynamic zerogpu duration & catch no audio file errors
Browse files
app.py
CHANGED
|
@@ -3,10 +3,60 @@ import gradio as gr
|
|
| 3 |
from audiosr import super_resolution, build_model
|
| 4 |
import torch
|
| 5 |
import gc # free up memory
|
|
|
|
|
|
|
| 6 |
|
| 7 |
|
| 8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
def inference(audio_file, model_name, guidance_scale, ddim_steps, seed):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
audiosr = build_model(model_name=model_name)
|
| 11 |
|
| 12 |
if torch.cuda.is_available():
|
|
@@ -45,7 +95,7 @@ iface = gr.Interface(
|
|
| 45 |
],
|
| 46 |
outputs=gr.Audio(type="numpy", label="Output Audio"),
|
| 47 |
title="AudioSR",
|
| 48 |
-
description="Audio Super Resolution with AudioSR"
|
| 49 |
)
|
| 50 |
|
| 51 |
-
iface.launch(share=False)
|
|
|
|
| 3 |
from audiosr import super_resolution, build_model
|
| 4 |
import torch
|
| 5 |
import gc # free up memory
|
| 6 |
+
import soundfile as sf # read audio
|
| 7 |
+
import math # For dynamic gpu duration calculation
|
| 8 |
|
| 9 |
|
| 10 |
+
# Estimate a dynamic gpu duration done by a private Benchmarking HuggingFace ZeroGPU (H200) Space on the 16th November 2025 for saving quota
|
| 11 |
+
def get_duration(audio_file, model_name, guidance_scale, ddim_steps, seed):
|
| 12 |
+
if not audio_file:
|
| 13 |
+
return 0
|
| 14 |
+
|
| 15 |
+
try:
|
| 16 |
+
info = sf.info(audio_file)
|
| 17 |
+
audio_duration = info.duration
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
# 1. Base overhead for model loading (using the higher 'speech' model value).
|
| 21 |
+
base_overhead = 24 # seconds
|
| 22 |
+
|
| 23 |
+
# 2. Multipliers for the core ML task.
|
| 24 |
+
# From benchmark: ~11s for 8s audio @ 50 steps.
|
| 25 |
+
# Formula: (8s * C1) + (50 steps * C2) = 11s.
|
| 26 |
+
# We'll estimate C1=1.0 and C2=0.06.
|
| 27 |
+
time_per_audio_second = 1.0
|
| 28 |
+
time_per_ddim_step = 0.06
|
| 29 |
+
|
| 30 |
+
# 3. Calculate the estimated processing time.
|
| 31 |
+
estimated_time = base_overhead + (audio_duration * time_per_audio_second) + (ddim_steps * time_per_ddim_step)
|
| 32 |
+
|
| 33 |
+
# 4. Add a safety buffer to prevent unexpected timeouts.
|
| 34 |
+
safety_buffer = 10
|
| 35 |
+
|
| 36 |
+
calculated_duration = estimated_time + safety_buffer
|
| 37 |
+
|
| 38 |
+
# 5. Apply min/max constraints.
|
| 39 |
+
min_duration = 50 # Must be enough for model load + buffer
|
| 40 |
+
max_duration = 180 # Current ZeroGPU maximum duration
|
| 41 |
+
|
| 42 |
+
final_duration = max(min_duration, min(max_duration, calculated_duration))
|
| 43 |
+
print("FINAL DURATION", final_duration)
|
| 44 |
+
return math.ceil(final_duration)
|
| 45 |
+
|
| 46 |
+
except Exception as e:
|
| 47 |
+
# Fallback to a safe default duration if reading the audio fails.
|
| 48 |
+
print(f"Error in get_duration, using fallback (60): {e}")
|
| 49 |
+
return 60
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
@spaces.GPU(duration=get_duration)
|
| 53 |
def inference(audio_file, model_name, guidance_scale, ddim_steps, seed):
|
| 54 |
+
|
| 55 |
+
if not audio_file:
|
| 56 |
+
print("No audio file provided, skipping inference.")
|
| 57 |
+
raise gr.Error(
|
| 58 |
+
"Please upload an audio file."
|
| 59 |
+
)
|
| 60 |
audiosr = build_model(model_name=model_name)
|
| 61 |
|
| 62 |
if torch.cuda.is_available():
|
|
|
|
| 95 |
],
|
| 96 |
outputs=gr.Audio(type="numpy", label="Output Audio"),
|
| 97 |
title="AudioSR",
|
| 98 |
+
description="Audio Super Resolution with AudioSR. <br> It estimates a dynamic gpu duration done by a private Benchmarking HuggingFace ZeroGPU (H200) Space on the 16th November 2025 for saving quota."
|
| 99 |
)
|
| 100 |
|
| 101 |
+
iface.launch(share=False)
|