Spaces:
Running
on
Zero
Running
on
Zero
Update gradio_app.py
Browse files- gradio_app.py +63 -61
gradio_app.py
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
import spaces
|
| 2 |
import os
|
| 3 |
import time
|
| 4 |
import threading
|
|
@@ -9,7 +9,7 @@ import io
|
|
| 9 |
import tempfile
|
| 10 |
import numpy as np
|
| 11 |
|
| 12 |
-
#
|
| 13 |
import torch
|
| 14 |
import soundfile as sf
|
| 15 |
from pydub import AudioSegment
|
|
@@ -18,16 +18,13 @@ from fastapi import FastAPI, HTTPException
|
|
| 18 |
from pydantic import BaseModel
|
| 19 |
from vieneu_tts import VieNeuTTS
|
| 20 |
|
| 21 |
-
# --- KHỞI TẠO
|
| 22 |
app = FastAPI()
|
|
|
|
| 23 |
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
# QUAN TRỌNG: Trên ZeroGPU, lúc khởi động PHẢI DÙNG CPU
|
| 28 |
-
# GPU chỉ được kích hoạt bên trong hàm @spaces.GPU
|
| 29 |
-
device = "cpu"
|
| 30 |
-
print(f"🖥️ Thiết bị khởi động (Global): {device.upper()} (Sẽ chuyển sang CUDA khi chạy)")
|
| 31 |
|
| 32 |
# Cache
|
| 33 |
CACHE_DIR = "./reference_cache"
|
|
@@ -53,21 +50,30 @@ def save_cache_to_disk(cache_key, ref_codes):
|
|
| 53 |
with open(cache_path, 'wb') as f: pickle.dump(ref_codes, f)
|
| 54 |
except Exception: pass
|
| 55 |
|
| 56 |
-
#
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
VOICE_SAMPLES = {
|
| 72 |
"Tuyên (nam miền Bắc)": {"audio": "./sample/Tuyên (nam miền Bắc).wav", "text": "./sample/Tuyên (nam miền Bắc).txt"},
|
| 73 |
"Vĩnh (nam miền Nam)": {"audio": "./sample/Vĩnh (nam miền Nam).wav", "text": "./sample/Vĩnh (nam miền Nam).txt"},
|
|
@@ -81,31 +87,22 @@ VOICE_SAMPLES = {
|
|
| 81 |
"Nhỏ Ngọt Ngào": {"audio": "./sample/Nhỏ Ngọt Ngào.wav", "text": "./sample/Nhỏ Ngọt Ngào.txt"},
|
| 82 |
}
|
| 83 |
|
| 84 |
-
# ---
|
| 85 |
|
| 86 |
-
|
| 87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
if torch.cuda.is_available():
|
| 89 |
-
# Kiểm tra xem model đã ở trên GPU chưa để tránh move thừa
|
| 90 |
-
# VieNeuTTS lưu model trong self.backbone và self.codec
|
| 91 |
try:
|
| 92 |
-
# Move backbone
|
| 93 |
if next(tts.backbone.parameters()).device.type != 'cuda':
|
| 94 |
-
print(" 🚀 Moving model to GPU...")
|
| 95 |
tts.backbone.to("cuda")
|
| 96 |
-
|
| 97 |
-
# Move codec
|
| 98 |
-
if next(tts.codec.parameters()).device.type != 'cuda':
|
| 99 |
tts.codec.to("cuda")
|
| 100 |
-
except
|
| 101 |
-
print(f"⚠️ Lỗi khi move model sang GPU: {e}")
|
| 102 |
|
| 103 |
-
|
| 104 |
-
def core_synthesize(text, voice_choice, speed_factor):
|
| 105 |
-
# 1. Đẩy model sang GPU (Chỉ làm việc này bên trong hàm @spaces.GPU)
|
| 106 |
-
move_model_to_cuda()
|
| 107 |
-
|
| 108 |
-
# 2. Lấy thông tin giọng
|
| 109 |
voice_info = VOICE_SAMPLES.get(voice_choice)
|
| 110 |
if not voice_info:
|
| 111 |
raise ValueError("Giọng không tồn tại")
|
|
@@ -116,38 +113,38 @@ def core_synthesize(text, voice_choice, speed_factor):
|
|
| 116 |
with open(ref_text_path, "r", encoding="utf-8") as f:
|
| 117 |
ref_text_raw = f.read()
|
| 118 |
|
| 119 |
-
#
|
| 120 |
cache_key = f"preset:{voice_choice}"
|
| 121 |
with reference_cache_lock:
|
| 122 |
if cache_key in reference_cache:
|
| 123 |
ref_codes = reference_cache[cache_key]
|
| 124 |
-
|
| 125 |
-
if isinstance(ref_codes, torch.Tensor):
|
| 126 |
ref_codes = ref_codes.to("cuda")
|
| 127 |
else:
|
| 128 |
ref_codes = load_cache_from_disk(cache_key)
|
| 129 |
if ref_codes is None:
|
| 130 |
-
ref_codes = tts.encode_reference(ref_audio_path)
|
| 131 |
-
#
|
| 132 |
save_cache_to_disk(cache_key, ref_codes.cpu() if isinstance(ref_codes, torch.Tensor) else ref_codes)
|
| 133 |
|
| 134 |
-
|
| 135 |
-
if isinstance(ref_codes, torch.Tensor):
|
| 136 |
ref_codes = ref_codes.to("cuda")
|
| 137 |
reference_cache[cache_key] = ref_codes
|
| 138 |
|
| 139 |
-
#
|
| 140 |
wav = tts.infer(text, ref_codes, ref_text_raw)
|
| 141 |
|
| 142 |
-
#
|
| 143 |
if speed_factor != 1.0:
|
| 144 |
with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as tmp:
|
| 145 |
sf.write(tmp.name, wav, 24000)
|
| 146 |
tmp_path = tmp.name
|
|
|
|
| 147 |
sound = AudioSegment.from_wav(tmp_path)
|
| 148 |
new_frame_rate = int(sound.frame_rate * speed_factor)
|
| 149 |
sound_stretched = sound._spawn(sound.raw_data, overrides={'frame_rate': new_frame_rate})
|
| 150 |
sound_stretched = sound_stretched.set_frame_rate(24000)
|
|
|
|
| 151 |
wav = np.array(sound_stretched.get_array_of_samples()).astype(np.float32) / 32768.0
|
| 152 |
if sound_stretched.channels == 2:
|
| 153 |
wav = wav.reshape((-1, 2)).mean(axis=1)
|
|
@@ -155,17 +152,21 @@ def core_synthesize(text, voice_choice, speed_factor):
|
|
| 155 |
|
| 156 |
return wav
|
| 157 |
|
| 158 |
-
@spaces.GPU
|
| 159 |
def custom_synthesize_logic(text, ref_audio_path, ref_text_raw):
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 164 |
ref_codes = tts.encode_reference(ref_audio_path)
|
| 165 |
wav = tts.infer(text, ref_codes, ref_text_raw)
|
| 166 |
return wav
|
| 167 |
|
| 168 |
-
# ---
|
| 169 |
class FastTTSRequest(BaseModel):
|
| 170 |
text: str
|
| 171 |
voice_choice: str
|
|
@@ -180,11 +181,10 @@ async def get_voices():
|
|
| 180 |
async def fast_tts(request: FastTTSRequest):
|
| 181 |
try:
|
| 182 |
start = time.time()
|
| 183 |
-
# Gọi hàm
|
| 184 |
wav = core_synthesize(request.text, request.voice_choice, request.speed_factor)
|
| 185 |
process_time = time.time() - start
|
| 186 |
|
| 187 |
-
# Base64
|
| 188 |
audio_buffer = io.BytesIO()
|
| 189 |
sf.write(audio_buffer, wav, 24000, format='WAV')
|
| 190 |
audio_base64 = base64.b64encode(audio_buffer.getvalue()).decode('utf-8')
|
|
@@ -197,7 +197,7 @@ async def fast_tts(request: FastTTSRequest):
|
|
| 197 |
except Exception as e:
|
| 198 |
raise HTTPException(status_code=500, detail=str(e))
|
| 199 |
|
| 200 |
-
# ---
|
| 201 |
theme = gr.themes.Soft()
|
| 202 |
css = ".container { max-width: 900px; margin: auto; }"
|
| 203 |
|
|
@@ -239,8 +239,10 @@ with gr.Blocks(theme=theme, css=css, title="VieNeu-TTS") as demo:
|
|
| 239 |
tabs.children[1].select(lambda: "custom_mode", None, mode_state)
|
| 240 |
btn.click(ui_synthesize, [inp_text, inp_voice, inp_audio, inp_ref_text, mode_state, inp_speed], [out_audio, out_status])
|
| 241 |
|
|
|
|
| 242 |
app = gr.mount_gradio_app(app, demo, path="/")
|
| 243 |
|
| 244 |
if __name__ == "__main__":
|
| 245 |
import uvicorn
|
|
|
|
| 246 |
uvicorn.run(app, host="0.0.0.0", port=7860)
|
|
|
|
| 1 |
+
import spaces # <--- BẮT BUỘC DÒNG 1
|
| 2 |
import os
|
| 3 |
import time
|
| 4 |
import threading
|
|
|
|
| 9 |
import tempfile
|
| 10 |
import numpy as np
|
| 11 |
|
| 12 |
+
# Các thư viện khác
|
| 13 |
import torch
|
| 14 |
import soundfile as sf
|
| 15 |
from pydub import AudioSegment
|
|
|
|
| 18 |
from pydantic import BaseModel
|
| 19 |
from vieneu_tts import VieNeuTTS
|
| 20 |
|
| 21 |
+
# --- KHỞI TẠO ---
|
| 22 |
app = FastAPI()
|
| 23 |
+
print("⏳ Đang khởi động Server...")
|
| 24 |
|
| 25 |
+
# Biến toàn cục để lưu model (Lazy Load)
|
| 26 |
+
tts_model = None
|
| 27 |
+
model_lock = threading.Lock()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
# Cache
|
| 30 |
CACHE_DIR = "./reference_cache"
|
|
|
|
| 50 |
with open(cache_path, 'wb') as f: pickle.dump(ref_codes, f)
|
| 51 |
except Exception: pass
|
| 52 |
|
| 53 |
+
# --- HELPER: LOAD MODEL AN TOÀN ---
|
| 54 |
+
def get_tts_model():
|
| 55 |
+
"""Hàm này chỉ tải model khi được gọi lần đầu tiên"""
|
| 56 |
+
global tts_model
|
| 57 |
+
with model_lock:
|
| 58 |
+
if tts_model is None:
|
| 59 |
+
print("📦 Đang khởi tạo model lần đầu (Lazy Load)...")
|
| 60 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 61 |
+
print(f" 🖥️ Device: {device}")
|
| 62 |
+
try:
|
| 63 |
+
# Load model
|
| 64 |
+
tts_model = VieNeuTTS(
|
| 65 |
+
backbone_repo="pnnbao-ump/VieNeu-TTS",
|
| 66 |
+
backbone_device=device,
|
| 67 |
+
codec_repo="neuphonic/neucodec",
|
| 68 |
+
codec_device=device
|
| 69 |
+
)
|
| 70 |
+
print(" ✅ Model tải thành công!")
|
| 71 |
+
except Exception as e:
|
| 72 |
+
print(f" ❌ Lỗi tải model: {e}")
|
| 73 |
+
raise e
|
| 74 |
+
return tts_model
|
| 75 |
+
|
| 76 |
+
# --- DATA ---
|
| 77 |
VOICE_SAMPLES = {
|
| 78 |
"Tuyên (nam miền Bắc)": {"audio": "./sample/Tuyên (nam miền Bắc).wav", "text": "./sample/Tuyên (nam miền Bắc).txt"},
|
| 79 |
"Vĩnh (nam miền Nam)": {"audio": "./sample/Vĩnh (nam miền Nam).wav", "text": "./sample/Vĩnh (nam miền Nam).txt"},
|
|
|
|
| 87 |
"Nhỏ Ngọt Ngào": {"audio": "./sample/Nhỏ Ngọt Ngào.wav", "text": "./sample/Nhỏ Ngọt Ngào.txt"},
|
| 88 |
}
|
| 89 |
|
| 90 |
+
# --- CORE LOGIC (DECORATED WITH @spaces.GPU) ---
|
| 91 |
|
| 92 |
+
@spaces.GPU(duration=120) # Tăng thời gian timeout lên 120s cho lần đầu load model
|
| 93 |
+
def core_synthesize(text, voice_choice, speed_factor):
|
| 94 |
+
# 1. Lấy model (Sẽ tải nếu chưa có)
|
| 95 |
+
tts = get_tts_model()
|
| 96 |
+
|
| 97 |
+
# 2. Đảm bảo model ở đúng device (GPU)
|
| 98 |
if torch.cuda.is_available():
|
|
|
|
|
|
|
| 99 |
try:
|
|
|
|
| 100 |
if next(tts.backbone.parameters()).device.type != 'cuda':
|
|
|
|
| 101 |
tts.backbone.to("cuda")
|
|
|
|
|
|
|
|
|
|
| 102 |
tts.codec.to("cuda")
|
| 103 |
+
except: pass
|
|
|
|
| 104 |
|
| 105 |
+
# 3. Lấy thông tin giọng
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
voice_info = VOICE_SAMPLES.get(voice_choice)
|
| 107 |
if not voice_info:
|
| 108 |
raise ValueError("Giọng không tồn tại")
|
|
|
|
| 113 |
with open(ref_text_path, "r", encoding="utf-8") as f:
|
| 114 |
ref_text_raw = f.read()
|
| 115 |
|
| 116 |
+
# 4. Encode Reference
|
| 117 |
cache_key = f"preset:{voice_choice}"
|
| 118 |
with reference_cache_lock:
|
| 119 |
if cache_key in reference_cache:
|
| 120 |
ref_codes = reference_cache[cache_key]
|
| 121 |
+
if isinstance(ref_codes, torch.Tensor) and torch.cuda.is_available():
|
|
|
|
| 122 |
ref_codes = ref_codes.to("cuda")
|
| 123 |
else:
|
| 124 |
ref_codes = load_cache_from_disk(cache_key)
|
| 125 |
if ref_codes is None:
|
| 126 |
+
ref_codes = tts.encode_reference(ref_audio_path)
|
| 127 |
+
# Cache trên CPU
|
| 128 |
save_cache_to_disk(cache_key, ref_codes.cpu() if isinstance(ref_codes, torch.Tensor) else ref_codes)
|
| 129 |
|
| 130 |
+
if isinstance(ref_codes, torch.Tensor) and torch.cuda.is_available():
|
|
|
|
| 131 |
ref_codes = ref_codes.to("cuda")
|
| 132 |
reference_cache[cache_key] = ref_codes
|
| 133 |
|
| 134 |
+
# 5. Infer
|
| 135 |
wav = tts.infer(text, ref_codes, ref_text_raw)
|
| 136 |
|
| 137 |
+
# 6. Speed Control (CPU Processing)
|
| 138 |
if speed_factor != 1.0:
|
| 139 |
with tempfile.NamedTemporaryFile(suffix='.wav', delete=False) as tmp:
|
| 140 |
sf.write(tmp.name, wav, 24000)
|
| 141 |
tmp_path = tmp.name
|
| 142 |
+
|
| 143 |
sound = AudioSegment.from_wav(tmp_path)
|
| 144 |
new_frame_rate = int(sound.frame_rate * speed_factor)
|
| 145 |
sound_stretched = sound._spawn(sound.raw_data, overrides={'frame_rate': new_frame_rate})
|
| 146 |
sound_stretched = sound_stretched.set_frame_rate(24000)
|
| 147 |
+
|
| 148 |
wav = np.array(sound_stretched.get_array_of_samples()).astype(np.float32) / 32768.0
|
| 149 |
if sound_stretched.channels == 2:
|
| 150 |
wav = wav.reshape((-1, 2)).mean(axis=1)
|
|
|
|
| 152 |
|
| 153 |
return wav
|
| 154 |
|
| 155 |
+
@spaces.GPU(duration=120)
|
| 156 |
def custom_synthesize_logic(text, ref_audio_path, ref_text_raw):
|
| 157 |
+
tts = get_tts_model()
|
| 158 |
+
if torch.cuda.is_available():
|
| 159 |
+
try:
|
| 160 |
+
if next(tts.backbone.parameters()).device.type != 'cuda':
|
| 161 |
+
tts.backbone.to("cuda")
|
| 162 |
+
tts.codec.to("cuda")
|
| 163 |
+
except: pass
|
| 164 |
+
|
| 165 |
ref_codes = tts.encode_reference(ref_audio_path)
|
| 166 |
wav = tts.infer(text, ref_codes, ref_text_raw)
|
| 167 |
return wav
|
| 168 |
|
| 169 |
+
# --- API ---
|
| 170 |
class FastTTSRequest(BaseModel):
|
| 171 |
text: str
|
| 172 |
voice_choice: str
|
|
|
|
| 181 |
async def fast_tts(request: FastTTSRequest):
|
| 182 |
try:
|
| 183 |
start = time.time()
|
| 184 |
+
# Gọi hàm GPU
|
| 185 |
wav = core_synthesize(request.text, request.voice_choice, request.speed_factor)
|
| 186 |
process_time = time.time() - start
|
| 187 |
|
|
|
|
| 188 |
audio_buffer = io.BytesIO()
|
| 189 |
sf.write(audio_buffer, wav, 24000, format='WAV')
|
| 190 |
audio_base64 = base64.b64encode(audio_buffer.getvalue()).decode('utf-8')
|
|
|
|
| 197 |
except Exception as e:
|
| 198 |
raise HTTPException(status_code=500, detail=str(e))
|
| 199 |
|
| 200 |
+
# --- GRADIO UI ---
|
| 201 |
theme = gr.themes.Soft()
|
| 202 |
css = ".container { max-width: 900px; margin: auto; }"
|
| 203 |
|
|
|
|
| 239 |
tabs.children[1].select(lambda: "custom_mode", None, mode_state)
|
| 240 |
btn.click(ui_synthesize, [inp_text, inp_voice, inp_audio, inp_ref_text, mode_state, inp_speed], [out_audio, out_status])
|
| 241 |
|
| 242 |
+
# Mount Gradio vào FastAPI
|
| 243 |
app = gr.mount_gradio_app(app, demo, path="/")
|
| 244 |
|
| 245 |
if __name__ == "__main__":
|
| 246 |
import uvicorn
|
| 247 |
+
# Mở port 7860 để Hugging Face truy cập
|
| 248 |
uvicorn.run(app, host="0.0.0.0", port=7860)
|