Spaces:
Runtime error
Runtime error
back to basics
Browse files- app.py +91 -247
- app.py.bak +247 -92
- tts/model/simple_gla.py +222 -235
app.py
CHANGED
|
@@ -1,176 +1,54 @@
|
|
| 1 |
import os
|
| 2 |
-
import re
|
| 3 |
-
import json
|
| 4 |
-
import sys
|
| 5 |
-
import time
|
| 6 |
-
import threading
|
| 7 |
-
import traceback
|
| 8 |
-
|
| 9 |
import gradio as gr
|
| 10 |
import numpy as np
|
| 11 |
-
import soundfile as sf
|
| 12 |
import torch
|
|
|
|
| 13 |
import spaces
|
| 14 |
-
from huggingface_hub import login, snapshot_download
|
| 15 |
-
|
| 16 |
-
# --------- Environnement / stabilité ----------
|
| 17 |
-
os.environ.setdefault("FLA_CONV_BACKEND", "torch") # éviter les kernels Triton
|
| 18 |
-
os.environ.setdefault("FLA_USE_FAST_OPS", "0")
|
| 19 |
-
os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
|
| 20 |
-
torch.backends.cuda.matmul.allow_tf32 = True
|
| 21 |
-
try:
|
| 22 |
-
torch.set_float32_matmul_precision("high")
|
| 23 |
-
except Exception:
|
| 24 |
-
pass
|
| 25 |
|
|
|
|
| 26 |
from pardi_speech import PardiSpeech, VelocityHeadSamplingParams # présent dans ce repo
|
| 27 |
|
| 28 |
MODEL_REPO_ID = os.environ.get("MODEL_REPO_ID", "theodorr/pardi-speech-enfr-forbidden")
|
| 29 |
-
HF_TOKEN = os.environ.get("HF_TOKEN")
|
| 30 |
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
def _log(msg: str):
|
| 35 |
-
_MODEL["logs"].append(str(msg))
|
| 36 |
-
# borne la taille
|
| 37 |
-
if len(_MODEL["logs"]) > 2000:
|
| 38 |
-
_MODEL["logs"] = _MODEL["logs"][-2000:]
|
| 39 |
-
|
| 40 |
-
def _env_diag() -> str:
|
| 41 |
-
parts = []
|
| 42 |
try:
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
import triton # type: ignore
|
| 46 |
-
parts.append(f"triton={getattr(triton, '__version__', 'unknown')}")
|
| 47 |
-
except Exception:
|
| 48 |
-
parts.append("triton=not_importable")
|
| 49 |
-
parts.append(f"cuda.is_available={torch.cuda.is_available()}")
|
| 50 |
-
if torch.cuda.is_available():
|
| 51 |
-
parts.append(f"cuda.version={torch.version.cuda}")
|
| 52 |
-
try:
|
| 53 |
-
free, total = torch.cuda.mem_get_info()
|
| 54 |
-
parts.append(f"mem_free={free/1e9:.2f}GB/{total/1e9:.2f}GB")
|
| 55 |
-
except Exception:
|
| 56 |
-
pass
|
| 57 |
except Exception as e:
|
| 58 |
-
|
| 59 |
-
|
|
|
|
|
|
|
| 60 |
|
| 61 |
def _normalize_text(s: str, lang_hint: str = "fr") -> str:
|
| 62 |
-
s = (s or "").strip()
|
| 63 |
try:
|
| 64 |
-
import re
|
| 65 |
from num2words import num2words
|
| 66 |
-
def repl(m):
|
| 67 |
-
|
| 68 |
-
return num2words(int(m.group()), lang=lang_hint)
|
| 69 |
-
except Exception:
|
| 70 |
-
return m.group()
|
| 71 |
-
s = _re.sub(r"\d+", repl, s)
|
| 72 |
except Exception:
|
| 73 |
pass
|
| 74 |
return s
|
| 75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
def _to_mono_float32(arr: np.ndarray) -> np.ndarray:
|
| 77 |
-
arr = np.
|
| 78 |
if arr.ndim == 2:
|
| 79 |
arr = arr.mean(axis=1)
|
| 80 |
-
return arr
|
| 81 |
-
|
| 82 |
-
def _extract_repo_ids_from_config(config_path: str):
|
| 83 |
-
repo_ids = set()
|
| 84 |
-
preview = None
|
| 85 |
-
try:
|
| 86 |
-
with open(config_path, "r", encoding="utf-8") as f:
|
| 87 |
-
cfg = json.load(f)
|
| 88 |
-
pattern = re.compile(r"^[\w\-]+\/[\w\.\-]+$") # org/name
|
| 89 |
-
def rec(obj):
|
| 90 |
-
if isinstance(obj, dict):
|
| 91 |
-
for v in obj.values(): rec(v)
|
| 92 |
-
elif isinstance(obj, list):
|
| 93 |
-
for v in obj: rec(v)
|
| 94 |
-
elif isinstance(obj, str):
|
| 95 |
-
if pattern.match(obj): repo_ids.add(obj)
|
| 96 |
-
rec(cfg)
|
| 97 |
-
try:
|
| 98 |
-
subset_keys = list(cfg)[:5] if isinstance(cfg, dict) else []
|
| 99 |
-
preview = json.dumps({k: cfg[k] for k in subset_keys}, ensure_ascii=False)[:600]
|
| 100 |
-
except Exception:
|
| 101 |
-
pass
|
| 102 |
-
except Exception:
|
| 103 |
-
pass
|
| 104 |
-
return sorted(repo_ids), preview
|
| 105 |
-
|
| 106 |
-
def _prefetch_and_load_cpu():
|
| 107 |
-
"""Exécuté dans un thread au démarrage du Space (hors worker GPU)."""
|
| 108 |
-
try:
|
| 109 |
-
_log("[prefetch] snapshot_download (main)...")
|
| 110 |
-
local_dir = snapshot_download(
|
| 111 |
-
repo_id=MODEL_REPO_ID,
|
| 112 |
-
token=HF_TOKEN,
|
| 113 |
-
local_dir=None,
|
| 114 |
-
local_files_only=False,
|
| 115 |
-
)
|
| 116 |
-
_log(f"[prefetch] main done -> {local_dir}")
|
| 117 |
-
|
| 118 |
-
cfg_path = os.path.join(local_dir, "config.json")
|
| 119 |
-
nested, cfg_preview = _extract_repo_ids_from_config(cfg_path)
|
| 120 |
-
if cfg_preview:
|
| 121 |
-
_log(f"[config] preview: {cfg_preview}")
|
| 122 |
-
for rid in nested:
|
| 123 |
-
if rid == MODEL_REPO_ID:
|
| 124 |
-
continue
|
| 125 |
-
_log(f"[prefetch] nested repo: {rid} ...")
|
| 126 |
-
snapshot_download(repo_id=rid, token=HF_TOKEN, local_dir=None, local_files_only=False)
|
| 127 |
-
_log(f"[prefetch] nested repo: {rid} done")
|
| 128 |
-
|
| 129 |
-
# Forcer offline pendant le vrai chargement
|
| 130 |
-
old_off = os.environ.get("HF_HUB_OFFLINE")
|
| 131 |
-
os.environ["HF_HUB_OFFLINE"] = "1"
|
| 132 |
-
try:
|
| 133 |
-
_log("[load] from_pretrained(map_location='cpu')...")
|
| 134 |
-
m = PardiSpeech.from_pretrained(local_dir, map_location="cpu")
|
| 135 |
-
m.eval()
|
| 136 |
-
_MODEL["pardi"] = m
|
| 137 |
-
_MODEL["sr"] = getattr(m, "sampling_rate", 24000)
|
| 138 |
-
_log(f"[load] cpu OK (sr={_MODEL['sr']})")
|
| 139 |
-
finally:
|
| 140 |
-
if old_off is None:
|
| 141 |
-
os.environ.pop("HF_HUB_OFFLINE", None)
|
| 142 |
-
else:
|
| 143 |
-
os.environ["HF_HUB_OFFLINE"] = old_off
|
| 144 |
-
|
| 145 |
-
except BaseException as e:
|
| 146 |
-
_MODEL["err"] = e
|
| 147 |
-
_log(f"[EXC@preload] {type(e).__name__}: {e}")
|
| 148 |
-
_log(traceback.format_exc())
|
| 149 |
|
| 150 |
-
|
| 151 |
-
if _MODEL["thread"] is None:
|
| 152 |
-
_MODEL["thread"] = threading.Thread(target=_prefetch_and_load_cpu, daemon=True)
|
| 153 |
-
_MODEL["thread"].start()
|
| 154 |
-
|
| 155 |
-
def _move_to_cuda_if_available(m, logs_acc):
|
| 156 |
-
def L(msg): logs_acc.append(str(msg))
|
| 157 |
-
if torch.cuda.is_available():
|
| 158 |
-
L("[move] moving model to cuda...")
|
| 159 |
-
try:
|
| 160 |
-
m = m.to("cuda") # type: ignore[attr-defined]
|
| 161 |
-
L("[move] cuda OK")
|
| 162 |
-
except Exception as e:
|
| 163 |
-
L(f"[move] .to('cuda') failed: {e}. Keeping on CPU.")
|
| 164 |
-
else:
|
| 165 |
-
L("[move] cuda not available, keep CPU")
|
| 166 |
-
return m
|
| 167 |
-
|
| 168 |
-
# --------- UI callback (GPU) ----------
|
| 169 |
-
@spaces.GPU(duration=200)
|
| 170 |
def synthesize(
|
| 171 |
text: str,
|
| 172 |
-
debug: bool,
|
| 173 |
-
adv_sampling: bool, # Velocity Head sampling
|
| 174 |
ref_audio,
|
| 175 |
ref_text: str,
|
| 176 |
steps: int,
|
|
@@ -179,112 +57,83 @@ def synthesize(
|
|
| 179 |
temperature: float,
|
| 180 |
max_seq_len: int,
|
| 181 |
seed: int,
|
| 182 |
-
lang_hint: str
|
| 183 |
):
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
logs.append(str(msg))
|
| 187 |
-
joined = "\n".join(logs + _MODEL["logs"][-50:]) # mêle quelques logs de préchargement
|
| 188 |
-
if len(joined) > 12000:
|
| 189 |
-
joined = joined[-12000:]
|
| 190 |
-
return joined
|
| 191 |
-
|
| 192 |
-
try:
|
| 193 |
-
if HF_TOKEN:
|
| 194 |
-
try:
|
| 195 |
-
login(token=HF_TOKEN)
|
| 196 |
-
yield None, LOG("✅ HF login ok")
|
| 197 |
-
except Exception as e:
|
| 198 |
-
yield None, LOG(f"⚠️ HF login failed: {e}")
|
| 199 |
-
|
| 200 |
-
yield None, LOG("[env] " + _env_diag())
|
| 201 |
-
torch.manual_seed(int(seed))
|
| 202 |
-
os.environ.setdefault("CUDA_LAUNCH_BLOCKING", "1")
|
| 203 |
-
|
| 204 |
-
# Si le modèle n’est pas encore prêt, on attend jusqu’à 180s max ici
|
| 205 |
-
t0 = time.perf_counter()
|
| 206 |
-
while _MODEL["pardi"] is None and _MODEL["err"] is None:
|
| 207 |
-
elapsed = time.perf_counter() - t0
|
| 208 |
-
yield None, LOG(f"[init] still loading on CPU… {elapsed:.1f}s")
|
| 209 |
-
if elapsed > 180:
|
| 210 |
-
# dump de la stack du thread de préchargement pour debug
|
| 211 |
-
tid = _MODEL["thread"].ident if _MODEL["thread"] else None
|
| 212 |
-
if tid is not None:
|
| 213 |
-
frame = sys._current_frames().get(tid)
|
| 214 |
-
if frame is not None:
|
| 215 |
-
stack_txt = "".join(traceback.format_stack(frame))
|
| 216 |
-
yield None, LOG("[stack-final]\n" + stack_txt)
|
| 217 |
-
raise TimeoutError("Preload timeout (>180s)")
|
| 218 |
-
time.sleep(2.0)
|
| 219 |
-
|
| 220 |
-
if _MODEL["err"]:
|
| 221 |
-
raise _MODEL["err"]
|
| 222 |
-
|
| 223 |
-
pardi = _MODEL["pardi"]
|
| 224 |
-
sr_out = _MODEL["sr"]
|
| 225 |
-
|
| 226 |
-
# Déplacement vers CUDA si possible
|
| 227 |
-
pardi = _move_to_cuda_if_available(pardi, logs)
|
| 228 |
-
yield None, LOG(f"[init] model ready on {'cuda' if torch.cuda.is_available() else 'cpu'}, sr={sr_out}")
|
| 229 |
-
|
| 230 |
-
# ---- Texte + prefix optionnel ----
|
| 231 |
-
txt = _normalize_text(text or "", lang_hint=lang_hint)
|
| 232 |
-
yield None, LOG(f"[text] {txt[:120]}{'...' if len(txt) > 120 else ''}")
|
| 233 |
|
| 234 |
-
|
| 235 |
-
|
| 236 |
|
| 237 |
-
|
| 238 |
-
if ref_audio is not None:
|
| 239 |
-
yield None, LOG("[prefix] encoding reference audio...")
|
| 240 |
-
if isinstance(ref_audio, str):
|
| 241 |
-
wav, sr = sf.read(ref_audio)
|
| 242 |
-
else:
|
| 243 |
-
sr, wav = ref_audio
|
| 244 |
-
wav = _to_mono_float32(wav)
|
| 245 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 246 |
-
wav_t = torch.from_numpy(wav).to(device).unsqueeze(0)
|
| 247 |
-
with torch.inference_mode():
|
| 248 |
-
prefix_tokens = pardi.patchvae.encode(wav_t) # type: ignore[attr-defined]
|
| 249 |
-
prefix = (ref_text or "", prefix_tokens[0])
|
| 250 |
-
yield None, LOG("[prefix] done.")
|
| 251 |
|
| 252 |
-
|
| 253 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 254 |
|
| 255 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 256 |
with torch.inference_mode():
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
vparams = VelocityHeadSamplingParams(cfg_ref=float(cfg_ref), cfg=float(cfg), num_steps=int(steps))
|
| 260 |
-
except TypeError:
|
| 261 |
-
vparams = VelocityHeadSamplingParams(cfg_ref=float(cfg_ref), cfg=float(cfg),
|
| 262 |
-
num_steps=int(steps), temperature=float(temperature))
|
| 263 |
-
wavs, _ = pardi.text_to_speech([txt], prefix, max_seq_len=int(max_seq_len),
|
| 264 |
-
velocity_head_sampling_params=vparams)
|
| 265 |
-
else:
|
| 266 |
-
wavs, _ = pardi.text_to_speech([txt], prefix, max_seq_len=int(max_seq_len))
|
| 267 |
|
| 268 |
-
|
| 269 |
-
yield (sr_out, wav), LOG("[ok] done.")
|
| 270 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 271 |
except Exception as e:
|
| 272 |
-
|
| 273 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 274 |
|
| 275 |
-
# --------- UI ----------
|
| 276 |
def build_demo():
|
| 277 |
with gr.Blocks(title="Lina-speech / pardi-speech Demo") as demo:
|
| 278 |
gr.Markdown(
|
| 279 |
-
"
|
| 280 |
-
"Génère de l'audio à partir de texte, avec ou sans prefix (audio de référence).\n"
|
| 281 |
-
"
|
| 282 |
)
|
|
|
|
| 283 |
with gr.Row():
|
| 284 |
text = gr.Textbox(label="Texte à synthétiser", lines=4, placeholder="Tape ton texte ici…")
|
| 285 |
with gr.Accordion("Prefix (optionnel)", open=False):
|
| 286 |
ref_audio = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Audio de référence")
|
| 287 |
-
ref_text
|
| 288 |
with gr.Accordion("Options avancées", open=False):
|
| 289 |
with gr.Row():
|
| 290 |
steps = gr.Slider(1, 50, value=10, step=1, label="num_steps")
|
|
@@ -293,26 +142,21 @@ def build_demo():
|
|
| 293 |
with gr.Row():
|
| 294 |
temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.05, label="Température")
|
| 295 |
max_seq_len = gr.Slider(50, 1200, value=300, step=10, label="max_seq_len (tokens audio)")
|
| 296 |
-
seed = gr.Number(value=0, precision=0, label="Seed")
|
| 297 |
-
|
| 298 |
-
with gr.Row():
|
| 299 |
-
debug = gr.Checkbox(value=False, label="Mode debug")
|
| 300 |
-
adv_sampling = gr.Checkbox(value=False, label="Sampling avancé (Velocity Head)")
|
| 301 |
|
| 302 |
btn = gr.Button("Synthétiser")
|
| 303 |
out_audio = gr.Audio(label="Sortie audio", type="numpy")
|
| 304 |
-
logs_box = gr.Textbox(label="Logs (live)", lines=28)
|
| 305 |
|
| 306 |
demo.queue(default_concurrency_limit=1, max_size=32)
|
|
|
|
| 307 |
btn.click(
|
| 308 |
fn=synthesize,
|
| 309 |
-
inputs=[text,
|
| 310 |
-
outputs=[out_audio
|
| 311 |
-
api_name="synthesize",
|
| 312 |
)
|
| 313 |
return demo
|
| 314 |
|
| 315 |
if __name__ == "__main__":
|
| 316 |
-
build_demo()
|
| 317 |
-
|
| 318 |
-
# retrigger 2025-10-30T16:37:47+01:00
|
|
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import gradio as gr
|
| 3 |
import numpy as np
|
|
|
|
| 4 |
import torch
|
| 5 |
+
import soundfile as sf
|
| 6 |
import spaces
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
| 8 |
+
from huggingface_hub import login
|
| 9 |
from pardi_speech import PardiSpeech, VelocityHeadSamplingParams # présent dans ce repo
|
| 10 |
|
| 11 |
MODEL_REPO_ID = os.environ.get("MODEL_REPO_ID", "theodorr/pardi-speech-enfr-forbidden")
|
|
|
|
| 12 |
|
| 13 |
+
HF_TOKEN = os.environ.get("HF_TOKEN")
|
| 14 |
+
if HF_TOKEN:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
try:
|
| 16 |
+
login(token=HF_TOKEN)
|
| 17 |
+
print("✅ Logged to Hugging Face Hub.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
except Exception as e:
|
| 19 |
+
print("⚠️ HF login failed:", e)
|
| 20 |
+
|
| 21 |
+
_pardi = None
|
| 22 |
+
_sampling_rate = 24000
|
| 23 |
|
| 24 |
def _normalize_text(s: str, lang_hint: str = "fr") -> str:
|
| 25 |
+
s = (s or "").strip().lower()
|
| 26 |
try:
|
| 27 |
+
import re
|
| 28 |
from num2words import num2words
|
| 29 |
+
def repl(m): return num2words(int(m.group()), lang=lang_hint)
|
| 30 |
+
s = re.sub(r"\d+", repl, s)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
except Exception:
|
| 32 |
pass
|
| 33 |
return s
|
| 34 |
|
| 35 |
+
def _load_model(device: str = "cuda"):
|
| 36 |
+
global _pardi, _sampling_rate
|
| 37 |
+
if _pardi is None:
|
| 38 |
+
_pardi = PardiSpeech.from_pretrained(MODEL_REPO_ID, map_location=device)
|
| 39 |
+
_sampling_rate = getattr(_pardi, "sampling_rate", 24000)
|
| 40 |
+
print(f"✅ PardiSpeech loaded on {device} (sr={_sampling_rate}).")
|
| 41 |
+
return _pardi
|
| 42 |
+
|
| 43 |
def _to_mono_float32(arr: np.ndarray) -> np.ndarray:
|
| 44 |
+
arr = arr.astype(np.float32)
|
| 45 |
if arr.ndim == 2:
|
| 46 |
arr = arr.mean(axis=1)
|
| 47 |
+
return arr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
+
@spaces.GPU(duration=120)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
def synthesize(
|
| 51 |
text: str,
|
|
|
|
|
|
|
| 52 |
ref_audio,
|
| 53 |
ref_text: str,
|
| 54 |
steps: int,
|
|
|
|
| 57 |
temperature: float,
|
| 58 |
max_seq_len: int,
|
| 59 |
seed: int,
|
| 60 |
+
lang_hint: str
|
| 61 |
):
|
| 62 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 63 |
+
torch.manual_seed(int(seed))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
|
| 65 |
+
pardi = _load_model(device)
|
| 66 |
+
txt = _normalize_text(text, lang_hint=lang_hint)
|
| 67 |
|
| 68 |
+
cache = pardi.tts.audio_decoder.init_cache(int(max_seq_len), device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
+
# --- IMPORTANT : signature de VelocityHeadSamplingParams ---
|
| 71 |
+
# Dans ton notebook d’inférence, la classe attend (cfg_ref, cfg, num_steps) SANS 'temperature'.
|
| 72 |
+
# On essaie d’abord sans temperature, puis fallback si la classe en accepte une.
|
| 73 |
+
try:
|
| 74 |
+
vel_params = VelocityHeadSamplingParams(
|
| 75 |
+
cfg_ref=float(cfg_ref),
|
| 76 |
+
cfg=float(cfg),
|
| 77 |
+
num_steps=int(steps)
|
| 78 |
+
)
|
| 79 |
+
except TypeError:
|
| 80 |
+
vel_params = VelocityHeadSamplingParams(
|
| 81 |
+
cfg_ref=float(cfg_ref),
|
| 82 |
+
cfg=float(cfg),
|
| 83 |
+
num_steps=int(steps),
|
| 84 |
+
temperature=float(temperature)
|
| 85 |
+
)
|
| 86 |
|
| 87 |
+
# Prefix optionnel
|
| 88 |
+
prefix = None
|
| 89 |
+
if ref_audio is not None:
|
| 90 |
+
if isinstance(ref_audio, str):
|
| 91 |
+
wav, sr = sf.read(ref_audio)
|
| 92 |
+
else:
|
| 93 |
+
sr, wav = ref_audio
|
| 94 |
+
wav = _to_mono_float32(np.array(wav))
|
| 95 |
+
wav_t = torch.from_numpy(wav).to(device)
|
| 96 |
+
import torchaudio
|
| 97 |
+
if sr != pardi.sampling_rate:
|
| 98 |
+
wav_t = torchaudio.functional.resample(wav_t, sr, pardi.sampling_rate)
|
| 99 |
+
wav_t = wav_t.unsqueeze(0)
|
| 100 |
with torch.inference_mode():
|
| 101 |
+
prefix_tokens = pardi.patchvae.encode(wav_t)
|
| 102 |
+
prefix = (ref_text or "", prefix_tokens[0])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
+
print(f"[debug] has_prefix={prefix is not None}, steps={steps}, cfg={cfg}, cfg_ref={cfg_ref}, T={temperature}, max_seq_len={max_seq_len}, seed={seed}")
|
|
|
|
| 105 |
|
| 106 |
+
try:
|
| 107 |
+
with torch.inference_mode():
|
| 108 |
+
wavs, _ = pardi.text_to_speech(
|
| 109 |
+
[txt],
|
| 110 |
+
prefix,
|
| 111 |
+
max_seq_len=int(max_seq_len),
|
| 112 |
+
velocity_head_sampling_params=vel_params,
|
| 113 |
+
cache=cache
|
| 114 |
+
)
|
| 115 |
except Exception as e:
|
| 116 |
+
import traceback, sys
|
| 117 |
+
print("❌ text_to_speech failed:", e, file=sys.stderr)
|
| 118 |
+
traceback.print_exc()
|
| 119 |
+
raise gr.Error(f"Synthèse échouée: {type(e).__name__}: {e}")
|
| 120 |
+
|
| 121 |
+
wav = wavs[0].detach().cpu().numpy()
|
| 122 |
+
return (_sampling_rate, wav)
|
| 123 |
|
|
|
|
| 124 |
def build_demo():
|
| 125 |
with gr.Blocks(title="Lina-speech / pardi-speech Demo") as demo:
|
| 126 |
gr.Markdown(
|
| 127 |
+
"## Lina-speech (pardi-speech) – Démo TTS\n"
|
| 128 |
+
"Génère de l'audio à partir de texte, avec ou sans *prefix* (audio de référence).\n"
|
| 129 |
+
"Paramètres avancés: *num_steps*, *CFG*, *température*, *max_seq_len*, *seed*."
|
| 130 |
)
|
| 131 |
+
|
| 132 |
with gr.Row():
|
| 133 |
text = gr.Textbox(label="Texte à synthétiser", lines=4, placeholder="Tape ton texte ici…")
|
| 134 |
with gr.Accordion("Prefix (optionnel)", open=False):
|
| 135 |
ref_audio = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Audio de référence")
|
| 136 |
+
ref_text = gr.Textbox(label="Texte du prefix (si connu)", placeholder="Transcription du prefix (optionnel)")
|
| 137 |
with gr.Accordion("Options avancées", open=False):
|
| 138 |
with gr.Row():
|
| 139 |
steps = gr.Slider(1, 50, value=10, step=1, label="num_steps")
|
|
|
|
| 142 |
with gr.Row():
|
| 143 |
temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.05, label="Température")
|
| 144 |
max_seq_len = gr.Slider(50, 1200, value=300, step=10, label="max_seq_len (tokens audio)")
|
| 145 |
+
seed = gr.Number(value=0, precision=0, label="Seed (reproductibilité)")
|
| 146 |
+
lang_hint = gr.Dropdown(choices=["fr", "en"], value="fr", label="Langue (normalisation)")
|
|
|
|
|
|
|
|
|
|
| 147 |
|
| 148 |
btn = gr.Button("Synthétiser")
|
| 149 |
out_audio = gr.Audio(label="Sortie audio", type="numpy")
|
|
|
|
| 150 |
|
| 151 |
demo.queue(default_concurrency_limit=1, max_size=32)
|
| 152 |
+
|
| 153 |
btn.click(
|
| 154 |
fn=synthesize,
|
| 155 |
+
inputs=[text, ref_audio, ref_text, steps, cfg, cfg_ref, temperature, max_seq_len, seed, lang_hint],
|
| 156 |
+
outputs=[out_audio]
|
|
|
|
| 157 |
)
|
| 158 |
return demo
|
| 159 |
|
| 160 |
if __name__ == "__main__":
|
| 161 |
+
demo = build_demo()
|
| 162 |
+
demo.launch()
|
|
|
app.py.bak
CHANGED
|
@@ -1,54 +1,176 @@
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import gradio as gr
|
| 3 |
import numpy as np
|
| 4 |
-
import torch
|
| 5 |
import soundfile as sf
|
|
|
|
| 6 |
import spaces
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
| 8 |
-
from huggingface_hub import login
|
| 9 |
from pardi_speech import PardiSpeech, VelocityHeadSamplingParams # présent dans ce repo
|
| 10 |
|
| 11 |
MODEL_REPO_ID = os.environ.get("MODEL_REPO_ID", "theodorr/pardi-speech-enfr-forbidden")
|
| 12 |
-
|
| 13 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
try:
|
| 16 |
-
|
| 17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
except Exception as e:
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
_pardi = None
|
| 22 |
-
_sampling_rate = 24000
|
| 23 |
|
| 24 |
def _normalize_text(s: str, lang_hint: str = "fr") -> str:
|
| 25 |
-
s = (s or "").strip()
|
| 26 |
try:
|
| 27 |
-
import re
|
| 28 |
from num2words import num2words
|
| 29 |
-
def repl(m):
|
| 30 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
except Exception:
|
| 32 |
pass
|
| 33 |
return s
|
| 34 |
|
| 35 |
-
def _load_model(device: str = "cuda"):
|
| 36 |
-
global _pardi, _sampling_rate
|
| 37 |
-
if _pardi is None:
|
| 38 |
-
_pardi = PardiSpeech.from_pretrained(MODEL_REPO_ID, map_location=device)
|
| 39 |
-
_sampling_rate = getattr(_pardi, "sampling_rate", 24000)
|
| 40 |
-
print(f"✅ PardiSpeech loaded on {device} (sr={_sampling_rate}).")
|
| 41 |
-
return _pardi
|
| 42 |
-
|
| 43 |
def _to_mono_float32(arr: np.ndarray) -> np.ndarray:
|
| 44 |
-
arr =
|
| 45 |
if arr.ndim == 2:
|
| 46 |
arr = arr.mean(axis=1)
|
| 47 |
-
return arr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
def synthesize(
|
| 51 |
text: str,
|
|
|
|
|
|
|
| 52 |
ref_audio,
|
| 53 |
ref_text: str,
|
| 54 |
steps: int,
|
|
@@ -57,83 +179,112 @@ def synthesize(
|
|
| 57 |
temperature: float,
|
| 58 |
max_seq_len: int,
|
| 59 |
seed: int,
|
| 60 |
-
lang_hint: str
|
| 61 |
):
|
| 62 |
-
|
| 63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
|
| 65 |
-
|
| 66 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
|
| 68 |
-
|
|
|
|
|
|
|
| 69 |
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
)
|
| 86 |
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
if ref_audio is not None:
|
| 90 |
-
if isinstance(ref_audio, str):
|
| 91 |
-
wav, sr = sf.read(ref_audio)
|
| 92 |
-
else:
|
| 93 |
-
sr, wav = ref_audio
|
| 94 |
-
wav = _to_mono_float32(np.array(wav))
|
| 95 |
-
wav_t = torch.from_numpy(wav).to(device)
|
| 96 |
-
import torchaudio
|
| 97 |
-
if sr != pardi.sampling_rate:
|
| 98 |
-
wav_t = torchaudio.functional.resample(wav_t, sr, pardi.sampling_rate)
|
| 99 |
-
wav_t = wav_t.unsqueeze(0)
|
| 100 |
-
with torch.inference_mode():
|
| 101 |
-
prefix_tokens = pardi.patchvae.encode(wav_t)
|
| 102 |
-
prefix = (ref_text or "", prefix_tokens[0])
|
| 103 |
|
| 104 |
-
|
|
|
|
| 105 |
|
| 106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
with torch.inference_mode():
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
traceback.print_exc()
|
| 119 |
-
raise gr.Error(f"Synthèse échouée: {type(e).__name__}: {e}")
|
| 120 |
|
| 121 |
-
|
| 122 |
-
|
| 123 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 124 |
def build_demo():
|
| 125 |
with gr.Blocks(title="Lina-speech / pardi-speech Demo") as demo:
|
| 126 |
gr.Markdown(
|
| 127 |
-
"
|
| 128 |
-
"Génère de l'audio à partir de texte, avec ou sans
|
| 129 |
-
"
|
| 130 |
)
|
| 131 |
-
|
| 132 |
with gr.Row():
|
| 133 |
text = gr.Textbox(label="Texte à synthétiser", lines=4, placeholder="Tape ton texte ici…")
|
| 134 |
with gr.Accordion("Prefix (optionnel)", open=False):
|
| 135 |
ref_audio = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Audio de référence")
|
| 136 |
-
ref_text
|
| 137 |
with gr.Accordion("Options avancées", open=False):
|
| 138 |
with gr.Row():
|
| 139 |
steps = gr.Slider(1, 50, value=10, step=1, label="num_steps")
|
|
@@ -142,22 +293,26 @@ def build_demo():
|
|
| 142 |
with gr.Row():
|
| 143 |
temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.05, label="Température")
|
| 144 |
max_seq_len = gr.Slider(50, 1200, value=300, step=10, label="max_seq_len (tokens audio)")
|
| 145 |
-
seed = gr.Number(value=0, precision=0, label="Seed
|
| 146 |
-
|
|
|
|
|
|
|
|
|
|
| 147 |
|
| 148 |
btn = gr.Button("Synthétiser")
|
| 149 |
out_audio = gr.Audio(label="Sortie audio", type="numpy")
|
|
|
|
| 150 |
|
| 151 |
demo.queue(default_concurrency_limit=1, max_size=32)
|
| 152 |
-
|
| 153 |
btn.click(
|
| 154 |
fn=synthesize,
|
| 155 |
-
inputs=[text, ref_audio, ref_text, steps, cfg, cfg_ref, temperature, max_seq_len, seed, lang_hint],
|
| 156 |
-
outputs=[out_audio]
|
|
|
|
| 157 |
)
|
| 158 |
return demo
|
| 159 |
|
| 160 |
if __name__ == "__main__":
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
# retrigger 2025-10-
|
|
|
|
| 1 |
import os
|
| 2 |
+
import re
|
| 3 |
+
import json
|
| 4 |
+
import sys
|
| 5 |
+
import time
|
| 6 |
+
import threading
|
| 7 |
+
import traceback
|
| 8 |
+
|
| 9 |
import gradio as gr
|
| 10 |
import numpy as np
|
|
|
|
| 11 |
import soundfile as sf
|
| 12 |
+
import torch
|
| 13 |
import spaces
|
| 14 |
+
from huggingface_hub import login, snapshot_download
|
| 15 |
+
|
| 16 |
+
# --------- Environnement / stabilité ----------
|
| 17 |
+
os.environ.setdefault("FLA_CONV_BACKEND", "torch") # éviter les kernels Triton
|
| 18 |
+
os.environ.setdefault("FLA_USE_FAST_OPS", "0")
|
| 19 |
+
os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
|
| 20 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
| 21 |
+
try:
|
| 22 |
+
torch.set_float32_matmul_precision("high")
|
| 23 |
+
except Exception:
|
| 24 |
+
pass
|
| 25 |
|
|
|
|
| 26 |
from pardi_speech import PardiSpeech, VelocityHeadSamplingParams # présent dans ce repo
|
| 27 |
|
| 28 |
MODEL_REPO_ID = os.environ.get("MODEL_REPO_ID", "theodorr/pardi-speech-enfr-forbidden")
|
|
|
|
| 29 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
| 30 |
+
|
| 31 |
+
# --------- Cache global (préchargement au démarrage) ----------
|
| 32 |
+
_MODEL = {"pardi": None, "sr": 24000, "err": None, "logs": [], "thread": None}
|
| 33 |
+
|
| 34 |
+
def _log(msg: str):
|
| 35 |
+
_MODEL["logs"].append(str(msg))
|
| 36 |
+
# borne la taille
|
| 37 |
+
if len(_MODEL["logs"]) > 2000:
|
| 38 |
+
_MODEL["logs"] = _MODEL["logs"][-2000:]
|
| 39 |
+
|
| 40 |
+
def _env_diag() -> str:
|
| 41 |
+
parts = []
|
| 42 |
try:
|
| 43 |
+
parts.append(f"torch={torch.__version__}")
|
| 44 |
+
try:
|
| 45 |
+
import triton # type: ignore
|
| 46 |
+
parts.append(f"triton={getattr(triton, '__version__', 'unknown')}")
|
| 47 |
+
except Exception:
|
| 48 |
+
parts.append("triton=not_importable")
|
| 49 |
+
parts.append(f"cuda.is_available={torch.cuda.is_available()}")
|
| 50 |
+
if torch.cuda.is_available():
|
| 51 |
+
parts.append(f"cuda.version={torch.version.cuda}")
|
| 52 |
+
try:
|
| 53 |
+
free, total = torch.cuda.mem_get_info()
|
| 54 |
+
parts.append(f"mem_free={free/1e9:.2f}GB/{total/1e9:.2f}GB")
|
| 55 |
+
except Exception:
|
| 56 |
+
pass
|
| 57 |
except Exception as e:
|
| 58 |
+
parts.append(f"env_diag_error={e}")
|
| 59 |
+
return " | ".join(parts)
|
|
|
|
|
|
|
| 60 |
|
| 61 |
def _normalize_text(s: str, lang_hint: str = "fr") -> str:
|
| 62 |
+
s = (s or "").strip()
|
| 63 |
try:
|
| 64 |
+
import re as _re
|
| 65 |
from num2words import num2words
|
| 66 |
+
def repl(m):
|
| 67 |
+
try:
|
| 68 |
+
return num2words(int(m.group()), lang=lang_hint)
|
| 69 |
+
except Exception:
|
| 70 |
+
return m.group()
|
| 71 |
+
s = _re.sub(r"\d+", repl, s)
|
| 72 |
except Exception:
|
| 73 |
pass
|
| 74 |
return s
|
| 75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
def _to_mono_float32(arr: np.ndarray) -> np.ndarray:
|
| 77 |
+
arr = np.asarray(arr)
|
| 78 |
if arr.ndim == 2:
|
| 79 |
arr = arr.mean(axis=1)
|
| 80 |
+
return arr.astype(np.float32)
|
| 81 |
+
|
| 82 |
+
def _extract_repo_ids_from_config(config_path: str):
|
| 83 |
+
repo_ids = set()
|
| 84 |
+
preview = None
|
| 85 |
+
try:
|
| 86 |
+
with open(config_path, "r", encoding="utf-8") as f:
|
| 87 |
+
cfg = json.load(f)
|
| 88 |
+
pattern = re.compile(r"^[\w\-]+\/[\w\.\-]+$") # org/name
|
| 89 |
+
def rec(obj):
|
| 90 |
+
if isinstance(obj, dict):
|
| 91 |
+
for v in obj.values(): rec(v)
|
| 92 |
+
elif isinstance(obj, list):
|
| 93 |
+
for v in obj: rec(v)
|
| 94 |
+
elif isinstance(obj, str):
|
| 95 |
+
if pattern.match(obj): repo_ids.add(obj)
|
| 96 |
+
rec(cfg)
|
| 97 |
+
try:
|
| 98 |
+
subset_keys = list(cfg)[:5] if isinstance(cfg, dict) else []
|
| 99 |
+
preview = json.dumps({k: cfg[k] for k in subset_keys}, ensure_ascii=False)[:600]
|
| 100 |
+
except Exception:
|
| 101 |
+
pass
|
| 102 |
+
except Exception:
|
| 103 |
+
pass
|
| 104 |
+
return sorted(repo_ids), preview
|
| 105 |
+
|
| 106 |
+
def _prefetch_and_load_cpu():
|
| 107 |
+
"""Exécuté dans un thread au démarrage du Space (hors worker GPU)."""
|
| 108 |
+
try:
|
| 109 |
+
_log("[prefetch] snapshot_download (main)...")
|
| 110 |
+
local_dir = snapshot_download(
|
| 111 |
+
repo_id=MODEL_REPO_ID,
|
| 112 |
+
token=HF_TOKEN,
|
| 113 |
+
local_dir=None,
|
| 114 |
+
local_files_only=False,
|
| 115 |
+
)
|
| 116 |
+
_log(f"[prefetch] main done -> {local_dir}")
|
| 117 |
+
|
| 118 |
+
cfg_path = os.path.join(local_dir, "config.json")
|
| 119 |
+
nested, cfg_preview = _extract_repo_ids_from_config(cfg_path)
|
| 120 |
+
if cfg_preview:
|
| 121 |
+
_log(f"[config] preview: {cfg_preview}")
|
| 122 |
+
for rid in nested:
|
| 123 |
+
if rid == MODEL_REPO_ID:
|
| 124 |
+
continue
|
| 125 |
+
_log(f"[prefetch] nested repo: {rid} ...")
|
| 126 |
+
snapshot_download(repo_id=rid, token=HF_TOKEN, local_dir=None, local_files_only=False)
|
| 127 |
+
_log(f"[prefetch] nested repo: {rid} done")
|
| 128 |
+
|
| 129 |
+
# Forcer offline pendant le vrai chargement
|
| 130 |
+
old_off = os.environ.get("HF_HUB_OFFLINE")
|
| 131 |
+
os.environ["HF_HUB_OFFLINE"] = "1"
|
| 132 |
+
try:
|
| 133 |
+
_log("[load] from_pretrained(map_location='cpu')...")
|
| 134 |
+
m = PardiSpeech.from_pretrained(local_dir, map_location="cpu")
|
| 135 |
+
m.eval()
|
| 136 |
+
_MODEL["pardi"] = m
|
| 137 |
+
_MODEL["sr"] = getattr(m, "sampling_rate", 24000)
|
| 138 |
+
_log(f"[load] cpu OK (sr={_MODEL['sr']})")
|
| 139 |
+
finally:
|
| 140 |
+
if old_off is None:
|
| 141 |
+
os.environ.pop("HF_HUB_OFFLINE", None)
|
| 142 |
+
else:
|
| 143 |
+
os.environ["HF_HUB_OFFLINE"] = old_off
|
| 144 |
+
|
| 145 |
+
except BaseException as e:
|
| 146 |
+
_MODEL["err"] = e
|
| 147 |
+
_log(f"[EXC@preload] {type(e).__name__}: {e}")
|
| 148 |
+
_log(traceback.format_exc())
|
| 149 |
|
| 150 |
+
# Lance le préchargement (hors GPU) dès l’import
|
| 151 |
+
if _MODEL["thread"] is None:
|
| 152 |
+
_MODEL["thread"] = threading.Thread(target=_prefetch_and_load_cpu, daemon=True)
|
| 153 |
+
_MODEL["thread"].start()
|
| 154 |
+
|
| 155 |
+
def _move_to_cuda_if_available(m, logs_acc):
|
| 156 |
+
def L(msg): logs_acc.append(str(msg))
|
| 157 |
+
if torch.cuda.is_available():
|
| 158 |
+
L("[move] moving model to cuda...")
|
| 159 |
+
try:
|
| 160 |
+
m = m.to("cuda") # type: ignore[attr-defined]
|
| 161 |
+
L("[move] cuda OK")
|
| 162 |
+
except Exception as e:
|
| 163 |
+
L(f"[move] .to('cuda') failed: {e}. Keeping on CPU.")
|
| 164 |
+
else:
|
| 165 |
+
L("[move] cuda not available, keep CPU")
|
| 166 |
+
return m
|
| 167 |
+
|
| 168 |
+
# --------- UI callback (GPU) ----------
|
| 169 |
+
@spaces.GPU(duration=200)
|
| 170 |
def synthesize(
|
| 171 |
text: str,
|
| 172 |
+
debug: bool,
|
| 173 |
+
adv_sampling: bool, # Velocity Head sampling
|
| 174 |
ref_audio,
|
| 175 |
ref_text: str,
|
| 176 |
steps: int,
|
|
|
|
| 179 |
temperature: float,
|
| 180 |
max_seq_len: int,
|
| 181 |
seed: int,
|
| 182 |
+
lang_hint: str,
|
| 183 |
):
|
| 184 |
+
logs = []
|
| 185 |
+
def LOG(msg: str):
|
| 186 |
+
logs.append(str(msg))
|
| 187 |
+
joined = "\n".join(logs + _MODEL["logs"][-50:]) # mêle quelques logs de préchargement
|
| 188 |
+
if len(joined) > 12000:
|
| 189 |
+
joined = joined[-12000:]
|
| 190 |
+
return joined
|
| 191 |
|
| 192 |
+
try:
|
| 193 |
+
if HF_TOKEN:
|
| 194 |
+
try:
|
| 195 |
+
login(token=HF_TOKEN)
|
| 196 |
+
yield None, LOG("✅ HF login ok")
|
| 197 |
+
except Exception as e:
|
| 198 |
+
yield None, LOG(f"⚠️ HF login failed: {e}")
|
| 199 |
|
| 200 |
+
yield None, LOG("[env] " + _env_diag())
|
| 201 |
+
torch.manual_seed(int(seed))
|
| 202 |
+
os.environ.setdefault("CUDA_LAUNCH_BLOCKING", "1")
|
| 203 |
|
| 204 |
+
# Si le modèle n’est pas encore prêt, on attend jusqu’à 180s max ici
|
| 205 |
+
t0 = time.perf_counter()
|
| 206 |
+
while _MODEL["pardi"] is None and _MODEL["err"] is None:
|
| 207 |
+
elapsed = time.perf_counter() - t0
|
| 208 |
+
yield None, LOG(f"[init] still loading on CPU… {elapsed:.1f}s")
|
| 209 |
+
if elapsed > 180:
|
| 210 |
+
# dump de la stack du thread de préchargement pour debug
|
| 211 |
+
tid = _MODEL["thread"].ident if _MODEL["thread"] else None
|
| 212 |
+
if tid is not None:
|
| 213 |
+
frame = sys._current_frames().get(tid)
|
| 214 |
+
if frame is not None:
|
| 215 |
+
stack_txt = "".join(traceback.format_stack(frame))
|
| 216 |
+
yield None, LOG("[stack-final]\n" + stack_txt)
|
| 217 |
+
raise TimeoutError("Preload timeout (>180s)")
|
| 218 |
+
time.sleep(2.0)
|
|
|
|
| 219 |
|
| 220 |
+
if _MODEL["err"]:
|
| 221 |
+
raise _MODEL["err"]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 222 |
|
| 223 |
+
pardi = _MODEL["pardi"]
|
| 224 |
+
sr_out = _MODEL["sr"]
|
| 225 |
|
| 226 |
+
# Déplacement vers CUDA si possible
|
| 227 |
+
pardi = _move_to_cuda_if_available(pardi, logs)
|
| 228 |
+
yield None, LOG(f"[init] model ready on {'cuda' if torch.cuda.is_available() else 'cpu'}, sr={sr_out}")
|
| 229 |
+
|
| 230 |
+
# ---- Texte + prefix optionnel ----
|
| 231 |
+
txt = _normalize_text(text or "", lang_hint=lang_hint)
|
| 232 |
+
yield None, LOG(f"[text] {txt[:120]}{'...' if len(txt) > 120 else ''}")
|
| 233 |
+
|
| 234 |
+
steps = int(min(max(1, int(steps)), 16))
|
| 235 |
+
max_seq_len = int(min(max(50, int(max_seq_len)), 600))
|
| 236 |
+
|
| 237 |
+
prefix = None
|
| 238 |
+
if ref_audio is not None:
|
| 239 |
+
yield None, LOG("[prefix] encoding reference audio...")
|
| 240 |
+
if isinstance(ref_audio, str):
|
| 241 |
+
wav, sr = sf.read(ref_audio)
|
| 242 |
+
else:
|
| 243 |
+
sr, wav = ref_audio
|
| 244 |
+
wav = _to_mono_float32(wav)
|
| 245 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 246 |
+
wav_t = torch.from_numpy(wav).to(device).unsqueeze(0)
|
| 247 |
+
with torch.inference_mode():
|
| 248 |
+
prefix_tokens = pardi.patchvae.encode(wav_t) # type: ignore[attr-defined]
|
| 249 |
+
prefix = (ref_text or "", prefix_tokens[0])
|
| 250 |
+
yield None, LOG("[prefix] done.")
|
| 251 |
+
|
| 252 |
+
yield None, LOG(f"[run] has_prefix={prefix is not None}, steps={steps}, cfg={cfg}, cfg_ref={cfg_ref}, "
|
| 253 |
+
f"T={temperature}, max_seq_len={max_seq_len}, seed={seed}, adv_sampling={adv_sampling}")
|
| 254 |
+
|
| 255 |
+
# ---- Chemin rapide (comme le notebook) ----
|
| 256 |
with torch.inference_mode():
|
| 257 |
+
if adv_sampling:
|
| 258 |
+
try:
|
| 259 |
+
vparams = VelocityHeadSamplingParams(cfg_ref=float(cfg_ref), cfg=float(cfg), num_steps=int(steps))
|
| 260 |
+
except TypeError:
|
| 261 |
+
vparams = VelocityHeadSamplingParams(cfg_ref=float(cfg_ref), cfg=float(cfg),
|
| 262 |
+
num_steps=int(steps), temperature=float(temperature))
|
| 263 |
+
wavs, _ = pardi.text_to_speech([txt], prefix, max_seq_len=int(max_seq_len),
|
| 264 |
+
velocity_head_sampling_params=vparams)
|
| 265 |
+
else:
|
| 266 |
+
wavs, _ = pardi.text_to_speech([txt], prefix, max_seq_len=int(max_seq_len))
|
|
|
|
|
|
|
| 267 |
|
| 268 |
+
wav = wavs[0].detach().cpu().numpy().astype(np.float32)
|
| 269 |
+
yield (sr_out, wav), LOG("[ok] done.")
|
| 270 |
|
| 271 |
+
except Exception as e:
|
| 272 |
+
tb = traceback.format_exc()
|
| 273 |
+
yield None, LOG(f"[EXC] {type(e).__name__}: {e}\n{tb}")
|
| 274 |
+
|
| 275 |
+
# --------- UI ----------
|
| 276 |
def build_demo():
|
| 277 |
with gr.Blocks(title="Lina-speech / pardi-speech Demo") as demo:
|
| 278 |
gr.Markdown(
|
| 279 |
+
"### Lina-speech (pardi-speech) – Démo TTS\n"
|
| 280 |
+
"Génère de l'audio à partir de texte, avec ou sans prefix (audio de référence).\n"
|
| 281 |
+
"Chemin rapide par défaut (comme le notebook)."
|
| 282 |
)
|
|
|
|
| 283 |
with gr.Row():
|
| 284 |
text = gr.Textbox(label="Texte à synthétiser", lines=4, placeholder="Tape ton texte ici…")
|
| 285 |
with gr.Accordion("Prefix (optionnel)", open=False):
|
| 286 |
ref_audio = gr.Audio(sources=["upload", "microphone"], type="numpy", label="Audio de référence")
|
| 287 |
+
ref_text = gr.Textbox(label="Texte du prefix (si connu)", placeholder="Transcription du prefix (optionnel)")
|
| 288 |
with gr.Accordion("Options avancées", open=False):
|
| 289 |
with gr.Row():
|
| 290 |
steps = gr.Slider(1, 50, value=10, step=1, label="num_steps")
|
|
|
|
| 293 |
with gr.Row():
|
| 294 |
temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.05, label="Température")
|
| 295 |
max_seq_len = gr.Slider(50, 1200, value=300, step=10, label="max_seq_len (tokens audio)")
|
| 296 |
+
seed = gr.Number(value=0, precision=0, label="Seed")
|
| 297 |
+
lang_hint = gr.Dropdown(choices=["fr", "en"], value="fr", label="Langue (normalisation)")
|
| 298 |
+
with gr.Row():
|
| 299 |
+
debug = gr.Checkbox(value=False, label="Mode debug")
|
| 300 |
+
adv_sampling = gr.Checkbox(value=False, label="Sampling avancé (Velocity Head)")
|
| 301 |
|
| 302 |
btn = gr.Button("Synthétiser")
|
| 303 |
out_audio = gr.Audio(label="Sortie audio", type="numpy")
|
| 304 |
+
logs_box = gr.Textbox(label="Logs (live)", lines=28)
|
| 305 |
|
| 306 |
demo.queue(default_concurrency_limit=1, max_size=32)
|
|
|
|
| 307 |
btn.click(
|
| 308 |
fn=synthesize,
|
| 309 |
+
inputs=[text, debug, adv_sampling, ref_audio, ref_text, steps, cfg, cfg_ref, temperature, max_seq_len, seed, lang_hint],
|
| 310 |
+
outputs=[out_audio, logs_box],
|
| 311 |
+
api_name="synthesize",
|
| 312 |
)
|
| 313 |
return demo
|
| 314 |
|
| 315 |
if __name__ == "__main__":
|
| 316 |
+
build_demo().launch(ssr_mode=False)
|
| 317 |
+
# retrigger 2025-10-30T15:17:49+01:00
|
| 318 |
+
# retrigger 2025-10-30T16:37:47+01:00
|
tts/model/simple_gla.py
CHANGED
|
@@ -1,304 +1,291 @@
|
|
| 1 |
-
"""
|
| 2 |
-
Patched Simple GLA decoder for HF Spaces (ZeroGPU) — safe PyTorch-only paths.
|
| 3 |
-
|
| 4 |
-
- Forces FLA (flash-linear-attention) to avoid fused/Triton kernels during __init__ & forward
|
| 5 |
-
- Adds tolerant construction of SimpleGatedLinearAttention (backend="torch", fused=False)
|
| 6 |
-
- Falls back to a no-op GLA stub if FLA construction fails (for demo resilience)
|
| 7 |
-
- Keeps cache handling defensive to avoid NoneType unpack errors
|
| 8 |
-
|
| 9 |
-
Drop-in replacement for: tts/model/simple_gla.py
|
| 10 |
-
"""
|
| 11 |
-
|
| 12 |
import os
|
| 13 |
-
from typing import Optional, Dict, Any, Tuple, List, Union
|
| 14 |
-
|
| 15 |
-
# ---- Force safe runtime defaults (no Triton / no compile) ----
|
| 16 |
-
os.environ.setdefault("FLA_CONV_BACKEND", "torch")
|
| 17 |
-
os.environ.setdefault("FLA_USE_FAST_OPS", "0")
|
| 18 |
-
os.environ.setdefault("FLA_DISABLE_TRITON", "1") # ignored if not recognized
|
| 19 |
-
os.environ.setdefault("TORCH_COMPILE_DISABLE", "1")
|
| 20 |
-
os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
|
| 21 |
|
| 22 |
import torch
|
| 23 |
import torch.nn.functional as F
|
| 24 |
-
from torch import nn
|
| 25 |
-
|
| 26 |
from einops import rearrange
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
|
| 28 |
-
# ---------- Try importing FLA; otherwise define a stub ----------
|
| 29 |
-
try:
|
| 30 |
-
from fla.layers.simple_gla import SimpleGatedLinearAttention # type: ignore
|
| 31 |
-
from fla.models.utils import Cache # type: ignore
|
| 32 |
-
_FLA_AVAILABLE = True
|
| 33 |
-
except Exception:
|
| 34 |
-
_FLA_AVAILABLE = False
|
| 35 |
-
|
| 36 |
-
class SimpleGatedLinearAttention(nn.Module): # minimal stub (identity)
|
| 37 |
-
def __init__(self, *args, **kwargs):
|
| 38 |
-
super().__init__()
|
| 39 |
-
|
| 40 |
-
def forward(self, x, past_key_values=None, use_cache: bool = False, **kwargs):
|
| 41 |
-
# Match tuple output convention used by callers: (x, kv)
|
| 42 |
-
return x, None
|
| 43 |
-
|
| 44 |
-
# Fallback Cache typing
|
| 45 |
-
Cache = Dict[str, Any] # type: ignore
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
# Local layers / utils
|
| 49 |
from tts.layers.attention import CrossAttention
|
| 50 |
from tts.layers.ffn import SwiGLU
|
| 51 |
|
| 52 |
from .cache_utils import FLACache
|
| 53 |
from .config import SimpleGLADecoderConfig
|
| 54 |
from .registry import register_decoder
|
|
|
|
| 55 |
|
|
|
|
| 56 |
|
| 57 |
-
def
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
"""
|
| 63 |
-
# Prefer explicit mode to avoid backend auto-selection
|
| 64 |
-
try:
|
| 65 |
-
if hasattr(m, "mode"):
|
| 66 |
-
m.mode = "chunk" # safer than "recurrent" fused paths
|
| 67 |
-
except Exception:
|
| 68 |
-
pass
|
| 69 |
-
|
| 70 |
-
# For recent versions exposing implementation switches:
|
| 71 |
-
for attr, val in (("recurrent_impl", "naive"),
|
| 72 |
-
("chunk_impl", "naive"),
|
| 73 |
-
("fused", False),
|
| 74 |
-
("backend", "torch")):
|
| 75 |
-
if hasattr(m, attr):
|
| 76 |
-
try:
|
| 77 |
-
setattr(m, attr, val)
|
| 78 |
-
except Exception:
|
| 79 |
-
pass
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
def _make_tmix(dim: int, num_heads: int) -> SimpleGatedLinearAttention:
|
| 83 |
-
"""
|
| 84 |
-
Construct SimpleGatedLinearAttention using the safest available signature.
|
| 85 |
-
Falls back gracefully if kwargs are not supported by the installed FLA version.
|
| 86 |
-
"""
|
| 87 |
-
# Try most explicit signature first
|
| 88 |
-
try:
|
| 89 |
-
tmix = SimpleGatedLinearAttention(
|
| 90 |
-
hidden_size=dim,
|
| 91 |
-
num_heads=num_heads,
|
| 92 |
-
causal=True,
|
| 93 |
-
backend="torch", # key to avoid Triton
|
| 94 |
-
fused=False,
|
| 95 |
-
)
|
| 96 |
-
_force_safe_fla_impl(tmix)
|
| 97 |
-
return tmix
|
| 98 |
-
except TypeError:
|
| 99 |
-
pass
|
| 100 |
-
except Exception:
|
| 101 |
-
# If constructing with explicit kwargs fails for another reason,
|
| 102 |
-
# we will try progressively simpler signatures below.
|
| 103 |
-
pass
|
| 104 |
-
|
| 105 |
-
# Try without fused/backends but keep causal if supported
|
| 106 |
-
try:
|
| 107 |
-
tmix = SimpleGatedLinearAttention(
|
| 108 |
-
hidden_size=dim,
|
| 109 |
-
num_heads=num_heads,
|
| 110 |
-
causal=True,
|
| 111 |
-
)
|
| 112 |
-
_force_safe_fla_impl(tmix)
|
| 113 |
-
return tmix
|
| 114 |
-
except TypeError:
|
| 115 |
-
pass
|
| 116 |
-
except Exception:
|
| 117 |
-
pass
|
| 118 |
-
|
| 119 |
-
# Try minimal signature
|
| 120 |
-
try:
|
| 121 |
-
tmix = SimpleGatedLinearAttention(
|
| 122 |
-
hidden_size=dim,
|
| 123 |
-
num_heads=num_heads,
|
| 124 |
-
)
|
| 125 |
-
_force_safe_fla_impl(tmix)
|
| 126 |
-
return tmix
|
| 127 |
-
except Exception:
|
| 128 |
-
# Last resort: identity stub
|
| 129 |
-
return SimpleGatedLinearAttention()
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
def _cache_for_layer(cache: Optional[Cache], idx: int) -> Optional[Cache]:
|
| 133 |
-
"""
|
| 134 |
-
Extract per-layer cache if present; return None if structure is not compatible.
|
| 135 |
-
FLA expects either:
|
| 136 |
-
- cache["layers"][i]["conv_state"] being a tuple/list
|
| 137 |
-
- or a top-level cache dict with "conv_state" key
|
| 138 |
-
"""
|
| 139 |
-
if isinstance(cache, dict):
|
| 140 |
-
if "layers" in cache and isinstance(cache["layers"], (list, tuple)):
|
| 141 |
-
if idx < len(cache["layers"]) and isinstance(cache["layers"][idx], dict):
|
| 142 |
-
# Layer-specific cache entry
|
| 143 |
-
c = cache["layers"][idx]
|
| 144 |
-
# Validate conv_state shape
|
| 145 |
-
if isinstance(c.get("conv_state", None), (list, tuple)):
|
| 146 |
-
return c
|
| 147 |
-
# If not valid, ignore layer cache to prevent NoneType errors
|
| 148 |
-
return None
|
| 149 |
-
# Some layouts put conv states directly at top-level
|
| 150 |
-
if isinstance(cache.get("conv_state", None), (list, tuple)):
|
| 151 |
-
return cache
|
| 152 |
-
return None
|
| 153 |
|
|
|
|
|
|
|
| 154 |
|
| 155 |
-
|
| 156 |
-
|
| 157 |
|
|
|
|
|
|
|
| 158 |
def __init__(
|
| 159 |
self,
|
| 160 |
dim: int,
|
| 161 |
num_heads: int,
|
| 162 |
-
layer_idx: int
|
| 163 |
-
expand_k: float
|
| 164 |
-
expand_v: float
|
| 165 |
-
use_short_conv: bool
|
| 166 |
-
ffn_expansion_factor: int
|
| 167 |
):
|
| 168 |
super().__init__()
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
self.cmix = SwiGLU(dim,
|
| 175 |
-
|
| 176 |
-
# Norms
|
| 177 |
self.norm1 = nn.LayerNorm(dim)
|
| 178 |
self.norm2 = nn.LayerNorm(dim)
|
| 179 |
|
| 180 |
-
# (Optional) short conv placeholder
|
| 181 |
-
self.use_short_conv = use_short_conv
|
| 182 |
-
|
| 183 |
def forward(
|
| 184 |
self,
|
| 185 |
-
x
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
past_key_values=pkv,
|
| 198 |
-
use_cache=use_cache_flag,
|
| 199 |
)
|
| 200 |
-
x = y + x
|
| 201 |
x = self.cmix(self.norm2(x)) + x
|
| 202 |
return x
|
| 203 |
|
| 204 |
|
| 205 |
class DecoderBlockWithOptionalCrossAttention(nn.Module):
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
def __init__(self, decoder_block: nn.Module, crossatt: Optional[nn.Module] = None):
|
| 209 |
super().__init__()
|
|
|
|
| 210 |
self.decoder_block = decoder_block
|
| 211 |
self.crossatt = crossatt
|
| 212 |
|
| 213 |
def forward(
|
| 214 |
self,
|
| 215 |
x: torch.Tensor,
|
| 216 |
-
encoder_output:
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
|
|
|
|
|
|
| 220 |
) -> torch.Tensor:
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 224 |
x,
|
| 225 |
-
|
|
|
|
| 226 |
mask=crossatt_mask,
|
|
|
|
| 227 |
)
|
| 228 |
-
|
| 229 |
return x
|
| 230 |
|
| 231 |
|
| 232 |
@register_decoder("simple_gla")
|
| 233 |
class SimpleGLADecoder(nn.Module):
|
| 234 |
config = SimpleGLADecoderConfig
|
| 235 |
-
"""Decoder composed of a stack of SimpleGLABlock (+ optional cross-attention)."""
|
| 236 |
|
| 237 |
-
def __init__(self,
|
| 238 |
super().__init__()
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 260 |
)
|
| 261 |
-
|
| 262 |
-
if cross_every and (i % cross_every == 0):
|
| 263 |
-
# CrossAttention(dim, num_heads=num_heads) -> module expects (x, context, mask)
|
| 264 |
-
cross = CrossAttention(dim, num_heads=num_heads)
|
| 265 |
-
decoder_layers.append(DecoderBlockWithOptionalCrossAttention(block, cross))
|
| 266 |
|
| 267 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 268 |
|
| 269 |
-
# Backward compatibility with code expecting "prefill" API
|
| 270 |
def prefill(
|
| 271 |
self,
|
| 272 |
-
encoder_output:
|
| 273 |
decoder_input: torch.Tensor,
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
return self(
|
| 279 |
-
encoder_output=encoder_output,
|
| 280 |
-
decoder_input=decoder_input,
|
| 281 |
-
cache=cache,
|
| 282 |
-
text_freqs=text_freqs,
|
| 283 |
-
crossatt_mask=crossatt_mask,
|
| 284 |
-
)
|
| 285 |
|
| 286 |
-
def
|
| 287 |
self,
|
| 288 |
-
encoder_output:
|
| 289 |
decoder_input: torch.Tensor,
|
| 290 |
-
cache:
|
| 291 |
-
text_freqs:
|
| 292 |
-
crossatt_mask:
|
| 293 |
-
)
|
| 294 |
x = decoder_input
|
| 295 |
-
for
|
| 296 |
-
layer_cache = _cache_for_layer(cache, idx)
|
| 297 |
x = layer(
|
| 298 |
x,
|
| 299 |
-
encoder_output
|
| 300 |
text_freqs=text_freqs,
|
| 301 |
-
cache=
|
| 302 |
crossatt_mask=crossatt_mask,
|
| 303 |
)
|
| 304 |
return x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
import torch
|
| 4 |
import torch.nn.functional as F
|
|
|
|
|
|
|
| 5 |
from einops import rearrange
|
| 6 |
+
from fla.layers.simple_gla import SimpleGatedLinearAttention
|
| 7 |
+
from fla.models.utils import Cache
|
| 8 |
+
from sympy import num_digits
|
| 9 |
+
from torch import nn
|
| 10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
from tts.layers.attention import CrossAttention
|
| 12 |
from tts.layers.ffn import SwiGLU
|
| 13 |
|
| 14 |
from .cache_utils import FLACache
|
| 15 |
from .config import SimpleGLADecoderConfig
|
| 16 |
from .registry import register_decoder
|
| 17 |
+
from .shortconv import ShortConvBlock
|
| 18 |
|
| 19 |
+
if "GRAD_CKPT" in os.environ:
|
| 20 |
|
| 21 |
+
def maybe_grad_ckpt(f):
|
| 22 |
+
def grad_ckpt_f(*args, **kwargs):
|
| 23 |
+
return torch.utils.checkpoint.checkpoint(
|
| 24 |
+
f, *args, **kwargs, use_reentrant=False
|
| 25 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
+
return grad_ckpt_f
|
| 28 |
+
else:
|
| 29 |
|
| 30 |
+
def maybe_grad_ckpt(f):
|
| 31 |
+
return f
|
| 32 |
|
| 33 |
+
|
| 34 |
+
class SimpleGLABlock(nn.Module):
|
| 35 |
def __init__(
|
| 36 |
self,
|
| 37 |
dim: int,
|
| 38 |
num_heads: int,
|
| 39 |
+
layer_idx: int,
|
| 40 |
+
expand_k: float,
|
| 41 |
+
expand_v: float,
|
| 42 |
+
use_short_conv: bool,
|
| 43 |
+
ffn_expansion_factor: int,
|
| 44 |
):
|
| 45 |
super().__init__()
|
| 46 |
+
self.tmix = SimpleGatedLinearAttention(
|
| 47 |
+
hidden_size=dim,
|
| 48 |
+
num_heads=num_heads,
|
| 49 |
+
layer_idx=layer_idx,
|
| 50 |
+
)
|
| 51 |
+
self.cmix = SwiGLU(dim, ffn_expansion_factor)
|
|
|
|
|
|
|
| 52 |
self.norm1 = nn.LayerNorm(dim)
|
| 53 |
self.norm2 = nn.LayerNorm(dim)
|
| 54 |
|
|
|
|
|
|
|
|
|
|
| 55 |
def forward(
|
| 56 |
self,
|
| 57 |
+
x,
|
| 58 |
+
freqs: torch.Tensor | None = None,
|
| 59 |
+
text_freqs: torch.Tensor | None = None,
|
| 60 |
+
cache: Cache | None = None,
|
| 61 |
+
):
|
| 62 |
+
x = (
|
| 63 |
+
self.tmix(
|
| 64 |
+
self.norm1(x),
|
| 65 |
+
past_key_values=cache,
|
| 66 |
+
use_cache=cache is not None,
|
| 67 |
+
)[0]
|
| 68 |
+
+ x
|
|
|
|
|
|
|
| 69 |
)
|
|
|
|
| 70 |
x = self.cmix(self.norm2(x)) + x
|
| 71 |
return x
|
| 72 |
|
| 73 |
|
| 74 |
class DecoderBlockWithOptionalCrossAttention(nn.Module):
|
| 75 |
+
def __init__(self, decoder_block: nn.Module, crossatt: nn.Module | None = None):
|
|
|
|
|
|
|
| 76 |
super().__init__()
|
| 77 |
+
|
| 78 |
self.decoder_block = decoder_block
|
| 79 |
self.crossatt = crossatt
|
| 80 |
|
| 81 |
def forward(
|
| 82 |
self,
|
| 83 |
x: torch.Tensor,
|
| 84 |
+
encoder_output: torch.Tensor | None = None,
|
| 85 |
+
freqs: torch.Tensor | None = None,
|
| 86 |
+
text_freqs: torch.Tensor | None = None,
|
| 87 |
+
cache: Cache | None = None,
|
| 88 |
+
selfatt_mask: torch.Tensor | None = None,
|
| 89 |
+
crossatt_mask: torch.Tensor | list[torch.Tensor] | None = None,
|
| 90 |
) -> torch.Tensor:
|
| 91 |
+
x = self.decoder_block(
|
| 92 |
+
x,
|
| 93 |
+
freqs=freqs,
|
| 94 |
+
cache=cache,
|
| 95 |
+
)
|
| 96 |
+
if type(crossatt_mask) is list:
|
| 97 |
+
crossatt_mask = crossatt_mask[self.decoder_block.tmix.layer_idx]
|
| 98 |
+
if self.crossatt is not None:
|
| 99 |
+
x = x + self.crossatt(
|
| 100 |
x,
|
| 101 |
+
k=encoder_output,
|
| 102 |
+
text_freqs=text_freqs,
|
| 103 |
mask=crossatt_mask,
|
| 104 |
+
cache=cache,
|
| 105 |
)
|
| 106 |
+
|
| 107 |
return x
|
| 108 |
|
| 109 |
|
| 110 |
@register_decoder("simple_gla")
|
| 111 |
class SimpleGLADecoder(nn.Module):
|
| 112 |
config = SimpleGLADecoderConfig
|
|
|
|
| 113 |
|
| 114 |
+
def __init__(self, cfg: SimpleGLADecoderConfig):
|
| 115 |
super().__init__()
|
| 116 |
+
|
| 117 |
+
assert cfg.dim % cfg.num_heads == 0, "num_heads should divide dim"
|
| 118 |
+
assert cfg.blind_crossatt + (cfg.listen_read_crossatt is not None) < 2, (
|
| 119 |
+
"at most one specialized cross-attention"
|
| 120 |
+
)
|
| 121 |
+
|
| 122 |
+
self.head_dim = cfg.dim // cfg.num_heads
|
| 123 |
+
self.num_heads = cfg.num_heads
|
| 124 |
+
|
| 125 |
+
def simple_gla_block(i):
|
| 126 |
+
conv_layers = [] if cfg.conv_layers is None else cfg.conv_layers
|
| 127 |
+
if i in conv_layers:
|
| 128 |
+
return ShortConvBlock(
|
| 129 |
+
dim=cfg.dim,
|
| 130 |
+
kernel_size=4,
|
| 131 |
+
ffn_expansion_factor=cfg.ffn_expansion_factor,
|
| 132 |
+
layer_idx=i,
|
| 133 |
+
use_fast_conv1d=True,
|
| 134 |
+
)
|
| 135 |
+
|
| 136 |
+
else:
|
| 137 |
+
return SimpleGLABlock(
|
| 138 |
+
dim=cfg.dim,
|
| 139 |
+
num_heads=cfg.num_heads,
|
| 140 |
+
layer_idx=i,
|
| 141 |
+
expand_k=cfg.expand_k,
|
| 142 |
+
expand_v=cfg.expand_v,
|
| 143 |
+
use_short_conv=cfg.use_short_conv,
|
| 144 |
+
ffn_expansion_factor=cfg.ffn_expansion_factor,
|
| 145 |
+
)
|
| 146 |
+
|
| 147 |
+
def crossatt_block(i):
|
| 148 |
+
if i in cfg.crossatt_layer_idx:
|
| 149 |
+
return CrossAttention(
|
| 150 |
+
dim=cfg.dim,
|
| 151 |
+
num_heads=cfg.crossatt_num_heads,
|
| 152 |
+
dropout=cfg.crossatt_dropout,
|
| 153 |
+
layer_idx=i,
|
| 154 |
+
)
|
| 155 |
+
else:
|
| 156 |
+
return None
|
| 157 |
+
|
| 158 |
+
self.decoder_layers = nn.ModuleList(
|
| 159 |
+
[
|
| 160 |
+
DecoderBlockWithOptionalCrossAttention(
|
| 161 |
+
simple_gla_block(i),
|
| 162 |
+
crossatt_block(i),
|
| 163 |
+
)
|
| 164 |
+
for i in range(cfg.num_layers)
|
| 165 |
+
]
|
| 166 |
+
)
|
| 167 |
+
|
| 168 |
+
def forward(
|
| 169 |
+
self,
|
| 170 |
+
encoder_output: torch.Tensor,
|
| 171 |
+
decoder_input: torch.Tensor,
|
| 172 |
+
crossatt_mask: torch.Tensor | list[torch.Tensor] | None = None,
|
| 173 |
+
text_ids: torch.Tensor | None = None,
|
| 174 |
+
cache: FLACache | None = None,
|
| 175 |
+
):
|
| 176 |
+
x = decoder_input
|
| 177 |
+
text_freqs = None
|
| 178 |
+
|
| 179 |
+
for layer in self.decoder_layers:
|
| 180 |
+
x = maybe_grad_ckpt(layer)(
|
| 181 |
+
x,
|
| 182 |
+
encoder_output,
|
| 183 |
+
text_freqs=text_freqs,
|
| 184 |
+
cache=cache,
|
| 185 |
+
crossatt_mask=crossatt_mask,
|
| 186 |
)
|
| 187 |
+
return x
|
|
|
|
|
|
|
|
|
|
|
|
|
| 188 |
|
| 189 |
+
def init_cache(self, max_seq_len, device):
|
| 190 |
+
return FLACache(num_states=len(self.decoder_layers) + 1)
|
| 191 |
+
|
| 192 |
+
def init_initial_state(self, batch_size=1, scale=1e-2, device="cpu"):
|
| 193 |
+
return tuple(
|
| 194 |
+
nn.Parameter(
|
| 195 |
+
torch.randn(
|
| 196 |
+
batch_size,
|
| 197 |
+
self.num_heads,
|
| 198 |
+
self.head_dim,
|
| 199 |
+
self.head_dim,
|
| 200 |
+
device=device,
|
| 201 |
+
)
|
| 202 |
+
* scale
|
| 203 |
+
)
|
| 204 |
+
for _ in range(len(self.decoder_layers))
|
| 205 |
+
)
|
| 206 |
+
def init_initial_state_lora(self, lora:int=1, batch_size: int = 1, scale: float=1e-2, device: str="cpu"):
|
| 207 |
+
return tuple(
|
| 208 |
+
(
|
| 209 |
+
nn.Parameter(
|
| 210 |
+
torch.randn(
|
| 211 |
+
batch_size,
|
| 212 |
+
self.num_heads,
|
| 213 |
+
self.head_dim,
|
| 214 |
+
lora,
|
| 215 |
+
device=device,
|
| 216 |
+
)
|
| 217 |
+
* scale
|
| 218 |
+
),
|
| 219 |
+
nn.Parameter(
|
| 220 |
+
torch.randn(
|
| 221 |
+
batch_size,
|
| 222 |
+
self.num_heads,
|
| 223 |
+
lora,
|
| 224 |
+
self.head_dim,
|
| 225 |
+
device=device,
|
| 226 |
+
)
|
| 227 |
+
* scale
|
| 228 |
+
)
|
| 229 |
+
)
|
| 230 |
+
for _ in range(len(self.decoder_layers))
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
def _get_query(self, audio_inputs: torch.Tensor, layer_idx: int):
|
| 234 |
+
assert self.decoder_layers[layer_idx].crossatt is not None
|
| 235 |
+
x = audio_inputs
|
| 236 |
+
for _, layer in zip(range(layer_idx - 1), self.decoder_layers):
|
| 237 |
+
x = layer(x, None)
|
| 238 |
+
return self.decoder_layers[layer_idx].crossatt._query(x)
|
| 239 |
+
|
| 240 |
+
def forward_first_n_layers(
|
| 241 |
+
self,
|
| 242 |
+
encoder_output: torch.Tensor,
|
| 243 |
+
decoder_input: torch.Tensor,
|
| 244 |
+
n_first_layers: int,
|
| 245 |
+
crossatt_mask: torch.Tensor | None = None,
|
| 246 |
+
cache: FLACache | None = None,
|
| 247 |
+
):
|
| 248 |
+
x = decoder_input
|
| 249 |
+
if self.text_freqs_embd is not None:
|
| 250 |
+
text_freqs = torch.arange(encoder_output.shape[1], device=x.device)[None, :]
|
| 251 |
+
text_freqs = self.text_freqs_embd(text_freqs)
|
| 252 |
+
else:
|
| 253 |
+
text_freqs = None
|
| 254 |
+
|
| 255 |
+
for layer in self.decoder_layers[:n_first_layers]:
|
| 256 |
+
x = maybe_grad_ckpt(layer)(
|
| 257 |
+
x,
|
| 258 |
+
encoder_output,
|
| 259 |
+
text_freqs=text_freqs,
|
| 260 |
+
cache=cache,
|
| 261 |
+
crossatt_mask=crossatt_mask,
|
| 262 |
+
)
|
| 263 |
+
return x
|
| 264 |
|
|
|
|
| 265 |
def prefill(
|
| 266 |
self,
|
| 267 |
+
encoder_output: torch.Tensor,
|
| 268 |
decoder_input: torch.Tensor,
|
| 269 |
+
crossatt_mask: torch.Tensor | None = None,
|
| 270 |
+
cache: FLACache | None = None,
|
| 271 |
+
):
|
| 272 |
+
return self(encoder_output, decoder_input, cache=cache, crossatt_mask=crossatt_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 273 |
|
| 274 |
+
def decode_one(
|
| 275 |
self,
|
| 276 |
+
encoder_output: torch.Tensor,
|
| 277 |
decoder_input: torch.Tensor,
|
| 278 |
+
cache: Cache,
|
| 279 |
+
text_freqs: torch.Tensor | None = None,
|
| 280 |
+
crossatt_mask: torch.Tensor | None = None,
|
| 281 |
+
):
|
| 282 |
x = decoder_input
|
| 283 |
+
for layer in self.decoder_layers:
|
|
|
|
| 284 |
x = layer(
|
| 285 |
x,
|
| 286 |
+
encoder_output,
|
| 287 |
text_freqs=text_freqs,
|
| 288 |
+
cache=cache,
|
| 289 |
crossatt_mask=crossatt_mask,
|
| 290 |
)
|
| 291 |
return x
|