File size: 28,201 Bytes
579ca46 81a1074 579ca46 81a1074 579ca46 81a1074 f5eb735 579ca46 81a1074 579ca46 81a1074 f5eb735 81a1074 f5eb735 81a1074 579ca46 81a1074 f5eb735 81a1074 f5eb735 81a1074 f5eb735 81a1074 f5eb735 81a1074 f5eb735 81a1074 f5eb735 81a1074 cf28c2e 81a1074 579ca46 81a1074 579ca46 81a1074 579ca46 81a1074 579ca46 81a1074 f5eb735 81a1074 d737d8e 81a1074 f5eb735 81a1074 f5eb735 81a1074 2961d4e 81a1074 f5eb735 81a1074 2961d4e 81a1074 2961d4e 81a1074 2961d4e 81a1074 f5eb735 81a1074 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 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 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 | """CuteTTS-distill AX650 runtime (numpy + axengine, no torch/onnxruntime).
Pipeline (all heavy parts run on the NPU):
reference wav --vae_enc--> latents --locenc--> speech embeddings \
reference wav --logmel-> speaker_enc --> speaker embedding ---------> prompt embeds
prompt embeds --lm_prefill--> hidden / kv cache
loop:
hidden --dit_head (4 Euler steps)--> latent patch
latent patch --locenc_step--> next token embedding --lm_decode--\
latents --vae_dec--> waveform
"""
from __future__ import annotations
import json
import os
import time
import numpy as np
from . import audio as A
ASSETS = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "assets")
# --------------------------------------------------------------------- engines
class OnnxEngine:
"""Host-side verification engine (onnxruntime)."""
def __init__(self, path):
import onnxruntime as ort
self.sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"])
self.output_names = [o.name for o in self.sess.get_outputs()]
def run(self, feed):
return self.sess.run(self.output_names, feed)
def close(self):
pass
class AxEngine:
"""AX650 NPU engine (pyaxengine)."""
def __init__(self, path):
import axengine as axe
self.sess = axe.InferenceSession(path, providers=["AxEngineExecutionProvider"])
self.input_names = [i.name for i in self.sess.get_inputs()]
self.output_names = [o.name for o in self.sess.get_outputs()]
def run(self, feed):
return self.sess.run(self.output_names, {k: np.ascontiguousarray(feed[k]) for k in self.input_names})
def close(self):
self.sess = None
# ------------------------------------------------------------------ tokenizer
class Tokenizer:
"""SentencePiece with HuggingFace-compatible added-token splitting."""
def __init__(self, model_path):
import sentencepiece as spm
self.sp = spm.SentencePieceProcessor()
self.sp.Load(model_path)
folder = os.path.dirname(model_path)
self.specials = {"<unk>": 0, "<s>": 1, "</s>": 2, "<pad>": 3}
cfg_path = os.path.join(folder, "tokenizer_config.json")
if os.path.exists(cfg_path):
cfg = json.load(open(cfg_path))
for tid, entry in cfg.get("added_tokens_decoder", {}).items():
self.specials[entry["content"]] = int(tid)
added = os.path.join(folder, "added_tokens.json")
if os.path.exists(added):
for k, v in json.load(open(added)).items():
self.specials[k] = int(v)
self._keys = sorted(self.specials, key=len, reverse=True)
def encode(self, text):
out, buf, i = [], [], 0
while i < len(text):
hit = None
for k in self._keys:
if text.startswith(k, i):
hit = k
break
if hit is None:
buf.append(text[i])
i += 1
continue
chunk = "".join(buf).strip()
if chunk:
out += list(self.sp.encode(chunk, out_type=int))
out.append(self.specials[hit])
buf = []
i += len(hit)
chunk = "".join(buf).strip()
if chunk:
out += list(self.sp.encode(chunk, out_type=int))
return out
class CuteTTSAX:
def __init__(self, model_dir=None, assets_dir=ASSETS, engine="ax", verbose=True):
t_init = time.time()
self.root = os.path.abspath(model_dir or os.environ.get("CUTETTS_MODEL_DIR", "."))
self.assets = os.path.abspath(assets_dir)
self.cfg = json.load(open(os.path.join(self.assets, "config.json")))
self.verbose = verbose
base_mk = AxEngine if engine == "ax" else OnnxEngine
def mk(path_):
# AX CMM is a shared pool; retry briefly on transient load failures
import gc
last = None
for _ in range(4):
try:
return base_mk(path_)
except Exception as exc: # noqa: BLE001
last = exc
gc.collect()
time.sleep(0.4)
raise last
onnx_dir = os.path.join(self.root, "onnx")
def path(name):
for cand in (os.path.join(self.root, name + ".axmodel"),
os.path.join(self.root, "models", name + ".axmodel"),
os.path.join(onnx_dir, name + ".onnx")):
if os.path.exists(cand):
return cand
raise FileNotFoundError(name)
self._find = path
self._vae_enc = mk(path("vae_enc"))
self.m_vae_stream = None
self.vae_state_shapes = None
self._stream_path = None
sp = os.path.join(self.assets, "vae_stream.json")
if os.path.exists(sp):
try:
self._stream_path = path("vae_dec_stream")
self.vae_state_shapes = [tuple(int(v) for v in sh)
for sh in json.load(open(sp))["states"]]
except FileNotFoundError:
self._stream_path = None
self.m_spk = mk(path("spk_body"))
self.m_loc_step = mk(path("locenc_step"))
self.m_dit = mk(path("dit_step"))
self.m_pref = mk(path("lm_prefill"))
self.m_dec = mk(path("lm_decode"))
self.m_vae_dec = {}
self._dec_paths = {}
for lb in (80, 128, 256):
try:
self._dec_paths[lb] = path(f"vae_dec_l{lb}")
except FileNotFoundError:
pass
self.embed = np.load(os.path.join(self.assets, "embed_tokens.npy")).astype(np.float32)
self.spk_proj = np.load(os.path.join(self.assets, "lm_speaker_proj.npy"))
self.mel_filter = np.load(os.path.join(self.assets, "mel_filter.npy"))
self.spk_bn = (np.load(os.path.join(self.assets, "spk_bn_mean.npy")),
np.load(os.path.join(self.assets, "spk_bn_var.npy")),
np.load(os.path.join(self.assets, "spk_bn_w.npy")),
np.load(os.path.join(self.assets, "spk_bn_b.npy")),
np.load(os.path.join(self.assets, "spk_linear_w.npy")),
np.load(os.path.join(self.assets, "spk_linear_b.npy")))
self.mel_window = np.load(os.path.join(self.assets, "mel_window.npy"))
self.tok = Tokenizer(self._find_tokenizer())
self.scale = float(self.cfg["speech_scaling_factor"])
self.bias = float(self.cfg["speech_bias_factor"])
self.L = int(self.cfg["lm_prefill_len"])
self.B = int(self.cfg["lm_kv_capacity"])
self.sr = int(self.cfg["sample_rate"])
self.rng = np.random.default_rng(42)
self.last_timing = {}
self.init_seconds = time.time() - t_init
def _find_tokenizer(self):
for p in (os.path.join(self.assets, "tokenizer", "tokenizer.model"),
os.path.join(self.root, "tokenizer", "tokenizer.model"),
os.path.join(self.root, "tokenizer.model")):
if os.path.exists(p):
return p
raise FileNotFoundError("tokenizer.model")
# ------------------------------------------------------------------ helpers
@staticmethod
def _rms(x):
return np.sqrt(np.mean(x * x) + 1e-12)
def _ref_latents(self, wav24):
cap = int(self.cfg["vae_enc_samples"])
hop = int(self.cfg["vae_hop"])
x = np.asarray(wav24[:cap], dtype=np.float32)
nreal = len(x) // hop
if len(x) < cap:
x = np.pad(x, (0, cap - len(x)))
lat = self._vae_enc.run({"wav": x.reshape(1, 1, -1)})[0].reshape(-1, 64)
return lat[:max(nreal, 1)].astype(np.float32)
def _speaker_embedding(self, wav16):
mel = A.logmel(wav16, self.mel_filter, self.mel_window,
n_fft=self.cfg["mel"]["n_fft"], hop=self.cfg["mel"]["hop"],
win_length=self.cfg["mel"]["win_length"], eps=self.cfg["mel"]["eps"])
T = int(self.cfg["speaker_mel_frames"])
if mel.shape[1] < T:
mel = np.pad(mel, ((0, 0), (0, T - mel.shape[1])))
else:
mel = mel[:, :T]
stats = self.m_spk.run({"logmel": mel[None]})[0]
# BN + linear + L2-normalize on host (tiny, keeps the NPU graph conv-only)
mean, var, w, b, lw, lb = self.spk_bn
y = (stats - mean) / np.sqrt(var + 1e-5) * w + b
emb = y @ lw.T + lb
return (emb / np.linalg.norm(emb, axis=1, keepdims=True)).astype(np.float32)
def _prompt_embeds(self, text, ref_lat_scaled, spk_emb):
p = self.cfg["prompt"]
ids_prefix = self.tok.encode(p["voice_prefix"])
ids_mid = self.tok.encode(p["voice_mid"])
ids_suffix = self.tok.encode(p["voice_suffix_fmt"].format(text))
# speech latents -> patches (pad to even)
T = ref_lat_scaled.shape[0]
if T % 2:
ref_lat_scaled = np.concatenate([ref_lat_scaled, np.zeros((1, 64), np.float32)], 0)
patches = ref_lat_scaled.reshape(-1, 2, 64)
n_patch = patches.shape[0]
if n_patch > int(self.cfg["locenc_prompt_patches"]):
n_patch = int(self.cfg["locenc_prompt_patches"])
patches = patches[:n_patch]
# LocEnc is per-patch: reuse the single-patch graph (cheap, 23..63 calls)
speech_emb = np.concatenate(
[self.m_loc_step.run({"patches": patches[i:i + 1][None].astype(np.float32)})[0][0]
for i in range(n_patch)], axis=0)
seq_ids = ids_prefix + [0] + ids_mid + [0] * n_patch + ids_suffix
n = len(seq_ids)
if n > self.L:
raise ValueError(f"prompt too long ({n} > {self.L}); use a shorter reference/text")
emb = self.embed[np.asarray(seq_ids, dtype=np.int64)]
spk_pos = len(ids_prefix)
emb[spk_pos] = (spk_emb @ self.spk_proj.T)[0]
s0 = spk_pos + 1 + len(ids_mid)
emb[s0:s0 + n_patch] = speech_emb
out = np.zeros((1, self.L, 1024), np.float32)
out[0, :n] = emb
return out, n
@staticmethod
def _rope(pos):
inv = 1.0 / (1e6 ** (np.arange(0, 128, 2, dtype=np.float32) / 128.0))
freqs = np.float32(pos) * inv
emb = np.concatenate([freqs, freqs])
return (np.cos(emb).astype(np.float32).reshape(1, 1, 1, 128),
np.sin(emb).astype(np.float32).reshape(1, 1, 1, 128))
def _get_dec(self, bucket):
if bucket not in self.m_vae_dec:
if bucket not in self._dec_paths:
raise ValueError(f"no VAE decoder for {bucket} latents")
self.m_vae_dec[bucket] = self._find_engine(self._dec_paths[bucket])
return self.m_vae_dec[bucket]
def _find_engine(self, path_):
mk = AxEngine if path_.endswith(".axmodel") else OnnxEngine
import gc
last = None
for _ in range(4):
try:
return mk(path_)
except Exception as exc: # noqa: BLE001
last = exc
gc.collect()
time.sleep(0.4)
raise last
def _get_stream_dec(self):
if self.m_vae_stream is None and self._stream_path is not None:
self.m_vae_stream = self._find_engine(self._stream_path)
return self.m_vae_stream
def warmup(self):
"""Run one tiny synthesis so the first real request does not pay NPU init cost."""
try:
import tempfile, wave as _wave # noqa: F401
ref = os.path.join(self.assets, "warmup.wav")
if not os.path.exists(ref):
import wave
pcm = np.zeros(24000, np.int16)
with wave.open(ref, "wb") as wf:
wf.setnchannels(1); wf.setsampwidth(2); wf.setframerate(self.sr)
wf.writeframes(pcm.tobytes())
for _ in self.synthesize_stream("warm up", ref, max_new=2, seed=1):
pass
return True
except Exception: # noqa: BLE001
return False
def _prepare_reference(self, reference_audio, ref_max_seconds=5.0):
"""Reference latents + speaker embedding, cached per (path, size, seconds)."""
try:
key = (os.path.abspath(reference_audio), os.path.getsize(reference_audio),
float(ref_max_seconds))
except OSError:
key = None
cache = getattr(self, "_ref_cache", None)
if cache is None:
cache = self._ref_cache = {}
if key is not None and key in cache:
self.ref_cache_hit = True
return cache[key]
# disk cache (shared across processes, like the C++ runtime)
disk_path = None
if key is not None:
import hashlib
h = hashlib.md5("|".join(str(k) for k in key).encode()).hexdigest()[:16]
disk_path = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
"..", "cache", "ref", h + ".npz")
disk_path = os.path.normpath(disk_path)
if os.path.exists(disk_path):
try:
z = np.load(disk_path)
cached = (z["lat"].astype(np.float32), z["spk"].astype(np.float32))
cache[key] = cached
self.ref_cache_hit = True
return cached
except Exception: # noqa: BLE001
pass
self.ref_cache_hit = False
wav, sr = A.read_wav(reference_audio)
wav24 = A.resample_to(wav, sr, self.sr)
cap = int(self.cfg["vae_enc_samples"]) / float(self.sr)
wav24 = wav24[: int(min(ref_max_seconds, cap) * self.sr)]
if len(wav24) < 2 * self.sr:
wav24 = np.tile(wav24, int(np.ceil(2 * self.sr / len(wav24))))
wav16 = A.resample_to(wav, sr, 16000)[: 8 * 16000]
if len(wav16) < 32000:
wav16 = np.tile(wav16, int(np.ceil(32000 / len(wav16))))
lat = self._ref_latents(wav24)
spk = self._speaker_embedding(wav16)
if key is not None:
cache[key] = (lat, spk)
if disk_path is not None:
try:
os.makedirs(os.path.dirname(disk_path), exist_ok=True)
np.savez(disk_path, lat=lat, spk=spk)
except Exception: # noqa: BLE001
pass
return lat, spk
# ------------------------------------------------------------------ streaming
def _decode_window(self, vae_lat, start, end, length=128):
"""Exact VAE decode of latents [start, end) with a causal sliding window.
The DAC decoder is causal with a ~60-latent receptive field (measured: zeroing
the first 60 latents leaves all later samples bit-identical), so any window
with >=64 latents of context reproduces the offline decode exactly.
"""
w0 = max(0, end - length)
win = np.zeros((1, 64, length), np.float32)
n = end - w0
win[0, :, :n] = np.stack(vae_lat[w0:end], axis=1)
out = self._get_dec(length).run({"latent": win})[0][0, 0]
return out[(start - w0) * 1920:(end - w0) * 1920]
def _stream_window(self, chunk=16):
"""Smallest exact window that fits 64 context + ``chunk`` new latents."""
if chunk <= 16 and 80 in self._dec_paths:
return 80
if chunk <= 64 and 128 in self._dec_paths:
return 128
return max(self._dec_paths)
# --------------------------------------------------------------------- api
@staticmethod
def _merge_timing(agg, tm):
"""多段合成时聚合计时:TTFT/TTFA 取第一段,其余累加。"""
if tm is None:
return agg
if agg is None:
return dict(tm)
for k in ("total", "audio_seconds", "steps", "chunks", "dit", "lm_decode"):
if k in tm:
agg[k] = agg.get(k, 0) + tm[k]
return agg
# ---------------------------------------------------------------- long text
def synthesize_long(self, text, reference_audio, max_seconds=18.0, **kw):
"""长文本:按句切分后逐段合成,返回拼接好的 PCM。
单次请求上限约 20.3 s(KV 256 prompt + 127 patch),超出会被截断;
这里按 ``max_seconds``(默认 18 s)切段,逐段生成后拼起来。
"""
from .text import split_text
parts = split_text(text, max_seconds)
self.segments = len(parts)
outs, agg = [], None
for p in parts:
outs.append(self.synthesize(p, reference_audio, **kw))
agg = self._merge_timing(agg, self.last_timing)
if agg is not None:
agg["segments"] = len(parts)
self.last_timing = agg
return np.concatenate(outs) if outs else np.zeros(0, np.float32)
def synthesize_stream_long(self, text, reference_audio, max_seconds=18.0, **kw):
"""长文本流式:按句切段,一段接一段地流式输出(音色缓存后每段 ≈0.5 s 首包)。"""
from .text import split_text
parts = split_text(text, max_seconds)
self.segments = len(parts)
agg = None
for p in parts:
yield from self.synthesize_stream(p, reference_audio, **kw)
agg = self._merge_timing(agg, self.last_timing)
if agg is not None:
agg["segments"] = len(parts)
agg["streaming"] = True
self.last_timing = agg
def synthesize(self, text, reference_audio, max_new=None, cfg_strength=None, seed=42,
ref_max_seconds=5.0, rng=None):
t0 = time.time()
cfg_strength = self.cfg["cfg_strength"] if cfg_strength is None else float(cfg_strength)
max_new = int(self.cfg["max_new_patches"] if max_new is None else max_new)
self.rng = rng if rng is not None else np.random.default_rng(seed)
# ---- reference features -------------------------------------------------
t1 = time.time()
lat, spk = self._prepare_reference(reference_audio, ref_max_seconds)
t2 = t3 = time.time()
lat_scaled = (lat + self.bias) * self.scale
embeds, plen = self._prompt_embeds(text, lat_scaled, spk)
t4 = time.time()
# ---- prefill ------------------------------------------------------------
pre = self.m_pref.run({"embeds": embeds})
hidden = pre[0][0, plen - 1].reshape(1, 1024)
stop = pre[1].reshape(2)
kvs = []
for i in range(7):
kk = np.zeros((1, 8, self.B, 128), np.float32)
vv = np.zeros((1, 8, self.B, 128), np.float32)
kk[0, :, :plen] = pre[2 + 2 * i][0, :, :plen]
vv[0, :, :plen] = pre[3 + 2 * i][0, :, :plen]
kvs += [kk, vv]
t5 = time.time()
# ---- autoregressive loop ------------------------------------------------
mask = np.full((1, 1, 1, self.B + 1), -30.0, np.float32)
mask[..., :plen] = 0.0
cond = np.zeros((1, 2, 64), np.float32)
cfg_t = np.array([cfg_strength], np.float32)
latents = []
pos = plen
n_steps = 0
t_dit = t_dec = 0.0
for step in range(max_new):
stop_now = int(np.argmax(stop)) == 1
x0 = self.rng.standard_normal((1, 2, 64)).astype(np.float32)
t_a = time.time()
x = x0
for k in range(int(self.cfg["diffusion_steps"])):
v = self.m_dit.run({"x": x, "z": hidden, "cond": cond, "speaker": spk,
"cfg_strength": cfg_t, "t": np.array([k * 0.25], np.float32),
"dt": np.array([0.25], np.float32)})[0]
x = x + v * 0.25
pred = x
t_dit += time.time() - t_a
latents.append(pred.copy())
n_steps += 1
if stop_now:
break
cond = pred
emb = self.m_loc_step.run({"patches": pred[None].astype(np.float32)})[0]
if pos >= self.B - 1:
break
mask[..., pos] = 0.0
mask[..., self.B] = 0.0
cos, sin = self._rope(pos)
feed = {"embed": emb, "mask_add": mask, "cos": cos, "sin": sin}
for i in range(7):
feed[f"pk{i}"] = kvs[2 * i]
feed[f"pv{i}"] = kvs[2 * i + 1]
t_a = time.time()
d = self.m_dec.run(feed)
t_dec += time.time() - t_a
hidden = d[0].reshape(1, 1024)
stop = d[1].reshape(2)
for i in range(7):
kvs[2 * i][0, :, pos] = d[2 + 2 * i][0, :, 0]
kvs[2 * i + 1][0, :, pos] = d[3 + 2 * i][0, :, 0]
pos += 1
t6 = time.time()
# ---- decode -------------------------------------------------------------
lat = np.concatenate(latents, axis=0) # [N,2,64]
self.last_patches = lat.copy()
vae_lat = (lat / self.scale - self.bias).reshape(-1, 64).T[None] # [1,64,2N]
n = vae_lat.shape[-1]
bucket = next((b for b in sorted(self._dec_paths) if n <= b), max(self._dec_paths))
if n > bucket:
raise ValueError(f"audio too long for available decoders ({n} > {bucket} latents)")
padded = np.zeros((1, 64, bucket), np.float32)
padded[:, :, :n] = vae_lat
self.last_latents = vae_lat
self.last_bucket = bucket
wav_out = self._get_dec(bucket).run({"latent": padded})[0][0, 0, : n * int(self.cfg["vae_hop"])]
t7 = time.time()
self.last_timing = {
"vae_enc": t2 - t1, "speaker_enc": t3 - t2, "prompt": t4 - t3,
"prefill": t5 - t4, "ar_loop": t6 - t5, "dit": t_dit, "lm_decode": t_dec,
"vae_dec": t7 - t6, "total": t7 - t0, "steps": n_steps,
"audio_seconds": len(wav_out) / self.sr,
}
return wav_out.astype(np.float32)
def synthesize_stream(self, text, reference_audio, max_new=None, cfg_strength=None, seed=42,
ref_max_seconds=5.0, rng=None, first_chunk=4, chunk=16, exact=True):
"""True streaming synthesis: yields float32 PCM chunks while the AR loop runs.
Nothing is sentence-split: the AR loop keeps generating latents and audio is
decoded/emitted every ``chunk`` latents (1.28 s) after an initial
``first_chunk`` latents (0.32 s), using an exact 80-latent sliding window
(64 latents of causal context + the new ones).
"""
t0 = time.time()
cfg_strength = self.cfg["cfg_strength"] if cfg_strength is None else float(cfg_strength)
max_new = int(self.cfg["max_new_patches"] if max_new is None else max_new)
self.rng = rng if rng is not None else np.random.default_rng(seed)
lat, spk = self._prepare_reference(reference_audio, ref_max_seconds)
lat_scaled = (lat + self.bias) * self.scale
embeds, plen = self._prompt_embeds(text, lat_scaled, spk)
pre = self.m_pref.run({"embeds": embeds})
hidden = pre[0][0, plen - 1].reshape(1, 1024)
stop = pre[1].reshape(2)
kvs = []
for i in range(7):
kk = np.zeros((1, 8, self.B, 128), np.float32)
vv = np.zeros((1, 8, self.B, 128), np.float32)
kk[0, :, :plen] = pre[2 + 2 * i][0, :, :plen]
vv[0, :, :plen] = pre[3 + 2 * i][0, :, :plen]
kvs += [kk, vv]
mask = np.full((1, 1, 1, self.B + 1), -30.0, np.float32)
mask[..., :plen] = 0.0
cond = np.zeros((1, 2, 64), np.float32)
cfg_t = np.array([cfg_strength], np.float32)
vae_lat = [] # VAE-space latents, 64 floats each
emitted = 0
win_len = self._stream_window(chunk)
if exact or self._get_stream_dec() is None:
states = None
else:
states = [np.zeros(sh, np.float32) for sh in self.vae_state_shapes]
pos = plen
n_steps = 0
ttft = None
ttft_latent = None
n_chunks = 0
t_dit = t_dec = 0.0
def emit_patch(pred):
"""stateful per-patch decode (160 ms per call)"""
nonlocal states, ttft, n_chunks
vae = pred[0] / self.scale - self.bias # scaled (LM) space -> VAE space
lat2 = np.ascontiguousarray(vae.T[None].astype(np.float32)) # [1,64,2]
feed = {"latent": lat2}
for i, st in enumerate(states):
feed[f"s{i}"] = st
out = self.m_vae_stream.run(feed)
# copy: the engine reuses its output buffers, and we feed them back next step
states = [np.array(o, copy=True) for o in out[1:]]
n_chunks += 1
if ttft is None:
ttft = time.time() - t0
return np.array(out[0][0, 0], copy=True)
def emit(upto):
nonlocal emitted, ttft, n_chunks
audio = self._decode_window(vae_lat, emitted, upto, win_len)
emitted = upto
n_chunks += 1
if ttft is None:
ttft = time.time() - t0
return audio
for step in range(max_new):
stop_now = int(np.argmax(stop)) == 1
x0 = self.rng.standard_normal((1, 2, 64)).astype(np.float32)
t_a = time.time()
x = x0
for k in range(int(self.cfg["diffusion_steps"])):
v = self.m_dit.run({"x": x, "z": hidden, "cond": cond, "speaker": spk,
"cfg_strength": cfg_t, "t": np.array([k * 0.25], np.float32),
"dt": np.array([0.25], np.float32)})[0]
x = x + v * 0.25
t_dit += time.time() - t_a
cond = x
n_steps += 1
for c in range(2):
vae_lat.append((x[0, c] / self.scale - self.bias).astype(np.float32))
if ttft_latent is None:
ttft_latent = time.time() - t0
if states is not None:
yield emit_patch(x)
if stop_now:
if states is None and emitted < len(vae_lat):
yield emit(len(vae_lat))
break
emb = self.m_loc_step.run({"patches": x[None].astype(np.float32)})[0]
if pos >= self.B - 1:
if emitted < len(vae_lat):
yield emit(len(vae_lat))
break
mask[..., pos] = 0.0
mask[..., self.B] = 0.0
cos, sin = self._rope(pos)
feed = {"embed": emb, "mask_add": mask, "cos": cos, "sin": sin}
for i in range(7):
feed[f"pk{i}"] = kvs[2 * i]
feed[f"pv{i}"] = kvs[2 * i + 1]
t_a = time.time()
d = self.m_dec.run(feed)
t_dec += time.time() - t_a
hidden = d[0].reshape(1, 1024)
stop = d[1].reshape(2)
for i in range(7):
kvs[2 * i][0, :, pos] = d[2 + 2 * i][0, :, 0]
kvs[2 * i + 1][0, :, pos] = d[3 + 2 * i][0, :, 0]
pos += 1
if states is None:
need = first_chunk if emitted == 0 else chunk
if len(vae_lat) - emitted >= need:
yield emit(len(vae_lat))
if states is None and emitted < len(vae_lat):
yield emit(len(vae_lat))
ttfa = ttft or (time.time() - t0)
self.last_timing = {
# TTFT = 首个 latent 产出;TTFA = 首个音频包产出("ttft" 保留为 TTFA 别名)
"ttft_latent": ttft_latent or ttfa, "ttfa": ttfa, "ttft": ttfa,
"total": time.time() - t0,
"audio_seconds": len(vae_lat) / 12.5, "steps": n_steps,
"chunks": n_chunks, "dit": t_dit, "lm_decode": t_dec,
"streaming": True, "streaming_mode": ("stateful" if states is not None else f"window{win_len}"),
}
|