File size: 16,790 Bytes
40ba644 | 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 | #!/usr/bin/env python3
"""
Phase 1: Deep Architecture Analysis of Qwen3-TTS for ExecuTorch Export
======================================================================
Loads the model, maps all modules with parameter counts, traces a real
voice-clone inference to capture shapes, and identifies export blockers.
"""
import sys
import os
import time
import json
import numpy as np
import torch
import torch.nn as nn
# ββ paths ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
MODEL_PATH = os.path.expanduser("~/Documents/Qwen3-TTS/models/1.7B-Base")
VENV_SITE = os.path.expanduser("~/Documents/Qwen3-TTS/.venv/lib/python3.10/site-packages")
QWEN_TTS_SRC = os.path.expanduser("~/Documents/Qwen3-TTS")
# Ensure the venv's site-packages is on the path so qwen_tts can be imported
if VENV_SITE not in sys.path:
sys.path.insert(0, VENV_SITE)
if QWEN_TTS_SRC not in sys.path:
sys.path.insert(0, QWEN_TTS_SRC)
# ββ helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def count_params(module: nn.Module) -> int:
return sum(p.numel() for p in module.parameters())
def fmt(n: int) -> str:
if n >= 1e9:
return f"{n / 1e9:.1f}B"
if n >= 1e6:
return f"{n / 1e6:.1f}M"
if n >= 1e3:
return f"{n / 1e3:.1f}K"
return str(n)
def param_table(module: nn.Module, prefix: str = "", depth: int = 0, max_depth: int = 3):
"""Print a hierarchical parameter table."""
total = count_params(module)
indent = " " * depth
name = prefix or module.__class__.__name__
print(f"{indent}{name}: {fmt(total)} params")
if depth < max_depth:
for child_name, child in module.named_children():
child_prefix = f"{prefix}.{child_name}" if prefix else child_name
param_table(child, child_prefix, depth + 1, max_depth)
# ββ 1. Load Model βββββββββββββββββββββββββββββββββββββββββββββββββββ
print("=" * 70)
print("PHASE 1: Deep Architecture Analysis β Qwen3-TTS 1.7B-Base")
print("=" * 70)
print("\n[1/5] Loading model from", MODEL_PATH)
t0 = time.time()
from qwen_tts.core.models.configuration_qwen3_tts import Qwen3TTSConfig
from qwen_tts.core.models.modeling_qwen3_tts import (
Qwen3TTSForConditionalGeneration,
mel_spectrogram,
)
config = Qwen3TTSConfig.from_pretrained(MODEL_PATH)
# Force SDPA attention for exportability
model = Qwen3TTSForConditionalGeneration.from_pretrained(
MODEL_PATH,
config=config,
torch_dtype=torch.float32,
attn_implementation="sdpa",
device_map="cpu",
)
model.eval()
print(f" Loaded in {time.time() - t0:.1f}s")
# ββ 2. Parameter Map ββββββββββββββββββββββββββββββββββββββββββββββββ
print("\n[2/5] Parameter Map (hierarchical)")
print("-" * 60)
param_table(model, "Qwen3TTSForConditionalGeneration", max_depth=4)
print("\n--- Top-level component sizes ---")
components = {
"speaker_encoder": model.speaker_encoder,
"talker": model.talker,
"talker.model": model.talker.model,
"talker.text_projection": model.talker.text_projection,
"talker.codec_head": model.talker.codec_head,
"talker.code_predictor": model.talker.code_predictor,
}
for name, mod in components.items():
print(f" {name:40s}: {fmt(count_params(mod)):>8s} params")
if model.speech_tokenizer is not None and hasattr(model.speech_tokenizer, 'model'):
st = model.speech_tokenizer.model # Qwen3TTSTokenizerV2Model (nn.Module)
print(f" {'speech_tokenizer.model':40s}: {fmt(count_params(st)):>8s} params")
if hasattr(st, 'encoder'):
print(f" {'speech_tokenizer.model.encoder':40s}: {fmt(count_params(st.encoder)):>8s} params")
if hasattr(st, 'decoder'):
print(f" {'speech_tokenizer.model.decoder':40s}: {fmt(count_params(st.decoder)):>8s} params")
# ββ 3. Config Summary βββββββββββββββββββββββββββββββββββββββββββββββ
print("\n[3/5] Key Config Values")
print("-" * 60)
tc = config.talker_config
cpc = tc.code_predictor_config
sec = config.speaker_encoder_config
info = {
"Speaker Encoder": {
"mel_dim": sec.mel_dim,
"enc_dim (output)": sec.enc_dim,
"enc_channels": sec.enc_channels,
"sample_rate": sec.sample_rate,
},
"Talker (Main LM)": {
"hidden_size": tc.hidden_size,
"num_hidden_layers": tc.num_hidden_layers,
"num_attention_heads": tc.num_attention_heads,
"num_key_value_heads": tc.num_key_value_heads,
"head_dim": tc.head_dim,
"intermediate_size": tc.intermediate_size,
"text_vocab_size": tc.text_vocab_size,
"codec_vocab_size": tc.vocab_size,
"num_code_groups": tc.num_code_groups,
"max_position_embeddings": tc.max_position_embeddings,
"rope_scaling": tc.rope_scaling,
},
"Code Predictor": {
"hidden_size": cpc.hidden_size,
"num_hidden_layers": cpc.num_hidden_layers,
"num_attention_heads": cpc.num_attention_heads,
"num_key_value_heads": cpc.num_key_value_heads,
"num_code_groups": cpc.num_code_groups,
"vocab_size": cpc.vocab_size,
},
}
for section, kvs in info.items():
print(f"\n {section}:")
for k, v in kvs.items():
print(f" {k:35s}: {v}")
# ββ 4. Trace Real Inference βββββββββββββββββββββββββββββββββββββββββ
print("\n[4/5] Tracing Real Voice-Clone Inference")
print("-" * 60)
# Create synthetic reference audio: 3 seconds of white noise at 24kHz
ref_sr = 24000
ref_duration = 3.0
ref_audio = np.random.randn(int(ref_sr * ref_duration)).astype(np.float32) * 0.1
# --- 4a. Speaker Encoder ---
print("\n === Speaker Encoder ===")
mels = mel_spectrogram(
torch.from_numpy(ref_audio).unsqueeze(0),
n_fft=1024,
num_mels=128,
sampling_rate=24000,
hop_size=256,
win_size=1024,
fmin=0,
fmax=12000,
).transpose(1, 2)
print(f" Mel input shape: {list(mels.shape)}") # [1, T, 128]
with torch.no_grad():
spk_embed = model.speaker_encoder(mels)
print(f" Speaker embedding shape: {list(spk_embed.shape)}") # [1, enc_dim]
x_vector = spk_embed[0]
print(f" X-vector (per sample): {list(x_vector.shape)}") # [enc_dim]
# --- 4b. Speech Tokenizer Encode (ref audio -> codes) ---
print("\n === Speech Tokenizer Encode ===")
if model.speech_tokenizer is not None:
st_model = model.speech_tokenizer.model
ref_wav_tensor = torch.from_numpy(ref_audio).unsqueeze(0).float() # [1, samples]
padding_mask = torch.ones_like(ref_wav_tensor, dtype=torch.long)
with torch.no_grad():
enc_out = st_model.encode(ref_wav_tensor, padding_mask=padding_mask, return_dict=True)
ref_codes = enc_out.audio_codes
print(f" Ref audio samples: {ref_wav_tensor.shape[1]}")
print(f" Number of code tensors: {len(ref_codes)}")
for i, c in enumerate(ref_codes):
print(f" ref_codes[{i}] shape: {list(c.shape)}") # [T, num_quantizers]
else:
print(" Speech tokenizer not loaded (will skip encode)")
ref_codes = None
# --- 4c. Talker Prefill Input Construction ---
print("\n === Talker Input Construction ===")
# Simulate tokenized text: "<|im_start|>assistant\nHello world<|im_end|>\n<|im_start|>assistant\n"
# Using config token IDs
from transformers import AutoTokenizer
try:
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
text = "Hello world."
chat_text = f"<|im_start|>assistant\n{text}<|im_end|>\n<|im_start|>assistant\n"
input_ids = tokenizer(chat_text, return_tensors="pt", add_special_tokens=False).input_ids
print(f" Text input_ids shape: {list(input_ids.shape)}")
print(f" Text input_ids: {input_ids[0].tolist()[:20]}...")
except Exception as e:
print(f" Tokenizer load failed: {e}")
# Fallback: synthetic token IDs
input_ids = torch.tensor([[config.im_start_token_id, 77091, 198, 9707, 1879, 13,
config.im_end_token_id, 198,
config.im_start_token_id, 77091, 198]])
print(f" Fallback input_ids shape: {list(input_ids.shape)}")
# --- 4d. Talker Key Shapes ---
print("\n === Talker Architecture Key Shapes ===")
talker = model.talker
# Text embedding
text_emb = talker.get_text_embeddings()
print(f" text_embedding: {text_emb.weight.shape}") # [text_vocab, hidden]
# Codec embedding
codec_emb = talker.get_input_embeddings()
print(f" codec_embedding: {codec_emb.weight.shape}") # [codec_vocab, hidden]
# text_projection (ResizeMLP)
print(f" text_projection type: {type(talker.text_projection).__name__}")
with torch.no_grad():
sample_text_hidden = text_emb(torch.tensor([[0]]))
proj_out = talker.text_projection(sample_text_hidden)
print(f" text_projection in/out: {list(sample_text_hidden.shape)} -> {list(proj_out.shape)}")
# codec_head
print(f" codec_head: Linear({talker.codec_head.in_features} -> {talker.codec_head.out_features})")
# KV cache dimensions
num_layers = tc.num_hidden_layers
num_kv_heads = tc.num_key_value_heads
head_dim = tc.head_dim
print(f"\n Static KV cache per layer: 2 x [B, {num_kv_heads}, max_seq_len, {head_dim}]")
print(f" Total KV layers: {num_layers}")
print(f" Total KV cache (fp32, B=1, seq=2048): "
f"{2 * num_layers * num_kv_heads * 2048 * head_dim * 4 / 1e6:.1f} MB")
# --- 4e. Code Predictor Key Shapes ---
print("\n === Code Predictor Key Shapes ===")
cp = talker.code_predictor
print(f" small_to_mtp_projection: {type(cp.small_to_mtp_projection).__name__}")
if hasattr(cp.small_to_mtp_projection, 'weight'):
print(f" weight shape: {list(cp.small_to_mtp_projection.weight.shape)}")
print(f" lm_heads: {len(cp.lm_head)} heads")
for i, head in enumerate(cp.lm_head):
print(f" lm_head[{i}]: Linear({head.in_features} -> {head.out_features})")
print(f" codec_embeddings: {len(cp.model.codec_embedding)} embeddings")
for i, emb in enumerate(cp.model.codec_embedding):
print(f" codec_embedding[{i}]: {emb.weight.shape}")
cp_layers = cpc.num_hidden_layers
cp_kv_heads = cpc.num_key_value_heads
cp_head_dim = cpc.head_dim
print(f"\n Static KV cache per layer: 2 x [B, {cp_kv_heads}, max_seq_len, {cp_head_dim}]")
print(f" Total KV layers: {cp_layers}")
# --- 4f. Speech Tokenizer Decoder Key Shapes ---
print("\n === Speech Tokenizer Decoder Key Shapes ===")
if model.speech_tokenizer is not None:
st_dec = model.speech_tokenizer.model.decoder
print(f" Decoder type: {type(st_dec).__name__}")
print(f" Total params: {fmt(count_params(st_dec))}")
# Test decode with synthetic codes
# codes shape: [batch, num_quantizers, seq_len]
test_codes = torch.randint(0, 2048, (1, 16, 10))
with torch.no_grad():
test_wav = st_dec(test_codes)
print(f" Test input codes: {list(test_codes.shape)}")
print(f" Test output wav: {list(test_wav.shape)}")
upsample_factor = test_wav.shape[-1] // test_codes.shape[-1]
print(f" Upsample factor: {upsample_factor}x")
# ββ 5. Export Blocker Analysis βββββββββββββββββββββββββββββββββββββββ
print("\n[5/5] Export Blocker Analysis")
print("-" * 60)
blockers = []
# Check speaker encoder
print("\n === Speaker Encoder Export Blockers ===")
se_issues = []
# Conv1d with padding="same" and padding_mode="reflect"
for name, mod in model.speaker_encoder.named_modules():
if isinstance(mod, nn.Conv1d):
if hasattr(mod, 'padding') and mod.padding == 'same':
se_issues.append(f"Conv1d '{name}' uses padding='same' (dynamic pad calc)")
if hasattr(mod, 'padding_mode') and mod.padding_mode == 'reflect':
se_issues.append(f"Conv1d '{name}' uses padding_mode='reflect'")
# AttentiveStatisticsPooling dynamic masking
se_issues.append("AttentiveStatisticsPooling: dynamic _length_to_mask(), .repeat(), masked_fill_")
se_issues.append("Res2NetBlock: torch.chunk + for loop (but fixed scale=8, should be OK)")
for issue in se_issues:
print(f" [!] {issue}")
blockers.extend([("speaker_encoder", i) for i in se_issues])
# Check talker
print("\n === Talker Export Blockers ===")
t_issues = []
t_issues.append("MROPE: 3D rotary embedding with sections [24,20,20] β need custom handling")
t_issues.append("DynamicCache: must replace with static KV cache tensors")
t_issues.append("create_causal_mask/create_sliding_window_causal_mask from transformers")
t_issues.append("Two embedding tables (text + codec) with interleaving logic")
t_issues.append("code_predictor.generate() called inside forward() β autoregressive sub-loop")
t_issues.append("trailing_text_hidden conditional addition in decode step")
t_issues.append("@can_return_tuple decorator")
t_issues.append("@use_kernel_forward_from_hub on RMSNorm")
for issue in t_issues:
print(f" [!] {issue}")
blockers.extend([("talker", i) for i in t_issues])
# Check code predictor
print("\n === Code Predictor Export Blockers ===")
cp_issues = []
cp_issues.append("Uses GenerationMixin.generate() β full autoregressive loop")
cp_issues.append("generation_steps counter used to index into lm_head ModuleList")
cp_issues.append("DynamicCache")
cp_issues.append("get_input_embeddings() returns ModuleList (indexed by generation step)")
for issue in cp_issues:
print(f" [!] {issue}")
blockers.extend([("code_predictor", i) for i in cp_issues])
# Check speech tokenizer
print("\n === Speech Tokenizer Export Blockers ===")
st_issues = []
if model.speech_tokenizer is not None:
st_issues.append("chunked_decode: while loop with dynamic chunk boundaries")
st_issues.append("ConvTranspose1d with dynamic slicing (right_pad removal)")
st_issues.append("CausalConv1d: dynamic padding calculation")
st_issues.append("SnakeBeta: custom activation (should be OK)")
st_issues.append("SplitResidualVectorQuantizer: F.embedding based (OK)")
st_issues.append("Transformer decoder with @dynamic_rope_update and torch.autocast")
st_issues.append("Sliding window attention (window=72)")
for issue in st_issues:
print(f" [!] {issue}")
blockers.extend([("speech_tokenizer", i) for i in st_issues])
# ββ Summary ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print("\n" + "=" * 70)
print("SUMMARY")
print("=" * 70)
print(f"""
Model: Qwen3TTSForConditionalGeneration (1.7B-Base)
Total params: {fmt(count_params(model))}
Export Targets (4 modules):
1. Speaker Encoder ({fmt(count_params(model.speaker_encoder))} params) β ECAPA-TDNN
2. Talker (Main LM) ({fmt(count_params(model.talker.model))} + heads) β Qwen3 28L
3. Code Predictor ({fmt(count_params(model.talker.code_predictor))} params) β 5L transformer
4. Speech Tokenizer Dec ({fmt(count_params(model.speech_tokenizer.model.decoder)) if model.speech_tokenizer else 'N/A'} params) β Transformer + ConvTranspose
Voice Clone Pipeline:
ref_audio (24kHz)
-> mel_spectrogram -> [B, T, 128]
-> speaker_encoder -> x_vector [B, {sec.enc_dim}]
ref_audio -> speech_tokenizer.encode -> ref_codes [T, 16]
text -> tokenizer -> input_ids
[x_vector, ref_codes, input_ids]
-> talker.generate() -> codec_tokens [T', 16]
(internally calls code_predictor.generate() per step)
codec_tokens -> speech_tokenizer.decode -> PCM waveform
Key Dimensions:
Talker: hidden=2048, layers=28, heads=16, kv_heads=8, head_dim=128
Code Predictor: hidden=1024, layers=5, heads=16, kv_heads=8
Codec: vocab=3072 (talker), 2048 (code_predictor), 16 code groups
Speaker: enc_dim={sec.enc_dim}
Export Strategy:
Phase 2: Speaker encoder β fixed mel length, handle Conv1d padding
Phase 3: Talker β static KV cache, unrolled MROPE, separate prefill/decode
Phase 4: Code predictor β static KV, unroll 15-step generation
Phase 5: Vocoder (decoder only) β fixed code length, handle ConvTranspose1d
Phase 6: INT8 via torchao int8_weight_only (instant, no calibration)
Total export blockers found: {len(blockers)}
""")
print("Phase 1 analysis complete!")
|