1068 Commits

Author SHA1 Message Date
RVC-Boss
6d51ffceaf Update ASR configuration, Faster Whisper, FunASR and WebUI from update_asr package 2026-11-12 00:26:17 +08:00
Kakaru
59c44686f5
feat: add cache-dit (DBCache) block caching to the DiT sampler (#2868)
The DiT samples with the full block stack on every diffusion step. cache-dit
(vipshop/cache-dit, DBCache) skips the middle of the stack when the front residual
barely moved:

    front : trunk(x) -> blocks[0:Fn] -> x_F,   R = x_F - trunk(x)
            fd = sum|R - R_prev| / (sum|R_prev| + eps)
    hit   : x_mid = x_F + MnResidual_prev, then norm_out/proj_out only
    miss  : x_F -> blocks[Fn:depth] -> x_M, store MnResidual = x_M - x_F, R -> R_prev

The head (norm_out + proj_out) runs unconditionally, hits included. A hit additionally
needs `fd < threshold` outside the warmup steps and must pass the max_cached / max_cont /
max_accum gates.

The conditional and unconditional CFG branches are two separate forward passes, so each
keeps its own R_prev, MnResidual, hit/miss counters and gate counters.

Implementation:

* new `module/cache_dit.py`: configuration, per-sampling state, the CFM-side mixin;
* `DiT.forward` split into `trunk()` / `run_blocks(start, end)` / `head()` (weights are
  shared, nothing is copied or duplicated);
* `CFM.inference` creates one cache state per sampling run, advances it per step and reads
  back a single scalar per step (the only device/host sync);
* `cache_dit=off|auto|f4|f8|f16` plus dbcache_fn/dbcache_threshold/dbcache_warmup/
  dbcache_downsample on TTS_infer_pack.run, api.py, api_v2.py, both WebUIs and the CLI,
  with GSV_DBCACHE* environment fallbacks for scripts that build no request dict.

Default is off and the off path is the previous one: the split is bit-identical to the old
`forward()` on 24 configurations (fp16/fp32 x long skip x static cache x inference/training),
and `CFM.inference` with the switch off reproduces the pre-change output bit for bit with the
real weights. `auto` resolves to F8 from 16 sampling steps and stays off below that, so
4-step runs are never cached (the warmup budget already covers them).

Measured on a 32-step v5dev request with F8 / threshold 0.02 / warmup 8 / downsample 4:
18.8% of the DiT forwards served from cache, mel relL1 1.2e-3 against the uncached run,
DiT stage ~1.12x faster; F4 at the same threshold gives 21.9% and ~1.18x, threshold 0.05
gives 45.3% and ~1.34x.
2026-11-09 21:59:04 +08:00
RVC-Boss
04a44eb977 Sync 20261031 v5 code update and remove unused v5_inference module 2026-10-31 23:56:56 +08:00
游手好闲的言柒
b4570b876f
fix(streaming): 支持API v2语义流式文本切分 (#2850) 2026-10-19 18:05:29 +08:00
游手好闲的言柒
1d9fe63d28
feat(streaming): 为 API v2 增加 V5 和 V2PP 语义流式加速 (#2848)
* feat(streaming): 为官方API v2接入V5和V2PP语义流式

* feat(streaming): 为API v2添加渐进混合窗口
2026-10-18 21:14:20 +08:00
RVC-Boss
1b47ca07ef build: add triton-windows to Python 3.12 Windows requirements 2026-10-09 00:27:04 +08:00
RVC-Boss
e7cd61ec3d docs: synchronize multilingual changelogs for ASR, single-GPU training and V5 2026-10-08 00:18:27 +08:00
RVC-Boss
b16b45653b
Update Changelog_CN.md 2026-10-08 00:09:23 +08:00
RVC-Boss
9bbd80ab33 docs: localize SeedTTS speaker similarity chart and sync latest README updates 2026-10-07 20:45:43 +08:00
RVC-Boss
7b35028ca0
Update Zero-Shot TTS Speaker Similarity Comparison Image on SeedTTS Eval Testset_CN2020
Update Zero-Shot TTS Speaker Similarity Comparison Image on SeedTTS Eval Testset_CN2020
2026-10-07 20:42:34 +08:00
RVC-Boss
6a82d32744 docs: add Python 3.12 Torch 2.7.1 CUDA tested environments 2026-10-07 17:19:38 +08:00
RVC-Boss
a3035089c6 docs: restore multilingual V5 release notes and acceleration credit 2026-10-06 16:49:22 +08:00
RVC-Boss
3b95794c49 Integrate v5 training and inference updates with single-GPU training support 2026-10-06 16:39:59 +08:00
RVC-Boss
a9beed7c44 Update Windows acceleration requirements: add resampy and constrain setuptools 2026-10-05 22:39:01 +08:00
RVC-Boss
d8607fa97c docs: add V5 release notes to all README languages 2026-10-05 17:54:26 +08:00
RVC-Boss
522c6e92ca docs: add V5 release notes to all README languages 2026-10-05 17:42:33 +08:00
RVC-Boss
ac846f5bc7 Add v5 inference support to CUDA graph acceleration branch 2026-10-05 17:28:06 +08:00
RVC-Boss
25937fe2d8 Sync CUDA graph acceleration update from GPT-SoVITS_add_accel package 2026-10-05 15:22:55 +08:00
zhifu gao
48b1a0169a
Fix Fun-ASR-Nano Transformers requirement (#2824)
Signed-off-by: LauraGPT <LauraGPT@users.noreply.github.com>
Co-authored-by: LauraGPT <LauraGPT@users.noreply.github.com>
2026-08-18 17:16:25 +08:00
RVC-Boss
d523079fc0 Remove redundant code 2026-07-22 16:20:45 +08:00
zhifu gao
be6a4f1e9d
docs: document FunASR backend options (#2803)
Co-authored-by: LauraGPT <LauraGPT@users.noreply.github.com>
2026-07-13 10:41:22 +08:00
zhifu gao
551918539c
Fix Fun-ASR-Nano fallback for GPT-SoVITS ASR (#2801)
Co-authored-by: LauraGPT <LauraGPT@users.noreply.github.com>
2026-07-10 17:36:10 +08:00
zhifu gao
bf81cdb14a
feat: upgrade FunASR + add Fun-ASR-Nano & SenseVoice ASR backends (#2782)
- Upgrade funasr from ==1.0.27 to >=1.3.7
- Add Fun-ASR-Nano (31 languages, Chinese dialects, recommended default)
- Add SenseVoice (ultra-fast 170x realtime, 5 languages)
- Keep original Paraformer as '达摩 ASR (中文经典)' for backward compat
- WebUI shows 3 FunASR options + Faster Whisper

Tested: routing logic verified for all backends (zh/en/ja/ko).
Resolves #2777

Co-authored-by: xiaoyunchong.xyc <xiaoyunchong.xyc@alibaba-inc.com>
2026-06-20 15:44:59 +08:00
SapphireLab
b2cff0cd0a
fix: 多音字修改 (#2791)
* fix: 固定唑的读音

* add: 增加md5检测以更新缓存
2026-06-16 21:46:53 +08:00
RVC-Boss
08d627c333
增加cuda graph支持,普通推理模式推理速度原地翻倍,效果不变。2
增加cuda graph支持,普通推理模式推理速度原地翻倍,效果不变。2
2026-04-30 15:01:45 +08:00
RVC-Boss
6d95b559e8
增加cuda graph支持,普通推理模式推理速度原地翻倍,效果不变。1
增加cuda graph支持,普通推理模式推理速度原地翻倍,效果不变。1
2026-04-30 15:01:11 +08:00
RVC-Boss
ea2d2a8166
Update README.md 2026-04-19 21:02:57 +08:00
SapphireLab
d9f03dad3e
Update Documentation (#2768)
* 调整日志格式

* docs: Update other languages' changelogs
2026-04-18 22:33:55 +08:00
RVC-Boss
647935357a
Update Changelog_CN.md 2026-04-18 19:01:11 +08:00
RVC-Boss
02425ea256
Fixed issues such as missing imports for types like Optional.
Fixed issues such as missing imports for types like `Optional`.
2026-04-18 17:33:53 +08:00
Harikrishna KP
938f05fce8
fix: correct torch.randint upper bound to include both values (#2733) 2026-04-18 17:19:55 +08:00
huang yutong
445d18ccce
fix: 修复 TTS 音频后处理中的多个缺陷 (#2753)
1. 修复音频超采样时 int16 双重转换导致整数溢出(CRITICAL)
   - audio_postprocess 中 `audio = (audio * 32768).astype(np.int16)` 位于
     if/else 块之外无条件执行,当 super_sampling=True 时音频已在分支内
     转为 int16,再次乘以 32768 导致溢出和音频完全失真
   - 同时修复 super_sampling=True 但超分模型不存在时 torch.Tensor 调用
     .astype() 的 AttributeError

2. 修复 batched vocoder 推理中 padding_len=0 导致音频丢失(HIGH)
   - 当 padding_len 恰好为 0 时,`-0 * upsample_rate == 0`,切片
     `audio[x:0]` 返回空张量,导致整段音频丢失

3. 修复文件不存在时错误地抛出 FileExistsError(LOW)
   - 应为 FileNotFoundError

Made-with: Cursor
2026-04-18 17:16:24 +08:00
Mushroomcowisheggs
00ce973412
feat: 添加数据集的错误处理提示 (#2758)
Co-authored-by: moomushroom <107208254+moomushroom@users.noreply.github.com>
2026-04-18 17:13:30 +08:00
huang yutong
14191901cd
fix: 修复多个模块中的独立 bug (#2755)
1. 修复 sync_buffer 中除以函数对象而非调用结果(distrib.py)
   - `buffer.data /= world_size` 中 world_size 是函数,缺少 (),
     导致 TypeError 使分布式训练 buffer 同步失败

2. 修复 istft 函数缺少 return 语句(spec_utils.py)
   - 函数计算了结果但未返回,调用者始终得到 None

3. 修复 cut0 返回字面量 "/n" 而非换行符 "\n"(text_segmentation_method.py)
   - 导致后续 text.split("\n") 无法正确切分,字面 /n 被当作文本内容

4. 修复粤语 ASR 的 vad/punc model_revision 被无条件覆盖(funasr_asr.py)
   - 粤语分支将 vad_model_revision 设为空(因不使用 VAD/标点模型),
     但 if/else 外的赋值将其覆盖为 "v2.0.4",传入错误的 revision 参数

Made-with: Cursor
2026-04-18 17:10:56 +08:00
东云
780383d5bd
[codex] Improve Windows single-GPU v3 LoRA training / 改进 Windows 单卡 v3 LoRA 训练流程 (#2767)
* Improve Windows single-GPU v3 LoRA training

* Drop unrelated checkpoint helper change from PR

* Tighten PR scope to single-GPU training path fixes
2026-04-18 16:54:26 +08:00
白菜工厂1145号员工
ba8de9b760
优化 G2PW 的推理输入构造与多音字处理流程,减少重复计算,降低长句场景下的推理开销 (#2763)
* Enhance G2P processing by implementing batch input handling in _g2p function, improving efficiency. Update prepare_onnx_input to utilize caching for tokenization and add optional parameters for character ID mapping and phoneme masks. Refactor G2PWOnnxConverter to streamline model loading and configuration management.

* Enhance G2PW model input handling by introducing polyphonic context character support and updating the data preparation method to return additional query IDs. This improves the processing of polyphonic characters in sentences.
2026-04-18 16:52:32 +08:00
XXXXRT666
2d9193b0d3
Migrate to miniforge, add missing dependencies, update docker file, remove deprecated files (#2732)
* Migrate to miniforge, add missing dependencies, update docker file, remove deprecated files

* Add Env Vars and Secrets
2026-02-09 15:05:25 +08:00
Oarora
9986880b3f
fix Conda 条款未同意导致的构建失败 (#2727) 2026-02-08 23:52:04 +08:00
ChasonJiang
c767f0b83b
修复bug (#2704)
* 修复bug

* fallbak and bug fix
2025-12-30 16:00:21 +08:00
ChasonJiang
9080a967d5
修复采样错误 (#2703) 2025-12-30 15:21:03 +08:00
sushistack
51df9f7384
Fix model file name in README instructions (#2700) 2025-12-25 16:44:21 +08:00
ChasonJiang
bfca0f6b2d
对齐naive_infer的解码策略,防止吞句 (#2697) 2025-12-19 17:37:19 +08:00
ChasonJiang
abe984395c
对齐gpt topk默认采样参数 (#2696) 2025-12-19 16:05:36 +08:00
RVC-Boss
cc89c3660e
Update requirements.txt 2025-12-19 15:54:54 +08:00
ChasonJiang
36b3231c6f
bug fix (#2689) 2025-12-15 14:23:06 +08:00
RVC-Boss
9ec3a60f30
Update config.py 2025-12-01 20:23:49 +08:00
RVC-Boss
fc533b6fb7
Update fasterwhisper_asr.py 2025-12-01 11:38:37 +08:00
XXXXRT666
857799276c
Fix Modelscope (#2679) 2025-12-01 11:13:15 +08:00
Spr_Aachen
92d2d337fd
Fix training error caused by float type of default_batch_size parameter (#2662) 2025-11-28 22:53:43 +08:00
ChasonJiang
6fb441f65e
更友好的流模式选项 (#2678) 2025-11-28 22:13:48 +08:00