2ae2c9bd9b
Merge pull request #1298 from Kaihui-AMD/docs/update-amd-rocm-install
...
Update AMD ROCm install to 7.2 for RDNA 3.5/4 support
2026-05-18 20:46:11 +08:00
0556b2c1a5
update AMD ROCm install to 7.2 for RDNA 3.5/4 support
...
Signed-off-by: Kaihui-AMD <Kaihui.Tang@amd.com >
2026-05-18 16:39:43 +08:00
2f53ded68e
Merge pull request #1294 from AAtomical/fix/path-traversal-finetune-gradio
...
fix: path traversal in finetune Gradio handlers
2026-05-13 10:06:36 +08:00
25dc4e8668
fix: path traversal in finetune Gradio handlers ( closes #1293 )
...
Add _safe_project_path() helper that rejects absolute paths, null bytes,
and path separators, then verifies the resolved path stays within the
intended base directory via realpath + startswith check.
Apply to all 10 sinks in save_settings, load_settings,
create_data_project, vocab_extend, transcribe_all, create_metadata,
and related functions.
2026-05-12 16:04:28 -04:00
6f91022519
v1.1.20: refactor cache handling in DiT, MMDiT, and UNetT classes (lazyinit), to fix training bug (EMA deepcopy failure)
2026-04-20 15:28:10 +08:00
650c177b14
Merge pull request #1290 from will422-l/fix/gradio-version-cap
...
fix: cap Gradio version to <6.11 to prevent UI freeze
Fixes #1289 thanks to @gmtrnv
2026-04-17 12:01:39 +08:00
761643b1ff
fix: cap Gradio version to <6.11 to prevent UI freeze
...
Gradio 6.11.0 has a bug in gr.Tabs that causes the UI to completely
freeze when switching tabs. This caps the version to prevent users
from hitting this issue until Gradio fixes it upstream.
Ref: #1289
2026-04-17 09:11:15 +08:00
25874ca255
Bump version from 1.1.18 to 1.1.19
2026-04-16 10:56:40 +08:00
428050aa80
Fixes #1287 #1288
2026-04-16 10:49:31 +08:00
22299b38f7
Merge pull request #1285 from ZhikangNiu/main
...
reuse resamplers and cache vocos MelSpectrogram instances
2026-04-04 16:50:43 +08:00
5486a158d4
reuse resamplers and cache vocos MelSpectrogram instances, it will reduce some training cost
2026-04-04 14:31:00 +08:00
82fc4fe622
Bump version from 1.1.17 to 1.1.18
2026-03-24 20:20:15 +08:00
1a63dda3df
Several fixes for utils_infer.py; separate streaming and non-streaming func and add back parallelism
2026-03-24 20:03:01 +08:00
2414e3d492
Merge pull request #1281 from zhuxiaoxuhit/fix/remove-ineffective-threadpoolexecutor
...
remove ineffective ThreadPoolExecutor in infer_batch_process
2026-03-24 19:31:29 +08:00
543fe4facf
remove ineffective ThreadPoolExecutor in infer_batch_process
...
process_batch is a generator function, so submitting it to a thread pool
only creates a generator object without running any inference code.
all actual work happens sequentially in the main thread when next() is called.
also removes the always-true 'if result:' guard on a generator object.
2026-03-24 16:10:46 +08:00
deb5540edb
Merge pull request #1280 from ZhikangNiu/main
...
Add F5TTS v1 Small + LibriTTS training config
2026-03-23 20:14:40 +08:00
a25de67cbd
F5TTS v1 Small + LibriTTS training config
2026-03-23 16:16:52 +08:00
623c96c294
Add Arabic model details to SHARED.md ( #1279 )
...
Added Arabic section with details for F5-TTS Small model.
2026-03-16 16:13:45 +08:00
5005714c4c
Remove pydantic<=2.10.6 restriction to fit latest gradio version
2026-03-07 19:37:59 +08:00
4533426c72
Bump version from 1.1.16 to 1.1.17
2026-03-04 19:34:03 +08:00
b5ab1afa16
Merge pull request #1270 from ZhikangNiu/main
...
- Use fused=True for AdamW by default
- Warn on torch attention mask memory usage `if attn_backend == "torch" and attn_mask_enabled`
---------
Co-authored-by: SWivid <swivid@qq.com >
2026-03-04 19:31:52 +08:00
ab75dc2837
Merge pull request #1271 from mlxu995/patch-1
...
Add show_info parameter to preprocess_ref_audio_text
2026-03-04 18:56:46 +08:00
4361b0b94f
Add show_info parameter to preprocess_ref_audio_text
2026-03-04 17:04:20 +08:00
097772c917
Merge pull request #1269 from ZhikangNiu/main
...
feat:add mmdit flash attn support
fix: autocast when use flash_attn to enable log_sample
2026-02-27 01:20:09 +08:00
76c00b127e
when use flash_attn, log_sample should under autocast context
2026-02-25 08:10:00 +08:00
d7c7a117fa
feat:add mmdit flash attn support
2026-02-23 20:39:35 +08:00
54c50eb8f6
Bump version from 1.1.15 to 1.1.16
2026-02-16 12:37:19 +08:00
65250152da
Merge pull request #1267 from QingyuLiu0521/qyl/pr-dit-only
...
Optimize DiT text embedding with batched per-sample seq handling
2026-02-16 12:28:17 +08:00
c817d6a21d
Unify seq_len naming in DiT get_input_embed
2026-02-15 23:24:11 -05:00
04459f71e6
Merge pull request #1266 from ZhikangNiu/main
...
Make wandb project/run_name/resume_id configurable via Hydra yaml, backward compatible with defaults
2026-02-16 11:10:17 +08:00
57dc698c16
Apply ruff formatting
2026-02-15 21:41:17 -05:00
6b6ce47d2e
Optimize DiT text embedding with batched per-sample seq handling
2026-02-15 21:31:19 -05:00
6768b1bcff
Make wandb project/run_name/resume_id configurable via Hydra yaml, backward compatible with defaults
...
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com >
2026-02-16 10:05:02 +08:00
ecfdccb890
Merge pull request #1265 from ZhikangNiu/main
...
Use torch.utils.checkpoint in mmdit forward loop when enabled to reduce memory usage.
2026-02-14 11:14:57 +08:00
bb5526fc5b
Use torch.utils.checkpoint in mmdit forward loop when enabled to reduce memory usage.
2026-02-14 11:05:08 +08:00
655fbca552
Update run_asr_wer method in utils_eval.py for compat with jiwer>=4.0.0
2026-02-02 17:48:16 +08:00
fc0fa67a03
Update eval README with ctranslate2 installation instructions
...
Added installation instructions for ctranslate2 based on CUDA and cuDNN versions.
2026-01-28 19:15:57 +08:00
a3c2ea9784
Merge pull request #1261 from ZhikangNiu/main
...
Ignore padding at the end of the GT mel spectrogram when training sample
2026-01-26 18:58:47 +08:00
d71a69d528
Ignore padding at the end of the ground truth mel spectrogram when training sample
2026-01-26 09:40:37 +08:00
b9d923088c
Increase default max_duration from 4096 to 65536 in cfm.py ( #1260 )
2026-01-25 23:58:21 +08:00
c279a2b7d5
Merge pull request #1256 from ZhikangNiu/main
...
change prepare_csv_wavs from relative path to absolute path and get duration info with soundfile and torchaudio
2026-01-22 20:00:54 +08:00
5d473e980c
add tqdm in convert text to pinyin
2026-01-22 16:34:39 +08:00
2aefa7c5f7
fix many tensorboard writer and only log in main_process
2026-01-22 13:36:09 +08:00
97fdc7fbb4
change prepare_csv_wavs from relative path to absolute path and get duration info with soundfile and torchaudio
2026-01-22 12:27:23 +08:00
1d2f7c5389
Formatting
2026-01-21 13:42:19 +00:00
37a2633d35
Fix epoch updates count logic as drop_last
2026-01-21 21:36:57 +08:00
bc1df7a4fa
Adding support for hf:// links on CLI ( #1252 )
2026-01-21 16:38:30 +08:00
eca786ee0c
Merge pull request #1250 from raivisdejus/add-latvian-community-model
...
Adding Latvian model to shared community models list
2026-01-17 19:29:23 +08:00
27e20fcf39
Merge pull request #1242 from acadarmeria/fix-speech-edit-mel-domain
...
Fix speech editing boundary artifacts by working in mel domain
2025-12-26 17:35:57 +08:00
dff57ebd2a
Fix speech editing boundary artifacts by working in mel domain
...
Previously, speech_edit.py worked in wav domain (inserting zeros into the waveform before computing mel spectrogram), which caused boundary artifacts when mel spectrogram windows straddled zeros and real audio.
This commit refactors the approach to work in mel domain:
- Compute mel spectrogram on the clean original audio first
- Insert zero frames in mel domain instead of zero samples in wav domain
- Use frame-level granularity throughout for consistency
Benefits:
- Eliminates boundary artifacts
- More consistent behavior regardless of small float variations in input times
- Cleaner edit boundaries
Changes to speech_edit.py (lines 148-220):
- Convert audio to mel using model.mel_spec() before editing
- Build mel_cond by concatenating original mel frames + zero frames
- Calculate all time-based values at frame level first, then convert to samples
- Pass mel_cond directly to model.sample() instead of raw audio
2025-12-26 08:49:57 +00:00
Yushen CHEN