1.1.20

What's Changed

• fix: refactor cache handling in DiT, MMDiT, and UNetT classes (lazyinit to avoid EMA deepcopy failure while training)
• fix: cap Gradio version to <6.11 to prevent UI freeze by @will422-l in #1290

New Contributors

• @will422-l made their first contribution in #1290

Full Changelog: 1.1.19...1.1.20

1.1.19

What's Changed

• reuse resamplers and cache vocos MelSpectrogram instances by @ZhikangNiu in #1285
• fix #1287 in v1.1.18

Full Changelog: 1.1.18...1.1.19

1.1.18

What's Changed

• Add Arabic model details to SHARED.md by @karimouda in #1279
• Add F5TTS v1 Small + LibriTTS training config by @ZhikangNiu in #1280
• remove ineffective ThreadPoolExecutor in infer_batch_process by @zhuxiaoxuhit in #1281
• Several fixes for utils_infer.py; separate streaming and non-streaming functions, and add back parallelism 82fc4fe

New Contributors

• @karimouda made their first contribution in #1279
• @zhuxiaoxuhit made their first contribution in #1281

Full Changelog: 1.1.17...1.1.18

1.1.17

What's Changed

• feat:add mmdit flash attn support by @ZhikangNiu in #1269
• Add show_info parameter to preprocess_ref_audio_text by @mlxu995 in #1271
• Add fused AdamW option and warn on torch attention mask memory usage by @ZhikangNiu in #1270

New Contributors

• @mlxu995 made their first contribution in #1271

Full Changelog: 1.1.16...1.1.17

1.1.16

What's Changed

• Fix speech editing boundary artifacts by working in mel domain by @acadarmeria in #1242
• Adding Latvian model to shared community models list by @raivisdejus in #1250
• Adding support for hf:// links on CLI by @raivisdejus in #1252
• change prepare_csv_wavs from relative path to absolute path and get d… by @ZhikangNiu in #1256
• Ignore padding at the end of the GT mel spectrogram when training sample by @ZhikangNiu in #1261
• Use torch.utils.checkpoint in mmdit forward loop when enabled to redu… by @ZhikangNiu in #1265
• Make wandb project/run_name/resume_id configurable via yaml by @ZhikangNiu in #1266
• Optimize DiT text embedding with batched per-sample seq handling by @QingyuLiu0521 in #1267

New Contributors

• @acadarmeria made their first contribution in #1242
• @raivisdejus made their first contribution in #1250

Full Changelog: 1.1.15...1.1.16

1.1.15

• Workaround for gr.Accordion default open=False bug, #1239 (comment)

Full Changelog: 1.1.12...1.1.15

1.1.12

• Bump gradio version from 5.0 to 6.0, with several fixes to ensure compatibility with f5-tts_infer-gradio

Full Changelog: 1.1.10...1.1.12

1.1.10

• Fix issues of skipping last words in shorter sentences when batch inference #1039 #1179,
  in both runtime TRT-LLM and PyTorch implementation.
• Clean-up and make advanced eval scripts 529d856
• Add torchcodec dependency to pyproject.toml
• Replace jieba package with rjieba #1230

Full Changelog: 1.1.9...1.1.10

1.1.9

• Update numpy version constraint to support numpy 2.x for python version >= 3.11 #1162

Full Changelog: 1.1.8...1.1.9

1.1.8

Full Changelog: 1.1.7...1.1.8