How should an IndexTTS2 run survive crashes and checkpoint uncertainty?
By Wei Jie Chee · 07 Feb 2026, 00:00 Z
Download printable cheat-sheet (CC-BY 4.0)Checkpoint retention and listening are part of the experiment because the final checkpoint is not automatically the best one.
Do this first
Preserve every recoverable checkpoint and verify the exact loader and dependency versions before resuming or evaluating.
Keep this boundary in view
The recovery notes come from one full-training path and corpus; they do not establish a universal best checkpoint schedule.
Inspect the evidence and what would change the answer
First-party training and recovery record
What supports it
The article records crash recovery, checkpoint retention behavior, dependency constraints, and checkpoint-selection consequences.
Change our mind if
Repeated controlled runs show final-checkpoint selection is stable and intermediate retention adds no decision value.
Claim TTS-INDEXTTS2-CHECKPOINT-2026-08 / v1, checked 7 Aug 2026
60-second takeaway
IndexTTS2 gave us a usable full-SFT baseline with strong operational predictability once we stabilized restart behavior.
In this run,model_step14000.pthwas the practical checkpoint to keep.
The major challenge was process reliability and checkpoint retention policy, not core output quality.
Where this fits
- For founders: IndexTTS2 is a steady full-finetune option in this benchmark.
- For engineers: this page focuses on run recovery and checkpoint management as much as quality.
Series overview:
The fine-tuning pipeline used in this benchmark is open-source: instavar/indextts2-finetuning - the first public fine-tuning code for IndexTTS2 (the official repo is inference-only).
Experiment setup
- Model: IndexTTS2
- Dataset: IMDA NSC
FEMALE_01_44kprocessed manifests - Hardware: RTX 3090 Ti 24 GB
- Training mode: full SFT with resume
Best checkpoint logic
- Best validation region was around step ~13800.
- Saved checkpoints available around that region were
model_step14000.pthand latermodel_step15949.pth. - We treated
model_step14000.pthas the practical best anchor for this run.
Audio evidence
Representative sample
Settings: long-text prompt comparison path, step 14000 checkpoint.
Failure modes we saw
- Training runs were interrupted multiple times and required explicit resume management.
- Some crashes were low-level (
pt_autograd_0segfault signs), which made clean logs critical. - Retention policy kept only recent checkpoint windows, so older steps disappeared automatically.
Recommended inference settings
- Keep checkpoint selection tied to both listening and nearest validation region.
- Prefer explicit run logs and resume metadata over implicit state.