Ethanfel 8166c56552 perf: gradient checkpointing on vocoder forward to reduce activation memory ...

BigVGAN's 512x upsampling stack stores huge intermediate activations for
backward even in snake_alpha_only mode (only 5K trainable params, but
activation graph runs through the full network after each snake op).

Wrapping vocoder() in checkpoint(use_reentrant=False) recomputes activations
during backward instead of storing them — ~2x compute cost, large reduction
in peak VRAM. Should allow batch_size > 1 on 96 GB without OOM.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-04-09 13:45:24 +02:00

..

__init__.py

feat: add DITTO optimizer, upgrade BigVGAN trainer, document all nodes

2026-04-09 12:04:05 +02:00

selva_activation_steering_extractor.py

feat(steering): conditional-only injection + per-position vectors

2026-04-09 01:02:51 +02:00

selva_activation_steering_loader.py

feat: add activation steering pipeline (extractor, loader, sampler injection)

2026-04-09 00:38:26 +02:00

selva_audio_preprocessors.py

feat: add SelVA HF Smoother and Spectral Matcher preprocessing nodes

2026-04-08 20:28:16 +02:00

selva_bigvgan_loader.py

fix(bigvgan): add 44k/BigVGANv2 support to trainer and loader

2026-04-09 01:28:32 +02:00

selva_bigvgan_trainer.py

perf: gradient checkpointing on vocoder forward to reduce activation memory

2026-04-09 13:45:24 +02:00

selva_dataset_browser.py

feat: thorough overnight sweep + dataset browser updates

2026-04-08 00:38:19 +02:00

selva_ditto_optimizer.py

fix: strip inference tensor flags in DITTO optimizer

2026-04-09 12:18:20 +02:00

selva_feature_extractor.py

fix: set OUTPUT_NODE=True on SelVA Feature Extractor so it runs without connected outputs

2026-04-06 00:11:16 +02:00

selva_lora_evaluator.py

feat: evaluate adapters on all dataset clips, not just clip_001

2026-04-08 17:42:55 +02:00

selva_lora_loader.py

debug: print lora_A norm after loading to confirm adapter applied

2026-04-05 23:05:23 +02:00

selva_lora_scheduler.py

feat: add cosine LR decay schedule to trainer and scheduler

2026-04-08 13:25:01 +02:00

selva_lora_trainer.py

feat: evaluate adapters on all dataset clips, not just clip_001

2026-04-08 17:42:55 +02:00

selva_model_loader.py

feat: comprehensive node improvements

2026-04-04 18:16:03 +02:00

selva_sampler.py

feat(steering): conditional-only injection + per-position vectors

2026-04-09 01:02:51 +02:00

selva_skip_experiment.py

feat: SelVA Skip Experiment node + save partial scalars on skip

2026-04-08 13:10:43 +02:00

selva_textual_inversion_loader.py

feat: add inject_mode (suffix/prefix) to TI pipeline

2026-04-08 23:31:52 +02:00

selva_textual_inversion_trainer.py

fix(ti-trainer): clamp token norm to CLIP manifold to prevent buzz artifacts

2026-04-08 23:54:23 +02:00

selva_ti_scheduler.py

fix(ti): lower default lr/batch, add lr_batch sweep group

2026-04-08 23:42:22 +02:00

selva_vae_roundtrip.py

fix(vae-roundtrip): pass bigvgan path to encoder-only FeaturesUtils

2026-04-08 20:05:44 +02:00

utils.py

chore: remove all PrismAudio code from main branch

2026-04-04 17:58:31 +02:00