Circuit-Forge/avocet
2
0
Fork 0
Code Issues 9 Pull requests 1 Projects Releases Packages Wiki Activity Actions
avocet/scripts
History
pyr0ball 5dee23f53c fix(avocet): reduce deberta-small VRAM + auto-select freest GPU for training 
- deberta-small: batch_size 16→8 + grad_accum 1→2 (same effective batch),
  gradient_checkpointing=True (fp16 stays off: DeBERTa v3 disentangled
  attention overflows fp16 at the gather step)
- api: _best_cuda_device() picks highest free-VRAM GPU via nvidia-smi;
  sets CUDA_VISIBLE_DEVICES in subprocess env to prevent DataParallel
  replication across both GPUs; adds PYTORCH_ALLOC_CONF=expandable_segments
- SSE log now reports which GPU was selected
2026-03-15 17:09:06 -07:00
..
__init__.py feat: initial avocet repo — email classifier training tool 2026-02-27 14:07:38 -08:00
benchmark_classifier.py fix(avocet): guard discover_finetuned_models against malformed/incomplete training_info.json 2026-03-15 15:18:13 -07:00
classifier_adapters.py fix(avocet): FineTunedAdapter GPU device routing + precise body truncation test 2026-03-15 10:56:47 -07:00
finetune_classifier.py fix(avocet): reduce deberta-small VRAM + auto-select freest GPU for training 2026-03-15 17:09:06 -07:00