cmedina

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 21:14:32 -04:00

815115e97a feat: add BnB 4-bit quantization strategy + use bf16 model

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 18:16:53 -04:00

e56ae5e002 fix: restore all strategies with CPU-first as primary

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 18:00:58 -04:00

f18a9f0973 fix: use aggressive CPU loading for memory optimization

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 17:01:29 -04:00

83b45c2a5c feat: add PYTORCH_CUDA_ALLOC_CONF for memory optimization

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:50:56 -04:00

19f405255c remove auto device map in training strategies

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:39:57 -04:00

cc353b3dea refactor: restructure project - scripts at root level

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:35:49 -04:00

293f9caf65 fix: use correct repo_root path for dataset files

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:24:32 -04:00

1095032801 feat: add ai none to unload AI model before training

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:22:57 -04:00

79aa1e4061 fix: don't reinstall PyTorch (already installed)

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:22:10 -04:00

3b771c0db5 fix: add venv setup and proper pip install

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:21:49 -04:00

ff0d6e8736 fix: use correct path for training script

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:06:59 -04:00

5eae235180 executable permissions

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 16:06:10 -04:00

7e13680e27 additional fixes

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 15:57:43 -04:00

088521094e feat: add 4 loading strategies (no BnB for already-quantized model)

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 15:56:01 -04:00

3089f27901 fix: load 4-bit model AS-IS without BnB (already quantized)

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 15:18:51 -04:00

1197235b4d fix: skip prepare_model_for_kbit_training to avoid OOM

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 15:14:02 -04:00

d9fd079686 fix: add accelerate CPU offload to bf16 strategies

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 15:11:47 -04:00

4ed80ad3a3 feat: restore all 6 loading strategies

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 15:10:56 -04:00

313b44381f fix: remove DeepSpeed, use torchrun without distributed training

cmedina pushed to main at cmedina/agenx-lora-training

2026-07-01 14:45:32 -04:00

7de54746a2 feat: add 6 loading variants with different offload strategies