cmedina

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

2026-07-01 07:45:35 -04:00

da5eb3abed feat: implement QLoRA with 4-bit BitsAndBytes quantization

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

2026-07-01 07:38:37 -04:00

f0ee6bc9a2 fix: torch_dtype -> dtype (fix deprecation warning)

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

2026-07-01 07:37:48 -04:00

f651ca030c fix: convert warmup_ratio to warmup_steps (fix deprecation warning)

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

2026-07-01 07:36:39 -04:00

f47944ee75 fix: remove quantization_config after loading to bypass SFTTrainer check

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

2026-07-01 07:30:12 -04:00

86b401c12f fix: convert FP8 to bf16 for PEFT compatibility

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

2026-07-01 07:29:33 -04:00

6f631f00ae fix: use FP8 model with 32GB CPU offload

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

2026-07-01 07:28:23 -04:00

9d92531c55 fix: use Ornith-1.0-35B bf16 base model (PEFT compatible)

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

2026-07-01 07:27:14 -04:00

05ace2d38f fix: load NVFP4 as bf16 by overriding quantization config

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

2026-07-01 00:34:35 -04:00

be7b589c6b fix: use BitsAndBytes INT4 for PEFT compatibility

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

2026-07-01 00:08:43 -04:00

d7301d37dd fix: use float16 dtype to preserve NVFP4 quantization

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

2026-07-01 00:03:26 -04:00

af09d47c6f fix: load model to CPU first, let DeepSpeed distribute

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

2026-06-30 23:57:01 -04:00

2c1251788d fix: add compressed-tensors to training dependencies

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

2026-06-30 23:50:29 -04:00

905c2ba30d fix: preserve NVFP4 quantization, don't force bf16

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

2026-06-30 23:45:58 -04:00

114e25fcfb fix: use device_map=auto to distribute model across GPUs

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

2026-06-30 23:34:35 -04:00

8d4cdca3bb feat: use Ornith-1.0-35B-NVFP4 for training

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

2026-06-30 22:47:21 -04:00

5713bf0285 fix: add CPU offload back to DeepSpeed ZeRO-3

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

2026-06-30 22:41:29 -04:00

3bb7328e2a fix: disable quantization config for fp8 model training

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

2026-06-30 22:36:52 -04:00

a8c9cd682e fix: simplify DeepSpeed config, remove offload settings

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

2026-06-30 22:32:15 -04:00

b760ba09e7 fix: use CUDA 13.0 nightly for RTX 5090 support

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

2026-06-30 21:54:07 -04:00

9cbf6549e5 fix: use output_file parameter in prepare_dataset