cmedina

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

2026-07-02 20:22:40 -04:00

b624e90b02 fix: use BnB quantize_batch to actually quantize weights

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

2026-07-02 20:18:15 -04:00

ec2b1be864 chore: clean up temp files

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

2026-07-02 20:17:25 -04:00

3716051abc fix: remove PEFT prepare (causes OOM)

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

2026-07-02 20:13:20 -04:00

ac891fe25b fix: replace with proper BnB 4-bit quantization script

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

2026-07-02 20:11:41 -04:00

7984ae93e7 feat: proper BnB 4-bit quantization with PEFT prepare

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

2026-07-02 20:04:56 -04:00

57b169f202 feat: use BnB 4-bit model for QLoRA training

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

2026-07-02 20:04:46 -04:00

76ff09b140 feat: add BnB 4-bit quantization script

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

2026-07-02 19:55:05 -04:00

3c31e4dfb0 feat: load bf16 model, quantize with BnB on CPU, move to GPU

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

2026-07-02 19:33:40 -04:00

a0f4f644b0 fix: load 4-bit model with device_map=auto (let transformers distribute)

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

2026-07-02 18:46:07 -04:00

20626e7a78 feat: manual 4-bit quantization with PEFT prepare + Linear4bit

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

2026-07-02 18:23:42 -04:00

42c25962b2 feat: load model to CPU with BnB 4-bit, then move to GPU

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

2026-07-02 18:20:44 -04:00

1c7e73e927 fix: load already-quantized 4-bit model directly (no BnB)

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

2026-07-02 18:05:29 -04:00

199176c8d0 feat: load model with BnB 4-bit quantization (fit in single GPU)

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

2026-07-02 17:53:58 -04:00

5a84b0eee5 feat: add CPU offload for optimizer states

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

2026-07-02 17:53:12 -04:00

a51a744d7c fix: reduce seq_length to 512 and increase gradient_accumulation

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

2026-07-02 17:53:07 -04:00

691e5d3611 feat: use 8-bit AdamW optimizer to save VRAM

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

2026-07-02 17:50:02 -04:00

92999b7783 fix: load model to single GPU (MoE requires all layers on same device)

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

2026-07-02 17:29:20 -04:00

0456d55c8a fix: disable gradient checkpointing (device mismatch with distributed MoE)

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

2026-07-02 17:09:41 -04:00

d9c3781242 fix: quote 'no' in eval_strategy (YAML boolean issue)

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

2026-07-02 16:49:39 -04:00

565ddd416e fix: disable eval (no test split in dataset)