cmedina

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

2026-07-01 14:43:33 -04:00

22d1bfd573 fix: use 4-bit model first, then bf16 with CPU offload

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

2026-07-01 14:41:50 -04:00

3fa2cfe4cc fix: use DeepSpeed ZeRO-3 directly (skip CPU loading)

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

2026-07-01 14:19:44 -04:00

3760aaf316 fix: rename output dir to ornith-35b-lora

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

2026-07-01 14:16:07 -04:00

4287e42148 fix: remove screen, run training directly

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

2026-07-01 14:09:35 -04:00

3bd51ec7db feat: add screen support to bypass cgroup limits

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

2026-07-01 14:00:42 -04:00

ce758f7eb5 feat: improve loading strategies with low_cpu_mem_usage and proper DeepSpeed config

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

2026-07-01 13:55:25 -04:00

2d13e63c25 feat: add 4 loading strategies with automatic fallback

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

2026-07-01 13:53:44 -04:00

b300235be0 fix: load 4-bit model AS-IS without additional quantization

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

2026-07-01 13:53:08 -04:00

e967bd74f1 fix: skip quantization, use DeepSpeed ZeRO-3 CPU offload directly

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

2026-07-01 13:39:53 -04:00

7539e39a93 fix: use accelerate load_checkpoint_and_dispatch for proper CPU offload

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

2026-07-01 13:32:06 -04:00

528e7002ef feat: add 8-bit CPU offload fallback between 4-bit and bf16

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

2026-07-01 13:29:24 -04:00

59ddfb45ed feat: add fallback to bf16 with CPU offload if 4-bit fails

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

2026-07-01 13:12:59 -04:00

b41d99f956 fix: remove broken quantization_config before loading

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

2026-07-01 13:11:16 -04:00

1228a1f38b fix: load model first then quantize with BnB

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

2026-07-01 12:35:22 -04:00

14042dbb44 fix: use pre-quantized 4bit model

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

2026-07-01 12:33:23 -04:00

f784d06f0a fix: load QLoRA directly to GPU (skip CPU)

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

2026-07-01 11:19:02 -04:00

9471d5cc77 fix: cast learning_rate to float (YAML parses 2e-4 as string)

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

2026-07-01 09:37:22 -04:00

2fa52cd9d4 fix: load QLoRA model to CPU first, DeepSpeed distributes

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

2026-07-01 07:48:45 -04:00

2fa5173c56 feat: add prepare_model_for_kbit_training and update LoRA r=64 alpha=128

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

2026-07-01 07:47:44 -04:00

357c9fa781 fix: use standard QLoRA pattern with BitsAndBytesConfig