cmedina

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

2026-07-02 16:39:24 -04:00

dd86a4df37 fix: rename dataset column to 'text' for SFTTrainer compatibility

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

2026-07-02 16:24:36 -04:00

36fe2c8d45 fix: pass dataset_text_field to SFTTrainer from config

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

2026-07-02 16:14:55 -04:00

728073ef6a fix: use 'output' column (not 'text')

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

2026-07-02 16:04:22 -04:00

0752c03b83 fix: use absolute path for dataset (no ~ expansion)

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

2026-07-02 15:53:52 -04:00

9e1c5120fb fix: single-process training, remove FSDP (model pre-distributed via device_map)

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

2026-07-02 15:23:29 -04:00

b8214f64cb fix: use 4-bit model, remove DeepSpeed, fix dataset path

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

2026-07-02 15:19:05 -04:00

4348323116 feat: use accelerate device_map for DISTRIBUTED model loading (Test 7 method)

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

2026-07-02 14:30:21 -04:00

48e44cdf1c feat: add Test 11 - PEFT prepare + manual 4-bit quantization

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

2026-07-02 14:17:54 -04:00

da8658ae73 fix: switch to CompressedTensors 4-bit checkpoint (BnB on CPU not working)

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

2026-07-02 13:59:26 -04:00

81dea9b9c8 docs: add model class verification to test output

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

2026-07-02 13:56:47 -04:00

f608c7656f feat: dynamically detect layer names from model config

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

2026-07-02 13:54:20 -04:00

a55c39c818 fix: use device_map=cpu with BnB config for quantization

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

2026-07-02 13:53:47 -04:00

3480dd9fbd docs: add model inspection script and comment failing tests

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

2026-07-02 13:45:19 -04:00

048e68f91a refactor: comment out failing tests, add Test 7-10 variations

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

2026-07-02 13:34:36 -04:00

e2b351e04c fix: actually quantize model with quantize_4bit()

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

2026-07-02 13:26:53 -04:00

aff91c990b feat: add Test 6 - bf16 to CPU → BnB 4-bit quantize → GPU

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

2026-07-02 13:20:27 -04:00

ffd1d29c35 fix: use bf16 model with BnB 4-bit (ON-THE-FLY quantization)

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

2026-07-02 12:33:01 -04:00

48a2518b4c feat: comprehensive test of 5 loading strategies

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

2026-07-02 12:23:36 -04:00

3892ea7fec feat: test with BnB 4-bit quantization to force proper loading

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

2026-07-02 12:19:07 -04:00

c5a3b87eed feat: test both strategies - device_map=auto and FSDP