cmedina

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

2026-07-03 04:19:45 -04:00

e25f146fb8 feat: add proper BnB quantization script for 2x RTX 5090

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

2026-07-03 04:10:56 -04:00

d2dde1079f fix: Path conversion for shard_path

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

2026-07-03 04:09:25 -04:00

3498a40390 fix: add argparser and fix Path issue

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

2026-07-03 04:05:15 -04:00

c84a3b6252 feat: add safetensors conversion script

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

2026-07-03 03:58:26 -04:00

1ebad2365c fix: use torch.load for quantized shards index creation

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

2026-07-03 03:58:17 -04:00

5078d7c8dc feat: test all quantized shards loading

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

2026-07-03 02:07:12 -04:00

1c576aac5d fix: weights_only=False for QuantState

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

2026-07-03 02:04:52 -04:00

5b2d629d11 feat: test quantized tensor shapes directly

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

2026-07-03 02:00:31 -04:00

1c792d23e4 fix: save as model.safetensors for from_pretrained

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

2026-07-03 01:57:32 -04:00

7115d56723 fix: copy config.json for load test

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

2026-07-03 01:52:49 -04:00

a549084013 feat: add test for quantized shard loading

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

2026-07-03 01:52:09 -04:00

4c7ca01d7d feat: add single shard test script

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

2026-07-03 01:51:24 -04:00

d8393969fc fix: save quantization states with torch.save

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

2026-07-03 01:50:03 -04:00

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

2026-07-03 01:41:54 -04:00

0878b11e98 fix: restore streaming with device_map=auto and max_memory

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

2026-07-03 01:40:35 -04:00

5018be86ec fix: use accelerate device_map for proper MoE quantization

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

2026-07-03 01:39:29 -04:00

49b99889ef feat: add proper BnB quantization script (loads to CPU, saves)

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

2026-07-03 01:38:56 -04:00

b82be559dc fix: use Linear4bit to preserve weight shapes during quantization

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

2026-07-03 01:25:19 -04:00

aa7961605f fix: add index file creation for sharded models

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

2026-07-03 01:25:12 -04:00

5d0ff6c8af simplify: only train, no quantization logic