cmedina/agenx-lora-training
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
175 Commits 1 Branch 0 Tags
fa2f21562f9f2335eb36e85efd88800c170c4f7f
Commit Graph

50 Commits

Author SHA1 Message Date
Christian Medina
fa2f21562f fix: load BnB 4-bit model to single GPU (no distribution) 2026-07-02 20:42:05 -04:00
Christian Medina
57b169f202 feat: use BnB 4-bit model for QLoRA training 2026-07-02 20:04:55 -04:00
Christian Medina
3c31e4dfb0 feat: load bf16 model, quantize with BnB on CPU, move to GPU 2026-07-02 19:55:01 -04:00
Christian Medina
a0f4f644b0 fix: load 4-bit model with device_map=auto (let transformers distribute) 2026-07-02 19:33:37 -04:00
Christian Medina
20626e7a78 feat: manual 4-bit quantization with PEFT prepare + Linear4bit 2026-07-02 18:46:05 -04:00
Christian Medina
42c25962b2 feat: load model to CPU with BnB 4-bit, then move to GPU 2026-07-02 18:23:39 -04:00
Christian Medina
1c7e73e927 fix: load already-quantized 4-bit model directly (no BnB) 2026-07-02 18:20:43 -04:00
Christian Medina
199176c8d0 feat: load model with BnB 4-bit quantization (fit in single GPU) 2026-07-02 18:05:27 -04:00
Christian Medina
5a84b0eee5 feat: add CPU offload for optimizer states 2026-07-02 17:53:56 -04:00
Christian Medina
691e5d3611 feat: use 8-bit AdamW optimizer to save VRAM 2026-07-02 17:53:05 -04:00
Christian Medina
92999b7783 fix: load model to single GPU (MoE requires all layers on same device) 2026-07-02 17:50:01 -04:00
Christian Medina
dd86a4df37 fix: rename dataset column to 'text' for SFTTrainer compatibility 2026-07-02 16:39:21 -04:00
Christian Medina
36fe2c8d45 fix: pass dataset_text_field to SFTTrainer from config 2026-07-02 16:24:34 -04:00
Christian Medina
9e1c5120fb fix: single-process training, remove FSDP (model pre-distributed via device_map) 2026-07-02 15:53:51 -04:00
Christian Medina
4348323116 feat: use accelerate device_map for DISTRIBUTED model loading (Test 7 method) 2026-07-02 15:19:04 -04:00
Christian Medina
7f498134f2 fix: FSDP loads model to GPU first, then shards across GPUs 2026-07-02 12:17:44 -04:00
Christian Medina
f9c748706f fix: use already-quantized 4-bit model for training (no BnB needed) 2026-07-02 12:09:59 -04:00
Christian Medina
14ef1a07c4 fix: remove sync_module_states=True (model is on CPU) 2026-07-02 11:55:21 -04:00
Christian Medina
5a13ca1d1c feat: FSDP training failures now fallback to next strategy 2026-07-02 11:48:03 -04:00
Christian Medina
ac1417567c fix: fix another indentation error 2026-07-02 11:38:54 -04:00
Christian Medina
243223d899 fix: fix indentation error in strategy 3 2026-07-02 11:38:37 -04:00
Christian Medina
1b3f678b50 feat: add Strategy 1 - 4-bit QLoRA with device_map=auto (distributed across GPUs) 2026-07-02 11:35:09 -04:00
Christian Medina
93ac7391dc fix: clarify strategy naming + init distributed process group for FSDP 2026-07-02 11:30:34 -04:00
Christian Medina
8cabc0e986 fix: correct transformer_auto_wrap_policy API signature 2026-07-02 11:12:31 -04:00
Christian Medina
651965e844 feat: manually wrap model with FSDP on CPU before trainer 2026-07-02 11:03:41 -04:00
Christian Medina
b056ec0306 fix: force FSDP1 with string value to avoid FSDP2 gather spike 2026-07-02 10:54:17 -04:00
Christian Medina
f1e016c8af fix: set gradient_checkpointing=False (FSDP handles it) 2026-07-02 09:38:51 -04:00
Christian Medina
361a58addc fix: simplify FSDP config, comment out mixed_precision 2026-07-02 09:38:10 -04:00
Christian Medina
ab1988705b fix: add use_orig_params=True for LoRA/PEFT compatibility 2026-07-02 09:37:37 -04:00
Christian Medina
b7d680966e feat: restore FSDP with SHARD_GRAD_OP + sync_module_states 2026-07-02 09:36:08 -04:00
Christian Medina
fe7df7c92f fix: disable FSDP, use standard accelerate data parallelism 2026-07-02 09:22:28 -04:00
Christian Medina
42d61c15b0 fix: disable sync_module_states + add mixed precision for FSDP 2026-07-02 09:14:18 -04:00
Christian Medina
4eb06a625e fix: add bnb_4bit_quant_storage for FSDP sharding of 4-bit weights 2026-07-02 09:05:00 -04:00
Christian Medina
4c9522072e fix: add BitsAndBytesConfig import + remove bf16 GPU strategy 2026-07-02 08:54:41 -04:00
Christian Medina
70a2bf0dd6 fix: use FSDP1 with sync_module_states to avoid GPU gather spike 2026-07-02 08:47:56 -04:00
Christian Medina
c607888d8f fix: load QLoRA model to CPU first, FSDP shards later 2026-07-02 08:26:23 -04:00
Christian Medina
dbc97138e9 fix: use SHARD_GRAD_OP instead of FULL_SHARD for 2-GPU setup 2026-07-02 00:58:41 -04:00
Christian Medina
e3ea45fe96 fix: remove gradient_checkpointing (FSDP activation_checkpointing handles it) 2026-07-02 00:41:57 -04:00
Christian Medina
dfe85c3c82 fix: fsdp=True instead of list + add activation_checkpointing 2026-07-02 00:16:33 -04:00
Christian Medina
d5ccd28e39 chore: remove temporary debugging code 2026-07-01 23:46:34 -04:00
Christian Medina
8a5edb25e0 fix: specify correct transformer layer class for FSDP auto-wrap 2026-07-01 23:45:05 -04:00
Christian Medina
05cbd7b6b2 show layer class names 2026-07-01 23:27:25 -04:00
Christian Medina
e64515e3ca refactor: switch from DeepSpeed ZeRO-3 to FSDP for QLoRA compatibility 2026-07-01 22:33:06 -04:00
Christian Medina
a655040db2 feat: use DeepSpeed ZeRO-3 for proper model sharding across GPUs 2026-07-01 22:08:37 -04:00
Christian Medina
9d8c2b6cf0 refactor: cleaner loading strategies with error tracking 2026-07-01 21:38:38 -04:00
Christian Medina
815115e97a feat: add BnB 4-bit quantization strategy + use bf16 model 2026-07-01 21:14:31 -04:00
Christian Medina
e56ae5e002 fix: restore all strategies with CPU-first as primary 2026-07-01 18:16:51 -04:00
Christian Medina
f18a9f0973 fix: use aggressive CPU loading for memory optimization 2026-07-01 18:00:55 -04:00
Christian Medina
19f405255c remove auto device map in training strategies 2026-07-01 16:50:48 -04:00
Christian Medina
cc353b3dea refactor: restructure project - scripts at root level 2026-07-01 16:39:54 -04:00