cmedina/agenx-lora-training
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Commit Graph

  • 14ef1a07c4 fix: remove sync_module_states=True (model is on CPU) Christian Medina 2026-07-02 11:55:21 -04:00
  • 5a13ca1d1c feat: FSDP training failures now fallback to next strategy Christian Medina 2026-07-02 11:48:03 -04:00
  • ac1417567c fix: fix another indentation error Christian Medina 2026-07-02 11:38:54 -04:00
  • 243223d899 fix: fix indentation error in strategy 3 Christian Medina 2026-07-02 11:38:37 -04:00
  • 1b3f678b50 feat: add Strategy 1 - 4-bit QLoRA with device_map=auto (distributed across GPUs) Christian Medina 2026-07-02 11:35:09 -04:00
  • 93ac7391dc fix: clarify strategy naming + init distributed process group for FSDP Christian Medina 2026-07-02 11:30:34 -04:00
  • 8cabc0e986 fix: correct transformer_auto_wrap_policy API signature Christian Medina 2026-07-02 11:12:31 -04:00
  • 651965e844 feat: manually wrap model with FSDP on CPU before trainer Christian Medina 2026-07-02 11:03:41 -04:00
  • b056ec0306 fix: force FSDP1 with string value to avoid FSDP2 gather spike Christian Medina 2026-07-02 10:54:17 -04:00
  • f1e016c8af fix: set gradient_checkpointing=False (FSDP handles it) Christian Medina 2026-07-02 09:38:51 -04:00
  • 361a58addc fix: simplify FSDP config, comment out mixed_precision Christian Medina 2026-07-02 09:38:10 -04:00
  • ab1988705b fix: add use_orig_params=True for LoRA/PEFT compatibility Christian Medina 2026-07-02 09:37:37 -04:00
  • b7d680966e feat: restore FSDP with SHARD_GRAD_OP + sync_module_states Christian Medina 2026-07-02 09:36:08 -04:00
  • 7b39fc3a1b fix: add --mixed_precision bf16 to accelerate launch Christian Medina 2026-07-02 09:26:37 -04:00
  • fe7df7c92f fix: disable FSDP, use standard accelerate data parallelism Christian Medina 2026-07-02 09:22:28 -04:00
  • 42d61c15b0 fix: disable sync_module_states + add mixed precision for FSDP Christian Medina 2026-07-02 09:14:18 -04:00
  • 4eb06a625e fix: add bnb_4bit_quant_storage for FSDP sharding of 4-bit weights Christian Medina 2026-07-02 09:05:00 -04:00
  • 4c9522072e fix: add BitsAndBytesConfig import + remove bf16 GPU strategy Christian Medina 2026-07-02 08:54:41 -04:00
  • 70a2bf0dd6 fix: use FSDP1 with sync_module_states to avoid GPU gather spike Christian Medina 2026-07-02 08:47:56 -04:00
  • c607888d8f fix: load QLoRA model to CPU first, FSDP shards later Christian Medina 2026-07-02 08:26:23 -04:00
  • dbc97138e9 fix: use SHARD_GRAD_OP instead of FULL_SHARD for 2-GPU setup Christian Medina 2026-07-02 00:58:41 -04:00
  • e3ea45fe96 fix: remove gradient_checkpointing (FSDP activation_checkpointing handles it) Christian Medina 2026-07-02 00:41:57 -04:00
  • dfe85c3c82 fix: fsdp=True instead of list + add activation_checkpointing Christian Medina 2026-07-02 00:16:33 -04:00
  • 8747111671 fix: update config comment to reflect Ornith-35B Christian Medina 2026-07-02 00:11:10 -04:00
  • d5ccd28e39 chore: remove temporary debugging code Christian Medina 2026-07-01 23:46:34 -04:00
  • 8a5edb25e0 fix: specify correct transformer layer class for FSDP auto-wrap Christian Medina 2026-07-01 23:45:05 -04:00
  • 05cbd7b6b2 show layer class names Christian Medina 2026-07-01 23:12:28 -04:00
  • e64515e3ca refactor: switch from DeepSpeed ZeRO-3 to FSDP for QLoRA compatibility Christian Medina 2026-07-01 22:33:06 -04:00
  • a655040db2 feat: use DeepSpeed ZeRO-3 for proper model sharding across GPUs Christian Medina 2026-07-01 22:08:37 -04:00
  • 9d8c2b6cf0 refactor: cleaner loading strategies with error tracking Christian Medina 2026-07-01 21:38:38 -04:00
  • 815115e97a feat: add BnB 4-bit quantization strategy + use bf16 model Christian Medina 2026-07-01 21:14:31 -04:00
  • e56ae5e002 fix: restore all strategies with CPU-first as primary Christian Medina 2026-07-01 18:16:51 -04:00
  • f18a9f0973 fix: use aggressive CPU loading for memory optimization Christian Medina 2026-07-01 18:00:55 -04:00
  • 83b45c2a5c feat: add PYTORCH_CUDA_ALLOC_CONF for memory optimization Christian Medina 2026-07-01 17:01:28 -04:00
  • 19f405255c remove auto device map in training strategies Christian Medina 2026-07-01 16:50:48 -04:00
  • cc353b3dea refactor: restructure project - scripts at root level Christian Medina 2026-07-01 16:39:54 -04:00
  • 293f9caf65 fix: use correct repo_root path for dataset files Christian Medina 2026-07-01 16:35:47 -04:00
  • 1095032801 feat: add ai none to unload AI model before training Christian Medina 2026-07-01 16:24:30 -04:00
  • 79aa1e4061 fix: don't reinstall PyTorch (already installed) Christian Medina 2026-07-01 16:22:55 -04:00
  • 3b771c0db5 fix: add venv setup and proper pip install Christian Medina 2026-07-01 16:22:09 -04:00
  • ff0d6e8736 fix: use correct path for training script Christian Medina 2026-07-01 16:21:47 -04:00
  • 5eae235180 executable permissions Christian Medina 2026-07-01 16:06:52 -04:00
  • 7e13680e27 additional fixes Christian Medina 2026-07-01 16:05:52 -04:00
  • 088521094e feat: add 4 loading strategies (no BnB for already-quantized model) Christian Medina 2026-07-01 15:57:42 -04:00
  • 3089f27901 fix: load 4-bit model AS-IS without BnB (already quantized) Christian Medina 2026-07-01 15:56:00 -04:00
  • 1197235b4d fix: skip prepare_model_for_kbit_training to avoid OOM Christian Medina 2026-07-01 15:18:48 -04:00
  • d9fd079686 fix: add accelerate CPU offload to bf16 strategies Christian Medina 2026-07-01 15:14:00 -04:00
  • 4ed80ad3a3 feat: restore all 6 loading strategies Christian Medina 2026-07-01 15:11:45 -04:00
  • 313b44381f fix: remove DeepSpeed, use torchrun without distributed training Christian Medina 2026-07-01 15:10:54 -04:00
  • 7de54746a2 feat: add 6 loading variants with different offload strategies Christian Medina 2026-07-01 14:45:29 -04:00
  • 22d1bfd573 fix: use 4-bit model first, then bf16 with CPU offload Christian Medina 2026-07-01 14:43:30 -04:00
  • 3fa2cfe4cc fix: use DeepSpeed ZeRO-3 directly (skip CPU loading) Christian Medina 2026-07-01 14:41:48 -04:00
  • 3760aaf316 fix: rename output dir to ornith-35b-lora Christian Medina 2026-07-01 14:19:43 -04:00
  • 4287e42148 fix: remove screen, run training directly Christian Medina 2026-07-01 14:16:05 -04:00
  • 3bd51ec7db feat: add screen support to bypass cgroup limits Christian Medina 2026-07-01 14:09:34 -04:00
  • ce758f7eb5 feat: improve loading strategies with low_cpu_mem_usage and proper DeepSpeed config Christian Medina 2026-07-01 14:00:41 -04:00
  • 2d13e63c25 feat: add 4 loading strategies with automatic fallback Christian Medina 2026-07-01 13:55:23 -04:00
  • b300235be0 fix: load 4-bit model AS-IS without additional quantization Christian Medina 2026-07-01 13:53:43 -04:00
  • e967bd74f1 fix: skip quantization, use DeepSpeed ZeRO-3 CPU offload directly Christian Medina 2026-07-01 13:53:06 -04:00
  • 7539e39a93 fix: use accelerate load_checkpoint_and_dispatch for proper CPU offload Christian Medina 2026-07-01 13:39:51 -04:00
  • 528e7002ef feat: add 8-bit CPU offload fallback between 4-bit and bf16 Christian Medina 2026-07-01 13:32:04 -04:00
  • 59ddfb45ed feat: add fallback to bf16 with CPU offload if 4-bit fails Christian Medina 2026-07-01 13:29:23 -04:00
  • b41d99f956 fix: remove broken quantization_config before loading Christian Medina 2026-07-01 13:12:57 -04:00
  • 1228a1f38b fix: load model first then quantize with BnB Christian Medina 2026-07-01 13:11:14 -04:00
  • 14042dbb44 fix: use pre-quantized 4bit model Christian Medina 2026-07-01 12:35:19 -04:00
  • f784d06f0a fix: load QLoRA directly to GPU (skip CPU) Christian Medina 2026-07-01 12:33:21 -04:00
  • 9471d5cc77 fix: cast learning_rate to float (YAML parses 2e-4 as string) Christian Medina 2026-07-01 11:19:00 -04:00
  • 2fa52cd9d4 fix: load QLoRA model to CPU first, DeepSpeed distributes Christian Medina 2026-07-01 09:37:20 -04:00
  • 2fa5173c56 feat: add prepare_model_for_kbit_training and update LoRA r=64 alpha=128 Christian Medina 2026-07-01 07:48:44 -04:00
  • 357c9fa781 fix: use standard QLoRA pattern with BitsAndBytesConfig Christian Medina 2026-07-01 07:47:42 -04:00
  • da5eb3abed feat: implement QLoRA with 4-bit BitsAndBytes quantization Christian Medina 2026-07-01 07:45:33 -04:00
  • f0ee6bc9a2 fix: torch_dtype -> dtype (fix deprecation warning) Christian Medina 2026-07-01 07:38:36 -04:00
  • f651ca030c fix: convert warmup_ratio to warmup_steps (fix deprecation warning) Christian Medina 2026-07-01 07:37:47 -04:00
  • f47944ee75 fix: remove quantization_config after loading to bypass SFTTrainer check Christian Medina 2026-07-01 07:36:38 -04:00
  • 86b401c12f fix: convert FP8 to bf16 for PEFT compatibility Christian Medina 2026-07-01 07:30:10 -04:00
  • 6f631f00ae fix: use FP8 model with 32GB CPU offload Christian Medina 2026-07-01 07:29:32 -04:00
  • 9d92531c55 fix: use Ornith-1.0-35B bf16 base model (PEFT compatible) Christian Medina 2026-07-01 07:28:21 -04:00
  • 05ace2d38f fix: load NVFP4 as bf16 by overriding quantization config Christian Medina 2026-07-01 07:27:13 -04:00
  • be7b589c6b fix: use BitsAndBytes INT4 for PEFT compatibility Christian Medina 2026-07-01 00:34:34 -04:00
  • d7301d37dd fix: use float16 dtype to preserve NVFP4 quantization Christian Medina 2026-07-01 00:08:42 -04:00
  • af09d47c6f fix: load model to CPU first, let DeepSpeed distribute Christian Medina 2026-07-01 00:03:24 -04:00
  • 2c1251788d fix: add compressed-tensors to training dependencies Christian Medina 2026-06-30 23:56:59 -04:00
  • 905c2ba30d fix: preserve NVFP4 quantization, don't force bf16 Christian Medina 2026-06-30 23:50:27 -04:00
  • 114e25fcfb fix: use device_map=auto to distribute model across GPUs Christian Medina 2026-06-30 23:45:57 -04:00
  • 8d4cdca3bb feat: use Ornith-1.0-35B-NVFP4 for training Christian Medina 2026-06-30 23:34:33 -04:00
  • 5713bf0285 fix: add CPU offload back to DeepSpeed ZeRO-3 Christian Medina 2026-06-30 22:47:19 -04:00
  • 3bb7328e2a fix: disable quantization config for fp8 model training Christian Medina 2026-06-30 22:41:28 -04:00
  • a8c9cd682e fix: simplify DeepSpeed config, remove offload settings Christian Medina 2026-06-30 22:36:49 -04:00
  • b760ba09e7 fix: use CUDA 13.0 nightly for RTX 5090 support Christian Medina 2026-06-30 22:32:14 -04:00
  • 9cbf6549e5 fix: use output_file parameter in prepare_dataset Christian Medina 2026-06-30 21:54:05 -04:00
  • ed0b1eddf8 fix: use messages format for SFTTrainer (system/user/assistant) Christian Medina 2026-06-30 21:05:48 -04:00
  • d8ad31fd03 fix: remove max_seq_length from SFTTrainer (newer TRL API) Christian Medina 2026-06-30 20:59:24 -04:00
  • 8ad87f0e7e fix: remove dataset_text_field (dataset pre-formatted) Christian Medina 2026-06-30 20:48:07 -04:00
  • 6bd6662f5f docs: rename config and docs to reflect Ornith-35B model Christian Medina 2026-06-30 20:41:53 -04:00
  • 6f31fa0197 feat: add --check-only flag to validate setup without training Christian Medina 2026-06-30 20:39:29 -04:00
  • 83ba987840 fix: tokenizer -> processing_class SFTTrainer API change Christian Medina 2026-06-30 20:38:31 -04:00
  • 9d5d4eebc7 fix: mixed_precision -> bf16/api change Christian Medina 2026-06-30 20:34:05 -04:00
  • 9280103260 fix: evaluation_strategy -> eval_strategy API change Christian Medina 2026-06-30 20:29:03 -04:00
  • 50fa171f68 fix: use local model path instead of HuggingFace download Christian Medina 2026-06-30 20:24:44 -04:00
  • 7f7467ef89 fix: remove max_seq_length from TrainingArguments (passed to SFTTrainer) Christian Medina 2026-06-30 20:17:04 -04:00