frozenleaves/openmind
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

!190 Update datasets loader and pissa algorithm related

Browse Source 
Merge pull request !190 from 金勇旭/master
This commit is contained in:
金勇旭
2025-04-30 06:32:51 +00:00
committed by i-robot
parent 6624b5ce3b
commit 4fe36aeeef
2 changed files with 11 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/openmind/flow/datasets/loader.py
View File

@ -51,7 +51,7 @@ class DatasetFormatConfig:
DATASET_FORMAT_REGISTRY: Dict[str, DatasetFormatConfig] = {
"alpaca": DatasetFormatConfig(
required_columns=["instruction", "output"],
allowed_columns=["instruction", "input", "output", "history", "system", "tools"],
allowed_columns=["instruction", "input", "output", "history", "system", "tools", "text"],
),
"sharegpt": DatasetFormatConfig(
required_columns=["conversations"], allowed_columns=["conversations", "system", "tools"]

15
src/openmind/flow/model/adapter.py
View File

@ -60,9 +60,6 @@ def apply_lora(
if not is_trainable:
return model
logger.info_rank0(f"Fine-tuning method: {'DoRA' if args.use_dora else 'LoRA'}")
if args.init_lora_weights:
logger.info_rank0(f"Initializing lora weights method: {args.init_lora_weights}")
all_target_modules = get_all_lora_target_modules(model)
if not args.lora_target_modules:
@ -77,8 +74,6 @@ def apply_lora(
f"The input lora modules {input_target_modules} is not supported. The possible lora modules list is {all_target_modules}"
)
logger.info_rank0(f"Lora target modeules {target_modules} are applied to model")
peft_kwargs = {
"r": args.lora_rank,
"target_modules": target_modules,
@ -98,6 +93,16 @@ def apply_lora(
model.enable_input_require_grads()
model = get_peft_model(model, lora_config)
finetuning_method = "DoRA" if model.get_peft_config_as_dict()["use_dora"] else "LoRA"
logger.info_rank0(f"Fine-tuning method: {finetuning_method}")
if args.init_lora_weights:
init_lora_weights_method = model.get_peft_config_as_dict()["init_lora_weights"]
logger.info_rank0(f"Initializing lora weights method: {init_lora_weights_method}")
logger.info_rank0(f"Lora target modeules {target_modules} are applied to model")
model.print_trainable_parameters()
# Cast trainable parameters to fp32. According to the source code of PEFT, it has an improvement in stability.