ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

tiny fix about badam

Browse Source 
Former-commit-id: 03f49267c7406e36aee35639f86e6e0383897090
This commit is contained in:
hiyouga
2024-06-25 01:54:53 +08:00
parent 98fb3d015a
commit 9fd7a410bb
15 changed files with 31 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/llamafactory/hparams/parser.py
View File

@@ -121,7 +121,7 @@ def _check_extra_dependencies(
require_version("galore_torch", "To fix: pip install galore_torch")
if finetuning_args.use_badam:
require_version("badam", "To fix: pip install badam")
require_version("badam>=1.2.1", "To fix: pip install badam>=1.2.1")
if finetuning_args.plot_loss:
require_version("matplotlib", "To fix: pip install matplotlib")
@@ -214,15 +214,15 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
if (
finetuning_args.use_badam
and training_args.parallel_mode.value == "distributed"
and training_args.parallel_mode == ParallelMode.DISTRIBUTED
):
if finetuning_args.badam_mode == "ratio":
raise ValueError("Ratio-wise BAdam does not yet support distributed training, use layer-wise BAdam: --badam_mode layer")
if finetuning_args.badam_mode == "layer" and (not is_deepspeed_zero3_enabled()):
raise ValueError(f"Layer-wise BAdam only supports DeepSpeed ZeRO 3 stage.")
raise ValueError("Radio-based BAdam does not yet support distributed training, use layer-wise BAdam.")
elif not is_deepspeed_zero3_enabled():
raise ValueError("Layer-wise BAdam only supports DeepSpeed ZeRO-3 training.")
if (finetuning_args.use_galore) and training_args.deepspeed is not None:
raise ValueError("GaLore are incompatible with DeepSpeed yet.")
if finetuning_args.use_galore and training_args.deepspeed is not None:
raise ValueError("GaLore is incompatible with DeepSpeed yet.")
if model_args.infer_backend == "vllm":
raise ValueError("vLLM backend is only available for API, CLI and Web.")

3
src/llamafactory/train/dpo/trainer.py
View File

@@ -96,7 +96,8 @@ class CustomDPOTrainer(DPOTrainer):
self.save_model(os.path.join(self.args.output_dir, "pissa_init"))
if finetuning_args.use_badam:
from badam import clip_grad_norm_old_version, BAdamCallback
from badam import BAdamCallback, clip_grad_norm_old_version
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.callback_handler.add_callback(BAdamCallback)

3
src/llamafactory/train/kto/trainer.py
View File

@@ -91,7 +91,8 @@ class CustomKTOTrainer(KTOTrainer):
self.ref_model.eval()
if finetuning_args.use_badam:
from badam import clip_grad_norm_old_version, BAdamCallback
from badam import BAdamCallback, clip_grad_norm_old_version
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.callback_handler.add_callback(BAdamCallback)

3
src/llamafactory/train/ppo/trainer.py
View File

@@ -166,7 +166,8 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
self.reward_model = self.accelerator.prepare_model(self.reward_model, evaluation_mode=True)
if finetuning_args.use_badam:
from badam import clip_grad_norm_old_version, BAdamCallback
from badam import BAdamCallback, clip_grad_norm_old_version
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.callback_handler.add_callback(BAdamCallback)

3
src/llamafactory/train/pt/trainer.py
View File

@@ -48,7 +48,8 @@ class CustomTrainer(Trainer):
self.save_model(os.path.join(self.args.output_dir, "pissa_init"))
if finetuning_args.use_badam:
from badam import clip_grad_norm_old_version, BAdamCallback
from badam import BAdamCallback, clip_grad_norm_old_version
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.callback_handler.add_callback(BAdamCallback)

3
src/llamafactory/train/rm/trainer.py
View File

@@ -72,7 +72,8 @@ class PairwiseTrainer(Trainer):
self.processor = processor
self.can_return_loss = True # override property to return eval_loss
if finetuning_args.use_badam:
from badam import clip_grad_norm_old_version, BAdamCallback
from badam import BAdamCallback, clip_grad_norm_old_version
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.callback_handler.add_callback(BAdamCallback)

3
src/llamafactory/train/sft/trainer.py
View File

@@ -56,7 +56,8 @@ class CustomSeq2SeqTrainer(Seq2SeqTrainer):
self.save_model(os.path.join(self.args.output_dir, "pissa_init"))
if finetuning_args.use_badam:
from badam import clip_grad_norm_old_version, BAdamCallback
from badam import BAdamCallback, clip_grad_norm_old_version
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
self.callback_handler.add_callback(BAdamCallback)

7
src/llamafactory/train/trainer_utils.py
View File

@@ -23,6 +23,7 @@ from typing import TYPE_CHECKING, Callable, Dict, List, Optional, Tuple, Union
import torch
from peft import PeftModel
from transformers import Trainer
from transformers.integrations import is_deepspeed_zero3_enabled
from transformers.optimization import get_scheduler
from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
from transformers.trainer_pt_utils import get_parameter_names
@@ -372,9 +373,6 @@ def _create_badam_optimizer(
dict(params=decay_params, weight_decay=training_args.weight_decay),
]
from transformers.integrations import is_deepspeed_zero3_enabled
ds_zero3_enabled = is_deepspeed_zero3_enabled()
if finetuning_args.badam_mode == "layer":
from badam import BlockOptimizer
@@ -387,7 +385,7 @@ def _create_badam_optimizer(
start_block=finetuning_args.badam_start_block,
switch_mode=finetuning_args.badam_switch_mode,
verbose=finetuning_args.badam_verbose,
ds_zero3_enabled=ds_zero3_enabled
ds_zero3_enabled=is_deepspeed_zero3_enabled(),
)
logger.info(
f"Using BAdam optimizer with layer-wise update, switch mode is {finetuning_args.badam_switch_mode}, "
@@ -398,7 +396,6 @@ def _create_badam_optimizer(
elif finetuning_args.badam_mode == "ratio":
from badam import BlockOptimizerRatio
assert not ds_zero3_enabled, "BAdam with ratio-based update does not support Deepspeed ZeRO-3 yet, use layer-wise update instead: --badam_mode layer."
assert finetuning_args.badam_update_ratio > 1e-6
optimizer = BlockOptimizerRatio(
param_groups=param_groups,