fix ppo args

This commit is contained in:
hiyouga 2023-10-11 23:40:50 +08:00
parent 2818af0b09
commit 11bd271364
4 changed files with 18 additions and 9 deletions

View File

@ -57,7 +57,15 @@ class FinetuningArguments:
)
ppo_score_norm: Optional[bool] = field(
default=False,
metadata={"help": "Use score normalization in PPO Training."}
metadata={"help": "Use score normalization in PPO training."}
)
ppo_logger: Optional[str] = field(
default=None,
metadata={"help": "Log with either 'wandb' or 'tensorboard' in PPO training."}
)
ppo_target: Optional[float] = field(
default=6.0,
metadata={"help": "Target KL value for adaptive KL control in PPO training."}
)
dpo_beta: Optional[float] = field(
default=0.1,

View File

@ -201,7 +201,9 @@ def get_train_args(
)
# postprocess model_args
model_args.compute_dtype = torch.bfloat16 if training_args.bf16 else (torch.float16 if training_args.fp16 else None)
model_args.compute_dtype = (
torch.bfloat16 if training_args.bf16 else (torch.float16 if training_args.fp16 else torch.float32)
)
model_args.model_max_length = data_args.cutoff_len
# Log on each process the small summary:

View File

@ -206,7 +206,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
replace_model(unwrapped_model, target="reward")
batch = self.prepare_model_inputs(queries, responses)
with torch.cuda.amp.autocast(dtype=self.compute_dtype): # support bf16
with torch.cuda.amp.autocast(dtype=self.model_args.compute_dtype): # support bf16
_, _, values = self.model(**batch, output_hidden_states=True, return_dict=True)
if values.size(0) != batch["input_ids"].size(0): # adapt to chatglm2
@ -251,7 +251,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
input_ids = input_kwargs["input_ids"]
attention_mask = input_kwargs["attention_mask"]
with torch.cuda.amp.autocast(dtype=self.compute_dtype): # support bf16
with torch.cuda.amp.autocast(dtype=self.model_args.compute_dtype): # support bf16
logits, _, values = model(**input_kwargs)
if values.size(0) != input_ids.size(0): # adapt to chatglm2

View File

@ -42,15 +42,14 @@ def run_ppo(
ppo_epochs=1,
max_grad_norm=training_args.max_grad_norm,
seed=training_args.seed,
log_with=training_args.report_to,
optimize_cuda_cache=True,
target=finetuning_args.ppo_target,
log_with=finetuning_args.ppo_logger,
use_score_scaling=finetuning_args.ppo_score_norm,
use_score_norm=finetuning_args.ppo_score_norm,
accelerator_kwargs={"step_scheduler_with_optimizer": False}
)
if finetuning_args.ppo_score_norm:
ppo_config.use_score_scaling = True
ppo_config.use_score_norm = True
optimizer = AdamW(filter(lambda p: p.requires_grad, model.parameters()), lr=training_args.learning_rate)
total_train_batch_size = (
training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps * training_args.world_size