fix galore
Former-commit-id: 62a3ceeef8f60caef43ccc7f971a0c9184e21296
This commit is contained in:
@@ -66,10 +66,6 @@ class LoraArguments:
|
||||
Others choices: the same as LLaMA."""
|
||||
},
|
||||
)
|
||||
lora_bf16_mode: bool = field(
|
||||
default=False,
|
||||
metadata={"help": "Whether or not to train lora adapters in bf16 precision."},
|
||||
)
|
||||
use_rslora: bool = field(
|
||||
default=False,
|
||||
metadata={"help": "Whether or not to use the rank stabilization scaling factor for LoRA layer."},
|
||||
@@ -194,6 +190,10 @@ class FinetuningArguments(FreezeArguments, LoraArguments, RLHFArguments, GaloreA
|
||||
Arguments pertaining to which techniques we are going to fine-tuning with.
|
||||
"""
|
||||
|
||||
pure_bf16: bool = field(
|
||||
default=False,
|
||||
metadata={"help": "Whether or not to train model in purely bf16 precision (without AMP)."},
|
||||
)
|
||||
stage: Literal["pt", "sft", "rm", "ppo", "dpo"] = field(
|
||||
default="sft",
|
||||
metadata={"help": "Which stage will be performed in training."},
|
||||
|
||||
Reference in New Issue
Block a user