mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-09-22 18:09:06 +00:00
[chat] use official transformers and fix some issues (#4117)
* feat: remove on_learn_epoch fn as not used * revert: add _on_learn_epoch fn * feat: remove NaiveStrategy * test: update train_prompts tests * fix: remove prepare_llama_tokenizer_and_embedding * test: add lora arg * feat: remove roberta support in train_prompts due to runtime errs * feat: remove deberta & roberta in rm as not used * test: remove deberta and roberta tests * feat: remove deberta and roberta models as not used * fix: remove calls to roberta * fix: remove prepare_llama_tokenizer_and_embedding * chore: update transformers version * docs: update transformers version * fix: fix actor inference * fix: fix ci * feat: change llama pad token to unk * revert: revert ddp setup_distributed * fix: change llama pad token to unk * revert: undo unnecessary changes * fix: use pip to install transformers
This commit is contained in:
Wenhao Chen
GitHub
@@ -1,4 +0,0 @@
|
||||
from .deberta_critic import DebertaCritic
|
||||
from .deberta_rm import DebertaRM
|
||||
|
||||
__all__ = ['DebertaCritic', 'DebertaRM']
|
||||
@@ -1,36 +0,0 @@
|
||||
from typing import Optional
|
||||
|
||||
import torch.nn as nn
|
||||
from transformers import DebertaV2Config, DebertaV2Model
|
||||
|
||||
from ..base import Critic
|
||||
|
||||
|
||||
class DebertaCritic(Critic):
|
||||
"""
|
||||
Deberta Critic model.
|
||||
|
||||
Args:
|
||||
pretrained (str): Pretrained model name or path.
|
||||
config (DebertaV2Config): Model config.
|
||||
checkpoint (bool): Enable gradient checkpointing.
|
||||
lora_rank (int): Rank of the LO-RA decomposition.
|
||||
lora_train_bias (str): LoRA bias training mode.
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
pretrained: Optional[str] = None,
|
||||
config: Optional[DebertaV2Config] = None,
|
||||
checkpoint: bool = False,
|
||||
lora_rank: int = 0,
|
||||
lora_train_bias: str = 'none') -> None:
|
||||
if pretrained is not None:
|
||||
model = DebertaV2Model.from_pretrained(pretrained)
|
||||
elif config is not None:
|
||||
model = DebertaV2Model(config)
|
||||
else:
|
||||
model = DebertaV2Model(DebertaV2Config())
|
||||
if checkpoint:
|
||||
model.gradient_checkpointing_enable()
|
||||
value_head = nn.Linear(model.config.hidden_size, 1)
|
||||
super().__init__(model, value_head, lora_rank, lora_train_bias)
|
||||
@@ -1,37 +0,0 @@
|
||||
from typing import Optional
|
||||
|
||||
import torch.nn as nn
|
||||
from transformers import DebertaV2Config, DebertaV2Model
|
||||
|
||||
from ..base import RewardModel
|
||||
|
||||
|
||||
class DebertaRM(RewardModel):
|
||||
"""
|
||||
Deberta Reward model.
|
||||
|
||||
Args:
|
||||
pretrained (str): Pretrained model name or path.
|
||||
config (DebertaV2Config): Model config.
|
||||
checkpoint (bool): Enable gradient checkpointing.
|
||||
lora_rank (int): Rank of the LO-RA decomposition.
|
||||
lora_train_bias (str): LoRA bias training mode.
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
pretrained: str = None,
|
||||
config: Optional[DebertaV2Config] = None,
|
||||
checkpoint: bool = False,
|
||||
lora_rank: int = 0,
|
||||
lora_train_bias: str = 'none') -> None:
|
||||
if pretrained is not None:
|
||||
model = DebertaV2Model.from_pretrained(pretrained)
|
||||
elif config is not None:
|
||||
model = DebertaV2Model(config)
|
||||
else:
|
||||
model = DebertaV2Model(DebertaV2Config())
|
||||
if checkpoint:
|
||||
model.gradient_checkpointing_enable()
|
||||
value_head = nn.Linear(model.config.hidden_size, 1)
|
||||
value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.hidden_size + 1))
|
||||
super().__init__(model, value_head, lora_rank, lora_train_bias)
|
||||
@@ -1,5 +0,0 @@
|
||||
from .roberta_actor import RoBERTaActor
|
||||
from .roberta_critic import RoBERTaCritic
|
||||
from .roberta_rm import RoBERTaRM
|
||||
|
||||
__all__ = ['RoBERTaActor', 'RoBERTaCritic', 'RoBERTaRM']
|
||||
@@ -1,35 +0,0 @@
|
||||
from typing import Optional
|
||||
|
||||
from transformers.models.roberta.configuration_roberta import RobertaConfig
|
||||
from transformers.models.roberta.modeling_roberta import RobertaForCausalLM
|
||||
|
||||
from ..base import Actor
|
||||
|
||||
class RoBERTaActor(Actor):
|
||||
"""
|
||||
RoBERTa Actor model.
|
||||
|
||||
Args:
|
||||
pretrained (str): Pretrained model name or path.
|
||||
config (RoBERTaConfig): Model config.
|
||||
checkpoint (bool): Enable gradient checkpointing.
|
||||
lora_rank (int): Rank of the low-rank approximation.
|
||||
lora_train_bias (str): LoRA bias training mode.
|
||||
"""
|
||||
|
||||
|
||||
def __init__(self,
|
||||
pretrained: Optional[str] = None,
|
||||
config: Optional[RobertaConfig] = None,
|
||||
checkpoint: bool = False,
|
||||
lora_rank: int = 0,
|
||||
lora_train_bias: str = 'none') -> None:
|
||||
if pretrained is not None:
|
||||
model = RobertaForCausalLM.from_pretrained(pretrained)
|
||||
elif config is not None:
|
||||
model = RobertaForCausalLM(config)
|
||||
else:
|
||||
model = RobertaForCausalLM(RobertaConfig())
|
||||
if checkpoint:
|
||||
model.gradient_checkpointing_enable()
|
||||
super().__init__(model, lora_rank, lora_train_bias)
|
||||
@@ -1,38 +0,0 @@
|
||||
from typing import Optional
|
||||
|
||||
import torch.nn as nn
|
||||
from transformers.models.roberta.configuration_roberta import RobertaConfig
|
||||
from transformers.models.roberta.modeling_roberta import RobertaModel
|
||||
|
||||
from ..base import Critic
|
||||
|
||||
|
||||
class RoBERTaCritic(Critic):
|
||||
"""
|
||||
RoBERTa Critic model.
|
||||
|
||||
Args:
|
||||
pretrained (str): Pretrained model name or path.
|
||||
config (RoBERTa Config): Model config.
|
||||
checkpoint (bool): Enable gradient checkpointing.
|
||||
lora_rank (int): Rank of the low-rank approximation.
|
||||
lora_train_bias (str): LoRA bias training mode.
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
pretrained: Optional[str] = None,
|
||||
config: Optional[RobertaConfig] = None,
|
||||
checkpoint: bool = False,
|
||||
lora_rank: int = 0,
|
||||
lora_train_bias: str = 'none',
|
||||
**kwargs) -> None:
|
||||
if pretrained is not None:
|
||||
model = RobertaModel.from_pretrained(pretrained, add_pooling_layer=False)
|
||||
elif config is not None:
|
||||
model = RobertaModel(config)
|
||||
else:
|
||||
model = RobertaModel(RobertaConfig())
|
||||
if checkpoint:
|
||||
model.gradient_checkpointing_enable()
|
||||
value_head = nn.Linear(model.config.hidden_size, 1)
|
||||
super().__init__(model, value_head, lora_rank, lora_train_bias, **kwargs)
|
||||
@@ -1,39 +0,0 @@
|
||||
from typing import Optional
|
||||
|
||||
import torch.nn as nn
|
||||
from transformers import RobertaConfig, RobertaModel
|
||||
|
||||
|
||||
from ..base import RewardModel
|
||||
|
||||
|
||||
class RoBERTaRM(RewardModel):
|
||||
"""
|
||||
RoBERTa Reward model.
|
||||
|
||||
Args:
|
||||
pretrained (str): Pretrained model name or path.
|
||||
config (RoBERTaConfig): Model config.
|
||||
checkpoint (bool): Enable gradient checkpointing.
|
||||
lora_rank (int): Rank of the low-rank approximation.
|
||||
lora_train_bias (str): LoRA bias training mode.
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
pretrained: Optional[str] = None,
|
||||
config: Optional[RobertaConfig] = None,
|
||||
checkpoint: bool = False,
|
||||
lora_rank: int = 0,
|
||||
lora_train_bias: str = 'none') -> None:
|
||||
if pretrained is not None:
|
||||
model = RobertaModel.from_pretrained(pretrained, add_pooling_layer=False)
|
||||
elif config is not None:
|
||||
model = RobertaModel(config)
|
||||
else:
|
||||
model = RobertaModel(RobertaConfig())
|
||||
if checkpoint:
|
||||
model.gradient_checkpointing_enable()
|
||||
|
||||
value_head = nn.Linear(model.config.hidden_size, 1)
|
||||
value_head.weight.data.normal_(mean=0.0, std=1/(model.config.hidden_size + 1))
|
||||
super().__init__(model, value_head, lora_rank, lora_train_bias)
|
||||
Reference in New Issue
Block a user