biG
d dl Z d dlmZ d dlmZ d dlmZ d dlmZm Z m
Z
m Z m Z d dl
mZ d dlmZ d dlmZmZmZ dd lmZmZ eefZerd d
lmZ d d lm Z d d l
m Z d d
l m Z e G d d Z de iZ! d&dedede e
d de e" de#eef f
dZ$ d'dedede%de e" de#eef f
dZ&d(dZ'd)dZ(d)dZ)d(d Z*e d*de d d d d e+fd! Z,d+d"Z-d# Z. G d$ d% Z/y), N)contextmanager)deepcopy) dataclass)
TYPE_CHECKINGAnyLiteralOptionalUnion)version)
AutoTokenizerPreTrainedModelPreTrainedTokenizer )!AutoModelForCausalLMWithValueHead"AutoModelForSeq2SeqLMWithValueHead) Accelerator)DeepSpeedEngine)Module)DistributedDataParallelc | e Zd ZU dZdZeed< dZeed< dZeed< e d Z
e d Z e d Z e d
Z
y ) ChatMlSpecialTokensziDataclass for special tokens used in ChatML, including system, user, assistant, bos, eos, and pad tokens.z <|im_start|> bos_tokenz
<|im_end|> eos_token pad_tokenc | j dS )Nsystemr selfs K/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/trl/models/utils.pyr zChatMlSpecialTokens.system1 s ..!(( c | j dS )Nuserr r s r r# zChatMlSpecialTokens.user5 s ..!&&r! c | j dS )N assistantr r s r r% z ChatMlSpecialTokens.assistant9 s ..!++r! c V d| j d| j d| j dS )Nz {% for message in messages %}{{'z2' + message['role'] + '
' + message['content'] + 'z7' + '
'}}{% endfor %}{% if add_generation_prompt %}{{ 'z
' }}{% endif %})r r r% r s r
chat_templatez!ChatMlSpecialTokens.chat_template= sA
NN##VW[WeWeVf g ^^$ %
r! N)__name__
__module__ __qualname____doc__r str__annotations__r r propertyr r# r% r' r! r r r ) sq s#Is#!Is!!Is!
) ) ' ' , ,
r! r chatmlmodel tokenizerformatresize_to_multiple_ofreturnc N |j t d |t vr#t d| dt j t | }|j |_ |j
|_ |j |_ |j d|j |j gi |j |_ | j t |j ||nd t | dd Q|j | j _ |j | j _ |j | j _ t | dd Q|j | j _ |j | j _ |j | j _ | |fS ) a
Setup chat format by adding special tokens to the tokenizer, setting the correct format, and extending the
embedding layer of the model based on the new special tokens.
We recommend using [`clone_chat_template`] instead of this function.
If the model already has a chat template, this will throw an error. If you want to overwrite it, please set
`tokenizer.chat_template` to `None`.
Args:
model (`~transformers.PreTrainedModel`): The model to be modified.
tokenizer (`~transformers.PreTrainedTokenizer`): The tokenizer to be modified.
format (`Optional[Literal["chatml"]]`): The format to be set. Defaults to "chatml".
resize_to_multiple_of (`int` or `None`): Number to resize the embedding layer to. Defaults to None.
Returns:
model (`~transformers.PreTrainedModel`):
The modified model.
tokenizer (`~transformers.PreTrainedTokenizer`):
The modified tokenizer.
NzcChat template is already added to the tokenizer. If you want to overwrite it, please set it to NonezFormat z" not available. Please use one of additional_special_tokensnew_num_tokenspad_to_multiple_ofconfiggeneration_config)r'
ValueErrorFORMAT_MAPPINGkeysr r r add_special_tokensresize_token_embeddingslenvocabgetattr pad_token_idr; bos_token_id eos_token_idr< )r1 r2 r3 r4 chat_formats r setup_chat_formatrI L s < * q
^#76(*L^M`M`MbLcdee (*K &//I%//I%//I
"= @U@UWbWlWl?m!no ) 7 7I
!! 9??+4I4U0[_ " uh%1$-$:$: !$-$:$: !$-$:$: !u)40 >?M - .w 7E u||^T:
)) * }d;
"uz
!>!c+BcBc&22 + < < * A A"-"@"@!,!>!> + < < * : : ( 6 6 * : :!,!>!>$++
U%f% JJL Lr! c
e Zd ZdZdej
dej
dededef
dZdej
dej
dd fd
Z dej
dej
dd fd Z
y ) _ForwardRedirectionaI Implements the `forward-redirection`.
Taken from Pytorch-lightning:
https://github.com/Lightning-AI/pytorch-lightning/blob/02311d03fb982560246eead7c08104481fac9579/src/lightning/pytorch/strategies/strategy.py#L602
A method call to a wrapped module gets rerouted through the wrapper's `forward` method instead.
wrapper_moduleoriginal_modulemethodargsr c j dt dt dt f fd }|_ |i |} j |S )a Reroutes a method call through the `wrapper_module`'s `forward` method.
Args:
wrapper_module: The module that has `original_module` wrapped.
original_module: The module that was wrapped inside `wrapper_module`.
method_name: The name of the method that should be called on the `original_module` after inputs get
redirected through the `wrapper_module`'s `forward` method.
*args: The positional arguments to the method `method_name`. They will get passed to a patched
`forward` method instead.
**kwargs: The keyword arguments to the method `method_name`. They will get passed to a patched
`forward` method instead.
_args_kwargsr5 c J _ | i |}j |S ro )forwardon_after_inner_forward)r r outr original_forwardr r r s r wrapped_forwardz5_ForwardRedirection.__call__..wrapped_forward s1 '7O #%+7+C ' ' HJr! )r r on_after_outer_forward) r r r r r r r wrapper_outputr s ```` @r __call__z _ForwardRedirection.__call__ s_ +22 C C C #2 '88 ##NOD r! r5 Nc y ro r/ r r r s r r z*_ForwardRedirection.on_after_inner_forward r! c y ro r/ r s r r z*_ForwardRedirection.on_after_outer_forward r r! ) r( r) r* r+ nnr callabler r r r r/ r! r r r x s ii :<)) MU ^a mp D
RYY
QSQZQZ
_c
RYY
QSQZQZ
_c
r! r )r0 N)@ )r1 r r5 N)F)T)r1 r rx r )0rh
contextlibr copyr dataclassesr typingr r r r r
torch.nnr packagingr transformersr r
r modeling_value_headr r SUPPORTED_ARCHITECTURES
accelerater deepspeed.runtime.enginer r torch.nn.parallel.distributedr r r> inttuplerI r, rS rf rm r\ rw boolr r r r r/ r! r r s~ % ! ? ? L L f &&
&8 E
> /0 +3+/ D D "D
WX& 'D $C= D
? / /0 D V ,. B B "B B $C= B
? / /0 B J*.nGP& &*& = >& & #& & R%P80
0
r!