qiOS N d Z ddlmZ ddlZddlmZ ddlmZmZ ddlm Z m
Z
ddl m Z m
Z
dd lmZmZ dd
lmZ dd lmZ dd lmZ dd
lmZmZ ddlmZ ddl m Z ddl m Z ddl m!Z!m"Z"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z) ddl*m+Z+m,Z, ejZ e. Z/dZ0dZ1 G d de Z2 G d de' Z3 G d de Z4 G d d ejj Z6 G d d e, Z7 G d d e+ Z8 G d! d"e& Z9 G d# d$e9e% Z: G d% d&e! Z; G d' d(e# Z< G d) d*e$ Z= G d+ d,e" Z>g d-Z?y).zLG AI Research EXAONE Lab )CallableN)nn )Cache DynamicCache)PreTrainedConfiglayer_type_validation)create_causal_mask!create_sliding_window_causal_mask)BaseModelOutputWithPastCausalLMOutputWithPast)RopeParameters)ALL_ATTENTION_FUNCTIONS)Unpack)TransformersKwargslogging)merge_with_config_defaults)capture_outputs )Gemma2RotaryEmbedding) LlamaForCausalLMLlamaForQuestionAnswering LlamaForSequenceClassificationLlamaForTokenClassification
LlamaModelLlamaPreTrainedModel LlamaRMSNormapply_rotary_pos_embeager_attention_forward)Olmo2DecoderLayerOlmo2MLPzLGAI-EXAONE/EXAONE-4.0-32B
Exaone4Configc * t e Zd ZdZdZdgZdddddddddd Zdgd gfd
d gd
gfd
gd
gfd Z d#ded
z ded
z ded
z ded
z ded
z ded
z de d
z ded
z de
d
z ded
z de d
z ded
z ded
z ded
z d e d
z d e e
e e f z d
z d e
d
z d ed
z d ed
z d!ee d
z f( fd"
Z xZS )$r" a
This is the configuration class to store the configuration of a [`Exaone4Model`]. It is used to
instantiate a EXAONE 4.0 model according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the EXAONE-4.0-32B [LGAI-EXAONE/EXAONE-4.0-32B](https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B)
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model
outputs. Read the documentation from [`PreTrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 102400):
Vocabulary size of the EXAONE 4.0 model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`Exaone4Model`].
hidden_size (`int`, *optional*, defaults to 4096):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to `hidden_size * 4`):
Dimensionality of the MLP representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer decoder.
num_key_value_heads (`int`, *optional*):
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
by meanpooling all the original heads within that group. For more details checkout [this
paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to
`num_attention_heads`.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
max_position_embeddings (`int`, *optional*, defaults to 2048):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 32768 for EXAONE 3.5).
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
rms_norm_eps (`float`, *optional*, defaults to 1e-05):
The epsilon used by the layer normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if ``config.is_decoder=True``.
bos_token_id (`int`, *optional*, defaults to 0):
Beginning of stream token id.
eos_token_id (`int`, *optional*, defaults to 2):
End of stream token id.
pad_token_id (`int`, *optional*):
The id of the padding token.
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether to tie weight embeddings
rope_parameters (`RopeParameters`, *optional*):
Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain
a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE
with longer `max_position_embeddings`.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
sliding_window (`int`, *optional*):
The size of the sliding window for the sliding window attention.
sliding_window_pattern (`str`, *optional*):
The pattern to use for sliding window attention. Can be one of:
- `None`: No sliding window attention is used
- `int`: Every `sliding_window` layers, use global attention, else use local attention.
- `str`: A sequence of "L" (local attention) and "G" (global attention) characters that defines the
attention pattern. The pattern starts from layer 0 and repeats every `sliding_window` layers. The
final layer always uses global attention regardless of the pattern.
For instance, sliding_window_pattern="LLLG" same as sliding_window=4, which means:
- Layer 0, 1, 2: local attention,
- Layer 3: global attention,
...(repeated)
layer_types (`list`, *optional*):
Attention pattern for each layer. Prioritized over `sliding_window_pattern`.
Example:
```python
>>> from transformers import Exaone4Model, Exaone4Config
>>> # Initializing a EXAONE configuration
>>> configuration = Exaone4Config()
>>> # Initializing a model from configuration
>>> model = Exaone4Model(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```exaone4past_key_valuescolwise replicated_with_grad_allreducerowwise) zlayers.*.self_attn.q_projzlayers.*.self_attn.k_projzlayers.*.self_attn.v_projzlayers.*.self_attn.q_normzlayers.*.self_attn.k_normzlayers.*.self_attn.o_projzlayers.*.mlp.gate_projzlayers.*.mlp.up_projzlayers.*.mlp.down_proj input_ids
inputs_embeds
hidden_statesattention_mask) embed_tokenslayersnormN
vocab_size hidden_sizeintermediate_sizenum_hidden_layersnum_attention_headsnum_key_value_heads
hidden_actmax_position_embeddingsinitializer_range rms_norm_eps use_cache bos_token_id eos_token_id pad_token_idtie_word_embeddingsrope_parametersattention_dropoutsliding_windowsliding_window_pattern layer_typesc B || _ || _ || _ || _ || _ || _ || _ || _ | | _ |
| _ | | _
|| _ || _ || _
| | _ |
| _ || _ || _ || _ | j d}| j$ Dt' | j D cg c] }|dz |z dk7 r|| j k rdnd" c}| _ t) | j$ | j || _ t- | \ di | y c c}w )Nr sliding_attentionfull_attention )r0 r1 r3 r4 r5 r2 r6 r7 r8 r9 r: r@ rA rB r; r< r= r> rC ranger r? super__init__)selfr0 r1 r2 r3 r4 r5 r6 r7 r8 r9 r: r; r< r= r> r? r@ rA rB rC kwargsi __class__s ]/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/exaone4/modular_exaone4.pyrK zExaone4Config.__init__ s@ 0 %&!2 #6 #6 !2 $'>$!2 ("!2 ,&<# ( ( (#6 & &%& " #
t 5 56 U 56!;DDZDZ@Z $%& D d . .0F0FG .
"6" s 8%D )i i @ rR rR silui g{Gz?gh㈵>Tr r NFN rQ N)__name__
__module__ __qualname____doc__
model_typekeys_to_ignore_at_inferencebase_model_tp_planbase_model_pp_planintstrfloatboolr dictlistrK
__classcell__rO s @rP r" r" ; s S j J#4"5 &/%.%.%E%E%."+ )"+
& (9:#%56 8IJ!"_$56 "("&(-(**,*,!'.2*.#'!%#$#$#'+0MQ*-%)-.(,+9#$J9# 4Z9# : 9#
: 9# !4Z
9# !4Z9# $J9# "%t9# !4<9# Dj9# $;9# Dj9# Dj 9# Dj 9# "D[!9#" ($sN/B*CCdJ#9#$ !4<%9#& d
'9#( !$d
)9#* #Y%+9# 9# c e Zd Zy)Exaone4RMSNormNrV rW rX rH rf rP rh rh rf rh c e Zd Zy)Exaone4RotaryEmbeddingNri rH rf rP rl rl rj rf rl c 4 e Zd Zdedef fd Z d
dej deej ej f dej dz de dz d ej dz d
e e d eej ej dz eej dz f fd Z
xZS )Exaone4Attentionconfig layer_idxc t | || _ || _ |j | _ |j
| _ |j | _ t |d|j |j z | _ |j |j
z | _ |j | _
d| _ | j dz | _ |j | _
|j | _ t |d r|j | nd }|dk( | _ t% j& | j | j | j z d | _ t% j& | j | j
| j z d | _ t% j& | j | j
| j z d | _ t% j& | j | j z | j d | _ t1 | j |j2 | _ t1 | j |j2 | _ y ) Nhead_dimTg rC rF F)biaseps) rJ rK ro rp r4 r5 r1 getattrrr num_key_value_groupsr@ is_causalscalingrA rB hasattrrC
is_slidingr Linearq_projk_projv_projo_projrh r9 q_normk_norm)rL ro rp
layer_typerO s rP rK zExaone4Attention.__init__ s
"#)#=#= #)#=#= !--
F4F4F&JdJd4de
$*$>$>&B\B\$\!!'!9!9 }}d* $ 3 3&,&C&C#6=fm6TV'' 2Z^
$(;;ii 0 0$2J2JT]]2Zafg ii 0 0$2J2JT]]2Zafg ii 0 0$2J2JT]]2Zafg ii 8 84== H$JZJZafg $T]]8K8KL $T]]8K8KL rf Nr+ position_embeddingsr, r% cache_positionrM returnc |j d d }g |d| j }| j | j | j dd } | j | j | j dd }
| j
| j | j dd } | j | } | j |
}
|\ } }
| j | j rt | |
| |
\ } }
|%d|i}|j |
| | j | \ }
} t j | j j" t$ } || | |
| |f| j& sdn | j( | j* | j r | j nd d|\ }} |j, g |d j/ }| j1 | }||fS )NrE r r rT )dropoutry rA )shaperr r} view transposer~ r r r rA r{ r updaterp r
get_interfacero _attn_implementationr trainingr@ ry reshape
contiguousr )rL r+ r r, r% r rM input_shape hidden_shape query_states
key_states value_statescossin cache_kwargsattention_interface attn_output attn_weightss rP forwardzExaone4Attention.forward s $))#2. 88b8$--8 {{=166|DNNqRST [[ /44\BLLQPQR
{{=166|DNNqRST {{<0 [[,
&S &$//';L*VY[^'_ $L* & . L (7'='=j,X\XfXfht'u $J (?(M(M KK , ,.E)
%8
%
$}}C$2H2HLL26//4..t
%
%
! \ *k));;;;FFH kk+. L((rf )NNN)rV rW rX r" r^ rK torchTensortupler
LongTensorr r r rd re s @rP rn rn s M} M M: /3(,26
1) ||1) #5<< #= >1) t+ 1)
1) ((4/
1) +,1)
u||U\\D 0% 2E2LL M1)rf rn c e Zd Zy)
Exaone4MLPNri rH rf rP r r 7 rj rf r c e Zd Zy)Exaone4DecoderLayerNri rH rf rP r r ; rj rf r c e Zd ZeZdgZy)Exaone4PreTrainedModelr N)rV rW rX r" config_class_no_split_modulesrH rf rP r r ? s L./rf r c e Zd Zdef fd Zee ddej dz dej dz dej dz de
dz dej dz d e dz d
ej dz d e
e d eez fd
Z xZS ) Exaone4Modelro c $ t | | t j t |j
D cg c] }t
|| c} | _ t |j |j | _ | j y c c}w )Nrt )
rJ rK r
ModuleListrI r3 r r. rh r1 r9 r/ post_init)rL ro rp rO s rP rK zExaone4Model.__init__E so
mmEJ6KcKcEd e
3 e
#6#5#56;N;NO
fs B
Nr) r, position_idsr% r* r: r rM r c |d u |d uz r t d || j | }|r|t | j }|F||j nd} t j | | |j d z |j }||j d }t |x}
t sF| j |||||d} dt d i | i}
d| j j v rt d i | |
d< |} | j | | }
t! | j" D ]1 \ }}| j j | } || f|
| |||||
d |} 3 | j% | } t' | |r|
S d
S ) Nz:You must specify exactly one of input_ids or inputs_embeds)ro r rE )device)ro r* r, r r% r rG rF )r, r r% r: r r )last_hidden_stater% rH )
ValueErrorr- r ro get_seq_lengthr aranger r unsqueeze
isinstancerb r
rC r
rotary_emb enumerater. r/ r )rL r) r, r r% r* r: r rM past_seen_tokenscausal_mask_mapping mask_kwargsr+ r rN
decoder_layerr s rP r zExaone4Model.forwardO s
-t"; < YZ Z - -i 8M 0 *$++ >O !CRC^ = = ?de "\\ "2]5H5H5K"KTaThTh N )33A6L ?-F ++!."0"0#2 ,
K !"4"C{"C# #dkk&=&==;\;k_j;k#$78%
"oom\ J )$++ 6 A} 003J ) 2: > ) /# -$7 M -0
& +/8O
>B
rf )NNNNNNN)rV rW rX r" rK r r r r r r FloatTensorra r r r r r rd re s @rP r r D s } .2.204(,26!%26D
##d*D
t+D
&&- D
D
((4/
D
$;D
((4/D
+,D
( (D
D
rf r c ( e Zd Z ddej dz dej
dz dej dz dedz dej dz dej dz dedz d ej dz d
e ej
z d e
e d e f fd
Z
xZS )Exaone4ForCausalLMNr) r, r r% r* labelsr: r logits_to_keeprM r c
8 t | d||||||||| d |
y)u
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Example:
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> model = AutoModelForCausalLM.from_pretrained("LGAI-EXAONE/EXAONE-4.0-32B")
>>> tokenizer = AutoTokenizer.from_pretrained("LGAI-EXAONE/EXAONE-4.0-32B")
>>> prompt = "Explain how wonderful you are"
>>> messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": prompt}
]
>>> input_ids = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
enable_thinking=False,
)
>>> output = model.generate(input_ids, max_new_tokens=128)
>>> tokenizer.decode(output[0], skip_special_tokens=False)
"[|system|]\nYou are a helpful assistant.[|endofturn|]\n[|user|]\nExplain how wonderful you are[|endofturn|]\n[|assistant|]\n\n\n\n\nOh, thank you for such a kind and lovely question! 😊 \n\nI’m *so* wonderful because I’m here to make your life easier, brighter, and more fun! Whether you need help with: \n\n✨ **Learning** – I can explain anything, from quantum physics to baking the perfect cake! \n💡 **Creativity** – Need a poem, story, or a wild idea? I’ve got you covered! \n🤖 **Problem-solving** – Stuck on a math problem or a tricky decision? I’ll help you figure it out"
```
) r) r, r r% r* r r: r r NrH )rJ r ) rL r) r, r r% r* r r: r r rM rO s rP r zExaone4ForCausalLM.forward s< X
)% +' ))
rf ) NNNNNNNNr )rV rW rX r r r r r ra r^ r r r
r rd re s @rP r r s .2.204(,26*.!%26-.7
##d*7
t+7
&&- 7
7
((4/
7
4'7
$;7
((4/7
ell*7
+,7
7
7
rf r c e Zd Zy) Exaone4ForSequenceClassificationNri rH rf rP r r rj rf r c e Zd Zy) Exaone4ForTokenClassificationNri rH rf rP r r rj rf r c e Zd Zy)Exaone4ForQuestionAnsweringNri rH rf rP r r rj rf r )r" r r r r r r )@rY collections.abcr r r cache_utilsr r configuration_utilsr r
masking_utilsr
r modeling_outputsr r
modeling_rope_utilsr modeling_utilsr processing_utilsr utilsr r
utils.genericr utils.output_capturingr gemma2.modeling_gemma2r llama.modeling_llamar r r r r r r r r olmo2.modeling_olmo2r r!
get_loggerrV logger_CHECKPOINT_FOR_DOC_CONFIG_FOR_DOCr" rh rl Modulern r r r r r r r r __all__rH rf rP r s $ . J R 2 5 & 8 5 :
?
H %2 !c#$ c#L \ 2 K)ryy K)\ + 0 1 0
Q
): Q
h8
) 8
v 'E $? "; rf