Upload ExaoneForCausalLM

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,33 @@
+{
+  "_name_or_path": "/data/LLM_Checkpoint/EXAONE_8b_SFT_1ep/checkpoint-3033",
+  "activation_function": "silu",
+  "architectures": [
+    "ExaoneForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_exaone.ExaoneConfig",
+    "AutoModelForCausalLM": "modeling_exaone.ExaoneForCausalLM",
+    "AutoModelForSequenceClassification": "LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct--modeling_exaone.ExaoneForSequenceClassification"
+  },
+  "bos_token_id": 1,
+  "embed_dropout": 0.0,
+  "eos_token_id": 361,
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "layer_norm_epsilon": 1e-05,
+  "max_position_embeddings": 4096,
+  "model_type": "exaone",
+  "num_attention_heads": 32,
+  "num_key_value_heads": 8,
+  "num_layers": 32,
+  "pad_token_id": 0,
+  "rope_scaling": null,
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.44.0",
+  "use_cache": false,
+  "vocab_size": 102400
+}

configuration_exaone.py

@@ -0,0 +1,186 @@
+# coding=utf-8
+# Copyright 2021 The LG AI Research EXAONE Lab. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" EXAONE model configuration """
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+EXAONE_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+}
+
+
+class ExaoneConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.ExaoneModel`. It is used to
+    instantiate a EXAONE 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
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 102400):
+            Vocabulary size of the EXAONE model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.ExaoneModel`. Vocabulary size of the model.
+            Defines the different tokens that can be represented by the `inputs_ids` passed to the forward method of
+            :class:`~transformers.EXAONEModel`.
+        max_position_embeddings (:obj:`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., 512 or 1024 or 2048).
+        hidden_size (:obj:`int`, `optional`, defaults to 2048):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_layers (:obj:`int`, `optional`, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, `optional`, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (:obj:`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://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        intermediate_size (:obj:`int`, `optional`, defaults to `hidden_size * 4`):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        rope_theta (:obj:`float`, `optional`, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (:obj:`Dict`, `optional`):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (:obj:`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (:obj:`float`, `optional`):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (:obj:`int`, `optional`):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (:obj:`float`, `optional`):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (:obj:`float`, `optional`):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (:obj:`float`, `optional`):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (:obj:`List[float]`, `optional`):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (:obj:`List[float]`, `optional`):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (:obj:`float`, `optional`):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (:obj:`float`, `optional`):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        embed_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        layer_norm_epsilon (:obj:`float`, `optional`, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`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 (:obj:`int`, `optional`, defaults to 0):
+            Beginning of stream token id.
+        eos_token_id (:obj:`int`, `optional`, defaults to 2):
+            End of stream token id.
+        tie_word_embeddings (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to tie weight embeddings
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+
+        Example::
+
+            >>> from transformers import EXAONEModel, ExaoneConfig
+
+            >>> # Initializing a EXAONE configuration
+            >>> configuration = ExaoneConfig()
+
+            >>> # Initializing a model from configuration
+            >>> model = EXAONEModel(configuration)
+
+            >>> # Accessing the model configuration
+            >>> configuration = model.config
+    """
+    model_type = "exaone"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_hidden_layers": "num_layers"}
+
+    def __init__(
+        self,
+        vocab_size=102400,
+        max_position_embeddings=2048,
+        hidden_size=2048,
+        num_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        intermediate_size=None,
+        activation_function="silu",
+        rope_theta=10000.0,
+        rope_scaling=None,
+        embed_dropout=0.0,
+        attention_dropout=0.0,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=0,
+        eos_token_id=2,
+        tie_word_embeddings=True,
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.num_layers = num_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_hidden_layers = num_layers
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        if intermediate_size:
+            self.intermediate_size = intermediate_size
+        else:
+            self.intermediate_size = hidden_size * 4
+        self.activation_function = activation_function
+        self.embed_dropout = embed_dropout
+        self.attention_dropout = attention_dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs)

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 361,
+  "pad_token_id": 0,
+  "transformers_version": "4.44.0"
+}

model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b037922d7a42018c598dcf78492baa78fe15bc4c142f9b584486c009f37d1c45
+size 4966229976

model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ab952c7f41e380ca6f067fc91a85576633e9aa4564ae5bfdf174abec21a0130
+size 4915916080

model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e0fa8b3561f1b5a984d08a12702076d1cc4325caf1891224e6da7aafc07a95b
+size 4915924392

model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:938e45b6ac5b8fa9b80f130bd87c821c272913fea4f1619cd21a3361b30d4c51
+size 838860928

model.safetensors.index.json

@@ -0,0 +1,298 @@
+{
+  "metadata": {
+    "total_size": 15636897792
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00004-of-00004.safetensors",
+    "transformer.h.0.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.0.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.1.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.10.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.10.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.11.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.12.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.13.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.14.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.15.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.16.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.17.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.18.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.19.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.2.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.2.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.20.attn.attention.k_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.attn.attention.out_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.attn.attention.q_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.attn.attention.v_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.ln_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.ln_2.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.mlp.c_fc_0.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.20.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.21.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.22.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.23.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.24.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.25.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.26.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.27.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.28.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.29.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.3.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.3.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.30.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.30.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.attn.attention.k_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.attn.attention.out_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.attn.attention.q_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.attn.attention.v_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.ln_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.ln_2.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.mlp.c_fc_0.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.mlp.c_fc_1.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.31.mlp.c_proj.weight": "model-00003-of-00004.safetensors",
+    "transformer.h.4.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.4.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.5.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.6.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.7.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.mlp.c_fc_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.8.mlp.c_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.attn.attention.k_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.attn.attention.out_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.attn.attention.q_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.attn.attention.v_proj.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.ln_1.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.ln_2.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.mlp.c_fc_0.weight": "model-00001-of-00004.safetensors",
+    "transformer.h.9.mlp.c_fc_1.weight": "model-00002-of-00004.safetensors",
+    "transformer.h.9.mlp.c_proj.weight": "model-00002-of-00004.safetensors",
+    "transformer.ln_f.weight": "model-00003-of-00004.safetensors",
+    "transformer.wte.weight": "model-00001-of-00004.safetensors"
+  }
+}

modeling_exaone.py

@@ -0,0 +1,1747 @@
+# coding=utf-8
+# Copyright 2021 The LG AI Research EXAONE Lab
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" LG AI Research EXAONE Lab"""
+import sys
+import os
+from typing import List, Optional, Tuple, Union
+from packaging import version
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, StaticCache
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    CausalLMOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    QuestionAnsweringModelOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    logging,
+)
+from .configuration_exaone import ExaoneConfig
+from torch.nn.utils import skip_init
+import math
+import numpy as np
+from typing import List, Optional, Tuple, Union
+
+
+if is_flash_attn_2_available():
+    try:
+        import inspect
+        from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
+        from flash_attn import flash_attn_func, flash_attn_varlen_func
+
+        _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
+
+        import flash_attn
+        if version.parse(flash_attn.__version__) > version.parse('2.4.2'):
+            from flash_attn.ops.triton.layer_norm import rms_norm_fn
+        else:
+            from flash_attn.ops.triton.layernorm import rms_norm_fn
+    except:
+        pass
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "exaone"
+_CONFIG_FOR_DOC = "ExaoneConfig"
+
+EXAONE_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "exaone",
+]
+
+
+@torch.jit.script
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def rotate_half(x):
+    """ Rotates half the hidden dims of the input. """
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# copied from llama
+def _prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    min_dtype: float,
+    cache_position: torch.Tensor,
+    batch_size: int,
+):
+    """
+    Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+    `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+    Args:
+        attention_mask (`torch.Tensor`):
+            A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+        sequence_length (`int`):
+            The sequence length being processed.
+        target_length (`int`):
+            The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+        dtype (`torch.dtype`):
+            The dtype to use for the 4D attention mask.
+        device (`torch.device`):
+            The device to plcae the 4D attention mask on.
+        min_dtype (`float`):
+            The minimum value representable with the dtype `dtype`.
+        cache_position (`torch.Tensor`):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        batch_size (`torch.Tensor`):
+            Batch size.
+    """
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask
+    else:
+        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask
+
+
+class ExaoneRMSNorm(torch.nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = torch.nn.Parameter(torch.ones(hidden_size))
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class ExaoneTritonRMSNorm(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 0,
+        eps: float = 1e-5,
+    ):
+        super().__init__()
+        self.eps = eps
+        self.drop = None
+        self.weight = torch.nn.Parameter(torch.empty(hidden_size))
+        self.register_parameter("bias", None)
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        torch.nn.init.ones_(self.weight)
+
+    def forward(self, x, residual=None, prenorm=False, residual_in_fp32=False):
+        return rms_norm_fn(
+            x,
+            self.weight,
+            self.bias,
+            residual=residual,
+            eps=self.eps,
+            dropout_p=self.drop.p if self.drop is not None and self.training else 0.0,
+            prenorm=prenorm,
+            residual_in_fp32=residual_in_fp32,
+        )
+
+
+ALL_LAYERNORM_LAYERS.append(ExaoneRMSNorm)
+ALL_LAYERNORM_LAYERS.append(ExaoneTritonRMSNorm)
+
+
+class ExaoneRotaryEmbedding(nn.Module):
+    """
+    Common description for the functions named `_compute_XXX_rope_parameters()`
+        - Copied from `transformers.modeling_rope_utils` in v4.43, with some modifications.
+
+        Computes the inverse frequencies with linear scaling. 
+        The EXAONE model supports 'default', 'linear', 'dynamic', and 'yarn'.
+        
+        Args:
+            config (:obj:`~transformers.PretrainedConfig`):
+                The model configuration.
+            device (:obj:`torch.device`):
+                The device to use for initialization of the inverse frequencies.
+            seq_len (:obj:`int`, `optional`):
+                The current sequence length. Unused for this type of RoPE.
+        Returns:
+            Tuple of (:obj:`torch.Tensor`, :obj:`float`), containing the inverse frequencies for the RoPE embeddings and the
+            post-processing scaling factor applied to the computed cos/sin (unused in some types of RoPE).
+    """
+
+    def _compute_default_rope_parameters(
+        self,
+        config: Optional[PretrainedConfig],
+        device: Optional["torch.device"] = None,
+        seq_len: Optional[int] = None,
+    ) -> Tuple["torch.Tensor", float]:
+        base = config.rope_theta
+        partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
+        dim = int((config.hidden_size // config.num_attention_heads) * partial_rotary_factor)
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).float().to(device) / dim))
+        return inv_freq, attention_factor
+
+    def _compute_linear_scaling_rope_parameters(
+        self,
+        config: Optional[PretrainedConfig],
+        device: Optional["torch.device"] = None,
+        seq_len: Optional[int] = None,
+    ) -> Tuple["torch.Tensor", float]:
+        factor = config.rope_scaling["factor"]
+        if factor < 1.0:
+            logger.warning_once(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+
+        inv_freq, attention_factor = self._compute_default_rope_parameters(config, device, seq_len)
+        inv_freq /= factor
+        return inv_freq, attention_factor
+
+    def _compute_dynamic_ntk_parameters(
+        self,
+        config: Optional[PretrainedConfig],
+        device: Optional["torch.device"] = None,
+        seq_len: Optional[int] = None,
+    ) -> Tuple["torch.Tensor", float]:
+        base = config.rope_theta
+        partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
+        dim = int((config.hidden_size // config.num_attention_heads) * partial_rotary_factor)
+        max_position_embeddings = config.max_position_embeddings
+        factor = config.rope_scaling["factor"]
+        if factor < 1.0:
+            logger.warning_once(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+
+        attention_factor = 1.0  # Unused in this type of RoPE
+        seq_len = seq_len if seq_len is not None else max_position_embeddings
+
+        base = base * ((factor * seq_len / max_position_embeddings) - (factor - 1)) ** (dim / (dim - 2))
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).float().to(device) / dim))
+        return inv_freq, attention_factor
+    
+    def _compute_yarn_parameters(
+        self,
+        config: PretrainedConfig, 
+        device: "torch.device", 
+        seq_len: Optional[int] = None, 
+    ) -> Tuple["torch.Tensor", float]:
+        base = config.rope_theta
+        partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
+        dim = int((config.hidden_size // config.num_attention_heads) * partial_rotary_factor)
+        max_position_embeddings = config.max_position_embeddings
+        factor = config.rope_scaling["factor"]
+        if factor < 1.0:
+            logger.warning_once(f"`rope_scaling`'s factor field must be a float >= 1, got {factor}")
+
+        # Sets the attention factor as suggested in the paper
+        attention_factor = config.rope_scaling.get("attention_factor")
+        if attention_factor is None:
+            attention_factor = 0.1 * math.log(factor) + 1.0
+        if attention_factor < 0:
+            logger.warning_once(
+                f"`rope_scaling`'s attention_factor field must be a float greater than 0, got {attention_factor}"
+            )
+
+        # Optional config options
+        # beta_fast/beta_slow: as suggested in the paper, default to 32/1 (correspondingly)
+        beta_fast = config.rope_scaling.get("beta_fast") or 32
+        beta_slow = config.rope_scaling.get("beta_slow") or 1
+        if not isinstance(beta_fast, float):
+            logger.warning_once(f"`rope_scaling`'s beta_fast field must be a float, got {beta_fast}")
+        if not isinstance(beta_slow, float):
+            logger.warning_once(f"`rope_scaling`'s beta_slow field must be a float, got {beta_fast}")
+        if beta_fast < beta_slow:
+            logger.warning_once(
+                f"`rope_scaling`'s beta_fast field must be greater than beta_slow, got beta_fast={beta_fast} "
+                f"(defaults to 32 if None) and beta_slow={beta_slow} (defaults to 1 if None)"
+            )
+
+        # Compute the inverse frequencies
+        def find_correction_dim(num_rotations, dim, base, max_position_embeddings):
+            """Inverse dimension formula to find the dimension based on the number of rotations"""
+            return (dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))) / (2 * math.log(base))
+
+        def find_correction_range(low_rot, high_rot, dim, base, max_position_embeddings):
+            """Find dimension range bounds based on rotations"""
+            low = math.floor(find_correction_dim(low_rot, dim, base, max_position_embeddings))
+            high = math.ceil(find_correction_dim(high_rot, dim, base, max_position_embeddings))
+            return max(low, 0), min(high, dim - 1)
+
+        def linear_ramp_mask(min, max, dim):
+            if min == max:
+                max += 0.001  # Prevent singularity
+
+            linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
+            ramp_func = torch.clamp(linear_func, 0, 1)
+            return ramp_func
+
+        pos_freqs = base ** (torch.arange(0, dim, 2).float().to(device) / dim)
+        inv_freq_extrapolation = 1.0 / pos_freqs
+        inv_freq_interpolation = 1.0 / (factor * pos_freqs)
+
+        low, high = find_correction_range(beta_fast, beta_slow, dim, base, max_position_embeddings)
+
+        # Get n-dimensional rotational scaling corrected for extrapolation
+        inv_freq_mask = 1 - linear_ramp_mask(low, high, dim // 2).float().to(device)
+        inv_freq = inv_freq_interpolation * (1 - inv_freq_mask) + inv_freq_extrapolation * inv_freq_mask
+
+        return inv_freq, attention_factor
+
+    def __init__(self, config: ExaoneConfig, device=None):
+        ROPE_INIT_FUNCTIONS = {
+            "default": self._compute_default_rope_parameters,
+            "linear": self._compute_linear_scaling_rope_parameters,
+            "dynamic": self._compute_dynamic_ntk_parameters,
+            "yarn": self._compute_yarn_parameters,
+        }
+
+        super().__init__()
+        if config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        if self.rope_type not in ROPE_INIT_FUNCTIONS:
+            raise KeyError(f"The EXAONE model does not support RoPE type: {self.rope_type}")
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _update_freq(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len:  # expand to seq_len
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)
+            self.max_seq_len = seq_len
+        
+        if seq_len < self.original_max_seq_len and self.max_seq_len > self.original_max_seq_len:    # reset to original
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._update_freq(position_ids, device=x.device)
+
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos, sin = emb.cos(), emb.sin()
+        
+        cos, sin = cos * self.attention_scaling, sin * self.attention_scaling
+        return cos.to(x.dtype), sin.to(x.dtype)
+
+
+class ExaoneSelfAttention(nn.Module):
+    def __init__(self, config: ExaoneConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.attention_dropout_rate = config.attention_dropout
+
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads})."
+            )
+
+        self.rotary = ExaoneRotaryEmbedding(config)
+        
+        self.k_proj = nn.Linear(self.embed_dim, self.num_key_value_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.embed_dim, self.num_key_value_heads * self.head_dim, bias=False)
+        self.q_proj = nn.Linear(self.embed_dim, self.num_heads * self.head_dim, bias=False)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+
+        bsz, q_len, _ = hidden_states.size()
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if position_embeddings is None:
+            cos, sin = self.rotary(value_states, position_ids=position_ids)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attention_mask is not None:
+            causal_mask = attention_mask[:, :, :, :key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout_rate, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"Attention outputs should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.embed_dim).contiguous()
+
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class ExaoneFlashAttention(ExaoneSelfAttention):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if isinstance(past_key_value, StaticCache):
+            raise ValueError(
+                "`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
+                "make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
+            )
+
+        output_attentions = False
+
+        bsz, q_len, h_size = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if position_embeddings is None:
+            cos, sin = self.rotary(value_states, position_ids=position_ids)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            # Only update cache as shape of [bsz, n_head, q_len, head_dim]
+            # TODO: need to be fixed when transformers' KV cache layout is changed
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        dropout_rate = self.attention_dropout_rate if self.training else 0.0
+
+        attn_output = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate, is_causal=True
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.embed_dim).contiguous()
+        attn_output = self.out_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+    
+    @staticmethod
+    def _flash_attention_forward(
+        query_states: torch.Tensor,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        query_length: int,
+        is_causal: bool,
+        dropout: float = 0.0,
+        softmax_scale: Optional[float] = None,
+        sliding_window: Optional[int] = None,
+        use_top_left_mask: bool = False,
+        softcap: Optional[float] = None,
+        deterministic: bool = os.environ.get("FLASH_ATTENTION_DETERMINISTIC", "0") == "1",
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+            use_top_left_mask (`bool`, defaults to `False`):
+                flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference.
+            softcap (`float`, *optional*):
+                Softcap for the attention logits, used e.g. in gemma2.
+            deterministic (`bool`, *optional*):
+                Determines if the deterministic option introduced in flash_attn>=2.4.1 is enabled.
+        """
+        if not use_top_left_mask:
+            causal = is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__.
+            causal = is_causal and query_length != 1
+
+        # Assuming 4D tensors, key_states.shape[1] is the key/value sequence length (source length).
+        use_sliding_windows = (
+            _flash_supports_window_size and sliding_window is not None and key_states.shape[1] > sliding_window
+        )
+        flash_kwargs = {"window_size": (sliding_window, sliding_window)} if use_sliding_windows else {}
+
+        if softcap is not None:
+            flash_kwargs["softcap"] = softcap
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = ExaoneFlashAttention._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+                **flash_kwargs,
+            )
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal, **flash_kwargs
+            )
+
+        return attn_output
+    
+    @staticmethod
+    def _upad_input(
+        query_layer: torch.Tensor,
+        key_layer: torch.Tensor,
+        value_layer: torch.Tensor,
+        attention_mask: torch.Tensor,
+        query_length: int,
+    ):
+        """
+        Unpads query, key, and values tensors, using a single dimension for all tokens even though they belong to different batches.
+
+        This function is used instead of `flash_attn.bert_padding.unpad_input` in order to avoid the recomputation of the same intermediary
+        tensors for query, key, value tensors.
+
+        Arguments:
+            query_layer (`torch.Tensor`):
+                Query state with padding. Shape: (batch_size, query_length, num_heads, head_dim).
+            key_layer (`torch.Tensor`):
+                Key state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
+            value_layer (`torch.Tensor`):
+                Value state with padding. Shape: (batch_size, kv_seq_len, num_key_value_heads, head_dim).
+            attention_mask (`torch.Tensor`):
+                Boolean or int tensor of shape (batch_size, sequence_length), 1 means valid and 0 means not valid.
+            query_length (`int`):
+                Target length.
+
+        Return:
+            query_layer (`torch.Tensor):
+                Query state without padding. Shape: (total_target_length, num_heads, head_dim).
+            key_layer (`torch.Tensor`):
+                Key state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
+            value_layer (`torch.Tensor`):
+                Value state with padding. Shape: (total_source_length, num_key_value_heads, head_dim).
+            indices_q (`torch.Tensor`):
+                The indices of non-masked tokens from the flattened input target sequence.
+            (cu_seqlens_q, cu_seqlens_k) (`Tuple[int]`):
+                The cumulative sequence lengths for the target (query) and source (key, value), used to index into ragged (unpadded) tensors. `cu_seqlens` shape is (batch_size + 1,).
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k) (`Tuple[int]`):
+                Maximum sequence length in batch (`max_seqlen_in_batch_q` for the target sequence i.e. query, `max_seqlen_in_batch_k` for the source sequence i.e. key/value).
+        """
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = ExaoneFlashAttention._get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k)
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(query_layer.reshape(batch_size * kv_seq_len, -1, head_dim), indices_k)
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+    
+    @staticmethod
+    def _get_unpad_data(attention_mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, int]:
+        """
+        Retrieves indexing data required to repad unpadded (ragged) tensors.
+
+        Arguments:
+            attention_mask (`torch.Tensor`):
+                Boolean or int tensor of shape (batch_size, sequence_length), 1 means valid and 0 means not valid.
+
+        Return:
+            indices (`torch.Tensor):
+                The indices of non-masked tokens from the flattened input sequence.
+            cu_seqlens (`torch.Tensor`):
+                The cumulative sequence lengths, used to index into ragged (unpadded) tensors. `cu_seqlens` shape is (batch_size + 1,).
+            max_seqlen_in_batch (`int`):
+                Maximum sequence length in batch.
+        """
+        seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+        indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+        max_seqlen_in_batch = seqlens_in_batch.max().item()
+        cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+        return (
+            indices,
+            cu_seqlens,
+            max_seqlen_in_batch,
+        )
+
+
+class ExaoneSdpaAttention(ExaoneSelfAttention):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+    
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+
+        if output_attentions:
+            logger.warning_once(
+                "ExaoneModel is using ExaoneSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if position_embeddings is None:
+            cos, sin = self.rotary(value_states, position_ids=position_ids)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, :key_states.shape[-2]]        
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and causal_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout_rate if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.embed_dim).contiguous()
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+class ExaoneAttention(nn.Module):
+    def __init__(self, config, layer_id=0):
+        super().__init__()
+        self.layer_id = layer_id
+        if 'flash' in config._attn_implementation:
+            self.attention = ExaoneFlashAttention(config, self.layer_id)
+        elif 'sdpa' in config._attn_implementation:
+            self.attention = ExaoneSdpaAttention(config, self.layer_id)
+        else:
+            self.attention = ExaoneSelfAttention(config, self.layer_id)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+
+        return self.attention(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+
+
+class ExaoneGatedMLP(nn.Module):
+    def __init__(self, intermediate_size, config):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+        self.c_fc_0 = nn.Linear(embed_dim, intermediate_size, bias=False)
+        self.c_fc_1 = nn.Linear(embed_dim, intermediate_size, bias=False)
+        self.c_proj = nn.Linear(intermediate_size, embed_dim, bias=False)
+        self.act = ACT2FN[config.activation_function]
+
+    def forward(self, hidden_states):
+        output_proj = self.c_proj(self.act(self.c_fc_0(hidden_states)) * self.c_fc_1(hidden_states))
+        return output_proj
+
+
+class ExaoneBlock(nn.Module):
+    def __init__(self, config, layer_id):
+        super().__init__()
+        self.config = config
+        hidden_size = config.hidden_size
+        inner_dim = config.intermediate_size if config.intermediate_size is not None else 4 * hidden_size
+        self.ln_1 = ExaoneRMSNorm(hidden_size = hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = ExaoneAttention(config, layer_id)
+        self.ln_2 = ExaoneRMSNorm(hidden_size = hidden_size, eps=config.layer_norm_epsilon)
+        self.mlp = ExaoneGatedMLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+
+        hidden_states, self_attn_weights, present_key_value = self.attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        # residual connection
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class ExaonePreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ExaoneConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["ExaoneBlock"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear,)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, ExaoneRMSNorm):
+            module.weight.data.fill_(1.0)
+
+
+EXAONE_START_DOCSTRING = r"""
+
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config (:class:`~transformers.ExaoneConfig`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+EXAONE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
+            ``past_key_values.get_seq_length()`` (``sequence_length`` of input past key value states). Indices of input
+            sequence tokens in the vocabulary.
+
+            If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be
+            passed as ``input_ids``.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
+            config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        past_key_values (:obj:`Cache`, `optional`):
+            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
+            :obj:`past_key_values` output below). Can be used to speed up sequential decoding. This typically consists 
+            in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or 
+            `config.use_cache=True`.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+
+            If :obj:`past_key_values` is used, optionally only the last :obj:`inputs_embeds` have to be input (see
+            :obj:`past_key_values`).
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+        cache_position (:obj:`torch.LongTensor` of shape :obj:`(sequence_length)`, `optional`):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare EXAONE Model transformer outputting raw hidden-states without any specific head on top.",
+    EXAONE_START_DOCSTRING,
+)
+class ExaoneModel(ExaonePreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.wte = nn.Embedding(config.vocab_size, self.embed_dim, self.config.pad_token_id)
+        self.drop = nn.Dropout(float(config.embed_dropout))
+        self.h = nn.ModuleList([ExaoneBlock(config, layer_id=i) for i in range(config.num_layers)])
+        self.ln_f = ExaoneRMSNorm(hidden_size=self.embed_dim, eps=config.layer_norm_epsilon)
+        self.rotary = ExaoneRotaryEmbedding(config)
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    @add_start_docstrings_to_model_forward(EXAONE_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        return_legacy_cache = False
+        if (
+            use_cache and not isinstance(past_key_values, Cache) and not self.training
+        ):  # kept for BC (non `Cache` `past_key_values` inputs)
+            return_legacy_cache = True
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            logger.warning_once(
+                "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "
+                "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"
+            )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+        hidden_states = self.drop(hidden_states)
+
+        position_embeddings = self.rotary(hidden_states, position_ids)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for block in self.h:
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                outputs = self._gradient_checkpointing_func(
+                    block.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache:
+                next_decoder_cache = outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (outputs[1],)
+
+        hidden_states = self.ln_f(hidden_states)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if return_legacy_cache else next_decoder_cache
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+    
+    # copied from llama
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
+        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
+        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_length()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            min_dtype=min_dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+
+@add_start_docstrings(
+    """
+    The EXAONE Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """,
+    EXAONE_START_DOCSTRING,
+)
+class ExaoneForCausalLM(ExaonePreTrainedModel):
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = ExaoneModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.config = config
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    @add_start_docstrings_to_model_forward(EXAONE_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+                `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+                are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+        >>> model = AutoModelForCausalLM.from_pretrained("LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct",
+                                                         trust_remote_code=True)
+        >>> tokenizer = AutoTokenizer.from_pretrained("LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct")
+
+        >>> 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"
+        )
+
+        >>> output = model.generate(input_ids, max_new_tokens=128)
+        >>> tokenizer.decode(output[0], skip_special_tokens=True)
+        "[|system|]You are a helpful assistant.\n[|user|]Explain how wonderful you are\n[|assistant|]Thank you for your kind words! I'm here to assist you with information, answer questions, and help you in any way I can. My goal is to provide accurate, helpful, and timely responses. Whether you need help with a specific task, want to learn something new, or just need someone to talk to, I'm here for you. How can I assist you today?"
+        ```
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        transformer_outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+        hidden_states = transformer_outputs[0]
+        lm_logits = self.lm_head(hidden_states)
+        lm_logits = lm_logits.float()
+        loss = None
+        if labels is not None:
+            lm_logits = lm_logits.to(torch.float32)
+
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+            lm_logits = lm_logits.to(hidden_states.dtype)
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        **kwargs,
+    ):
+        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+        # Exception 1: when passing input_embeds, input_ids may be missing entries
+        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+        if past_key_values is not None:
+            if inputs_embeds is not None:  # Exception 1
+                input_ids = input_ids[:, -cache_position.shape[0] :]
+            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
+                input_ids = input_ids[:, cache_position]
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                position_ids = position_ids.clone(memory_format=torch.contiguous_format)
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and cache_position[0] == 0:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+            if inputs_embeds is not None:
+                batch_size, sequence_length = inputs_embeds.shape
+                device = inputs_embeds.device
+            else:
+                batch_size, sequence_length = input_ids.shape
+                device = input_ids.device
+
+            dtype = self.lm_head.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "use_cache": use_cache,
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The EXAONE Model transformer with a sequence classification head on top (linear layer).
+
+    :class:`~transformers.ExaoneForSequenceClassification` uses the last token in order to do the classification, as
+    other causal models (e.g. GPT-1) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    :obj:`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each
+    row. If no :obj:`pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot
+    guess the padding tokens when :obj:`inputs_embeds` are passed instead of :obj:`input_ids`, it does the same (take
+    the last value in each row of the batch).
+    """,
+    EXAONE_START_DOCSTRING,
+)
+class ExaoneForSequenceClassification(ExaonePreTrainedModel):
+    _keys_to_ignore_on_load_missing = ["lm_head.weight"]
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = ExaoneModel(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(EXAONE_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size, sequence_length = input_ids.shape[:2]
+        else:
+            batch_size, sequence_length = inputs_embeds.shape[:2]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+                logger.warning(
+                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                    "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+                )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The EXAONE Model transformer with a span classification head on top for extractive question-answering tasks like
+    SQuAD (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    EXAONE_START_DOCSTRING,
+)
+class ExaoneForQuestionAnswering(ExaonePreTrainedModel):
+    _keys_to_ignore_on_load_missing = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.transformer = ExaoneModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], QuestionAnsweringModelOutput]:
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1).to(start_logits.device)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1).to(end_logits.device)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions = start_positions.clamp(0, ignored_index)
+            end_positions = end_positions.clamp(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )