[Model] EXAONE 4.0 model support (#21060)

Signed-off-by: Deepfocused <rlawhdrhs27@gmail.com>
Signed-off-by: woongsik <rlawhdrhs27@gmail.com>

docs/models/supported_models.md

@@ -331,6 +331,7 @@ Specified using `--task generate`.
 | `Ernie4_5_ForCausalLM` | Ernie4.5 | `baidu/ERNIE-4.5-0.3B-PT`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Ernie4_5_MoeForCausalLM` | Ernie4.5MoE | `baidu/ERNIE-4.5-21B-A3B-PT`, `baidu/ERNIE-4.5-300B-A47B-PT`, etc. |✅︎| ✅︎ | ✅︎ |
 | `ExaoneForCausalLM` | EXAONE-3 | `LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |
+| `Exaone4ForCausalLM` | EXAONE-4 | `LGAI-EXAONE/EXAONE-4.0-32B`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Fairseq2LlamaForCausalLM` | Llama (fairseq2 format) | `mgleize/fairseq2-dummy-Llama-3.2-1B`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `FalconForCausalLM` | Falcon | `tiiuae/falcon-7b`, `tiiuae/falcon-40b`, `tiiuae/falcon-rw-7b`, etc. | | ✅︎ | ✅︎ |
 | `FalconMambaForCausalLM` | FalconMamba | `tiiuae/falcon-mamba-7b`, `tiiuae/falcon-mamba-7b-instruct`, etc. | | ✅︎ | ✅︎ |

tests/models/registry.py

@@ -169,6 +169,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "Ernie4_5_MoeForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-21B-A3B-PT",
                                         trust_remote_code=True),
     "ExaoneForCausalLM": _HfExamplesInfo("LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct"),  # noqa: E501
+    "Exaone4ForCausalLM": _HfExamplesInfo("LGAI-EXAONE/EXAONE-4.0-32B"),  # noqa: E501
     "Fairseq2LlamaForCausalLM": _HfExamplesInfo("mgleize/fairseq2-dummy-Llama-3.2-1B"),  # noqa: E501
     "FalconForCausalLM": _HfExamplesInfo("tiiuae/falcon-7b"),
     "FalconH1ForCausalLM":_HfExamplesInfo("tiiuae/Falcon-H1-0.5B-Base",

vllm/model_executor/models/registry.py

@@ -57,6 +57,7 @@ _TEXT_GENERATION_MODELS = {
     "Ernie4_5_ForCausalLM": ("ernie45", "Ernie4_5_ForCausalLM"),
     "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
     "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
+    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
     "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
     "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
     "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),

vllm/transformers_utils/config.py

@@ -31,9 +31,10 @@ from vllm.logger import init_logger
 # yapf: disable
 from vllm.transformers_utils.configs import (ChatGLMConfig, Cohere2Config,
                                              DbrxConfig, DeepseekVLV2Config,
-                                             EAGLEConfig, ExaoneConfig,
-                                             JAISConfig, KimiVLConfig,
-                                             MedusaConfig, MiniMaxText01Config,
+                                             EAGLEConfig, Exaone4Config,
+                                             ExaoneConfig, JAISConfig,
+                                             KimiVLConfig, MedusaConfig,
+                                             MiniMaxText01Config,
                                              MiniMaxVL01Config, MllamaConfig,
                                              MLPSpeculatorConfig, MPTConfig,
                                              NemotronConfig, NVLM_D_Config,
@@ -87,6 +88,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = {
     "medusa": MedusaConfig,
     "eagle": EAGLEConfig,
     "exaone": ExaoneConfig,
+    "exaone4": Exaone4Config,
     "minimax_text_01": MiniMaxText01Config,
     "minimax_vl_01": MiniMaxVL01Config,
     "nemotron": NemotronConfig,

vllm/transformers_utils/configs/__init__.py

@@ -7,6 +7,7 @@ from vllm.transformers_utils.configs.dbrx import DbrxConfig
 from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekVLV2Config
 from vllm.transformers_utils.configs.eagle import EAGLEConfig
 from vllm.transformers_utils.configs.exaone import ExaoneConfig
+from vllm.transformers_utils.configs.exaone4 import Exaone4Config
 # RWConfig is for the original tiiuae/falcon-40b(-instruct) and
 # tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the
 # `FalconConfig` class from the official HuggingFace transformers library.
@@ -40,6 +41,7 @@ __all__ = [
     "MedusaConfig",
     "EAGLEConfig",
     "ExaoneConfig",
+    "Exaone4Config",
     "MiniMaxText01Config",
     "MiniMaxVL01Config",
     "MllamaConfig",

vllm/transformers_utils/configs/exaone4.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# ruff: noqa: E501
+
+# Copied from
+# https://github.com/lgai-exaone/transformers/blob/add-exaone4/src/transformers/models/exaone4/configuration_exaone4.py
+# Copyright 2025 The LG CNS Gen AI Solution Delivery Team.
+# Copyright 2025 The LG AI Research and HuggingFace Inc. team. 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.
+from transformers.configuration_utils import (PretrainedConfig,
+                                              layer_type_validation)
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+def check_is_sliding(config, layer_idx):
+    """
+    Check if the current layer is a sliding window attention (local attention) layer.
+    """
+    if config.sliding_window is None:
+        return False
+    if config.layer_types is not None:
+        return config.layer_types[layer_idx] == "sliding_attention"
+    if isinstance(config.sliding_window_pattern, int):
+        return ((layer_idx + 1) % config.sliding_window_pattern) != 0
+    elif isinstance(config.sliding_window_pattern, str):
+        assert isinstance(config.sliding_window, int), (
+            f"Sliding window must be positive integer, but got {config.sliding_window}"
+        )
+        return (layer_idx != config.num_hidden_layers - 1
+                and config.sliding_window_pattern[layer_idx % len(
+                    config.sliding_window_pattern)] == "L")
+    else:
+        logger.warning_once(
+            "Sliding window is set, but none of `sliding_window_pattern` or `layer_types` is set. "
+            "Defaulting to use 'full_attention' for all layers.")
+    return False
+
+
+class Exaone4Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Exaone4Model`]. It is used to
+    instantiate a EXAONE 4.0 model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the EXAONE-4.0-Instruct [LGAI-EXAONE/EXAONE-4.0-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-Instruct)
+    NOTE: `EXAONE-4.0-Instruct` is a placeholder model ID. The exact model ID will be updated in the future.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model
+    outputs. Read the documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 102400):
+            Vocabulary size of the EXAONE 4.0 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Exaone4Model`].
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to `hidden_size * 4`):
+            Dimensionality of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 32768 for EXAONE 3.5).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if ``config.is_decoder=True``.
+        bos_token_id (`int`, *optional*, defaults to 0):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`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` (`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` (`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` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`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` (`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` (`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` (`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` (`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` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        sliding_window (`int`, *optional*):
+            The size of the sliding window for the sliding window attention.
+        sliding_window_pattern (`str`, *optional*):
+            The pattern to use for sliding window attention. Can be one of:
+                - `None`: No sliding window attention is used
+                - `int`: Every `sliding_window` layers, use global attention, else use local attention.
+                - `str`: A sequence of "L" (local attention) and "G" (global attention) characters that defines the
+                  attention pattern. The pattern starts from layer 0 and repeats every `sliding_window` layers. The
+                  final layer always uses global attention regardless of the pattern.
+            For instance, sliding_window_pattern="LLLG" same as sliding_window=4, which means:
+                - Layer 0, 1, 2: local attention,
+                - Layer 3: global attention,
+                ...(repeated)
+        layer_types (`list`, *optional*):
+            Attention pattern for each layer. Prioritized over `sliding_window_pattern`.
+
+    Example:
+
+    ```python
+    >>> from transformers import Exaone4Model, Exaone4Config
+
+    >>> # Initializing a EXAONE configuration
+    >>> configuration = Exaone4Config()
+
+    >>> # Initializing a model from configuration
+    >>> model = Exaone4Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "exaone4"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    # Default tensor parallel plan for base model `LlamaModel`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=102400,
+        hidden_size=4096,
+        intermediate_size=None,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        bos_token_id=0,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_dropout=0.0,
+        sliding_window=None,
+        sliding_window_pattern=None,
+        layer_types=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        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.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.attention_dropout = attention_dropout
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.sliding_window = sliding_window
+        self.sliding_window_pattern = sliding_window_pattern
+
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention"
+                if check_is_sliding(self, i) else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+        super().__init__(bos_token_id=bos_token_id,
+                         eos_token_id=eos_token_id,
+                         tie_word_embeddings=tie_word_embeddings,
+                         **kwargs)
+
+
+__all__ = ["Exaone4Config"]