[None][feat] EXAONE-4.5 Support

docs/source/models/supported-models.md

@@ -82,6 +82,7 @@ Note: Support for other models may vary. Features marked "N/A" are not applicabl
 
 | Model Architecture/Feature           | Overlap Scheduler | CUDA Graph | Chunked Prefill | Torch Sampler | TLLM C++ Sampler | KV Cache Reuse | Logits Post Processor | EPD Disaggregated Serving | Modality  |
 | ------------------------------------ | ----------------- | ---------- | --------------- | ------------- | ---------------- | -------------- | --------------------- | ------------------------- | --------- |
+| `Exaone4_5_ForConditionalGeneration` | Yes               | Yes        | Yes             | Yes           | Yes              | Yes            | Yes                   | No                        | L + I + V |
 | `Gemma3ForConditionalGeneration`     | Yes               | Yes        | N/A             | Yes           | Yes              | N/A            | Yes                   | No                        | L + I     |
 | `Gemma4ForConditionalGeneration`     | Untested          | Yes        | No              | Yes           | Untested         | No             | Untested              | No                        | L + I + A [^9] |
 | `HCXVisionForCausalLM`               | Yes               | Yes        | No              | Yes           | Yes              | Yes            | Yes                   | No                        | L + I     |

examples/models/core/exaone/README.md

@@ -14,9 +14,11 @@ This document shows how to build and run [EXAONE](https://huggingface.co/LGAI-EX
     - [EXAONE-3.0](#exaone-30)
     - [EXAONE-Deep](#exaone-deep)
     - [EXAONE-4.0](#exaone-40)
+    - [EXAONE-4.5](#exaone-45)
     - [K-EXAONE](#k-exaone)
   - [PyTorch flow](#pytorch-flow)
     - [Running EXAONE-4.0](#running-exaone-40)
+    - [Running EXAONE-4.5](#running-exaone-45)
     - [Running K-EXAONE](#running-k-exaone)
       - [MoE Backend Options](#moe-backend-options)
     - [PyTorch flow Quantization](#pytorch-flow-quantization)
@@ -45,6 +47,7 @@ This document shows how to build and run [EXAONE](https://huggingface.co/LGAI-EX
   * FP16
   * BF16
   * Tensor Parallel (TP)
+  * Multimodal (EXAONE-4.5 only)
   * Expert Parallel (EP) (K-EXAONE only)
   * Attention Data Parallel (ADP) (K-EXAONE only)
   * Disaggregated Serving
@@ -59,7 +62,7 @@ This document shows how to build and run [EXAONE](https://huggingface.co/LGAI-EX
 
 **Note:**
 - **EXAONE-3.0** & **EXAONE-Deep** are supported using the [TRT Flow](#trt-flow).
-- **EXAONE-4.0** & **K-EXAONE** are supported using the [PyTorch flow](#pytorch-flow).
+- **EXAONE-4.0**, **EXAONE-4.5**, & **K-EXAONE** are supported using the [PyTorch flow](#pytorch-flow).
 
 Please refer to the corresponding sections below for usage instructions and examples for each model.
 
@@ -90,6 +93,17 @@ export HF_MODEL_DIR=hf_models/exaone4
 git clone https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B $HF_MODEL_DIR
 ```
 
+### EXAONE-4.5
+
+EXAONE-4.5 is a multimodal model. It is supported only via the [PyTorch flow](#pytorch-flow).
+
+Download the HuggingFace checkpoint for your EXAONE-4.5 variant from the [LGAI-EXAONE](https://huggingface.co/LGAI-EXAONE) organization. The example below uses `EXAONE-4.5-33B`; replace it with the variant you want to run.
+
+```bash
+export HF_MODEL_DIR=hf_models/exaone4_5
+huggingface-cli download LGAI-EXAONE/EXAONE-4.5-33B --local-dir $HF_MODEL_DIR
+```
+
 ### K-EXAONE
 
 K-EXAONE is a Mixture of Experts (MoE) model based on the EXAONE architecture. It features a hybrid architecture with both dense and MoE layers, sliding window attention, and supports FP8 and NVFP4 quantization for efficient inference.
@@ -98,7 +112,7 @@ Download the HuggingFace checkpoints of the K-EXAONE model:
 
 ```bash
 export HF_MODEL_DIR=hf_models/kexaone
-git clone https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B $HF_MODEL_DIR
+huggingface-cli download LGAI-EXAONE/K-EXAONE-236B-A23B --local-dir $HF_MODEL_DIR
 ```
 
 ## PyTorch flow
@@ -117,6 +131,21 @@ The output will be like:
 [2] Prompt: 'The future of AI is', Generated text: ' not just about technology but also about how we choose to use it. We must ensure that AI is developed and deployed in a way that benefits all of humanity, not just a select few. This means prioritizing ethical considerations, transparency, and accountability in AI development. It also means involving diverse stakeholders in the conversation about AI'
 ```
 
+### Running EXAONE-4.5
+
+To quickly run EXAONE-4.5 models, you can use [examples/llm-api/quickstart_multimodal.py](../../../llm-api/quickstart_multimodal.py):
+
+```bash
+python ../../../llm-api/quickstart_multimodal.py --model_dir $HF_MODEL_DIR
+```
+
+The output will be like:
+```bash
+[0] Prompt: 'Describe the natural environment in the image.', Generated text: 'Okay, the user asked me to describe the natural environment in the image. But wait, there's no image provided here. Hmm, that's a problem. How can I describe something I can't see?\n\nFirst, I need to check if there's any image attached. The user mentioned "the image," but in the current context, there's no image data. Maybe they forgot'
+[1] Prompt: 'Describe the object and the weather condition in the image.', Generated text: 'Okay, the user asked me to describe the object and the weather condition in the image. But wait, there's no image provided here. Hmm, that's a problem. How can I describe something I can't see?\n\nFirst, I need to check if there's any image attached. The user mentioned "the image," but in the current context, there's no image data. Maybe'
+[2] Prompt: 'Describe the traffic condition on the road in the image.', Generated text: 'Okay, the user is asking me to describe the traffic condition on the road in the image. But wait, there's a problem here—I don't actually see any image. The user mentioned "the image," but in this text-based interface, there's no visual content provided. \n\nHmm, I need to handle this carefully. The user might have forgotten to attach the image or assumed I could see it'
+```
+
 ### Running K-EXAONE
 
 K-EXAONE is a Mixture of Experts model that benefits from multiple parallelism strategies. You can run it with tensor parallelism (TP), expert parallelism (EP), and attention data parallelism (ADP):

tensorrt_llm/_torch/model_config.py

@@ -608,8 +608,14 @@ def cached_file(path_or_repo_id, file_name):
                 return None
 
         # Some checkpoints lack torch_dtype, populate with dtype
-        pretrained_config.torch_dtype = getattr(pretrained_config, 'dtype',
-                                                None)
+        dtype = getattr(pretrained_config, 'dtype', None)
+        # For composite VLM configs the dtype lives inside ``text_config``
+        # because the top-level config has no ``dtype`` field.
+        if dtype is None:
+            text_config = getattr(pretrained_config, 'text_config', None)
+            if text_config is not None:
+                dtype = getattr(text_config, 'dtype', None)
+        pretrained_config.torch_dtype = dtype
 
         # Prior to transformers 5, composite configs (e.g. Qwen2_5_VLConfig) delegated attribute
         # lookups to their text sub-config, so accesses like `config.vocab_size` /

tensorrt_llm/_torch/models/__init__.py

@@ -11,6 +11,7 @@
 from .modeling_cohere2 import Cohere2ForCausalLM
 from .modeling_deepseekv3 import DeepseekV3ForCausalLM
 from .modeling_exaone4 import Exaone4ForCausalLM
+from .modeling_exaone4_5 import Exaone4_5_ForConditionalGeneration
 from .modeling_exaone_moe import ExaoneMoeForCausalLM
 from .modeling_gemma3 import Gemma3ForCausalLM
 from .modeling_gemma3vl import Gemma3VLM
@@ -54,6 +55,7 @@
     "CLIPVisionModel",
     "DeepseekV3ForCausalLM",
     "Exaone4ForCausalLM",
+    "Exaone4_5_ForConditionalGeneration",
     "ExaoneMoeForCausalLM",
     "Gemma3ForCausalLM",
     "Gemma3VLM",

tensorrt_llm/_torch/models/checkpoints/hf/exaone4_5_weight_mapper.py

@@ -0,0 +1,38 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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 tensorrt_llm._torch.models.checkpoints.base_weight_loader import ConsumableWeightsDict
+from tensorrt_llm._torch.models.checkpoints.hf.weight_mapper import HfWeightMapper
+from tensorrt_llm._torch.models.modeling_utils import register_mapper
+
+
+@register_mapper("HF", "Exaone4_5_ForConditionalGeneration")
+class Exaone4_5HfWeightMapper(HfWeightMapper):
+    def preprocess_weights(self, weights: dict):
+        """Rename HF checkpoint prefixes; supports plain dict and ConsumableWeightsDict."""
+        is_consumable = isinstance(weights, ConsumableWeightsDict)
+        renamed = {}
+        for key, value in weights.items():
+            if key.startswith("model.visual."):
+                new_key = key.replace("model.visual.", "visual.")
+                renamed[new_key] = value
+            elif key.startswith("model.language_model."):
+                new_key = key.replace("model.language_model.", "model.")
+                renamed[new_key] = value
+            else:
+                renamed[key] = value
+        if is_consumable:
+            return ConsumableWeightsDict(renamed)
+        return renamed