---
language:
- ko
- en
license: llama3
library_name: transformers
tags:
- moe
- awq
- quantized
- w4a16
- compressed-tensors
- vllm
- llm-compressor
base_model: LGAI-EXAONE/K-EXAONE-236B-A23B
---

# K-EXAONE-236B-A23B-W4A16-G128
### 🔄 (2026-06-04) Change Calibration Dataset
* use mixed calibration datasets (2,048 samples, sequence length 8,192)
  * [`beomi/KoAlpaca-RealQA`](https://huggingface.co/datasets/beomi/KoAlpaca-RealQA): 512 samples
  * [`ChuGyouk/HFH4_ultrachat_200k_ko`](https://huggingface.co/datasets/ChuGyouk/HFH4_ultrachat_200k_ko): 512 samples
  * [`HuggingFaceH4/ultrachat_200k`](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k): 1,024 samples
### 🔄 (2026-05-04) Change Calibration Dataset
* use [`beomi/KoAlpaca-RealQA`](https://huggingface.co/datasets/beomi/KoAlpaca-RealQA) (512 samples, sequence length 2048)
### 🔄 (2026-04-15) Improved Quantization - scale-up calibration dataset
* (# of Calibration Dataset 512, Sequence len 2048)
### 🔄 (2026-04-13) Bug fixed
* the vLLM rms_norm contiguous-buffer patch is no longer required.
* the tensor-parallel issue is resolved, and inference now works correctly on other GPUs(a100, ...) as well.
### 🔄 (2026-04-13) Improved Quantization - scale-up calibration dataset
* (# of Calibration Dataset 512, Sequence len 512)
### 🔄 (2026-04-10) Initial commit
* (# of Calibration Dataset 32, Sequence len 128)

---

**W4A16 AWQ quantization** of [`LGAI-EXAONE/K-EXAONE-236B-A23B`](https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B), produced with [llm-compressor](https://github.com/vllm-project/llm-compressor).

This is the **first W4A16 AWQ checkpoint** for K-EXAONE-236B-A23B publicly available — the original model only has FP8 and GGUF variants on HuggingFace.

---

## Model Details

| Property | Value |
|----------|-------|
| Base model | LGAI-EXAONE/K-EXAONE-236B-A23B |
| Architecture | ExaoneMoEForCausalLM |
| Total parameters | ~236B |
| Active parameters | ~23B per token |
| Quantization method | AWQ (Activation-aware Weight Quantization) |
| Weight precision | INT4 (packed) |
| Activation precision | BF16 |
| Group size | 128 |
| Quantization scope | All `Linear` layers except `lm_head` and gate projections |
| Compressed-tensors version | 0.15.0 |
| Context length | 262,144 tokens |
| Languages | Korean, English |

### Architecture Highlights

* **48 transformer layers** with mixed sliding-window (`LLLG` pattern) and full attention
* **MoE layers**: 47 sparse MoE layers + 1 dense MLP (layer 0)
* **128 routed experts** + 1 shared expert per MoE layer; top-8 experts activated per token
* **Sigmoid scoring** with `norm_topk_prob=True`
* **Hidden size**: 6144, **MoE intermediate size**: 2048

---

## Quantization Details

Quantization was performed using [llm-compressor](https://github.com/vllm-project/llm-compressor) with a **MoE-aware AWQ** recipe.

The EXAONE specific MoE-aware AWQ recipe was developed in [SqueezeBits/llm-compressor-K-EXAONE](https://github.com/SqueezeBits/llm-compressor-K-EXAONE).

**Method:** AWQ applies channel-wise scaling to minimize quantization error by protecting salient weights, using a calibration dataset to determine optimal scales.

**Recipe highlights:**

* `scheme`: W4A16 (INT4 weights, BF16 activations)
* `group_size`: 128
* `n_grid`: 20 (search resolution for AWQ scale optimization)
* `duo_scaling`: True
* Smooth mappings cover all MoE expert layers (layers 1–47) independently, plus attention and MLP projections
* Layer 0 (dense MLP) and `lm_head` are excluded from quantization
* Gate weight tensors are excluded from quantization

**Calibration dataset:** [`beomi/KoAlpaca-RealQA`](https://huggingface.co/datasets/beomi/KoAlpaca-RealQA) (512 samples, sequence length 2048)

---

## Hardware Requirements

| Precision | VRAM |
|-----------|------|
| This model (W4A16) | ~120 GB |
| Original BF16 | ~480 GB |

**CUDA / driver requirement:** for H200(sm_90), vLLM 0.19.0 wheels are compiled with the CUDA 12.9 toolkit, so you need **CUDA ≥ 12.9** (NVIDIA driver ≥ 575.x) to run without issues. If your driver is older, follow the monkey-patch workaround in the inference section below.
* It is okay for A100(sm_80)

---

## Setup

```bash
# 1. Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Create a Python 3.12 virtual environment
uv venv --python 3.12

# 3. Activate it
source .venv/bin/activate

# 4. Install vLLM and Transformers
uv pip install "vllm==0.19.0"
uv pip install "transformers==5.5.0"
```

---

## Running Inference

Save the script below as `vllm_inference.py` and run:

```bash
python vllm_inference.py
```

```python
from vllm import LLM, SamplingParams

MODEL_PATH = "Hyun9junn/K-EXAONE-236B-A23B-W4A16-G128"


"""# ── Monkey-patch required if NVIDIA driver < 575.x (CUDA < 12.9) ─────────────
# for H200, vLLM 0.19.0 is compiled with CUDA 12.9; older drivers cannot JIT-compile its
# PTX and crash with "cudaErrorUnsupportedPtxVersion" during weight loading.
# This patch forces vLLM to use WNA16MoEMethod (no Marlin CUDA kernels) instead
# of MarlinMoEMethod.  Safe to keep even after upgrading the driver.
import vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe as _ct_moe
_ct_moe.check_moe_marlin_supports_layer = lambda *args, **kwargs: False
# ─────────────────────────────────────────────────────────────────────────────"""


def main():
    llm = LLM(
        model=MODEL_PATH,
        max_model_len=8192,
        trust_remote_code=True,     # K-EXAONE uses custom modeling code
        tensor_parallel_size=2,     # 2x H200; or 4x A100
        enforce_eager=True,
    )

    sampling_params = SamplingParams(
        temperature=0.2,
        top_p=0.9,
        max_tokens=2048,
    )

    prompts = [
        "What is the capital of South Korea?",
        "Explain the difference between MoE and dense transformer models.",
        "Write a short Python function to compute Fibonacci numbers.",
    ]

    tokenizer = llm.get_tokenizer()
    formatted_prompts = [
        tokenizer.apply_chat_template(
            [{"role": "user", "content": p}],
            tokenize=False,
            add_generation_prompt=True,
        )
        for p in prompts
    ]

    outputs = llm.generate(formatted_prompts, sampling_params)

    for prompt, output in zip(prompts, outputs):
        print(f"Prompt : {prompt}")
        print(f"Response: {output.outputs[0].text.strip()}")
        print("-" * 60)


if __name__ == "__main__":
    main()
```

---

## Files

| File | Description |
|------|-------------|
| `model-00001-of-00003.safetensors` | Model weights shard 1/3 |
| `model-00002-of-00003.safetensors` | Model weights shard 2/3 |
| `model-00003-of-00003.safetensors` | Model weights shard 3/3 |
| `model.safetensors.index.json` | Weight shard index |
| `config.json` | Model config with quantization metadata |
| `recipe.yaml` | llm-compressor AWQ recipe used for quantization |
| `tokenizer.json` | Tokenizer |
| `tokenizer_config.json` | Tokenizer config |
| `chat_template.jinja` | Chat template |
| `generation_config.json` | Default generation config |

---

## License

This model inherits the license of the base model [`LGAI-EXAONE/K-EXAONE-236B-A23B`](https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B). Please refer to the original model page for license details.

---

## Citation

If you use this model, please cite the original K-EXAONE work:

```bibtex
@misc{k-exaone-236b,
  title  = {K-EXAONE-236B-A23B},
  author = {LG AI Research},
  year   = {2025},
  url    = {https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B}
}
```