Calvin806/EXAONE-3.5-32B-Instruct-GGUF

GGUF quantizations for EXAONE-3.5-32B-Instruct.

Contents

This folder typically contains:

• EXAONE-3.5-32B-Instruct.F16.gguf
• EXAONE-3.5-32B-Instruct.Q4_K_M.gguf
• EXAONE-3.5-32B-Instruct.Q5_K_M.gguf
• EXAONE-3.5-32B-Instruct.Q8_0.gguf (optional)

---

🔧 llama.cpp patch (EXAONE GGUF quantize compatibility)

EXAONE GGUF 변환/양자화 과정에서 일부 모델(예: 2.4B / 7.8B) 간 KV key 네이밍 불일치가 발견되었습니다.

• 어떤 GGUF는 exaone.attention.layer_norm_epsilon만 존재
• 어떤 GGUF는 exaone.attention.layer_norm_rms_epsilon만 존재

이 상태에서 vanilla llama.cpp의 llama-quantize가 특정 키를 찾지 못해 실패할 수 있어, llama.cpp의 model loader에서 gguf key lookup에 fallback을 추가하는 패치를 적용했습니다.

What was patched

src/llama-model-loader.cpp에서 gguf_find_key() lookup에 다음 fallback을 수행하도록 수정:

• key가 exaone.attention.layer_norm_epsilon이고 찾지 못하면 → exaone.attention.layer_norm_rms_epsilon로 재시도
• key가 exaone.attention.layer_norm_rms_epsilon이고 찾지 못하면 → exaone.attention.layer_norm_epsilon로 재시도

이 패치를 통해 EXAONE 3.5 / EXAONE-Deep 2.4B, 7.8B, 32B 계열을 동일 파이프라인으로 GGUF+quantize할 수 있습니다.

Patch note (minimal diff summary)

• Added a fallback wrapper/hook for gguf_find_key() inside llama-model-loader.cpp
• Ensured all lookups in that translation unit route through the fallback

This repo includes:

• exaone-gguf-fallback.patch

Tested llama.cpp commit

• 021cc28bef4dd7d0bf9c91dbbd0803caa6cb15f2

---

Build (CUDA)

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git apply ../exaone-gguf-fallback.patch

rm -rf build
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON
cmake --build build -j

Convert / Quantize

# Convert HF snapshot -> GGUF(F16)
python3 llama.cpp/convert_hf_to_gguf.py <LOCAL_SNAPSHOT_DIR> --outtype f16 --outfile model.F16.gguf

# Quantize (example: Q4_K_M)
llama.cpp/build/bin/llama-quantize model.F16.gguf model.Q4_K_M.gguf Q4_K_M