Hi! I'm Hong Jae Hyeong, an AI Engineer from Korea.
I work on AI systems, backend services, and scalable infrastructure, always aiming to build reliable and efficient solutions.
I also enjoy contributing to the open source ecosystem and collaborating with developers to make tools and frameworks even better.
| library | stars | summary | link |
|---|---|---|---|
I requested review PR that fix inspect.getcoroutinefunction() can break testing with unittest.mock.patch() |
PR | ||
| I requested review PR that set encodings when reading config files | PR | ||
| I requested review PR that update repository references from 'encode' to 'Kludex' | PR |
- [flashinfer] Cake 결정적 MoE Grouped GEMM: FP32 wgrad 누적을 커널에 fuse해 scratch를 없애고 최대 1.29x 가속 (2026.11.08)
- [flashinfer] Hopper FP8 GEMM+AllReduce 융합 커널의 Grid Lockstep Throttle: CTA가 앞서 달리지 못하게 해서 최대 1.32배 빨라지기 (2026.11.08)
- [flashinfer] GB300 NVFP4 Context Attention: scale slot 조기 반납과 8단 packed K/V ring으로 최대 1.26x 가속 (2026.11.08)
- [vllm] vLLM 멀티모달 인코더 CUDA Graph 재설계: 배치당 한 번의 replay와 5120 토큰 상한 (2026.11.08)
- [sglang] Kimi-K3.1 전용 MXFP4 Decode MoE 커널을 Shape 템플릿으로 일반화해 DeepSeek-V4까지 확장하기 (2026.11.08)




