Thisish3

Follow

Thisish3

Follow

0 followers · 2 following

Achievements

Achievements

Thisish3/README.md

Hi, I'm Gunhee Lee (이건희)

I work on LLM inference infrastructure: serving engines, kernels, and the systems around them.

Open-source contributions

NVIDIA Dynamo #16399: pick a complete tool call when parallel_tool_calls is false while streaming (Rust, Chat Completions and Responses APIs). Under review
vLLM #56040: honor an explicit VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=0 on AMD RDNA GPUs. Under review
SGLang #39075: traced a decode context parallel crash to the aiter decode path returning no LSE; closed after AMD landed the full fix in #34432

Projects

Serving and deployment

vllm-serving-tuning-practice: continuous batching throughput, prefix caching TTFT, and decode memory-bandwidth utilization on an RTX 3060 Ti
vllm-k8s-deploy-practice: vLLM on Kubernetes with rolling updates, HPA, and a CI latency regression gate, then real vLLM on an RTX 3060 Ti node (k3s on WSL2, throughput within 0.4% of bare Docker over 5 repeated runs)

Agent workloads

agentic-loop-bottleneck-practice: real vLLM tool-calling loops; structured decoding reliability, a LangGraph agent profiled with NeMo Agent Toolkit and Nsight Systems, fp16 vs fp8 KV cache
harness-cpu-overhead-practice: CPU overhead in the agent harness, framed with Amdahl's Law
gil-scheduling-practice: does LangGraph's parallel node execution use multiple cores under the GIL?

Kernels, communication, and hardware

vllm-kernel-practice: Triton versions of vLLM CUDA kernels
allreduce-practice: naive, ring, and recursive-doubling allreduce in C with a cost model, plus transfer/compute overlap on a real GPU
lpu-compiler-practice: toy instruction-block DSL and register allocator modeled on an LPU
bloom-filter-ngram-practice: Bloom filter n-gram pre-filter modeled on vLLM's n-gram speculative decoding

Writing

Blog: thisish3.tistory.com

Pinned Loading

vllm-kernel-practice vllm-kernel-practice Public

Triton kernel practice exercises based on vLLM's CUDA kernels

Jupyter Notebook
agentic-loop-bottleneck-practice agentic-loop-bottleneck-practice Public

LLM agent loops on real vLLM: structured decoding reliability, GPU vs CPU time, a LangGraph agent profiled with NeMo Agent Toolkit and Nsight, fp16 vs fp8 KV on RTX 3060 Ti

Python
allreduce-practice allreduce-practice Public

naive / ring / recursive-doubling allreduce in C (pthreads) with a cost model, plus transfer/compute overlap measured on a real GPU

C
vllm-serving-tuning-practice vllm-serving-tuning-practice Public

vLLM serving tuning benchmarks: concurrency vs throughput, prefix caching TTFT, agentic multi-turn caching (Colab T4), decode memory-bandwidth utilization (RTX 3060 Ti)

Jupyter Notebook
vllm-k8s-deploy-practice vllm-k8s-deploy-practice Public

vLLM on Kubernetes: rolling updates, HPA, CI latency regression gate, and real vLLM on an RTX 3060 Ti node (Docker and k3s on WSL2)

Python
lpu-compiler-practice lpu-compiler-practice Public

Toy instruction-block DSL and register allocator modeled on HyperAccel's LPU/HyperDex architecture

Python