Skip to content
View 100-JM's full-sized avatar

Block or report 100-JM

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
100-JM/README.md

백지명 (Jimyeong Baek)

AI Infrastructure Engineer — GPU serving on Kubernetes, LLM inference runtimes, and MLPerf Inference benchmarking.

Kubernetes Go Python CUDA TensorRT-LLM vLLM Slurm Ansible

Most of my day job lives in private repositories, so this page tracks what I have been able to contribute upstream.

Upstream contributions

7 merged · 10 in review — NVIDIA (TensorRT-LLM · DeepOps · GPU Operator) · Kubeflow (Trainer · Katib) · MLCommons Inference · vLLM · Triton Inference Server

Merged

kubeflow/trainer#4159 Do not panic on a nil numProcPerNode in the Flux runtime plugin
NVIDIA/TensorRT-LLM#18865 Warn when a tool parser detects markup but extracts no tool calls a silently dropped tool call now leaves a diagnostic naming the configured parser
NVIDIA/TensorRT-LLM#18856 Cover /v1/responses in per-request perf metrics tests closed the regression gap left by an earlier fix
NVIDIA/deepops#1397 Run nvidia-smi tasks outside the ssh cgroup GPU tasks no longer die with the login session that started them
NVIDIA/deepops#1394 Do not replace the DCGM that DGX OS already ships
NVIDIA/deepops#1390 Prefer in-tree nvidia_peermem over legacy nv_peer_mem
NVIDIA/deepops#1389 Use ssh.service on Debian-family when hiding GPUs from login shells

In review

triton-inference-server/server#8995 Fix avg_time_queue_us under-reporting queue time in the k8s-onprem autoscaling rule
vllm-project/vllm#56582 Keep a plain CUDA arch requested alongside its a/f variant sm_100a-only kernels failed to load on B300 (CC 10.3)
NVIDIA/gpu-operator#2870 Configurable ServiceAccount for the DCGM Exporter lets IRSA / Workload Identity bind to the operator-managed exporter
mlcommons/inference#2668 Bound and pre-load the DeepSeek-R1 LiveCodeBench grader workers a grader that OOM-died was scoring its own failure as a wrong answer
mlcommons/inference#2669 Pin anthropic<1.0 in the DeepSeek-R1 evaluation requirements
mlcommons/inference#2670 Fix the llama3.1-8b Offline vLLM SUT on current vLLM
mlcommons/inference#2671 Make the llama3.1-8b run scripts and README accuracy flow runnable
mlcommons/inference#2672 Default --audit-conf to audit.config in the reference entry points
kubeflow/trainer#4045 Do not allocate a worker when runLauncherAsNode is set
kubeflow/katib#2720 Keep the skopt suggestion service from stalling and OOMing

What I work on

  • GPU serving on Kubernetes — operators, autoscaling (KEDA), multi-instance GPU, DCGM metrics
  • LLM inference runtimes — TensorRT-LLM and vLLM, quantization (FP8 / NVFP4), disaggregated serving
  • MLPerf Inference — benchmarking and submission pipelines on NVIDIA B300

Pinned Loading

  1. NVIDIA/TensorRT-LLM NVIDIA/TensorRT-LLM Public

    TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. Tensor…

    Python 15.1k 2.9k

  2. NVIDIA/gpu-operator NVIDIA/gpu-operator Public

    NVIDIA GPU Operator creates, configures, and manages GPUs in Kubernetes

    Go 3k 598

  3. kubeflow/trainer kubeflow/trainer Public

    Distributed AI Model Training and LLM Fine-Tuning on Kubernetes

    Go 2.3k 1.1k

  4. NVIDIA/deepops NVIDIA/deepops Public

    Tools for building GPU clusters

    Python 1.5k 363