Welcome to my page!

I'm Jungwoo Song, solutions architect from Seoul, Korea.

Contribute to LMCache, Valkey, KServe, Karmada

• 2026.10 Growing padded full attention blocks in vLLM to fill the unified KV page, adding about 16% KV cache capacity for hybrid models like Gemma-4 and Qwen3.6 (vllm#63347)
• 2026.10 Adding a POSIX fd delivery channel so LMCache MP mode can register VMM KV pools on GPUs without IMEX (LMCache#5867)
• 2026.08 Adding a claim-scoped local IPC directory to the NVIDIA GPU DRA driver, enabling cross-Pod CUDA IPC without hostIPC or hostPath (dra-driver-nvidia-gpu#1391)

• 2026.10 Fixed KV registration for hybrid models with padded KV pools under CUDA IPC (#5693, #5855)
• 2026.10 Supported KV events under MultiConnector for P/D disaggregation (#5524)
• 2026.08 Made hostIPC optional in the LMCache operator, part of the 2026 Q3 roadmap to remove the hostIPC and privileged mode requirements (#4399)
• 2026.07 Attached the SageMaker HyperPod daemon's shared memory read-only, so an LMCache pod restart no longer wipes the node-wide cache (#4181)
• 2026.07 Added a SageMaker HyperPod L2 adapter for MP mode (#4138)
• 2026.03 Supported numbered databases in the Valkey connector, including Valkey 9.0 cluster mode (#2307)
• 2025.10 Contributed a Valkey connector with pipelining, achieving at least 20% RTT improvement over the existing Redis connector (#1743)

• 2026.09 Guarded HNSW prefetches against empty neighbor lists in valkey-search (valkey-search#1320)
• 2025.03 Documented the valkey-benchmark --rfr option (valkey-doc#254)
• 2024.12 Enabled reads from replicas in valkey-benchmark for more accurate cluster-mode testing (valkey#1392)

• 2026.02 Parallelized S3 file downloads, making model downloads about 2x faster (#5102)

• 2026.08 Built seven agent skills for Karmada users through the LFX Mentorship program (community#216)

[email protected]