PyTorchKR and XCENA Hosted KVCache Meetup Korea 2026

What happened

On September 30, 2026, PyTorchKR and XCENA co-hosted the KVCache Meetup Korea 2026 in Seoul, bringing together industry experts and developers to explore advancements in large language model (LLM) inference. The event featured presentations from XCENA, SK hynix, FuriosaAI, and LMCache, with topics covering KV Cache optimization and related technical challenges.

Attendees were given the option to select up to two interest categories during registration, including Model, Serving, Infra, Hardware, and Memory. The session included welcoming remarks from Junghwan Park of PyTorchKR and Juhyun Kim, CPO at XCENA.

Why it matters

The discussion highlighted the growing importance of efficient memory management and cache storage in LLM inference, particularly as models become larger and more complex. KV Cache plays a critical role in storing attention data for reuse, and optimizing this process is essential for improving performance and reducing resource usage.

Key details

Jooho Lee from XCENA explained how CXL technology expands the memory accessible by CPUs and introduced Maru, an open-source KV Cache storage engine. Sejin Oh from SK hynix presented an analysis of GPU compute and HBM bandwidth in LLM serving systems, using two H100 GPUs and the Llama 3.1 8B model.

Donghun Lee from FuriosaAI discussed request scheduling and KV Cache management in Furiosa-LLM, emphasizing techniques like PagedAttention, Continuous batching, and Chunked prefill. MP Mode was also introduced as a method using CUDA IPC to share GPU memory across processes.