WTAKO Network

Production serving stack for Qwen3.8-Flash-Next (~176B params, 6B active) on a single NVIDIA DGX Spark (GB10). vLLM 0.32 with native Qwen4Exp support plus DFlash2 speculative decoding (3 draft tokens) hits 57-71 tok/s decode on a table-stripped hybrid checkpoint: INT4 routed experts, blockwise-FP8 side layers and INT8 lm_head. The fp8 n-gram PLE table (320M rows) is served from local SSD through a graph-capturable mmap host func, so CUDA graphs stay whole, with one-sided RDMA READ still available via PLE_RDMA. 262k context, 16 streams, prefix cache + chunked prefill, never-evict KV pinning for system prompts, and an fp8 KV path that reaches 1.1M tokens for 1M-token requests. OpenAI-compatible API on :8000 as `qwen` with tool calling and reasoning parsers. magi-v5 branch of blazux/qwen3.8-Flash-DGX.

WTAKO Network's Frontend (this website) is written in Next.js 16 and MobX, with the help of a local LLM in my AI core inside my home. I just hope it looks good enough because I don't have much design sense. This Next.js project is exported as a static site and it is being hosted on my own server. I never liked the idea of Server-Side Rendering (SSR), and I also avoid Vercel as much as possible.

Hardware & IoT Monitoring Server with real-time Linux hardware monitoring and IoT device tracking via WebSocket. Collects CPU, memory, temperatures, fan speeds, storage health, and network statistics.