Fankserver

Plain-language glossary of LLM & inference-serving terminology (MoE, MLA, DSA, KV cache, speculative decoding, ...)

Benchmarks for poolside Laguna-S-2.1 (118B-A8B coder) NVFP4 on a single NVIDIA GB10 / DGX Spark (vLLM 0.25.1, TP=1) — DFlash speculative decoding, KV dtype, TP scaling; llama-benchy sweeps with the Qwen-122B depth set for comparison

Benchmarks for GLM-4.7 355B AWQ (QuantTrio W4A16) with vLLM TP=2 across 2x NVIDIA GB10 / DGX Spark over 200G fabric — llama-benchy sweeps, datasets, graphs

DGX Spark (GB10) llama-benchy benchmarks: NVIDIA Nemotron-Puzzle-75B-A9B NVFP4 with MTP speculative decoding on stock vLLM

Benchmarks for NVIDIA Nemotron-3-Super-120B-A12B (NVFP4, hybrid MoE) with vLLM+MTP on a single NVIDIA GB10 / DGX Spark — llama-benchy sweeps, datasets, graphs

Benchmarks for Tencent Hy3-295B (Hunyuan 3, NVFP4 MoE) with vLLM+Ray at TP=2 across 2x NVIDIA GB10 / DGX Spark — llama-benchy sweeps, datasets, graphs

DGX Spark (GB10) llama-benchy benchmarks: Qwen3.6-27B-AEON — 3 vLLM engines, BF16 vs NVFP4, DSpark-DFlash draft head

LLM inference benchmarks, tuning results, and production notes for AMD MI50 (gfx906 / Vega 20) on ROCm — covering Ornith-1.0-35B, Gemma-4-26B, Qwen3.6-35B, MTP, DFlash, EAGLE3, and Docker vs bare-metal comparisons.

Benchmarks for Qwen3.5-122B-A10B (hybrid MoE) with vLLM+DFlash on a single NVIDIA GB10 / DGX Spark — llama-benchy sweeps, datasets, graphs

BepInEx plugin that dumps Vanguard Galaxy's live catalogue (ships, turrets, armor, aspects, factions, conquest ranks) to JSON, plus a Python merger for the wiki's Module:ShipData.