doug.sh

Building things with keyboards and end mills. Rabbit holes included.

Qwen

2026

Nov 11
gpu-unlock on Two RTX 3090s: Faster Prefill, No Decode Gain
linux gpu nvidia kernel llm vllm qwen self-hosted
Nov 11
GPU P2P On vs Off: Tensor Parallel LLM Serving on Four R9700s
llm self-hosted gpu pcie amd rocm qwen linux
Nov 8
Six GPUs on One x16 Slot: R9700s and RTX 3090s Behind a PCIe Switch
llm ai self-hosted gpu pcie amd nvidia rocm vllm qwen linux
Sep 23
Custom Models in Oh My Pi: vLLM, llama.cpp, SGLang and More
llm ai agents self-hosted oh-my-pi qwen vllm llama-cpp sglang bifrost
Sep 15
Keeping vLLM's Prefix Cache Warm Between Agent Turns
llm ai agents self-hosted vllm gpu qwen oh-my-pi mcp bifrost
Sep 13
Tuning a Local Coding Agent: Oh My Pi and Qwen3.8-27B on Two RTX 3090s
llm ai agents self-hosted oh-my-pi qwen vllm gpu