Tuning small LLM on my ThinkPad's Arch iGPU: from backend, to fast(-ish) long context
The xe driver, dense vs MoE, and speculative decoding at depth. What I measured running small models on an Arc 140T with llama.cpp SYCL.
The Notebook
A working log of testing, tinkering, occasional detours and everything I broke trying.
Trying OpenVINO on my ThinkPad's Arc iGPU
Intel's own AI toolkit on Intel's own laptop chip should be a perfect match. I tried three ways in, here's why I went back to SYCL.
Making a split tunnel out of eduVPN for mail and intranet only
My institute wants me on a VPN to send mail, but I made that VPN carry only my mails and nothing more.
Fusion IRD‑CNRS : les spécificités de l’IRD qui seraient perdues à la suite du projet de réorganisation↗
France’s proposed merger of the IRD with the CNRS could put at risk the IRD’s unique role in sustainable development research and eight decades of partnerships with countries in the Global South.
My $50 flea-market Dell with 7+ hours of video on one charge
How I turned a $50 flea-market Dell Inspiron 11 into my travel laptop, and got 7+ hours of video out of its damaged battery.
Three LFM2.5-2.6B in parallel on an iGPU
Ollama queued every request in a single slot because lfm2 is blocklisted from parallelism. llama-server fixed it.
