Arc B70 Struggles

Ran Qwen3.5-27B on Intel Arc B70 and summarized the perceived performance and errors by backend.

• llama-server + Vulkan: Works normally but very slow. Roughly 300/10 tokens/sec level.
• llama-server + OpenVINO: Doesn't work at all. A CPY error related to pre-allocated tensor occurs.
• llama-server + SYCL: Performance improves noticeably to around 800/20 tokens/sec, but there's a problem where output breaks on large requests.
• vLLM INT4: Local build failed, but the intel/llm-scaler-vllm docker image runs. Ingestion is fast at up to 2200 tokens/sec, but generation is only about 10 tokens/sec.

Overall, this is a real-world account of testing various stacks on the Intel Arc B70, with the key point being that performance varies greatly by backend and stability is still lacking. The author especially noted a problem where tool calling doesn't work properly in vLLM, and asked other users for their success or failure stories.