The XPU additions in this release

The SGLang v0.5.15 release notes include speculative-decoding support on XPU, Intel quantization support and work on XPU graphs. They also list XPU profiling support in the offline throughput benchmark. These are separate changes, so the selected model and execution path matter when evaluating the release.

Graph capture for decode and prefill

Pull request #29053 introduces capture and replay for a full decode graph and piecewise prefill on Intel XPU. It adds XPU graph runners, attention metadata buffers and a test using Qwen2.5-1.5B. Its stated purpose is to reuse captured execution instead of repeatedly dispatching the same work from the host.

The release also records a subsequent change that makes XPU graphs opt-in. The existence of the graph implementation therefore needs to be distinguished from the defaults of a particular build.

The test limitation is part of the story

Pull request #30048 restores a flag disabling decode graphs in the Intel XPU backend test. The author traces startup failures to a scratch-memory extension used by its attention kernel that was unsupported during SYCL graph capture. The patch changes the test configuration; the underlying kernel limitation requires separate work.

For an Arc comparison, my focus would be the exact graph mode and attention backend alongside the model and dependencies. An enabled feature, a passing startup and measured performance are different observations. This note describes the upstream changes and their documented scope.