hotpath 0.3.0
hotpath 0.3.0 focuses on serving-timing correctness and the local profiling flow.
Highlights:
serve-profilenow honors configured concurrency and stops dispatching new requests when the requested duration window closes.- client TTFT is now measured from the first streamed token chunk instead of HTTP first-byte timing.
- local vLLM DEBUG log autodiscovery now follows the real live listener stdout/stderr files instead of stale
.hotpath/video-server/*logs. - vLLM v1 parsing now merges exact request IDs from
Added request ...lines with anonymousRunning batch ... BatchDescriptor(...)lines. - vLLM v1 server queue, prefill, and decode timings are now refined from Prometheus when the raw DEBUG timestamps are only second-resolution, and the report explicitly discloses that.
serve-reportno longer renders missing timing as fake0.0 msvalues.- the live
serve-profiledashboard redraw path was hardened for Ghostty-style terminals so updates happen in place instead of stacking duplicate frames. logs.mdis now included in the repo with a detailed UTC log of this pass.
Focused verification run before release:
cmake --build build --parallelctest --test-dir build -R 'test_log_parser|test_serve_profiler|test_serve_report|test_traffic_replayer|test_cli|test_audit' --output-on-failure- live local smoke against
Qwen/Qwen3.5-4Bonlocalhost:8000