A chronological record of releases and completed evaluations. Detailed measurements live in the benchmark pages.
Publication feed, not live GPU telemetry. Dates identify the recorded milestone; historical entries retain their original results.
Showing the published update history.
Expert reuse, root reductions and scalar attention broadcasts were tested on both V100s. The matched comparisons did not earn a runtime promotion. Correctness, regressions, controls and excluded harness attempts are retained in the study.
View supporting evidence →Rust and Kairo now load the verified 47f0b50 runtime. The complete live before/after evaluation preserves matching outputs and shows substantial Rust prefill variation, without a demonstrated serving speedup.
View supporting evidence →The a078 CPU expert port was followed by 0982 parallel activation preparation. Kairo retained its existing attention/MTP profile. The measured effects and retained failures are documented in the engine study.
View supporting evidence →The first Rust evaluation covered Titan Xp and V100 configurations, offload boundaries, and unsupported engine paths. It established the historical baseline retained in the configuration explorer.
View supporting evidence →