Repository navigation
performance_parity GPU-010 asserts a llama.cpp reference it never measures #5072
Description
crates/aprender-serve/examples/performance_parity.rs GPU-010 (bench_production_parity, around :1132-1142) is framed on a llama.cpp reference (256 tok/s for a 7B model on CUDA, an 80% parity target) that this example never measures. #4756 / PR #5068 removed the same kind of asserted comparator from GPU-018 (F12, PERF-008) and left GPU-010 alone as out of scope.
Done when: GPU-010 either measures its comparator on the same config and model, or reports only its own figure with no parity claim; PERF-003's M7 row says which.
Found by the #4756 quorum review.
Activity
fix(examples): GPU-018 drops its fabricated llama.cpp parity; include…
fix(examples): examples-nightly 36 rows to pass or a skip class, fail…
noahgift commentedon Oct 12, 2026
Branch
fix/5072-gpu010-no-asserted-parity@1d33e87d88d6053e988f23231013d05a4c52456c (branch-only, no PR).Which way: GPU-010 now reports only its own figure and makes no parity claim. llama.cpp is not run.
parity_pct >= 80was that same check written a second way.GPU-010: Sustained Gen, compared against our own 50 tok/s floor. The llama.cpp text andparity_pctare gone. This is the same rule GPU-018 got in fix(examples): examples-nightly 36 rows to pass or a skip class, fail 0 timeout 0 (#4756) #5068 (PERF-008, F12).PERF-003-performance-parity-tracking.md: added a dated 2026-10-12 correction. Every M7 row and line now says it is our own sustained floor and that llama.cpp parity is not measured.Measured (release,
cargo run --release --example performance_parity):EXIT=0. The suite was 37/38. The one red is GPU-020 (Optimized Gen speedup 0.60x vs 0.90x), a separate row this change does not touch.
Other checks
check_no_fabricated_baselines.sh: PASS.cargo fmt --check: clean.