facebookresearch/faiss
1
0
Fork 0
mirror of https://github.com/facebookresearch/faiss.git synced 2026-10-11 22:50:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

De-flake bench_transposed_encode (CI timeout) (#5360)

Browse Source 
Summary:
Pull Request resolved: https://github.com/facebookresearch/faiss/pull/5360

`bench_transposed_encode` is an ad-hoc microbenchmark wired into CI as a plain `python_unittest`. It encodes 1,000,000 vectors with `ProductQuantizer.compute_codes`, single-threaded, repeated 5x, across all SIMD levels (`NONE`/`AVX2`/`AVX512`). Under `-c faiss.dynamic_dispatch=true` all three levels are available (the slow scalar `NONE` path is always available), so all three are timed — roughly tripling the work and pushing wall-clock right at the 600s test deadline (TestInfra showed `duration_min == duration_max == 600s`, avg 424s). It intermittently tripped the deadline, so TestWarden auto-disabled it (T277674127, dups T277525456 / T277525454).

Fix: shrink the deterministic workload (`nb` 1,000,000 -> 50,000, `ntrain` 100,000 -> 20,000, `n_repeat` 5 -> 2) so the run finishes well under the deadline even when DD mode times all three SIMD levels, and add the `tpx_labels.long_running` label for extra timeout headroom. The cross-level `compute_codes` coverage is preserved. The "codes identical across levels" line stays a diagnostic print (not an assertion), because scalar vs FMA reduction-order differences legitimately change nearest-centroid tie-breaks, so exact code equality across SIMD levels is not guaranteed.

___

Differential Revision: D110092205

fbshipit-source-id: 2d4cabaf53e1dba0a1f74b7670f4cbbbebea1dbb
This commit is contained in:
Tia Lim
2026-06-29 15:07:10 -07:00
committed by meta-codesync[bot]
parent 027054f40f
commit 54f83a4131
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/bench_transposed_encode.py
+6 -3
View File
@@ -103,11 +103,14 @@ def main() -> None:
# Single-threaded for a clean per-core kernel signal.
faiss.omp_set_num_threads(1)
# Sizes stay small so the run finishes well under the CI test deadline even
# in dynamic-dispatch mode, where the slow scalar NONE level is also timed
# on every case.
# dsub in {1,2,4,8} -> the fast _D specialization is taken.
run_case(d=128, m=32, nbits=8, nb=1_000_000, ntrain=100_000, n_repeat=5)
run_case(d=64, m=16, nbits=8, nb=1_000_000, ntrain=100_000, n_repeat=5)
run_case(d=128, m=32, nbits=8, nb=50_000, ntrain=20_000, n_repeat=2)
run_case(d=64, m=16, nbits=8, nb=50_000, ntrain=20_000, n_repeat=2)
# dsub=16 -> _D fast path NOT taken (control: no AVX512 vs NONE gap).
run_case(d=128, m=8, nbits=8, nb=1_000_000, ntrain=100_000, n_repeat=5)
run_case(d=128, m=8, nbits=8, nb=50_000, ntrain=20_000, n_repeat=2)
sys.stdout.flush()