The existing backward tests stop at 16 kv blocks. The TMEM race fixed in
88319b7 never fired there but corrupted dk/dv on every rerun from 32K
tokens, and the nb >= 1024 device-computed work order had no test at all.
This case (1024 kv blocks, 12.5% density) covers both: sm_100a route vs
all-Triton within the existing tolerances, and two sm_100a runs within
summation-order noise (invert_indices orders each kv row's q list with
atomics, so exact bitwise equality is not available through this path;
measured delta 5e-3 rel_max / 7e-6 mean, the race gave 0.6-1.0 / 4e-2).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01G4KbxfFobc4FUXLZV3H4QF
Tests
tests/local_tests/holds local-only parity, smoke, and component tests organized by model family (e.g.tests/local_tests/ltx2/,tests/local_tests/stable_audio/). They typically require a checked-out reference repo and/or model weights and are skipped in CI; see each family'sREADME.mdfor setup. The top-leveltests/local_tests/README.mdis the family index.- The CI-backed test suite still lives in
fastvideo/tests/. - Eventually, all tests will move under
tests/.