Skip to content

fix: unblock Thor validation (vocab-ce API, sageattention2 gate, fp8-cross-attention docs) - #8

Open
shideqin wants to merge 3 commits into
flashrt-project:mainfrom
shideqin:feat/thor-test-and-doc-fixes
Open

fix: unblock Thor validation (vocab-ce API, sageattention2 gate, fp8-cross-attention docs)#8
shideqin wants to merge 3 commits into
flashrt-project:mainfrom
shideqin:feat/thor-test-and-doc-fixes

Conversation

@shideqin

@shideqin shideqin commented Aug 4, 2026

Copy link
Copy Markdown

Summary

Three small changes that unblock the on-device Thor validation of previously green packages:

  1. flashrt-vocab-ce-train: expose the documented vocab_ce_loss public API (README/CARD list it but the package only exposed vocab_ce). Installed smoke/full 2/2 pass.
  2. sageattention2-blackwell: widen the test capability gate (was major>=12, rejecting Thor sm_110) to the full Blackwell family. With the gate open, 9/9 full rows pass on Thor (fp8v cos>=0.9992).
  3. fp8-cross-attention-blackwell: record the installed-artifact SM110 validation (built with CUTLASS 4.4; 9/9 rows + graph/compile pass, production B1,Sq786,Sk7984 cos=0.999786).

The README/CARD document vocab_ce_loss as the public training API but the
package only exposed vocab_ce. Add the documented alias (delegating to
vocab_ce) and re-run the installed smoke/full suites (2/2 pass).
The test gate required major>=12, rejecting Thor (sm_110, major 11) even
though all six kernels pass on-device. Accept the full Blackwell family
(sm_100/103/110/120); the widened gate passes 9/9 full rows on Thor
(fp8v rows cos>=0.9992).
…validation

Built for sm_110a with CUTLASS 4.4 (the declared cutlass_4_4 dependency;
4.5.x moved SM100_MMA_F8F6F4_SS to a class template and breaks csrc/fmha77)
and closed the installed-artifact gate on Thor: 9/9 rows plus rejection,
CUDA Graph replay, and torch.compile pass; production B1,Sq786,Sk7984 row
cos=0.999786.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants