Repository navigation
tests: MUL_MAT_ID cases for the one-column mat-vec store (small K, several expert rows per block) - #295
Conversation
|
Thanks. This still doesn't build on prism: #258 added its own nrows_dst bound in mul_mat_vec_q (ids ? stride_channel_dst : stride_col_dst), so the new declaration at mmvq.cu:620 collides with it. That bound already covers the one-column case: I applied only your test hunk to current prism and all 50 new MUL_MAT_ID cases pass (1085/1085). Could you drop the kernel change and keep the test? Then it's ready to merge as a regression test. |
4cdb698 to
ed56a5c
Compare
|
Thank you. Done in RTX 4070 (sm_89), CUDA 13: |
…veral expert rows per block) The kernel bound from the first version of this PR is dropped: prism PrismML-Eng#258 already bounds the store (nrows_dst = ids ? stride_channel_dst : stride_col_dst), and these cases pass with it. They stay as a regression test.
ed56a5c to
01982f5
Compare
|
I rebased this PR onto prism Conflict: RTX 4070 (sm_89), CUDA 13, Windows: |
Regression test only. The kernel bound from the first version is dropped: prism #258 already bounds the one-column MUL_MAT_ID store (
nrows_dst = ids ? stride_channel_dst : stride_col_dst). These cases pass with it:test-backend-ops -o MUL_MAT_ID1085/1085 on an RTX 4070 (prism c96baf4).