PR #35511 fix for by_channel

fix for by_channel

ceciliapeng2011 committed 59 days ago

fix gemm_qk fp16 uncompressed: remove double dot in Xe2 path, split reduce2d for non-square input

ceciliapeng2011 committed 59 days ago

xattention fix

ceciliapeng2011 committed 58 days ago

fix for xe1

ceciliapeng2011 committed 58 days ago

Merge branch 'master' into cecilia/fix/by_channel

ceciliapeng2011 committed 56 days ago

fix: SUB_BLOCK_SIZE for by_channel only

ceciliapeng2011 committed 56 days ago

SUB_BLOCK_SIZE for all compressions.

ceciliapeng2011 committed 56 days ago

refactor: load scale_zp without splitting to halves.

ceciliapeng2011 committed 56 days ago

more test coverages.

ceciliapeng2011 committed 55 days ago

1. fix A770 cm_pa_xe1 build issue due to causal/mask tailing; 2. enable xe1 xattention unit tests. 3. fix xe1 cm_pa_decoding jit const issue.

ceciliapeng2011 committed 55 days ago

fix kvcache_update for Xe1.

ceciliapeng2011 committed 55 days ago

fix: initialize rO matrix and skip xattention tests on Xe1

ceciliapeng2011 committed 54 days ago

Merge branch 'master' into cecilia/fix/by_channel

ceciliapeng2011 committed 54 days ago

openvino fix for by_channel #35511 Merged