DeepSpeed
64bd29d6 - fix: materialize concat-sliced k/v for decode_attn contiguity (Loop 9b)

Commit
6 days ago
fix: materialize concat-sliced k/v for decode_attn contiguity (Loop 9b) Signed-off-by: Guokai Ma <guokai.ma@intel.com>
Author
Parents
Loading