Skip to content

Commit 2d9d422

Browse files
rahulbatra85Rahul Batrak50112113okakarpa
authored
[TRITON] Triton RoPE Fwd Kernels (ROCm#299)
* Triton RoPE Fwd Kernels * [Triton]: Add SGLANG/DS RoPE Kernel * set sequence length block upper limit to 128 on _rope_fwd_kernel_gptj_cached_thd_position_2c * clean up * bypass addtional global memory loads and tune SPLIT_SEQ_SIZE * add NotImplementedError for NEOX style --------- Co-authored-by: Rahul Batra <[email protected]> Co-authored-by: k50112113 <[email protected]> Co-authored-by: omkar kakarparthi <[email protected]>
1 parent 3994bf4 commit 2d9d422

File tree

3 files changed

+2765
-0
lines changed

3 files changed

+2765
-0
lines changed

0 commit comments

Comments
 (0)