mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-29 19:28:33 +00:00
5bebfd460f
Add tile size for FMHA batch prefill bf16 for MI308X ## Motivation Adding a tile size adapted to MI308X, for the FMHA Batch Prefill BF16 input type case ## Technical Details N/A ## Test Plan Benchmarking from the Aiter side with: ``` python3 op_tests/test_batch_prefill.py -s 8000 -p 1 -q 4 -k 1 --head_dim 256 -c true -d bf16 --input_dtype bf16 --quant_method none --kv_layout linear -t sglang -l 0.0 --return_lse false --profile ``` ## Test Result We see an improvement with the new tile size on MI308X (both with PLT mode OFF and ON) ## Submission Checklist - [X] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests. Co-authored-by: Damien Lejeune <damien.lejeune@amd.com>