SYCL: improve MoE prefill throughput (llama/23142)
- change `k_copy_src1_to_contiguous` so that uses a precomputed contiguous mapping where all rows "owned" by an expert are in one slice with a know starts and ends - switch the `O(n_as * n_routed_rows)` contraption to a counting sort-based procedure with `O(n_as + n_routed_rows)` complexity
A
Alexey Kopytko committed
b0c9f9005926301e8a8306bd300e86dbd09db41d
Parent: 21c65a7
Committed by Georgi Gerganov <ggerganov@gmail.com>
on 5/25/2026, 9:26:07 AM