-
-
Notifications
You must be signed in to change notification settings - Fork 69
Open
Description
As today the benchmarker benchs 4 cases of partialsort:
MY_BENCHMARK_CAPTURE(func, type, k10, 10000, 10); \
MY_BENCHMARK_CAPTURE(func, type, k100, 10000, 100); \
MY_BENCHMARK_CAPTURE(func, type, k1000, 10000, 1000); \
MY_BENCHMARK_CAPTURE(func, type, k5000, 10000, 5000);
https://github.com/intel/x86-simd-sort/blob/87486d17a7d916f2992800b0219e28d0af29e765/benchmarks/bench.h#L52
For LLM inference the values are more extreme: N (vocab size) = 100000 or more, and k could be even smaller than 10.
Would you mind if I add a CLI arg to set at least N (ARRSIZE) ?
Metadata
Metadata
Assignees
Labels
No labels