Commit Graph
1 Commits
Author SHA1 Message Date
TrevorSandGitHub 4b2a4778f8 arg: allow -kvu flag for llama-perplexity (#18117)
The -kvu (--kv-unified) flag is required for hellaswag and winogrande
benchmarks which use coupled sequences. Without unified KV cache,
these benchmarks fail with:

  split_equal: sequential split is not supported when there are
  coupled sequences in the input batch (you may need to use the -kvu flag)

This change adds LLAMA_EXAMPLE_PERPLEXITY to the allowed examples for
the -kvu argument, enabling its use with llama-perplexity.
2025-12-17 08:33:02 +02:00