]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
arg: allow -kvu flag for llama-perplexity (#18117)
authorTrevorS <redacted>
Wed, 17 Dec 2025 06:33:02 +0000 (22:33 -0800)
committerGitHub <redacted>
Wed, 17 Dec 2025 06:33:02 +0000 (08:33 +0200)
commit4b2a4778f81f222c12271ce3b1997990b3071faf
tree768012bf1543772825b7b944e93e0ae83bdaa6a9
parent58062860afb88e555857c1266d3a17e1b65b5eb9
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.
common/arg.cpp