From 853634afd11f79da6fca0f7b6048ea35e1b683a8 Mon Sep 17 00:00:00 2001 From: Trevor Strieber Date: Tue, 16 Dec 2025 15:40:48 -0800 Subject: [PATCH] arg: allow -kvu flag for llama-perplexity 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 | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/common/arg.cpp b/common/arg.cpp index f2aec895ba..a4ffd5586c 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -1140,7 +1140,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.kv_unified = true; } - ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER})); + ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY})); add_opt(common_arg( {"--context-shift"}, {"--no-context-shift"},