Bug: llama-bench: split-mode flag doesn't recognize argument 'none'

[letter-v]

What happened?

Example command:

llama-bench -m /path/to/model.gguf -p 0 -ctk q4_0 -ctv q4_0 -fa -sm none -mg 0 -ngl 50 -n 128,256,512

Example output:

error: invalid parameter for argument: none
usage: llama-bench [options]

Setting -sm=none doesn't give any error output, but llama-bench seems to disregard this and offloads to the internal GPU anyway. For what it's worth, using llama-cli and llama-server with -sm none works fine:

llama-server -c 64000 -fa -ctk q4_0 -ctv q4_0 -ngl 50 -m /path/to/model.gguf --port 8188 -sm none -mg 0

Name and Version

llama-bench doesn't seem to have a --version flag, but llama-cli --version prints out the following:

version: 3761 (6262d13e)
built with cc (GCC) 14.2.1 20240801 (Red Hat 14.2.1-1) for x86_64-redhat-linux

What operating system are you seeing the problem on?

Linux

Relevant log output

No response

[ggerganov]

llama-bench requires to specify -fa [values] in contrast to llama-server that uses just -fa.

Try this:

llama-bench -m /path/to/model.gguf -p 0 -ctk q4_0 -ctv q4_0 -fa 1 -sm none -mg 0 -ngl 50 -n 128,256,512