mtmd: add --mmproj-device argument (#23255)

* feat: add --mmproj-device arg & backwards compatible MTMD_BACKEND_DEVICE env var

* feat: load mmproj device backend immediately, add -mmdev shortflag

* fix: its a pointer now get the name

* clean up

* gen docs

* nits

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
This commit is contained in:
John-Henry Lim
2026-08-20 18:45:37 +02:00
committed by GitHub
co-authored by Xuan Son Nguyen
parent 07822bddf8
commit 681c29d36a
12 changed files with 41 additions and 11 deletions
+20
View File
@@ -2595,6 +2595,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.mmproj_use_gpu = value;
}
).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
add_opt(common_arg(
// note: "-mmdev" must sort after "--rpc" in the preset map, else RPC devices are not registered yet
{"-mmdev", "--mmproj-device"}, "DEVICE",
"device to use for multimodal projector (none = don't offload, default: auto)\n"
"use --list-devices to see a list of available devices",
[](common_params & params, const std::string & value) {
if (value == "none") {
params.mmproj_use_gpu = false;
params.mmproj_device = nullptr;
return;
}
auto devices = parse_device_list(value);
// parse_device_list pushes nullptr at back so devices is length 2 for single device.
if (devices.size() > 2) {
throw std::invalid_argument("only one device may be specified for mmproj");
}
params.mmproj_use_gpu = true;
params.mmproj_device = devices.front();
}
).set_examples(mmproj_examples).set_env("MTMD_BACKEND_DEVICE")); // no LLAMA_ARG_ prefix for backward compatibility reason
add_opt(common_arg(
{"--image", "--audio", "--video"}, "FILE",
"path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files\n",