fix multimodal backend

common/arg.cpp

@@ -2338,6 +2338,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.mmproj_use_gpu = false;
         }
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD"));
+    add_opt(common_arg(
+        {"--mmproj-backend"}, "NAME",
+        "GPU backend for multimodal projector (e.g. CUDA, Metal, Vulkan)\n"
+        "if not specified, will use MTMD_BACKEND_DEVICE env var or default GPU backend",
+        [](common_params & params, const std::string & value) {
+            params.mmproj_backend = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"--image", "--audio"}, "FILE",
         "path to an image or audio file. use with multimodal models, can be repeated if you have multiple files\n",

common/common.h

@@ -402,6 +402,7 @@ struct common_params {
     // multimodal models (see tools/mtmd)
     struct common_params_model mmproj;
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
+    std::string mmproj_backend = "";    // GPU backend for multimodal model (e.g. "CUDA", "Metal", "Vulkan")
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 

tools/mtmd/clip.cpp

@@ -28,11 +28,7 @@
 #include <numeric>
 #include <functional>
 
-struct clip_logger_state g_logger_state = {
-    GGML_LOG_LEVEL_CONT,           // verbosity_thold
-    clip_log_callback_default,     // log_callback
-    NULL                           // log_callback_user_data
-};
+struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL};
 
 enum ffn_op_type {
     FFN_GELU,
@@ -401,7 +397,7 @@ struct clip_ctx {
             throw std::runtime_error("failed to initialize CPU backend");
         }
         if (ctx_params.use_gpu) {
-            auto backend_name = std::getenv("MTMD_BACKEND_DEVICE");
+            auto backend_name = ctx_params.backend_device ? ctx_params.backend_device : std::getenv("MTMD_BACKEND_DEVICE");
             if (backend_name != nullptr) {
                 backend = ggml_backend_init_by_name(backend_name, nullptr);
                 if (!backend) {

tools/mtmd/clip.h

@@ -25,6 +25,7 @@ enum clip_modality {
 struct clip_context_params {
     bool use_gpu;
     enum ggml_log_level verbosity;
+    const char * backend_device; // optional, if null will use env var or default GPU backend
 };
 
 struct clip_init_result {

tools/mtmd/mtmd.cpp

@@ -91,6 +91,7 @@ mtmd_context_params mtmd_context_params_default() {
     params.verbosity = GGML_LOG_LEVEL_INFO;
     params.image_marker = MTMD_DEFAULT_IMAGE_MARKER;
     params.media_marker = mtmd_default_marker();
+    params.backend_device = nullptr;
     return params;
 }
 
@@ -152,6 +153,7 @@ struct mtmd_context {
         clip_context_params ctx_clip_params;
         ctx_clip_params.use_gpu   = ctx_params.use_gpu;
         ctx_clip_params.verbosity = ctx_params.verbosity;
+        ctx_clip_params.backend_device = ctx_params.backend_device;
         auto res = clip_init(mmproj_fname, ctx_clip_params);
         ctx_v = res.ctx_v;
         ctx_a = res.ctx_a;

tools/mtmd/mtmd.h

@@ -82,6 +82,7 @@ struct mtmd_context_params {
     enum ggml_log_level verbosity;
     const char * image_marker; // deprecated, use media_marker instead
     const char * media_marker;
+    const char * backend_device; // optional GPU backend name (e.g. "CUDA", "Metal", "Vulkan"), if null will use env var or default
 };
 
 MTMD_API const char * mtmd_default_marker(void);