run, server: add option to enable/disable multimodal support

docs/options/mtmd.md

@@ -0,0 +1,9 @@
+####> This option file is used in:
+####>   ramalama run, ramalama sandbox goose, ramalama sandbox opencode, ramalama sandbox pi, ramalama serve
+####> If this file is edited, make sure the changes
+####> are applicable to all of those.
+#### **--mtmd**=*BOOL*
+Enable or disable multimodal support.
+If enabled and a multimodal projector file is detected, `--mmproj` is
+added to the llama-server command-line.
+When omitted, multimodal support is automatically detected.

docs/ramalama-run.1.md

@@ -275,6 +275,16 @@ Make sure the sampling parameters like top_k on the web UI are set correctly.
 [//]: # (END   included file options/model-draft.md)
 
 
+[//]: # (BEGIN included file options/mtmd.md)
+#### **--mtmd**=*BOOL*
+Enable or disable multimodal support.
+If enabled and a multimodal projector file is detected, `--mmproj` is
+added to the llama-server command-line.
+When omitted, multimodal support is automatically detected.
+
+[//]: # (END   included file options/mtmd.md)
+
+
 [//]: # (BEGIN included file options/name.md)
 #### **--name**, **-n**
 Name of the container to run the Model in.

docs/ramalama-run.1.md.in

@@ -50,6 +50,8 @@ the execution. URL support means if a model is on a web site or even on your loc
 
 @@option model-draft
 
+@@option mtmd
+
 @@option name
 
 @@option ncmoe

docs/ramalama-sandbox-goose.1.md

@@ -263,6 +263,16 @@ Only used when invoked in router mode (zero or multiple models).
 [//]: # (END   included file options/models-max.md)
 
 
+[//]: # (BEGIN included file options/mtmd.md)
+#### **--mtmd**=*BOOL*
+Enable or disable multimodal support.
+If enabled and a multimodal projector file is detected, `--mmproj` is
+added to the llama-server command-line.
+When omitted, multimodal support is automatically detected.
+
+[//]: # (END   included file options/mtmd.md)
+
+
 [//]: # (BEGIN included file options/name.md)
 #### **--name**, **-n**
 Name of the container to run the Model in.

docs/ramalama-sandbox-goose.1.md.in

@@ -63,6 +63,8 @@ Otherwise only the agent container starts using the given url for the openai com
 
 @@option models-max
 
+@@option mtmd
+
 @@option name
 
 @@option ncmoe

docs/ramalama-sandbox-opencode.1.md

@@ -256,6 +256,16 @@ Only used when invoked in router mode (zero or multiple models).
 [//]: # (END   included file options/models-max.md)
 
 
+[//]: # (BEGIN included file options/mtmd.md)
+#### **--mtmd**=*BOOL*
+Enable or disable multimodal support.
+If enabled and a multimodal projector file is detected, `--mmproj` is
+added to the llama-server command-line.
+When omitted, multimodal support is automatically detected.
+
+[//]: # (END   included file options/mtmd.md)
+
+
 [//]: # (BEGIN included file options/name.md)
 #### **--name**, **-n**
 Name of the container to run the Model in.

docs/ramalama-sandbox-opencode.1.md.in

@@ -61,6 +61,8 @@ Otherwise only the agent container starts using the given url for the openai com
 
 @@option models-max
 
+@@option mtmd
+
 @@option name
 
 @@option ncmoe

docs/ramalama-sandbox-pi.1.md

@@ -260,6 +260,16 @@ Only used when invoked in router mode (zero or multiple models).
 [//]: # (END   included file options/models-max.md)
 
 
+[//]: # (BEGIN included file options/mtmd.md)
+#### **--mtmd**=*BOOL*
+Enable or disable multimodal support.
+If enabled and a multimodal projector file is detected, `--mmproj` is
+added to the llama-server command-line.
+When omitted, multimodal support is automatically detected.
+
+[//]: # (END   included file options/mtmd.md)
+
+
 [//]: # (BEGIN included file options/name.md)
 #### **--name**, **-n**
 Name of the container to run the Model in.

docs/ramalama-sandbox-pi.1.md.in

@@ -65,6 +65,8 @@ Otherwise only the agent container starts using the given url for the openai com
 
 @@option models-max
 
+@@option mtmd
+
 @@option name
 
 @@option ncmoe

docs/ramalama-serve.1.md

@@ -312,6 +312,16 @@ Only used when invoked in router mode (zero or multiple models).
 [//]: # (END   included file options/models-max.md)
 
 
+[//]: # (BEGIN included file options/mtmd.md)
+#### **--mtmd**=*BOOL*
+Enable or disable multimodal support.
+If enabled and a multimodal projector file is detected, `--mmproj` is
+added to the llama-server command-line.
+When omitted, multimodal support is automatically detected.
+
+[//]: # (END   included file options/mtmd.md)
+
+
 [//]: # (BEGIN included file options/name.md)
 #### **--name**, **-n**
 Name of the container to run the Model in.

docs/ramalama-serve.1.md.in

@@ -98,6 +98,8 @@ appending the path to the type, e.g. `--generate kube:/etc/containers/systemd`.
 
 @@option models-max
 
+@@option mtmd
+
 @@option name
 
 @@option ncmoe

ramalama/plugins/runtimes/inference/llama_cpp.py

@@ -87,6 +87,7 @@ class LlamaCppConfig:
     backend: Literal["auto", "vulkan", "rocm", "cuda", "sycl", "openvino", "cann", "musa"] = "auto"
     cache_reuse: Optional[int] = None
     gguf_quantization_mode: GGUF_QUANTIZATION_MODES = DEFAULT_GGUF_QUANTIZATION_MODE  # type: ignore[assignment]
+    mtmd: Optional[bool] = None
     ngl: Optional[str] = None
     ncmoe: Optional[int] = None
     spec_type: Optional[str] = None
@@ -116,6 +117,8 @@ def __post_init__(self):
         self.threads = int(self.threads)
         if self.thinking is not None:
             self.thinking = coerce_to_bool(self.thinking)
+        if self.mtmd is not None:
+            self.mtmd = coerce_to_bool(self.mtmd)
 
 
 def _positive_int(value: str) -> int:
@@ -506,6 +509,12 @@ def _add_inference_args(self, parser: "argparse.ArgumentParser", command: str) -
                 help="min speculative decoding probability (default: 0.0)",
                 completer=suppressCompleter,
             )
+            parser.add_argument(
+                "--mtmd",
+                default=None,
+                help="enable/disable multimodal mode (default: on)",
+                action=CoerceToBool,
+            )
         self._add_threads_arg(parser)
         if command == "serve":
             parser.add_argument(

ramalama/plugins/runtimes/inference/llama_cpp_commands.py

@@ -72,7 +72,8 @@ def _cmd_run(self, args: argparse.Namespace) -> list[str]:
             model_path = model._get_entry_model_path(is_container, should_generate, dry_run)
             cmd += ["--model", model_path]
 
-            mmproj_path = model._get_mmproj_path(is_container, should_generate, dry_run)
+            mtmd = getattr(args, 'mtmd', None)
+            mmproj_path = model._get_mmproj_path(is_container, should_generate, dry_run) if mtmd is not False else None
             if mmproj_path:
                 cmd += ["--mmproj", str(mmproj_path)]
 

test/unit/test_inference_engine_plugins.py

@@ -51,6 +51,7 @@ def make_ns(
     spec_draft_n_max=None,
     spec_draft_n_min=None,
     spec_draft_p_min=None,
+    mtmd=None,
     runtime_args=None,
     engine_args=None,
     gguf=None,
@@ -80,6 +81,7 @@ def make_ns(
         spec_draft_n_max=spec_draft_n_max,
         spec_draft_n_min=spec_draft_n_min,
         spec_draft_p_min=spec_draft_p_min,
+        mtmd=mtmd,
         runtime_args=runtime_args or [],
         engine_args=engine_args or [],
         gguf=gguf,
@@ -159,6 +161,7 @@ def test_defaults(self):
         assert config.temp == 0.8
         assert config.thinking is None
         assert config.threads > 0
+        assert config.mtmd is None
 
     def test_coerces_string_values(self):
         config = LlamaCppConfig(ngl="4", ncmoe="128", cache_reuse="512", temp="0.5", threads="8", thinking="false")
@@ -338,6 +341,50 @@ def test_serve_thinking_default(self, mock_colorize):
 
         assert "--reasoning" not in cmd
 
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.New")
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.should_colorize", return_value=False)
+    def test_serve_mtmd_disabled(self, mock_colorize, mock_new):
+        mock_model = make_transport_model(mmproj_path="/mnt/models/mmproj.file")
+        mock_new.return_value = mock_model
+
+        ns = make_ns(MODEL="ollama://mymodel", mtmd=False)
+        cmd = self.plugin.handle_subcommand("serve", ns)
+
+        assert "--mmproj" not in cmd
+
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.New")
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.should_colorize", return_value=False)
+    def test_serve_mtmd_enabled(self, mock_colorize, mock_new):
+        mock_model = make_transport_model(mmproj_path="/mnt/models/mmproj.file")
+        mock_new.return_value = mock_model
+
+        ns = make_ns(MODEL="ollama://mymodel", mtmd=True)
+        cmd = self.plugin.handle_subcommand("serve", ns)
+
+        assert "--mmproj" in cmd
+
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.New")
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.should_colorize", return_value=False)
+    def test_serve_mtmd_enabled_without_mmproj(self, mock_colorize, mock_new):
+        mock_model = make_transport_model(mmproj_path=None)
+        mock_new.return_value = mock_model
+
+        ns = make_ns(MODEL="ollama://mymodel", mtmd=True)
+        cmd = self.plugin.handle_subcommand("serve", ns)
+
+        assert "--mmproj" not in cmd
+
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.New")
+    @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.should_colorize", return_value=False)
+    def test_serve_mtmd_default(self, mock_colorize, mock_new):
+        mock_model = make_transport_model(mmproj_path="/mnt/models/mmproj.file")
+        mock_new.return_value = mock_model
+
+        ns = make_ns(MODEL="ollama://mymodel", mtmd=None)
+        cmd = self.plugin.handle_subcommand("serve", ns)
+
+        assert "--mmproj" in cmd
+
     @patch("ramalama.plugins.runtimes.inference.llama_cpp_commands.should_colorize", return_value=False)
     def test_serve_ctx_size(self, mock_colorize):
         ns = make_ns(ctx_size=4096)