fix: requesting gpu from docker runtime only for loaders that need it

dyana/cli.py

@@ -98,8 +98,8 @@ def trace(
 ) -> None:
     try:
         # disable GPU on non-Linux systems
-        if not no_gpu and platform_pkg.system() != "Linux":
-            no_gpu = True
+        # if not no_gpu and platform_pkg.system() != "Linux":
+        #    no_gpu = True
 
         # check if policy is either a file or a directory
         if policy and not policy.exists():

dyana/loaders/automodel/settings.yml

@@ -1,5 +1,6 @@
 description: Loads and profiles machine learning models compatible with AutoModel and AutoTokenizer.
 
+gpu: true
 build_args:
   extra-requirements: EXTRA_REQUIREMENTS
 

dyana/loaders/loader.py

@@ -220,7 +220,7 @@ def run(self, allow_network: bool = False, allow_gpus: bool = True, allow_volume
                 volumes,
                 environment=environment,
                 allow_network=allow_network,
-                allow_gpus=allow_gpus,
+                allow_gpus=allow_gpus and self.settings.gpu,
                 allow_volume_write=allow_volume_write,
             )
             self.container_id = self.container.id

dyana/loaders/lora/settings.yml

@@ -1,5 +1,7 @@
 description: Loads LoRA adapters via PEFT.
 
+gpu: true
+
 build_args:
   extra-requirements: EXTRA_REQUIREMENTS
 

dyana/loaders/ollama/settings.yml

@@ -1,5 +1,7 @@
 description: Loads and profiles models via an Ollama server. Local models on the host machine are mounted and shared with the container.
 
+gpu: true
+
 volumes:
   # on macOS
   - host: ~/.ollama/models

dyana/loaders/settings.py

@@ -32,6 +32,7 @@ class LoaderSettings(BaseModel):
     build_args: dict[str, str] | None = None
     args: list[LoaderArgument] | None = None
     network: bool | None = False
+    gpu: bool = False
     volumes: list[Volume] | None = None
     examples: list[Example] | None = None