Load constants

app.py

@@ -15,6 +15,8 @@ import gradio as gr
 import spaces
 import torch
 import torch._inductor
+from torch.export.pt2_archive._package import AOTICompiledModel
+from torch.export.pt2_archive._package_weights import Weights
 from torchvision.models import ResNet18_Weights, resnet18
 
 
@@ -33,24 +35,29 @@ def compile_model():
             model,
             example_inputs,
         )
-        torch._inductor.aoti_compile_and_package(
-            exported_program,
-            package_path=package_path,
-            inductor_configs=inductor_configs
-        )
-    return "compiled"
+        artifacts = torch._inductor.aot_compile(exported_program.module(), *exported_program.example_inputs, options={
+            'aot_inductor.package_constants_in_so': False,
+            'aot_inductor.package_constants_on_disk': True,
+            'aot_inductor.package': True,
+            'max_autotune': True,
+        })
+    weights, = (artifact for artifact in artifacts if isinstance(artifact, Weights))
+    weights_: dict[str, torch.Tensor] = {}
+    for name in weights:
+        tensor, _properties = weights.get_weight(name)
+        tensor_ = torch.empty_like(tensor, device='cpu').pin_memory()
+        weights_[name] = tensor_.copy_(tensor).share_memory_()
+    return weights_
+
+weights = compile_model()
+del model
 
 @spaces.GPU
 def run_model():
-    compiled_model = torch._inductor.aoti_load_package(package_path)
+    compiled_model: AOTICompiledModel = torch._inductor.aoti_load_package(package_path)
+    compiled_model.load_constants(weights, check_full_update=True, user_managed=True)
     with torch.inference_mode():
         return str(compiled_model(example_inputs))
 
 
-gr.TabbedInterface([
-    gr.Interface(compile_model, [], "text", clear_btn=None, flagging_mode='never'),
-    gr.Interface(run_model, [], "text", clear_btn=None, flagging_mode='never'),
-], [
-    "Compile",
-    "Run"
-]).launch(show_error=True)
+gr.Interface(run_model, [], 'text').launch(show_error=True)