use_precompiled = True

optimization.py

@@ -22,6 +22,10 @@ from optimization_utils import ZeroGPUCompiledModel
 
 P = ParamSpec('P')
 
+# Expose compiled models so app.py can offer them for download
+COMPILED_TRANSFORMER_1 = None
+COMPILED_TRANSFORMER_2 = None
+
 LATENT_FRAMES_DIM = torch.export.Dim('num_latent_frames', min=8, max=81)
 LATENT_PATCHED_HEIGHT_DIM = torch.export.Dim('latent_patched_height', min=30, max=52)
 LATENT_PATCHED_WIDTH_DIM = torch.export.Dim('latent_patched_width', min=30, max=52)
@@ -79,6 +83,7 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
       Si ça échoue (format incompatible, etc.), fallback compilation GPU.
     - Sinon -> compile sur GPU.
     """
+    global COMPILED_TRANSFORMER_1, COMPILED_TRANSFORMER_2
 
     @spaces.GPU(duration=1500)
     def compile_transformer():
@@ -128,7 +133,7 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
     # Quantize text encoder
     quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
 
-    use_precompiled = _strtobool(os.getenv("WAN_USE_PRECOMPILED"), default=False)
+    use_precompiled = True
     precompiled_repo = os.getenv("WAN_PRECOMPILED_REPO", "Fabrice-TIERCELIN/Wan_2.2_compiled")
 
     compiled_transformer_1 = None
@@ -147,6 +152,10 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
     else:
         compiled_transformer_1, compiled_transformer_2 = compile_transformer()
 
+    # expose for downloads
+    COMPILED_TRANSFORMER_1 = compiled_transformer_1
+    COMPILED_TRANSFORMER_2 = compiled_transformer_2
+
     pipeline.transformer.forward = compiled_transformer_1
     drain_module_parameters(pipeline.transformer)