Fix, doc

Author: BenjaminBossan

tests/pipelines/run_compiled_model_hotswap.py

@@ -123,7 +123,7 @@ def check_hotswap(do_hotswap):
         unet = get_small_unet()
         file_name = os.path.join(tmp_dirname, "pytorch_lora_weights.safetensors")
         unet.load_attn_procs(file_name)
-        # unet = torch.compile(unet, mode="reduce-overhead")
+        unet = torch.compile(unet, mode="reduce-overhead")
 
         torch.manual_seed(42)
         out0 = unet(**dummy_input)["sample"]

tests/pipelines/test_pipelines.py

@@ -2070,6 +2070,9 @@ class TestLoraHotSwapping:
     tested there. The goal of this test is specifically to ensure that hotswapping with diffusers does not require
     recompilation.
 
+    The reason why we need to shell out instead of just running the script inside of the test is that shelling out is
+    required to collect the torch.compile logs.
+
     """
 
     @slow