Upload handler.py

handler.py

@@ -9,7 +9,7 @@ import torch
 from torchao.quantization import quantize_, autoquant, int8_dynamic_activation_int8_weight
 from huggingface_hub import hf_hub_download
 
-IS_COMPILE = False
+IS_COMPILE = True
 IS_TURBO = False
 
 if IS_COMPILE:
@@ -34,9 +34,9 @@ def load_pipeline_compile(repo_id: str, dtype: torch.dtype) -> Any:
     pipe.transformer.fuse_qkv_projections()
     pipe.vae.fuse_qkv_projections()
     pipe.transformer.to(memory_format=torch.channels_last)
-    pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="inductor")
+    pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="eager")
     pipe.vae.to(memory_format=torch.channels_last)
-    pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="inductor")
+    pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="eager")
     pipe.to("cuda")
     return pipe
 
@@ -45,9 +45,9 @@ def load_pipeline_autoquant(repo_id: str, dtype: torch.dtype) -> Any:
     pipe.transformer.fuse_qkv_projections()
     pipe.vae.fuse_qkv_projections()
     pipe.transformer.to(memory_format=torch.channels_last)
-    pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
+    pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True, backend="eager")
     pipe.vae.to(memory_format=torch.channels_last)
-    pipe.vae = torch.compile(pipe.vae, mode="max-autotune", fullgraph=True)
+    pipe.vae = torch.compile(pipe.vae, mode="max-autotune", fullgraph=True, backend="eager")
     pipe.transformer = autoquant(pipe.transformer, error_on_unseen=False)
     pipe.vae = autoquant(pipe.vae, error_on_unseen=False)
     pipe.to("cuda")
@@ -75,9 +75,9 @@ def load_pipeline_turbo_compile(repo_id: str, dtype: torch.dtype) -> Any:
     quantize_(pipe.transformer, int8_dynamic_activation_int8_weight(), device="cuda")
     quantize_(pipe.vae, int8_dynamic_activation_int8_weight(), device="cuda")
     pipe.transformer.to(memory_format=torch.channels_last)
-    pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="inductor")
+    pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="eager")
     pipe.vae.to(memory_format=torch.channels_last)
-    pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="inductor")
+    pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=False, dynamic=False, backend="eager")
     pipe.to("cuda")
     return pipe