fix: Destory cuda graphs before setting weight streaming (#3461)

Author: keehyuna

core/runtime/TRTEngine.cpp

@@ -453,6 +453,10 @@ std::vector<std::string> TRTEngine::serialize() {
   return serialized_info;
 }
 
+void TRTEngine::reset_captured_graph() {
+  cudagraph.reset();
+}
+
 } // namespace runtime
 } // namespace core
 } // namespace torch_tensorrt

core/runtime/TRTEngine.h

@@ -185,6 +185,7 @@ struct TRTEngine : torch::CustomClassHolder {
   // c10::List<at::Tensor> Run(c10::List<at::Tensor> inputs);
 
   void set_profiling_paths();
+  void reset_captured_graph();
 #ifndef NDEBUG
   bool profile_execution = true;
 #else

core/runtime/register_jit_hooks.cpp

@@ -88,6 +88,7 @@ static auto TORCHTRT_UNUSED TRTEngineTSRegistrtion =
         .def("dump_engine_layer_info", &TRTEngine::dump_engine_layer_info)
         .def("get_engine_layer_info", &TRTEngine::get_engine_layer_info)
         .def("infer_outputs", &TRTEngine::infer_outputs)
+        .def("reset_captured_graph", &TRTEngine::reset_captured_graph)
         .def_readwrite("use_pre_allocated_outputs", &TRTEngine::use_pre_allocated_outputs)
         .def_readwrite("use_output_allocator_outputs", &TRTEngine::use_output_allocator_outputs)
         .def_property(

py/torch_tensorrt/dynamo/runtime/_CudaGraphsTorchTensorRTModule.py

@@ -103,9 +103,13 @@ def validate_input_shapes(self, inputs: Sequence[torch.Tensor]) -> bool:
 
         return False
 
-    def __del__(self) -> None:
+    def _reset_captured_graph(self) -> None:
         if self.cudagraph:
             self.cudagraph.reset()
+            self.cudagraph = None
+
+    def __del__(self) -> None:
+        self._reset_captured_graph()
 
     def set_use_output_allocator(self, enable: bool) -> None:
         self.use_output_allocator_outputs = enable
@@ -119,8 +123,7 @@ def forward(
             shape_changed = self.validate_input_shapes(inputs)
             need_cudagraphs_record = shape_changed or self.is_weight_streaming_set
             if need_cudagraphs_record:
-                if self.cudagraph:
-                    self.cudagraph.reset()
+                self._reset_captured_graph()
                 self._input_buffers = [None] * len(inputs)
 
             self.is_weight_streaming_set = False
@@ -196,7 +199,5 @@ def forward(
                 return outputs[0]
             return outputs
         else:
-            if self.cudagraph:
-                self.cudagraph.reset()
-                self.cudagraph = None
+            self._reset_captured_graph()
             return self.compiled_module(*args, **kwargs)

py/torch_tensorrt/dynamo/runtime/_PythonTorchTensorRTModule.py

@@ -333,9 +333,13 @@ def __deepcopy__(self, memo: Any) -> PythonTorchTensorRTModule:
         result.__setstate__(self.__getstate__())
         return result
 
-    def __del__(self) -> None:
+    def _reset_captured_graph(self) -> None:
         if self.cudagraph:
             self.cudagraph.reset()
+            self.cudagraph = None
+
+    def __del__(self) -> None:
+        self._reset_captured_graph()
 
     def setup_input_tensors(
         self,
@@ -426,9 +430,8 @@ def run_standard_execution() -> torch.Tensor | Tuple[torch.Tensor, ...]:
                 self.cudagraphs_enabled, self.use_pre_allocated_outputs, shape_changed
             )
 
-            if need_cudagraphs_reset and self.cudagraph:
-                self.cudagraph.reset()
-                self.cudagraph = None
+            if need_cudagraphs_reset:
+                self._reset_captured_graph()
 
             if need_cudagraphs_record:
                 self._input_buffers = [None] * len(self.input_names)

py/torch_tensorrt/dynamo/runtime/_TorchTensorRTModule.py

@@ -209,6 +209,9 @@ def set_device_memory_budget(self, budget_bytes: int) -> int:
 
         return budget_bytes
 
+    def _reset_captured_graph(self) -> None:
+        self.engine.reset_captured_graph()
+
     def setup_engine(self) -> None:
         """
         Setup engine for a module which has deferred engine setup.

py/torch_tensorrt/dynamo/runtime/meta_ops/register_meta_ops.py

@@ -142,6 +142,9 @@ def automatic_device_memory_budget_getter(self) -> Any:
     def infer_outputs(self, input_shapes: List[Any]) -> Any:
         pass
 
+    def reset_captured_graph(self) -> Any:
+        pass
+
     def __setstate__(self, serialized_state: List[str]) -> Any:
         pass
 

py/torch_tensorrt/runtime/_cudagraphs.py

@@ -1,5 +1,5 @@
 import logging
-from typing import Any, Union
+from typing import Any, Optional, Union
 
 import torch
 import torch_tensorrt
@@ -68,6 +68,7 @@ def __init__(self, compiled_module: torch.nn.Module) -> None:
         global _PY_RT_CUDAGRAPHS
         self.old_mode = _PY_RT_CUDAGRAPHS
         self.compiled_module = compiled_module
+        self.cudagraphs_module: Optional[CudaGraphsTorchTensorRTModule] = None
 
     def __enter__(self) -> torch.nn.Module:
         global _PY_RT_CUDAGRAPHS
@@ -98,7 +99,8 @@ def __enter__(self) -> torch.nn.Module:
             logger.debug(
                 "Found pytorch subgraphs in module, wrapping module in CudaGraphsTorchTensorRTModule"
             )
-            return CudaGraphsTorchTensorRTModule(self.compiled_module)
+            self.cudagraphs_module = CudaGraphsTorchTensorRTModule(self.compiled_module)
+            return self.cudagraphs_module
         else:
             if num_trt_module > 0:
                 logger.debug("No graph breaks detected, using runtime cudagraphs mode")
@@ -113,6 +115,9 @@ def __enter__(self) -> torch.nn.Module:
     def __exit__(self, *args: Any) -> None:
         # Set cudagraphs back to old mode
         set_cudagraphs_mode(self.old_mode)
+        # __del__ is not entirely predictable, so we reset cudagraph here
+        if self.cudagraphs_module:
+            self.cudagraphs_module._reset_captured_graph()
 
 
 def enable_cudagraphs(

py/torch_tensorrt/runtime/_weight_streaming.py

@@ -76,12 +76,15 @@ def _set_streamable_weight_bytes(self, requested_budget: int) -> int:
             int(streamable_bytes / total_bytes * requested_budget)
             for streamable_bytes in self.streamable_budget
         ]
+        if self.cuda_graphs_module:
+            self.cuda_graphs_module.is_weight_streaming_set = True
+            self.cuda_graphs_module._reset_captured_graph()
+
         for i, (name, rt_mod) in enumerate(self.rt_mods):
+            rt_mod._reset_captured_graph()
             ws_budget_bytes += rt_mod.set_device_memory_budget(normalized_size[i])
             logger.debug(f"Set weight streaming size {normalized_size[i]} for {name}")
 
-        if self.cuda_graphs_module:
-            self.cuda_graphs_module.is_weight_streaming_set = True
         return ws_budget_bytes
 
     def __setattr__(self, name: str, value: Any) -> None: