Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix for warning as default stream was used in enqueueV3 #3191

Merged
merged 2 commits into from
Oct 15, 2024
Merged

Fix for warning as default stream was used in enqueueV3 #3191

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
691cf1e
chore: set stream when new stream is acquired
keehyuna Sep 27, 2024
134ffc6
chore: Keep using created engine stream
keehyuna Oct 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 0 additions & 2 deletions core/runtime/TRTEngine.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,8 +73,6 @@ TRTEngine::TRTEngine(
<< get_current_platform() << ")");
this->target_platform = target_platform;

this->cudagraph_mempool_id = at::cuda::graph_pool_handle();

this->hardware_compatible = hardware_compatible;
auto most_compatible_device = get_most_compatible_device(cuda_device, RTDevice(), hardware_compatible);
TORCHTRT_CHECK(most_compatible_device, "No compatible device was found for instantiating TensorRT engine");
Expand Down
1 change: 0 additions & 1 deletion core/runtime/TRTEngine.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,7 +81,6 @@ struct TRTEngine : torch::CustomClassHolder {
std::vector<at::Tensor> input_buffers = {};
std::vector<at::Tensor> output_buffers = {};
std::string shape_key;
at::cuda::MempoolId_t cudagraph_mempool_id;

// TODO: Implement a call method
// c10::List<at::Tensor> Run(c10::List<at::Tensor> inputs);
Expand Down
4 changes: 1 addition & 3 deletions core/runtime/execute_engine.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -305,8 +305,6 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
if (compiled_engine->engine_stream == c10::cuda::getDefaultCUDAStream(current_device_id)) {
// Create a new stream if the engine stream is the default stream
compiled_engine->engine_stream = c10::cuda::getStreamFromPool(false, current_device_id);
} else {
compiled_engine->engine_stream = compiled_engine->caller_stream;
}

// nvinfer1::IExecutionContext::enqueue is not thread safe and we need a mutex for it.
Expand All @@ -333,7 +331,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
if (need_cudagraphs_record) {
// If cudagraphs needs to record a graph, capture the enqueueV3 call in a graph
c10::cuda::CUDAStream recording_stream = compiled_engine->engine_stream;
compiled_engine->cudagraph.capture_begin(compiled_engine->cudagraph_mempool_id);
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This was reverted to fix below assert() from torch. We don't share memory across captures, I think we can use internally created pool.

https://pytorch.org/docs/stable/notes/cuda.html#graph-memory-management

  File "/root/trt/TensorRT/py/torch_tensorrt/dynamo/runtime/_TorchTensorRTModule.py", line 274, in forward
    outputs: List[torch.Tensor] = torch.ops.tensorrt.execute_engine(
  File "/root/.pyenv/versions/3.10.14/lib/python3.10/site-packages/torch/_ops.py", line 1113, in __call__
    return self._op(*args, **(kwargs or {}))
RuntimeError: it->second->use_count > 0 INTERNAL ASSERT FAILED at "../c10/cuda/CUDACachingAllocator.cpp":2056, please report a bug to PyTorch

compiled_engine->cudagraph.capture_begin();
compiled_engine->exec_ctx->enqueueV3(recording_stream);
compiled_engine->cudagraph.capture_end();

Expand Down
8 changes: 3 additions & 5 deletions py/torch_tensorrt/dynamo/runtime/_PythonTorchTensorRTModule.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
from tempfile import tempdir
from typing import Any, Dict, List, Optional, Sequence, Tuple

import tensorrt as trt
import torch
import torch_tensorrt
from torch.nn import Module
Expand All @@ -19,8 +20,6 @@
multi_gpu_device_check,
)

import tensorrt as trt

logger = logging.getLogger(__name__)


Expand Down Expand Up @@ -372,8 +371,6 @@ def forward(self, *inputs: torch.Tensor) -> torch.Tensor | Tuple[torch.Tensor, .
or self._engine_stream is None
):
self._engine_stream = torch.cuda.Stream()
else:
self._engine_stream = self._caller_stream

self._engine_stream.wait_stream(self._caller_stream)

Expand Down Expand Up @@ -464,7 +461,8 @@ def cudagraphs_validate_shapes(self, inputs: Sequence[torch.Tensor]) -> bool:
if new_shape_key != self.shape_key:
logger.debug(f"Resetting Cudagraph on new shape key {new_shape_key}")
self.shape_key = new_shape_key
self.cudagraph.reset() # type: ignore
if self.cudagraph:
self.cudagraph.reset()
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

self.cudagraph can be None when torch_compile backend is used.

self.cudagraph is initialized when cudagraphs mode is enabled. But this init was called at compile()
https://github.com/pytorch/TensorRT/blob/main/py/torch_tensorrt/dynamo/runtime/_PythonTorchTensorRTModule.py#L144-L145

return False

return True
Loading