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

Jump to bottom

🐛 [Bug] #2881 regression #2991

Closed
HolyWu opened this issue Jul 9, 2024 · 8 comments
Closed

🐛 [Bug] #2881 regression #2991

HolyWu opened this issue Jul 9, 2024 · 8 comments
Labels
bug Something isn't working

Comments

@HolyWu
Copy link
Contributor

HolyWu commented Jul 9, 2024

Bug Description

Since #2881, if the inference is performed in its own stream, the output randomly becomes all zeros.

cc: @gs-olive

To Reproduce

import torch
import torch.nn as nn
import torch_tensorrt


class MyModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.relu = nn.ReLU()

    def forward(self, x):
        return self.relu(x)


with torch.inference_mode():
    dtype = torch.half
    device = torch.device("cuda", 0)
    model = MyModule().eval().to(device)
    inputs = [torch_tensorrt.Input(shape=(1, 3, 5), dtype=dtype)]

    optimized_model = torch_tensorrt.compile(
        model,
        ir="dynamo",
        inputs=inputs,
        enabled_precisions={dtype},
        min_block_size=1,
        device=device,
    )

    for _ in range(10):
        new_input = torch.randn((1, 3, 5), dtype=dtype, device=device)

        eager_output = model(new_input)

        stream = torch.cuda.Stream(device=device)
        stream.wait_stream(torch.cuda.current_stream(device=device))
        with torch.cuda.stream(stream):
            trt_output_with_stream = optimized_model(new_input)
        torch.cuda.current_stream(device=device).wait_stream(stream)

        trt_output_without_stream = optimized_model(new_input)

        print("")
        print(f"{torch.allclose(eager_output, trt_output_with_stream)=}")
        print(f"{torch.allclose(eager_output, trt_output_without_stream)=}")
        print(f"{trt_output_with_stream=}")
        print(f"{trt_output_without_stream=}")
torch.allclose(eager_output, trt_output_with_stream)=True
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[1.4609, 1.7363, 2.2363, 0.5176, 1.1904],
         [0.0000, 0.0000, 0.9243, 0.4497, 0.0000],
         [0.3186, 0.0000, 0.0000, 0.2537, 0.0707]]], device='cuda:0',
       dtype=torch.float16)
trt_output_without_stream=tensor([[[1.4609, 1.7363, 2.2363, 0.5176, 1.1904],
         [0.0000, 0.0000, 0.9243, 0.4497, 0.0000],
         [0.3186, 0.0000, 0.0000, 0.2537, 0.0707]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=False
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]], device='cuda:0', dtype=torch.float16)
trt_output_without_stream=tensor([[[0.6279, 0.6396, 0.9829, 0.2437, 0.0855],
         [0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [0.9268, 0.2942, 0.1624, 0.2323, 0.5444]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=False
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]], device='cuda:0', dtype=torch.float16)
trt_output_without_stream=tensor([[[0.4104, 0.9424, 0.9624, 0.0000, 0.6167],
         [0.2534, 0.2571, 0.0000, 0.0000, 0.0000],
         [0.4695, 0.0136, 0.9429, 0.2498, 0.6011]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=False
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]], device='cuda:0', dtype=torch.float16)
trt_output_without_stream=tensor([[[0.3899, 0.0000, 0.0000, 0.0000, 0.0000],
         [1.1504, 1.1680, 1.5703, 0.7173, 0.0000],
         [0.4333, 0.1777, 2.0723, 1.0098, 0.0000]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=True
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0.0000, 0.0000, 0.0000, 0.0000, 1.1924],
         [0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [1.0889, 0.0027, 0.0000, 1.1680, 0.9487]]], device='cuda:0',
       dtype=torch.float16)
trt_output_without_stream=tensor([[[0.0000, 0.0000, 0.0000, 0.0000, 1.1924],
         [0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [1.0889, 0.0027, 0.0000, 1.1680, 0.9487]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=False
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]], device='cuda:0', dtype=torch.float16)
trt_output_without_stream=tensor([[[0.0000, 0.0765, 0.4006, 0.0000, 2.0410],
         [2.2793, 0.0000, 0.0000, 0.0000, 0.0000],
         [1.6064, 0.0000, 0.6040, 0.4290, 0.0000]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=False
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]], device='cuda:0', dtype=torch.float16)
trt_output_without_stream=tensor([[[0.8057, 0.9116, 0.3110, 0.0000, 0.0000],
         [1.6016, 0.0000, 0.3757, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0048, 0.9331, 1.1475]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=False
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]], device='cuda:0', dtype=torch.float16)
trt_output_without_stream=tensor([[[0.0000, 0.7251, 0.0000, 0.0000, 0.9985],
         [0.0000, 1.3789, 0.6831, 0.0000, 0.7051],
         [0.0000, 0.0000, 0.0000, 0.6748, 0.0000]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=True
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000, 0.0000, 1.0420],
         [0.3943, 0.0000, 0.0000, 0.5200, 1.4277]]], device='cuda:0',
       dtype=torch.float16)
trt_output_without_stream=tensor([[[0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
         [0.0000, 0.0000, 0.0000, 0.0000, 1.0420],
         [0.3943, 0.0000, 0.0000, 0.5200, 1.4277]]], device='cuda:0',
       dtype=torch.float16)

torch.allclose(eager_output, trt_output_with_stream)=False
torch.allclose(eager_output, trt_output_without_stream)=True
trt_output_with_stream=tensor([[[0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.],
         [0., 0., 0., 0., 0.]]], device='cuda:0', dtype=torch.float16)
trt_output_without_stream=tensor([[[1.1484, 0.0000, 0.0000, 0.0000, 0.0000],
         [0.0000, 0.5659, 0.0000, 1.3525, 0.5669],
         [1.7070, 0.0000, 2.0957, 0.0000, 1.1602]]], device='cuda:0',
       dtype=torch.float16)

Environment

  • Torch-TensorRT Version (e.g. 1.0.0): 2.5.0.dev20240709+cu124
  • PyTorch Version (e.g. 1.0): 2.5.0.dev20240709+cu124
  • CPU Architecture: x86_64
  • OS (e.g., Linux): Ubuntu 24.04 LTS
  • How you installed PyTorch (conda, pip, libtorch, source): pip
  • Build command you used (if compiling from source):
  • Are you using local sources or building from archives:
  • Python version: 3.12
  • CUDA version: 12.4
  • GPU models and configuration: RTX 3050
  • Any other relevant information:

Additional context

Interestingly, it's only reproducible when using dtype=torch.half, but not for dtype=torch.float.
@HolyWu HolyWu added the bug Something isn't working label Jul 9, 2024
@narendasan
Copy link
Collaborator

@HolyWu what driver version are you using?

@HolyWu
Copy link
Contributor Author

HolyWu commented Jul 9, 2024

I'm using GeForce Game Ready driver 555.99 (I run Ubuntu under WSL 2). Note that the CI (https://github.com/pytorch/TensorRT/actions/runs/9848088305/job/27193531360) also has failure, but I'm not sure whether it's related to my issue.

@peri044
Copy link
Collaborator

peri044 commented Jul 12, 2024

I couldn't reproduce this issue on my RTX 3080 Ti with driver 545.29.06. I have a hunch though.
@HolyWu could you try the following and let me know if you're still seeing the issue

  1. Use Python runtime by specifying use_python_runtime=True in the compilation setting
  2. Comment out

@narendasan
Copy link
Collaborator

Just as an additional datapoint:
latest main, latest pyt, Driver Version: 550.90.07 (Ubuntu Native) CUDA Version: 12.4 Titan V
No issues over 10000 iterations

Do let us know if the python runtime patch works, but I'm inclined to think its related to the WSL driver

@HolyWu
Copy link
Contributor Author

HolyWu commented Jul 13, 2024

I had run my test on Windows and Ubuntu-in-WSL, and both had the same issue. Hence it's probably a quirk on Windows and has nothing to do with the WSL driver.

Using use_python_runtime=True alone without modifying _PythonTorchTensorRTModule.py can already produce correct result. Ultimately, I found that adding the following codes above this line can make C++ runtime produce correct result for me.

    auto current_stream = c10::cuda::getCurrentCUDAStream(inputs[0].device().index());
    if (compiled_engine->active_stream != current_stream) {
      compiled_engine->active_stream = current_stream;
    }

@HolyWu
Copy link
Contributor Author

HolyWu commented Jul 15, 2024

Hmm...I managed to reproduce this issue on Google Colab. But I had to use a tensor with big enough size like (1, 3, 2160, 3840) and also a few more iterations to make it happen.

@lanluo-nvidia
Copy link
Collaborator

@HolyWu @narendasan @peri044
I had tried on both windows and Linux, both in the local and GitHub workflow, I was not able to reproduce the issue.
I also tried both use_python_run_time=True and False, it is no issues.

Here is the workflow run with use_python_run_time=False for linux:
https://github.com/pytorch/TensorRT/actions/runs/9943098674/job/27466803783
Here is the workflow run with use_python_run_time=False for window:
https://github.com/pytorch/TensorRT/actions/runs/9943098676/job/27467208309

Let me also try on the big enough size then.

@HolyWu
Copy link
Contributor Author

HolyWu commented Jul 21, 2024

@narendasan @peri044 @lanluo-nvidia
Just wanted to make sure if any of you could reproduce it with big size tensor? I thought it's not Windows-specific now since it's also reproducible on Google Colab.

@HolyWu HolyWu closed this as not planned Won't fix, can't repro, duplicate, stale Jul 27, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
bug Something isn't working
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
4 participants
@narendasan @HolyWu @peri044 @lanluo-nvidia