Different efficiencies in windows and wsl

[showarp]

I deployed EfficientTAM using WSL2 to run with good results, but I tried to deploy it to windows but I encountered problems, first of all windows doesn't support triton causing the runtime to make an error report, then I went to look for the community implementation of triton and eventually the program runs, but the efficiency is not good, I use efficienttam_s_512x512.pt to get only 3 frames per second on a RTX3060 12GB, have you guys tried to experiment on windows?

Below is a variety of information about my system:

nvidia-smi

nvcc -V

pip list

aiofiles 23.2.1
annotated-types 0.7.0
antlr4-python3-runtime 4.9.3
anyio 4.8.0
asttokens 3.0.0
certifi 2025.1.31
charset-normalizer 3.4.1
click 8.1.8
colorama 0.4.6
comm 0.2.2
contourpy 1.3.1
cycler 0.12.1
debugpy 1.8.11
decorator 4.4.2
defusedxml 0.7.1
efficient_track_anything 1.0 D:\YstSpace\EfficientTAM
executing 2.1.0
fastapi 0.115.8
ffmpy 0.5.0
filelock 3.17.0
fonttools 4.56.0
fsspec 2025.2.0
gradio 4.44.0
gradio_client 1.3.0
gradio_image_prompter 0.1.0
h11 0.14.0
httpcore 1.0.7
httpx 0.28.1
huggingface-hub 0.26.2
hydra-core 1.3.2
idna 3.10
imageio 2.9.0
imageio-ffmpeg 0.5.1
importlib_metadata 8.6.1
importlib_resources 6.5.2
iopath 0.1.10
ipykernel 6.29.5
ipython 8.30.0
jedi 0.19.2
Jinja2 3.1.5
jupyter_client 8.6.3
jupyter_core 5.7.2
kiwisolver 1.4.8
markdown-it-py 3.0.0
MarkupSafe 2.1.5
matplotlib 3.10.0
matplotlib-inline 0.1.7
mdurl 0.1.2
moviepy 1.0.3
mpmath 1.3.0
nest_asyncio 1.6.0
networkx 3.4.2
numpy 2.2.3
omegaconf 2.3.0
opencv-python 4.11.0.86
orjson 3.10.15
packaging 24.2
pandas 2.2.3
parso 0.8.4
pillow 10.4.0
pip 25.0
platformdirs 4.3.6
portalocker 3.1.1
proglog 0.1.10
prompt_toolkit 3.0.50
psutil 5.9.0
pure_eval 0.2.3
pydantic 2.10.6
pydantic_core 2.27.2
pydub 0.25.1
Pygments 2.19.1
pyparsing 3.2.1
python-dateutil 2.9.0.post0
python-multipart 0.0.20
pytz 2025.1
pywin32 308
PyYAML 6.0.2
pyzmq 24.0.1
requests 2.32.3
rich 13.9.4
ruff 0.9.6
safehttpx 0.1.6
scipy 1.15.1
semantic-version 2.10.0
setuptools 75.8.0
shellingham 1.5.4
six 1.17.0
sniffio 1.3.1
stack_data 0.6.3
starlette 0.45.3
supervision 0.25.0
sympy 1.13.1
tomlkit 0.12.0
torch 2.6.0+cu124
torchaudio 2.6.0+cu124
torchvision 0.21.0+cu124
tornado 6.2
tqdm 4.67.1
traitlets 5.14.3
triton 3.2.0
typer 0.15.1
typing_extensions 4.12.2
tzdata 2025.1
urllib3 2.3.0
uvicorn 0.34.0
wcwidth 0.2.13
websockets 12.0
wheel 0.45.1
zipp 3.21.0

[rodneyML]

Did you use benchmark.py to run EfficientTAM? @showarp

They key that solved that issue for me was to build the predictor like this:

predictor = build_efficienttam_video_predictor(
        model_cfg, 
        model_checkpoint, 
        device=torch.device("cuda"), 
        vos_optimized=False,                                             # Set vos_optimized to false
        hydra_overrides_extra=["++model.compile_image_encoder=False"]    # Add this argument
)

Note that by using this method, triton wont be used. When I run the s512x512 model, I get about 14 FPS on a RTX 4050 Laptop GPU.

[showarp]

Did you use benchmark.py to run EfficientTAM? @showarp

They key that solved that issue for me was to build the predictor like this:

predictor = build_efficienttam_video_predictor(
model_cfg,
model_checkpoint,
device=torch.device("cuda"),
vos_optimized=False, # Set vos_optimized to false
hydra_overrides_extra=["++model.compile_image_encoder=False"] # Add this argument
)
Note that by using this method, triton wont be used. When I run the s512x512 model, I get about 14 FPS on a RTX 4050 Laptop GPU.

Thank you very much! I used your method to run benchmark.py on Windows and got to 50 frames per second. You saved me!

And I found that using triton-windows under windows and building the following predictor

predictor = build_efficienttam_video_predictor(
        model_cfg, 
        model_checkpoint, 
        device=torch.device("cuda"), 
        vos_optimized=True,
)

Benchmark scores are only 2 - 3fps
If you add hydra_overrides_extra=["++model.compile_image_encoder=False"], you will get a score of 50FPS (I guess this is how triton-windows really works). If you set vos_optimized=False, the number of frames will only be about 20FPS.
Anyway, thank you very much for your help!

[rodneyML]

Glad I could help! 😎 And thanks for the tip to use vos_optimized=True, I am however having trouble to use vos_optimized=True, because of some torch._dynamo errors:

torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
TypeError: must be called with a dataclass type or instance

[yformer]

Did you use benchmark.py to run EfficientTAM? @showarp
They key that solved that issue for me was to build the predictor like this:
predictor = build_efficienttam_video_predictor(
model_cfg,
model_checkpoint,
device=torch.device("cuda"),
vos_optimized=False, # Set vos_optimized to false
hydra_overrides_extra=["++model.compile_image_encoder=False"] # Add this argument
)
Note that by using this method, triton wont be used. When I run the s512x512 model, I get about 14 FPS on a RTX 4050 Laptop GPU.

Thank you very much! I used your method to run benchmark.py on Windows and got to 50 frames per second. You saved me!

And I found that using triton-windows under windows and building the following predictor

predictor = build_efficienttam_video_predictor(
model_cfg,
model_checkpoint,
device=torch.device("cuda"),
vos_optimized=True,
)
Benchmark scores are only 2 - 3fps If you add hydra_overrides_extra=["++model.compile_image_encoder=False"], you will get a score of 50FPS (I guess this is how triton-windows really works). If you set vos_optimized=False, the number of frames will only be about 20FPS. Anyway, thank you very much for your help!

@showarp, glad to see you use the flag for enabling triton on Windows.