Make it optional to build CUDA extension for SAM 2; also fallback to all available kernels if Flash Attention fails #155
+173
−33
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ronghanghu
force-pushed
the
ronghanghu/cuda_kernel_optional
branch
3 times, most recently
from
509f0b1
to
268ad1c
Compare
ronghanghu
changed the title
ronghanghu
force-pushed
the
ronghanghu/cuda_kernel_optional
branch
3 times, most recently
from
8522a19
to
6943cf6
Compare
|
Do you think that Kornia like pure pytorch connected components will be too much numerically misaligned? |
@bhack Thanks for the suggestion! We have also tried this kornia implementation before, but it was too slow for video applications (as it's using an iteration loop in Python and its algorithm has not been carefully optimized for GPUs), so we added a custom CUDA kernel in connected_components.cu instead, which is much faster. |
…all available kernels if Flash Attention fails In this PR, we make it optional to build the SAM 2 CUDA extension, in observation that many users encounter difficulties with the CUDA compilation step. 1. During installation, we catch build errors and print a warning message. We also allow explicitly turning off the CUDA extension building with `SAM2_BUILD_CUDA=0`. 2. At runtime, we catch CUDA kernel errors from connected components and print a warning on skipping the post processing step. We also fall back to the all available kernels if the Flash Attention kernel fails.
ronghanghu
force-pushed
the
ronghanghu/cuda_kernel_optional
branch
from
6943cf6
to
1757177
Compare
|
Yes I know that it has loops. It is not easy to implement with pytorch ops. Have you benchmarked how is the pytorch compiler behaving with these loops? |
|
Quite funny that... |
haithamkhedr
approved these changes
Aug 6, 2024
@bhack In our internal benchmarking, the custom CUDA kernel is much (~100x) faster than the kornia implementation even if we try to optimize the latter (e.g. via torch compilation). Another user also reported similar observations (prittt/YACCLAB#28 (comment)). |
This was referenced Aug 6, 2024
Closed
Closed
This was referenced Aug 7, 2024
This was referenced Aug 7, 2024
Closed
Closed
4 tasks
fbcotter
added a commit
to wayveai/segment-anything-2
that referenced
this pull request
Sep 6, 2024
… into facebookresearch-main * 'main' of github.com:facebookresearch/segment-anything-2: (40 commits) open `README.md` with unicode (to support Hugging Face emoji); fix various typos (facebookresearch#218) accept kwargs in auto_mask_generator Fix HF image predictor improving warning message and adding further tips for installation (facebookresearch#204) better support for non-CUDA devices (CPU, MPS) (facebookresearch#192) Update hieradet.py add Colab support to the notebooks; pack config files in `sam2_configs` package during installation (facebookresearch#176) also catch errors during installation in case `CUDAExtension` cannot be loaded (facebookresearch#175) Add interface for box prompt in SAM 2 video predictor (facebookresearch#174) Address comment Update hieradet.py Update docstrings Revert code snippet Updated INSTALL.md with CUDA_HOME-related troubleshooting (facebookresearch#140) Format using ufmt Update INSTALL.md (facebookresearch#156) Update README Make it optional to build CUDA extension for SAM 2; also fallback to all available kernels if Flash Attention fails (facebookresearch#155) Clean up Address comment ...
xydy666
pushed a commit
to xydy666/segment-anything-2
that referenced
this pull request
Sep 17, 2024
…all available kernels if Flash Attention fails (facebookresearch#155) In this PR, we make it optional to build the SAM 2 CUDA extension, in observation that many users encounter difficulties with the CUDA compilation step. 1. During installation, we catch build errors and print a warning message. We also allow explicitly turning off the CUDA extension building with `SAM2_BUILD_CUDA=0`. 2. At runtime, we catch CUDA kernel errors from connected components and print a warning on skipping the post processing step. We also fall back to the all available kernels if the Flash Attention kernel fails.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
In this PR, we make it optional to build the SAM 2 CUDA extension, in observation that many users encounter difficulties with the CUDA compilation step.
SAM2_BUILD_CUDA=0.We also fall back to the all available kernels if the Flash Attention kernel fails.