Use spawn for multiprocessing start method #3284
Closed
Conversation
This file contains hidden or 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
meta-cla
bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D80305233 |
zhuhan0
force-pushed
the
export-D80305233
branch
from
fb5995c to
249f44b
Compare
zhuhan0
added a commit
to zhuhan0/torchrec
that referenced
this pull request
Aug 15, 2025
Summary:
CUDA context initialization is not fork-safe. If a CUDA context is created in a parent process, and then the process is forked (using `os.fork()`), the child process may encounter errors or undefined behavior when using CUDA. This is because the CUDA driver and runtime are not designed to be safely duplicated via `fork()`. It's recommended to use `spawn` or `forkserver`.
Among the two, `forkserver` needs to be use carefully and specifically, it's recommended to call `multiprocessing.set_start_method('forkserver')` at the very start of the program, and the parent process also needs to avoid initializing the CUDA context. When upgrading APS to CUDA 12.8, we encountered a test failure, and the test is apparently initializing the CUDA context before starting up two children processes, and I suspect that caused the test to hang - [post](https://fb.workplace.com/groups/319878845696681/posts/1494595861558301).
It's hard to avoid initializing the CUDA context early in this test, because it checks the GPU count in the test method's decorator - [code](https://fburl.com/code/27naz2eg). Among the `spawn` and `forkserver` start methods, `spawn` is less efficient but it's the most robust. Let's switch to that instead to avoid any potential undefined behaviors with CUDA 12.8 and multiprocessing.
Reviewed By: adamomainz, weifengpy
Differential Revision: D80305233
|
This pull request was exported from Phabricator. Differential Revision: D80305233 |
zhuhan0
added a commit
to zhuhan0/torchrec
that referenced
this pull request
Aug 15, 2025
Summary: Pull Request resolved: meta-pytorch#3284 CUDA context initialization is not fork-safe. If a CUDA context is created in a parent process, and then the process is forked (using `os.fork()`), the child process may encounter errors or undefined behavior when using CUDA. This is because the CUDA driver and runtime are not designed to be safely duplicated via `fork()`. It's recommended to use `spawn` or `forkserver`. Among the two, `forkserver` needs to be use carefully and specifically, it's recommended to call `multiprocessing.set_start_method('forkserver')` at the very start of the program, and the parent process also needs to avoid initializing the CUDA context. When upgrading APS to CUDA 12.8, we encountered a test failure, and the test is apparently initializing the CUDA context before starting up two children processes, and I suspect that caused the test to hang - [post](https://fb.workplace.com/groups/319878845696681/posts/1494595861558301). It's hard to avoid initializing the CUDA context early in this test, because it checks the GPU count in the test method's decorator - [code](https://fburl.com/code/27naz2eg). Among the `spawn` and `forkserver` start methods, `spawn` is less efficient but it's the most robust. Let's switch to that instead to avoid any potential undefined behaviors with CUDA 12.8 and multiprocessing. Reviewed By: adamomainz, weifengpy Differential Revision: D80305233
zhuhan0
force-pushed
the
export-D80305233
branch
2 times, most recently
from
c9a654b to
528871e
Compare
zhuhan0
added a commit
to zhuhan0/torchrec
that referenced
this pull request
Aug 15, 2025
Summary:
CUDA context initialization is not fork-safe. If a CUDA context is created in a parent process, and then the process is forked (using `os.fork()`), the child process may encounter errors or undefined behavior when using CUDA. This is because the CUDA driver and runtime are not designed to be safely duplicated via `fork()`. It's recommended to use `spawn` or `forkserver`.
Among the two, `forkserver` needs to be use carefully and specifically, it's recommended to call `multiprocessing.set_start_method('forkserver')` at the very start of the program, and the parent process also needs to avoid initializing the CUDA context. When upgrading APS to CUDA 12.8, we encountered a test failure, and the test is apparently initializing the CUDA context before starting up two children processes, and I suspect that caused the test to hang - [post](https://fb.workplace.com/groups/319878845696681/posts/1494595861558301).
It's hard to avoid initializing the CUDA context early in this test, because it checks the GPU count in the test method's decorator - [code](https://fburl.com/code/27naz2eg). Among the `spawn` and `forkserver` start methods, `spawn` is less efficient but it's the most robust. Let's switch to that instead to avoid any potential undefined behaviors with CUDA 12.8 and multiprocessing.
Reviewed By: adamomainz, weifengpy
Differential Revision: D80305233
Summary: Pull Request resolved: meta-pytorch#3284 CUDA context initialization is not fork-safe. If a CUDA context is created in a parent process, and then the process is forked (using `os.fork()`), the child process may encounter errors or undefined behavior when using CUDA. This is because the CUDA driver and runtime are not designed to be safely duplicated via `fork()`. It's recommended to use `spawn` or `forkserver`. Among the two, `forkserver` needs to be use carefully and specifically, it's recommended to call `multiprocessing.set_start_method('forkserver')` at the very start of the program, and the parent process also needs to avoid initializing the CUDA context. When upgrading APS to CUDA 12.8, we encountered a test failure, and the test is apparently initializing the CUDA context before starting up two children processes, and I suspect that caused the test to hang - [post](https://fb.workplace.com/groups/319878845696681/posts/1494595861558301). It's hard to avoid initializing the CUDA context early in this test, because it checks the GPU count in the test method's decorator - [code](https://fburl.com/code/27naz2eg). Among the `spawn` and `forkserver` start methods, `spawn` is less efficient but it's the most robust. Let's switch to that instead to avoid any potential undefined behaviors with CUDA 12.8 and multiprocessing. Reviewed By: adamomainz, weifengpy Differential Revision: D80305233
|
This pull request was exported from Phabricator. Differential Revision: D80305233 |
zhuhan0
force-pushed
the
export-D80305233
branch
from
528871e to
125ee00
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Summary:
CUDA context initialization is not fork-safe. If a CUDA context is created in a parent process, and then the process is forked (using
os.fork()), the child process may encounter errors or undefined behavior when using CUDA. This is because the CUDA driver and runtime are not designed to be safely duplicated viafork(). It's recommended to usespawnorforkserver.Among the two,
forkserverneeds to be use carefully and specifically, it's recommended to callmultiprocessing.set_start_method('forkserver')at the very start of the program, and the parent process also needs to avoid initializing the CUDA context. When upgrading APS to CUDA 12.8, we encountered a test failure, and the test is apparently initializing the CUDA context before starting up two children processes, and I suspect that caused the test to hang - post.It's hard to avoid initializing the CUDA context early in this test, because it checks the GPU count in the test method's decorator - code. Among the
spawnandforkserverstart methods,spawnis less efficient but it's the most robust. Let's switch to that instead to avoid any potential undefined behaviors with CUDA 12.8 and multiprocessing.Differential Revision: D80305233