Made it possible to disable GPU detection#2424
Open
philip-paul-mueller wants to merge 10 commits into
Open
Conversation
ThrudPrimrose
requested changes
Jul 6, 2026
However, I have not yet tested it fully.
Collaborator
Author
|
cscs-ci run |
havogt
added a commit
to havogt/gt4py
that referenced
this pull request
Jul 7, 2026
- CompileJob -> LoadTask (a task yielding a loaded ExecutableProgram), with fields run -> compile_and_load and offload -> compilation - OffloadableWork -> CompilationTask (the decomposed part producing the CompilationArtifact), matching the compile/load vocabulary of the artifact split - CompilationRunner -> Runner - modules: compilation_runner.py -> runners.py, compile_jobs.py -> load_tasks.py (make_compile_job -> make_load_task) - TODO on the worker initializer for spcl/dace#2424 (scope of the CUDAARCHS workaround) - sharpen the connectivity-registry comment (main-process only, shared across the tasks of a run)
havogt
added a commit
to havogt/gt4py
that referenced
this pull request
Jul 7, 2026
Workers need the architecture from CUDAARCHS for any GPU build as long as they cannot query a device — spcl/dace#2424 only changes how the value is passed to dace, which is already marked in set_dace_config.
ThrudPrimrose
approved these changes
Jul 7, 2026
ThrudPrimrose
left a comment
ThrudPrimrose
left a comment
Collaborator
There was a problem hiding this comment.
LGTM with the changes
edopao
reviewed
Jul 7, 2026
Co-authored-by: Edoardo Paone <edoardo16@gmail.com>
Co-authored-by: Edoardo Paone <edoardo16@gmail.com>
tbennun
requested changes
Jul 14, 2026
Collaborator
There was a problem hiding this comment.
Need to also change setup.py: https://github.com/spcl/dace/blob/main/setup.py#L35
for the minimum cmake version required
Otherwise lgtm
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Currently DaCe will always compile and run a test program to determine the CUDA/HIP architecture.
This is problematic because this will create a CUDA-Context and prevents parallel compilations.
This PR changes this by changing the meaning of
compiler.cuda.{hip,cuda}_archs.Before they were the additional architectures to compile for, since they carried a value, in the case for CUDA the value was
60, we would always compile for two architectures the one that was detected and for60.The new meaning of these keys are "the architectures we want to compile for" and if they are empty, the current default, the architecture is auto detected.
Also the auto detection changed.
Before a test program was compiled to probe for the architecture.
The auto detection for CUDA was switched to
native, i.e. the compiler does it itself.However, this does not work for HIP thus the program is retained.