This commit extends the DSL that can be used in packages
to allow declaring that a package uses different build-systems
under different conditions.
It requires each spec to have a `build_system` single valued
variant. The variant can be used in many context to query, manipulate
or select the build system associated with a concrete spec.
The knowledge to build a package has been moved out of the
PackageBase hierarchy, into a new Builder hierarchy. Customization
of the default behavior for a given builder can be obtained by
coding a new derived builder in package.py.
The "run_after" and "run_before" decorators are now applied to
methods on the builder. They can also incorporate a "when="
argument to specify that a method is run only when certain
conditions apply.
For packages that do not define their own builder, forwarding logic
is added between the builder and package (methods not found in one
will be retrieved from the other); this PR is expected to be fully
backwards compatible with unmodified packages that use a single
build system.
* backtraces without --debug
Currently `--debug` is too verbose and not-`--debug` gives to little
context about where exceptions are coming from.
So, instead, it'd be nice to have `spack --backtrace` and
`SPACK_BACKTRACE=1` as methods to get something inbetween: no verbose
debug messages, but always a full backtrace.
This is useful for CI, where we don't want to drown in debug messages
when installing deps, but we do want to get details where something goes
wrong if it goes wrong.
* completion
When we lose a running pod (possibly loss of spot instance) or encounter
some other infrastructure-related failure of this job, we need to retry
it. This retries the job the maximum number of times in those cases.
`reuse` and `when_possible` concretization broke the invariant that
`spec[pkg_name]` has unique keys. This invariant is relied on in tons of
places, such as when setting up the build environment.
When using `when_possible` concretization, one may end up with two or
more `perl`s or `python`s among the transitive deps of a spec, because
concretization does not consider build-only deps of reusable specs.
Until the code base is fixed not to rely on this broken property of
`__getitem__`, we should disable reuse in CI.
When installing some/all specs from a buildcache, build edges are pruned
from those specs. This can result in a much smaller effective DAG. Until
now, `spack env depfile` would always generate a full DAG.
Ths PR adds the `spack env depfile --use-buildcache` flag that was
introduced for `spack install` before. This way, not only can we drop
build edges, but also we can automatically set the right buildcache
related flags on the specific specs that are gonna get installed.
This way we get parallel installs of binary deps without redundancy,
which is useful for Gitlab CI.
Currently "spack ci generate" chooses the first matching entry in
gitlab-ci:mappings to fill attributes for a generated build-job,
requiring that the entire configuration matrix is listed out
explicitly. This unfortunately causes significant problems in
environments with large configuration spaces, for example the
environment in #31598 (spack.yaml) supports 5 operating systems,
3 architectures and 130 packages with explicit size requirements,
resulting in 1300 lines of configuration YAML.
This patch adds a configuraiton option to the gitlab-ci schema called
"match_behavior"; when it is set to "merge", all matching entries
are applied in order to the final build-job, allowing a few entries
to cover an entire matrix of configurations.
The default for "match_behavior" is "first", which behaves as before
this commit (only the runner attributes of the first match are used).
In addition, match entries may now include a "remove-attributes"
configuration, which allows matches to remove tags that have been
aggregated by prior matches. This only makes sense to use with
"match_behavior:merge". You can combine "runner-attributes" with
"remove-attributes" to effectively override prior tags.
* env depfile: allow deps only install
- Refactor `spack env depfile` to use a Jinja template, making it a bit
easier to follow as a human being.
- Add a layer of indirection in the generated Makefile through an
`<prefix>/.install-deps/<hash>` target, which allows one to specify
different options when installing dependencies. For example, only
verbose/debug mode on when installing some particular spec:
```
$ spack -e my_env env depfile -o Makefile --make-target-prefix example
$ make example/.install-deps/<hash> -j16
$ make example/.install/<hash> SPACK="spack -d" SPACK_INSTALL_FLAGS=--verbose -j16
```
This could be used to speed up `spack ci rebuild`:
- Parallel install of dependencies from buildcache
- Better readability of logs, e.g. reducing verbosity when installing
dependencies, and splitting logs into deps.log and current_spec.log
* Silence please!
Caches used by repositories don't reference the global spack.repo.path instance
anymore, but get the repository they refer to during initialization.
Spec.virtual now use the index, and computation done to compute the index
use Repository.is_virtual_safe.
Code to construct mock packages and mock repository has been factored into
a unique MockRepositoryBuilder that is used throughout the codebase.
Add debug print for pushing and popping config scopes.
Changed spack.repo.use_repositories so that it can override or not previous repos
spack.repo.use_repositories updates spack.config.config according to the modifications done
Removed a peculiar behavior from spack.config.Configuration where push would always
bubble-up a scope named command_line if it existed
Basic stack of ML packages we would like to test and generate binaries for in CI.
Spack now has a large CI framework in GitLab for PR testing and public binary generation.
We should take advantage of this to test and distribute optimized binaries for popular ML
frameworks.
This is a pretty extensive initial set, including CPU, ROCm, and CUDA versions of a core
`x96_64_v4` stack.
### Core ML frameworks
These are all popular core ML frameworks already available in Spack.
- [x] PyTorch
- [x] TensorFlow
- [x] Scikit-learn
- [x] MXNet
- [x] CNTK
- [x] Caffe
- [x] Chainer
- [x] XGBoost
- [x] Theano
### ML extensions
These are domain libraries and wrappers that build on top of core ML libraries
- [x] Keras
- [x] TensorBoard
- [x] torchvision
- [x] torchtext
- [x] torchaudio
- [x] TorchGeo
- [x] PyTorch Lightning
- [x] torchmetrics
- [x] GPyTorch
- [x] Horovod
### ML-adjacent libraries
These are libraries that aren't specific to ML but are still core libraries used in ML pipelines
- [x] numpy
- [x] scipy
- [x] pandas
- [x] ONNX
- [x] bazel
Co-authored-by: Jonathon Anderson <17242663+blue42u@users.noreply.github.com>
Remove `module-info mode load` condition that prevents auto-unloading when autoloading is enabled. It looks like this condition was added to work around an issue in environment-modules that is no longer necessary.
Add quotes to make is-loaded happy
Install: Add use-buildcache option to install
* Allow differentiating between top level packages and dependencies when
determining whether to install from the cache or not.
* Add unit test for --use-buildcache
* Use metavar to display use-buildcache options.
* Update spack-completion
PR #32615 deprecated Python versions up to 3.6.X. Since
the "build-systems" pipeline requires Python 3.6.15 to
build "tut", it will fail on the first rebuild that
involves Python.
The "tut" package is meant to perform an end-to-end
test of the "Waf" build-system, which is scarcely
used. The fix therefore is just to remove it from
the pipeline.
amazon linux 2 ships a glibc that is too old to work with cuda toolkit
for aarch64.
For example:
`libcurand.so.10.2.10.50` requires the symbol `logf@@GLIBC_2.27`, but
glibc is at 2.26.
So, these specs are removed.
* ci: restore coverage computation
* Mark "test_foreground_background" as xfail
* Mark "test_foreground_background_output" as xfail
* Make number of processes explicit, remove verbosity on linux
* Run coverage on just 3 Python jobs for linux
* Run coverage on just 3 Python jobs for linux
* Run coverage on just 2 Python jobs for linux
* Add back verbose, since before we didn't encounter the xdist internal error
* Reduce the workers to 2
* Try to use command line
