[iree.build] Add turbine_generate iree.build rule.
#249
Conversation
stellaraccident
force-pushed
the
turbine_generate
branch
from
e76aa63
to
8ca6c64
Compare
Signed-off-by: Stella Laurenzo <[email protected]>
Signed-off-by: Stella Laurenzo <[email protected]>
stellaraccident
force-pushed
the
turbine_generate
branch
from
8ca6c64
to
1f039fe
Compare
Signed-off-by: Stella Laurenzo <[email protected]>
Signed-off-by: Stella Laurenzo <[email protected]>
There was a problem hiding this comment.
@marbre any ideas for how we could include iree-turbine docs as part of https://iree-python-api.readthedocs.io/ ?
I'd like to include this page under https://iree-python-api.readthedocs.io/en/latest/compiler/build.html as part of iree-org/iree#19019 , and we can also put the rest of the iree-turbine docs up too. We should limit the number of doc sites we have to maintain, even if this crosses repository boundaries a bit :P
| * `ExportOutput`: The result of calling `aot.export(...)` will result in | ||
| `save_mlir()` being called on it while still in the subprocess to write to | ||
| a file names `{name}.mlir` if there is one return or `{name}_{n}.mlir` if | ||
| multiple. |
There was a problem hiding this comment.
Can this output .mlirbc? What about .irpa files?
There was a problem hiding this comment.
It needs to grow some options for mlirbc. Have some thoughts on how to plug parameter generation through in an ergonomic way. One of:
- Support another return type and then the generator just returns something that causes the export.
- Have an
extra_outputs=dict that lets you set up additional output files and have that passed into the generator function to do with as you please.
There was a problem hiding this comment.
(but wanted to land this before adding more)
| # Do an exhaustive subclass check. | ||
| for k, m in RETURN_MARSHALLERS_BY_TYPE.items(): | ||
| if issubclass(t, k): | ||
| # Cache it. | ||
| RETURN_MARSHALLERS_BY_TYPE[t] = m | ||
| return m | ||
| raise ValueError( | ||
| f"In order to use a callable as a generator in turbine_generate, it must be " | ||
| f" annotated with specific return types. Found '{t}' but only " | ||
| f"{EXPLICIT_MARSHALLER_TYPES} are supported" | ||
| ) |
| return export(fxb) | ||
|
|
||
|
|
||
| @entrypoint(description="Builds an awesome pipeline") |
iree-requirements-ci.txt
Outdated
| @@ -10,5 +10,5 @@ | |||
| # Uncomment to skip versions from PyPI (so _only_ nightly versions). | |||
| # --no-index | |||
|
|
|||
| iree-base-compiler==3.1.0rc20241122 | |||
| iree-base-runtime==3.1.0rc20241122 | |||
| iree-base-compiler==3.1.0rc20241126 | |||
There was a problem hiding this comment.
One of the actions is failing on this
https://github.com/iree-org/iree-turbine/actions/runs/12040949027/job/33571834502?pr=249#step:4:99
Ah, because the macOS release build failed to upload with failed, reason: write ECONNRESET 🤦 : https://github.com/iree-org/iree/actions/runs/12024344562/job/33519713662
We should
- Change that workflow to not need all packages to exist
- Fix the release upload (this has happened a few times lately)
There was a problem hiding this comment.
I'll update it with tomorrow's build before landing. Technically the one from today had a bug with this stuff on Python 3.10.
| out_of_process: bool = True, | ||
| **kwargs, | ||
| ): | ||
| """Invokes a user-defined generator callable as an action, performing turbine |
There was a problem hiding this comment.
consider renaming "generator callable" to "exporter callable" and/or explain that it is a callable that returns an exported torch module.
For me, "generator callable" resolves to the python language feature, "generator functions", like
def fib():
prev = 0
cur = 1
while True:
yield cur
prev, cur = cur, cur+prevThere was a problem hiding this comment.
I'm running out of daylight before vacation to land changes, but sounds like a reasonable idea. Maybe you could send a patch next week and work with Ean/Scott on a better name? I'll land it like this but it has no use yet -- easy to rename.
Signed-off-by: Stella Laurenzo <[email protected]>
This supports performing parallelizable aot model export as part of the iree.build tool. By default, it runs actions out of process, meaning that multiple torch.export and torch-mlir imports can be running with true concurrency.