leandron commented on a change in pull request #7823:
URL: https://github.com/apache/tvm/pull/7823#discussion_r614905434
##########
File path: python/tvm/driver/tvmc/runner.py
##########
@@ -112,8 +112,10 @@ def drive_run(args):
except IOError as ex:
raise TVMCException("Error loading inputs file: %s" % ex)
- outputs, times = run_module(
- args.FILE,
+ tvmc_package = TVMCPackage(package_path=args.FILE)
Review comment:
Thanks @jwfromm I'll have a look.
(the comment got a bit long, but I thought to put the idea here and see what
you think, as we probably got everybody in the PR anyway).
@areusch I think we have the opportunity to make Model Library Format (MLF)
to be the way we serialize TVM outputs. Obviously, due to the nature of the
outputs we produce from different backends (object files, shared libraries,
sources), it can contain various artifacts within a tarball, and for that I
think we could make better use of the metadata descriptor file, to describe
what is in there in a way that we programmatically understand, for a given MLF
package.
This would open the door to, for example, on Project API, we check whether a
given tarball is created from a "micro" target and can be used to integrate
with an RTOS. In TVMC, we can define whether the package aims, cpu, cuda, cl,
etc. and whether it includes the shared library, etc.
With an abstraction like `TVMCPackage` (maybe renamed) we could make those
features to be easily available for target applications already using TVM, to
make use of the packages.
> the one I mentioned, which is that Model Library Format doesn't support
non-DSO-exportable Modules e.g. CUDA, Vulkan, etc.
I think I saw you mentioning this limitation in the RFC as well, but I'm
curious to understand a bit more on where is the limitation? is that on TVM
side, or on the targets, such as Vulkan, that we can't export them?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
