search

iree-org / iree-turbine

IREE's PyTorch Frontend, based on Torch Dynamo.
Apache License 2.0
44 stars 24 forks source link

How to customize cpy.py for my own backend? #38

Open bailuan opened 3 months ago

bailuan commented 3 months ago

I notice there is some description like below in dynamo/backends/cpu.py: # IREE runtime globals. For the CPU right now, there is no device selection, so it is easy. @functools.lru_cache(maxsize=None) def _get_device_state() -> DeviceState: return DeviceState(driver="local-task"),

I know the "local-task" means llvm-cpu backend. What should I do if I want use my own customized backend abm? Create a abm.py like cpu.py? In complile stage, I can use same api in cpu.py and add some flags of my own backend.

but in runtime stage, I am not very clear how to do that: becauese I have write a specific custom-module-run to take place of iree-run-module, the custom-module-run's main.cc contains some specific device created named abm and some other function like instance create, etc. How to customize the code here to fit my backend, right? But if I only imitate to construct a DeviceState. Really need some guide, thx.

stellaraccident commented 3 months ago

The dynamo torch.compile side needs some work. The end state were heading to is to only have one "turbine" torch compile backend and have it get its runtime device and compile flags based on looking at the devices that its arguments are on. The is how the other parts of turbine work, but the torch.compile side is old code that was never retrofitted.

I have someone who was going to start refactoring this. Let me check with them tomorrow and see where it is in the queue.

For your reference, the newer code is getting a device instance (which includes compile flags) from here: https://github.com/iree-org/iree-turbine/blob/b489d8b37f2ca0c7a658f2df47467d8c394f0e2a/shark_turbine/runtime/device.py#L419

There are some hard coded mappings in there for known device types that would need to be extended. That is the right thing to be doing if you have an iree device that is also natively supported by pytorch.

I think you might be asking for something slightly different: supporting offload to a device not otherwise supported by pytorch. For that, indeed, you would follow the example of the CPU.py and create your own. You will need to build iree python bindings that have compiled in support for your device and use that.

If you get past this just being a toy, there are ways to support a completely out of tree device in the iree Python API. There are a couple of options but I don't think anyone outside of projects under active development have crossed that bridge yet.

bailuan commented 3 months ago

Very thanks, @stellaraccident. I think you might be asking for something slightly different: supporting offload to a device not otherwise supported by pytorch. For that, indeed, you would follow the example of the CPU.py and create your own. You will need to build iree python bindings that have compiled in support for your device and use that. this is my request. My precise question is: if I create a abm.py file based on cpu.py, how to make the torch.compile program goes into abm.py but not cpu.py. Do you mind give me a email or any other contact information? Maybe a chat is better for efficiency.

stellaraccident commented 3 months ago

See the discord link on the main iree page for an invite to the server. The devs hang out there and try to answer questions.

With the caveat that we will be working in this area again at some point and may reorganize some things: it's just a torch thing done through setup.py entrypoints. For example: https://github.com/iree-org/iree-turbine/blob/3e678c7aa0e8ddb5ef004f18b0fda6af82134aec/setup.py#L103

That is how torch associates a backend name with an implementation.

bailuan commented 3 months ago

Thanks a lot. I will have a try.