Is there an easy way to specify a that a task should recieve the serialized version of an output from a previous task? Like say I have a train task that outputs a

nn.Module

with the pytorch plugin, but I want a downstream task to receive the filepath that the

nn.Module

is saved to. Is there a way to do that?

The serialization/deserialization has to happen automatically. https://docs.flyte.org/projects/cookbook/en/latest/auto/core/type_system/pytorch_types.html#tensors-modules

Say I want to report the location of the serialized object to a third party tracking system. How would you suggest I do that? I could subclass the serialize and reregister it with the typeengine, but that seems like overkill.

You can manually fetch the path. It should be available on the Flyte UI (in the inputs/outputs pane). But if you'd like to automate this process, I believe you'll need to re-register it.

So you'd suggest I have a separate task that manually serializes it as a flyte file and then fetches the remote path of that flyte file.

That isn't possible cause workflow is essentially a DSL and you can't perform any operations as such. The best way forward IMO is to re-register. Here's the existing type transformer code for your reference: https://github.com/flyteorg/flytekit/blob/152080fa27c6d07edda9fcc2a602ca3ca9f4739a/flytekit/extras/pytorch/native.py#L24-L64.

Yes I understand it's not possible now. It would essentially be a flag to the flytekit engine informing it not to deserialize in the consuming input task.

@tall-lock-23197 you do have to manually fetch, @melodic-magician-71351 you are right, you can receive as a FlyteFile[format]

Can you clarify? Do I annotate as:

def reporter(location: FlyteFile[nn.Module]):

Format=PyTorchModule

Awesome thanks!

You are in the far trenches

Ya no worries. I'm trying to integrate with a legacy system so that has me sprinting to the edges of the surface 😉

you are right, you can receive as a FlyteFile[format]

Oh yeah. This has to work!

Heyo, I'm trying this with a huggingface module and am getting:

TypeError: No automatic conversion found from type <class 'transformers.models.esm.modeling_esm.EsmForMaskedLM'> to FlyteFile.Supported (os.PathLike, str, Flytefile)

Looks like you're trying to use two different types. Would you mind sharing your Flyte tasks?

Sure. Task in question is:

@wandb_task
def register_model(model: FlyteFile[PyTorchModuleTransformer.PYTORCH_FORMAT],
                   model_name: str,
                   tokenizer: Optional[TokenizerSpec]) -> str:
    api = wandb.Api()
    artifact_id = f'{commons.FLYTE_CONTEXT.project}/{commons.FLYTE_CONTEXT.workflow_name}/{model_name}:{commons.FLYTE_CONTEXT.version}'
    artifact = api.artifact(artifact_id, type='model')
    artifact.add_reference(uri=model.remote_path)
    if tokenizer:
        artifact.metadata['tokenizer'] = tokenizer.to_dict()

    artifact.save()

    return artifact_id

I want to be able to pass it a

nn.Module

or a

FlyteFile

that represents an nn.Module, and I thought what @freezing-airport-6809 was suggesting is I could automatically do conversion/don't deserialize by using an annotated FlyteFile. e.g.

@workflow
def my_workflow(model: nn.Module) -> str:
    model = train_task(model=model)
    # here the pipelines still sees "model" as a `nn.Module`
    return register_model(model=model) # This task can take an `nn.Module` but will always receive it as a `FlyteFile`

And register_model would get the raw

FlyteFile

. As you can see the goal for the "register" task is to simply update a remote server with a reference to the remote location of the artifact. It would be wasteful to download that and then upload it to another location, just to convert the type from

nn.Module

to

FlyteFile

when there is already a remote location. I couldn't find anything in the logic for

FlyteFile

that would indicate this is the case, so perhaps I misunderstood. I'm wondering what the right solution to this pattern is (ideally one that doesn't involve redownloading the

nn.Module

just to upload it to a different location where I have the URI....my other solution was to essentially rewrite the

PyTorchModuleTransformer

to do this registration for me once it gets the remote_uri, but this feels a little wrong.

yes union type wont work