I’m encountering a problem during workflow registration when task arguments (e.g. an `nn.Module`) ne...

cool-lifeguard-49380

03/07/2023, 11:51 AM

I’m encountering a problem during workflow registration when task arguments (e.g. an

nn.Module

) need to be serialized into blob storage. Instead of a bucket uri, a

/tmp/flyte-xyz

path is shown in the console: I expect that my flytekit/remote config is not correct and would appreciate some help to figure out where exactly. Thanks 🙏 More in the 🧵

cool-lifeguard-49380

03/07/2023, 11:52 AM

I didn’t notice this problem so far since we typically pass dataclass_jsons with references to models in a bucket to our workflows/tasks instead of actual files. But this minimal working example results in the screenshot shown above:

Copy code

from flytekit import task, workflow
from torch import nn

class Config:  # enforce pickle transport
    def __init__(self):
        self.a = 1

@task
def train(cfg: Config, model: nn.Module):
    print(cfg)

@workflow
def wf():
    train(cfg=Config(), model=nn.Linear(10, 10))

cool-lifeguard-49380

03/07/2023, 11:54 AM

I put a print statement into the

to_literal

method of the pickle transformer (here). It results in:

Copy code

execution state: ExecutionState(mode=None, working_dir=PosixPath('/tmp/flyte-z15pwqet/sandbox/local_flytekit'), engine_dir='/tmp/flyte-z15pwqet/sandbox/local_flytekit/engine_dir', branch_eval_mode=None, user_space_params=<flytekit.core.context_manager.ExecutionParameters object at 0x7f7a445d6520>)

cool-lifeguard-49380

03/07/2023, 11:54 AM

So the

working_dir

and the

engine_dir

of the

ExecutionState

are the same as in the Flyte console screenshot.

cool-lifeguard-49380

03/07/2023, 11:55 AM

I’m registering with this command:

Copy code

pyflyte register --project sandbox     --version ...     --destination-dir /home/flyte     --image ... test_pickle.py

cool-lifeguard-49380

03/07/2023, 11:56 AM

~/.flyte/config.yaml

contains this:

Copy code

admin:
  ...
logger:
  show-source: true
  level: 0
storage:
  type: stow
  stow:
    kind: google
    config:
      json: ""
      project_id: <project_id>
      scopes: <https://www.googleapis.com/auth/devstorage.read_write>

cool-lifeguard-49380

03/07/2023, 4:59 PM

Here and here it looks like the

default_local_file_access_provider

always uses a

/tmp/flyte-

path, without taking any config into consideration 🤔

tall-lock-23197

03/08/2023, 10:42 AM

Have you tried setting

--raw-data-prefix

in your

pyflyte register

command to the s3 bucket?

cool-lifeguard-49380

03/09/2023, 2:11 PM

Yes,

--raw-data-prefix

(with a

gs://

uri) doesn’t have an effect on this, the console still shows

/tmp/flyte-

inputs 🤔

cool-lifeguard-49380

03/09/2023, 2:12 PM

Should

--raw-data-prefix

have changed this? In the links I pasted here, it looks like the

/tmp/flyte-

paths are hardcoded.

tall-lock-23197

03/10/2023, 6:16 AM

Yeah.

--raw-data-prefix

should have changed this. What's the raw output prefix in the propeller config? cc @high-accountant-32689

cool-lifeguard-49380

03/10/2023, 8:33 AM

It is set to the same bucket as `userSettings.bucketName`:

rawoutput-prefix: gs://...

tall-lock-23197

03/10/2023, 9:01 AM

Eduardo / @glamorous-carpet-83516, any idea what might be causing this issue?

tall-lock-23197

03/14/2023, 3:45 AM

Hey @high-accountant-32689 / @thankful-minister-83577, can you help Fabio please?

high-accountant-32689

03/15/2023, 10:14 PM

@cool-lifeguard-49380, can you separate the default values for those inputs into a separate task? The thing is that at serialization time we have to evaluate those values (

Config

and

nn.Module

in the definition of the workflow in your example) locally. Essentially I'm saying:

Copy code

import pandas as pd
from flytekit import task, workflow
from torch import nn

class Config:  # enforce pickle transport
    def __init__(self):
        self.a = 1

@task
def get_default_config() -> Config:
    return Config()

@task
def get_default_model() -> nn.Module:
    return nn.Linear(10, 10)

@task
def train(cfg: Config, model: nn.Module):
    print(cfg)

@workflow
def wf():
    train(cfg=get_default_config(), model=get_default_model())

cool-lifeguard-49380

03/16/2023, 5:24 PM

Hey @high-accountant-32689, yes that would of course circumvent the problem but it is not what the engineer is trying to do 🤔 Their goal was to take a file that exists locally (e.g. a config file or a model checkpoint) and during registration have the type transformer serialize it into blob storage so that it is then available during remote execution. Especially for experimentation this can be helpful. My example with the config was just to enforce the pickle transformer. I feel it would also be totally valid if this behaviour wasn’t supported but in this case the registration should fail since any

/tmp/…

dir will not be available in the cluster and it takes some amount of understanding what is happening to figure out why this fails. Is there any way in the type transformer to identify from the

ctx

that we are in registration mode and not local execution?

high-accountant-32689

03/16/2023, 10:22 PM

yeah, we could certainly detect that, @cool-lifeguard-49380. Mind filling an issue?

cool-lifeguard-49380

03/17/2023, 9:00 AM

Will do 👍

thx 1

152 Views

Open in Slack

Previous Next

Flyte

Flyte enables production-grade orchestration for machine learning workflows and data processing created to accelerate local workflows to production.