import inspect
from pathlib import Path
from typing import Any, Iterable, List, Optional, Tuple, Union
from runhouse import globals
from runhouse.logger import get_logger
from runhouse.resources.envs import Env
from runhouse.resources.hardware import Cluster
from runhouse.resources.module import Module
from runhouse.resources.resource import Resource
logger = get_logger(__name__)
[docs]class Function(Module):
RESOURCE_TYPE = "function"
[docs] def __init__(
self,
fn_pointers: Optional[Tuple] = None,
name: Optional[str] = None,
system: Optional[Cluster] = None,
env: Optional[Env] = None,
dryrun: bool = False,
**kwargs, # We have this here to ignore extra arguments when calling from from_config
):
"""
Runhouse Function object. It is comprised of the entrypoint, system/cluster,
and dependencies necessary to run the service.
.. note::
To create a Function, please use the factory method :func:`function`.
"""
self.fn_pointers = fn_pointers
self._loaded_fn = None
super().__init__(name=name, dryrun=dryrun, system=system, env=env, **kwargs)
# ----------------- Constructor helper methods -----------------
[docs] @classmethod
def from_config(
cls, config: dict, dryrun: bool = False, _resolve_children: bool = True
):
if isinstance(config["system"], dict):
config["system"] = Cluster.from_config(
config["system"], dryrun=dryrun, _resolve_children=_resolve_children
)
if isinstance(config["env"], dict):
config["env"] = Env.from_config(
config["env"], dryrun=dryrun, _resolve_children=_resolve_children
)
config.pop("resource_subtype", None)
return Function(**config, dryrun=dryrun)
[docs] def share(self, *args, visibility=None, **kwargs):
if visibility and not visibility == self.visibility:
self.visibility = visibility
super().remote.visibility = (
visibility # do this to avoid hitting Function's .remote
)
return super().share(*args, **kwargs, visibility=visibility)
def default_name(self):
return (
self.fn_pointers[2] if self.fn_pointers else None
) or super().default_name()
[docs] def to(
self,
system: Union[str, Cluster],
env: Optional[Union[str, List[str], Env]] = None,
name: Optional[str] = None,
force_install: bool = False,
):
"""
Send the function to the specified env on the cluster. This will sync over relevant code and packages
onto the cluster, and set up the environment if it does not yet exist on the cluster.
Args:
system (str or Cluster): The system to setup the function and env on.
env (str, List[str], or Env, optional): The environment where the function lives on in the cluster,
or the set of requirements necessary to run the function. (Default: ``None``)
name (Optional[str], optional): Name to give to the function resource, if you wish to rename it.
(Default: ``None``)
force_install (bool, optional): Whether to re-install and perform the environment setup steps, even
if it may already exist on the cluster. (Defualt: ``False``)
Example:
>>> rh.function(fn=local_fn).to(gpu_cluster)
>>> rh.function(fn=local_fn).to(system=gpu_cluster, env=my_conda_env)
>>> rh.function(fn=local_fn).to(system='aws_lambda') # will deploy the rh.function to AWS as a Lambda.
""" # noqa: E501
if isinstance(system, str) and system.lower() == "lambda_function":
from runhouse.resources.functions.aws_lambda_factory import aws_lambda_fn
return aws_lambda_fn(
fn=self._get_obj_from_pointers(*self.fn_pointers), env=env
)
return super().to(
system=system, env=env, name=name, force_install=force_install
)
# ----------------- Function call methods -----------------
def __call__(self, *args, **kwargs) -> Any:
"""Call the function on its system
Args:
*args: Optional args for the Function
stream_logs (bool): Whether to stream the logs from the Function's execution.
Defaults to ``True``.
run_name (Optional[str]): Name of the Run to create. If provided, a Run will be created
for this function call, which will be executed synchronously on the cluster before returning its result
**kwargs: Optional kwargs for the Function
Returns:
The Function's return value
"""
return self.call(*args, **kwargs)
def call(self, *args, **kwargs) -> Any:
# We need this strictly because Module's __getattribute__ overload can't pick up the __call__ method
if not self._loaded_fn:
self._loaded_fn = self._get_obj_from_pointers(*self.fn_pointers)
return self._loaded_fn(*args, **kwargs)
[docs] def method_signature(self, method):
if callable(method) and method.__name__ == "call":
return self.method_signature(self._get_obj_from_pointers(*self.fn_pointers))
return super().method_signature(method)
def map(self, *args, **kwargs):
"""Map a function over a list of arguments.
Example:
>>> def local_sum(arg1, arg2, arg3):
>>> return arg1 + arg2 + arg3
>>>
>>> remote_fn = rh.function(local_fn).to(gpu)
>>> remote_fn.map([1, 2], [1, 4], [2, 3])
>>> # output: [4, 9]
"""
import ray
fn = self._get_obj_from_pointers(*self.fn_pointers)
ray_wrapped_fn = ray.remote(fn)
return ray.get([ray_wrapped_fn.remote(*args, **kwargs) for args in zip(*args)])
def starmap(self, args_lists: List[Iterable], **kwargs):
"""Like :func:`map` except that the elements of the iterable are expected to be iterables
that are unpacked as arguments. An iterable of [(1,2), (3, 4)] results in [func(1,2), func(3,4)].
Args:
arg_lists (List[Iterable]): List containing iterbles of arguments to be passed into the function.
Example:
>>> arg_list = [(1, 2), (3, 4)]
>>> # runs the function twice, once with args (1, 2) and once with args (3, 4)
>>> remote_fn.starmap(arg_list)
"""
import ray
fn = self._get_obj_from_pointers(*self.fn_pointers)
ray_wrapped_fn = ray.remote(fn)
return ray.get([ray_wrapped_fn.remote(*args, **kwargs) for args in args_lists])
def run(self, *args, local=True, **kwargs):
key = self.call.run(*args, **kwargs)
return key
[docs] def get(self, run_key):
"""Get the result of a Function call that was submitted as async using `run`.
Args:
run_key: A single or list of runhouse run_key strings returned by calling ``.call.remote()`` on the
Function. The ObjectRefs must be from the cluster that this Function is running on.
Example:
>>> remote_fn = rh.function(local_fn).to(gpu)
>>> remote_fn_run = remote_fn.run()
>>> remote_fn.get(remote_fn_run.name)
"""
return self.system.get(run_key)
[docs] def config(self, condensed=True):
"""The config of the function.
Args:
condensed (bool, optional): Whether to return the condensed config without expanding children subresources,
or return the whole expanded config. (Default: ``True``)
"""
config = super().config(condensed)
config.update(
{
"fn_pointers": self.fn_pointers,
}
)
return config
def get_or_call(
self, run_name: str, load_from_den: bool = True, *args, **kwargs
) -> Any:
"""Check if object already exists on cluster or rns, and if so return the result. If not, run the function.
Keep in mind this can be called with any of the usual method call modifiers - `remote=True`, `run_async=True`,
`stream_logs=False`, etc.
Args:
run_name (str): Name of a particular run for this function.
load_from_den (bool, optional): Whether to try loading the run name from Den. (Default: ``True``)
*args: Arguments to pass to the function for the run (relevant if creating a new run).
**kwargs: Keyword arguments to pass to the function for the run (relevant if creating a new run).
Returns:
Any: Result of the Run
Example:
>>> # previously, remote_fn.run(arg1, arg2, run_name="my_async_run")
>>> remote_fn.get_or_call()
"""
# TODO let's just do this for functions initially, and decide if we want to support it for calls on modules
# as well. Right now this only works with remote=True, we should decide if we want to fix that later.
resource = globals.rns_client.load_config(
name=run_name, load_from_den=load_from_den
)
if resource:
return Resource.from_name(
name=run_name, load_from_den=load_from_den, dryrun=self.dryrun
)
try:
return self.system.get(run_name, default=KeyError)
except KeyError:
logger.info(f"Item {run_name} not found on cluster. Running function.")
return self.call(*args, **kwargs, run_name=run_name)
@staticmethod
def _handle_nb_fn(fn, fn_pointers, serialize_notebook_fn, name):
"""Handle the case where the user passes in a notebook function"""
if serialize_notebook_fn:
# This will all be cloudpickled by the RPC client and unpickled by the RPC server
# Note that this means the function cannot be saved, and it's better that way because
# pickling functions is not meant for long term storage. Case in point, this method will be
# sensitive to differences in minor Python versions between the serializing and deserializing envs.
return "", "notebook", fn
else:
module_path = Path.cwd() / (f"{name}_fn.py" if name else "sent_fn.py")
logger.info(
f"Because this function is defined in a notebook, writing it out to {str(module_path)} "
f"to make it importable. Please make sure the function does not rely on any local variables, "
f"including imports (which should be moved inside the function body). "
f"This restriction does not apply to functions defined in normal Python files."
)
if not name:
logger.warning(
"You should name Functions that are created in notebooks to avoid naming collisions "
"between the modules that are created to hold their functions "
'(i.e. "sent_fn.py" errors.'
)
source = inspect.getsource(fn).strip()
with module_path.open("w") as f:
f.write(source)
return fn_pointers[0], module_path.stem, fn_pointers[2]