torchx-nightly 2025.7.9__py3-none-any.whl → 2025.11.12__py3-none-any.whl

torchx/specs/__init__.py

@@ -1,4 +1,3 @@
-#!/usr/bin/env python3
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
 #
@@ -13,6 +12,8 @@ used by components to define the apps which can then be launched via a TorchX
 scheduler or pipeline adapter.
 """
 import difflib
+
+import os
 from typing import Callable, Dict, Mapping, Optional
 
 from torchx.specs.api import (
@@ -42,9 +43,11 @@ from torchx.specs.api import (
     RoleStatus,
     runopt,
     runopts,
+    TORCHX_HOME,
     UnknownAppException,
     UnknownSchedulerException,
     VolumeMount,
+    Workspace,
 )
 from torchx.specs.builders import make_app_handle, materialize_appdef, parse_mounts
 
@@ -52,14 +55,22 @@ from torchx.util.entrypoints import load_group
 
 from torchx.util.modules import import_attr
 
-AWS_NAMED_RESOURCES: Mapping[str, Callable[[], Resource]] = import_attr(
+GiB: int = 1024
+
+
+ResourceFactory = Callable[[], Resource]
+
+AWS_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
     "torchx.specs.named_resources_aws", "NAMED_RESOURCES", default={}
 )
-GENERIC_NAMED_RESOURCES: Mapping[str, Callable[[], Resource]] = import_attr(
+GENERIC_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
     "torchx.specs.named_resources_generic", "NAMED_RESOURCES", default={}
 )
-
-GiB: int = 1024
+CUSTOM_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
+    os.environ.get("TORCHX_CUSTOM_NAMED_RESOURCES", "torchx.specs.fb.named_resources"),
+    "NAMED_RESOURCES",
+    default={},
+)
 
 
 def _load_named_resources() -> Dict[str, Callable[[], Resource]]:
@@ -69,6 +80,7 @@ def _load_named_resources() -> Dict[str, Callable[[], Resource]]:
     for name, resource in {
         **GENERIC_NAMED_RESOURCES,
         **AWS_NAMED_RESOURCES,
+        **CUSTOM_NAMED_RESOURCES,
         **resource_methods,
     }.items():
         materialized_resources[name] = resource
@@ -122,7 +134,7 @@ def resource(
 
     If ``h`` is specified then it is used to look up the
     resource specs from the list of registered named resources.
-    See `registering named resource <https://pytorch.org/torchx/latest/advanced.html#registering-named-resources>`_.
+    See `registering named resource <https://meta-pytorch.org/torchx/latest/advanced.html#registering-named-resources>`_.
 
     Otherwise a ``Resource`` object is created from the raw resource specs.
 
@@ -225,5 +237,10 @@ __all__ = [
     "make_app_handle",
     "materialize_appdef",
     "parse_mounts",
+    "torchx_run_args_from_argparse",
+    "torchx_run_args_from_json",
+    "TorchXRunArgs",
     "ALL",
+    "TORCHX_HOME",
+    "Workspace",
 ]

torchx/specs/api.py

@@ -1,4 +1,3 @@
-#!/usr/bin/env python3
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
 #
@@ -12,11 +11,15 @@ import copy
 import inspect
 import json
 import logging as logger
+import os
+import pathlib
 import re
+import shutil
 import typing
+import warnings
 from dataclasses import asdict, dataclass, field
 from datetime import datetime
-from enum import Enum
+from enum import Enum, IntEnum
 from json import JSONDecodeError
 from string import Template
 from typing import (
@@ -67,6 +70,32 @@ YELLOW_BOLD = "\033[1;33m"
 RESET = "\033[0m"
 
 
+def TORCHX_HOME(*subdir_paths: str) -> pathlib.Path:
+    """
+    Path to the "dot-directory" for torchx.
+    Defaults to `~/.torchx` and is overridable via the `TORCHX_HOME` environment variable.
+
+    Usage:
+
+    .. doc-test::
+
+        from pathlib import Path
+        from torchx.specs import TORCHX_HOME
+
+        assert TORCHX_HOME() == Path.home() / ".torchx"
+        assert TORCHX_HOME("conda-pack-out") ==  Path.home() / ".torchx" / "conda-pack-out"
+    ```
+    """
+
+    default_dir = str(pathlib.Path.home() / ".torchx")
+    torchx_home = pathlib.Path(os.getenv("TORCHX_HOME", default_dir))
+
+    torchx_home = torchx_home / os.path.sep.join(subdir_paths)
+    torchx_home.mkdir(parents=True, exist_ok=True)
+
+    return torchx_home
+
+
 # ========================================
 # ==== Distributed AppDef API =======
 # ========================================
@@ -83,6 +112,8 @@ class Resource:
         memMB: MB of ram
         capabilities: additional hardware specs (interpreted by scheduler)
         devices: a list of named devices with their quantities
+        tags: metadata tags for the resource (not interpreted by schedulers)
+          used to add non-functional information about resources (e.g. whether it is an alias of another resource)
 
     Note: you should prefer to use named_resources instead of specifying the raw
     resource requirement directly.
@@ -93,6 +124,7 @@ class Resource:
     memMB: int
     capabilities: Dict[str, Any] = field(default_factory=dict)
     devices: Dict[str, int] = field(default_factory=dict)
+    tags: Dict[str, object] = field(default_factory=dict)
 
     @staticmethod
     def copy(original: "Resource", **capabilities: Any) -> "Resource":
@@ -101,6 +133,7 @@ class Resource:
         are present in the original resource and as parameter, the one from parameter
         will be used.
         """
+
         res_capabilities = dict(original.capabilities)
         res_capabilities.update(capabilities)
         return Resource(
@@ -319,6 +352,121 @@ class DeviceMount:
     permissions: str = "rwm"
 
 
+@dataclass
+class Workspace:
+    """
+    Specifies a local "workspace" (a set of directories). Workspaces are ad-hoc built
+    into an (usually ephemeral) image. This effectively mirrors the local code changes
+    at job submission time.
+
+    For example:
+
+      1. ``projects={"~/github/torch": "torch"}`` copies ``~/github/torch/**`` into ``$REMOTE_WORKSPACE_ROOT/torch/**``
+      2. ``projects={"~/github/torch": ""}`` copies ``~/github/torch/**`` into ``$REMOTE_WORKSPACE_ROOT/**``
+
+    The exact location of ``$REMOTE_WORKSPACE_ROOT`` is implementation dependent and varies between
+    different implementations of :py:class:`~torchx.workspace.api.WorkspaceMixin`.
+    Check the scheduler documentation for details on which workspace it supports.
+
+    Note: ``projects`` maps the location of the local project to a sub-directory in the remote workspace root directory.
+    Typically the local project location is a directory path (e.g. ``/home/foo/github/torch``).
+
+
+    Attributes:
+        projects: mapping of local project to the sub-dir in the remote workspace dir.
+    """
+
+    projects: dict[str, str]
+
+    def __bool__(self) -> bool:
+        """False if no projects mapping. Lets us use workspace object in an if-statement"""
+        return bool(self.projects)
+
+    def __eq__(self, other: object) -> bool:
+        if not isinstance(other, Workspace):
+            return False
+        return self.projects == other.projects
+
+    def __hash__(self) -> int:
+        # makes it possible to use Workspace as the key in the workspace build cache
+        # see WorkspaceMixin.caching_build_workspace_and_update_role
+        return hash(frozenset(self.projects.items()))
+
+    def is_unmapped_single_project(self) -> bool:
+        """
+        Returns ``True`` if this workspace only has 1 project
+        and its target mapping is an empty string.
+        """
+        return len(self.projects) == 1 and not next(iter(self.projects.values()))
+
+    def merge_into(self, outdir: str | pathlib.Path) -> None:
+        """
+        Copies each project dir of this workspace into the specified ``outdir``.
+        Each project dir is copied into ``{outdir}/{target}`` where ``target`` is
+        the target mapping of the project dir.
+
+        For example:
+
+        .. code-block:: python
+            from os.path import expanduser
+
+            workspace = Workspace(
+                projects={
+                    expanduser("~/workspace/torch"): "torch",
+                    expanduser("~/workspace/my_project": "")
+                }
+            )
+            workspace.merge_into(expanduser("~/tmp"))
+
+        Copies:
+
+            * ``~/workspace/torch/**`` into ``~/tmp/torch/**``
+            * ``~/workspace/my_project/**`` into ``~/tmp/**``
+
+        """
+
+        for src, dst in self.projects.items():
+            dst_path = pathlib.Path(outdir) / dst
+            if pathlib.Path(src).is_file():
+                shutil.copy2(src, dst_path)
+            else:  # src is dir
+                shutil.copytree(src, dst_path, dirs_exist_ok=True)
+
+    @staticmethod
+    def from_str(workspace: str | None) -> "Workspace":
+        import yaml
+
+        if not workspace:
+            return Workspace({})
+
+        projects = yaml.safe_load(workspace)
+        if isinstance(projects, str):  # single project workspace
+            projects = {projects: ""}
+        else:  # multi-project workspace
+            # Replace None mappings with "" (empty string)
+            projects = {k: ("" if v is None else v) for k, v in projects.items()}
+
+        return Workspace(projects)
+
+    def __str__(self) -> str:
+        """
+        Returns a string representation of the Workspace by concatenating
+        the project mappings using ';' as a delimiter and ':' between key and value.
+        If the single-project workspace with no target mapping, then simply
+        returns the src (local project dir)
+
+        NOTE: meant to be used for logging purposes not serde.
+          Therefore not symmetric with :py:func:`Workspace.from_str`.
+
+        """
+        if self.is_unmapped_single_project():
+            return next(iter(self.projects))
+        else:
+            return ";".join(
+                k if not v else f"{k}:{v}" for k, v in self.projects.items()
+            )
+
+
 @dataclass
 class Role:
     """
@@ -371,12 +519,15 @@ class Role:
             metadata: Free form information that is associated with the role, for example
                 scheduler specific data. The key should follow the pattern: ``$scheduler.$key``
             mounts: a list of mounts on the machine
+            workspace: local project directories to be mirrored on the remote job.
+              NOTE: The workspace argument provided to the :py:class:`~torchx.runner.api.Runner` APIs
+              only takes effect on ``appdef.role[0]`` and overrides this attribute.
+
     """
 
     name: str
     image: str
     min_replicas: Optional[int] = None
-    base_image: Optional[str] = None  # DEPRECATED DO NOT SET, WILL BE REMOVED SOON
     entrypoint: str = MISSING
     args: List[str] = field(default_factory=list)
     env: Dict[str, str] = field(default_factory=dict)
@@ -386,9 +537,10 @@ class Role:
     resource: Resource = field(default_factory=_null_resource)
     port_map: Dict[str, int] = field(default_factory=dict)
     metadata: Dict[str, Any] = field(default_factory=dict)
-    mounts: List[Union[BindMount, VolumeMount, DeviceMount]] = field(
-        default_factory=list
-    )
+    mounts: List[BindMount | VolumeMount | DeviceMount] = field(default_factory=list)
+    workspace: Workspace | None = None
+
+    # DEPRECATED DO NOT SET, WILL BE REMOVED SOON
     overrides: Dict[str, Any] = field(default_factory=dict)
 
     # pyre-ignore
@@ -788,6 +940,62 @@ class runopt:
     opt_type: Type[CfgVal]
     is_required: bool
     help: str
+    aliases: list[str] | None = None
+    deprecated_aliases: list[str] | None = None
+
+    @property
+    def is_type_list_of_str(self) -> bool:
+        """
+        Checks if the option type is a list of strings.
+
+        Returns:
+            bool: True if the option type is either List[str] or list[str], False otherwise.
+        """
+        return self.opt_type in (List[str], list[str])
+
+    @property
+    def is_type_dict_of_str(self) -> bool:
+        """
+        Checks if the option type is a dict of string keys to string values.
+
+        Returns:
+            bool: True if the option type is either Dict[str, str] or dict[str, str], False otherwise.
+        """
+        return self.opt_type in (Dict[str, str], dict[str, str])
+
+    def cast_to_type(self, value: str) -> CfgVal:
+        """Casts the given `value` (in its string representation) to the type of this run option.
+        Below are the cast rules for each option type and value literal:
+
+        1. opt_type=str, value="foo" -> "foo"
+        1. opt_type=bool, value="True"/"False" -> True/False
+        1. opt_type=int, value="1" -> 1
+        1. opt_type=float, value="1.1" -> 1.1
+        1. opt_type=list[str]/List[str], value="a,b,c" or value="a;b;c" -> ["a", "b", "c"]
+        1. opt_type=dict[str,str]/Dict[str,str],
+           value="key1:val1,key2:val2" or value="key1:val1;key2:val2" -> {"key1": "val1", "key2": "val2"}
+
+        NOTE: dict parsing uses ":" as the kv separator (rather than the standard "=") because "=" is used
+        at the top-level cfg to parse runopts (notice the plural) from the CLI. Originally torchx only supported
+        primitives and list[str] as CfgVal but dict[str,str] was added in https://github.com/meta-pytorch/torchx/pull/855
+        """
+
+        if self.opt_type is None:
+            raise ValueError("runopt's opt_type cannot be `None`")
+        elif self.opt_type == bool:
+            return value.lower() == "true"
+        elif self.opt_type in (List[str], list[str]):
+            # lists may be ; or , delimited
+            # also deal with trailing "," by removing empty strings
+            return [v for v in value.replace(";", ",").split(",") if v]
+        elif self.opt_type in (Dict[str, str], dict[str, str]):
+            return {
+                s.split(":", 1)[0]: s.split(":", 1)[1]
+                for s in value.replace(";", ",").split(",")
+            }
+        else:
+            assert self.opt_type in (str, int, float)
+            return self.opt_type(value)
 
 
 class runopts:
@@ -825,6 +1033,7 @@ class runopts:
 
     def __init__(self) -> None:
         self._opts: Dict[str, runopt] = {}
+        self._alias_to_key: dict[str, str] = {}
 
     def __iter__(self) -> Iterator[Tuple[str, runopt]]:
         return self._opts.items().__iter__()
@@ -852,9 +1061,16 @@ class runopts:
 
     def get(self, name: str) -> Optional[runopt]:
         """
-        Returns option if any was registered, or None otherwise
+        Returns option if any was registered, or None otherwise.
+        First searches for the option by ``name``, then falls-back to matching ``name`` with any
+        registered aliases.
+
         """
-        return self._opts.get(name, None)
+        if name in self._opts:
+            return self._opts[name]
+        if name in self._alias_to_key:
+            return self._opts[self._alias_to_key[name]]
+        return None
 
     def resolve(self, cfg: Mapping[str, CfgVal]) -> Dict[str, CfgVal]:
         """
@@ -869,6 +1085,36 @@ class runopts:
 
         for cfg_key, runopt in self._opts.items():
             val = resolved_cfg.get(cfg_key)
+            resolved_name = None
+            aliases = runopt.aliases or []
+            deprecated_aliases = runopt.deprecated_aliases or []
+            if val is None:
+                for alias in aliases:
+                    val = resolved_cfg.get(alias)
+                    if alias in cfg or val is not None:
+                        resolved_name = alias
+                        break
+                for alias in deprecated_aliases:
+                    val = resolved_cfg.get(alias)
+                    if val is not None:
+                        resolved_name = alias
+                        use_instead = self._alias_to_key.get(alias)
+                        warnings.warn(
+                            f"Run option `{alias}` is deprecated, use `{use_instead}` instead",
+                            UserWarning,
+                            stacklevel=2,
+                        )
+                        break
+            else:
+                resolved_name = cfg_key
+                for alias in aliases:
+                    duplicate_val = resolved_cfg.get(alias)
+                    if alias in cfg or duplicate_val is not None:
+                        raise InvalidRunConfigException(
+                            f"Duplicate opt name. runopt: `{resolved_name}``, is an alias of runopt: `{alias}`",
+                            resolved_name,
+                            cfg,
+                        )
 
             # check required opt
             if runopt.is_required and val is None:
@@ -888,7 +1134,7 @@ class runopts:
                 )
 
             # not required and not set, set to default
-            if val is None:
+            if val is None and resolved_name is None:
                 resolved_cfg[cfg_key] = runopt.default
         return resolved_cfg
 
@@ -948,27 +1194,11 @@ class runopts:
 
         """
 
-        def _cast_to_type(value: str, opt_type: Type[CfgVal]) -> CfgVal:
-            if opt_type == bool:
-                return value.lower() == "true"
-            elif opt_type == List[str]:
-                # lists may be ; or , delimited
-                # also deal with trailing "," by removing empty strings
-                return [v for v in value.replace(";", ",").split(",") if v]
-            elif opt_type == Dict[str, str]:
-                return {
-                    s.split(":", 1)[0]: s.split(":", 1)[1]
-                    for s in value.replace(";", ",").split(",")
-                }
-            else:
-                # pyre-ignore[19, 6] type won't be dict here as we handled it above
-                return opt_type(value)
-
         cfg: Dict[str, CfgVal] = {}
         for key, val in to_dict(cfg_str).items():
-            runopt_ = self.get(key)
-            if runopt_:
-                cfg[key] = _cast_to_type(val, runopt_.opt_type)
+            opt = self.get(key)
+            if opt:
+                cfg[key] = opt.cast_to_type(val)
             else:
                 logger.warning(
                     f"{YELLOW_BOLD}Unknown run option passed to scheduler: {key}={val}{RESET}"
@@ -982,16 +1212,16 @@ class runopts:
         cfg: Dict[str, CfgVal] = {}
         cfg_dict = json.loads(json_repr)
         for key, val in cfg_dict.items():
-            runopt_ = self.get(key)
-            if runopt_:
+            opt = self.get(key)
+            if opt:
                 # Optional runopt cfg values default their value to None,
                 # but use `_type` to specify their type when provided.
                 # Make sure not to treat None's as lists/dictionaries
                 if val is None:
                     cfg[key] = val
-                elif runopt_.opt_type == List[str]:
+                elif opt.is_type_list_of_str:
                     cfg[key] = [str(v) for v in val]
-                elif runopt_.opt_type == Dict[str, str]:
+                elif opt.is_type_dict_of_str:
                     cfg[key] = {str(k): str(v) for k, v in val.items()}
                 else:
                     cfg[key] = val
@@ -1004,12 +1234,16 @@ class runopts:
         help: str,
         default: CfgVal = None,
         required: bool = False,
+        aliases: Optional[list[str]] = None,
+        deprecated_aliases: Optional[list[str]] = None,
     ) -> None:
         """
         Adds the ``config`` option with the given help string and ``default``
         value (if any). If the ``default`` is not specified then this option
         is a required option.
         """
+        aliases = aliases or []
+        deprecated_aliases = deprecated_aliases or []
         if required and default is not None:
             raise ValueError(
                 f"Required option: {cfg_key} must not specify default value. Given: {default}"
@@ -1021,7 +1255,19 @@ class runopts:
                     f" Given: {default} ({type(default).__name__})"
                 )
 
-        self._opts[cfg_key] = runopt(default, type_, required, help)
+        opt = runopt(
+            default,
+            type_,
+            required,
+            help,
+            list(set(aliases)),
+            list(set(deprecated_aliases)),
+        )
+        for alias in aliases:
+            self._alias_to_key[alias] = cfg_key
+        for deprecated_alias in deprecated_aliases:
+            self._alias_to_key[deprecated_alias] = cfg_key
+        self._opts[cfg_key] = opt
 
     def update(self, other: "runopts") -> None:
         self._opts.update(other._opts)