Allow aliases for run_opt (meta-pytorch#1141)

Summary:

Lets allow aliases for a runopt. This will give downstream users to have multiple ways of accessing the same runopt. 
1. Modify add() to accept this and add a new optional aliases List[str] that will store this. 
2. Modify resolve() to check if a different alias is already used in cfg i.e if the "run_as" and "Run_As" are aliases for the same one, we dont allow for both to be present in the cfg.

Differential Revision: D84157870

Author: AbishekS

torchx/specs/api.py

@@ -886,6 +886,10 @@ def get_type_name(tp: Type[CfgVal]) -> str:
         return str(tp)
 
 
+class RunOptAlias(str):
+    pass
+
+
 @dataclass
 class runopt:
     """
@@ -896,6 +900,7 @@ class runopt:
     opt_type: Type[CfgVal]
     is_required: bool
     help: str
+    aliases: list[RunOptAlias] | None = None
 
     @property
     def is_type_list_of_str(self) -> bool:
@@ -987,6 +992,7 @@ class runopts:
 
     def __init__(self) -> None:
         self._opts: Dict[str, runopt] = {}
+        self._alias_to_key: Dict[RunOptAlias, str] = {}
 
     def __iter__(self) -> Iterator[Tuple[str, runopt]]:
         return self._opts.items().__iter__()
@@ -1016,7 +1022,11 @@ def get(self, name: str) -> Optional[runopt]:
         """
         Returns option if any was registered, or None otherwise
         """
-        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]:
         """
@@ -1031,6 +1041,24 @@ def resolve(self, cfg: Mapping[str, CfgVal]) -> Dict[str, CfgVal]:
 
         for cfg_key, runopt in self._opts.items():
             val = resolved_cfg.get(cfg_key)
+            resolved_name = None
+            aliases = runopt.aliases or []
+            if val is None:
+                for alias in aliases:
+                    val = resolved_cfg.get(alias)
+                    if val is not None:
+                        resolved_name = alias
+                        break
+            else:
+                resolved_name = cfg_key
+                for alias in aliases:
+                    duplicate_val = resolved_cfg.get(alias)
+                    if duplicate_val is not None:
+                        raise InvalidRunConfigException(
+                            f"Run option: {resolved_name}, is an alias of another run option already used in the cfg.",
+                            resolved_name,
+                            cfg,
+                        )
 
             # check required opt
             if runopt.is_required and val is None:
@@ -1143,9 +1171,38 @@ def cfg_from_json_repr(self, json_repr: str) -> Dict[str, CfgVal]:
                     cfg[key] = val
         return cfg
 
+    def _get_primary_key_and_aliases(
+        self,
+        cfg_key: list[str] | str,
+    ) -> tuple[str, list[RunOptAlias]]:
+        """
+        Returns the primary key and aliases for the given cfg_key.
+        """
+        if isinstance(cfg_key, str):
+            return cfg_key, []
+
+        if len(cfg_key) == 0:
+            raise ValueError("cfg_key must be a non-empty list")
+        primary_key = None
+        aliases = list[RunOptAlias]()
+        for name in cfg_key:
+            if isinstance(name, RunOptAlias):
+                aliases.append(name)
+            else:
+                if primary_key is not None:
+                    raise ValueError(
+                        f"cfg_key must contain a single primary key. Given more than one primary keys: {primary_key}, {name}. If one of them is an alias please use RunOptAlias type instead of str. "
+                    )
+                primary_key = name
+        if primary_key is None or primary_key == "":
+            raise ValueError(
+                "Missing cfg_key. Please provide one other than the aliases."
+            )
+        return primary_key, aliases
+
     def add(
         self,
-        cfg_key: str,
+        cfg_key: str | list[str],
         type_: Type[CfgVal],
         help: str,
         default: CfgVal = None,
@@ -1156,6 +1213,7 @@ def add(
         value (if any). If the ``default`` is not specified then this option
         is a required option.
         """
+        primary_key, aliases = self._get_primary_key_and_aliases(cfg_key)
         if required and default is not None:
             raise ValueError(
                 f"Required option: {cfg_key} must not specify default value. Given: {default}"
@@ -1166,8 +1224,10 @@ def add(
                     f"Option: {cfg_key}, must be of type: {type_}."
                     f" Given: {default} ({type(default).__name__})"
                 )
-
-        self._opts[cfg_key] = runopt(default, type_, required, help)
+        opt = runopt(default, type_, required, help, aliases)
+        for alias in aliases:
+            self._alias_to_key[alias] = primary_key
+        self._opts[primary_key] = opt
 
     def update(self, other: "runopts") -> None:
         self._opts.update(other._opts)

torchx/specs/test/api_test.py

@@ -41,6 +41,7 @@
     Role,
     RoleStatus,
     runopt,
+    RunOptAlias,
     runopts,
     TORCHX_HOME,
     Workspace,
@@ -578,6 +579,29 @@ def test_runopts_add(self) -> None:
         # this print is intentional (demonstrates the intended usecase)
         print(opts)
 
+    def test_runopts_add_with_aliases(self) -> None:
+        opts = runopts()
+        opts.add(
+            ["job_priority", RunOptAlias("jobPriority")],
+            type_=str,
+            help="priority for the job",
+        )
+        self.assertEqual(1, len(opts._opts))
+        self.assertIsNotNone(opts.get("job_priority"))
+        self.assertIsNotNone(opts.get("jobPriority"))
+
+    def test_runopts_resolve_with_aliases(self) -> None:
+        opts = runopts()
+        opts.add(
+            ["job_priority", RunOptAlias("jobPriority")],
+            type_=str,
+            help="priority for the job",
+        )
+        opts.resolve({"job_priority": "high"})
+        opts.resolve({"jobPriority": "low"})
+        with self.assertRaises(InvalidRunConfigException):
+            opts.resolve({"job_priority": "high", "jobPriority": "low"})
+
     def get_runopts(self) -> runopts:
         opts = runopts()
         opts.add("run_as", type_=str, help="run as user", required=True)