run par as an entrypoint if there is no patch or jetter patch. (pytor…

…ch#994)

Summary:
Pull Request resolved: pytorch#994

# Context:
Currently, when running torchx local job, we are using penv_python as entrypoint. That means we pass the actual .par or .xar file as argument to penv_python.  within penv_python, the par/xar is executed as a new process.

# Old way to run torchx local job.

For example, if the local job is running "jetter --help", torchx runs it like:
  PENV_PAR='/data/users/yikai/fbsource/buck-out/v2/gen/fbcode/a6cb9616985b22b0/jetter/__jetter-bin__/jetter-bin-inplace.par' penv_python -m jetter.main --help
It passes the par file as an environment variable called "PENV_PAR"(There is another way to pass this to penv_python, which is passing 'PENV_PARNAME' as env variable then get the par file's path using it. But it is very very rare, only 0.1% of total usage.)

# New way to run torchx local job
After migration, We will run it like:
  PAR_MAIN_OVERRIDE=jetter.main /data/users/yikai/fbsource/buck-out/v2/gen/fbcode/a6cb9616985b22b0/jetter/__jetter-bin__/jetter-bin-inplace.par --help

NOTE: This diff only migrates one of the most common use cases, which: 1. There are no patch or jetter patch. 2. it's a par not xar. 3. the par file is passed via "PENV_PAR" env variable.  For other use cases, we still run penv_python as entrypoint.

Reviewed By: Sanjay-Ganeshan

Differential Revision: D66621649

torchx/schedulers/local_scheduler.py

@@ -28,6 +28,7 @@
 import warnings
 from dataclasses import asdict, dataclass
 from datetime import datetime
+from subprocess import Popen
 from types import FrameType
 from typing import (
     Any,
@@ -696,12 +697,11 @@ def _popen(
         log.debug(f"Running {role_name} (replica {replica_id}):\n {args_pfmt}")
         env = self._get_replica_env(replica_params)
 
-        proc = subprocess.Popen(
+        proc = self.run_local_job(
             args=replica_params.args,
             env=env,
             stdout=stdout_,
             stderr=stderr_,
-            start_new_session=True,
             cwd=replica_params.cwd,
         )
         return _LocalReplica(
@@ -714,6 +714,23 @@ def _popen(
             error_file=env.get("TORCHELASTIC_ERROR_FILE", "<N/A>"),
         )
 
+    def run_local_job(
+        self,
+        args: List[str],
+        env: Dict[str, str],
+        stdout: Optional[io.FileIO],
+        stderr: Optional[io.FileIO],
+        cwd: Optional[str] = None,
+    ) -> Popen[bytes]:
+        return subprocess.Popen(
+            args=args,
+            env=env,
+            stdout=stdout,
+            stderr=stderr,
+            start_new_session=True,
+            cwd=cwd,
+        )
+
     def _get_replica_output_handles(
         self,
         replica_params: ReplicaParam,