reset flight recorder trace (meta-pytorch#283)

Summary:

- call FR api to reset the trace after every quorum - we reset so that after every quorum, we start a fresh FR trace since the pg's could have changed and we already dumped FR trace from previous errors
- change the env var that's used to determine the file after every quorum

Differential Revision: D84260745

Author: tushar00jain

torchft/manager.py

@@ -88,6 +88,8 @@
 # crash if call to quorum fails, all replicas will crash.
 QUORUM_RETRIES_ENV: str = "TORCHFT_QUORUM_RETRIES"
 
+TORCH_FR_DUMP_TEMP_FILE_ENV: str = "TORCH_FR_DUMP_TEMP_FILE"
+
 T = TypeVar("T")
 
 
@@ -109,6 +111,17 @@ def get_timeout(
     return default_timeout_sec
 
 
+def extract_trailing_digits(s: str) -> int:
+    """
+    Extracts the trailing digits from the end of the string s.
+    Returns an empty string if no trailing digits are found.
+    """
+    i = len(s) - 1
+    while i >= 0 and s[i].isdigit():
+        i -= 1
+    return int(s[i + 1 :]) if i < len(s) - 1 else 0
+
+
 class WorldSizeMode(Enum):
     """
     This controls the numerics for the job when doing allreduces across replicas
@@ -223,6 +236,9 @@ def __init__(
         self._load_state_dict_fns: Dict[str, Callable[[object], None]] = {}
         self._user_state_dicts: Dict[str, Callable[[], object]] = {}
 
+        self._original_fr_dump_temp_file: Optional[str] = os.environ.get(
+            TORCH_FR_DUMP_TEMP_FILE_ENV
+        )
         self._replica_id = replica_id
 
         # Protects state dict
@@ -257,7 +273,7 @@ def __init__(
         store_port = store_port or int(os.environ["MASTER_PORT"])
         self._group_rank: int = rank if rank is not None else int(os.environ["RANK"])
         group_rank = self._group_rank
-        group_world_size = world_size or int(os.environ["WORLD_SIZE"])
+        self._group_world_size: int = world_size or int(os.environ["WORLD_SIZE"])
         self._min_replica_size = min_replica_size
 
         if checkpoint_transport is None:
@@ -310,7 +326,7 @@ def __init__(
                 hostname=hostname,
                 bind=bind,
                 store_addr=f"{store_addr}:{store_port}",
-                world_size=group_world_size,
+                world_size=self._group_world_size,
                 heartbeat_interval=heartbeat_interval,
                 connect_timeout=connect_timeout,
                 quorum_retries=self._quorum_retries,
@@ -338,6 +354,8 @@ def __init__(
         self._participating_replica_world_size: int = 0
         self._is_state_dict_read_allowed = True
 
+        self._update_fr_path()
+
     def allow_state_dict_read(self) -> None:
         if self._is_state_dict_read_allowed:
             return
@@ -665,16 +683,20 @@ def _async_quorum(
             self._logger.info(f"reconfiguring for {quorum_id=} {store_prefixed_addr=}")
             # We use the replica rank and world as we want all replicas in the PG.
             try:
+                self._quorum_id = quorum_id
                 with torch.profiler.record_function("torchft::manager::_pg::configure"):
+                    # Reset GPU state for Flight Recorder
                     if torch.accelerator.is_available():
                         torch.accelerator.synchronize()
+                    torch._C._distributed_c10d._reset_fr_recording_nccl()
+                    self._update_fr_path()
+
                     self._pg.configure(
                         store_prefixed_addr,
                         self._replica_id if self._replica_id is not None else "0",
                         replica_rank,
                         replica_world_size,
                     )
-                self._quorum_id = quorum_id
             except Exception as e:
                 self._logger.exception(f"got exception in pg configure: {e}")
                 self.report_error(e)
@@ -749,6 +771,21 @@ def _async_quorum(
                     else None
                 )
 
+    def _update_fr_path(self) -> None:
+        if self._original_fr_dump_temp_file is not None:
+            folder = f"{self._original_fr_dump_temp_file}_quorum_{self._quorum_id}"
+            os.makedirs(folder, exist_ok=True)
+
+            filename = (
+                self._group_rank
+                if self._replica_id is None
+                else (
+                    extract_trailing_digits(self._replica_id) * self._group_world_size
+                    + self._group_rank
+                )
+            )
+            os.environ[TORCH_FR_DUMP_TEMP_FILE_ENV] = f"{folder}/{filename}"
+
     def _apply_pending_state_dict(self) -> None:
         assert self._healing, "must be in healing state"