improve profiler

torchtitan/config/job_config.py

@@ -34,6 +34,12 @@ class Profiling:
     profile_freq: int = 10
     """How often to collect profile traces, in iterations"""
 
+    profiler_active: int = 1
+    """The steps profiler is active for"""
+
+    profiler_warmup: int = 3
+    """The number of warmup steps before the active step in each profiling cycle"""
+
     enable_memory_snapshot: bool = False
     """Whether to dump memory snapshot"""
 

torchtitan/tools/profiling.py

@@ -14,9 +14,6 @@
 from torchtitan.config import Profiling as ProfilingConfig
 from torchtitan.tools.logging import logger
 
-# the number of warmup steps before the active step in each profiling cycle
-WARMUP = 3
-
 # how much memory allocation/free ops to record in memory snapshots
 MEMORY_SNAPSHOT_MAX_ENTRIES = 100000
 
@@ -58,7 +55,10 @@ def trace_handler(prof):
         if not os.path.exists(trace_dir):
             os.makedirs(trace_dir, exist_ok=True)
 
-        warmup, active = WARMUP, 1
+        warmup, active = (
+            profiling_config.profiler_warmup,
+            profiling_config.profiler_active,
+        )
         wait = profile_freq - (active + warmup)
         assert (
             wait >= 0

torchtitan/train.py

@@ -6,6 +6,7 @@
 
 import importlib
 import os
+import signal
 import time
 from datetime import timedelta
 from typing import Any, Generator, Iterable, Optional
@@ -588,6 +589,17 @@ def train(self):
                 ),
             ),
         ):
+            if torch_profiler:
+
+                def sigabrt_handler(signal, frame):
+                    logger.info("SIGABRT received. Stopping profiler")
+                    for _ in range(config.profiling.profiler_active):
+                        # Step the profiler enough times to trigger a trace
+                        torch_profiler.step()
+                    torch_profiler.stop()
+
+                signal.signal(signal.SIGABRT, sigabrt_handler)
+
             data_iterator = self.batch_generator(self.dataloader)
             while self.should_continue_training():
                 self.step += 1