Avoid splitting files if num_threads is 1 (#787)

LucasLLC · facebook-github-bot · commit 34b04aea1aef · 2024-04-17T12:41:20.000-07:00
Summary: Pull Request resolved: #787 Splitting files only provides performance gains in the threaded case, additionally it's been noted that splitting files increases likelihood of hitting manifold quotas ghstack-source-id: 222855206 exported-using-ghexport Reviewed By: JKSenthil Differential Revision: D56065261 fbshipit-source-id: 65d037286a7ced0f2f968ae6ed095e360c6c435c
diff --git a/torchtnt/framework/callbacks/dcp_saver.py b/torchtnt/framework/callbacks/dcp_saver.py
@@ -272,7 +272,8 @@ def default_writer_options(self) -> Dict[str, Any]:
         dcp_options = {
             "thread_count": self._knob_options.max_per_rank_io_concurrency or 16,
             "sync_files": False,
-            "single_file_per_rank": False,
         }
+        if dcp_options["thread_count"] > 1:
+            dcp_options["single_file_per_rank"] = False
 
         return dcp_options

Original file line number	Diff line number	Diff line change
`@@ -272,7 +272,8 @@ def default_writer_options(self) -> Dict[str, Any]:`
`272`	`272`	`dcp_options = {`
`273`	`273`	`"thread_count": self._knob_options.max_per_rank_io_concurrency or 16,`
`274`	`274`	`"sync_files": False,`
`275`		`- "single_file_per_rank": False,`
`276`	`275`	`}`
	`276`	`+ if dcp_options["thread_count"] > 1:`
	`277`	`+ dcp_options["single_file_per_rank"] = False`
`277`	`278`
`278`	`279`	`return dcp_options`