From f3393633e4230c84dfe260cd739c4cb83a1d1057 Mon Sep 17 00:00:00 2001 From: Howard Huang Date: Thu, 15 Aug 2024 23:30:41 -0400 Subject: [PATCH] Fail when using tracer made without seed checkpoint (#522) Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #473 * __->__ #522 --- train.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/train.py b/train.py index d297b8a7..a6bcbdd8 100644 --- a/train.py +++ b/train.py @@ -204,6 +204,11 @@ def loss_fn(pred, labels): checkpoint_loaded = checkpoint.load() if parallel_dims.pp_enabled and not checkpoint_loaded: + if job_config.experimental.pipeline_parallel_split_mode == "tracer": + raise RuntimeError( + "Pipeline parallelism with tracer mode is not supported without a seed checkpoint." + ) + # TODO: fix this by allowing each rank to set their own seed logger.warning( "Pipeline Parallelism is being used without a seed checkpoint. "