From d5779c7900b63e7df791645e6a5a1302f177ad00 Mon Sep 17 00:00:00 2001 From: Eitan Turok Date: Thu, 26 Sep 2024 22:27:34 +0000 Subject: [PATCH] warn checkpointing does not work --- llmfoundry/utils/builders.py | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-) diff --git a/llmfoundry/utils/builders.py b/llmfoundry/utils/builders.py index 12e734446c..b1e1a303fe 100644 --- a/llmfoundry/utils/builders.py +++ b/llmfoundry/utils/builders.py @@ -7,14 +7,9 @@ import logging import os import re +import warnings from collections import OrderedDict -from typing import ( - Any, - ContextManager, - Iterable, - Optional, - Union, -) +from typing import Any, ContextManager, Iterable, Optional, Union import torch from composer.core import Algorithm, Callback, Evaluator @@ -711,6 +706,10 @@ def build_tp_strategies( name: str, model: ComposerModel, ) -> dict[str, ParallelStyle]: + + warnings.warn( + 'Checkpointing is not currently supported for tensor parallelism due to this pytorch bug: https://github.com/pytorch/pytorch/issues/134095#issuecomment-2345018244' + ) return construct_from_registry( name=name, registry=registry.tp_strategies,