SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Decouple device_map='auto' and tp_plan='auto' (#38942)

Browse Source 
* dissociate

* better place

* fix
This commit is contained in:
Marc Sun
2025-07-03 11:07:11 +02:00
committed by GitHub
parent 8178c43112
commit bff964c429
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/transformers/modeling_utils.py
View File

@@ -4431,10 +4431,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, PushToHubMixin, PeftAdapterMi
"`tp_plan` and `device_map` are mutually exclusive. Choose either one for parallelization." "`tp_plan` and `device_map` are mutually exclusive. Choose either one for parallelization."
) )
# If torchrun was used, make sure to TP by default. This way people don't need to change tp or device map if device_map == "auto" and int(os.environ.get("WORLD_SIZE", 0)):
if device_map == "auto" and tp_plan is None and int(os.environ.get("WORLD_SIZE", 0)): logger.info(
tp_plan = "auto" # device_map = "auto" in torchrun equivalent to TP plan = AUTO! "You've set device_map=`auto` while triggering a distributed run with torchrun. This might lead to unexpected behavior. "
device_map = None "If your plan is to load the model on each device, you should set device_map={"
": PartialState().process_index} where PartialState comes from accelerate library"
)
# We need to correctly dispatch the model on the current process device. The easiest way for this is to use a simple # We need to correctly dispatch the model on the current process device. The easiest way for this is to use a simple
# `device_map` pointing to the correct device # `device_map` pointing to the correct device