Skip to content

Skip layerwise reload for non-quantized models #48284

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
yuchenwang3 wants to merge 1 commit into
base: main
Choose a base branch
from
+10 −4
Open

Skip layerwise reload for non-quantized models #48284

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 10 additions & 4 deletions vllm/v1/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5492,10 +5492,16 @@ def reload_weights(
# begin loading weights
logger.info_once("Reloading weights inplace...")
if is_checkpoint_format:
# load weights from checkpoint/ original model format
initialize_layerwise_reload(model)
loaded_weights = model.load_weights(weights_iterator)
finalize_layerwise_reload(model, self.model_config)
# Layerwise reload only exists to defer quantization repacking;
# for non-quantized models it is unnecessary, and its per-layer
# reprocessing corrupts hybrid (Mamba) weights on reload. Use the
# plain startup load path instead.
if getattr(self.model_config, "quantization", None) is None:
loaded_weights = model.load_weights(weights_iterator)
else:
initialize_layerwise_reload(model)
loaded_weights = model.load_weights(weights_iterator)
finalize_layerwise_reload(model, self.model_config)

else:
# load weights from kernel format
Expand Down
Loading