Skip to content

[vllm] feat: add model saving func to dump vllm state_dict before and after update weights #5278

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
RobotGF wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[vllm] feat: add model saving func to dump vllm state_dict before and after update weights #5278

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7a7217f
[vllm] feat: add model saving functionality to safetensors before and…
RobotGF Feb 11, 2026
987cbbd
[vllm] feat: enhance model saving path to include device UUID
RobotGF Feb 11, 2026
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
25 changes: 25 additions & 0 deletions verl/workers/rollout/vllm_rollout/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -218,6 +218,9 @@ def update_weights_from_ipc(self, peft_config: dict = None, base_sync_done=False
if use_standard_weight_load:
patch_vllm_moe_model_weight_loader(self.model_runner.model)

# save vllm model to safetensors before weight update
self._save_vllm_model_to_safetensors(suffix="before")

# receive bucket and update weights
while True:
metadata = socket.recv_pyobj()
Expand Down Expand Up @@ -249,6 +252,9 @@ def update_weights_from_ipc(self, peft_config: dict = None, base_sync_done=False
model_config = self.model_runner.vllm_config.model_config
process_weights_after_loading(model, model_config, self.device)

# save vllm model to safetensors after weight update
self._save_vllm_model_to_safetensors(suffix="after")

# clean up
socket.close()
del buffer
Expand Down Expand Up @@ -284,6 +290,25 @@ def _update_weights(self, weights: list[tuple[str, torch.Tensor]], peft_config:
logger.info("Loading standard weights (non-FP8, async)")
self.model_runner.model.load_weights(weights)

def _save_vllm_model_to_safetensors(self, suffix: str):
"""Save vLLM model to safetensors via save_model.
Dir from env VERL_VLLM_WEIGHT_SAVE_DIR (default None for not saving).
Args:
suffix: The suffix of the saved model.
"""
save_dir = os.environ.get("VERL_VLLM_WEIGHT_SAVE_DIR", None)
if save_dir is None:
return
os.makedirs(save_dir, exist_ok=True)
model = self.model_runner.model
rank = getattr(self, "local_rank", 0)
device_uuid = get_device_uuid(self.device.index)
path = os.path.join(save_dir, f"vllm_weights_{suffix}_rank{rank}_device{device_uuid}.safetensors")
from safetensors.torch import save_model

save_model(model, path)
logger.info(f"vLLM model saved to {path}")

def _get_zmq_handle(self) -> str:
"""Get ZMQ handle for communication."""
if not hasattr(self, "device_uuid") or not self.device_uuid:
Expand Down