Skip to content

Save bf16 model #414

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

Merged
pan-x-c merged 1 commit into from
Nov 27, 2025
Merged

Save bf16 model #414

Show file tree
Hide file tree
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
2 changes: 2 additions & 0 deletions trinity/trainer/verl/fsdp_checkpoint_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -360,6 +360,8 @@ def save_checkpoint( # noqa: C901
if self._save_model_thread is not None:
self._save_model_thread.join()

state_dict = {k: v.to(torch.bfloat16) for k, v in state_dict.items()}

def _save_model():
runtime_context = ray.get_runtime_context()
node_id = runtime_context.get_node_id()
Expand Down
5 changes: 3 additions & 2 deletions trinity/trainer/verl/megatron_checkpoint_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -257,15 +257,16 @@ def save_checkpoint( # noqa: C901
from transformers import MistralForSequenceClassification

model = MistralForSequenceClassification.from_pretrained(
self.config.model.path
self.config.model.path, torch_dtype=torch.bfloat16
) # use score head instead of lm_head
state_dict["score.weight"] = state_dict["score.weight"]
else:
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
self.config.model.path, torch_dtype="auto"
self.config.model.path, torch_dtype=torch.bfloat16
)
state_dict = {k: v.to(torch.bfloat16) for k, v in state_dict.items()}
model.save_pretrained(hf_model_ckpt_path, state_dict=state_dict)
log_with_rank(
f"Saved Huggingface config and tokenizer to {hf_model_ckpt_path}",
Expand Down