Skip to content

Fix multimodal hidden-state preparation for Qwen3-VL models #526

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
liusy58 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix multimodal hidden-state preparation for Qwen3-VL models #526

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4f43509
fix
liusy58 Apr 8, 2026
dab3ff9
fix lint
liusy58 Apr 9, 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
21 changes: 16 additions & 5 deletions scripts/prepare_hidden_states.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,6 +72,8 @@ class DataPoint:
loss_mask: torch.Tensor
hidden_state: torch.Tensor
aux_hidden_state: Optional[torch.Tensor] = None
pixel_values: Optional[torch.Tensor] = None
image_grid_thw: Optional[torch.Tensor] = None


def parse_args():
Expand Down Expand Up @@ -187,14 +189,19 @@ def build_target_model(
)
else:
target_model_kwargs = SGLangBackendArgs.from_args(args).to_kwargs()
if hasattr(model_config, "dtype") and model_config.dtype is not None:
torch_dtype = model_config.dtype
elif hasattr(model_config, "text_config") and hasattr(
model_config.text_config, "dtype"
):
torch_dtype = model_config.text_config.dtype
else:
torch_dtype = getattr(model_config, "torch_dtype", "bfloat16")
Comment on lines +192 to +199
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Apr 8, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The logic for determining torch_dtype can be simplified and made more robust by using getattr with defaults and explicitly checking for None values in nested configurations.

Suggested change
if hasattr(model_config, "dtype") and model_config.dtype is not None:
torch_dtype = model_config.dtype
elif hasattr(model_config, "text_config") and hasattr(model_config.text_config, "dtype"):
torch_dtype = model_config.text_config.dtype
else:
torch_dtype = getattr(model_config, "torch_dtype", "bfloat16")
torch_dtype = getattr(model_config, "dtype", None)
if torch_dtype is None and hasattr(model_config, "text_config"):
torch_dtype = getattr(model_config.text_config, "dtype", None)
if torch_dtype is None:
torch_dtype = getattr(model_config, "torch_dtype", "bfloat16")


target_model = get_eagle3_target_model(
pretrained_model_name_or_path=args.target_model_path,
backend="sglang", # we set this as the default backend to minimize precision mismatch in training and serving
torch_dtype=(
model_config.dtype
if hasattr(model_config, "dtype")
else model_config.torch_dtype
),
torch_dtype=torch_dtype,
device="cuda",
cache_dir=args.model_download_dir,
trust_remote_code=args.trust_remote_code,
Expand Down Expand Up @@ -488,6 +495,8 @@ def generate(
"attention_mask": batch["attention_mask"][valid_indices_in_batch],
"loss_mask": batch["loss_mask"][valid_indices_in_batch],
}
pixel_values = batch["pixel_values"][valid_indices_in_batch]
image_grid_thw = batch["image_grid_thw"][valid_indices_in_batch]
del batch
if num_valid == 0:
# Data has already been generated, no sample processing, update progress bar.
Expand Down Expand Up @@ -542,6 +551,8 @@ def generate(
data_point = DataPoint(
input_ids=filtered_batch["input_ids"][i].clone(),
loss_mask=filtered_batch["loss_mask"][i].clone(),
pixel_values=pixel_values[i].clone(),
image_grid_thw=image_grid_thw[i].clone(),
Comment on lines +554 to +555
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot Apr 8, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

Accessing pixel_values[i] and image_grid_thw[i] will cause a TypeError or AttributeError if the variables are None (which occurs for non-VLM models).

Suggested change
pixel_values=pixel_values[i].clone(),
image_grid_thw=image_grid_thw[i].clone(),
pixel_values=pixel_values[i].clone() if pixel_values is not None else None,
image_grid_thw=image_grid_thw[i].clone() if image_grid_thw is not None else None,

hidden_state=last_hidden_states,
aux_hidden_state=aux_hidden_states,
)
Expand Down
Loading