Skip to content

support pixels range at prepare_hidden_states stage #438

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

support pixels range at prepare_hidden_states stage #438

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ec04cee
support pixels range at prepare_hidden_states stage
Jan 19, 2026
d061774
Apply suggestions from code review
justadogistaken Jan 19, 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
18 changes: 17 additions & 1 deletion scripts/prepare_hidden_states.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,8 +138,18 @@ def parse_args():
help="Number of files per subdirectory.",
)

# vlm related args
vlm_group = parser.add_argument_group("vlm")
vlm_group.add_argument(
"--min-pixels", type=int, default=50176
) # 64*28*28 for qwen2.5-vl
vlm_group.add_argument(
"--max-pixels", type=int, default=802816
) # 1024*28*28 for qwen2.5-vl
Comment on lines +143 to +148
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Jan 19, 2026

Choose a reason for hiding this comment

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

medium

It's good practice to add help strings to command-line arguments to explain their purpose. This improves the script's usability and maintainability. The formatting of these calls can also be improved to be more conventional and readable, similar to what code formatters like black would produce.

Suggested change
vlm_group.add_argument(
"--min-pixels", type=int, default=50176
) # 64*28*28 for qwen2.5-vl
vlm_group.add_argument(
"--max-pixels", type=int, default=802816
) # 1024*28*28 for qwen2.5-vl
vlm_group.add_argument(
"--min-pixels",
type=int,
default=50176,
help="The minimum number of pixels for VLM processing.",
) # 64*28*28 for qwen2.5-vl
vlm_group.add_argument(
"--max-pixels",
type=int,
default=802816,
help="The maximum number of pixels for VLM processing.",
) # 1024*28*28 for qwen2.5-vl


sglang_group = parser.add_argument_group("sglang")
SGLangBackendArgs.add_args(sglang_group)

return parser.parse_args()


Expand Down Expand Up @@ -187,7 +197,11 @@ def build_target_model(
target_model.set_aux_hidden_states_layers(args.aux_hidden_states_layers)

if args.is_vlm:
processor = AutoProcessor.from_pretrained(args.target_model_path)
processor = AutoProcessor.from_pretrained(
args.target_model_path,
min_pixels=args.min_pixels,
max_pixels=args.max_pixels,
)
else:
processor = None

Expand Down Expand Up @@ -583,6 +597,8 @@ def main():
args.target_model_path, trust_remote_code=True
)
cache_params_string = f"{args.data_path}-{args.max_length}-{args.chat_template}-{args.target_model_path}-{args.num_samples}-{args.is_preformatted}"
if args.is_vlm:
cache_params_string = f'{cache_params_string}-{args.min_pixels}-{args.max_pixels}'
cache_key = hashlib.md5(cache_params_string.encode()).hexdigest()

# Preprocess on complete, un-sharded dataset
Expand Down
Loading