Skip to content

update mustdrop #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
zhangbilang merged 1 commit into from
Jul 14, 2025
Merged

update mustdrop #414

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
22 changes: 14 additions & 8 deletions llmc/compression/token_reduction/mustdrop.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@ def __init__(self, config, model, blocks):
self.register_reduction_modules()

def add_sparse_config(self):
self.pruning_loc = self.special_config['pruning_loc']
Copy link

@gemini-code-assist gemini-code-assist bot Jul 14, 2025

Choose a reason for hiding this comment

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

medium

Directly accessing self.special_config['pruning_loc'] can raise a KeyError if the key is missing. Consider adding a check to ensure the key exists before accessing it to prevent potential crashes.

self.pruning_paras = self.special_config

def register_reduction_modules(self):
Expand All @@ -30,6 +31,7 @@ def conditional_pooling(
feat: torch.Tensor,
threshold: float,
window_size: Tuple[int, int],
fix_r: int = 0,
) -> Tuple[Callable, Callable]:

with torch.no_grad():
Expand Down Expand Up @@ -91,7 +93,8 @@ def conditional_pooling(
node_mean = node_mean.repeat(1, n_H)
r = torch.ge(similarity_map, node_mean).sum(dim=1).min()
# -------------#

if fix_r != 0:
r = fix_r
# get top k similar super patches
_, sim_super_patch_idxs = similarity_map.topk(r, dim=-1)

Expand Down Expand Up @@ -184,17 +187,20 @@ def merge_wavg(

return x, size

def spatial_merge_hook(module, args, kwargs, pruning_paras):
def spatial_merge_hook(module, args, kwargs, layer_outs, pruning_paras):
spatial_threshold = pruning_paras['spatial_threshold']
window_size = pruning_paras['window_size']
hidden_states = args[0]
merge = conditional_pooling(hidden_states, spatial_threshold, window_size)
hidden_states = layer_outs[0]
fix_r = 0
if pruning_paras.get('retained_tokens', None) is not None:
retained_tokens = pruning_paras['retained_tokens']
fix_r = (pruning_paras['vision_token_length'] - retained_tokens) \
// (window_size[0] * window_size[1] - 1)
Comment on lines +197 to +198
Copy link

@gemini-code-assist gemini-code-assist bot Jul 14, 2025

Choose a reason for hiding this comment

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

high

The denominator window_size[0] * window_size[1] - 1 in the fix_r calculation can be zero if window_size is (1, 1), leading to a ZeroDivisionError. Add a check to prevent this.

if (window_size[0] * window_size[1] - 1) > 0:
    fix_r = (pruning_paras['vision_token_length'] - retained_tokens) // (window_size[0] * window_size[1] - 1)
else:
    fix_r = 0

merge = conditional_pooling(hidden_states, spatial_threshold, window_size, fix_r)
hidden_states, size = merge_wavg(merge, hidden_states, None)
return (hidden_states,) + args[1:], kwargs
return (hidden_states,)
Copy link

@gemini-code-assist gemini-code-assist bot Jul 14, 2025

Choose a reason for hiding this comment

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

critical

The spatial_merge_hook is registered as a forward_hook. Returning only (hidden_states,) replaces the block's entire output, potentially discarding other outputs like attention weights, which can cause errors in subsequent layers. Ensure the hook preserves other elements of the output tuple.

return (hidden_states,) + layer_outs[1:]


self.model.set_modality('vision')
self.model.find_blocks()
self.model.blocks[1].register_forward_pre_hook(
self.blocks[self.pruning_loc - 1].register_forward_hook(
functools.partial(spatial_merge_hook, pruning_paras=self.pruning_paras),
with_kwargs=True,
)