Skip to content

update sparsevlm for llava1.6 #427

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
llmc-reviewer merged 1 commit into from
Jul 27, 2025
Merged

update sparsevlm for llava1.6 #427

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
47 changes: 35 additions & 12 deletions llmc/compression/token_reduction/sparsevlm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,9 @@
sparse_token_list_192 = []
sparse_token_list_128 = []
sparse_token_list_64 = []
sparse_token_list_640 = []
sparse_token_list_320 = []
sparse_token_list_160 = []
sparse_token_dict = {}


Expand Down Expand Up @@ -55,7 +58,7 @@ def input_hook(module, args, pruning_paras):
pre_prompt_length_list.append(0)
pruning_paras['pre_prompt_length_list'] = pre_prompt_length_list

def input_hook_llava(fn, pruning_paras):
def input_hook_llava(fn, pruning_paras, llava_next=False):
@wraps(fn)
def wrapper(self, *args, **kwargs):
if args[0].shape[1] == 1:
Expand All @@ -81,11 +84,14 @@ def wrapper(self, *args, **kwargs):

pruning_paras['pre_prompt_length_list'] = pre_prompt_length_list

return fn(*args, **kwargs)
outs = fn(*args, **kwargs)
if llava_next:
pruning_paras['vision_token_length'] = outs[-1]
return outs
return wrapper

@prefill_wrapper_model
def register_module_pars(module, args, kwargs, pruning_paras):
def register_module_paras(module, args, kwargs, pruning_paras):
pre_prompt_length_list = pruning_paras['pre_prompt_length_list']
hidden_states = kwargs['inputs_embeds']
if hidden_states is None:
Expand Down Expand Up @@ -227,7 +233,8 @@ def decoder_attn_hook(module, inputs, kwargs, layer_outputs, pruning_paras, laye
text_token_start,
t_token_idx,
layer_idx,
retained_tokens
retained_tokens,
pruning_paras['reduction_ratio']
)
if not prune_flag:
pred_score_vis = torch.zeros_like(relation_vis_text, dtype=bool)
Expand Down Expand Up @@ -353,7 +360,8 @@ def read_parameter_hook(module, args, kwargs, pruning_paras):
self.model.vlm_model.prepare_inputs_labels_for_multimodal = MethodType(
input_hook_llava(
self.model.vlm_model.prepare_inputs_labels_for_multimodal,
self.pruning_paras
self.pruning_paras,
llava_next=self.special_config['vision_token_length'] is None
), self.model.vlm_model
)

Expand All @@ -362,7 +370,7 @@ def read_parameter_hook(module, args, kwargs, pruning_paras):
elif self.model.__class__.__name__ == 'Llava':
llama_model = self.model.model.model
llama_model.register_forward_pre_hook(
functools.partial(register_module_pars, pruning_paras=self.pruning_paras),
functools.partial(register_module_paras, pruning_paras=self.pruning_paras),
with_kwargs=True
)

Expand Down Expand Up @@ -417,6 +425,7 @@ def read_parameter_hook(module, args, kwargs, pruning_paras):

def update_list():
global sparse_token_list_192, sparse_token_list_128, sparse_token_list_64
global sparse_token_list_640, sparse_token_list_320, sparse_token_list_160
global prune_flag, merge_flag, sparse_token_dict

if layer_dict == {2: 0, 6: 1, 15: 2}: # 2*576 4*300 10*200 16*110
Expand All @@ -428,10 +437,16 @@ def update_list():
sparse_token_list_192 = [180]
sparse_token_list_128 = [114]
sparse_token_list_64 = [48]
sparse_token_list_640 = [0.1979]
sparse_token_list_320 = [0.0833]
sparse_token_list_160 = [0.0261]
Comment on lines +440 to +442
Copy link

@gemini-code-assist gemini-code-assist bot Jul 27, 2025

Choose a reason for hiding this comment

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

medium

The newly added floating-point values for sparse_token_list_* are magic numbers. Their purpose and origin are not clear from the code, which harms readability and maintainability.

Please add comments to explain what these ratios represent and how they were determined. For example:

# Ratios for prune_flag=True, merge_flag=True, derived from internal experiments
sparse_token_list_640 = [0.1979]
# ...

This will help future developers understand and modify this code.

elif prune_flag:
sparse_token_list_192 = [192]
sparse_token_list_128 = [128]
sparse_token_list_64 = [64]
sparse_token_list_640 = [0.2222]
sparse_token_list_320 = [0.1111]
sparse_token_list_160 = [0.0555]
elif merge_flag:
sparse_token_list_192 = [149]
sparse_token_list_128 = [78]
Expand All @@ -444,7 +459,10 @@ def update_list():
sparse_token_dict = {
192: sparse_token_list_192,
128: sparse_token_list_128,
64: sparse_token_list_64
64: sparse_token_list_64,
640: sparse_token_list_640,
320: sparse_token_list_320,
160: sparse_token_list_160
}


Expand All @@ -455,7 +473,8 @@ def attn_postprocess_topk(
text_token_start,
t_token_idx,
layer_idx,
retained_tokens):
retained_tokens,
reduction_ratio):
'''
self_attn_weights: [B, H, L, L]
'''
Expand All @@ -470,13 +489,17 @@ def attn_postprocess_topk(

relation_vis = relation_vis_text
s_flag = True # s_flag controls whether token merge is needed.

sparse_token_list = sparse_token_dict[retained_tokens]

if retained_tokens in [192, 128, 64]:
sparse_token_list = sparse_token_dict[retained_tokens]
else:
sparse_token_list = sparse_token_dict[round((1 - reduction_ratio) * 2880)]
Comment on lines +494 to +495
Copy link

@gemini-code-assist gemini-code-assist bot Jul 27, 2025

Choose a reason for hiding this comment

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

high

This logic is brittle and can lead to a KeyError at runtime. The number 2880 is a magic number, which hurts readability and maintainability.

I suggest refactoring this block to be more robust by finding the closest available key.

    if retained_tokens in [192, 128, 64]:
        sparse_token_list = sparse_token_dict[retained_tokens]
    else:
        # TODO: Replace 2880 with a named constant e.g. LLAVA_1_6_MAX_VISION_TOKENS
        target_key = round((1 - reduction_ratio) * 2880)
        available_new_keys = [160, 320, 640]
        # Find the closest key to handle any reduction_ratio and prevent KeyErrors
        closest_key = min(available_new_keys, key=lambda k: abs(k - target_key))
        sparse_token_list = sparse_token_dict[closest_key]

retained_tokens_prune = sparse_token_list[layer_dict[layer_idx]]
if retained_tokens_prune < 1:
retained_tokens_prune = round(retained_tokens_prune * v_token_num)
if v_token_num != 0:
mask = torch.zeros_like(relation_vis, dtype=bool)
_, indices = torch.topk(relation_vis, min(
sparse_token_list[layer_dict[layer_idx]], v_token_num - 1), dim=1)
retained_tokens_prune, v_token_num - 1), dim=1)
mask[0][indices] = 1
else:
mask = torch.ones_like(relation_vis_text, dtype=bool)
Expand Down
2 changes: 1 addition & 1 deletion llmc/compression/token_reduction/token_reduction_module.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,7 +42,7 @@ def wrapper(self, *args, **kwargs):

message = (
'To obtain the vision_token_length for LLaVA-1.6, you should append '
'`image_features.shape[1]` to the return value of the function '
'`image_features[0].shape[0]` to the return value of the function '
'`prepare_inputs_labels_for_multimodal`, and modify the related code accordingly.'
)
outs = fn(*args, **kwargs)
Expand Down