xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 14:25:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Model][Pixtral] Optimizations for input_processor_for_pixtral_hf (#9514)

Browse Source
This commit is contained in:
Michael Goin 2024-10-19 10:44:29 -04:00 committed by GitHub
parent 263d8ee150
commit 8e3e7f2713
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 40 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

79
vllm/model_executor/models/pixtral.py
View File

@ -701,63 +701,64 @@ def input_processor_for_pixtral_hf(
new_prompt = inputs.get("prompt") new_prompt = inputs.get("prompt")
new_token_ids = inputs["prompt_token_ids"] new_token_ids = inputs["prompt_token_ids"]
image_token = processor.image_token
image_break_token = processor.image_break_token
image_end_token = processor.image_end_token
# Update new_prompt if present # Update new_prompt if present
if new_prompt: if new_prompt:
replace_strings = [] parts = new_prompt.split(image_token)
for image in image_data: assert len(parts) - 1 == len(image_data)
w, h = image.size new_parts = [parts[0]] # Start with the part before any image tokens
for image, next_part in zip(image_data, parts[1:]):
w, h = image.size
(num_width_tokens, (num_width_tokens,
num_height_tokens) = get_pixtral_hf_image_feature_size( num_height_tokens) = get_pixtral_hf_image_feature_size(
hf_config, image_width=w, image_height=h) hf_config, image_width=w, image_height=h)
replace_tokens = [[processor.image_token] * num_width_tokens + replace_tokens = [image_token] * num_width_tokens + [
[processor.image_break_token] image_break_token
] * num_height_tokens
# Flatten list
replace_tokens = [
item for sublist in replace_tokens for item in sublist
] ]
replace_tokens[-1] = processor.image_end_token replace_tokens = replace_tokens * num_height_tokens
replace_str = "".join(replace_tokens) replace_tokens[-1] = image_end_token
replace_strings.append(replace_str)
new_prompt = new_prompt.replace(processor.image_token,
"<placeholder>", 1)
while "<placeholder>" in new_prompt: new_parts.append("".join(replace_tokens))
replace_str = replace_strings.pop(0) new_parts.append(next_part)
new_prompt = new_prompt.replace("<placeholder>", replace_str, 1)
new_prompt = "".join(new_parts)
# Update new_token_ids # Update new_token_ids
image_token_id = 10 convert_tokens_to_ids = processor.tokenizer.convert_tokens_to_ids
image_break_id = 12 image_token_id = convert_tokens_to_ids(image_token)
image_end_id = 13 image_break_id = convert_tokens_to_ids(image_break_token)
image_end_id = convert_tokens_to_ids(image_end_token)
placeholder_token_id = -999 placeholder_token_id = -999
# Find all image token indices at once
placeholder_indices = [
idx for idx, token_id in enumerate(new_token_ids)
if token_id == image_token_id
]
assert len(placeholder_indices) == len(image_data)
replace_tokens_list = [] replace_tokens_list = []
for image in image_data: for placeholder_idx, image in zip(placeholder_indices, image_data):
new_token_ids[placeholder_idx] = placeholder_token_id
w, h = image.size w, h = image.size
(num_width_tokens,
num_height_tokens) = get_pixtral_hf_image_feature_size(hf_config,
image_width=w,
image_height=h)
num_width_tokens, num_height_tokens = get_pixtral_hf_image_feature_size( replace_tokens = [image_token_id] * num_width_tokens + [image_break_id]
hf_config, image_width=w, image_height=h) replace_tokens = replace_tokens * num_height_tokens
replace_tokens = [[image_token_id] * num_width_tokens +
[image_break_id]] * num_height_tokens
# Flatten list
replace_tokens = [
item for sublist in replace_tokens for item in sublist
]
replace_tokens[-1] = image_end_id replace_tokens[-1] = image_end_id
replace_tokens_list.append(replace_tokens) replace_tokens_list.append(replace_tokens)
# Replace image id with placeholder id
next_image_index = new_token_ids.index(image_token_id)
new_token_ids[next_image_index] = placeholder_token_id
while placeholder_token_id in new_token_ids: # Backward iteration for replacement without affecting known indices
replace_tokens = replace_tokens_list.pop(0) for placeholder_idx, replace_tokens in zip(reversed(placeholder_indices),
next_image_index = new_token_ids.index(placeholder_token_id) reversed(replace_tokens_list)):
prefix = new_token_ids[:next_image_index] new_token_ids[placeholder_idx:placeholder_idx + 1] = replace_tokens
postfix = new_token_ids[next_image_index + 1:]
new_token_ids = prefix + replace_tokens + postfix
# NOTE: Create a defensive copy of the original inputs # NOTE: Create a defensive copy of the original inputs
return token_inputs(prompt_token_ids=new_token_ids, return token_inputs(prompt_token_ids=new_token_ids,