[Model] Add multi-image input support for LLaVA-Next offline inference (#7230)
This commit is contained in:
@@ -1,7 +1,7 @@
|
||||
"""Minimal implementation of CLIPVisionModel intended to be only used
|
||||
"""Minimal implementation of CLIPVisionModel intended to be only used
|
||||
within a vision language model."""
|
||||
from array import array
|
||||
from typing import Iterable, Optional, Tuple
|
||||
from typing import Iterable, List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
@@ -84,7 +84,7 @@ def input_processor_for_clip(
|
||||
llm_inputs: LLMInputs,
|
||||
*,
|
||||
image_token_id: int,
|
||||
image_feature_size_override: Optional[int] = None,
|
||||
image_feature_size_override: Optional[Union[int, List[int]]] = None,
|
||||
):
|
||||
multi_modal_data = llm_inputs.get("multi_modal_data")
|
||||
if multi_modal_data is None or "image" not in multi_modal_data:
|
||||
@@ -217,7 +217,7 @@ class CLIPEncoderLayer(nn.Module):
|
||||
|
||||
class CLIPEncoder(nn.Module):
|
||||
"""
|
||||
Transformer encoder consisting of `config.num_hidden_layers` self
|
||||
Transformer encoder consisting of `config.num_hidden_layers` self
|
||||
attention layers. Each layer is a [`CLIPEncoderLayer`].
|
||||
|
||||
Args:
|
||||
|
||||
@@ -19,6 +19,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
||||
from vllm.multimodal import MULTIMODAL_REGISTRY
|
||||
from vllm.sequence import IntermediateTensors, SamplerOutput
|
||||
from vllm.utils import is_list_of
|
||||
|
||||
from .clip import (CLIPVisionModel, dummy_image_for_clip,
|
||||
dummy_seq_data_for_clip, get_clip_image_feature_size,
|
||||
@@ -223,6 +224,13 @@ def input_processor_for_llava_next(ctx: InputContext, llm_inputs: LLMInputs):
|
||||
input_height=height,
|
||||
input_width=width,
|
||||
)
|
||||
elif is_list_of(image_data, Image.Image):
|
||||
image_feature_size = [
|
||||
get_llava_next_image_feature_size(hf_config,
|
||||
input_height=img.height,
|
||||
input_width=img.width)
|
||||
for img in image_data
|
||||
]
|
||||
elif isinstance(image_data, torch.Tensor):
|
||||
image_feature_size = image_data.shape[0]
|
||||
else:
|
||||
@@ -425,7 +433,10 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
self.config.image_grid_pinpoints,
|
||||
self.config.vision_config.image_size,
|
||||
)
|
||||
other_patch_embeds = other_patch_embeds \
|
||||
num_patches = num_patch_height * num_patch_width
|
||||
|
||||
# Image patches might be padded for batch processing
|
||||
other_patch_embeds = other_patch_embeds[:num_patches] \
|
||||
.view(num_patch_height, num_patch_width, height, width, -1)
|
||||
|
||||
if "unpad" in strategy:
|
||||
@@ -496,7 +507,6 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
self,
|
||||
image_input: LlavaNextImageInputs,
|
||||
) -> Union[torch.Tensor, List[torch.Tensor]]:
|
||||
|
||||
if image_input["type"] == "image_embeds":
|
||||
return [image_input["data"]]
|
||||
|
||||
|
||||
@@ -3,7 +3,7 @@ within a vision language model."""
|
||||
|
||||
import math
|
||||
from array import array
|
||||
from typing import Iterable, Optional, Tuple
|
||||
from typing import Iterable, List, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
from PIL import Image
|
||||
@@ -93,7 +93,7 @@ def input_processor_for_siglip(
|
||||
llm_inputs: LLMInputs,
|
||||
*,
|
||||
image_token_id: int,
|
||||
image_feature_size_override: Optional[int] = None,
|
||||
image_feature_size_override: Optional[Union[int, List[int]]] = None,
|
||||
):
|
||||
multi_modal_data = llm_inputs.get("multi_modal_data")
|
||||
if multi_modal_data is None or "image" not in multi_modal_data:
|
||||
|
||||
+31
-18
@@ -189,10 +189,13 @@ def repeat_and_pad_placeholder_tokens(
|
||||
prompt_token_ids: List[int],
|
||||
*,
|
||||
placeholder_token_id: int,
|
||||
repeat_count: int = 1,
|
||||
repeat_count: Union[int, List[int]],
|
||||
pad_token_left: Optional[int] = None,
|
||||
pad_token_right: Optional[int] = None,
|
||||
) -> Tuple[Optional[str], List[int]]:
|
||||
if isinstance(repeat_count, int):
|
||||
repeat_count = [repeat_count]
|
||||
|
||||
if prompt is None:
|
||||
new_prompt = None
|
||||
else:
|
||||
@@ -201,13 +204,6 @@ def repeat_and_pad_placeholder_tokens(
|
||||
tokenizer.decode(pad_token_left))
|
||||
pad_token_str_right = (None if pad_token_right is None else
|
||||
tokenizer.decode(pad_token_right))
|
||||
replacement_str = "".join(
|
||||
repeat_and_pad_token(
|
||||
placeholder_token_str,
|
||||
repeat_count=repeat_count,
|
||||
pad_token_left=pad_token_str_left,
|
||||
pad_token_right=pad_token_str_right,
|
||||
))
|
||||
|
||||
placeholder_token_count = prompt.count(placeholder_token_str)
|
||||
# This is an arbitrary number to distinguish between the two cases
|
||||
@@ -216,28 +212,45 @@ def repeat_and_pad_placeholder_tokens(
|
||||
"Please follow the prompt format that is "
|
||||
"documented on HuggingFace which does not involve "
|
||||
"repeating %s tokens.", placeholder_token_str)
|
||||
elif placeholder_token_count > 1:
|
||||
logger.warning("Multiple multi-modal input is not supported yet, "
|
||||
"so any extra placeholder tokens will be treated "
|
||||
"as plain text.")
|
||||
if placeholder_token_count < len(repeat_count):
|
||||
logger.warning(
|
||||
"The number of multi-modal placeholder tokens in the prompt "
|
||||
"is less than the number of multi-modal inputs. Extra "
|
||||
"placeholder tokens will be treated as plain text")
|
||||
repeat_count = repeat_count[:placeholder_token_count]
|
||||
|
||||
# The image tokens are removed to be consistent with HuggingFace
|
||||
new_prompt = prompt.replace(placeholder_token_str, replacement_str, 1)
|
||||
prompt_parts = prompt.split(placeholder_token_str,
|
||||
maxsplit=len(repeat_count))
|
||||
new_prompt = ""
|
||||
for i, repeat_count_item in enumerate(repeat_count):
|
||||
replacement_str = "".join(
|
||||
repeat_and_pad_token(
|
||||
placeholder_token_str,
|
||||
repeat_count=repeat_count_item,
|
||||
pad_token_left=pad_token_str_left,
|
||||
pad_token_right=pad_token_str_right,
|
||||
))
|
||||
# The image tokens are removed to be consistent with HuggingFace
|
||||
new_prompt += prompt_parts[i] + replacement_str
|
||||
new_prompt += prompt_parts[-1]
|
||||
|
||||
new_token_ids: List[int] = []
|
||||
placeholder_token_idx = 0
|
||||
for i, token in enumerate(prompt_token_ids):
|
||||
if token == placeholder_token_id:
|
||||
replacement_ids = repeat_and_pad_token(
|
||||
placeholder_token_id,
|
||||
repeat_count=repeat_count,
|
||||
repeat_count=repeat_count[placeholder_token_idx],
|
||||
pad_token_left=pad_token_left,
|
||||
pad_token_right=pad_token_right,
|
||||
)
|
||||
new_token_ids.extend(replacement_ids)
|
||||
placeholder_token_idx += 1
|
||||
|
||||
# No need to further scan the list since we only replace once
|
||||
new_token_ids.extend(prompt_token_ids[i + 1:])
|
||||
break
|
||||
# No need to further scan the list since we replaced all tokens
|
||||
if placeholder_token_idx >= len(repeat_count):
|
||||
new_token_ids.extend(prompt_token_ids[i + 1:])
|
||||
break
|
||||
else:
|
||||
new_token_ids.append(token)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user