[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:
|
||||
|
||||
Reference in New Issue
Block a user