[Model] support MiniMax-VL-01 model (#16328)
Signed-off-by: qingjun <qingjun@minimaxi.com>
This commit is contained in:
qscqesze
615
vllm/model_executor/models/minimax_vl_01.py
Normal file
615
vllm/model_executor/models/minimax_vl_01.py
Normal file
@@ -0,0 +1,615 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
|
||||
from abc import abstractmethod
|
||||
from collections.abc import Iterable, Mapping, Sequence
|
||||
from dataclasses import dataclass
|
||||
from typing import (Final, Literal, Optional, Protocol, Set, Tuple, TypedDict,
|
||||
TypeVar, Union, cast)
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from transformers import BatchFeature, CLIPVisionConfig, PretrainedConfig
|
||||
from transformers.image_processing_utils import select_best_resolution
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.jsontree import json_map_leaves
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.activation import get_act_fn
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.quantization import QuantizationConfig
|
||||
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
||||
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
|
||||
from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
|
||||
from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
|
||||
ImageSize, MultiModalDataItems)
|
||||
from vllm.multimodal.processing import (BaseMultiModalProcessor,
|
||||
BaseProcessingInfo, PromptReplacement,
|
||||
PromptUpdate)
|
||||
from vllm.multimodal.profiling import BaseDummyInputsBuilder
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.transformers_utils.configs.minimax_vl_01 import MiniMaxVL01Config
|
||||
|
||||
from .clip import CLIPVisionModel
|
||||
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
|
||||
from .pixtral import PixtralHFVisionModel
|
||||
from .siglip import SiglipVisionModel
|
||||
from .utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model,
|
||||
maybe_prefix, merge_multimodal_embeddings)
|
||||
from .vision import get_vision_encoder_info
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
# For dummy input only
|
||||
@dataclass
|
||||
class MaxImageTokenMeta:
|
||||
width: int = 1024
|
||||
height: int = 1024
|
||||
|
||||
|
||||
class MiniMaxVL01ImagePixelInputs(TypedDict):
|
||||
type: Literal["pixel_values"]
|
||||
pixel_values: torch.Tensor
|
||||
"""
|
||||
Shape: `(batch_size * num_images, num_channels, height, width)`
|
||||
|
||||
Note that `height` or `width` may be different per batch and image,
|
||||
in which case the data is passed as a list instead of a batched tensor.
|
||||
"""
|
||||
|
||||
|
||||
class MiniMaxVL01ImageEmbeddingInputs(TypedDict):
|
||||
type: Literal["image_embeds"]
|
||||
data: torch.Tensor
|
||||
"""Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
|
||||
|
||||
`hidden_size` must match the hidden size of language model backbone.
|
||||
"""
|
||||
|
||||
|
||||
def image_size_to_num_patches(image_size, grid_pinpoints, patch_size: int):
|
||||
if not isinstance(grid_pinpoints, list):
|
||||
raise TypeError("grid_pinpoints should be a list of tuples or lists")
|
||||
|
||||
# ! VERY IMPORTANT if image_size is tensor, must convert to into tuple,
|
||||
# otherwise it will cause wrong calculate
|
||||
if not isinstance(image_size, (list, tuple)):
|
||||
if not isinstance(image_size, (torch.Tensor, np.ndarray)):
|
||||
raise TypeError("image_size invalid type " +
|
||||
f"{type(image_size)} with value {image_size}")
|
||||
image_size = image_size.tolist()
|
||||
|
||||
best_resolution = select_best_resolution(image_size, grid_pinpoints)
|
||||
height, width = best_resolution
|
||||
num_patches = 0
|
||||
# consider change to ceil(height/patch_size)*ceil(width/patch_size) + 1
|
||||
for i in range(0, height, patch_size):
|
||||
for j in range(0, width, patch_size):
|
||||
num_patches += 1
|
||||
# add the base patch
|
||||
num_patches += 1
|
||||
return num_patches
|
||||
|
||||
|
||||
def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
|
||||
if not isinstance(grid_pinpoints, list):
|
||||
raise TypeError("grid_pinpoints should be a list of tuples or lists")
|
||||
|
||||
# ! VERY IMPORTANT if image_size is tensor,
|
||||
# must convert to into tuple,
|
||||
# otherwise it will cause wrong calculate
|
||||
if not isinstance(image_size, (list, tuple)):
|
||||
if not isinstance(image_size, (torch.Tensor, np.ndarray)):
|
||||
raise TypeError(
|
||||
"image_size invalid type " +
|
||||
f"{type(image_size)} not valid, " +
|
||||
"should be either list, tuple, np.ndarray or tensor")
|
||||
image_size = image_size.tolist()
|
||||
|
||||
height, width = select_best_resolution(image_size, grid_pinpoints)
|
||||
return height // patch_size, width // patch_size
|
||||
|
||||
|
||||
def unpad_image(tensor, original_size):
|
||||
original_height, original_width = original_size
|
||||
current_height, current_width = tensor.shape[1:]
|
||||
|
||||
original_aspect_ratio = original_width / original_height
|
||||
current_aspect_ratio = current_width / current_height
|
||||
|
||||
if original_aspect_ratio > current_aspect_ratio:
|
||||
new_height = int(original_height * current_width) // original_width
|
||||
padding = (current_height - new_height) // 2
|
||||
unpadded_tensor = tensor[:, padding:current_height - padding, :]
|
||||
else:
|
||||
new_width = int(original_width * current_height) // original_height
|
||||
padding = (current_width - new_width) // 2
|
||||
unpadded_tensor = tensor[:, :, padding:current_width - padding]
|
||||
|
||||
return unpadded_tensor
|
||||
|
||||
|
||||
class MiniMaxVL01MultiModalProjector(nn.Module):
|
||||
|
||||
def __init__(self,
|
||||
vision_hidden_size: int,
|
||||
text_hidden_size: int,
|
||||
projector_hidden_act: str,
|
||||
multimodal_projector_bias: bool,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
prefix: str = ""):
|
||||
super().__init__()
|
||||
|
||||
self.linear_1 = ColumnParallelLinear(vision_hidden_size,
|
||||
text_hidden_size,
|
||||
bias=multimodal_projector_bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.linear_1")
|
||||
self.act = get_act_fn(projector_hidden_act)
|
||||
self.linear_2 = RowParallelLinear(text_hidden_size,
|
||||
text_hidden_size,
|
||||
bias=multimodal_projector_bias,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.linear_2")
|
||||
|
||||
def forward(self, image_features: torch.Tensor) -> torch.Tensor:
|
||||
hidden_states, _ = self.linear_1(image_features)
|
||||
hidden_states = self.act(hidden_states)
|
||||
hidden_states, _ = self.linear_2(hidden_states)
|
||||
return hidden_states
|
||||
|
||||
|
||||
class MiniMaxVL01LikeConfig(Protocol):
|
||||
vision_config: Final[PretrainedConfig]
|
||||
image_token_index: Final[int]
|
||||
vision_feature_select_strategy: Final[str]
|
||||
vision_feature_layer: Final[Union[int, list[int]]]
|
||||
|
||||
|
||||
class MiniMaxVL01LikeProcessor(Protocol):
|
||||
image_token: Final[str]
|
||||
|
||||
|
||||
_I = TypeVar("_I", bound=BaseProcessingInfo)
|
||||
|
||||
|
||||
class MiniMaxVL01DummyInputsBuilder(BaseDummyInputsBuilder[_I]):
|
||||
|
||||
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
|
||||
num_images = mm_counts.get("image", 0)
|
||||
processor = self.info.get_hf_processor()
|
||||
image_token = processor.image_token
|
||||
return image_token * num_images
|
||||
|
||||
def get_dummy_mm_data(
|
||||
self,
|
||||
seq_len: int,
|
||||
mm_counts: Mapping[str, int],
|
||||
) -> MultiModalDataDict:
|
||||
num_images = mm_counts.get("image", 0)
|
||||
|
||||
return {
|
||||
"image":
|
||||
self._get_dummy_images(width=MaxImageTokenMeta.width,
|
||||
height=MaxImageTokenMeta.height,
|
||||
num_images=num_images)
|
||||
}
|
||||
|
||||
|
||||
class MiniMaxVL01ProcessingInfo(BaseProcessingInfo):
|
||||
|
||||
def get_hf_config(self):
|
||||
return self.ctx.get_hf_config(MiniMaxVL01Config)
|
||||
|
||||
def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
|
||||
return {"image": None}
|
||||
|
||||
def get_vision_encoder_info(self):
|
||||
return get_vision_encoder_info(self.get_hf_config())
|
||||
|
||||
def _apply_feature_select_strategy(
|
||||
self,
|
||||
strategy: str,
|
||||
encoder_num_image_tokens: int,
|
||||
) -> int:
|
||||
if strategy == "default":
|
||||
return encoder_num_image_tokens - 1
|
||||
if strategy == "full":
|
||||
return encoder_num_image_tokens
|
||||
|
||||
msg = f"Unexpected feature select strategy: {strategy!r}"
|
||||
raise NotImplementedError(msg)
|
||||
|
||||
def get_num_image_tokens(
|
||||
self,
|
||||
*,
|
||||
image_width: int,
|
||||
image_height: int,
|
||||
) -> int:
|
||||
hf_config = self.get_hf_config()
|
||||
vision_encoder_info = self.get_vision_encoder_info()
|
||||
|
||||
return self._apply_feature_select_strategy(
|
||||
hf_config.vision_feature_select_strategy,
|
||||
vision_encoder_info.get_num_image_tokens(
|
||||
image_width=image_width,
|
||||
image_height=image_height,
|
||||
),
|
||||
)
|
||||
|
||||
def get_image_size_with_most_features(self) -> ImageSize:
|
||||
vision_encoder_info = self.get_vision_encoder_info()
|
||||
width = height = vision_encoder_info.get_image_size()
|
||||
return ImageSize(width=width, height=height)
|
||||
|
||||
def get_max_image_tokens(self) -> int:
|
||||
target_width, target_height = self.get_image_size_with_most_features()
|
||||
|
||||
return self.get_num_image_tokens(
|
||||
image_width=target_width,
|
||||
image_height=target_height,
|
||||
)
|
||||
|
||||
|
||||
class BaseMiniMaxVL01MultiModalProcessor(BaseMultiModalProcessor[_I]):
|
||||
|
||||
# Copied from BaseMultiModalProcessor
|
||||
@abstractmethod
|
||||
def _get_mm_fields_config(
|
||||
self,
|
||||
hf_inputs: BatchFeature,
|
||||
hf_processor_mm_kwargs: Mapping[str, object],
|
||||
) -> Mapping[str, MultiModalFieldConfig]:
|
||||
raise NotImplementedError
|
||||
|
||||
def _get_prompt_updates(
|
||||
self,
|
||||
mm_items: MultiModalDataItems,
|
||||
hf_processor_mm_kwargs: Mapping[str, object],
|
||||
out_mm_kwargs: MultiModalKwargs,
|
||||
) -> Sequence[PromptUpdate]:
|
||||
hf_config = self.info.get_hf_config()
|
||||
image_token_id = hf_config.image_token_index
|
||||
|
||||
def get_replacement(item_idx: int):
|
||||
images = mm_items.get_items(
|
||||
"image", (ImageEmbeddingItems, ImageProcessorItems))
|
||||
|
||||
if isinstance(images, ImageEmbeddingItems):
|
||||
num_image_tokens = images.get_feature_size(item_idx)
|
||||
else:
|
||||
image_size = images.get_image_size(item_idx)
|
||||
num_image_tokens = self.info.get_num_image_tokens(
|
||||
image_width=image_size.width,
|
||||
image_height=image_size.height,
|
||||
)
|
||||
|
||||
return [image_token_id] * num_image_tokens
|
||||
|
||||
return [
|
||||
PromptReplacement(
|
||||
modality="image",
|
||||
target=[image_token_id],
|
||||
replacement=get_replacement,
|
||||
),
|
||||
]
|
||||
|
||||
|
||||
class MiniMaxVL01MultiModalProcessor(
|
||||
BaseMiniMaxVL01MultiModalProcessor[MiniMaxVL01ProcessingInfo]):
|
||||
|
||||
def _call_hf_processor(
|
||||
self,
|
||||
prompt: str,
|
||||
mm_data: Mapping[str, object],
|
||||
mm_kwargs: Mapping[str, object],
|
||||
) -> BatchFeature:
|
||||
processed_outputs = super()._call_hf_processor(
|
||||
prompt=prompt,
|
||||
mm_data=mm_data,
|
||||
mm_kwargs=mm_kwargs,
|
||||
)
|
||||
|
||||
pixel_values = processed_outputs.get("pixel_values")
|
||||
if pixel_values is not None:
|
||||
image_sizes = processed_outputs["image_sizes"]
|
||||
min_len = min(len(pixel_values), len(image_sizes))
|
||||
pixel_values = pixel_values[:min_len]
|
||||
image_sizes = image_sizes[:min_len]
|
||||
assert len(pixel_values) == len(image_sizes)
|
||||
|
||||
processed_outputs["pixel_values"] = [
|
||||
p[:, :h, :w] for p, (h, w) in zip(pixel_values, image_sizes)
|
||||
]
|
||||
|
||||
return processed_outputs
|
||||
|
||||
def _get_mm_fields_config(
|
||||
self,
|
||||
hf_inputs: BatchFeature,
|
||||
hf_processor_mm_kwargs: Mapping[str, object],
|
||||
) -> Mapping[str, MultiModalFieldConfig]:
|
||||
return {
|
||||
"pixel_values": MultiModalFieldConfig.batched("image"),
|
||||
"image_embeds": MultiModalFieldConfig.batched("image"),
|
||||
}
|
||||
|
||||
|
||||
def _get_num_hidden_layers(hf_config: MiniMaxVL01LikeConfig) -> int:
|
||||
"""Determine the number of hidden layers to initialize up to in the
|
||||
visual encoder.
|
||||
|
||||
Args:
|
||||
hf_config: Model config with vision feature layer(s).
|
||||
"""
|
||||
feature_layers = hf_config.vision_feature_layer
|
||||
num_hidden_layers = hf_config.vision_config.num_hidden_layers
|
||||
# If we have one feature layer, initialize up to that layer
|
||||
if isinstance(feature_layers, int):
|
||||
return _get_layer_index(feature_layers, num_hidden_layers)
|
||||
# If we have multiple feature layers, initialize up to the deepest one
|
||||
elif isinstance(feature_layers, (list, tuple)):
|
||||
return max(
|
||||
_get_layer_index(idx, num_hidden_layers) for idx in feature_layers)
|
||||
raise TypeError(f"vision_layer_feature type: {type(feature_layers)}"
|
||||
" is not supported")
|
||||
|
||||
|
||||
def _get_layer_index(feature_layer_index: int, num_hidden_layers: int) -> int:
|
||||
"""Given a signed vision feature layer, get the number of hidden layers
|
||||
needed to leverage it.
|
||||
|
||||
Args:
|
||||
feature_layer_index: Index of a required layer in the visual encoder.
|
||||
num_hidden_layers: The total number of hidden layers in the visual
|
||||
encoder.
|
||||
"""
|
||||
if feature_layer_index < 0:
|
||||
return num_hidden_layers + feature_layer_index + 1
|
||||
return feature_layer_index
|
||||
|
||||
|
||||
def init_vision_tower_for_MiniMaxVL01(
|
||||
hf_config: MiniMaxVL01LikeConfig,
|
||||
quant_config: Optional[QuantizationConfig],
|
||||
*,
|
||||
require_post_norm: Optional[bool] = None,
|
||||
prefix: str = "",
|
||||
) -> Union[CLIPVisionModel, SiglipVisionModel, PixtralHFVisionModel]:
|
||||
vision_config = hf_config.vision_config
|
||||
|
||||
# Initialize the vision tower only up to the deepest required feature layer
|
||||
num_hidden_layers = _get_num_hidden_layers(hf_config)
|
||||
|
||||
if isinstance(vision_config, CLIPVisionConfig):
|
||||
return CLIPVisionModel(
|
||||
vision_config,
|
||||
quant_config=quant_config,
|
||||
num_hidden_layers_override=num_hidden_layers,
|
||||
require_post_norm=require_post_norm,
|
||||
prefix=prefix,
|
||||
)
|
||||
|
||||
msg = f"Unsupported vision config: {type(vision_config)}"
|
||||
raise NotImplementedError(msg)
|
||||
|
||||
|
||||
@MULTIMODAL_REGISTRY.register_processor(
|
||||
MiniMaxVL01MultiModalProcessor,
|
||||
info=MiniMaxVL01ProcessingInfo,
|
||||
dummy_inputs=MiniMaxVL01DummyInputsBuilder)
|
||||
class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
SupportsPP):
|
||||
|
||||
packed_modules_mapping = {
|
||||
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
|
||||
"gate_up_proj": ["gate_proj", "up_proj"]
|
||||
}
|
||||
|
||||
def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
|
||||
super().__init__()
|
||||
|
||||
config = vllm_config.model_config.hf_config
|
||||
quant_config = vllm_config.quant_config
|
||||
multimodal_config = vllm_config.model_config.multimodal_config
|
||||
|
||||
self.config = config
|
||||
self.multimodal_config = multimodal_config
|
||||
|
||||
# TODO: Optionally initializes this for supporting embeddings.
|
||||
self.vision_tower = init_vision_tower_for_MiniMaxVL01(
|
||||
config,
|
||||
quant_config,
|
||||
require_post_norm=False,
|
||||
prefix=maybe_prefix(prefix, "vision_tower"))
|
||||
self.multi_modal_projector = MiniMaxVL01MultiModalProjector(
|
||||
vision_hidden_size=config.vision_config.hidden_size,
|
||||
text_hidden_size=config.text_config.hidden_size,
|
||||
projector_hidden_act=config.projector_hidden_act,
|
||||
multimodal_projector_bias=True,
|
||||
quant_config=quant_config,
|
||||
prefix=maybe_prefix(prefix, "multi_modal_projector"))
|
||||
self.image_newline = nn.Parameter(
|
||||
torch.empty(config.text_config.hidden_size))
|
||||
self.language_model = init_vllm_registered_model(
|
||||
vllm_config=vllm_config,
|
||||
hf_config=config.text_config,
|
||||
prefix=maybe_prefix(prefix, "language_model"),
|
||||
)
|
||||
self.vision_feature_layer = config.vision_feature_layer
|
||||
self.vocab_size = config.text_config.vocab_size
|
||||
self.pad_token_id = -1
|
||||
if self.config.pad_token_id is not None:
|
||||
self.pad_token_id = self.config.pad_token_id
|
||||
|
||||
self.make_empty_intermediate_tensors = (
|
||||
self.language_model.make_empty_intermediate_tensors)
|
||||
|
||||
def get_input_embeddings(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
|
||||
) -> torch.Tensor:
|
||||
inputs_embeds = self.language_model.get_input_embeddings(input_ids)
|
||||
if multimodal_embeddings is not None:
|
||||
inputs_embeds = merge_multimodal_embeddings(
|
||||
input_ids,
|
||||
inputs_embeds,
|
||||
multimodal_embeddings,
|
||||
self.config.image_token_index,
|
||||
)
|
||||
return inputs_embeds
|
||||
|
||||
def get_language_model(self) -> torch.nn.Module:
|
||||
return self.language_model
|
||||
|
||||
def _select_image_features(self, image_features: torch.Tensor, *,
|
||||
strategy: str) -> torch.Tensor:
|
||||
if strategy == "default":
|
||||
return image_features[:, 1:]
|
||||
elif strategy == "full":
|
||||
return image_features
|
||||
|
||||
raise ValueError(f"Unexpected select feature strategy: {strategy}")
|
||||
|
||||
def _image_pixels_to_features(
|
||||
self,
|
||||
vision_tower: Union[CLIPVisionModel],
|
||||
pixel_values: Union[torch.Tensor, list[torch.Tensor]],
|
||||
) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
|
||||
# NOTE: we skip the step to select the vision feature layer since
|
||||
# this is already done inside the vision tower
|
||||
image_features = vision_tower(pixel_values)
|
||||
|
||||
def select_features(leaf: torch.Tensor):
|
||||
return self._select_image_features(
|
||||
leaf,
|
||||
strategy=self.config.vision_feature_select_strategy,
|
||||
)
|
||||
|
||||
return cast(
|
||||
Union[torch.Tensor, tuple[torch.Tensor, ...]],
|
||||
json_map_leaves(select_features, image_features),
|
||||
)
|
||||
|
||||
def _process_image_pixels(
|
||||
self,
|
||||
inputs: Union[MiniMaxVL01ImagePixelInputs],
|
||||
) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
|
||||
assert self.vision_tower is not None
|
||||
|
||||
pixel_values = inputs["pixel_values"]
|
||||
|
||||
return self._image_pixels_to_features(self.vision_tower, pixel_values)
|
||||
|
||||
def _process_image_input(
|
||||
self,
|
||||
image_input: MiniMaxVL01ImagePixelInputs,
|
||||
) -> Union[torch.Tensor, tuple[torch.Tensor, ...]]:
|
||||
if image_input["type"] == "image_embeds":
|
||||
return image_input["data"]
|
||||
|
||||
assert self.vision_tower is not None
|
||||
image_features = self._process_image_pixels(image_input)
|
||||
|
||||
if isinstance(image_features, torch.Tensor):
|
||||
return self.multi_modal_projector(image_features)
|
||||
|
||||
feature_sizes = [
|
||||
image_feature.shape[0] for image_feature in image_features
|
||||
]
|
||||
|
||||
image_embeds = self.multi_modal_projector(torch.cat(image_features))
|
||||
image_embeds = torch.split(image_embeds, feature_sizes)
|
||||
return image_embeds
|
||||
|
||||
def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
|
||||
h = w = self.config.vision_config.image_size
|
||||
expected_dims = (3, h, w)
|
||||
actual_dims = tuple(data.shape[1:])
|
||||
|
||||
if actual_dims != expected_dims:
|
||||
expected_expr = ("batch_size", *map(str, expected_dims))
|
||||
raise ValueError(
|
||||
f"The expected shape of pixel values is {expected_expr}. "
|
||||
f"You supplied {tuple(data.shape)}.")
|
||||
|
||||
return data
|
||||
|
||||
def _parse_and_validate_image_input(
|
||||
self, **kwargs: object) -> Optional[MiniMaxVL01ImagePixelInputs]:
|
||||
pixel_values = kwargs.pop("pixel_values", None)
|
||||
image_embeds = kwargs.pop("image_embeds", None)
|
||||
|
||||
if pixel_values is None and image_embeds is None:
|
||||
return None
|
||||
|
||||
if pixel_values is not None:
|
||||
if not isinstance(pixel_values, (torch.Tensor, list)):
|
||||
raise ValueError("Incorrect type of pixel values. "
|
||||
f"Got type: {type(pixel_values)}")
|
||||
|
||||
return MiniMaxVL01ImagePixelInputs(
|
||||
type="pixel_values",
|
||||
pixel_values=self._validate_pixel_values(
|
||||
flatten_bn(pixel_values, concat=True)),
|
||||
)
|
||||
|
||||
if image_embeds is not None:
|
||||
if not isinstance(image_embeds, (torch.Tensor, list)):
|
||||
raise ValueError("Incorrect type of image embeddings. "
|
||||
f"Got type: {type(image_embeds)}")
|
||||
|
||||
return MiniMaxVL01ImageEmbeddingInputs(
|
||||
type="image_embeds",
|
||||
data=flatten_bn(image_embeds, concat=True),
|
||||
)
|
||||
|
||||
raise AssertionError("This line should be unreachable.")
|
||||
|
||||
def get_multimodal_embeddings(
|
||||
self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
|
||||
image_input = self._parse_and_validate_image_input(**kwargs)
|
||||
if image_input is None:
|
||||
return None
|
||||
|
||||
return self._process_image_input(image_input)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
positions: torch.Tensor,
|
||||
intermediate_tensors: Optional[IntermediateTensors] = None,
|
||||
inputs_embeds: Optional[torch.Tensor] = None,
|
||||
**kwargs: object,
|
||||
) -> Union[torch.Tensor, IntermediateTensors]:
|
||||
|
||||
if intermediate_tensors is not None:
|
||||
inputs_embeds = None
|
||||
elif inputs_embeds is None:
|
||||
vision_embeddings = self.get_multimodal_embeddings(**kwargs)
|
||||
inputs_embeds = self.get_input_embeddings(input_ids,
|
||||
vision_embeddings)
|
||||
input_ids = None
|
||||
|
||||
hidden_states = self.language_model.model(input_ids,
|
||||
positions,
|
||||
intermediate_tensors,
|
||||
inputs_embeds=inputs_embeds)
|
||||
|
||||
return hidden_states
|
||||
|
||||
def compute_logits(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> Optional[torch.Tensor]:
|
||||
return self.language_model.compute_logits(hidden_states,
|
||||
sampling_metadata)
|
||||
|
||||
def load_weights(self, weights: Iterable[Tuple[str,
|
||||
torch.Tensor]]) -> Set[str]:
|
||||
loader = AutoWeightsLoader(self)
|
||||
return loader.load_weights(weights)
|
||||
Reference in New Issue
Block a user