Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -10,30 +10,36 @@ from transformers import BatchFeature, GotOcr2ImageProcessor
|
||||
from transformers.activations import ACT2FN
|
||||
from transformers.image_processing_utils import get_size_dict
|
||||
from transformers.models.aya_vision import AyaVisionConfig
|
||||
from transformers.models.aya_vision.processing_aya_vision import (
|
||||
AyaVisionProcessor)
|
||||
from transformers.models.aya_vision.processing_aya_vision import AyaVisionProcessor
|
||||
from transformers.models.got_ocr2.image_processing_got_ocr2 import (
|
||||
get_optimal_tiled_canvas)
|
||||
get_optimal_tiled_canvas,
|
||||
)
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.config.multimodal import BaseDummyOptions
|
||||
from vllm.multimodal import MULTIMODAL_REGISTRY
|
||||
from vllm.multimodal.inputs import MultiModalDataDict, MultiModalKwargsItems
|
||||
from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
|
||||
MultiModalDataItems)
|
||||
from vllm.multimodal.processing import (BaseMultiModalProcessor,
|
||||
BaseProcessingInfo,
|
||||
MultiModalFieldConfig,
|
||||
PromptReplacement, PromptUpdate,
|
||||
PromptUpdateDetails)
|
||||
from vllm.multimodal.parse import ImageProcessorItems, ImageSize, MultiModalDataItems
|
||||
from vllm.multimodal.processing import (
|
||||
BaseMultiModalProcessor,
|
||||
BaseProcessingInfo,
|
||||
MultiModalFieldConfig,
|
||||
PromptReplacement,
|
||||
PromptUpdate,
|
||||
PromptUpdateDetails,
|
||||
)
|
||||
from vllm.multimodal.profiling import BaseDummyInputsBuilder
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.utils.tensor_schema import TensorSchema, TensorShape
|
||||
|
||||
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
|
||||
from .siglip import SiglipVisionModel
|
||||
from .utils import (AutoWeightsLoader, WeightsMapper,
|
||||
init_vllm_registered_model, maybe_prefix)
|
||||
from .utils import (
|
||||
AutoWeightsLoader,
|
||||
WeightsMapper,
|
||||
init_vllm_registered_model,
|
||||
maybe_prefix,
|
||||
)
|
||||
|
||||
|
||||
class AyaVisionImagePixelInputs(TensorSchema):
|
||||
@@ -61,17 +67,17 @@ class AyaVisionImagePixelInputs(TensorSchema):
|
||||
|
||||
|
||||
class AyaVisionMultiModalProjector(nn.Module):
|
||||
|
||||
def __init__(self, config: AyaVisionConfig):
|
||||
super().__init__()
|
||||
self.config = config
|
||||
self.downsample_factor = config.downsample_factor
|
||||
self.alignment_intermediate_size = getattr(
|
||||
config, "alignment_intermediate_size",
|
||||
config.text_config.hidden_size)
|
||||
self.layernorm = nn.LayerNorm(config.vision_config.hidden_size *
|
||||
(config.downsample_factor**2),
|
||||
eps=config.adapter_layer_norm_eps)
|
||||
config, "alignment_intermediate_size", config.text_config.hidden_size
|
||||
)
|
||||
self.layernorm = nn.LayerNorm(
|
||||
config.vision_config.hidden_size * (config.downsample_factor**2),
|
||||
eps=config.adapter_layer_norm_eps,
|
||||
)
|
||||
|
||||
self.linear_1 = nn.Linear(
|
||||
config.vision_config.hidden_size * (config.downsample_factor**2),
|
||||
@@ -81,9 +87,11 @@ class AyaVisionMultiModalProjector(nn.Module):
|
||||
|
||||
self.act = ACT2FN["silu"] # SwiGLU uses SiLU activation
|
||||
# For SwiGLU, project down to half size since we split intermediate dim
|
||||
self.linear_2 = nn.Linear(self.alignment_intermediate_size // 2,
|
||||
config.text_config.hidden_size,
|
||||
bias=True)
|
||||
self.linear_2 = nn.Linear(
|
||||
self.alignment_intermediate_size // 2,
|
||||
config.text_config.hidden_size,
|
||||
bias=True,
|
||||
)
|
||||
|
||||
def forward(self, image_features: torch.Tensor) -> torch.Tensor:
|
||||
image_features = self.pixel_shuffle(image_features)
|
||||
@@ -97,26 +105,31 @@ class AyaVisionMultiModalProjector(nn.Module):
|
||||
hidden_states = self.linear_2(hidden_states)
|
||||
return hidden_states
|
||||
|
||||
def pixel_shuffle(self,
|
||||
image_features: torch.Tensor) -> torch.Tensor: # B, S, D
|
||||
def pixel_shuffle(self, image_features: torch.Tensor) -> torch.Tensor: # B, S, D
|
||||
batch_size, seq_length, _ = image_features.shape
|
||||
height = width = int(seq_length**0.5)
|
||||
image_features = image_features.reshape(image_features.shape[0], width,
|
||||
height, -1)
|
||||
image_features = image_features.reshape(
|
||||
image_features.shape[0], width, height, -1
|
||||
)
|
||||
channels = image_features.shape[-1]
|
||||
image_features = image_features.reshape(
|
||||
batch_size, width, int(height / self.downsample_factor),
|
||||
int(channels * self.downsample_factor))
|
||||
batch_size,
|
||||
width,
|
||||
int(height / self.downsample_factor),
|
||||
int(channels * self.downsample_factor),
|
||||
)
|
||||
image_features = image_features.permute(0, 2, 1, 3)
|
||||
image_features = image_features.reshape(
|
||||
batch_size, int(height / self.downsample_factor),
|
||||
int(width / self.downsample_factor), -1)
|
||||
batch_size,
|
||||
int(height / self.downsample_factor),
|
||||
int(width / self.downsample_factor),
|
||||
-1,
|
||||
)
|
||||
image_features = image_features.permute(0, 2, 1, 3)
|
||||
return image_features
|
||||
|
||||
|
||||
class AyaVisionProcessingInfo(BaseProcessingInfo):
|
||||
|
||||
def get_hf_config(self) -> AyaVisionConfig:
|
||||
return self.ctx.get_hf_config(AyaVisionConfig)
|
||||
|
||||
@@ -131,14 +144,20 @@ class AyaVisionProcessingInfo(BaseProcessingInfo):
|
||||
|
||||
def get_image_size_with_most_features(self) -> ImageSize:
|
||||
image_processor = self.get_image_processor()
|
||||
height = image_processor.size['height']
|
||||
width = image_processor.size['width']
|
||||
height = image_processor.size["height"]
|
||||
width = image_processor.size["width"]
|
||||
max_patches = image_processor.max_patches
|
||||
return ImageSize(height=height * max_patches,
|
||||
width=width * max_patches)
|
||||
return ImageSize(height=height * max_patches, width=width * max_patches)
|
||||
|
||||
def get_num_patches(self, *, image_width: int, image_height: int,
|
||||
size: dict, min_patches: int, max_patches: int) -> int:
|
||||
def get_num_patches(
|
||||
self,
|
||||
*,
|
||||
image_width: int,
|
||||
image_height: int,
|
||||
size: dict,
|
||||
min_patches: int,
|
||||
max_patches: int,
|
||||
) -> int:
|
||||
"""
|
||||
Calculate the number of patches needed for a given image based on size
|
||||
constraints. This method replicates and adjusts the logic from:
|
||||
@@ -146,15 +165,16 @@ class AyaVisionProcessingInfo(BaseProcessingInfo):
|
||||
"""
|
||||
size = get_size_dict(size, default_to_square=False)
|
||||
num_columns, num_rows = get_optimal_tiled_canvas(
|
||||
(image_height, image_width), (size["height"], size["width"]),
|
||||
min_patches, max_patches)
|
||||
(image_height, image_width),
|
||||
(size["height"], size["width"]),
|
||||
min_patches,
|
||||
max_patches,
|
||||
)
|
||||
num_blocks = num_columns * num_rows
|
||||
return num_blocks if num_blocks == 1 else num_blocks + 1
|
||||
|
||||
|
||||
class AyaVisionDummyInputsBuilder(
|
||||
BaseDummyInputsBuilder[AyaVisionProcessingInfo]):
|
||||
|
||||
class AyaVisionDummyInputsBuilder(BaseDummyInputsBuilder[AyaVisionProcessingInfo]):
|
||||
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
|
||||
num_images = mm_counts.get("image", 0)
|
||||
|
||||
@@ -170,23 +190,21 @@ class AyaVisionDummyInputsBuilder(
|
||||
mm_options: Optional[Mapping[str, BaseDummyOptions]] = None,
|
||||
) -> MultiModalDataDict:
|
||||
num_images = mm_counts.get("image", 0)
|
||||
image_size = \
|
||||
self.info.get_image_size_with_most_features()
|
||||
image_size = self.info.get_image_size_with_most_features()
|
||||
|
||||
image_overrides = mm_options.get("image") if mm_options else None
|
||||
|
||||
return {
|
||||
"image":
|
||||
self._get_dummy_images(width=image_size.width,
|
||||
height=image_size.height,
|
||||
num_images=num_images,
|
||||
overrides=image_overrides)
|
||||
"image": self._get_dummy_images(
|
||||
width=image_size.width,
|
||||
height=image_size.height,
|
||||
num_images=num_images,
|
||||
overrides=image_overrides,
|
||||
)
|
||||
}
|
||||
|
||||
|
||||
class AyaVisionMultiModalProcessor(
|
||||
BaseMultiModalProcessor[AyaVisionProcessingInfo]):
|
||||
|
||||
class AyaVisionMultiModalProcessor(BaseMultiModalProcessor[AyaVisionProcessingInfo]):
|
||||
def _call_hf_processor(
|
||||
self,
|
||||
prompt: str,
|
||||
@@ -205,13 +223,13 @@ class AyaVisionMultiModalProcessor(
|
||||
|
||||
# HF processor pops the `num_patches` kwarg, which is needed by vLLM
|
||||
if (images := mm_data.get("images")) is not None:
|
||||
parsed_images = (self._get_data_parser().parse_mm_data({
|
||||
"image":
|
||||
images
|
||||
}).get_items("image", ImageProcessorItems))
|
||||
parsed_images = (
|
||||
self._get_data_parser()
|
||||
.parse_mm_data({"image": images})
|
||||
.get_items("image", ImageProcessorItems)
|
||||
)
|
||||
image_sizes = [
|
||||
parsed_images.get_image_size(i)
|
||||
for i in range(len(parsed_images))
|
||||
parsed_images.get_image_size(i) for i in range(len(parsed_images))
|
||||
]
|
||||
|
||||
num_patches = [
|
||||
@@ -220,7 +238,8 @@ class AyaVisionMultiModalProcessor(
|
||||
image_height=image_size.height,
|
||||
size=image_processor.size,
|
||||
min_patches=image_processor.min_patches,
|
||||
max_patches=image_processor.max_patches)
|
||||
max_patches=image_processor.max_patches,
|
||||
)
|
||||
for image_size in image_sizes
|
||||
]
|
||||
processed_outputs["num_patches"] = torch.tensor(num_patches)
|
||||
@@ -234,8 +253,7 @@ class AyaVisionMultiModalProcessor(
|
||||
) -> Mapping[str, MultiModalFieldConfig]:
|
||||
num_patches = hf_inputs.get("num_patches", torch.empty(0))
|
||||
return dict(
|
||||
pixel_values=MultiModalFieldConfig.flat_from_sizes(
|
||||
"image", num_patches),
|
||||
pixel_values=MultiModalFieldConfig.flat_from_sizes("image", num_patches),
|
||||
num_patches=MultiModalFieldConfig.batched("image"),
|
||||
image_embeds=MultiModalFieldConfig.batched("image"),
|
||||
)
|
||||
@@ -282,10 +300,10 @@ def _get_num_hidden_layers(hf_config: AyaVisionConfig) -> int:
|
||||
return _get_layer_index(feature_layers, num_hidden_layers)
|
||||
# If we have multiple feature layers, initialize up to the deepest m
|
||||
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")
|
||||
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:
|
||||
@@ -297,9 +315,9 @@ def _get_layer_index(feature_layer_index: int, num_hidden_layers: int) -> int:
|
||||
@MULTIMODAL_REGISTRY.register_processor(
|
||||
AyaVisionMultiModalProcessor,
|
||||
info=AyaVisionProcessingInfo,
|
||||
dummy_inputs=AyaVisionDummyInputsBuilder)
|
||||
class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
SupportsPP):
|
||||
dummy_inputs=AyaVisionDummyInputsBuilder,
|
||||
)
|
||||
class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
|
||||
merge_by_field_config = True
|
||||
|
||||
hf_to_vllm_mapper = WeightsMapper(
|
||||
@@ -309,7 +327,8 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
"model.vision_tower.": "vision_tower.",
|
||||
"model.multi_modal_projector.": "multi_modal_projector.",
|
||||
"lm_head.": "language_model.lm_head.",
|
||||
})
|
||||
}
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
|
||||
@@ -332,7 +351,8 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
config.vision_config,
|
||||
quant_config,
|
||||
num_hidden_layers_override=num_hidden_layers,
|
||||
prefix=maybe_prefix(prefix, "vision_model"))
|
||||
prefix=maybe_prefix(prefix, "vision_model"),
|
||||
)
|
||||
self.vocab_size = config.text_config.vocab_size
|
||||
self.multi_modal_projector = AyaVisionMultiModalProjector(config)
|
||||
self.language_model = init_vllm_registered_model(
|
||||
@@ -340,14 +360,14 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
hf_config=config.text_config,
|
||||
prefix=maybe_prefix(prefix, "model"),
|
||||
# Cohere2ForCausalLM and CohereForCausalLM are the same on vllm
|
||||
architectures=["Cohere2ForCausalLM"])
|
||||
architectures=["Cohere2ForCausalLM"],
|
||||
)
|
||||
|
||||
@property
|
||||
def dtype(self):
|
||||
return next(self.parameters()).dtype
|
||||
|
||||
def load_weights(self, weights: Iterable[tuple[str,
|
||||
torch.Tensor]]) -> set[str]:
|
||||
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
|
||||
loader = AutoWeightsLoader(self)
|
||||
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|
||||
|
||||
@@ -361,20 +381,21 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
feature_select_strategy=self.config.vision_feature_select_strategy,
|
||||
)
|
||||
|
||||
def _process_image_input(self, image_input: AyaVisionImagePixelInputs,
|
||||
**kwargs) -> list[torch.Tensor]:
|
||||
def _process_image_input(
|
||||
self, image_input: AyaVisionImagePixelInputs, **kwargs
|
||||
) -> list[torch.Tensor]:
|
||||
assert self.vision_tower is not None
|
||||
pixel_values = image_input["pixel_values"]
|
||||
num_patches = image_input["num_patches"]
|
||||
image_features = self._image_pixels_to_features(
|
||||
self.vision_tower, pixel_values=pixel_values)
|
||||
self.vision_tower, pixel_values=pixel_values
|
||||
)
|
||||
image_embeds = self.multi_modal_projector(image_features)
|
||||
return [
|
||||
e.flatten(0, 2) for e in image_embeds.split(num_patches.tolist())
|
||||
]
|
||||
return [e.flatten(0, 2) for e in image_embeds.split(num_patches.tolist())]
|
||||
|
||||
def _parse_and_validate_image_input(
|
||||
self, **kwargs: object) -> Optional[AyaVisionImagePixelInputs]:
|
||||
self, **kwargs: object
|
||||
) -> Optional[AyaVisionImagePixelInputs]:
|
||||
pixel_values = kwargs.pop("pixel_values", None)
|
||||
num_patches = kwargs.pop("num_patches", None)
|
||||
image_embeds = kwargs.pop("image_embeds", None)
|
||||
@@ -390,13 +411,13 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
resolve_bindings={
|
||||
"h": self.config.vision_config.image_size,
|
||||
"w": self.config.vision_config.image_size,
|
||||
})
|
||||
},
|
||||
)
|
||||
|
||||
def get_language_model(self) -> torch.nn.Module:
|
||||
return self.language_model
|
||||
|
||||
def get_multimodal_embeddings(self,
|
||||
**kwargs: object) -> MultiModalEmbeddings:
|
||||
def get_multimodal_embeddings(self, **kwargs: object) -> MultiModalEmbeddings:
|
||||
image_input = self._parse_and_validate_image_input(**kwargs)
|
||||
if image_input is None:
|
||||
return []
|
||||
|
||||
Reference in New Issue
Block a user