@@ -29,7 +29,7 @@ from .llava import LlavaMultiModalProjector
|
||||
from .siglip import (SiglipVisionModel, dummy_image_for_siglip,
|
||||
dummy_seq_data_for_siglip, get_siglip_image_feature_size,
|
||||
get_siglip_patch_grid_length, input_processor_for_siglip)
|
||||
from .utils import (filter_weights, init_vllm_registered_model,
|
||||
from .utils import (filter_weights, flatten_bn, init_vllm_registered_model,
|
||||
merge_multimodal_embeddings)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
@@ -47,15 +47,16 @@ class LlavaNextImagePixelInputs(TypedDict):
|
||||
type: Literal["pixel_values"]
|
||||
data: Union[torch.Tensor, List[torch.Tensor]]
|
||||
"""
|
||||
Shape: `(batch_size, 1 + num_patches, num_channels, height, width)`
|
||||
Shape:
|
||||
`(batch_size * num_images, 1 + num_patches, num_channels, height, width)`
|
||||
|
||||
Note that `num_patches` may be different for each batch, in which case
|
||||
the data is passed as a list instead of a batched tensor.
|
||||
Note that `num_patches` may be different per batch and image,
|
||||
in which case the data is passed as a list instead of a batched tensor.
|
||||
"""
|
||||
|
||||
image_sizes: NotRequired[torch.Tensor]
|
||||
"""
|
||||
Shape: `(batch_size, 2)`
|
||||
Shape: `(batch_size * num_images, 2)`
|
||||
|
||||
This should be in `(height, width)` format.
|
||||
"""
|
||||
@@ -64,7 +65,7 @@ class LlavaNextImagePixelInputs(TypedDict):
|
||||
class LlavaNextImageEmbeddingInputs(TypedDict):
|
||||
type: Literal["image_embeds"]
|
||||
data: torch.Tensor
|
||||
"""Shape: `(batch_size, image_feature_size, hidden_size)`
|
||||
"""Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
|
||||
|
||||
`hidden_size` must match the hidden size of language model backbone.
|
||||
"""
|
||||
@@ -315,10 +316,19 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
torch.empty(config.text_config.hidden_size))
|
||||
|
||||
def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
|
||||
if list(data.shape[1:]) != [2]:
|
||||
raise ValueError(
|
||||
f"The expected image sizes shape is batch dimension plus "
|
||||
f"{[2]}. You supplied {data.shape}.")
|
||||
expected_dims = (2, )
|
||||
|
||||
def _validate_shape(d: torch.Tensor):
|
||||
actual_dims = tuple(d.shape)
|
||||
|
||||
if actual_dims != expected_dims:
|
||||
expected_expr = str(expected_dims)
|
||||
raise ValueError(
|
||||
f"The expected shape of image sizes per image per batch "
|
||||
f"is {expected_expr}. You supplied {tuple(d.shape)}.")
|
||||
|
||||
for d in data:
|
||||
_validate_shape(d)
|
||||
|
||||
return data
|
||||
|
||||
@@ -335,7 +345,7 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
if actual_dims != expected_dims:
|
||||
expected_expr = ("num_patches", *map(str, expected_dims))
|
||||
raise ValueError(
|
||||
"The expected shape of pixel values in each batch element "
|
||||
"The expected shape of pixel values per image per batch "
|
||||
f"is {expected_expr}. You supplied {tuple(d.shape)}.")
|
||||
|
||||
for d in data:
|
||||
@@ -357,22 +367,15 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
raise ValueError("Incorrect type of pixel values. "
|
||||
f"Got type: {type(pixel_values)}")
|
||||
|
||||
if not isinstance(image_sizes, torch.Tensor):
|
||||
if not isinstance(image_sizes, (torch.Tensor, list)):
|
||||
raise ValueError("Incorrect type of image sizes. "
|
||||
f"Got type: {type(image_sizes)}")
|
||||
|
||||
# Remove the N dimension until multiple images are supported.
|
||||
if isinstance(pixel_values, torch.Tensor):
|
||||
pixel_values = pixel_values.squeeze(1)
|
||||
else:
|
||||
pixel_values = [t.squeeze(0) for t in pixel_values]
|
||||
|
||||
image_sizes = image_sizes.squeeze(1)
|
||||
|
||||
return LlavaNextImagePixelInputs(
|
||||
type="pixel_values",
|
||||
data=self._validate_pixel_values(pixel_values),
|
||||
image_sizes=self._validate_image_sizes(image_sizes),
|
||||
data=self._validate_pixel_values(flatten_bn(pixel_values)),
|
||||
image_sizes=self._validate_image_sizes(
|
||||
flatten_bn(image_sizes, concat=True)),
|
||||
)
|
||||
|
||||
if image_embeds is not None:
|
||||
@@ -380,12 +383,9 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
raise ValueError("Incorrect type of image embeds. "
|
||||
f"Got type: {type(image_embeds)}")
|
||||
|
||||
# Remove the N dimension until multiple images are supported.
|
||||
image_embeds = image_embeds.squeeze(1)
|
||||
|
||||
return LlavaNextImageEmbeddingInputs(
|
||||
type="image_embeds",
|
||||
data=image_embeds,
|
||||
data=flatten_bn(image_embeds),
|
||||
)
|
||||
|
||||
raise AssertionError("This line should be unreachable.")
|
||||
|
||||
Reference in New Issue
Block a user