@@ -44,7 +44,7 @@ from vllm.utils import is_list_of
|
||||
|
||||
from .clip import dummy_image_for_clip, dummy_seq_data_for_clip
|
||||
from .interfaces import SupportsMultiModal
|
||||
from .utils import merge_multimodal_embeddings
|
||||
from .utils import flatten_bn, merge_multimodal_embeddings
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -75,15 +75,16 @@ class Phi3VImagePixelInputs(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: torch.Tensor
|
||||
"""
|
||||
Shape: `(batch_size, 2)`
|
||||
Shape: `(batch_size * num_images, 2)`
|
||||
|
||||
This should be in `(height, width)` format.
|
||||
"""
|
||||
@@ -92,7 +93,7 @@ class Phi3VImagePixelInputs(TypedDict):
|
||||
class Phi3VImageEmbeddingInputs(TypedDict):
|
||||
type: Literal["image_embeds"]
|
||||
data: Union[torch.Tensor, List[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.
|
||||
"""
|
||||
@@ -511,10 +512,19 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal):
|
||||
self.sampler = Sampler()
|
||||
|
||||
def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
|
||||
if list(data.shape[1:]) != [2]:
|
||||
raise ValueError(
|
||||
f"The expected shape of image sizes is batch dimension plus "
|
||||
f"{[2]}. You supplied {tuple(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
|
||||
|
||||
@@ -531,7 +541,7 @@ class Phi3VForCausalLM(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:
|
||||
@@ -556,30 +566,24 @@ class Phi3VForCausalLM(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)}")
|
||||
|
||||
# Merge the B and N dimensions.
|
||||
if isinstance(pixel_values, torch.Tensor):
|
||||
pixel_values = pixel_values.flatten(0, 1)
|
||||
else:
|
||||
pixel_values = torch.cat(pixel_values)
|
||||
|
||||
image_sizes = image_sizes.flatten(0, 1)
|
||||
|
||||
return Phi3VImagePixelInputs(
|
||||
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:
|
||||
if not isinstance(image_embeds, torch.Tensor):
|
||||
raise ValueError("Incorrect type of image embeddings. "
|
||||
f"Got type: {type(image_embeds)}")
|
||||
|
||||
return Phi3VImageEmbeddingInputs(
|
||||
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