[Model] Add Support for Multimodal Granite Models (#10291)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
Alex Brooks
2024-11-21 10:46:20 +00:00
committed by GitHub
co-authored by Cyrus Leung
parent f0e0238016
commit 1cfde82ffd
6 changed files with 191 additions and 35 deletions
+44
View File
@@ -6,6 +6,7 @@ from typing import Any, List, Optional, Tuple, TypeVar, Union
import numpy as np
import numpy.typing as npt
import torch
from PIL import Image
import vllm.envs as envs
@@ -392,6 +393,49 @@ def encode_video_base64(frames: npt.NDArray):
return ",".join(base64_frames)
def resolve_visual_encoder_outputs(
encoder_outputs: Union[torch.Tensor, list[torch.Tensor]],
feature_sample_layers: Optional[list[int]],
post_layer_norm: Optional[torch.nn.LayerNorm],
max_possible_layers: int,
) -> torch.Tensor:
"""Given the outputs a visual encoder module that may correspond to the
output of the last layer, or a list of hidden states to be stacked,
handle post normalization and resolve it into a single output tensor.
Args:
encoder_outputs: Output of encoder's last layer or all hidden states.
feature_sample_layers: Optional layer indices to grab from the encoder
outputs; if provided, encoder outputs must be a list.
post_layer_norm: Post norm to apply to the output of the encoder.
max_possible_layers: Total layers in the fully loaded visual encoder.
"""
if feature_sample_layers is None:
if post_layer_norm is not None:
return post_layer_norm(encoder_outputs)
return encoder_outputs
# Get the hidden states corresponding to the layer indices.
# Negative values are relative to the full visual encoder,
# so offset them depending on how many layers were loaded.
# NOTE: this assumes that encoder_outputs contains a list
# of hidden states in the same order as the encoder layers
# that produced them.
offset = max_possible_layers - len(encoder_outputs)
hs_pool = [
encoder_outputs[layer_idx]
if layer_idx >= 0 else encoder_outputs[layer_idx + offset]
for layer_idx in feature_sample_layers
]
# Apply post-norm on the final hidden state if we are using it
uses_last_layer = feature_sample_layers[-1] in (len(hs_pool) - 1, -1)
if post_layer_norm is not None and uses_last_layer:
hs_pool[-1] = post_layer_norm(encoder_outputs)
return torch.cat(hs_pool, dim=-1)
# Utilities for input processors
_T = TypeVar("_T", str, int)