biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[6/N] pass whole config to inner model (#10205)

Browse Source 
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2024-11-10 22:41:46 -08:00
committed by GitHub
parent f0f2e5638e
commit f89d18ff74
69 changed files with 681 additions and 963 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/llava_onevision.py
View File

@@ -35,7 +35,7 @@ from .siglip import (SiglipVisionModel, dummy_seq_data_for_siglip,
dummy_video_for_siglip, get_siglip_image_feature_size,
get_siglip_patch_grid_length, input_processor_for_siglip)
from .utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model,
merge_multimodal_embeddings)
maybe_prefix, merge_multimodal_embeddings)
# Result in the max possible feature size (2x2 grid of 336x336px tiles)
MAX_IMAGE_FEATURE_SIZE_HEIGHT = MAX_IMAGE_FEATURE_SIZE_WIDTH = 448
@@ -418,12 +418,12 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal,
config,
quant_config,
require_post_norm=False,
prefix="vision_tower")
prefix=maybe_prefix(prefix, "vision_tower"))
self.multi_modal_projector = LlavaOnevisionMultiModalProjector(config)
self.language_model = init_vllm_registered_model(
config.text_config,
vllm_config=vllm_config,
prefix="language_model")
prefix=maybe_prefix(prefix, "language_model"))
self.image_newline = nn.Parameter(
torch.empty(config.text_config.hidden_size))