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_next_video.py
View File

@@ -29,7 +29,7 @@ from .llava import init_vision_tower_for_llava
from .siglip import (SiglipVisionModel, dummy_image_for_siglip,
dummy_seq_data_for_siglip)
from .utils import (AutoWeightsLoader, init_vllm_registered_model,
merge_multimodal_embeddings)
maybe_prefix, merge_multimodal_embeddings)
# For profile run
_MAX_FRAMES_PER_VIDEO = 32
@@ -267,7 +267,7 @@ class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal,
config,
quant_config,
require_post_norm=False,
prefix="vision_tower")
prefix=maybe_prefix(prefix, "vision_tower"))
self.vision_resampler = LlavaNextVideoPooler(config)
self.multi_modal_projector = LlavaNextMultiModalProjector(
vision_hidden_size=config.vision_config.hidden_size,
@@ -276,7 +276,7 @@ class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal,
self.language_model = init_vllm_registered_model(
config.text_config,
vllm_config=vllm_config,
prefix="language_model")
prefix=maybe_prefix(prefix, "language_model"))
self.make_empty_intermediate_tensors = (
self.language_model.model.make_empty_intermediate_tensors)