[VLM] Enable overriding whether post layernorm is used in vision encoder + fix quant args (#9217)

Co-authored-by: Isotr0py <2037008807@qq.com>

vllm/model_executor/models/llava_onevision.py

@@ -31,6 +31,7 @@ from .clip import (CLIPVisionModel, dummy_seq_data_for_clip,
                    dummy_video_for_clip, get_clip_image_feature_size,
                    get_clip_patch_grid_length, input_processor_for_clip)
 from .interfaces import SupportsMultiModal, SupportsPP
+from .llava import init_vision_tower_for_llava
 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)
@@ -357,32 +358,6 @@ def input_processor_for_llava_onevision(ctx: InputContext,
     raise NotImplementedError(msg)
 
 
-def _init_vision_tower(hf_config: LlavaOnevisionConfig):
-    vision_config = hf_config.vision_config
-
-    # Initialize the vision tower only up to the required feature layer
-    vision_feature_layer = hf_config.vision_feature_layer
-    if vision_feature_layer < 0:
-        num_hidden_layers = hf_config.vision_config.num_hidden_layers \
-            + vision_feature_layer + 1
-    else:
-        num_hidden_layers = vision_feature_layer + 1
-
-    if isinstance(vision_config, CLIPVisionConfig):
-        return CLIPVisionModel(
-            vision_config,
-            num_hidden_layers_override=num_hidden_layers,
-        )
-    elif isinstance(vision_config, SiglipVisionConfig):
-        return SiglipVisionModel(
-            vision_config,
-            num_hidden_layers_override=num_hidden_layers,
-        )
-
-    msg = f"Unsupported vision config: {type(vision_config)}"
-    raise NotImplementedError(msg)
-
-
 class LlavaOnevisionMultiModalProjector(nn.Module):
 
     def __init__(self, config: LlavaOnevisionConfig):
@@ -425,7 +400,7 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal,
         self.multimodal_config = multimodal_config
 
         # Initialize the vision tower only up to the required feature layer
-        self.vision_tower = _init_vision_tower(config)
+        self.vision_tower = init_vision_tower_for_llava(config, quant_config)
         self.multi_modal_projector = LlavaOnevisionMultiModalProjector(config)
         self.language_model = init_vllm_registered_model(
             config.text_config, cache_config, quant_config)