[Bugfix] Fix prefix strings for quantized VLMs (#9772)

2024-10-29 19:02:59 -04:00
parent 8d7724104a
commit bc73e9821c
20 changed files with 288 additions and 97 deletions
--- a/vllm/model_executor/models/llava_onevision.py
+++ b/vllm/model_executor/models/llava_onevision.py
@@ -415,10 +415,16 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal,

        # Initialize the vision tower only up to the required feature layer
        self.vision_tower = init_vision_tower_for_llava(
-            config, quant_config, require_post_norm=False)
+            config,
+            quant_config,
+            require_post_norm=False,
+            prefix="vision_tower")
        self.multi_modal_projector = LlavaOnevisionMultiModalProjector(config)
        self.language_model = init_vllm_registered_model(
-            config.text_config, cache_config, quant_config)
+            config.text_config,
+            cache_config,
+            quant_config,
+            prefix="language_model")
        self.image_newline = nn.Parameter(
            torch.empty(config.text_config.hidden_size))