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

This commit is contained in:
Michael Goin
2024-10-29 16:02:59 -07:00
committed by GitHub
parent 8d7724104a
commit bc73e9821c
20 changed files with 288 additions and 97 deletions
+8 -2
View File
@@ -293,7 +293,10 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal,
# TODO: Optionally initializes this for supporting embeddings.
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.image_newline = nn.Parameter(
torch.empty(config.text_config.hidden_size))
self.multi_modal_projector = LlavaMultiModalProjector(
@@ -302,7 +305,10 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal,
projector_hidden_act=config.projector_hidden_act)
self.language_model = init_vllm_registered_model(
config.text_config, cache_config, quant_config)
config.text_config,
cache_config,
quant_config,
prefix="language_model")
# The same model class supports both language generation and embedding
# because the architecture name is the same