[Model] Replace embedding models with pooling adapter (#10769)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
2024-12-01 08:02:54 +08:00
parent 7e4bbda573
commit 133707123e
32 changed files with 383 additions and 319 deletions
--- a/vllm/model_executor/models/paligemma.py
+++ b/vllm/model_executor/models/paligemma.py
@@ -151,9 +151,10 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal,
        self.quant_config = quant_config
        config.text_config.architectures = ["GemmaForCausalLM"]
        self.language_model = init_vllm_registered_model(
-            config.text_config,
            vllm_config=vllm_config,
-            prefix=maybe_prefix(prefix, "language_model"))
+            hf_config=config.text_config,
+            prefix=maybe_prefix(prefix, "language_model"),
+        )
        logit_scale = getattr(config, "logit_scale", 1.0)
        self.language_model.logits_processor.scale *= logit_scale