1 files changed, 9 insertions, 0 deletions
diff --git a/src/personalization/config/registry.py b/src/personalization/config/registry.py
index d825ad3..6048044 100644
--- a/src/personalization/config/registry.py
+++ b/src/personalization/config/registry.py
@@ -49,6 +49,7 @@ def get_chat_model(name: str, device_override: Optional[str] = None):
     from personalization.models.llm.base import ChatModel
     from personalization.models.llm.qwen_instruct import QwenInstruct
     from personalization.models.llm.llama_instruct import LlamaChatModel
+    from personalization.models.llm.vllm_chat import VLLMChatModel
     
     cfg = settings.load_local_models_config()
     
@@ -81,6 +82,14 @@ def get_chat_model(name: str, device_override: Optional[str] = None):
                 dtype=choose_dtype(dtype), # Converts string to torch.dtype
                 max_context_length=max_len
             )
+        elif backend == "vllm":
+            # Use vLLM HTTP API for high-throughput inference
+            vllm_url = spec.get("vllm_url", "http://localhost:8003/v1")
+            return VLLMChatModel(
+                vllm_url=vllm_url,
+                model_name=spec.get("model_name"),
+                max_context_length=max_len
+            )
     
     # Fallback to legacy single config
     return QwenInstruct.from_config(cfg)