debugging

vllm-v0.6.2/vllm/model_executor/models/registry.py

@@ -28,6 +28,9 @@ from .interfaces_base import is_embedding_model, is_text_generation_model
 
 logger = init_logger(__name__)
 
+# Cache for architectures that have already been logged
+_logged_transformers_architectures: set = set()
+
 # yapf: disable
 _TEXT_GENERATION_MODELS = {
     # [Decoder-only]
@@ -403,11 +406,14 @@ class _ModelRegistry:
         model_module = getattr(transformers, architecture, None)
         if model_module is not None:
             # Model exists in transformers, can use TransformersForCausalLM wrapper
-            logger.info(
-                "Architecture %s found in transformers library, "
-                "using TransformersForCausalLM wrapper",
-                architecture
-            )
+            # Only log once per architecture to avoid spam
+            if architecture not in _logged_transformers_architectures:
+                _logged_transformers_architectures.add(architecture)
+                logger.info(
+                    "Architecture %s found in transformers library, "
+                    "using TransformersForCausalLM wrapper",
+                    architecture
+                )
             return "TransformersForCausalLM"
         
         # Get auto_map from hf_config