Upgrade transformers and add fallbacks

backend/architecture_parser.py

@@ -11,6 +11,14 @@ import torch
 import torch.nn as nn
 from transformers import AutoConfig, AutoModel, AutoModelForCausalLM, AutoModelForSeq2SeqLM
 
+# Monkeypatch for transformers import issues in some environment/model combinations
+try:
+    import transformers.utils.import_utils as import_utils
+    if not hasattr(import_utils, "is_torch_fx_available"):
+        import_utils.is_torch_fx_available = lambda: False
+except (ImportError, AttributeError):
+    pass
+
 
 def format_params(count: int) -> str:
     """Format parameter count in human-readable form."""
@@ -623,22 +631,31 @@ def load_model_for_inspection(model_id: str) -> Tuple[nn.Module, AutoConfig]:
         model = None
         errors = []
 
-        try:
-            model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
-        except Exception as e:
-            errors.append(f"CausalLM: {e}")
-
-        if model is None:
-            try:
-                model = AutoModelForSeq2SeqLM.from_config(config, trust_remote_code=True)
-            except Exception as e:
-                errors.append(f"Seq2SeqLM: {e}")
+        # Try to guess the model class from config
+        archs = getattr(config, "architectures", [])
+        is_encoder_decoder = getattr(config, "is_encoder_decoder", False)
+        
+        # Determine order of AutoModel classes to try
+        if is_encoder_decoder or any("Seq2Seq" in a or "ConditionalGeneration" in a for a in archs):
+            model_classes = [
+                (AutoModelForSeq2SeqLM, "Seq2SeqLM"),
+                (AutoModelForCausalLM, "CausalLM"),
+                (AutoModel, "AutoModel")
+            ]
+        else:
+            model_classes = [
+                (AutoModelForCausalLM, "CausalLM"),
+                (AutoModel, "AutoModel"),
+                (AutoModelForSeq2SeqLM, "Seq2SeqLM")
+            ]
 
-        if model is None:
+        for model_class, label in model_classes:
             try:
-                model = AutoModel.from_config(config, trust_remote_code=True)
+                model = model_class.from_config(config, trust_remote_code=True)
+                if model is not None:
+                    break
             except Exception as e:
-                errors.append(f"AutoModel: {e}")
+                errors.append(f"{label}: {e}")
 
     if model is None:
         raise ValueError(f"Could not load model architecture. Errors: {errors}")
@@ -902,22 +919,31 @@ def load_model_from_config(config_dict: Dict[str, Any]) -> Tuple[nn.Module, Auto
         model = None
         errors = []
 
-        try:
-            model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
-        except Exception as e:
-            errors.append(f"CausalLM: {e}")
-
-        if model is None:
-            try:
-                model = AutoModelForSeq2SeqLM.from_config(config, trust_remote_code=True)
-            except Exception as e:
-                errors.append(f"Seq2SeqLM: {e}")
+        # Try to guess the model class from config
+        archs = getattr(config, "architectures", [])
+        is_encoder_decoder = getattr(config, "is_encoder_decoder", False)
+        
+        # Determine order of AutoModel classes to try
+        if is_encoder_decoder or any("Seq2Seq" in a or "ConditionalGeneration" in a for a in archs):
+            model_classes = [
+                (AutoModelForSeq2SeqLM, "Seq2SeqLM"),
+                (AutoModelForCausalLM, "CausalLM"),
+                (AutoModel, "AutoModel")
+            ]
+        else:
+            model_classes = [
+                (AutoModelForCausalLM, "CausalLM"),
+                (AutoModel, "AutoModel"),
+                (AutoModelForSeq2SeqLM, "Seq2SeqLM")
+            ]
 
-        if model is None:
+        for model_class, label in model_classes:
             try:
-                model = AutoModel.from_config(config, trust_remote_code=True)
+                model = model_class.from_config(config, trust_remote_code=True)
+                if model is not None:
+                    break
             except Exception as e:
-                errors.append(f"AutoModel: {e}")
+                errors.append(f"{label}: {e}")
 
     if model is None:
         raise ValueError(f"Could not load model from config. Errors: {errors}")

backend/requirements.txt

@@ -3,7 +3,7 @@ uvicorn[standard]==0.27.0
 httpx==0.26.0
 pydantic==2.5.3
 python-multipart==0.0.6
-transformers>=4.36.0
+transformers>=4.54.0
 torch>=2.0.0
 accelerate>=0.25.0
 huggingface_hub>=0.20.0