DeepSeek Update

src/audit/engine.py

@@ -314,18 +314,12 @@ def silence_specific_warnings():
 
 
 # -----------------------------------------------------------------------------
-# 4. Main Transformers and Auditor Classes
+# 4. Safe Transformers Import (Guaranteed not to fail on deprecated classes)
 # -----------------------------------------------------------------------------
 try:
     import torch
-    from transformers import (
-        AutoConfig,
-        AutoModel,
-        AutoModelForCausalLM,
-        AutoModelForSeq2SeqLM,
-        AutoModelForVision2Seq,
-        AutoTokenizer,
-    )
+    import transformers
+    from transformers import AutoConfig, AutoModel, AutoModelForCausalLM, AutoTokenizer
     HAS_TRANSFORMERS = True
 except ImportError:
     HAS_TRANSFORMERS = False
@@ -663,12 +657,16 @@ class TransformerAuditorBackend:
                         except Exception as e_final:
                             raise AuditError(f"Unable to load tokenizer for '{model_id}': {e_final}")
 
-        # 2. Universal Multi-Architecture Model Loader (CausalLM -> AutoModel -> Vision2Seq -> Seq2Seq)
+        # 2. Dynamically gather available model loaders (CausalLM -> AutoModel -> Multimodal)
         dtype = torch.bfloat16 if self.device == "cuda" else torch.float32
         model_loaded = None
         load_errors = []
 
-        loader_classes = [AutoModelForCausalLM, AutoModel, AutoModelForVision2Seq, AutoModelForSeq2SeqLM]
+        loader_classes = [AutoModelForCausalLM, AutoModel]
+        for extra_loader_name in ["AutoModelForImageTextToText", "AutoModelForVision2Seq", "AutoModelForSeq2SeqLM"]:
+            extra_cls = getattr(transformers, extra_loader_name, None)
+            if extra_cls is not None and extra_cls not in loader_classes:
+                loader_classes.append(extra_cls)
 
         with silence_specific_warnings():
             for loader_cls in loader_classes:
@@ -752,7 +750,7 @@ class TransformerAuditorBackend:
                 conds.append(prof["cond"])
 
         return {
-            "mean_srank": float(np.mean(sranks)) if sranks else 0.0,
+             "mean_srank": float(np.mean(sranks)) if sranks else 0.0,
             "mean_eff_rank": float(np.mean(eff_ranks)) if eff_ranks else 0.0,
             "mean_cond": float(np.mean(conds)) if conds else 0.0,
             "matrices_sampled": len(sranks),