Enhance model loading process with improved error handling and alternative loading method; configure Hugging Face endpoint and disable offline mode

app.py

@@ -9,6 +9,10 @@ import sys
 os.environ['HF_HOME'] = '/tmp/huggingface'
 os.environ['TRANSFORMERS_CACHE'] = '/tmp/huggingface/transformers'
 os.environ['HF_DATASETS_CACHE'] = '/tmp/huggingface/datasets'
+# Ensure proper Hugging Face endpoint
+os.environ['HF_ENDPOINT'] = 'https://huggingface.co'
+# Disable offline mode to allow downloads
+os.environ['TRANSFORMERS_OFFLINE'] = '0'
 import torch
 import transformers
 from transformers import PreTrainedTokenizerFast
@@ -103,22 +107,49 @@ def load_model_cached(model_type):
         cache_dir = "/tmp/huggingface/transformers"
         os.makedirs(cache_dir, exist_ok=True)
         
+        # Clear any potential proxy issues
+        import requests
+        session = requests.Session()
+        session.trust_env = False
+        
+        # Try loading with explicit parameters
         model = tranception.model_pytorch.TranceptionLMHeadModel.from_pretrained(
             pretrained_model_name_or_path=model_path,
             cache_dir=cache_dir,
             local_files_only=False,  # Allow downloading if not cached
-            resume_download=True     # Resume incomplete downloads
+            resume_download=True,    # Resume incomplete downloads
+            force_download=False,    # Don't force re-download if cached
+            proxies=None,           # Explicitly set no proxies
+            use_auth_token=None,    # No auth token needed for public models
+            revision="main"         # Use main branch
         )
         MODEL_CACHE[model_type] = model
         print(f"{model_type} model loaded and cached")
         return model
     except Exception as e:
         print(f"Error loading {model_type} model: {e}")
-        # Fallback to Medium if requested model fails
-        if model_type != "Medium":
-            print("Falling back to Medium model...")
-            return load_model_cached("Medium")
-        raise
+        print(f"Attempting alternative loading method...")
+        
+        # Try alternative loading approach
+        try:
+            # Manually specify the full model ID
+            full_model_id = f"PascalNotin/Tranception_{model_type}"
+            model = tranception.model_pytorch.TranceptionLMHeadModel.from_pretrained(
+                full_model_id,
+                cache_dir=cache_dir,
+                local_files_only=False,
+                trust_remote_code=True  # Allow custom model code
+            )
+            MODEL_CACHE[model_type] = model
+            print(f"{model_type} model loaded successfully with alternative method")
+            return model
+        except Exception as e2:
+            print(f"Alternative loading also failed: {e2}")
+            # Fallback to Medium if requested model fails
+            if model_type != "Medium":
+                print("Falling back to Medium model...")
+                return load_model_cached("Medium")
+            raise
 
 AA_vocab = "ACDEFGHIKLMNPQRSTVWY"
 tokenizer = PreTrainedTokenizerFast(tokenizer_file="./tranception/utils/tokenizers/Basic_tokenizer",