read-lfs-files

backend/app/services/models.py

@@ -14,7 +14,6 @@ import sys
 import contextlib
 from concurrent.futures import ThreadPoolExecutor
 import tempfile
-
 from app.config import (
     QUEUE_REPO,
     HF_TOKEN,
@@ -481,10 +480,16 @@ class ModelService(HuggingFaceService):
                 raise Exception(error)
             logger.info(LogFormatter.success("Chat template validation passed"))
 
-
-        architectures = model_info.config.get("architectures", "")     
-        if architectures:
-            architectures = ";".join(architectures)
+        try:
+            architectures = "unknown"
+            if hasattr(model_info, 'config') and model_info.config:
+                config_architectures = model_info.config.get("architectures", [])
+                if config_architectures:
+                    architectures = ";".join(config_architectures)
+            logger.info(LogFormatter.info(f"Model architectures: {architectures}"))
+        except Exception as e:
+            logger.warning(LogFormatter.warning(f"Could not determine architectures: {str(e)}"))
+            architectures = "unknown"
 
         # Create eval entry
         eval_entry = {

backend/app/utils/model_validation.py

@@ -1,10 +1,12 @@
 import json
 import logging
 import asyncio
+import aiohttp
 from typing import Tuple, Optional, Dict, Any
 from datasets import load_dataset
 from huggingface_hub import HfApi, ModelCard, hf_hub_download
 from huggingface_hub import hf_api
+from huggingface_hub.utils import build_hf_headers
 from transformers import AutoConfig, AutoTokenizer
 from app.config.base import HF_TOKEN
 from app.config.hf_config import OFFICIAL_PROVIDERS_REPO
@@ -176,15 +178,35 @@ class ModelValidator:
     ) -> Tuple[bool, Optional[str], Optional[Any]]:
         """Check if model exists and is properly configured on the Hub"""
         try:
-            config = await asyncio.to_thread(
-                AutoConfig.from_pretrained,
-                model_name,
-                revision=revision,
-                trust_remote_code=trust_remote_code,
-                token=self.token,
-                force_download=True
-            )
-            
+            # First try regular API approach
+            try:
+                config = await asyncio.to_thread(
+                    AutoConfig.from_pretrained,
+                    model_name,
+                    revision=revision,
+                    trust_remote_code=trust_remote_code,
+                    token=self.token,
+                    force_download=True
+                )
+            except Exception as e:
+                logger.info(f"Standard config loading failed, attempting LFS fallback: {str(e)}")
+                
+                # Fallback to direct LFS file access
+                async with aiohttp.ClientSession() as session:
+                    config_url = f"https://huggingface.co/{model_name}/raw/{revision}/config.json"
+                    headers = build_hf_headers(token=self.token)
+                    
+                    async with session.get(config_url, headers=headers) as response:
+                        if response.status == 200:
+                            config_json = await response.json()
+                            logger.info("Successfully loaded config.json from LFS")
+                            config = AutoConfig.from_dict(config_json)
+                        else:
+                            error_msg = f"Failed to load config.json (Status {response.status})"
+                            logger.error(error_msg)
+                            raise Exception(error_msg)
+
+            # Test tokenizer if requested
             if test_tokenizer:
                 try:
                     await asyncio.to_thread(
@@ -198,9 +220,18 @@ class ModelValidator:
                     return False, f"The tokenizer is not available in an official Transformers release: {e}", None
                 except Exception:
                     return False, "The tokenizer cannot be loaded. Ensure the tokenizer class is part of a stable Transformers release and correctly configured.", None
-            
+
             return True, None, config
-            
+
+        except ValueError as e:
+            if "trust_remote_code" in str(e):
+                return False, "The model requires `trust_remote_code=True` to launch, and for safety reasons, we don't accept such models automatically.", None
+            return False, str(e), None
+        except Exception as e:
+            if "You are trying to access a gated repo." in str(e):
+                return True, "The model is gated and requires special access permissions.", None
+            return False, f"The model was not found or is misconfigured on the Hub. Error: {e}", None
+                
         except ValueError:
             return False, "The model requires `trust_remote_code=True` to launch, and for safety reasons, we don't accept such models automatically.", None
         except Exception as e: