fix: make model loading offline-safe for Spaces runtime

Dockerfile

@@ -11,6 +11,9 @@ RUN mkdir -p /app/data/uploads /app/static/images
 
 ENV TRANSFORMERS_CACHE=/app/.cache/huggingface/transformers
 ENV HF_HOME=/app/.cache/huggingface
+ENV HF_HUB_OFFLINE=1
+ENV TRANSFORMERS_OFFLINE=1
+ENV HF_HUB_DISABLE_TELEMETRY=1
 
 EXPOSE 7860
 

README.md

@@ -6,6 +6,7 @@ colorTo: purple
 sdk: docker
 sdk_version: "1.0"
 app_file: app.py
+app_port: 7860
 pinned: false
 ---
 

app.py

@@ -54,12 +54,12 @@ def index():
 def models():
     return jsonify({
         "models": [
-            {"id": "qwen",   "name": "Qwen2.5-1.5B",    "provider": "Local (transformers)", "free": True},
+            {"id": "qwen",   "name": "Qwen2.5-1.5B",    "provider": "Local (optional path)", "free": True},
             {"id": "bart",   "name": "BART (fine-tuned)", "provider": "Local (transformers)", "free": True},
             {"id": "gemini", "name": "Gemini 2.0 Flash", "provider": "Google AI (API key)",  "free": False},
             {"id": "grok",   "name": "Grok-3 Mini",      "provider": "xAI (API key)",       "free": False},
         ],
-        "default": "qwen"
+        "default": "bart"
     })
 
 
@@ -70,7 +70,7 @@ def plot():
     if not data or not data.get('query'):
         return jsonify({'error': 'Missing required field: query'}), 400
 
-    logging.info(f"Plot request: model={data.get('model','qwen')} query={data.get('query')[:80]}")
+    logging.info(f"Plot request: model={data.get('model','bart')} query={data.get('query')[:80]}")
     result = agent.process_request(data)
     logging.info(f"Plot completed in {time.time() - t0:.2f}s")
     return jsonify(result)

llm_agent.py

@@ -13,6 +13,10 @@ load_dotenv()
 
 logger = logging.getLogger(__name__)
 
+
+def _model_dir(dirname: str) -> str:
+    return os.path.join(os.path.dirname(os.path.abspath(__file__)), dirname)
+
 # ---------------------------------------------------------------------------
 # Prompt templates
 # ---------------------------------------------------------------------------
@@ -103,10 +107,13 @@ class LLM_Agent:
     def _run_qwen(self, user_msg: str) -> str:
         if self._qwen_model is None:
             from transformers import AutoModelForCausalLM, AutoTokenizer
-            model_id = "Qwen/Qwen2.5-1.5B-Instruct"
+            # Prefer a local model path in Spaces to avoid any runtime network dependency.
+            model_id = os.getenv("QWEN_LOCAL_PATH", "")
+            if not model_id:
+                raise ValueError("Qwen local model is not configured in this Space")
             logger.info("Loading Qwen model (first request)...")
-            self._qwen_tokenizer = AutoTokenizer.from_pretrained(model_id)
-            self._qwen_model = AutoModelForCausalLM.from_pretrained(model_id)
+            self._qwen_tokenizer = AutoTokenizer.from_pretrained(model_id, local_files_only=True)
+            self._qwen_model = AutoModelForCausalLM.from_pretrained(model_id, local_files_only=True)
             logger.info("Qwen model loaded.")
         messages = [
             {"role": "system", "content": _SYSTEM_PROMPT},
@@ -155,10 +162,10 @@ class LLM_Agent:
     def _run_bart(self, query: str) -> str:
         if self._bart_model is None:
             from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
-            model_id = "ArchCoder/fine-tuned-bart-large"
+            model_id = os.getenv("BART_LOCAL_PATH", _model_dir("fine-tuned-bart-large"))
             logger.info("Loading BART model (first request)...")
-            self._bart_tokenizer = AutoTokenizer.from_pretrained(model_id)
-            self._bart_model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
+            self._bart_tokenizer = AutoTokenizer.from_pretrained(model_id, local_files_only=True)
+            self._bart_model = AutoModelForSeq2SeqLM.from_pretrained(model_id, local_files_only=True)
             logger.info("BART model loaded.")
         inputs = self._bart_tokenizer(
             query, return_tensors="pt", max_length=512, truncation=True
@@ -172,7 +179,7 @@ class LLM_Agent:
         t0        = time.time()
         query     = data.get("query", "")
         data_path = data.get("file_path")
-        model     = data.get("model", "qwen")
+        model     = data.get("model", "bart")
 
         if data_path and os.path.exists(data_path):
             self.data_processor  = DataProcessor(data_path)
@@ -194,8 +201,14 @@ class LLM_Agent:
             user_msg = _user_message(query, columns, dtypes, sample_rows)
             if   model == "gemini": raw_text = self._run_gemini(user_msg)
             elif model == "grok":   raw_text = self._run_grok(user_msg)
-            elif model == "bart":   raw_text = self._run_bart(query)
-            else:                   raw_text = self._run_qwen(user_msg)
+            elif model == "qwen":
+                try:
+                    raw_text = self._run_qwen(user_msg)
+                except Exception as qwen_exc:
+                    logger.warning(f"Qwen unavailable, falling back to BART: {qwen_exc}")
+                    raw_text = self._run_bart(query)
+            else:
+                raw_text = self._run_bart(query)
 
             logger.info(f"LLM [{model}] output: {raw_text}")
             parsed    = _parse_output(raw_text)