final

src/question_generator.py

@@ -1,46 +1,35 @@
 # ─────────────────────────────────────────────
-#  src/question_generator.py  (v4)
-#  Key fix: validate that the generated question
-#  actually targets the intended answer.
-#  Also filters circular questions like
-#  "What is the name of X?" when answer IS X.
+#  src/question_generator.py  (v5)
+#  Fix: replaced removed pipeline("text2text-generation")
+#  with direct AutoTokenizer + T5ForConditionalGeneration
+#  inference, compatible with transformers >= 4.46.
 # ─────────────────────────────────────────────
 
-from transformers import pipeline
 import re
 import sys, os
+import warnings
+import torch
 from transformers import AutoTokenizer, T5ForConditionalGeneration
-import streamlit as st
-
-@st.cache_resource
-def load_model():
-    tokenizer = AutoTokenizer.from_pretrained("valhalla/t5-small-qg-hl", use_fast=False)
-    model     = T5ForConditionalGeneration.from_pretrained("valhalla/t5-small-qg-hl")
-    model.eval()
-    return tokenizer, model
-
-tokenizer, qg_model = load_model()
-
 
 sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 from config import QG_MODEL_NAME, MAX_QUESTIONS
 
-print(f"[INFO] Loading QG model: {QG_MODEL_NAME} ...")
-import warnings
 warnings.filterwarnings("ignore")   # suppress HuggingFace FutureWarnings
 
+# ── Lazy model cache (avoids re-loading on every Streamlit rerun) ──
+_tokenizer = None
+_qg_model  = None
 
-
-try:
-    qg_pipeline = pipeline(
-        "text2text-generation",
-        model     = QG_MODEL_NAME,
-        tokenizer = QG_MODEL_NAME,
-    )
-    print("[INFO] Model loaded.")
-except Exception as e:
-    print(f"[ERROR] {e}")
-    raise
+def _get_model():
+    """Load model once and cache in module-level globals."""
+    global _tokenizer, _qg_model
+    if _tokenizer is None or _qg_model is None:
+        print(f"[INFO] Loading QG model: {QG_MODEL_NAME} ...")
+        _tokenizer = AutoTokenizer.from_pretrained(QG_MODEL_NAME, use_fast=False)
+        _qg_model  = T5ForConditionalGeneration.from_pretrained(QG_MODEL_NAME)
+        _qg_model.eval()
+        print("[INFO] Model loaded.")
+    return _tokenizer, _qg_model
 
 
 def highlight_answer(sentence: str, answer: str) -> str:
@@ -108,24 +97,38 @@ def generate_question(sentence: str, answer: str) -> str | None:
     """
     Generate a question for a (sentence, answer) pair.
     Returns the best valid question string, or None.
+    Uses direct tokenizer + model inference (no pipeline()).
     """
+    tokenizer, qg_model = _get_model()
+
     highlighted = highlight_answer(sentence, answer)
     input_text  = f"generate question: {highlighted}"
 
     try:
-        outputs = qg_pipeline(
+        inputs = tokenizer(
             input_text,
-            max_new_tokens       = 64,
-            num_beams            = 5,
-            num_return_sequences = 3,
-            early_stopping       = True,
+            return_tensors = "pt",
+            max_length     = 512,
+            truncation     = True,
         )
+        with torch.no_grad():
+            output_ids = qg_model.generate(
+                inputs["input_ids"],
+                max_new_tokens       = 64,
+                num_beams            = 5,
+                num_return_sequences = 3,
+                early_stopping       = True,
+            )
+        outputs = [
+            tokenizer.decode(ids, skip_special_tokens=True)
+            for ids in output_ids
+        ]
     except Exception as e:
         print(f"  [QG] Generation error: {e}")
         return None
 
-    for output in outputs:
-        q = output["generated_text"].strip()
+    for q in outputs:
+        q = q.strip()
         if not q.endswith("?"):
             q += "?"
         if answer_is_addressable(q, answer):