Update app.py

app.py

@@ -10,7 +10,7 @@ def install_package(package, version=None):
     print(f"Installing {package_spec}...")
     try:
         if package == "torch":
-            # CPU-only build is ~200MB vs ~900MB for CUDA — avoids disk quota errors
+            # CPU-only build (~190MB) instead of CUDA (~900MB) to avoid disk quota errors
             subprocess.check_call([
                 sys.executable, "-m", "pip", "install", "--no-cache-dir",
                 "torch", "--index-url", "https://download.pytorch.org/whl/cpu"
@@ -24,14 +24,19 @@ def install_package(package, version=None):
         raise
 
 
-# Required packages
+# Required packages — ORDER MATTERS:
+#   tokenizers must be pinned before transformers so pip does not upgrade it.
+#   transformers 5.x dropped the "summarization" pipeline task entirely,
+#   so we pin to 4.46.3 (last v4 release) which still has it.
+#   tokenizers 0.21.3 is required by transformers 4.46.3 and has
+#   pre-built abi3 wheels that work on Python 3.13 without Rust compilation.
 required_packages = {
     "gradio": None,
     "torch": None,
-    "transformers": None,
+    "tokenizers": "0.21.3",    # must come before transformers; abi3 wheel works on Py3.13
+    "transformers": "4.46.3",  # last v4 release; still supports the summarization pipeline
 }
 
-# Install missing packages using importlib (pkg_resources is deprecated)
 for package, version in required_packages.items():
     try:
         __import__(package)
@@ -92,7 +97,7 @@ def summarize_text(text, model_name, summary_length, num_beams):
 
     try:
         global summarizer
-        # Reload pipeline only if the model has changed, always on CPU
+        # Reload pipeline on CPU whenever the model changes
         summarizer = pipeline("summarization", model=model_name, device=-1)
 
         length_mapping = {