Fix version sort + add GGUF file validation

app.py

@@ -1,10 +1,10 @@
 import os
 import sys
 import subprocess
+import re
 
 # =============================================================
-# 1. Bootstrap: Install a KNOWN-GOOD prebuilt CPU-only wheel
-#    Versions >= 0.3.20 use "+cpu" tag = zero CUDA dependency
+# 1. Bootstrap: Install latest prebuilt CPU-only llama-cpp-python
 # =============================================================
 def bootstrap_llama_cpp():
     print("Bootstrapping llama-cpp-python (CPU-only, manylinux)...")
@@ -15,33 +15,33 @@ def bootstrap_llama_cpp():
 
     py_tag = f"cp{sys.version_info.major}{sys.version_info.minor}"
 
-    # Filter: linux/x86_64/cpu, version 0.3.x, "+cpu" in name, py3-none or matching cpXXX
-    candidates = [
-        p for p in files
-        if p.endswith(".whl")
-        and "linux/x86_64/cpu/" in p
-        and "/0.3." in p
-        and "+cpu" in p
-        and ("py3-none" in p or py_tag in p)
-    ]
+    # Collect all CPU wheels for linux x86_64, version 0.3.x
+    candidates = []
+    for p in files:
+        if not p.endswith(".whl"):
+            continue
+        if "linux/x86_64/cpu/" not in p:
+            continue
+        if "cuda" in p.lower():
+            continue
+        if py_tag not in p and "py3-none" not in p:
+            continue
+        # Extract version number
+        m = re.search(r"/0\.3\.(\d+)/", p)
+        if not m:
+            continue
+        minor = int(m.group(1))
+        if minor < 20:  # Need >= 0.3.20 for Qwen3.5 support
+            continue
+        candidates.append((minor, p))
+
+    # Sort by version number descending → pick latest
+    candidates.sort(key=lambda x: x[0], reverse=True)
 
     if not candidates:
-        # Fallback: any 0.3.x cpu wheel matching our python
-        candidates = [
-            p for p in files
-            if p.endswith(".whl")
-            and "linux/x86_64/cpu/" in p
-            and "/0.3." in p
-            and ("py3-none" in p or py_tag in p)
-            and "cuda" not in p.lower()
-        ]
-
-    # Sort descending to pick latest version
-    candidates.sort(reverse=True)
-    if not candidates:
-        raise RuntimeError("No compatible CPU wheel found!")
+        raise RuntimeError("No compatible CPU wheel found for >= 0.3.20!")
 
-    chosen = candidates[0]
+    chosen = candidates[0][1]
     print(f"Selected wheel: {chosen}")
     whl_path = hf_hub_download("AIencoder/llama-cpp-wheels", chosen, repo_type="dataset")
     subprocess.check_call([
@@ -51,7 +51,6 @@ def bootstrap_llama_cpp():
 
 try:
     from llama_cpp import Llama
-    # Verify it actually loads (catches CUDA linkage errors)
     print("llama-cpp-python already available.")
 except (ImportError, RuntimeError, OSError):
     bootstrap_llama_cpp()
@@ -63,11 +62,27 @@ from fastapi import FastAPI, HTTPException, Header
 import gradio as gr
 
 # =============================================================
-# 2. Load GGUF Model
+# 2. Verify & Load GGUF Model
 # =============================================================
 MODEL_PATH = "Qwen3.5-0.8B.Q4_K_M.gguf"
 API_KEY = os.getenv("API_KEY", "commitcon-secret-2026")
 
+# Diagnostic: check if file exists and is real (not LFS pointer)
+if not os.path.exists(MODEL_PATH):
+    raise FileNotFoundError(f"Model file {MODEL_PATH} not found in /app/")
+
+file_size = os.path.getsize(MODEL_PATH)
+print(f"Model file size: {file_size / 1e6:.1f} MB")
+
+if file_size < 1_000_000:  # < 1MB means it's an LFS pointer, not the real file
+    with open(MODEL_PATH, "r", errors="ignore") as f:
+        content = f.read(200)
+    raise RuntimeError(
+        f"Model file is only {file_size} bytes — likely a Git LFS pointer, not the real weights.\n"
+        f"File content: {content}\n"
+        f"Fix: Re-upload the .gguf file using the Hugging Face web UI (Files tab → Add file → Upload)."
+    )
+
 print(f"Loading {MODEL_PATH} via llama-cpp-python ...")
 llm = Llama(
     model_path=MODEL_PATH,