Update app.py

app.py

@@ -14,27 +14,26 @@ except Exception as e:
     raise RuntimeError("llama-cpp-python import failed: " + str(e))
 
 MODEL_DIR = "model"
-EXPECTED_TARGET = os.path.join(MODEL_DIR, "llama-joycaption-q4_k_m.gguf")
+MODEL_MAIN = os.path.join(MODEL_DIR, "llama-joycaption-q4_k_m.gguf")
+MODEL_FALLBACK = os.path.join(MODEL_DIR, "llama-joycaption-q4_k_s.gguf")
 
 # Candidate direct-download URLs (try in order)
 CANDIDATES = [
-    # Jasaga7818 copy (often a direct GGUF)
+    # Primary Q4_K_M (Jasaga then mradermacher)
     ("https://huggingface.co/Jasaga7818/llama-joycaption-beta-one-hf-llava-Q4_K_M-GGUF/resolve/main/llama-joycaption-beta-one-hf-llava-q4_k_m.gguf",
-     EXPECTED_TARGET),
-    # mradermacher (alternate host)
+     MODEL_MAIN),
     ("https://huggingface.co/mradermacher/llama-joycaption-beta-one-hf-llava-GGUF/resolve/main/llama-joycaption-beta-one-hf-llava-q4_k_m.gguf",
-     EXPECTED_TARGET),
-    # Fallback to Q4_K_S (Jasaga)
-    ("https://huggingface.co/Jasaga7818/llama-joycaption-beta-one-hf-llava-Q4_K_M-GGUF/resolve/main/llama-joycaption-beta-one-hf-llava-q4_k_s.gguf",
-     os.path.join(MODEL_DIR, "llama-joycaption-q4_k_s.gguf")),
+     MODEL_MAIN),
+    # Fallback Q4_K_S (mradermacher / Jasaga)
     ("https://huggingface.co/mradermacher/llama-joycaption-beta-one-hf-llava-GGUF/resolve/main/llama-joycaption-beta-one-hf-llava-q4_k_s.gguf",
-     os.path.join(MODEL_DIR, "llama-joycaption-q4_k_s.gguf")),
+     MODEL_FALLBACK),
+    ("https://huggingface.co/Jasaga7818/llama-joycaption-beta-one-hf-llava-Q4_K_M-GGUF/resolve/main/llama-joycaption-beta-one-hf-llava-q4_k_s.gguf",
+     MODEL_FALLBACK),
 ]
 
 def download_curl(url: str, path: str) -> bool:
     os.makedirs(os.path.dirname(path), exist_ok=True)
     try:
-        # Use curl for resume support and progress in logs
         subprocess.check_call(["curl", "-L", "-C", "-", "-o", path, url])
         return True
     except Exception:
@@ -46,8 +45,6 @@ def download_curl(url: str, path: str) -> bool:
         return False
 
 def is_valid_gguf(path: str) -> bool:
-    # GGUF files start with "GGUF" in ASCII at offset 0 (0x47 0x47 0x55 0x46).
-    # Some converted uploads may be HTML pages or redirects; check header.
     try:
         with open(path, "rb") as f:
             head = f.read(8)
@@ -55,50 +52,85 @@ def is_valid_gguf(path: str) -> bool:
     except Exception:
         return False
 
-def ensure_model() -> str:
-    # If already present (and valid), use it.
-    if os.path.exists(EXPECTED_TARGET) and is_valid_gguf(EXPECTED_TARGET):
-        sys.stderr.write(f"Model already present and valid at {EXPECTED_TARGET}\n")
-        return EXPECTED_TARGET
-
-    sys.stderr.write("Model not found locally or invalid, attempting download (several GB)...\n")
+def ensure_models_downloaded():
+    # If main present and valid, done.
+    if os.path.exists(MODEL_MAIN) and is_valid_gguf(MODEL_MAIN):
+        sys.stderr.write(f"Found valid main model: {MODEL_MAIN}\n")
+        return
+    # If fallback present and valid, done.
+    if os.path.exists(MODEL_FALLBACK) and is_valid_gguf(MODEL_FALLBACK):
+        sys.stderr.write(f"Found valid fallback model: {MODEL_FALLBACK}\n")
+        return
+
+    sys.stderr.write("Model(s) missing or invalid; attempting downloads...\n")
     for url, dest in CANDIDATES:
-        sys.stderr.write(f"Attempting download: {url} -> {dest}\n")
-        if download_curl(url, dest):
-            sys.stderr.write(f"Downloaded candidate to {dest}; verifying header...\n")
-            if is_valid_gguf(dest):
-                # If candidate wasn't the expected filename, create symlink so rest of code can use EXPECTED_TARGET.
-                if os.path.abspath(dest) != os.path.abspath(EXPECTED_TARGET):
-                    try:
-                        if os.path.exists(EXPECTED_TARGET):
-                            os.remove(EXPECTED_TARGET)
-                        os.symlink(os.path.basename(dest), EXPECTED_TARGET)
-                        sys.stderr.write(f"Created symlink {EXPECTED_TARGET} -> {os.path.basename(dest)}\n")
-                    except Exception:
-                        # fallback: copy
-                        try:
-                            import shutil
-                            shutil.copyfile(dest, EXPECTED_TARGET)
-                            sys.stderr.write(f"Copied {dest} to {EXPECTED_TARGET}\n")
-                        except Exception:
-                            sys.stderr.write("Warning: failed to symlink or copy candidate to expected filename.\n")
-                sys.stderr.write("Model verified as GGUF and ready.\n")
-                return EXPECTED_TARGET
-            else:
-                sys.stderr.write("Downloaded file is not a valid GGUF (header mismatch). Removing and trying next.\n")
-                try:
-                    os.remove(dest)
-                except Exception:
-                    pass
+        sys.stderr.write(f"Downloading {url} -> {dest}\n")
+        ok = download_curl(url, dest)
+        if not ok:
+            sys.stderr.write(f"Download failed for {url}\n")
+            continue
+        if is_valid_gguf(dest):
+            sys.stderr.write(f"Downloaded and verified GGUF at {dest}\n")
+            # if we downloaded fallback but main missing, don't copy; we'll try to load fallback later
+            if dest == MODEL_MAIN:
+                return
+            # if dest is fallback, still continue loop to attempt main first (if available)
         else:
-            sys.stderr.write("Download failed for candidate; trying next.\n")
+            sys.stderr.write(f"Downloaded file at {dest} is not a valid GGUF (header mismatch). Removing.\n")
+            try:
+                os.remove(dest)
+            except Exception:
+                pass
 
-    raise FileNotFoundError("Failed to download a valid GGUF model from candidates. Check URLs and repo availability.")
+    sys.stderr.write("Download attempts finished.\n")
 
-# Ensure model exists and is a GGUF before importing/initializing Llama
-MODEL_PATH = ensure_model()
-if not os.path.exists(MODEL_PATH):
-    raise FileNotFoundError(f"Model not found at {MODEL_PATH} after download attempt.")
+def try_load_model(path: str, n_ctx: int = 2048, n_threads: int = 2):
+    try:
+        sys.stderr.write(f"Initializing Llama with model {path}...\n")
+        llm = Llama(model_path=path, n_ctx=n_ctx, n_threads=n_threads)
+        sys.stderr.write("Model loaded successfully.\n")
+        return llm
+    except Exception as e:
+        sys.stderr.write(f"Failed to load model {path}: {e}\n")
+        return None
+
+# Ensure at least one model file is present (download if needed)
+ensure_models_downloaded()
+
+# Prefer main, then fallback
+model_to_try = None
+if os.path.exists(MODEL_MAIN) and is_valid_gguf(MODEL_MAIN):
+    model_to_try = MODEL_MAIN
+elif os.path.exists(MODEL_FALLBACK) and is_valid_gguf(MODEL_FALLBACK):
+    model_to_try = MODEL_FALLBACK
+else:
+    # attempt to download again and pick whatever exists
+    ensure_models_downloaded()
+    if os.path.exists(MODEL_MAIN) and is_valid_gguf(MODEL_MAIN):
+        model_to_try = MODEL_MAIN
+    elif os.path.exists(MODEL_FALLBACK) and is_valid_gguf(MODEL_FALLBACK):
+        model_to_try = MODEL_FALLBACK
+
+if model_to_try is None:
+    raise FileNotFoundError("No valid GGUF model found. Place a compatible GGUF under model/ with filename\n"
+                            "model/llama-joycaption-q4_k_m.gguf or model/llama-joycaption-q4_k_s.gguf.")
+
+# Attempt to load chosen model; if load fails for magic/version, try fallback (if different)
+llm = try_load_model(model_to_try, n_ctx=2048, n_threads=2)
+if llm is None and model_to_try == MODEL_MAIN and os.path.exists(MODEL_FALLBACK) and is_valid_gguf(MODEL_FALLBACK):
+    sys.stderr.write("Primary model failed to load; attempting fallback model.\n")
+    llm = try_load_model(MODEL_FALLBACK, n_ctx=2048, n_threads=2)
+
+if llm is None:
+    # Provide clear diagnostic and exit
+    sys.stderr.write("\nERROR: All model load attempts failed. Likely causes:\n"
+                     " - The GGUF uses a newer GGUF version not supported by the installed llama.cpp/llama-cpp-python.\n"
+                     " - The file is corrupted despite the header check.\n\n"
+                     "Recommended fixes:\n"
+                     " - Install a newer llama.cpp/llama-cpp-python built from main/master (supports newer GGUF versions).\n"
+                     " - Or place a known-compatible GGUF (Q4_K_S from mradermacher or older GGUF) at model/llama-joycaption-q4_k_m.gguf\n"
+                     " - To inspect the header run: hexdump -n4 model/llama-joycaption-q4_k_m.gguf\n")
+    raise RuntimeError("Model load failed for all candidates.")
 
 def download_bytes(url: str, timeout: int = 30) -> bytes:
     with requests.get(url, stream=True, timeout=timeout) as r:
@@ -117,11 +149,6 @@ def make_prompt_for_image(image_path: str, user_prompt: str = "Describe the imag
     # JoyCaption-style multimodal GGUFs accept <img>{path}</img>
     return f"<img>{image_path}</img>\nUser: {user_prompt}\nAssistant:"
 
-# Initialize model (low-resource options)
-print("Loading GGUF model (this can take 30–120s)...", file=sys.stderr)
-# Adjust n_threads for the Space CPU; increase if you know you have more cores available.
-llm = Llama(model_path=MODEL_PATH, n_ctx=2048, n_threads=2)
-
 def generate_caption_from_url(url: str, prompt: str = "Describe the image."):
     if not url:
         return "No URL provided."