Update app.py

app.py

@@ -2,8 +2,8 @@ import os
 import io
 import time
 import sys
-import requests
 import subprocess
+import requests
 from PIL import Image, ImageSequence
 import gradio as gr
 
@@ -13,52 +13,89 @@ try:
 except Exception as e:
     raise RuntimeError("llama-cpp-python import failed: " + str(e))
 
-# Ensure model is present: try Q4_K_M, fall back to Q4_K_S
-def download_with_curl(url: str, path: str) -> bool:
+MODEL_DIR = "model"
+EXPECTED_TARGET = os.path.join(MODEL_DIR, "llama-joycaption-q4_k_m.gguf")
+
+# Candidate direct-download URLs (try in order)
+CANDIDATES = [
+    # Jasaga7818 copy (often a direct GGUF)
+    ("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)
+    ("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")),
+    ("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")),
+]
+
+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:
-        if os.path.exists(path):
-            try:
+        try:
+            if os.path.exists(path):
                 os.remove(path)
-            except Exception:
-                pass
+        except Exception:
+            pass
+        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)
+        return head.startswith(b"GGUF")
+    except Exception:
         return False
 
 def ensure_model() -> str:
-    target = os.path.join("model", "llama-joycaption-q4_k_m.gguf")
-    alt = os.path.join("model", "llama-joycaption-q4_k_s.gguf")
-    if os.path.exists(target):
-        return target
-    # URLs (update if upstream changes)
-    url_m = "https://huggingface.co/mradermacher/llama-joycaption-beta-one-hf-llava-GGUF/resolve/main/llama-joycaption-beta-one-hf-llava-q4_k_m.gguf"
-    url_s = "https://huggingface.co/mradermacher/llama-joycaption-beta-one-hf-llava-GGUF/resolve/main/llama-joycaption-beta-one-hf-llava-q4_k_s.gguf"
-    sys.stderr.write("Model not found locally, attempting download (this can be several GB)...\n")
-    # Try M first
-    if download_with_curl(url_m, target):
-        sys.stderr.write("Downloaded Q4_K_M model.\n")
-        return target
-    # Try S
-    if download_with_curl(url_s, alt):
-        # create symlink so code expecting the M filename still works
-        try:
-            if os.path.exists(target):
-                os.remove(target)
-            os.symlink(os.path.basename(alt), target)
-        except Exception:
-            # fallback: copy file
-            try:
-                import shutil
-                shutil.copyfile(alt, target)
-            except Exception:
-                pass
-        sys.stderr.write("Downloaded Q4_K_S model and linked as Q4_K_M filename.\n")
-        return target
-    raise FileNotFoundError("Failed to download both Q4_K_M and Q4_K_S GGUF models. Check logs and URLs.")
-
-# Ensure model is available before loading llama.cpp
+    # 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")
+    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
+        else:
+            sys.stderr.write("Download failed for candidate; trying next.\n")
+
+    raise FileNotFoundError("Failed to download a valid GGUF model from candidates. Check URLs and repo availability.")
+
+# 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.")
@@ -82,7 +119,7 @@ def make_prompt_for_image(image_path: str, user_prompt: str = "Describe the imag
 
 # Initialize model (low-resource options)
 print("Loading GGUF model (this can take 30–120s)...", file=sys.stderr)
-# Adjust n_threads as appropriate for the Space CPU available
+# 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."):