Update app.py

app.py

@@ -4,24 +4,43 @@ from pydantic import BaseModel
 from huggingface_hub import snapshot_download
 from llama_cpp import Llama
 
-# -------- Writable cache/weights dirs (HF Docker Spaces) --------
-HOME = os.path.expanduser("~") or "/home/user"
-BASE_DIR = os.getenv("SPACE_CACHE_DIR", os.path.join(HOME, ".cache"))
-HF_HOME = os.path.join(BASE_DIR, "huggingface")
-MODELS_DIR = os.path.join(BASE_DIR, "models")
+# ---------- pick a writable cache dir ----------
+def first_writable(paths):
+    for p in paths:
+        if not p:
+            continue
+        try:
+            os.makedirs(p, exist_ok=True)
+            testfile = os.path.join(p, ".write_test")
+            with open(testfile, "w") as f:
+                f.write("ok")
+            os.remove(testfile)
+            return p
+        except Exception:
+            continue
+    raise RuntimeError("No writable cache dir found")
 
-os.environ["HF_HOME"] = HF_HOME
-os.environ["HF_HUB_CACHE"] = os.path.join(HF_HOME, "hub")
+CACHE_BASE = first_writable([
+    os.getenv("SPACE_CACHE_DIR"),   # optional env override
+    "/app/.cache",                  # WORKDIR is usually writable on Spaces
+    "/tmp/app_cache",               # always writable fallback
+])
+
+HF_HOME = os.path.join(CACHE_BASE, "huggingface")
+MODELS_DIR = os.path.join(CACHE_BASE, "models")
 os.makedirs(HF_HOME, exist_ok=True)
 os.makedirs(MODELS_DIR, exist_ok=True)
 
+os.environ["HF_HOME"] = HF_HOME
+os.environ["HF_HUB_CACHE"] = os.path.join(HF_HOME, "hub")
+
 # ---- Model selection (override in Settings → Variables if needed) ----
 MODEL_REPO = os.getenv("MODEL_REPO", "Qwen/Qwen2.5-3B-Instruct-GGUF")
 MODEL_FILE = os.getenv("MODEL_FILE", "qwen2.5-3b-instruct-q4_k_m.gguf")
 
 # Inference knobs
-N_CTX   = int(os.getenv("N_CTX", 2048))
-N_BATCH = int(os.getenv("N_BATCH", 64))
+N_CTX    = int(os.getenv("N_CTX", 2048))
+N_BATCH  = int(os.getenv("N_BATCH", 64))
 N_THREADS = os.cpu_count() or 2
 
 app = FastAPI(title="Qwen Planner API (CPU)")
@@ -51,7 +70,7 @@ def ensure_model():
 
 @app.get("/healthz")
 def healthz():
-    return {"status": "ok", "loaded": model_loaded, "model_file": MODEL_FILE}
+    return {"status": "ok", "loaded": model_loaded, "cache_base": CACHE_BASE, "model_file": MODEL_FILE}
 
 SYSTEM_PROMPT = "You are a concise assistant. Reply briefly in plain text."