Inspect blob cache sizes; fail fast with local_files_only=True

Replaces existence-only check with blob count + GB so we can see if
the cache is partial. Adds local_files_only=True to both from_pretrained
calls so a missing/incomplete cache raises an immediate OSError instead
of hanging on an indefinite download.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -45,20 +45,32 @@ print("[startup] all imports done", flush=True)
 
 dtype = torch.bfloat16
 
+def _inspect_cache(label, path):
+    if not os.path.isdir(path):
+        print(f"[startup] {label}: MISSING", flush=True)
+        return
+    blobs_dir = os.path.join(path, "blobs")
+    if os.path.isdir(blobs_dir):
+        blobs = os.listdir(blobs_dir)
+        total = sum(os.path.getsize(os.path.join(blobs_dir, b)) for b in blobs
+                    if os.path.isfile(os.path.join(blobs_dir, b)))
+        print(f"[startup] {label}: {len(blobs)} blobs, {total/1024**3:.2f} GB", flush=True)
+    else:
+        print(f"[startup] {label}: dir exists but no blobs/ subdirectory", flush=True)
+
 _hf_cache = os.path.join(os.environ.get("HF_HOME", "/data/hf_cache"), "hub")
-_transformer_cache = os.path.join(_hf_cache, "models--prithivMLmods--Qwen-Image-Edit-Rapid-AIO-V23")
-_pipeline_cache  = os.path.join(_hf_cache, "models--FireRedTeam--FireRed-Image-Edit-1.1")
 print(f"[startup] HF cache dir : {_hf_cache}", flush=True)
-print(f"[startup] transformer cache exists: {os.path.isdir(_transformer_cache)}", flush=True)
-print(f"[startup] pipeline cache exists   : {os.path.isdir(_pipeline_cache)}", flush=True)
 if os.path.isdir(_hf_cache):
     print(f"[startup] /hub contents: {os.listdir(_hf_cache)}", flush=True)
+_inspect_cache("transformer (V23)", os.path.join(_hf_cache, "models--prithivMLmods--Qwen-Image-Edit-Rapid-AIO-V23"))
+_inspect_cache("pipeline (FireRed-Image-Edit-1.1)", os.path.join(_hf_cache, "models--FireRedTeam--FireRed-Image-Edit-1.1"))
 
 print("[startup] loading transformer from_pretrained (prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V23)...", flush=True)
 _transformer = QwenImageTransformer2DModel.from_pretrained(
     "prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V23",
     torch_dtype=dtype,
     device_map="cuda",
+    local_files_only=True,
 )
 print("[startup] transformer loaded", flush=True)
 
@@ -67,6 +79,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     "FireRedTeam/FireRed-Image-Edit-1.1",
     transformer=_transformer,
     torch_dtype=dtype,
+    local_files_only=True,
 ).to(device)
 print("[startup] pipeline loaded and moved to device", flush=True)