Add granular startup logging to diagnose hang

Adds flush=True and [startup] breadcrumbs between every import and
from_pretrained call so the log stream shows exactly where startup
stalls. Also splits from_pretrained into two named steps and prints
CUDA device_count at boot.

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

app.py

@@ -21,33 +21,47 @@ MAX_OUTPUT_DIM = 2048
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES"))
-print("torch.__version__ =", torch.__version__)
-print("Using device:", device)
+print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES"), flush=True)
+print("torch.__version__ =", torch.__version__, flush=True)
+print("Using device:", device, flush=True)
+print(f"CUDA device_count={torch.cuda.device_count()}, is_available={torch.cuda.is_available()}", flush=True)
 
 # TF32 matmul: ~10-15% free speedup on Ampere/Hopper (bfloat16 accumulation paths benefit too)
 torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cudnn.allow_tf32 = True
+print("[startup] TF32 enabled", flush=True)
 
+print("[startup] importing dimensions...", flush=True)
 from dimensions import compute_output_dimensions
+print("[startup] importing diffusers...", flush=True)
 from diffusers import FlowMatchEulerDiscreteScheduler
+print("[startup] importing QwenImageEditPlusPipeline...", flush=True)
 from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
+print("[startup] importing QwenImageTransformer2DModel...", flush=True)
 from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
+print("[startup] importing QwenDoubleStreamAttnProcessorFA3...", flush=True)
 from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
+print("[startup] all imports done", flush=True)
 
 dtype = torch.bfloat16
 
+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",
+)
+print("[startup] transformer loaded", flush=True)
+
+print("[startup] loading pipeline from_pretrained (FireRedTeam/FireRed-Image-Edit-1.1)...", flush=True)
 pipe = QwenImageEditPlusPipeline.from_pretrained(
     "FireRedTeam/FireRed-Image-Edit-1.1",
-    transformer=QwenImageTransformer2DModel.from_pretrained(
-        "prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V23",
-        torch_dtype=dtype,
-        device_map="cuda",
-    ),
+    transformer=_transformer,
     torch_dtype=dtype,
 ).to(device)
+print("[startup] pipeline loaded and moved to device", flush=True)
 
-print("Using default attention processor (FA3 skipped for ZeroGPU GPU-arch compatibility).")
+print("Using default attention processor (FA3 skipped for ZeroGPU GPU-arch compatibility).", flush=True)
 
 print("torch.compile skipped: lazy Triton kernel compilation inside @spaces.GPU always exceeds ZeroGPU's task timeout.")