Stabilize FP8 kernel loading on ZeroGPU

app.py

@@ -14,6 +14,14 @@ import time
 import uuid
 from typing import Any
 
+# Hub kernel downloads can create concurrent tqdm bars inside a ZeroGPU worker;
+# tqdm's shared position bookkeeping is racy there.  Progress is irrelevant in
+# a Space runtime, so disable it before importing Hugging Face libraries.
+os.environ.setdefault("HF_HUB_DISABLE_PROGRESS_BARS", "1")
+# DeepGEMM does not yet recognize the ZeroGPU Blackwell SM120 device.  The
+# fine-grained FP8 Triton kernel supports it and is the correct fallback.
+os.environ.setdefault("TRANSFORMERS_DISABLE_DEEPGEMM_LINEAR", "1")
+
 import gradio as gr
 import spaces
 import torch