fix: Configure CUDA precision to disable BF16 on older GPUs.

app_hf.py

@@ -84,6 +84,19 @@ threading.Thread(target=_warmup_hf_cache, daemon=True).start()
 device = "cuda" if torch.cuda.is_available() else "cpu"
 dtype = torch.float16 if torch.cuda.is_available() else torch.float32
 
+def _configure_cuda_precision():
+    if not torch.cuda.is_available():
+        return
+    # Avoid BF16 on GPUs that don't support it (sm80+).
+    try:
+        major, minor = torch.cuda.get_device_capability()
+        if (major, minor) < (8, 0):
+            torch.backends.cuda.matmul.allow_bf16 = False
+    except Exception:
+        torch.backends.cuda.matmul.allow_bf16 = False
+
+_configure_cuda_precision()
+
 class ModelManager:
     def __init__(self):
         self.models = {}