add find bs func2

pipeline.py

@@ -1820,6 +1820,32 @@ def ensemble_depth(
     return depth, uncertainty  # [1,1,H,W], [1,1,H,W]
 
 
+# Search table for suggested max. inference batch size
+bs_search_table = [
+    # tested on A100-PCIE-80GB
+    {"res": 768, "total_vram": 79, "bs": 35, "dtype": torch.float32},
+    {"res": 1024, "total_vram": 79, "bs": 20, "dtype": torch.float32},
+    # tested on A100-PCIE-40GB
+    {"res": 768, "total_vram": 39, "bs": 15, "dtype": torch.float32},
+    {"res": 1024, "total_vram": 39, "bs": 8, "dtype": torch.float32},
+    {"res": 768, "total_vram": 39, "bs": 30, "dtype": torch.float16},
+    {"res": 1024, "total_vram": 39, "bs": 15, "dtype": torch.float16},
+    # tested on RTX3090, RTX4090
+    {"res": 512, "total_vram": 23, "bs": 20, "dtype": torch.float32},
+    {"res": 768, "total_vram": 23, "bs": 7, "dtype": torch.float32},
+    {"res": 1024, "total_vram": 23, "bs": 3, "dtype": torch.float32},
+    {"res": 512, "total_vram": 23, "bs": 40, "dtype": torch.float16},
+    {"res": 768, "total_vram": 23, "bs": 18, "dtype": torch.float16},
+    {"res": 1024, "total_vram": 23, "bs": 10, "dtype": torch.float16},
+    # tested on GTX1080Ti
+    {"res": 512, "total_vram": 10, "bs": 5, "dtype": torch.float32},
+    {"res": 768, "total_vram": 10, "bs": 2, "dtype": torch.float32},
+    {"res": 512, "total_vram": 10, "bs": 10, "dtype": torch.float16},
+    {"res": 768, "total_vram": 10, "bs": 5, "dtype": torch.float16},
+    {"res": 1024, "total_vram": 10, "bs": 3, "dtype": torch.float16},
+]
+
+
 def find_batch_size(ensemble_size: int, input_res: int, dtype: torch.dtype) -> int:
     """
     Automatically search for suitable operating batch size.