recognize and issue error if GPU does not support bf16

Address #1298 which causes models to fail on T4 (and other pre-9.0 arch level GPUs) by selecting an alternate dtype when possible, and issue an error otherwise

Author: mikekgfb

torchchat/utils/build_utils.py

@@ -148,22 +148,51 @@ def get_precision():
 ###               dtype name to torch.dtype mapping                    ###
 
 
+def get_cuda_architecture(device=None):
+    device_str = get_device_str(device)
+    if "cuda" is in device_str and torch.cuda.is_available():
+        # Get the compute capability as (major, minor) tuple
+        capability = torch.cuda.get_device_capability(device)
+        return capability[0], capability[1]
+    else:
+        return 0, 0
+        
+        
+##########################################################################
+###               dtype name to torch.dtype mapping                    ###
+
+
 def name_to_dtype(name, device):
+    device_str = get_device_str(device)
+    # if it's CUDA, the architecture level indicates whether we can use bf16
+    major, minor = get_cuda_architecture(device)  
+    
     if (name == "fast") or (name == "fast16"):
         # MacOS now supports bfloat16
         import platform
 
         if platform.processor() == "arm":
-            device = get_device_str(device)
             # ARM CPU is faster with float16, MPS with bf16 if supported
-            if device == "cpu" or int(platform.mac_ver()[0].split(".")[0]) < 14:
+            if device_str == "cpu" or int(platform.mac_ver()[0].split(".")[0]) < 14:
                 return torch.float16
-        return torch.bfloat16
+                
+        # if it's not CUDA, we know it's bfloat16
+        if "cuda" is not in device_str:
+            return torch.bfloat16
+            
+        if major >= 9:
+            return torch.bfloat16
+        else:
+            return torch.float16
 
     try:
-        return name_to_dtype_dict[name]
+        dtype = name_to_dtype_dict[name]
     except KeyError:
         raise RuntimeError(f"unsupported dtype name {name} specified")
+        
+    if ("cuda" is in device_str) and (dtype == torch.bfloat16) and (major < 9):
+        raise RuntimeError(f"target device {device_str} does not support the bfloat16 data type")
+    return dtype
 
 
 def allowable_dtype_names() -> List[str]: