update per comments

vllm/model_executor/layers/quantization/bitsandbytes.py

@@ -35,7 +35,7 @@ def get_name(self) -> str:
 
     @classmethod
     def get_supported_act_dtypes(self) -> List[torch.dtype]:
-        return [torch.half]
+        return [torch.float32, torch.float16, torch.bfloat16]
 
     @classmethod
     def get_min_capability(self) -> int:
@@ -123,7 +123,7 @@ def create_weights(self, layer: torch.nn.Module,
             qweight,
             {
                 "input_dim": 0,
-                # In bitsandbytes, a tensor of shape [n,m] is qunatized to
+                # In bitsandbytes, a tensor of shape [n,m] is quantized to
                 #[n*m/pack_ratio, 1],so the output_dim is 0
                 "output_dim": 0,
                 "pack_factor": quant_ratio,
@@ -140,7 +140,7 @@ def apply(self,
         # only load the bitsandbytes module when needed
         from bitsandbytes import matmul_4bit
 
-        orginal_type = x.dtype
+        original_type = x.dtype
         bf_x = x.to(torch.bfloat16)
 
         qweight = layer.qweight
@@ -167,7 +167,7 @@ def apply(self,
 
             current_index += output_size
 
-        out = out.to(orginal_type)
+        out = out.to(original_type)
 
         if bias is not None:
             out += bias

vllm/model_executor/model_loader/loader.py

@@ -547,7 +547,7 @@ def save_model(
 
 
 class BitsAndBytesModelLoader(BaseModelLoader):
-    """Model loader to load model weights with BitAndBytes quabntization."""
+    """Model loader to load model weights with BitAndBytes quantization."""
 
     default_target_modules = [
         "gate_proj", "down_proj", "up_proj", "q_proj", "k_proj", "v_proj",
@@ -654,7 +654,6 @@ def _get_quantized_weights_iterator(
     ) -> Tuple[Generator[Tuple[str, torch.Tensor], None, None], Dict[str,
                                                                      Any]]:
         """Get an iterator to the model weights with bitsandbytes quantization,
-
         as well as the quantization state dictionary."""
 
         # only load the bitsandbytes module when needed