truncate intermediate fp32 if converting bf16 to bf16

Author: CISC

convert-hf-to-gguf.py

@@ -295,8 +295,8 @@ def write_tensors(self):
 
                 if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
                     if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                        data = gguf.quantize_bf16(data)
-                        assert data.dtype == np.int16
+                        data = gguf.truncate_bf16(data) if old_dtype == torch.bfloat16 else gguf.quantize_bf16(data)
+                        assert data.dtype in (np.int16, np.uint16)
                         data_qtype = gguf.GGMLQuantizationType.BF16
 
                     elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):