fix enforced bf16 data type on SM75 and lower devices

modeling_dots_vision.py

@@ -489,9 +489,13 @@ class DotsVisionTransformer(PreTrainedModel):
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb
 
-    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, bf16=True) -> torch.Tensor:
-        if bf16:
-            hidden_states = hidden_states.bfloat16()
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, bf16=None) -> torch.Tensor:
+    	# 尝试修复SM75及之前不支持BF16设备的报错
+    	# 若未显式指定 bf16，则根据权重 dtype 推断
+        if bf16 is None:
+            bf16 = (self.dtype == torch.bfloat16)
+        # 始终将输入显式对齐到本模块的计算精度，避免 input/bias dtype 不一致
+        hidden_states = hidden_states.to(torch.bfloat16 if bf16 else self.dtype)
         hidden_states = self.patch_embed(hidden_states, grid_thw)
 
         rotary_pos_emb = self.rot_pos_emb(grid_thw)