Fix broken trtllm_mha attn backend with gpt-oss (#9161)

python/sglang/srt/models/gpt_oss.py

@@ -293,8 +293,12 @@ class GptOssAttention(nn.Module):
             prefix=add_prefix("qkv_proj", prefix),
         )
 
+        # Choose dtype of sinks based on attention backend: trtllm_mha requires float32,
+        # others can use bfloat16
+        attn_backend = global_server_args_dict.get("attention_backend")
+        sinks_dtype = torch.float32 if attn_backend == "trtllm_mha" else torch.bfloat16
         self.sinks = nn.Parameter(
-            torch.empty(self.num_heads, dtype=torch.bfloat16), requires_grad=False
+            torch.empty(self.num_heads, dtype=sinks_dtype), requires_grad=False
         )
 
         self.o_proj = RowParallelLinear(