Support DP attention with GPT-OSS (#9359)

python/sglang/srt/models/gpt_oss.py

@@ -1091,7 +1091,7 @@ class GptOssForCausalLM(nn.Module):
                     if name in params_dict.keys():
                         param = params_dict[name]
                         if "sinks" in name:
-                            start = tp_rank * param.numel()
+                            start = get_attention_tp_rank() * param.numel()
                             param.data.copy_(
                                 loaded_weight[start : start + param.numel()]
                             )

python/sglang/srt/server_args.py

@@ -2183,6 +2183,7 @@ class ServerArgs:
             ), f"GptOssForCausalLM requires one of {supported_backends} attention backend, but got '{self.attention_backend}'"
 
             if is_sm100_supported():
+                if not self.enable_dp_attention:
                     self.enable_flashinfer_allreduce_fusion = True
                     logger.info(
                         "Enable FlashInfer AllReduce Fusion on sm100 for GptOssForCausalLM"