Online weight updates from torch.distributed (#2279)

python/sglang/srt/managers/tp_worker.py

@@ -21,7 +21,9 @@ from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.hf_transformers_utils import get_processor, get_tokenizer
 from sglang.srt.managers.io_struct import (
     GetWeightsByNameReqInput,
+    InitWeightsUpdateGroupReqInput,
     UpdateWeightFromDiskReqInput,
+    UpdateWeightsFromDistributedReqInput,
 )
 from sglang.srt.managers.schedule_batch import ModelWorkerBatch, global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
@@ -164,6 +166,25 @@ class TpModelWorker:
         )
         return success, message
 
+    def init_weights_update_group(self, recv_req: InitWeightsUpdateGroupReqInput):
+        success, message = self.model_runner.init_weights_update_group(
+            recv_req.master_address,
+            recv_req.master_port,
+            recv_req.rank_offset,
+            recv_req.world_size,
+            recv_req.group_name,
+            recv_req.backend,
+        )
+        return success, message
+
+    def update_weights_from_distributed(
+        self, recv_req: UpdateWeightsFromDistributedReqInput
+    ):
+        success, message = self.model_runner.update_weights_from_distributed(
+            recv_req.name, recv_req.dtype, recv_req.shape
+        )
+        return success, message
+
     def get_weights_by_name(self, recv_req: GetWeightsByNameReqInput):
         parameter = self.model_runner.get_weights_by_name(
             recv_req.name, recv_req.truncate_size