support non disturbing remote instance weight loader v2 (#14997)

Signed-off-by: Anqi Shen <amy.saq@antgroup.com>

docs/advanced_features/rfork.md

@@ -0,0 +1,49 @@
+# R-Fork
+
+R-Fork (Tensor Remote Fork) is a novel weight loading methodology that leverages efficient inter-node GPU-to-GPU data transfer path to load tensors from a running SGLang instance to a new instance with zero-copy. It can significantly optimize the SGLang instance boot-up time by reducing model weights loading from several minutes to mere seconds.
+
+To learn more details about R-Fork, please check **<a href=https://lmsys.org/blog/2025-12-10-rfork/> R-Fork blog </a>**
+
+## Usage
+
+| Argument     | Usage                                      |
+|--------------|--------------------------------------------|
+| load-format  | set to `remote_instance` to enable R-Fork. |
+| remote-instance-weight-loader-backend | `nccl` or `transfer_engine`, default value is `nccl` |
+| remote-instance-weight-loader-seed-instance-ip | IP address of the seed instance who will provide the model weight |
+| remote-instance-weight-loader-seed-instance-service-port | the port that the seed instance's HTTP server is listening on |
+| remote-instance-weight-loader-send-weights-group-ports | the list of available ports on the seed instance that will be used to build NCCL communication groups between seed and client instance. This argument is only needed by `nccl` backend.  |
+| remote-instance-weight-loader-start-seed-via-transfer-engine | set to start seed service that supports TransferEngine as backend. It is needed for seed instances when using `transfer_engine` as backend. |
+
+### NCCL as backend
+
+seed instance:
+```shell
+python -m sglang.launch_server [args]
+```
+
+client instance:
+```shell
+python -m sglang.launch_server [args] \
+  --load-format remote_instance \
+  --remote-instance-weight-loader-seed-instance-ip [seed_instance_ip] \
+  --remote-instance-weight-loader-seed-instance-service-port [seed_instance_service_port] \
+  --remote-instance-weight-loader-send-weights-group-ports [send_weights_nccl_group_ports_list]  \
+  --remote-instance-weight-loader-backend nccl
+```
+
+### TransferEngine as backend
+
+seed instance:
+```shell
+python -m sglang.launch_server [args] \
+  --remote-instance-weight-loader-start-seed-via-transfer-engine
+```
+
+```shell
+python -m sglang.launch_server [args] \
+  --load-format remote_instance \
+  --remote-instance-weight-loader-seed-instance-ip [seed_instance_ip] \
+  --remote-instance-weight-loader-seed-instance-service-port [seed_instance_service_port] \
+  --remote-instance-weight-loader-backend transfer_engine
+```

python/sglang/srt/configs/load_config.py

@@ -2,7 +2,7 @@
 import enum
 import logging
 from dataclasses import dataclass, field
-from typing import List, Optional, Union
+from typing import Any, List, Optional, Union
 
 import orjson
 
@@ -73,6 +73,8 @@ class LoadConfig:
     remote_instance_weight_loader_seed_instance_ip: Optional[str] = None
     remote_instance_weight_loader_seed_instance_service_port: Optional[int] = None
     remote_instance_weight_loader_send_weights_group_ports: Optional[List[int]] = None
+    remote_instance_weight_loader_backend: Optional[str] = None
+    remote_instance_weight_loader_transfer_engine: Optional[Any] = None
 
     # ModelOpt-specific loading options
     modelopt_checkpoint_restore_path: Optional[str] = None

python/sglang/srt/entrypoints/engine.py

@@ -63,6 +63,9 @@ from sglang.srt.managers.multi_tokenizer_mixin import MultiTokenizerRouter
 from sglang.srt.managers.scheduler import run_scheduler_process
 from sglang.srt.managers.template_manager import TemplateManager
 from sglang.srt.managers.tokenizer_manager import TokenizerManager
+from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
+    parse_remote_instance_transfer_engine_info_from_scheduler_infos,
+)
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.tracing.trace import process_tracing_init, trace_set_thread_info
 from sglang.srt.utils import (
@@ -173,10 +176,9 @@ def _launch_subprocesses(
         scheduler_infos.append(data)
 
     # Get back some info from scheduler to tokenizer_manager
-    scheduler_info = scheduler_infos[0]
+    tokenizer_manager.max_req_input_len = scheduler_infos[0]["max_req_input_len"]
-    tokenizer_manager.max_req_input_len = scheduler_info["max_req_input_len"]
 
-    return tokenizer_manager, template_manager, scheduler_info, port_args
+    return tokenizer_manager, template_manager, scheduler_infos, port_args
 
 
 class Engine(EngineBase):
@@ -221,13 +223,21 @@ class Engine(EngineBase):
         atexit.register(self.shutdown)
 
         # Launch subprocesses
-        tokenizer_manager, template_manager, scheduler_info, port_args = (
+        tokenizer_manager, template_manager, scheduler_infos, port_args = (
             self.launch_subprocesses_func(server_args=server_args)
         )
         self.tokenizer_manager = tokenizer_manager
         self.template_manager = template_manager
+
+        scheduler_info = scheduler_infos[0]
         self.scheduler_info = scheduler_info
+
         self.port_args = port_args
+        self.remote_instance_transfer_engine_info = (
+            parse_remote_instance_transfer_engine_info_from_scheduler_infos(
+                scheduler_infos
+            )
+        )
 
         # Initialize ZMQ sockets
         context = zmq.Context(2)

python/sglang/srt/entrypoints/http_server.py

@@ -123,6 +123,9 @@ from sglang.srt.managers.multi_tokenizer_mixin import (
 from sglang.srt.managers.template_manager import TemplateManager
 from sglang.srt.managers.tokenizer_manager import ServerStatus, TokenizerManager
 from sglang.srt.metrics.func_timer import enable_func_timer
+from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
+    parse_remote_instance_transfer_engine_info_from_scheduler_infos,
+)
 from sglang.srt.parser.reasoning_parser import ReasoningParser
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.tracing.trace import process_tracing_init, trace_set_thread_info
@@ -152,6 +155,15 @@ class _GlobalState:
     tokenizer_manager: Union[TokenizerManager, MultiTokenizerRouter, TokenizerWorker]
     template_manager: TemplateManager
     scheduler_info: Dict
+    # Dict{
+    #   rank: Tuple(
+    #           session_id,
+    #           Dict{
+    #               name: Tuple (d_ptr, numel, element_size)
+    #           }
+    #         )
+    # }
+    remote_instance_transfer_engine_info: Optional[Dict] = None
 
 
 _global_state: Optional[_GlobalState] = None
@@ -825,6 +837,30 @@ async def send_weights_to_remote_instance(
         return ORJSONResponse(content, status_code=HTTPStatus.BAD_REQUEST)
 
 
+@app.get("/get_remote_instance_transfer_engine_info")
+async def get_remote_instance_transfer_engine_info(rank: int = None):
+    if rank is None or rank < 0:
+        return Response(status_code=HTTPStatus.BAD_REQUEST)
+
+    if (
+        _global_state.remote_instance_transfer_engine_info is None
+        or len(_global_state.remote_instance_transfer_engine_info) == 0
+    ):
+        return Response(status_code=HTTPStatus.BAD_REQUEST)
+
+    try:
+        result = {
+            "rank": rank,
+            "remote_instance_transfer_engine_info": _global_state.remote_instance_transfer_engine_info[
+                rank
+            ],
+        }
+        return result
+    except Exception as e:
+        logger.error(f"Exception: {e}")
+        return Response(status_code=HTTPStatus.BAD_REQUEST)
+
+
 @app.post("/init_weights_update_group")
 async def init_weights_update_group(
     obj: InitWeightsUpdateGroupReqInput, request: Request
@@ -1615,15 +1651,24 @@ def launch_server(
     1. The HTTP server, Engine, and TokenizerManager all run in the main process.
     2. Inter-process communication is done through IPC (each process uses a different port) via the ZMQ library.
     """
-    tokenizer_manager, template_manager, scheduler_info, port_args = (
+    tokenizer_manager, template_manager, scheduler_infos, port_args = (
         launch_subprocesses_func(server_args=server_args)
     )
 
+    scheduler_info = scheduler_infos[0]
+    remote_instance_transfer_engine_info = None
+    if server_args.remote_instance_weight_loader_use_transfer_engine():
+        remote_instance_transfer_engine_info = (
+            parse_remote_instance_transfer_engine_info_from_scheduler_infos(
+                scheduler_infos
+            )
+        )
     set_global_state(
         _GlobalState(
             tokenizer_manager=tokenizer_manager,
             template_manager=template_manager,
             scheduler_info=scheduler_info,
+            remote_instance_transfer_engine_info=remote_instance_transfer_engine_info,
         )
     )
 

python/sglang/srt/managers/scheduler.py

@@ -2656,6 +2656,9 @@ class Scheduler(
         self.send_to_detokenizer.send_output(recv_req, recv_req)
         return None
 
+    def get_remote_instance_transfer_engine_info(self):
+        return self.tp_worker.get_remote_instance_transfer_engine_info()
+
 
 class IdleSleeper:
     """
@@ -2769,14 +2772,25 @@ def run_scheduler_process(
             pp_rank,
             dp_rank,
         )
-        pipe_writer.send(
+        result_dict = {
-            {
             "status": "ready",
             "max_total_num_tokens": scheduler.max_total_num_tokens,
             "max_req_input_len": scheduler.max_req_input_len,
         }
+        if server_args.remote_instance_weight_loader_use_transfer_engine():
+            (
+                remote_instance_transfer_engine_session_id,
+                remote_instance_transfer_engine_weights_info_dict,
+            ) = scheduler.get_remote_instance_transfer_engine_info()
+            result_dict.update(
+                {
+                    "tp_rank": tp_rank,
+                    "remote_instance_transfer_engine_session_id": remote_instance_transfer_engine_session_id,
+                    "remote_instance_transfer_engine_weights_info_dict": remote_instance_transfer_engine_weights_info_dict,
+                }
             )
 
+        pipe_writer.send(result_dict)
         disaggregation_mode: DisaggregationMode = scheduler.disaggregation_mode
         if disaggregation_mode == DisaggregationMode.NULL:
             if scheduler.enable_pdmux:

python/sglang/srt/managers/tp_worker.py

@@ -366,6 +366,12 @@ class TpModelWorker(BaseTpWorker):
             can_run_cuda_graph=can_run_cuda_graph,
         )
 
+    def get_remote_instance_transfer_engine_info(self):
+        return (
+            self.model_runner.remote_instance_transfer_engine_session_id,
+            self.model_runner.remote_instance_transfer_engine_weight_info,
+        )
+
     def forward_batch_generation(
         self,
         model_worker_batch: ModelWorkerBatch,

python/sglang/srt/model_executor/model_runner.py

@@ -136,9 +136,10 @@ from sglang.srt.model_executor.input_buffers import GraphInputBuffers
 from sglang.srt.model_executor.piecewise_cuda_graph_runner import (
     PiecewiseCudaGraphRunner,
 )
-from sglang.srt.model_loader import get_model
 from sglang.srt.model_loader.loader import DefaultModelLoader, get_model_loader
 from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
+    RemoteInstanceWeightLoaderBackend,
+    register_memory_region,
     trigger_init_weights_send_group_for_remote_instance_request,
 )
 from sglang.srt.model_loader.utils import set_default_torch_dtype
@@ -158,6 +159,7 @@ from sglang.srt.utils import (
     get_available_gpu_memory,
     get_bool_env_var,
     get_cpu_ids_by_node,
+    get_local_ip_auto,
     init_custom_process_group,
     is_cuda,
     is_float4_e2m1fn_x2,
@@ -319,6 +321,10 @@ class ModelRunner:
         self.forward_pass_id = 0
         self.init_new_workspace = False
 
+        self.remote_instance_transfer_engine = None
+        self.remote_instance_transfer_engine_session_id = ""
+        self.remote_instance_transfer_engine_weight_info = None
+
         # Apply the rank zero filter to logger
         if server_args.show_time_cost:
             enable_show_time_cost()
@@ -393,6 +399,9 @@ class ModelRunner:
             enable=self.server_args.enable_memory_saver
         )
 
+        if self.server_args.remote_instance_weight_loader_use_transfer_engine():
+            self.remote_instance_init_transfer_engine()
+
         if not self.is_draft_worker:
             set_global_expert_location_metadata(
                 compute_initial_expert_location_metadata(
@@ -433,6 +442,15 @@ class ModelRunner:
         self.sampler = Sampler()
         self.load_model()
 
+        if (
+            self.server_args.remote_instance_weight_loader_use_transfer_engine()
+            and self.remote_instance_transfer_engine is not None
+            and self.remote_instance_transfer_engine_weight_info is None
+        ):
+            self.remote_instance_transfer_engine_weight_info = register_memory_region(
+                self.model, self.remote_instance_transfer_engine
+            )
+
         # Check if the model is using hybrid SWA
         if (
             not self.server_args.disable_hybrid_swa_memory
@@ -547,6 +565,23 @@ class ModelRunner:
         # Initialize piecewise CUDA graph
         self.init_piecewise_cuda_graphs()
 
+    def remote_instance_init_transfer_engine(self):
+        try:
+            from mooncake.engine import TransferEngine
+        except ImportError as e:
+            logger.warning(
+                "Please install mooncake for using remote instance transfer engine: pip install mooncake"
+            )
+            return
+        self.remote_instance_transfer_engine = TransferEngine()
+        local_ip = get_local_ip_auto()
+        self.remote_instance_transfer_engine.initialize(
+            local_ip, "P2PHANDSHAKE", "rdma", envs.MOONCAKE_DEVICE.value
+        )
+        self.remote_instance_transfer_engine_session_id = (
+            f"{local_ip}:{self.remote_instance_transfer_engine.get_rpc_port()}"
+        )
+
     def model_specific_adjustment(self):
         server_args = self.server_args
 
@@ -764,6 +799,8 @@ class ModelRunner:
             remote_instance_weight_loader_seed_instance_ip=self.server_args.remote_instance_weight_loader_seed_instance_ip,
             remote_instance_weight_loader_seed_instance_service_port=self.server_args.remote_instance_weight_loader_seed_instance_service_port,
             remote_instance_weight_loader_send_weights_group_ports=self.server_args.remote_instance_weight_loader_send_weights_group_ports,
+            remote_instance_weight_loader_backend=self.server_args.remote_instance_weight_loader_backend,
+            remote_instance_weight_loader_transfer_engine=self.remote_instance_transfer_engine,
             modelopt_config=modelopt_config,
             rl_quant_profile=self.server_args.rl_quant_profile,
         )
@@ -772,7 +809,11 @@ class ModelRunner:
                 self.model_config, self.load_config, self.tp_size
             )
 
-        if self.server_args.load_format == LoadFormat.REMOTE_INSTANCE:
+        if (
+            self.server_args.load_format == LoadFormat.REMOTE_INSTANCE
+            and self.server_args.remote_instance_weight_loader_backend
+            == RemoteInstanceWeightLoaderBackend.NCCL
+        ):
             if self.tp_rank == 0:
                 instance_ip = socket.gethostbyname(socket.gethostname())
                 t = threading.Thread(
@@ -797,11 +838,18 @@ class ModelRunner:
             GPU_MEMORY_TYPE_WEIGHTS,
             enable_cpu_backup=enable_cpu_backup,
         ):
-            self.model = get_model(
+            self.loader = get_model_loader(
-                model_config=self.model_config,
                 load_config=self.load_config,
+                model_config=self.model_config,
+            )
+            self.model = self.loader.load_model(
+                model_config=self.model_config,
                 device_config=DeviceConfig(self.device, self.gpu_id),
             )
+            if hasattr(self.loader, "remote_instance_transfer_engine_weight_info"):
+                self.remote_instance_transfer_engine_weight_info = (
+                    self.loader.remote_instance_transfer_engine_weight_info
+                )
         monkey_patch_vllm_parallel_state(reverse=True)
 
         get_offloader().post_init()

python/sglang/srt/model_loader/loader.py

@@ -34,6 +34,11 @@ import huggingface_hub
 import numpy as np
 import torch
 
+from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
+    RemoteInstanceWeightLoaderBackend,
+    get_remote_instance_transfer_engine_info_per_rank,
+    register_memory_region,
+)
 from sglang.srt.server_args import get_global_server_args
 
 # Try to import accelerate (optional dependency)
@@ -1987,6 +1992,7 @@ class RemoteInstanceModelLoader(BaseModelLoader):
                 f"Model loader extra config is not supported for "
                 f"load format {load_config.load_format}"
             )
+        self.remote_instance_transfer_engine_weight_info = None
 
     def download_model(self, model_config: ModelConfig) -> None:
         raise NotImplementedError
@@ -2005,16 +2011,19 @@ class RemoteInstanceModelLoader(BaseModelLoader):
             f"load format {load_config.load_format}"
         )
 
-        model_weights = f"instance://{load_config.remote_instance_weight_loader_seed_instance_ip}:{load_config.remote_instance_weight_loader_send_weights_group_ports[load_config.tp_rank]}"
-
         with set_default_torch_dtype(model_config.dtype):
             with torch.device(device_config.device):
                 model = _initialize_model(model_config, self.load_config)
 
+        if (
+            load_config.remote_instance_weight_loader_backend
+            == RemoteInstanceWeightLoaderBackend.NCCL
+        ):
+            model_weights = f"instance://{load_config.remote_instance_weight_loader_seed_instance_ip}:{load_config.remote_instance_weight_loader_send_weights_group_ports[load_config.tp_rank]}"
             with create_remote_connector(model_weights, device_config.device) as client:
                 connector_type = get_connector_type(client)
                 if connector_type == ConnectorType.INSTANCE:
-                    self.load_model_from_remote_instance(
+                    self.load_model_from_remote_instance_by_nccl(
                         model, client, model_config, device_config
                     )
                 else:
@@ -2022,9 +2031,43 @@ class RemoteInstanceModelLoader(BaseModelLoader):
                         f"Unsupported connector type {connector_type} for "
                         f"remote tensor model loading."
                     )
+        elif (
+            load_config.remote_instance_weight_loader_backend
+            == RemoteInstanceWeightLoaderBackend.TRANSFER_ENGINE
+        ):
+            if load_config.remote_instance_weight_loader_transfer_engine is None:
+                raise RuntimeError(
+                    "Transfer engine is not initialized for remote instance "
+                    "model loader with `transfer_engine` backend. "
+                )
+            logger.info(
+                "TransferEngine registering memory regions (this may take a few seconds)..."
+            )
+            # register memory region
+            self.remote_instance_transfer_engine_weight_info = register_memory_region(
+                model, load_config.remote_instance_weight_loader_transfer_engine
+            )
+            logger.info(
+                "TransferEngine memory regions have been successfully registered."
+            )
+
+            # transfer weights
+            success = self.load_model_from_remote_instance_by_transfer_engine(
+                model,
+                load_config.remote_instance_weight_loader_transfer_engine,
+                f"http://{load_config.remote_instance_weight_loader_seed_instance_ip}:{load_config.remote_instance_weight_loader_seed_instance_service_port}",
+                load_config.tp_rank,
+            )
+            if not success:
+                raise RuntimeError(
+                    "Failed to load weights from remote instance via transfer engine."
+                )
+        else:
+            raise ValueError("Invalid remote instance weight loader backend.")
+
         return model.eval()
 
-    def load_model_from_remote_instance(
+    def load_model_from_remote_instance_by_nccl(
         self, model, client, model_config: ModelConfig, device_config: DeviceConfig
     ) -> nn.Module:
         load_config = self.load_config
@@ -2075,6 +2118,63 @@ class RemoteInstanceModelLoader(BaseModelLoader):
         )
         torch.cuda.empty_cache()
 
+    def load_model_from_remote_instance_by_transfer_engine(
+        self, model, transfer_engine, seed_url, tp_rank
+    ) -> bool:
+        # get remote weights metadata from source instance
+        seed_transfer_engine_session_id, seed_transfer_engine_weight_info = (
+            get_remote_instance_transfer_engine_info_per_rank(seed_url, tp_rank)
+        )
+        if (
+            seed_transfer_engine_session_id is None
+            or seed_transfer_engine_weight_info is None
+        ):
+            logger.error("Cannot get transfer engine session or weight info.")
+            return False
+
+        # prepare local/remote RDMA keys
+        seed_ptr_list = []
+        client_ptr_list = []
+        client_len_list = []
+        for name, tensor in model.named_parameters():
+            weight_info = seed_transfer_engine_weight_info.get(name, None)
+            if weight_info is None:
+                logger.error(f"Cannot find weight info for {name}.")
+                return False
+
+            seed_ptr, seed_numel, seed_element_size = weight_info
+            if (
+                seed_numel != tensor.numel()
+                or seed_element_size != tensor.element_size()
+            ):
+                logger.error(
+                    f"Weight info does not match for {name}, "
+                    f"expected ({seed_numel}, {seed_element_size}), "
+                    f"got ({tensor.numel()}, {tensor.element_size()})"
+                )
+                return False
+            client_ptr = tensor.data_ptr()
+            client_len = tensor.numel() * tensor.element_size()
+            seed_ptr_list.append(seed_ptr)
+            client_ptr_list.append(client_ptr)
+            client_len_list.append(client_len)
+
+        # load weights from source instance through TransferEngine
+        ret = transfer_engine.batch_transfer_sync_read(
+            seed_transfer_engine_session_id,
+            client_ptr_list,
+            seed_ptr_list,
+            client_len_list,
+        )
+        if ret < 0:
+            logger.error(f"batch transfer failed, error: {ret}")
+            return False
+
+        if hasattr(model, "post_load_weights"):
+            model.post_load_weights()
+
+        return True
+
 
 class RemoteModelLoader(BaseModelLoader):
     """Model loader that can load Tensors from remote database."""

python/sglang/srt/model_loader/remote_instance_weight_loader_utils.py

@@ -1,6 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import enum
+import importlib
+import importlib.util
 import logging
+import time
 from typing import List
 
 import requests
@@ -8,6 +12,11 @@ import requests
 logger = logging.getLogger(__name__)
 
 
+class RemoteInstanceWeightLoaderBackend(str, enum.Enum):
+    NCCL = "nccl"
+    TRANSFER_ENGINE = "transfer_engine"
+
+
 def trigger_init_weights_send_group_for_remote_instance_request(
     remote_instance_weight_loader_seed_instance_ip: str,
     remote_instance_weight_loader_seed_instance_service_port: int,
@@ -67,3 +76,133 @@ def trigger_transferring_weights_request(
     except Exception as e:
         logger.error(f"Failed to trigger send weights to remote instance request: {e}")
         raise
+
+
+def get_remote_instance_transfer_engine_info_per_rank(seed_url: str, rank: int):
+    try:
+        response = requests.get(
+            f"{seed_url}/get_remote_instance_transfer_engine_info",
+            params={
+                "rank": rank,
+            },
+        )
+
+        if response.status_code == 200:
+            data = response.json()
+
+            if "remote_instance_transfer_engine_info" in data:
+                return data["remote_instance_transfer_engine_info"]
+            else:
+                logger.error(
+                    "Failed to get `remote_instance_transfer_engine_info` in response."
+                )
+                return None, None
+        else:
+            logger.error(f"request.get failed: {response.status_code}")
+            return None, None
+    except Exception as e:
+        logger.error(f"Exception: {e}")
+        return None, None
+
+
+def parse_remote_instance_transfer_engine_info_from_scheduler_infos(scheduler_infos):
+    remote_instance_transfer_engine_info = {}
+    for data in scheduler_infos:
+        if (
+            "tp_rank" in data
+            and "remote_instance_transfer_engine_session_id" in data
+            and "remote_instance_transfer_engine_weights_info_dict" in data
+        ):
+            remote_instance_transfer_engine_info[data["tp_rank"]] = (
+                data["remote_instance_transfer_engine_session_id"],
+                data["remote_instance_transfer_engine_weights_info_dict"],
+            )
+    return remote_instance_transfer_engine_info
+
+
+def register_memory_region(model, transfer_engine):
+    if importlib.util.find_spec("torch") is None:
+        return register_memory_region_v1(model, transfer_engine)
+    else:
+        return register_memory_region_v2(model, transfer_engine)
+
+
+def register_memory_region_v1(model, transfer_engine):
+    start_tic = time.time()
+
+    weight_mr_dict = {}
+    for name, weight in model.named_parameters():
+        ret = transfer_engine.register_memory(
+            weight.data_ptr(), weight.numel() * weight.element_size()
+        )
+        if ret != 0:
+            raise RuntimeError(
+                f"register memory failed for weight {name}, error: {ret}"
+            )
+        weight_mr_dict[name] = (
+            weight.data_ptr(),
+            weight.numel(),
+            weight.element_size(),
+        )
+
+    end_tic = time.time()
+    logger.debug(f"Register memory region time: {(end_tic - start_tic):.4f}s")
+    return weight_mr_dict
+
+
+def register_memory_region_v2(model, transfer_engine):
+    start_tic = time.time()
+
+    weight_mr_dict = {}
+    weight_addr_set = set()
+    for name, weight in model.named_parameters():
+        weight_mr_dict[name] = (
+            weight.data_ptr(),
+            weight.numel(),
+            weight.element_size(),
+        )
+        weight_addr_set.add(weight.data_ptr())
+
+    import torch
+
+    memory_snapshot = torch.cuda.memory.memory_snapshot()
+    weight_blocks_for_reg_mr = []
+    # Blocks in each segment have continuous physical addresses,
+    # so they can be merged for memory registration.
+    for segment in memory_snapshot:
+        current_weight_block = None
+        blocks = segment.get("blocks", [])
+        for block in blocks:
+            address = block.get("address", -1)
+            size = block.get("size", -1)
+            state = block.get("state", "")
+            if address < 0 or size < 0 or state == "":
+                continue
+            # Only register active allocated memory blocks that hold weights.
+            if state == "active_allocated":
+                if address in weight_addr_set:
+                    if current_weight_block is None:
+                        current_weight_block = (address, size)
+                    elif current_weight_block[0] + current_weight_block[1] == address:
+                        current_weight_block = (
+                            current_weight_block[0],
+                            current_weight_block[1] + size,
+                        )
+                    else:
+                        weight_blocks_for_reg_mr.append(current_weight_block)
+                        current_weight_block = (address, size)
+        if current_weight_block is not None:
+            weight_blocks_for_reg_mr.append(current_weight_block)
+
+    # Register merged memory blocks that hold weights.
+    for weight_block in weight_blocks_for_reg_mr:
+        address, size = weight_block
+        ret = transfer_engine.register_memory(address, size)
+        if ret != 0:
+            raise RuntimeError(
+                f"register memory failed for weight block at address {address} with size {size}, error: {ret}"
+            )
+
+    end_tic = time.time()
+    logger.debug(f"Register memory region v2 time: {(end_tic - start_tic):.4f}s")
+    return weight_mr_dict

python/sglang/srt/server_args.py

@@ -18,6 +18,7 @@ from __future__ import annotations
 import argparse
 import dataclasses
 import importlib
+import importlib.util
 import json
 import logging
 import os
@@ -614,6 +615,8 @@ class ServerArgs:
     remote_instance_weight_loader_seed_instance_ip: Optional[str] = None
     remote_instance_weight_loader_seed_instance_service_port: Optional[int] = None
     remote_instance_weight_loader_send_weights_group_ports: Optional[List[int]] = None
+    remote_instance_weight_loader_backend: Literal["transfer_engine", "nccl"] = "nccl"
+    remote_instance_weight_loader_start_seed_via_transfer_engine: bool = False
 
     # For PD-Multiplexing
     enable_pdmux: bool = False
@@ -690,6 +693,9 @@ class ServerArgs:
         # Handle speculative decoding logic.
         self._handle_speculative_decoding()
 
+        # Handle remote instance weight loader.
+        self._handle_remote_instance_weight_loader_start_seed_via_transfer_engine()
+
         # Handle model loading format.
         self._handle_load_format()
 
@@ -2107,8 +2113,26 @@ class ServerArgs:
             if (
                 self.remote_instance_weight_loader_seed_instance_ip is None
                 or self.remote_instance_weight_loader_seed_instance_service_port is None
-                or self.remote_instance_weight_loader_send_weights_group_ports is None
             ):
+                logger.warning(
+                    "Fallback load_format to 'auto' due to incomplete remote instance weight loader settings."
+                )
+                self.load_format = "auto"
+            elif (
+                self.remote_instance_weight_loader_send_weights_group_ports is None
+                and self.remote_instance_weight_loader_backend == "nccl"
+            ):
+                logger.warning(
+                    "Fallback load_format to 'auto' due to incomplete remote instance weight loader NCCL group ports settings."
+                )
+                self.load_format = "auto"
+            elif (
+                not self.validate_transfer_engine()
+                and self.remote_instance_weight_loader_backend == "transfer_engine"
+            ):
+                logger.warning(
+                    "Fallback load_format to 'auto' due to 'transfer_engine' backend is not supported."
+                )
                 self.load_format = "auto"
 
     def _handle_encoder_disaggregation(self):
@@ -2366,19 +2390,12 @@ class ServerArgs:
             self.disable_cuda_graph = True
             self.skip_server_warmup = True
 
-    def _handle_remote_instance_weight_loader_support_transfer_engine(self):
+    def _handle_remote_instance_weight_loader_start_seed_via_transfer_engine(self):
-        if importlib.util.find_spec("mooncake.engine") is None:
+        # Check whether TransferEngine can be used when users want to start seed service that supports TransferEngine backend.
-            logger.warning(
+        if self.remote_instance_weight_loader_start_seed_via_transfer_engine:
-                f"Failed to import mooncake.engine. Does not support using TransferEngine as remote instance weight loader backend."
+            self.remote_instance_weight_loader_start_seed_via_transfer_engine = (
+                self.validate_transfer_engine()
             )
-            self.remote_instance_weight_loader_support_transfer_engine = False
-        elif self.enable_memory_saver:
-            logger.warning(
-                "Memory saver is enabled, which is not compatible with TransferEngine. Does not support using TransferEngine as remote instance weight loader backend."
-            )
-            self.remote_instance_weight_loader_support_transfer_engine = False
-        else:
-            self.remote_instance_weight_loader_support_transfer_engine = True
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -4277,6 +4294,18 @@ class ServerArgs:
             default=ServerArgs.remote_instance_weight_loader_send_weights_group_ports,
             help="The communication group ports for loading weights from remote instance.",
         )
+        parser.add_argument(
+            "--remote-instance-weight-loader-backend",
+            type=str,
+            choices=["transfer_engine", "nccl"],
+            default=ServerArgs.remote_instance_weight_loader_backend,
+            help="The backend for loading weights from remote instance. Can be 'transfer_engine' or 'nccl'. Default is 'nccl'.",
+        )
+        parser.add_argument(
+            "--remote-instance-weight-loader-start-seed-via-transfer-engine",
+            action="store_true",
+            help="Start seed server via transfer engine backend for remote instance weight loader.",
+        )
 
         # For PD-Multiplexing
         parser.add_argument(
@@ -4782,6 +4811,33 @@ class ServerArgs:
             original_server_arg_mem_fraction * final_overall_factor
         )
 
+    def validate_transfer_engine(self):
+        if importlib.util.find_spec("mooncake.engine") is None:
+            logger.warning(
+                f"Failed to import mooncake.engine. Does not support using TransferEngine as remote instance weight loader backend."
+            )
+            return False
+        elif self.enable_memory_saver:
+            logger.warning(
+                "Memory saver is enabled, which is not compatible with TransferEngine. Does not support using TransferEngine as remote instance weight loader backend."
+            )
+            return False
+        else:
+            return True
+
+    def remote_instance_weight_loader_use_transfer_engine(self):
+        # Use TransferEngine as seed backend.
+        if self.remote_instance_weight_loader_start_seed_via_transfer_engine:
+            return True
+        # Use TransferEngine as client backend.
+        elif (
+            self.load_format == "remote_instance"
+            and self.remote_instance_weight_loader_backend == "transfer_engine"
+        ):
+            return True
+        else:
+            return False
+
 
 # NOTE: This is a global variable to hold the server args for scheduler.
 _global_server_args: Optional[ServerArgs] = None

test/srt/test_load_weights_from_remote_instance.py

@@ -62,6 +62,7 @@ def init_process(
     seed_instance_group_base_port,
     event_seed_ready,
     event_dst_ready_list,
+    remote_instance_loader_backend,
 ):
     torch.cuda.set_device(rank)
 
@@ -90,6 +91,7 @@ def init_process(
             tp_size,
             event_seed_ready,
             event_dst_ready_list,
+            remote_instance_loader_backend,
         )
 
 
@@ -122,6 +124,7 @@ def init_process_seed(
             str(rank),
             "--tp-size",
             str(tp_size),
+            "--remote-instance-weight-loader-start-seed-via-transfer-engine",
         ),
     )
     torch.cuda.synchronize()
@@ -159,6 +162,7 @@ def init_process_dst(
     tp_size,
     event_seed_ready,
     event_dst_ready_list,
+    remote_instance_loader_backend,
 ):
     torch.cuda.set_device(rank * tp_size)
     torch.cuda.synchronize()
@@ -186,6 +190,10 @@ def init_process_dst(
             remote_instance_weight_loader_seed_instance_service_port=seed_instance_service_port,
             remote_instance_weight_loader_send_weights_group_ports=ports,
             load_format="remote_instance",
+            remote_instance_weight_loader_backend=remote_instance_loader_backend,
+            remote_instance_weight_loader_start_seed_via_transfer_engine=(
+                remote_instance_loader_backend == "transfer_engine"
+            ),
         )
     else:
         host, _, port = DEFAULT_URL_FOR_TEST.rpartition(":")
@@ -213,6 +221,9 @@ def init_process_dst(
                 f"[{','.join(str(port) for port in ports)}]",
                 "--load-format",
                 "remote_instance",
+                "--remote-instance-weight-loader-backend",
+                remote_instance_loader_backend,
+                "--remote-instance-weight-loader-start-seed-via-transfer-engine",
             ),
         )
     torch.cuda.synchronize()
@@ -250,9 +261,10 @@ def test_load_weights_from_remote_instance(
     seed_instance_ip,
     seed_instance_service_port,
     seed_instance_group_base_port,
+    remote_instance_loader_backend,
 ):
     print(
-        f"Testing model: {model_name} tp_size: {tp_size}, dp_size: {dp_size} backend: {backends}"
+        f"Testing model: {model_name} tp_size: {tp_size}, dp_size: {dp_size} backend: {backends} remote_instance_loader_backend: {remote_instance_loader_backend}"
     )
     param_queue = mp.Queue()
     results = {}
@@ -276,6 +288,7 @@ def test_load_weights_from_remote_instance(
             seed_instance_group_base_port,
             event_seed_ready,
             event_dst_ready_list,
+            remote_instance_loader_backend,
         ),
         nprocs=1 + dp_size,
         join=False,
@@ -340,14 +353,42 @@ class TestLoadWeightsFromRemoteInstance(CustomTestCase):
         # test_suits : tp, dp, model_name, backend, dst_instance_id
         if is_in_ci():
             mode = random.choice(["Engine", "Server"])
+            remote_instance_loader_backend = random.choice(["nccl", "transfer_engine"])
             test_suits = [
-                (1, 1, DEFAULT_SMALL_MODEL_NAME_FOR_TEST, [mode]),
+                (
+                    1,
+                    1,
+                    DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
+                    [mode],
+                    remote_instance_loader_backend,
+                ),
             ]
         else:
             test_suits = [
-                (1, 1, DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["Engine"]),
+                (1, 1, DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["Engine"], "nccl"),
-                (1, 1, DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["Sever"]),
+                (1, 1, DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["Server"], "nccl"),
-                (2, 2, DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["Engine", "Server"]),
+                (2, 2, DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["Engine", "Server"], "nccl"),
+                (
+                    1,
+                    1,
+                    DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
+                    ["Engine"],
+                    "transfer_engine",
+                ),
+                (
+                    1,
+                    1,
+                    DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
+                    ["Server"],
+                    "transfer_engine",
+                ),
+                (
+                    2,
+                    2,
+                    DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
+                    ["Engine", "Server"],
+                    "transfer_engine",
+                ),
             ]
 
         truncate_size = 10
@@ -365,7 +406,13 @@ class TestLoadWeightsFromRemoteInstance(CustomTestCase):
             "model.norm.weight",
         ]
 
-        for tp_size, dp_size, model_name, backends in test_suits:
+        for (
+            tp_size,
+            dp_size,
+            model_name,
+            backends,
+            remote_instance_loader_backend,
+        ) in test_suits:
             test_load_weights_from_remote_instance(
                 tp_size,
                 dp_size,
@@ -376,6 +423,7 @@ class TestLoadWeightsFromRemoteInstance(CustomTestCase):
                 "127.0.0.1",
                 DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 1000,
                 60000,
+                remote_instance_loader_backend,
             )