sglang 0.5.2rc1__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/managers/scheduler.py

@@ -79,13 +79,17 @@ from sglang.srt.managers.io_struct import (
     FreezeGCReq,
     GetInternalStateReq,
     GetInternalStateReqOutput,
+    GetLoadReqInput,
+    GetLoadReqOutput,
     GetWeightsByNameReqInput,
     HealthCheckOutput,
+    InitWeightsSendGroupForRemoteInstanceReqInput,
+    InitWeightsSendGroupForRemoteInstanceReqOutput,
     InitWeightsUpdateGroupReqInput,
     LoadLoRAAdapterReqInput,
     LoadLoRAAdapterReqOutput,
     MultiTokenizerRegisterReq,
-    MultiTokenizerWarpper,
+    MultiTokenizerWrapper,
     OpenSessionReqInput,
     OpenSessionReqOutput,
     ProfileReq,
@@ -93,6 +97,8 @@ from sglang.srt.managers.io_struct import (
     ResumeMemoryOccupationReqInput,
     RpcReqInput,
     RpcReqOutput,
+    SendWeightsToRemoteInstanceReqInput,
+    SendWeightsToRemoteInstanceReqOutput,
     SetInternalStateReq,
     SetInternalStateReqOutput,
     SlowDownReqInput,
@@ -141,10 +147,19 @@ from sglang.srt.mem_cache.lora_radix_cache import LoRARadixCache
 from sglang.srt.mem_cache.radix_cache import RadixCache
 from sglang.srt.mem_cache.swa_radix_cache import SWARadixCache
 from sglang.srt.model_executor.forward_batch_info import ForwardMode, PPProxyTensors
-from sglang.srt.reasoning_parser import ReasoningParser
+from sglang.srt.parser.reasoning_parser import ReasoningParser
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter
+from sglang.srt.tracing.trace import (
+    process_tracing_init,
+    trace_event,
+    trace_set_proc_propagate_context,
+    trace_set_thread_info,
+    trace_slice,
+    trace_slice_end,
+    trace_slice_start,
+)
 from sglang.srt.two_batch_overlap import TboDPAttentionPreparer
 from sglang.srt.utils import (
     DynamicGradMode,
@@ -158,6 +173,7 @@ from sglang.srt.utils import (
     get_zmq_socket,
     is_cpu,
     kill_itself_when_parent_died,
+    numa_bind_to_node,
     point_to_point_pyobj,
     pyspy_dump_schedulers,
     require_mlp_sync,
@@ -348,6 +364,18 @@ class Scheduler(
                 target_worker=self.tp_worker,
                 dp_rank=dp_rank,
             )
+        elif self.spec_algorithm.is_standalone():
+            from sglang.srt.speculative.standalone_worker import StandaloneWorker
+
+            self.draft_worker = StandaloneWorker(
+                gpu_id=gpu_id,
+                tp_rank=tp_rank,
+                moe_ep_rank=moe_ep_rank,
+                server_args=server_args,
+                nccl_port=port_args.nccl_port,
+                target_worker=self.tp_worker,
+                dp_rank=dp_rank,
+            )
         else:
             self.draft_worker = None
 
@@ -401,7 +429,7 @@ class Scheduler(
                 f"max_prefill_tokens={self.max_prefill_tokens}, "
                 f"max_running_requests={self.max_running_requests}, "
                 f"context_len={self.model_config.context_len}, "
-                f"available_gpu_mem={avail_mem:.2f} GB"
+                f"{'available_cpu_mem' if self.device == 'cpu' else 'available_gpu_mem'}={avail_mem:.2f} GB"
             )
 
         # Init memory pool and cache
@@ -488,7 +516,7 @@ class Scheduler(
             enable=server_args.enable_memory_saver
         )
         self.offload_tags = set()
-        self.init_profier()
+        self.init_profiler()
 
         self.recv_skipper = SchedulerRecvSkipper.maybe_create(server_args)
         self.input_blocker = (
@@ -500,6 +528,7 @@ class Scheduler(
         # Init metrics stats
         self.init_metrics(tp_rank, pp_rank, dp_rank)
         self.init_kv_events(server_args.kv_events_config)
+        self.init_dp_balance(dp_balance_meta)
 
         # Init disaggregation
         self.disaggregation_mode = DisaggregationMode(
@@ -524,6 +553,14 @@ class Scheduler(
                 (CloseSessionReqInput, self.close_session),
                 (UpdateWeightFromDiskReqInput, self.update_weights_from_disk),
                 (InitWeightsUpdateGroupReqInput, self.init_weights_update_group),
+                (
+                    InitWeightsSendGroupForRemoteInstanceReqInput,
+                    self.init_weights_send_group_for_remote_instance,
+                ),
+                (
+                    SendWeightsToRemoteInstanceReqInput,
+                    self.send_weights_to_remote_instance,
+                ),
                 (
                     UpdateWeightsFromDistributedReqInput,
                     self.update_weights_from_distributed,
@@ -542,18 +579,10 @@ class Scheduler(
                 (LoadLoRAAdapterReqInput, self.load_lora_adapter),
                 (UnloadLoRAAdapterReqInput, self.unload_lora_adapter),
                 (MultiTokenizerRegisterReq, self.register_multi_tokenizer),
+                (GetLoadReqInput, self.get_load),
             ]
         )
 
-        self.balance_meta = dp_balance_meta
-        if (
-            server_args.enable_dp_attention
-            and server_args.load_balance_method == "minimum_tokens"
-        ):
-            assert dp_balance_meta is not None
-
-        self.recv_dp_balance_id_this_term = []
-
     def init_tokenizer(self):
         server_args = self.server_args
         self.is_generation = self.model_config.is_generation
@@ -630,6 +659,7 @@ class Scheduler(
                     hicache_write_policy=server_args.hicache_write_policy,
                     hicache_io_backend=server_args.hicache_io_backend,
                     hicache_mem_layout=server_args.hicache_mem_layout,
+                    enable_metrics=self.enable_metrics,
                     hicache_storage_backend=server_args.hicache_storage_backend,
                     hicache_storage_prefetch_policy=server_args.hicache_storage_prefetch_policy,
                     model_name=server_args.served_model_name,
@@ -662,6 +692,21 @@ class Scheduler(
                     page_size=self.page_size,
                     disable=server_args.disable_radix_cache,
                 )
+            elif server_args.enable_lmcache:
+                from sglang.srt.mem_cache.storage.lmcache.lmc_radix_cache import (
+                    LMCRadixCache,
+                )
+
+                self.tree_cache = LMCRadixCache(
+                    req_to_token_pool=self.req_to_token_pool,
+                    token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
+                    page_size=self.page_size,
+                    disable=server_args.disable_radix_cache,
+                    model_config=self.model_config,
+                    tp_size=self.tp_size,
+                    rank=self.tp_rank,
+                    tp_group=self.tp_group,
+                )
             else:
                 self.tree_cache = RadixCache(
                     req_to_token_pool=self.req_to_token_pool,
@@ -793,6 +838,10 @@ class Scheduler(
             batch = self.get_next_batch_to_run()
             self.cur_batch = batch
 
+            if batch:
+                for req in batch.reqs:
+                    trace_event("schedule", req.rid)
+
             if batch:
                 result = self.run_batch(batch)
                 self.process_batch_result(batch, result)
@@ -814,6 +863,10 @@ class Scheduler(
             batch = self.get_next_batch_to_run()
             self.cur_batch = batch
 
+            if batch:
+                for req in batch.reqs:
+                    trace_event("schedule", req.rid)
+
             if batch:
                 batch.launch_done = threading.Event()
                 result = self.run_batch(batch)
@@ -1077,6 +1130,12 @@ class Scheduler(
                 self.tp_cpu_group,
                 src=self.tp_group.ranks[0],
             )
+
+        for req in recv_reqs:
+            if isinstance(req, (TokenizedGenerateReqInput, TokenizedEmbeddingReqInput)):
+                trace_set_proc_propagate_context(req.rid, req.trace_context)
+                trace_slice_start("", req.rid, anonymous=True)
+
         return recv_reqs
 
     def process_input_requests(self, recv_reqs: List):
@@ -1104,13 +1163,13 @@ class Scheduler(
                     self.send_to_tokenizer.send_pyobj(abort_req)
                     continue
 
-            # If it is a MultiTokenizerWarpper, unwrap it and handle the inner request.
-            if isinstance(recv_req, MultiTokenizerWarpper):
+            # If it is a MultiTokenizerWrapper, unwrap it and handle the inner request.
+            if isinstance(recv_req, MultiTokenizerWrapper):
                 worker_id = recv_req.worker_id
                 recv_req = recv_req.obj
                 output = self._request_dispatcher(recv_req)
                 if output is not None:
-                    output = MultiTokenizerWarpper(worker_id, output)
+                    output = MultiTokenizerWrapper(worker_id, output)
                     self.send_to_tokenizer.send_pyobj(output)
                 continue
 
@@ -1122,15 +1181,21 @@ class Scheduler(
                 else:
                     self.send_to_tokenizer.send_pyobj(output)
 
+    def init_req_max_new_tokens(self, req):
+        req.sampling_params.max_new_tokens = min(
+            (
+                req.sampling_params.max_new_tokens
+                if req.sampling_params.max_new_tokens is not None
+                else 1 << 30
+            ),
+            self.max_req_len - len(req.origin_input_ids) - 1,
+        )
+
     def handle_generate_request(
         self,
         recv_req: TokenizedGenerateReqInput,
     ):
-        if (
-            self.server_args.enable_dp_attention
-            and self.server_args.load_balance_method == "minimum_tokens"
-        ):
-            self.recv_dp_balance_id_this_term.append(recv_req.dp_balance_id)
+        self.maybe_update_dp_balance_data(recv_req)
 
         # Create a new request
         if (
@@ -1189,6 +1254,7 @@ class Scheduler(
                 req.set_finish_with_abort(
                     f"Invalid request: session id {recv_req.session_params.id} does not exist"
                 )
+                self.init_req_max_new_tokens(req)
                 self._add_request_to_queue(req)
                 return
         else:
@@ -1196,6 +1262,7 @@ class Scheduler(
             session = self.sessions[recv_req.session_params.id]
             req = session.create_req(recv_req, self.tokenizer)
             if isinstance(req.finished_reason, FINISH_ABORT):
+                self.init_req_max_new_tokens(req)
                 self._add_request_to_queue(req)
                 return
 
@@ -1215,9 +1282,13 @@ class Scheduler(
                         f"After expanding {len(req.origin_input_ids_unpadded)=} => {len(req.origin_input_ids)} >= {self.max_req_input_len}."
                     )
                 )
+                self.init_req_max_new_tokens(req)
                 self._add_request_to_queue(req)
                 return
 
+        # initialize before returning
+        self.init_req_max_new_tokens(req)
+
         # Validate prompt length
         error_msg = validate_input_length(
             req,
@@ -1232,26 +1303,25 @@ class Scheduler(
         # Copy more attributes
         if recv_req.logprob_start_len == -1 or not recv_req.return_logprob:
             # By default, only return the logprobs for output tokens
-            req.logprob_start_len = len(req.origin_input_ids) - 1
+            # For prefill-only requests with logprob_start_len == -1, set logprob_start_len beyond input sequence
+            # to skip input logprob computation entirely
+            if req.is_prefill_only:
+                req.logprob_start_len = len(req.origin_input_ids)
+            else:
+                # TODO: For text generation, evaluate setting logprob_start_len to len(req.origin_input_ids) as well
+                req.logprob_start_len = len(req.origin_input_ids) - 1
         else:
             req.logprob_start_len = recv_req.logprob_start_len
 
-        if req.logprob_start_len >= len(req.origin_input_ids):
+        if not req.is_prefill_only and req.logprob_start_len >= len(
+            req.origin_input_ids
+        ):
             error_msg = f"{req.logprob_start_len=} is higher than the number of input tokens {len(req.origin_input_ids)=}. Please use a smaller logprob_start_len."
             req.logprob_start_len = len(req.origin_input_ids) - 1
             req.set_finish_with_abort(error_msg)
             self._add_request_to_queue(req)
             return
 
-        req.sampling_params.max_new_tokens = min(
-            (
-                req.sampling_params.max_new_tokens
-                if req.sampling_params.max_new_tokens is not None
-                else 1 << 30
-            ),
-            self.max_req_len - len(req.origin_input_ids) - 1,
-        )
-
         # Init grammar cache for this request
         add_to_grammar_queue = False
         if (
@@ -1310,6 +1380,7 @@ class Scheduler(
         else:
             self._prefetch_kvcache(req)
             self.waiting_queue.append(req)
+            trace_slice_end("process req", req.rid, auto_next_anon=True)
 
     def _prefetch_kvcache(self, req: Req):
         if self.enable_hicache_storage:
@@ -1421,9 +1492,11 @@ class Scheduler(
             _, _, available_size, evictable_size = self._get_token_info()
             protected_size = self.tree_cache.protected_size()
             memory_leak = (available_size + evictable_size) != (
+                # self.max_total_num_tokens
+                # if not self.enable_hierarchical_cache
+                # else self.max_total_num_tokens - protected_size
                 self.max_total_num_tokens
-                if not self.enable_hierarchical_cache
-                else self.max_total_num_tokens - protected_size
+                - protected_size
             )
             token_msg = f"{self.max_total_num_tokens=}, {available_size=}, {evictable_size=}, {protected_size=}\n"
 
@@ -1474,6 +1547,20 @@ class Scheduler(
             self.stats.gen_throughput = 0
             self.stats.num_queue_reqs = len(self.waiting_queue)
             self.stats.num_grammar_queue_reqs = len(self.grammar_queue)
+            if self.disaggregation_mode == DisaggregationMode.PREFILL:
+                self.stats.num_prefill_prealloc_queue_reqs = len(
+                    self.disagg_prefill_bootstrap_queue.queue
+                )
+                self.stats.num_prefill_inflight_queue_reqs = len(
+                    self.disagg_prefill_inflight_queue
+                )
+            if self.disaggregation_mode == DisaggregationMode.DECODE:
+                self.stats.num_decode_prealloc_queue_reqs = len(
+                    self.disagg_decode_prealloc_queue.queue
+                )
+                self.stats.num_decode_transfer_queue_reqs = len(
+                    self.disagg_decode_transfer_queue.queue
+                )
             self.metrics_collector.log_stats(self.stats)
         self._publish_kv_events()
 
@@ -1521,7 +1608,12 @@ class Scheduler(
             chunked_req_to_exclude.add(self.chunked_req)
             self.tree_cache.cache_unfinished_req(self.chunked_req, chunked=True)
             # chunked request keeps its rid but will get a new req_pool_idx
-            self.req_to_token_pool.free(self.chunked_req.req_pool_idx)
+            if self.tp_worker.worker.model_runner.is_hybrid_gdn:
+                self.req_to_token_pool.free(
+                    self.chunked_req.req_pool_idx, free_mamba_cache=False
+                )
+            else:
+                self.req_to_token_pool.free(self.chunked_req.req_pool_idx)
         if self.last_batch and self.last_batch.forward_mode.is_extend():
             if self.last_batch.chunked_req is not None:
                 # In the context pipeline parallelism, after the last chunk, the current microbatch still track outdated chunked_req.
@@ -1568,11 +1660,7 @@ class Scheduler(
 
         # Handle DP attention
         if need_dp_attn_preparation:
-            if (
-                self.server_args.load_balance_method == "minimum_tokens"
-                and self.forward_ct % 40 == 0
-            ):
-                self.handle_dp_balance_data(ret)
+            self.maybe_handle_dp_balance_data()
             ret = self.prepare_mlp_sync_batch(ret)
 
         return ret
@@ -1792,10 +1880,6 @@ class Scheduler(
             if self.spec_algorithm.is_none():
                 model_worker_batch = batch.get_model_worker_batch()
 
-                # update the consumer index of hicache to the running batch
-                self.tp_worker.set_hicache_consumer(
-                    model_worker_batch.hicache_consumer_index
-                )
                 if self.pp_group.is_last_rank:
                     logits_output, next_token_ids, can_run_cuda_graph = (
                         self.tp_worker.forward_batch_generation(model_worker_batch)
@@ -1864,8 +1948,23 @@ class Scheduler(
     ):
         if batch.forward_mode.is_decode():
             self.process_batch_result_decode(batch, result, launch_done)
+            for req in batch.reqs:
+                trace_slice(
+                    "decode loop",
+                    req.rid,
+                    auto_next_anon=not req.finished(),
+                    thread_finish_flag=req.finished(),
+                )
+
         elif batch.forward_mode.is_extend():
             self.process_batch_result_prefill(batch, result, launch_done)
+            for req in batch.reqs:
+                trace_slice(
+                    "prefill",
+                    req.rid,
+                    auto_next_anon=not req.finished(),
+                    thread_finish_flag=req.finished(),
+                )
         elif batch.forward_mode.is_idle():
             if self.enable_overlap:
                 self.tp_worker.resolve_last_batch_result(launch_done)
@@ -1897,86 +1996,6 @@ class Scheduler(
             disable_overlap_schedule=self.server_args.disable_overlap_schedule,
         )
 
-    def handle_dp_balance_data(self, local_batch: ScheduleBatch):
-        def gather_dp_balance_info(holding_tokens_list) -> Union[None, List[List[int]]]:
-            """gather recv_dp_balance_id_this_term and holding tokens per worker for dp balance"""
-            recv_list = self.recv_dp_balance_id_this_term
-            assert len(recv_list) <= 511, (
-                "The number of requests received this round is too large. "
-                "Please increase gather_tensor_size and onfly_info_size."
-            )
-            # The maximum size of the tensor used for gathering data from all workers.
-            gather_tensor_size = 512
-
-            # recv_tensor: | holding_tokens | len(recv_dp_balance_id) | recv_dp_balance_ids
-            recv_tensor = torch.zeros(gather_tensor_size, dtype=torch.int32)
-            recv_tensor[0] = holding_tokens_list
-            recv_tensor[1] = len(
-                recv_list
-            )  # The first element is the length of the list.
-            recv_tensor[2 : len(recv_list) + 2] = torch.tensor(
-                recv_list, dtype=torch.int32
-            )
-
-            if self.tp_rank == 0:
-                gathered_list = [
-                    torch.zeros(gather_tensor_size, dtype=torch.int32)
-                    for _ in range(self.balance_meta.num_workers)
-                ]
-            else:
-                gathered_list = None
-
-            torch.distributed.gather(
-                recv_tensor, gathered_list, group=self.tp_cpu_group
-            )
-
-            gathered_id_list_per_worker = None
-            if self.tp_rank == 0:
-                gathered_id_list_per_worker = []
-                holding_tokens_list = []
-                for tensor in gathered_list:
-                    holding_tokens_list.append(tensor[0].item())
-                    list_length = tensor[1].item()
-                    gathered_id_list_per_worker.append(
-                        tensor[2 : list_length + 2].tolist()
-                    )
-
-            return gathered_id_list_per_worker, holding_tokens_list
-
-        def write_shared_dp_balance_info(new_recv_rid_lists, local_tokens):
-            meta = self.balance_meta
-
-            with meta.mutex:
-                onfly_list: List[Dict[int, int]] = meta.get_shared_onfly()
-                assert len(new_recv_rid_lists) == len(
-                    onfly_list
-                ), "num_worker not equal"
-                # 1.Check if the rid received by each worker this round is present in onfly.
-                #   If it is, remove the corresponding onfly item.
-                worker_id = 0
-                for new_recv_rids, on_fly_reqs in zip(new_recv_rid_lists, onfly_list):
-                    for new_recv_rid in new_recv_rids:
-                        assert (
-                            new_recv_rid in on_fly_reqs
-                        ), f"{new_recv_rid=} not in {worker_id=} {on_fly_reqs=}, data consistency is wrong"
-                        del on_fly_reqs[new_recv_rid]
-                    worker_id += 1
-                # 2. Atomically write local_tokens and onfly into shm under the mutex
-                meta.set_shared_onfly_info(onfly_list)
-                meta.set_shared_local_tokens(local_tokens)
-
-        holding_tokens = self.get_load()
-
-        new_recv_dp_balance_id_list, holding_token_list = gather_dp_balance_info(
-            holding_tokens
-        )
-
-        self.recv_dp_balance_id_this_term.clear()
-        if self.tp_rank == 0:  # only first worker write info
-            write_shared_dp_balance_info(
-                new_recv_dp_balance_id_list, holding_token_list
-            )
-
     @staticmethod
     def prepare_mlp_sync_batch_raw(
         local_batch: ScheduleBatch,
@@ -2270,39 +2289,50 @@ class Scheduler(
             if_success = False
         return if_success
 
-    def get_load(self):
+    def get_load(self, recv_req: GetLoadReqInput = None) -> GetLoadReqOutput:
         # TODO(lsyin): use dynamically maintained num_waiting_tokens
+
         if self.is_hybrid:
-            load_full = (
+            num_tokens_full = (
                 self.full_tokens_per_layer
                 - self.token_to_kv_pool_allocator.full_available_size()
                 - self.tree_cache.full_evictable_size()
             )
-            load_swa = (
+            num_tokens_swa = (
                 self.swa_tokens_per_layer
                 - self.token_to_kv_pool_allocator.swa_available_size()
                 - self.tree_cache.swa_evictable_size()
             )
-            load = max(load_full, load_swa)
+            num_tokens = max(num_tokens_full, num_tokens_swa)
         else:
-            load = (
+            num_tokens = (
                 self.max_total_num_tokens
                 - self.token_to_kv_pool_allocator.available_size()
                 - self.tree_cache.evictable_size()
             )
-        load += sum(len(req.origin_input_ids) for req in self.waiting_queue)
+
+        # Tokens in waiting queue, bootstrap queue, prealloc queue
+        num_tokens += sum(len(req.origin_input_ids) for req in self.waiting_queue)
+        num_waiting_reqs = len(self.waiting_queue)
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
-            load += sum(
+            num_tokens += sum(
                 len(req.origin_input_ids)
                 for req in self.disagg_prefill_bootstrap_queue.queue
             )
+            num_waiting_reqs += len(self.disagg_prefill_bootstrap_queue.queue)
         elif self.disaggregation_mode == DisaggregationMode.DECODE:
-            load += sum(
+            num_tokens += sum(
                 len(req.req.origin_input_ids)
                 for req in self.disagg_decode_prealloc_queue.queue
             )
+            num_waiting_reqs += len(self.disagg_decode_prealloc_queue.queue)
 
-        return load
+        return GetLoadReqOutput(
+            dp_rank=self.dp_rank,
+            num_reqs=len(self.running_batch.reqs) + num_waiting_reqs,
+            num_waiting_reqs=num_waiting_reqs,
+            num_tokens=num_tokens,
+        )
 
     def get_internal_state(self, recv_req: GetInternalStateReq):
         ret = dict(global_server_args_dict)
@@ -2317,10 +2347,9 @@ class Scheduler(
             "token_capacity": int(self.max_total_num_tokens),
         }
 
-        if not _is_cpu:
-            ret["memory_usage"]["cuda_graph"] = round(
-                self.tp_worker.worker.model_runner.cuda_graph_mem_usage, 2
-            )
+        ret["memory_usage"]["graph"] = round(
+            self.tp_worker.worker.model_runner.graph_mem_usage, 2
+        )
 
         if not self.spec_algorithm.is_none() and self.cum_spec_accept_count > 0:
             ret["avg_spec_accept_length"] = (
@@ -2329,8 +2358,6 @@ class Scheduler(
         if RECORD_STEP_TIME:
             ret["step_time_dict"] = self.step_time_dict
 
-        ret["load"] = self.get_load()
-
         return GetInternalStateReqOutput(internal_state=ret)
 
     def set_internal_state(self, recv_req: SetInternalStateReq):
@@ -2494,6 +2521,22 @@ class Scheduler(
         self.send_to_detokenizer.send_pyobj(recv_req)
         return recv_req
 
+    def init_weights_send_group_for_remote_instance(
+        self, recv_req: InitWeightsSendGroupForRemoteInstanceReqInput
+    ):
+        """Init the seed and client instance communication group."""
+        success, message = self.tp_worker.init_weights_send_group_for_remote_instance(
+            recv_req
+        )
+        return InitWeightsSendGroupForRemoteInstanceReqOutput(success, message)
+
+    def send_weights_to_remote_instance(
+        self, recv_req: SendWeightsToRemoteInstanceReqInput
+    ):
+        """Send the seed instance weights to the destination instance."""
+        success, message = self.tp_worker.send_weights_to_remote_instance(recv_req)
+        return SendWeightsToRemoteInstanceReqOutput(success, message)
+
     def slow_down(self, recv_req: SlowDownReqInput):
         t = recv_req.forward_sleep_time
         if t is not None and t <= 0:
@@ -2615,6 +2658,15 @@ def run_scheduler_process(
     pipe_writer,
     balance_meta: Optional[DPBalanceMeta] = None,
 ):
+    if server_args.enable_trace:
+        process_tracing_init(server_args.oltp_traces_endpoint, "sglang")
+        if server_args.disaggregation_mode == "null":
+            thread_label = "Scheduler"
+            trace_set_thread_info(thread_label, tp_rank, dp_rank)
+
+    if (numa_node := server_args.numa_node) is not None:
+        numa_bind_to_node(numa_node[gpu_id])
+
     # Generate the prefix
     prefix = ""
     if dp_rank is not None: