sglang 0.4.8__py3-none-any.whl → 0.4.8.post1__py3-none-any.whl

sglang/srt/disaggregation/nixl/conn.py

@@ -31,23 +31,19 @@ from sglang.srt.utils import get_local_ip_by_remote
 
 logger = logging.getLogger(__name__)
 
-NixlEngineInfo: TypeAlias = Dict[str, Union[str, int]]
-
 GUARD = "NixlMsgGuard".encode("ascii")
 
 
 @dataclasses.dataclass
 class TransferInfo:
+    """Contains indices for a transfer, sent by KVReceiver. Received by prefill bootstrap thread."""
+
     room: int
     endpoint: str
     dst_port: int
-    agent_metadata: bytes
     agent_name: str
-    dst_kv_ptrs: list[int]
     dst_kv_indices: npt.NDArray[np.int32]
-    dst_aux_ptrs: list[int]
     dst_aux_index: int
-    dst_gpu_id: int
     required_dst_info_num: int
 
     def is_dummy(self):
@@ -59,14 +55,37 @@ class TransferInfo:
             room=int(msg[0].decode("ascii")),
             endpoint=msg[1].decode("ascii"),
             dst_port=int(msg[2].decode("ascii")),
-            agent_metadata=msg[3],
-            agent_name=msg[4].decode("ascii"),
+            agent_name=msg[3].decode("ascii"),
+            dst_kv_indices=np.frombuffer(msg[4], dtype=np.int32),
+            dst_aux_index=int(msg[5].decode("ascii")),
+            required_dst_info_num=int(msg[6].decode("ascii")),
+        )
+
+
+@dataclasses.dataclass
+class KVArgsRegisterInfo:
+    """Contains base pointers and other info which only needs to be sent once by KVReceiver. Received by prefill bootstrap thread."""
+
+    room: str
+    endpoint: str
+    dst_port: int
+    agent_name: str
+    agent_metadata: bytes
+    dst_kv_ptrs: list[int]
+    dst_aux_ptrs: list[int]
+    gpu_id: int
+
+    @classmethod
+    def from_zmq(cls, msg: List[bytes]):
+        return cls(
+            room=str(msg[0].decode("ascii")),
+            endpoint=msg[1].decode("ascii"),
+            dst_port=int(msg[2].decode("ascii")),
+            agent_name=msg[3].decode("ascii"),
+            agent_metadata=msg[4],
             dst_kv_ptrs=list(struct.unpack(f"{len(msg[5])//8}Q", msg[5])),
-            dst_kv_indices=np.frombuffer(msg[6], dtype=np.int32),
-            dst_aux_ptrs=list(struct.unpack(f"{len(msg[7])//8}Q", msg[7])),
-            dst_aux_index=int(msg[8].decode("ascii")),
-            dst_gpu_id=int(msg[9].decode("ascii")),
-            required_dst_info_num=int(msg[10].decode("ascii")),
+            dst_aux_ptrs=list(struct.unpack(f"{len(msg[6])//8}Q", msg[6])),
+            gpu_id=int(msg[7].decode("ascii")),
         )
 
 
@@ -109,9 +128,9 @@ class NixlKVManager(CommonKVManager):
         self.register_buffer_to_engine()
 
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
-            self.request_status = {}
-            self.transfer_infos: Dict[int, TransferInfo] = {}
-            self.peer_names: Dict[str, str] = {}
+            self.request_status: Dict[int, KVPoll] = {}
+            self.transfer_infos: Dict[int, Dict[str, TransferInfo]] = {}
+            self.decode_kv_args_table: Dict[str, KVArgsRegisterInfo] = {}
             self._start_bootstrap_thread()
         elif self.disaggregation_mode == DisaggregationMode.DECODE:
             self.transfer_statuses: Dict[int, TransferStatus] = defaultdict(
@@ -154,10 +173,13 @@ class NixlKVManager(CommonKVManager):
         if not self.aux_descs:
             raise Exception("NIXL memory registration failed for aux tensors")
 
-    def _add_remote(self, agent_name: str, agent_metadata: bytes):
-        if agent_name not in self.peer_names:
-            self.peer_names[agent_name] = self.agent.add_remote_agent(agent_metadata)
-        return self.peer_names[agent_name]
+    def _add_remote_peer(self, decode_kv_args: KVArgsRegisterInfo):
+        agent_name = decode_kv_args.agent_name
+        if agent_name in self.decode_kv_args_table:
+            logger.info(f"Peer {agent_name} was already registered, ignoring.")
+            return
+        self.decode_kv_args_table[agent_name] = decode_kv_args
+        self.agent.add_remote_agent(decode_kv_args.agent_metadata)
 
     def send_kvcache(
         self,
@@ -262,17 +284,17 @@ class NixlKVManager(CommonKVManager):
             if req.is_dummy():
                 continue
 
-            peer_name = self._add_remote(req.agent_name, req.agent_metadata)
             chunked_dst_kv_indice = req.dst_kv_indices[index_slice]
             assert len(chunked_dst_kv_indice) == len(kv_indices)
+            assert req.agent_name in self.decode_kv_args_table
 
             notif = "_".join([str(req.room), "kv", str(chunk_id), str(int(is_last))])
             kv_xfer_handle = self.send_kvcache(
-                peer_name,
+                req.agent_name,
                 kv_indices,
-                req.dst_kv_ptrs,
+                self.decode_kv_args_table[req.agent_name].dst_kv_ptrs,
                 chunked_dst_kv_indice,
-                req.dst_gpu_id,
+                self.decode_kv_args_table[req.agent_name].gpu_id,
                 notif,
             )
             handles.append(kv_xfer_handle)
@@ -280,13 +302,15 @@ class NixlKVManager(CommonKVManager):
             if is_last:
                 assert aux_index is not None
                 aux_xfer_handle = self.send_aux(
-                    peer_name,
+                    req.agent_name,
                     aux_index,
-                    req.dst_aux_ptrs,
+                    self.decode_kv_args_table[req.agent_name].dst_aux_ptrs,
                     req.dst_aux_index,
                     str(req.room) + "_aux",
                 )
                 handles.append(aux_xfer_handle)
+        if is_last:
+            del self.transfer_infos[bootstrap_room]
         return handles
 
     def update_transfer_status(self):
@@ -328,16 +352,23 @@ class NixlKVManager(CommonKVManager):
                 ), f"First message should be {GUARD}. Foreign traffic?"
                 waiting_req_bytes = waiting_req_bytes[1:]
                 room = waiting_req_bytes[0].decode("ascii")
-
-                required_dst_info_num = int(waiting_req_bytes[10].decode("ascii"))
+                agent_name = waiting_req_bytes[3].decode("ascii")
+                if room == "None":
+                    # Register new peer and save KV base pointers.
+                    self._add_remote_peer(
+                        KVArgsRegisterInfo.from_zmq(waiting_req_bytes)
+                    )
+                    logger.debug(f"Register KVArgs from {agent_name} successfully")
+                    continue
                 room = int(room)
-                agent_name = waiting_req_bytes[4].decode("ascii")
                 if room not in self.transfer_infos:
                     self.transfer_infos[room] = {}
                 self.transfer_infos[room][agent_name] = TransferInfo.from_zmq(
                     waiting_req_bytes
                 )
-
+                required_dst_info_num = self.transfer_infos[room][
+                    agent_name
+                ].required_dst_info_num
                 logger.debug(f"got info {room=} {agent_name=} {required_dst_info_num=}")
                 if len(self.transfer_infos[room]) == required_dst_info_num:
                     logger.debug(f"{room=} is bootstrapped")
@@ -391,6 +422,7 @@ class NixlKVSender(BaseKVSender):
         self.chunk_id += 1
         if is_last:
             self.has_sent = True
+            del self.kv_mgr.request_status[self.bootstrap_room]
 
     def poll(self) -> KVPoll:
         if not self.has_sent:
@@ -415,6 +447,7 @@ class NixlKVReceiver(CommonKVReceiver):
         data_parallel_rank: Optional[int] = None,
     ):
         self.started_transfer = False
+        self.conclude_state = None
         super().__init__(mgr, bootstrap_addr, bootstrap_room, data_parallel_rank)
 
     def init(self, kv_indices: npt.NDArray[np.int32], aux_index: Optional[int] = None):
@@ -426,17 +459,8 @@ class NixlKVReceiver(CommonKVReceiver):
                 f"Fetched bootstrap info: {bootstrap_info} for engine rank: {self.kv_mgr.kv_args.engine_rank}"
             )
             is_dummy = bootstrap_info["is_dummy"]
-
-            # TODO: send_kv_args earlier
-            packed_kv_data_ptrs = b"".join(
-                struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.kv_data_ptrs
-            )
-            packed_aux_data_ptrs = b"".join(
-                struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.aux_data_ptrs
-            )
-
             logger.debug(
-                f"Sending to {self.prefill_server_url} with bootstrap room {self.bootstrap_room}"
+                f"Sending to {self.prefill_server_url} with bootstrap room {self.bootstrap_room} {is_dummy=}"
             )
             sock, lock = self._connect("tcp://" + self.prefill_server_url)
             with lock:
@@ -446,13 +470,9 @@ class NixlKVReceiver(CommonKVReceiver):
                         str(self.bootstrap_room).encode("ascii"),
                         get_local_ip_by_remote().encode("ascii"),
                         str(self.kv_mgr.rank_port).encode("ascii"),
-                        self.kv_mgr.agent.get_agent_metadata(),
                         self.kv_mgr.agent.name.encode("ascii"),
-                        packed_kv_data_ptrs,
                         kv_indices.tobytes() if not is_dummy else b"",
-                        packed_aux_data_ptrs,
                         str(aux_index).encode("ascii"),
-                        str(self.kv_mgr.kv_args.gpu_id).encode("ascii"),
                         str(self.required_dst_info_num).encode("ascii"),
                     ]
                 )
@@ -460,17 +480,45 @@ class NixlKVReceiver(CommonKVReceiver):
         self.started_transfer = True
 
     def poll(self) -> KVPoll:
+        if self.conclude_state is not None:
+            return self.conclude_state
         if not self.started_transfer:
             return KVPoll.WaitingForInput  # type: ignore
 
         self.kv_mgr.update_transfer_status()
-
         if self.kv_mgr.check_transfer_done(self.bootstrap_room):  # type: ignore
+            self.conclude_state = KVPoll.Success
+            del self.kv_mgr.transfer_statuses[self.bootstrap_room]
             return KVPoll.Success  # type: ignore
         return KVPoll.WaitingForInput  # type: ignore
 
     def _register_kv_args(self):
-        pass
+        for bootstrap_info in self.bootstrap_infos:
+            self.prefill_server_url = (
+                f"{bootstrap_info['rank_ip']}:{bootstrap_info['rank_port']}"
+            )
+            packed_kv_data_ptrs = b"".join(
+                struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.kv_data_ptrs
+            )
+            packed_aux_data_ptrs = b"".join(
+                struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.aux_data_ptrs
+            )
+
+            sock, lock = self._connect("tcp://" + self.prefill_server_url)
+            with lock:
+                sock.send_multipart(
+                    [
+                        GUARD,
+                        "None".encode("ascii"),
+                        get_local_ip_by_remote().encode("ascii"),
+                        str(self.kv_mgr.rank_port).encode("ascii"),
+                        self.kv_mgr.agent.name.encode("ascii"),
+                        self.kv_mgr.agent.get_agent_metadata(),
+                        packed_kv_data_ptrs,
+                        packed_aux_data_ptrs,
+                        str(self.kv_mgr.kv_args.gpu_id).encode("ascii"),
+                    ]
+                )
 
     def failure_exception(self):
         raise Exception("Fake KVReceiver Exception")

sglang/srt/disaggregation/prefill.py

@@ -93,8 +93,6 @@ class PrefillBootstrapQueue:
         self.gpu_id = gpu_id
         self.bootstrap_port = bootstrap_port
         self.queue: List[Req] = []
-        self.pp_rank = pp_rank
-        self.pp_size = pp_size
         self.gloo_group = gloo_group
         self.max_total_num_tokens = max_total_num_tokens
         self.scheduler = scheduler
@@ -124,6 +122,9 @@ class PrefillBootstrapQueue:
         kv_args.kv_data_ptrs = kv_data_ptrs
         kv_args.kv_data_lens = kv_data_lens
         kv_args.kv_item_lens = kv_item_lens
+        if not self.is_mla_backend:
+            kv_args.kv_head_num = self.token_to_kv_pool.head_num
+        kv_args.page_size = self.token_to_kv_pool.page_size
 
         kv_args.aux_data_ptrs, kv_args.aux_data_lens, kv_args.aux_item_lens = (
             self.metadata_buffers.get_buf_infos()

sglang/srt/disaggregation/utils.py

@@ -107,9 +107,6 @@ class MetadataBuffers:
             # The minimal size for RDMA is 64Bytes, so we pad it to > 64Bytes
             self.output_ids = torch.zeros((size, 16), dtype=torch.int32, device=device)
 
-            self.output_hidden_states = torch.zeros(
-                (size, hidden_size), dtype=dtype, device=device
-            )
             self.output_token_logprobs_val = torch.zeros(
                 (size, 16), dtype=torch.float32, device=device
             )
@@ -122,51 +119,50 @@ class MetadataBuffers:
             self.output_top_logprobs_idx = torch.zeros(
                 (size, max_top_logprobs_num), dtype=torch.int32, device=device
             )
+            self.output_hidden_states = torch.zeros(
+                (size, hidden_size), dtype=dtype, device=device
+            )
 
     def get_buf_infos(self):
         ptrs = [
             self.output_ids.data_ptr(),
-            self.output_hidden_states.data_ptr(),  # TODO: set None to avoid transfer hidden_states when spec_algorithm is None
             self.output_token_logprobs_val.data_ptr(),
             self.output_token_logprobs_idx.data_ptr(),
             self.output_top_logprobs_val.data_ptr(),
             self.output_top_logprobs_idx.data_ptr(),
+            self.output_hidden_states.data_ptr(),
         ]
         data_lens = [
             self.output_ids.nbytes,
-            self.output_hidden_states.nbytes,
             self.output_token_logprobs_val.nbytes,
             self.output_token_logprobs_idx.nbytes,
             self.output_top_logprobs_val.nbytes,
             self.output_top_logprobs_idx.nbytes,
+            self.output_hidden_states.nbytes,
         ]
         item_lens = [
             self.output_ids[0].nbytes,
-            self.output_hidden_states[0].nbytes,
             self.output_token_logprobs_val[0].nbytes,
             self.output_token_logprobs_idx[0].nbytes,
             self.output_top_logprobs_val[0].nbytes,
             self.output_top_logprobs_idx[0].nbytes,
+            self.output_hidden_states[0].nbytes,
         ]
         return ptrs, data_lens, item_lens
 
     def get_buf(self, idx: int):
         return (
             self.output_ids[idx],
-            self.output_hidden_states[idx],
             self.output_token_logprobs_val[idx],
             self.output_token_logprobs_idx[idx],
             self.output_top_logprobs_val[idx],
             self.output_top_logprobs_idx[idx],
+            self.output_hidden_states[idx],
         )
 
     def set_buf(self, req: Req):
 
         self.output_ids[req.metadata_buffer_index][0] = req.output_ids[0]
-        if req.hidden_states_tensor is not None:
-            self.output_hidden_states[req.metadata_buffer_index].copy_(
-                req.hidden_states_tensor
-            )
         if req.return_logprob:
             if req.output_token_logprobs_val:  # not none or empty list
                 self.output_token_logprobs_val[req.metadata_buffer_index][0] = (
@@ -189,6 +185,11 @@ class MetadataBuffers:
                 ] = torch.tensor(
                     req.output_top_logprobs_idx[0], dtype=torch.int32, device="cpu"
                 )
+        # for PD + spec decode
+        if req.hidden_states_tensor is not None:
+            self.output_hidden_states[req.metadata_buffer_index].copy_(
+                req.hidden_states_tensor
+            )
 
 
 #########################

sglang/srt/entrypoints/engine.py

@@ -115,13 +115,13 @@ class Engine(EngineBase):
         atexit.register(self.shutdown)
 
         # Allocate ports for inter-process communications
-        port_args = PortArgs.init_new(server_args)
+        self.port_args = PortArgs.init_new(server_args)
         logger.info(f"{server_args=}")
 
         # Launch subprocesses
         tokenizer_manager, template_manager, scheduler_info = _launch_subprocesses(
             server_args=server_args,
-            port_args=port_args,
+            port_args=self.port_args,
         )
         self.server_args = server_args
         self.tokenizer_manager = tokenizer_manager
@@ -130,7 +130,7 @@ class Engine(EngineBase):
 
         context = zmq.Context(2)
         self.send_to_rpc = get_zmq_socket(
-            context, zmq.DEALER, port_args.rpc_ipc_name, True
+            context, zmq.DEALER, self.port_args.rpc_ipc_name, True
         )
 
     def generate(
@@ -242,6 +242,7 @@ class Engine(EngineBase):
         token_ids_logprob: Optional[Union[List[List[int]], List[int]]] = None,
         lora_path: Optional[List[Optional[str]]] = None,
         custom_logit_processor: Optional[Union[List[str], str]] = None,
+        return_hidden_states: bool = False,
         stream: bool = False,
         bootstrap_host: Optional[Union[List[str], str]] = None,
         bootstrap_port: Optional[Union[List[int], int]] = None,
@@ -274,6 +275,7 @@ class Engine(EngineBase):
             top_logprobs_num=top_logprobs_num,
             token_ids_logprob=token_ids_logprob,
             lora_path=lora_path,
+            return_hidden_states=return_hidden_states,
             stream=stream,
             custom_logit_processor=custom_logit_processor,
             bootstrap_host=bootstrap_host,

sglang/srt/entrypoints/openai/protocol.py

@@ -14,7 +14,8 @@
 """Pydantic models for OpenAI API protocol"""
 
 import time
-from typing import Dict, List, Optional, Union
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Union
 
 from pydantic import (
     BaseModel,
@@ -195,6 +196,9 @@ class CompletionRequest(BaseModel):
     bootstrap_port: Optional[int] = None
     bootstrap_room: Optional[int] = None
 
+    # For request id
+    rid: Optional[Union[List[str], str]] = None
+
     @field_validator("max_tokens")
     @classmethod
     def validate_max_tokens_positive(cls, v):
@@ -309,6 +313,18 @@ class ChatCompletionMessageGenericParam(BaseModel):
     reasoning_content: Optional[str] = None
     tool_calls: Optional[List[ToolCall]] = Field(default=None, examples=[None])
 
+    @field_validator("role", mode="before")
+    @classmethod
+    def _normalize_role(cls, v):
+        if isinstance(v, str):
+            v_lower = v.lower()
+            if v_lower not in {"system", "assistant", "tool"}:
+                raise ValueError(
+                    "'role' must be one of 'system', 'assistant', or 'tool' (case-insensitive)."
+                )
+            return v_lower
+        raise ValueError("'role' must be a string")
+
 
 class ChatCompletionMessageUserParam(BaseModel):
     role: Literal["user"]
@@ -429,8 +445,8 @@ class ChatCompletionRequest(BaseModel):
     stream_reasoning: bool = True
     chat_template_kwargs: Optional[Dict] = None
 
-    # The request id.
-    rid: Optional[str] = None
+    # For request id
+    rid: Optional[Union[List[str], str]] = None
 
     # For PD disaggregation
     bootstrap_host: Optional[str] = None
@@ -528,7 +544,7 @@ class EmbeddingRequest(BaseModel):
     user: Optional[str] = None
 
     # The request id.
-    rid: Optional[str] = None
+    rid: Optional[Union[List[str], str]] = None
 
 
 class EmbeddingObject(BaseModel):
@@ -587,3 +603,30 @@ OpenAIServingRequest = Union[
     ScoringRequest,
     V1RerankReqInput,
 ]
+
+
+@dataclass
+class MessageProcessingResult:
+    """Result of processing chat messages and applying templates.
+
+    This dataclass encapsulates all the outputs from message processing including
+    prompt generation, multimodal data extraction, and constraint preparation.
+    Used internally by OpenAIServingChat to pass processed data between methods.
+
+    Args:
+        prompt: The final text prompt after applying chat template
+        prompt_ids: Either the text prompt (str) or tokenized IDs (List[int])
+        image_data: Extracted image data from messages, if any
+        audio_data: Extracted audio data from messages, if any
+        modalities: List of modality types present in the messages
+        stop: Combined stop strings from template and request
+        tool_call_constraint: Optional constraint for structured tool calls
+    """
+
+    prompt: str
+    prompt_ids: Union[str, List[int]]
+    image_data: Optional[Any]
+    audio_data: Optional[Any]
+    modalities: List[str]
+    stop: List[str]
+    tool_call_constraint: Optional[Any] = None