from __future__ import annotations import concurrent.futures import ctypes import dataclasses import logging import os import struct import threading import time from collections import defaultdict from typing import Dict, List, Optional, Set, Tuple import numpy as np import numpy.typing as npt import requests import zmq from sglang.srt.disaggregation.base.conn import KVArgs, KVPoll from sglang.srt.disaggregation.common.conn import ( CommonKVBootstrapServer, CommonKVManager, CommonKVReceiver, CommonKVSender, ) from sglang.srt.disaggregation.common.utils import ( FastQueue, group_concurrent_contiguous, ) from sglang.srt.disaggregation.mooncake.utils import ( check_mooncake_custom_mem_pool_enabled, ) from sglang.srt.disaggregation.utils import DisaggregationMode from sglang.srt.distributed.parallel_state import get_mooncake_transfer_engine from sglang.srt.environ import envs from sglang.srt.server_args import ServerArgs from sglang.srt.utils import format_tcp_address, is_valid_ipv6_address logger = logging.getLogger(__name__) class KVTransferError(Exception): def __init__(self, bootstrap_room: int, failure_reason: str): super().__init__(failure_reason) self.bootstrap_room = bootstrap_room self.failure_reason = failure_reason def __str__(self): return f"KVTransferError(bootstrap_room={self.bootstrap_room}): {self.failure_reason}" # prefill @dataclasses.dataclass class TransferKVChunk: room: int prefill_kv_indices: npt.NDArray[np.int32] index_slice: slice is_last: bool prefill_aux_index: Optional[int] state_indices: Optional[List[int]] # decode @dataclasses.dataclass class TransferInfo: room: int endpoint: str dst_port: int mooncake_session_id: str dst_kv_indices: npt.NDArray[np.int32] dst_aux_index: int dst_state_indices: List[int] required_dst_info_num: int is_dummy: bool @classmethod def from_zmq(cls, msg: List[bytes]): if msg[4] == b"" and msg[5] == b"": is_dummy = True dst_kv_indices = np.array([], dtype=np.int32) dst_aux_index = None dst_state_indices = [] else: dst_kv_indices = np.frombuffer(msg[4], dtype=np.int32) dst_aux_index = int(msg[5].decode("ascii")) if msg[6] == b"": dst_state_indices = [] else: dst_state_indices = list(np.frombuffer(msg[6], dtype=np.int32)) is_dummy = False return cls( room=int(msg[0].decode("ascii")), endpoint=msg[1].decode("ascii"), dst_port=int(msg[2].decode("ascii")), mooncake_session_id=msg[3].decode("ascii"), dst_kv_indices=dst_kv_indices, dst_aux_index=dst_aux_index, dst_state_indices=dst_state_indices, required_dst_info_num=int(msg[7].decode("ascii")), is_dummy=is_dummy, ) # decode @dataclasses.dataclass class KVArgsRegisterInfo: room: str endpoint: str dst_port: int mooncake_session_id: str dst_kv_ptrs: list[int] dst_aux_ptrs: list[int] dst_state_data_ptrs: list[int] dst_tp_rank: int dst_attn_tp_size: int dst_kv_item_len: int # for mamba state different tp slice transfer dst_state_item_lens: list[int] dst_state_dim_per_tensor: list[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")), mooncake_session_id=msg[3].decode("ascii"), dst_kv_ptrs=list(struct.unpack(f"{len(msg[4])//8}Q", msg[4])), dst_aux_ptrs=list(struct.unpack(f"{len(msg[5])//8}Q", msg[5])), dst_state_data_ptrs=list(struct.unpack(f"{len(msg[6])//8}Q", msg[6])), dst_tp_rank=int(msg[7].decode("ascii")), dst_attn_tp_size=int(msg[8].decode("ascii")), dst_kv_item_len=int(msg[9].decode("ascii")), dst_state_item_lens=( list(struct.unpack(f"{len(msg[10])//4}I", msg[10])) if len(msg) > 10 and len(msg[10]) > 0 else [] ), dst_state_dim_per_tensor=( list(struct.unpack(f"{len(msg[11])//4}I", msg[11])) if len(msg) > 11 and len(msg[11]) > 0 else [] ), ) class AuxDataCodec: """Handles serialization and deserialization of auxiliary data buffers""" @staticmethod def serialize_data_from_buffer(src_addr, data_length): """Serialize data from memory buffer to bytes""" buffer = (ctypes.c_byte * data_length).from_address(src_addr) return bytes(buffer) @staticmethod def deserialize_data_to_buffer(kv_args, buffer_index, aux_index, data): """Deserialize bytes into target memory buffer""" dst_aux_ptr = kv_args.aux_data_ptrs[buffer_index] item_len = kv_args.aux_item_lens[buffer_index] dst_addr = dst_aux_ptr + item_len * aux_index buffer = (ctypes.c_byte * len(data)).from_address(dst_addr) buffer[:] = data return class MooncakeKVManager(CommonKVManager): AUX_DATA_HEADER = b"AUX_DATA" def __init__( self, args: KVArgs, disaggregation_mode: DisaggregationMode, server_args: ServerArgs, is_mla_backend: Optional[bool] = False, ): super().__init__(args, disaggregation_mode, server_args, is_mla_backend) self.init_engine() self.register_buffer_to_engine() if self.disaggregation_mode == DisaggregationMode.PREFILL: self.start_prefill_thread() self.session_failures = defaultdict(int) self.failed_sessions = set() self.session_lock = threading.Lock() # Determine the number of threads to use for kv sender cpu_count = os.cpu_count() transfer_thread_pool_size = ( envs.SGLANG_DISAGGREGATION_THREAD_POOL_SIZE.get() ) if transfer_thread_pool_size is None: transfer_thread_pool_size = min(max(4, int(0.5 * cpu_count) // 8), 12) transfer_queue_size = envs.SGLANG_DISAGGREGATION_QUEUE_SIZE.get() self.transfer_queues: List[FastQueue] = [ FastQueue() for _ in range(transfer_queue_size) ] assert transfer_thread_pool_size >= transfer_queue_size, ( f"The environment variable SGLANG_DISAGGREGATION_THREAD_POOL_SIZE={transfer_thread_pool_size} must be " f"greater than or equal to SGLANG_DISAGGREGATION_QUEUE_SIZE={transfer_queue_size}." ) self.executors = [ concurrent.futures.ThreadPoolExecutor( transfer_thread_pool_size // transfer_queue_size ) for _ in range(transfer_queue_size) ] for queue, executor in zip(self.transfer_queues, self.executors): threading.Thread( target=self.transfer_worker, args=(queue, executor), daemon=True ).start() # If a timeout happens on the prefill side, it means prefill instances # fail to receive the KV indices from the decode instance of this request. # These timeout requests should be aborted to release the tree cache. self.bootstrap_timeout = envs.SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT.get() self.enable_custom_mem_pool, self.custom_mem_pool_type = ( check_mooncake_custom_mem_pool_enabled() ) elif self.disaggregation_mode == DisaggregationMode.DECODE: self.heartbeat_failures = {} self.session_pool = defaultdict(requests.Session) self.session_pool_lock = threading.Lock() self.addr_to_rooms_tracker = defaultdict(set) self.prefill_response_tracker: Dict[int, Set[int]] = defaultdict(set) # Heartbeat interval should be at least 2 seconds self.heartbeat_interval = max( envs.SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL.get(), 2.0 ) # Heartbeat failure should be at least 1 self.max_failures = max( envs.SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE.get(), 1 ) self.start_decode_thread() # If a timeout happens on the decode side, it means decode instances # fail to receive the KV Cache transfer done signal after bootstrapping. # These timeout requests should be aborted to release the tree cache. self.waiting_timeout = envs.SGLANG_DISAGGREGATION_WAITING_TIMEOUT.get() self.failure_records: Dict[int, str] = {} self.failure_lock = threading.Lock() def init_engine(self): self.engine = get_mooncake_transfer_engine() def register_buffer_to_engine(self): # Batch register KV data buffers if self.kv_args.kv_data_ptrs and self.kv_args.kv_data_lens: self.engine.batch_register( self.kv_args.kv_data_ptrs, self.kv_args.kv_data_lens ) # Batch register auxiliary data buffers if self.kv_args.aux_data_ptrs and self.kv_args.aux_data_lens: self.engine.batch_register( self.kv_args.aux_data_ptrs, self.kv_args.aux_data_lens ) # Batch register state/extra pool data buffers if self.kv_args.state_data_ptrs and self.kv_args.state_data_lens: self.engine.batch_register( self.kv_args.state_data_ptrs, self.kv_args.state_data_lens ) def _transfer_data(self, mooncake_session_id, transfer_blocks): if not transfer_blocks: return 0 src_addrs, dst_addrs, lengths = zip(*transfer_blocks) return self.engine.batch_transfer_sync( mooncake_session_id, list(src_addrs), list(dst_addrs), list(lengths) ) def _send_kvcache_generic( self, mooncake_session_id: str, src_data_ptrs: list[int], dst_data_ptrs: list[int], item_lens: list[int], prefill_data_indices: npt.NDArray[np.int32], dst_data_indices: npt.NDArray[np.int32], executor: concurrent.futures.ThreadPoolExecutor, ) -> int: """ Generic KV cache transfer supporting both MHA and MLA architectures. This method is used by both send_kvcache (full pool) and maybe_send_extra. """ # Group by indices for optimization prefill_kv_blocks, dst_kv_blocks = group_concurrent_contiguous( prefill_data_indices, dst_data_indices ) layers_params = None # Decode pp size should be equal to prefill pp size or 1 if self.is_mla_backend: src_kv_ptrs, dst_kv_ptrs, layers_current_pp_stage = ( self.get_mla_kv_ptrs_with_pp(src_data_ptrs, dst_data_ptrs) ) layers_params = [ ( src_kv_ptrs[layer_id], dst_kv_ptrs[layer_id], item_lens[layer_id], ) for layer_id in range(layers_current_pp_stage) ] else: src_k_ptrs, src_v_ptrs, dst_k_ptrs, dst_v_ptrs, layers_current_pp_stage = ( self.get_mha_kv_ptrs_with_pp(src_data_ptrs, dst_data_ptrs) ) # item_lens structure: [k_layer0, k_layer1, ..., k_layerN, v_layer0, v_layer1, ..., v_layerN] # Use correct item lengths for K and V separately if layers_current_pp_stage > len(dst_k_ptrs): logger.error( "Prefill transfer kvcache error, layers_current_pp_stage is out of range: " f"layers_current_pp_stage={layers_current_pp_stage}, len(dst_k_ptrs)={len(dst_k_ptrs)}" ) return -1 layers_params = [ ( src_k_ptrs[layer_id], dst_k_ptrs[layer_id], item_lens[layer_id], # K item length ) for layer_id in range(layers_current_pp_stage) ] + [ ( src_v_ptrs[layer_id], dst_v_ptrs[layer_id], item_lens[layers_current_pp_stage + layer_id], # V item length ) for layer_id in range(layers_current_pp_stage) ] assert layers_params is not None def set_transfer_blocks( src_ptr: int, dst_ptr: int, item_len: int ) -> List[Tuple[int, int, int]]: transfer_blocks = [] for prefill_index, decode_index in zip(prefill_kv_blocks, dst_kv_blocks): src_addr = src_ptr + int(prefill_index[0]) * item_len dst_addr = dst_ptr + int(decode_index[0]) * item_len length = item_len * len(prefill_index) transfer_blocks.append((src_addr, dst_addr, length)) return transfer_blocks # Worker function for processing a single layer def process_layer(src_ptr: int, dst_ptr: int, item_len: int) -> int: transfer_blocks = set_transfer_blocks(src_ptr, dst_ptr, item_len) return self._transfer_data(mooncake_session_id, transfer_blocks) # Worker function for processing all layers in a batch def process_layers(layers_params: List[Tuple[int, int, int]]) -> int: transfer_blocks = [] for src_ptr, dst_ptr, item_len in layers_params: transfer_blocks.extend(set_transfer_blocks(src_ptr, dst_ptr, item_len)) return self._transfer_data(mooncake_session_id, transfer_blocks) if self.enable_custom_mem_pool: futures = [ executor.submit( process_layer, src_ptr, dst_ptr, item_len, ) for (src_ptr, dst_ptr, item_len) in layers_params ] for future in concurrent.futures.as_completed(futures): status = future.result() if status != 0: for f in futures: f.cancel() return status return 0 else: # Combining all layers' params in one batch transfer is more efficient # compared to using multiple threads return process_layers(layers_params) def send_kvcache( self, mooncake_session_id: str, prefill_kv_indices: npt.NDArray[np.int32], dst_kv_ptrs: list[int], dst_kv_indices: npt.NDArray[np.int32], executor: concurrent.futures.ThreadPoolExecutor, ): return self._send_kvcache_generic( mooncake_session_id=mooncake_session_id, src_data_ptrs=self.kv_args.kv_data_ptrs, dst_data_ptrs=dst_kv_ptrs, item_lens=self.kv_args.kv_item_lens, prefill_data_indices=prefill_kv_indices, dst_data_indices=dst_kv_indices, executor=executor, ) def send_kvcache_slice( self, mooncake_session_id: str, prefill_kv_indices: npt.NDArray[np.int32], dst_kv_ptrs: list[int], dst_kv_indices: npt.NDArray[np.int32], dst_tp_rank: int, dst_attn_tp_size: int, dst_kv_item_len: int, executor: concurrent.futures.ThreadPoolExecutor, ): """ Sends KV cache slices from this Prefill rank to a target Decode rank, supporting generic M-to-N TP size configurations. NOTE: This implementation calls the transfer engine for each token slot within each page to ensure correctness for any page_size and head-slicing configuration. This may introduce performance overhead (increased TTFT) for long sequences. """ # Extract configuration local_tp_rank_in_group = self.kv_args.engine_rank % self.attn_tp_size src_kv_item_len = self.kv_args.kv_item_lens[0] dst_tp_rank_in_group = dst_tp_rank % dst_attn_tp_size num_kv_heads = self.kv_args.kv_head_num page_size = self.kv_args.page_size # Calculate head distribution src_heads_per_rank = num_kv_heads dst_heads_per_rank = num_kv_heads * self.attn_tp_size // dst_attn_tp_size bytes_per_head_slice_to_send = ( dst_kv_item_len // page_size // dst_heads_per_rank ) # Determine slicing parameters based on TP configuration if self.attn_tp_size > dst_attn_tp_size: # Send KVCache from multiple prefill instances to 1 decode instance src_head_start_offset = 0 num_heads_to_send = src_heads_per_rank dst_head_start_offset = local_tp_rank_in_group * src_heads_per_rank else: # Send KVCache from 1 prefill instance to multiple decode instances src_head_start_offset = ( dst_tp_rank_in_group * dst_heads_per_rank ) % src_heads_per_rank num_heads_to_send = dst_heads_per_rank dst_head_start_offset = 0 src_k_ptrs, src_v_ptrs, dst_k_ptrs, dst_v_ptrs, layers_current_pp_stage = ( self.get_mha_kv_ptrs_with_pp(self.kv_args.kv_data_ptrs, dst_kv_ptrs) ) # Calculate precise byte offset and length for the sub-slice within the token src_head_slice_offset = src_head_start_offset * bytes_per_head_slice_to_send dst_head_slice_offset = dst_head_start_offset * bytes_per_head_slice_to_send heads_bytes_per_token_to_send = num_heads_to_send * bytes_per_head_slice_to_send # Sanity check: The data sub-slice to be sent should fit into the dst buffer. # This means heads_bytes_per_token_to_send <= (dst_kv_item_len // page_size) if heads_bytes_per_token_to_send > (dst_kv_item_len // page_size): logger.error( f"[{mooncake_session_id}] slice size ({heads_bytes_per_token_to_send}) exceeds " f"target token slot size ({dst_kv_item_len // page_size})" ) return -1 prefill_page_indices = prefill_kv_indices.reshape(-1, 1).astype(np.int64) decode_page_indices = dst_kv_indices.reshape(-1, 1).astype(np.int64) tokens_per_page = np.arange(page_size, dtype=np.int64).reshape(1, -1) bytes_per_token_on_prefill = src_kv_item_len // page_size bytes_per_token_on_decode = dst_kv_item_len // page_size src_token_slot_offsets = ( tokens_per_page * bytes_per_token_on_prefill + src_head_slice_offset ) dst_token_slot_offsets = ( tokens_per_page * bytes_per_token_on_decode + dst_head_slice_offset ) def process_layer_tp_aware(src_layer_ptr, dst_layer_ptr): src_page_base_addrs = src_layer_ptr + prefill_page_indices * src_kv_item_len dst_page_base_addrs = dst_layer_ptr + decode_page_indices * dst_kv_item_len src_slice_addrs = src_page_base_addrs + src_token_slot_offsets dst_slice_addrs = dst_page_base_addrs + dst_token_slot_offsets src_addr_list = src_slice_addrs.reshape(-1).tolist() if not src_addr_list: # Nothing to transfer for this layer. return 0 dst_addr_list = dst_slice_addrs.reshape(-1).tolist() total_slices = len(src_addr_list) length_list = [heads_bytes_per_token_to_send] * total_slices return self.engine.batch_transfer_sync( mooncake_session_id, src_addr_list, dst_addr_list, length_list ) futures = [] for i in range(layers_current_pp_stage): futures.append( executor.submit(process_layer_tp_aware, src_k_ptrs[i], dst_k_ptrs[i]) ) for i in range(layers_current_pp_stage): futures.append( executor.submit(process_layer_tp_aware, src_v_ptrs[i], dst_v_ptrs[i]) ) for future in concurrent.futures.as_completed(futures): status = future.result() if status != 0: for f in futures: f.cancel() return status return 0 def send_aux( self, req: TransferInfo, prefill_aux_index: int, dst_aux_ptrs: list[int], ): # TODO(shangming): Fix me when nvlink_transport of Mooncake is bug-free if ( self.enable_custom_mem_pool and self.custom_mem_pool_type == "NVLINK" ) or envs.SGLANG_MOONCAKE_SEND_AUX_TCP.get(): return self.send_aux_tcp(req, prefill_aux_index, dst_aux_ptrs) transfer_blocks = [] prefill_aux_ptrs = self.kv_args.aux_data_ptrs prefill_aux_item_lens = self.kv_args.aux_item_lens for i, dst_aux_ptr in enumerate(dst_aux_ptrs): length = prefill_aux_item_lens[i] src_addr = prefill_aux_ptrs[i] + length * prefill_aux_index dst_addr = dst_aux_ptrs[i] + length * req.dst_aux_index transfer_blocks.append((src_addr, dst_addr, length)) return self._transfer_data(req.mooncake_session_id, transfer_blocks) def send_aux_tcp( self, req: TransferInfo, prefill_aux_index: int, dst_aux_ptrs: list[int], ): prefill_aux_ptrs = self.kv_args.aux_data_ptrs prefill_aux_item_lens = self.kv_args.aux_item_lens for i in range(len(prefill_aux_ptrs)): length = prefill_aux_item_lens[i] src_addr = prefill_aux_ptrs[i] + length * prefill_aux_index data = AuxDataCodec.serialize_data_from_buffer(src_addr, length) self.send_aux_data_to_endpoint( remote=req.endpoint, dst_port=req.dst_port, room=req.room, buffer_index=i, aux_index=req.dst_aux_index, data=data, ) return 0 def send_aux_data_to_endpoint( self, remote: str, dst_port: int, room: int, buffer_index: int, aux_index: int, data: bytes, ): socket = self._connect( format_tcp_address(remote, dst_port), is_ipv6=is_valid_ipv6_address(remote) ) socket.send_multipart( [ MooncakeKVManager.AUX_DATA_HEADER, str(room).encode("ascii"), str(buffer_index).encode("ascii"), str(aux_index).encode("ascii"), struct.pack(">I", len(data)), data, ] ) def _handle_aux_data(self, msg: List[bytes]): """Handle AUX_DATA messages received by the decode thread.""" room = int(msg[1].decode("ascii")) buffer_index = int(msg[2].decode("ascii")) aux_index = int(msg[3].decode("ascii")) data_length = struct.unpack(">I", msg[4])[0] data = msg[5] if len(data) != data_length: logger.error(f"AUX_DATA length mismatch for bootstrap_room {room}") return AuxDataCodec.deserialize_data_to_buffer( self.kv_args, buffer_index, aux_index, data ) logger.debug( f"Received AUX_DATA for bootstrap_room {room} with length:{len(data)}" ) def maybe_send_extra( self, req: TransferInfo, prefill_state_indices: list[int], dst_state_data_ptrs: list[int], executor: concurrent.futures.ThreadPoolExecutor, target_rank_registration_info: Optional[KVArgsRegisterInfo] = None, ): """Send state or extra pool data with type-specific handling.""" state_type = getattr(self.kv_args, "state_type", "none") if state_type == "mamba": # Check if we need slice transfer for different TP sizes if ( target_rank_registration_info is not None and self.attn_tp_size != target_rank_registration_info.dst_attn_tp_size ): return self._send_mamba_state_slice( req, prefill_state_indices, dst_state_data_ptrs, target_rank_registration_info.dst_state_item_lens, target_rank_registration_info.dst_state_dim_per_tensor, target_rank_registration_info.dst_tp_rank, target_rank_registration_info.dst_attn_tp_size, ) else: return self._send_mamba_state( req, prefill_state_indices, dst_state_data_ptrs, ) elif state_type in ["swa", "nsa"]: # SWA and NSA hybrid models do not support different TP sizes yet if ( target_rank_registration_info is not None and not self.is_mla_backend and self.attn_tp_size != target_rank_registration_info.dst_attn_tp_size ): raise RuntimeError( f"PD Disaggregation does NOT support PD different TP sizes for non-MLA {state_type.upper()} hybrid models yet." ) if len(prefill_state_indices) < len(req.dst_state_indices): logger.warning( f"len(prefill_state_indices) = {len(prefill_state_indices)}, len(dst_state_indices) = {len(req.dst_state_indices)}" ) prefill_state_indices = prefill_state_indices[ : len(req.dst_state_indices) ] # Reuse _send_kvcache_generic interface to send extra pool data prefill_state_indices = np.array(prefill_state_indices, dtype=np.int32) dst_state_indices = np.array(req.dst_state_indices, dtype=np.int32) return self._send_kvcache_generic( mooncake_session_id=req.mooncake_session_id, src_data_ptrs=self.kv_args.state_data_ptrs, dst_data_ptrs=dst_state_data_ptrs, item_lens=self.kv_args.state_item_lens, prefill_data_indices=prefill_state_indices, dst_data_indices=dst_state_indices, executor=executor, ) else: return 0 def _send_mamba_state( self, req: TransferInfo, prefill_mamba_index: list[int], dst_state_data_ptrs: list[int], ): """Transfer Mamba states.""" assert len(prefill_mamba_index) == 1, "Mamba should have single state index" transfer_blocks = [] prefill_state_data_ptrs = self.kv_args.state_data_ptrs prefill_state_item_lens = self.kv_args.state_item_lens for i, dst_state_ptr in enumerate(dst_state_data_ptrs): length = prefill_state_item_lens[i] src_addr = prefill_state_data_ptrs[i] + length * int(prefill_mamba_index[0]) dst_addr = dst_state_ptr + length * int(req.dst_state_indices[0]) transfer_blocks.append((src_addr, dst_addr, length)) return self._transfer_data(req.mooncake_session_id, transfer_blocks) def _send_mamba_state_slice( self, req: TransferInfo, prefill_mamba_index: list[int], dst_state_data_ptrs: list[int], dst_state_item_lens: list[int], dst_state_dim_per_tensor: list[int], dst_tp_rank: int, dst_attn_tp_size: int, ): """Transfer Mamba states with TP slice support. Mamba state layout: - conv_state: [num_layers, size+1, conv_dim/tp, conv_kernel-1] - temporal_state: [num_layers, size+1, num_heads/tp, head_dim, state_size] The 3rd dimension is sliced by TP. When prefill and decode have different attn_tp_size, we need to slice the state accordingly. """ logger.warning_once( "Using Mamba state slice transfer for different TP sizes between prefill and decode. " f"Prefill attn_tp_size={self.attn_tp_size}, Decode attn_tp_size={dst_attn_tp_size}. " "Performance may be affected." ) assert len(prefill_mamba_index) == 1, "Mamba should have single state index" transfer_blocks = [] prefill_state_data_ptrs = self.kv_args.state_data_ptrs prefill_state_item_lens = self.kv_args.state_item_lens src_state_dim_per_tensor = getattr(self.kv_args, "state_dim_per_tensor", []) # If no dimension info available, fall back to regular transfer if not src_state_dim_per_tensor or not dst_state_dim_per_tensor: return self._send_mamba_state(req, prefill_mamba_index, dst_state_data_ptrs) local_tp_rank_in_group = self.kv_args.engine_rank % self.attn_tp_size dst_tp_rank_in_group = dst_tp_rank % dst_attn_tp_size for i, dst_state_ptr in enumerate(dst_state_data_ptrs): src_item_len = prefill_state_item_lens[i] dst_item_len = dst_state_item_lens[i] src_dim = src_state_dim_per_tensor[i] dst_dim = dst_state_dim_per_tensor[i] # Calculate bytes per dimension slice # item_len = dim * trailing_dims_size, so trailing_dims_size = item_len / dim src_bytes_per_dim = src_item_len // src_dim dst_bytes_per_dim = dst_item_len // dst_dim # Determine slicing parameters based on TP configuration if self.attn_tp_size > dst_attn_tp_size: # Multiple prefill ranks send to 1 decode rank # Each prefill sends all its dims to the appropriate offset in decode src_dim_start = 0 num_dims_to_send = src_dim dst_dim_start = local_tp_rank_in_group * src_dim else: # 1 prefill rank sends to multiple decode ranks # Prefill sends a slice of its dims to each decode rank src_dim_start = (dst_tp_rank_in_group * dst_dim) % src_dim num_dims_to_send = dst_dim dst_dim_start = 0 # Calculate byte offsets src_dim_offset = src_dim_start * src_bytes_per_dim dst_dim_offset = dst_dim_start * dst_bytes_per_dim bytes_to_send = num_dims_to_send * src_bytes_per_dim # Calculate addresses for this state tensor src_addr = ( prefill_state_data_ptrs[i] + src_item_len * int(prefill_mamba_index[0]) + src_dim_offset ) dst_addr = ( dst_state_ptr + dst_item_len * int(req.dst_state_indices[0]) + dst_dim_offset ) transfer_blocks.append((src_addr, dst_addr, bytes_to_send)) return self._transfer_data(req.mooncake_session_id, transfer_blocks) def sync_status_to_decode_endpoint( self, remote: str, dst_port: int, room: int, status: int, prefill_rank: int ): self._connect( format_tcp_address(remote, dst_port), is_ipv6=is_valid_ipv6_address(remote) ).send_multipart( [ str(room).encode("ascii"), str(status).encode("ascii"), str(prefill_rank).encode("ascii"), ] ) def transfer_worker( self, queue: FastQueue, executor: concurrent.futures.ThreadPoolExecutor ): while True: try: kv_chunk: TransferKVChunk = queue.get() reqs_to_be_processed = ( self.transfer_infos[kv_chunk.room].values() if kv_chunk.room in self.transfer_infos else [] ) polls = [] dst_ranks_infos = [] local_rank = self.attn_tp_rank * self.pp_size + self.pp_rank for req in reqs_to_be_processed: if not req.is_dummy: # Early exit if the request has failed with self.session_lock: if req.mooncake_session_id in self.failed_sessions: self.record_failure( kv_chunk.room, f"Decode instance could be dead, remote mooncake session {req.mooncake_session_id} is not alive", ) self.update_status(kv_chunk.room, KVPoll.Failed) self.sync_status_to_decode_endpoint( req.endpoint, req.dst_port, req.room, KVPoll.Failed, local_rank, ) break chunked_dst_kv_indice = req.dst_kv_indices[kv_chunk.index_slice] # NOTE: This is temporarily a workaround to deal with the case where the prefill_kv_indices # is mismatched with the dst_kv_indices when page size > 1, this should never happen. if len(chunked_dst_kv_indice) < len( kv_chunk.prefill_kv_indices ): logger.warning( f"len(chunked_dst_kv_indice) = {len(chunked_dst_kv_indice)}, len(kv_chunk.prefill_kv_indices) = {len(kv_chunk.prefill_kv_indices)}" ) kv_chunk.prefill_kv_indices = kv_chunk.prefill_kv_indices[ : len(chunked_dst_kv_indice) ] target_rank_registration_info: KVArgsRegisterInfo = ( self.decode_kv_args_table[req.mooncake_session_id] ) if self.is_mla_backend or ( self.attn_tp_size == target_rank_registration_info.dst_attn_tp_size ): ret = self.send_kvcache( req.mooncake_session_id, kv_chunk.prefill_kv_indices, target_rank_registration_info.dst_kv_ptrs, chunked_dst_kv_indice, executor, ) else: ret = self.send_kvcache_slice( req.mooncake_session_id, kv_chunk.prefill_kv_indices, target_rank_registration_info.dst_kv_ptrs, chunked_dst_kv_indice, target_rank_registration_info.dst_tp_rank, target_rank_registration_info.dst_attn_tp_size, target_rank_registration_info.dst_kv_item_len, executor, ) if ret != 0: with self.session_lock: self.session_failures[req.mooncake_session_id] += 1 # Failures should never happen if the session is not dead, if the session fails once, mark it as failed if self.session_failures[req.mooncake_session_id] >= 1: self.failed_sessions.add(req.mooncake_session_id) logger.error( f"Session {req.mooncake_session_id} failed." ) self.record_failure( kv_chunk.room, f"Failed to send kv chunk of {kv_chunk.room} to {req.endpoint}:{req.dst_port}", ) self.update_status(kv_chunk.room, KVPoll.Failed) self.sync_status_to_decode_endpoint( req.endpoint, req.dst_port, req.room, KVPoll.Failed, local_rank, ) break if kv_chunk.is_last: if kv_chunk.state_indices is not None: self.maybe_send_extra( req, kv_chunk.state_indices, target_rank_registration_info.dst_state_data_ptrs, executor, target_rank_registration_info, ) # Only the last chunk we need to send the aux data ret = self.send_aux( req, kv_chunk.prefill_aux_index, target_rank_registration_info.dst_aux_ptrs, ) polls.append(True if ret == 0 else False) dst_ranks_infos.append( (req.endpoint, req.dst_port, req.room) ) # Only sync status when all the dst ranks have received the kvcache if len(polls) == req.required_dst_info_num: status = KVPoll.Success if all(polls) else KVPoll.Failed self.update_status(req.room, status) for endpoint, dst_port, room in dst_ranks_infos: self.sync_status_to_decode_endpoint( endpoint, dst_port, room, status, local_rank ) else: # Dummy request means the decode instance is not used, so its status can be marked as success directly # Dummy request does not need to sync status to decode endpoint if kv_chunk.is_last and req.room in self.request_status: self.update_status(req.room, KVPoll.Success) if ( kv_chunk.room not in self.request_status or self.check_status(kv_chunk.room) == KVPoll.Success ): if kv_chunk.room in self.transfer_infos: self.transfer_infos.pop(kv_chunk.room) except Exception as e: # NOTE(shangming): Remove this when we make sure the transfer thread is bug-free raise RuntimeError( f"Transfer thread failed because of {e}. Prefill instance with bootstrap_port={self.bootstrap_port} is dead." ) def start_prefill_thread(self): def bootstrap_thread(): """This thread recvs pre-alloc notification from the decode engine""" # KVPoll.Bootstrapping -> KVPoll.WaitingForInput while True: waiting_req_bytes = self.server_socket.recv_multipart() room = waiting_req_bytes[0].decode("ascii") mooncake_session_id = waiting_req_bytes[3].decode("ascii") if room == "None": self.decode_kv_args_table[mooncake_session_id] = ( KVArgsRegisterInfo.from_zmq(waiting_req_bytes) ) with self.session_lock: if mooncake_session_id in self.failed_sessions: self.failed_sessions.remove(mooncake_session_id) if mooncake_session_id in self.session_failures: del self.session_failures[mooncake_session_id] logger.debug( f"Register KVArgs from {mooncake_session_id} successfully" ) continue else: required_dst_info_num = int(waiting_req_bytes[7].decode("ascii")) room = int(room) if room not in self.transfer_infos: self.transfer_infos[room] = {} self.transfer_infos[room][mooncake_session_id] = ( TransferInfo.from_zmq(waiting_req_bytes) ) # NOTE: after bootstrapping we can mark the req as waiting for input if len(self.transfer_infos[room]) == required_dst_info_num: self.update_status(room, KVPoll.WaitingForInput) threading.Thread(target=bootstrap_thread).start() def start_decode_thread(self): def decode_thread(): while True: msg = self.server_socket.recv_multipart() if msg[0] == MooncakeKVManager.AUX_DATA_HEADER: self._handle_aux_data(msg) continue bootstrap_room, status, prefill_rank = msg status = int(status.decode("ascii")) bootstrap_room = int(bootstrap_room.decode("ascii")) prefill_rank = int(prefill_rank.decode("ascii")) if status == KVPoll.Success: if bootstrap_room in self.request_status: self.prefill_response_tracker[bootstrap_room].add(prefill_rank) expected_response_num = ( self.required_prefill_response_num_table[bootstrap_room] ) arrived_response_num = len( self.prefill_response_tracker[bootstrap_room] ) if arrived_response_num == expected_response_num: self.update_status(bootstrap_room, KVPoll.Success) elif status == KVPoll.Failed: self.record_failure( bootstrap_room, "Failed to get kvcache from prefill instance, it might be dead", ) self.update_status(bootstrap_room, status) def heartbeat_checker(): while True: time.sleep(self.heartbeat_interval) with self.connection_lock: addresses = list(self.prefill_dp_size_table.keys()) for bootstrap_addr in addresses: session = None try: with self.session_pool_lock: session = self.session_pool[bootstrap_addr] response = session.get( f"http://{bootstrap_addr}/health", timeout=(2, 3), headers={"Connection": "keep-alive"}, ) if response.status_code == 200: self.heartbeat_failures[bootstrap_addr] = 0 current_rooms = self.addr_to_rooms_tracker[ bootstrap_addr ].copy() for bootstrap_room in current_rooms: # Remove KVPoll.Success requests from the tracker if bootstrap_room not in self.request_status: self.addr_to_rooms_tracker[bootstrap_addr].discard( bootstrap_room ) else: logger.info( f"Attempting to reconnect to {bootstrap_addr}..." ) self.heartbeat_failures[bootstrap_addr] = ( self.heartbeat_failures.get(bootstrap_addr, 0) + 1 ) with self.session_pool_lock: if bootstrap_addr in self.session_pool: del self.session_pool[bootstrap_addr] except Exception: logger.info(f"Attempting to reconnect to {bootstrap_addr}...") self.heartbeat_failures[bootstrap_addr] = ( self.heartbeat_failures.get(bootstrap_addr, 0) + 1 ) if ( self.heartbeat_failures.get(bootstrap_addr, 0) >= self.max_failures ): self._handle_node_failure(bootstrap_addr) with self.session_pool_lock: if bootstrap_addr in self.session_pool: del self.session_pool[bootstrap_addr] threading.Thread(target=decode_thread).start() threading.Thread(target=heartbeat_checker).start() def add_transfer_request( self, bootstrap_room: int, kv_indices: npt.NDArray[np.int32], index_slice: slice, is_last: bool, aux_index: Optional[int] = None, state_indices: Optional[List[int]] = None, ): assert self.disaggregation_mode == DisaggregationMode.PREFILL assert not is_last or (is_last and aux_index is not None) if ( bootstrap_room not in self.request_status or self.check_status(bootstrap_room) == KVPoll.Failed ): logger.debug( "Request with bootstrap_room=%s already failed", bootstrap_room ) return if bootstrap_room not in self.transfer_infos: # This means that the current rank is a dummy rank for this request, # and it has already been marked as success, so there is no need to # add further chunks into the transfer queue. return # NOTE(shangming): sharding according to the dst_infos to make sure # requests with the same dst_sessions will be added into the same # queue, which enables early abort with failed sessions. dst_infos = self.transfer_infos[bootstrap_room].keys() session_port_sum = sum(int(session.rsplit(":", 1)[1]) for session in dst_infos) shard_idx = session_port_sum % len(self.transfer_queues) self.transfer_queues[shard_idx].put( TransferKVChunk( room=bootstrap_room, prefill_kv_indices=kv_indices, index_slice=index_slice, is_last=is_last, prefill_aux_index=aux_index, state_indices=state_indices, ) ) def check_status(self, bootstrap_room: int): return self.request_status[bootstrap_room] def update_status(self, bootstrap_room: int, status: KVPoll): if bootstrap_room not in self.request_status: self.request_status[bootstrap_room] = status else: # NOTE: status is only allowed to be incremented unless it is KVPoll.Failed if status == KVPoll.Failed: self.request_status[bootstrap_room] = KVPoll.Failed else: self.request_status[bootstrap_room] = max( self.request_status[bootstrap_room], status ) def record_failure(self, bootstrap_room: int, failure_reason: str): with self.failure_lock: self.failure_records[bootstrap_room] = failure_reason def get_session_id(self): return self.engine.get_session_id() def _handle_node_failure(self, failed_bootstrap_addr): with self.connection_lock: keys_to_remove = [ k for k in self.connection_pool if k.startswith(failed_bootstrap_addr) ] for k in keys_to_remove: del self.connection_pool[k] if failed_bootstrap_addr in self.prefill_attn_tp_size_table: del self.prefill_attn_tp_size_table[failed_bootstrap_addr] if failed_bootstrap_addr in self.prefill_dp_size_table: del self.prefill_dp_size_table[failed_bootstrap_addr] if failed_bootstrap_addr in self.prefill_pp_size_table: del self.prefill_pp_size_table[failed_bootstrap_addr] possible_affected_rooms = self.addr_to_rooms_tracker.get( failed_bootstrap_addr, [] ) if failed_bootstrap_addr in self.addr_to_rooms_tracker: del self.addr_to_rooms_tracker[failed_bootstrap_addr] # Report the requests associated with the failed bootstrap addr and mark their status as KVPoll.Failed affected_rooms = [] for room in possible_affected_rooms: if ( room in self.request_status and self.check_status(room) != KVPoll.Success ): self.record_failure( room, f"Losing connection with prefill instance (bootstrap_addr: {failed_bootstrap_addr})", ) self.update_status(room, KVPoll.Failed) affected_rooms.append(room) logger.error( f"Losing connection with prefill instance (bootstrap_addr: {failed_bootstrap_addr}), {len(affected_rooms)} requests affected" ) class MooncakeKVSender(CommonKVSender): def __init__( self, mgr: MooncakeKVManager, bootstrap_addr: str, bootstrap_room: int, dest_tp_ranks: List[int], pp_rank: int, ): super().__init__(mgr, bootstrap_addr, bootstrap_room, dest_tp_ranks, pp_rank) self.conclude_state = None self.init_time = time.time() def send( self, kv_indices: npt.NDArray[np.int32], state_indices: Optional[List[int]] = None, ): index_slice = slice(self.curr_idx, self.curr_idx + len(kv_indices)) self.curr_idx += len(kv_indices) is_last = self.curr_idx == self.num_kv_indices if not is_last: self.kv_mgr.add_transfer_request( self.bootstrap_room, kv_indices, index_slice, False, ) else: self.kv_mgr.add_transfer_request( self.bootstrap_room, kv_indices, index_slice, True, aux_index=self.aux_index, state_indices=state_indices, ) def poll(self) -> KVPoll: if self.conclude_state is None: status = self.kv_mgr.check_status(self.bootstrap_room) if status in (KVPoll.Success, KVPoll.Failed): self.conclude_state = status elif status == KVPoll.Bootstrapping: if self.init_time is not None: now = time.time() elapsed = now - self.init_time if elapsed >= self.kv_mgr.bootstrap_timeout: logger.warning_once( "Some requests timed out when bootstrapping, " "which means prefill instances fail to receive the KV indices from the decode instance of this request. " "If a greater mean TTFT is acceptable, you can 'export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=600' (10 minutes) to relax the timeout condition. " ) self.kv_mgr.record_failure( self.bootstrap_room, f"Request {self.bootstrap_room} timed out after {elapsed:.1f}s in KVPoll.Bootstrapping", ) self.conclude_state = KVPoll.Failed return KVPoll.Failed return status else: return self.conclude_state def clear(self) -> None: if self.bootstrap_room in self.kv_mgr.request_status: self.kv_mgr.request_status.pop(self.bootstrap_room) def failure_exception(self): # Explicitly set the status to failure since this request has failed in another rank if self.conclude_state is None: self.conclude_state = KVPoll.Failed self.clear() with self.kv_mgr.failure_lock: failure_reason = self.kv_mgr.failure_records.pop( self.bootstrap_room, "Failed due to an unknown reason from another rank" ) raise KVTransferError(self.bootstrap_room, failure_reason) def abort(self): self.kv_mgr.record_failure( self.bootstrap_room, "Aborted by AbortReq.", ) # Explicitly set the status to failure since this request has been aborted self.conclude_state = KVPoll.Failed class MooncakeKVReceiver(CommonKVReceiver): _ctx = zmq.Context() _socket_cache = {} _socket_locks = {} _global_lock = threading.Lock() def __init__( self, mgr: MooncakeKVManager, bootstrap_addr: str, bootstrap_room: Optional[int] = None, prefill_dp_rank: Optional[int] = None, ): self.session_id = mgr.get_session_id() self.conclude_state = None self.init_time = None super().__init__(mgr, bootstrap_addr, bootstrap_room, prefill_dp_rank) self.kv_mgr.addr_to_rooms_tracker[self.bootstrap_addr].add(self.bootstrap_room) self.kv_mgr.update_status(self.bootstrap_room, KVPoll.WaitingForInput) def _register_kv_args(self): for bootstrap_info in self.bootstrap_infos: 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 ) packed_state_data_ptrs = b"".join( struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.state_data_ptrs ) # Pack state_item_lens and state_dim_per_tensor for mamba state slice transfer packed_state_item_lens = b"".join( struct.pack("I", item_len) for item_len in self.kv_mgr.kv_args.state_item_lens ) state_dim_per_tensor = getattr( self.kv_mgr.kv_args, "state_dim_per_tensor", [] ) packed_state_dim_per_tensor = b"".join( struct.pack("I", dim) for dim in state_dim_per_tensor ) # Note(shangming): No need to add pp rank here since decode pp size should be equal to prefill pp size or 1 tp_rank = self.kv_mgr.kv_args.engine_rank kv_item_len = self.kv_mgr.kv_args.kv_item_lens[0] dst_tp_rank = str(tp_rank).encode("ascii") dst_attn_tp_size = str(self.kv_mgr.attn_tp_size).encode("ascii") dst_kv_item_len = str(kv_item_len).encode("ascii") sock, lock = self._connect_to_bootstrap_server(bootstrap_info) with lock: sock.send_multipart( [ "None".encode("ascii"), self.kv_mgr.local_ip.encode("ascii"), str(self.kv_mgr.rank_port).encode("ascii"), self.session_id.encode("ascii"), packed_kv_data_ptrs, packed_aux_data_ptrs, packed_state_data_ptrs, dst_tp_rank, dst_attn_tp_size, dst_kv_item_len, packed_state_item_lens, packed_state_dim_per_tensor, ] ) def init( self, kv_indices: npt.NDArray[np.int32], aux_index: Optional[int] = None, state_indices: Optional[List[int]] = None, ): if self.bootstrap_infos is None: self.kv_mgr.record_failure( self.bootstrap_room, f"Could not fetch prefill parallel info from bootstrap_addr: {self.bootstrap_addr}", ) self.kv_mgr.update_status(self.bootstrap_room, KVPoll.Failed) return for bootstrap_info in self.bootstrap_infos: sock, lock = self._connect_to_bootstrap_server(bootstrap_info) is_dummy = bootstrap_info["is_dummy"] with lock: sock.send_multipart( [ str(self.bootstrap_room).encode("ascii"), self.kv_mgr.local_ip.encode("ascii"), str(self.kv_mgr.rank_port).encode("ascii"), self.session_id.encode("ascii"), kv_indices.tobytes() if not is_dummy else b"", str(aux_index).encode("ascii") if not is_dummy else b"", ( np.array( state_indices, dtype=np.int32, ).tobytes() if not is_dummy and state_indices is not None else b"" ), str(self.required_dst_info_num).encode("ascii"), ] ) self.init_time = time.time() def poll(self) -> KVPoll: if self.conclude_state is None: status = self.kv_mgr.check_status(self.bootstrap_room) if status in (KVPoll.Success, KVPoll.Failed): self.conclude_state = status elif status == KVPoll.WaitingForInput: if self.init_time is not None: now = time.time() elapsed = now - self.init_time if elapsed >= self.kv_mgr.waiting_timeout: logger.warning_once( "Some requests fail to receive KV Cache transfer done signal after bootstrapping. " "If a greater mean TTFT is acceptable, you can 'export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=600' (10 minutes) to relax the timeout condition. " ) self.kv_mgr.record_failure( self.bootstrap_room, f"Request {self.bootstrap_room} timed out after {elapsed:.1f}s in KVPoll.WaitingForInput", ) self.conclude_state = KVPoll.Failed return KVPoll.Failed return status else: return self.conclude_state def clear(self) -> None: if self.bootstrap_room in self.kv_mgr.request_status: self.kv_mgr.request_status.pop(self.bootstrap_room) if self.bootstrap_room in self.kv_mgr.required_prefill_response_num_table: self.kv_mgr.required_prefill_response_num_table.pop(self.bootstrap_room) if self.bootstrap_room in self.kv_mgr.prefill_response_tracker: self.kv_mgr.prefill_response_tracker.pop(self.bootstrap_room) def failure_exception(self): # Explicitly set the status to failure since this request has failed in another rank if self.conclude_state is None: self.conclude_state = KVPoll.Failed self.clear() with self.kv_mgr.failure_lock: failure_reason = self.kv_mgr.failure_records.pop( self.bootstrap_room, "Failed due to an unknown reason from another rank" ) raise KVTransferError(self.bootstrap_room, failure_reason) def abort(self): self.kv_mgr.record_failure( self.bootstrap_room, "Aborted by AbortReq.", ) # Explicitly set the status to failure since this request has been aborted self.conclude_state = KVPoll.Failed class MooncakeKVBootstrapServer(CommonKVBootstrapServer): pass