Skip to content

GitLab

Commit 73032f48 authored by xuxz's avatar xuxz
Browse files

[PD]回退p2pncclconnector

parent a997359c
Expand all Show whitespace changes
Inline Side-by-side
Showing with 540 additions and 48 deletions
vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_engine.py
View file @ 73032f48
......@@ -13,6 +13,7 @@ from typing import TYPE_CHECKING, Any, Optional
import msgpack
import torch
import zmq
import regex
from vllm.config import KVTransferConfig
from vllm.distributed.device_communicators.pynccl_wrapper import (
......@@ -20,6 +21,13 @@ from vllm.distributed.device_communicators.pynccl_wrapper import (
from vllm.distributed.kv_transfer.kv_connector.v1.p2p.tensor_memory_pool import ( # noqa: E501
TensorMemoryPool)
from vllm.utils import current_stream, get_ip
from vllm import envs
from vllm.distributed.parallel_state import get_pp_group, get_tp_group
from dataclasses import dataclass
from vllm.model_executor.models.utils import extract_layer_index
from vllm.distributed.utils import get_pp_indices
from vllm.config import ModelConfig
if TYPE_CHECKING:
from vllm.forward_context import ForwardContext
......@@ -28,6 +36,11 @@ logger = logging.getLogger(__name__)
DEFAULT_MEM_POOL_SIZE_GB = 32
# @dataclass
# class SendQueueItem:
# tensor_id: str
# remote_address: str
# tensor: torch.Tensor
@contextmanager
def set_p2p_nccl_context(num_channels: str):
......@@ -59,17 +72,37 @@ def set_p2p_nccl_context(num_channels: str):
os.environ.pop(var, None)
@dataclass
class RemoteAddr:
pd_pair_id: str = ""
zmq_address: str = ""
comm_rank: int = 0
class P2pNcclEngine:
def __init__(self,
local_rank: int,
port_offset: int,
config: KVTransferConfig,
hostname: str = "",
port_offset: int = 0,
model_config: ModelConfig,
dp_rank: int = 0,
pp_rank: int = 0,
tp_rank: int = 0,
dp_size: int = 0,
pp_size: int = 0,
tp_size: int = 0,
library_path: Optional[str] = None) -> None:
self.config = config
self.model_config = model_config
self.rank = port_offset
self.local_rank = local_rank
self.dp_rank = dp_rank
self.pp_rank = pp_rank
self.tp_rank = tp_rank
self.dp_size = dp_size
self.pp_size = pp_size
self.tp_size = tp_size
self.device = torch.device(f"cuda:{self.local_rank}")
self.nccl = NCCLLibrary(library_path)
......@@ -95,7 +128,7 @@ class P2pNcclEngine:
port = int(self.config.kv_port) + port_offset
if port == 0:
raise ValueError("Port cannot be 0")
self._hostname = hostname
self._hostname = get_ip()
self._port = port
# Each card corresponds to a ZMQ address.
......@@ -129,6 +162,10 @@ class P2pNcclEngine:
self.send_stream = torch.cuda.Stream()
self.recv_stream = torch.cuda.Stream()
self.p2p_async_kv_tokens = envs.VLLM_P2P_BUF_TOKENS
self.p2p_async_buf = None
self.tensor_split_num: int = 0
mem_pool_size_gb = self.config.get_from_extra_config(
"mem_pool_size_gb", DEFAULT_MEM_POOL_SIZE_GB)
self.pool = TensorMemoryPool(max_block_size=int(mem_pool_size_gb) *
......@@ -167,11 +204,16 @@ class P2pNcclEngine:
self._listener_thread.start()
self._ping_thread = None
if self.multiple_machines:
if port_offset == 0 and self.proxy_address != "":
self._ping_thread = threading.Thread(target=self._ping,
daemon=True)
self._ping_thread.start()
else:
if self.proxy_address != "":
self._ping_thread = threading.Thread(target=self._ping_new,
daemon=True)
self._ping_thread.start()
logger.info(
"💯P2pNcclEngine init, rank:%d, local_rank:%d, http_address:%s, "
"zmq_address:%s, proxy_address:%s, send_type:%s, buffer_size_"
......@@ -179,6 +221,21 @@ class P2pNcclEngine:
self.http_address, self.zmq_address, self.proxy_address,
self.send_type, self.buffer_size_threshold, self.nccl_num_channels)
def _create_connect_new(self, remote_address: typing.Optional[str] = None):
assert remote_address is not None
if remote_address not in self.socks:
sock = self.context.socket(zmq.DEALER)
sock.setsockopt(zmq.SNDHWM, 10000)
sock.setsockopt(zmq.RCVHWM, 5000)
sock.setsockopt(zmq.LINGER, 0)
sock.setsockopt(zmq.TCP_KEEPALIVE, 1)
sock.setsockopt_string(zmq.IDENTITY, f"P-{self.zmq_address}")
sock.connect(f"tcp://{remote_address}")
self.socks[remote_address] = sock
return self.socks[remote_address]
def _create_connect(self, remote_address: typing.Optional[str] = None):
assert remote_address is not None
if remote_address not in self.socks:
......@@ -206,11 +263,73 @@ class P2pNcclEngine:
return self.socks[remote_address], self.comms[remote_address]
def get_send_queue_items(self, request_id: str, layer_name: str,
tensor: torch.Tensor,
is_mla: bool) -> list[any]:
tensor_id = self.get_tensor_id(request_id, layer_name)
remote_ip, remote_port = self.parse_request_id(request_id, True)
p_ip, p_port = self.parse_request_id(request_id, False)
pd_pair_id = p_ip + ":" + str(p_port) + "_" + remote_ip + ":" + str(remote_port)
if not self.enable_asymmetric_p2p:
remote_address = remote_ip + ":" + str(remote_port + self.rank)
remote_addr = RemoteAddr(pd_pair_id, remote_address, self.rank + self.pp_size * self.tp_size)
# logger.info(f"""+++++xiabo tensor_id:{tensor_id} request_id:{request_id} remote_address:{remote_address}""")
return [(tensor_id, remote_addr, tensor)]
if not is_mla:
logger.error(" P2PNCCL only support mla model symmetric PP/TP!!!!")
remote_pp_rank = self.compute_remote_pp_rank(layer_name)
items: list[Any] = []
for d_tp_rank in range(self.remote_tp_size):
for mul_tp in range(self.multp):
if self.tp_rank + mul_tp * self.tp_size == d_tp_rank:
remote_port_offset = remote_pp_rank * self.remote_tp_size + d_tp_rank
remote_address = remote_ip + ":" + str(remote_port + remote_port_offset)
remote_addr = RemoteAddr(pd_pair_id, remote_address, remote_port_offset + self.pp_size * self.tp_size)
logger.debug(
"Wait to send::%s, tensor_shape:%s, "
"(pp=%d, tp=%d) -> remote_address=%s(pp=%d, tp=%d) comm_rank (%d -> %d)", tensor_id,
tensor.shape, self.pp_rank, self.tp_rank, remote_address,
remote_pp_rank, self.rank * mul_tp + self.rank, self.rank, remote_port_offset + self.pp_size * self.tp_size)
items.append([tensor_id, remote_addr, tensor])
return items
def send_tensor_new(
self,
request_id: str,
layer_name: str,
tensor: torch.Tensor,
is_mla: bool = False,
) -> bool:
tensor_id = self.get_tensor_id(request_id, layer_name)
if self.send_type == "PUT":
return all(
self._send_sync_new(item) for item in self.get_send_queue_items(
request_id, layer_name, tensor, is_mla))
if self.send_type == "PUT_ASYNC":
with self.send_queue_cv:
for item in self.get_send_queue_items(request_id, layer_name,
tensor, is_mla):
self.send_queue.append(item)
self.send_queue_cv.notify()
return True
if self.send_type == "GET":
logger.error(" P2PNCCL new not support GET model, please set VLLM_P2PNCCL_NEW=0 use defalut model!!!!")
def send_tensor(
self,
tensor_id: str,
tensor: torch.Tensor,
remote_address: typing.Optional[str] = None,
tbo_evt = None,
) -> bool:
if remote_address is None:
with self.recv_store_cv:
......@@ -251,6 +370,53 @@ class P2pNcclEngine:
return True
def p2p_async_send_tensor(
self,
tensor_id: str,
tensor: torch.Tensor,
remote_address: typing.Optional[str] = None,
tbo_evt = None,
) -> bool:
if remote_address is None:
with self.recv_store_cv:
self.recv_store[tensor_id] = tensor
self.recv_store_cv.notify()
return True
else:
if self.send_type == "PUT":
return self._send_sync(tensor_id, tensor, remote_address)
elif self.send_type == "PUT_ASYNC":
with self.send_queue_cv:
kv_layer, slot_mapping = tensor # tesor (kv_layer, slot_mapping)
self.send_queue.append([tensor_id, remote_address, kv_layer, slot_mapping, tbo_evt])
self.send_queue_cv.notify()
else: # GET
with self.send_store_cv:
tensor_size = tensor.element_size() * tensor.numel()
while (self.buffer_size + tensor_size
> self.buffer_size_threshold):
oldest_tenser_id = next(iter(self.send_store))
oldest_tenser = self.send_store.pop(oldest_tenser_id)
oldest_tenser_size = oldest_tenser.element_size(
) * oldest_tenser.numel()
self.buffer_size -= oldest_tenser_size
logger.info(
"⛔[GET]Send to %s, tensor_id:%s, tensor_size:%d,"
" buffer_size:%d, oldest_tenser_size:%d, rank:%d",
remote_address, tensor_id, tensor_size,
self.buffer_size, oldest_tenser_size, self.rank)
self.send_store[tensor_id] = tensor
self.buffer_size += tensor_size
logger.debug(
"🔵[GET]Send to %s, tensor_id:%s, tensor_size:%d, "
"shape:%s, rank:%d, buffer_size:%d(%.2f%%)",
remote_address, tensor_id, tensor_size, tensor.shape,
self.rank, self.buffer_size,
self.buffer_size / self.buffer_size_threshold * 100)
return True
def recv_tensor(
self,
tensor_id: str,
......@@ -327,6 +493,8 @@ class P2pNcclEngine:
self.zmq_address, remote_address.decode(), rank)
elif data["cmd"] == "PUT":
tensor_id = data["tensor_id"]
if "tensor_split_num" in data:
self.tensor_split_num = data["tensor_split_num"]
try:
with torch.cuda.stream(self.recv_stream):
tensor = torch.empty(data["shape"],
......@@ -343,10 +511,6 @@ class P2pNcclEngine:
# Store Tensor in memory pool
addr = self.pool.store_tensor(tensor)
tensor = (addr, tensor.dtype, tensor.shape)
logger.warning(
"🔴[PUT]Recv Tensor, Out Of Threshold, "
"%s👈%s, data:%s, addr:%d", self.zmq_address,
remote_address.decode(), data, addr)
else:
self.buffer_size += tensor_size
......@@ -363,7 +527,56 @@ class P2pNcclEngine:
self.recv_store[tensor_id] = tensor
self._have_received_tensor_id(tensor_id)
self.recv_store_cv.notify()
elif data["cmd"] == "PUT_NEW":
tensor_id = data["tensor_id"]
if "tensor_split_num" in data:
self.tensor_split_num = data["tensor_split_num"]
try:
with torch.cuda.stream(self.recv_stream):
tensor = torch.empty(data["shape"],
dtype=getattr(
torch, data["dtype"]),
device=self.device)
self.router_socket.send_multipart(
[remote_address, b"0"])
# comm, rank = self.comms[remote_address.decode()]
# self._recv(comm, tensor, rank ^ 1, self.recv_stream)
comm, rank = self.comms[data["pd_pair_id"]]
self._recv(comm, tensor, int(data["comm_rank"]), self.recv_stream)
tensor_size = tensor.element_size() * tensor.numel()
if (self.buffer_size + tensor_size
> self.buffer_size_threshold):
# Store Tensor in memory pool
addr = self.pool.store_tensor(tensor)
tensor = (addr, tensor.dtype, tensor.shape)
else:
self.buffer_size += tensor_size
except torch.cuda.OutOfMemoryError:
self.router_socket.send_multipart(
[remote_address, b"1"])
tensor = None
logger.warning(
"🔴[PUT]Recv Tensor, Out Of Memory, %s👈%s, "
"data:%s", self.zmq_address,
remote_address.decode(), data)
with self.recv_store_cv:
self.recv_store[tensor_id] = tensor
self._have_received_tensor_id(tensor_id)
self.recv_store_cv.notify()
elif data["cmd"] == "comm_init":
unique_id = self.nccl.unique_id_from_bytes(
bytes(data["unique_id"]))
with torch.cuda.device(self.device):
rank = int(data["rank"])
world_size = int(data["world_size"])
with set_p2p_nccl_context(self.nccl_num_channels):
comm: ncclComm_t = self.nccl.ncclCommInitRank(
world_size, unique_id, rank)
self.comms[data["pd_pair_id"]] = (comm, rank)
logger.info(
"🤝ncclCommInitRank Success, %s👈%s, MyRank:%s",
self.zmq_address, data["pd_pair_id"], rank)
elif data["cmd"] == "GET":
tensor_id = data["tensor_id"]
with self.send_store_cv:
......@@ -410,10 +623,21 @@ class P2pNcclEngine:
with self.send_queue_cv:
while not self.send_queue:
self.send_queue_cv.wait()
if envs.VLLM_ENABLE_TBO or envs.VLLM_P2P_ASYNC:
tensor_id, remote_address, kv_layer, slot_mapping, tbo_evt = self.send_queue.popleft()
else:
tensor_id, remote_address, tensor = self.send_queue.popleft()
if not self.send_queue:
self.send_queue_cv.notify()
if (envs.VLLM_ENABLE_TBO or envs.VLLM_P2P_ASYNC) and tbo_evt is not None:
self.send_stream.wait_event(tbo_evt)
self._send_kv_p2p_sync(tensor_id, kv_layer, slot_mapping, remote_address)
else:
if self.multiple_machines:
self._send_sync(tensor_id, tensor, remote_address)
else:
# logger.info(f"""=============xiabo tensor_id:{tensor_id} remote_address:{remote_address}""")
self._send_sync_new(tensor_id, tensor, remote_address)
def wait_for_sent(self):
if self.send_type == "PUT_ASYNC":
......@@ -518,7 +742,7 @@ class P2pNcclEngine:
"pd_pair_id": remote_address.pd_pair_id,
"comm_rank": rank
}
# logger.info(f"""_send_sync_new:{data}""")
logger.info(f"""_send_sync_new:{data}""")
sock.send(msgpack.dumps(data))
response = sock.recv()
......@@ -627,6 +851,36 @@ class P2pNcclEngine:
sock.send(msgpack.dumps(data))
time.sleep(3)
def _ping_new(self):
sock = self.context.socket(zmq.DEALER)
sock.setsockopt_string(zmq.IDENTITY, self.zmq_address)
logger.debug("ping start, zmq_address:%s", self.zmq_address)
sock.connect(f"tcp://{self.proxy_address}")
if self.rank == 0:
data = {
"type": "P_init" if self.config.is_kv_producer else "D_init",
"http_address": self.http_address,
"zmq_address": self.zmq_address,
"dp_size" : self.dp_size,
"pp_size" : self.pp_size,
"tp_size" : self.tp_size
}
# logger.info(f"""_ping data:{data}""")
sock.send(msgpack.dumps(data))
data = {
"type": "P" if self.config.is_kv_producer else "D",
"http_address": self.http_address,
"dp_rank" : self.dp_rank,
"pp_rank" : self.pp_rank,
"tp_rank" : self.tp_rank,
"zmq_address": self.zmq_address
}
# while True:
# logger.info(f"""_ping data:{data}""")
sock.send(msgpack.dumps(data))
# time.sleep(3)
def _send(self, comm, tensor: torch.Tensor, dst: int, stream=None):
assert tensor.device == self.device, (
f"this nccl communicator is created to work on {self.device}, "
......
vllm/distributed/kv_transfer/kv_connector/v1/p2p/tensor_memory_pool.py
View file @ 73032f48
......@@ -63,7 +63,7 @@ class TensorMemoryPool:
than min_block_size
"""
def __init__(self, max_block_size: int, min_block_size: int = 512):
def __init__(self, max_block_size: int, min_block_size: int = 128):
if max_block_size <= 0 or min_block_size <= 0:
raise ValueError("Block sizes must be positive")
if max_block_size < min_block_size:
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment