Files
wehub-resource-sync 59a0a3844c
PR Test AMD / cancel-on-close (push) Has been skipped
PR Test NVIDIA ARM / scan (push) Has been skipped
PR Test NVIDIA / cancel-on-close (push) Has been skipped
PR Test AMD / scan (push) Has been skipped
PR Test NVIDIA ARM / cancel-on-close (push) Has been skipped
PR Test NVIDIA / scan (push) Has been skipped
Release Docker Images / build (cu129-torch-2.11.0) (push) Has been skipped
Release Docker Images / build (cu130-torch-2.11.0) (push) Has been skipped
Release PyPI / publish (push) Has been skipped
Scheduler Python Test / test (push) Successful in 27m19s
Docs / build (push) Successful in 28m8s
Scheduler C++ Test / test (push) Successful in 28m19s
Scheduler C++ Test / test-flat (push) Successful in 28m18s
Docs / deploy (push) Has been cancelled
PR Test AMD / finish (push) Has been cancelled
PR Test NVIDIA / finish (push) Has been cancelled
PR Test NVIDIA ARM / finish (push) Has been cancelled
PR Test NVIDIA ARM / ${{ matrix.name }} (${{ matrix.runner }}) (push) Has been cancelled
PR Test AMD / ${{ matrix.name }} (${{ matrix.runner }}) (push) Has been cancelled
PR Test NVIDIA / ${{ matrix.name }} (${{ matrix.runner }}) (push) Has been cancelled
chore: import upstream snapshot with attribution
2026-07-13 12:32:31 +08:00

485 lines
17 KiB
Python
Executable File

# Copyright (c) 2026 LightSeek Foundation
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
"""Shared helpers for PD transfer runtime components."""
from __future__ import annotations
import ctypes
import dataclasses
import os
import random
import threading
import warnings
from collections import deque
from enum import Enum
from typing import TYPE_CHECKING
import numpy as np
import numpy.typing as npt
import requests
import torch
import torch.distributed as dist
from tokenspeed.runtime.layers.logits_processor import LogitsProcessorOutput
from tokenspeed.runtime.utils import get_colorful_logger
from tokenspeed.runtime.utils.network import get_ip
if TYPE_CHECKING:
from tokenspeed.runtime.engine.request import Req
# env var for testing failure, convert to float explicitly
FAILURE_PROB = float(os.getenv("DISAGGREGATION_TEST_FAILURE_PROB", 0))
logger = get_colorful_logger(__name__)
class DisaggregationMode(Enum):
NULL = "null"
PREFILL = "prefill"
DECODE = "decode"
ENCODE = "encode"
class FastQueue:
class Empty(Exception):
"""Exception raised when the queue is empty."""
pass
def __init__(self):
self._buf = deque()
self._cond = threading.Condition()
def put(self, item):
with self._cond:
self._buf.append(item)
self._cond.notify()
def get(self):
with self._cond:
while not self._buf:
self._cond.wait()
return self._buf.popleft()
def get_nowait(self):
with self._cond:
if not self._buf:
raise FastQueue.Empty()
return self._buf.popleft()
def poll_and_all_reduce(pollers, gloo_group):
"""Poll transfer state and all-reduce the result across the gloo group."""
# At a certain probability, mark the poll as failed to simulate failure.
if FAILURE_PROB > 0:
from tokenspeed.runtime.pd.base.status import TransferPoll
polls = [
(
int(TransferPoll.Failed)
if random.random() < FAILURE_PROB
else int(poller.poll())
)
for poller in pollers
]
else:
polls = [int(poller.poll()) for poller in pollers]
tensor_to_reduce = torch.tensor(polls, dtype=torch.uint8, device="cpu")
dist.all_reduce(tensor_to_reduce, op=dist.ReduceOp.MIN, group=gloo_group)
return tensor_to_reduce.tolist()
class ReqToMetadataIdxAllocator:
"""A memory pool that maps a request to its first output token location."""
def __init__(
self,
size: int,
):
self.size = size
self.free_slots = deque(range(size))
def available_size(self) -> int:
return len(self.free_slots)
def alloc(self) -> int | None:
if not self.free_slots:
return None
return self.free_slots.popleft()
def free(self, free_index: int) -> None:
self.free_slots.append(free_index)
class TransferBackend(Enum):
MOONCAKE = "mooncake"
MOONCAKE_ASYNC = "mooncake_async"
class KVClassType(Enum):
MANAGER_PREFILL = "manager_prefill"
MANAGER_DECODE = "manager_decode"
# The async backend uses one role-agnostic manager (see get_kv_class's
# MOONCAKE_ASYNC branch); the sync backend splits into the prefill/decode
# managers above.
MANAGER = "manager"
SENDER = "sender"
RECEIVER = "receiver"
BOOTSTRAP_SERVER = "bootstrap_server"
def get_kv_class(transfer_backend: TransferBackend, class_type: KVClassType):
if transfer_backend == TransferBackend.MOONCAKE:
from tokenspeed.runtime.pd.mooncake.conn import (
MooncakeKVBootstrapServer,
)
from tokenspeed.runtime.pd.mooncake.decode import (
MooncakeKVManagerDecode,
)
from tokenspeed.runtime.pd.mooncake.prefill import (
MooncakeKVManagerPrefill,
)
from tokenspeed.runtime.pd.mooncake.receiver import (
MooncakeKVReceiver,
)
from tokenspeed.runtime.pd.mooncake.sender import (
MooncakeKVSender,
)
class_mapping = {
KVClassType.MANAGER_PREFILL: MooncakeKVManagerPrefill,
KVClassType.MANAGER_DECODE: MooncakeKVManagerDecode,
KVClassType.SENDER: MooncakeKVSender,
KVClassType.RECEIVER: (MooncakeKVReceiver),
KVClassType.BOOTSTRAP_SERVER: MooncakeKVBootstrapServer,
}
return class_mapping.get(class_type)
if transfer_backend == TransferBackend.MOONCAKE_ASYNC:
from tokenspeed.runtime.pd.mooncake.async_conn import (
MooncakeAsyncKVManager,
)
from tokenspeed.runtime.pd.mooncake.conn import (
MooncakeKVBootstrapServer,
)
from tokenspeed.runtime.pd.mooncake.receiver import (
MooncakeKVReceiver,
)
from tokenspeed.runtime.pd.mooncake.sender import (
MooncakeKVSender,
)
class_mapping = {
KVClassType.MANAGER: MooncakeAsyncKVManager,
KVClassType.SENDER: MooncakeKVSender,
KVClassType.RECEIVER: (MooncakeKVReceiver),
KVClassType.BOOTSTRAP_SERVER: MooncakeKVBootstrapServer,
}
return class_mapping.get(class_type)
raise ValueError(f"Unsupported transfer backend: {transfer_backend}")
def kv_to_page_indices(kv_indices: np.ndarray, page_size: int):
# 1. The page is guaranteed to be full except the last page.
# 2. page index = kv_index // page_size
# The return vector is kv_indices[::page_size] // page_size
if page_size == 1: # shortcut
return kv_indices
return kv_indices[::page_size] // page_size
def kv_to_page_num(num_kv_indices: int, page_size: int):
# ceil(num_kv_indices / page_size)
return (num_kv_indices + page_size - 1) // page_size
@dataclasses.dataclass
class PDRegistryRequest:
"""A request to register a machine itself to the LB."""
mode: str
registry_url: str
bootstrap_port: int | None = None
def __post_init__(self):
if self.mode == "prefill" and self.bootstrap_port is None:
raise ValueError("Bootstrap port must be set in PREFILL mode.")
if self.mode == "decode" and self.bootstrap_port is not None:
raise ValueError("Bootstrap port must not be set in DECODE mode.")
if self.mode not in {"prefill", "decode"}:
raise ValueError(
f"Invalid mode: {self.mode}. Must be 'prefill' or 'decode'."
)
def register_disaggregation_server(
mode: str, server_port: int, bootstrap_port: int, pdlb_url: str
):
pdlb_url = pdlb_url.rstrip("/")
registered_bootstrap_port = bootstrap_port if mode == "prefill" else None
registry_request = PDRegistryRequest(
mode=mode,
registry_url=f"http://{get_ip()}:{server_port}",
bootstrap_port=registered_bootstrap_port,
)
res = requests.post(
f"{pdlb_url}/register",
json=dataclasses.asdict(registry_request),
)
if res.status_code != 200:
warnings.warn(
f"Failed to register disaggregation server: {res.status_code} {res.text}"
)
else:
logger.info(
"Registered disaggregation server with %s: status_code=%s text=%s",
pdlb_url,
res.status_code,
res.text,
)
def is_mla_backend(target_kv_pool) -> bool:
from tokenspeed.runtime.layers.attention.kv_cache.mla import MLATokenToKVPool
return isinstance(target_kv_pool, MLATokenToKVPool)
def prepare_abort(req: Req, error_message: str, status_code=None, err_type=None):
from tokenspeed.runtime.engine.request_types import ABORT_CODE, FINISH_ABORT
if err_type is None:
err_type = ABORT_CODE.UnknownError
# populate finish metadata and stream output
req.finished_reason = FINISH_ABORT(
error_message, status_code=status_code, err_type=err_type
)
if req.return_logprob:
req.input_token_logprobs_val = []
req.input_token_logprobs_idx = []
req.input_top_logprobs_val = []
req.input_top_logprobs_idx = []
req.input_token_ids_logprobs_val = []
req.input_token_ids_logprobs_idx = []
class MetadataBuffers:
def __init__(self, size: int, max_top_logprobs_num: int = 128, device: str = "cpu"):
# We transfer the metadata of first output token to decode
# The minimal size for RDMA is 64Bytes, so we pad it to > 64Bytes
self.device = device
self.output_ids = torch.zeros((size, 16), dtype=torch.int32, device=device)
self.output_token_logprobs_val = torch.zeros(
(size, 16), dtype=torch.float32, device=device
)
self.output_token_logprobs_idx = torch.zeros(
(size, 16), dtype=torch.int32, device=device
)
self.output_top_logprobs_val = torch.zeros(
(size, max_top_logprobs_num), dtype=torch.float32, device=device
)
self.output_top_logprobs_idx = torch.zeros(
(size, max_top_logprobs_num), dtype=torch.int32, device=device
)
self.cached_tokens = torch.zeros((size, 1), dtype=torch.int32, device=device)
def get_buf_infos(self):
ptrs = [
self.output_ids.data_ptr(),
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.cached_tokens.data_ptr(),
]
data_lens = [
self.output_ids.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.cached_tokens.nbytes,
]
item_lens = [
self.output_ids[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.cached_tokens[0].nbytes,
]
return ptrs, data_lens, item_lens
def get_buf(self, idx: int):
return (
self.output_ids[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.cached_tokens[idx],
)
def set_buf_by_batch(
self,
output_ids: torch.Tensor,
output_buffer_indices: list[int],
logits_output: LogitsProcessorOutput,
output_token_logprobs_indices: list[tuple[int, int]] | None = None,
output_top_logprobs_indices: list[tuple[int, int]] | None = None,
cached_tokens: torch.Tensor | None = None,
):
self.output_ids[
torch.tensor(output_buffer_indices).to(self.device, non_blocking=True), 0
] = output_ids
if cached_tokens is not None:
self.cached_tokens[
torch.tensor(output_buffer_indices).to(self.device, non_blocking=True),
0,
] = cached_tokens
if output_token_logprobs_indices:
for src_idx, dst_idx in output_token_logprobs_indices:
self.output_token_logprobs_val[dst_idx][0] = (
logits_output.next_token_top_logprobs_val[src_idx]
)
self.output_token_logprobs_idx[dst_idx][0] = (
logits_output.next_token_top_logprobs_idx[src_idx]
)
if output_top_logprobs_indices:
for src_idx, dst_idx in output_top_logprobs_indices:
self.output_top_logprobs_val[dst_idx][
: len(logits_output.next_token_top_logprobs_val[src_idx][0])
] = torch.tensor(
logits_output.next_token_top_logprobs_val[src_idx][0],
dtype=torch.float32,
device="cpu",
)
self.output_top_logprobs_idx[dst_idx][
: len(logits_output.next_token_top_logprobs_idx[src_idx][0])
] = torch.tensor(
logits_output.next_token_top_logprobs_idx[src_idx][0],
dtype=torch.int32,
device="cpu",
)
def set_buf(self, req: Req):
self.output_ids[req.metadata_buffer_index][0] = req.output_ids[0]
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] = (
req.output_token_logprobs_val[0]
)
if req.output_token_logprobs_idx: # not none or empty list
self.output_token_logprobs_idx[req.metadata_buffer_index][0] = (
req.output_token_logprobs_idx[0]
)
if req.output_top_logprobs_val: # not none or empty list
self.output_top_logprobs_val[req.metadata_buffer_index][
: len(req.output_top_logprobs_val[0])
] = torch.tensor(
req.output_top_logprobs_val[0], dtype=torch.float32, device="cpu"
)
if req.output_top_logprobs_idx: # not none or empty list
self.output_top_logprobs_idx[req.metadata_buffer_index][
: len(req.output_top_logprobs_idx[0])
] = torch.tensor(
req.output_top_logprobs_idx[0], dtype=torch.int32, device="cpu"
)
def group_concurrent_contiguous(
src_indices: npt.NDArray[np.int64], dst_indices: npt.NDArray[np.int64]
) -> tuple[list[npt.NDArray[np.int64]], list[npt.NDArray[np.int64]]]:
"""Vectorised NumPy implementation."""
if src_indices.size == 0:
return [], []
brk = np.where((np.diff(src_indices) != 1) | (np.diff(dst_indices) != 1))[0] + 1
src_groups = np.split(src_indices, brk)
dst_groups = np.split(dst_indices, brk)
src_groups = [g.tolist() for g in src_groups]
dst_groups = [g.tolist() for g in dst_groups]
return src_groups, dst_groups
class StepCounter:
COUNT_NUM_MAX: int = 2**62
@classmethod
def is_step_ready(cls, current_step: int, target_step: int) -> bool:
# because COUNT_NUM_MAX is very large, we can make sure that if diff is > COUNT_NUM_MAX / 2 means the flush is finished
# and if the current_sent_count == task_stop_count also means the flush is not finished
# so if current_sent_count != task_stop_count and diff < COUNT_NUM_MAX / 2, the flush is not finished
return (
target_step != current_step
and (target_step + cls.COUNT_NUM_MAX - current_step) % cls.COUNT_NUM_MAX
> cls.COUNT_NUM_MAX / 2
)
def __init__(self, device: str, gpu_id: int):
# utilities for cache step
self.d_ready_cache_step = torch.tensor(0, dtype=torch.int64).cuda(gpu_id)
self.h_ready_cache_step = torch.tensor(0, dtype=torch.int64, pin_memory=True)
self.cache_step: int = 0
# utilities for aux step
self.d_ready_aux_step = torch.tensor(0, dtype=torch.int64).cuda(gpu_id)
self.h_ready_aux_step = torch.tensor(0, dtype=torch.int64, pin_memory=True)
self.aux_step: int = 0
def current_step(self) -> tuple[int, int]:
return self.cache_step, self.aux_step
def advance_step(self, delta_cache_step: int, delta_aux_step: int):
self.cache_step = (self.cache_step + delta_cache_step) % self.COUNT_NUM_MAX
self.aux_step = (self.aux_step + delta_aux_step) % self.COUNT_NUM_MAX
def record_cache(self):
self.d_ready_cache_step = (self.d_ready_cache_step + 1) % self.COUNT_NUM_MAX
self.h_ready_cache_step.copy_(self.d_ready_cache_step, non_blocking=True)
def record_aux(self):
self.d_ready_aux_step = (self.d_ready_aux_step + 1) % self.COUNT_NUM_MAX
self.h_ready_aux_step.copy_(self.d_ready_aux_step, non_blocking=True)
def query_ready_cache_step(self) -> int:
return ctypes.c_int64.from_address(self.h_ready_cache_step.data_ptr()).value
def query_ready_aux_step(self) -> int:
return ctypes.c_int64.from_address(self.h_ready_aux_step.data_ptr()).value
@dataclasses.dataclass
class PageTransferMetadata:
indices_are_local: bool
page_transfer_mask: npt.NDArray[np.bool_]
page_local_indices: npt.NDArray[np.int64] | None = None