Files
sglang/python/sglang/srt/managers/scheduler_dp_attn_mixin.py
T

202 lines
6.9 KiB
Python

from __future__ import annotations
from dataclasses import dataclass
from typing import TYPE_CHECKING, Callable
import torch
from sglang.srt.managers.schedule_batch import ScheduleBatch
from sglang.srt.two_batch_overlap import TboDPAttentionPreparer
from sglang.srt.utils.common import require_mlp_tp_gather
if TYPE_CHECKING:
from sglang.srt.distributed.parallel_state import GroupCoordinator
from sglang.srt.managers.scheduler import Scheduler
@dataclass
class MLPSyncBatchInfo:
dp_size: int
tp_size: int
num_tokens: int
num_tokens_for_logprob: int
can_cuda_graph: bool
is_extend_in_batch: bool
local_can_run_tbo: bool
local_forward_mode: int
# some gathered elements
tp0_info: torch.Tensor = None
global_num_tokens: list[int] = None
global_num_tokens_for_logprob: list[int] = None
tbo_split_seq_index: torch.Tensor = None
global_forward_mode: int = None
def _get_local_tensor(self, device, dtype=torch.int64) -> torch.Tensor:
return torch.tensor(
[
self.num_tokens,
self.num_tokens_for_logprob,
int(self.can_cuda_graph),
int(self.is_extend_in_batch),
int(self.local_can_run_tbo),
self.local_forward_mode,
],
device=device,
dtype=dtype,
)
def all_gather(self, device, group: torch.distributed.ProcessGroup):
local_info_tensor = self._get_local_tensor(device=device)
global_info_tensor = torch.empty(
(self.dp_size, self.tp_size, 6),
dtype=torch.int64,
device=device,
)
torch.distributed.all_gather_into_tensor(
global_info_tensor.flatten(),
local_info_tensor,
group=group,
)
tp0_info = global_info_tensor[:, 0, :]
self.tp0_info = tp0_info
self.global_num_tokens = tp0_info[:, 0].tolist()
self.global_num_tokens_for_logprob = tp0_info[:, 1].tolist()
self.can_cuda_graph = bool(tp0_info[:, 2].min().item())
self.is_extend_in_batch = bool(tp0_info[:, 3].max().item())
def _update_gather_batch(
batch: ScheduleBatch,
mlp_sync_info: MLPSyncBatchInfo,
require_mlp_tp_gather: bool,
):
# TODO: handle the case when moe_dense_tp_size != 1
if not require_mlp_tp_gather:
batch.global_num_tokens = [mlp_sync_info.num_tokens]
batch.global_num_tokens_for_logprob = [mlp_sync_info.num_tokens_for_logprob]
else:
batch.global_num_tokens = mlp_sync_info.global_num_tokens
batch.global_num_tokens_for_logprob = (
mlp_sync_info.global_num_tokens_for_logprob
)
batch.is_extend_in_batch = mlp_sync_info.is_extend_in_batch
batch.tbo_split_seq_index = mlp_sync_info.tbo_split_seq_index
batch.global_forward_mode = mlp_sync_info.global_forward_mode
# Check forward mode for cuda graph
batch.can_run_dp_cuda_graph = mlp_sync_info.can_cuda_graph
def prepare_mlp_sync_batch_raw(
local_batch: ScheduleBatch,
dp_size: int,
attn_tp_size: int,
tp_group: GroupCoordinator,
get_idle_batch: Callable[[], ScheduleBatch],
disable_cuda_graph: bool,
require_mlp_tp_gather: bool,
disable_overlap_schedule: bool,
offload_tags: set[str],
):
# Check if other DP workers have running batches
if local_batch is None or local_batch.forward_mode.is_prebuilt():
num_tokens = 0
num_tokens_for_logprob = 0
elif local_batch.forward_mode.is_decode():
num_tokens = local_batch.batch_size()
num_tokens_for_logprob = num_tokens
else:
num_tokens = local_batch.extend_num_tokens
if local_batch.return_logprob:
num_tokens_for_logprob = sum(
# We should have at least 1 token for sample in every case.
max(extend_len - logprob_start_len, 1)
for logprob_start_len, extend_len in zip(
local_batch.extend_logprob_start_lens,
local_batch.extend_lens,
)
)
else:
# When return_logprob = False, only need last token per request
num_tokens_for_logprob = local_batch.batch_size()
can_cuda_graph = (
local_batch is None
or local_batch.forward_mode.is_decode_or_idle()
or local_batch.forward_mode.is_prebuilt()
) and not disable_cuda_graph
is_extend_in_batch = local_batch.forward_mode.is_extend() if local_batch else False
tbo_preparer = TboDPAttentionPreparer()
if len(offload_tags) == 0 and disable_overlap_schedule:
group = tp_group.device_group
device = tp_group.device
else:
group = tp_group.cpu_group
device = "cpu"
local_can_run_tbo, local_forward_mode = tbo_preparer.prepare_all_gather(local_batch)
mlp_sync_info = MLPSyncBatchInfo(
dp_size=dp_size,
tp_size=attn_tp_size,
num_tokens=num_tokens,
num_tokens_for_logprob=num_tokens_for_logprob,
can_cuda_graph=can_cuda_graph,
is_extend_in_batch=is_extend_in_batch,
local_can_run_tbo=local_can_run_tbo,
local_forward_mode=local_forward_mode,
)
mlp_sync_info.all_gather(device=device, group=group)
mlp_sync_info.tbo_split_seq_index, mlp_sync_info.global_forward_mode = (
tbo_preparer.compute_output(
mlp_sync_info.tp0_info[:, 4:6],
)
)
need_idle_batch = max(mlp_sync_info.global_num_tokens) > 0
if need_idle_batch:
batch_to_gather = local_batch
if local_batch is None:
batch_to_gather = local_batch = get_idle_batch()
elif local_batch.forward_mode.is_prebuilt():
# NOTE: for prebuilt batch, we add an inner idle batch to run MLP sync
batch_to_gather = local_batch.inner_idle_batch = get_idle_batch()
_update_gather_batch(batch_to_gather, mlp_sync_info, require_mlp_tp_gather)
return local_batch
class SchedulerDPAttnMixin:
def prepare_mlp_sync_batch(self: Scheduler, local_batch: ScheduleBatch):
return prepare_mlp_sync_batch_raw(
local_batch,
dp_size=self.server_args.dp_size,
attn_tp_size=self.attn_tp_size,
tp_group=self.tp_group,
get_idle_batch=self.get_idle_batch,
disable_cuda_graph=self.server_args.disable_cuda_graph,
require_mlp_tp_gather=require_mlp_tp_gather(self.server_args),
disable_overlap_schedule=self.server_args.disable_overlap_schedule,
offload_tags=self.offload_tags,
)
def get_idle_batch(self: Scheduler) -> ScheduleBatch:
idle_batch = ScheduleBatch.init_new(
[],
self.req_to_token_pool,
self.token_to_kv_pool_allocator,
self.tree_cache,
self.model_config,
self.enable_overlap,
self.spec_algorithm,
)
idle_batch.prepare_for_idle()
return idle_batch