621 lines
21 KiB
Python
621 lines
21 KiB
Python
# Copyright 2023-2024 SGLang Team
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# ==============================================================================
|
|
import logging
|
|
import os
|
|
import time
|
|
from abc import ABC
|
|
from contextlib import contextmanager
|
|
from pathlib import Path
|
|
from typing import Dict, List, Literal, Optional, Tuple, Type
|
|
|
|
import torch
|
|
import torch.distributed
|
|
|
|
from sglang.srt.managers.expert_location import ExpertLocationMetadata
|
|
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
|
|
from sglang.srt.server_args import ServerArgs
|
|
from sglang.srt.utils import Withable
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
# --------------------------------------- Entrypoint -----------------------------------------
|
|
|
|
_OutputMode = Literal["file", "object"]
|
|
|
|
|
|
class ExpertDistributionRecorder(ABC):
|
|
"""Global expert distribution recording"""
|
|
|
|
@staticmethod
|
|
def init_new(
|
|
server_args: ServerArgs,
|
|
expert_location_metadata: "ExpertLocationMetadata",
|
|
rank: int,
|
|
):
|
|
if server_args.expert_distribution_recorder_mode is not None:
|
|
return _ExpertDistributionRecorderReal(
|
|
server_args, expert_location_metadata, rank
|
|
)
|
|
else:
|
|
return _ExpertDistributionRecorderNoop()
|
|
|
|
@contextmanager
|
|
def with_current_layer(self, layer_idx):
|
|
yield
|
|
|
|
@contextmanager
|
|
def with_debug_name(self, debug_name):
|
|
yield
|
|
|
|
@contextmanager
|
|
def with_forward_pass(self, forward_pass_id: int, forward_batch: ForwardBatch):
|
|
yield
|
|
|
|
def on_select_experts(self, topk_ids: torch.Tensor):
|
|
pass
|
|
|
|
def on_deepep_dispatch_normal(
|
|
self,
|
|
local_physical_count_of_layer: List[int],
|
|
num_tokens_per_rank,
|
|
num_tokens_per_rdma_rank,
|
|
num_tokens_per_expert,
|
|
):
|
|
pass
|
|
|
|
def on_deepep_dispatch_low_latency(
|
|
self, local_physical_count_of_layer: torch.Tensor
|
|
):
|
|
pass
|
|
|
|
def start_record(self):
|
|
self._on_not_implemented()
|
|
|
|
def stop_record(self):
|
|
self._on_not_implemented()
|
|
|
|
def dump_record(self, output_mode: _OutputMode = "file"):
|
|
self._on_not_implemented()
|
|
|
|
def _on_not_implemented(self):
|
|
raise Exception(
|
|
"Please set ServerArgs.expert_distribution_recorder_mode to use ExpertDistributionRecorder."
|
|
)
|
|
|
|
|
|
class _ExpertDistributionRecorderNoop(ExpertDistributionRecorder):
|
|
pass
|
|
|
|
|
|
class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
|
|
def __init__(
|
|
self,
|
|
server_args: ServerArgs,
|
|
expert_location_metadata: "ExpertLocationMetadata",
|
|
rank: int,
|
|
):
|
|
self._server_args = server_args
|
|
self._expert_location_metadata = expert_location_metadata
|
|
|
|
self._recording = False
|
|
self._current_forward_pass_id = Withable()
|
|
self._current_layer_idx = Withable()
|
|
self._current_debug_name = Withable()
|
|
self._accumulator = _Accumulator.init_new(
|
|
server_args, expert_location_metadata, rank
|
|
)
|
|
self._single_pass_gatherers = {
|
|
k: _SinglePassGatherer.init_new(server_args, expert_location_metadata, rank)
|
|
for k in self._accumulator.get_single_pass_gatherer_keys()
|
|
}
|
|
|
|
def with_current_layer(self, layer_idx):
|
|
return self._current_layer_idx.with_value(layer_idx)
|
|
|
|
def with_debug_name(self, debug_name):
|
|
return self._current_debug_name.with_value(debug_name)
|
|
|
|
@contextmanager
|
|
def with_forward_pass(self, forward_pass_id: int, forward_batch: ForwardBatch):
|
|
with self._current_forward_pass_id.with_value(forward_pass_id):
|
|
self._on_forward_pass_start(forward_batch)
|
|
try:
|
|
yield
|
|
finally:
|
|
self._on_forward_pass_end(forward_pass_id)
|
|
|
|
def _on_forward_pass_start(self, forward_batch: ForwardBatch):
|
|
if not self._recording:
|
|
return
|
|
for gatherer_key, gatherer in self._single_pass_gatherers.items():
|
|
gatherer.reset()
|
|
gatherer.on_forward_pass_start(forward_batch)
|
|
|
|
def _on_forward_pass_end(self, forward_pass_id: int):
|
|
if not self._recording:
|
|
return
|
|
for gatherer_key, gatherer in self._single_pass_gatherers.items():
|
|
single_pass_data = gatherer.collect()
|
|
self._accumulator.append(forward_pass_id, gatherer_key, single_pass_data)
|
|
|
|
def on_select_experts(self, topk_ids: torch.Tensor):
|
|
self._on_hook("on_select_experts", topk_ids=topk_ids)
|
|
|
|
def on_deepep_dispatch_normal(
|
|
self,
|
|
local_physical_count_of_layer: List[int],
|
|
num_tokens_per_rank,
|
|
num_tokens_per_rdma_rank,
|
|
num_tokens_per_expert,
|
|
):
|
|
self._on_hook(
|
|
"on_deepep_dispatch_normal",
|
|
local_physical_count_of_layer=local_physical_count_of_layer,
|
|
num_tokens_per_rank=num_tokens_per_rank,
|
|
num_tokens_per_rdma_rank=num_tokens_per_rdma_rank,
|
|
num_tokens_per_expert=num_tokens_per_expert,
|
|
)
|
|
|
|
def on_deepep_dispatch_low_latency(
|
|
self, local_physical_count_of_layer: torch.Tensor
|
|
):
|
|
self._on_hook(
|
|
"on_deepep_dispatch_low_latency",
|
|
local_physical_count_of_layer=local_physical_count_of_layer,
|
|
)
|
|
|
|
def _on_hook(self, hook_name: str, **kwargs):
|
|
if not (self._recording or torch.cuda.is_current_stream_capturing()):
|
|
return
|
|
gatherer = self._single_pass_gatherers[
|
|
self._accumulator.get_single_pass_gatherer_key(
|
|
self._current_debug_name.value
|
|
)
|
|
]
|
|
getattr(gatherer, hook_name)(layer_idx=self._current_layer_idx.value, **kwargs)
|
|
|
|
def _reset(self):
|
|
"""Reset the expert distribution recorder."""
|
|
logger.info("Resetting ExpertDistributionRecorder...")
|
|
assert (
|
|
self._current_layer_idx.value is None
|
|
), f"{self._current_layer_idx.value=}"
|
|
for gatherer in self._single_pass_gatherers.values():
|
|
gatherer.reset()
|
|
self._accumulator.reset()
|
|
|
|
def start_record(self):
|
|
"""Start recording the expert distribution."""
|
|
if self._recording:
|
|
logger.warning(
|
|
"SGLang server is already recording expert ids. Did you forget to dump the expert ids recorded so far by sending requests to the `/stop_expert_distribution_record` and `/dump_expert_distribution_record` endpoints?"
|
|
)
|
|
self._reset()
|
|
self._recording = True
|
|
|
|
def stop_record(self):
|
|
"""Stop recording the expert distribution."""
|
|
if not self._recording:
|
|
logger.warning(
|
|
"SGLang server has not been recording expert ids. Did you forget to start recording by sending request to the `/start_expert_distribution_record` endpoint?"
|
|
)
|
|
self._recording = False
|
|
|
|
def dump_record(self, output_mode: _OutputMode = "file"):
|
|
"""Dump the expert distribution record and reset the recorder after dumping."""
|
|
output = self._accumulator.dump(output_mode=output_mode)
|
|
self._reset()
|
|
return output
|
|
|
|
|
|
_global_expert_distribution_recorder: Optional[ExpertDistributionRecorder] = (
|
|
_ExpertDistributionRecorderNoop()
|
|
)
|
|
|
|
|
|
def get_global_expert_distribution_recorder():
|
|
return _global_expert_distribution_recorder
|
|
|
|
|
|
def set_global_expert_distribution_recorder(value):
|
|
global _global_expert_distribution_recorder
|
|
_global_expert_distribution_recorder = value
|
|
|
|
|
|
# --------------------------------------- SinglePassGatherer -----------------------------------------
|
|
|
|
|
|
class _SinglePassGatherer(ABC):
|
|
@staticmethod
|
|
def init_new(
|
|
server_args: ServerArgs,
|
|
expert_location_metadata: "ExpertLocationMetadata",
|
|
rank: int,
|
|
) -> "_SinglePassGatherer":
|
|
if server_args.expert_distribution_recorder_mode == "per_token":
|
|
return _DetailSinglePassGatherer(
|
|
server_args, expert_location_metadata, rank
|
|
)
|
|
if server_args.enable_deepep_moe:
|
|
if server_args.deepep_mode == "normal":
|
|
return _DeepepNormalSinglePassGatherer(expert_location_metadata, rank)
|
|
elif server_args.deepep_mode == "low_latency":
|
|
return _DeepepLowLatencySinglePassGatherer(
|
|
expert_location_metadata, rank
|
|
)
|
|
else:
|
|
raise NotImplementedError
|
|
return _SelectExpertsSinglePassGatherer(expert_location_metadata, rank)
|
|
|
|
def __init__(self, expert_location_metadata: "ExpertLocationMetadata", rank: int):
|
|
self._expert_location_metadata = expert_location_metadata
|
|
self._rank = rank
|
|
|
|
def on_forward_pass_start(self, forward_batch: ForwardBatch):
|
|
pass
|
|
|
|
def on_select_experts(self, layer_idx: int, topk_ids: torch.Tensor):
|
|
pass
|
|
|
|
def on_deepep_dispatch_normal(
|
|
self,
|
|
layer_idx: int,
|
|
local_physical_count_of_layer: List[int],
|
|
num_tokens_per_rank,
|
|
num_tokens_per_rdma_rank,
|
|
num_tokens_per_expert,
|
|
):
|
|
pass
|
|
|
|
def on_deepep_dispatch_low_latency(
|
|
self, layer_idx: int, local_physical_count_of_layer: torch.Tensor
|
|
):
|
|
pass
|
|
|
|
def reset(self):
|
|
raise NotImplementedError
|
|
|
|
def collect(self) -> Dict:
|
|
raise NotImplementedError
|
|
|
|
|
|
class _LayerBasedSinglePassGatherer(_SinglePassGatherer):
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__(*args, **kwargs)
|
|
self._objects_of_layer = {}
|
|
|
|
def _on_layer_data(self, layer_idx: int, objects: List[int]):
|
|
assert 0 <= layer_idx < self._expert_location_metadata.num_layers
|
|
if layer_idx in self._objects_of_layer:
|
|
self._objects_of_layer[layer_idx] = _list_sum(
|
|
self._objects_of_layer[layer_idx], objects
|
|
)
|
|
else:
|
|
self._objects_of_layer[layer_idx] = objects
|
|
|
|
def reset(self):
|
|
self._objects_of_layer.clear()
|
|
|
|
def _collect_objects(self, pad_len: int) -> torch.Tensor:
|
|
data = [
|
|
self._objects_of_layer.get(layer_index) or ([0] * pad_len)
|
|
for layer_index in range(self._expert_location_metadata.num_layers)
|
|
]
|
|
return torch.tensor(data)
|
|
|
|
|
|
def _list_sum(a: List, b: List) -> List:
|
|
return [x + y for x, y in zip(a, b, strict=True)]
|
|
|
|
|
|
class _SelectExpertsSinglePassGatherer(_LayerBasedSinglePassGatherer):
|
|
# pretty slow, but we will use the DeepEP Gatherer in production
|
|
def on_select_experts(self, layer_idx: int, topk_ids: torch.Tensor):
|
|
topk_ids_list = topk_ids.to("cpu", non_blocking=True).numpy().tolist()
|
|
torch.cuda.synchronize()
|
|
|
|
global_physical_count = [
|
|
0
|
|
] * self._expert_location_metadata.num_physical_experts
|
|
for token_record in topk_ids_list:
|
|
for global_physical_expert_idx in token_record:
|
|
global_physical_count[global_physical_expert_idx] += 1
|
|
|
|
self._on_layer_data(layer_idx, global_physical_count)
|
|
|
|
def collect(self) -> Dict:
|
|
global_physical_count = super()._collect_objects(
|
|
pad_len=self._expert_location_metadata.num_physical_experts
|
|
)
|
|
return dict(global_physical_count=global_physical_count)
|
|
|
|
|
|
class _DeepepNormalSinglePassGatherer(_LayerBasedSinglePassGatherer):
|
|
def on_deepep_dispatch_normal(
|
|
self,
|
|
layer_idx: int,
|
|
local_physical_count_of_layer: List[int],
|
|
num_tokens_per_rank,
|
|
num_tokens_per_rdma_rank,
|
|
num_tokens_per_expert,
|
|
):
|
|
assert isinstance(local_physical_count_of_layer, list)
|
|
self._on_layer_data(layer_idx, local_physical_count_of_layer)
|
|
|
|
def collect(self) -> Dict:
|
|
local_physical_count = super()._collect_objects(
|
|
pad_len=self._expert_location_metadata.num_local_physical_experts
|
|
)
|
|
global_physical_count = _convert_local_to_global_physical_count(
|
|
local_physical_count,
|
|
rank=self._rank,
|
|
num_local_physical_experts=self._expert_location_metadata.num_local_physical_experts,
|
|
num_physical_experts=self._expert_location_metadata.num_physical_experts,
|
|
)
|
|
return dict(global_physical_count=global_physical_count)
|
|
|
|
|
|
class _DeepepLowLatencySinglePassGatherer(_SinglePassGatherer):
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__(*args, **kwargs)
|
|
self._data = torch.zeros(
|
|
(
|
|
self._expert_location_metadata.num_layers,
|
|
self._expert_location_metadata.num_local_physical_experts,
|
|
),
|
|
dtype=torch.int,
|
|
device="cuda",
|
|
)
|
|
|
|
def on_deepep_dispatch_low_latency(
|
|
self, layer_idx: int, local_physical_count_of_layer: torch.Tensor
|
|
):
|
|
# Most naive implementation, can optimize later
|
|
self._data[layer_idx, :] += local_physical_count_of_layer
|
|
|
|
def reset(self):
|
|
self._data[...] = 0
|
|
|
|
def collect(self) -> Dict:
|
|
# Can optimize if bottleneck
|
|
global_physical_count = _convert_local_to_global_physical_count(
|
|
self._data,
|
|
rank=self._rank,
|
|
num_local_physical_experts=self._expert_location_metadata.num_local_physical_experts,
|
|
num_physical_experts=self._expert_location_metadata.num_physical_experts,
|
|
)
|
|
return dict(global_physical_count=global_physical_count)
|
|
|
|
|
|
def _convert_local_to_global_physical_count(
|
|
local_physical_count: torch.Tensor,
|
|
rank: int,
|
|
num_local_physical_experts: int,
|
|
num_physical_experts: int,
|
|
) -> torch.Tensor:
|
|
dtype = local_physical_count.dtype
|
|
device = local_physical_count.device
|
|
num_layers, _ = local_physical_count.shape
|
|
|
|
ans = torch.zeros((num_layers, num_physical_experts), dtype=dtype, device=device)
|
|
ans[
|
|
:, num_local_physical_experts * rank : num_local_physical_experts * (rank + 1)
|
|
] = local_physical_count
|
|
return ans
|
|
|
|
|
|
# --------------------------------------- Accumulator -----------------------------------------
|
|
|
|
_SINGLE_PASS_GATHERER_KEY_PRIMARY = "primary"
|
|
|
|
|
|
class _Accumulator(ABC):
|
|
@staticmethod
|
|
def init_new(
|
|
server_args: ServerArgs,
|
|
expert_location_metadata: "ExpertLocationMetadata",
|
|
rank: int,
|
|
) -> "_Accumulator":
|
|
return _Accumulator.get_class(server_args)(
|
|
server_args, expert_location_metadata, rank
|
|
)
|
|
|
|
@staticmethod
|
|
def get_class(server_args: ServerArgs) -> Type["_Accumulator"]:
|
|
return {
|
|
"stat": _StatAccumulator,
|
|
# TODO pr-chain: enable this later
|
|
# "per_pass": _DetailAccumulator,
|
|
# "per_token": _DetailAccumulator,
|
|
}[server_args.expert_distribution_recorder_mode]
|
|
|
|
def __init__(
|
|
self,
|
|
server_args: ServerArgs,
|
|
expert_location_metadata: "ExpertLocationMetadata",
|
|
rank: int,
|
|
):
|
|
self._server_args = server_args
|
|
self._expert_location_metadata = expert_location_metadata
|
|
self._rank = rank
|
|
|
|
def get_single_pass_gatherer_keys(self):
|
|
return [_SINGLE_PASS_GATHERER_KEY_PRIMARY]
|
|
|
|
def get_single_pass_gatherer_key(self, debug_name: Optional[str]):
|
|
return _SINGLE_PASS_GATHERER_KEY_PRIMARY
|
|
|
|
def append(
|
|
self,
|
|
forward_pass_id: int,
|
|
gatherer_key: str,
|
|
single_pass_data: Dict,
|
|
):
|
|
pass
|
|
|
|
def reset(self):
|
|
pass
|
|
|
|
def dump(self, output_mode: _OutputMode):
|
|
pass
|
|
|
|
|
|
class _StatAccumulator(_Accumulator):
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__(*args, **kwargs)
|
|
self._global_physical_count_of_buffered_step = _Buffer.init_new(
|
|
item_shape=(
|
|
self._expert_location_metadata.num_layers,
|
|
# Cannot use local_physical_count to support select_experts
|
|
self._expert_location_metadata.num_physical_experts,
|
|
),
|
|
buffer_size=self._server_args.expert_distribution_recorder_buffer_size,
|
|
dtype=torch.int32,
|
|
device=self._server_args.device,
|
|
)
|
|
|
|
def append(
|
|
self,
|
|
forward_pass_id: int,
|
|
gatherer_key: str,
|
|
single_pass_data: Dict,
|
|
):
|
|
super().append(forward_pass_id, gatherer_key, single_pass_data)
|
|
# Can optimize if overhead here is large
|
|
self._global_physical_count_of_buffered_step.append(
|
|
single_pass_data["global_physical_count"]
|
|
)
|
|
|
|
def reset(self):
|
|
super().reset()
|
|
self._global_physical_count_of_buffered_step.reset()
|
|
|
|
def dump(self, output_mode: _OutputMode):
|
|
logical_count_of_buffered_step = _convert_global_physical_count_to_logical_count(
|
|
self._global_physical_count_of_buffered_step.get_all(),
|
|
num_layers=self._expert_location_metadata.num_layers,
|
|
num_logical_experts=self._expert_location_metadata.num_logical_experts,
|
|
physical_to_logical_map=self._expert_location_metadata.physical_to_logical_map,
|
|
)
|
|
torch.distributed.all_reduce(
|
|
logical_count_of_buffered_step, op=torch.distributed.ReduceOp.SUM
|
|
)
|
|
output = dict(
|
|
rank=self._rank,
|
|
logical_count=logical_count_of_buffered_step,
|
|
)
|
|
|
|
if output_mode == "file":
|
|
if self._rank == 0:
|
|
_dump_to_file(f"expert_distribution_recorder_{time.time()}.pt", output)
|
|
elif output_mode == "object":
|
|
return output
|
|
else:
|
|
raise NotImplementedError
|
|
|
|
|
|
def _dump_to_file(name, data):
|
|
save_dir = Path(os.environ.get("SGLANG_EXPERT_DISTRIBUTION_RECORDER_DIR", "/tmp"))
|
|
path_output = save_dir / name
|
|
logger.info(f"Write expert distribution to {path_output}")
|
|
if not save_dir.exists():
|
|
save_dir.mkdir(parents=True, exist_ok=True)
|
|
torch.save(data, str(path_output))
|
|
|
|
|
|
class _Buffer:
|
|
@staticmethod
|
|
def init_new(item_shape: Tuple, buffer_size: int, dtype, device):
|
|
if buffer_size < 0:
|
|
return _InfiniteBuffer(item_shape, dtype=dtype, device=device)
|
|
else:
|
|
return _CircularBuffer(item_shape, buffer_size, dtype=dtype, device=device)
|
|
|
|
def append(self, value: torch.Tensor):
|
|
raise NotImplementedError
|
|
|
|
def get_all(self) -> torch.Tensor:
|
|
raise NotImplementedError
|
|
|
|
def reset(self):
|
|
raise NotImplementedError
|
|
|
|
|
|
class _CircularBuffer(_Buffer):
|
|
def __init__(self, item_shape: Tuple, buffer_size: int, dtype, device):
|
|
self._buffer = torch.zeros(
|
|
(buffer_size, *item_shape), dtype=dtype, device=device
|
|
)
|
|
self._curr_index = 0
|
|
|
|
def append(self, value: torch.Tensor):
|
|
self._buffer[self._curr_index] = value
|
|
self._curr_index = (self._curr_index + 1) % len(self._buffer)
|
|
|
|
def get_all(self) -> torch.Tensor:
|
|
return self._buffer
|
|
|
|
def reset(self):
|
|
self._buffer[...] = 0
|
|
|
|
|
|
class _InfiniteBuffer(_Buffer):
|
|
def __init__(self, item_shape: Tuple, dtype, device):
|
|
self._item_shape = item_shape
|
|
self._buffer = torch.zeros((128, *item_shape), dtype=dtype, device=device)
|
|
self._size = 0
|
|
|
|
def append(self, value: torch.Tensor):
|
|
curr_buffer_size = len(self._buffer)
|
|
dtype = self._buffer.dtype
|
|
device = self._buffer.device
|
|
|
|
if self._size == curr_buffer_size:
|
|
new_buffer = torch.zeros(
|
|
(2 * curr_buffer_size, *self._item_shape), dtype=dtype, device=device
|
|
)
|
|
new_buffer[:curr_buffer_size] = self._buffer
|
|
self._buffer = new_buffer
|
|
|
|
self._buffer[self._size] = value
|
|
self._size += 1
|
|
|
|
def get_all(self) -> torch.Tensor:
|
|
return self._buffer[: self._size]
|
|
|
|
def reset(self):
|
|
self._buffer[...] = 0
|
|
self._size = 0
|
|
|
|
|
|
def _convert_global_physical_count_to_logical_count(
|
|
# (whatever, num_layers, num_physical_experts)
|
|
global_physical_count: torch.Tensor,
|
|
num_layers: int,
|
|
num_logical_experts: int,
|
|
physical_to_logical_map: torch.Tensor,
|
|
):
|
|
dim_extra, _, _ = global_physical_count.shape
|
|
dtype = global_physical_count.dtype
|
|
device = global_physical_count.device
|
|
logical_count = torch.zeros(
|
|
(dim_extra, num_layers, num_logical_experts), dtype=dtype, device=device
|
|
)
|
|
logical_count.scatter_add_(
|
|
dim=2,
|
|
index=physical_to_logical_map.unsqueeze(0).expand(dim_extra, -1, -1),
|
|
src=global_physical_count,
|
|
)
|
|
return logical_count
|