[Metrics] Add per-rank DP attention imbalance metrics (#42755)

Co-authored-by: merrymercy <merrymercy@users.noreply.github.com>
Co-authored-by: hanming-lu <hanming-lu@users.noreply.github.com>
This commit is contained in:
metamergebot
2026-10-06 14:20:41 -07:00
committed by GitHub
co-authored by merrymercy hanming-lu
parent 38bdf25f8d
commit 4506927635
9 changed files with 532 additions and 3 deletions
@@ -132,6 +132,7 @@ from sglang.srt.multimodal.transport.cuda_ipc import (
CudaIpcTensorTransportProxy,
)
from sglang.srt.observability.metrics_collector import (
DPBalanceStats,
DPCooperationInfo,
SchedulerMetricsCollector,
)
@@ -2413,6 +2414,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
# Metrics
dp_cooperation_info: Optional[DPCooperationInfo] = None
dp_balance_stats: Optional[DPBalanceStats] = None
prefill_stats: Optional[PrefillStats] = None
forward_iter: Optional[int] = None
launch_ts: Optional[float] = None
@@ -3880,6 +3882,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
mamba_decode_batch_idx_cpu=self.mamba_decode_batch_idx_cpu,
mamba_lazy_spec_track_positions_cpu=self.mamba_lazy_spec_track_positions_cpu,
dp_cooperation_info=self.dp_cooperation_info,
dp_balance_stats=self.dp_balance_stats,
prefill_stats=self.prefill_stats,
fpm_start_time=self.fpm_start_time,
forward_iter=self.forward_iter,
+1
View File
@@ -4919,6 +4919,7 @@ class Scheduler(
@scheduler_stage_method(SCHEDULER_STAGE_IDLE)
def on_idle(self):
"""Idle housekeeping: guard, check, metrics, reset, sleep."""
self.dp_attn_adapter.drop_sync_wait_carry()
# Flush any health-check signal deferred while the engine was busy.
self.maybe_send_health_check_signal()
@@ -1,6 +1,7 @@
from __future__ import annotations
from dataclasses import dataclass
import time
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Callable, Optional
import torch
@@ -16,6 +17,7 @@ from sglang.srt.managers.schedule_batch import ScheduleBatch
from sglang.srt.managers.scheduler_components.recv_skipper import (
SchedulerRecvSkipper,
)
from sglang.srt.managers.utils import is_health_check_generate_req
from sglang.srt.mem_cache.allocator import BaseTokenToKVPoolAllocator
from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
from sglang.srt.mem_cache.kv_cache_builder import uses_ssm_state
@@ -28,7 +30,10 @@ from sglang.srt.model_executor.cuda_graph_config import (
)
from sglang.srt.model_executor.forward_batch_info import ForwardMode
from sglang.srt.model_executor.runner import PrefillCudaGraphRunner
from sglang.srt.observability.metrics_collector import DPCooperationInfo
from sglang.srt.observability.metrics_collector import (
DPBalanceStats,
DPCooperationInfo,
)
from sglang.srt.runtime_context import (
get_exec,
get_memory,
@@ -118,9 +123,13 @@ class MLPSyncBatchInfo:
local_can_run_tbo: bool
local_forward_mode: int
prefill_cuda_graph_max_prefix_len: int = 0
# Every request in the local batch is a health-check probe.
is_health_check: bool = False
# some gathered elements
tp0_info_cpu: torch.Tensor = None
any_health_check: bool = False
sync_wait_seconds: float = 0.0
global_num_tokens: list[int] = None
global_num_tokens_for_logprob: list[int] = None
tbo_split_seq_index: torch.Tensor = None
@@ -139,6 +148,7 @@ class MLPSyncBatchInfo:
int(self.can_run_prefill_cuda_graph),
self.prefill_cuda_graph_max_prefix_len,
int(self.can_run_draft_cuda_graph),
int(self.is_health_check),
],
device=device,
dtype=dtype,
@@ -156,6 +166,7 @@ class MLPSyncBatchInfo:
0, # can_run_prefill_cuda_graph
0, # prefill_cuda_graph_max_prefix_len
1, # can_run_draft_cuda_graph
0, # is_health_check
],
device=device,
dtype=dtype,
@@ -166,6 +177,7 @@ class MLPSyncBatchInfo:
self.tp0_info_cpu = self._get_local_tensor(device="cpu").view(1, -1)
self.global_num_tokens = [self.num_tokens]
self.global_num_tokens_for_logprob = [self.num_tokens_for_logprob]
self.any_health_check = self.is_health_check
if _ENABLE_METRICS_DP_ATTENTION:
self.dp_cooperation_info = DPCooperationInfo.create(
self.tp0_info_cpu[:, 5].tolist()
@@ -177,6 +189,7 @@ class MLPSyncBatchInfo:
group: torch.distributed.ProcessGroup,
use_all_reduce: bool = False,
):
sync_start = time.perf_counter()
local_info_tensor = self._get_local_tensor(device=device)
fallback_tensor = self._get_fallback_tensor(device=device)
info_width = local_info_tensor.numel()
@@ -239,6 +252,8 @@ class MLPSyncBatchInfo:
self.can_run_prefill_cuda_graph = bool(tp0_info_cpu[:, 6].min())
self.prefill_cuda_graph_max_prefix_len = int(tp0_info_cpu[:, 7].max())
self.can_run_draft_cuda_graph = bool(tp0_info_cpu[:, 8].min())
self.any_health_check = bool(tp0_info_cpu[:, 9].max())
self.sync_wait_seconds = time.perf_counter() - sync_start
if _ENABLE_METRICS_DP_ATTENTION:
self.dp_cooperation_info = DPCooperationInfo.create(
tp0_info_cpu[:, 5].tolist()
@@ -413,6 +428,24 @@ def _local_prefill_cuda_graph_vote(
)
class DPSyncWaitCarry:
"""MLP-sync wait of batch-less gathers, folded into the next recorded step.
One scheduler iteration can gather more than once (with speculative
decoding, get_next_batch_to_run gathers for the new prefill batch first,
then for the running batch); the first gather absorbs the cross-rank
arrival skew but yields no batch to record on. Only batch-less gathers carry: an iteration
that ends with no batch drops the carry in Scheduler.on_idle, and a step
that runs a batch without recording (prebuilt, health check) drops it too,
so an idle engine's gathers never pile onto the first busy step.
"""
__slots__ = ("seconds",)
def __init__(self) -> None:
self.seconds = 0.0
def prepare_mlp_sync_batch_raw(
local_batch: ScheduleBatch,
model_runner: ModelRunner,
@@ -423,6 +456,7 @@ def prepare_mlp_sync_batch_raw(
offload_tags: set[str],
draft_require_mlp_tp_gather: Optional[bool] = None,
dwdp: bool = False,
sync_wait_carry: Optional[DPSyncWaitCarry] = None,
):
parallel = get_parallel()
num_dp_ranks = parallel.num_dp_ranks
@@ -481,6 +515,11 @@ def prepare_mlp_sync_batch_raw(
)
is_extend_in_batch = local_batch.forward_mode.is_extend() if local_batch else False
is_health_check = bool(
local_batch is not None
and local_batch.reqs
and all(is_health_check_generate_req(req) for req in local_batch.reqs)
)
if local_batch is not None:
local_batch.is_extend_in_batch = is_extend_in_batch
@@ -523,6 +562,7 @@ def prepare_mlp_sync_batch_raw(
local_can_run_tbo=local_can_run_tbo,
local_forward_mode=local_forward_mode,
prefill_cuda_graph_max_prefix_len=prefill_cuda_graph_max_prefix_len,
is_health_check=is_health_check,
)
if num_dp_ranks == 1:
@@ -576,6 +616,23 @@ def prepare_mlp_sync_batch_raw(
if _ENABLE_METRICS_DP_ATTENTION and local_batch is not None:
local_batch.dp_cooperation_info = mlp_sync_info.dp_cooperation_info
local_batch.dp_balance_stats = None
if _ENABLE_METRICS_DP_ATTENTION and num_dp_ranks > 1 and not skip_all_gather:
sync_wait_seconds = mlp_sync_info.sync_wait_seconds
if sync_wait_carry is not None:
sync_wait_seconds += sync_wait_carry.seconds
sync_wait_carry.seconds = 0.0
# Health-check probes run on an otherwise idle engine; recording them
# would show maximal imbalance where there is no load.
if need_idle_batch and not mlp_sync_info.any_health_check:
local_batch.dp_balance_stats = DPBalanceStats.create(
local_tokens=num_tokens,
global_num_tokens=mlp_sync_info.global_num_tokens,
sync_wait_seconds=sync_wait_seconds,
)
elif local_batch is None and sync_wait_carry is not None:
# No batch yet: this iteration may gather again, or end in on_idle.
sync_wait_carry.seconds = sync_wait_seconds
return local_batch
@@ -591,6 +648,11 @@ class SchedulerDPAttnAdapter:
enable_overlap: bool
spec_algorithm: SpeculativeAlgorithm
get_require_mlp_sync: Callable[[], bool]
sync_wait_carry: DPSyncWaitCarry = field(default_factory=DPSyncWaitCarry)
def drop_sync_wait_carry(self) -> None:
"""The iteration produced no batch, so its gather waits have no step to land on."""
self.sync_wait_carry.seconds = 0.0
def prepare_mlp_sync_batch(self, local_batch: ScheduleBatch):
draft_require_mlp_tp_gather = None
@@ -612,6 +674,7 @@ class SchedulerDPAttnAdapter:
disable_overlap_schedule=get_schedule().disable_overlap_schedule,
offload_tags=self.offload_tags,
dwdp=get_parallel().dwdp_size > 1,
sync_wait_carry=self.sync_wait_carry,
)
def maybe_prepare_mlp_sync_batch(
@@ -219,6 +219,13 @@ class SchedulerMetricsReporter:
self.enable_kv_cache_events = (
self.metrics_collector_context.enable_kv_cache_events
)
parallel = get_parallel()
# The max/mean token ratio is engine-wide; one scheduler reports it.
self.reports_dp_token_imbalance_ratio = (
parallel.attn_dp_rank == 0
and parallel.attn_tp_rank == 0
and parallel.attn_cp_rank == 0
)
self._init_metrics()
self._install_device_timer_on_runners()
# Keep log history after the existing async result copy so reporting does
@@ -1116,6 +1123,16 @@ class SchedulerMetricsReporter:
batch: ScheduleBatch,
result: Union[GenerationBatchResult, EmbeddingBatchResult],
):
if (
self.current_scheduler_metrics_enabled
and (dp_balance_stats := batch.dp_balance_stats) is not None
):
self.metrics_collector.observe_dp_balance(dp_balance_stats)
if self.reports_dp_token_imbalance_ratio:
self.metrics_collector.observe_dp_token_imbalance_ratio(
dp_balance_stats
)
if not isinstance(result, GenerationBatchResult):
return
@@ -23,6 +23,8 @@ from collections import Counter
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Any, Dict, List, Mapping, Optional, Set, Union
import msgspec
from sglang.srt.disaggregation.utils import DisaggregationMode
from sglang.srt.model_executor.forward_batch_info import ForwardMode
from sglang.srt.observability.scheduler_stage_metrics import (
@@ -203,6 +205,52 @@ class DPCooperationInfo:
return dataclasses.asdict(self)
class DPBalanceStats(msgspec.Struct, frozen=True):
"""Token balance across the attention-DP ranks of one MLP-sync step.
DP attention runs the ranks in lockstep, so the step costs what the
busiest rank costs. ``max_tokens - local_tokens`` is the capacity this
rank gives up on the step: computed as padding rows under MAX_LEN
padding, or spent waiting for the busiest rank under SUM_LEN. Counts are
the gathered MLP-sync values, so on speculative-decode steps they are
batch rows (requests); the verify width is uniform across ranks, which
leaves the ratios unchanged.
"""
local_tokens: int
max_tokens: int
sum_tokens: int
num_ranks: int
# Wall time of this rank's scheduler MLP-sync all-gather. On the gloo path
# (the default with overlap scheduling) this is waiting for the slowest
# rank; on the device-group paths it also includes the gather's stream sync.
sync_wait_seconds: float
@classmethod
def create(
cls, local_tokens: int, global_num_tokens: List[int], sync_wait_seconds: float
) -> DPBalanceStats:
max_tokens = max(global_num_tokens)
if max_tokens == 0:
raise ValueError("DP balance stats need a rank with tokens")
return cls(
local_tokens=local_tokens,
max_tokens=max_tokens,
sum_tokens=sum(global_num_tokens),
num_ranks=len(global_num_tokens),
sync_wait_seconds=sync_wait_seconds,
)
@property
def imbalance_tokens(self) -> int:
return self.max_tokens - self.local_tokens
@property
def max_over_mean(self) -> float:
"""Busiest rank's tokens over the per-rank mean; 1.0 is perfectly balanced."""
return self.max_tokens * self.num_ranks / self.sum_tokens
# Role keys used by ServerArgs.stat_loggers to look up collector overrides.
# Embedded-use callers (e.g. Ray Serve LLM) pass {"scheduler": MyClass, ...} on
# ServerArgs and the five collector instantiation sites pick the right class.
@@ -1027,6 +1075,92 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
),
labelnames=list(labels.keys()) + ["category", "num_prefill_ranks"],
)
self.dp_attention_tokens_total = Counter(
name="sglang:dp_attention_tokens_total",
documentation=(
"Tokens of MLP-sync steps on this DP rank. kind=scheduled: tokens "
"this rank ran; kind=imbalance: tokens the busiest rank ran beyond "
"this rank, i.e. capacity lost to DP imbalance on this rank. "
"On speculative-decode steps the unit is batch rows (requests)."
),
labelnames=list(labels.keys()) + ["kind"],
)
self.dp_attention_steps_total = Counter(
name="sglang:dp_attention_steps_total",
documentation=(
"MLP-sync steps on this DP rank. rank_state=idle: this rank had no "
"tokens and only ran to keep the DP collectives in lockstep."
),
labelnames=list(labels.keys()) + ["rank_state"],
)
self.dp_attention_token_imbalance_ratio = Histogram(
name="sglang:dp_attention_token_imbalance_ratio",
documentation=(
"Per MLP-sync step, max tokens over mean tokens across the DP "
"ranks (1.0 = balanced, num_dp_ranks = one busy rank). Reported "
"once per engine, by the scheduler of DP rank 0."
),
labelnames=labels.keys(),
buckets=(
1.0,
1.05,
1.1,
1.25,
1.5,
2.0,
3.0,
4.0,
6.0,
8.0,
12.0,
16.0,
24.0,
32.0,
48.0,
64.0,
),
)
self.dp_attention_sync_wait_seconds = Histogram(
name="sglang:dp_attention_sync_wait_seconds",
documentation=(
"Wall time of this DP rank's scheduler MLP-sync all-gather per step. "
"On the gloo path (the default with overlap scheduling) this is "
"waiting for the slowest rank; on the device-group paths it also "
"includes the gather's stream sync."
),
labelnames=labels.keys(),
buckets=(
0.0001,
0.00025,
0.0005,
0.001,
0.0025,
0.005,
0.01,
0.025,
0.05,
0.1,
0.25,
0.5,
1.0,
),
)
# Pre-seed and cache the children: every step observes them, and a
# ratio chart needs both kinds / states present from the start.
self._dp_attention_tokens = {
kind: self.dp_attention_tokens_total.labels(**labels, kind=kind)
for kind in ("scheduled", "imbalance")
}
self._dp_attention_steps = {
state: self.dp_attention_steps_total.labels(**labels, rank_state=state)
for state in ("active", "idle")
}
self._dp_attention_token_imbalance_ratio = (
self.dp_attention_token_imbalance_ratio.labels(**labels)
)
self._dp_attention_sync_wait_seconds = (
self.dp_attention_sync_wait_seconds.labels(**labels)
)
# =================================================================
# Prefill delayer
@@ -1328,6 +1462,18 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
balancedness
)
def observe_dp_balance(self, stats: DPBalanceStats) -> None:
if stats.local_tokens:
self._dp_attention_tokens["scheduled"].inc(stats.local_tokens)
if stats.imbalance_tokens:
self._dp_attention_tokens["imbalance"].inc(stats.imbalance_tokens)
rank_state = "idle" if stats.local_tokens == 0 else "active"
self._dp_attention_steps[rank_state].inc(1)
self._dp_attention_sync_wait_seconds.observe(stats.sync_wait_seconds)
def observe_dp_token_imbalance_ratio(self, stats: DPBalanceStats) -> None:
self._dp_attention_token_imbalance_ratio.observe(stats.max_over_mean)
def increment_realtime_tokens(
self,
dp_cooperation_info: Optional[DPCooperationInfo],