Mode A (per-paper synthesis). Target: 2609.03335 (category: algorithm, arxivcs.DC). The related set assigned to this paper is unusually heterogeneous: it mixes the target's actual neighbors in the compound-AI stack (RAG 2005.11401, RAPTOR 2401.18059, GSM8K 2110.14168, Llama 3 2407.21783) with foundational algorithm papers from other sub-fields (Transformer 1706.03762, DreamerV3 2301.04104, StruQ 2402.06363, SageAttention 2410.02367). This L3 treats the relatedness as layered: workload/component neighbors first, then method-shape analogues, then orthogonal-but-instructive references. That layering is itself a finding (see §生态位).
The target is a systems scheduling paper whose object is an online
graph–schedule construction heuristic over a physical-execution state space
[2609.03335]. Against the assigned peers, relatedness
splits into three concentric rings:
Ring 1 — direct workload / component coupling (the true neighbors).
evaluation workloads: its ensemble DAG "runs five heterogeneous solvers
in parallel, aggregates their answers by majority vote, and refines the
result" [2609.03335]. The verifier paper is the
origin of the sample-many-then-rank pattern that produces exactly the
fan-out/fan-in agent graph the target must schedule
[2110.14168]. So 2110.14168 supplies the *workload
generator; 2609.03335 supplies the runtime* that serves it efficiently.
retrieval-augmented / tool-interleaved call patterns that the target's
MBPP "code generation, testing, diagnosis, review, repair" DAG and
QMSum hierarchical-summary DAG instantiate as multi-model, multi-parent
dependency graphs [2609.03335]. RAG's per-token latent
marginalization [2005.11401] and RAPTOR's collapsed-tree
multi-hop synthesis [2401.18059] are precisely the kind
of "interleave model calls with tool results and control-flow decisions"
the target's activation model abstracts over.
serves five Qwen3 models but its whole cost model (loading time, peak
memory, KV footprint) is parameterized by exactly the dense-Transformer
serving profile Llama 3 documents (GQA, PagedAttention rejection
sampling) [2407.21783]. Llama 3 also shares
the target's workloads (GSM8K 96.8, HumanEval, MATH)
Ring 2 — method-shape analogues (different domain, same algorithmic move).
world model that forecasts future latent state and plans in imagination
[2301.04104]. The target's Predictor→Scheduler loop is
structurally the same move — forecast future device/readiness state, then
optimize actions against the forecast [2609.03335].
Both replace exhaustive online search/profiling with a learned forecast.
model as an operator data-flow graph and uses a graph encoder + heads
[2609.03335]; the served models are Transformers, and
the KV/attention memory that dominates the target's cumulative-memory
constraint (Eq 11) is the direct downstream cost of the $O(n^2)$
attention Transformer introduced [1706.03762].
Ring 3 — orthogonal reference (shared "algorithm" bucket only).
method [2410.02367]. It shares no scheduling concern;
its only tie is that it changes the *per-activation run-time / memory
numbers* that the target's Predictor would have to re-profile. It is a
substitutable input to Eq 1, not a competitor.
[2402.06363]. Orthogonal to scheduling; relevant only
because the target's "authorized isolation" for cross-workflow replica
sharing [2609.03335] is exactly the trust boundary StruQ
formalizes at the prompt/data channel level.
Cited-related count: 8/8 peers carry at least one [ref:L2:...]
citation above.
| dimension | target 2609.03335 | closest peer | delta |
|---|---|---|---|
| object of optimization | physical execution graph + schedule jointly, online | DreamerV3 plans actions in a learned latent world model [2301.04104] | new: the "world state" is a live GPU pool snapshot $S_t$, not a learned RSSM; planning is combinatorial over lifecycle/placement, not gradient-based over a policy |
| prediction target | per-(model,request,device) run/peak-mem/load via graph encoder [2609.03335] | DreamerV3 predicts reward/return distribution [2301.04104] | incremental within prediction-guided planning, but the target predicts systems cost primitives, not task value — a different regression target with a decision-fidelity metric (ordering / window-fit) absent in DreamerV3 |
| workload it serves | fan-in/repair/summary agent DAGs [2609.03335] | GSM8K ensemble+refine [2110.14168], RAG marginalize-over-K [2005.11401], RAPTOR collapsed tree [2401.18059] | new layer: those papers produce the DAG shape and never model serving cost; the target is the first to make the DAG's physical realization a first-class decision |
| served model | 5 Qwen3 sizes, shared vLLM config [2609.03335] | Llama 3 recipe [2407.21783] | downstream/complementary: Llama 3 optimizes the model; the target takes the model as a fixed deployment identity $d$ and optimizes when/where/how it is loaded |
| memory reasoning | cumulative transition-boundary feasibility Eq 11 [2609.03335] | SageAttention shrinks per-op VRAM via INT8 [2410.02367] | orthogonal: SageAttention lowers $\widehat{M}^{\text{peak}}$; the target reasons about transient overlap of many replicas, which quantization does not address |
| formal guarantees | none — feasibility construction + lexicographic key, no competitive ratio [2609.03335] | Transformer (empirical only) [1706.03762]; DreamerV3 (empirical only) [2301.04104] | shared weakness: every peer here is "无形式化作者证明 — 仅实证"; the target is not an outlier, but it compares against an Oracle and still gives no bound relative to it |
Genuinely new (nothing in the peer set does this): the *joint,
cumulative-memory-feasible online construction of graph AND schedule
together*, evaluating each transformation's cumulative effect on ready work,
scarce devices, transient memory, and resident state at every predicted
transition boundary [2609.03335]. None of the eight peers
touches multi-workflow model-lifecycle coalescing on a heterogeneous pool.
Contradictory / tension. The target reports that model **loading
consumes more GPU time than token generation** and idle residency adds
28.1–33.0% [2609.03335]. This inverts the implicit cost
model of the algorithm-side peers, whose entire optimization budget assumes
compute is the bottleneck: SageAttention accelerates the matmuls
[2410.02367] and Llama 3's serving optimizations target
generation throughput [2407.21783].
矛盾根源: different regime. SageAttention/Llama 3 measure *single-model,
steady-residency, long-sequence* serving where attention compute dominates;
the target measures many small models on a fixed pool with churn, where
load/idle dominates. Both are correct in their own regime — but a reader who
imports "attention is the bottleneck" into the agent-serving-on-small-pool
setting would optimize the wrong thing.
reproducibility that its peers demonstrate.** The target's implementation
is unpublished [2609.03335], and it openly
flags two easy-to-miss tricks (two output buckets per unit; max-context /
summed-runtime asymmetry) whose omission "silently breaks either latency
accuracy or feasibility." Contrast with the peer baseline for
reproducibility: RAG, StruQ, and SageAttention all ship working code
[2410.02367]. A scheduling heuristic with
this many load-bearing undocumented constants is hard to trust without
the artifact.
paper concedes it loses some Poisson p95 tails (0.075 Overall p95 392 s
vs Kairos 233 s) [2609.03335]. Its global-drain
objective is not Pareto-dominant. This is the same "aggregate win hides
per-instance losses" honesty pattern DreamerV3 shows on Atari
(Breakout 384 vs MuZero 799) [2301.04104] — but for a
latency system, a worse tail under realistic (non-burst) arrivals is a
more serious concession than a lost game.
formula predictor beats the learned Predictor on peak-VRAM WAPE
(11.98% vs 14.08%) [2609.03335]. Since peak memory is
the feasibility gate (Eq 11), losing on exactly the safety-critical
target is awkward; the adaptive $\rho_\kappa(t)$ term only patches OOMs
after they occur. A skeptic asks whether the graph-encoder Predictor
earns its cost when a size/bandwidth formula is safer on memory.
Baselines Parrot and Kairos are re-run on the target's substrate
[2609.03335], which is the right control — but the two
predictor "baselines" are the target's own system with the predictor
swapped, so they isolate prediction quality, not scheduler quality. There
is no independent joint-scheduler baseline (e.g. an ILP solver over the
same $\mathcal{G}_t$), so the lexicographic heuristic's optimality gap
vs the Oracle is unquantified — the same missing competitive-ratio bound
noted in [2609.03335].
[2609.03335]. The "private small pool" framing is
internally consistent, but the cross-workflow-lifecycle benefit that
dominates makespan may not hold at larger scale where replica supply is
less contended — untested.
The target sits at a **paradigm boundary the peer set brackets on both
sides**. Read the eight peers as a timeline of where the "unit of
optimization" has moved:
[1706.03762]; SageAttention optimizes its kernels
[2410.02367]; Llama 3 optimizes the training recipe
[2407.21783]. Unit = the model.
[2005.11401] [2401.18059]; the GSM8K
verifier adds test-time search [2110.14168]. Unit = *one
inference call + its augmentation*.
realization of a whole concurrent DAG on a live pool*
So the target's niche is the serving-systems consequence of the
compound-AI shift that the algorithm papers created. It is a "shift from
models to compound AI systems" runtime — and, notably, it argues the
dominant cost in that regime is not the compute those algorithm papers
optimize, but model lifecycle [2609.03335].
Adoption evidence. Weak/none yet: unnamed, code pending
[2609.03335]. Contrast with strong-adoption
peers — Transformer is the substrate of "essentially all modern LLMs"
[1706.03762], SageAttention is "widely
integrated into video/image generation inference" with successors
[2410.02367], and RAG ships in HuggingFace
[2005.11401]. The target's niche is real and
timely, but its ecosystem position is aspirational, not established.
SageAttention-class INT8 attention changes per-activation run-time and
peak memory [2410.02367], the target's Eq 1 heads could
treat precision as an extra dimension of deployment identity $d$ — the
Scheduler could then trade accuracy for a smaller feasibility footprint
under KV pressure. Neither paper does this; it is a technically-clean
hybrid (quantization is orthogonal to lifecycle).
Scheduler is a hand-tuned lexicographic heuristic
[2609.03335]; DreamerV3 shows a learned world model can
plan in imagination across domains with fixed hyperparameters
[2301.04104]. A hybrid would replace the lexicographic
key $\Phi_t$ with a learned value over pool states, potentially closing
the Poisson-tail losses the heuristic currently eats.
from sampling 100 candidates and ranking [2110.14168],
yet the target schedules those samples as opaque activations. A scheduler
that knows early-terminated (low-verifier-score) branches can be reclaimed
would shrink both makespan and GPU-s — an unexplored coupling between
test-time-search semantics and physical lifecycle.
and RAG's frozen document index are large, reusable, cross-request state
[2401.18059] [2005.11401]. The
target's lifecycle model manages model weights + KV but not *retrieval
indices*; extending deployment identity $d$ to cover pinned index shards
would let the same fusion/coalescing logic amortize retrieval, not just
generation.
across workflows only under "authorized isolation"
[2609.03335] but never specifies the isolation
mechanism. StruQ's reserved-token channel separation
[2402.06363] is a concrete candidate: a formal
prompt/data boundary would let the Scheduler safely coalesce lifecycle
actions across mutually-distrusting tenants, turning an assumption into a
guarantee.