Systems

AI Systems는 모델이 실제 run, artifact, service, workflow가 되는 지점을 다룹니다. 모델 구조나 loss 자체가 아니라, 학습이 어떻게 기록되고, 추론이 어떤 계약으로 실행되며, 결과가 어떻게 재현 가능한 주장으로 남는지를 봅니다.

이 페이지는 Infra가 아닙니다. Infra는 GPU, Slurm, storage, server operation처럼 실행 환경의 제약을 다루고, Systems는 모델 run과 artifact가 어떤 계약으로 관리되는지를 다룹니다.

$$\text{AI system}=(\text{model},\text{data},\text{runtime},\text{artifact},\text{operation})$$

AI system note의 핵심은 “무슨 모델을 썼다”가 아니라 어떤 artifact가 어떤 contract로 재현되고 실행되는지입니다. 같은 model weight라도 dataset snapshot, tokenizer, preprocessing, precision, batching, serving policy가 달라지면 다른 system입니다.

System Contract

AI system은 모델 파일 하나가 아니라, 입력에서 artifact와 claim까지 이어지는 실행 계약입니다.

$$S=(D,f_{\theta },C,E,R,A,V)$$

Part	Meaning	Example
$D$	data contract	dataset snapshot, split, preprocessing, feature schema
$f_{\theta }$	model artifact	architecture, weights, tokenizer or featurizer
$C$	config	hyperparameters, precision, context length, batching rule
$E$	environment	package set, container/module, CUDA/runtime boundary
$R$	runtime policy	training loop, inference mode, serving policy, scheduler request
$A$	artifacts	checkpoint, logs, metrics, generated outputs, report
$V$	verification	tests, metrics, rendered outputs, reproducibility evidence

This gives a practical rule:

$$\text{same weights}\ne \text{same system}$$

if data preprocessing, tokenizer, environment, precision, batching, or serving policy changed.

Main Areas

Area	Use for	Key notes
Training	run state, checkpoint, scaling, failure recovery	Training run, Checkpoint state, Distributed training
Inference	prediction-time execution, batch/online mode, output contract	Inference, Batch and online inference, Inference contract
Serving	endpoint, batching, capacity, latency, rollout	Model serving, Inference serving, Inference capacity planning, Latency and throughput
Environment	module, container, dependency, runtime drift	Environment management, Environment modules and containers
Reproducibility	run record, artifact, versioning, evidence boundary	Reproducibility, Run artifact, Model versioning
Operations	monitoring, deployment, recovery, resource scheduling	Deployment strategy, Observability, Failure recovery, Resource scheduling

Route by Object

If you are describing	Start here	Then route to
model structure or inductive bias	[[ai/architectures	Architectures]]
objective, optimizer, supervision, transfer	[[ai/machine-learning	Machine Learning]], [[ai/learning-methods
evaluation metric or benchmark claim	[[ai/evaluation	Evaluation]]
serving endpoint or batch inference	this page	[[concepts/systems/inference-contract
GPU, Slurm, storage, account, server issue	[[infra/index	Infra]]
paper or project implementation	[[papers/index	Papers]], [[projects/index

Lifecycle

Stage	System question	Evidence
Data preparation	어떤 snapshot과 preprocessing contract를 썼는가?	manifest, split, script, checksum
Training	어떤 run state가 checkpoint를 만들었는가?	config, seed, optimizer, logs
Evaluation	어떤 metric과 selection rule로 claim을 만들었는가?	eval table, baseline, confidence interval
Inference	input/output contract가 무엇인가?	schema, example, batch policy
Serving	latency, throughput, capacity boundary는 무엇인가?	benchmark, monitoring, rollout record
Reproducibility	같은 claim을 다시 재구성할 수 있는가?	run artifact, version tag, environment

Boundary With Infra

Question	Route
training 또는 inference concept인가?	AI Systems, Systems concepts
hardware, GPU, Slurm, storage, server operation이 관련되는가?	Infra
run 또는 artifact가 reproducible하다는 public record가 필요한가?	Reproducibility, Run record
cluster 또는 shared machine에서 실패했는가?	HPC, GPU, Server operations

Boundary Examples

Note topic	Put in Systems	Put in Infra
training run	checkpoint contents, config, seed, resume contract	Slurm request, GPU allocation, shared filesystem bottleneck
inference	input/output schema, batching rule, output validation	GPU memory, server process, network, deployment machine
environment	dependency contract and reproducibility metadata	module path, container runtime issue, driver/runtime mismatch
serving	latency/throughput target, rollout, fallback behavior	node capacity, monitoring setup, firewall, process supervision
artifact	what proves the claim can be reconstructed	where artifacts are stored and backed up

Reading Path

1. Training run으로 run state와 checkpoint 경계를 봅니다.
2. Inference와 Model serving으로 실행/serving 경계를 봅니다.
3. Environment management로 dependency와 runtime drift를 정리합니다.
4. Reproducibility와 Run artifact로 공개 가능한 증거 형식을 맞춥니다.
5. 실제 GPU/HPC/storage 문제가 되면 Infra로 내려갑니다.

System Artifact Checklist

AI system note는 모델 이름보다 artifact boundary를 먼저 남깁니다.

Artifact	Include
Dataset snapshot	source, version, split, preprocessing contract
Model checkpoint	architecture, weight, tokenizer/featurizer, config
Training state	optimizer, scheduler, step, seed, environment
Inference contract	input schema, output schema, batching, limit
Evaluation record	metric, selection rule, baseline, uncertainty
Runtime environment	package version, container/module, hardware class

Claim Pattern

For public writing, reduce a systems claim to:

$$(\text{data},\text{model},\text{config},\text{environment},\text{runtime},\text{artifact},\text{verifier})\to \text{claim}$$

Examples:

Claim	Required evidence
training is reproducible	data snapshot, code revision, config, seed, environment, checkpoint, metric record
inference is stable	input schema, output schema, batch policy, error handling, validation examples
serving meets latency target	workload shape, hardware class, p50/p95/p99, concurrency, timeout policy
checkpoint can resume	optimizer, scheduler, scaler, RNG, step, data position, artifact integrity
result can be published	public-safe artifact, metric protocol, baseline, sanitization boundary

System Failure Modes

• checkpoint는 남았지만 tokenizer, featurizer, preprocessing rule이 없습니다.
• evaluation metric은 있지만 split, selection rule, confidence interval이 없습니다.
• local inference는 되지만 serving batch/precision/timeout 조건이 다릅니다.
• hardware 문제를 model quality 문제로 오해합니다.
• deployment는 성공했지만 public claim을 뒷받침하는 run record가 없습니다.
• inference example은 있지만 input/output schema와 error contract가 없습니다.
• serving benchmark는 있지만 request distribution과 concurrency assumption이 없습니다.
• reproducibility note가 private paths, server names, or unpublished results에 의존합니다.

Checks

• model, data, code, config, environment, hardware class가 분리되어 기록됐는가?
• training artifact와 inference artifact가 같은 preprocessing contract를 쓰는가?
• metric claim과 serving requirement가 서로 다른 evidence를 요구한다는 점을 구분했는가?
• public note에 private path, server name, account, unpublished result가 들어가지 않았는가?
• artifact가 “생겼다”와 claim이 “검증됐다”를 구분했는가?
• Infra incident를 model/system design claim으로 과도하게 일반화하지 않았는가?

Related

• AI
• AI systems concepts
• Infra
• Projects