한국어
이 페이지는 ENN-PyTorch 코드를 읽을 때 어디서부터 보면 되는지 정리한 참조 자료다.
본문 1~8장은 실행 흐름과 구조를 설명한다. 이 페이지는 그 설명을 반복하지 않고, 관심사별로 먼저 확인할 파일과 함께 볼 파일만 짧게 정리한다.
English
This page is a reference map for deciding where to start when reading the ENN-PyTorch codebase.
Chapters 1 through 8 explain execution flow and structure. This page does not repeat those explanations; instead, it briefly lists the first files to check and the related files to read by concern.
| 관심사 / Concern | 먼저 볼 파일 / First Files | 같이 볼 파일 / Also Check |
|---|---|---|
| 공개 API와 workflow | ||
| Public API and workflow | enn_torch/__init__.py |
enn_torch/runtime/workflows.py |
| 모델 생성 | ||
| Model creation | enn_torch/runtime/workflows.py |
enn_torch/core/config.py, enn_torch/nn/wrappers.py |
| 모델 구조 | ||
| Model structure | enn_torch/nn/wrappers.py |
enn_torch/nn/layers.py, enn_torch/nn/blocks.py |
assembled + p * delta |
enn_torch/nn/wrappers.py |
enn_torch/nn/layers.py |
| Scaler / SigmoidGate | enn_torch/nn/layers.py |
enn_torch/runtime/main.py |
| Fuser / Collector | enn_torch/nn/wrappers.py |
enn_torch/nn/blocks.py |
| attention backend | enn_torch/nn/kernels.py |
enn_torch/core/policies.py |
| autocast / dtype 협상 | ||
| autocast / dtype negotiation | enn_torch/core/precision.py |
enn_torch/core/policies.py |
| memmap staging | enn_torch/data/collate.py |
enn_torch/data/pipeline.py |
| sampler / loader / stream | enn_torch/data/nodes.py |
enn_torch/core/concurrency.py |
| 학습 worker | ||
| Training worker | enn_torch/runtime/main.py |
enn_torch/runtime/workflows.py |
| 예측 worker | ||
| Prediction worker | enn_torch/runtime/main.py |
enn_torch/nn/wrappers.py, enn_torch/nn/layers.py |
| OOM recovery | enn_torch/runtime/autobatch.py |
enn_torch/data/nodes.py |
| distributed / process group | enn_torch/runtime/distributed.py |
enn_torch/core/system.py |
| DCP checkpoint | enn_torch/runtime/distributed.py |
enn_torch/runtime/main.py |
| 저장과 내보내기 | ||
| Saving and export | enn_torch/runtime/io.py |
enn_torch/nn/graph.py |
| optimizer / EMA / SWA | enn_torch/runtime/optimizers.py |
enn_torch/runtime/main.py |
| loss 구성 | ||
| Loss construction | enn_torch/runtime/losses.py |
enn_torch/runtime/main.py |
한국어
모델이 예측을 만드는 핵심 흐름은 nn/wrappers.py에서 시작해 nn/layers.py로 이어진다. Fuser, Collector, SigmoidGate, Scaler의 연결을 보면 assembled + p * delta 구조를 따라갈 수 있다.
English
The core flow that produces predictions starts in nn/wrappers.py and continues into nn/layers.py. By following Fuser, Collector, SigmoidGate, and Scaler, you can trace the assembled + p * delta structure.
enn_torch/nn/wrappers.py
→ Model
→ Template
→ Fuser
→ Collector
enn_torch/nn/layers.py
→ Embedding
→ Scaler
→ SigmoidGate