Crane · Reproduction & Scheduling Lab

An independent Python reproduction and experiment lab for Crane's inter-layer DNN scheduling method. Inspect real network graphs, optimize ScT/MeT tables, compose nested schedules, and reproduce recorded experiments with native SET intra-layer cost profiles.

English · 简体中文 · Recorded results · Demo · Reproduction scope

Research status: inference experiments now use real network definitions and recorded SET core mappings. Training includes a checked uniform-cohort reference and a separate experimental MILP path. Full placement/traffic calibration and reproduction of all paper figures or headline speedups remain unestablished.

Paper and relationship to SET

Crane: Inter-Layer Scheduling Framework for DNN Inference and Training Co-Support on Tiled Architecture — Yu Gong, Lingyi Huang, Haodong Chang, Rongjian Liang, Cheng Yang, Zhexiang Tang, Jiang Hu, and Bo Yuan. MICRO 2025, pp. 1250–1263.

DOI / publisher · Paper PDF · Citation metadata

Crane and SET are different scheduling frameworks. The Crane paper uses SET for cost-model validation and as an inference baseline. This repository imports SET's compiled network definitions, can reuse its Polar intra-layer mapper, and runs its native inter-layer search separately as a reference. The Python Crane scheduler is an independent implementation; the paper's implementation is C++.

What is implemented

• 16 real network definitions, exported from a pinned SET revision with tensor shapes, operation counts, residual branches and activation/weight edges.
• ScT and MeT optimization, per-sample workload scaling, sample-index-aware traffic accounting, solver deadlines, termination status and bounds.
• Exact fixed-cost ScT EDP options: a normalized integer-product MILP and an equivalent vertex reduction for eligible canonical inference problems.
• Nested sub-batch composition: children process exactly one parent sub-batch within assigned tiles and conservative memory budgets. Serial micro-batch schedules provide a feasible alternative for small batches.
• Recorded native SET core profiles for ResNet-50, VGG-19, GoogLeNet and a Transformer cell, plus seeded native SET reference runs.
• Training memory reference: explicit FW/BW1/recomputation/BW2 cohorts, operation conservation and capacity checks, including the paper's Figure-6 cohort example.
• An experiment runner and offline demo with raw evidence, figures, timing, source hashes, configuration and dependency records.

flowchart LR
    A[Compiled network graph] --> B[Blocks and nested sub-batches]
    P[Recorded SET core mappings] --> C[ScT compute optimization]
    B --> C
    C --> D[MeT and tensor-interval traffic]
    D --> E[Capacity and dependency checks]
    E --> F[Raw records, figures and interactive demo]

Loading

These are scheduling and cost-model experiments. They run on a CPU without training model weights or downloading datasets.

Quick start

Use Python 3.11–3.13 and run commands from the repository root:

git clone https://github.com/aHappend/Crane.git
cd Crane
python -m venv .venv

Activate with source .venv/bin/activate on Linux/macOS or .\.venv\Scripts\Activate.ps1 in Windows PowerShell, then:

python -m pip install -r requirements.txt -c constraints.txt
python tools/doctor.py
python example/quickstart.py

The small synthetic example exercises both real SCIP table solvers and writes summary.json and schedule.html under outputs/experiments/quickstart_<timestamp>/. Use a new --output-dir to choose another destination. No GPU or commercial solver service is required.

Interactive demo

Open docs/demo/index.html locally after cloning. It is self-contained and works without a server or network access. Alternatively:

python -m http.server 8000 --bind 127.0.0.1 --directory docs/demo

Open http://127.0.0.1:8000 on that machine. For an SSH server, forward the port from your own computer with ssh -L 8000:127.0.0.1:8000 <your-host>.

The demo selects recorded runs: compare schedules, play through ScT states, inspect activation storage, explore training capacity and view seeded SET runs. Changing a selector does not run a new optimization in the browser.

Recorded experiments

The full report contains 10 inference runs, 9 training capacity configurations (7 feasible and 2 infeasible under the reference policy), and 12 native SET runs over three fixed seeds. Raw records are archived with hashes; negative results are retained.

The figure compares nested scheduling with a serial reference using the same SET core profiles and analytical traffic model, batch 64 and 16 tiles. VGG-19's slight regression is visible. Native SET results use a different outer traffic/placement evaluator and are reported separately.

Run the main suites:

python -m experiments.run --config experiments/configs/native_core_inference.json --output-dir outputs/experiments/my-inference
python -m experiments.run --config experiments/configs/training_memory.json --output-dir outputs/experiments/my-training

The checked-in cost profiles make these commands independent of a C++ compiler. Each case has a process deadline and records errors or infeasibility explicitly. See EXPERIMENTS.md for regeneration, native SET execution, plotting, units and result interpretation.

Code map

Directory	Responsibility
workloads/	Compiled graph metadata, loader and initial hierarchy
model/	Layer objects and DAG validation
scheduler/	ScT/MeT, exact objectives, traffic intervals, hardware profiles
search/	Flat search, nested composition and training policies
cost_model/	Analytical costs and recorded SET core adapter
experiments/	Configurations, isolated runs, profiles, raw results and reports
tools/	Environment check and pinned upstream extraction/baseline tools
example/	Quick start and retained exploratory examples
tests/	Mathematical oracles, graph/traffic regressions and integration tests
docs/demo/	Self-contained interactive recorded-run explorer

Read the architecture, mathematical audit and paper-to-code map before changing a cost assumption. The old official_nns examples remain historical proxy experiments; use the experiments/ suites for the real exported networks.

Development

python -m pip install -r requirements-dev.txt -c constraints.txt
python -m ruff check .
python -m pytest -q

CI runs on Linux/Python 3.11 and 3.13 and Windows/Python 3.11. It checks SCIP, mathematical/behavioral invariants and entrypoints, and uploads a quick-start report. Long experiment matrices are run separately with their recorded budgets.

See CONTRIBUTING.md, CHANGELOG.md and THIRD_PARTY_NOTICES.md. The repository has no blanket software license declared; reference material retains its own terms.

Citation and attribution

Cite the Crane paper for the method and identify this repository's commit and experiment manifest when using the implementation:

@inproceedings{gong2025crane,
  author = {Gong, Yu and Huang, Lingyi and Chang, Haodong and Liang, Rongjian
            and Yang, Cheng and Tang, Zhexiang and Hu, Jiang and Yuan, Bo},
  title = {Crane: Inter-Layer Scheduling Framework for {DNN} Inference and
           Training Co-Support on Tiled Architecture},
  booktitle = {Proceedings of the 58th IEEE/ACM International Symposium on Microarchitecture},
  year = {2025},
  pages = {1250--1263},
  doi = {10.1145/3725843.3756023}
}

For the imported network definitions, native cost profiles or SET baseline, also cite Cai et al., Inter-layer Scheduling Space Definition and Exploration for Tiled Accelerators, ISCA 2023, DOI: 10.1145/3579371.3589048.