Elastic Inference Methods

This tree is intentionally outside ran_benchmarks.

ran_benchmarks keeps original RAN benchmark tasks and original benchmark methods. This directory keeps cross-task elastic inference methods, their training/evaluation entrypoints, RAN-backbone adapters, retained assets, and paper references.

Standard Method Layout

Each method under methods/<method> uses this structure:

• core/: core elasticization implementation for the method.
• train/: training entrypoints and launch scripts.
• eval/: inference, evaluation, and plotting entrypoints.
• adapters/: task/backbone adapters. For RAN benchmarks, the path is adapters/ran_benchmarks/<task>/backbones/<backbone>/.
• assets/: retained or generated assets.
• assets/performance/: performance tables, plots, and comparison payloads.
• assets/runs/: retained run roots. Checkpoints, logs, resolved configs, and histories usually live inside these run roots.
• assets/logs/: logs generated by local launch scripts.
• assets/repositories/: retained upstream/source repository snapshots, when available.
• configs/: method-level config copies.
• papers/: original paper PDFs.
• README.md: method-level audit notes.

The RAN benchmark task directories keep lightweight comparison indexes at ${RAN_BENCHMARK_ROOT}/tasks/<task>/comparisons/elastic_methods.yaml.

Environment

Set these paths before training or evaluating:

export ELASTIC_INFERENCE_METHODS_ROOT=/path/to/elastic_inference_methods
export RAN_BENCHMARK_ROOT=/path/to/ran_benchmarks
export RAN_BENCHMARK_ASSET_ROOT=/path/to/ran_benchmark_assets
export STRUJEPA_ROOT=/path/to/StruJEPA
export STRUJEPA_RAN_RUN_ROOT=/path/to/strujepa_ran_runs

For a local asset mirror inside this checkout, use:

export ELASTIC_INFERENCE_METHODS_ROOT=/path/to/elastic_inference_methods
export RAN_BENCHMARK_ASSET_ROOT=/path/to/ran_benchmark_assets
export STRUJEPA_RAN_RUN_ROOT=/path/to/strujepa_ran_runs

RAN_BENCHMARK_ASSET_ROOT is also exported internally as AI_RAN_BENCHMARK_ROOT for older StruJEPA integration code.

Current Methods

• methods/dynabert: DynaBERT-style width/depth elasticization.
• methods/matformer: MatFormer-style nested-FFN elasticization.
• methods/ofa: Once-for-All-style progressive shrinking.

Current RAN Benchmark Adapters

• channel_estimation: AdaFortiTran and A-MMSE backbones.
• traffic_prediction: iTransformer backbone.
• channel_prediction_wifo_style: WiFo backbone.
• beam_management: BeamFormer backbone.

Large retained run roots may be symlinked from methods/*/assets/runs/ to an external ${RAN_BENCHMARK_ASSET_ROOT} because they contain checkpoints, logs, and full training histories. Adapter configs, upstream method-code snapshots, and performance CSV/plot payloads are kept as local files in this tree.

Final normalized CSVs are copied under assets/performance/final/latest_results/data/. The raw BS=1 latency measurement table used by the final comparison CSVs is under assets/performance/final/latency/.