mamba3

Focused benchmark and research workspace for comparing a ViT-5 attention baseline against a Mamba-3 variant on CIFAR-10, plus a small paper-processing utility and FALCON notes.

What is tracked

• benchmarks/: top-level experiment code for the CIFAR-10 ViT-5 vs Mamba-3 comparison.
• results/: compact benchmark outputs and markdown summaries that are useful to keep with the experiment code.
• scripts/: local utilities, currently including PDF-to-markdown conversion based on pdftotext -layout.
• papers/FALCON/*.md: local markdown notes derived from the FALCON paper.

What is intentionally not tracked

• data/: CIFAR-10 downloads and extracted dataset files.
• mamba_build.log: local build output.
• .venv/, __pycache__/, *.pyc: local environment and caches.
• papers/**/*.pdf: source PDFs can be re-downloaded and do not add much value to the repository history.
• state-spaces-mamba/ and vit5-baseline/: these are upstream dependency repos checked out locally, not source owned by this repo.

External dependencies

The benchmark script expects these sibling directories at the repository root:

git clone https://github.com/wangf3014/ViT-5.git vit5-baseline
git -C vit5-baseline checkout b7b0796

git clone https://github.com/state-spaces/mamba.git state-spaces-mamba
git -C state-spaces-mamba checkout mamba3-release

Notes on the current local dependency state:

• vit5-baseline is a clean checkout at commit b7b0796.
• state-spaces-mamba is based on origin/mamba3-release commit 5235bdc, with additional local FALCON work in commit 7fdaf53 plus further uncommitted edits. That nested repo is still in active development, so it is excluded from this initial top-level push instead of being snapshotted here.

Running the benchmark

Example:

python benchmarks/cifar10_vit5_mamba3.py \
  --data-dir data \
  --output-json results/cifar10_vit5_mamba3_results.json \
  --output-md results/cifar10_vit5_mamba3_results.md