Add nano-CellFM as a community reference implementation

[huynguyen250896]

What is nano-CellFM?

I recently built nano-CellFM, a lightweight and faithful PyTorch reimplementation of CellFM designed to make the core implementation easier to read, reproduce, benchmark, and extend while preserving the original architecture and inference behavior.

Repository:
https://github.com/huynguyen250896/nano-CellFM

Highlights

• Pure PyTorch implementation with minimal dependencies
• Reproduces the official CellFM cell- and gene-level embeddings from the released CellFM-80M checkpoint
• Cleaner, easy-to-read implementation suitable for learning, benchmarking, experimentation, and future development

Validation

I carefully validated nano-CellFM against the official implementation to ensure numerical consistency.

Specifically, nano-CellFM:

• reproduces the official cell embeddings with mean cosine similarity ≈ 1.000000
• reproduces the official gene embeddings with mean cosine similarity ≈ 1.000000
• preserves the representation geometry of the official embeddings
• includes reproducibility and inference benchmark notebooks in the repository

Since the official CellFM implementation currently depends on MindSpore and Ascend-oriented tooling, I did not benchmark inference runtime directly against the official implementation. Instead, the repository focuses on providing a fully reproducible PyTorch implementation while reporting inference performance within nano-CellFM itself.

Why this PR?

The goal of nano-CellFM is not to replace the official implementation, but to provide a lightweight community resource for users who prefer a pure PyTorch implementation for reproducibility, benchmarking, learning, and research. All credit belongs to the official CellFM authors for the original model and its development.

This PR only adds a link under Community Projects. It does not modify any code, pretrained models, datasets, checkpoints, or model behavior.

Thank you for your consideration.