Ishannaik

/

jax-efficient-training

Like 0

Hugging Face Community GPU Grant request

Applicant: Ishan Naik (independent researcher, India)
Affiliation: INDIGITAL
Space: https://huggingface.co/spaces/Ishannaik/jax-efficient-training-demo
Model repo: https://huggingface.co/Ishannaik/jax-efficient-training
License: Apache-2.0

What I am building

A public JAX/Flax demo of memory-efficient training: rematerialization / gradient checkpointing, mixed precision, and a small-model fine-tune so other independent researchers can reproduce the recipe without a university cluster.

Why a Community GPU is needed

Hugging Face now requires Pro for free Gradio CPU Spaces. A Community GPU grant would let me host an interactive JAX demo that visitors can run.

90-day plan

1. Ship a reproducible JAX/Flax training loop + memory/throughput notes.
2. Add a public notebook comparing checkpointed vs full backprop memory.
3. Keep the Space and code Apache-2.0 so others can fork it.

Thank you for considering this Community GPU grant.