remora-layer-lab / README.md
Gerald Corzo
feat: remora PyTorch ZeroGPU layer-control research space
446eaf9
|
Raw
History Blame Contribute Delete
2.18 kB

A newer version of the Gradio SDK is available: 6.26.0

Upgrade
metadata
title: Remora Layer Lab
emoji: 🐟
colorFrom: blue
colorTo: indigo
sdk: gradio
python_version: '3.12'
sdk_version: 5.49.1
app_file: app.py
models:
  - LiquidAI/LFM2.5-1.2B-Instruct
tags:
  - remora
  - layer-control
  - lfm2
  - zerogpu

Remora Layer Lab

A research-only Hugging Face ZeroGPU Space that runs LiquidAI/LFM2.5-1.2B-Instruct and applies explicit, reproducible controls to selected hidden-state layers.

What this is

  • A PyTorch experimental adapter for the existing Remora research method.

  • Captures per-layer hidden-state statistics for each generation run.

  • Applies a scalar gain to selected decoder-block outputs:

    controlled = baseline + strength × (baseline - token_mean)

    • strength = 0: observation only; output is unchanged.
    • negative strength: damp the selected residual signal.
    • positive strength: amplify it.
  • Produces a JSONL trace that can be compared with the Rust/llama.cpp Remora traces.

What this is not

This is not the Rust production lane and does not replace local remora on Janus. The HF ZeroGPU Space is a bounded research surface: Gradio/PyTorch only, shared GPU allocation, daily quota, and no persistent runtime disk. The production route remains octopus → local lfm-fast.

Run protocol

  1. Use the default strength = 0 to create a baseline trace.
  2. Repeat the same prompt with one selected layer and one non-zero strength.
  3. Compare output and trace statistics outside the Space. Save the downloaded JSONL with the prompt suite and run metadata.
  4. Keep a request below the declared GPU duration. Start with 32 output tokens.

Cost boundary

For a PRO account, ZeroGPU has a 40-minute daily quota. Usage after the quota costs $1 per 10 minutes from pre-paid HF credits. The Space uses large (48 GB, one quota unit), not xlarge.

Local checks

python -m unittest discover -s tests -v

Deploy

Create an HF Gradio ZeroGPU Space called gcorzo/remora-layer-lab, select ZeroGPU hardware in its Settings, then push this directory to the Space repository. The HF token used for this must have Space write permission; no token is stored in this repository.