mhc

AI Community

Skill notes

Contents

Examples Full JAX implementation of sinkhorn_knopp and mhc_layer_forward. Deep Theory Motivation, stability proofs, and scalability arguments.

Usage

Use this skill when implementing Deep Transformers (1000+ layers) where standard residual connections fail (Gradient Vanishing, Representation Collapse).

Fluxly profile Runtime and permissions are inferred by Fluxly static checks, not author declarations or runtime test results.
Domain
AI
Author-declared agents
No explicit declaration found; this is not inferred or tested compatibility
Static check
88 / 100 · heuristic scan, not runtime safety proof
Runtime evidence
Not run; no isolated runtime test evidence is currently recorded
Author / version / license
@diegosouzapw · no license declared
Token usage
Lean
Setup complexity
Plug-and-play
External API key
Not required
Operating systems
Unspecified (assume cross-platform)
Runtime requirements
Python
Detected file/system behavior
  • Read-only
Network behavior
Local-only
Install commands
None (reference only)

Heads up: Tool permissions are not declared; actual access depends on the runtime.

Author source file SKILL.md
# mHC Skill

Manifold-Constrained Hyper-Connections (mHC) uses Doubly Stochastic Matrices to improve Deep Learning stability.

## Contents

- [Examples](examples.md)
    - Full JAX implementation of `sinkhorn_knopp` and `mhc_layer_forward`.
- [Deep Theory](reference.md)
    - Motivation, stability proofs, and scalability arguments.

## Usage

Use this skill when implementing Deep Transformers (1000+ layers) where standard residual connections fail (Gradient Vanishing, Representation Collapse).

```python
# Quick Ref: Sinkhorn-Knopp (See examples.md for full context)
def sinkhorn_knopp(log_matrix, n_iters=20):
    M = jnp.exp(log_matrix)
    def body(i, m):
        m /= m.sum(axis=1, keepdims=True)
        m /= m.sum(axis=0, keepdims=True)
        return m
    return jax.lax.fori_loop(0, n_iters, body, M)
```

Discussion

Powered by GitHub Discussions. Sign in with GitHub to comment, react, or subscribe.