mhc-algorithm

Files (6)

• SKILL.md
• core-concepts.md · references/
• gpt-integration.md · references/
• module-implementation.md · references/
• pitfalls.md · references/
• sinkhorn-knopp.md · references/

SKILL.md (first lines)

# mHC: Manifold-Constrained Hyper-Connections

## Overview

mHC (Manifold-Constrained Hyper-Connections) stabilizes deep network training by constraining residual mixing matrices to be **doubly stochastic**. It provides:

- **Stable Training**: Lower gradient norm variance via doubly stochastic constraints
- **Multiple Streams**: Hyper-Connections with learnable mixing across residual streams
- **Sinkhorn Projection**: Log-space Sinkhorn-Knopp algorithm for doubly stochastic projection
- **GPT Integration**: Pattern for wrapping attention and MLP layers

Two components:
- **HyperConnections Module**: Core PyTorch module with H_res, H_pre, H_post matrices
- **Sinkhorn-Knopp**: Log-space projection to doubly stochastic manifold

## Quick Reference

| Topic | Reference |
|-------|-----------|
| Core Concepts & Math | [Core Concepts](references/core-concepts.md) |
| Sinkhorn Algorithm | [Sinkhorn-Knopp](references/sinkhorn-knopp.md) |
| HyperConnections Module | [Module Implementation](references/module-implementation.md) |
| GPT Integration | [GPT Integration](references/gpt-integration.md) |
| Common Pitfalls | [Pitfalls](references/pitfalls.md) |

## Installation

```python
# Required packages
pip install torch einops numpy
```

## Minimal Example

```python
import torch
import torch.nn as nn
from einops import rearrange, einsum

def sinkhorn_knopp(logits, num_iters=20, tau=0.05):
    log_alpha = logits / tau
    for _ in range(num_iters):
        log_alpha = log_alpha - torch.logsumexp(log_alpha, dim=-1, keepdim=True)
        log_alpha = log_alpha - torch.logsumexp(log_alpha, dim=-2, keepdim=True)
    return torch.exp(log_alpha)

class HyperConnections(nn.Module):
    def __init__(self, num_streams, dim, branch=None, layer_idx=0):
        super().__init__()
        self.num_streams = num_streams
        self.branch = branch

        # Initialize H_res near identity (use small negative for gradient flow)
        init_h_res = torch.full((num_streams, num_streams), -0.1)
        init_h_res.fill_diagonal_(0.0)
        self.H_res_logits = nn.Parameter(init_h_res)

        # H_pre/H_post for depth connections
        init_h_pre = torch.full((1, num_streams), -0.1)
        init_h_pre[0, layer_idx % num_streams] = 0.0
        self.H_pre_logits = nn.Parameter(init_h_pre)
        self.H_post_logits = nn.Parameter(torch.zeros(1, num_streams))

    def forward(self, x):
        s = self.num_streams
        x = rearrange(x, "(b s) t d -> b t s d", s=s)

        h_res = sinkhorn_knopp(self.H_res_logits)
        x_mixed = einsum(h_res, x, "s t, b n s d -> b n t d")

        h_pre = self.H_pre_logits.softmax(dim=-1)
        branch_in = einsum(h_pre, x, "v s, b n s d -> b n v d").squeeze(-2)

        branch_out = self.branch(branch_in) if self.branch else branch_in

        h_post = self.H_post_logits.softmax(dim=-1)
        depth_out = einsum(branch_out, h_post, "b t d, v s -> b t s d")

        output = x_mixed + depth_out
        return rearrange(output, "b t s d -> (b s) t d")
```

## Common Imports

```python
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, einsum, repeat, reduce
```

## When to Use What

| Scenario | Approach |
|----------|----------|
| Standard residual connection | No mHC needed |
| Deep networks (>12 layers) with stability issues | Use mHC with num_streams=4 |
| GPT/Transformer training | Wrap both attention and MLP with HyperConnections |
| Custom Sinkhorn iterations | Adjust num_iters (20 default) and tau (0.05 default) |
| Memory-constrained training | Reduce num_streams or batch size |

## External Resources

- mHC Paper: https://arxiv.org/abs/2512.24880
- Hyper-Connections: https://arxiv.org/abs/2409.19606
- Sinkhorn's Theorem: https://en.wikipedia.org/wiki/Sinkhorn%27s_theorem

From benchflow-ai/skillsbench, Apache-2.0.

More from benchflow-ai/skillsbench