Metadata-Version: 2.4
Name: mhc-mlx
Version: 0.4.7
Summary: High-performance MLX implementation of Manifold-Constrained Hyper-Connections (mHC)
Author-email: Salvador Escobedo <svdrecbd@gmail.com>
License: MIT License
        
        Copyright (c) 2026 Salvador 
        
        Permission is hereby granted, free of charge, to any person obtaining a copy
        of this software and associated documentation files (the "Software"), to deal
        in the Software without restriction, including without limitation the rights
        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
        copies of the Software, and to permit persons to whom the Software is
        furnished to do so, subject to the following conditions:
        
        The above copyright notice and this permission notice shall be included in all
        copies or substantial portions of the Software.
        
        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
        SOFTWARE.
        
Project-URL: Homepage, https://github.com/svdrecbd/mhc-mlx
Project-URL: Repository, https://github.com/svdrecbd/mhc-mlx
Project-URL: Issues, https://github.com/svdrecbd/mhc-mlx/issues
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: MacOS :: MacOS X
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.10
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: mlx>=0.30.0
Provides-Extra: dev
Requires-Dist: pytest>=7.4; extra == "dev"
Requires-Dist: build; extra == "dev"
Requires-Dist: twine; extra == "dev"
Requires-Dist: ruff; extra == "dev"
Provides-Extra: bench
Requires-Dist: matplotlib; extra == "bench"
Dynamic: license-file

# mhc-mlx

**High-performance MLX implementation of Manifold-Constrained Hyper-Connections (mHC)** for Apple Silicon.

This library provides a drop-in `MHCLayer` that fuses multiple operations into optimized Metal kernels, achieving massive speedups over compiled reference layers and standard Python-based implementations.

**Original Paper:** [mHC: Manifold-Constrained Hyper-Connections](https://arxiv.org/abs/2512.24880) (DeepSeek-AI)

## Installation

Install from PyPI:

```bash
pip install mhc-mlx
```

## Quick Start

### Option 1: Drop-in Layer (Recommended)
Use `MHCLayer` for maximum performance.

```python
import mlx.core as mx
from mhc_mlx import MHCLayer

layer = MHCLayer(n=32, C=64) # 32 streams, 64 channels each
x = mx.random.normal((1, 32, 64))
y = layer(x)
```

### Option 2: Universal Wrapper (MHCRewire)
Enhance **any** existing MLX module (Linear, Conv2d, Transformers) with manifold-constrained stability. *Note: optimizing arbitrary modules incurs some overhead compared to the fused MHCLayer.*

```python
import mlx.nn as nn
from mhc_mlx import MHCRewire

# Wrap a standard Linear layer
layer = MHCRewire(nn.Linear(512, 512), dims=512, n=16)
```

## Performance

We benchmarked on an Apple M4 Pro (macOS 15.6). `mhc-mlx` outperforms standard implementations across all scales.

### Head-to-Head: mhc-mlx vs mlx-mhc (Competitor)

| Scenario | mhc-mlx (ours) | mlx-mhc (them) | Speedup |
|---|---|---|---|
| **Latency** ($B=1, C=512$) | **435 us** | 1031 us | **2.37x** |
| **Throughput** ($B=32, C=512$) | **89 us/iter** | 940 us/iter | **10.53x** |
| **Throughput** ($B=32, C=2048$) | **243 us/iter** | 1122 us/iter | **4.61x** |

### Why We're Faster

| Implementation | Characteristics | Performance Impact |
|---|---|---|
| **Python / JIT** | Many small kernel launches | Higher overhead, low occupancy |
| **Fused Metal** | 1-3 highly optimized kernels | Minimal overhead, maximum bandwidth |

### Latency Floor ($B=1$, Sequence Length=32)

| Channels (C) | Kernel Strategy | Layer Speedup (vs Compiled MLX) |
|---|---|---|
| 256 | Fully Fused | **2.27x** |
| 1024 | Fully Fused | **1.57x** |
| 2048 | Fully Fused | **1.58x** |
| 4096 | Column Parallel | **1.41x** |
| 8192 | Column Parallel | **2.18x** |

### High Throughput ($B=32$, Sequence Length=32)

Maximum speedups for heavy data processing.

| Operation | Scale (n, C) | Peak Speedup |
|---|---|---|
| **Sinkhorn-Knopp** | n=4 | **26.99x** |
| **Mix + Add (Fused)** | n=32, C=2048 | **14.92x** |
| **Full MHCLayer** | n=4, C=4096 | **17.33x** |

### Training / Backward Pass

Optimized gradients ensure training is as fast as inference.

| Batch Size | Channels (C) | Speedup vs Compiled MLX |
|---|---|---|
| 1 | 2048 | **4.18x** |
| 1 | 4096 | **2.12x** |
| 32 | 2048 | **3.10x** |

*(Benchmarks run with bfloat16)*

## Key Optimizations

- **"Zero-Cost" Weight Folding:** `MHCRewire` folds scaling directly into `nn.Linear` weights, eliminating pre-scaling overhead.
- **Quantized Layer Support:** Seamlessly wraps `nn.QuantizedLinear` (4-bit/8-bit) for efficient local LLM inference.
- **Fully Fused Kernel:** Single kernel for Aggregate + RMS + Mix + Add.
- **Column-Parallel Mixing:** Vectorized kernel maximizing throughput for larger workloads.
- **Adaptive Dispatch:** Runtime heuristic selects the fastest kernel strategy.
- **Super-Fused Backward:** Fused gradients for maximum training efficiency.

## Advanced Usage

### Auto-Tuning for Your Hardware

`mhc-mlx` comes with a JIT kernel auto-tuner that finds the optimal threadgroup sizes for your specific Mac.

```bash
# Run the tuner (takes ~30s)
python scripts/tune.py
```
This generates a `mhc_tuning.json` file. The library will automatically load this config to squeeze out the last 5-10% of performance.

### Custom Blocks: Fused Residual Add + Aggregate

If you are building custom Transformer blocks, you can use `residual_add_agg` to fuse the residual connection with the mHC aggregation step. This saves a full memory read/write round-trip (~1.4x speedup).

```python
from mhc_mlx import residual_add_agg

# Standard: x = x + res; y_agg = aggregate(x)
# Fused:
x, y_agg = residual_add_agg(x, res, H_pre)
```

## Troubleshooting

Run diagnostics to check your environment:
```bash
mhc-mlx-info
```

## License

MIT
