mhc-mlx 0.4.7

High-performance MLX implementation of Manifold-Constrained Hyper-Connections (mHC)

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 (DeepSeek-AI)

Installation

Install from PyPI:

pip install mhc-mlx

Quick Start

Option 1: Drop-in Layer (Recommended)

Use MHCLayer for maximum performance.

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.

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.

# 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).

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:

mhc-mlx-info

License

MIT