chore: final hardening and documentation polish for release

- Robust kernel path discovery using importlib.resources
- Added platform guardrails and import-time checks
- Added MHC_MLX_DISABLE_METAL escape hatch
- Neutral README tone and reproducible benchmarks
- Added CONTRIBUTING.md and SECURITY.md
- Bumped version to 0.5.4

Author: svdrecbd

CONTRIBUTING.md

@@ -0,0 +1,39 @@
+# Contributing to mhc-mlx
+
+We welcome contributions to improve speed, stability, and compatibility!
+
+## Development Setup
+
+1. Clone the repo:
+   ```bash
+   git clone https://github.com/svdrecbd/mhc-mlx.git
+   cd mhc-mlx
+   ```
+
+2. Install dependencies:
+   ```bash
+   pip install -e ".[dev,bench]"
+   ```
+
+## Running Tests
+
+Run the full suite:
+```bash
+python -m pytest
+```
+
+Note: Some tests require Apple Silicon to execute the Metal paths.
+
+## Benchmarking
+
+Run the main benchmark:
+```bash
+PYTHONPATH=. python mhc_mlx/benchmark.py --mode latency
+```
+
+## Auto-Tuning
+
+If you've modified kernels, run the tuner:
+```bash
+PYTHONPATH=. python scripts/tune.py
+```

README.md

@@ -2,7 +2,7 @@
 
 **High-performance MLX implementation of Manifold-Constrained Hyper-Connections (mHC)** for Apple Silicon.
 
-mHC improves training stability and performance in deep architectures by constraining residual connections to the Birkhoff polytope (doubly stochastic matrices). This library provides optimized Metal kernels that achieve massive speedups over standard Python-based implementations.
+mHC improves training stability and performance in deep architectures by constraining residual connections to the Birkhoff polytope (doubly stochastic matrices). This library provides optimized Metal kernels that achieve significant speedups over standard baseline implementations.
 
 **Original Paper:** [mHC: Manifold-Constrained Hyper-Connections](https://arxiv.org/abs/2512.24880) (DeepSeek-AI)
 
@@ -15,7 +15,7 @@ pip install mhc-mlx
 ## Compatibility
 - **Hardware:** Apple Silicon (M1, M2, M3, M4).
 - **Software:** macOS, MLX >= 0.30.0.
-- **Fallback:** Automatically falls back to a compiled pure-MLX path if Metal kernels are unavailable.
+- **Fallback:** Automatically falls back to a compiled pure-MLX path on other platforms.
 
 ## Quick Start (30-second Demo)
 
@@ -25,14 +25,13 @@ import mlx.nn as nn
 from mhc_mlx import MHCRewire
 
 # 1. Take any standard MLX layer
-layer = nn.Linear(512, 512)
+layer = nn.Linear(2048, 2048)
 
 # 2. Wrap it with mHC stability (automatically uses optimized Metal kernels)
-# This computes: H_post * (Linear(H_pre * x) + M * H_pre * x)
-model = MHCRewire(layer, dims=512, n=16)
+model = MHCRewire(layer, dims=2048, n=32)
 
 # 3. Run forward pass
-x = mx.random.normal((1, 512))
+x = mx.random.normal((1, 2048))
 y = model(x)
 mx.eval(y)
 
@@ -41,7 +40,7 @@ loss_fn = lambda m, x: mx.sum(m(x))
 grads = mx.grad(loss_fn)(model, x)
 mx.eval(grads)
 
-print(f"Output shape: {y.shape}") # (1, 512)
+print(f"Output shape: {y.shape}") # (1, 2048)
 ```
 
 *Note: You can also use `from mlx_mhc import MHCRewire` for a community-friendly alias.*
@@ -52,9 +51,9 @@ print(f"Output shape: {y.shape}") # (1, 512)
 
 ### Comparative Benchmarks
 
-Comparison with other standard MLX implementations of mHC ($C=512$):
+Comparison with a standard MLX implementation of mHC ($C=512$):
 
-| Metric | mhc-mlx (Ours) | Standard Impl | Speedup |
+| Metric | mhc-mlx | Baseline Impl | Speedup |
 |---|---|---|---|
 | **Inference Latency** ($B=1$) | **392 us** | 1120 us | **2.86x** |
 | **Training Throughput** ($B=32$) | **105 us** | 866 us | **8.25x** |
@@ -63,7 +62,7 @@ Comparison with other standard MLX implementations of mHC ($C=512$):
 
 | Approach | Architecture | Impact |
 |---|---|---|
-| **Standard** | Multiple kernel launches | High memory overhead, low GPU occupancy |
+| **Baseline** | Multiple kernel launches | High memory overhead, low GPU occupancy |
 | **mhc-mlx** | Fused Metal Kernels | Minimal memory round-trips, maximal bandwidth |
 
 ### Reproduce Benchmarks
@@ -86,4 +85,4 @@ mhc-mlx-info
 ```
 
 ## License
-MIT
+MIT

SECURITY.md

@@ -0,0 +1,7 @@
+# Security Policy
+
+## Reporting a Vulnerability
+
+If you discover a security vulnerability within this project, please open an Issue on GitHub or contact the maintainer directly if sensitive.
+
+We aim to address all security concerns promptly. As this is a research-focused implementation, the primary surface is the Metal kernel execution and model weight handling.

mhc_mlx/__init__.py

@@ -3,6 +3,15 @@
 from .utils import residual_add_agg
 from .patching import AutoPatcher
 from importlib.metadata import version, PackageNotFoundError
+import platform
+import warnings
+
+# Platform check
+if platform.system() != "Darwin":
+    warnings.warn(
+        "mhc-mlx is optimized for macOS + Apple Silicon (Metal). "
+        "Falling back to pure-MLX path on this platform."
+    )
 
 try:
     __version__ = version("mhc-mlx")

mhc_mlx/kernels/__init__.py

@@ -22,6 +22,7 @@
 
 import mlx.core as mx
 import mlx.nn as nn
+import os
 
 from .metal import (
     mhc_forward_fused_metal_autograd,
@@ -152,6 +153,8 @@ def mixing_matrix(self) -> mx.array:
         return mixing_matrix_from_logits(self.H_res_raw, iters=self.sinkhorn_iters, eps=self.eps)
 
     def _should_use_metal(self, B: int, n: int, C: int) -> bool:
+        if os.getenv("MHC_MLX_DISABLE_METAL", "0") == "1":
+            return False
         if not self.use_metal:
             return False
         if not self.auto_dispatch:

mhc_mlx/layer.py

@@ -15,6 +15,7 @@
 from functools import lru_cache
 
 import mlx.core as mx
+from importlib import resources
 
 # Config loading
 _TUNING_CONFIG = {}
@@ -34,7 +35,13 @@
 def _get_tuned_tpg(kernel_name: str, default: int) -> int:
     return _TUNING_CONFIG.get(kernel_name, default)
 
-_KERNEL_DIR = os.path.join(os.path.dirname(__file__), "kernels")
+# Robust kernel path discovery using importlib.resources
+try:
+    _KERNEL_DIR = str(resources.files("mhc_mlx") / "kernels")
+except Exception:
+    # Fallback for older python or non-package installs
+    _KERNEL_DIR = os.path.join(os.path.dirname(__file__), "kernels")
+
 _STREAM_MIX_ADD_PATH = os.path.join(_KERNEL_DIR, "stream_mix_add.metal")
 _STREAM_MIX_ADD_RMS_PATH = os.path.join(_KERNEL_DIR, "stream_mix_add_rms.metal")
 _STREAM_MIX_ADD_RMS_FP16_PATH = os.path.join(_KERNEL_DIR, "stream_mix_add_rms_fp16.metal")

mhc_mlx/metal.py

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "mhc-mlx"
-version = "0.5.3"
+version = "0.5.4"
 description = "High-performance MLX implementation of Manifold-Constrained Hyper-Connections (mHC)"
 requires-python = ">=3.10"
 readme = "README.md"
@@ -23,7 +23,7 @@ classifiers = [
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
 dependencies = [
-  "mlx>=0.30.0",
+  "mlx>=0.30.0; platform_system == 'Darwin'",
 ]
 
 [project.urls]
@@ -47,7 +47,7 @@ mhc-mlx-info = "mhc_mlx.diagnostics:main"
 mhc-mlx-bench = "mhc_mlx.benchmark:main"
 
 [tool.setuptools]
-packages = ["mhc_mlx", "mlx_mhc"]
+packages = ["mhc_mlx", "mhc_mlx.kernels", "mlx_mhc"]
 include-package-data = true
 
 [tool.setuptools.package-data]

pyproject.toml

@@ -15,7 +15,7 @@ def test_kernels_present():
     # This works for python 3.10+
     try:
         kernel_files = [
-            p.name for p in resources.files("mhc_mlx.kernels").iterdir() 
+            p.name for p in (resources.files("mhc_mlx") / "kernels").iterdir() 
             if p.name.endswith(".metal")
         ]
     except (AttributeError, TypeError):