libane 0.8.2

Run ML graphs directly on the Apple Neural Engine from Python — Graph IR, op fusion, and ANE-native dispatch

libane

Direct access to the Apple Neural Engine from C++ and Python.

libane is a low-level ANE runtime and compiler interface. It exposes a Graph IR for describing full forward passes, compiles them into the minimum number of ANE dispatches via automatic op fusion, and executes them through a stable C ABI. Matmul is implemented as conv1×1 for the 3× throughput advantage over MIL matmul on ANE.

Private framework dependency. libane loads AppleNeuralEngine.framework via dlopen. It is intended for research use and low-level ANE experimentation, not production deployment.

---

Install

pip install libane

Requires Apple Silicon (M1 or later) and macOS 14+. The wheel is a compiled extension — no extra build steps needed.

---

Requirements

Hardware	Apple Silicon (M1 or later)
OS	macOS 14 Sonoma or later
Toolchain	Xcode 15+, CMake 3.24+, C++17
Python	3.10+ (optional)

---

Python quick-start

import ane
import numpy as np

print(ane.available())   # True on Apple Silicon
print(ane.version())     # "0.8.2"

# Single-op matmul
A = np.random.randn(128, 512).astype(np.float16)
B = np.random.randn(512, 256).astype(np.float16)
C = ane.matmul(A, B)    # shape (128, 256), fp16

# Graph API — fused FFN block
D, FFN, SEQ = 512, 2048, 128
W_up   = np.random.randn(D,   FFN).astype(np.float16)
W_down = np.random.randn(FFN, D  ).astype(np.float16)
scale  = np.ones(D, dtype=np.float16)

g = ane.Graph()
x   = g.add_input("x",  [1, D,   1, SEQ])
rn  = g.add_op(ane.RMSNORM, [x],   [1, D,   1, SEQ], weights=scale)
up  = g.add_op(ane.MATMUL,  [rn],  [1, FFN, 1, SEQ], weights=W_up)
act = g.add_op(ane.GELU,    [up],  [1, FFN, 1, SEQ])
out = g.add_op(ane.MATMUL,  [act], [1, D,   1, SEQ], weights=W_down)
g.mark_output(out)

cg = g.compile()
cg.set_output_shapes([[1, D, 1, SEQ]])

x_data = np.random.randn(D, SEQ).astype(np.float16)
result = cg(x_data)
print(result.shape)   # (1, 512, 1, 128)

See examples/ffn_inference.py for a timed end-to-end example.

---

Documentation

C API reference	All libane_* functions, types, and status codes
Python API reference	ane.*, Graph, CompiledGraph, CompiledMil
Graph IR	Tensor layout, op table, fusion rules, shape limits
Hardware introspection	Device info, shape limits, performance stats

---

Building from source

git clone https://github.com/AmiraniLabs/libane
cd libane
cmake -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo
cmake --build build -j$(sysctl -n hw.logicalcpu)
ctest --test-dir build --output-on-failure

To build the Python module from source:

pip install pybind11 numpy scikit-build-core
pip install ./bindings/python

---

C quick-start

Single op

#include "libane.h"

// Compile a matmul [1, 512, 1, 128] → [1, 256, 1, 128]
libane_shape_t shape = {.dims = {1, 256, 1, 128}, .ndim = 4};
libane_handle_t h = libane_compile(LIBANE_OP_MATMUL, shape,
                                    fp16_weights, sizeof(fp16_weights));
libane_execute(h, input, output, shape);
libane_release(h);

Graph API (fused multi-op)

#include "libane.h"

// Build: RMSNorm → Matmul → GELU
libane_graph_t g = libane_graph_create();

libane_shape_t in_shape  = {.dims={1,512,1,128}, .ndim=4};
libane_shape_t out_shape = {.dims={1,256,1,128}, .ndim=4};

uint32_t x  = libane_graph_add_input(g, "x", in_shape);
uint32_t rn = libane_graph_add_op(g, LIBANE_OP_RMSNORM, &x, 1,
                                   in_shape, rn_scale, rn_scale_len);
uint32_t mm = libane_graph_add_op(g, LIBANE_OP_MATMUL,  &rn, 1,
                                   out_shape, weights, weights_len);
uint32_t act= libane_graph_add_op(g, LIBANE_OP_GELU,    &mm, 1,
                                   out_shape, NULL, 0);
libane_graph_mark_output(g, act, "out");

libane_compiled_graph_t cg = libane_graph_compile(g);

// Execute
const void* in_ptrs[]   = { input_fp16 };
size_t      in_bytes[]  = { 512 * 128 * 2 };
void*       out_ptrs[]  = { output_fp16 };
size_t      out_bytes[] = { 256 * 128 * 2 };
libane_graph_execute(cg, in_ptrs, in_bytes, 1, out_ptrs, out_bytes, 1);

libane_compiled_graph_release(cg);
libane_graph_release(g);

---

Architecture

libane
├── include/libane.h          Stable C ABI
├── src/
│   ├── core/
│   │   ├── mil_builder       MIL text program generator
│   │   ├── compile_cache     LRU cache for compiled programs
│   │   └── buffer_manager    IOSurface-backed fp16 buffer pool
│   ├── graph/
│   │   ├── ane_graph         Graph IR (DAG builder)
│   │   ├── graph_validator   7-check validation pass
│   │   ├── fusion_rules      Greedy linear-chain fusion
│   │   ├── graph_compiler    build_plan() + compile()
│   │   ├── graph_executor    Per-group ANE dispatch
│   │   ├── mil_backend       MIL-path graph lowering
│   │   ├── hwx_backend       HWX-path graph lowering
│   │   ├── hwx_emitter       HWX bytecode emission
│   │   ├── espresso_backend  Espresso-path graph lowering
│   │   └── espresso_builder  Espresso program construction
│   ├── runtime/
│   │   └── ane_runtime.mm    AppleNeuralEngine.framework wrapper
│   └── fallback/             Accelerate BLAS CPU fallback
└── bindings/python/          pybind11 module (pip install libane)

Fusion rules compact linear chains into single ANE programs. A 6-op FFN (RMSNorm → Matmul → SiLU, Matmul → Mul → Matmul) typically becomes 3–4 ANE dispatches instead of 6, eliminating intermediate DRAM round-trips.

---

Supported ops

Category	Ops
Linear	MATMUL (conv1×1; ~3× faster than MIL matmul)
Normalization	RMSNORM, LAYERNORM
Activations	GELU, SILU, RELU, TANH, SIGMOID, HARDSWISH, LEAKY_RELU, ELU, PWL_ACTIVATION
Elementwise	ADD, SUB, MUL, REAL_DIV, NEG, MOD
Math	SQRT, LOG, RSQRT, SINH, COSH, TAN, ASIN, ACOS
Reduce	SOFTMAX, REDUCE_SUM, REDUCE_MEAN, REDUCE_MAX, REDUCE_PROD
Pooling	AVG_POOL, MAX_POOL
Structural	TRANSPOSE, RESHAPE, CONCAT, SLICE_BY_INDEX
Logical	LOGICAL_AND, LOGICAL_OR, LOGICAL_XOR
Scatter/Gather	SCATTER, GATHER, SCATTER_ND, SCATTER_ALONG_AXIS

Full op documentation with constraints and notes: docs/graph-ir.md.

---

Known Limitations

• Experimental / research-use only. Not production-supported.
• Private Apple framework dependency. Uses AppleNeuralEngine.framework via dlopen. Uses private API — intentional.
• Constrained tensor layout. Graph API requires [1, C, 1, S] (NCHW with batch=1, height=1). Arbitrary shapes are not supported.
• Channel cap. Graph API validation enforces C ≤ 16384. Larger channel counts (e.g. vocabulary projections) require raw MIL emission and are not exposed through the graph API.
• fp16 only. No quantization (int8, int4) support. Weights and activations are fp16 throughout.
• Some ops are compiler-sensitive. ANE's MIL compiler accepts a strict subset of MIL. Certain op combinations or shapes may require fallback paths. See the fallback module.
• Not a general-purpose model runner. libane is a programmable kernel/runtime/compiler interface for ANE-native experimentation, not a drop-in inference engine.
• macOS 14+ required. Older systems are not tested and not supported.

---

License

Apache 2.0. See LICENSE.

libane is not affiliated with or endorsed by Apple Inc.