litespark-inference 1.0.1

Efficient CPU inference for ternary language models

Litespark-Inference

Fast CPU inference for ternary neural networks

Litespark-Inference is a pip-installable Python library that enables efficient inference of ternary models on consumer CPUs. By exploiting the ternary weight structure ({-1, 0, +1}) with custom SIMD kernels, we eliminate floating-point multiplication entirely and achieve dramatic speedups over standard PyTorch inference.

Key Results

Apple Silicon (M1-M5)

Performance comparison on Apple Silicon M5 Max. Litespark-Inference achieves 6.03x memory reduction, 7.15x faster TTFT, and 18.15x higher throughput compared to PyTorch.

Metric	PyTorch	NEON	Improvement
Memory (MB)	4,868.22	806.81	6.03x
TTFT (ms)	4,213.92	589.02	7.15x
Throughput (tok/s)	2.20	39.92	18.15x

Intel Ice Lake / AMD Zen4 (AVX-512 VNNI)

Performance comparison on Intel Ice Lake / AMD Zen4 using AVX-512 VNNI kernels.

Metric	PyTorch	AVX-512 VNNI	Improvement
Memory (MB)	4,892.38	789.88	6.19x
TTFT (ms)	6,647.18	1,167.26	5.69x
Throughput (tok/s)	0.43	41.20	95.81x

Intel Core Ultra (AVX-VNNI)

Performance comparison on Intel Core Ultra using AVX-VNNI kernels.

Metric	PyTorch	AVX-VNNI	Improvement
Memory (MB)	4,601.55	775.84	5.93x
TTFT (ms)	7,173.05	1,134.48	6.32x
Throughput (tok/s)	0.41	39.96	97.46x

Cross-Platform Comparison

Cross-platform performance comparison showing TTFT, throughput, and memory-efficiency improvements across Apple Silicon, Intel, and AMD processors.

Energy Consumption

System	Metric	PyTorch	Litespark	Improvement
Apple M5 Max	Total energy (J)	606.46	101.45	5.98x
Apple M5 Max	Energy/token (J)	4.74	0.79	5.98x
AMD Ryzen Threadripper PRO 5965WX	Total energy (J)	12,173.53	957.44	12.71x
AMD Ryzen Threadripper PRO 5965WX	Energy/token (J)	95.11	7.48	12.71x

Thread Scaling

We also measured Litespark-Inference with the pp128+tg128 protocol across thread counts, separating prompt prefill from autoregressive token generation.

AMD EPYC 9R14 (AWS c7a.4xlarge)

Threads	Prefill pp128 (tok/s)	Generation tg128 (tok/s)
1	520.96	7.67
2	492.16	14.56
4	513.08	25.42
8	529.36	40.49
10	523.91	44.86
16	521.59	52.49

Intel Xeon Platinum 8488C (AWS c7i.4xlarge)

Threads	Prefill pp128 (tok/s)	Generation tg128 (tok/s)
1	102.30	6.32
2	105.11	11.02
4	112.38	16.93
8	120.44	25.70
10	135.73	23.25
16	131.43	30.43

Apple M5 Max (MacBook Pro)

Litespark-Inference scaling on Apple M5 Max. Prefill throughput continues scaling through 16 threads, while token generation gains quickly and then flattens out.

Threads	Prefill pp128 (tok/s)	Generation tg128 (tok/s)
1	50.52	10.98
2	91.55	19.46
4	152.09	33.29
8	194.23	35.81
10	218.88	37.44
16	262.59	37.92

Supported Platforms

• Apple Silicon (M1/M2/M3/M4/M5) — NEON SDOT instructions
• Intel Ice Lake+ — AVX-512 VNNI instructions
• AMD Zen4+ — AVX-512 VNNI instructions
• Intel Core Ultra — AVX-VNNI (256-bit) instructions
• AMD Zen 2–3 / pre-Skylake-X Intel — AVX2 + FMA fallback (256-bit, no VNNI fast path)

Installation

pip install litespark-inference

Requirements:

• Python 3.10+
• PyTorch 2.4+

macOS (recommended):

brew install libomp

OpenMP enables multi-threaded kernel execution. Without it, inference will run single-threaded.

The torchless runtime ships as a setuptools extension and is built during pip install. No JIT compile happens at first inference; you do need a C++ compiler available at install time (Xcode CLT on macOS, build-essential on Linux).

Torchless Runtime

litespark-inference ships with a torchless runtime for the supported BitNet and Falcon Edge models. It reads safetensors directly, stores ternary weights in the native packed format used by the SIMD kernels, owns the KV cache, and does not import torch for inference. The CLI dispatches to it automatically:

litespark-inference generate "Hello, how are you?"
# [litespark-inference] torchless runtime (model=bitnet-2b).

Or invoke the torchless CLI directly:

python -m litespark_inference.torchless generate "Hello, how are you?" --max-tokens 32 --raw
python -m litespark_inference.torchless info

Headline numbers vs the PyTorch baseline (Apple Silicon M5 Max, bitnet-2b)

Same prompt, same workload (pp128 + tg128), measured with benchmark_kernel.py --inference --pytorch:

Metric	PyTorch bf16	Litespark torchless (int4)	Speedup
Peak RSS	4,868.22 MB	806.81 MB	6.03×
TTFT (pp128)	4.21 s	0.59 s	7.15×
Throughput (tg128)	2.20 tok/s	39.92 tok/s	18.15×

Torchless generation is greedy-only today. Sampling flags are accepted by the CLI for compatibility but ignored on torchless routes; set LITESPARK_FORCE_TORCH=1 to force the legacy torch-backed path when sampling is required.

Usage

Supported Models

• bitnet-2b
• falcon-edge-1b
• falcon-edge-1b-instruct
• falcon-edge-3b
• falcon-edge-3b-instruct

Command Line

# Generate text with the default BitNet model
litespark-inference generate "The meaning of life is"

# Generate text with Falcon Edge instruct
litespark-inference generate "What is the capital of France?" --model falcon-edge-1b-instruct

# Interactive chat with Falcon Edge instruct
litespark-inference chat --model falcon-edge-1b-instruct

# Run benchmark on your hardware
litespark-inference benchmark

# Show system info and detected SIMD capabilities
litespark-inference info

Python API

from litespark_inference import load_model

# Load the default BitNet 2B model (auto-downloads from Hugging Face)
model, tokenizer = load_model("bitnet-2b")

# Generate text
input_ids = tokenizer.encode("Hello, world!", return_tensors="pt")
output = model.generate(input_ids, max_new_tokens=100)
print(tokenizer.decode(output[0]))

# Load a Falcon Edge instruct model
falcon_model, falcon_tokenizer = load_model("falcon-edge-1b-instruct")

Kernel Mode (Apple Silicon)

NEON is the default optimized path on Apple Silicon:

# NEON mode (default) — packed torchless inference, ~0.8 GB on BitNet-2B
litespark-inference generate "Hello" --mode neon

# In Python
model, tokenizer = load_model("bitnet-2b", mode="neon") # default, fast

How It Works

Ternary models use weights constrained to {-1, 0, +1}. This means matrix multiplication reduces to simple addition and subtraction:

y = Σ x_j · w_j  →  y = Σ(w=+1) x_j - Σ(w=-1) x_j

Litespark-Inference exploits this structure with custom SIMD kernels that:

1. Store weights as int8 — enabling direct use of hardware dot product instructions
2. Quantize activations per-row — converting float32 inputs to int8 with scale factors
3. Use hardware SIMD instructions — NEON SDOT (ARM) or AVX-512 VPDPBUSD (x86)
4. Apply zero-point correction — maintaining numerical accuracy

The library automatically detects your CPU's SIMD capabilities and dispatches to the optimal kernel.

Benchmarking

Run the built-in benchmark to measure performance on your hardware:

litespark-inference benchmark

Repo-level benchmarks for detailed profiling:

# Litespark (torchless by default for bitnet-2b) vs the PyTorch baseline,
# pp128 + tg128, both runtimes' RSS captured. ~6 min total.
python benchmark_kernel.py --inference --pytorch

# Quick torchless-only inference benchmark, no PyTorch baseline.
python benchmark_kernel.py --inference --no-matrix

# Full sweep: matrix + thread-scaling + inference (each torch-backed
# kernel phase runs in its own subprocess so torch's libomp never
# coexists with our libomp).
python benchmark_kernel.py --all

# Raw kernel benchmarks (matrix shapes, scaling) standalone.
python benchmark_kernel.py

ARM compatibility tests

The tests/arm_compat/ folder ships ARM's three validation scripts (benchmark_litespark.py, benchmark_transformers.py, benchmark_repeat_v2.py) plus their original instructions. After pip install -e . they run with no environment variables and no edits:

python tests/arm_compat/benchmark_repeat_v2.py

This produces the canonical ARM "transformers vs litespark" comparison under /usr/bin/time (Darwin or Linux), aggregated across 5 runs, with .log and .csv artefacts dropped in benchmark_logs/. See tests/arm_compat/README.md for details.

Citation

If you use Litespark-Inference in your research, please cite:

@article{litespark2026,
  title={Litespark Inference on Consumer CPUs: Custom SIMD Kernels for Ternary Neural Networks},
  author={Dade, Nii Osae Osae and Morri, Tony and Rahat, Moinul Hossain and Pal, Sayandip},
  year={2026}
}

License

Apache License 2.0. See LICENSE for details.