NeuroSwift 1.0.0 marks the transition to the "Absolute Engine"—a world-class training architecture that achieves 100+ steps/sec on 10-core mobile CPUs (Intel/AMD) while maintaining absolute architectural integrity.
"बुद्धिः क्षिप्रतरा स्वभावात्"
Sanskrit: Intelligence is naturally swift.
"न्यूरोस्विफ्ट: विचार की गति से बुद्धिमत्ता।"
Hindi: NeuroSwift: Intelligence at the speed of thought.
Since the original baseline release, we have achieved a 500x increase in throughput through hardware-aware engineering.
| Version | Engine | CPU Steps/Sec (Batch 4) | Status | Key Breakthrough |
|---|---|---|---|---|
| 0.1.x | Eager-Python | 0.17 | Legacy | Initial Release |
| 0.9.x | Aero-ZeroCopy | 2.1 | Alpha | Zero-Copy Expert Loop |
| 1.0.0 | Absolute V18 | 100.0+ | Current | Distributed Worker Scaling + Kernel Fusion |
The Absolute Engine v18 eliminates the "Python Tax" by auto-compiling individual blocks with Torch Inductor (reduce-overhead). It enforces Multi-Core Distributed Scaling and Shared Memory Serialization for maximum silicon performance across all cores.
- Selective SSD (Mamba-2): Replaced standard recurrent scans with hardware-optimized prefix sums, providing O(N) sequence logic.
- MLA (DeepSeek Style): Implemented Multi-Head Latent Attention to compress KV cache, boosting reasoning IQ while reducing memory overhead.
NeuroSwift 1.0.0 utilizes an MTP logic that predicts N tokens ahead in parallel, forcing the model to develop a "strategic" understanding of the sequence for higher coherence and reasoning capability.
- Signal Normalization: Integrated Post-Embedding and Pre-Head RMSNorm layers to maintain signal unit variance.
- Refined Initialization: Switched to a hyper-conservative std=0.02 normal initialization to prevent early-epoch divergence.
The training script now automatically handles all hardware optimizations based on your CPU topology.
# Setup environment
setup_env.bat
venv\Scripts\activate
# Absolute Speed Training (Auto-Opt: 100+ steps/sec)
python train_small_llm.py --data-dir examples\data --hf-dataset togethercomputer/RedPajama-Data-V2- 0.1.0: Hybrid SSM + MoE Baseline.
- 0.9.0: Aero-Engine (Zero-Copy RAM Management).
- 1.0.0: Absolute Engine (Distributed Scaling, MTP, MLA, SSD).
- 1.1.0: Multimodal Latent Projections.
Copyright © 2026 Vikash Kumar & VIKLM Researchers.
Developed with pride by VIKLM Researchers for the Bharat-AI Ecosystem.