VIDRAFT Trains LLM from Scratch on Inference-Only NPUs

VIDRAFT has achieved a significant milestone by pre-training a small language model named VKORE-46M-NPU, containing 46 million parameters, entirely on first-generation inference-only NPUs. Conventional machine learning wisdom dictates that such chips are unsuitable for training workloads due to the absence of hardware-level features like flexible memory access patterns and high-precision intermediate buffers required by GPUs.

This achievement serves as concrete proof-of-concept that backpropagation on inference-dedicated silicon is entirely feasible. The initiative is part of VIDRAFT's broader effort to develop VKORE, an NPU-native acceleration software layer designed to perform the equivalent role for NPUs that CUDA does for NVIDIA GPUs.

Quantitatively, VIDRAFT reported a key result showing a 0.99 loss trajectory correlation coefficient between the NPU-trained and GPU-trained runs. The company explicitly clarified that this measures training dynamics similarity rather than wall-clock throughput, indicating that the NPU-based training process converges along a statistically consistent path.

Stock photo for illustration only, not from the actual event

"The press release explicitly states that VIDRAFT implemented backpropagation on the NPU. It is described as standard reverse-mode gradient computation, not a gradient-free alternative."

VIDRAFT

For developers and engineers interested in exploring the model, VKORE-46M-NPU is publicly available on Hugging Face. Users can download it using the Hugging Face CLI command pip install huggingface_hub huggingface-cli download vidraft/VKORE-46M-NPU. However, the core training code remains proprietary to protect intellectual property, with a technical paper planned for future release.