VIDRAFT Pre-Trains a 46M-Parameter Language Model on Inference-Only NPUs, Without GPUs

·8 views

VIDRAFT has pre-trained a small language model called VKORE-46M-NPU entirely on a first-generation inference-only NPU, avoiding GPUs at every training stage. The 46-million-parameter model was trained from random initialization on a public dataset and can generate short English stories. The company implemented backpropagation through its own VKORE software layer, since the chip's vendor offers no training stack for it. A matched GPU run showed a 0.99 correlation in loss trajectories, indicating numerically consistent training, though the company stresses this is not a claim of matching GPU training speed. The model and comparison materials are publicly available on Hugging Face, with detailed implementation methods promised in a forthcoming paper.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Open-source Rust service wrapper runs Java apps on Windows, Mac and Linux

A new MIT-licensed Java service wrapper has been released on GitHub under the theonedev project, written in Rust. It has been tested intensively on Windows and Linux (amd64 and arm64) as well as Mac Silicon. OneDev uses the wrapper to run its Java application as a service on these platforms. Its configuration is a subset of the Java Service Wrapper (JSW), focusing solely on service wrapping and omitting JVM monitoring.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

BillCraft AI: Privacy-First Invoice Generator and Receipt Reader for Freelancers

BillCraft AI is an open-source invoice tool aimed at freelancers, contractors, and small business owners who want to avoid manual bookkeeping. It offers an invoice generator with saved business profiles, automatic totals, AI-assisted item descriptions, and branded PDF export. A companion reader uses local OCR and the open-weight Qwen 2.5 7B model to extract line items from receipts and PDFs, then checks the math on the server. The developers say the design keeps financial data off third-party cloud servers and contrasts it with subscription-based SaaS accounting tools costing $30 to $70 per month. The project was built for the Hacktoberfest Open-Source AI Challenge and is available as a live web app and on GitHub.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Guide Outlines Backend Exception Tracking for Express Pricing Rule Rollouts

A technical guide describes how a small Node.js Express service should capture backend exceptions while rolling out a property-pricing rule. It recommends attaching release, environment, request context, and privacy-reviewed identifiers at the Express error boundary, and triaging by error group rather than paging on every event. The author argues that silent failures, such as scheduled rollouts that never start, need separate heartbeat and counter monitoring beyond exception capture. The piece also presents a runnable Go transport that posts validated error events with key-based authentication, bounded retries, Retry-After handling, and exponential backoff. It notes that the Infrai provider lacks alert routing, source-map decoding, session replay, trace queries, and heartbeat monitoring, so those tasks must be assigned elsewhere.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Causal RL Framework Aims to Explain and Schedule Maintenance for Soft Robots

A developer describes building an explainable causal reinforcement learning pipeline to maintain bio-inspired soft robots, such as a silicone octopus-like arm that kept developing micro-tears. Standard deep RL agents maximize task reward but cannot identify why hardware is degrading, since soft-body failures like fatigue, hysteresis, creep, and delamination are not directly observable. The approach uses embodied sensing and a structural causal model linking sensor readings, actions, latent degradation, and maintenance events, so the agent can reason about interventions on the causes of wear. The author reports that the rate of change in strain sensor readings was a stronger early failure predictor than absolute values. The write-up presents the method as a way to make maintenance policies both effective and explainable.

0 comments Read more at DEV Community