Two Nvidia DGX Stations Serve GLM-5.3 at 607.8 Tokens per Second Using vLLM

·9 views

Nvidia has published a blog post describing how it serves the GLM-5.3 model across two DGX Station systems. The setup uses the vLLM inference engine to distribute the workload between the two machines. The reported throughput reaches 607.8 tokens per second. The post is framed around the limits of relying on a single GPU for large models.

Read the full story at Hacker News

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

Discussion (0)

Method for diffing large 834 enrollment files without crashing Excel

Working with enrollment files that run to hundreds of megabytes can freeze Notepad++ or cause Excel to truncate data, which makes routine month-over-month comparisons hard. The article proposes a member-level approach: split each file into member records at INS segments, build a stable key per member, and store a hash and file offset for each. Streaming the new file against that index classifies members as added, unchanged, changed, or possibly termed while keeping memory tied to member count rather than file size. It also warns that full and change 834 files must be treated differently, since an absent member means termination in one but not the other. A small free command-line tool, x12diff, is presented as an implementation that detects the file type from BGN08.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Cursor AI can receive your workspace folder path before any tool runs

A new Cursor chat can be given the absolute path of the opened workspace folder, even before the agent opens any file. Parent directory names may reveal details such as a local account name, a client relationship, or business activity. Cursor staff confirmed this in a November 12, 2026 forum reply, noting that project-specific directories under the user's home folder can also enter context. Knowing a location does not grant permission to read it, and .cursorignore only controls indexing rather than acting as a data-loss-prevention control. The article advises checking what a fresh chat reveals, reviewing the chat context, and renaming or relocating folders whose names leak sensitive details.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

OpenID4VCI 1.0 Final Errata, 1.1 Draft and Server-to-Server Specs Diverge in Late 2026

An analysis of the OpenID4VCI credential issuance standard finds that the approved 1.0 Final text leaves gaps: a Wallet receives the same invalid_grant error for a wrong PIN and an expired Pre-Authorized Code, and the text sets no limit on PIN guesses. Fixes have since appeared in separate documents, including a 1.0 errata set, a 1.1 editor's draft, and a server-to-server draft merged in August 2026, so implementers may follow different texts. The author checked the specification against its repository at a commit dated 27 October 2026 and reviewed the cited pull requests on 15 November 2026. The article also walks through the protocol's roles, endpoints and pre-authorized code flow, including the credential offer format and the token request.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Cómo construir un agente de IA con Google ADK

Un tutorial explica cómo pasar de los conceptos de la IA agéntica a una implementación práctica usando Google Agent Development Kit (ADK), un framework de código abierto para crear agentes mediante código. El ejemplo consiste en un Asistente de Políticas Corporativas que responde preguntas en lenguaje natural sobre documentos internos de áreas como Recursos Humanos, Finanzas y Seguridad. El agente consulta documentos almacenados en Google Drive, recupera la información relevante y responde citando sus fuentes originales. La solución combina Gemini como modelo de lenguaje, Streamlit para una interfaz tipo chatbot y una estrategia RAG basada en búsqueda léxica sobre un índice en memoria, sin necesidad de base de datos vectorial.

0 comments Read more at DEV Community