SShortSingh.
Back to feed

RoPE: cómo la mayoría de los LLM abiertos codifican la posición de cada token

0
·1 views

Una red neuronal no percibe el orden de las palabras por defecto, por lo que los modelos de lenguaje necesitan embeddings posicionales para distinguir oraciones con las mismas palabras. El método dominante en los LLM abiertos de 2026 es RoPE (Rotary Position Embedding), descrito originalmente en el paper RoFormer. A diferencia de GPT-2, que suma un vector de posición una sola vez al inicio, RoPE aplica la información posicional dentro de cada capa de atención, como hace MiniMax-M3. Su rotación de vectores de consulta y clave hace que la relación entre dos tokens dependa solo de su distancia, lo que favorece la generalización a secuencias más largas que las vistas en entrenamiento.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Developer Documents DevOps Learning Journey Through Hands-On Building and Troubleshooting

A developer has started publishing a public log of their DevOps and cloud engineering studies on DEV Community. So far they have experimented with Python and FastAPI, Docker, Kubernetes, GitHub Actions, and Jenkins, building a small employee API and CI pipelines. Their next phase covers Linux and networking, AWS, Terraform, monitoring with Prometheus and Grafana, and incident response practices. They plan to use the Floci local AWS sandbox to practise without incurring cloud costs, and will post weekly updates that include their errors and fixes.

0
ProgrammingDEV Community ·

Preventing double bookings in a Telegram appointment bot using Python and SQLite

A developer describes how they built a small appointment bot for barbershops and salons that avoids double bookings without external services, relying only on SQLite. Rather than maintaining a table of slot availability that can go stale, the design stores only bookings and computes free slots on request based on working hours and service duration. An overlap check ensures a slot is free only if it does not conflict with any existing booking, and a final re-check inside a single transaction guards against races when the booking is written. Reminders are prevented from repeating through per-reminder flags, and the booking logic takes the current time as a parameter so it can be tested against an in-memory database. The author also mentions selling a complete bot template for $19 as a commercial offering.

0
ProgrammingDEV Community ·

Who Should Pay for Client Website Software: Client, Agency, or Free Tools?

A discussion among freelancers, agency owners and designers asks who should buy the paid software that websites depend on, such as premium plugins, page builder licences and per-seat editor access. One approach has clients buy every licence in their own name, which keeps costs visible and makes leaving an agency simple. Another has agencies hold multi-site licences and bundle the cost into care plans, which is simpler for clients and often cheaper overall. A third camp favours free or open-source tools, or platforms without per-person editing fees, to avoid recurring licence management. Each model carries trade-offs, including renewal risks when agency relationships end and the added setup or maintenance effort that free tools can require.

0
ProgrammingDEV Community ·

Python Script Finds Duplicate and Near-Duplicate Pages in a Sitemap

A tutorial explains how to detect duplicate pages across a website using only Python's standard library. Duplicates often arise unintentionally from CMS URL changes, tracking parameters, print views, or template-based pages. The script reads a sitemap, fetches each page, strips the shared template, and compares content by identical HTML, identical main content, and similarity scored with difflib. Matching pages are grouped into clusters with canonical checks and a suggested keeper URL, and results are exported to CSV with a console summary and a CI-friendly exit code. The author notes that Google does not treat duplicate content as a penalty in itself, but it can complicate search tracking and user experience.