Skip to content
View gonzabotmdp's full-sized avatar

Block or report gonzabotmdp

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
gonzabotmdp/README.md

G. Arroyo

🇬🇧 English · 🇦🇷 Español

HPC cluster administrator at IFIMAR-UNMDP (Institute of Physics Research of Mar del Plata, CONICET), Argentina.

This profile documents real contributions of AI-assisted work in HPC system administration: software stack migrations, benchmarking, and upstream fixes to open-source projects used by the scientific community.

🤖 Project: gonzabot

An AI assistant for HPC users (Slurm + Spack) that checks an sbatch before it runs and fixes configuration mistakes, resource misuse and oversized resource requests, combining deterministic rules with a local LLM (served with vLLM), so no data leaves the cluster.

In production since 2026 on the IFIMAR cluster. Measured real cases: up to 2.7x faster after applying its suggestion (same node, before/after). No external Python dependencies beyond the standard library. Designed to be gentle on the scheduler: see Load on Slurm.

→ Repository · README in English · README en español

Contributions

  • potfit/potfit#121: compatibility fix for modern LAPACK/LAPACKE headers (LAPACK_FORTRAN_STRLEN_END convention), found while packaging potfit for Spack against OpenBLAS (no Intel/MKL) on AMD Zen4. Merged (adapted by the maintainer in potfit/potfit#122).
  • openxla/xla#47794: fix for a duplicate include of cupti_driver_cbid.h (enum redefinition) when building TensorFlow with HERMETIC_CUDA_VERSION=12.4.1, to run on GPUs whose driver does not support CUDA 12.5+. Redirected from tensorflow/tensorflow#125963 (the file is vendored from openxla/xla). Approved, awaiting merge.
  • Twelve fixes to recipes in spack/spack-packages, found while packaging the full HPC stack with AOCC + AMD Zen4:
    • #6202 octave: boolean logic bug + missing GL/GLU dependencies for +qt. Approved.
    • #6203 hpl: nvblas false positive in the BLAS auto-probe.
    • #6204 scafacos: malformed .pc file (root cause in configure.ac, with our own upstream fix in scafacos/scafacos#44).
    • #6205 gnuplot: build broken with ld.lld (clang/AOCC).
    • #6206 grace: Motif detection broken on builds with embedded Xpm + GCC 14+ strictness. Approved.
    • #6207 lammps: CUDA_HOST_COMPILER (a legacy FindCUDA variable) was never set, so the chosen CUDA host compiler was ignored. Merged.
    • #6269 spades: missing <cstdint> under GCC 14 (111 files affected, solved by forcing the include at compiler level).
    • #6270 metaeuk: same <cstdint> bug under GCC 14 in the vendored mmseqs2 (47 files affected).
    • #6271 boost: bjam bootstrap broken under GCC 14 in @1.56.0 (undeclared file_query, a real missing include).
    • #6272 sepp: build broken on Python 3.12+ because distutils was removed (obsolete 2010 bootstrap).
    • #6273 py-quast: vestigial pin to boost@1.56.0 removed, Python 3.12+ compatibility, and permissions of a binary (minimap2) that was compiled on first use instead of at install time.
    • #6274 augustus: TABIX path never patched in bam2wig, which broke under Spack.

Institutional links


Español

Administrador del cluster HPC de IFIMAR-UNMDP (Instituto de Investigaciones Físicas de Mar del Plata, CONICET).

Este perfil documenta contribuciones reales de IA asistida en tareas de administración de sistemas HPC: migraciones de stack de software, benchmarking, y fixes upstream a proyectos de código abierto usados por la comunidad científica.

🤖 Proyecto: gonzabot

Asistente de IA para usuarios de HPC (Slurm + Spack) que revisa un sbatch antes de correrlo y corrige errores de configuración, mal uso de recursos, y pedidos de recursos sobredimensionados — combinando reglas determinísticas con un LLM local (vía vLLM).

En producción desde 2026 en el cluster de IFIMAR. Casos reales medidos: hasta 2.7x más rápido tras aplicar su sugerencia (mismo nodo, antes/después). Sin dependencias externas de Python más allá de la librería estándar.

→ Repositorio · README en español

Contribuciones

  • potfit/potfit#121 — fix de compatibilidad con headers LAPACK/LAPACKE modernos (convención LAPACK_FORTRAN_STRLEN_END), encontrado al empaquetar potfit para Spack contra OpenBLAS (sin Intel/MKL) en AMD Zen4. Mergeado (adaptado por el maintainer en potfit/potfit#122).
  • openxla/xla#47794 — fix de inclusión duplicada de cupti_driver_cbid.h (redefinición de enum) al compilar TensorFlow con HERMETIC_CUDA_VERSION=12.4.1, para correr en GPUs con driver que no soporta CUDA 12.5+. Redirigido desde tensorflow/tensorflow#125963 (el archivo es vendored desde openxla/xla). Aprobado, esperando merge.
  • Doce fixes a recipes de spack/spack-packages, encontrados empaquetando el stack HPC completo con AOCC + AMD Zen4:
    • #6202 octave — bug de lógica booleana + deps GL/GLU faltantes para +qt. Aprobado.
    • #6203 hpl — falso positivo de nvblas en el auto-probe de BLAS.
    • #6204 scafacos — .pc mal formado (causa raíz en configure.ac, con fix upstream propio en scafacos/scafacos#44).
    • #6205 gnuplot — build roto con ld.lld (clang/AOCC).
    • #6206 grace — detección de Motif rota en builds con Xpm embebido + strictness de GCC14+. Aprobado.
    • #6207 lammps — CUDA_HOST_COMPILER (variable legacy de FindCUDA) nunca se seteaba, ignorando el compilador host de CUDA elegido. Mergeado.
    • #6269 spades — <cstdint> faltante bajo GCC 14 (111 archivos afectados, resuelto forzando el include a nivel de compilador).
    • #6270 metaeuk — mismo bug de <cstdint> bajo GCC 14 en el mmseqs2 vendorizado (47 archivos afectados).
    • #6271 boost — bootstrap de bjam roto bajo GCC 14 en @1.56.0 (file_query sin declarar, include faltante real).
    • #6272 sepp — build roto en Python 3.12+ por remoción de distutils (bootstrap obsoleto de 2010).
    • #6273 py-quast — pin vestigial a boost@1.56.0 eliminado, compat Python 3.12+, y permisos de un binario (minimap2) que se compilaba en el primer uso en vez de en el install.
    • #6274 augustus — path de TABIX nunca parcheado en bam2wig, rompía bajo Spack.

Enlaces institucionales

Popular repositories Loading

  1. gonzabot gonzabot Public

    AI assistant for Slurm + Spack HPC users: checks your sbatch before it runs and fixes it with deterministic rules + a local LLM (no data leaves the cluster). Open source (MIT). In production at IFI…

    Python 7

  2. potfit potfit Public

    Forked from potfit/potfit

    potfit force-matching code

    C

  3. gonzabotmdp gonzabotmdp Public

    Perfil / README de gonzabot — IFIMAR-CONICET

  4. tensorflow tensorflow Public

    Forked from tensorflow/tensorflow

    An Open Source Machine Learning Framework for Everyone

    C++

  5. spack-packages spack-packages Public

    Forked from spack/spack-packages

    Spack's community package recipes

    Python

  6. scafacos scafacos Public

    Forked from scafacos/scafacos

    ScaFaCoS library for Fast Coulomb Solvers

    C