https://blog.defectiv.es/en/tags/slurm/Recent content in slurm on Gaia Lab · BlogHugoen-GBWed, 23 Sep 2026 09:00:00 +0200https://blog.defectiv.es/en/posts/servir-modelos-y-cuidar-el-cluster-inferencia-burn-in-y-sondas/Tue, 22 Sep 2026 14:15:00 +0200https://blog.defectiv.es/en/posts/servir-modelos-y-cuidar-el-cluster-inferencia-burn-in-y-sondas/<p>Ninth and final instalment of <strong>Cluster X-ray</strong>. Here we group what does not fit into a research line but shows up in almost everyone&rsquo;s scripts.</p> <h2 id="serving-models-so-that-others-can-work">Serving models so that others can work <a class="hanchor" href="#serving-models-so-that-others-can-work" aria-label="Enlace a esta sección">#</a></h2> <p>Several lines in the series need a large model up and running: a judge to score responses, a synthetic data generator, a coding assistant. On a cluster with one GPU per node, the natural way to serve it is a Slurm job that brings up <strong>vLLM</strong>, waits for it to respond, registers with a shared proxy and stays there until its time runs out. On finishing, it deregisters: the model appears in and disappears from the catalogue along with the job.</p>https://blog.defectiv.es/en/posts/4400-complejos-de-proteinas-con-boltz-2/Tue, 22 Sep 2026 13:00:00 +0200https://blog.defectiv.es/en/posts/4400-complejos-de-proteinas-con-boltz-2/<p>Fourth instalment of <strong>Cluster X-ray</strong> and the line with the fewest jobs: fifteen. Also one of those that reserves the most GPU, because each job runs for days.</p> <h2 id="the-question">The question <a class="hanchor" href="#the-question" aria-label="Enlace a esta sección">#</a></h2> <p>Knowing whether two proteins interact, and how, is one of the central questions of structural biology: it explains signalling pathways, drug targets and the effect of mutations. Determining it experimentally costs months per pair. The line uses structure prediction to <strong>estimate, at the scale of thousands of pairs, the structure of the complex that two proteins form and the confidence that the binding is real.</strong></p>https://blog.defectiv.es/en/posts/auto-alineamiento-de-un-llm-de-3b-con-lora-y-harmbench/Tue, 22 Sep 2026 12:15:00 +0200https://blog.defectiv.es/en/posts/auto-alineamiento-de-un-llm-de-3b-con-lora-y-harmbench/<p>First instalment of <strong>Cluster X-ray</strong>, the series that goes line by line through <a href="https://blog.defectiv.es/en/posts/que-corre-de-verdad-en-nuestro-cluster/">what really runs on the ANTS group&rsquo;s cluster</a> . We start with the first of three lines on language model safety.</p> <h2 id="the-question">The question <a class="hanchor" href="#the-question" aria-label="Enlace a esta sección">#</a></h2> <p>Fine-tuning a language model so that it refuses harmful requests has a cost: if it is taught too thoroughly to say no, it starts refusing harmless things too, and loses usefulness. The line studies that balance: <strong>what proportion of safety data has to be mixed with the utility data, and whether the model itself can generate its own refusal examples and regulate the dose on its own.</strong></p>https://blog.defectiv.es/en/posts/que-corre-de-verdad-en-nuestro-cluster/Tue, 22 Sep 2026 10:00:00 +0200https://blog.defectiv.es/en/posts/que-corre-de-verdad-en-nuestro-cluster/<p>A cluster&rsquo;s job history is usually a wall of six-digit identifiers and cryptic names —<code>mcpfw_cc</code>, <code>eval_dynamic_pid</code>, <code>boltz2</code>— that say everything to whoever launched them and nothing to anyone else. But if you keep the <em>scripts</em>, the wall turns into a document: people write down <strong>why</strong> they do things, especially when a batch costs money and a night of waiting.</p>