{"article":{"annotation":{"article_id":"compilemask-transformer-attention-masking-2026","available_at":"2026-10-13T01:40:21.308279Z","geography_revision":"ef6ee0010dda5fe7b240fd514749b43e214c57da73222d2e40ec83b276dd0977","id":"5a89817510cbf460605095b2b19f78a5fc2fab0aedeb08e5d6e5f14c22c2a2cb","input_sha256":"46154885e980e79d9d3608bcbe68ac21d3719a323c7d9d1afe3a77486b0befc0","limitations":["Rule-derived subjects are candidates, not verified editorial labels or calibrated model probabilities.","Coverage is distinct from event location. Unknown country/city/coordinates remain absent; no automatic geocoding.","Article geometries keep their stated precision and methodology; reproducing them is not independent coordinate verification."],"method":"controlled-rules-v1","places":[{"admin1":null,"city":null,"coordinate_note":null,"coordinate_reference_system":"OGC:CRS84","coordinate_source":null,"country":"CN","evidence":[{"field":"country","language":"en","note":"Explicit preserved editorial country/region; no city or national centroid inferred","source_urls":["https://arxiv.org/abs/2610.13474"]}],"geometry":null,"id":"iso3166-1:CN","kind":"country-or-area","labels":{"ar":"الصين","en":"China","es":"China","fr":"Chine"},"match_status":"candidate","precision":"regional","role":"coverage","uncertainty_m":null}],"rules_sha256":"fb5ad3c27c91cb840f7a0a05f7f64a5cb26ad0374409ff9457f277c7092e22b4","schema_version":1,"subjects":[{"concept_id":"medtop:20000233","confidence":null,"confidence_method":null,"evidence":[{"field":"en.title+en.summary","language":"en","note":"Subject suggested from the title or summary: Telecommunications","source_urls":["https://arxiv.org/abs/2610.13474"]}],"method":"controlled-rules-v1","rule_id":"telecommunications","status":"candidate"},{"concept_id":"medtop:13000000","confidence":null,"confidence_method":null,"evidence":[{"field":"category+topics","language":"en","note":"Preserved editorial category/topic; broad candidate only","source_urls":["https://arxiv.org/abs/2610.13474"]}],"method":"controlled-rules-v1","rule_id":"technology-broad","status":"candidate"}],"taxonomy_revision":"1f2db31d811fca07e99b16d402c10e2cc2965c9df1c7b41d9be348b40a24e7fb","taxonomy_version":"IPTC Media Topics 2026-Q2"},"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=es","category":"technology","content":{"key_points":[],"paragraphs":["Las cargas de los transformadores usan distintas máscaras para definir qué tokens pueden atenderse entre sí. CompileMask permite describir las reglas en un lenguaje basado en regiones; el compilador deduce de antemano los intervalos visibles y proporciona un descriptor compacto al núcleo de atención GPU durante la ejecución.","En NVIDIA Blackwell y con longitud de secuencia 32.000, el estudio comunica un rendimiento frente a Flex-FA4 de 1,44 veces para HSTU, 2,13 para un modelo de recompensa con prefijo compartido y 2,87 para CLEAR. En atención causal conservó el 85–89 % del rendimiento de FA4 nativo. Las comparaciones se limitan a máscaras y hardware concretos y no prueban la misma mejora para todas las cargas de transformadores."],"sections":[]},"corrections":[],"country":"CN","data":{"blocks":[{"columns":[{"key":"attention_mode","label":{"ar":"نوع الانتباه","de":"Aufmerksamkeitsmodus","en":"Attention mode","es":"Tipo de atención","fr":"Type d’attention","pt":"Modo de atenção","tr":"Dikkat türü"},"language":"en","type":"text"},{"key":"native_fa4_percent","label":{"ar":"النسبة المحتفظ بها من أداء FA4 الأصلي","de":"Erhaltener Anteil des nativen FA4-Durchsatzes","en":"Native FA4 throughput retained","es":"Proporción del rendimiento FA4 nativo conservada","fr":"Part du débit FA4 natif conservée","pt":"Parcela do débito FA4 nativo mantida","tr":"Yerel FA4 veriminin korunan bölümü"},"type":"number","unit":"percent"}],"description":{"ar":"أعباء انتباه قِيست عند طول 32 ألفاً؛ القيم خاصة بالدراسة.","de":"Aufmerksamkeitslasten bei einer Länge von 32.000; die Werte gelten für diese Studie.","en":"Attention workloads measured at length 32K; values are specific to the study.","es":"Cargas de atención medidas con longitud 32.000; los valores son específicos del estudio.","fr":"Charges d’attention mesurées à une longueur de 32 000; les valeurs sont propres à l’étude.","pt":"Cargas de atenção medidas com comprimento de 32 mil; os valores são específicos do estudo.","tr":"32 bin uzunlukta ölçülen dikkat iş yükleri; değerler çalışmaya özgüdür."},"id":"attention-benchmark","methodology":{"ar":"اختُبرت أربعة أنماط للأقنعة على وحدات NVIDIA Blackwell بأطوال تسلسل من 1K إلى 32K؛ والتسريعات مقارنة بخط أساس Flex-FA4 في الورقة.","de":"Vier Maskenmuster wurden auf NVIDIA-Blackwell-GPUs bei Sequenzlängen von 1K bis 32K getestet; die Beschleunigungen beziehen sich auf die Flex-FA4-Referenz der Studie.","en":"Four mask patterns were tested on NVIDIA Blackwell GPUs at sequence lengths from 1K to 32K; speedups are relative to the paper’s Flex-FA4 baseline.","es":"Se probaron cuatro patrones de máscara en GPU NVIDIA Blackwell, con secuencias de 1K a 32K; las aceleraciones se comparan con la referencia Flex-FA4 del artículo.","fr":"Quatre motifs de masque ont été testés sur GPU NVIDIA Blackwell, pour des séquences de 1K à 32K; les accélérations sont relatives à la référence Flex-FA4 de l’article.","pt":"Foram testados quatro padrões de máscara em GPUs NVIDIA Blackwell, com sequências de 1K a 32K; os ganhos são relativos à referência Flex-FA4 do artigo.","tr":"NVIDIA Blackwell GPU’larda dört maske deseni, uzunlukları 1K–32K arasında olan iş yükleriyle karşılaştırıldı; hızlanmalar makalede belirtilen Flex-FA4 taban çizgisine göredir."},"rowTranslations":[{"attention_mode":{"ar":"الانتباه السببي، الحد الأدنى","de":"Kausale Aufmerksamkeit, Untergrenze","en":"Causal attention, lower bound","es":"Atención causal, límite inferior","fr":"Attention causale, borne basse","pt":"Atenção causal, limite inferior","tr":"Nedensel dikkat, alt sınır"}},{"attention_mode":{"ar":"الانتباه السببي، الحد الأعلى","de":"Kausale Aufmerksamkeit, Obergrenze","en":"Causal attention, upper bound","es":"Atención causal, límite superior","fr":"Attention causale, borne haute","pt":"Atenção causal, limite superior","tr":"Nedensel dikkat, üst sınır"}}],"rows":[{"attention_mode":"Causal attention, lower bound","native_fa4_percent":85},{"attention_mode":"Causal attention, upper bound","native_fa4_percent":89}],"sourceUrls":["https://arxiv.org/abs/2610.13474"],"title":{"ar":"مقارنة أقنعة الانتباه","de":"Vergleich von Aufmerksamkeitsmasken","en":"Attention-mask comparison","es":"Comparación de máscaras de atención","fr":"Comparaison des masques d’attention","pt":"Comparação de máscaras de atenção","tr":"Dikkat maskesi kıyaslaması"},"type":"table","updatedAt":"2026-10-13T01:31:44Z"}],"version":1},"data_downloads":[{"formats":{"csv":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026/data/attention-benchmark?lang=es&format=csv","json":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026/data/attention-benchmark?lang=es&format=json"},"id":"attention-benchmark"}],"data_types":["table"],"editorial_illustration":{"alt":"Ilustración conceptual de tecnología, procesadores e información conectada.","asset_id":"newsverum-v1-technology","avif_variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.avif","width":480},{"height":480,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-720.avif","width":720},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.avif","width":960},{"height":800,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1200.avif","width":1200},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.avif","width":1536}],"credit":"Newsverum","documentary":false,"height":1024,"kind":"symbolic-editorial-illustration","language":"es","scope":"topic","url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.webp","width":480},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.webp","width":960},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","width":1536}],"width":1536},"event_id":"event-5aecc060-5685-58ec-867e-61619bd57b62","event_period":"2026-10-09","event_period_basis":"source-review","id":"compilemask-transformer-attention-masking-2026","image":null,"importance":0.6,"language":"es","modified_at":"2026-10-13T01:40:12+00:00","publication_basis":"publication-record","published_at":"2026-10-13T01:40:12+00:00","reading_minutes":1,"region":"world","section_illustration":{"alt":"Ilustración conceptual de tecnología, procesadores e información conectada.","asset_id":"newsverum-v1-technology","avif_variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.avif","width":480},{"height":480,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-720.avif","width":720},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.avif","width":960},{"height":800,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1200.avif","width":1200},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.avif","width":1536}],"credit":"Newsverum","documentary":false,"height":1024,"kind":"symbolic-editorial-illustration","language":"es","scope":"section","url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.webp","width":480},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.webp","width":960},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","width":1536}],"width":1536},"slug":"compilemask-transformer-attention-masking-2026","sources":[{"checkedAt":"2026-10-13","name":"arXiv","publishedAt":"2026-10-09","url":"https://arxiv.org/abs/2610.13474"}],"summary":"Investigadores de JD.com y la Universidad de Correos y Telecomunicaciones de Pekín desarrollaron CompileMask, un compilador que conecta máscaras de atención flexibles con núcleos GPU eficientes. Las pruebas en GPU Blackwell midieron más rendimiento que Flex-FA4 en varios patrones.","text":"Investigadores de JD.com y la Universidad de Correos y Telecomunicaciones de Pekín desarrollaron CompileMask, un compilador que conecta máscaras de atención flexibles con núcleos GPU eficientes. Las pruebas en GPU Blackwell midieron más rendimiento que Flex-FA4 en varios patrones.\n\nLas cargas de los transformadores usan distintas máscaras para definir qué tokens pueden atenderse entre sí. CompileMask permite describir las reglas en un lenguaje basado en regiones; el compilador deduce de antemano los intervalos visibles y proporciona un descriptor compacto al núcleo de atención GPU durante la ejecución.\n\nEn NVIDIA Blackwell y con longitud de secuencia 32.000, el estudio comunica un rendimiento frente a Flex-FA4 de 1,44 veces para HSTU, 2,13 para un modelo de recompensa con prefijo compartido y 2,87 para CLEAR. En atención causal conservó el 85–89 % del rendimiento de FA4 nativo. Las comparaciones se limitan a máscaras y hardware concretos y no prueban la misma mejora para todas las cargas de transformadores.","title":"CompileMask mejora el rendimiento de máscaras personalizadas de atención","topics":[],"translations":[{"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=tr","language":"tr","url":"https://newsverum.com/tr/news/compilemask-transformer-attention-masking-2026"},{"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=en","language":"en","url":"https://newsverum.com/en/news/compilemask-transformer-attention-masking-2026"},{"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=ar","language":"ar","url":"https://newsverum.com/ar/news/compilemask-transformer-attention-masking-2026"},{"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=fr","language":"fr","url":"https://newsverum.com/fr/news/compilemask-transformer-attention-masking-2026"},{"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=es","language":"es","url":"https://newsverum.com/es/news/compilemask-transformer-attention-masking-2026"},{"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=de","language":"de","url":"https://newsverum.com/de/news/compilemask-transformer-attention-masking-2026"},{"api_url":"https://newsverum.com/api/v1/articles/compilemask-transformer-attention-masking-2026?lang=pt","language":"pt","url":"https://newsverum.com/pt/news/compilemask-transformer-attention-masking-2026"}],"url":"https://newsverum.com/es/news/compilemask-transformer-attention-masking-2026"},"version":"1"}