An NVFP4 (ModelOpt format) checkpoint of ukisai/Swift1.5-Qwen3.8-Flash-Next, produced for the dgpp inference engine on 2× NVIDIA DGX Spark (GB10), TP=2 — the same format the nvidia/Qwen3.8-Flash-Next-NVFP4 release ships, so dgpp serves it without any engine changes. Swift 1.5 is UkisAI’s reasoning-efficiency fine-tune of Qwen3.8-Flash-Next: fewer thinking tokens at ~equal accuracy. This artifact exists so that fine-tune can run on the Spark pair (the BF16 source is 336 GB; this is 126 GiB). dgpp’s default Qwen NVFP4 template ships the nvidia base. This is the same family, format, engine and decode speed — only the weights differ: ukisai’s Swift 1.5, which thinks less. For the agentic / dev / chat workload this gateway targets, that is the whole value: — No speed penalty — decode matches the base within noise (C1 54.9→59.7, C4 116.6→114.4). — Fewer tokens on chat/agentic → cheaper, faster answers — chat −24% / −24.6% wall, rancube −48% / −45%. — The only Swift 1.5 that runs here — the source is a 336 GB gated BF16 repo; this 126 GiB NVFP4 is what boots on the Spark pair. Not a speed or accuracy win. The low-token behavior is inherited from Swift 1.5 (ukisai); our contribution is…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Языки программирования:
Swift
Задача: Генерация текста
Автор: kaushikvira
Теги: dgpp, qwen4_exp, nvfp4, quantized, qwen3_8, moe, dgx-spark, modelopt
Лайков: 3 | Загрузок: 252
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
