1. Layer replacement. The 8 full-attention layers listed above were replaced with GDN/linear-attention layers, each initialized from a neighboring native GDN layer. 2. Distillation. The replaced layers were optimized one after another with layer-wise distillation, using 20 optimization steps per converted layer in this build. The checkpoint must be loaded with the matching qwen35unimaxengine runtime. Install it first. The repository is tagged customcode; loading through plain transformers requires trustremotecode=True` and has not been verified here. Use the UNI MAX runtime above as the supported path. With mode=»auto», the runtime may choose a different execution path depending on your hardware and the quality-verification result. The machine-readable record is included as benchmarkresults.json`. Throughput, latency, kernel selection, quantization support, and memory use depend heavily on the GPU, CUDA/PyTorch versions, context length, and runtime settings. Do not treat these numbers as hardware-independent performance claims. Before publication, the checkpoint was checked with the UNI MAX checkpoint validator and the benchmark path. These layer-wise agreement checks are narrower…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: summerAIO
Теги: qwen3_5_gdn24, qwen, qwen3.5, recurrent, linear-attention, gdn, cuda, custom_code
Лайков: 3 | Загрузок: 303
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
