trohrbaugh/Nemotron-Cascade-2-30B-A3B-heretic-uncensored - Каталог нейросетей
Генерация текста

trohrbaugh/Nemotron-Cascade-2-30B-A3B-heretic-uncensored

Добавлено:
trohrbaugh/Nemotron-Cascade-2-30B-A3B-heretic-uncensored

Nemotron-Cascade-2-30B-A3B ships with custom model code (modelingnemotronh.py) that has two bugs exposed by causalconv1d 1.6.x. These are bugs in NVIDIA’s model code, not in Heretic. Fix 1: Conv State Cache Dimension Mismatch File: modelingnemotronh.py (in the NemotronHCache.init method) What was wrong: The cache allocated convstates tensors with intermediatesize (4096 = mambanumheads × mambaheaddim), but the Mamba mixer’s conv1d layer operates on the full projected dimension (6144 = intermediatesize + 2 × ngroups × ssmstatesize), which includes the B and C state channels concatenated with the hidden states. When causalconv1dupdate compared convstate.shape[1] (4096) against weight.shape[0] (6144), it failed with «weight must have shape (dim, width)». The fix: Added a convdim variable matching the model’s own formula from NemotronHMamba2Mixer.init, and used it for the cache allocation: intermediatesize = config.mambanumheads config.mambaheaddim convdim = intermediatesize + 2 config.ngroups config.ssmstatesize and torch.zeros(batchsize, convdim, convkernelsize, …) Fix 2: Prefill Sequence Length Guard File: modelingnemotronh.py (in cudakernelsforward) What was wrong: The…

Модальности:
Генерация текста

Области применения:
Логика и рассуждение Диалог / чат


Задача: Генерация текста
Автор: trohrbaugh
Теги: nemotron_h, nvidia, nemotron-cascade-2, reasoning, general-purpose, SFT, RL, heretic
Лайков: 4  |  Загрузок: 17

Открыть на HuggingFace →

Описание основано на материалах HuggingFace. Перевод выполнен автоматически.