QSA-grafted build of llmfan46/Qwen3.6-35B-A3B-uncensored-heretic, made from the existing Qwen3.6-35B-A3B-uncensored-heretic-No-MTP-custom.gguf BPW4.75 base. The base file is preserved byte-for-byte: all 733 original tensors keep their exact payload and offset. Only these are added: — 50 F32 indexer tensors (5 per full-attention layer): blk.{N}.indexer.qproj.weight, kproj.weight, qnorm.weight, knorm.preweight, knorm.weight — metadata: — qwen35moe.attention.indexer.headcount = 4 — qwen35moe.attention.indexer.keylength = 128 — qwen35moe.attention.indexer.topk = 2048 — qwen35moe.attention.compressratios = 4 on layers [3, 7, 11, 15, 19, 23, 27, 31, 35, 39], else 0 This model needs the Qwen3.5/3.6-MoE QSA k-pool indexer support in llama.cpp. LLAMAQSATOPK is the main thing to tune. The GGUF metadata ships the trained budget (2048`), which is right for voice/chat-length prompts; for long documents do not just accept the default: For long prompts, keep the context window large enough that the prompt actually fits in one pass; the QSA path only kicks in above the indexer top_k. By default, the GGUF already carries the indexer metadata, so there is nothing else to enable on the model side.…
Модальности:
Генерация текста Компьютерное зрение
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: tvall43
Теги: gguf, qwen, llama.cpp, vision, en, endpoints_compatible, imatrix, conversational
Лайков: 3 | Загрузок: 974
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
