已获得支持mtk npu 的最新仓库mnn3.6.0,使用neuropilot-sdk-basic-7.0.8-build20240807进行天玑9300/8300 llmexport.py导出后,使用neruonsdk 对第一层prompt编译报错
Author: dr3334Created Aug 20, 2026Updated Sep 11, 2026
Labelstype:bugplatform:androidmodule:llm
https://mnn-docs.readthedocs.io/en/latest/transformers/llm.html#mtk-llm step follow above link rope op not support
sin/cos
cast(int32 -> fp16)
mul
unsupport operror occur because io function _nupSupport
static bool _npuSupportOp(const Op* op) {
if (gMaxKVSize > 0) {
return true;
}
if (op->type() == OpType_Attention && attn->kv_cache())
return false;
if (op->type() == OpType_LinearAttention)
return false;
return true;
}prefill -> dynamic rope decode -> static rope
Source: alibaba/MNN