github-featured/alibaba--mnn
2
0
Fork 0
Code Issues Pull Requests Deployments Models Agent Notes Packages Projects Releases Wiki Activity

MNN:Sync: Fix bug for llama2/llama3 attention fuse, refract llm usage

Browse Source
This commit is contained in:
xiaying
2024-06-15 15:39:59 +08:00
parent 226f1bc1b7
commit 65ec0ea406
110 changed files with 12600 additions and 2786 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/cpp/ModuleBasic.cpp
+7
View File
@@ -232,6 +232,13 @@ int main(int argc, char *argv[]) {
// Need tensor static for each op, open debug
rtmgr->setMode(Interpreter::Session_Debug);
}
// For Debug
if (false) {
int geometryMask = Interpreter::GeometryComputeMask::GEOMETRCOMPUTEMASK_ALL;
geometryMask -= Interpreter::GeometryComputeMask::GEOMETRCOMPUTEMASK_FUSEREGION;
geometryMask -= Interpreter::GeometryComputeMask::GEOMETRCOMPUTEMASK_OPENCACHE;
rtmgr->setHint(Interpreter::GEOMETRY_COMPUTE_MASK, geometryMask);
}
if (runMask & 4) {
// Need time trace for each op, open debug
rtmgr->setMode(Interpreter::Session_Debug);