You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit cf08087
Browse filesBrowse the repository at this point in the historyBrowse files
Fix JNI reading mTypeCode instead of mData for scalar EValue inputs (pytorch#17603)
- The `execute_method` paths in `jni_layer.cpp` and
`jni_layer_training.cpp` extracted Int/Double/Bool EValue data by
re-reading the
`mTypeCode` field instead of the actual `mData`. This meant every scalar
input was silently replaced by its type code constant (e.g. Int always
became 4, Double became 3.0, Bool was always true).
- Call the Java accessor methods (`toInt`, `toDouble`, `toBool`)
instead, matching the pattern already used for Tensor inputs via
`toTensor()`.
0 commit comments