duplicate token_embd to output

This commit is contained in:
Michael Yang 2025-03-09 11:29:08 -07:00
parent d368c039f0
commit 6b0486c216

View File

@ -240,6 +240,9 @@ func New(r *os.File, params ml.BackendParams) (ml.Backend, error) {
switch {
case contains(t.Name, "position_embd", "token_embd", "token_norm_embd", "token_types"):
createTensor(tensor{source: t}, input.bts)
if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
createTensor(tensor{source: t, target: "output.weight"}, output.bts)
}
case contains(t.Name, "cls", "output", "output_norm"):
createTensor(tensor{source: t}, output.bts)
case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):