llama : add Gemma4 MTP (#23398)

2026-06-07 20:50:54 +08:00
parent 8a091c47ab
commit 04eb4c446d
31 changed files with 644 additions and 142 deletions
@@ -185,6 +185,9 @@ struct llama_hparams {
    // for Classifiers
    uint32_t n_cls_out = 1;

+    // input embedding dimension (0 = use n_embd)
+    uint32_t n_embd_inp_impl = 0;
+
    // output embedding dimension (0 = use n_embd)
    uint32_t n_embd_out_impl = 0;

@@ -224,6 +227,7 @@ struct llama_hparams {
    // complex mapping. If using deepstack_mapping_arr, also make sure to set
    // n_deepstack_layers to the number of unique deepstack layers so that
    // n_embd_imp is accurate (see granite.cpp).
+    // TODO: can be expressed via the `new n_embd_inp_impl` and remove this param
    uint32_t n_deepstack_layers = 0;

    // deepstack layer array (Granite4 Vision)