llama: add n_embd_inp_enc encoder-input hparams (DFlash dependency; extracted from upstream ggml-org#24707 minus the unrelated eagle3 fix)

Author: TheTom

src/llama-context.cpp

@@ -1383,7 +1383,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     // eagle3/DFlash: features as encoder input, and non-draft paths fall back to model's input dim
-    const int64_t n_embd = hparams.n_embd_inp();
+    const int64_t n_embd = hparams.n_embd_inp_enc();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0

src/llama-hparams.cpp

@@ -104,6 +104,10 @@ uint32_t llama_hparams::n_embd_inp() const {
     return n_embd_inp;
 }
 
+uint32_t llama_hparams::n_embd_inp_enc() const {
+    return n_embd_inp_enc_impl > 0 ? n_embd_inp_enc_impl : n_embd_inp();
+}
+
 uint32_t llama_hparams::n_embd_out() const {
     return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
 }

src/llama-hparams.h

@@ -189,6 +189,10 @@ struct llama_hparams {
     // input embedding dimension (0 = use n_embd)
     uint32_t n_embd_inp_impl = 0;
 
+    // encoder input embedding dimension (0 = use n_embd_inp())
+    // e.g. the eagle3 encoder fuses target_layers * target_hidden features
+    uint32_t n_embd_inp_enc_impl = 0;
+
     // output embedding dimension (0 = use n_embd)
     uint32_t n_embd_out_impl = 0;
 
@@ -303,6 +307,9 @@ struct llama_hparams {
     // dimension of main + auxiliary input embeddings
     uint32_t n_embd_inp() const;
 
+    // dimension of the encoder input embeddings
+    uint32_t n_embd_inp_enc() const;
+
     // dimension of output embeddings
     uint32_t n_embd_out() const;