talk-llama : sync llama.cpp

2025-08-09 10:35:51 +02:00 · 2025-01-14 09:53:50 +02:00
parent 19d95f9f9a
commit 99b011a9f5
26 changed files with 5788 additions and 5093 deletions
--- a/examples/talk-llama/llama-hparams.cpp
+++ b/examples/talk-llama/llama-hparams.cpp
@ -52,7 +52,7 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
 uint32_t llama_hparams::n_embd_k_s() const {
    if (wkv_head_size != 0) {
        // for RWKV models
-        return 2 * n_embd;
+        return token_shift_count * n_embd;
    }

    // TODO: maybe support other convolution strides than 1