DeepseekV4 MTP + DSpark (#25784)

2026-08-02 20:55:34 +08:00
parent f5919bf458
commit 596a5795bd
14 changed files with 1534 additions and 225 deletions
@@ -1291,7 +1291,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
        GGML_ASSERT(ctx_tgt && ctx_dft && "MTP requires ctx_tgt and ctx_dft to be set");

        n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
-        GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
+        GGML_ASSERT(n_embd == llama_model_n_embd_out(llama_get_model(ctx_tgt)) &&
                "MTP input row width must match the target h_nextn width");
        n_mtp_layers = std::max(1, (int) llama_model_n_layer_nextn(llama_get_model(ctx_dft)));