vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-08-19 05:15:21 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

spec: fix segfault error on long prompts for eagle3 (#24707)

Browse source
This commit is contained in:
Ruixiang Wang 2026-06-17 16:29:49 +02:00 • committed by GitHub
parent 74a80dd9c0
commit 1a2dea29b9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 16 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

8
src/models/eagle3.cpp
View file

@ -19,7 +19,7 @@ void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_TARGET_HIDDEN_SIZE, n_embd_tgt);
LLAMA_LOG_INFO("%s: EAGLE3 n_embd_tgt = %u (draft n_embd = %u)\n", __func__, n_embd_tgt, hparams.n_embd);
hparams.n_embd_inp_impl = (uint32_t) target_layer_ids.size() * n_embd_tgt;
hparams.n_embd_inp_enc_impl = (uint32_t) target_layer_ids.size() * n_embd_tgt;
// eagle3 norm_before_residual (optional, default false)
// compatible with Readhat eagle3 speculator model
@ -34,7 +34,7 @@ void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {
void llama_model_eagle3::load_arch_tensors(llama_model_loader &) {
LLAMA_LOAD_LOCALS;
const int64_t n_embd_inp = hparams.n_embd_inp();
const int64_t n_embd_inp = hparams.n_embd_inp_enc();
const int64_t n_embd_attn_input = 2 * n_embd;
// Get vocab size from the d2t tensor in the GGUF file (optional - only needed if eagle3 has different vocab_size than target)
@ -109,8 +109,8 @@ ggml_tensor * llama_model_eagle3::graph<true>::build_inp_embd_enc() const {
// Input: Target model features (3 layers concatenated: low, mid, high)
// Data will be provided via ubatch->embd in encode_eagle3_features()
auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp());
inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32,hparams.n_embd_inp(), n_tokens);
auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp_enc());
inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp_enc(), n_tokens);
ggml_set_input(inp_target->embd);
cur = inp_target->embd;