Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

spec: add eagle3-v3 support for gpt-oss model (#25794)

Browse Source
This commit is contained in:
Ruixiang Wang
2026-07-28 09:58:16 +03:00
committed by GitHub
parent c6292cfb8e
commit f87067841b
9 changed files with 59 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/speculative.cpp
+13 -3
View File
@@ -437,6 +437,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
int32_t n_embd_dec = 0; // draft hidden size
int32_t n_embd_enc = 0; // target_layer_ids_n * target_hidden_size
int32_t n_embd_tgt = 0; // target model hidden size
int32_t n_layer_tgt = 0; // target model layer count
const int32_t * target_layer_ids = nullptr; // model_dft's extract layer indices
uint32_t target_layer_ids_n = 0;
@@ -478,6 +479,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
n_embd_tgt = llama_model_n_embd(model_tgt);
n_embd_dec = llama_model_n_embd(model_dft);
n_embd_enc = (int32_t) target_layer_ids_n * n_embd_tgt;
n_layer_tgt = llama_model_n_layer(model_tgt);
const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd_dec, /*n_seq_max=*/ 1);
@@ -510,9 +512,15 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
}
}
// turn on extraction of the target layers' input embeddings
// turn on extraction of the target layers' hidden states
for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
if (target_layer_ids[k] < n_layer_tgt) {
llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
} else if (target_layer_ids[k] == n_layer_tgt) {
llama_set_embeddings_nextn(ctx_tgt, true, /*masked*/ false);
} else {
GGML_ABORT("EAGLE3: target layer id %d exceeds target n_layer %d", target_layer_ids[k], n_layer_tgt);
}
}
// turn on extraction of the draft model's pre-norm hidden state
@@ -600,7 +608,9 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
features_buf.resize((size_t) n_tokens * n_embd_enc, 0.0f);
for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
const float * layer = target_layer_ids[k] < n_layer_tgt
? llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k])
: llama_get_embeddings_nextn(ctx_tgt);
if (!layer) {
GGML_ABORT("EAGLE3: target layer %d input not extracted.", target_layer_ids[k]);
}