LLM-Inference/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-12 22:31:11 +04:00
Code Issues Packages Projects Releases Wiki Activity

tests : cover exaone4 hparams ordering

Browse Source 
Builds a synthetic exaone4 model with the layout the shipped EXAONE 4.5
GGUFs use (block_count 65 + nextn 1). The swa_type check is the one that
catches the ordering bug; the n_layer_nextn and n_layer() checks only tell
a broken fixture apart from a real regression.

Fails before the ordering fix with "swa_type is not STANDARD", passes after.
This commit is contained in:
Junmo Kim
2026-08-10 23:29:04 +09:00
parent aecb9bc0c7
commit d2f3bafeee
1 changed files with 60 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-llama-archs.cpp
+60
View File
@@ -9,6 +9,7 @@
// TODO: replace with #include "llama-ext.h" in the future
#include "../src/llama-arch.h"
#include "../src/llama-model.h"
#include "../src/llama-model-saver.h"
#include <cinttypes>
@@ -264,6 +265,62 @@ static bool silent_model_load_progress(float /*progress*/, void * /*user_data*/)
return true;
}
static int check_exaone4_swa(const size_t seed, const uint32_t block_count, const uint32_t nextn,
const uint32_t n_layer_expected, const llama_swa_type swa_expected) {
gguf_context_ptr gguf_ctx = get_gguf_ctx(LLM_ARCH_EXAONE4, false);
// Applied here rather than in get_gguf_ctx because that fixture is shared with the backend tests.
llama_model_saver ms(LLM_ARCH_EXAONE4, gguf_ctx.get());
ms.add_kv(LLM_KV_BLOCK_COUNT, block_count);
ms.add_kv(LLM_KV_NEXTN_PREDICT_LAYERS, nextn);
llama_model_params model_params = llama_model_default_params();
model_params.progress_callback = silent_model_load_progress;
size_t tmp = seed;
llama_model_ptr model(llama_model_init_from_user(gguf_ctx.get(), set_tensor_data, &tmp, model_params));
if (!model) {
printf("%s: failed to create model with block_count %u\n", __func__, block_count);
return 1;
}
const auto & hparams = model->hparams;
int nfail = 0;
if (hparams.n_layer_nextn != nextn) {
printf("%s: block_count %u: n_layer_nextn is %u, expected %u\n",
__func__, block_count, hparams.n_layer_nextn, nextn);
nfail++;
}
if (hparams.n_layer() != n_layer_expected) {
printf("%s: block_count %u: n_layer() is %u, expected %u\n",
__func__, block_count, hparams.n_layer(), n_layer_expected);
nfail++;
}
if (hparams.swa_type != swa_expected) {
printf("%s: block_count %u: swa_type is %d, expected %d\n",
__func__, block_count, hparams.swa_type, swa_expected);
nfail++;
}
return nfail;
}
// exaone4 enables SWA only when n_layer() reads 64, and n_layer() subtracts the nextn layers.
// Shipped EXAONE 4.5 GGUFs carry an MTP head, so block_count is n_layer + 1.
static int test_exaone4_mtp_swa(const size_t seed) {
int nfail = 0;
nfail += check_exaone4_swa(seed, 65, 1, 64, LLAMA_SWA_TYPE_STANDARD); // 32B
nfail += check_exaone4_swa(seed, 64, 0, 64, LLAMA_SWA_TYPE_STANDARD); // 32B without an MTP head
nfail += check_exaone4_swa(seed, 31, 1, 30, LLAMA_SWA_TYPE_NONE); // 1.2B
if (nfail != 0) {
printf("%s: FAIL\n", __func__);
}
return nfail;
}
static std::pair<llama_model_ptr, llama_context_ptr> get_model_and_ctx(
struct gguf_context * gguf_ctx, FILE * file, const size_t seed, const std::vector<ggml_backend_dev_t> & devs,
const llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER, bool encode = false) {
@@ -718,6 +775,9 @@ int main(int argc, char ** argv) {
if (!out.empty()) {
return save_models(arch, seed, log_level, out);
}
if (test_exaone4_mtp_swa(seed) != 0) {
return 1;
}
return test_backends(arch, seed, log_level);
} catch (const std::exception & err) {
fprintf(stderr, "encountered runtime error: %s\n", err.what());