Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-04 02:37:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

tests : rename fusion api helpers to match stats_init signature

Browse Source 
Align the test with the ad-hoc fusion stats API: fusion_stats_init no
longer takes an enable bool (stats are turned on by calling it), so the
proc-address wrappers and typedefs are renamed to the api_* convention.

Assisted-by: pi:llama.cpp/DeepSeek-V4-Flash-0731
This commit is contained in:
Georgi Gerganov
2026-09-03 17:19:49 +03:00
parent c9c0ceca0a
commit 25c68498f9
1 changed files with 22 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-fusion.cpp
+22 -20
View File
@@ -35,9 +35,9 @@
// generic fusion debugging API, resolved through the ad-hoc get_proc_address mechanism
// (not part of the official ggml backend interface yet). a backend that adopts fusion debugging
// exports these exact names.
typedef void ( * fusion_enable_t) (ggml_backend_dev_t, bool);
typedef void ( * fusion_reset_t) (ggml_backend_dev_t);
typedef int ( * fusion_get_stats_t)(ggml_backend_dev_t, const char **, uint64_t *, int);
typedef void ( * fusion_stats_init_t) (ggml_backend_dev_t);
typedef void ( * fusion_stats_reset_t)(ggml_backend_dev_t);
typedef int ( * fusion_stats_get_t) (ggml_backend_dev_t, const char **, uint64_t *, int);
typedef void ( * fusion_set_enabled_t)(ggml_backend_dev_t, bool);
static bool silent_model_load_progress(float, void *) {
@@ -172,12 +172,12 @@ static std::vector<float> decode_token_by_token(llama_model * model, llama_conte
return ret;
}
static void read_counts(fusion_get_stats_t get_stats, ggml_backend_dev_t dev,
static void read_counts(fusion_stats_get_t api_stats_get, ggml_backend_dev_t dev,
std::vector<const char *> & labels, std::vector<uint64_t> & counts) {
const int n = get_stats(dev, nullptr, nullptr, 0);
const int n = api_stats_get(dev, nullptr, nullptr, 0);
labels.assign(n, nullptr);
counts.assign(n, 0);
get_stats(dev, labels.data(), counts.data(), n);
api_stats_get(dev, labels.data(), counts.data(), n);
}
// one row of the per-label report
@@ -261,21 +261,21 @@ int main(int argc, char ** argv) {
// resolve the generic fusion debugging functions through the ad-hoc get_proc_address
// mechanism; a backend that does not adopt fusion debugging exports none of them
auto * reg = ggml_backend_dev_backend_reg(dev);
auto enable = (fusion_enable_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_stats_init");
auto reset = (fusion_reset_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_stats_reset");
auto get_stats = (fusion_get_stats_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_stats_get");
auto set_enabled = (fusion_set_enabled_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_set_enabled");
if (!enable || !set_enabled || !reset || !get_stats) {
auto api_stats_init = (fusion_stats_init_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_stats_init");
auto api_stats_reset = (fusion_stats_reset_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_stats_reset");
auto api_stats_get = (fusion_stats_get_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_stats_get");
auto api_set_enabled = (fusion_set_enabled_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_fusion_set_enabled");
if (!api_stats_init || !api_set_enabled || !api_stats_reset || !api_stats_get) {
LOG_ERR("%s: backend '%s' does not export the generic fusion debugging API "
"(ggml_backend_fusion_*) - cannot run the fusion regression test\n",
__func__, backend_name.c_str());
return 1;
}
// enabling fusion debugging also forces n_cb == 0 (single-threaded encoding) for every
// backend context created afterwards, so the counters are race-free
enable(dev, true);
// enable fusions stats
api_stats_init(dev);
const bool has_counts = true;
@@ -357,12 +357,12 @@ int main(int argc, char ** argv) {
{
llama_context_ptr ctx = create_ctx(model.get(), 32);
if (has_counts) {
set_enabled(dev, true);
reset(dev);
api_set_enabled(dev, true);
api_stats_reset(dev);
}
logits_fused = mode.decode(model.get(), ctx.get(), tokens);
if (has_counts) {
read_counts(get_stats, dev, labels, counts_fused);
read_counts(api_stats_get, dev, labels, counts_fused);
}
}
@@ -372,20 +372,22 @@ int main(int argc, char ** argv) {
{
llama_context_ptr ctx = create_ctx(model.get(), 32);
if (has_counts) {
set_enabled(dev, false);
reset(dev);
api_set_enabled(dev, false);
api_stats_reset(dev);
}
logits_unfused = mode.decode(model.get(), ctx.get(), tokens);
if (has_counts) {
read_counts(get_stats, dev, labels, counts_unfused);
read_counts(api_stats_get, dev, labels, counts_unfused);
}
}
const double nmse_fus = nmse(logits_fused, logits_unfused);
const double nmse_dev = logits_cpu.empty() ? 0.0 : nmse(logits_fused, logits_cpu);
// an arch that is already broken on the device (huge device-vs-CPU NMSE, e.g. plamo2
// on Metal) produces garbage regardless of fusion, so the fused-vs-unfused NMSE is not
// a meaningful signal - skip it and rely on the count regression check only
// TODO: run in Debug build to get an assert in `ggml-alloc.c` and fix it
const bool dev_broken = !std::isnan(nmse_dev) && nmse_dev > 1e-4;
// build the per-label rows