reanimator/bee
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity

Disable fp64/fp4 in mixed gpu burn

Browse Source
This commit is contained in:
Mikhail Chusavitin
2026-04-16 10:00:03 +03:00
parent c5b2081ac9
commit e306250da7
1 changed files with 17 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
iso/builder/bee-gpu-stress.c
View File

@@ -713,6 +713,19 @@ static const struct profile_desc k_profiles[] = {
#define PROFILE_COUNT ((int)(sizeof(k_profiles) / sizeof(k_profiles[0])))
static int profile_allowed_for_run(const struct profile_desc *desc, int cc, const char *precision_filter) {
if (!(desc->enabled && cc >= desc->min_cc)) {
return 0;
}
if (precision_filter != NULL) {
return strcmp(desc->block_label, precision_filter) == 0;
}
/* Mixed/all phases intentionally exclude fp64/fp4 for now: both paths are
* unstable on the current benchmark fleet and can abort the whole mixed
* pass after earlier phases already collected useful telemetry. */
return strcmp(desc->block_label, "fp64") != 0 && strcmp(desc->block_label, "fp4") != 0;
}
static int load_cublaslt(struct cublaslt_api *api) {
memset(api, 0, sizeof(*api));
api->lib = dlopen("libcublasLt.so.13", RTLD_NOW | RTLD_LOCAL);
@@ -1222,8 +1235,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
/* Count profiles matching the filter (for deciding what to run). */
for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
if (k_profiles[i].enabled && cc >= k_profiles[i].min_cc &&
(precision_filter == NULL || strcmp(k_profiles[i].block_label, precision_filter) == 0)) {
if (profile_allowed_for_run(&k_profiles[i], cc, precision_filter)) {
planned++;
}
}
@@ -1240,7 +1252,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
* profiles matching precision_filter. */
int planned_total = 0;
for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
if (k_profiles[i].enabled && cc >= k_profiles[i].min_cc) {
if (profile_allowed_for_run(&k_profiles[i], cc, precision_filter)) {
planned_total++;
}
}
@@ -1310,10 +1322,10 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
desc->min_cc);
continue;
}
if (precision_filter != NULL && strcmp(desc->block_label, precision_filter) != 0) {
if (!profile_allowed_for_run(desc, cc, precision_filter)) {
append_detail(report->details,
sizeof(report->details),
"%s=SKIPPED precision_filter\n",
"%s=SKIPPED benchmark_disabled\n",
desc->name);
continue;
}