Skip to content

Fix native CPU-env training hang (OMP regions + missing log keys) #668

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
IlyaasK wants to merge 1 commit into
base: 5.0
Choose a base branch
from
Open

Fix native CPU-env training hang (OMP regions + missing log keys) #668

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions src/ini.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,6 +104,11 @@ static inline DictItem* dict_set(Dict* dict, const char* key, double value) {
return item;
}

static inline double dict_get_default(Dict* dict, const char* key, double def) {
DictItem* item = dict_find(dict, key);
return item ? item->value : def;
}

static inline const char* dict_get_str(Dict* dict, const char* key) {
DictItem* item = dict_find(dict, key);
if (!item || !item->str) {
Expand Down
27 changes: 14 additions & 13 deletions src/pufferl.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1177,7 +1177,6 @@ static void* vec_thread_main(void* arg) {
memset(&vec->rewards[agent_start], 0, apb * sizeof(float));
memset(&vec->terminals[agent_start], 0, apb * sizeof(float));
clock_gettime(CLOCK_MONOTONIC, &t0);
#pragma omp parallel for schedule(static) num_threads(vec->num_workers)
for (int i = env_start; i < env_start + env_count; i++) {
puf_step(&envs[i]);
}
Expand Down Expand Up @@ -1210,7 +1209,6 @@ static void env_start(PuffeRL* p) {
VecThreadArg* args = (VecThreadArg*)calloc(1,
vec->buffers * sizeof(VecThreadArg));
vec->accum = (float*)calloc(1, vec->buffers * NUM_PROF * sizeof(float));
#pragma omp parallel for schedule(static) num_threads(vec->num_workers)
for (int i = 0; i < vec->size; i++) {
puf_reset(&vec->envs[i]);
}
Expand Down Expand Up @@ -2239,11 +2237,11 @@ void puf_dashboard_print(Ini* ini, PuffeRL* p, Dict* log, int epoch) {
}

char gpu[16], vram[32], ram[16];
snprintf(gpu, sizeof(gpu), "%3.0f%%", dict_get(log, "util/gpu_percent"));
snprintf(gpu, sizeof(gpu), "%3.0f%%", dict_get_default(log, "util/gpu_percent", 0.0));
snprintf(vram, sizeof(vram), "%.1f/%.0fG",
dict_get(log, "util/vram_used_gb"),
dict_get(log, "util/vram_total_gb"));
snprintf(ram, sizeof(ram), "%.1fG", dict_get(log, "util/cpu_mem_gb"));
dict_get_default(log, "util/vram_used_gb", 0.0),
dict_get_default(log, "util/vram_total_gb", 0.0));
snprintf(ram, sizeof(ram), "%.1fG", dict_get_default(log, "util/cpu_mem_gb", 0.0));

int fish_span = 18;
int fish_pos = (fish_span - 3) - (puf_dashboard_frame++ % (fish_span - 2));
Expand Down Expand Up @@ -2407,20 +2405,21 @@ static void log_history_bin_mean(PufLogHistory* h, const char* key,
int points, double* out) {
assert(h->size > 0 && points >= 1);
if (points == 1) {
out[0] = dict_get(&h->items[h->size - 1], key);
out[0] = dict_get_default(&h->items[h->size - 1], key, 0.0);
return;
}
double final_steps = dict_get(&h->items[h->size - 1], "agent_steps");
double final_steps = dict_get_default(
&h->items[h->size - 1], "agent_steps", 0.0);
int out_idx = 0;
int bin_n = 0;
double bin_sum = 0;
double fallback = dict_get(&h->items[0], key);
double fallback = dict_get_default(&h->items[0], key, 0.0);
double next_bin = final_steps / (points - 1);
for (int i = 0; i < h->size; i++) {
Dict* log = &h->items[i];
bin_sum += dict_get(log, key);
bin_n++;
double steps = dict_get(log, "agent_steps");
double steps = dict_get_default(log, "agent_steps", 0.0);
if (steps < next_bin || out_idx >= points - 1) {
continue;
}
Expand All @@ -2430,7 +2429,7 @@ static void log_history_bin_mean(PufLogHistory* h, const char* key,
bin_sum = 0;
next_bin += final_steps / (points - 1);
}
out[points - 1] = dict_get(&h->items[h->size - 1], key);
out[points - 1] = dict_get_default(&h->items[h->size - 1], key, 0.0);
while (out_idx < points - 1) {
out[out_idx++] = fallback;
}
Expand Down Expand Up @@ -3155,8 +3154,10 @@ TrainResult run_train(Ini* ini, TrainContext* ctx) {
}

// TrainResult curve: bin-mean over log_history (same as artifact metrics).
result.cost = dict_get(&last_log, "uptime");
result.steps = dict_get(&last_log, "agent_steps");
DictItem* cost_item = dict_find(&last_log, "uptime");
result.cost = cost_item ? cost_item->value : 0.0;
DictItem* steps_item = dict_find(&last_log, "agent_steps");
result.steps = steps_item ? steps_item->value : 0.0;
DictItem* target = dict_find(&last_log, target_key);
result.score = target ? (float)target->value : 0;

Expand Down