llama.cpp on a jailbroken PS5: run any GGUF model natively on the console, starting with Qwen 3.8
App: a decision model's speed as time per decision
Mert Cobanovb4ca329dc27e
diff --git a/ps5/app/activity.cpp b/ps5/app/activity.cppindex b8079bc..e925ad8 100644--- a/ps5/app/activity.cpp+++ b/ps5/app/activity.cpp@@ -167,7 +167,8 @@ void ps5lm_activity_begin(int id, const char * method, const char * path, constreturn;}const json j = json::parse(body, body + body_len, nullptr, false);- g_sink->activity_begin(id, remote ? remote : "", path ? path : "", one_line(asked_from(path ? path : "", j), 400));+ const int questions = j.is_object() && j.contains("questions") && j["questions"].is_object() ? (int) j["questions"].size() : 0;+ g_sink->activity_begin(id, remote ? remote : "", path ? path : "", one_line(asked_from(path ? path : "", j), 400), questions);}void ps5lm_activity_data(int id, const char * data, size_t len) {diff --git a/ps5/app/dashboard.cpp b/ps5/app/dashboard.cppindex 27c5a17..cbd94a7 100644--- a/ps5/app/dashboard.cpp+++ b/ps5/app/dashboard.cpp@@ -1077,9 +1077,19 @@ void Dashboard::draw_home(gfx::DrawList & l) const {big = std::max(0.0f, live_.load_progress) * 100.0f;unit = "% on the GPU";} else if (decision) {- big = live_.prompt_tps; // as measured, once a second: an eased number flickers on a TV- unit = "tokens read per second";- known = live_.prompt_tps > 0;+ // Time per decision (one question answered) over the latest requests:+ // what a decision model is judged by, as Ollaya reports it.+ double seconds = 0;+ int decisions = 0;+ for (const auto & e : live_.activity) {+ if (e.status == 200 && e.questions > 0) {+ seconds += e.seconds;+ decisions += e.questions;+ }+ }+ big = decisions ? (float) (1000.0 * seconds / decisions) : 0;+ unit = "ms per decision";+ known = decisions > 0;} else {big = live_.gen_tps;unit = "tokens per second";diff --git a/ps5/app/preview_main.cpp b/ps5/app/preview_main.cppindex 943e09b..55dd9d2 100644--- a/ps5/app/preview_main.cpp+++ b/ps5/app/preview_main.cpp@@ -187,12 +187,12 @@ ps5lm::Live sample(float seconds, bool generating) {L.free_gib = 8.6;L.activity.clear();ps5lm::Exchange a;- a.id = 130; a.client = "192.168.1.40"; a.path = "/v1/systemone"; a.status = 200; a.seconds = 0.3;+ a.id = 130; a.client = "192.168.1.40"; a.path = "/v1/systemone"; a.status = 200; a.seconds = 0.3; a.questions = 3;a.asked = "3 questions: {\"from\": \"[email protected]\", \"subject\": \"URGENT: you won a free iPhone!!!\"}";a.answered = "spam yes 1.00, intent sell, urgency 1.02";L.activity.push_back(a);a.id = 131; a.asked = "7 questions, 1 image"; a.answered = "text no 0.01, color black, ad no 0.01, people no 0.00, mood 1.00";- a.seconds = 2.9; L.activity.push_back(a);+ a.seconds = 2.9; a.questions = 7; L.activity.push_back(a);}if (g_empty) {L.state = ps5lm::ServerState::no_model;diff --git a/ps5/app/stats.cpp b/ps5/app/stats.cppindex 483b0d6..902d864 100644--- a/ps5/app/stats.cpp+++ b/ps5/app/stats.cpp@@ -265,13 +265,14 @@ double steady_s() {} // namespace-void StatsCollector::activity_begin(int id, std::string client, std::string path, std::string asked) {+void StatsCollector::activity_begin(int id, std::string client, std::string path, std::string asked, int questions) {std::lock_guard<std::mutex> lock(mutex_);Exchange e;e.id = id;e.client = std::move(client);e.path = std::move(path);e.asked = std::move(asked);+ e.questions = questions;e.started = steady_s();live_.activity.push_back(std::move(e));if (live_.activity.size() > 6) {diff --git a/ps5/app/stats.hpp b/ps5/app/stats.hppindex 7e5c9e5..876f588 100644--- a/ps5/app/stats.hpp+++ b/ps5/app/stats.hpp@@ -32,6 +32,7 @@ struct Exchange {std::string path; // "/v1/chat/completions"std::string asked; // the last user message, a decision's state, or the promptstd::string answered; // the text answered so far+ int questions = 0; // a decision request's questions, each one decisionint status = 0; // the HTTP status once done, 0 while it runsdouble seconds = 0; // how long it took, once donedouble started = 0; // steady clock, seconds@@ -129,7 +130,7 @@ class StatsCollector {void set_state(ServerState s);void set_scratch(double gib);// From llama-server's HTTP threads, through activity.cpp.- void activity_begin(int id, std::string client, std::string path, std::string asked);+ void activity_begin(int id, std::string client, std::string path, std::string asked, int questions);void activity_text(int id, const std::string & text);void activity_end(int id, int status);