Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

DeepSeek Metal: DS4_BUSY_ALL experiment (off), progress lines

Niko122190359513e3b1
diff --git a/metal/src/mtl_rt.mm b/metal/src/mtl_rt.mm
index 8a0ba7978..bd0fd8499 100644
--- a/metal/src/mtl_rt.mm
+++ b/metal/src/mtl_rt.mm
@@ -218,6 +218,7 @@ struct Busy {
id<MTLComputePipelineState> ps = nil;
id<MTLBuffer> flag = nil;
id<MTLCommandBuffer> last = nil;
+ bool running = false;
};
Busy g_busy;
} // namespace
@@ -250,7 +251,8 @@ kernel void busy_spin(device volatile atomic_uint* flag [[buffer(0)]], constant
g_busy.ok = true;
}
}
- if (!g_busy.ok) return;
+ if (!g_busy.ok || g_busy.running) return;
+ g_busy.running = true;
if (g_busy.last != nil) [g_busy.last waitUntilCompleted]; // (the previous spin has left)
*(volatile uint32_t*) [g_busy.flag contents] = 0;
uint32_t cap = std::getenv("DS4_BUSY_CAP") ? (uint32_t) std::atoi(std::getenv("DS4_BUSY_CAP")) : 20000u; // ~ tens of ms at most
@@ -266,8 +268,9 @@ kernel void busy_spin(device volatile atomic_uint* flag [[buffer(0)]], constant
}
}
extern "C" void ds4_busy_end() {
- if (!g_busy.ok || g_busy.last == nil) return;
+ if (!g_busy.ok || g_busy.last == nil || !g_busy.running) return;
*(volatile uint32_t*) [g_busy.flag contents] = 1;
+ g_busy.running = false;
}
Context::Context() : impl_(new Impl) {
diff --git a/src/dsv41/ds4_main.cpp b/src/dsv41/ds4_main.cpp
index 1ecf4fa2a..9a997e77a 100644
--- a/src/dsv41/ds4_main.cpp
+++ b/src/dsv41/ds4_main.cpp
@@ -991,6 +991,14 @@ int main(int argc, char** argv) {
out.push_back(tok);
std::printf("%d,", tok);
std::fflush(stdout);
+ if (std::getenv("DS4_PROGRESS") && out.size() % 200 == 0) { // (sustained-run time series)
+ static double tp = now_ms();
+ static size_t np = 0;
+ const double tn = now_ms();
+ std::fprintf(stderr, "progress %zu tokens: %.2f tok/s over the last %zu\n", out.size(), 1000.0 * (double) (out.size() - np) / (tn - tp), out.size() - np);
+ tp = tn;
+ np = out.size();
+ }
if ((tok == 1 && std::getenv("DS4_NO_EOS") == nullptr) || (int) out.size() >= n) break; // (DS4_NO_EOS: decode n tokens whatever)
if (got.empty()) {
int off = 0, acc = 0;
diff --git a/src/dsv41/ds4_model.cu b/src/dsv41/ds4_model.cu
index 60f56c5ba..341536959 100644
--- a/src/dsv41/ds4_model.cu
+++ b/src/dsv41/ds4_model.cu
@@ -7584,6 +7584,13 @@ struct Model::Impl {
}
bool forward(const int* tk, int T, int logits_rows, std::string& err, int seed_rows = -1) {
+#if defined(STRATA_METAL)
+ struct BusyAll { // DS4_BUSY_ALL: the GPU is held busy for the whole step (the host gaps between its bursts too)
+ bool on;
+ BusyAll() : on(std::getenv("DS4_BUSY_ALL") != nullptr) { if (on) ds4_busy_begin(); }
+ ~BusyAll() { if (on) ds4_busy_end(); }
+ } busy_all;
+#endif
if (T < 1 || T > kWin) {
err = "a window of " + std::to_string(T) + " tokens";
return false;