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++

metal: histogram off the hot path, dedupe and host_view experiments reverted, chunk sweep recorded

Niko12211a80ccc98133
diff --git a/metal/src/engine/ds4_finish.cpp b/metal/src/engine/ds4_finish.cpp
index 93b1be059..2ca4f6700 100644
--- a/metal/src/engine/ds4_finish.cpp
+++ b/metal/src/engine/ds4_finish.cpp
@@ -18,6 +18,8 @@
// called by the gpu layer's synchronisation: everything batched must reach the GPU before a host read
extern "C" void ds4_metal_flush();
+extern "C" unsigned long long ds4_dispatch_seq();
+static unsigned long long g_q8_deduped = 0;
#define DSV41_PROBE(fmt, ...) do { static bool once_ = false; if (!once_) { once_ = true; std::fprintf(stderr, "[probe] " fmt, __VA_ARGS__); } } while (0)
diff --git a/metal/src/mtl_rt.mm b/metal/src/mtl_rt.mm
index 9b5fb03df..fb6cfe11b 100644
--- a/metal/src/mtl_rt.mm
+++ b/metal/src/mtl_rt.mm
@@ -262,10 +262,6 @@ static std::unordered_map<void*, std::string> g_pipe_names;
static std::unordered_map<std::string, uint64_t> g_kernel_counts;
static void encode(id<MTLComputeCommandEncoder> enc, const Dispatch& d) {
require_fit(d, "dispatch");
- {
- std::lock_guard<std::mutex> lk(g_hist_mu);
- ++g_kernel_counts[d.p ? d.p->name : std::string("<null>")];
- }
[enc setComputePipelineState:P(d.p)];
for (size_t i = 0; i < d.args.size(); ++i) {
const Arg& a = d.args[i];
@@ -301,6 +297,8 @@ static void ds4_report_counts() {
std::fprintf(stderr, " %6llu %s\n", (unsigned long long) v[i].first, v[i].second.c_str());
}
+extern "C" unsigned long long ds4_dispatch_seq() { return g_disp_count.load(); }
+
void Context::flush_runs() {
if (!g_count_reg) { g_count_reg = true; std::atexit(ds4_report_counts); }
if (impl_->cb_def == nil) return;