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: the DSpark drafter stays off on a Mac unless asked for (its quantised Markov bias has no Metal kernel; the setup test stopped with kernel launch with no Metal dispatch)

Niko1221d50f6a77ed7f
diff --git a/src/dsv41/ds4_model.cu b/src/dsv41/ds4_model.cu
index 83a97ac58..980bc0deb 100644
--- a/src/dsv41/ds4_model.cu
+++ b/src/dsv41/ds4_model.cu
@@ -2229,6 +2229,14 @@ struct Model::Impl {
if (o.vision_only) return true;
// layers: roles and the sources they read
+#if defined(STRATA_METAL)
+ // On a Mac the drafter is off unless it is asked for outright: it gave the same answers and no speed-up there, and the head of its
+ // quantised Markov bias has no Metal kernel (the engine stopped with "kernel launch with no Metal dispatch" on the setup test).
+ if (o.dspark_auto && o.dspark && g.n_mtp > 0) {
+ o.dspark = false;
+ std::fprintf(stderr, "ds4: DSpark not loaded on a Mac (--dspark on loads it anyway)\n");
+ }
+#endif
if (o.dspark_auto && o.dspark && g.n_mtp > 0) { // (see Options::dspark_auto)
double exp_b = 0, room_b = 0;
for (const auto& kt : pk.t)
diff --git a/tools/dsv41/smoke.py b/tools/dsv41/smoke.py
index 3a3e19562..7a2b94e9c 100644
--- a/tools/dsv41/smoke.py
+++ b/tools/dsv41/smoke.py
@@ -77,8 +77,8 @@ def main() -> int:
with tempfile.TemporaryDirectory() as d:
f = Path(d) / "ids.txt"
f.write_text(",".join(str(i) for i in ids), encoding="ascii")
- cmd = [a.exe, "--pack", str(pack), "--ids", "@" + str(f), "--n", str(a.tokens), "--max-context", "1024", "--dspark", "auto",
- "--window", "16"]
+ cmd = [a.exe, "--pack", str(pack), "--ids", "@" + str(f), "--n", str(a.tokens), "--max-context", "1024", "--dspark", "none" if sys.platform == "darwin" else "auto",
+ "--window", "16"] # (a Mac runs without the drafter: its Metal engine has no kernels for it)
if a.gpus:
cmd += ["--gpus", a.gpus]
try: