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++
dsv41 tune: with the drafter on, also try a draft block of 3 against the pack's own
Niko12211904c0261bcc
diff --git a/tests/terminal/test_dsv41_tune.py b/tests/terminal/test_dsv41_tune.pyindex 1881b9420..aa81fb399 100644--- a/tests/terminal/test_dsv41_tune.py+++ b/tests/terminal/test_dsv41_tune.py@@ -48,7 +48,7 @@ class Tune(unittest.TestCase):args = json.loads(self.cfg.read_text())["args"]self.assertEqual(args[args.index("--dspark") + 1], "on")self.assertIn("faster", " ".join(res["lines"]))- self.assertEqual(len(calls), 3) # baseline, drafter on, threads+ self.assertEqual(len(calls), 4) # baseline, drafter on, block 3, threadsself.assertIn("tuned", json.loads(self.cfg.read_text()))def test_noise_changes_nothing(self):diff --git a/tools/dsv41/tune.py b/tools/dsv41/tune.pyindex 96bfa0347..1b286ba8c 100644--- a/tools/dsv41/tune.py+++ b/tools/dsv41/tune.py@@ -5,7 +5,7 @@What it does (a few short decode runs through the engine's own command line; each loads the model, so about 1-3 minutes each):1. the launch's own settings (the baseline);2. the same with the DSpark drafter forced on (`--dspark on --dspark-conf 0.65`) against `--dspark auto`, which leaves it unloaded on a- small card: kept only when it is clearly faster;+ small card: kept only when it is clearly faster; with it on, a draft block of 3 against the pack's own (5);3. the CPU threads: the number of physical cores against every logical thread (SMT), when they differ: kept only when clearly faster.The RAM store (`--ssd-cache`) sizes itself from the free memory at start, so it is left as the engine's own; the line says so.A setting is changed only for a gain of more than 5% (run-to-run noise on this engine is about 5%). Whatever is kept is written into@@ -119,12 +119,18 @@ def tune(config: Path, mode: str = "quick", n: int = 48, timeout: int = 1500, mebest, best_args = base, base_argssay(f" standard settings: {base:.2f} tokens a second")trials = [("the drafter on (DSpark)", lambda a: with_flag(with_flag(a, "--dspark", "on"), "--dspark-conf", "0.65"))]+ trials.append(("a shorter draft block (3)", lambda a: with_flag(a, "--dspark-block", "3")))cores, logical = physical_cores(), os.cpu_count() or 0if cores and logical and logical > cores and "--cpu-threads" not in base_args:trials.append((f"{cores} CPU threads (one for each physical core)", lambda a, c=cores: with_flag(a, "--cpu-threads", str(c))))if mode == "thorough" and cores:trials.append(("every logical CPU thread", lambda a, c=logical: with_flag(a, "--cpu-threads", str(c))))- for label, make in trials:+ queue = list(trials)+ while queue:+ label, make = queue.pop(0)+ if label == "a shorter draft block (3)":+ if "on" != (best_args[best_args.index("--dspark") + 1] if "--dspark" in best_args else ""):+ continue # only with the drafter on: with `auto` it may not even be loadedcand = make(best_args)say(f" trying {label} ...")v = measure_fn(exe, cand, f, n, timeout, env)