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++
Fix #3490: tuning stops waiting for memory when none is being freed instead of ten minutes per candidate, and says how much is free and needed
Niko122175d0d0b96654
diff --git a/strata_control/autotune.py b/strata_control/autotune.pyindex a6d37c3cd..083f23591 100644--- a/strata_control/autotune.py+++ b/strata_control/autotune.py@@ -37,6 +37,10 @@ EARLY_DROP = 0.05 # a candidate this much slower than its own rSTAGE_GAP_NOISY = 0.04 # an A/A gap above this: the machine is noisy (6.5)+STAGNANT_WAIT_S = 60.0 # free memory that has not grown by STAGNANT_MIB after this long is not going to+STAGNANT_MIB = 512++class Cancelled(Exception):"""The user stopped the tune; what was measured so far is kept in the result."""@@ -193,6 +197,7 @@ class Tuner:self.progress = progress or (lambda p: None)self.cancel = cancel or threading.Event()self.history_bad = history_bad or set()+ self._ram_gave_up: tuple[float, float] | None = Noneself.clock = clockself.sleep = sleepself.system = system or session.system@@ -261,8 +266,10 @@ class Tuner:return Nonedef _ram_guard(self, delta_mib: int, label: str, wait_s: float | None = None) -> str | None:- """None when a start fits; otherwise the plain reason it does not."""+ """None when a start fits; otherwise the plain reason it does not. Waiting ends early when the free memory does not+ grow (a loaded model keeps it): ten minutes per candidate made a tuning run last for hours (#3490)."""waited = 0.0+ first_avail: float | None = Nonewhile True:mem = self.system().get("memory", {})avail = mem.get("available_mib")@@ -276,10 +283,19 @@ class Tuner:need = max(need, self.o.min_free_ram_mib)if avail >= need:return None- if waited >= (self.o.wait_ram_s if wait_s is None else wait_s):- return (f"not enough free memory to start it safely ({avail / 1024:.1f} GiB free, "- f"{need / 1024:.1f} GiB needed)")- self._say("waiting", f"Waiting for memory to be freed before trying {label}")+ reason = (f"not enough free memory to start it safely ({avail / 1024:.1f} GiB free, "+ f"{need / 1024:.1f} GiB needed)")+ prior = self._ram_gave_up+ if prior is not None and need >= prior[0] and avail <= prior[1] + STAGNANT_MIB:+ return reason # an earlier wait found nothing freed and this needs no less+ if first_avail is None:+ first_avail = avail+ stagnant = waited >= STAGNANT_WAIT_S and avail <= first_avail + STAGNANT_MIB+ if stagnant or waited >= (self.o.wait_ram_s if wait_s is None else wait_s):+ self._ram_gave_up = (need, avail)+ return reason+ self._say("waiting", f"Waiting for memory to be freed before trying {label} "+ f"({avail / 1024:.1f} GiB free, {need / 1024:.1f} GiB needed)")self._check_cancel()self.sleep(10.0)waited += 10.0diff --git a/tests/tuning/test_autotune.py b/tests/tuning/test_autotune.pyindex 3a99b111d..2efd9556e 100644--- a/tests/tuning/test_autotune.py+++ b/tests/tuning/test_autotune.py@@ -267,6 +267,22 @@ class TunerTest(unittest.TestCase):self.assertEqual(r.outcome, "default_confirmed")self.assertTrue(any(a["status"] == "skipped" and "memory" in a["reason"] for a in r.arms))+ def test_memory_wait_gives_up_when_nothing_is_freed(self):+ # #3490: a loaded model keeps the memory; ten minutes of waiting per candidate made a tuning last for hours+ model = speed_with({})+ clock = FakeClock()+ session = StubSession(model=model, clock=clock)++ def system():+ out = session.system()+ out["memory"]["available_mib"] = 40000 if not session.runs else 3000+ return out+ t = autotune.Tuner(session, BASE, {}, make_ctx(), autotune.Options(per_request=False), clock=clock,+ sleep=clock.sleep, system=system, total_ram_mib=64000, quiet=lambda: None)+ before = clock()+ t.run()+ self.assertLess(clock() - before, 600.0, "waited for the whole ten minutes per candidate")+def test_nothing_compared_is_inconclusive_not_default_confirmed(self):model = speed_with({})clock = FakeClock()