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++

launch: the 16 GiB rule for the Speed label is compared in whole MiB; the catalog's vram_gb is already GiB (nvidia-smi MiB / 1024, AMD bytes / 2**30), so a 16 GB card (16376 MiB) counts as 16 GiB or less; tests with exact MiB values

Niko1221bd7acc0e6dd0
diff --git a/strata_cli/launch/settings.py b/strata_cli/launch/settings.py
index d653098e7..0345ce1b8 100644
--- a/strata_cli/launch/settings.py
+++ b/strata_cli/launch/settings.py
@@ -36,7 +36,9 @@ FASTER_NOTE = "Faster uses 5 of 6 experts per word: a small quality change, usua
def faster_default(total_vram_gb: float | None) -> bool:
"""Mirrors the engine's `--experts auto`: Faster (5 experts) when the cards the launch uses total 16 GiB (17,179,869,184 bytes) or less.
The engine decides at its start; this only says the same thing in the screens and in --plan, so the two never disagree."""
- return bool(total_vram_gb) and float(total_vram_gb) <= FASTER_VRAM_GIB
+ # The catalog's vram_gb is already GiB: nvidia-smi's memory.total in MiB / 1024 (setup.py gpus(), tools/strata_mcp.py), an AMD card's bytes / 2**30,
+ # Windows registry bytes / 2**30. Compared in whole MiB so a 16 GB card (RTX 4080 / 4060 Ti 16 GB: 16376 MiB = 15.99 GiB) counts as 16 GiB or less.
+ return bool(total_vram_gb) and round(float(total_vram_gb) * 1024) <= round(FASTER_VRAM_GIB * 1024)
def speed_words(speed: str, default: bool) -> str:
diff --git a/tests/terminal/test_launch_speed.py b/tests/terminal/test_launch_speed.py
index 5d902b06f..ca8853584 100644
--- a/tests/terminal/test_launch_speed.py
+++ b/tests/terminal/test_launch_speed.py
@@ -37,6 +37,46 @@ class Words(unittest.TestCase):
self.assertIn("--speed faster", S.flags_text(S.LaunchSettings(family="deepseek41", model="IST-DAS-2.00bpw", speed="faster")))
+class Units(unittest.TestCase):
+ """The catalog's vram_gb is GiB, the engine's unit: checked through the code that reads the cards (#3843-style: exact MiB values)."""
+
+ def nvidia(self, *mib):
+ text = "\n".join(f"{i}, NVIDIA GeForce RTX {i}, {m}, 8.9, 580.97" for i, m in enumerate(mib))
+ import setup
+ from tools import strata_mcp
+ with mock.patch.object(setup, "out", return_value=text), mock.patch.object(strata_mcp, "run_quiet", return_value=text):
+ return setup.gpus(), strata_mcp.own_nvidia_gpus()
+
+ def total(self, cards):
+ return sum(round(c["vram_gb"], 1) for c in cards) # the catalog rounds each card to 0.1 GiB
+
+ def test_nvidia_mib_is_divided_by_1024(self):
+ for cards in self.nvidia(16376):
+ self.assertAlmostEqual(cards[0]["vram_gb"], 16376 / 1024)
+
+ def test_16_gb_cards_count_as_16_gib_or_less(self):
+ for mib in (16376, 16384, 16380): # RTX 4080 / 4060 Ti 16 GB, an exact 16 GiB card, in between
+ for cards in self.nvidia(mib):
+ self.assertTrue(S.faster_default(self.total(cards)), mib)
+ for mib in (12288, 12282, 8188): # RTX 3060, 12 GiB, 8 GB
+ for cards in self.nvidia(mib):
+ self.assertTrue(S.faster_default(self.total(cards)), mib)
+
+ def test_bigger_cards_and_sums_over_16_gib_do_not(self):
+ for mib in (24564, 24576, 20475, 32607):
+ for cards in self.nvidia(mib):
+ self.assertFalse(S.faster_default(self.total(cards)), mib)
+ for cards in self.nvidia(8188, 8188): # two 8 GB cards: 15.99 GiB together
+ self.assertTrue(S.faster_default(self.total(cards)))
+ for cards in self.nvidia(12288, 8188): # 12 + 8: over
+ self.assertFalse(S.faster_default(self.total(cards)))
+
+ def test_an_amd_16_gb_card_counts(self):
+ for size in (17179869184, 17163091968): # 16 GiB exactly, and a 16 GB card's 16,368 MiB
+ self.assertTrue(S.faster_default(round(size / 2 ** 30, 1)), size)
+ self.assertFalse(S.faster_default(round(34342961152 / 2 ** 30, 1))) # a 32 GB card
+
+
class EngineFlag(unittest.TestCase):
def test_experts_replaces_the_saved_value_and_goes_last(self):
cfg = {"args": ["--pack", "p", "--experts", "6", "--dspark", "auto"]}