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++

setup: say when the BIOS gives a big carve-out to a Strix Halo this run does not use (the OS loses that RAM; the experts and the desktop need it)

Niko122130a67051d002
diff --git a/setup.py b/setup.py
index 514290d99..cd0c8fbda 100644
--- a/setup.py
+++ b/setup.py
@@ -3625,6 +3625,19 @@ def igpu_notes(gpu, ram) -> list[str]:
return []
+def unused_apu_notes(chosen, amd, ram) -> list[str]:
+ """A Strix Halo / APU that is NOT in the run but sits beside the card that is (a Radeon AI PRO R9700 and the 8060S): its BIOS carve-out is memory
+ the OS does not have, and it is the RAM the experts live in. Nothing is changed: a "!" line says what is lost and what to set."""
+ out = []
+ used = {g["index"] for g in chosen}
+ for g in amd:
+ if g.get("uma") and g["index"] not in used and g.get("dedicated_gb", 0.0) > 16:
+ out.append(f"!the BIOS gives {g['dedicated_gb']:.0f} GB of your memory to {g['name']}, which this run does not use: the OS has "
+ f"{ram:.0f} GB left for the experts and the desktop, and the engine can run out of RAM. A small carve-out (512 MB - 2 GB, "
+ "'UMA Frame Buffer Size' in the BIOS) gives it back; the GPU reaches the shared memory anyway")
+ return out
+
+
def strix_halo_notes(gpu, ram) -> list[str]:
"""What setup tells a Strix Halo owner (a leading "!" = a warning): the unified memory, the build, the guide, and the
memory settings that give the GPU room. Nothing is changed or refused: recommend, never force."""
@@ -8086,6 +8099,8 @@ def main() -> int:
if gpu.get("uma"):
for line in igpu_notes(gpu, ram_gb()):
(warn if line.startswith("!") else say)(line.lstrip("!"))
+ for line in unused_apu_notes(chosen, amd, ram_gb()):
+ (warn if line.startswith("!") else say)(line.lstrip("!"))
for line in rdna2_notes(gpu):
say(line)
else:
diff --git a/tools/test_setup_amd.py b/tools/test_setup_amd.py
index f612246a0..276653284 100644
--- a/tools/test_setup_amd.py
+++ b/tools/test_setup_amd.py
@@ -146,6 +146,16 @@ class GpuLists(unittest.TestCase):
self.assertEqual([g["index"] for g in setup.amd_parse_gpus("all", [apu, r9700])], [1, 0])
self.assertEqual([g["index"] for g in setup.amd_parse_gpus("0,1", [apu, r9700, r9070])], [0, 1])
+ def test_an_unused_apu_with_a_big_carve_out_is_said(self):
+ apu = {"index": 0, "name": "AMD Radeon 8060S", "vram_gb": 96.0, "arch": "gfx1151", "uma": True, "dedicated_gb": 64.0}
+ r9700 = {"index": 1, "name": "AMD Radeon AI PRO R9700", "vram_gb": 32.0, "arch": "gfx1201"}
+ out = setup.unused_apu_notes([r9700], [apu, r9700], 62.0)
+ self.assertEqual(len(out), 1)
+ self.assertTrue(out[0].startswith("!") and "64 GB" in out[0] and "62 GB" in out[0])
+ self.assertEqual(setup.unused_apu_notes([apu, r9700], [apu, r9700], 62.0), []) # in the run: not "unused"
+ self.assertEqual(setup.unused_apu_notes([r9700], [dict(apu, dedicated_gb=2.0), r9700], 120.0), []) # a small carve-out: nothing to say
+ self.assertEqual(setup.unused_apu_notes([r9700], [r9700], 120.0), [])
+
def test_refusals(self):
for text in ("1,2", "1,7", "1", "1,1", "x,y"):
with self.assertRaises(SystemExit, msg=text):