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++
gpu_fit: a chip that shares the system memory adds only its BIOS carve-out beside the RAM (a Strix Halo model was counted as VRAM and as RAM)
Niko12218895949ef710
diff --git a/strata_control/gpu_fit.py b/strata_control/gpu_fit.pyindex adf6630fa..9a34002e3 100644--- a/strata_control/gpu_fit.py+++ b/strata_control/gpu_fit.py@@ -175,9 +175,18 @@ def _name(c: dict) -> str:return n + (f" (card {c.get('index')})" if c.get("_dup") else "")+def own_memory_gb(c: dict) -> float:+ """The graphics memory a card adds BESIDE the RAM. A discrete card: its VRAM. A chip that shares the system memory (Strix Halo, an APU): only+ its BIOS carve-out (`dedicated_gb`; 0 when the report does not give it) - its "shared" pool is the RAM itself, and counting it as VRAM as well as+ as RAM made a model look like it fit twice (setup.py low_ram_vram does the same)."""+ if c.get("uma") or c.get("integrated"):+ return float(c.get("dedicated_gb") or 0)+ return float(c.get("vram_gb") or 0)++def _option(key: str, cards: list[dict], memory_gb: float, ram_gb: float | None) -> dict[str, Any]:n = len(cards)- total = sum(float(c.get("vram_gb") or 0) for c in cards)+ total = sum(own_memory_gb(c) for c in cards)eff = total - SPLIT_CARD_OVERHEAD_GB * (n - 1)f = fit(memory_gb, ram_gb, eff)v = verdict(f)diff --git a/tests/terminal/test_gpu_fit.py b/tests/terminal/test_gpu_fit.pyindex e788af89e..88e001678 100644--- a/tests/terminal/test_gpu_fit.py+++ b/tests/terminal/test_gpu_fit.py@@ -68,6 +68,19 @@ class GpuFitRule(unittest.TestCase):self.assertEqual([c["index"] for c in keep], [0] if keep[0]["name"] == "RTX 3070 Ti" else [])self.assertEqual(len(ignored), 1)+ def test_a_strix_halo_counts_its_memory_once(self):+ """The 'shared' pool of an APU is the RAM: a 70 GB model on a 128 GB Strix Halo with a 2 GB carve-out is RAM-resident, not 'VRAM + RAM'."""+ igpu = {"index": 0, "name": "AMD Radeon 8060S", "vram_gb": 66.0, "dedicated_gb": 2.0, "uma": True, "usable": True}+ self.assertEqual(G.own_memory_gb(igpu), 2.0)+ self.assertEqual(G.own_memory_gb({"index": 1, "vram_gb": 32.0}), 32.0)+ self.assertEqual(G.own_memory_gb({"index": 0, "vram_gb": 66.0, "uma": True}), 0.0) # no carve-out figure: none counted+ # 64 GB PC, a 90 GB model: counted twice (66 GB 'VRAM' + 54 GB RAM) it looked like it fit; once, 36 GB come from the SSD+ r = G.options(90.0, 64.0, [dict(igpu, vram_gb=34.0)])+ self.assertEqual(r["options"][0]["verdict"], "spills")+ self.assertGreater(r["options"][0]["spill_gb"], 30.0)+ # 128 GB, 90 GB model: fits in RAM+ self.assertEqual(G.options(90.0, 128.0, [dict(igpu, vram_gb=66.0)])["options"][0]["verdict"], "fits")+def test_an_unusable_card_is_ignored_with_its_own_reason(self):r = G.options(44.0, 62.0, [card(0, "RTX 3070 Ti", 8), card(1, "GTX 1060", 6, usable=False, problem="older than the RTX 20 series")])self.assertEqual([o["key"] for o in r["options"]], ["single"])