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++

Control: the NVIDIA driver version also from /proc/driver/nvidia/version (Fedora has no nvidia-smi without its cuda package); an integrated AMD GPU beside a usable NVIDIA card is not offered

Niko1221254ee2b9cd6f
diff --git a/strata_control/engine_component.py b/strata_control/engine_component.py
index 902711277..5b784155a 100644
--- a/strata_control/engine_component.py
+++ b/strata_control/engine_component.py
@@ -474,6 +474,12 @@ def _smi_query(field_name: str) -> list[str]:
def nvidia_driver_version() -> str | None:
"""The NVIDIA driver version nvidia-smi reports (the oldest one when several cards disagree); None when unreadable."""
versions = _smi_query("driver_version")
+ if not versions: # Linux without nvidia-smi (Fedora: it is in xorg-x11-drv-nvidia-cuda): the kernel module says
+ try:
+ found = re.search(r"Kernel Module\s+(\d+(?:\.\d+)+)", Path("/proc/driver/nvidia/version").read_text())
+ versions = [found.group(1)] if found else []
+ except OSError:
+ versions = []
return min(versions, key=_version_key) if versions else None
diff --git a/tests/terminal/test_engine_component.py b/tests/terminal/test_engine_component.py
index 52064a864..49396d289 100644
--- a/tests/terminal/test_engine_component.py
+++ b/tests/terminal/test_engine_component.py
@@ -309,6 +309,15 @@ class NeverAcquires(unittest.TestCase):
with mock.patch.object(ec, "nvidia_driver_version", return_value="580.178.04"):
self.assertEqual(ec.resolve(self.root, backend="cuda", runtime="1").root.name, a.name, "the newer-driver build wins")
+ def test_the_driver_version_comes_from_the_kernel_module_when_there_is_no_nvidia_smi(self):
+ text = "NVRM version: NVIDIA UNIX x86_64 Kernel Module 580.95.05 Thu Sep 11 2026 GCC version: gcc 15"
+ with mock.patch.object(ec, "_smi_query", return_value=[]), mock.patch.object(Path, "read_text", return_value=text):
+ self.assertEqual(ec.nvidia_driver_version(), "580.95.05")
+ with mock.patch.object(ec, "_smi_query", return_value=["575.57.08", "580.95.05"]):
+ self.assertEqual(ec.nvidia_driver_version(), "575.57.08", "the oldest when cards disagree")
+ with mock.patch.object(ec, "_smi_query", return_value=[]), mock.patch.object(Path, "read_text", side_effect=OSError):
+ self.assertIsNone(ec.nvidia_driver_version())
+
def test_the_component_fetcher_refuses_where_acquiring_is_forbidden(self):
store = components.ComponentStore(self.root / "store")
artifact = components.Artifact(name="engine.tar.gz", url="https://example.invalid/e.tgz",
diff --git a/tools/strata_mcp.py b/tools/strata_mcp.py
index 1ed5e25bb..93fb8f9b8 100644
--- a/tools/strata_mcp.py
+++ b/tools/strata_mcp.py
@@ -759,6 +759,10 @@ class Strata:
amd = []
if not amd and shutil.which("rocm-smi"):
amd = rocm_smi_gpus()
+ if any(g.get("usable") for g in nv):
+ for g in amd: # a Ryzen's integrated graphics beside an NVIDIA card: the engine runs on the NVIDIA card
+ if g.get("uma") and g.get("usable"):
+ g["usable"], g["problem"] = False, "integrated graphics: Strata uses the NVIDIA card in this computer"
hw["gpus"] = nv + amd
if WIN:
hw["display_adapters"] = windows_video_controllers()
diff --git a/tools/test_strata_mcp.py b/tools/test_strata_mcp.py
index 0ce2afe46..fbcfff056 100644
--- a/tools/test_strata_mcp.py
+++ b/tools/test_strata_mcp.py
@@ -692,6 +692,25 @@ class MacCatalog(unittest.TestCase):
self.assertIn("unsloth", [f["family"] for f in s.model_table({"ram_gb": 96, "gpus": []})])
+class MixedGraphics(unittest.TestCase):
+ def test_an_integrated_amd_gpu_beside_an_nvidia_card_is_not_offered(self):
+ s = M.Strata(HERE.parent)
+ S = s.setup_module()
+ nv = [{"index": 0, "name": "NVIDIA GeForce RTX 5090", "vram_gb": 31.8, "arch": "120", "driver": "580.95"}]
+ igpu = [{"index": 0, "name": "AMD Radeon Graphics", "vram_gb": 0.5, "arch": "gfx1036", "uma": True}]
+ with mock.patch.object(M.sys, "platform", "linux"), mock.patch.object(M, "WIN", False), mock.patch.object(S, "gpus", lambda: [dict(g) for g in nv]), mock.patch.object(S, "gpu_problem", lambda g: None), mock.patch.object(S, "amd_gpus", lambda: [dict(g) for g in igpu]), mock.patch.object(S, "amd_problem", lambda g: None):
+ hw = s._hardware()
+ by = {g["vendor"]: g for g in hw["gpus"]}
+ self.assertTrue(by["nvidia"]["usable"])
+ self.assertFalse(by["amd"]["usable"])
+ self.assertIn("NVIDIA card", by["amd"]["problem"])
+ rec = s.recommend({**hw, "ram_gb": 92.0})
+ self.assertEqual((rec["backend"], rec["gpu"]["name"]), ("cuda", "NVIDIA GeForce RTX 5090"))
+ # with no NVIDIA card the integrated GPU stays what it was
+ with mock.patch.object(M.sys, "platform", "linux"), mock.patch.object(M, "WIN", False), mock.patch.object(S, "gpus", lambda: []), mock.patch.object(S, "amd_gpus", lambda: [dict(g) for g in igpu]), mock.patch.object(S, "amd_problem", lambda g: None):
+ self.assertTrue(s._hardware()["gpus"][0]["usable"])
+
+
if __name__ == "__main__":
unittest.main()