Mirrors/ProxMenux
1
0
Fork 0
mirror of https://github.com/MacRimi/ProxMenux.git synced 2026-11-02 10:52:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(oci): run Ollama and llama.cpp on Intel and AMD integrated GPUs

Browse Source
This commit is contained in:
MacRimi
2026-10-31 00:25:21 +01:00
parent 81cc746e19
commit f9fe4d2793
6 changed files with 226 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
oci/tests/test_acceleration_profiles_of_the_host.py
+20 -3
View File
@@ -48,8 +48,25 @@ class HostProfileTests(unittest.TestCase):
def test_an_amd_host_is_not_offered_nvidia(self, *_):
amd = {"intel": [], "amd": [NODE], "nvidia": False}
self.assertEqual(self.offered("frigate", amd, kfd=True)[0], ["none", "vaapi", "rocm"])
self.assertEqual(self.offered("ollama", amd, kfd=True)[0], ["cpu", "rocm"])
self.assertEqual(self.offered("llamacpp", amd, kfd=True)[0], ["cpu", "rocm"])
self.assertEqual(self.offered("ollama", amd, kfd=True)[0], ["cpu", "vulkan", "rocm"])
self.assertEqual(self.offered("llamacpp", amd, kfd=True)[0], ["cpu", "vulkan", "rocm"])
def test_a_radeon_780m_gets_rocm_as_experimental_with_its_generation(self, *_):
# gfx1103 is not among the kernels of the Ollama and llama.cpp ROCm
# images; it runs presented as gfx1100. Without its compute device,
# Vulkan still uses it.
amd = {"intel": [], "amd": [NODE], "nvidia": False, "amd_gfx_target": 110003}
for app in ("ollama", "llamacpp"):
with patch("proxmenux_oci.installer.confirm_experimental_rocm", return_value=True):
plan = self.offered(app, amd, kfd=True, answer="rocm")[2]
override = [item["value"] for item in plan["environment"] if item["name"] == "HSA_OVERRIDE_GFX_VERSION"]
self.assertEqual(override, ["11.0.0"], app)
if app == "ollama":
self.assertIn({"name": "OLLAMA_IGPU_ENABLE", "value": "1", "sensitive": False}, plan["environment"])
self.assertEqual(self.offered(app, amd, kfd=False)[0], ["cpu", "vulkan"], app)
# Ollama leaves an integrated GPU out unless it is told to use it.
plan = self.offered("ollama", amd, kfd=False, answer="vulkan")[2]
self.assertIn({"name": "OLLAMA_IGPU_ENABLE", "value": "1", "sensitive": False}, plan["environment"])
def test_rocm_is_not_offered_without_its_compute_device(self, *_):
amd = {"intel": [], "amd": [NODE], "nvidia": False}
@@ -58,7 +75,7 @@ class HostProfileTests(unittest.TestCase):
def test_an_intel_and_nvidia_host_is_not_offered_amd(self, *_):
both = {"intel": [NODE], "amd": [], "nvidia": True}
self.assertEqual(self.offered("frigate", both)[0], ["none", "vaapi", "nvidia"])
self.assertEqual(self.offered("llamacpp", both)[0], ["cpu", "nvidia", "intel"])
self.assertEqual(self.offered("llamacpp", both)[0], ["cpu", "nvidia", "vulkan", "intel"])
def test_a_host_without_gpu_is_told_and_not_asked(self, *_):
nothing = {"intel": [], "amd": [], "nvidia": False}