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

oci/catalog/apps/llamacpp.json

@@ -305,9 +305,53 @@
               }
             ]
           },
+          {
+            "id": "vulkan",
+            "label": "Intel/AMD (Vulkan)",
+            "architectures": [
+              "amd64"
+            ],
+            "image": {
+              "reference": "ghcr.io/ggml-org/llama.cpp:server-vulkan",
+              "registry": "ghcr.io",
+              "repository": "ghcr.io/ggml-org/llama.cpp",
+              "tag": "server-vulkan",
+              "digest": null,
+              "pull_policy": "resolve-selected-tag-to-architecture-digest-at-install"
+            },
+            "device_requests": [
+              {
+                "id": "gpu-render",
+                "path_prompt": "GPU render device",
+                "host_path_default": "/dev/dri/renderD128",
+                "mode": "0660",
+                "deny_write": false,
+                "gid_strategy": "host-device-gid",
+                "kind": "character-device",
+                "container_path_strategy": "same-as-host",
+                "drm_vendor_ids": [
+                  "0x8086",
+                  "0x1002"
+                ]
+              }
+            ]
+          },
           {
             "id": "rocm",
             "label": "AMD (ROCm)",
+            "amd_gfx_targets": [
+              90008,
+              90010,
+              90402,
+              100300,
+              110000,
+              110001,
+              110002,
+              110500,
+              110501,
+              120000,
+              120001
+            ],
             "architectures": [
               "amd64"
             ],

oci/catalog/apps/ollama.json

@@ -410,9 +410,60 @@
               }
             ]
           },
+          {
+            "id": "vulkan",
+            "label": "Intel/AMD (Vulkan)",
+            "architectures": [
+              "amd64"
+            ],
+            "image": {
+              "reference": "ollama/ollama:latest",
+              "registry": "docker.io",
+              "repository": "ollama/ollama",
+              "tag": "latest",
+              "digest": null,
+              "pull_policy": "resolve-selected-tag-to-architecture-digest-at-install"
+            },
+            "device_requests": [
+              {
+                "id": "gpu-render",
+                "path_prompt": "GPU render device",
+                "host_path_default": "/dev/dri/renderD128",
+                "mode": "0660",
+                "deny_write": false,
+                "gid_strategy": "host-device-gid",
+                "kind": "character-device",
+                "container_path_strategy": "same-as-host",
+                "drm_vendor_ids": [
+                  "0x8086",
+                  "0x1002"
+                ]
+              }
+            ],
+            "environment": [
+              {
+                "name": "OLLAMA_IGPU_ENABLE",
+                "value": "1"
+              }
+            ]
+          },
           {
             "id": "rocm",
             "label": "AMD (ROCm)",
+            "amd_gfx_targets": [
+              90008,
+              90010,
+              90402,
+              90500,
+              100300,
+              110000,
+              110001,
+              110002,
+              110500,
+              110501,
+              120000,
+              120001
+            ],
             "architectures": [
               "amd64"
             ],
@@ -448,6 +499,12 @@
                 "kind": "character-device",
                 "container_path_strategy": "same-as-host"
               }
+            ],
+            "environment": [
+              {
+                "name": "OLLAMA_IGPU_ENABLE",
+                "value": "1"
+              }
             ]
           }
         ]

oci/catalog/curated/llamacpp.json

@@ -305,9 +305,53 @@
               }
             ]
           },
+          {
+            "id": "vulkan",
+            "label": "Intel/AMD (Vulkan)",
+            "architectures": [
+              "amd64"
+            ],
+            "image": {
+              "reference": "ghcr.io/ggml-org/llama.cpp:server-vulkan",
+              "registry": "ghcr.io",
+              "repository": "ghcr.io/ggml-org/llama.cpp",
+              "tag": "server-vulkan",
+              "digest": null,
+              "pull_policy": "resolve-selected-tag-to-architecture-digest-at-install"
+            },
+            "device_requests": [
+              {
+                "id": "gpu-render",
+                "path_prompt": "GPU render device",
+                "host_path_default": "/dev/dri/renderD128",
+                "mode": "0660",
+                "deny_write": false,
+                "gid_strategy": "host-device-gid",
+                "kind": "character-device",
+                "container_path_strategy": "same-as-host",
+                "drm_vendor_ids": [
+                  "0x8086",
+                  "0x1002"
+                ]
+              }
+            ]
+          },
           {
             "id": "rocm",
             "label": "AMD (ROCm)",
+            "amd_gfx_targets": [
+              90008,
+              90010,
+              90402,
+              100300,
+              110000,
+              110001,
+              110002,
+              110500,
+              110501,
+              120000,
+              120001
+            ],
             "architectures": [
               "amd64"
             ],

oci/catalog/overlays/ollama.json

@@ -51,9 +51,60 @@
               }
             ]
           },
+          {
+            "id": "vulkan",
+            "label": "Intel/AMD (Vulkan)",
+            "architectures": [
+              "amd64"
+            ],
+            "image": {
+              "reference": "ollama/ollama:latest",
+              "registry": "docker.io",
+              "repository": "ollama/ollama",
+              "tag": "latest",
+              "digest": null,
+              "pull_policy": "resolve-selected-tag-to-architecture-digest-at-install"
+            },
+            "device_requests": [
+              {
+                "id": "gpu-render",
+                "path_prompt": "GPU render device",
+                "host_path_default": "/dev/dri/renderD128",
+                "mode": "0660",
+                "deny_write": false,
+                "gid_strategy": "host-device-gid",
+                "kind": "character-device",
+                "container_path_strategy": "same-as-host",
+                "drm_vendor_ids": [
+                  "0x8086",
+                  "0x1002"
+                ]
+              }
+            ],
+            "environment": [
+              {
+                "name": "OLLAMA_IGPU_ENABLE",
+                "value": "1"
+              }
+            ]
+          },
           {
             "id": "rocm",
             "label": "AMD (ROCm)",
+            "amd_gfx_targets": [
+              90008,
+              90010,
+              90402,
+              90500,
+              100300,
+              110000,
+              110001,
+              110002,
+              110500,
+              110501,
+              120000,
+              120001
+            ],
             "architectures": [
               "amd64"
             ],
@@ -89,6 +140,12 @@
                 "kind": "character-device",
                 "container_path_strategy": "same-as-host"
               }
+            ],
+            "environment": [
+              {
+                "name": "OLLAMA_IGPU_ENABLE",
+                "value": "1"
+              }
             ]
           }
         ]

oci/tests/test_acceleration_profiles_of_the_host.py

@@ -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("ollama", amd, kfd=True)[0], ["cpu", "vulkan", "rocm"])
-        self.assertEqual(self.offered("llamacpp", amd, kfd=True)[0], ["cpu", "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}

oci/tests/test_ai_gpu_profiles.py

@@ -17,9 +17,11 @@ from test_advanced_flow_order import RecordingUI, storages
 
 RENDER, KFD = "/dev/dri/renderD128", "/dev/kfd"
 EXPECTED = {
-    "ollama": {"cpu": (":latest", []), "nvidia": (":latest", ["nvidia-runtime"]), "rocm": (":rocm", [RENDER, KFD])},
+    "ollama": {"cpu": (":latest", []), "nvidia": (":latest", ["nvidia-runtime"]), "vulkan": (":latest", [RENDER]),
+               "rocm": (":rocm", [RENDER, KFD])},
     "llamacpp": {"cpu": (":server", []), "nvidia": (":server-cuda", ["nvidia-runtime"]),
-                 "rocm": (":server-rocm", [RENDER, KFD]), "intel": (":server-intel", [RENDER])},
+                 "vulkan": (":server-vulkan", [RENDER]), "rocm": (":server-rocm", [RENDER, KFD]),
+                 "intel": (":server-intel", [RENDER])},
     "faster-whisper": {"cpu": (":latest", []), "nvidia": (":gpu", ["nvidia-runtime"])},
     "piper": {"cpu": (":latest", []), "nvidia": (":gpu", ["nvidia-runtime"])},
 }