diff --git a/bundles/llamacpp-vulkan-qwen36-35b-a3b/docker-compose.yml b/bundles/llamacpp-vulkan-qwen36-35b-a3b/docker-compose.yml index 3284d7a46..c42605e2f 100644 --- a/bundles/llamacpp-vulkan-qwen36-35b-a3b/docker-compose.yml +++ b/bundles/llamacpp-vulkan-qwen36-35b-a3b/docker-compose.yml @@ -65,5 +65,8 @@ services: # that was running at shutdown. The two policies differ only for containers # that were deliberately stopped, which is exactly the window's case. restart: unless-stopped - ipc: host + # No ipc: host. llama-server is one process: weights load with --no-mmap into GPU + # memory through Vulkan, and nothing shares POSIX shared memory with another process, + # so the host IPC namespace buys nothing (the install security check refuses it + # without privileged consent). A private /dev/shm of shm_size is kept as headroom. shm_size: 16g diff --git a/bundles/llamacpp-vulkan-qwen36-35b-a3b/manifest.json b/bundles/llamacpp-vulkan-qwen36-35b-a3b/manifest.json index dd610aa6c..72f59b527 100644 --- a/bundles/llamacpp-vulkan-qwen36-35b-a3b/manifest.json +++ b/bundles/llamacpp-vulkan-qwen36-35b-a3b/manifest.json @@ -1,5 +1,6 @@ { "id": "llamacpp-vulkan-qwen36-35b-a3b", + "version": "1.0.1", "type": "bundle", "name": "llama.cpp Qwen3.6-35B-A3B UD-Q6_K + Vision (Vulkan AMDVLK)", "category": "ai", diff --git a/registry/add-ons.json b/registry/add-ons.json index 064cf1f42..cd2679084 100644 --- a/registry/add-ons.json +++ b/registry/add-ons.json @@ -2701,6 +2701,7 @@ }, { "id": "llamacpp-vulkan-qwen36-35b-a3b", + "version": "1.0.1", "type": "bundle", "name": "llama.cpp Qwen3.6-35B-A3B UD-Q6_K + Vision (Vulkan AMDVLK)", "category": "ai", diff --git a/tests/compose-guard.test.js b/tests/compose-guard.test.js index c442340ea..fa2dcc483 100644 --- a/tests/compose-guard.test.js +++ b/tests/compose-guard.test.js @@ -684,3 +684,9 @@ test("no tracked bundle id ends in an owned-place suffix (so the suffix rule can const bad = [...trackedBundleSet()].filter((id) => G.OWN_SUFFIXES.some((x) => id.endsWith(x))); assert.deepEqual(bad, []); }); + +test("the 35B model bundle no longer asks for the host IPC namespace", () => { + const compose = readFileSync(join(REPO, "bundles", "llamacpp-vulkan-qwen36-35b-a3b", "docker-compose.yml"), "utf8"); + assert.doesNotMatch(compose.split("\n").filter((l) => !/^\s*#/.test(l)).join("\n"), /^\s*ipc:\s*host/m); + assert.match(compose, /^\s*shm_size:\s*16g/m); +});