llama.cpp: Embedding mit -ub 4096 (ub 8192 braucht 5 GB Rechenpuffer auf der P2000)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

deploy/compose/gpu-llamacpp/docker-compose.yml

@@ -39,7 +39,7 @@ services:
     container_name: llamacpp-embed
     profiles: ["embed"]
     command: ["-m", "/models/${EMBED_GGUF:-Qwen3-Embedding-0.6B-Q8_0.gguf}", "--embedding",
-              "--pooling", "${EMBED_POOLING:-last}", "-c", "8192", "-b", "8192", "-ub", "8192", "-np", "1",
+              "--pooling", "${EMBED_POOLING:-last}", "-c", "8192", "-b", "4096", "-ub", "4096", "-np", "1",   # ub 8192 sprengt die P2000 (Task 6.1)
               "-ngl", "99", "--alias", "embed-default", "--metrics", "--no-slots", "--no-webui",
               "--host", "0.0.0.0", "--port", "8081"]
     deploy:

platform/proxmox/ansible/roles/gpu-llamacpp/defaults/main.yml

@@ -60,8 +60,10 @@ llamacpp_k3s_services:
       port: 8081
       gpu_var: EMBED_GPU
       memory_limit: 3Gi
+      # -ub 4096 statt 8192: Der Rechenpuffer wächst mit ub × Vokabular (151k) und bräuchte bei 8192 allein 5 GB
+      # (19.10. gemessen). Mit 4096: 2,5 GB Puffer, P2000 mit Rerank 4,4–4,6 von 5 GB, 5.467 Tokens in einem Aufruf ok.
       args: ["-m", "/models/{{ llamacpp_env.EMBED_GGUF }}", "--embedding", "--pooling", "{{ llamacpp_env.EMBED_POOLING }}",
-             "-c", "8192", "-b", "8192", "-ub", "8192", "-np", "1", "-ngl", "99", "--alias", "embed-default",
+             "-c", "8192", "-b", "4096", "-ub", "4096", "-np", "1", "-ngl", "99", "--alias", "embed-default",
              "--metrics", "--no-slots", "--no-webui", "--host", "0.0.0.0", "--port", "8081"]
     - name: llamacpp-rerank
       port: 8082

platform/proxmox/ansible/roles/gpu-llamacpp/tasks/prepare.yml

@@ -36,4 +36,3 @@
   loop: "{{ llamacpp_profiles | map('extract', llamacpp_models) | flatten }}"
   loop_control:
     label: "{{ item.dest }}"
-