Update ISO installer tiers and backend deployment

ansible/roles/vllm/defaults/main.yml

@@ -0,0 +1,7 @@
+---
+# vLLM role defaults — overridden per-tier in the tier playbook vars block
+vllm_model: "meta-llama/Meta-Llama-3.1-8B-Instruct"
+vllm_tensor_parallel: 1
+vllm_gpu_memory_util: "0.70"
+vllm_max_model_len: 8192
+vllm_quantization: ""              # blank = full precision; set to "awq" for 4-bit

ansible/roles/vllm/tasks/main.yml

@@ -1,6 +1,12 @@
 ---
 # vLLM — high-performance LLM inference with OpenAI-compatible API
 # Skipped automatically if no GPU is present.
+# Variables (set defaults in defaults/main.yml, override per-tier in the playbook):
+#   vllm_model              HuggingFace model ID to load on start
+#   vllm_tensor_parallel    Number of GPUs for tensor parallelism (1 for Starter/Basic)
+#   vllm_gpu_memory_util    Fraction of VRAM to reserve for vLLM (0.0–1.0)
+#   vllm_max_model_len      Maximum context length in tokens
+#   vllm_quantization       Quantization method: "" (none) | "awq" | "gptq" | "fp8"
 
 - name: Check for NVIDIA GPU
   shell: nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1
@@ -10,7 +16,9 @@
 
 - name: Skip vLLM if no GPU detected
   debug:
-    msg: "No GPU detected — skipping vLLM image pull. Run manually when GPU is available: docker pull vllm/vllm-openai:latest"
+    msg: >
+      No GPU detected — skipping vLLM image pull.
+      Run manually when GPU is available: docker pull vllm/vllm-openai:latest
   when: gpu_check.stdout == "" or gpu_check.rc != 0
 
 - name: Pull vLLM Docker image
@@ -19,18 +27,23 @@
   delay: 15
   when: gpu_check.stdout != "" and gpu_check.rc == 0
 
+- name: Build vLLM quantization flag
+  set_fact:
+    vllm_quant_flag: "{{ '--quantization ' + vllm_quantization if vllm_quantization != '' else '' }}"
+
 - name: Create vLLM systemd service
   copy:
     dest: /etc/systemd/system/vllm.service
+    mode: "0644"
     content: |
       [Unit]
-      Description=vLLM OpenAI-Compatible Inference Server
-      After=docker.service ollama.service
+      Description=vLLM OpenAI-Compatible Inference Server ({{ vllm_model }})
+      After=docker.service
       Requires=docker.service
 
       [Service]
       Restart=always
-      RestartSec=5
+      RestartSec=10
       ExecStartPre=-/usr/bin/docker stop vllm
       ExecStartPre=-/usr/bin/docker rm vllm
       ExecStart=/usr/bin/docker run \
@@ -41,15 +54,16 @@
         -v /opt/cezen/models:/root/.cache/huggingface \
         -e HF_HOME=/root/.cache/huggingface \
         vllm/vllm-openai:latest \
-        --model meta-llama/Meta-Llama-3.1-8B-Instruct \
-        --gpu-memory-utilization 0.7 \
-        --max-model-len 8192 \
-        --tensor-parallel-size 1
+        --model {{ vllm_model }} \
+        --gpu-memory-utilization {{ vllm_gpu_memory_util }} \
+        --max-model-len {{ vllm_max_model_len }} \
+        --tensor-parallel-size {{ vllm_tensor_parallel }} \
+        {{ vllm_quant_flag }}
       ExecStop=/usr/bin/docker stop vllm
+      TimeoutStartSec=300
 
       [Install]
       WantedBy=multi-user.target
-    mode: "0644"
 
 - name: Create vLLM model directory
   file:
@@ -57,3 +71,26 @@
     state: directory
     owner: cezen
     group: cezen
+    mode: "0755"
+
+- name: Write vLLM tier config file (for portal reference)
+  copy:
+    dest: /opt/cezen/vllm-config.json
+    owner: cezen
+    group: cezen
+    mode: "0644"
+    content: |
+      {
+        "model": "{{ vllm_model }}",
+        "tensor_parallel_size": {{ vllm_tensor_parallel }},
+        "gpu_memory_utilization": {{ vllm_gpu_memory_util }},
+        "max_model_len": {{ vllm_max_model_len }},
+        "quantization": "{{ vllm_quantization }}"
+      }
+
+- name: Enable and start vLLM service
+  systemd:
+    name: vllm
+    enabled: true
+    daemon_reload: true
+  when: gpu_check.stdout != "" and gpu_check.rc == 0