fix(llm-inference): switch serve phase from vLLM+bitsandbytes to llama.cpp+GGUF

bitsandbytes peak RAM ~54GB (bf16 load before quantize) — kills 40GB OptiPlex.
llama.cpp Q4_K_M GGUF loads pre-quantized: peak RAM ~15.5GB, fits cleanly.

Changes:
- serve.yml: build llama.cpp with CUDA, download Q4_K_M GGUF from bartowski,
  disable vllm-serve, deploy llama-server.service
- llama-server.service.j2: OpenAI-compatible server on same port 8000,
  --n-gpu-layers 99 (full GPU offload), --parallel 4, gemma chat template
- defaults: llm_gguf_dir, llm_gguf_path, llm_gpu_layers, llm_parallel_slots
- handlers: restart llama-server, vllm-serve failed_when=false (may not exist)

GGUF: bartowski/gemma-2-27b-it-Q4_K_M.gguf (15.5GB, 24GB VRAM fits w/ ~8GB headroom)

ansible/roles/llm-inference/defaults/main.yml

@@ -17,12 +17,19 @@ llm_venv_owner: jarvis
 llm_hf_cache_dir: /home/jarvis/.cache/huggingface
 llm_hf_model: google/gemma-2-27b-it
 
-# vLLM serve
+# vLLM serve (deprecated — replaced by llama-server)
+# llama-server serve
 llm_serve_port: 8000
 llm_serve_host: "0.0.0.0"
-llm_quantization: "bitsandbytes"   # int4 — fits 27B in 24GB VRAM (~14GB vs ~54GB bf16)
-llm_gpu_memory_utilization: "0.92" # higher utilization to give KV cache room
-llm_max_model_len: 4096            # safe ceiling given 15.96GB model + int4
+llm_max_model_len: 8192
+llm_gpu_layers: 99              # offload all layers to GPU
+llm_parallel_slots: 4           # concurrent request slots
+llm_gguf_dir: /home/jarvis/models
+llm_gguf_path: /home/jarvis/models/gemma-2-27b-it-Q4_K_M.gguf
+
+# Legacy vLLM vars (kept for role documentation, not used by llama-server)
+llm_quantization: "bitsandbytes"
+llm_gpu_memory_utilization: "0.92"
 
 # Monitoring
 llm_gpu_exporter_version: "1.13.1"