command:
      - "/usr/local/bin/llama-server"
      - "-m"
      - "/models/Qwen3.6-27B-MTP-Q6_K.gguf"

      # === CONTEXT & OFFLOAD ===
      - "-c"
      - "196608"    # was 196608 - 262144 for kvarn6
      - "-ngl"
      - "99"
      - "--cache-ram"
      - "32768"     # 32GB RAM budget for prompt cache / checkpoint spill
      - "--no-mmap"

      # === MTP SPECULATIVE DECODING ===
      - "--spec-type"
      - "draft-mtp"
      - "--spec-draft-n-max"
      - "10"       #was 6, 10 gives 10-20% more throughput
      - "--spec-draft-p-min"
      - "0.5"               # pmin=0.5 best across 2k/15k/64k depths at 6 drafts (sweep: 119/122/118 t/s vs 106-126 range for others)

      # === PERFORMANCE & BATCHING ===
      - "--flash-attn"
      - "on"
      - "-b"
      - "512"  #1024, 2048 optimal uses vram  512 needed for 192k context
      - "-ub"
      - "512"  #1024, 2048 optimal
      - "--parallel"
      - "1"

      # === CHECKPOINT TUNING ===
      - "--checkpoint-min-step"
      - "256"   #was 512.  tune for use case
      - "--ctx-checkpoints"
      - "8"

      # === KV CACHE ===
      - "-ctk"
      - "q8_0" # q8_0 - kvarn6
      - "-ctv"
      - "q8_0" # q8_0 - kvarn6
      - "--kv-unified"

      # === SAMPLING ===
      - "--temp"
      - "0.6"      #was 0.6
      - "--top-k"
      - "20"       #was 20
      - "--top-p"
      - "0.95"       # was 0.95
      - "--min-p"
      - "0.0"       # was 0.0
      - "--presence-penalty"
      - "0.0"
      - "--repeat-penalty"
      - "1.0"       #was 1.0
      - "--no-mmproj"

      # === THINKING MODE & CACHE CONTINUITY ===
      - "--jinja"
      # - "--chat-template-file"
      # - "/models/qwen3.6-chat-template.jinja"
      - "--reasoning"
      - "on"
      - "--reasoning-budget"
      - "16384"

      # === SERVER ===
      - "--perf"
      - "--metrics"
      - "--port"
      - "8080"
      - "--host"
      - "0.0.0.0"
      - "--alias"
      - "chat"