{
  "schema_version": "1.0",
  "as_of": "2026-10-12",
  "href": "https://amdai.net/stacks",
  "default_backend": "llama.cpp",
  "test_rig": {
    "label": "Our test rig",
    "gpus": "Two AMD Radeon AI PRO R9700 (32 GB GDDR6, gfx1201)",
    "ram": "256 GB DDR5",
    "note": "Daily driver uses ONE card. The second card is there; do not assume a split."
  },
  "stacks": [
    {
      "id": "daily-driver",
      "title": "Daily driver",
      "kicker": "One R9700 · llama.cpp Vulkan",
      "model": "Qwen3.6-35B-A3B",
      "official": "Qwen/Qwen3.6-35B-A3B",
      "gguf": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "quant": "UD-Q5_K_XL",
      "engine": "llama.cpp",
      "backend_id": "llamacpp",
      "gpus": "1 × Radeon AI PRO R9700 (32 GB, gfx1201)",
      "context": 262144,
      "kv": "f16",
      "offload": "0% system RAM",
      "why": "Full native context on one 32 GB card, Q5-family quality, no RAM crawl.",
      "measured": [
        "Publisher GGUF listing 26.6 GB (Hugging Face) — file size, not VRAM",
        "Context 262144 and f16 KV confirmed on our rig",
        "UD-Q5_K_XL decode: 102.39 tok/s at short context, 90.52 tok/s at 32K, 66.43 tok/s at 128K, 44.4 tok/s at 256K. Measured on our rig, 2026-10. One R9700, llama.cpp, f16 KV.",
        "Q5_K_M sibling: 109.73 tok/s, 24.68 GB VRAM. Measured on our rig, 2026-10. One R9700, llama.cpp, f16 KV.",
        "Q4_K_M sibling: 117.69 tok/s decode; 1902.1 tok/s prefill at 32K. Measured on our rig, 2026-10. One R9700, llama.cpp, f16 KV.",
        "UD-Q5_K_XL peak VRAM 25.97 GB at 4K / 29.34 GB at 256K from official llama-bench c060ca974."
      ],
      "calculator": "https://amdai.net/calculator?gpu=r9700&vram=32&n=1&split=0&ram=256&model=qwen36&tp=35&ap=3&layers=40&kvl=10&kvh=2&hd=256&quant=q5_k_m&ctx=262144&kv=f16&oh=1",
      "href": "https://amdai.net/stacks",
      "launch": {
        "family": "llamacpp",
        "todo_flags": true,
        "env": [],
        "lines": [
          "# Shape only — exact flags TODO on stacks.",
          "./build-vk/bin/llama-server \\",
          "  -m ~/models/Qwen3.6-35B-A3B-Q5_K_M.gguf \\",
          "  -ngl 99 \\",
          "  -c 4096 \\",
          "  --cache-type-k f16 \\",
          "  --cache-type-v f16 \\",
          "  --host 127.0.0.1 --port 8080"
        ]
      }
    },
    {
      "id": "fast-agent",
      "title": "Fast agent",
      "kicker": "Strata · Qwen 3.8 Flash",
      "model": "Qwen 3.8 Flash",
      "official": "Qwen/Qwen3.8-Flash-Next",
      "gguf": "unsloth/Qwen3.8-Flash-Next-GGUF",
      "quant": "UD-Q6_K_XL",
      "engine": "llama.cpp",
      "backend_id": "llamacpp",
      "gpus": "Our test rig (R9700 class)",
      "context": 262144,
      "kv": "f16 (llama.cpp fallback; Strata flags not recorded)",
      "offload": "Publisher file is ~169 GB — expect RAM to hold most of it on one 32 GB card",
      "why": "Faster agent loop when you already have the daily driver proven.",
      "measured": [
        "~32.6 tok/s on Strata v0.1.42 vs ~16 tok/s on llama.cpp — Measured on our rig, 2026-10",
        "Publisher UD-Q6_K_XL folder ~169 GB (Hugging Face, six shards)",
        "not recorded: exact command line / measured VRAM+RAM from rig owner."
      ],
      "calculator": "https://amdai.net/calculator?gpu=r9700&vram=32&n=1&split=0&ram=256&model=flash&tp=125&ap=6&layers=48&kvl=12&kvh=2&hd=256&quant=q6_k&ctx=262144&kv=f16&oh=1",
      "href": "https://amdai.net/stacks",
      "launch": {
        "family": "llamacpp",
        "todo_flags": true,
        "env": [],
        "lines": [
          "# Shape only — llama.cpp Q6, 8 threads on the same Flash file. Exact flags TODO.",
          "./build-vk/bin/llama-server \\",
          "  -m ~/models/Qwen3.8-Flash-Next-UD-Q6_K_XL.gguf \\",
          "  -t 8 \\",
          "  -c 262144 \\",
          "  --host 127.0.0.1 --port 8080"
        ]
      }
    },
    {
      "id": "vllm-try",
      "title": "vLLM serve try",
      "kicker": "Not the daily driver",
      "model": "Same official Qwen you already trust on llama.cpp",
      "official": "Pin whatever the weekly card names",
      "gguf": "HF weights, not GGUF",
      "quant": "Framework default — Q3 listed as estimate/community; Q1/Q2 stay off",
      "engine": "Community / ROCm vLLM 0.31 try (0.30 previous; AMD 0.27 on ROCm 10.0 is the conservative pin)",
      "backend_id": "vllm",
      "gpus": "One GPU first",
      "context": "Confirm the engine, do not assume 262K",
      "kv": "Engine default",
      "offload": "Do not treat vLLM as a RAM-offload strategy",
      "why": "Continuous batching when several clients hit /v1. Consumers still start on llama.cpp.",
      "measured": [
        "AMD-supported combo on the 10.0 matrix remains vLLM 0.27 + PyTorch 2.13 (gfx1201 listed)",
        "not recorded: exact image tag / measured VRAM from rig owner if 0.31 is pinned."
      ],
      "calculator": "https://amdai.net/calculator",
      "href": "https://amdai.net/stacks",
      "launch": null
    }
  ]
}