{
  "date": "2026-09-16",
  "hardware": {
    "system": "GMKtec EVO-X3",
    "cpu": "AMD Ryzen AI Max+ 395",
    "gpu": "Radeon 8060S / gfx1151",
    "installed_ram_gb": 128,
    "linux_visible_ram_gib_approx": 124
  },
  "os": "Ubuntu 24.04.4; kernel 6.17.0-40-generic",
  "runtime": {
    "rocm": "7.14.0",
    "torch": "2.11.0+rocm7.14.0",
    "triton": "3.7.1+git0263a6a6.rocm7.14.0",
    "transformers": "5.14.1",
    "compressed_tensors": "0.17.0",
    "safetensors": "0.8.0"
  },
  "checkpoint": {
    "repository": "wtdcode/Qwen3.8-Flash-Next-AWQ-W4A16",
    "revision": "0939125b929543a783ce700c90e36dd1a575c00c",
    "text_backbone_gib": 67.225,
    "ple_bf16_gib": 95.368
  },
  "vllm": {
    "base_commit": "d4d0f73ef171154eac6f1914dca47001d662cfbb",
    "local_changes": [
      "HIP discovery fallback",
      "disk-backed BF16 PLE lookup and loader adaptation",
      "AMD QSA optional FlashAttention dependency fixes"
    ],
    "reused_native_extensions": true,
    "max_model_len": 33280,
    "max_num_seqs": 2,
    "max_num_batched_tokens": 256,
    "kv_cache_gib": 4,
    "kv_dtype": "BF16",
    "eager": true,
    "tensor_parallel_size": 1
  },
  "baseline": {
    "engine": "llama.cpp Vulkan v0.7.1",
    "checkpoint": "Qwen3.8-Flash-Next-UD-Q4_K_XL",
    "checkpoint_gib": 103.688,
    "gpu_layers": 40,
    "context_per_slot": 33280,
    "slots": 2,
    "kv_dtype": "Q8",
    "batch": 512,
    "microbatch": 128
  },
  "protocol": {
    "input_tokens": [
      512,
      8192,
      32768
    ],
    "clients": [
      1,
      2
    ],
    "waves_per_case": 3,
    "reported_waves": [
      1,
      2
    ],
    "max_output_tokens": 128,
    "temperature": 0,
    "identical_input_token_ids": true,
    "prefix_reuse": false,
    "speculation": false,
    "npu_unloaded": true
  },
  "validation": {
    "retrieval_correct_per_engine": "27/27",
    "http_errors_per_engine": 0,
    "amd_qsa_tests_passed": 10,
    "production_configuration_hashes_unchanged": 4,
    "production_restored": true
  },
  "limitations": [
    "Different quantisation, KV and placement: deployment comparison, not engine-only.",
    "Two reported repeats, not confidence intervals.",
    "Repeated filler does not qualify cold diverse-text disk access.",
    "Text only; vision and MTP disabled.",
    "No simultaneous vLLM/NPU qualification.",
    "Development source with local patches; no clean stock-wheel build validation."
  ]
}