{
  "date": "2026-06-07",
  "host": "4x Intel Arc Pro B70",
  "model": "Intel/gemma-4-12B-it-int4-AutoRound",
  "engine": "vLLM/XPU",
  "endpoint": "OpenAI-compatible no-auth LAN endpoint on 0.0.0.0:8000",
  "production_after_test": {
    "profile": "gemma4-12b-it-int4-autoround-c8",
    "max_model_len": 32768,
    "max_active_generations": 8,
    "prefix_caching": true,
    "xpu_graph": true,
    "quality_canary": "pass"
  },
  "raw_result_dir": "/mnt/fast-ai/bench-results/gemma4-12b-it-int4-autoround/concurrency32k-20260607T190702Z",
  "harness_changes": {
    "script": "scripts/bench-openai-concurrency.py",
    "added_options": [
      "--shared-prefix-tokens",
      "--prompt-salt"
    ],
    "purpose": "Measure fixed-prefix plus unique-tail workloads without accidentally reusing full cached prompts."
  },
  "c12": {
    "profile": "configs/model-slots/gemma4-12b-it-int4-autoround-c12.env",
    "status": "rejected-device-lost",
    "max_model_len": 32768,
    "max_num_seqs": 12,
    "startup": {
      "gpu_kv_cache_tokens": 991437,
      "theoretical_full_32768_concurrency": 30.26,
      "ocloc_error_code": 245,
      "igc_internal_compiler_error": "Floating point exception",
      "compile_range_1_1_s": 79.24,
      "compile_range_1_4096_s": 99.77
    },
    "failure": [
      "UR_RESULT_ERROR_OUT_OF_RESOURCES",
      "UR_RESULT_ERROR_DEVICE_LOST"
    ],
    "interpretation": "The KV pool estimate alone is not enough to declare higher full-context active concurrency production-safe."
  },
  "c10": {
    "profile": "configs/model-slots/gemma4-12b-it-int4-autoround-c10.env",
    "status": "research-only",
    "max_model_len": 32768,
    "max_num_seqs": 10,
    "startup": {
      "gpu_kv_cache_tokens": 991437,
      "theoretical_full_32768_concurrency": 30.26,
      "compile_range_1_1_s": 83.14,
      "compile_range_1_4096_s": 100.85
    },
    "quality_canary": {
      "expected_pass_all": true,
      "baseline_match_all": true,
      "cases": [
        "exact_ok",
        "copy_phrase",
        "small_arithmetic",
        "red_image"
      ]
    },
    "short_prompt_100p_512o": {
      "c8_load": {
        "actual_prompt_tokens_each": 119,
        "output_tokens_each": 512,
        "mean_ttft_s": 0.140,
        "wall_aggregate_output_tok_s": 754.88
      },
      "c10_load": {
        "actual_prompt_tokens_each": 120,
        "output_tokens_each": 512,
        "mean_ttft_s": 0.221,
        "wall_aggregate_output_tok_s": 849.59
      }
    },
    "unique_near_32k_1o": {
      "c8_load": {
        "actual_prompt_tokens_each": 30690,
        "output_tokens_each": 1,
        "mean_ttft_s": 22.20,
        "wall_aggregate_output_tok_s": 0.205
      },
      "c10_load": {
        "actual_prompt_tokens_each": 30744,
        "output_tokens_each": 1,
        "mean_ttft_s": 27.21,
        "wall_aggregate_output_tok_s": 0.204
      }
    },
    "unique_near_32k_128o": {
      "c8_load": {
        "actual_prompt_tokens_each": 30690,
        "output_tokens_each": 128,
        "mean_ttft_s": 23.61,
        "wall_aggregate_output_tok_s": 22.73
      },
      "c10_load": {
        "actual_prompt_tokens_each": 30744,
        "output_tokens_each": 128,
        "mean_ttft_s": 29.25,
        "wall_aggregate_output_tok_s": 22.76
      }
    },
    "half_shared_prefix_near_32k_1o": {
      "shared_prefix_tokens_requested": 16000,
      "prompt_salt": "unique-a",
      "c8_load": {
        "actual_prompt_tokens_each": 31021,
        "output_tokens_each": 1,
        "mean_ttft_s": 12.45,
        "wall_aggregate_output_tok_s": 0.371
      },
      "c10_load": {
        "actual_prompt_tokens_each": 31034,
        "output_tokens_each": 1,
        "mean_ttft_s": 15.11,
        "wall_aggregate_output_tok_s": 0.372
      }
    },
    "half_shared_prefix_near_32k_128o": {
      "shared_prefix_tokens_requested": 16000,
      "prompt_salt": "unique-b",
      "c8_load": {
        "actual_prompt_tokens_each": 31021,
        "output_tokens_each": 128,
        "mean_ttft_s": 13.24,
        "wall_aggregate_output_tok_s": 39.20
      },
      "c10_load": {
        "actual_prompt_tokens_each": 31034,
        "output_tokens_each": 128,
        "mean_ttft_s": 16.04,
        "wall_aggregate_output_tok_s": 39.67
      }
    }
  },
  "recommendation": "Keep production on c8 for 32K request support. c10 is useful research for short-prompt aggregate throughput but does not improve near-32K throughput. c12 is rejected until the Level Zero/vLLM/XPU device-lost boundary is understood."
}