{
  "date": "2026-05-07",
  "model": {
    "base": "/home/steve/models/qwen3.6-27b-q4_0-gguf/Qwen3.6-27B-Q4_0.gguf",
    "fused_beta_alpha": "/home/steve/models/qwen3.6-27b-q4_0-fused-ba-gguf/Qwen3.6-27B-Q4_0-fused-ba.gguf",
    "architecture": "qwen35",
    "quantization": "Q4_0",
    "fused_tensors_added": 48,
    "fused_layout": "flat: all beta rows followed by all alpha rows",
    "weights_changed": false,
    "extra_tensors_added": true
  },
  "hardware": {
    "gpu": "Intel Arc Pro B70 / BMG-G31",
    "gpu_count_used": 3,
    "gpu_count_installed": 4,
    "vram_gb_each": 32
  },
  "command_shape": {
    "devices": "SYCL2/SYCL1/SYCL3",
    "split_mode": "tensor",
    "tensor_split": "1/1/1",
    "ubatch": 128,
    "poll": 25,
    "kv_cache": "f16",
    "flash_attention": true,
    "safe_flags": [
      "GGML_SYCL_FORCE_MMVQ_Q4_1=0",
      "GGML_SYCL_DISABLE_DNN=1",
      "GGML_SYCL_Q8_CACHE=1",
      "GGML_SYCL_ASYNC_CPY_TENSOR=0",
      "GGML_SYCL_ASYNC_PEER_COPY=1",
      "GGML_SYCL_COMM_ALLREDUCE=1",
      "GGML_SYCL_COMM_SINGLE_KERNEL=1",
      "GGML_SYCL_COMM_EVENT_BARRIER=1",
      "GGML_SYCL_COMM_SYNC_AFTER=2",
      "GGML_SYCL_COMM_FUSEADD_ROOT_RESIDUAL=0",
      "GGML_META_FUSE_ALLREDUCE_ADD=1",
      "GGML_META_FUSE_ALLREDUCE_GET_ROWS=1",
      "GGML_SYCL_FUSE_MMVQ2=1",
      "GGML_SYCL_FUSE_MMVQ2_SWIGLU=1",
      "GGML_SYCL_FUSE_RMS_NORM_MUL=1"
    ]
  },
  "results": {
    "flat_layout_safe_no_root_p512_n512_r3": {
      "original_prompt_tok_s": 193.872714,
      "original_decode_tok_s": 49.52499,
      "original_total_tok_s": 78.89593093386667,
      "fused_prompt_tok_s": 200.480796,
      "fused_decode_tok_s": 50.1299,
      "fused_total_tok_s": 80.20473519933404,
      "decode_delta_tok_s": 0.60491,
      "decode_delta_percent": 1.22142,
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/tp3-refresh-20260507/fused-ba-model/fused-ba-flat-no-root-p512n512-r3-20260507T065943Z.tsv"
    },
    "flat_layout_safe_flags_short_p0_n128_r3": {
      "original_no_root_tok_s": 48.252739,
      "fused_no_root_tok_s": 49.369187,
      "original_no_meta_add_tok_s": 27.442101,
      "fused_no_meta_add_tok_s": 27.941093,
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/tp3-refresh-20260507/fused-ba-model/fused-ba-flat-safe-flags-p0n128-r3-20260507T065456Z.tsv"
    },
    "transient_host_wait_all_current": {
      "original_decode_tok_s": 28.974221,
      "fused_decode_tok_s": 29.247443,
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/tp3-refresh-20260507/fused-ba-model/fused-ba-flat-all-current-after-wait-p512n512-r3-20260507T071628Z.tsv",
      "decision": "Reverted. Host wait fixed correctness but destroyed throughput."
    },
    "final_rebuild_root_rerun_p512_n512_r3": {
      "status": "promising_not_promoted",
      "original_prompt_tok_s": 199.802378,
      "original_decode_tok_s": 50.372777,
      "original_total_tok_s": 80.45683822686443,
      "fused_prompt_tok_s": 200.046542,
      "fused_decode_tok_s": 50.687317,
      "fused_total_tok_s": 80.87871651470131,
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/tp3-refresh-20260507/fused-ba-model/root-rerun-20260507T081227Z",
      "decision": "Do not submit/promote yet because a two-token prompt follow-up timed out."
    }
  },
  "quality": {
    "status": "cleared_with_root_residual_disabled",
    "final_no_root_logit_probe": {
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fused-ba-no-root-final-20260507T073839Z",
      "same_token": true,
      "token": 321,
      "hash_match": true,
      "hash": "4f77fb8d90ed49e6"
    },
    "single_gpu_flat_probe": {
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fused-ba-logit-probe-flat-single-20260507T061507Z",
      "same_token": true,
      "token": 321,
      "hash_match": true,
      "hash": "681ad69c58fffd49"
    },
    "tp3_no_fusions_flat_probe": {
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fused-ba-logit-probe-flat-tp3-nofusions-20260507T061721Z",
      "same_token": true,
      "token": 321,
      "hash_match": true,
      "hash": "f97bd3230bbf35f5"
    },
    "unsafe_interaction": {
      "status": "root_residual_plus_meta_allreduce_add_not_cleared",
      "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fused-ba-meta-pair-isolation-20260507T064644Z",
      "root_meta_add": {
        "passed": false,
        "original_token": 220,
        "original_hash": "c57597537930eab9",
        "fused_token": 271,
        "fused_hash": "f8be73fee293172b"
      },
      "root_meta_rows": {
        "passed": true
      },
      "meta_add_meta_rows": {
        "passed": true
      }
    },
    "final_rebuild_root_rerun": {
      "default_prompt_root_off_vs_on_original": {
        "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/root-on-vs-off-final-nostats-20260507T081002Z",
        "same_token": true,
        "token": 271,
        "hash_match": true,
        "hash": "00e1eaeb550b0025"
      },
      "default_prompt_root_on_original_vs_fused": {
        "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/fused-ba-root-final-rerun-20260507T080518Z",
        "same_token": true,
        "token": 271,
        "hash_match": true,
        "hash": "00e1eaeb550b0025"
      },
      "two_token_prompt_followup": {
        "artifact": "/home/steve/bench-results/qwen36-q4_0-gguf/correctness/root-thecapital-final-rerun-20260507T082522Z",
        "prompt": "The capital",
        "n_prompt": 2,
        "status": "timeout",
        "timeout_seconds": 240
      }
    }
  },
  "localmaxxing": {
    "submitted": true,
    "id": "cmov6p4r7007tqr01yi8ug4un",
    "status": "APPROVED",
    "createdAt": "2026-05-07T07:48:07.891Z",
    "tokSOut": 50.1299,
    "tokSTotal": 80.20473519933404,
    "detailedPayloadHttpStatus": 500,
    "reducedPayloadHttpStatus": 201,
    "notes": "Submitted as Q4_0+fused-ssm_ba experimental augmented GGUF. Root-residual disabled because the root-residual plus meta allreduce-add interaction is not quality-cleared."
  },
  "decision": "Use as the current quality-cleared no-root Q4_0 TP3 experimental result. Do not use root-residual for quality claims until the ordering issue is fixed."
}