{
  "schemaVersion": 3,
  "retrievedAt": "2026-09-12",
  "fullPeakCalibrationCount": 0,
  "componentReferences": [
    {
      "id": "qwen35-08b-attention",
      "model": "Qwen3.5-0.8B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/20072",
      "runtimeBuild": "8191 / 24350fdf9",
      "hardware": "NVIDIA L40",
      "component": "attention",
      "contextPerSequence": 4032,
      "sequences": 4,
      "kvType": "f16",
      "runtime": {
        "backend": "cuda",
        "logicalBatch": 2048,
        "microBatch": 512,
        "recurrentSnapshots": 0
      },
      "observedMiB": 189,
      "config": {
        "model_type": "qwen3_5_text",
        "num_hidden_layers": 24,
        "hidden_size": 1024,
        "num_attention_heads": 8,
        "num_key_value_heads": 2,
        "head_dim": 256,
        "full_attention_interval": 4,
        "linear_conv_kernel_dim": 4,
        "linear_key_head_dim": 128,
        "linear_value_head_dim": 128,
        "linear_num_key_heads": 16,
        "linear_num_value_heads": 16
      },
      "eligibleForCalibration": false,
      "notes": "The normalized 4032 x 4 context matches 16128 allocated unified cells, not the reported per-slot limit of 16128. A multimodal projector is loaded, other device memory consumers are present, and native fitting fails its free-memory target. Only attention/recurrent allocation arithmetic is compared; full-device peak is excluded."
    },
    {
      "id": "qwen35-08b-recurrent",
      "model": "Qwen3.5-0.8B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/20072",
      "runtimeBuild": "8191 / 24350fdf9",
      "hardware": "NVIDIA L40",
      "component": "recurrent",
      "contextPerSequence": 4032,
      "sequences": 4,
      "kvType": "f16",
      "runtime": {
        "backend": "cuda",
        "logicalBatch": 2048,
        "microBatch": 512,
        "recurrentSnapshots": 0
      },
      "observedMiB": 77.06,
      "config": {
        "model_type": "qwen3_5_text",
        "num_hidden_layers": 24,
        "hidden_size": 1024,
        "num_attention_heads": 8,
        "num_key_value_heads": 2,
        "head_dim": 256,
        "full_attention_interval": 4,
        "linear_conv_kernel_dim": 4,
        "linear_key_head_dim": 128,
        "linear_value_head_dim": 128,
        "linear_num_key_heads": 16,
        "linear_num_value_heads": 16
      },
      "eligibleForCalibration": false,
      "notes": "The normalized 4032 x 4 context matches 16128 allocated unified cells, not the reported per-slot limit of 16128. A multimodal projector is loaded, other device memory consumers are present, and native fitting fails its free-memory target. Only attention/recurrent allocation arithmetic is compared; full-device peak is excluded."
    },
    {
      "id": "llama31-8b-kv",
      "model": "Llama 3.1 8B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/10664",
      "runtimeBuild": "4265 / 59f4db10",
      "hardware": "A100 80GB",
      "measurement": "allocated cache buffer",
      "component": "attention",
      "contextPerSequence": 8192,
      "sequences": 1,
      "observedMiB": 1024,
      "config": {
        "model_type": "llama",
        "num_hidden_layers": 32,
        "hidden_size": 4096,
        "num_attention_heads": 32,
        "num_key_value_heads": 8
      },
      "notes": "Main-model cache only. The process also loaded a draft model and another server; no whole-process peak inference."
    },
    {
      "id": "llama32-1b-kv",
      "model": "Llama 3.2 1B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/10664",
      "runtimeBuild": "4265 / 59f4db10",
      "hardware": "A100 80GB",
      "measurement": "allocated cache buffer",
      "component": "attention",
      "contextPerSequence": 8192,
      "sequences": 1,
      "observedMiB": 256,
      "config": {
        "model_type": "llama",
        "num_hidden_layers": 16,
        "hidden_size": 2048,
        "num_attention_heads": 32,
        "num_key_value_heads": 8
      },
      "notes": "Draft-model cache only, not total device use."
    },
    {
      "id": "qwen35-27b-attention",
      "model": "Qwen3.5 27B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/19910",
      "runtimeBuild": "8148 / 244641955",
      "hardware": "Intel Arc B580 + CPU",
      "measurement": "allocated cache across CPU and GPU",
      "component": "attention",
      "contextPerSequence": 1024,
      "sequences": 4,
      "observedMiB": 256,
      "config": {
        "model_type": "qwen3_5_text",
        "num_hidden_layers": 64,
        "hidden_size": 5120,
        "num_attention_heads": 24,
        "num_key_value_heads": 4,
        "head_dim": 256,
        "full_attention_interval": 4,
        "linear_conv_kernel_dim": 4,
        "linear_key_head_dim": 128,
        "linear_value_head_dim": 128,
        "linear_num_key_heads": 16,
        "linear_num_value_heads": 48
      },
      "notes": "Partial offload, Flash Attention disabled, later crash. Context input normalizes the shared 4096-cell pool to 1024 x 4; it does not reproduce the reported 4096-token limit of each slot. CPU + GPU cache arithmetic only, not execution or peak equivalence."
    },
    {
      "id": "qwen35-27b-recurrent",
      "model": "Qwen3.5 27B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/19910",
      "runtimeBuild": "8148 / 244641955",
      "hardware": "Intel Arc B580 + CPU",
      "measurement": "allocated cache across CPU and GPU",
      "component": "recurrent",
      "contextPerSequence": 1024,
      "sequences": 4,
      "observedMiB": 598.5,
      "config": {
        "model_type": "qwen3_5_text",
        "num_hidden_layers": 64,
        "hidden_size": 5120,
        "num_attention_heads": 24,
        "num_key_value_heads": 4,
        "head_dim": 256,
        "full_attention_interval": 4,
        "linear_conv_kernel_dim": 4,
        "linear_key_head_dim": 128,
        "linear_value_head_dim": 128,
        "linear_num_key_heads": 16,
        "linear_num_value_heads": 48
      },
      "notes": "Partial offload, Flash Attention disabled, later crash. Context input normalizes the shared 4096-cell pool to 1024 x 4; it does not reproduce the reported 4096-token limit of each slot. CPU + GPU cache arithmetic only, not execution or peak equivalence."
    }
  ],
  "cacheMechanismReferences": [
    {
      "id": "deepseek-v4-flash-0731-cache",
      "model": "DeepSeek-V4-Flash-0731",
      "source": "https://github.com/ggml-org/llama.cpp/issues/27872",
      "runtimeBuild": "10621 / c1d0e7a",
      "hardware": "RTX 3090 + CPU experts",
      "contextPerSequence": 524288,
      "runtime": {"microBatch":8192, "logicalBatch":8192, "recurrentSnapshots":0},
      "config": {
        "model_type":"deepseek_v4", "num_hidden_layers":43, "hidden_size":4096,
        "num_attention_heads":64, "num_key_value_heads":1, "head_dim":512,
        "index_head_dim":128, "index_n_heads":64, "index_topk":512, "sliding_window":128,
        "compress_ratios":[0,0,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4,128,4]
      },
      "observedMiB": {"raw":354.75, "compressed4":2688, "compressed128":80, "indexer":672,
        "compressor4":1.31, "compressor128":10, "indexerState":0.33},
      "eligibleForCalibration":false,
      "notes":"F16 K-only caches; one sequence and zero rollback planes. CPU expert offload and later compute-buffer OOM exclude total/peak calibration. The failed compute request was 14656.12 MiB; it is not a measured device peak or a universal overhead."
    }
  ],
  "allocationRuns": [
    {
      "id": "gemma3-12b-historical",
      "model": "Gemma 3 12B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/12352",
      "runtimeBuild": "Reported b4876; binary prints build 0 / unknown",
      "hardware": "RTX 3060 12GB",
      "artifact": "gemma-3-12b-it-Q5_K_M.gguf",
      "artifactRevision": null,
      "params": 11.77,
      "paramsAreRounded": true,
      "scenarioBpw": 5.69,
      "contextPerSequence": 4096,
      "sequences": 1,
      "kvType": "f16",
      "runtime": { "microBatch": 512, "runtimeBudgetGiB": 1 },
      "config": {
        "model_type": "gemma3_text", "num_hidden_layers": 48, "hidden_size": 3840,
        "num_attention_heads": 16, "num_key_value_heads": 8, "head_dim": 256,
        "sliding_window": 1024, "sliding_window_pattern": 6
      },
      "observedGpuMiB": { "weights": 8047.63, "attention": 1536, "recurrent": 0, "compute": 519.62 },
      "comparableComponents": [],
      "comparePlanningTotal": true,
      "eligibleForCalibration": false,
      "exclusionReasons": ["Historical full-size sliding cache differs from the current profile", "Unpinned artifact and rounded text parameter count", "Logged buffer sum is not measured device peak"],
      "notes": "Requested context 4000 was allocated as 4096. Successful generation is logged. Current scenario uses windowed cache, not this historical full allocation. CPU/host buffers excluded. Default Q5 scenario BPW is independent of the observed weights."
    },
    {
      "id": "qwen3-30b-a3b-shared-cache",
      "model": "Qwen3-30B-A3B-Instruct-2507",
      "source": "https://github.com/ggml-org/llama.cpp/issues/20063",
      "runtimeBuild": "8191 / 24350fdf9",
      "hardware": "RTX 3090 + CPU",
      "artifact": "Qwen3-30B-A3B-Instruct-2507-Q4_K_M.gguf",
      "artifactRevision": null,
      "params": 30.53,
      "paramsAreRounded": true,
      "scenarioBpw": 4.85,
      "contextPerSequence": 20000,
      "sequences": 3,
      "kvType": "f16",
      "runtime": { "microBatch": 512, "runtimeBudgetGiB": 1 },
      "config": {
        "model_type": "qwen3_moe", "num_hidden_layers": 48, "hidden_size": 2048,
        "num_attention_heads": 32, "num_key_value_heads": 4, "head_dim": 128, "num_experts": 128
      },
      "observedGpuMiB": { "weights": 16510.93, "attention": 5640, "recurrent": 0, "compute": 358.15 },
      "comparableComponents": ["attention"],
      "comparePlanningTotal": false,
      "eligibleForCalibration": false,
      "exclusionReasons": ["Automatic fitting overflows expert weights to CPU despite the offloaded-layer count", "Shared context pool is not an independent full context per slot", "No measured device peak or artifact revision"],
      "notes": "60000 requested unified cells pad to 60160. The 20000 x 3 input matches allocated cells only, not the source's 60160-token per-slot maximum. Never multiply that logged maximum by three or treat 49/49 offloaded layers as proof of full weight residency."
    },
    {
      "id": "qwen35-35b-a3b-hybrid",
      "model": "Qwen3.5-35B-A3B",
      "source": "https://github.com/ggml-org/llama.cpp/issues/20093",
      "runtimeBuild": "1 / ecd99d6",
      "hardware": "RTX 5090",
      "artifact": "Qwen3.5-35B-A3B-UD-Q3_K_M.gguf",
      "artifactRevision": null,
      "params": 34.66,
      "paramsAreRounded": true,
      "scenarioBpw": 3.9,
      "contextPerSequence": 8192,
      "sequences": 1,
      "kvType": "f16",
      "runtime": { "microBatch": 512, "runtimeBudgetGiB": 1 },
      "config": {
        "model_type": "qwen3_5_moe_text", "num_hidden_layers": 40, "hidden_size": 2048,
        "num_attention_heads": 16, "num_key_value_heads": 2, "head_dim": 256, "num_experts": 256,
        "full_attention_interval": 4, "linear_conv_kernel_dim": 4,
        "linear_key_head_dim": 128, "linear_value_head_dim": 128,
        "linear_num_key_heads": 16, "linear_num_value_heads": 32
      },
      "observedGpuMiB": { "weights": 15572.55, "attention": 160, "recurrent": 62.81, "compute": 493 },
      "comparableComponents": ["attention", "recurrent"],
      "comparePlanningTotal": true,
      "eligibleForCalibration": false,
      "exclusionReasons": ["Dynamic mixed quantization differs from the standard Q3 conversion scenario", "Unpinned artifact and rounded parameter count", "Later cache-restore crash; buffer sum is not a measured peak"],
      "notes": "FA enabled, 10 attention layers and 30 recurrent layers. GGUF SSM fields map to 16 key heads, 32 value heads, 128 dimensions and convolution width 4. Recurrent output is rounded to two decimals. Q3 scenario BPW is not inferred from this dynamic artifact's filename or file type."
    }
  ],
  "crossRuntimeReferences": [
    {
      "model": "Qwen2.5-1.5B-Instruct",
      "source": "https://qwen.readthedocs.io/en/v2.5/benchmark/speed_benchmark.html",
      "runtime": "Transformers 4.46.0",
      "hardware": "A100 80GB",
      "cuda": "12.1",
      "torch": "2.3.1",
      "flashAttention": "2.5.8",
      "precision": "BF16",
      "sequences": 1,
      "generatedTokens": 2048,
      "reportedUnit": "GB (binary/decimal convention not specified on page)",
      "params": 1.543714304,
      "config": {
        "model_type": "qwen2",
        "num_hidden_layers": 28,
        "hidden_size": 1536,
        "num_attention_heads": 12,
        "num_key_value_heads": 2
      },
      "rows": [
        {
          "inputTokens": 1,
          "observed": 2.95
        },
        {
          "inputTokens": 6144,
          "observed": 3.43
        },
        {
          "inputTokens": 14336,
          "observed": 4.16
        },
        {
          "inputTokens": 30720,
          "observed": 5.62
        }
      ],
      "use": "Contextual cross-check only. Different runtime and unpinned historical artifact; do not calibrate llama.cpp or impose a peak accuracy threshold from these rows."
    }
  ],
  "liveArtifactReference": {
    "repo": "bartowski/Qwen2.5-0.5B-Instruct-GGUF",
    "revision": "41ba88dbac95fed2528c92514c131d73eb5a174b",
    "file": "Qwen2.5-0.5B-Instruct-Q4_K_M.gguf",
    "parameterCount": 494032768,
    "tensorBytes": 391859712,
    "fileBytes": 397808192,
    "headerPrefixBytesRead": 8000000,
    "use": "Real metadata integration, not an inference benchmark."
  }
}