{
  "artifacts": [
    {
      "file": "gliner25_multi_s128_wfp16.tflite",
      "sha256": "a4603fbcead732e6d50a0716ce80a79fbecd9e3763c07e7d2463f827a1c8b87d",
      "size_mb": 182.988
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-10-02",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": null,
        "latency_p50_ms": 1308.485,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": true
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-10-02",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": true,
        "latency_p50_ms": 21.68,
        "loads": true,
        "max_rel_diff": 1.5858585858585859,
        "output_match": false,
        "provenance": "measured",
        "runs": true
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-10-02/gliner2.5-multi__gliner25_multi_s128_wfp16.json"
  },
  "conversion": {
    "command": "export_round3.py --seq 128 (unified dense-prefix recipe with --mask-neg -1e4 --ln-shift 3) then quantize_shipment.py --seq 128 (ai-edge-quantizer 0.8.0 FLOAT_CASTING, FULLY_CONNECTED weights only); REPRODUCE.md, graph contract graph_contract_s128.json",
    "quantization": "float16 weight storage (ai-edge-quantizer 0.8.0 FLOAT_CASTING, weight-only 16-bit): the 96 FULLY_CONNECTED weight tensors stored as float16 with a DEQUANTIZE to float32; inputs, outputs, activations and every other constant stay float32 (card 'Files and supported configuration').",
    "tool": "litert-torch",
    "tool_version": "0.9.3 (ai-edge-litert 2.1.6; torch 2.12.1; transformers 4.57.6; gliner2 2.0.0; ai-edge-quantizer 0.8.0 for the float16 weight storage)"
  },
  "cross_runtime": [],
  "delegation": {
    "backend": "gpu_mldrift",
    "blocking_ops": [
      "DEQUANTIZE",
      "SQUARED_DIFFERENCE",
      "LOGISTIC"
    ],
    "coverage_ops_pct": 89.0,
    "lint_report_version": "1.1",
    "litert_version": "2.2.0",
    "matched_provenance_counts": {
      "measured": 1118,
      "unmatched": 30
    },
    "partitions": 127
  },
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu_xnnpack",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1147,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliner25multi-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.gliner25multi.gate (android/gate/, Kotlin CompiledModel 2.2.0, Environment.create(context, BuiltinNpuAcceleratorProvider(Qualcomm)) with the Qualcomm dispatch/compiler libraries in jniLibs), one process per (graph, accelerator), inputs pushed as raw float32 files built on the Mac from the float16 host table (rows upcast to float32); timed interval = the five input writes + run + packed-output readback per input, 2 warm-ups + 5 timed repetitions, median per input.",
            "conditions: job start thermal status 0, battery 37.6 C; job end status 1, battery 38.8 C; first timed input at status 0, 37.6 C; screen on (FLAG_KEEP_SCREEN_ON, setShowWhenLocked), USB powered, gate activity top-resumed (warm).",
            "parity vs the official gliner2 2.0.0 fp32 CPU implementation (whitespace splitter for English / char splitter for Japanese, threshold 0.5): identical span sets on 70/70 inputs, span micro-F1 1.000, max |dconfidence| 0.001377, non-finite inputs 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over the packed output (5467700 elements over 70 inputs) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same float16-table inputs (recomputed by scripts/litert_compat_records.py 2026-10-02): 1572 violating elements, max |d| 0.0002432, max |d|/|ref| 2.098.",
            "Reference row: on-device CPU with the same inputs; XNNPACK takes all but one node (the DEQUANTIZE-fed FULLY_CONNECTED weights are float16 constants).",
            "10-02 13:17:44.021 28777 28800 I litert  : [cpu_registry.cc:75] XNNPACK CPU accelerator registered.",
            "10-02 13:17:44.151 28777 28800 I tflite  : Created TensorFlow Lite XNNPACK delegate for CPU.",
            "10-02 13:17:44.152 28777 28800 I tflite  : Replacing 1147 out of 1148 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 3 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": 49.4759895,
          "loads": true,
          "max_abs_diff": 0.00024318695068359375,
          "max_rel_diff": 2.09808349609375,
          "metrics": {
            "battery_temp_c_end": 38.8,
            "battery_temp_c_first_input": 37.6,
            "battery_temp_c_start": 37.6,
            "element_rule_elements": 5467700.0,
            "element_rule_violations": 1572.0,
            "first_input_median_ms": 32.720156,
            "fixtures": 70.0,
            "identical_span_sets": 70.0,
            "load_compile_ms": 332.425989,
            "max_abs_dconfidence_vs_reference": 0.0013766884803771973,
            "max_ms": 59.023803,
            "min_ms": 31.200469,
            "nonfinite_inputs": 0.0,
            "span_f1": 1.0,
            "spans_fn": 0.0,
            "spans_fp": 0.0,
            "spans_tp": 277.0,
            "sustained_median_ms": 49.4759895,
            "thermal_status_end": 1.0,
            "thermal_status_first_input": 0.0,
            "thermal_status_start": 0.0,
            "timed_repetitions_per_input": 5.0,
            "warmups_per_input": 2.0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "CpuOptions numThreads=4 (XNNPACK); debug gate app com.gliner25multi.gate",
          "total_ops": 1148,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliner2.5-multi__gliner25_multi_s128_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1148,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliner25multi-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.gliner25multi.gate (android/gate/, Kotlin CompiledModel 2.2.0, Environment.create(context, BuiltinNpuAcceleratorProvider(Qualcomm)) with the Qualcomm dispatch/compiler libraries in jniLibs), one process per (graph, accelerator), inputs pushed as raw float32 files built on the Mac from the float16 host table (rows upcast to float32); timed interval = the five input writes + run + packed-output readback per input, 2 warm-ups + 5 timed repetitions, median per input.",
            "conditions: job start thermal status 0, battery 36.6 C; job end status 0, battery 36.6 C; first timed input at status 0, 36.6 C; screen on (FLAG_KEEP_SCREEN_ON, setShowWhenLocked), USB powered, gate activity top-resumed (foreground).",
            "parity vs the official gliner2 2.0.0 fp32 CPU implementation (whitespace splitter for English / char splitter for Japanese, threshold 0.5): identical span sets on 69/70 inputs, span micro-F1 0.998, max |dconfidence| 0.03256, non-finite inputs 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over the packed output (5467700 elements over 70 inputs) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same float16-table inputs (recomputed by scripts/litert_compat_records.py 2026-10-02): 2682186 violating elements, max |d| 2.281, max |d|/|ref| 2.031e+04.",
            "Default GPU precision: finite outputs, one 0.5-threshold span flipped (fixture 30 person 'Marco', official confidence 0.517), max |dconfidence| 0.033 — the reason the card requires explicit FP32.",
            "10-02 13:46:45.866  3116  3138 I litert  : [npu_registry.cc:30] NPU accelerator registered.",
            "10-02 13:46:46.004  3116  3138 I tflite  : Replacing 1148 out of 1148 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 15.035599000000001,
          "loads": true,
          "max_abs_diff": 2.2813587188720703,
          "max_rel_diff": 20312.8828125,
          "metrics": {
            "battery_temp_c_end": 36.6,
            "battery_temp_c_first_input": 36.6,
            "battery_temp_c_start": 36.6,
            "element_rule_elements": 5467700.0,
            "element_rule_violations": 2682186.0,
            "first_input_median_ms": 13.904739,
            "fixtures": 70.0,
            "identical_span_sets": 69.0,
            "load_compile_ms": 1225.536405,
            "max_abs_dconfidence_vs_reference": 0.032560884952545166,
            "max_ms": 19.388073,
            "min_ms": 13.821771,
            "nonfinite_inputs": 0.0,
            "span_f1": 0.9981916817359855,
            "spans_fn": 1.0,
            "spans_fp": 0.0,
            "spans_tp": 276.0,
            "sustained_median_ms": 15.035599000000001,
            "thermal_status_end": 0.0,
            "thermal_status_first_input": 0.0,
            "thermal_status_start": 0.0,
            "timed_repetitions_per_input": 5.0,
            "warmups_per_input": 2.0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=DEFAULT (fp16 compute) — diagnostic probe, not a shipping setting; debug gate app com.gliner25multi.gate",
          "total_ops": 1148,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliner2.5-multi__gliner25_multi_s128_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1148,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliner25multi-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.gliner25multi.gate (android/gate/, Kotlin CompiledModel 2.2.0, Environment.create(context, BuiltinNpuAcceleratorProvider(Qualcomm)) with the Qualcomm dispatch/compiler libraries in jniLibs), one process per (graph, accelerator), inputs pushed as raw float32 files built on the Mac from the float16 host table (rows upcast to float32); timed interval = the five input writes + run + packed-output readback per input, 2 warm-ups + 5 timed repetitions, median per input.",
            "conditions: job start thermal status 0, battery 34.1 C; job end status 0, battery 34.1 C; first timed input at status 0, 34.1 C; screen on (FLAG_KEEP_SCREEN_ON, setShowWhenLocked), USB powered, gate activity top-resumed (foreground).",
            "parity vs the official gliner2 2.0.0 fp32 CPU implementation (whitespace splitter for English / char splitter for Japanese, threshold 0.5): identical span sets on 70/70 inputs, span micro-F1 1.000, max |dconfidence| 0.001369, non-finite inputs 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over the packed output (5467700 elements over 70 inputs) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same float16-table inputs (recomputed by scripts/litert_compat_records.py 2026-10-02): 2152 violating elements, max |d| 0.0001841, max |d|/|ref| 3.052.",
            "GPU explicit FP32 is the validated setting of this family (the default precision stays finite but flips a 0.5-threshold span: s128 69/70, see the test row for s128).",
            "10-02 10:57:10.007 27340 27364 I litert  : [npu_registry.cc:30] NPU accelerator registered.",
            "10-02 10:57:10.307 27340 27364 I tflite  : Replacing 1148 out of 1148 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 24.286589,
          "loads": true,
          "max_abs_diff": 0.00018405914306640625,
          "max_rel_diff": 3.0517578125,
          "metrics": {
            "battery_temp_c_end": 34.1,
            "battery_temp_c_first_input": 34.1,
            "battery_temp_c_start": 34.1,
            "element_rule_elements": 5467700.0,
            "element_rule_violations": 2152.0,
            "first_input_median_ms": 24.468073,
            "fixtures": 70.0,
            "identical_span_sets": 70.0,
            "load_compile_ms": 1489.946718,
            "max_abs_dconfidence_vs_reference": 0.0013687610626220703,
            "max_ms": 31.088385,
            "min_ms": 22.834375,
            "nonfinite_inputs": 0.0,
            "span_f1": 1.0,
            "spans_fn": 0.0,
            "spans_fp": 0.0,
            "spans_tp": 277.0,
            "sustained_median_ms": 24.286589,
            "thermal_status_end": 0.0,
            "thermal_status_first_input": 0.0,
            "thermal_status_start": 0.0,
            "timed_repetitions_per_input": 5.0,
            "warmups_per_input": 2.0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "GpuOptions precision=FP32; debug gate app com.gliner25multi.gate",
          "total_ops": 1148,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliner2.5-multi__gliner25_multi_s128_wfp16__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-10-02",
          "decode_tokens_per_s": null,
          "delegated_ops": 1148,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-gliner25multi-gate",
            "os_build": "Android 16 (SDK 36)",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Debug gate app com.gliner25multi.gate (android/gate/, Kotlin CompiledModel 2.2.0, Environment.create(context, BuiltinNpuAcceleratorProvider(Qualcomm)) with the Qualcomm dispatch/compiler libraries in jniLibs), one process per (graph, accelerator), inputs pushed as raw float32 files built on the Mac from the float16 host table (rows upcast to float32); timed interval = the five input writes + run + packed-output readback per input, 2 warm-ups + 5 timed repetitions, median per input.",
            "conditions: job start thermal status 0, battery 36 C; job end status 0, battery 36 C; first timed input at status 0, 36 C; screen on (FLAG_KEEP_SCREEN_ON, setShowWhenLocked), USB powered, gate activity top-resumed (foreground).",
            "parity vs the official gliner2 2.0.0 fp32 CPU implementation (whitespace splitter for English / char splitter for Japanese, threshold 0.5): identical span sets on 69/70 inputs, span micro-F1 0.998, max |dconfidence| 0.04781, non-finite inputs 0.",
            "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) over the packed output (5467700 elements over 70 inputs) against the Mac ai-edge-litert 2.1.6 CPU run of the same graph bytes on the same float16-table inputs (recomputed by scripts/litert_compat_records.py 2026-10-02): 2655666 violating elements, max |d| 1.06, max |d|/|ref| 1.936e+04.",
            "The whole graph ran as one DispatchDelegate node ('Replacing 1 out of 1 node(s) with delegate (DispatchDelegate)' after 'Partitioned subgraph<0>, selected N ops, from a total of N ops'); load_compile_ms is the on-device JIT compile. Every span-set difference is a 0.5-threshold near-tie: official confidences 0.502 (fixture 71 product) and 0.503 (fixture 46 location) dropped, one extra span at NPU confidence 0.519 (fixture 72 location); the work reports the NPU on the card but does not call it a validated path.",
            "10-02 12:47:28.583 20944 20971 I litert  : [npu_registry.cc:30] NPU accelerator registered.",
            "10-02 12:47:28.851 20944 20971 I litert  : [compiler_plugin.cc:741] Partitioned subgraph<0>, selected 1148 ops, from a total of 1148 ops. resulted in 1 partitions.",
            "10-02 12:47:33.943 20944 20971 I tflite  : Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": 8.253828,
          "loads": true,
          "max_abs_diff": 1.0596599578857422,
          "max_rel_diff": 19361.908203125,
          "metrics": {
            "battery_temp_c_end": 36.0,
            "battery_temp_c_first_input": 36.0,
            "battery_temp_c_start": 36.0,
            "element_rule_elements": 5467700.0,
            "element_rule_violations": 2655666.0,
            "first_input_median_ms": 8.348594,
            "fixtures": 70.0,
            "identical_span_sets": 69.0,
            "load_compile_ms": 5402.455311,
            "max_abs_dconfidence_vs_reference": 0.04780912399291992,
            "max_ms": 9.721094,
            "min_ms": 8.061354,
            "nonfinite_inputs": 0.0,
            "span_f1": 0.9981916817359855,
            "spans_fn": 1.0,
            "spans_fp": 0.0,
            "spans_tp": 276.0,
            "sustained_median_ms": 8.253828,
            "thermal_status_end": 0.0,
            "thermal_status_first_input": 0.0,
            "thermal_status_start": 0.0,
            "timed_repetitions_per_input": 5.0,
            "warmups_per_input": 2.0
          },
          "output_match": false,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "QualcommOptions htpPerformanceMode=BURST, JIT (first compile, cache dir of a fresh app process); debug gate app com.gliner25multi.gate",
          "total_ops": 1148,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-10-02/gliner2.5-multi__gliner25_multi_s128_wfp16__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "gliner2.5-multi",
    "id": "gliner2.5-multi__gliner25_multi_s128_wfp16",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/GLiNER2.5-Multi-LiteRT",
    "task": "token-classification"
  },
  "pitfalls": [
    "Weight storage does not set the GPU compute precision: request GpuOptions(precision = FP32). At the default precision the outputs stay finite but one of 70 span sets changes (card 'Measured quality and performance', s128 row 69/70, max confidence drift 3.3e-2).",
    "The graph is the dense prefix of the upstream BoundaryExtractor (mDeBERTa-v3-base encoder, boundary encoder, boundary query head, per-token projections): the host looks up the word-embedding rows before the graph (host_assets/word_embeddings_fp16.bin [250112,768] float16, rows upcast to float32; fp32 table as reference) and runs the upstream gliner2 sparse decoder after it (host_assets/runtime/, 16 decoder tensors); the five inputs are args_0..args_4 = inputs_embeds [1,128,768], attention_mask [1,128], text_routing [1,48,128], query_routing [1,5,128], text_mask [1,48]; one packed float32 output [1,1,1,78110] with 17 logical slices (graph_contract_s128.json).",
    "The caller selects the word splitter per request — official whitespace for space-delimited text such as English, official char for Japanese / CJK; there is no language detector, and capacity counts processed text slots and encoded tokens (text ending in 。 takes one extra slot). Fixed window N=128 / T=48: longer inputs are rejected, not truncated (card 'Files and supported configuration', HOST_CONTRACT.md).",
    "Two fp32-exact rewrites are inside the shipped graph (attention-mask fill -10000 instead of the float32 minimum; each of the 29 active LayerNorms computed on x*2^-3 with eps*2^-6; 60/60 bit-identical comparisons with the stock graph) so the same files stay finite under fp16 delegates — finite, not exact: the Galaxy S26 Hexagon NPU (JIT) matched 69/70, 72/75 and 77/80 span sets with every difference a 0.5-threshold near-tie (max confidence drift 4.8e-2), so the card reports the NPU and does not offer it as a supported path (card 'Measured quality and performance').",
    "Back-to-back requests heat the phone: the s256 and s512 GPU jobs rose from thermal status 0 to 1-2 and their sustained medians (93.0 / 422.5 ms) are 1.5-2.1x the cool opening values (63.5 / 200.9 ms); the card states both with their condition. Latency is a single-device sample (card 'Measured quality and performance').",
    "Validated only for the five fixed labels in that order, threshold 0.5, English (whitespace) and Japanese (char), one window per call, on the Galaxy S26 (card 'Limits'); conversion fidelity to the official gliner2 fp32 implementation (span F1 1.000 on 80 bilingual inputs), not a general accuracy claim — upstream misses carry over (4 of 10 expected organizations on ten early Japanese sentences)."
  ],
  "schema_version": "1.2"
}
