{ "artifacts": [ { "file": "gliner25_decide_s512_wfp16.tflite", "sha256": "bbe150fa6f7bf29cf4b12a6eff630b3a8fb59f41b41ea14a68b1b693fea6fc71", "size_mb": 773.791 } ], "benchmarks": [], "browser": { "backends": [ { "backend": "wasm_xnnpack", "date": "2026-09-26", "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": null, "loads": true, "max_rel_diff": null, "output_match": null, "provenance": "measured", "runs": false }, { "backend": "webgpu_mldrift", "date": "2026-09-26", "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": null, "loads": false, "max_rel_diff": null, "output_match": null, "provenance": "measured", "runs": false } ], "demo_url": null, "sweep_source": "data/sweep/2.5.3/2026-09-26/gliner2.5-decide__gliner25_decide_s512_wfp16.json" }, "conversion": { "command": "cd conversion && python -B export_shape.py --seq 512 && python -B quantize_wfp16.py --seq 512 (conversion/README.md 'Order': the fp32 export first, then the weight-only float16 cast of that gated export; exports/gliner25_decide_s512_wfp16.tflite)", "quantization": "float16 weight storage (ai-edge-quantizer 0.8.0 FLOAT_CASTING, weight-only 16-bit): the 146 FULLY_CONNECTED weight tensors stored as float16 with a DEQUANTIZE to float32; activations and every other constant stay float32; 1,780 operators vs 1,634 in the fp32 reference graph; no INT8 file is shipped (card 'Files', 'Limits'; conversion/README.md)", "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" ], "coverage_ops_pct": 89.0, "lint_report_version": "1.1", "litert_version": "2.2.0", "matched_provenance_counts": { "measured": 1731, "unmatched": 49 }, "partitions": 196 }, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu_xnnpack", "context_length": null, "date": "2026-09-26", "decode_tokens_per_s": null, "delegated_ops": 1780, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-gliner25decide-gate", "os_build": "Android 16 (SDK 36, build BP4A.251205.006)", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "Native LiteRT 2.2.0 CompiledModel C-API runner (gpu_runner, build gliner25-decide-round2-v1, the GLiNER2.5-Small runner rebuilt for three inputs; vendor libLiteRt.so + libLiteRtClGlAccelerator.so sha256-pinned in android/provenance.json), one process per job, fixture inputs pushed as .f32 files (inputs_embeds, attention_mask, label_routing), sha256-checked on the device. Round-3 inputs: rows of the float16 embedding table upcast to float32 on the host = the shipping path (results/device/inputs_s512/inputs.json table.kind 'fp16', word_embeddings_fp16.bin sha256 f90f5d83...). Sources: ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/gpu_gate_r3.json (job g: 'not a gate: S26 CompiledModel CPU (card CPU column)'), raw outputs ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r3_g_s512_wfp16_cpu/NNN.f32, runner stderr ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r3_g_s512_wfp16_cpu/runner.stderr.log, supervisor recount ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/SUPERVISOR.md.", "logcat/stderr: 'Replacing 1780 out of 1780 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions'; compile PASS in 692.4 ms; is_fully_accelerated=true; hardware_accelerators 'CPU only (bitmask 1)'", "conditions (round 3, card CPU column, not a gate): cool start (cool after 121.5 s wait; rule kgsl <= 50 C and thermal_status 0); screen kept awake via 'svc power stayon usb' (restored to false afterwards), USB powered; before: kgsl 41.5 C, max clock 1300 MHz, thermal_pwrlevel 0, thermal_status 0, battery 37.5 C 86 %, wakefulness Awake; after: kgsl 50.8 C, max clock 500 MHz, thermal_pwrlevel 10, thermal_status 2, battery 41.2 C; runner 2026-09-26T13:14:40+09:00 -> 2026-09-26T13:22:27+09:00.", "parity: 42/42 decisions equal the official gliner2 2.0.0 fp32 classify_text result (42 requests = 21 model-card examples + 21 fast-decisions dev rows that fit the window); max |dlogit| vs official 3.311e-03 (ticket_route_019), max |dprob| 2.256e-04; NaN/Inf in any repetition: 0.", "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CompiledModel CPU run of the same graph on the same inputs (~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/table_gate_raw/s512_graph-wfp16_table-fp16), recomputed 2026-09-26 from the saved .f32 outputs: 42/42 fixtures pass over all 32 slots (0 violating elements; valid label slots: 42/42, 0 violations); max |d| 1.192e-05, max rel 2.239e-04.", "latency scope: all input lock/write/unlock + CompiledModel run + output lock/read/unlock; no file I/O; 2 warm-ups then 5 timed runs per fixture, 42 fixtures = 210 timed runs; median 1665.98 / min 863.07 / max 1788.94 ms; per-fixture medians 879.92 ms (first) -> 1785.20 ms (last) in run order." ], "failure_class": null, "full_delegation": true, "latency_p50_ms": 1665.9826295, "loads": true, "max_abs_diff": 1.1920928955078125e-05, "max_rel_diff": 0.0002239291206933558, "metrics": { "battery_temp_c_end": 41.2, "battery_temp_c_start": 37.5, "cool_start_wait_s": 121.5, "decisions_equal_official": 42, "fixtures": 42, "gpu_max_clock_mhz_end": 500, "iterations": 210, "kgsl_temp_c_end": 50.8, "kgsl_temp_c_start": 41.5, "latency_max_ms": 1788.941093, "latency_min_ms": 863.065, "load_compile_ms": 692.383593, "max_abs_dlogit_vs_official": 0.0033109188079833984, "max_abs_dprob_vs_official": 0.00022557377815246582, "per_fixture_median_first_ms": 879.916354, "per_fixture_median_last_ms": 1785.198124, "repetitions_per_fixture": 5, "shared_rule_pass_fixtures": 42, "shared_rule_violating_elements": 0, "thermal_status_end": 2, "thermal_status_start": 0, "warmups_per_fixture": 2 }, "output_match": true, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "signature": "CPU XNNPACK (native runner, default threads)", "total_ops": 1780, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-09-26/gliner2.5-decide__gliner25_decide_s512_wfp16__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "cpu_xnnpack", "context_length": null, "date": "2026-09-26", "decode_tokens_per_s": null, "delegated_ops": 1780, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-gliner25decide-gate", "os_build": "Android 16 (SDK 36, build BP4A.251205.006)", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "Android sample app gate (litert-community/GLiNER2.5-Decide-LiteRT android/, package com.gliner25decide, debug build APK sha256 0c5f08b04413..., com.google.ai.edge.litert:litert:2.2.0): one Environment, Accelerator.CPU, CpuOptions 4 threads (XNNPACK); the app builds the inputs itself (SentencePiece tokenizer, gliner2 schema, memory-mapped float16 table) and the on-device token ids, [L] positions and padding equal the captured Python batches on all 126 (request, window) pairs (126/126). Sources: ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/app_gate_r5.json (gate.CPU.windows.s512), ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r5_gate_cpu/pulled/files/gate/gate_CPU.json (per-pair logits), logcat ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r5_gate_cpu/logcat.txt.", "logcat (three compiles in one process, one per window): '09-26 14:45:50.856 9723 9775 I tflite : Replacing 1780 out of 1780 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main).' | '09-26 14:45:52.430 9723 9775 I tflite : Replacing 1780 out of 1780 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main).' | '09-26 14:45:54.399 9723 9775 I tflite : Replacing 1780 out of 1780 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 0 (main).'.", "conditions: screen behind the secure keyguard, activity shown over it (setShowWhenLocked, debug diagnostics only), verified top-resumed / cpuset top-app; cool start (cool after 192.3 s wait); before: kgsl 41.1 C, max clock 1300 MHz, thermal_status 0, battery 37.2 C 83 %; after the whole gate (all three windows, 2026-09-26T14:45:50+09:00 + 332 s): kgsl 51.9 C, max clock 500 MHz, thermal_pwrlevel 10, thermal_status 2, battery 43.4 C. This window's numbers are one third of that run; the per-window thermal state is not recorded separately.", "parity: 42/42 pairs' decisions equal the official gliner2 2.0.0 fp32 result, 42/42 all logits finite; max |dlogit| vs official 3.311e-03; CPU logits bit-identical to the native-runner round-3 job (max |d| 0.0). output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CompiledModel CPU run of the same graph on the same inputs (~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/table_gate_raw/s512_graph-wfp16_table-fp16), recomputed 2026-09-26 from the per-pair logits: 42/42 pairs pass over all 32 slots (0 violating elements); max |d| 1.192e-05, max rel 2.239e-04.", "latency: graph phase = First input write through synchronized output readback; 1 warm-up + 3 timed runs per pair, 42 pairs = 126 timed runs; median 1242.72 / min 711.42 / max 1405.62 ms over all timed runs; pair medians: median 1243.16, first three 735.08, 982.06, 1065.04 -> last three 1244.27, 1390.12, 1402.48 ms; tokenize+embed pair-median 22.73 ms, decode 0.051 ms (host phases, outside latency_p50_ms)." ], "failure_class": null, "full_delegation": true, "latency_p50_ms": 1242.723984, "loads": true, "max_abs_diff": 1.1920928955078125e-05, "max_rel_diff": 0.0002239291206933558, "metrics": { "cpu_threads": 4, "decisions_equal_official": 42, "decode_pair_median_ms": 0.0506515, "gpu_max_clock_mhz_end_whole_gate": 500, "inputs_identical_to_python_pairs": 126, "iterations": 126, "kgsl_temp_c_end_whole_gate": 51.9, "kgsl_temp_c_start": 41.1, "latency_max_ms": 1405.6192179999998, "latency_min_ms": 711.421094, "max_abs_dlogit_vs_native_runner": 0.0, "max_abs_dlogit_vs_official": 0.0033109188079833984, "pair_median_first_ms": 735.0803129999999, "pair_median_last_ms": 1402.4848440000003, "pair_median_ms": 1243.1611715, "pairs": 42, "shared_rule_pass_fixtures": 42, "shared_rule_violating_elements": 0, "thermal_status_end_whole_gate": 2, "thermal_status_start": 0, "timed_runs_per_pair": 3, "tokenize_embed_pair_median_ms": 22.7265885, "warmup_runs_per_pair": 1 }, "output_match": true, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "signature": "CPU XNNPACK 4 threads; Android sample app (debug build)", "total_ops": 1780, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-09-26/gliner2.5-decide__gliner25_decide_s512_wfp16__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu_mldrift", "context_length": null, "date": "2026-09-26", "decode_tokens_per_s": null, "delegated_ops": 1780, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-gliner25decide-gate", "os_build": "Android 16 (SDK 36, build BP4A.251205.006)", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "Native LiteRT 2.2.0 CompiledModel C-API runner (gpu_runner, build gliner25-decide-round2-v1, the GLiNER2.5-Small runner rebuilt for three inputs; vendor libLiteRt.so + libLiteRtClGlAccelerator.so sha256-pinned in android/provenance.json), one process per job, fixture inputs pushed as .f32 files (inputs_embeds, attention_mask, label_routing), sha256-checked on the device. Round-3 inputs: rows of the float16 embedding table upcast to float32 on the host = the shipping path (results/device/inputs_s512/inputs.json table.kind 'fp16', word_embeddings_fp16.bin sha256 f90f5d83...). Sources: ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/gpu_gate_r3.json (job c: 'G3 S26 GPU precision FP32, wfp16 storage, fp16 table inputs'), raw outputs ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r3_c_s512_wfp16_fp32/NNN.f32, runner stderr ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r3_c_s512_wfp16_fp32/runner.stderr.log, supervisor recount ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/SUPERVISOR.md.", "logcat/stderr: 'Replacing 1780 out of 1780 node(s) with delegate (LITERT_CL) node, yielding 1 partitions'; compile PASS in 2766.0 ms; is_fully_accelerated=true; hardware_accelerators 'GPU only (bitmask 2)'; gpu_options_toml 'precision = 2'", "conditions (round 3, gate G3): cool start (cool after 151.8 s wait; rule kgsl <= 50 C and thermal_status 0); screen kept awake via 'svc power stayon usb' (restored to false afterwards), USB powered; before: kgsl 41.8 C, max clock 1300 MHz, thermal_pwrlevel 0, thermal_status 0, battery 37.6 C 92 %, wakefulness Awake; after: kgsl 58.5 C, max clock 646 MHz, thermal_pwrlevel 8, thermal_status 2, battery 44.7 C; runner 2026-09-26T12:48:20+09:00 -> 2026-09-26T12:52:00+09:00.", "parity: 42/42 decisions equal the official gliner2 2.0.0 fp32 classify_text result (42 requests = 21 model-card examples + 21 fast-decisions dev rows that fit the window); max |dlogit| vs official 3.301e-03 (ticket_route_019), max |dprob| 2.258e-04; NaN/Inf in any repetition: 0.", "output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CompiledModel CPU run of the same graph on the same inputs (~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/table_gate_raw/s512_graph-wfp16_table-fp16), recomputed 2026-09-26 from the saved .f32 outputs: 42/42 fixtures pass over all 32 slots (0 violating elements; valid label slots: 42/42, 0 violations); max |d| 2.575e-05, max rel 7.895e-05.", "latency scope: all input lock/write/unlock + CompiledModel run + output lock/read/unlock; no file I/O; 2 warm-ups then 5 timed runs per fixture, 42 fixtures = 210 timed runs; median 733.12 / min 573.49 / max 936.55 ms; per-fixture medians 575.23 ms (first) -> 912.12 ms (last) in run order." ], "failure_class": null, "full_delegation": true, "latency_p50_ms": 733.1154160000001, "loads": true, "max_abs_diff": 2.574920654296875e-05, "max_rel_diff": 7.895220187492669e-05, "metrics": { "battery_temp_c_end": 44.7, "battery_temp_c_start": 37.6, "cool_start_wait_s": 151.8, "decisions_equal_official": 42, "fixtures": 42, "gpu_max_clock_mhz_end": 646, "iterations": 210, "kgsl_temp_c_end": 58.5, "kgsl_temp_c_start": 41.8, "latency_max_ms": 936.549166, "latency_min_ms": 573.492812, "load_compile_ms": 2765.954582, "max_abs_dlogit_vs_official": 0.003300905227661133, "max_abs_dprob_vs_official": 0.00022584199905395508, "per_fixture_median_first_ms": 575.227343, "per_fixture_median_last_ms": 912.124947, "repetitions_per_fixture": 5, "shared_rule_pass_fixtures": 42, "shared_rule_violating_elements": 0, "thermal_status_end": 2, "thermal_status_start": 0, "warmups_per_fixture": 2 }, "output_match": true, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "signature": "GpuOptions precision=FP32", "total_ops": 1780, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-09-26/gliner2.5-decide__gliner25_decide_s512_wfp16__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu_mldrift", "context_length": null, "date": "2026-09-26", "decode_tokens_per_s": null, "delegated_ops": 1780, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-gliner25decide-gate", "os_build": "Android 16 (SDK 36, build BP4A.251205.006)", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "Android sample app gate (litert-community/GLiNER2.5-Decide-LiteRT android/, package com.gliner25decide, debug build APK sha256 0c5f08b04413..., com.google.ai.edge.litert:litert:2.2.0): one Environment, Accelerator.GPU with GpuOptions(precision = FP32) (explicit); the app builds the inputs itself (SentencePiece tokenizer, gliner2 schema, memory-mapped float16 table) and the on-device token ids, [L] positions and padding equal the captured Python batches on all 126 (request, window) pairs (126/126). Sources: ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/app_gate_r5.json (gate.GPU.windows.s512), ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r5_gate_gpu/pulled/files/gate/gate_GPU.json (per-pair logits), logcat ~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/device/r5_gate_gpu/logcat.txt.", "logcat (three compiles in one process, one per window): '09-26 14:38:51.610 6980 7036 I tflite : Replacing 1780 out of 1780 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).' | '09-26 14:38:54.126 6980 7036 I tflite : Replacing 1780 out of 1780 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).' | '09-26 14:38:58.145 6980 7036 I tflite : Replacing 1780 out of 1780 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).'.", "conditions: screen behind the secure keyguard, activity shown over it (setShowWhenLocked, debug diagnostics only), verified top-resumed / cpuset top-app; cool start (cool after 0.1 s wait); before: kgsl 41.1 C, max clock 1300 MHz, thermal_status 0, battery 32.3 C 85 %; after the whole gate (all three windows, 2026-09-26T14:38:50+09:00 + 212 s): kgsl 49.6 C, max clock 1200 MHz, thermal_pwrlevel 1, thermal_status 2, battery 43.1 C. This window's numbers are one third of that run; the per-window thermal state is not recorded separately.", "parity: 42/42 pairs' decisions equal the official gliner2 2.0.0 fp32 result, 42/42 all logits finite; max |dlogit| vs official 3.301e-03; GPU logits bit-identical to the native-runner round-3 job (max |d| 0.0). output_match = shared element rule |out - ref| <= max(1e-5, 1e-3 * |ref|) against the Mac ai-edge-litert 2.1.6 CompiledModel CPU run of the same graph on the same inputs (~/code/codex-conversions/2026-09-26/gliner25-decide-litert/results/table_gate_raw/s512_graph-wfp16_table-fp16), recomputed 2026-09-26 from the per-pair logits: 42/42 pairs pass over all 32 slots (0 violating elements); max |d| 2.575e-05, max rel 7.895e-05.", "latency: graph phase = First input write through synchronized output readback; 1 warm-up + 3 timed runs per pair, 42 pairs = 126 timed runs; median 639.24 / min 576.02 / max 1047.52 ms over all timed runs; pair medians: median 638.77, first three 579.43, 578.33, 582.49 -> last three 672.03, 704.54, 695.11 ms; tokenize+embed pair-median 43.09 ms, decode 0.072 ms (host phases, outside latency_p50_ms)." ], "failure_class": null, "full_delegation": true, "latency_p50_ms": 639.23664, "loads": true, "max_abs_diff": 2.574920654296875e-05, "max_rel_diff": 7.895220187492669e-05, "metrics": { "decisions_equal_official": 42, "decode_pair_median_ms": 0.07210949999999999, "gpu_max_clock_mhz_end_whole_gate": 1200, "inputs_identical_to_python_pairs": 126, "iterations": 126, "kgsl_temp_c_end_whole_gate": 49.6, "kgsl_temp_c_start": 41.1, "latency_max_ms": 1047.515729, "latency_min_ms": 576.0178639999999, "max_abs_dlogit_vs_native_runner": 0.0, "max_abs_dlogit_vs_official": 0.003300905227661133, "pair_median_first_ms": 579.433229, "pair_median_last_ms": 695.1102599999999, "pair_median_ms": 638.7681769999999, "pairs": 42, "shared_rule_pass_fixtures": 42, "shared_rule_violating_elements": 0, "thermal_status_end_whole_gate": 2, "thermal_status_start": 0, "timed_runs_per_pair": 3, "tokenize_embed_pair_median_ms": 43.090417, "warmup_runs_per_pair": 1 }, "output_match": true, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "signature": "GpuOptions precision=FP32; Android sample app (debug build)", "total_ops": 1780, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-09-26/gliner2.5-decide__gliner25_decide_s512_wfp16__galaxy-s26.json" } ] }, "model": { "family": "gliner2.5-decide", "id": "gliner2.5-decide__gliner25_decide_s512_wfp16", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/GLiNER2.5-Decide-LiteRT", "task": "text-classification" }, "pitfalls": [ "Weight storage and GPU computation precision are separate settings: at the runtime's default GPU precision the s128 graph (float32 weights) compiled and returned finite logits on the Galaxy S26, but only 14 of 42 decisions matched the official result; with GpuOptions(precision = FP32) all 42 matched. Use the explicit FP32 option for every file (card 'GPU precision').", "The graph is the DeBERTa-v3-large encoder plus the classification head only: the host looks up the word embeddings before the graph (host_assets/word_embeddings_fp16.bin, [128011,1024] float16, upcast to float32 on lookup; the graph applies the embedding LayerNorm itself, so feed the raw rows) and turns the logits into decisions after it (softmax or sigmoid per task, thresholds). HOST_CONTRACT.md specifies the encoded sequence, the [L] marker positions and the decision rules (card 'Files'; HOST_CONTRACT.md).", "Assign the inputs by shape, not by name: the converter names them args_0..args_2 (inputs_embeds [1,N,1024], attention_mask [1,N] with 1.0 for encoded tokens, label_routing [1,32,N] one-hot at the j-th [L] marker); output_0 [1,1,1,32] holds one logit per label slot in request order, and slots past the label count hold a constant that is ignored (HOST_CONTRACT.md 'Graph signature').", "Fixed windows N = 128 / 256 / 512: the task schemas plus the text must fit in N encoded tokens with at most 32 labels in total; longer requests are rejected, never truncated, and gliner2's long-text chunking (classify_text_long) is not ported. Use the smallest window that holds the request (card 'Files', 'Limits').", "Back-to-back requests heat the phone: on the Galaxy S26 the GPU clock limit steps down from 1,300 MHz (s256 902 MHz, s512 646 MHz at the end of the job) and each s256 / s512 request gets slower during the job — the opening request is the cool value, the median mixes it with the throttled end (card 'Latency').", "GPU validation covers the Galaxy S26 (Adreno) only; NPU execution was not evaluated. No INT8 file is shipped: on GLiNER2.5 Small, the same family of graphs, dynamic-range INT8 FULLY_CONNECTED weights did not compile on the LiteRT 2.2.0 GPU, and no INT8 variant of this model was built or tested (card 'Limits').", "English only, as the source model; one text per call, no batching; few-shot examples and gliner2's extraction tasks (entities, relations, structures) are not part of the graphs (card 'Limits')." ], "schema_version": "1.2" }