{
  "artifacts": [
    {
      "file": "granite-speech-5.0-2b_ekv2048.litertlm",
      "sha256": "bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642",
      "size_mb": 2693.554
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "ibm-granite/granite-speech-5.0-2b rev 98160169 -> the language model: export_hf, prefill signatures 64, 256 and 1,024, recipe dynamic_wi8_afp32, KV cache 2,048; the audio encoder: written with tensors of rank 4 or less (the relative-position shift as a skew, the convolutions as 2-D convolutions, no GATHER_ND, BROADCAST_TO or DIV), exported for a 24-second window, float16 weights; the source's tokenizer.json byte for byte, a chat template rendering `USER: <|audio|>{instruction}\\n ASSISTANT:`, stop token <|end_of_text|> (100257), model type generic_model (card L59, L202, L204-L205)",
    "quantization": "language model: int8 weights (one scale per output channel), KV cache of 2,048 tokens, no activation-type declaration (float16 activations on the GPU, the runtime's default); token embedding table int8; audio encoder: float16 weights computed in float32, one 24-second window (card L48, L54-L56)",
    "tool": "litert-torch 0.10.0 (export_hf for the Granite language model, the token embedding exported separately; the audio encoder as a PyTorch module traced with litert-torch) packed with litert-lm-builder 0.19.1; repro = hf-to-litertlm REPRODUCE.md, the granite-speech-5.0-2b_ekv2048.litertlm entry (card L204-L207)",
    "tool_version": "litert-torch 0.10.0, litert-lm-builder 0.19.1 (litert-torch 0.10.0's own packing calls a builder API that 0.19 no longer has); LiteRT-LM 0.19.1 for every measurement except the encoder-alone times (ai-edge-litert 2.3.0) (card L31, L203, L206)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-tcd1 (run s26e, 2026-11-07 09:10:05 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 78s since 09:08:47): 9.9 s clip first text 1.5349 s, end 2.9905 s (5.9 s 1.5099 / 2.5088 s, 29.4 s 3.0236 / 7.4008 s), encoder in-call 1023.8 ms on the 9.9 s clip (mean over the 3 clips 1300.7667 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4204508 kB; samples 16, CPU capped 12, GPU capped 12, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-cpu-tcd2 (run s26e, 2026-11-07 09:17:53 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 150s since 09:15:23): 9.9 s clip first text 1.7302 s, end 3.5202 s (5.9 s 1.7138 / 2.935 s, 29.4 s 3.8545 / 9.7342 s), encoder in-call 1073.5 ms on the 9.9 s clip (mean over the 3 clips 1492.7333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4204316 kB; samples 19, CPU capped 18, GPU capped 12, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-cpu-tcd2.err):",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 8.5675,
            "clip29_4s_first_text_s": 3.439,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.7219,
            "clip5_9s_first_text_s": 1.6119,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 3.2553,
            "clip9_9s_first_text_s": 1.6325,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 1048.65,
            "encoder_in_call_ms_mean_3_clips": 1396.75,
            "first_text_sum_3_clips_s": 6.6835,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.5349,
            "run1_cpu_capped_samples": 12.0,
            "run1_samples": 16.0,
            "run1_vmhwm_kb": 4204508.0,
            "run2_clip9_9s_first_text_s": 1.7302,
            "run2_cpu_capped_samples": 18.0,
            "run2_samples": 19.0,
            "run2_vmhwm_kb": 4204316.0
          },
          "output_match": null,
          "peak_mem_mb": 4305.4162,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the CPU at the default thread count, in the thread-count session of 2026-11-07 09:10-09:18 JST (the default and thread_count=6 alternating, 2 processes each), audio encoder on the GPU; ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1632.55
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu(thread_count=6), audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-tc1 (run s8x, 2026-11-07 15:17:59 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 114s since 15:16:05): 9.9 s clip first text 1.6702 s, end 3.4836 s (5.9 s 1.5982 / 2.8049 s, 29.4 s 3.6853 / 9.5414 s), encoder in-call 1025.5 ms on the 9.9 s clip (mean over the 3 clips 1388.8333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4245056 kB; samples 22, CPU capped 21, GPU capped 14, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-cpu-tc2 (run s8x, 2026-11-07 15:20:18 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 88s since 15:18:50): 9.9 s clip first text 1.5598 s, end 3.1185 s (5.9 s 1.5151 / 2.5973 s, 29.4 s 3.1218 / 8.4209 s), encoder in-call 1018.1 ms on the 9.9 s clip (mean over the 3 clips 1300.3333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4130516 kB; samples 18, CPU capped 14, GPU capped 14, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-cpu-tc2.err):",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 8.9811,
            "clip29_4s_first_text_s": 3.4036,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.7011,
            "clip5_9s_first_text_s": 1.5566,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 3.3011,
            "clip9_9s_first_text_s": 1.615,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 1021.8,
            "encoder_in_call_ms_mean_3_clips": 1344.5833,
            "first_text_sum_3_clips_s": 6.5752,
            "lm_cpu_thread_count": 6.0,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.6702,
            "run1_cpu_capped_samples": 21.0,
            "run1_samples": 22.0,
            "run1_vmhwm_kb": 4245056.0,
            "run2_clip9_9s_first_text_s": 1.5598,
            "run2_cpu_capped_samples": 14.0,
            "run2_samples": 18.0,
            "run2_vmhwm_kb": 4130516.0
          },
          "output_match": null,
          "peak_mem_mb": 4346.9373,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the CPU at thread_count=6, audio encoder on the GPU, in the session of 2026-11-07 15:15-15:33 JST (this file and an unpublished build with the audio encoder on the NPU alternating per language-model backend, A B B A, 2 processes each); ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1615.0
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu(thread_count=6), audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-tc6a (run s26e, 2026-11-07 09:12:21 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 99s since 09:10:42): 9.9 s clip first text 1.4617 s, end 2.9325 s (5.9 s 1.4123 / 2.4349 s, 29.4 s 2.9245 / 7.2703 s), encoder in-call 964.7 ms on the 9.9 s clip (mean over the 3 clips 1245.9333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4203828 kB; samples 16, CPU capped 11, GPU capped 10, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-cpu-tc6b (run s26e, 2026-11-07 09:14:46 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 108s since 09:12:58): 9.9 s clip first text 1.407 s, end 2.8813 s (5.9 s 1.3992 / 2.4123 s, 29.4 s 2.7784 / 7.2108 s), encoder in-call 923.8 ms on the 9.9 s clip (mean over the 3 clips 1205.1333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4204136 kB; samples 16, CPU capped 10, GPU capped 9, thermal max 3; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-cpu-tc6b.err):",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 7.2405,
            "clip29_4s_first_text_s": 2.8514,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.4236,
            "clip5_9s_first_text_s": 1.4058,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 2.9069,
            "clip9_9s_first_text_s": 1.4344,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 944.25,
            "encoder_in_call_ms_mean_3_clips": 1225.5333,
            "first_text_sum_3_clips_s": 5.6916,
            "lm_cpu_thread_count": 6.0,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.4617,
            "run1_cpu_capped_samples": 11.0,
            "run1_samples": 16.0,
            "run1_vmhwm_kb": 4203828.0,
            "run2_clip9_9s_first_text_s": 1.407,
            "run2_cpu_capped_samples": 10.0,
            "run2_samples": 16.0,
            "run2_vmhwm_kb": 4204136.0
          },
          "output_match": null,
          "peak_mem_mb": 4305.0353,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the CPU at thread_count=6, in the thread-count session of 2026-11-07 09:10-09:18 JST (the default and thread_count=6 alternating, 2 processes each), audio encoder on the GPU; ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1434.35
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu(thread_count=6), audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-t6a (run s8mt, 2026-11-07 15:01:29 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 119s since 14:59:30): 9.9 s clip first text 1.7385 s, end 3.6002 s (5.9 s 1.7212 / 2.9932 s, 29.4 s 4.0099 / 10.0128 s), encoder in-call 1083.7 ms on the 9.9 s clip (mean over the 3 clips 1564.4333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4204700 kB; samples 20, CPU capped 18, GPU capped 13, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-cpu-t6b (run s8mt, 2026-11-07 15:09:25 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 140s since 15:07:05): 9.9 s clip first text 1.4745 s, end 2.9731 s (5.9 s 1.3941 / 2.4434 s, 29.4 s 2.8459 / 7.5124 s), encoder in-call 980.3 ms on the 9.9 s clip (mean over the 3 clips 1222.9667 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4203896 kB; samples 16, CPU capped 12, GPU capped 9, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-cpu-t6b.err):",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 8.7626,
            "clip29_4s_first_text_s": 3.4279,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.7183,
            "clip5_9s_first_text_s": 1.5576,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 3.2866,
            "clip9_9s_first_text_s": 1.6065,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 1032.0,
            "encoder_in_call_ms_mean_3_clips": 1393.7,
            "first_text_sum_3_clips_s": 6.592,
            "lm_cpu_thread_count": 6.0,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.7385,
            "run1_cpu_capped_samples": 18.0,
            "run1_samples": 20.0,
            "run1_vmhwm_kb": 4204700.0,
            "run2_clip9_9s_first_text_s": 1.4745,
            "run2_cpu_capped_samples": 12.0,
            "run2_samples": 16.0,
            "run2_vmhwm_kb": 4203896.0
          },
          "output_match": null,
          "peak_mem_mb": 4305.6128,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the CPU at thread_count=6, in the thread-count session of 2026-11-07 15:01-15:10 JST (thread_count=6 and 8 alternating, 2 processes each), audio encoder on the GPU; ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1606.5
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu(thread_count=8), audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-t8a (run s8mt, 2026-11-07 15:04:04 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 109s since 15:02:15): 9.9 s clip first text 1.6444 s, end 3.4829 s (5.9 s 1.6162 / 2.8689 s, 29.4 s 3.5936 / 9.4207 s), encoder in-call 1028.8 ms on the 9.9 s clip (mean over the 3 clips 1383.2667 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4204596 kB; samples 18, CPU capped 17, GPU capped 12, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-cpu-t8b (run s8mt, 2026-11-07 15:06:27 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 99s since 15:04:48): 9.9 s clip first text 1.486 s, end 3.0124 s (5.9 s 1.4157 / 2.4945 s, 29.4 s 2.9167 / 7.7524 s), encoder in-call 979.2 ms on the 9.9 s clip (mean over the 3 clips 1252.5333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4203348 kB; samples 16, CPU capped 11, GPU capped 9, thermal max 3; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-cpu-t8b.err):",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 8.5865,
            "clip29_4s_first_text_s": 3.2551,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.6817,
            "clip5_9s_first_text_s": 1.516,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 3.2477,
            "clip9_9s_first_text_s": 1.5652,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 1004.0,
            "encoder_in_call_ms_mean_3_clips": 1317.9,
            "first_text_sum_3_clips_s": 6.3363,
            "lm_cpu_thread_count": 8.0,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.6444,
            "run1_cpu_capped_samples": 17.0,
            "run1_samples": 18.0,
            "run1_vmhwm_kb": 4204596.0,
            "run2_clip9_9s_first_text_s": 1.486,
            "run2_cpu_capped_samples": 11.0,
            "run2_samples": 16.0,
            "run2_vmhwm_kb": 4203348.0
          },
          "output_match": null,
          "peak_mem_mb": 4305.5063,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the CPU at thread_count=8, in the thread-count session of 2026-11-07 15:01-15:10 JST (thread_count=6 and 8 alternating, 2 processes each), audio encoder on the GPU; ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1565.2
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-tc (run s26a, 2026-11-07 06:31:23 JST at the run's start, clean; rested 80s since 06:30:03): 9.9 s clip first text 1.6193 s, end 3.2721 s (5.9 s 1.6095 / 2.7547 s, 29.4 s 3.0446 / 8.0221 s), encoder in-call 1030.0 ms on the 9.9 s clip (mean over the 3 clips 1284.8667 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4127080 kB; samples 21, CPU capped 19, GPU capped 4, thermal max 0; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-cpu-tc2 (run s26a, 2026-11-07 06:41:54 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 68s since 06:40:46): 9.9 s clip first text 1.7502 s, end 3.4768 s (5.9 s 1.7268 / 2.9806 s, 29.4 s 3.6144 / 9.1433 s), encoder in-call 1088.1 ms on the 9.9 s clip (mean over the 3 clips 1455.9 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 4207388 kB; samples 20, CPU capped 18, GPU capped 12, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-cpu-tc2.err):",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 8.5827,
            "clip29_4s_first_text_s": 3.3295,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.8677,
            "clip5_9s_first_text_s": 1.6681,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 3.3744,
            "clip9_9s_first_text_s": 1.6847,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 1059.05,
            "encoder_in_call_ms_mean_3_clips": 1370.3834,
            "first_text_sum_3_clips_s": 6.6824,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.6193,
            "run1_cpu_capped_samples": 19.0,
            "run1_samples": 21.0,
            "run1_vmhwm_kb": 4127080.0,
            "run2_clip9_9s_first_text_s": 1.7502,
            "run2_cpu_capped_samples": 18.0,
            "run2_samples": 20.0,
            "run2_vmhwm_kb": 4207388.0
          },
          "output_match": null,
          "peak_mem_mb": 4308.3653,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the CPU, audio encoder on the GPU; ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1684.75
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-ml (run s26d, 2026-11-07 09:04:32 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 105s since 09:02:47): item ml (24.94 s): prefill 270 tokens, 98 tokens out, stopped True, cap hit False, empty False, loop False",
            "u1w24-cpu-ml: d8_multilingual: WER 0.0 % (66 words), CER 0.0 % (280 characters), byte-equal to the source model's text: True; rule: PASS = stopped, not empty, WER <= 10.0 % against R's one pass (BasicTextNormalizer); verdict PASS",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "multilingual_cer_pct": 0.0,
            "multilingual_decode_tokens": 98.0,
            "multilingual_text_equal_source": 1.0,
            "multilingual_wer_pct": 0.0,
            "multilingual_words": 66.0,
            "wer_line_pct": 10.0
          },
          "output_match": true,
          "peak_mem_mb": 4509.1881,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the source card's multilingual_sample.wav (24.9 s, English and then French) with the instruction `transcribe the speech with proper punctuation and capitalization.`, one call without streaming (then a text-only control), greedy, at most 256 new tokens; output_match = the reply stops, is not empty and is within 10 % word error rate of the source model's transcript (the line, set before the runs); audio encoder on the GPU",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": 2048,
          "date": "2026-11-07",
          "decode_tokens_per_s": 15.3363,
          "delegated_ops": 2173,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Benchmark(model, cpu, prefill_tokens=1024, decode_tokens=256, max_num_tokens=2048); 3 x run on one engine (no audio backend: the LM alone)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-cpu-td1 (run s26e, 2026-11-07 09:20:15 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 98s since 09:18:37): runs 2-3 prefill 226.86 / 222.81, decode 15.29 / 15.34 tok/s, TTFT 4.579 / 4.661 s, Init Total 491.25 ms; VmHWM 2569136 kB (the 1 s sampler), CPU-capped samples per run [18, 18, 18]",
            "u1w24-cpu-td2 (run s26e, 2026-11-07 09:28:34 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 109s since 09:26:45): runs 2-3 prefill 244.14 / 222.88, decode 15.34 / 15.38 tok/s, TTFT 4.260 / 4.659 s, Init Total 552.03 ms; VmHWM 2568872 kB (the 1 s sampler), CPU-capped samples per run [19, 19, 19]",
            "(the same lines in u1w24-cpu-td2.err):",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 0 (prefill_1024).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318",
            "delegation: XNNPACK; the record's counts = the last prefill_1024 line (XNNPACK's second pass over the graph): 2173 of 2259 nodes, the rest on the CPU kernels"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 256.0,
            "init_total_ms": 521.64,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "prefill_tokens": 1024.0,
            "run1_decode_tok_s": 15.314,
            "run1_decode_tok_s_it1": 21.9543,
            "run1_init_total_ms": 491.25,
            "run1_prefill_tok_s": 224.837,
            "run1_ttft_s": 4.6201,
            "run1_vmhwm_kb": 2569136.0,
            "run2_decode_tok_s": 15.3586,
            "run2_decode_tok_s_it1": 22.3134,
            "run2_init_total_ms": 552.03,
            "run2_prefill_tok_s": 233.5107,
            "run2_ttft_s": 4.4594,
            "run2_vmhwm_kb": 2568872.0,
            "runs": 2.0
          },
          "output_match": null,
          "peak_mem_mb": 2630.7953,
          "prefill_tokens_per_s": 229.1739,
          "provenance": "measured",
          "runs": true,
          "total_ops": 2259,
          "ttft_ms": 4539.75
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, gpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-gpu-tc1 (run s8x, 2026-11-07 15:27:53 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 88s since 15:26:25): 9.9 s clip first text 1.3796 s, end 2.6852 s (5.9 s 1.3667 / 2.2657 s, 29.4 s 2.8882 / 7.1216 s), encoder in-call 829.0 ms on the 9.9 s clip (mean over the 3 clips 1218.4 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 2509732 kB; samples 20, CPU capped 18, GPU capped 10, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-gpu-tc2 (run s8x, 2026-11-07 15:30:17 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 99s since 15:28:38): 9.9 s clip first text 1.2589 s, end 2.4886 s (5.9 s 1.2298 / 2.0877 s, 29.4 s 2.1816 / 5.9556 s), encoder in-call 801.9 ms on the 9.9 s clip (mean over the 3 clips 1047.1667 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 2323976 kB; samples 16, CPU capped 7, GPU capped 7, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-gpu-tc2.err):",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1616 out of 1616 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 6.5386,
            "clip29_4s_first_text_s": 2.5349,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.1767,
            "clip5_9s_first_text_s": 1.2982,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 2.5869,
            "clip9_9s_first_text_s": 1.3192,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 815.45,
            "encoder_in_call_ms_mean_3_clips": 1132.7834,
            "first_text_sum_3_clips_s": 5.1524,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.3796,
            "run1_cpu_capped_samples": 18.0,
            "run1_samples": 20.0,
            "run1_vmhwm_kb": 2509732.0,
            "run2_clip9_9s_first_text_s": 1.2589,
            "run2_cpu_capped_samples": 7.0,
            "run2_samples": 16.0,
            "run2_vmhwm_kb": 2323976.0
          },
          "output_match": null,
          "peak_mem_mb": 2569.9656,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the GPU, audio encoder on the GPU, in the session of 2026-11-07 15:15-15:33 JST (this file and an unpublished build with the audio encoder on the NPU alternating per language-model backend, A B B A, 2 processes each); ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1319.25
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, gpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message_async(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-gpu-tc (run s26b, 2026-11-07 06:49:02 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 124s since 06:46:58): 9.9 s clip first text 1.2086 s, end 2.4405 s (5.9 s 1.21 / 2.0672 s, 29.4 s 2.2793 / 6.1492 s), encoder in-call 738.2 ms on the 9.9 s clip (mean over the 3 clips 1037.5333 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 2373956 kB; samples 17, CPU capped 10, GPU capped 10, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "u1w24-gpu-tc2 (run s26b, 2026-11-07 07:03:31 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 119s since 07:01:32): 9.9 s clip first text 1.2152 s, end 2.4339 s (5.9 s 1.2593 / 2.109 s, 29.4 s 2.329 / 6.0636 s), encoder in-call 768.6 ms on the 9.9 s clip (mean over the 3 clips 1088.7 ms; rule: encoder in-call ms = a class's mean (call wall - prefill turn s - decode turn s) - the text control's); VmHWM 2485692 kB; samples 16, CPU capped 12, GPU capped 6, thermal max 2; every transcript equal to the Mac reference of the same backend: True",
            "(the same lines in u1w24-gpu-tc2.err):",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1616 out of 1616 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "calls_per_clip": 2.0,
            "clip29_4s_decode_tokens_per_call": 93.0,
            "clip29_4s_end_s": 6.1064,
            "clip29_4s_first_text_s": 2.3041,
            "clip29_4s_prefill_tokens_per_call": 312.0,
            "clip5_9s_decode_tokens_per_call": 23.0,
            "clip5_9s_end_s": 2.0881,
            "clip5_9s_first_text_s": 1.2347,
            "clip5_9s_prefill_tokens_per_call": 78.0,
            "clip9_9s_decode_tokens_per_call": 32.0,
            "clip9_9s_end_s": 2.4372,
            "clip9_9s_first_text_s": 1.2119,
            "clip9_9s_prefill_tokens_per_call": 117.0,
            "encoder_in_call_ms_clip9_9s": 753.4,
            "encoder_in_call_ms_mean_3_clips": 1063.1166,
            "first_text_sum_3_clips_s": 4.7507,
            "max_num_tokens": 2048.0,
            "max_output_tokens": 256.0,
            "processes": 2.0,
            "run1_clip9_9s_first_text_s": 1.2086,
            "run1_cpu_capped_samples": 10.0,
            "run1_samples": 17.0,
            "run1_vmhwm_kb": 2373956.0,
            "run2_clip9_9s_first_text_s": 1.2152,
            "run2_cpu_capped_samples": 12.0,
            "run2_samples": 16.0,
            "run2_vmhwm_kb": 2485692.0
          },
          "output_match": null,
          "peak_mem_mb": 2545.3486,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing: one streaming call (send_message_async) per clip with a new conversation each, after a warm-up clip, for the 5.9, 9.9 and 29.4 s LibriSpeech clips (en00 / en03 / en04 of the en 20) and a text-only control, the test's rendered prompt (`transcribe the speech with proper punctuation and capitalization.`), greedy, at most 256 new tokens; the mean of 2 calls per clip in each of 2 processes; language model on the GPU, audio encoder on the GPU; ttft_ms = the 9.9 s clip's first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 1211.9
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-granite-speech-devicebench",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, gpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys",
            "u1w24-gpu-ml (run s26d, 2026-11-07 09:05:55 JST at the run's start, uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 58s since 09:04:57): item ml (24.94 s): prefill 270 tokens, 100 tokens out, stopped True, cap hit False, empty False, loop False",
            "u1w24-gpu-ml: d8_multilingual: WER 0.0 % (66 words), CER 0.0 % (280 characters), byte-equal to the source model's text: False; rule: PASS = stopped, not empty, WER <= 10.0 % against R's one pass (BasicTextNormalizer); verdict PASS",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1616 out of 1616 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "multilingual_cer_pct": 0.0,
            "multilingual_decode_tokens": 100.0,
            "multilingual_text_equal_source": 0.0,
            "multilingual_wer_pct": 0.0,
            "multilingual_words": 66.0,
            "wer_line_pct": 10.0
          },
          "output_match": true,
          "peak_mem_mb": 2533.7037,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the source card's multilingual_sample.wav (24.9 s, English and then French) with the instruction `transcribe the speech with proper punctuation and capitalization.`, one call without streaming (then a text-only control), greedy, at most 256 new tokens; output_match = the reply stops, is not empty and is within 10 % word error rate of the source model's transcript (the line, set before the runs); audio encoder on the GPU",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__galaxy-s26.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: the PyPI litert-lm 0.19.1 wheel on the Mac (Python 3.13.13; GPU = WebGPU on Metal); the model loaded once per process, `Engine(bundle, backend=CPU, audio_backend=GPU)`, CPU threads 4",
            "Bundle granite-speech-5.0-2b_r7u1w24_enc6-fp16-w24_lm-int8-kv2048_nodecl.litertlm, 2824395760 B; sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642 for that file = the Hub LFS oid",
            "os_build: sw_vers and hw.model read 2026-11-07 at this entry (macOS 27.0, 26A428; Mac16,9); its own records name the host 'Mac M4 Max 128 GB, macOS 27.0'",
            "u1w24_cpu: the medians over passes p1-p5 (its first_s lists = these five files' values; p0 not in the table)",
            "ls_clip01 first chunk 0.5293 s, wall 0.868 s; ls_clip00 first chunk 0.4058 s, wall 0.963 s; en_us_1671 first chunk 0.4063 s, wall 1.045 s; en_us_1670 first chunk 0.7725 s, wall 2.316 s",
            "ls_clip01 first chunk 0.5243 s, wall 0.872 s; ls_clip00 first chunk 0.41 s, wall 0.968 s; en_us_1671 first chunk 0.4097 s, wall 2.529 s; en_us_1670 first chunk 0.7675 s, wall 2.305 s",
            "ls_clip01 first chunk 0.5267 s, wall 0.875 s; ls_clip00 first chunk 0.4082 s, wall 0.967 s; en_us_1671 first chunk 0.4095 s, wall 1.049 s; en_us_1670 first chunk 0.7666 s, wall 2.286 s",
            "ls_clip01 first chunk 0.5327 s, wall 0.88 s; ls_clip00 first chunk 0.4145 s, wall 0.978 s; en_us_1671 first chunk 0.4106 s, wall 1.052 s; en_us_1670 first chunk 0.7621 s, wall 2.3 s",
            "ls_clip01 first chunk 1.3304 s, wall 1.656 s; ls_clip00 first chunk 0.4248 s, wall 0.95 s; en_us_1671 first chunk 0.4225 s, wall 1.016 s; en_us_1670 first chunk 0.7048 s, wall 1.908 s"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "clip10_0s_first_text_s": 0.4097,
            "clip10_0s_total_s": 1.049,
            "clip29_3s_first_text_s": 0.7666,
            "clip29_3s_total_s": 2.3,
            "clip5_9s_first_text_s": 0.41,
            "clip5_9s_total_s": 0.967,
            "cpu_threads": 4.0,
            "max_output_tokens": 256.0,
            "passes": 5.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing on the Mac: a warm-up clip, then a 5.9 s LibriSpeech clip, a 10.0 s and a 29.3 s FLEURS en_us clip in that order, one streaming call per clip with a new conversation each, the model loaded once per process, the transcription instruction, greedy, at most 256 new tokens; the median of 5 processes; audio encoder on the GPU; ttft_ms = the 10.0 s clip's (the third conversation's) first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 409.7
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: the PyPI litert-lm 0.19.1 wheel on the Mac (Python 3.13.13; GPU = WebGPU on Metal); `Engine(bundle, backend=CPU, audio_backend=GPU)`, CPU threads 4, `SamplerConfig(top_k=1)`, a new conversation per clip, at most 256 new tokens",
            "Bundle granite-speech-5.0-2b_r7u1w24_enc6-fp16-w24_lm-int8-kv2048_nodecl.litertlm, 2824395760 B (bf8eb5a8... for that file = the Hub LFS oid)",
            "run: 2026-11-07 06:10:12-06:16:21 JST",
            "os_build: sw_vers and hw.model read 2026-11-07 at this entry (macOS 27.0, 26A428; Mac16,9); its own records name the host 'Mac M4 Max 128 GB, macOS 27.0'",
            "En 4.408 % (70/1588); ja 5.206 % (140/2689); eu 11.934 % (282/2363) | source: en 4.597 %; ja 5.132 %; eu 12.103 % | lines: en 5.097; ja 6.132; eu 13.103 | clips 220, empty 0, looping 0, not ending at the stop token 0, character-equal to the source model 180; pass True"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "clips": 220.0,
            "empty": 0.0,
            "en_wer_line_pct": 5.097,
            "en_wer_pct": 4.408,
            "en_wer_source_pct": 4.597,
            "fr_de_es_pt_wer_line_pct": 13.103,
            "fr_de_es_pt_wer_pct": 11.934,
            "fr_de_es_pt_wer_source_pct": 12.103,
            "ja_cer_line_pct": 6.132,
            "ja_cer_pct": 5.206,
            "ja_cer_source_pct": 5.132,
            "looping": 0.0,
            "not_eos": 0.0,
            "text_equal_source": 180.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR gate: the 220 clips (20 LibriSpeech dev-clean, 50 FLEURS en_us, 50 FLEURS ja_jp and 25 FLEURS clips each of French, German, Spanish and Portuguese), the instruction `transcribe the speech with proper punctuation and capitalization.`, greedy, at most 256 new tokens, one conversation per clip; scored against the source model in transformers 5.22.1 on the CPU (English with Whisper's English normalizer, the other European languages with its basic normalizer, Japanese by characters); line = the source model + 0.5 pt (English WER) / + 1.0 pt (Japanese CER) / + 1.0 pt (French + German + Spanish + Portuguese WER), no empty, looping or unstopped reply; audio encoder on the GPU",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": 2048,
          "date": "2026-11-07",
          "decode_tokens_per_s": 36.21,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: the PyPI litert-lm 0.19.1 wheel on the Mac (Python 3.13.13; GPU = WebGPU on Metal); `litert-lm benchmark <file> -p 1024 -d 256 --max-num-tokens 2048 --runs 3 --backend cpu --cpu-thread-count 4 --cache no`, one process: a warm-up and 3 runs; the CLI prints the mean of the 3 runs",
            "Benchmarking model: granite-speech-5.0-2b_r7u1w24_enc6-fp16-w24_lm-int8-kv2048_nodecl.litertlm",
            "Names that local file with sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642 = the Hub LFS oid (at 16:13: the Mac copy re-hashed equal)",
            "window: this run 07:46:14-07:46:54 ('bench u1w24 cpu')",
            "os_build: sw_vers and hw.model read 2026-11-07 at this entry (macOS 27.0, 26A428; Mac16,9); its own records name the host 'Mac M4 Max 128 GB, macOS 27.0'",
            "Prefill speed:        841.62 tokens/s | Decode speed:         36.21 tokens/s | Init time:            11.0036 s | Time to first token:  1.2452 s"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 256.0,
            "init_s": 11.0036,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "prefill_tokens": 1024.0,
            "runs": 1.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 841.62,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 1245.2
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: the PyPI litert-lm 0.19.1 wheel on the Mac (Python 3.13.13; GPU = WebGPU on Metal); the model loaded once per process, `Engine(bundle, backend=GPU, audio_backend=GPU)`, CPU threads 4",
            "Bundle granite-speech-5.0-2b_r7u1w24_enc6-fp16-w24_lm-int8-kv2048_nodecl.litertlm, 2824395760 B; sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642 for that file = the Hub LFS oid",
            "os_build: sw_vers and hw.model read 2026-11-07 at this entry (macOS 27.0, 26A428; Mac16,9); its own records name the host 'Mac M4 Max 128 GB, macOS 27.0'",
            "u1w24_gpu: the medians over passes p1-p5 (its first_s lists = these five files' values; p0 not in the table)",
            "ls_clip01 first chunk 0.3054 s, wall 0.425 s; ls_clip00 first chunk 0.2259 s, wall 0.418 s; en_us_1671 first chunk 0.2249 s, wall 0.445 s; en_us_1670 first chunk 0.3964 s, wall 0.93 s",
            "ls_clip01 first chunk 0.8071 s, wall 0.926 s; ls_clip00 first chunk 0.2274 s, wall 0.417 s; en_us_1671 first chunk 0.229 s, wall 0.449 s; en_us_1670 first chunk 0.3953 s, wall 0.932 s",
            "ls_clip01 first chunk 0.3086 s, wall 0.426 s; ls_clip00 first chunk 0.2262 s, wall 0.416 s; en_us_1671 first chunk 0.2258 s, wall 0.441 s; en_us_1670 first chunk 0.3977 s, wall 0.943 s",
            "ls_clip01 first chunk 0.3072 s, wall 0.423 s; ls_clip00 first chunk 0.226 s, wall 0.417 s; en_us_1671 first chunk 0.2255 s, wall 0.442 s; en_us_1670 first chunk 0.3978 s, wall 0.93 s",
            "ls_clip01 first chunk 0.8076 s, wall 0.925 s; ls_clip00 first chunk 0.2254 s, wall 0.415 s; en_us_1671 first chunk 0.226 s, wall 0.449 s; en_us_1670 first chunk 0.3992 s, wall 0.936 s"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "clip10_0s_first_text_s": 0.2258,
            "clip10_0s_total_s": 0.445,
            "clip29_3s_first_text_s": 0.3977,
            "clip29_3s_total_s": 0.932,
            "clip5_9s_first_text_s": 0.226,
            "clip5_9s_total_s": 0.417,
            "cpu_threads": 4.0,
            "max_output_tokens": 256.0,
            "passes": 5.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR clip timing on the Mac: a warm-up clip, then a 5.9 s LibriSpeech clip, a 10.0 s and a 29.3 s FLEURS en_us clip in that order, one streaming call per clip with a new conversation each, the model loaded once per process, the transcription instruction, greedy, at most 256 new tokens; the median of 5 processes; audio encoder on the GPU; ttft_ms = the 10.0 s clip's (the third conversation's) first streamed text, audio encoding included",
          "total_ops": null,
          "ttft_ms": 225.8
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: the PyPI litert-lm 0.19.1 wheel on the Mac (Python 3.13.13; GPU = WebGPU on Metal); `Engine(bundle, backend=GPU, audio_backend=GPU)`, CPU threads 4, `SamplerConfig(top_k=1)`, a new conversation per clip, at most 256 new tokens",
            "Bundle granite-speech-5.0-2b_r7u1w24_enc6-fp16-w24_lm-int8-kv2048_nodecl.litertlm, 2824395760 B (bf8eb5a8... for that file = the Hub LFS oid)",
            "run: 2026-11-07 06:17:04-06:19:24 JST",
            "os_build: sw_vers and hw.model read 2026-11-07 at this entry (macOS 27.0, 26A428; Mac16,9); its own records name the host 'Mac M4 Max 128 GB, macOS 27.0'",
            "En 4.534 % (72/1588); ja 5.244 % (141/2689); eu 12.146 % (287/2363) | source: en 4.597 %; ja 5.132 %; eu 12.103 % | lines: en 5.097; ja 6.132; eu 13.103 | clips 220, empty 0, looping 0, not ending at the stop token 0, character-equal to the source model 195; pass True"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "clips": 220.0,
            "empty": 0.0,
            "en_wer_line_pct": 5.097,
            "en_wer_pct": 4.534,
            "en_wer_source_pct": 4.597,
            "fr_de_es_pt_wer_line_pct": 13.103,
            "fr_de_es_pt_wer_pct": 12.146,
            "fr_de_es_pt_wer_source_pct": 12.103,
            "ja_cer_line_pct": 6.132,
            "ja_cer_pct": 5.244,
            "ja_cer_source_pct": 5.132,
            "looping": 0.0,
            "not_eos": 0.0,
            "text_equal_source": 195.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR gate: the 220 clips (20 LibriSpeech dev-clean, 50 FLEURS en_us, 50 FLEURS ja_jp and 25 FLEURS clips each of French, German, Spanish and Portuguese), the instruction `transcribe the speech with proper punctuation and capitalization.`, greedy, at most 256 new tokens, one conversation per clip; scored against the source model in transformers 5.22.1 on the CPU (English with Whisper's English normalizer, the other European languages with its basic normalizer, Japanese by characters); line = the source model + 0.5 pt (English WER) / + 1.0 pt (Japanese CER) / + 1.0 pt (French + German + Spanish + Portuguese WER), no empty, looping or unstopped reply; audio encoder on the GPU",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-11-07",
          "decode_tokens_per_s": 109.87,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: the PyPI litert-lm 0.19.1 wheel on the Mac (Python 3.13.13; GPU = WebGPU on Metal); `litert-lm benchmark <file> -p 1024 -d 256 --max-num-tokens 2048 --runs 3 --backend gpu --cache no`, one process: a warm-up and 3 runs; the CLI prints the mean of the 3 runs",
            "Benchmarking model: granite-speech-5.0-2b_r7u1w24_enc6-fp16-w24_lm-int8-kv2048_nodecl.litertlm",
            "Names that local file with sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642 = the Hub LFS oid (at 16:13: the Mac copy re-hashed equal)",
            "window: this run 07:46:54-07:47:06 ('bench u1w24 gpu')",
            "os_build: sw_vers and hw.model read 2026-11-07 at this entry (macOS 27.0, 26A428; Mac16,9); its own records name the host 'Mac M4 Max 128 GB, macOS 27.0'",
            "Prefill speed:        3373.13 tokens/s | Decode speed:         109.87 tokens/s | Init time:            2.0224 s | Time to first token:  0.3127 s",
            "(the run's one error-level line): E0000 00:00:1794005216.071558 125648876 gpu_model_builder.cc:4150] TopK requires src tensor C dimension to be divisible by 4."
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 256.0,
            "init_s": 2.0224,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "prefill_tokens": 1024.0,
            "runs": 1.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 3373.13,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 312.7
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "pixel-11-pro",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Pixel 11 Pro",
            "machine_label": "pixel-11-pro-granite-speech-devicebench",
            "os_build": "Android 17 (CD1A.260905.001.B1)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Google Tensor G6",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, cpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=Pixel 11 Pro release=17 build_id=CD1A.260905.001.B1 fingerprint=google/grizzly/grizzly:17/CD1A.260905.001.B1/16238327:user/release-keys",
            "u1w24-cpu-g0 (p11a, 2026-11-07 06:42 JST): en 2.208 % (source 2.428, line 2.928); n 20, stopped 20, empty 0, looping 0, transcripts equal to the Mac reference of the same backend 17 of 20; set verdict PASS",
            "u1w24-cpu-g1 (p11a, 2026-11-07 06:46 JST): ja 7.591 % (source 7.261, line 8.261); eu 2.027 % (source 2.027, line 3.027); n 18, stopped 18, empty 0, looping 0, transcripts equal to the Mac reference of the same backend 7 of 18; set verdict PASS",
            "Acceptance rule: en20 and xx18 runs PASS; verdict PASS",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1620 out of 1785 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 238 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1450 out of 1616 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 243 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 1936 out of 2021 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 83 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 2173 out of 2259 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 82 partitions for subgraph 1 (prefill_256).",
            "plus one XNNPACK line per composite subgraph: odml.rms_norm 318"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "en20_empty": 0.0,
            "en20_looping": 0.0,
            "en20_n": 20.0,
            "en20_stopped": 20.0,
            "en20_text_equal_mac": 17.0,
            "en_wer_line_pct": 2.928,
            "en_wer_pct": 2.208,
            "en_wer_source_pct": 2.428,
            "fr_de_es_pt_wer_line_pct": 3.027,
            "fr_de_es_pt_wer_pct": 2.027,
            "fr_de_es_pt_wer_source_pct": 2.027,
            "ja_cer_line_pct": 8.261,
            "ja_cer_pct": 7.591,
            "ja_cer_source_pct": 7.261,
            "xx18_empty": 0.0,
            "xx18_looping": 0.0,
            "xx18_n": 18.0,
            "xx18_stopped": 18.0,
            "xx18_text_equal_mac": 7.0
          },
          "output_match": true,
          "peak_mem_mb": 2633.9615,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR gate on the phone: 38 of the 220 clips (the 20 LibriSpeech clips, the first 10 Japanese and the first 2 French, German, Spanish and Portuguese FLEURS clips), the instruction `transcribe the speech with proper punctuation and capitalization.`, greedy, at most 256 new tokens, one conversation per clip; line = the source model's rate on the same clips + 0.5 pt (English WER) / + 1.0 pt (Japanese CER) / + 1.0 pt (French + German + Spanish + Portuguese WER), no empty, looping or unstopped reply; audio encoder on the GPU",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__pixel-11-pro.json"
      },
      {
        "device": "pixel-11-pro",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-07",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Pixel 11 Pro",
            "machine_label": "pixel-11-pro-granite-speech-devicebench",
            "os_build": "Android 17 (CD1A.260905.001.B1)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Google Tensor G6",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "artifact: litert-community/granite-speech-5.0-2b `granite-speech-5.0-2b_ekv2048.litertlm` at main 53f901f0: LFS sha256 bf8eb5a8075b4fd3cdb1109a9edf0c247ff2bfb52b95699c31626aae80f3a642, 2824395760 B (Hub API read 2026-11-07 16:18:12 JST); the file was added in Hub commit be28d417 (2026-11-07T01:48:52Z)",
            "runtime: PyPI litert-lm-api 0.19.1 Android wheel on the phone (CPython 3.14.8; liblitert-lm.so sha256 3bc61c88a5998a41cc0eb35c5b133c14e5635790290839452d35cc936231d903), the launcher ltlm_py0191_granite with the run adapter, one process per run: `Engine(model, gpu, audio_backend=gpu, max_num_tokens=2048, enable_benchmark=True); per item: render on a throwaway conversation, then create_conversation(SamplerConfig(top_k=1)).send_message(message, max_output_tokens=256)`; protocol: a benchmark script, the phone's state and a ready check before each run",
            "Device: model=Pixel 11 Pro release=17 build_id=CD1A.260905.001.B1 fingerprint=google/grizzly/grizzly:17/CD1A.260905.001.B1/16238327:user/release-keys",
            "u1w24-gpu-g0 (p11a, 2026-11-07 06:51 JST): en 1.987 % (source 2.428, line 2.928); n 20, stopped 20, empty 0, looping 0, transcripts equal to the Mac reference of the same backend 20 of 20; set verdict PASS",
            "u1w24-gpu-g1 (p11b, 2026-11-07 06:56 JST): ja 7.591 % (source 7.261, line 8.261); eu 2.027 % (source 2.027, line 3.027); n 18, stopped 18, empty 0, looping 0, transcripts equal to the Mac reference of the same backend 17 of 18; set verdict PASS",
            "Acceptance rule: en20 and xx18 runs PASS; verdict PASS",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1616 out of 1616 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).",
            "VERBOSE: Replacing 1785 out of 1785 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64).",
            "VERBOSE: Replacing 1616 out of 1616 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 1 (prefill_embedder_1024).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 2 (prefill_embedder_256).",
            "VERBOSE: Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 3 (prefill_embedder_64).",
            "VERBOSE: Replacing 2674 out of 2674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "en20_empty": 0.0,
            "en20_looping": 0.0,
            "en20_n": 20.0,
            "en20_stopped": 20.0,
            "en20_text_equal_mac": 20.0,
            "en_wer_line_pct": 2.928,
            "en_wer_pct": 1.987,
            "en_wer_source_pct": 2.428,
            "fr_de_es_pt_wer_line_pct": 3.027,
            "fr_de_es_pt_wer_pct": 2.027,
            "fr_de_es_pt_wer_source_pct": 2.027,
            "ja_cer_line_pct": 8.261,
            "ja_cer_pct": 7.591,
            "ja_cer_source_pct": 7.261,
            "xx18_empty": 0.0,
            "xx18_looping": 0.0,
            "xx18_n": 18.0,
            "xx18_stopped": 18.0,
            "xx18_text_equal_mac": 17.0
          },
          "output_match": true,
          "peak_mem_mb": 1474.4863,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "ASR gate on the phone: 38 of the 220 clips (the 20 LibriSpeech clips, the first 10 Japanese and the first 2 French, German, Spanish and Portuguese FLEURS clips), the instruction `transcribe the speech with proper punctuation and capitalization.`, greedy, at most 256 new tokens, one conversation per clip; line = the source model's rate on the same clips + 0.5 pt (English WER) / + 1.0 pt (Japanese CER) / + 1.0 pt (French + German + Spanish + Portuguese WER), no empty, looping or unstopped reply; audio encoder on the GPU",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-07/granite-speech-5.0-2b-ekv2048__pixel-11-pro.json"
      }
    ]
  },
  "model": {
    "family": "granite-speech",
    "id": "granite-speech-5.0-2b-ekv2048",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/granite-speech-5.0-2b",
    "task": "automatic-speech-recognition"
  },
  "pitfalls": [
    "Context is 2,048 tokens, prompt and reply together: one message fits about 116 s of audio at the highest reply length of the card's 220 test clips, 131 s at their 95th percentile and 152 s at their median; a message that does not fit is refused with INVALID_ARGUMENT, not cut. Take granite-speech-5.0-2b.litertlm (4,096 tokens) when one message has to hold more than about 2 minutes of audio (card L42, L65, L198).",
    "One message holds one audio item, then the instruction as text (for example `transcribe the speech with proper punctuation and capitalization.`); send 16 kHz mono WAV (16-bit PCM and 32-bit float were tested; other sample rates, stereo and other containers were not) (card L63-L64).",
    "The audio encoder runs on the GPU (audio_backend GPU, `--audio-backend gpu`); the card did not test this file with the audio encoder on the CPU. The encoder always computes a whole 24-second window, also for a short clip (card L64, L67, L195).",
    "Use a new conversation for each clip: from the second message of a conversation on, LiteRT-LM 0.19.1 puts the previous answer's stop token into the context with a zero embedding, so a second transcript in one conversation would start from a context the model never saw; not measured (card L102).",
    "A file that ends in exact digital silence is read as slightly shorter (the encoder takes the last non-zero sample as the end); on the card's 5 test files with 173 to 274 trailing zero samples the source language model's transcript was unchanged (card L66).",
    "With the language model on the CPU, LiteRT-LM quantizes the activations to int8 at run time: the transcripts were the source model's character for character on 180 of 220 clips against 195 on the GPU, at error rates within 0.3 points of each other (card L197).",
    "Galaxy S26 CPU threads: Backend.CPU(thread_count=6) reached the first text sooner than the default (4 in the LiteRT-LM 0.19.1 source): 1.43 s against 1.63 s on the 9.9-second clip in one session; the LM benchmark was not faster with 6 threads (card L100, L117).",
    "No NPU file and no int4 file: an audio encoder compiled for the Galaxy S26 NPU turned the English part of the source card's sample into French (the NPU's sigmoid returns 0 below about -5.7); the fixed encoder was slower than this file's GPU window on clips up to 24 s; an int4 language model prefilled at 42 % of int8's rate on a Mac CPU (card L196)."
  ],
  "schema_version": "1.2"
}
