{
  "artifacts": [
    {
      "file": "NVIDIA-Nemotron-Labs-Dory1.0-4B_int4_1loop.litertlm",
      "sha256": "53f935834360de82ba553b630e124578fb1384c6ff59675f035ae7023450dab4",
      "size_mb": 2289.351
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 rev b366b87f and this repository's int4 file (both downloaded at pinned commits and checked) -> the graph exported again with n_recurrent_loops set to 1 in the config (the same arguments otherwise: prefill signatures 128 / 512 / 1024 plus decode over a 2048-token KV budget, 36 KV caches, the MLP normalization rewrite of the same 11 blocks) -> data-free int4 with the int4 recipe -> the int4 file's GPTQ codes written weight by weight (each weight found by its block scales, which the data-free quantization reproduces; 289 of 289) -> GPU composites (426 in its four graphs) -> packed with the int4 file's metadata and embedding sections; run twice on 2026-10-23, each in a fresh directory, the script rebuilt every section of the file byte for byte (card L136)",
    "quantization": "the int4 file's weights byte for byte: int4 blockwise-32 on the linear layers with the int4 file's GPTQ codes and block scales; the int4 file's int8 embedding section; GPU composites; no activation type declared (the GPU runs the runtime's default fp16 activations) (card L32, L34)",
    "tool": "litert-torch 0.9.4 `export_hf` with a model definition of its own (transformers has no Dory class; the formulas follow NVIDIA's vLLM pull request); repro = REPRODUCE.md \"NVIDIA-Nemotron-Labs-Dory1.0-4B\" (card L24, L130, L136)",
    "tool_version": "litert-torch 0.9.4, ai-edge-quantizer 0.9.0, litert-converter 0.4.0, torch 2.13.0, transformers 5.14.1; packed with litert-lm-builder 0.18.0 (card L130)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-24",
          "decode_tokens_per_s": null,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory-loops-app",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.16.1",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: hfmodels-android samples/loops (main e93b645; the APK built from 034366c, whose difference to e93b645 is the sample's README only), hfmodels SDK 0.2.1-SNAPSHOT over litertlm-android 0.16.1; Run with 'Both' = the one-loop file, then the int4 file, one model loaded at a time, one conversation per answer, thinking off, the engine's default sampler (greedy); total_ms / first_token_ms = walls from send to the last / first streamed chunk, app_chunks_per_s = (chunks - 1) / (last - first chunk) as the app prints it ('tok/s' on the screen; tokenizer counts are 1.09-1.60 times the chunks, the sample's README)",
            "int4_1loop final_value 55 (mismatch, GSM8K answer 40), total_ms 19160.1, first_token_ms 1037.5, chunks 303, tok_s 16.72, load_ms 13562, thermal 0 -> 0; runtime_version 0.16.1, sdk 0.2.1-SNAPSHOT, device SM-S942Q BP4A.251205.006.S942QOPS1AZH9",
            "10-24 09:35:34.711 20880 20880 I loops   : ANSWER loop=1 variant=int4_1loop q=gsm8k-54 first_token_ms=1037.5 tok_s=16.72 total_ms=19160.1 chunks=303 thought_chars=0 load_ms=13562 backend=GPU mark=mismatch text=\"**Step-by-step reasoning:**\\n\\n1. **Identify the existing cats and their current kittens:**\\n   - The family already has two adult cats: Patchy and Trixie.\\n   - Trixie has had 12 kittens.\\n\\n2. **Calculate Patchy's kittens:**\\n   - The problem states Patchy has had thrice the number of kittens that Trixie has.\\n   - Trixie's kittens = 12.\\n   - Patchy's kittens = 3 × 12 = 36.\\n\\n3. **Calculate the total number of kittens from the cats:**\\n   - Trixie's kittens: 12\\n   - Patchy's kittens: 36\\n   - Total kittens from the cats = 12 + 36 = 48.\\n\\n4. **Account for the kittens adopted from the shelter:**\\n   - The family is driving home with 7 kittens adopted from the local animal shelter.\\n   - Total kittens = 48 (from the cats) + 7 (adopted) = 55.\\n\\n5. **Determine the total number of kittens the family now has:**\\n   - The question asks for the total number of kittens the family now has.\\n   - The 7 kittens from the shelter are kittens, and the 48 kittens born to the cats are also kittens.\\n   - Total kittens = 55.\\n\\n#### 55\"",
            "Hold taken: uptime_s=729901.57 memavail_kb=6129188 thermal=0 battery_dC=267",
            "10-24 09:35:04.483 20880 20947 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "10-24 09:35:12.165 20880 20947 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "10-24 09:35:12.788 20880 20947 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "10-24 09:35:13.355 20880 20947 I tflite  : Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answer_matches_gsm8k": 0.0,
            "app_chunks_per_s": 16.72,
            "chunks": 303.0,
            "final_value": 55.0,
            "first_token_ms": 1037.5,
            "gsm8k_answer": 40.0,
            "load_ms": 13562.0,
            "screen_recording": 0.0,
            "thermal_status_end": 0.0,
            "thermal_status_start": 0.0,
            "total_ms": 19160.1
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "Android app (hfmodels-android samples/loops, debug build); gsm8k-54 (GSM8K test question 54, answer 40), thinking off, the one-loop file then the int4 file, screen recording off",
          "total_ops": 3998,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.16.1/2026-10-24/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-24",
          "decode_tokens_per_s": null,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory-loops-app",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.16.1",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: hfmodels-android samples/loops (main e93b645; the APK built from 034366c, whose difference to e93b645 is the sample's README only), hfmodels SDK 0.2.1-SNAPSHOT over litertlm-android 0.16.1; Run with 'Both' = the one-loop file, then the int4 file, one model loaded at a time, one conversation per answer, thinking off, the engine's default sampler (greedy); total_ms / first_token_ms = walls from send to the last / first streamed chunk, app_chunks_per_s = (chunks - 1) / (last - first chunk) as the app prints it ('tok/s' on the screen; tokenizer counts are 1.09-1.60 times the chunks, the sample's README)",
            "int4_1loop final_value 55 (mismatch, GSM8K answer 40), total_ms 18813.9, first_token_ms 1030, chunks 303, tok_s 17.04, load_ms 12783, thermal 0 -> 0; runtime_version 0.16.1, sdk 0.2.1-SNAPSHOT, device SM-S942Q BP4A.251205.006.S942QOPS1AZH9",
            "10-24 09:43:46.171 24447 24447 I loops   : ANSWER loop=1 variant=int4_1loop q=gsm8k-54 first_token_ms=1030.0 tok_s=17.04 total_ms=18813.9 chunks=303 thought_chars=0 load_ms=12783 backend=GPU mark=mismatch text=\"**Step-by-step reasoning:**\\n\\n1. **Identify the existing cats and their current kittens:**\\n   - The family already has two adult cats: Patchy and Trixie.\\n   - Trixie has had 12 kittens.\\n\\n2. **Calculate Patchy's kittens:**\\n   - The problem states Patchy has had thrice the number of kittens that Trixie has.\\n   - Trixie's kittens = 12.\\n   - Patchy's kittens = 3 × 12 = 36.\\n\\n3. **Calculate the total number of kittens from the cats:**\\n   - Trixie's kittens: 12\\n   - Patchy's kittens: 36\\n   - Total kittens from the cats = 12 + 36 = 48.\\n\\n4. **Account for the kittens adopted from the shelter:**\\n   - The family is driving home with 7 kittens adopted from the local animal shelter.\\n   - Total kittens = 48 (from the cats) + 7 (adopted) = 55.\\n\\n5. **Determine the total number of kittens the family now has:**\\n   - The question asks for the total number of kittens the family now has.\\n   - The 7 kittens from the shelter are kittens, and the 48 kittens born to the cats are also kittens.\\n   - Total kittens = 55.\\n\\n#### 55\"",
            "Hold taken: uptime_s=730395.17 memavail_kb=6840184 thermal=0 battery_dC=302",
            "10-24 09:43:17.040 24447 24484 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "10-24 09:43:24.326 24447 24484 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "10-24 09:43:24.909 24447 24484 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "10-24 09:43:25.419 24447 24484 I tflite  : Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "screen recording: the phone's screenrecord at 12 Mbit/s ran through the run (the clip was cut from it)"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answer_matches_gsm8k": 0.0,
            "app_chunks_per_s": 17.04,
            "chunks": 303.0,
            "final_value": 55.0,
            "first_token_ms": 1030.0,
            "gsm8k_answer": 40.0,
            "load_ms": 12783.0,
            "screen_recording": 1.0,
            "thermal_status_end": 0.0,
            "thermal_status_start": 0.0,
            "total_ms": 18813.9
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "Android app (hfmodels-android samples/loops, debug build); gsm8k-54 (GSM8K test question 54, answer 40), thinking off, the one-loop file then the int4 file, screen recording on",
          "total_ops": 3998,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.16.1/2026-10-24/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-26",
          "decode_tokens_per_s": null,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory-loops-app",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.16.1",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: hfmodels-android samples/loops (the APK built from b4ab57f, the commit that ships GSM8K test question 75 among the examples; main a582f5c adds only the sample's README and results to it), hfmodels SDK 0.2.1-SNAPSHOT over litertlm-android 0.16.1, both files pushed with adb and imported after the SDK's sha256 check; one model loaded at a time, one conversation per answer, thinking off, the engine's default sampler (greedy); total_ms / first_token_ms = walls from send to the last / first streamed chunk, app_chunks_per_s = the rate the app prints ('tok/s' on the screen; chunks are not tokens, the sample's README); load_ms = the file's load in that run",
            "int4_1loop gsm8k-75 final_value 7.5 (mismatch, GSM8K answer 60), total_ms 18544, first_token_ms 725.6, chunks 264, tok_s 14.81, load_ms 15041, thermal 0 -> 1; files in the run ['int4_1loop', 'int4'], questions 1; runtime_version 0.16.1, sdk 0.2.1-SNAPSHOT, device SM-S942Q BP4A.251205.006.S942QOPS1AZH9",
            "10-26 14:20:43.377 23967 23967 I loops   : ANSWER loop=1 variant=int4_1loop q=gsm8k-75 first_token_ms=725.6 tok_s=14.81 total_ms=18544.0 chunks=264 thought_chars=0 load_ms=15041 backend=GPU mark=mismatch text=\"To find the average square foo…",
            "2026-10-26 14:12:16 hold taken: uptime_s=919469.73 memavail_kb=7640420 thermal=2 battery_dC=416",
            "10-26 14:20:13.779 23967 24003 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "10-26 14:20:21.563 23967 24003 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "10-26 14:20:22.159 23967 24003 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "10-26 14:20:22.662 23967 24003 I tflite  : Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "hfmodels-android origin/main a582f5c3 samples/loops/results/2026-10-26-s26/t1.json and .log = / .log byte for byte",
            "screen recording: the phone's screenrecord ran through the run (the demo clip was cut from it; the clip's times checked against this file's walls)"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answer_matches_gsm8k": 0.0,
            "app_chunks_per_s": 14.81,
            "chunks": 264.0,
            "file_order_in_run": 1.0,
            "final_value": 7.5,
            "first_token_ms": 725.6,
            "gsm8k_answer": 60.0,
            "load_ms": 15041.0,
            "question_position": 1.0,
            "screen_recording": 1.0,
            "thermal_status_end": 1.0,
            "thermal_status_start": 0.0,
            "total_ms": 18544.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "Android app (hfmodels-android samples/loops, debug build); gsm8k-75 (GSM8K test question 75, answer 60), thinking off, gsm8k-75 alone, picked from the Examples list with 'Both' (the one-loop file, then the int4 file), run t1, screen recording on",
          "total_ops": 3998,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.16.1/2026-10-26/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-26",
          "decode_tokens_per_s": null,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory-loops-app",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.16.1",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: hfmodels-android samples/loops (the APK built from b4ab57f, the commit that ships GSM8K test question 75 among the examples; main a582f5c adds only the sample's README and results to it), hfmodels SDK 0.2.1-SNAPSHOT over litertlm-android 0.16.1, both files pushed with adb and imported after the SDK's sha256 check; one model loaded at a time, one conversation per answer, thinking off, the engine's default sampler (greedy); total_ms / first_token_ms = walls from send to the last / first streamed chunk, app_chunks_per_s = the rate the app prints ('tok/s' on the screen; chunks are not tokens, the sample's README); load_ms = the file's load in that run",
            "int4_1loop gsm8k-75 final_value 7.5 (mismatch, GSM8K answer 60), total_ms 18177.3, first_token_ms 686.2, chunks 264, tok_s 15.09, load_ms 13511, thermal 0 -> 1; files in the run ['int4_1loop', 'int4'], questions 1; runtime_version 0.16.1, sdk 0.2.1-SNAPSHOT, device SM-S942Q BP4A.251205.006.S942QOPS1AZH9",
            "10-26 14:11:06.698 19155 19155 I loops   : ANSWER loop=1 variant=int4_1loop q=gsm8k-75 first_token_ms=686.2 tok_s=15.09 total_ms=18177.3 chunks=264 thought_chars=0 load_ms=13511 backend=GPU mark=mismatch text=\"To find the average square foo…",
            "2026-10-26 14:03:44 hold taken: uptime_s=918957.98 memavail_kb=8189648 thermal=1 battery_dC=399",
            "10-26 14:10:37.526 19155 19188 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "10-26 14:10:45.326 19155 19188 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "10-26 14:10:45.922 19155 19188 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "10-26 14:10:46.444 19155 19188 I tflite  : Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answer_matches_gsm8k": 0.0,
            "app_chunks_per_s": 15.09,
            "chunks": 264.0,
            "file_order_in_run": 1.0,
            "final_value": 7.5,
            "first_token_ms": 686.2,
            "gsm8k_answer": 60.0,
            "load_ms": 13511.0,
            "question_position": 1.0,
            "screen_recording": 0.0,
            "thermal_status_end": 1.0,
            "thermal_status_start": 0.0,
            "total_ms": 18177.3
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "Android app (hfmodels-android samples/loops, debug build); gsm8k-75 (GSM8K test question 75, answer 60), thinking off, gsm8k-75 alone, picked from the Examples list with 'Both' (the one-loop file, then the int4 file), run u2, screen recording off",
          "total_ops": 3998,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.16.1/2026-10-26/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-26",
          "decode_tokens_per_s": null,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory-loops-app",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.16.1",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: hfmodels-android samples/loops (the APK built from b4ab57f, the commit that ships GSM8K test question 75 among the examples; main a582f5c adds only the sample's README and results to it), hfmodels SDK 0.2.1-SNAPSHOT over litertlm-android 0.16.1, both files pushed with adb and imported after the SDK's sha256 check; one model loaded at a time, one conversation per answer, thinking off, the engine's default sampler (greedy); total_ms / first_token_ms = walls from send to the last / first streamed chunk, app_chunks_per_s = the rate the app prints ('tok/s' on the screen; chunks are not tokens, the sample's README); load_ms = the file's load in that run",
            "int4_1loop gsm8k-75 final_value 7.5 (mismatch, GSM8K answer 60), total_ms 30798.7, first_token_ms 982.2, chunks 264, tok_s 8.85, load_ms 26776, thermal 2 -> 2; files in the run ['int4', 'int4_1loop'], questions 8; runtime_version 0.16.1, sdk 0.2.1-SNAPSHOT, device SM-S942Q BP4A.251205.006.S942QOPS1AZH9",
            "10-26 13:11:00.556 13427 13427 I loops   : ANSWER loop=1 variant=int4_1loop q=gsm8k-75 first_token_ms=982.2 tok_s=8.85 total_ms=30798.7 chunks=264 thought_chars=0 load_ms=none backend=GPU mark=mismatch text=\"To find the average square foota…",
            "2026-10-26 12:53:33 hold taken: uptime_s=914746.99 memavail_kb=7545512 thermal=3 battery_dC=440",
            "10-26 13:09:36.246 13427 13461 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "10-26 13:09:56.772 13427 13461 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "10-26 13:09:57.347 13427 13461 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "10-26 13:09:57.958 13427 13461 I tflite  : Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "hfmodels-android origin/main a582f5c3 samples/loops/results/2026-10-26-s26/b2.json and .log = / .log byte for byte",
            "hfmodels-android origin/main samples/loops/README.md (2026-10-26, each file's first-run pass): | `gsm8k-75` | `#### 7.5` (GSM8K's is 60) | 22.9 | 11.8 | `#### 60` = GSM8K's | 53.7 | 6.1 |"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answer_matches_gsm8k": 0.0,
            "app_chunks_per_s": 8.85,
            "chunks": 264.0,
            "file_order_in_run": 2.0,
            "final_value": 7.5,
            "first_token_ms": 982.2,
            "gsm8k_answer": 60.0,
            "load_ms": 26776.0,
            "question_position": 6.0,
            "screen_recording": 0.0,
            "thermal_status_end": 2.0,
            "thermal_status_start": 2.0,
            "total_ms": 30798.7
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "Android app (hfmodels-android samples/loops, debug build); gsm8k-75 (GSM8K test question 75, answer 60), thinking off, the sample's eight bundled examples with the int4 file first (this row = gsm8k-75, the sixth), run b2, screen recording off",
          "total_ops": 3998,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.16.1/2026-10-26/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-26",
          "decode_tokens_per_s": null,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory-loops-app",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.16.1",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: hfmodels-android samples/loops (the APK built from b4ab57f, the commit that ships GSM8K test question 75 among the examples; main a582f5c adds only the sample's README and results to it), hfmodels SDK 0.2.1-SNAPSHOT over litertlm-android 0.16.1, both files pushed with adb and imported after the SDK's sha256 check; one model loaded at a time, one conversation per answer, thinking off, the engine's default sampler (greedy); total_ms / first_token_ms = walls from send to the last / first streamed chunk, app_chunks_per_s = the rate the app prints ('tok/s' on the screen; chunks are not tokens, the sample's README); load_ms = the file's load in that run",
            "int4_1loop gsm8k-75 final_value 7.5 (mismatch, GSM8K answer 60), total_ms 22893.1, first_token_ms 559.6, chunks 264, tok_s 11.82, load_ms 13823, thermal 1 -> 1; files in the run ['int4_1loop', 'int4'], questions 8; runtime_version 0.16.1, sdk 0.2.1-SNAPSHOT, device SM-S942Q BP4A.251205.006.S942QOPS1AZH9",
            "10-26 12:59:34.742  9812  9812 I loops   : ANSWER loop=1 variant=int4_1loop q=gsm8k-75 first_token_ms=559.6 tok_s=11.82 total_ms=22893.1 chunks=264 thought_chars=0 load_ms=none backend=GPU mark=mismatch text=\"To find the average square foot…",
            "2026-10-26 12:53:33 hold taken: uptime_s=914746.99 memavail_kb=7545512 thermal=3 battery_dC=440",
            "10-26 12:58:42.070  9812  9859 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "10-26 12:58:49.731  9812  9859 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "10-26 12:58:50.324  9812  9859 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "10-26 12:58:50.834  9812  9859 I tflite  : Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "hfmodels-android origin/main a582f5c3 samples/loops/results/2026-10-26-s26/b1.json and .log = / .log byte for byte",
            "hfmodels-android origin/main samples/loops/README.md (2026-10-26, each file's first-run pass): | `gsm8k-75` | `#### 7.5` (GSM8K's is 60) | 22.9 | 11.8 | `#### 60` = GSM8K's | 53.7 | 6.1 |"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answer_matches_gsm8k": 0.0,
            "app_chunks_per_s": 11.82,
            "chunks": 264.0,
            "file_order_in_run": 1.0,
            "final_value": 7.5,
            "first_token_ms": 559.6,
            "gsm8k_answer": 60.0,
            "load_ms": 13823.0,
            "question_position": 6.0,
            "screen_recording": 0.0,
            "thermal_status_end": 1.0,
            "thermal_status_start": 1.0,
            "total_ms": 22893.1
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "Android app (hfmodels-android samples/loops, debug build); gsm8k-75 (GSM8K test question 75, answer 60), thinking off, the sample's eight bundled examples with the one-loop file first (this row = gsm8k-75, the sixth), run b1, screen recording off",
          "total_ops": 3998,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.16.1/2026-10-26/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-23",
          "decode_tokens_per_s": null,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory-chat-check",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.16.1",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "harness: hfmodels-android samples/chat ChatDeviceCheck (androidTest), hfmodels SDK 0.2.1-SNAPSHOT over com.google.ai.edge.litertlm:litertlm-android 0.16.1 (the SDK's pin), debug build; one model load and ONE conversation per run, the descriptor's thinking default (on), the engine's default sampler (the bundle declares none: greedy); load_ms = the wall of fromPretrained (an adb-pushed copy imported after its sha256 check), first_chunk_ms / total_ms = the walls from send to the first and the last streamed chunk; chunks are the streamed pieces, not tokens (the SDK states no token count); hfmodels-android origin/main e93b645a litertlm/results/2026-10-23-RFGL80R6A6H-litertlm0.16.1-dory-1loop-chat-check.log: # hfmodels on litertlm-android 0.16.1 (at commit edc8fbf): samples/chat's ChatDeviceCheck with litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B, variant int4_1loop on gpu (17 + 25), log_thought=true / # serial=RFGL80R6A6H device=SM-S942Q (Galaxy S26) build=BP4A.251205.006.S942QOPS1AZH9 android=16; chat-debug.apk sha256 d5af1b25…, chat-debug-androidTest.apk sha256 2345a28b…; the int4_1loop file pushed with adb into the app's external files dir before the check (the SDK checked its sha256 and imported it); the run's start / end line (uptime, MemAvailable, thermal status, free /data)",
            "litertlm/results/2026-10-23-RFGL80R6A6H-litertlm0.16.1-dory-1loop-chat-check.log run s1 (check litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B gpu int4_1loop): start 2026-10-23 18:14:34, end 2026-10-23 18:15:03, rc 0",
            "10-23 18:14:35.444 27504 27520 I hfmodels: resolve litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B: commit=ff0f6e32 via RESOLVED_BRANCH, descriptor=john-rocky/hfmodels-android@63d6bd40/catalog/entries/litert-community__NVIDIA-Nemotron-Labs-Dory1.0-4B.json sha=35ec8a84",
            "10-23 18:14:37.637 27504 27520 I hfmodels: imported NVIDIA-Nemotron-Labs-Dory1.0-4B_int4_1loop.litertlm: 2400558400 bytes, sha256 ok, from /storage/emulated/0/Android/data/io.github.johnrocky.hfmodels.samples.chat/files/NVIDIA-Nemotron-Labs-Dory1.0-4B_int4_1loop.litertlm",
            "10-23 18:14:53.896 27504 27517 I hfmodels: ready litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B@ff0f6e32 variant=int4_1loop profile=gpu language=GPU prepare_ms=16253",
            "10-23 18:14:53.897 27504 27517 I hfmodels-check: RESULT step=load ok=true model=litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B variant=int4_1loop profile=gpu initialized=[language=GPU] thinking=descriptor load_ms=18939 last_event=Ready",
            "10-23 18:15:01.156 27504 27517 I hfmodels-check: RESULT step=answer ok=true model=litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B expect=\"42\" chunks=108 thought_chunks=106 first_chunk_ms=775 total_ms=7256 state=READY thought_chars=453 reply=\"42\"",
            "10-23 18:15:02.477 27504 27517 I hfmodels-check: RESULT step=stop ok=true model=litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B chunks_at_stop=5 state=INVALID",
            "10-23 18:15:02.847 27504 27517 I hfmodels-check: RESULT step=release ok=true model=litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B close_ms=367",
            "10-23 18:15:02.847 27504 27517 I hfmodels-check: RESULT ok=true model=litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B failed=[] sdk=0.2.1-SNAPSHOT runtime=litert_lm 0.16.1 device=SM-S942Q build=BP4A.251205.006.S942QOPS1AZH9",
            "10-23 18:14:38.191 27504 27520 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "10-23 18:14:49.237 27504 27520 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "10-23 18:14:50.272 27504 27520 I tflite  : Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "10-23 18:14:51.071 27504 27520 I tflite  : Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "10-23 18:14:53.883 27504 27520 I tflite  : Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder)."
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answer_matches_expect": 1.0,
            "chunks": 108.0,
            "close_ms": 367.0,
            "first_chunk_ms": 775.0,
            "load_ms": 18939.0,
            "memavailable_end_kb": 7139528.0,
            "memavailable_start_kb": 6986676.0,
            "thermal_status_end": 1.0,
            "thermal_status_start": 1.0,
            "thought_chars": 453.0,
            "thought_chunks": 106.0,
            "total_ms": 7256.0,
            "uptime_end_s": 674836.67,
            "uptime_start_s": 674807.73
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "Android library (hfmodels-android samples/chat ChatDeviceCheck, debug build); one conversation, 'What is 17 + 25? Answer briefly.' (expect '42')",
          "total_ops": 3998,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.16.1/2026-10-23/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-23",
          "decode_tokens_per_s": 17.3225,
          "delegated_ops": 3998,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-dory",
            "os_build": "Android 16 (BP4A.251205.006)",
            "runtime": "litert-lm",
            "runtime_version": "0.18.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: LiteRT-LM v0.18.0 litert_lm_advanced_main sha256 cf3636070a8e51b0ea13d9a4a8b26fc9a2f9aa38f9e1c7687fe009ce7550f815 + 7 .so (the edge-llm-bench source build of the v0.18.0 tag, 2026-10-07, android/engine-pins.json v0.18.0), every file sha256-checked on the phone at each hold's start (s0_kit_check.txt: 8 OK)",
            "hold s26-h2 (r17): HOLD_SEEN 2026-10-23 07:03:29: acquired 07:03:29, skip_after 07:25:29, run_end 07:29:29, released by 07:31:29; keeper released the hold at 2026-10-23 07:11:05",
            "hold s26-h3 (r17): HOLD_SEEN 2026-10-23 07:11:20: acquired 07:11:18, skip_after 07:33:18, run_end 07:37:18, released by 07:39:18; keeper released the hold at 2026-10-23 07:20:18",
            "(s26-h2, 1792706996 device epoch, ready=clean: clean; rested 172s since 07:07:03; thermal status 0 -> 2): iterations 2-3 prefill 455.91 / 426.58, decode 17.71 / 17.05 tok/s, TTFT 2.3 / 2.46 s, Init Total 11976.68 ms (cold); .adb.txt EXIT=0 WALL=63 START=1792706996 END=1792707059; .meta.json memory 'VmHWM_max_kB 1515236 VmSwap_max_kB 133220 MemAvailable_min_kB 3240156 kgsl_page_alloc_max_B 4516253696'; sampler 29 samples, CPU capped in 20, GPU capped in 19",
            "(s26-h3, 1792707261 device epoch, ready=clean: clean; rested 181s since 07:11:20; thermal status 0 -> 2): iterations 2-3 prefill 447.48 / 428.44, decode 17.71 / 16.82 tok/s, TTFT 2.34 / 2.45 s, Init Total 11602.0 ms (cold); .adb.txt EXIT=0 WALL=64 START=1792707261 END=1792707325; .meta.json memory 'VmHWM_max_kB 1516572 VmSwap_max_kB 89716 MemAvailable_min_kB 2974708 kgsl_page_alloc_max_B 4518596608'; sampler 29 samples, CPU capped in 20, GPU capped in 19",
            "command: --backend=gpu --model_path=/data/local/tmp/dory_r17/f4cl1.litertlm --input_prompt_file=/data/local/tmp/dory_r17/prompts/q1.txt --max_num_tokens=2048 --max_output_tokens=2048 --async=false --benchmark --benchmark_prefill_tokens=0 --benchmark_decode_tokens=0 --use_session=false --use_session=true --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256 --num_iterations=3",
            "activation_data_type: Not set",
            "Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).",
            "Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "Replacing 3998 out of 3998 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_1024).",
            "Replacing 3621 out of 3621 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (decode).",
            "Replacing 1 out of 4 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (decode_embedder).",
            "Device: model=SM-S942Q release=16 build_id=BP4A.251205.006 fingerprint=samsung/m1qjpnx/m1q:16/BP4A.251205.006/S942QOPS1AZH9_SJP1AZH9:user/release-keys"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 256.0,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "prefill_tokens": 1024.0,
            "run1_cold_start": 1.0,
            "run1_cpu_capped_samples": 20.0,
            "run1_decode_tok_s": 17.38,
            "run1_decode_tok_s_it1": 17.85,
            "run1_decode_tok_s_it2": 17.71,
            "run1_decode_tok_s_it3": 17.05,
            "run1_gpu_capped_samples": 19.0,
            "run1_init_total_ms": 11976.68,
            "run1_kgsl_page_alloc_max_b": 4516253696.0,
            "run1_memavail_min_kb": 3240156.0,
            "run1_prefill_tok_s": 441.245,
            "run1_prefill_tok_s_it1": 451.8,
            "run1_prefill_tok_s_it2": 455.91,
            "run1_prefill_tok_s_it3": 426.58,
            "run1_sampler_samples": 29.0,
            "run1_t_s": 17.0503,
            "run1_thermal_status_end": 2.0,
            "run1_thermal_status_start": 0.0,
            "run1_ttft_s": 2.38,
            "run1_vmhwm_kb": 1515236.0,
            "run1_vmswap_max_kb": 133220.0,
            "run2_cold_start": 1.0,
            "run2_cpu_capped_samples": 20.0,
            "run2_decode_tok_s": 17.265,
            "run2_decode_tok_s_it1": 17.79,
            "run2_decode_tok_s_it2": 17.71,
            "run2_decode_tok_s_it3": 16.82,
            "run2_gpu_capped_samples": 19.0,
            "run2_init_total_ms": 11602.0,
            "run2_kgsl_page_alloc_max_b": 4518596608.0,
            "run2_memavail_min_kb": 2974708.0,
            "run2_prefill_tok_s": 437.96,
            "run2_prefill_tok_s_it1": 451.91,
            "run2_prefill_tok_s_it2": 447.48,
            "run2_prefill_tok_s_it3": 428.44,
            "run2_sampler_samples": 29.0,
            "run2_t_s": 17.1658,
            "run2_thermal_status_end": 2.0,
            "run2_thermal_status_start": 0.0,
            "run2_ttft_s": 2.395,
            "run2_vmhwm_kb": 1516572.0,
            "run2_vmswap_max_kb": 89716.0,
            "runs": 2.0,
            "t_1024_256_s": 17.1078
          },
          "output_match": null,
          "peak_mem_mb": 1552.286,
          "prefill_tokens_per_s": 439.6025,
          "provenance": "measured",
          "runs": true,
          "total_ops": 3998,
          "ttft_ms": 2387.5
        },
        "source": "data/device_runs/0.18.0/2026-10-23/nemotron-labs-dory1.0-4b-int4-1loop__galaxy-s26.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-23",
          "decode_tokens_per_s": 88.175,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0",
            "runtime": "litert-lm",
            "runtime_version": "0.18.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "Settle: quiet after 0 s; warm-up + 3 runs; prefill 1617.72 / decode 88.1 tok/s, TTFT 0.6443 s, init 4.8756 s (the CLI prints the mean of the 3 runs; the log's R17_IT lines carry each run); maximum resident set size 3098083328 B, peak memory footprint 3665857632 B; 0 'Validation error' lines",
            "2026-10-23 06:41:47 [w5] START l6_gpu_b1 cap 320s (disk 265 GB free) / 2026-10-23 06:42:14 [w5] DONE l6_gpu_b1 rc=0 27s",
            "Settle: quiet after 0 s; warm-up + 3 runs; prefill 1621.5 / decode 88.25 tok/s, TTFT 0.6428 s, init 4.899 s (the CLI prints the mean of the 3 runs; the log's R17_IT lines carry each run); maximum resident set size 3088449536 B, peak memory footprint 3672083552 B; 0 'Validation error' lines",
            "2026-10-23 06:42:14 [w5] START l6_gpu_b2 cap 320s (disk 265 GB free) / 2026-10-23 06:42:38 [w5] DONE l6_gpu_b2 rc=0 24s",
            "2026-10-23 06:43:17 [w5] JUDGE R17_C l6_gpu: card (it 2-3) A F4c 1034.9968 / 58.5142 tok/s TTFT 1.0065 s, B F4c-L1 1618.885 / 88.1338 tok/s TTFT 0.6439 s, decode B/A 1.5062, prefill B/A 1.5641; CLI mean (it 1-3) decode B/A 1.5084"
          ],
          "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": 4.8873,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "prefill_tokens": 1024.0,
            "run1_decode_tok_s": 88.1,
            "run1_init_s": 4.8756,
            "run1_max_rss_mb": 3098.083,
            "run1_peak_footprint_mb": 3665.858,
            "run1_prefill_tok_s": 1617.72,
            "run1_t_s": 3.5388,
            "run1_ttft_s": 0.6443,
            "run2_decode_tok_s": 88.25,
            "run2_init_s": 4.899,
            "run2_max_rss_mb": 3088.45,
            "run2_peak_footprint_mb": 3672.084,
            "run2_prefill_tok_s": 1621.5,
            "run2_t_s": 3.5324,
            "run2_ttft_s": 0.6428,
            "runs": 2.0,
            "t_1024_256_s": 3.5356,
            "warmup_runs": 1.0
          },
          "output_match": null,
          "peak_mem_mb": 3668.971,
          "prefill_tokens_per_s": 1619.61,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 643.6
        },
        "source": "data/device_runs/0.18.0/2026-10-23/nemotron-labs-dory1.0-4b-int4-1loop__mac-studio-m4-max.json"
      }
    ]
  },
  "model": {
    "family": "dory",
    "id": "nemotron-labs-dory1.0-4b-int4-1loop",
    "license": "other (openmdw-1.1)",
    "source_url": "https://huggingface.co/litert-community/NVIDIA-Nemotron-Labs-Dory1.0-4B",
    "task": "text-generation"
  },
  "pitfalls": [
    "One loop instead of two: the same int4 weights with the shared block of layers applied once per token; on GSM8K (questions 0-99, thinking off, Mac GPU) 86 / 100 against the int4 file's 90, the two files disagreeing on 10 questions (7 right with two loops only, 3 with one loop only); the BF16 reference shows the trade too: 86 with one loop against 95 with two, 10 questions right with two loops only and 1 with one loop only, and on question 75 it also answers wrong with one loop and right with two; NVIDIA recommends two loops and the card names the int4 file as the one to start with (card L36, L63, L81, L83, L85).",
    "Speed: in runs that alternated the two int4 files (2026-10-23), decode 88.18 against 58.46 tok/s on the Mac M4 Max GPU (x 1.51) and 17.32 against 11.08 on the Galaxy S26 GPU (x 1.56), prefill x 1.57 / x 1.60, and 4.52 against 5.38 GB of GPU memory on the Galaxy S26 (card L78-L80).",
    "Near-tie answers move with the device: asked the rhyme question ('Roses are red, violets are ___'), the file answered 'red' on the Mac GPU (with fp32 activations too), 'purple' on the Mac CPU and 'blue' on the Galaxy S26 GPU; the PyTorch reference with the same one-loop int4 weights answers 'blue', and along the Mac GPU's fp32-activation text it picks the same next token at 932 of 935 steps, the first difference at the 39th token of the thought where its own two best tokens were 0.016 apart (card L52, L83).",
    "GPU only: the card measures the file on GPUs (Mac WebGPU on Metal, the Galaxy S26 OpenCL delegate with every node of the four graphs on it); the Pixel 8a row reads not run, and the int8 file is the CPU file (card L36, L42, L52).",
    "fp16 activations: the same MLP normalization rewrite covers the one-loop graph: in an fp16 emulation over the same 25 prompts no row overflows with the rewrite, and without it every overflowing row is in one of the 11 rewritten blocks (card L136).",
    "Phone heat: both Galaxy S26 GPU runs of the file (2026-10-23) started cool (thermal status 0) and ended at status 2, the GPU clock under its cap in 19 of 29 samples each (card L122).",
    "Thinking and template: thinking is on by default (the generation prompt ends with <think> and a newline, the bundle declares the thought channel); rendered with enable_thinking false the prompt ends with <think></think>; NVIDIA recommends temperature 1.0 and top-p 0.95, the bundle sets no sampler; stop tokens </s> and <|im_end|> plus 14 stop sequences; model type generic_model (tool calls not parsed; tool calling not tested) (card L97).",
    "Final mark on LiteRT-LM 0.16.1 and earlier: of the 14 stop sequences, 13 put one of 13 characters (12 punctuation marks and a space) in front of <|im_end|> and a newline; an answer whose last mark is a token of its own, such as a ? or . after a word, arrives without that mark on LiteRT-LM 0.16.1 (the runtime of the Android app under Loops) and on 0.11.0 to 0.16.0 (source read): the runtime holds the mark back, the answer ends on <|im_end|> alone and the held mark is never written; a mark the tokenizer merges with the text before it (**., ).) is kept; 0.17.0 and later write the held mark (card L97).",
    "Second turn with thinking off: from the second turn of a conversation on, LiteRT-LM puts the previous answer's stop token into the KV cache with a zero embedding, because the file holds the embedding table in its own section and the decoder takes embeddings (read in the source of 0.16.0 to 0.18.1; measured with 0.18.1 on WeVisOne-VL-2B and Qwen2.5-Coder-1.5B-Instruct, not on these files), so later answers can differ from the answer to the same messages in one prompt; with thinking on the runtime rebuilds the cache at each turn, and the source shows a zero embedding at another token from the third turn on (not measured); one question per conversation, or a new conversation that starts with the earlier messages and the new question, avoids both (card L101).",
    "Conversation preface with tools and no messages: the template reads messages[0] without checking that the list is non-empty, so LiteRT-LM's render fails with 'Failed to apply template: undefined value' when the preface is prefilled at creation (prefillPrefaceOnInit, default false) and in the preface render (renderPrefaceIntoString and its C++ and C API forms); the same with extra context (such as enable_thinking) and no messages; read in the source of 0.17.0, 0.18.1 and main, measured with the 0.18.1 PyPI package through its C API on these files' template text; the default path, where the first message is rendered with the preface, is not affected; a system message or any initial message in the conversation config avoids it (card L141).",
    "Cache: without --cache no, the CLI can write a compiled cache file next to the model on its first run; without --backend, `run` uses the CPU (card L95)."
  ],
  "schema_version": "1.2"
}
