{
  "artifacts": [
    {
      "file": "nemotron3_nano_30b_a3b_mixed_int4_ekv2048.litertlm",
      "sha256": "b0bd9a61f9c168197f9c1135762f14d56c39e7f3feb3d71ef8f4b904ed5923f4",
      "size_mb": 19670.354
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "TODO (the card gives no conversion command); source nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16, quantized (card L13-L14, L27)",
    "quantization": "experts int4 in blocks of 32 values; the other weight matrices int8; router fp32; fp32 activations declared; a 2,048-token context; 20.6 GB (card L37, L40)",
    "tool": "TODO (the card names no converter). The graph, as the card states it: each layer's experts are one `moe` custom op; relu² is written as its `gelu_tanh` (the activation was checked through the logits); each `moe` output has a second consumer (`moe * 0`) so ML Drift does not fuse the next op into it (card L40)",
    "tool_version": "TODO (the converter's version is not stated); runtime LiteRT-LM 0.19.1: the CPU through the PyPI package, the GPU through the v0.19.1 command-line tool built from source (7a03724c) (card L27, L76-L80)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-10",
          "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": [
            "bytes: these runs read as the bundle step wrote it (done 2026-11-10 13:59:31 JST; the build log at 13:59:51 gives its sha256 b0bd9a61... at the build), hashed by the pre-upload after them (done 2026-11-10 20:39:53 JST, sha256 b0bd9a61f9c168197f9c1135762f14d56c39e7f3feb3d71ef8f4b904ed5923f4, 20625860800 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit ac7bc38a (the commit's readback)",
            "runtime: the PyPI litert-lm 0.19.1 macOS arm64 wheel's Python API (tools/line3_rt.py; each record names litert-lm 0.19.1), CPU (XNNPACK) 8 threads, one Engine per process and one Conversation with the file's own template, SamplerConfig(top_k=1), ThinkingConfig(enable_thinking=False), max_output_tokens 2,048; the 3 turns in one conversation (the runtime keeps the history); the file's CPU runs used one shared engine cache directory (tools/line3_chain.sh)",
            "output_match = line 3 (a) and (b) as written before any output: (a) the file's verdict equals the reference's on 12 of 12 (at least 11 required; both answer 'violet' to the rhyme question, so 11 are right), (b) every question the reference ends with a stop token ends with one here (12 / 12). Reference = the float32 implementation of the model on the bf16 weights, thinking off, greedy, its replies built by draft-and-verify (every appended token its argmax), held to NVIDIA's modeling_nemotron_h.py by lines 0 and R",
            "the 3-turn conversation: 'Kenji', then '401', all three ending with a stop token; the reference's own turns: 'Kenji', then '401'. Line 3 (c) as written asks for 391 in turn 3 (not met); line 3' (written after these outputs were seen: a turn the reference gets wrong is not required of the file) is met; in metrics, not part of output_match",
            "(cpu_off: 2026-11-10 14:41:12-15:24:06 JST (the CPU step's start to the GPU step's)): reply tokens up to 351; = MEASUREMENTS.md's Accuracy row for 'mixed_int4 ekv2048' and the Hub card's Accuracy table at main 5792549c"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answers_right": 11.0,
            "conversation_c_as_written": 0.0,
            "conversation_c_line3_prime": 1.0,
            "conversation_turn2_name_right": 1.0,
            "conversation_turn3_answer": 401.0,
            "conversation_turns": 3.0,
            "conversation_turns_ended_with_stop_token": 3.0,
            "cpu_threads": 8.0,
            "ended_with_stop_token": 12.0,
            "max_new_tokens": 2048.0,
            "questions": 12.0,
            "reference_ended_with_stop_token": 12.0,
            "reference_turn3_answer": 401.0,
            "reply_tokens_max": 351.0,
            "verdicts_equal_reference": 12.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the 12 check questions of line 3 (the 8 short questions of hf-to-litertlm's scripts/verify_quality.py, a price word problem, a syllogism, an average speed, a question in Japanese) and its 3-turn conversation, thinking off (enable_thinking false), greedy (top-k 1), at most 2,048 new tokens, one process per question; verdicts against the source model's replies (float32 on the bf16 weights, thinking off)",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-10/nemotron-3-nano-30b-a3b-mixed-int4-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-10",
          "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": [
            "bytes: these runs read as the bundle step wrote it (done 2026-11-10 13:59:31 JST; the build log at 13:59:51 gives its sha256 b0bd9a61... at the build), hashed by the pre-upload after them (done 2026-11-10 20:39:53 JST, sha256 b0bd9a61f9c168197f9c1135762f14d56c39e7f3feb3d71ef8f4b904ed5923f4, 20625860800 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit ac7bc38a (the commit's readback)",
            "runtime: the PyPI litert-lm 0.19.1 macOS arm64 wheel's Python API (tools/line3_rt.py; each record names litert-lm 0.19.1), CPU (XNNPACK) 8 threads, one Engine per process and one Conversation with the file's own template, SamplerConfig(top_k=1), ThinkingConfig(enable_thinking=True), max_output_tokens 2,048; the 3 turns in one conversation (the runtime keeps the history); the file's CPU runs used one shared engine cache directory (tools/line3_chain.sh)",
            "output_match = line 3's rule for thinking on, written before any output (the fallback taken on 2026-11-10 because the reference generator ran at 77.8 s per step, so no reference ran with thinking on): the line's regex right on at least 11 of the 12 and every reply ending with a stop token within 2,048 new tokens: 12 / 12 right, 12 / 12 stopped",
            "the 3-turn conversation ('My name is Kenji.' / 'What is my name? Answer with the name only.' / 'What is 17 times 23? Answer with the number only.'): 'Kenji', then '391', all three ending with a stop token = line 3 (c) as written; in metrics, not part of output_match",
            "(cpu_on: 2026-11-10 15:50:14-17:15:55 JST (the CPU step's start to the GPU step's)): reply tokens (the thought included) up to 97; = MEASUREMENTS.md's Accuracy row for 'mixed_int4 ekv2048' and the Hub card's Accuracy table at main 5792549c"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answers_right": 12.0,
            "conversation_c_as_written": 1.0,
            "conversation_turn2_name_right": 1.0,
            "conversation_turn3_answer": 391.0,
            "conversation_turns": 3.0,
            "conversation_turns_ended_with_stop_token": 3.0,
            "cpu_threads": 8.0,
            "ended_with_stop_token": 12.0,
            "max_new_tokens": 2048.0,
            "questions": 12.0,
            "reply_tokens_max": 97.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the 12 check questions of line 3 (the 8 short questions of hf-to-litertlm's scripts/verify_quality.py, a price word problem, a syllogism, an average speed, a question in Japanese) and its 3-turn conversation, thinking on (the template's default), greedy (top-k 1), at most 2,048 new tokens, one process per question; verdicts by the line's regex on the text after the thought",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-10/nemotron-3-nano-30b-a3b-mixed-int4-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": 2048,
          "date": "2026-11-12",
          "decode_tokens_per_s": 8.0,
          "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": [
            "bytes: windows W22 and W23 read, rebuilt as an APFS clone of the mixed_int4 ekv4096 file with the 387 differing byte runs written in (logged back at 2026-11-11 18:23:15 JST) and checked before use (tools/unpark_bundles.sh compares size and sha256 before it logs the file back): sha256 b0bd9a61f9c168197f9c1135762f14d56c39e7f3feb3d71ef8f4b904ed5923f4, 20625860800 B, the value the work had checked against the Hub's before it removed the first copy (2026-11-11 15:03) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit ac7bc38a (the commit's readback); the copy was removed after the windows (2026-11-12 03:31:28, deleted_paths.txt)",
            "runtime: litert_lm_advanced_main built from the LiteRT-LM v0.19.1 tag (7a03724c; sha256 6f0eba84..., with the tag's prebuilt macos_arm64 libwebgpu_dawn.dylib and libGemmaModelConstraintProvider.dylib), one process at a time: `--backend=cpu --benchmark --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256 --num_iterations=2 --max_num_tokens=2048 --disable_cache=true --report_peak_memory_footprint --use_session=true --async=false --num_cpu_threads=8`; CPU = XNNPACK, number_of_threads 8 (each log prints it); the --use_session=true --async=false path gives one 1,024-token prefill per iteration (the CLI's conversation path prefilled the 1,024 tokens twice per iteration on this bundle, MEASUREMENTS.md); the logs print no partition or node count",
            "window W23 of 2026-11-12: this arm's processes ran 02:25:01-02:43:21 JST between the other arms' (one process at a time); every process's meta says 'overlap clean'; both exclusion lists of the window are empty; the cell (the benchmark rule): per process iteration 2 of 2 of one engine (iteration 1 = the warm-up), the cell = the median of the 3 processes, prefill and decode each; a process counts when it exits 0, its log has no Validation error, Shape mismatch or not-fully-delegated line, every iteration is one 1,024-token prefill turn and one 256-token decode turn, and the window's 1 s record shows no burst of other work at 300 % or more in its counted spans (every span ok); ttft_ms = the median of the processes' mean TTFT over the same iterations; peak_mem_mb = the largest of the processes' last 'Peak private footprint' print (--report_peak_memory_footprint; the block of the median of the iterations), the reading of the card and MEASUREMENTS.md, whose GB = MB x 2^20 / 10^9 (tools/speed_parse.py reads the print as MiB); each iteration's 'Peak system ram usage' print is in metrics",
            "context_length = the file's KV cache, 2,048 tokens (the card's Context column; the writer's --cache=2048); --max_num_tokens=2048 caps the run's tokens",
            "(started 2026-11-12 02:25:01 JST, CPU 70 % idle): 54.76 / 8.64 tok/s, TTFT 18.82 s; 114.69 / 7.89 tok/s, TTFT 9.06 s; Init Total 4956.67 ms; the last Peak private footprint print 6070.53 MB",
            "(started 2026-11-12 02:29:56 JST, CPU 77 % idle): 85.22 / 8.03 tok/s, TTFT 12.14 s; 116.17 / 8.0 tok/s, TTFT 8.94 s; Init Total 5074.66 ms; the last Peak private footprint print 6070.33 MB",
            "(started 2026-11-12 02:41:34 JST, CPU 81 % idle): 53.58 / 7.94 tok/s, TTFT 19.24 s; 117.91 / 8.05 tok/s, TTFT 8.81 s; Init Total 4754.63 ms; the last Peak private footprint print 6069.42 MB",
            "= MEASUREMENTS.md section 2026-11-12 ('| mixed_int4 ekv2048 | CPU (XNNPACK, 8 threads) | 116.17 tokens/s | 8.00 tokens/s | 3 | 6.37 GB |' and its per-process lists) and the Hub card's Performance row ('| mixed_int4 ekv2048 | CPU (8 threads) | 116 tokens/s | 8.00 tokens/s | 6.4 GB |') at main 5792549c"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "cpu_threads": 8.0,
            "decode_tokens": 256.0,
            "iterations": 2.0,
            "max_num_tokens": 2048.0,
            "peak_private_footprint_gb_max": 6.37,
            "prefill_tokens": 1024.0,
            "proc1_cpu_idle_at_start_pct": 70.0,
            "proc1_init_total_ms": 4956.67,
            "proc1_iter1_decode_tok_s": 8.64,
            "proc1_iter1_prefill_tok_s": 54.76,
            "proc1_iter1_ttft_s": 18.82,
            "proc1_iter2_decode_tok_s": 7.89,
            "proc1_iter2_prefill_tok_s": 114.69,
            "proc1_iter2_ttft_s": 9.06,
            "proc1_peak_private_footprint_mb": 6070.53,
            "proc1_peak_system_ram_mb_max": 6070.57,
            "proc2_cpu_idle_at_start_pct": 77.0,
            "proc2_init_total_ms": 5074.66,
            "proc2_iter1_decode_tok_s": 8.03,
            "proc2_iter1_prefill_tok_s": 85.22,
            "proc2_iter1_ttft_s": 12.14,
            "proc2_iter2_decode_tok_s": 8.0,
            "proc2_iter2_prefill_tok_s": 116.17,
            "proc2_iter2_ttft_s": 8.94,
            "proc2_peak_private_footprint_mb": 6070.33,
            "proc2_peak_system_ram_mb_max": 6070.35,
            "proc3_cpu_idle_at_start_pct": 81.0,
            "proc3_init_total_ms": 4754.63,
            "proc3_iter1_decode_tok_s": 7.94,
            "proc3_iter1_prefill_tok_s": 53.58,
            "proc3_iter1_ttft_s": 19.24,
            "proc3_iter2_decode_tok_s": 8.05,
            "proc3_iter2_prefill_tok_s": 117.91,
            "proc3_iter2_ttft_s": 8.81,
            "proc3_peak_private_footprint_mb": 6069.42,
            "proc3_peak_system_ram_mb_max": 6069.42,
            "processes": 3.0
          },
          "output_match": null,
          "peak_mem_mb": 6070.53,
          "prefill_tokens_per_s": 116.17,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 8940.0
        },
        "source": "data/device_runs/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-mixed-int4-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-10",
          "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": [
            "bytes: these runs read as the bundle step wrote it (done 2026-11-10 13:59:31 JST; the build log at 13:59:51 gives its sha256 b0bd9a61... at the build), hashed by the pre-upload after them (done 2026-11-10 20:39:53 JST, sha256 b0bd9a61f9c168197f9c1135762f14d56c39e7f3feb3d71ef8f4b904ed5923f4, 20625860800 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit ac7bc38a (the commit's readback)",
            "runtime: litert_lm_advanced_main built from the LiteRT-LM v0.19.1 tag (7a03724c; sha256 6f0eba84...), WebGPU on Metal, one process per question: `--backend=gpu --max_output_tokens=2048 --min_log_severity=2 --sampler_top_k=1 --sampler_handles_input=false --disable_cache=true --use_session=true --async=false --input_prompt_file=<the vendor template's render with enable_thinking false>` (the CLI has no thinking switch; the render equals the file's template for these prompts, line T); the 3 turns = 3 processes, each render carrying the file's own earlier answers",
            "output_match = line 3 (a) and (b) as written before any output: (a) the file's verdict equals the reference's on 12 of 12 (at least 11 required; both answer 'violet' to the rhyme question, so 11 are right), (b) every question the reference ends with a stop token ends with one here (12 / 12). Reference = the float32 implementation of the model on the bf16 weights, thinking off, greedy, its replies built by draft-and-verify (every appended token its argmax), held to NVIDIA's modeling_nemotron_h.py by lines 0 and R",
            "the 3-turn conversation: 'Kenji', then '401', all three ending with a stop token; the reference's own turns: 'Kenji', then '401'. Line 3 (c) as written asks for 391 in turn 3 (not met); line 3' (written after these outputs were seen: a turn the reference gets wrong is not required of the file) is met; in metrics, not part of output_match",
            "(gpu_off: 2026-11-10 15:37:02-15:50:14 JST (each run's end time in the jsonl)): reply tokens up to 286; = MEASUREMENTS.md's Accuracy row for 'mixed_int4 ekv2048' and the Hub card's Accuracy table at main 5792549c"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answers_right": 11.0,
            "conversation_c_as_written": 0.0,
            "conversation_c_line3_prime": 1.0,
            "conversation_turn2_name_right": 1.0,
            "conversation_turn3_answer": 401.0,
            "conversation_turns": 3.0,
            "conversation_turns_ended_with_stop_token": 3.0,
            "ended_with_stop_token": 12.0,
            "max_new_tokens": 2048.0,
            "questions": 12.0,
            "reference_ended_with_stop_token": 12.0,
            "reference_turn3_answer": 401.0,
            "reply_tokens_max": 286.0,
            "verdicts_equal_reference": 12.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the 12 check questions of line 3 (the 8 short questions of hf-to-litertlm's scripts/verify_quality.py, a price word problem, a syllogism, an average speed, a question in Japanese) and its 3-turn conversation, thinking off (enable_thinking false), greedy (top-k 1), at most 2,048 new tokens, one process per question; verdicts against the source model's replies (float32 on the bf16 weights, thinking off)",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-10/nemotron-3-nano-30b-a3b-mixed-int4-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-10",
          "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": [
            "bytes: these runs read as the bundle step wrote it (done 2026-11-10 13:59:31 JST; the build log at 13:59:51 gives its sha256 b0bd9a61... at the build), hashed by the pre-upload after them (done 2026-11-10 20:39:53 JST, sha256 b0bd9a61f9c168197f9c1135762f14d56c39e7f3feb3d71ef8f4b904ed5923f4, 20625860800 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit ac7bc38a (the commit's readback)",
            "runtime: litert_lm_advanced_main built from the LiteRT-LM v0.19.1 tag (7a03724c; sha256 6f0eba84...), WebGPU on Metal, one process per question: `--backend=gpu --max_output_tokens=2048 --min_log_severity=2 --sampler_top_k=1 --sampler_handles_input=false --disable_cache=true --input_prompt=<question>` (the file's template, thinking on by default); the 3 turns through `--multi_turns` in one process",
            "output_match = line 3's rule for thinking on, written before any output (the fallback taken on 2026-11-10 because the reference generator ran at 77.8 s per step, so no reference ran with thinking on): the line's regex right on at least 11 of the 12 and every reply ending with a stop token within 2,048 new tokens: 12 / 12 right, 12 / 12 stopped",
            "the 3-turn conversation ('My name is Kenji.' / 'What is my name? Answer with the name only.' / 'What is 17 times 23? Answer with the number only.'): 'Kenji', then '391', all three ending with a stop token = line 3 (c) as written; in metrics, not part of output_match",
            "(gpu_on: 2026-11-10 17:24:38-17:59:28 JST (each run's end time in the jsonl)): reply tokens (the thought included) up to 244; = MEASUREMENTS.md's Accuracy row for 'mixed_int4 ekv2048' and the Hub card's Accuracy table at main 5792549c"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "answers_right": 12.0,
            "conversation_c_as_written": 1.0,
            "conversation_turn2_name_right": 1.0,
            "conversation_turn3_answer": 391.0,
            "conversation_turns": 3.0,
            "conversation_turns_ended_with_stop_token": 3.0,
            "ended_with_stop_token": 12.0,
            "max_new_tokens": 2048.0,
            "questions": 12.0,
            "reply_tokens_max": 244.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the 12 check questions of line 3 (the 8 short questions of hf-to-litertlm's scripts/verify_quality.py, a price word problem, a syllogism, an average speed, a question in Japanese) and its 3-turn conversation, thinking on (the template's default), greedy (top-k 1), at most 2,048 new tokens, one process per question; verdicts by the line's regex on the text after the thought",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-10/nemotron-3-nano-30b-a3b-mixed-int4-ekv2048__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-11-12",
          "decode_tokens_per_s": 71.387,
          "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": [
            "bytes: windows W22 and W23 read, rebuilt as an APFS clone of the mixed_int4 ekv4096 file with the 387 differing byte runs written in (logged back at 2026-11-11 18:23:15 JST) and checked before use (tools/unpark_bundles.sh compares size and sha256 before it logs the file back): sha256 b0bd9a61f9c168197f9c1135762f14d56c39e7f3feb3d71ef8f4b904ed5923f4, 20625860800 B, the value the work had checked against the Hub's before it removed the first copy (2026-11-11 15:03) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit ac7bc38a (the commit's readback); the copy was removed after the windows (2026-11-12 03:31:28, deleted_paths.txt)",
            "runtime: litert_lm_advanced_main built from the LiteRT-LM v0.19.1 tag (7a03724c; sha256 6f0eba84..., with the tag's prebuilt macos_arm64 libwebgpu_dawn.dylib and libGemmaModelConstraintProvider.dylib), one process at a time: `--backend=gpu --benchmark --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256 --num_iterations=4 --max_num_tokens=2048 --disable_cache=true --report_peak_memory_footprint --sampler_handles_input=false --use_session=true --async=false`; GPU = WebGPU on Metal (each log selects 'Apple M4 Max, arch=metal-3, vendor=apple, backend=Metal'); the --use_session=true --async=false path gives one 1,024-token prefill per iteration (the CLI's conversation path prefilled the 1,024 tokens twice per iteration on this bundle, MEASUREMENTS.md); the logs print no partition or node count",
            "window W22 of 2026-11-12: this arm's processes ran 01:09:35-01:30:16 JST between the other arms' (one process at a time); every process's meta says 'overlap clean'; both exclusion lists of the window are empty; the cell (the benchmark rule): per process the mean of iterations 2-4 of one engine (iteration 1 = the warm-up), the cell = the median of the 5 processes, prefill and decode each; a process counts when it exits 0, its log has no Validation error, Shape mismatch or not-fully-delegated line, every iteration is one 1,024-token prefill turn and one 256-token decode turn, and the window's 1 s record shows no burst of other work at 300 % or more in its counted spans (every span ok); ttft_ms = the median of the processes' mean TTFT over the same iterations; peak_mem_mb = the largest of the processes' last 'Peak private footprint' print (--report_peak_memory_footprint; the block of the median of the iterations), the reading of the card and MEASUREMENTS.md, whose GB = MB x 2^20 / 10^9 (tools/speed_parse.py reads the print as MiB); each iteration's 'Peak system ram usage' print is in metrics",
            "context_length = the file's KV cache, 2,048 tokens (the card's Context column; the writer's --cache=2048); --max_num_tokens=2048 caps the run's tokens",
            "(started 2026-11-12 01:09:35 JST, CPU 77 % idle): 819.9 / 66.49 tok/s, TTFT 1.26 s; 847.27 / 71.51 tok/s, TTFT 1.22 s; 842.17 / 71.86 tok/s, TTFT 1.23 s; 847.25 / 70.67 tok/s, TTFT 1.22 s; Init Total 40022.38 ms; the last Peak private footprint print 21503.86 MB",
            "(started 2026-11-12 01:13:14 JST, CPU 70 % idle): 818.75 / 71.25 tok/s, TTFT 1.26 s; 845.09 / 71.76 tok/s, TTFT 1.23 s; 843.97 / 71.36 tok/s, TTFT 1.23 s; 847.38 / 71.81 tok/s, TTFT 1.22 s; Init Total 39285.69 ms; the last Peak private footprint print 21502.7 MB",
            "(started 2026-11-12 01:16:54 JST, CPU 68 % idle): 824.3 / 70.31 tok/s, TTFT 1.26 s; 847.46 / 69.43 tok/s, TTFT 1.22 s; 841.6 / 71.87 tok/s, TTFT 1.23 s; 843.07 / 71.33 tok/s, TTFT 1.23 s; Init Total 38831.55 ms; the last Peak private footprint print 21506.75 MB",
            "(started 2026-11-12 01:25:26 JST, CPU 72 % idle): 820.98 / 71.39 tok/s, TTFT 1.26 s; 846.04 / 72.14 tok/s, TTFT 1.22 s; 846.44 / 71.72 tok/s, TTFT 1.22 s; 847.26 / 71.61 tok/s, TTFT 1.22 s; Init Total 38669.75 ms; the last Peak private footprint print 21489.49 MB",
            "(started 2026-11-12 01:29:05 JST, CPU 87 % idle): 826.9 / 71.27 tok/s, TTFT 1.25 s; 844.1 / 70.68 tok/s, TTFT 1.23 s; 845.38 / 71.9 tok/s, TTFT 1.23 s; 844.03 / 71.58 tok/s, TTFT 1.23 s; Init Total 38436.01 ms; the last Peak private footprint print 21504.37 MB",
            "= MEASUREMENTS.md section 2026-11-12 ('| mixed_int4 ekv2048 | GPU (WebGPU) | 845.48 tokens/s | 71.39 tokens/s | 5 | 22.55 GB |' and its per-process lists) and the Hub card's Performance row ('| mixed_int4 ekv2048 | GPU (WebGPU) | 845 tokens/s | 71.4 tokens/s | 22.6 GB |') at main 5792549c"
          ],
          "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,
            "iterations": 4.0,
            "max_num_tokens": 2048.0,
            "peak_private_footprint_gb_max": 22.55,
            "prefill_tokens": 1024.0,
            "proc1_cpu_idle_at_start_pct": 77.0,
            "proc1_init_total_ms": 40022.38,
            "proc1_iter1_decode_tok_s": 66.49,
            "proc1_iter1_prefill_tok_s": 819.9,
            "proc1_iter1_ttft_s": 1.26,
            "proc1_iter2_decode_tok_s": 71.51,
            "proc1_iter2_prefill_tok_s": 847.27,
            "proc1_iter2_ttft_s": 1.22,
            "proc1_iter3_decode_tok_s": 71.86,
            "proc1_iter3_prefill_tok_s": 842.17,
            "proc1_iter3_ttft_s": 1.23,
            "proc1_iter4_decode_tok_s": 70.67,
            "proc1_iter4_prefill_tok_s": 847.25,
            "proc1_iter4_ttft_s": 1.22,
            "proc1_peak_private_footprint_mb": 21503.86,
            "proc1_peak_system_ram_mb_max": 21517.8,
            "proc2_cpu_idle_at_start_pct": 70.0,
            "proc2_init_total_ms": 39285.69,
            "proc2_iter1_decode_tok_s": 71.25,
            "proc2_iter1_prefill_tok_s": 818.75,
            "proc2_iter1_ttft_s": 1.26,
            "proc2_iter2_decode_tok_s": 71.76,
            "proc2_iter2_prefill_tok_s": 845.09,
            "proc2_iter2_ttft_s": 1.23,
            "proc2_iter3_decode_tok_s": 71.36,
            "proc2_iter3_prefill_tok_s": 843.97,
            "proc2_iter3_ttft_s": 1.23,
            "proc2_iter4_decode_tok_s": 71.81,
            "proc2_iter4_prefill_tok_s": 847.38,
            "proc2_iter4_ttft_s": 1.22,
            "proc2_peak_private_footprint_mb": 21502.7,
            "proc2_peak_system_ram_mb_max": 21510.6,
            "proc3_cpu_idle_at_start_pct": 68.0,
            "proc3_init_total_ms": 38831.55,
            "proc3_iter1_decode_tok_s": 70.31,
            "proc3_iter1_prefill_tok_s": 824.3,
            "proc3_iter1_ttft_s": 1.26,
            "proc3_iter2_decode_tok_s": 69.43,
            "proc3_iter2_prefill_tok_s": 847.46,
            "proc3_iter2_ttft_s": 1.22,
            "proc3_iter3_decode_tok_s": 71.87,
            "proc3_iter3_prefill_tok_s": 841.6,
            "proc3_iter3_ttft_s": 1.23,
            "proc3_iter4_decode_tok_s": 71.33,
            "proc3_iter4_prefill_tok_s": 843.07,
            "proc3_iter4_ttft_s": 1.23,
            "proc3_peak_private_footprint_mb": 21506.75,
            "proc3_peak_system_ram_mb_max": 21512.8,
            "proc4_cpu_idle_at_start_pct": 72.0,
            "proc4_init_total_ms": 38669.75,
            "proc4_iter1_decode_tok_s": 71.39,
            "proc4_iter1_prefill_tok_s": 820.98,
            "proc4_iter1_ttft_s": 1.26,
            "proc4_iter2_decode_tok_s": 72.14,
            "proc4_iter2_prefill_tok_s": 846.04,
            "proc4_iter2_ttft_s": 1.22,
            "proc4_iter3_decode_tok_s": 71.72,
            "proc4_iter3_prefill_tok_s": 846.44,
            "proc4_iter3_ttft_s": 1.22,
            "proc4_iter4_decode_tok_s": 71.61,
            "proc4_iter4_prefill_tok_s": 847.26,
            "proc4_iter4_ttft_s": 1.22,
            "proc4_peak_private_footprint_mb": 21489.49,
            "proc4_peak_system_ram_mb_max": 21491.6,
            "proc5_cpu_idle_at_start_pct": 87.0,
            "proc5_init_total_ms": 38436.01,
            "proc5_iter1_decode_tok_s": 71.27,
            "proc5_iter1_prefill_tok_s": 826.9,
            "proc5_iter1_ttft_s": 1.25,
            "proc5_iter2_decode_tok_s": 70.68,
            "proc5_iter2_prefill_tok_s": 844.1,
            "proc5_iter2_ttft_s": 1.23,
            "proc5_iter3_decode_tok_s": 71.9,
            "proc5_iter3_prefill_tok_s": 845.38,
            "proc5_iter3_ttft_s": 1.23,
            "proc5_iter4_decode_tok_s": 71.58,
            "proc5_iter4_prefill_tok_s": 844.03,
            "proc5_iter4_ttft_s": 1.23,
            "proc5_peak_private_footprint_mb": 21504.37,
            "proc5_peak_system_ram_mb_max": 21505.9,
            "processes": 5.0
          },
          "output_match": null,
          "peak_mem_mb": 21506.75,
          "prefill_tokens_per_s": 845.48,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 1226.7
        },
        "source": "data/device_runs/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-mixed-int4-ekv2048__mac-studio-m4-max.json"
      }
    ]
  },
  "model": {
    "family": "nemotron-3-nano",
    "id": "nemotron-3-nano-30b-a3b-mixed-int4-ekv2048",
    "license": "other (nvidia-nemotron-open-model-license)",
    "source_url": "https://huggingface.co/litert-community/Nemotron-3-Nano-30B-A3B",
    "task": "text-generation"
  },
  "pitfalls": [
    "Keep the GPU cache off (cache_dir=\":nocache\" in Python, --disable_cache=true in the command-line tool): on a Mac, runs that loaded an earlier run's GPU program cache answered wrongly (card L68).",
    "One Engine per conversation: the Mamba-2 state is not reset between conversations (LiteRT-LM #3165) (card L69).",
    "Tool calls come back as text (the model's <tool_call> form) (card L70).",
    "Long replies: a blocking send_message still generating after 10 minutes fails with DEADLINE_EXCEEDED; use send_message_async or cap max_output_tokens (card L71).",
    "After a message overflows the context window, every later message in that conversation fails (Task failed with state: 7); start a new conversation (card L72).",
    "GPU: the card runs it with the v0.19.1 command-line tool built from source, which has --sampler_handles_input=false (the PyPI package does not); LiteRT-LM 0.18.0's GPU moe op only supports activation 'gelu'. CPU: the PyPI package, litert-lm 0.19.1 (card L27, L76-L93).",
    "Thinking is on by default; thinking_config=litert_lm.ThinkingConfig(enable_thinking=False) in create_conversation turns it off (card L29, L105).",
    "The card's checks (12 questions and a three-turn chat, a name then 17 × 23, greedy, against NVIDIA's bf16 weights): with thinking on, 12/12 right and 391 for 17 × 23 on all four files on the CPU and the GPU (the source model was not run with thinking on); with thinking off, 11/12 right, as the source model, and 17 × 23 comes back as 401 for the mixed_int4 files (this one among them) and for the source model, 391 on the CPU and 411 on the GPU for the int4 files (card L59-L64).",
    "Which file: this one gives faster CPU prefill than int4 ekv2048 (card L37). On an Apple M4 Max (128 GB) its benchmark peaked at 22.6 GB on the GPU and 6.4 GB on the CPU (2026-11-12); speed still being tuned (updates follow) (card L27, L50-L51, L55)."
  ],
  "schema_version": "1.2"
}
