{
  "artifacts": [
    {
      "file": "nemotron3_nano_30b_a3b_int4_ekv4096.litertlm",
      "sha256": "cf6aa824ae6623355ffc82b4d844706517fdd61887f6a72cc9e5e9f32f449282",
      "size_mb": 17017.559
    }
  ],
  "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 128 values (64 in their down projection); the other weight matrices int4, except the token table and lm_head (int8); router fp32; fp32 activations declared; a 4,096-token context; 17.8 GB (card L36, 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": "2026-11-12",
          "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 2026-11-12 05:20:39 JST), hashed after them at 05:34:05 (sha256 cf6aa824ae6623355ffc82b4d844706517fdd61887f6a72cc9e5e9f32f449282); at 05:34:48 that path became an APFS clone with the 389 differing byte runs written in, which hashed the same; the pre-upload hashed the clone (done 2026-11-12 07:09:36 JST, preupload_nemotron3_nano_30b_a3b_int4_ekv4096.litertlm.json: 17844203904 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit 5792549c (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 '391', 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 (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: 2026-11-12 05:20:39-05:22:29 JST (the CPU step's start to the GPU step's)): reply tokens up to 180; = MEASUREMENTS.md's Accuracy row for 'int4 ekv4096' 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": 1.0,
            "conversation_c_line3_prime": 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,
            "reference_ended_with_stop_token": 12.0,
            "reference_turn3_answer": 401.0,
            "reply_tokens_max": 180.0,
            "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/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-int4-ekv4096__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-11-12",
          "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 2026-11-12 05:20:39 JST), hashed after them at 05:34:05 (sha256 cf6aa824ae6623355ffc82b4d844706517fdd61887f6a72cc9e5e9f32f449282); at 05:34:48 that path became an APFS clone with the 389 differing byte runs written in, which hashed the same; the pre-upload hashed the clone (done 2026-11-12 07:09:36 JST, preupload_nemotron3_nano_30b_a3b_int4_ekv4096.litertlm.json: 17844203904 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit 5792549c (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: 2026-11-12 05:26:46-05:29:42 JST (the CPU step's start to the GPU step's)): reply tokens (the thought included) up to 75; = MEASUREMENTS.md's Accuracy row for 'int4 ekv4096' 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": 75.0
          },
          "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/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-int4-ekv4096__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": 4096,
          "date": "2026-11-12",
          "decode_tokens_per_s": 12.16,
          "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 after 05:34:48 JST, when that path became an APFS clone with the 389 differing byte runs written in (mtime 2026-11-12 05:34:11), whose sha256 cf6aa824ae6623355ffc82b4d844706517fdd61887f6a72cc9e5e9f32f449282 equals the built file's (05:34:05 and 05:34:48); the pre-upload hashed it again (done 2026-11-12 07:09:36 JST, preupload_nemotron3_nano_30b_a3b_int4_ekv4096.litertlm.json: 17844203904 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit 5792549c (the commit's readback)",
            "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 W27 of 2026-11-12: this arm's processes ran 06:08:55-06:15:59 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, 4,096 tokens (the card's Context column; the writer's --cache=4096); --max_num_tokens=2048 caps the run's tokens",
            "(started 2026-11-12 06:08:55 JST, CPU 86 % idle): 60.28 / 12.34 tok/s, TTFT 17.07 s; 104.92 / 12.16 tok/s, TTFT 9.84 s; Init Total 2800.73 ms; the last Peak private footprint print 4768.59 MB",
            "(started 2026-11-12 06:10:28 JST, CPU 79 % idle): 101.6 / 12.69 tok/s, TTFT 10.16 s; 105.99 / 12.62 tok/s, TTFT 9.74 s; Init Total 1202.97 ms; the last Peak private footprint print 4775.69 MB",
            "(started 2026-11-12 06:14:48 JST, CPU 86 % idle): 102.78 / 12.82 tok/s, TTFT 10.04 s; 106.33 / 12.04 tok/s, TTFT 9.71 s; Init Total 1225.2 ms; the last Peak private footprint print 4775.66 MB",
            "= MEASUREMENTS.md section 2026-11-12 ('| int4 ekv4096 | CPU (XNNPACK, 8 threads) | 105.99 tokens/s | 12.16 tokens/s | 3 | 5.01 GB |' and its per-process lists) and the Hub card's Performance row ('| int4 ekv4096 | CPU (8 threads) | 106 tokens/s | 12.2 tokens/s | 5.0 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": 5.01,
            "prefill_tokens": 1024.0,
            "proc1_cpu_idle_at_start_pct": 86.0,
            "proc1_init_total_ms": 2800.73,
            "proc1_iter1_decode_tok_s": 12.34,
            "proc1_iter1_prefill_tok_s": 60.28,
            "proc1_iter1_ttft_s": 17.07,
            "proc1_iter2_decode_tok_s": 12.16,
            "proc1_iter2_prefill_tok_s": 104.92,
            "proc1_iter2_ttft_s": 9.84,
            "proc1_peak_private_footprint_mb": 4768.59,
            "proc1_peak_system_ram_mb_max": 4768.69,
            "proc2_cpu_idle_at_start_pct": 79.0,
            "proc2_init_total_ms": 1202.97,
            "proc2_iter1_decode_tok_s": 12.69,
            "proc2_iter1_prefill_tok_s": 101.6,
            "proc2_iter1_ttft_s": 10.16,
            "proc2_iter2_decode_tok_s": 12.62,
            "proc2_iter2_prefill_tok_s": 105.99,
            "proc2_iter2_ttft_s": 9.74,
            "proc2_peak_private_footprint_mb": 4775.69,
            "proc2_peak_system_ram_mb_max": 4775.74,
            "proc3_cpu_idle_at_start_pct": 86.0,
            "proc3_init_total_ms": 1225.2,
            "proc3_iter1_decode_tok_s": 12.82,
            "proc3_iter1_prefill_tok_s": 102.78,
            "proc3_iter1_ttft_s": 10.04,
            "proc3_iter2_decode_tok_s": 12.04,
            "proc3_iter2_prefill_tok_s": 106.33,
            "proc3_iter2_ttft_s": 9.71,
            "proc3_peak_private_footprint_mb": 4775.66,
            "proc3_peak_system_ram_mb_max": 4775.7,
            "processes": 3.0
          },
          "output_match": null,
          "peak_mem_mb": 4775.69,
          "prefill_tokens_per_s": 105.99,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 9740.0
        },
        "source": "data/device_runs/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-int4-ekv4096__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-12",
          "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 2026-11-12 05:20:39 JST), hashed after them at 05:34:05 (sha256 cf6aa824ae6623355ffc82b4d844706517fdd61887f6a72cc9e5e9f32f449282); at 05:34:48 that path became an APFS clone with the 389 differing byte runs written in, which hashed the same; the pre-upload hashed the clone (done 2026-11-12 07:09:36 JST, preupload_nemotron3_nano_30b_a3b_int4_ekv4096.litertlm.json: 17844203904 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit 5792549c (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 '411', 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: 2026-11-12 05:22:49-05:26:46 JST (each run's end time in the jsonl)): reply tokens up to 133; = MEASUREMENTS.md's Accuracy row for 'int4 ekv4096' 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": 411.0,
            "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": 133.0,
            "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/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-int4-ekv4096__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-11-12",
          "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 2026-11-12 05:20:39 JST), hashed after them at 05:34:05 (sha256 cf6aa824ae6623355ffc82b4d844706517fdd61887f6a72cc9e5e9f32f449282); at 05:34:48 that path became an APFS clone with the 389 differing byte runs written in, which hashed the same; the pre-upload hashed the clone (done 2026-11-12 07:09:36 JST, preupload_nemotron3_nano_30b_a3b_int4_ekv4096.litertlm.json: 17844203904 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit 5792549c (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: 2026-11-12 05:30:01-05:33:23 JST (each run's end time in the jsonl)): reply tokens (the thought included) up to 77; = MEASUREMENTS.md's Accuracy row for 'int4 ekv4096' 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": 77.0
          },
          "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/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-int4-ekv4096__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": 4096,
          "date": "2026-11-12",
          "decode_tokens_per_s": 70.657,
          "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 after 05:34:48 JST, when that path became an APFS clone with the 389 differing byte runs written in (mtime 2026-11-12 05:34:11), whose sha256 cf6aa824ae6623355ffc82b4d844706517fdd61887f6a72cc9e5e9f32f449282 equals the built file's (05:34:05 and 05:34:48); the pre-upload hashed it again (done 2026-11-12 07:09:36 JST, preupload_nemotron3_nano_30b_a3b_int4_ekv4096.litertlm.json: 17844203904 B) = the Hub's LFS oid and size at main 5792549c (anonymous) and at commit 5792549c (the commit's readback)",
            "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 W26 of 2026-11-12: this arm's processes ran 05:47:16-05:55:44 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, 4,096 tokens (the card's Context column; the writer's --cache=4096); --max_num_tokens=2048 caps the run's tokens",
            "(started 2026-11-12 05:47:16 JST, CPU 58 % idle): 794.2 / 71.02 tok/s, TTFT 1.3 s; 815.69 / 70.2 tok/s, TTFT 1.27 s; 814.41 / 71.05 tok/s, TTFT 1.27 s; 815.86 / 70.72 tok/s, TTFT 1.27 s; Init Total 29442.62 ms; the last Peak private footprint print 19480.72 MB",
            "(started 2026-11-12 05:48:14 JST, CPU 73 % idle): 792.14 / 71.11 tok/s, TTFT 1.31 s; 814.73 / 70.32 tok/s, TTFT 1.27 s; 818.0 / 70.28 tok/s, TTFT 1.27 s; 814.32 / 71.09 tok/s, TTFT 1.27 s; Init Total 9164.04 ms; the last Peak private footprint print 19478.87 MB",
            "(started 2026-11-12 05:51:02 JST, CPU 80 % idle): 791.75 / 71.05 tok/s, TTFT 1.31 s; 813.13 / 70.5 tok/s, TTFT 1.27 s; 810.01 / 70.78 tok/s, TTFT 1.28 s; 811.42 / 71.17 tok/s, TTFT 1.28 s; Init Total 29544.66 ms; the last Peak private footprint print 19472.19 MB",
            "(started 2026-11-12 05:52:00 JST, CPU 79 % idle): 796.0 / 70.48 tok/s, TTFT 1.3 s; 816.69 / 71.62 tok/s, TTFT 1.27 s; 817.9 / 69.75 tok/s, TTFT 1.27 s; 816.81 / 71.25 tok/s, TTFT 1.27 s; Init Total 8812.66 ms; the last Peak private footprint print 19473.83 MB",
            "(started 2026-11-12 05:54:49 JST, CPU 78 % idle): 790.98 / 69.5 tok/s, TTFT 1.31 s; 812.32 / 71.34 tok/s, TTFT 1.27 s; 816.87 / 71.04 tok/s, TTFT 1.27 s; 813.98 / 69.53 tok/s, TTFT 1.27 s; Init Total 29767.95 ms; the last Peak private footprint print 19470.67 MB",
            "= MEASUREMENTS.md section 2026-11-12 ('| int4 ekv4096 | GPU (WebGPU) | 815.32 tokens/s | 70.66 tokens/s | 5 | 20.43 GB |' and its per-process lists) and the Hub card's Performance row ('| int4 ekv4096 | GPU (WebGPU) | 815 tokens/s | 70.7 tokens/s | 20.4 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": {
            "decode_tokens": 256.0,
            "iterations": 4.0,
            "max_num_tokens": 2048.0,
            "peak_private_footprint_gb_max": 20.43,
            "prefill_tokens": 1024.0,
            "proc1_cpu_idle_at_start_pct": 58.0,
            "proc1_init_total_ms": 29442.62,
            "proc1_iter1_decode_tok_s": 71.02,
            "proc1_iter1_prefill_tok_s": 794.2,
            "proc1_iter1_ttft_s": 1.3,
            "proc1_iter2_decode_tok_s": 70.2,
            "proc1_iter2_prefill_tok_s": 815.69,
            "proc1_iter2_ttft_s": 1.27,
            "proc1_iter3_decode_tok_s": 71.05,
            "proc1_iter3_prefill_tok_s": 814.41,
            "proc1_iter3_ttft_s": 1.27,
            "proc1_iter4_decode_tok_s": 70.72,
            "proc1_iter4_prefill_tok_s": 815.86,
            "proc1_iter4_ttft_s": 1.27,
            "proc1_peak_private_footprint_mb": 19480.72,
            "proc1_peak_system_ram_mb_max": 19482.3,
            "proc2_cpu_idle_at_start_pct": 73.0,
            "proc2_init_total_ms": 9164.04,
            "proc2_iter1_decode_tok_s": 71.11,
            "proc2_iter1_prefill_tok_s": 792.14,
            "proc2_iter1_ttft_s": 1.31,
            "proc2_iter2_decode_tok_s": 70.32,
            "proc2_iter2_prefill_tok_s": 814.73,
            "proc2_iter2_ttft_s": 1.27,
            "proc2_iter3_decode_tok_s": 70.28,
            "proc2_iter3_prefill_tok_s": 818.0,
            "proc2_iter3_ttft_s": 1.27,
            "proc2_iter4_decode_tok_s": 71.09,
            "proc2_iter4_prefill_tok_s": 814.32,
            "proc2_iter4_ttft_s": 1.27,
            "proc2_peak_private_footprint_mb": 19478.87,
            "proc2_peak_system_ram_mb_max": 19488.6,
            "proc3_cpu_idle_at_start_pct": 80.0,
            "proc3_init_total_ms": 29544.66,
            "proc3_iter1_decode_tok_s": 71.05,
            "proc3_iter1_prefill_tok_s": 791.75,
            "proc3_iter1_ttft_s": 1.31,
            "proc3_iter2_decode_tok_s": 70.5,
            "proc3_iter2_prefill_tok_s": 813.13,
            "proc3_iter2_ttft_s": 1.27,
            "proc3_iter3_decode_tok_s": 70.78,
            "proc3_iter3_prefill_tok_s": 810.01,
            "proc3_iter3_ttft_s": 1.28,
            "proc3_iter4_decode_tok_s": 71.17,
            "proc3_iter4_prefill_tok_s": 811.42,
            "proc3_iter4_ttft_s": 1.28,
            "proc3_peak_private_footprint_mb": 19472.19,
            "proc3_peak_system_ram_mb_max": 19474.3,
            "proc4_cpu_idle_at_start_pct": 79.0,
            "proc4_init_total_ms": 8812.66,
            "proc4_iter1_decode_tok_s": 70.48,
            "proc4_iter1_prefill_tok_s": 796.0,
            "proc4_iter1_ttft_s": 1.3,
            "proc4_iter2_decode_tok_s": 71.62,
            "proc4_iter2_prefill_tok_s": 816.69,
            "proc4_iter2_ttft_s": 1.27,
            "proc4_iter3_decode_tok_s": 69.75,
            "proc4_iter3_prefill_tok_s": 817.9,
            "proc4_iter3_ttft_s": 1.27,
            "proc4_iter4_decode_tok_s": 71.25,
            "proc4_iter4_prefill_tok_s": 816.81,
            "proc4_iter4_ttft_s": 1.27,
            "proc4_peak_private_footprint_mb": 19473.83,
            "proc4_peak_system_ram_mb_max": 19476.0,
            "proc5_cpu_idle_at_start_pct": 78.0,
            "proc5_init_total_ms": 29767.95,
            "proc5_iter1_decode_tok_s": 69.5,
            "proc5_iter1_prefill_tok_s": 790.98,
            "proc5_iter1_ttft_s": 1.31,
            "proc5_iter2_decode_tok_s": 71.34,
            "proc5_iter2_prefill_tok_s": 812.32,
            "proc5_iter2_ttft_s": 1.27,
            "proc5_iter3_decode_tok_s": 71.04,
            "proc5_iter3_prefill_tok_s": 816.87,
            "proc5_iter3_ttft_s": 1.27,
            "proc5_iter4_decode_tok_s": 69.53,
            "proc5_iter4_prefill_tok_s": 813.98,
            "proc5_iter4_ttft_s": 1.27,
            "proc5_peak_private_footprint_mb": 19470.67,
            "proc5_peak_system_ram_mb_max": 19474.0,
            "processes": 5.0
          },
          "output_match": null,
          "peak_mem_mb": 19480.72,
          "prefill_tokens_per_s": 815.32,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 1270.0
        },
        "source": "data/device_runs/0.19.1/2026-11-12/nemotron-3-nano-30b-a3b-int4-ekv4096__mac-studio-m4-max.json"
      }
    ]
  },
  "model": {
    "family": "nemotron-3-nano",
    "id": "nemotron-3-nano-30b-a3b-int4-ekv4096",
    "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 391 on the CPU and 411 on the GPU for the int4 files (this one among them), 401 for the mixed_int4 files and for the source model (card L59-L64).",
    "Which file: this one gives faster decode than mixed_int4 ekv4096 (card L36). On an Apple M4 Max (128 GB) its benchmark peaked at 20.4 GB on the GPU and 5.0 GB on the CPU (2026-11-12); speed still being tuned (updates follow) (card L27, L48-L49, L55)."
  ],
  "schema_version": "1.2"
}
