{ "artifacts": [ { "file": "Nemotron-3-Nano-4B_int8.litertlm", "sha256": "ba73d2ad1d878bc2b796eb3a1d993eb1b3bd5105743cd85bd6fa9d86e58a013d", "size_mb": 3935.525 } ], "benchmarks": [], "conversion": { "command": "python scripts/convert.py nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 (hf-to-litertlm)", "quantization": "int8 dynamic on linears + embedding (convs and the selective scan stay float, which keeps the hybrid state numerically sane); fp32 activations declared (README file table + Honest notes)", "tool": "litert-torch plus a hybrid-cache patch, one command via hf-to-litertlm (python scripts/convert.py nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16; 1645 s on an M4 Max) (README Conversion notes)", "tool_version": "TODO (the README names the converter but not its version)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 12.87, "delegated_ops": 3012, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-cold-cache-cooled", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 2643 out of 2827 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 218 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 2622 out of 2806 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 218 partitions for subgraph 1 (prefill_256).", "VERBOSE: Replacing 2725 out of 2909 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 218 partitions for subgraph 2 (prefill_64).", "VERBOSE: Replacing 2725 out of 2909 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 218 partitions for subgraph 3 (prefill_16).", "results block: prefill=52.94 decode=12.87 tokens/s" ], "failure_class": null, "full_delegation": false, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 706.0, "init_s": 8.98674, "prefill_tokens": 211.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 52.94, "provenance": "measured", "runs": true, "total_ops": 3155, "ttft_ms": 4060.0 }, "source": "data/device_runs/0.16.0/2026-09-05/nemotron-3-nano-4b-int8__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": null, "delegated_ops": 2827, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-cold-cache-cooled", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": "exit 139 during the gpu generation gate; log tail: local_attention_mask_policy: Not set | sliding_window_size: Not set | advanced_settings: Not set", "evidence": [ "VERBOSE: Replacing 2827 out of 2827 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "community_accel_work/s7_manifest_backfill/logs/Nemotron-3-Nano-4B__Nemotron-3-Nano-4B_int8.gate.log: generation gate verdict FAIL (journal rows.jsonl, fail_class=run_error, exit=139, generated=True, timeout=False); the same journal's --benchmark numbers for this leg are withheld because the gate output is not a real answer", "protocol: each bench run cold (caches deleted between runs), device cooled <42 C before every run; runtime litert_lm_advanced_main v0.16.0 (S4 kit)", "gate process peak 1280 MB (journal gate_peak_mb)" ], "failure_class": "run_error", "full_delegation": true, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": false, "total_ops": 2827, "ttft_ms": null }, "source": "data/device_runs/0.16.0/2026-09-05/nemotron-3-nano-4b-int8__galaxy-s26.json" }, { "device": "mac-studio-m4-max", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-27", "decode_tokens_per_s": 22.73, "delegated_ops": null, "env": { "device": "mac-studio-m4-max", "machine_label": "m4max-loaded-internvl-export", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Apple M4 Max", "vendor_sdk": null }, "error": null, "evidence": [ "results block: prefill=99.57 decode=22.73 tokens/s" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 256.0, "prefill_tokens": 256.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 99.57, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 2643.2 }, "source": "data/device_runs/0.16.0/2026-08-27/nemotron-3-nano-4b-int8__mac-studio-m4-max.json" }, { "device": "mac-studio-m4-max", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-27", "decode_tokens_per_s": 83.3, "delegated_ops": null, "env": { "device": "mac-studio-m4-max", "machine_label": "m4max-loaded-internvl-export", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Apple M4 Max", "vendor_sdk": null }, "error": null, "evidence": [ "results block: prefill=803.37 decode=83.3 tokens/s" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 256.0, "prefill_tokens": 256.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 803.37, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 330.7 }, "source": "data/device_runs/0.16.0/2026-08-27/nemotron-3-nano-4b-int8__mac-studio-m4-max.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-02", "decode_tokens_per_s": 2.2, "delegated_ops": null, "env": { "device": "Raspberry Pi 5 Model B Rev 1.1", "machine_label": "raspberry-pi-5-cooled-52c", "os_build": "Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41", "runtime": "litert-lm", "runtime_version": "0.16.1", "soc": "Broadcom BCM2712", "vendor_sdk": null }, "error": null, "evidence": [ "pi5 LLM sweep row: `litert-lm benchmark --backend cpu --cpu-thread-count 4 -p 256 -d 256 --runs 1 --cache memory` (wave-2 driver pi5_llm_bench.py; --cache memory rather than the house --cache no, which OOM-kills every >=1.2B file on the 8 GB Pi — equivalence measured on granite-350m int8, +2-3%), 3 invocations per file with cool-down to <=52 C between them, vcgencmd measure_temp + get_throttled logged per invocation, peak RSS polled from /proc; throughput = median of the three invocations (spread in metrics); a row counts as measured only when the real-generation gate (`litert-lm run`, degenerate-output check) passed and every invocation exited 0 with get_throttled 0x0", "versions: cpu=Raspberry Pi 5 Model B Rev 1.1, litert-lm=0.16.1, litert-lm-api=0.16.1, platform=Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41, python=3.13.5", "cache mode 'memory'; -p 256 -d 256 --runs 1 --cpu-thread-count 4", "gate ('What is 17 plus 26? Answer with the number only.'): status pass, exit 0, wall 144.5 s, output head '[thought] The user asks: \"What is 17 plus 26? Answer with the number only.\" So a'", "invocation 0: exit=0 wall_s=356.2 temp 48.3->51.6C throttled=0x0 prefill_tps=22.75 decode_tps=2.2 ttft_s=11.7051 init_s=94.7673 peak_rss_mb=5150", "invocation 1: exit=0 wall_s=350.2 temp 49.4->55.4C throttled=0x0 prefill_tps=22.59 decode_tps=2.2 ttft_s=11.7869 init_s=94.7575 peak_rss_mb=5150", "invocation 2: exit=0 wall_s=354.2 temp 48.3->52.1C throttled=0x0 prefill_tps=22.72 decode_tps=2.2 ttft_s=11.7215 init_s=94.6848 peak_rss_mb=5167" ], "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, "decode_tps_max": 2.2, "decode_tps_min": 2.2, "init_s": 94.7575, "init_s_max": 94.7673, "init_s_min": 94.6848, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 22.75, "prefill_tps_min": 22.59, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 11.7869, "ttft_s_min": 11.7051 }, "output_match": null, "peak_mem_mb": 5167.0, "prefill_tokens_per_s": 22.72, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 11721.5 }, "source": "data/device_runs/0.16.1/2026-09-02/nemotron-3-nano-4b-int8__raspberry-pi-5.json" } ] }, "model": { "family": "nemotron-3-nano", "id": "nemotron-3-nano-4b-int8", "license": "other (nvidia-nemotron-open-model-license)", "source_url": "https://huggingface.co/litert-community/Nemotron-3-Nano-4B", "task": "text-generation" }, "pitfalls": [ "Requires litert-lm >= 0.15 (README intro).", "GPU requires --cache no on this bundle: with the compiled-graph cache enabled, litert-lm run --backend gpu fails with WebGPU 'Invalid BindGroup' validation errors and the 8-question sweep returns token soup (0/8); the same file with --cache no answers 8/8 (README Honest notes).", "8-question gate 7/8 on CPU, 8/8 on GPU (litert-lm 0.16.0, M4 Max); the single CPU miss is the rhyme item ('violets are purple'); chat template byte-equal to the source (10,504/10,504 bytes — the source tokenizer_config.json carries a different 10,497-byte copy); stops <|im_end|> (id 11) + id 2; no spurious start token (README Correctness).", "Reasoning model — answers arrive after a block, so give it a token budget that fits the thought (the gate used 3200); >=3B exports use a reduced 7-signature prefill ladder because every signature costs engine RAM (README Honest + Conversion notes).", "Not measured on a phone by the README (Mac-only there; a 4B hybrid did not fit an 8 GB Android phone when Nemotron-H-4B was measured); this repo's Galaxy S26 rows come from the S7 backfill: CPU pass, GPU run_error exit 139 after full delegation (device runs).", "auto_map in the config is not proof of remote code: transformers registers nemotron_h natively, so the library implementation loads without trust_remote_code (README Conversion notes).", "LiteRT-LM .litertlm bundle: LiteRT.js cannot run it, so delegation stays null and there is no browser block; device rows come from data/device_runs/." ], "schema_version": "1.2" }