{ "artifacts": [ { "file": "Ministral-3-3B-Instruct-2512_q4_block32_ekv4096.litertlm", "sha256": "e68fcf0d52e5e9d86663c819b8ab53998694a633190b9102b456e12de4f16270", "size_mb": 2232.535 } ], "benchmarks": [], "conversion": { "command": "extract_ministral3_text.py (drops the Pixtral vision tower) -> CACHE=4096 EXTERNALIZE_EMBEDDER=1 export_simple_template.py templates/mistral_simple.jinja BOCTAV4", "quantization": "int4 weights, BLOCKWISE_32 symmetric + OCTAV; tied embedding/lm_head INT8", "tool": "litert-torch (hf-to-litertlm reproduce_llm.sh key `ministral3-3b`)", "tool_version": "TODO" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 11.26, "delegated_ops": 1187, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-cl-pinned", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 1187 out of 1187 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).", "VERBOSE: Replacing 1087 out of 1087 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (decode).", "results block: prefill=223.58 decode=11.26 tokens/s" ], "failure_class": null, "full_delegation": true, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 1429.0, "init_s": 16.6479, "prefill_tokens": 198.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 223.58, "provenance": "measured", "runs": true, "total_ops": 1187, "ttft_ms": 970.0 }, "source": "data/device_runs/0.16.0/2026-08-24/ministral3-3b__galaxy-s26.json" }, { "device": "mac-studio-m4-max", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-07-22", "decode_tokens_per_s": null, "delegated_ops": null, "env": { "device": "Mac Studio (M4 Max)", "machine_label": "mac-studio-m4-max", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.14.0", "soc": "Apple M4 Max", "vendor_sdk": null }, "error": null, "evidence": [ "speed figures withheld: this measurement batch was retracted as contaminated (parallel GPU load; owner correction in gpu_audit MATRIX.md, 2026-07-23) — the PASS/FAIL verdict is load-independent and stands" ], "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, "max_num_tokens": 1024.0, "prefill_tokens": 256.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": null }, "source": "data/device_runs/0.14.0/2026-07-22/ministral3-3b__mac-studio-m4-max.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-07-23", "decode_tokens_per_s": null, "delegated_ops": null, "env": { "device": "Pixel 8a", "machine_label": "pixel-8a-gallery", "os_build": null, "runtime": "litert-lm", "runtime_version": "gallery-1.0.15", "soc": "Google Tensor G3", "vendor_sdk": null }, "error": "app process died ~90 s into the benchmark (no Java exception) — suspected 3B-class GPU memory exhaustion on Pixel 8a", "evidence": [ "hand-recorded device phase: litertlm-convert reports/gpu_audit/MATRIX.md 実機フェーズ table (committed 0559ad3, 2026-07-23; Gallery 1.0.15 official Benchmark, ML Drift GPU, p256/d256 x3 runs; runtime version axis = the Gallery app version — the embedded litert-lm version is not recorded)", "row: '❌ アプリごと native 死 — ベンチ~90秒でプロセス死 = 3B 級 GPU メモリ枯渇の疑い'" ], "failure_class": "process_crash", "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": {}, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": false, "total_ops": null, "ttft_ms": null }, "source": "data/device_runs/gallery-1.0.15/2026-07-23/ministral3-3b__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-17", "decode_tokens_per_s": 6.8, "delegated_ops": 1187, "env": { "device": "Pixel 8a", "machine_label": "pixel-8a", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Tensor G3", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 1187 out of 1187 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).", "VERBOSE: Replacing 1087 out of 1087 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (decode).", "results block: prefill=8.49 decode=6.8 tokens/s" ], "failure_class": null, "full_delegation": true, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 124.0, "init_s": 20.85413, "prefill_tokens": 13.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 8.49, "provenance": "measured", "runs": true, "total_ops": 1187, "ttft_ms": 1680.0 }, "source": "data/device_runs/0.16.0/2026-08-17/ministral3-3b__pixel-8a.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 1.91, "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 102.5 s, output head '43'", "invocation 0: exit=0 wall_s=408.3 temp 51.0->54.9C throttled=0x0 prefill_tps=13.74 decode_tps=1.89 ttft_s=21.6413 init_s=56.9749 peak_rss_mb=3812", "invocation 1: exit=0 wall_s=404.3 temp 49.9->54.3C throttled=0x0 prefill_tps=13.73 decode_tps=1.91 ttft_s=21.2617 init_s=55.1212 peak_rss_mb=3828", "invocation 2: exit=0 wall_s=402.3 temp 49.4->56.5C throttled=0x0 prefill_tps=13.85 decode_tps=1.91 ttft_s=21.1142 init_s=55.255 peak_rss_mb=3796" ], "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": 1.91, "decode_tps_min": 1.89, "init_s": 55.255, "init_s_max": 56.9749, "init_s_min": 55.1212, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 13.85, "prefill_tps_min": 13.73, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 21.6413, "ttft_s_min": 21.1142 }, "output_match": null, "peak_mem_mb": 3828.0, "prefill_tokens_per_s": 13.74, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 21261.7 }, "source": "data/device_runs/0.16.1/2026-09-01/ministral3-3b__raspberry-pi-5.json" } ] }, "model": { "family": "mistral", "id": "ministral3-3b", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/Ministral-3-3B-Instruct-2512", "task": "text-generation" }, "pitfalls": [ "Text-only conversion: the Ministral-3 text decoder, with the Pixtral vision tower dropped (HF card).", "Embedding externalized so every section stays under 2 GiB, which is what lets it load on iOS (HF card).", "The bundle carries a start_token (), and here that is CORRECT rather than the Granite defect: upstream bos is and eos is , different tokens, so prepending BOS is the Mistral convention. The Granite trap needs bos == eos (or add_bos_token False) to bite (CATALOG_REGATE.md).", "Two true statements that must not be collapsed: the Gallery only OFFERS GPU on ~12 GB+ devices, because its accelerator choice is fixed at import time from device RAM; the MODEL itself runs fully delegated on an 8 GB Pixel 8a GPU when litert_lm_main drives it directly — 1187/1187 prefill, 1087/1087 decode, zero rejections, correct output, measured 2026-08-17. The first is a property of the app, the second of the runtime (HF card + CATALOG_REGATE.md)." ], "schema_version": "1.2" }