{ "artifacts": [ { "file": "model.litertlm", "sha256": "e1edad5042c8354bec1d6187c22cad9e460a38428da517fc7ef22b3aa828eb3b", "size_mb": 2359.731 } ], "benchmarks": [], "conversion": { "command": "FORCE_SPM=1 EXTERNALIZE_EMBEDDER=1 CACHE=4096 python scripts/export_simple_template.py POLARIS-Project/Polaris-4B-Preview out/polaris-4b templates/qwen3_think.jinja BOCTAV4_128 ; python qa/add_thought_channel.py in.litertlm out.litertlm --start '' --end '' (2026-08-25 staged meta; FORCE_SPM because the published bundle carries an SP_Tokenizer section)", "quantization": "int4 blockwise (block 128) + OCTAV, symmetric; embedding INT8 (externalized section); KV cache 4096; ChatML template (HF card Conversion)", "tool": "litert-torch via litertlm-convert scripts/export_simple_template.py (reproduce_llm.sh key polaris-4b), then the 2026-08-22 metadata-only thought-channel re-ship (qa/add_thought_channel.py through the litert-lm 0.16.0 CLI; every tflite section byte-identical)", "tool_version": "TODO (owner to fill — searched 2026-08-25 four ways: no convert log, reproduce_llm.sh pins no version, the bundle header carries no converter version, the tflite sections carry only 'MLIR Converted.'; the RE-SHIP step is pinned to the litert-lm 0.16.0 CLI)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "mac-studio-m4-max", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-07-23", "decode_tokens_per_s": 68.87, "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": [ "results block: prefill=971.19 decode=68.87 tokens/s, init=1.9649 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, "init_s": 1.9649, "max_num_tokens": 1024.0, "prefill_tokens": 256.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 971.19, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 278.1 }, "source": "data/device_runs/0.14.0/2026-07-23/polaris-4b__mac-studio-m4-max.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-02", "decode_tokens_per_s": 1.49, "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 536.3 s, output head '[thought] Okay, so I need to figure out what 17 plus 26 is. Let me start by brea'", "invocation 0: exit=0 wall_s=500.4 temp 51.0->54.3C throttled=0x0 prefill_tps=10.54 decode_tps=1.49 ttft_s=25.7184 init_s=56.6486 peak_rss_mb=3959", "invocation 1: exit=0 wall_s=496.4 temp 51.6->53.2C throttled=0x0 prefill_tps=10.77 decode_tps=1.49 ttft_s=25.1082 init_s=55.4097 peak_rss_mb=3927", "invocation 2: exit=0 wall_s=496.3 temp 49.4->52.1C throttled=0x0 prefill_tps=10.54 decode_tps=1.5 ttft_s=25.6566 init_s=55.3922 peak_rss_mb=3958" ], "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": 1.5, "decode_tps_min": 1.49, "init_s": 55.4097, "init_s_max": 56.6486, "init_s_min": 55.3922, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 10.77, "prefill_tps_min": 10.54, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 25.7184, "ttft_s_min": 25.1082 }, "output_match": null, "peak_mem_mb": 3959.0, "prefill_tokens_per_s": 10.54, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 25656.6 }, "source": "data/device_runs/0.16.1/2026-09-02/polaris-4b__raspberry-pi-5.json" } ] }, "model": { "family": "qwen3", "id": "polaris-4b", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/Polaris-4B-Preview", "task": "text-generation" }, "pitfalls": [ "Standard dense Qwen3ForCausalLM with rope_scaling: yarn, exported with a cache within original_max_position_embeddings so base RoPE is exact; externalize_embedder keeps every section under the iOS ~2 GiB single-section mmap limit so it loads on iPhone (HF card Conversion).", "License Apache-2.0, inherited from POLARIS-Project/Polaris-4B-Preview (HF card License).", "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/ (Pi 5 / S26 / Pixel 8a / Mac / iPhone as measured)." ], "schema_version": "1.2" }