{ "artifacts": [ { "file": "model_block128.litertlm", "sha256": "3842a564a1f8db2ed53992ec6e2d5e2579aff72131dfac664c19cf2ec90542ce", "size_mb": 2359.731 } ], "benchmarks": [], "conversion": { "command": "from litert_torch.generative.export_hf.export import export; export(model='microsoft/FastContext-1.0-4B-SFT', output_dir='out', quantization_recipe='qwen3_int4_block32_octav.json' | 'qwen3_int4_block128_octav.json', cache_length=4096, externalize_embedder=True) (HF card Conversion snippet)", "quantization": "int4 blockwise (block 128) + OCTAV, symmetric; INT8 embedding externalized; KV cache 4096 (HF card)", "tool": "litert-torch export_hf (official converter; FastContext is a standard Qwen3ForCausalLM, existing Qwen3 path, no custom graph code) (HF card Conversion)", "tool_version": "TODO (the HF card names the converter but not its version; no export log in the sources)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 7.49, "delegated_ops": 1674, "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 1674 out of 1674 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).", "VERBOSE: Replacing 1596 out of 1596 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (decode).", "results block: prefill=86.28 decode=7.49 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": 727.0, "init_s": 12.40631, "prefill_tokens": 202.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 86.28, "provenance": "measured", "runs": true, "total_ops": 1674, "ttft_ms": 2470.0 }, "source": "data/device_runs/0.16.0/2026-08-24/fastcontext-1.0-4b-sft-model-block128__galaxy-s26.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 101.4 s, output head '43'", "invocation 0: exit=0 wall_s=506.4 temp 49.4->52.7C throttled=0x0 prefill_tps=10.32 decode_tps=1.49 ttft_s=26.2791 init_s=55.9504 peak_rss_mb=4009", "invocation 1: exit=0 wall_s=500.3 temp 51.6->52.1C throttled=0x0 prefill_tps=10.47 decode_tps=1.48 ttft_s=25.7707 init_s=55.4192 peak_rss_mb=3951", "invocation 2: exit=0 wall_s=494.3 temp 49.9->54.9C throttled=0x0 prefill_tps=10.33 decode_tps=1.5 ttft_s=26.0396 init_s=55.3038 peak_rss_mb=3968" ], "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": 1.48, "init_s": 55.4192, "init_s_max": 55.9504, "init_s_min": 55.3038, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 10.47, "prefill_tps_min": 10.32, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 26.2791, "ttft_s_min": 25.7707 }, "output_match": null, "peak_mem_mb": 4009.0, "prefill_tokens_per_s": 10.33, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 26039.6 }, "source": "data/device_runs/0.16.1/2026-09-02/fastcontext-1.0-4b-sft-model-block128__raspberry-pi-5.json" } ] }, "model": { "family": "qwen3", "id": "fastcontext-1.0-4b-sft-model-block128", "license": "mit", "source_url": "https://huggingface.co/litert-community/FastContext-1.0-4B-SFT", "task": "text-generation" }, "pitfalls": [ "Which file: block 32 (model.litertlm) is the quality pick — GSM8K 88.0% vs 81.0% for block 128 — while block 128 (model_block128.litertlm) decodes ~40% faster on the iPhone 17 Pro (14 vs 10 tok/s; Mac M-series GPU 66-68 vs 64-73 tok/s) (HF card 'Which file?').", "Blockwise (not the tool's default channelwise) int4 is what preserves accuracy; embeddings kept at INT8 and externalized into their own section (dedups the tied matrix) (HF card Conversion).", "License MIT, inherited from microsoft/FastContext-1.0-4B-SFT (itself built on Qwen3-4B-Instruct) (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" }