{ "artifacts": [ { "file": "LFM2.5-2.6B_int4.litertlm", "sha256": "f4797d5a16812231deb6d505b4f417dd835b9b3dc47bb74d9ec71facfd0c91a4", "size_mb": 1590.873 } ], "benchmarks": [], "conversion": { "command": "TODO (the HF card states the recipe and post-steps, not the invocation)", "quantization": "int4 blockwise-32 + OCTAV on linears, int8 embedding, convs float; the checkpoint is sparse — OCTAV int4-b32 produced 746k all-zero weight blocks whose zero scales XNNPACK rejects at load, patched to the tensor's smallest nonzero scale (dequantization unchanged) (HF card file table + Conversion notes)", "tool": "litert-torch (released 0.9.2 with upstream lfm2 hybrid support incl. the ShortConv prefill-pad fix), packaged for litert-lm >= 0.14 with the executor-metadata section binding the 22 conv states + 16 KV caches (HF card Conversion notes)", "tool_version": "0.9.2 (HF card Conversion notes)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 25.31, "delegated_ops": 936, "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 936 out of 936 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 936 out of 936 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 936 out of 936 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 936 out of 936 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).", "results block: prefill=297.42 decode=25.31 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": 3304.0, "init_s": 14.68967, "prefill_tokens": 206.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 297.42, "provenance": "measured", "runs": true, "total_ops": 936, "ttft_ms": 730.0 }, "source": "data/device_runs/0.16.0/2026-08-24/lfm2.5-2.6b-int4__galaxy-s26.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 4.91, "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 101.3 s, output head '[thought] The user is asking for the sum of 17 and 26.\\nCalculation: 17 + 26 = 43'", "invocation 0: exit=0 wall_s=202.1 temp 49.4->56.0C throttled=0x0 prefill_tps=28.5 decode_tps=4.91 ttft_s=9.1855 init_s=79.6785 peak_rss_mb=2965", "invocation 1: exit=0 wall_s=202.1 temp 51.0->54.9C throttled=0x0 prefill_tps=28.46 decode_tps=4.91 ttft_s=9.1992 init_s=79.1251 peak_rss_mb=2964", "invocation 2: exit=0 wall_s=202.1 temp 51.0->55.4C throttled=0x0 prefill_tps=28.49 decode_tps=4.9 ttft_s=9.1895 init_s=79.073 peak_rss_mb=2964" ], "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": 4.91, "decode_tps_min": 4.9, "init_s": 79.1251, "init_s_max": 79.6785, "init_s_min": 79.073, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 28.5, "prefill_tps_min": 28.46, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 9.1992, "ttft_s_min": 9.1855 }, "output_match": null, "peak_mem_mb": 2965.0, "prefill_tokens_per_s": 28.49, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 9189.5 }, "source": "data/device_runs/0.16.1/2026-09-01/lfm2.5-2.6b-int4__raspberry-pi-5.json" } ] }, "model": { "family": "lfm2.5", "id": "lfm2.5-2.6b-int4", "license": "other (lfm-open-license-v1.0)", "source_url": "https://huggingface.co/litert-community/LFM2.5-2.6B", "task": "text-generation" }, "pitfalls": [ "Requires litert-lm >= 0.14: the bundled executor-metadata section binds the 22 conv states + 16 KV caches (files exported without it do not run) and makes the same file run on 0.15's new state binding (HF card intro + Conversion notes).", "The generation prompt pre-fills exactly as the vendor chat template does; with a bare assistant prompt, think-block emission becomes the model's choice, and int4 degrades that discipline first (unscaffolded rambling in place of answers on multi-turn) (HF card Conversion notes).", "GSM8K (greedy, 0-shot CoT, max-tokens 2048, n=100, same harness for all rows): bf16 reference 92%, int8 88%, int4 83%; both files pass an 8-question sanity gate 8/8 with zero degenerate outputs (CPU), a 42-length prefill sweep with zero corrupt first tokens, and a 3-turn conversation gate (HF card file table + Quality).", "iPhone 17 Pro (CPU backend, cold first runs): the int4 file answers the 8-question gate 8/8 and decodes ~20 tok/s warm; the int8 file loads and runs only when the host app carries the extended-virtual-addressing / increased-memory entitlements (its 2.87 GB single weight section exceeds a default-entitlement app's mmap) — int4 is the recommended phone variant (HF card Speed).", "Multi-length prefill signatures (1-1024) are exported so the runtime picks tight chunks (HF card 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/ (Pi 5 / S26 / Pixel 8a / Mac / iPhone as measured)." ], "schema_version": "1.2" }