{ "artifacts": [ { "file": "LFM2.5-1.2B-Thinking_int8_gpu.litertlm", "sha256": "11faec13dca0699dae1a43da30f0b5d57c64627edf6c4a6fa16b45b716db1a92", "size_mb": 1187.099 } ], "benchmarks": [], "conversion": { "command": "minicpm5_work/ship_lfm25_int8gpu_20260825/build_int8_gpu.sh", "quantization": "int8 dynamic (linears + embedding via wi8fc; convs float) — GSM8K 77% vs 76% for export-time conv-int8 on this finetune (build_int8_gpu.sh primary pointers)", "tool": "litert-torch 0.9.3 lineage export + post-hoc ai-edge-quantizer (wi8fc recipe), per build_int8_gpu.sh", "tool_version": "same toolchain as ship_lfm25_gpu_variant_20260812/build_gpu_variants.sh (0.9.3-lineage GPU-safe ShortConv); quantizer recipe wi8fc" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-25", "decode_tokens_per_s": 28.83, "delegated_ops": 542, "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 542 out of 542 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 542 out of 542 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 542 out of 542 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 542 out of 542 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).", "results block: prefill=860.26 decode=28.83 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": 3058.0, "init_s": 7.34249, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 860.26, "provenance": "measured", "runs": true, "total_ops": 542, "ttft_ms": 270.0 }, "source": "data/device_runs/0.16.0/2026-08-25/lfm25-12b-thinking-int8-gpu__galaxy-s26.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 7.03, "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 98.5 s, output head '[thought] Okay, let me see. The question is asking \"What is 17 plus 26? Answer '", "invocation 0: exit=0 wall_s=108.1 temp 50.5->54.9C throttled=0x0 prefill_tps=75.12 decode_tps=7.03 ttft_s=3.5504 init_s=30.9851 peak_rss_mb=1884", "invocation 1: exit=0 wall_s=112.1 temp 51.6->54.3C throttled=0x0 prefill_tps=75.47 decode_tps=7.04 ttft_s=3.534 init_s=30.7704 peak_rss_mb=1893", "invocation 2: exit=0 wall_s=112.1 temp 49.9->53.8C throttled=0x0 prefill_tps=74.91 decode_tps=6.99 ttft_s=3.5604 init_s=30.6476 peak_rss_mb=1885" ], "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": 7.04, "decode_tps_min": 6.99, "init_s": 30.7704, "init_s_max": 30.9851, "init_s_min": 30.6476, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 75.47, "prefill_tps_min": 74.91, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 3.5604, "ttft_s_min": 3.534 }, "output_match": null, "peak_mem_mb": 1893.0, "prefill_tokens_per_s": 75.12, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 3550.4 }, "source": "data/device_runs/0.16.1/2026-09-01/lfm25-12b-thinking-int8-gpu__raspberry-pi-5.json" } ] }, "model": { "family": "lfm2.5", "id": "lfm25-12b-thinking-int8-gpu", "license": "lfm-open-license-v1.0", "source_url": "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking", "task": "text-generation" }, "pitfalls": [ "The published _int4/_int8 (non-gpu) siblings stall GPU delegation at 536/579 ops on Adreno (0.9.1-era ShortConv INT64 block; s4_gpu_gate/RESULTS.md rows 5/12). This _int8_gpu delegates 542/542 on the Galaxy S26 (device_runs 2026-08-25).", "Thinking bundle: the runtime needs the think opener prefilled and the bundle's declared thought channel to show reasoning; gate multi-turn behaviour at sizes where the thinking completes (family records; memory thinking-models-need-think-prefill).", "add_executor_metadata is required after post-hoc quantization (0.9.3 emits no ExecutorMetadataProto; litert-lm >=0.15 needs it) (build_int8_gpu.sh).", "One arm of the 08-25 Mac Metal sieve refused engine creation for a JP/Thinking variant while S26 OpenCL passed the same file — see ship_lfm25_int8gpu_20260825/RESULTS.md before quoting Mac Metal numbers for this family." ], "schema_version": "1.2" }