{ "artifacts": [ { "file": "LFM2.5-1.2B-Instruct_int8.litertlm", "sha256": "e002a91545cec6328be2a86e9b8b4fccb2dd9dde7e6c8b29d730f915a0a697fe", "size_mb": 1189.319 } ], "benchmarks": [], "conversion": { "command": "python convert_lfm25.py LiquidAI/LFM2.5-1.2B-Instruct out_lfm25_12b", "quantization": "int8 dynamic, export-time recipe (linears + convs + embedding)", "tool": "litert-torch", "tool_version": "0.9.1" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 40.96, "delegated_ops": 772, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-cold-cache-cooled", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 3 (prefill_128).", "results block: prefill=491.11 decode=40.96 tokens/s" ], "failure_class": null, "full_delegation": false, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 105.0, "init_s": 1.93164, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 491.11, "provenance": "measured", "runs": true, "total_ops": 837, "ttft_ms": 440.0 }, "source": "data/device_runs/0.16.0/2026-09-05/lfm2.5-1.2b-instruct-int8__galaxy-s26.json" }, { "device": "mac-studio-m4-max", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-06", "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": "0.17.0", "soc": null, "vendor_sdk": null }, "error": null, "evidence": [ "compat_check status ok (runtime /Users/USER/code/litertlm-convert/.qa-venvs/litert-lm-0.17.0/bin/litert-lm 0.17.0); fixed-question answer: '17 + 25 = 42'" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "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/0.17.0/2026-09-06/lfm2.5-1.2b-instruct-int8__mac-studio-m4-max.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 20.4, "delegated_ops": 772, "env": { "device": "Pixel 8a", "machine_label": "pixel-8a-cold-cache-cooled", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Tensor G3", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 471 out of 579 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 109 partitions for subgraph 3 (prefill_128).", "results block: prefill=143.38 decode=20.4 tokens/s" ], "failure_class": null, "full_delegation": false, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 105.0, "init_s": 7.9068, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 143.38, "provenance": "measured", "runs": true, "total_ops": 837, "ttft_ms": 1480.0 }, "source": "data/device_runs/0.16.0/2026-09-05/lfm2.5-1.2b-instruct-int8__pixel-8a.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 7.02, "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 15.1 s, output head '43'", "invocation 0: exit=0 wall_s=112.1 temp 51.0->53.8C throttled=0x0 prefill_tps=74.45 decode_tps=7.02 ttft_s=3.581 init_s=30.5554 peak_rss_mb=1894", "invocation 1: exit=0 wall_s=112.1 temp 49.4->53.8C throttled=0x0 prefill_tps=75.73 decode_tps=7.06 ttft_s=3.5221 init_s=30.6025 peak_rss_mb=1895", "invocation 2: exit=0 wall_s=112.1 temp 51.0->54.9C throttled=0x0 prefill_tps=74.95 decode_tps=7.0 ttft_s=3.5582 init_s=30.6072 peak_rss_mb=1895" ], "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": 7.06, "decode_tps_min": 7.0, "init_s": 30.6025, "init_s_max": 30.6072, "init_s_min": 30.5554, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 75.73, "prefill_tps_min": 74.45, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 3.581, "ttft_s_min": 3.5221 }, "output_match": null, "peak_mem_mb": 1895.0, "prefill_tokens_per_s": 74.95, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 3558.2 }, "source": "data/device_runs/0.16.1/2026-09-01/lfm2.5-1.2b-instruct-int8__raspberry-pi-5.json" } ] }, "model": { "family": "lfm2.5", "id": "lfm2.5-1.2b-instruct-int8", "license": "lfm-open-license-v1.0", "source_url": "https://huggingface.co/litert-community/LFM2.5-1.2B-Instruct", "task": "text-generation" }, "pitfalls": [ "Export-time int8 is what makes this file work: it is the only path that safely quantizes the conv layers. Post-hoc ALL_SUPPORTED int8 through ai-edge-quantizer kills them outright (no output), so post-hoc recipes must stay on linears and the embedding (wi8fc, wi4b32_wi8) — REPRODUCE.md, LFM2.5 family.", "Conv-int8 sensitivity is per-finetune, not a family property: export-time conv-int8 is free on this Instruct tune (+2 GSM8K) but costs the JP tune 9 points, which is why the published JP int8 file uses a linears-only recipe instead. A/B the two before reusing this command on a new finetune.", "This artifact cannot use a GPU delegate. It is litert-torch 0.9.1 lineage, whose ShortConv patch emits GATHER_ND and INT64 ops that GPU delegates reject; the delegate takes 536 of 579 operations and engine creation then aborts. That count was measured on the int4 sibling (Pixel 8a and Galaxy S26, litert-lm v0.16.0) — the same export lineage, but this specific file has not been separately gated on GPU. The repo ships no int8 GPU variant; the GPU re-export exists only for int4.", "The 0.9.1 exporter needs the ShortConv prefill-pad fix that convert_lfm25.py applies: the stock block saves its conv state from the padded columns of a prefill chunk, corrupting the first generated token of nearly every reply. Without it this int8 file scores 59 on GSM8K instead of 81 — that is the bug, not the quantization.", "litert-lm >= 0.15 needs an ExecutorMetadata section for this hybrid: files exported before that run on 0.14 but fail at inference on 0.15 with 'missing some output TensorBuffers'. The published files were repaired in place on 2026-08-04.", "Decode speed depends strongly on the KV budget: the HF card measures int8 decode at 101 tok/s with --max-num-tokens 1024 and 77 tok/s at 4096 on an M4 Max. Set the smallest budget the use case needs." ], "schema_version": "1.2" }