{ "artifacts": [ { "file": "SmolLM3-3B_q4_block32_ekv4096.litertlm", "sha256": "38d7bf55e243f5a95d8b16c21b1f7ef7debe6ce6bb2ddb66eb54f4bc2be8e6eb", "size_mb": 1909.502 } ], "benchmarks": [], "conversion": { "command": "CACHE=4096 EXTERNALIZE_EMBEDDER=1 export_simple_template.py HuggingFaceTB/SmolLM3-3B templates/smollm3_think.jinja BOCTAV4", "quantization": "int4 weights, blockwise-32 + OCTAV optimal clipping, symmetric; embedding INT8", "tool": "litert-torch (hf-to-litertlm reproduce_llm.sh key `smollm3-3b`)", "tool_version": "TODO" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 9.81, "delegated_ops": 1894, "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 1230 out of 1382 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 220 partitions for subgraph 0 (decode).", "VERBOSE: Replacing 1397 out of 1548 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 215 partitions for subgraph 1 (prefill_256).", "VERBOSE: Replacing 6 out of 6 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 2 (odml.rms_norm.impl).", "VERBOSE: Replacing 6 out of 6 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 3 (odml.rms_norm.impl_0).", "results block: prefill=147.93 decode=9.81 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": 912.0, "init_s": 9.12393, "prefill_tokens": 442.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 147.93, "provenance": "measured", "runs": true, "total_ops": 1974, "ttft_ms": 3090.0 }, "source": "data/device_runs/0.16.0/2026-09-05/smollm3-3b__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 14.84, "delegated_ops": 1476, "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 1476 out of 1476 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).", "VERBOSE: Replacing 1308 out of 1308 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (decode).", "results block: prefill=267.97 decode=14.84 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": 358.0, "init_s": 16.74028, "prefill_tokens": 202.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 267.97, "provenance": "measured", "runs": true, "total_ops": 1476, "ttft_ms": 820.0 }, "source": "data/device_runs/0.16.0/2026-08-24/smollm3-3b__galaxy-s26.json" }, { "device": "mac-studio-m4-max", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-07-22", "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": "0.14.0", "soc": "Apple M4 Max", "vendor_sdk": null }, "error": null, "evidence": [ "speed figures withheld: this measurement batch was retracted as contaminated (parallel GPU load; owner correction in gpu_audit MATRIX.md, 2026-07-23) — the PASS/FAIL verdict is load-independent and stands" ], "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, "max_num_tokens": 1024.0, "prefill_tokens": 256.0 }, "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/0.14.0/2026-07-22/smollm3-3b__mac-studio-m4-max.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-17", "decode_tokens_per_s": 7.69, "delegated_ops": 1476, "env": { "device": "Pixel 8a", "machine_label": "pixel-8a", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Tensor G3", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 1476 out of 1476 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).", "VERBOSE: Replacing 1308 out of 1308 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (decode).", "results block: prefill=11.89 decode=7.69 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": 104.0, "init_s": 18.34859, "prefill_tokens": 17.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 11.89, "provenance": "measured", "runs": true, "total_ops": 1476, "ttft_ms": 1560.0 }, "source": "data/device_runs/0.16.0/2026-08-17/smollm3-3b__pixel-8a.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 2.46, "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 193.1 s, output head \"[thought] Okay, let's see. The user is asking for the sum of 17 and 26. I need t\"", "invocation 0: exit=0 wall_s=316.2 temp 51.6->52.7C throttled=0x0 prefill_tps=18.26 decode_tps=2.48 ttft_s=16.3785 init_s=49.1304 peak_rss_mb=3193", "invocation 1: exit=0 wall_s=316.2 temp 48.8->54.9C throttled=0x0 prefill_tps=18.38 decode_tps=2.46 ttft_s=16.2956 init_s=48.9598 peak_rss_mb=3194", "invocation 2: exit=0 wall_s=318.2 temp 51.0->52.7C throttled=0x0 prefill_tps=18.32 decode_tps=2.45 ttft_s=16.3289 init_s=49.0204 peak_rss_mb=3194" ], "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": 2.48, "decode_tps_min": 2.45, "init_s": 49.0204, "init_s_max": 49.1304, "init_s_min": 48.9598, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 18.38, "prefill_tps_min": 18.26, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 16.3785, "ttft_s_min": 16.2956 }, "output_match": null, "peak_mem_mb": 3194.0, "prefill_tokens_per_s": 18.32, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 16328.9 }, "source": "data/device_runs/0.16.1/2026-09-01/smollm3-3b__raspberry-pi-5.json" } ] }, "model": { "family": "smollm3", "id": "smollm3-3b", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/SmolLM3-3B", "task": "text-generation" }, "pitfalls": [ "Thinking model: the bundle carries the SmolLM3 template with its /think and /no_think controls, not plain ChatML (HF card).", "Gallery import: only v1.0.16+ accepts .litertlm directly from Hugging Face; older 1.0.x builds (package com.google.aiedge.gallery) do not (HF card).", "Measured on an 8 GB phone 2026-08-17: fully delegated on Mali OpenCL — 1476/1476 prefill nodes, 1308/1308 decode, zero rejections — and answers correctly. This refutes the earlier standing assumption that a ~2B-and-up dense model cannot use the Pixel 8a GPU at all (CATALOG_REGATE.md)." ], "schema_version": "1.2" }