{ "artifacts": [ { "file": "Qwen2.5-Coder-1.5B-Instruct_int4.litertlm", "sha256": "273ecc7771ba2dd5fe1bb6d4d4726ad0353102f04ad094082ccf59bca9f21213", "size_mb": 1065.622 } ], "benchmarks": [], "conversion": { "command": "CACHE=4096 PREFILL=1024,256,64,16,4,1 EXTERNALIZE_EMBEDDER=1 python scripts/export_simple_template.py src_models/qwen25-coder-1.5b out_qwen25_coder qwen25coder_work/qwen25_coder_simple.jinja BOCTAV4", "quantization": "int4 weights, blockwise-32 + OCTAV, symmetric (BOCTAV4) on linears; int8 embedding, externalised (tied lm_head)", "tool": "litert-torch (hf-to-litertlm `qwen25coder_work/convert_qwen25_coder.sh`; export_simple_template.py)", "tool_version": "litert-torch 0.9.3 (litert-converter 0.3.1, ai-edge-quantizer 0.8.0, litert-lm-builder 0.16.0) — stated on the published HF card" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 29.73, "delegated_ops": 1510, "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 1125 out of 1243 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 165 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 1125 out of 1243 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 165 partitions for subgraph 1 (prefill_256).", "VERBOSE: Replacing 1125 out of 1243 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 165 partitions for subgraph 2 (prefill_64).", "VERBOSE: Replacing 1125 out of 1243 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 165 partitions for subgraph 3 (prefill_16).", "results block: prefill=106.0 decode=29.73 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": 409.0, "init_s": 2.73147, "prefill_tokens": 223.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 106.0, "provenance": "measured", "runs": true, "total_ops": 1573, "ttft_ms": 2140.0 }, "source": "data/device_runs/0.16.0/2026-09-05/qwen25-coder-1.5b-int4__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 10.07, "delegated_ops": 1243, "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 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).", "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64).", "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_16).", "results block: prefill=205.89 decode=10.07 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": 360.0, "init_s": 6.9454, "prefill_tokens": 223.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 205.89, "provenance": "measured", "runs": true, "total_ops": 1243, "ttft_ms": 1180.0 }, "source": "data/device_runs/0.16.0/2026-08-24/qwen25-coder-1.5b-int4__galaxy-s26.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-19", "decode_tokens_per_s": 12.57, "delegated_ops": null, "env": { "device": "Pixel 8a", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Tensor G3", "vendor_sdk": null }, "error": null, "evidence": [ "results block: prefill=16.09 decode=12.57 tokens/s" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 32.0, "init_s": 4.94288, "prefill_tokens": 38.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 16.09, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 2440.0 }, "source": "data/device_runs/0.16.0/2026-08-19/qwen25-coder-1.5b-int4__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-19", "decode_tokens_per_s": 12.0, "delegated_ops": 1243, "env": { "device": "Pixel 8a", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Tensor G3", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256).", "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64).", "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_16).", "results block: prefill=83.7 decode=12.0 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": 32.0, "init_s": 15.11825, "prefill_tokens": 38.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 83.7, "provenance": "measured", "runs": true, "total_ops": 1243, "ttft_ms": 540.0 }, "source": "data/device_runs/0.16.0/2026-08-19/qwen25-coder-1.5b-int4__pixel-8a.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 5.61, "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 31.0 s, output head '43'", "invocation 0: exit=0 wall_s=162.1 temp 46.6->54.3C throttled=0x0 prefill_tps=40.64 decode_tps=5.5 ttft_s=7.3265 init_s=33.504 peak_rss_mb=1842", "invocation 1: exit=0 wall_s=152.1 temp 49.4->53.2C throttled=0x0 prefill_tps=41.91 decode_tps=5.61 ttft_s=7.1984 init_s=33.6273 peak_rss_mb=1840", "invocation 2: exit=0 wall_s=152.1 temp 49.9->51.6C throttled=0x0 prefill_tps=41.69 decode_tps=5.67 ttft_s=7.2692 init_s=33.6107 peak_rss_mb=1835" ], "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": 5.67, "decode_tps_min": 5.5, "init_s": 33.6107, "init_s_max": 33.6273, "init_s_min": 33.504, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 41.91, "prefill_tps_min": 40.64, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 7.3265, "ttft_s_min": 7.1984 }, "output_match": null, "peak_mem_mb": 1842.0, "prefill_tokens_per_s": 41.69, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 7269.2 }, "source": "data/device_runs/0.16.1/2026-09-01/qwen25-coder-1.5b-int4__raspberry-pi-5.json" } ] }, "model": { "family": "qwen2.5-coder", "id": "qwen25-coder-1.5b-int4", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/Qwen2.5-Coder-1.5B-Instruct", "task": "text-generation" }, "pitfalls": [ "EXTERNALIZE_EMBEDDER=1 is required, not cosmetic: the model ties embedding and lm_head, and BOCTAV4 (int4 linears + int8 embedding) describes one tensor two ways — the quantizer resolves it by copying the 151936x1536 table once per prefill signature (7 copies, 1.63 GB, 65% of the file, no error). Externalising removes the conflict (HF card / REPRODUCE.md).", "The upstream chat template injects a default system prompt ('You are Qwen, created by Alibaba Cloud…') when none is supplied; a generic ChatML template omits it and runs the model outside its tuned state — the bundle carries a template that reproduces the default (FINDINGS.md / REPRODUCE.md).", "Requires litert-lm >= 0.16 (HF card).", "Pixel 8a 2026-08-19: fully delegated on Mali OpenCL (1243/1243 per prefill signature, 1132/1132 decode, zero rejections); GPU buys prefill/TTFT (0.54 s vs 2.44 s on CPU), not decode (12.0 vs 12.6 tok/s — same LPDDR) (FINDINGS.md)." ], "schema_version": "1.2" }