{
  "artifacts": [
    {
      "file": "LFM2.5-1.2B-JP_int4_gpu.litertlm",
      "sha256": "2c826a9cdeefc614c9b7de1d1c5565c9a64f704e5f3291ad2cd5d7da396790d5",
      "size_mb": 702.115
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "python convert_lfm25.py LiquidAI/LFM2.5-1.2B-JP out_jp_fp --fp && python quantize_litertlm.py apply out_jp_fp/model.litertlm LFM2.5-1.2B-JP_int4_gpu.litertlm --recipe wi4b32_wi8 --algo octav && python scripts/add_executor_metadata.py  # RESULTS.md pipeline; upstream ShortConv fix, composite lowered by converter 0.3.1",
    "quantization": "int4 blockwise-32 + OCTAV linears, int8 embedding, convs float (same recipe as the CPU int4 file, re-exported so it runs on the GPU — HF card)",
    "tool": "litert-torch (via litertlm-convert minicpm5_work convert_lfm25.py + quantize_litertlm.py + add_executor_metadata.py)",
    "tool_version": "0.9.3 (~/venvs/ltconv040dev: litert-converter 0.3.1, ai-edge-quantizer 0.8.0, litert-lm 0.15.0 builder; ship_lfm25_gpu_variant_20260812/RESULTS.md)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-08-24",
          "decode_tokens_per_s": 54.63,
          "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=918.98 decode=54.63 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": 298.0,
            "init_s": 3.23858,
            "prefill_tokens": 205.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 918.98,
          "provenance": "measured",
          "runs": true,
          "total_ops": 542,
          "ttft_ms": 240.0
        },
        "source": "data/device_runs/0.16.0/2026-08-24/lfm25-12b-jp-int4-gpu-093__galaxy-s26.json"
      },
      {
        "device": "iphone-17-pro",
        "run": {
          "accelerator": "gpu",
          "context_length": 1024,
          "date": "2026-08-25",
          "decode_tokens_per_s": 70.02,
          "delegated_ops": null,
          "env": {
            "device": "iphone-17-pro",
            "machine_label": "iphone-17-pro-charging",
            "os_build": "iOS 27.0",
            "runtime": "litert-lm",
            "runtime_version": "0.16.0",
            "soc": "A19-Pro",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "maxNumTokens=1024; all Metal kernels initialize, zero half4/float4, coherent 128-token output (distinct phrasing from the Instruct file = no cache contamination across the slot swap)",
            "prefill counter returned 0 on this cell (harness metric gap; generation itself normal); device charging",
            "litertlm-convert/minicpm5_work/ship_lfm25_gpu_0160/device_gate_20260825_maxtok/console_jp_maxtok1024.log"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "short-chat.decode_tokens_per_s": 70.02,
            "short-chat.generated_tokens": 128.0,
            "short-chat.peak_mem_mb": 556.0,
            "short-chat.ttft_ms": 91.0
          },
          "output_match": null,
          "peak_mem_mb": 556.0,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 91.0
        },
        "source": "data/device_runs/0.16.0/2026-08-25/lfm25-12b-jp-int4-gpu-093__iphone-17-pro.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-08-12",
          "decode_tokens_per_s": 20.48,
          "delegated_ops": 542,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-cl-pinned",
            "os_build": null,
            "runtime": "litert-lm",
            "runtime_version": "0.16.0",
            "soc": "Google Tensor G3",
            "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=68.05 decode=20.48 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": 86.0,
            "init_s": 9.83055,
            "prefill_tokens": 19.0
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 68.05,
          "provenance": "measured",
          "runs": true,
          "total_ops": 542,
          "ttft_ms": 330.0
        },
        "source": "data/device_runs/0.16.0/2026-08-12/lfm25-12b-jp-int4-gpu-093__pixel-8a.json"
      },
      {
        "device": "raspberry-pi-5",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-09-01",
          "decode_tokens_per_s": 9.24,
          "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 30.8 s, output head '53'",
            "invocation 0: exit=0 wall_s=102.0 temp 47.7->54.3C throttled=0x0 prefill_tps=54.73 decode_tps=9.33 ttft_s=4.7846 init_s=32.7389 peak_rss_mb=1442",
            "invocation 1: exit=0 wall_s=100.0 temp 51.0->52.7C throttled=0x0 prefill_tps=54.66 decode_tps=9.24 ttft_s=4.7916 init_s=33.571 peak_rss_mb=1454",
            "invocation 2: exit=0 wall_s=102.0 temp 50.5->58.2C throttled=0x0 prefill_tps=53.01 decode_tps=9.2 ttft_s=4.9376 init_s=36.4797 peak_rss_mb=1454"
          ],
          "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": 9.33,
            "decode_tps_min": 9.2,
            "init_s": 33.571,
            "init_s_max": 36.4797,
            "init_s_min": 32.7389,
            "invocations": 3.0,
            "prefill_tokens": 256.0,
            "prefill_tps_max": 54.73,
            "prefill_tps_min": 53.01,
            "runs_per_invocation": 1.0,
            "threads": 4.0,
            "ttft_s_max": 4.9376,
            "ttft_s_min": 4.7846
          },
          "output_match": null,
          "peak_mem_mb": 1454.0,
          "prefill_tokens_per_s": 54.66,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 4791.6
        },
        "source": "data/device_runs/0.16.1/2026-09-01/lfm25-12b-jp-int4-gpu-093__raspberry-pi-5.json"
      }
    ]
  },
  "model": {
    "family": "lfm2.5",
    "id": "lfm25-12b-jp-int4-gpu-093",
    "license": "lfm-open-license-v1.0",
    "source_url": "https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP",
    "task": "text-generation"
  },
  "pitfalls": [
    "GPU needs litert-lm >= 0.16.0 (Android OpenCL and macOS, verified by generation); iOS Metal fails at engine creation for this family (LiteRT-LM#3129) — use CPU on iOS (HF card).",
    "On phone-class hardware the GPU's win is prefill and TTFT (3-6x); decode is bandwidth-bound and roughly a wash (HF card, measured).",
    "This tune's convs must stay float in int8 recipes: quantizing the JP tune's convs costs ~9pt GSM8K, unlike the Instruct sibling where conv-int8 is free; the int4 recipe here keeps convs float anyway (HF card).",
    "English GSM8K undersells a Japanese-optimized tune (int4 55% vs bf16 63%) — reported for quantization-fidelity transparency; Japanese conversation quality was verified by inspection (HF card).",
    "0.9.3 does not write ExecutorMetadata itself — add_executor_metadata.py is a mandatory post-step or litert-lm >= 0.15 fails at inference (RESULTS.md)."
  ],
  "schema_version": "1.2"
}