{ "artifacts": [ { "file": "LFM2.5-VL-1.6B_int8.litertlm", "sha256": "d6a254aa3edefab51684918e7c717d096c60e495c085fcc35fd845fb8c80c8bd", "size_mb": 1725.193 } ], "benchmarks": [], "conversion": { "command": "python convert_lfm25_vl3b.py ../src_models/lfm25-vl-1.6b out_vl16b_int8 && python quantize_vl.py out_vl16b_int8/model.litertlm LFM2.5-VL-1.6B_int8.litertlm --recipe none # export-time dynamic int8 is the convert default (no --fp); the quantize_vl pass adds ExecutorMetadata + the externalized-embedder int8 recipe (RESULTS.md)", "quantization": "int8 dynamic (text linears + convs + embedding, vision tower) — export-time dynamic_wi8_afp32; externalized embedder int8 via quantize_vl", "tool": "litert-torch export_hf --task image_text_to_text (via litertlm-convert lfm25vl_work/convert_lfm25_vl3b.py + quantize_vl.py; public mirror hf-to-litertlm lfm_work/convert_lfm25_vl.py + quantize_lfm25_vl.py)", "tool_version": "0.9.3 (.venv-vl093: transformers 5.14.1, torch 2.12.1, torchvision 0.27.1)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 41.66, "delegated_ops": 760, "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 459 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 104 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 459 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 104 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 459 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 104 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 459 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 104 partitions for subgraph 3 (prefill_128).", "results block: prefill=525.76 decode=41.66 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": 122.0, "init_s": 1.94013, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 525.76, "provenance": "measured", "runs": true, "total_ops": 801, "ttft_ms": 410.0 }, "source": "data/device_runs/0.16.0/2026-09-05/lfm25-vl-1.6b-int8__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 42.27, "delegated_ops": 543, "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 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).", "results block: prefill=1518.86 decode=42.27 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": 113.0, "init_s": 2.15942, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 1518.86, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 160.0 }, "source": "data/device_runs/0.16.0/2026-08-24/lfm25-vl-1.6b-int8__galaxy-s26.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 8.38, "delegated_ops": null, "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": [ "results block: prefill=60.42 decode=8.38 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": 1515.0, "init_s": 16.7728, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 60.42, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 5020.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-1.6b-int8__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 15.91, "delegated_ops": 543, "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 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 543 out of 543 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).", "results block: prefill=403.31 decode=15.91 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": 3758.0, "init_s": 80.45954, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 403.31, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 800.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-1.6b-int8__pixel-8a.json" } ] }, "model": { "family": "lfm2_vl", "id": "lfm25-vl-1.6b-int8", "license": "lfm-open-license-v1.0", "source_url": "https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B", "task": "image-text-to-text" }, "pitfalls": [ "int4 vs int8 trade on this model: int8 prefills faster on CPU, int4 decodes markedly faster (blockwise-int4 repacking cost sits in prefill) — pick by whether prompts or outputs dominate; text graph delegates fully on Android OpenCL (543/543, zero rejections) (HF card, measured).", "Same VLM-bundle mechanics as the siblings: transformers==5.14.1 pin; llm_model_type { lfm2 {} } with bare template; externalized-embedder int8 pass; pack/unpack post-processing (single-tflite tools drop vision sections) (REPRODUCE.md).", "iPhone GPU blocked (LiteRT-LM#3129) — iOS runs CPU; use litert-lm >= 0.16.0 for Android OpenCL / macOS GPU (HF card).", "Android v0.16.0 litert_lm_main has no image flag (--max_num_images only) — image input is API-only on Android for now, so device rows are text-graph numbers (RESULTS.md)." ], "schema_version": "1.2" }