{ "artifacts": [ { "file": "LFM2.5-VL-450M_int8.litertlm", "sha256": "f941a5f9482e31c1b7980456ac94269f2e45fffa28cd49f50c8435475c106dda", "size_mb": 537.443 } ], "benchmarks": [], "conversion": { "command": "python convert_lfm25_vl3b.py ../src_models/lfm25-vl-450m out_vl450m_int8_dir && python quantize_vl.py out_vl450m_int8_dir/model.litertlm LFM2.5-VL-450M_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 (embedder 268 -> 68 MB)", "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": 70.53, "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=766.74 decode=70.53 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": 149.0, "init_s": 1.12031, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 766.74, "provenance": "measured", "runs": true, "total_ops": 801, "ttft_ms": 280.0 }, "source": "data/device_runs/0.16.0/2026-09-05/lfm25-vl-450m-int8__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 97.57, "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=2345.63 decode=97.57 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": 139.0, "init_s": 1.01434, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 2345.63, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 100.0 }, "source": "data/device_runs/0.16.0/2026-08-24/lfm25-vl-450m-int8__galaxy-s26.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 15.75, "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=177.25 decode=15.75 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": 3800.0, "init_s": 2.7301, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 177.25, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 1730.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-450m-int8__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 31.08, "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=577.98 decode=31.08 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": 941.0, "init_s": 39.44901, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 577.98, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 540.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-450m-int8__pixel-8a.json" } ] }, "model": { "family": "lfm2_vl", "id": "lfm25-vl-450m-int8", "license": "lfm-open-license-v1.0", "source_url": "https://huggingface.co/LiquidAI/LFM2.5-VL-450M", "task": "image-text-to-text" }, "pitfalls": [ "int4 is the better variant of this model on our gates — this int8 scores 6/8 text and 3/5 image vs the int4's 8/8 and 4/5; on Pixel 8a CPU int8 prefills faster than int4 but decodes ~40% slower — pick by whether prompts or outputs dominate (HF card; lfm25vl_work/RESULTS.md gate table, measured).", "Fine-grained shape questions can miss on-device at this scale across all quant levels including an unquantized probe, while the torch reference answers them; conversion exonerated by experiment (vision tflite cosine 1.0000, token-identical streams). Treat as a fast small VLM for coarse visual tasks and large-text OCR; the 3B answers all fixtures (RESULTS.md; HF card wording).", "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; Android OpenCL delegates the text graph fully (543/543, zero rejections).", "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" }