{ "artifacts": [ { "file": "LFM2.5-VL-1.6B_int4.litertlm", "sha256": "44816005b42bef8887bc3dc40b7d6483ea4dd8569fb1591f00faa33ef3043e8d", "size_mb": 1238.005 } ], "benchmarks": [], "conversion": { "command": "python convert_lfm25_vl3b.py ../src_models/lfm25-vl-16b out_vl16b_fp --fp && python quantize_vl.py out_vl16b_fp/model.litertlm LFM2.5-VL-1.6B_int4.litertlm", "quantization": "text int4 blockwise-32 OCTAV linears + int8 embedding/lm_head; vision tower int8 dynamic (wi4b32_wi8 + embedder wi8 pass; embedder 537 -> 135 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": 50.85, "delegated_ops": 769, "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 468 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 86 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 468 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 86 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 468 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 86 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 468 out of 543 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 86 partitions for subgraph 3 (prefill_128).", "results block: prefill=127.27 decode=50.85 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": 110.0, "init_s": 2.9921, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 127.27, "provenance": "measured", "runs": true, "total_ops": 801, "ttft_ms": 1630.0 }, "source": "data/device_runs/0.16.0/2026-09-05/lfm25-vl-1.6b-int4__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 54.1, "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=1020.13 decode=54.1 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": 126.0, "init_s": 6.4125, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 1020.13, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 220.0 }, "source": "data/device_runs/0.16.0/2026-08-24/lfm25-vl-1.6b-int4__galaxy-s26.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 13.64, "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=36.92 decode=13.64 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": 13.85961, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 36.92, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 8090.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-1.6b-int4__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 22.08, "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=201.04 decode=22.08 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": 3800.0, "init_s": 17.30064, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 201.04, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 1520.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-1.6b-int4__pixel-8a.json" } ] }, "model": { "family": "lfm2_vl", "id": "lfm25-vl-1.6b-int4", "license": "lfm-open-license-v1.0", "source_url": "https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B", "task": "image-text-to-text" }, "pitfalls": [ "Vocab split inside one family: the 1.6B and 450M use the old 64k vocab (stop ids 7/2, same as the 1.2B text family), the 3B the 128k one (124900/124895) — verify generation_config before reusing a metadata pbtext across sizes (REPRODUCE.md).", "Text is the family's best on our gates (8/8 on all four int4/int8 x cpu/gpu configs), but fine-grained shape/counting image fixtures deterministically miss on-device across ALL quant levels including an unquantized probe, while the torch reference answers them — with conversion exonerated by experiment (vision+projector tflite vs torch cosine 1.0000; token-identical prompt streams; on-device geometry probes correct; chunk-boundary alignment probe negative). Engine-internal residue; the 3B is unaffected. Card as coarse-understanding/OCR/localization-strong, defer fine shape/counting to the 3B (RESULTS.md elimination log; REPRODUCE.md warning block).", "Same VLM-bundle mechanics as the 3B: 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).", "int4 zero-scale fix found 0 all-zero blocks on this checkpoint (like the 450M and the 1.2B family) — only the 3B/2.6B dense lineage carries the 746k-block signature; run the fix regardless, it is a no-op when clean (RESULTS.md).", "iPhone GPU blocked (LiteRT-LM#3129) — iOS runs CPU; Android OpenCL delegates fully (543/543) and macOS GPU is measured on the HF card." ], "schema_version": "1.2" }