{ "artifacts": [ { "file": "LFM2.5-VL-450M_int4.litertlm", "sha256": "45a0224180c89a25299078524e5a8dfcce131686fd0f941cdcf3507095db4797", "size_mb": 387.974 } ], "benchmarks": [], "conversion": { "command": "python convert_lfm25_vl3b.py ../src_models/lfm25-vl-450m out_vl450m_fp2 --fp && python quantize_vl.py out_vl450m_fp2/model.litertlm LFM2.5-VL-450M_int4.litertlm", "quantization": "text int4 blockwise-32 OCTAV linears + int8 embedding/lm_head; vision tower int8 dynamic (wi4b32_wi8 + embedder wi8 pass; 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": 99.25, "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=403.63 decode=99.25 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": 192.0, "init_s": 0.99613, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 403.63, "provenance": "measured", "runs": true, "total_ops": 801, "ttft_ms": 520.0 }, "source": "data/device_runs/0.16.0/2026-09-05/lfm25-vl-450m-int4__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 103.02, "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=2699.59 decode=103.02 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": 117.0, "init_s": 3.79212, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 2699.59, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 90.0 }, "source": "data/device_runs/0.16.0/2026-08-24/lfm25-vl-450m-int4__galaxy-s26.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 22.62, "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=92.12 decode=22.62 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": 5.1952, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 92.12, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 3260.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-450m-int4__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 33.05, "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=528.07 decode=33.05 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": 8.20607, "prefill_tokens": 296.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 528.07, "provenance": "measured", "runs": true, "total_ops": 543, "ttft_ms": 590.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-450m-int4__pixel-8a.json" } ] }, "model": { "family": "lfm2_vl", "id": "lfm25-vl-450m-int4", "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 — 8/8 text and 4/5 image vs the int8's 6/8 and 3/5; on Pixel 8a CPU int4 prefills slower than int8 (blockwise repacking) but decodes ~40% faster — pick by whether prompts or outputs dominate (HF card, measured).", "64k vocab (stop ids 7/2, same as the 1.2B text family) — the 3B's 128k pbtext does not transfer; verify generation_config before reusing metadata across sizes (REPRODUCE.md).", "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, geometry probes correct). 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).", "Zero-scale fix is a no-op on this checkpoint (0 all-zero blocks, like the 1.2B family); harmless to keep in the pipeline (RESULTS.md).", "iPhone GPU blocked (LiteRT-LM#3129) — iOS runs CPU; Android OpenCL delegates fully (543/543)." ], "schema_version": "1.2" }