{ "artifacts": [ { "file": "LFM2.5-VL-3B_int4.litertlm", "sha256": "ebb563f3587feb0cfa6867ff53fba556043ee459410a6a0276f3245e20ba0a59", "size_mb": 2243.068 } ], "benchmarks": [], "conversion": { "command": "python convert_lfm25_vl3b.py ../src_models/lfm25-vl-3b out_vl3b_fp --fp && python quantize_vl.py out_vl3b_fp/model.litertlm LFM2.5-VL-3B_int4.litertlm", "quantization": "text int4 blockwise-32 OCTAV linears + int8 embedding/lm_head; vision tower int8 dynamic (wi4b32_wi8 + embedder wi8 pass)", "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": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 27.04, "delegated_ops": 937, "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 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).", "results block: prefill=471.56 decode=27.04 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": 161.0, "init_s": 10.11537, "prefill_tokens": 205.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 471.56, "provenance": "measured", "runs": true, "total_ops": 937, "ttft_ms": 470.0 }, "source": "data/device_runs/0.16.0/2026-08-24/lfm25-vl-3b-int4__galaxy-s26.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 5.77, "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=14.12 decode=5.77 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": 1109.0, "init_s": 44.67743, "prefill_tokens": 292.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 14.12, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 20850.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-3b-int4__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-13", "decode_tokens_per_s": 10.79, "delegated_ops": 937, "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 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 937 out of 937 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).", "results block: prefill=91.63 decode=10.79 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": 1168.0, "init_s": 39.1153, "prefill_tokens": 292.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 91.63, "provenance": "measured", "runs": true, "total_ops": 937, "ttft_ms": 3280.0 }, "source": "data/device_runs/0.16.0/2026-08-13/lfm25-vl-3b-int4__pixel-8a.json" } ] }, "model": { "family": "lfm2_vl", "id": "lfm25-vl-3b-int4", "license": "lfm-open-license-v1.0", "source_url": "https://huggingface.co/LiquidAI/LFM2.5-VL-3B", "task": "image-text-to-text" }, "pitfalls": [ "Pin transformers==5.14.1 for the export: 5.15.0 renamed Lfm2ShortConv.L_cache -> conv_kernel_size and litert-torch 0.9.3's subclass still reads the old attribute (AttributeError at model load). torchvision must stay 0.27.x to keep torch under litert-torch's <2.13 pin (REPRODUCE.md).", "The metadata pbtext needs llm_model_type { lfm2 {} } — the empty message is correct (runtime Lfm2DataProcessor proto defaults equal this family's processor config). The chat template must render a bare per image part and never boi/eoi: the runtime splits on the marker and inserts <|image_start|> + pixels + <|image_end|> itself (REPRODUCE.md).", "The externalized VLM embedder dodges every text quantization recipe: an --fp text export leaves it float32 (1 GB at the 128k vocab) and a post-hoc recipe applied to prefill_decode never reaches it — quantize it int8 as a separate pass (1049 -> 264 MB; REPRODUCE.md / quantize_vl.py).", "VLM bundles have >1 tflite section: single-tflite post-processing tools (quantize_litertlm.py, fix_zero_block_scales.py main()) rebuild via litert-lm-builder with one tflite and silently DROP the vision sections — post-process through litert-lm pack/unpack instead (RESULTS.md trap).", "The zero-scale int4 wall is inherited from the dense LFM2.5 lineage: this 3B checkpoint carries 746,432 all-zero 32-blocks across 26 tensors, so the zero-scale fix is mandatory or XNNPACK refuses the file (REPRODUCE.md; same signature as LFM2.5-2.6B).", "iPhone GPU is blocked for the ShortConv family at engine creation (upstream LiteRT-LM#3129, re-verified on 0.16.0) — iOS runs CPU; Android OpenCL and macOS GPU delegate fully (937/937 nodes, HF card).", "On a nearly-full Android device the GPU path's multi-GB compile-cache write can fail mid-initialization and masquerade as a kernel failure — run with --disable_cache=true (bare --disable_cache does not take) or free storage (RESULTS.md trap, hit on Pixel 8a)." ], "schema_version": "1.2" }