{ "artifacts": [ { "file": "InternVL3-2B.litertlm", "sha256": "cb7d63cbf2f5d9b3eb307012b54dc4fc4d68fef78dc3a5d4fbba41c59812f9b6", "size_mb": 1363.411 } ], "benchmarks": [], "conversion": { "command": "TODO (the HF card states the bundle layout and rewrites, not the invocation)", "quantization": "TODO — the HF card states no quantization recipe for this bundle; sections: VISION_ENCODER [1,448,448,3]->[1,256,4096], VISION_ADAPTER [1,256,4096]->[1,256,1536], externalized embedder, PREFILL_DECODE (HF card Conversion notes)", "tool": "litert-torch (LiteRT-LM fast_vlm bundle: VISION_ENCODER + VISION_ADAPTER + single-token EMBEDDER + PREFILL_DECODE with embeddings input) (HF card Conversion notes)", "tool_version": "TODO (the HF card names the converter but not its version; no export log in the sources)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 43.67, "delegated_ops": 1510, "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 1125 out of 1243 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 165 partitions for subgraph 0 (prefill_128).", "VERBOSE: Replacing 1125 out of 1243 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 165 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 1013 out of 1132 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 170 partitions for subgraph 2 (decode).", "VERBOSE: Replacing 7 out of 7 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 1 partitions for subgraph 3 (odml.rms_norm.impl).", "results block: prefill=207.94 decode=43.67 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": 182.0, "init_s": 1.72102, "prefill_tokens": 239.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 207.94, "provenance": "measured", "runs": true, "total_ops": 1573, "ttft_ms": 1170.0 }, "source": "data/device_runs/0.16.0/2026-09-05/internvl3-2b__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 34.4, "delegated_ops": 1243, "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 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_128).", "VERBOSE: Replacing 1243 out of 1243 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 1132 out of 1132 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (decode).", "results block: prefill=573.22 decode=34.4 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": 1845.0, "init_s": 6.47372, "prefill_tokens": 203.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 573.22, "provenance": "measured", "runs": true, "total_ops": 1243, "ttft_ms": 380.0 }, "source": "data/device_runs/0.16.0/2026-08-24/internvl3-2b__galaxy-s26.json" } ] }, "model": { "family": "internvl3", "id": "internvl3-2b", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/InternVL3-2B", "task": "image-text-to-text" }, "pitfalls": [ "Known limitation — one image per conversation on the GPU backend: on the GPU (Metal) backend a second image in the same conversation truncates the answer; on the CPU backend multi-image works (verified); the truncation reproduces with other fast_vlm models including litert-community/FastVLM-0.5B (ruled out as causes: max_num_images, and the vision encoder's 5D reshape — a 4D-clean rebuild still truncates on GPU); single-image VQA on the iPhone 17 Pro GPU runs ~45 tok/s, so it is the runtime's GPU fast_vlm path, not this bundle. For reliable multi-image, run on the CPU backend (HF card Known limitation).", "The vision encoder bakes ImageNet normalization + the NCHW transpose into the graph (the runtime feeds a [0,1] NHWC image), and the InternViT attention is rewritten 4D-clean (qkv split before the head reshape — no GPU-rejected 5D reshape), numerically identical (corr ≈ 1.0) (HF card Conversion notes).", "Decoder exported with an externalized embedder; InternVL's dynamic-NTK rope_scaling is stripped to base RoPE, valid since the export cache <= the base context window (HF card Conversion notes).", "License: MIT (the InternVL model) + Apache-2.0 (the Qwen2.5 language component); converted artifacts released under the same terms (HF card License).", "LiteRT-LM .litertlm bundle: LiteRT.js cannot run it, so delegation stays null and there is no browser block; device rows come from data/device_runs/ (Pi 5 / S26 / Pixel 8a / Mac / iPhone as measured)." ], "schema_version": "1.2" }