litert-community/LFM2.5-Encoder-350M-Prompt-Router

LFM2.5-Encoder-350M-Prompt-Router — LiteRT

LiquidAI/LFM2.5-Encoder-350M-Prompt-Router converted to LiteRT (.tflite) for on-device inference. Zero-shot prompt routing: define your routing lanes as free text and the model scores the whole prompt against every lane in one CPU pass (demo Space).

Model description

Two signatures, route_128 and route_512 (S = 128 / 512, batch 1, right-padded, up to 8 lane slots):

Softmax over the first N (real) lanes only — an all-zero pool row produces a constant bias logit that must be ignored.

How to use

1. Install dependencies

pip install ai-edge-litert numpy tokenizers huggingface_hub

2. Save the script below as route_prompt.py:

#!/usr/bin/env python3
"""Route a prompt to one of your lanes with litert-community/LFM2.5-Encoder-350M-Prompt-Router."""
import argparse

import numpy as np
from ai_edge_litert.interpreter import Interpreter
from huggingface_hub import hf_hub_download
from tokenizers import Tokenizer

REPO = "litert-community/LFM2.5-Encoder-350M-Prompt-Router"
MAX_LANES = 8


def build_inputs(text, lanes, tokenizer, seq_len):
    """Builds input_ids/attention_mask plus the two mean-pool matrices."""
    body = "\n".join(f"- {lane}" for lane in lanes)
    prefix = f"Categories:\n{body}\n\nText:\n"
    encoding = tokenizer.encode(prefix + text)
    ids, offsets = encoding.ids, encoding.offsets
    if len(ids) > seq_len:
        raise SystemExit(f"{len(ids)} tokens exceed --seq-len {seq_len}")

    input_ids = np.zeros((1, seq_len), np.int32)
    attention_mask = np.zeros((1, seq_len), np.int32)
    input_ids[0, : len(ids)] = ids
    attention_mask[0, : len(ids)] = 1

    # Mean-pool over the document's own tokens.
    text_pool = np.zeros((1, 1, seq_len), np.float32)
    text_idx = [i for i, (a, b) in enumerate(offsets) if b > len(prefix) and a != b]
    text_pool[0, 0, text_idx] = 1 / len(text_idx)

    # Mean-pool over each lane's tokens; unused lane rows stay all-zero.
    category_pool = np.zeros((1, MAX_LANES, seq_len), np.float32)
    pos = len("Categories:\n")
    for r, lane in enumerate(lanes):
        start, end = pos + 2, pos + 2 + len(lane)
        pos = end + 1
        idx = [i for i, (a, b) in enumerate(offsets) if a < end and b > start and a != b]
        category_pool[0, r, idx] = 1 / len(idx)

    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "text_pool": text_pool,
        "category_pool": category_pool,
    }


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--text", required=True, help="The prompt to route.")
    parser.add_argument("--lane", action="append", required=True,
                        help="A routing lane, repeatable (up to 8).")
    parser.add_argument("--seq-len", type=int, default=512, choices=[128, 512])
    args = parser.parse_args()
    if len(args.lane) > MAX_LANES:
        raise SystemExit(f"at most {MAX_LANES} lanes")

    model_path = hf_hub_download(REPO, "LFM2.5-Encoder-350M-Prompt-Router_wi8fc.tflite")
    tokenizer = Tokenizer.from_file(hf_hub_download(REPO, "tokenizer.json"))

    feed = build_inputs(args.text, args.lane, tokenizer, args.seq_len)
    interpreter = Interpreter(model_path=model_path)
    runner = interpreter.get_signature_runner(f"route_{args.seq_len}")
    logits = list(runner(**feed).values())[0][0]

    # Softmax over the real lanes only — unused rows carry a constant bias logit.
    real = logits[: len(args.lane)]
    probs = np.exp(real - real.max())
    probs /= probs.sum()
    for lane, p in sorted(zip(args.lane, probs), key=lambda x: -x[1]):
        print(f"{p:6.3f}  {lane}")


if __name__ == "__main__":
    main()

3. Run it

python route_prompt.py \
  --text "My Python script throws a KeyError on a dict lookup, how do I fix it?" \
  --lane "coding question" --lane "travel planning" \
  --lane "medical advice" --lane "small talk"

 0.838  coding question
 0.054  small talk
 0.054  travel planning
 0.054  medical advice

On Android/iOS use the LiteRT runtime's SignatureRunner APIs with the same signature names; the tokenizer is the standard Hugging Face tokenizer.json.

Performance

One pass over a padded sequence with the int8 (wi8fc) file, CPU only.

Mac figures are the median of 20 warm runs (ai-edge-litert 2.1.6, XNNPACK, otherwise idle machine). The iPhone figure comes from the on-device gate (TFLite C API + SignatureRunner + XNNPACK) and is a single run, not a median.

Budget for one slow first call. The first inference after loading pays a one-time graph preparation: on the Mac it took 372 ms against a 34.5 ms steady state. Later signatures on the same loaded model do not pay it again — route_512 measured 110 ms cold against 112 ms warm. Model load itself was 0.38 s on the iPhone, with a peak footprint of 649 MiB.

One pass scores the prompt against all eight lane slots at once, so the cost does not grow with the number of lanes. The signatures are fixed-shape, so input language or content does not change the time.

Accuracy note

Task-level parity against the PyTorch reference on the demo prompt with four lanes: fp32, fp16 and int8 all reproduce the reference lane probabilities to four decimal places — 0.838 for "coding question". That is a single-prompt spot check, not a benchmark over a labelled corpus.

On the iPhone 17 Pro the int8 file reproduces the desktop outputs bit-exactly — cosine 1.000000, max absolute difference 0.0.

Android (Pixel 8a)

Android figures use the standard TFLite `benchmark_model` on a Pixel 8a (Tensor G3, Android 16) — 5 warm-up runs then 20 timed runs, the signature selected explicitly with --signature_to_run_for, CPU at 4 threads.

GPU works as of the 2026-08-13 re-export. The re-export respells the one idiom mobile GPU delegates refuse — transformers' rank-5 repeat_kv expand — into an equivalent rank-4 matmul (outputs bitwise-identical on CPU); the OpenCL delegate now takes the whole graph. Measured with the LiteRT CompiledModel API (fp32 GPU precision, real inputs incl. pooling matrices, best of 3 warm runs): route_512 18.7 ms on the Pixel 8a, cosine 0.9949 vs the fp32 desktop reference; iPhone 17 Pro Metal route_512 172 ms, cosine 1.000000. Set the GPU precision to fp32 — at fp16 GPU precision this family's norm reductions overflow and every output is NaN. These CompiledModel timings are not comparable to the classic-delegate benchmark_model timings above (different GPU runtime).

Snapdragon NPU (Hexagon)

• —LFM2.5-Encoder-350M-Prompt-Router_fp16.tflite — the NPU runs it at 176.1 ms. The GPU does not — LiteRtException: Failed to compile model.
• —LFM2.5-Encoder-350M-Prompt-Router_wi8fc.tflite — the GPU runs it at 80.59 ms. The NPU does not — LiteRtException: Failed to compile model.

Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16) with LiteRT CompiledModel 2.2.0, one accelerator per process, 5 warm-up runs then N=50 timed runs, median reported. Every run held thermal status NONE throughout. Headroom 0.75–0.78, where 1.0 is the throttling threshold.

The NPU row marked AOT ran an artifact compiled ahead of time for SM8850 (ai-edge-litert 2.2.0 + QAIRT 2.47.0), not the published file. That artifact is not distributed here; the compile is one command in the NPU guide.

GPU wiring: GPU guide.

License

LFM Open License v1.0 (see LICENSE, unchanged from the base model). Note the license's commercial-use threshold (Section 5). This repository redistributes converted Derivative Works of LiquidAI/LFM2.5-Encoder-350M-Prompt-Router with modification notices per Section 4; all credit for the model to Liquid AI.