Sandhi-1.0-8M-int8-executorch

The int8 ExecuTorch build of Sandhi-1.0-8M, a small character-level model that analyses a Malayalam word offline on a phone: sandhi split, morphemes, root lemma, grammatical features, IPA and syllables, and a confidence level.

Files

File	What
sandhi_int8.pte	int8 ExecuTorch program (XNNPACK), 9,594,204 bytes. Three graphs: encoder, cross, step
sandhi_meta.json	The runtime contract: vocabulary, label sets, tensor shapes and dtypes of each graph, thresholds, limits, checksums
code/normalize.py	The one normalize() function. Input must be normalized with this before the model sees it.
golden/	Test vectors: 1,017 full analyses (golden_analyse.jsonl) and normalization cases (golden_normalize.jsonl), for checking a port against the reference
reports/	The quantization and export reports, as measured
LICENSE, NOTICE	Apache-2.0, and attribution for the training sources

How to run it

This is a .pte program, not a transformers model. Load it with the ExecuTorch runtime (built with the XNNPACK backend) and follow sandhi_meta.json:

• encoder: takes [<FULL>] + one id per character (int64) and returns boundary logits, syllable logits, the multi-label other head, the classifier-head logits, and a memory tensor.
• cross: takes memory and returns the cross-attention keys and values.
• step: one decoder step with a fixed-size KV cache. The caller drives greedy decoding for two rows, root.lemma and ipa, up to 40 steps.

Words are limited to 39 characters. Thresholds and the confidence wording are in sandhi_meta.json. The first thing to do with any port is to run golden/: a correct port reproduces golden_analyse.jsonl. A reference Kotlin implementation lives in the app source at https://github.com/nithinmanoj10/sandhi (android/).

Example

The code that drives this .pte through the ExecuTorch runtime is the package ml-Sandhi. It takes its vocabulary and shapes from sandhi_meta.json, so it needs only the files in this repository.

pip install "ml-sandhi[int8] @ git+https://github.com/nithinmanoj10/ml-Sandhi"

import json
from ml_sandhi import Sandhi

sandhi = Sandhi.from_pretrained_int8()     # downloads sandhi_int8.pte and sandhi_meta.json on first use
result = sandhi.analyse("കമ്പ്യൂട്ടറുകളിലൂടെ")
print(json.dumps(result, ensure_ascii=False, indent=2))

Output (fields shown; the rest is omitted). On this word it is identical to the fp32 model's:

{
  "input": "കമ്പ്യൂട്ടറുകളിലൂടെ",
  "source": "model",
  "split": "കമ്പ്യൂട്ടറ + ുകള + ിലൂടെ",
  "root": {"lemma": "കമ്പ്യൂട്ടർ", "pos": "n", "type": "known"},
  "features": {"case": "perlative"},
  "ipa": "kampjuːʈʈarukaɭiluːʈe",
  "syllables": ["ക", "മ്പ്യൂ", "ട്ട", "റു", "ക", "ളി", "ലൂ", "ടെ"],
  "confidence_text": "Fairly sure"
}

കമ്പ്യൂട്ടറുകളിലൂടെ ("through the computers") is a loanword plus two endings, and is not in the dictionary, so source is model. The lemma കമ്പ്യൂട്ടർ is not a prefix of the surface: sandhi rewrote the stem. The full result also carries morphemes, display and alternatives.

What quantization cost

Measured against the fp32 model (reports/phase3_quantization_v3.md, reports/phase3_export.md):

• On 3,000 dev words the int8 program agrees with fp32 on 99.93% of splits, 99.87% of lemmas, 99.90% of IPA, 99.93% of feature sets and 99.87% of confidence labels.
• The torch.ao estimate is more pessimistic: split −0.24 points overall and −1.97 points on words of four or more morphemes. Treat that as the conservative bound.
• Size: 33.1 MiB fp32 → 9.1 MiB int8.

License

Apache-2.0. Training sources are attributed in NOTICE. mlmorph and Mlphon were dev-time label generators and are not distributed here.