morphoformer 3.2.0

MorphFormer: multilingual morphological reinflection with character-level Transformer (GQA, RoPE, SwiGLU, language adapters)

MorphFormer

Character-level Transformer for multilingual morphological reinflection.

Version 3.1 — modular multi-package architecture by Voluntas Progressus.

Installation

pip install morphoformer

Requires Python >= 3.14 and PyTorch >= 2.0.

Dependencies are installed automatically:

• chartoken-vp — character-level tokenizer
• torchblocks-vp — pluggable Transformer blocks
• sigmorphon-vp — SigMorphon dataset tools
• trainkit-vp — training utilities

Quick Start

# Download SigMorphon data
morphoformer download --lang rus,deu,fra --merge

# Train with preset
morphoformer train --preset medium --data "data/collections/*_train.tsv" --device cuda

# Single-word inference
morphoformer infer --checkpoint checkpoints/morph_v3.pt --word "laufen" --morph "V;IND;PST;3;SG" --lang deu

# Interactive REPL
morphoformer serve --checkpoint checkpoints/morph_v3.pt

Architecture

Component	Implementation
Attention	Grouped Query Attention (GQA) with KV cache
Positions	Rotary Position Embeddings (RoPE)
Feed-forward	SwiGLU
Normalization	RMSNorm (pre-norm)
Encoder conv	Conformer-style depthwise separable conv1d
Adapters	Language-conditioned bottleneck adapters
Morph features	Pooled or attention-based structured encoder

Presets

Preset	d_model	Encoder	Decoder	~Params	VRAM
small	384	4 layers	3 layers	~7M	< 4 GB
medium	512	8 layers	6 layers	~45M	4-8 GB
large	768	10 layers	8 layers	~120M	>= 8 GB

CLI Commands

Command	Description
train	Train model from TSV data
infer	Single-word inference
serve	Interactive REPL
download	Download SigMorphon 2021 datasets
modules	List registered NN building blocks
init-config	Generate TOML config template

Data Format

TSV with columns: lemma, features, surface_form, language

laufen	V;IND;PST;3;SG	lief	deu

Python API

import torch
from chartoken import CharVocab, FeatureVocab
from morphoformer.model import MorphFormer
from morphoformer.inference import greedy_decode

checkpoint = torch.load("checkpoints/morph_v3.pt", map_location="cpu", weights_only=False)
char_vocab = CharVocab.from_dict(checkpoint["char_vocab"])
feature_vocab = FeatureVocab.from_dict(checkpoint["feature_vocab"])
lang_to_id = checkpoint["lang_to_id"]

# Build model, load state_dict, call greedy_decode()

Supported Devices

Device	Flag
Auto-detect	--device auto
NVIDIA GPU	--device cuda
AMD GPU	--device rocm
Intel Arc	--device xpu
Apple Silicon	--device mps
CPU	--device cpu

License

MIT