Floret Torch

Floret Torch is a Python library that trains fastText-style word vectors on a GPU using PyTorch.

Floret-torch trains fastText-style subword vectors on a graphics processing unit using PyTorch. It reimplements the training loop of Facebook AI Research's fastText library and the floret variant, achieving high performance because memory bandwidth limits this workload rather than arithmetic. The library writes standard .floret and .vec files that load straight into spaCy. It uses per-batch word deduplication and hand-written gradients to accelerate training without custom CUDA kernels. It requires Python 3.9 or newer and PyTorch, and it runs on systems equipped with compatible GPUs.

Floret Torch at a glance

Best for: Software engineer, Researcher
Runs on: Python, PyTorch
Pricing: Free
Released: October 2026
Category: AI coding tool
Website: github.com/fazel94/floret-torch

Key features

Trains fastText-style word vectors on a GPU
Writes spaCy-compatible .floret and .vec files
Implements per-batch word deduplication
Uses hand-written gradients for optimization
Supports FP16 table storage options
Includes byte-exact hashing against CPU floret

Use it for

Training word embeddings for large text corpora
Generating vectors for rare or unseen words
Building custom spaCy NLP pipelines

Alternatives to Floret Torch

torchao — It speeds up training and inference while reducing memory usage with native PyTorch.
AirLLM — Runs massive 70B+ parameter models on small 4GB consumer GPUs.
FastEmbed — It uses ONNX Runtime for high speed and does not require a GPU or PyTorch.
Finetuning With Sampling — Official implementation of the Finetuning with Sampling research paper from Harvard.
Soup — Fine-tune and align models like Llama-3.1-8B on a 4 GB laptop GPU.
GLiNER — Enables zero-shot entity extraction without labeled training data on consumer CPUs.
fal — Access 1,000+ generative models and high-performance serverless GPUs.
SwanLab — Tracks training metrics and system performance with minimal code integration.
TurboVec — Fits a 10 million document corpus in 4 GB of RAM and searches faster than FAISS.
Galahad — Eliminates recomputing tokens and keeps GPU memory flat up to 50M tokens.
LangExtract — It maps every extraction to its exact location in the source text.
Unsloth — Runs and trains models locally with lower memory usage.

All Floret Torch alternatives