README.md

---
license: mit
task_categories:
- feature-extraction
language:
- en
tags:
- embeddings
- sentence-transformers
- all-MiniLM-L6-v2
- feature-extraction
pretty_name: Distributed Embedding Generation Queue Sample
size_categories:
- n<1K
---

# Distributed Embedding Generation Queue - Sample Embeddings

Sample text embeddings produced by a durable producer/consumer GPU queue with
resume-on-crash support. Source code:
[github.com/narinzar/distributed-embedding-generation-queue](https://github.com/narinzar/distributed-embedding-generation-queue).

## Generation method

- **Model:** `sentence-transformers/all-MiniLM-L6-v2` (384-dimensional vectors).
- **Pipeline:** a durable SQLite task queue (WAL mode) feeds a GPU worker pool.
  Items are claimed atomically, embedded in batches, written as `.npy` files, and
  marked done. Orphaned `in_progress` items are re-queued on restart, so a run
  resumes without re-embedding finished items. Batch size is tuned to GPU headroom
  by an autobatcher (grows with free VRAM, shrinks on a caught OOM).
- **Run:** 500 text items embedded on an RTX 5090 at 33.9 items/s (wall 14.76s,
  single worker). This was a small-scale single-worker run; the architecture
  supports scaling the worker count, and multi-worker throughput scaling is
  reproducible on Linux.

## Contents

- `sample_embeddings.npy` - a 200 x 384 `float32` array, the first 200 vectors of
  the 500-item run.
- `sample_ids.txt` - the item ids for those 200 vectors, one per line, aligned by
  row order.
- `throughput.json` - the run summary: model, device, autobatch configuration,
  queue transition counts, and measured throughput.

## Usage

```python
import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="narinzar/distributed-embedding-generation-queue",
    filename="sample_embeddings.npy",
    repo_type="dataset",
)
vecs = np.load(path)  # (200, 384) float32
```

## License

MIT.