silvic
Train and serve gradient-boosted trees in Rust with LightGBM-compatible model files.
Training runs on CPU, Apple Metal, NVIDIA CUDA and AWS Neuron. Serving runs on CPU, CUDA and Neuron (float32).
Training
End-to-end seconds
Serving
Host input to host output · milliseconds
Benchmark details
Training. 1M synthetic learning-to-rank rows and 100k validation rows, 100 numerical float32 features, queries of 50; 1,000 lambdarank trees, 53 leaves, 63 bins. All engines ran sequentially on the same two Xeon Platinum 8259CL / Tesla T4 nodes. Native rows time direct single-process dataset construction, binning, fit, validation and serialization; Ray rows also include partitioning, transfer, actor or Trainer setup, TCP and teardown. Two CPU threads per process and one GPU per CUDA process. Median of three fits after a one-tree warmup; data generation and cluster provisioning excluded.
Serving. One fixed 200-tree, 53-leaf model with 100 features. Median of 100 calls cycling through eight held-out 3,000-row batches after 10 warmups. Engines ran sequentially on one Xeon Platinum 8259CL / Tesla T4 host; CPU engines used two threads. Host input to host output includes CUDA transfers, allocation and synchronization; model loading and compilation excluded.
Python API
pip install silvic
pip install "silvic[ray]"import silvic
train_set = silvic.Dataset(X, label=y, group=query_sizes)
booster = silvic.train({"objective": "lambdarank"}, train_set, num_boost_round=1500)
booster.save_model("model.txt")import silvic.ray
booster = silvic.ray.train(
params,
ray.data.read_parquet("s3://bucket/train/"),
label_column="label",
group_column="query_id",
num_workers=5,
use_gpu=True,
)import silvic.serve
predictor = silvic.serve.Predictor("model.txt") # LightGBM 3.x or 4.x model
scores = predictor.predict(candidates, num_threads=1)