github-featured / ludwig-ai--ludwig

发布

0 发布 85 Git 标签

RSS 订阅源
  • v0.14.0 8ff5163a09
    比较

    Ludwig 0.14.0

    frostbyte_neo 发布于 2026-04-15 05:04:45 +00:00

    Major feature release. Highlights:

    Encoders

    • Text: ModernBERT (Flash Attention 2, RoPE, 8192 tokens); remove deprecated
      TransformerXL / CTRL / FlauBERT
    • Image: CLIP, DINOv2, SigLIP, ConvNeXt V2
    • Audio: Wav2Vec2, Whisper, HuBERT
    • Sequence: Mamba (linear-time state space model); RoPE for StackedTransformer
    • Category: target (mean-target encoding), hash (feature hashing)
    • Number: bins (discretization)

    Decoders

    • Sequence: TransformerDecoder, scheduled sampling, beam search
    • Category: MLPClassifier, temperature scaling calibration, MC dropout
    • Image segmentation: configurable U-Net depth, SegFormer, FPN
    • LLM: real logits extraction, structured output parsing, constrained decoding
      via category_extractor (regex / json_schema / constrain_to_vocabulary)

    Losses

    • Focal, Dice, Lovasz-Softmax, NT-Xent, PolyLoss, Entmax-1.5 registration
    • Open-set recognition: Entropic Open-Set, Objectosphere
    • Anomaly detection: Deep SVDD, Deep SAD, DROCC

    Trainer

    • New optimizers: RAdam, Adafactor, Schedule-Free AdamW, Muon, SOAP
    • New LR schedulers: OneCycleLR, inverse_sqrt, polynomial, WSD
    • Preference-based LLM training: DPO, KTO, ORPO, GRPO
    • Loss balancing strategies: uncertainty, FAMO, GradNorm, log_transform
    • Quality presets (medium / high / best), model soup, modality dropout
    • Ftrl optimizer removed

    Serving

    • Auto-generated Pydantic request/response schemas
    • Prometheus /metrics endpoint, structured logging, request timeouts
    • OpenAI-compatible vLLM server (PagedAttention, continuous batching)
    • Model export: SafeTensors (default), torch.export (.pt2), dynamo-based ONNX

    Infrastructure

    • Python 3.12, PyTorch 2.7+, Pydantic 2, Transformers 5, NumPy 2
    • Ray 2.54 with modern ray.data.Dataset pipeline, Dask 2026.1.2, MLflow 3.10
    • uv-based test runner, GitHub Actions Docker workflow modernization
    • Distributed metric aggregation fix for Ray eval workers
    • Horovod, Neuropod, and GBM/LightGBM backends removed
    下载附件