MakazhanAlpamys/Soup v0.16.0

v0.16.0 — Embedding Models + Export

on GitHub

What's New

Embedding Model Fine-Tuning

• task: embedding — fine-tune sentence embedding models (BGE, E5, GTE, INSTRUCTOR)
• Contrastive loss (InfoNCE with configurable temperature)
• Triplet loss with configurable margin
• Cosine loss for simple similarity training
• Configurable pooling: mean, cls, last token
• New template: soup init --template embedding

ONNX Export

• soup export --format onnx — export models via optimum
• Supports --onnx-task for text-generation or feature-extraction
• Install: pip install 'soup-cli[onnx]'

TensorRT-LLM Export

• soup export --format tensorrt — export for high-throughput GPU inference
• Two-step pipeline: HF checkpoint → TensorRT engine
• Install: pip install 'soup-cli[tensorrt]'

Speculative Decoding

• soup serve --speculative-decoding <draft-model> — 2-3x faster generation
• Transformers backend: HuggingFace assisted generation
• vLLM backend: native speculative decoding
• SSRF protection on draft model paths

Security

• ONNX export: no unconditional trust_remote_code
• Speculative decoding: URL validation, warning panel before draft model load
• TensorRT export: separate error handling per subprocess call
• Embedding config: Literal constraints, margin validation

Stats

• 1270 tests across 52 test files
• 58% code coverage
• All ruff lint rules pass

Install / Upgrade

pip install --upgrade soup-cli