v0.4.0

What's Changed

• feat: USE_FLASH_ATTENTION env var by @OlivierDehaene in #57
• docs: The initial version of the TEI docs for the hf.co/docs/ by @MKhalusova in #60
• feat: support roberta by @kozistr in #62
• fix: GH workflows update: added --not_python_module flag by @MKhalusova in #66
• docs: Images links updated by @MKhalusova in #72
• feat: add normalize option by @OlivierDehaene in #70
• ci: Migrate CI to new Runners by @glegendre01 in #74
• feat: add support for classification models by @OlivierDehaene in #76

New Contributors

• @MKhalusova made their first contribution in #60
• @kozistr made their first contribution in #62
• @glegendre01 made their first contribution in #74

Full Changelog: v0.3.0...v0.4.0

v0.3.0

What's Changed

• feat: faster CPU image on AMD in #35
• feat: support camembert in #42
• feat: support float32 on cuda in #41
• feat: support jinaAI variant in #48

Full Changelog: v0.2.2...v0.3.0

v0.2.2

What's Changed

fix: max_input_length should take into account position_offset (aec5efd)

Full Changelog: v0.2.1...v0.2.2

v0.2.1

What's Changed

• fix: only use position offset for xlm-roberta (8c507c3)

Full Changelog: v0.2.0...v0.2.1

v0.2.0

What's Changed

• add support for XLM-RoBERTa in #5
• get number of tokenization workers from the number of CPU cores in #8
• prefetch batch in #10
• support loading from .pth in #12
• add --pooling arg in #14
• fix compute cap matching in #21

Full Changelog: v0.1.0...v0.2.0

v0.1.0

• No compilation step
• Dynamic shapes
• Small docker images and fast boot times. Get ready for true serverless!
• Token based dynamic batching
• Optimized transformers code for inference using Flash Attention, Candle and cuBLASLt
• Safetensors weight loading
• Production ready (distributed tracing with Open Telemetry, Prometheus metrics)