Supported Features¶

This table summarize the status of features on Spyre. By default, those features are planned to be developed using vLLM engine V1.

Feature	Status
Chunked Prefill	✅
Automatic Prefix Caching	✅
LoRA	⛔
Speculative Decoding	⛔
Guided Decoding	⛔
Enc-dec	⛔
Multi Modality	⚠️
LogProbs	✅
Prompt logProbs	⛔
Beam search	✅
Tensor Parallel	✅
Pipeline Parallel	⛔
Expert Parallel	⛔
Data Parallel	⛔
Prefill Decode Disaggregation	⛔
Quantization	⚠️
Sleep Mode	⛔
Embedding models	✅