- common : read ngram cache parts by reference instead of copying9/10ggml-org/llama.cpp·performance·+7-9Referencesstd::unordered_mapPerformance optimizationAvoiding unnecessary copying
- today's pick · read the full storyIntegrate the UltraBERT architecture9/10huggingface/transformers·feature·+1666-0Class inheritance and method overridingCustom model architecture integrationAttention mechanisms with variable window sizesRotary positional embeddings
- [Feature][PCP] Add DeepSeek-V4.1-Flash support8/10vllm-project/vllm·feature·+873-37CUDA programmingSparse attention mechanismsPrefill Context Parallelism (PCP)Model caching and key-value compression
- FSDP x TP Saving / Loading8/10huggingface/transformers·feature·+611-85Distributed checkpointingTensor parallelismFully sharded data parallelism (FSDP)Data redistribution for tensor shards
- Mangle personality symbol8/10rust-lang/rust·feature·+80-51Symbol manglingStatic librariesCompiler version hashingInter-project coordination (Rust and LLVM)