- [Core] shm tensor arena for efficient cpu->gpu worker broadcast9/10vllm-project/vllm·performance·+1290-8Shared memory managementZero-copy data accessPinned (page-locked) memoryMemory copy optimization
- [ROCm][Perf][DSv4.1] Mono decode layer for DeepSeek-V4.1-Flash on MI355X9/10vllm-project/vllm·performance·+9119-2Environment variables for feature togglingGPU kernel launches and optimizationPersistent kernel executionIn-kernel collective operations (all-reduce)
- today's pick · read the full storyConditional Authz [3/n]: Make the webhook authorizer support conditions9/10kubernetes/kubernetes·feature·+16260-287InterfacesConditional logicAuthorization patternsModular design
- Add Molmo28/10huggingface/transformers·feature·+5917-11Python classes and inheritanceModel configuration and initializationIntegration of new model architecturesLarge-scale codebase extension
- KEP-3104: promote kuberc to stable8/10kubernetes/kubernetes·feature·+1437-254API versioningDeprecation strategyConfiguration managementKubernetes Custom Resource Definitions (CRDs)