Your daily OSS contribution digest
Hi,
Here's today's OSS contribution digest — the highest-signal merged PRs from the last 24 hours, picked out of 20+ tracked repos.
Added support for NVIDIA's Dory1.0-4B BF16 large language model in vLLM.
This update enhances batch-sharded sampling by adding support for adaptive verification, ensuring that the sampling process dynamically adjusts to the actual admitted draft counts rather than relying on pre-scheduled estimates. It synchronizes draft counts between device and host to optimize resource usage and maintain accuracy during sampling.
vllm-project/vllm · Feature · impact 8/10 · +160 -25Asynchronous device-to-host memory copyConditional logic based on runtime stateBatch processing and shardingDynamic resource allocationThis update enhances the handling of special floating-point values like NaNs and their payloads for an emulated 128-bit floating-point type. It modifies relevant functions, adjusts build configurations, and adds tests to ensure correct behavior in this extended precision format.
llvm/llvm-project · Feature · impact 8/10 · +329 -154Floating-point representationFunction overloadingBuild system integration (CMake and Bazel)Unit testingThis update improves query caching for table functions that read files or object storage by enabling cache usage even when tables lack unique identifiers. It ensures that repeated queries over file-based data sources like Parquet files benefit from caching by deriving cache keys from file metadata and reader definitions rather than relying solely on table UUIDs.
ClickHouse/ClickHouse · Feature · impact 8/10 · +920 -32Cache key derivationFile metadata handlingTable function abstractionConditional logic for cache entriesThis update adds support for loading workflow definitions directly from Azure Blob Storage in Apache Airflow, enabling users to manage their workflows stored in Azure containers similarly to existing support for other cloud storage services. It introduces a new component to handle Azure blobs, integrates it with existing hooks, and includes tests and documentation to ensure smooth operation.
apache/airflow · Feature · impact 8/10 · +828 -0Class inheritance and extensionCloud storage SDK integrationUnit testing in PythonConfiguration and registration of plugins
Today's highest-signal arXiv papers, ranked within each topic.
This paper introduces a method that enables web agents to self-verify their actions during training and testing, improving performance without relying on expensive external judges at runtime.
This paper introduces a new offline metric that predicts how well a robot model will perform by measuring how closely its sensory representations align with real physical motion, enabling cheaper and earlier evaluation of embodied AI models.
Ai Ml · relevance 8/10This paper develops algorithms that decide when to trust AI decisions automatically and when to involve human experts, minimizing human effort while ensuring AI errors are caught below a set threshold.
Ai Ml · relevance 8/10
This paper presents a new algorithm that significantly speeds up solving the ARRIVAL problem, which involves determining the path of a token through a special graph, improving the runtime from roughly exponential in the square root of the graph size to a subexponential cube-root time.
Algorithms Theory · relevance 7/10The paper provides a new explicit and constructive algorithm that multiplies matrices faster than classical methods, proving an improved upper bound on the matrix multiplication exponent.
Algorithms Theory · relevance 7/10This paper develops algorithms to produce multiple high-quality and diverse job schedules on unrelated machines, providing alternatives that balance makespan efficiency with varied job assignments.
Algorithms Theory · relevance 7/10
This paper analyzes how to efficiently combine large data values across network nodes when messages have limited size, providing optimal algorithms for common network topologies and approximation methods for general graphs.
Distributed Systems · relevance 6/10
This paper fixes a key flaw in a recent quaternion tensor decomposition method and introduces a faster, accurate algorithm that improves processing of multidimensional quaternion data, with practical benefits demonstrated in color video encryption.
Math · relevance 7/10This paper offers a new way to understand and construct faster matrix multiplication algorithms by combining tensor contractions, achieving better performance than Strassen's method and potentially useful for practical implementations.
Math · relevance 7/10This paper presents a new method to compress complex spatial data by adaptively breaking it into regions and applying tailored low-rank approximations, ensuring controlled error while reducing storage needs.
Math · relevance 7/10
This paper analyzes how different tools used to group MQTT network packets into flows impact the structure and clustering of traffic data, showing that tool choice significantly affects analysis results.
Networking · relevance 7/10
Tulva introduces a practical and modular way to add asynchronous threshold decryption to fully homomorphic encryption schemes, enabling efficient and scalable privacy-preserving computations compatible with real-world platforms like Ethereum.
Security · relevance 8/10This paper introduces a new stealthy backdoor attack on spiking neural networks that manipulates only the timing of spikes, making it hard to detect while maintaining high attack success and minimal impact on normal performance.
Security · relevance 8/10This paper introduces Phantomware, a method to generate diverse synthetic malware examples in a learned feature space to improve detection of new malware variants when only few samples are available, avoiding costly binary generation and validation.
Security · relevance 8/10