Loading / 加载中
Thinkgap feed
Curated items are read from the processed items table and served as a bilingual feed.
8 items
A commit in the PyTorch repository under the 'ciflow/torchtitan/187383' release addresses a fix in the DTensor math_ops module. The update targets the handling of single-dimension strategies. The author is anshul-si. No further implementation details are provided in the release note.
PyTorch trunk now enables symmetric communication operations for Intel's XPU backend, allowing computation and communication to overlap and reduce overhead on Intel client GPUs. The symmetric ops are designed for asynchronous tensor parallelism (async TP). The implementation involved backend changes in intel/torch-xpu-ops#2041 and Python op enabling in this pull request (#185102). Operation correctness was verified through tests in intel/torch-xpu-ops#3747, and the PR was approved by multiple reviewers.
The hexo-ai/sia repository releases SIA, a self-improving AI framework. SIA is designed to autonomously enhance the performance of any AI model or agent on a given benchmark task. It targets automatic performance gain without manual tuning or retraining by human engineers. The framework is open-source but the description provides no further implementation details.
The PyTorch DTensor component updated its operation registration system. Before the change, there were 158 direct op_strategy registrations and 1013 single_dim_strategy registrations, totaling 1164 registered operations. After migration, op_strategy dropped to 114 while single_dim_strategy rose to 1068, for a total of 1176. This reallocates 44 op_strategy entries into the unified single_dim_strategy framework and nets 12 new operations. The refactor simplifies DTensor's op registration maintenance. Testing coverage was exercised via pytest in test/distributed/tensor/test_tensor_ops.py.
This repository provides open-source tools for healthcare AI applications. It aims to democratize access to medical AI models. The project includes resources for model training and deployment. It is suitable for researchers and developers in healthcare.
This release note details a commit that folds the decomposed gelu operation back into the native CUTLASS GELU implementation. The change is part of the inductor and cutlass backend for PyTorch. It aims to improve performance by reducing overhead from the decomposition. This update is likely to enhance efficiency in models using GELU activations.