Loading / 加载中
This commit introduces index coercion in PyTorch Dynamo for range() arguments and slice members. The change improves Dynamo’s ability to handle Python’s dynamic integer types during tracing, preventing type mismatches in compiled graphs. Authored by guilhermeleobas, it is associated with issue/PR #18. It is a targeted compiler enhancement for PyTorch’s just-in-time compilation pipeline.
PyTorch Dynamo previously lacked support for collections.deque.__init__, causing graph breaks when re-initializing an existing deque. This commit adds handling in DequeVariable.call_method by mirroring CPython's deque_init: it resets maxlen, clears the deque, and extends it with the provided iterable. The implementation also validates maxlen as a non-negative integer, matching CPython's error behavior. This change removes an expected-failure sentinel for CPython 3.13's test_basics, which now passes. The fix improves Dynamo's coverage of Python built-in types, reducing friction for tracing code that uses deques.
A new pull request (#187377) merged into PyTorch adds three helper functions to Dynamo: `long_as_ssize_t`, `number_as_ssize_t`, and `number_index`. These helpers mirror the CPython API functions `PyLong_AsSsize_t`, `PyNumber_AsSsize_t`, and `PyNumber_Index` respectively. The addition improves Dynamo's ability to correctly handle Python size (ssize_t) and index operations during graph tracing, enhancing the compiler's accuracy and compatibility. The change was co-authored by the AI assistant Claude Opus 4.8 (1M context) and approved by a maintainer.
Ollama version 0.30.10 has been released. It introduces support for the Cohere2MoE model, contributed by jmorganca. The underlying llama.cpp backend has been updated to commit b9672 by pdevine. These changes expand Ollama's model compatibility and incorporate recent improvements from the llama.cpp project.
On June 18, Apple's developer channel posted an 88-minute WWDC26 special session video recorded at Steve Jobs Theater. The closing segment demonstrates local inference of the 1-trillion-parameter Kimi K2.6 model across four Mac Studios using LM Studio. This setup leverages RDMA over Thunderbolt, a new macOS Tahoe 26.2 feature that provides microsecond-latency direct memory access between machines. Kimi K2.6, released by Chinese startup Moonshot AI in April 2026, features enhanced coding, long-horizon task execution, and agentic clustering. The video also shows an AI app builder creating a WWDC badge tracker with 3D animation and Visual Intelligence from a single prompt.
In llama.cpp release b9692, the llava_uhd component of the multimodal vision encoder no longer uses the batch dimension. This change resolves issue #24732 by eliminating batch-dimension handling in llava_uhd processing. The release includes platform build statuses for various OS and backend combinations, but the core change is the batch-dimension removal. The fix likely simplifies vision encoding and avoids potential errors in batch processing.
