github-featured / jundot--omlx

发布

0 发布 132 Git 标签

RSS 订阅源
  • v0.2.19.dev3 f00d752ea1
    比较

    bump version to v0.2.19.dev3

    frostbyte_neo 发布于 2026-03-18 15:24:35 +00:00

    revert commandBuffer() retained references due to ~60% prefill
    performance regression on large models. dev3 ships per-stream-lock
    patch only, with mx.set_cache_limit(total_mem) to prevent allocator
    from releasing buffers while GPU is still using them.

    下载附件