本期总览
本期跟踪 6 个仓库,其中 6 个仓库有有效动态,筛出 18 项重点变化。整体重点是:硬件/架构支持 4 项、版本发布 4 项、工具链/开发体验 4 项、安全修复 3 项。
本期跟踪 6 个仓库,其中 6 个仓库有有效动态,筛出 18 项重点变化。整体重点是:硬件/架构支持 4 项、版本发布 4 项、工具链/开发体验 4 项、安全修复 3 项。
AI 与开发工具 本周期有 3 个活跃仓库,筛出 8 项重点变化,主要集中在版本发布、硬件/架构支持、工具链/开发体验、安全修复、性能优化、功能变化。代表仓库包括 llama.cpp、vLLM、Transformers。这类变化更适合从本地推理、开发者工作流和模型应用集成角度跟进。
Linux 与基础软件 本周期有 2 个活跃仓库,筛出 8 项重点变化,主要集中在安全修复、硬件/架构支持、版本发布、工具链/开发体验、功能变化。代表仓库包括 systemd、LLVM。这类变化更适合从发行版同步、系统稳定性、网络/安全边界和长期维护成本角度评估。
RTOS 与嵌入式 本周期有 1 个活跃仓库,筛出 2 项重点变化,主要集中在硬件/架构支持、安全修复。代表仓库包括 Zephyr。这类变化更适合从板级适配、驱动成熟度、实时系统可维护性和量产风险角度评估。
硬件/架构支持出现 4 项。对嵌入式开发的意义主要在于芯片、板级、驱动和架构覆盖继续补齐。
版本发布出现 4 项。适合检查新版本说明、兼容性变化和是否值得进入下一轮评估。
工具链/开发体验出现 4 项。对日常工程效率、调试链路、构建系统和 CI 稳定性更有参考价值。
安全修复出现 3 项。优先关注是否影响现有系统的升级窗口、补丁回归和下游发行版同步节奏。
代表性合并:#108498 drivers: counter: mcux_lptimer: add reset API support;#109087 scripts: dashboard: Add memory report search and fix top-ten type style;#108767 drivers: sensor: adltc2990: Add support for fetching SENSOR_CHAN_ALL
代表性合并:#109098 doc: security: Additional info for CVE-2025-10456
代表性合并:#42120 Revert "shared/pager: add support for more(1) pager in secure mode"
代表性合并:#196825 [SLP] Enable full non-power-of-2 vectorization by default;#167783 [AArch64] Fix handling of x29/x30 in inline assembly clobbers;#197606 [NewPM] Add newpm port for AArch64LowerHomogeneousPrologEpilog
代表性合并:#198086 [libc++] Introduce a private version of in_out_result and use it for copy/move algorithms
代表性合并:#194414 [llvm-ir2vec] Breaking up llvm-ir2vec lib implementation to clean up MIR deps from ir2vec python bindings;#198155 [clang-tidy] Fix crash in misc-static-initialization-cycle;#197633 [compiler-rt] Fix StackDepot benchmark thread barrier
代表性合并:#42140 dhcp-client: clear overloaded sname/file fields after parsing;#42127 Final batch of conversions to option and verb macros;#42141 hwdb: Add keymap for Lenovo 15IRX9
代表性合并:#196240 [Clang][ItaniumMangle][NFC] Refactor FunctionTypeDepthState;#187860 Add clang warning if fp exception functions are called without appropriate flags/pragmas;#198178 [InstCombine] Convert logical and/or with trunc nuw to i1 into bitwise ops
代表性合并:#42121 socket-util: extend the buffer in sockaddr_union to make it consistent with hw_addr_data;#41065 Introduce several more helper functions to read sysfs attributes and use them where applicable
代表性合并:#42138 sd-dhcp-client: use dhcp_message_send_{udp,raw}() and update BPF code;#42135 dhcp: cleanups for sd-dhcp-lease.c and rewrite test-dhcp-client;#42129 sd-dhcp-client: two tiny cleanups
代表性合并:#22009 ggml-vulkan/CMakeLists: add a check for SPIRV-Headers;#22673 llama + spec: MTP Support ;#23204 cmake : do not install conversion script
代表性合并:#23125 server : honor --embd-normalize CLI arg;#22677 vulkan: add cpy bf16 -> f32 pipelines;#23017 feat: Support d_conv=15 for ssm-conv.cu
代表性合并:#42710 [MRV2][XPU] add Model Runner V2 log;#42725 [XPU] fix weight scale shape;#42716 Fix Weight loading for Qwen3.5-MTP and Qwen3-VL using runai_streamer
代表性合并:#45983 bugfix(ci): avoid E2BIG in pr_slow_ci_suggestion
代表性合并:#42686 [torch.compile] Add patch for fullgraph compilation;#42810 [ROCm] [Bugfix] Fix DeepSeek V4 Functionality and Accuracy
代表性合并:#22637 vulkan: Support unaligned tensors for ROPE;#23127 ui: untrack settings sync in props effect to prevent reactive loop;#23173 common : enable streaming JSON argument values
代表性合并:#42304 [Experimental] Breakable CUDA graph
代表性合并:#23102 CUDA: Continue directly including cuda/iterator;#23185 ngram : reduce noisy logs;#23143 sync : ggml
评分:82
工具链/开发体验:6 项,sd-dhcp-client: use dhcp_message_send_{udp,raw}() and update BPF code;dhcp: cleanups for sd-dhcp-lease.c and rewrite test-dhcp-client
功能变化:4 项,networkctl: introduce 'dhcp-lease' command to dump acquired DHCP lease;sd-dhcp-client: use new message parser
版本发布:3 项,dhcp-client: clear overloaded sname/file fields after parsing;Final batch of conversions to option and verb macros
硬件/架构支持:2 项,socket-util: extend the buffer in sockaddr_union to make it consistent with hw_addr_data;Introduce several more helper functions to read sysfs attributes and use them where applicable
功能变化:13 项,[Clang][ItaniumMangle][NFC] Refactor FunctionTypeDepthState;Add clang warning if fp exception functions are called without appropriate flags/pragmas
工具链/开发体验:12 项,[llvm-ir2vec] Breaking up llvm-ir2vec lib implementation to clean up MIR deps from ir2vec python bindings;[clang-tidy] Fix crash in misc-static-initialization-cycle
硬件/架构支持:3 项,[SLP] Enable full non-power-of-2 vectorization by default;[AArch64] Fix handling of x29/x30 in inline assembly clobbers
性能优化:3 项,[AMDGPU] Add dot product patterns with saturating add (clamp);[X86] LowerBUILD_VECTORvXi1 - attempt to fold as VPTESTMB(BUILD_VECTOR_vXi8(X),1)
硬件/架构支持:6 项,drivers: counter: mcux_lptimer: add reset API support;scripts: dashboard: Add memory report search and fix top-ten type style
安全修复:1 项,doc: security: Additional info for CVE-2025-10456
硬件/架构支持:5 项,server : honor --embd-normalize CLI arg;vulkan: add cpy bf16 -> f32 pipelines
版本发布:4 项,ggml-vulkan/CMakeLists: add a check for SPIRV-Headers;llama + spec: MTP Support
工具链/开发体验:3 项,vulkan: Support unaligned tensors for ROPE;ui: untrack settings sync in props effect to prevent reactive loop
功能变化:3 项,CUDA: Continue directly including cuda/iterator;ngram : reduce noisy logs
工具链/开发体验:5 项,[MRV2][XPU] add Model Runner V2 log;[XPU] fix weight scale shape
版本发布:2 项,[torch.compile] Add patch for fullgraph compilation;[ROCm] [Bugfix] Fix DeepSeek V4 Functionality and Accuracy
性能优化:1 项,[Experimental] Breakable CUDA graph
硬件/架构支持:1 项,[KV Connector] Support disk offloading in MooncakeStoreConnector
安全修复:1 项,bugfix(ci): avoid E2BIG in pr_slow_ci_suggestion