本期总览
本期跟踪 9 个仓库,其中 9 个仓库有有效动态,筛出 24 项重点变化。整体重点是:版本发布 7 项、工具链/开发体验 5 项、硬件/架构支持 4 项、性能优化 4 项。
本期跟踪 9 个仓库,其中 9 个仓库有有效动态,筛出 24 项重点变化。整体重点是:版本发布 7 项、工具链/开发体验 5 项、硬件/架构支持 4 项、性能优化 4 项。
AI 与开发工具 本周期有 3 个活跃仓库,筛出 11 项重点变化,主要集中在硬件/架构支持、版本发布、性能优化、工具链/开发体验、功能变化、安全修复。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,主要集中在版本发布、工具链/开发体验、功能变化、硬件/架构支持、性能优化。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 4 个活跃仓库,筛出 5 项重点变化,主要集中在硬件/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 Zephyr、ThreadX、LVGL、Mbed TLS。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。
版本发布出现 7 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 5 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。
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。
性能优化出现 4 项。建议关注是否会改变资源占用、延迟、吞吐或实时性表现。
代表性合并:#109592 [Backport v4.4-branch] dts: nxp: mcxw23: Fix faults caused by unaligned access on shared memory;#107415 [Backport v4.4-branch] drivers: gpio: esp32: fix BIT overflow for pins >= 32;#109683 drivers: eeprom: fake: install default delegate to avoid uninitialized mem
代表性合并:#108694 debug: cpu_load: reset measurement window on timer (re)start;#109451 Add watchdog driver support for microchip pic32cmjh family;#109720 docs: mcumgr: Update mcumgr-toolkit
代表性合并:#10116 refactor(debugging): split profiler and test files into their respective folders;#10183 fix(dma2d): ensure to also enable LP DMA2D clock;#10184 fix(dma2d): fix the way end of dispatch is being handled
代表性合并:#109655 scripts: west_commands: zspdx: handle ValueError of os.path.commonpath;#109640 scripts: twister: tolerate qemu self-cleanup of qemu.pid;#109371 hostap: Fix HostAP legacy crypto for enterprise (CRYPTO_EXT, oid.h)
代表性合并:#529 Added win64 ports of ThreadX and ThreadX SMP
代表性合并:#42585 [Bugfix][V1] Fix TOCTOU race causing intermittent `EADDRINUSE` on multi-API-server DP startup;#43402 [Reasoning] [Bugfix] Reject invalid thinking_token_budget values;#43482 [Bugfix] Apply fc_norm in Eagle3DeepseekV2 combine_hidden_states
代表性合并:#23648 hexagon: add support for CONCAT op;#22596 model : add support for talkie-1930-13b;#23728 ci : add `[no release]` keyword + fix sanitizer builds
代表性合并:#42124 Add LM head quantization support for ModelOpt;#43584 Add CuTe DSL sparse compressor support;#41847 [KV Transfer] Enable HMA by default for connectors that support it
代表性合并:#43709 [CI] Soft-fail AMD entrypoints mirror tests;#43162 [Feat][DSV4] Fuse q pad into deepseek v4 fused kernel;#42789 [MoE Refactor] W4a8 int8 oracle
代表性合并:#23690 CUDA: missing PDL sync for FWHT, better fallback;#23457 [WebGPU] Check batch_compute_passes before sending passes when not doing GPU profiling
代表性合并:#45524 utils: handle flash_attn missing from importlib packages_distributions without crashing;#45534 🚨 [ALM] Add base model without head
代表性合并:#46157 chore: enable Dependabot weekly GitHub Actions bumps
代表性合并:#46221 fix model parallel device mismatch issue in `create_bidirectional_mask`
代表性合并:#43273 [GDN] GDN Prefill kernel for SM100;#43410 [Kernel] Porting fuse_minimax_qk_norm to manual fusion;#38278 [Model] Use AutoWeightsLoader for InternLM2
代表性合并:#23594 ggml-webgpu: Add MMVQ path for Q4/Q8/Q2_K/Q4_K and clean up legacy MUL_MAT pipeline;#22862 SYCL: implement ggml_sycl_pool_vmm;#22620 vulkan: optimize conv2d and implement coopmat1 support
代表性合并:#23721 ci : move macos jobs to the apple workflow + fix names;#23713 ci : move sanitizer jobs to self-hosted runners;#23718 ci : remove vulkan SDK dep from webgpu job
CHANGES WITH 261 in spe: Announcements of Future Feature Removals and Incompatible Changes: systemd logind's integration with the UAPI.1 Boot Loader Specification which allows the systemctl reboot boot loader entry= switch to work so far has supported a special directory /run/boot loader entries/ which allowed defining boot loader entries outside of the ESP/XBOOTLDR partition for compatibility with legacy systems that do not natively implement UAPI.1. However, it appears that to our knowledge it is not actually being used by any project quite unlike UAPI.1 itself, which found adoption far beyond systemd , and its implementation is incomplete. With the future 262 release we intend to remove support for /run/boot loader entries/ and related interfaces, in order to simplify our codebase. Support for UAPI.1 is – of course – kept in place. The experimental "systemd sysupdated" D Bus API is go
代表性合并:#175800 [WebAssembly] WASIP3 Library Call Thread Context Support;#199137 [AArch64] Fix hasNearbyPairedStore to handle non-inbounds GEPs;#197620 [Clang][AArch64] Fix crash with large arguments to MTE built-ins
代表性合并:#183153 [amdgpu-cfi: 5/9]: [AMDGPU] Implement CFI for non-kernel functions;#193367 [lld][macho] Restructure thunk generation algorithm;#198667 [Clang] Emit prefix map normalization before generating hashes for the unique linkage names.
代表性合并:#42295 dissect-image: reject empty partitions array;#42324 meson: fix building without logind;#42306 test: skip TEST-55-OOMD entirely if stress-ng is broken on this host
代表性合并:#42316 udev: ignore errors in adding property with spurious value;#42313 v260 stable batch up to 39a2cc966b34a222584f251757866953fbc86220;#42293 v260 stable batch up to c310106c15ea83913e2765dcb0d7c81d83f08a0e
代表性合并:#199116 [lldb] Fix vtable support on arm64e;#182457 [LV] Add support for partial alias masking with tail folding;#199756 Revert "[clang][driver][darwin] Hold onto full triples in Darwin SDKP…
代表性合并:#42321 test: disable ManagedOOMMemoryPressure on /system.slice in TEST-55-OOMD
代表性合并:#183152 [amdgpu-cfi: 4/9]: [AMDGPU] Emit entry function Dwarf CFI
工具链/开发体验:7 项,dissect-image: reject empty partitions array;meson: fix building without logind
功能变化:6 项,udev: ignore errors in adding property with spurious value;v260 stable batch up to 39a2cc966b34a222584f251757866953fbc86220
性能优化:1 项,test: disable ManagedOOMMemoryPressure on /system.slice in TEST-55-OOMD
工具链/开发体验:9 项,[amdgpu-cfi: 5/9]: [AMDGPU] Implement CFI for non-kernel functions;[lld][macho] Restructure thunk generation algorithm
硬件/架构支持:9 项,[lldb] Fix vtable support on arm64e;[LV] Add support for partial alias masking with tail folding
版本发布:4 项,[WebAssembly] WASIP3 Library Call Thread Context Support;[AArch64] Fix hasNearbyPairedStore to handle non-inbounds GEPs
功能变化:4 项,[MLIR][XeGPU] Fix pass name in RUN command;[lld][WebAssembly] Only include __stack_pointer when needed
硬件/架构支持:24 项,[Backport v4.4-branch] dts: nxp: mcxw23: Fix faults caused by unaligned access on shared memory;[Backport v4.4-branch] drivers: gpio: esp32: fix BIT overflow for pins >= 32
工具链/开发体验:4 项,scripts: west_commands: zspdx: handle ValueError of os.path.commonpath;scripts: twister: tolerate qemu self-cleanup of qemu.pid
版本发布:3 项,debug: cpu_load: reset measurement window on timer (re)start;Add watchdog driver support for microchip pic32cmjh family
性能优化:1 项,doc: kernel: Memory blocks allocator clarifications
功能变化:1 项,Added win64 ports of ThreadX and ThreadX SMP
版本发布:7 项,refactor(debugging): split profiler and test files into their respective folders;fix(dma2d): ensure to also enable LP DMA2D clock
工具链/开发体验:1 项,fix(tjpgd): support EXIF images(#10084)
硬件/架构支持:1 项,docs(xml): flatten XML section into single redirect page to Pro Docs
工具链/开发体验:2 项,mbedtls_config.c missing mbedtls_platform_requirements.h;[backport 4.1] mbedtls_config.c missing mbedtls_platform_requirements.h
版本发布:7 项,hexagon: add support for CONCAT op;model : add support for talkie-1930-13b
性能优化:4 项,ggml-webgpu: Add MMVQ path for Q4/Q8/Q2_K/Q4_K and clean up legacy MUL_MAT pipeline;SYCL: implement ggml_sycl_pool_vmm
工具链/开发体验:3 项,ci : move macos jobs to the apple workflow + fix names;ci : move sanitizer jobs to self-hosted runners
功能变化:2 项,CUDA: missing PDL sync for FWHT, better fallback;[WebGPU] Check batch_compute_passes before sending passes when not doing GPU profiling
工具链/开发体验:9 项,[CI] Soft-fail AMD entrypoints mirror tests;[Feat][DSV4] Fuse q pad into deepseek v4 fused kernel
硬件/架构支持:6 项,[Bugfix][V1] Fix TOCTOU race causing intermittent `EADDRINUSE` on multi-API-server DP startup;[Reasoning] [Bugfix] Reject invalid thinking_token_budget values
版本发布:4 项,[GDN] GDN Prefill kernel for SM100;[Kernel] Porting fuse_minimax_qk_norm to manual fusion
性能优化:3 项,Add LM head quantization support for ModelOpt;Add CuTe DSL sparse compressor support
功能变化:4 项,Fix loading with only 1 device or distributed config;init FSDP through from_pretrained
工具链/开发体验:3 项,Fix a regression in encoder-decoder generation cache initialization;[AMD CI] revert AMD mi325 hf-workflows ref from SHA back to @main
版本发布:2 项,utils: handle flash_attn missing from importlib packages_distributions without crashing;🚨 [ALM] Add base model without head
安全修复:1 项,chore: enable Dependabot weekly GitHub Actions bumps