代表性合并：#108874 net: ssh: replace removed CONFIG_NET_SOCKETS_POLL_MAX；#108640 bluetooth: mesh: fix BLOB IO flash write alignment for non-erase devices；#108593 tests: tf-m: Add exclusion for Edge board in tf-m regression test
代表性合并：#107110 Add SSH support；#108641 modem: cellular: group scripts into struct
代表性合并：#10095 fix(api_map): correctly map v8's display rotation types and defines to v9；#10103 refactor(tjpgd): move 4kb stack allocation to heap；#10104 fix(obj): fix segfault when obj_remove_flag is called for screen
代表性合并：#10093 refactor(include): pass through lvgl_public.h to include public api from src；#10102 fix(api_map): correctly map v8's display rotation types and defines to v9；#10101 fix(api_map): correctly map v8's display rotation types and defines to v9
代表性合并：#10105 fix(examples/stm32): fix deprecated path；#10066 arch(lvgl_private): add lvgl_private inc folder and add missing includes for private headers
代表性合并：#10030 feat(obj): add delete_cb utility；#10079 fix(sdl): remove extraneous function declarations
代表性合并：#41119 [ROCm][Bugfix]: dynamically align BLOCK_DMODEL with Lv in MLA decode kernel；#42181 [Bugfix] Accept canonicalized `modelopt_*` quant_method in `_extract_modelopt_quant_algo`；#41806 fix nixl side-channel host selection
代表性合并：#42244 Avoid silent weights corruption when loading Nemotron Nano VL with reusable-buffer loaders like runai distributed streaming；#42272 [Frontend]Responses API supports chat_template_kwargs；#42188 [Bugfix] Gemma 4 chat template crash with missing tool name and tool id
代表性合并：#38895 bugfix(flashinfer,dcp): remove kv_cache_layout for BatchDCPPrefillWrapper._new_tokens.；#41812 [ROCm][DSv4] implement flash sparse mla with triton kernels；#41942 [ROCm] Clean up a bit the AITER FA backend
代表性合并：#40364 [KV Connector][NIXL][Bugfix] Fix NIXL handshake failures not honoring kv_load_failure_policy
代表性合并：#20275 model: add sarvam_moe architecture support；#21391 [SYCL] Add BF16 support to GET_ROWS operation；#22936 CUDA: directly include cuda/iterator
代表性合并：#45351 fix(testing_utils): guard get_device_capability with torch.cuda.is_available()
代表性合并：#22888 vendor : update cpp-httplib to 0.43.4
代表性合并：#45884 Remove deprecation cycle for `cache_position` in masking primitives；#45885 Remove deprecation cycle for inputs embeds；#45881 [deepseek_v4] fix CSA per-query masking in eager path
代表性合并：#22622 backend sampling: support returning post-sampling probs；#22838 spec : parallel drafting support；#22914 convert : add image break token fallback
代表性合并：#13532 test: integration test hardening
代表性合并：#15431 mlx: refined model push behavior
代表性合并：#195966 [libc] C11 threads: Add cnd_timedwait and mtx_trylock.；#187623 [clangd] Add InsertReplaceEdit for code completion；#196658 [SPIRV] Make sure spirv-tools are copied with exe suffix
代表性合并：#42007 hwdb/keyboard: fix KP_Enter on Clevo PA70ES
代表性合并：#196687 [lldb] Handle SIGINT via the MainLoop signal thread (on POSIX) (#195959)；#191655 [CIR][AArch64] Upstream for vector vector shl；#195725 [mlir][ABI] Add Test target + classification injection helper
代表性合并：#41999 test: fix flaky TEST-07-PID1.socket-defer.sh；#42023 vmspawn: Attach a USB keyboard in GUI mode；#41998 units: enable systemd-report-basic.socket by default
代表性合并：#41775 sysupdate: Allow vacuuming partially downloaded updates and improve error reporting in that case；#42016 elf-util: pass exectable path to dwfl_core_file_report；#42009 test-path-util: add coverage for path edge cases
代表性合并：#195916 [CIR] Use SymbolTableCollection in CallOp/TryCallOp LLVM lowering；#196912 [llvm][tools][llvm-objdump] Fix nested-offload-binary.test ；#196957 [libc] Fix BigInt shift on big-endian platforms
代表性合并：#192469 [mlir] Add erase sub-region dominate tree logic in DominanceInfoBase::invalidate method；#196994 [mlir][SPIR-V] Lower math.{exp2,log2} to GLSL.std.450 ops；#193786 XCOFF no inline ptr glue
工具链/开发体验：9 项，sysupdate: Allow vacuuming partially downloaded updates and improve error reporting in that case；elf-util: pass exectable path to dwfl_core_file_report
功能变化：5 项，curl-util: Fix clang-tidy warnings；repart: make definitions varlink parameter actually optional
硬件/架构支持：3 项，test: fix flaky TEST-07-PID1.socket-defer.sh；vmspawn: Attach a USB keyboard in GUI mode
版本发布：1 项，hwdb/keyboard: fix KP_Enter on Clevo PA70ES
工具链/开发体验：12 项，[CIR] Use SymbolTableCollection in CallOp/TryCallOp LLVM lowering；[llvm][tools][llvm-objdump] Fix nested-offload-binary.test
功能变化：10 项，[mlir] Add erase sub-region dominate tree logic in DominanceInfoBase::invalidate method；[mlir][SPIR-V] Lower math.{exp2,log2} to GLSL.std.450 ops
硬件/架构支持：9 项，[lldb] Handle SIGINT via the MainLoop signal thread (on POSIX) (#195959)；[CIR][AArch64] Upstream for vector vector shl
版本发布：5 项，[libc] C11 threads: Add cnd_timedwait and mtx_trylock.；[clangd] Add InsertReplaceEdit for code completion
硬件/架构支持：21 项，net: ssh: replace removed CONFIG_NET_SOCKETS_POLL_MAX；bluetooth: mesh: fix BLOB IO flash write alignment for non-erase devices
功能变化：3 项，net: net_if: don't zero the iface name；usb: uac2: Adjust log level
工具链/开发体验：3 项，scripts: build: handle duplicate exports in llext exptab prep；Bluetooth: HCI: include addr.h from util_hci_evt
版本发布：2 项，Add SSH support；modem: cellular: group scripts into struct
工具链/开发体验：9 项，refactor(include): pass through lvgl_public.h to include public api from src；fix(api_map): correctly map v8's display rotation types and defines to v9
版本发布：5 项，fix(api_map): correctly map v8's display rotation types and defines to v9；refactor(tjpgd): move 4kb stack allocation to heap
硬件/架构支持：2 项，fix(examples/stm32): fix deprecated path；arch(lvgl_private): add lvgl_private inc folder and add missing includes for private headers
功能变化：2 项，feat(obj): add delete_cb utility；fix(sdl): remove extraneous function declarations
工具链/开发体验：1 项，add probe-rs-meta
性能优化：1 项，Fix Problems with Flashing MCXN947
版本发布：5 项，model: add sarvam_moe architecture support；[SYCL] Add BF16 support to GET_ROWS operation
工具链/开发体验：4 项，backend sampling: support returning post-sampling probs；spec : parallel drafting support
安全修复：1 项，vendor : update cpp-httplib to 0.43.4
性能优化：1 项，vulkan: Support asymmetric FA in scalar/mmq/coopmat1 paths
工具链/开发体验：9 项，Avoid silent weights corruption when loading Nemotron Nano VL with reusable-buffer loaders like runai distributed streaming；[Frontend]Responses API supports chat_template_kwargs
版本发布：6 项，bugfix(flashinfer,dcp): remove kv_cache_layout for BatchDCPPrefillWrapper._new_tokens.；[ROCm][DSv4] implement flash sparse mla with triton kernels
硬件/架构支持：4 项，[ROCm][Bugfix]: dynamically align BLOCK_DMODEL with Lv in MLA decode kernel；[Bugfix] Accept canonicalized `modelopt_*` quant_method in `_extract_modelopt_quant_algo`
性能优化：4 项，add fused mhc_post_pre kernel；[kv_offload] Move `FilterReusedOffloadingManager` logic to `CPUOffloadingManager`
功能变化：3 项，Remove deprecation cycle for `cache_position` in masking primitives；Remove deprecation cycle for inputs embeds
工具链/开发体验：3 项，fix(minicpmv4_6): skip invalid failing tests；[Weight converter] Revert unnecessary changes to `rename_source_key`
版本发布：1 项，fix(testing_utils): guard get_device_capability with torch.cuda.is_available()
工具链/开发体验：1 项，test: integration test hardening
性能优化：1 项，mlx: refined model push behavior