Skip to content

v0.3.13

Latest
Latest

Choose a tag to compare

View all tags
@ShangmingCai ShangmingCai released this 26 Aug 09:17
· 2 commits to release/v0.3.13 since this release
v0.3.13
e5598b0
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's Changed

  • [TENT] Fix io_uring stale status in getTransferStatus by @jfeng18 in #2379
  • ci: add manual "Cancel Queued and Running CI" workflow by @ShangmingCai in #3117
  • ci: fix cancel-ci listing that matched no runs by @ShangmingCai in #3123
  • [TENT] Fix req_map_ leak for single-task batches in ascend transport by @jfeng18 in #3116
  • [PG] Forward collectives through the backend shim for PyTorch 2.13 by @mmangkad in #3122
  • [Store] implement RFC #2756 batch-record HA OpLog writer by @Icedcoco in #2826
  • Fix: allocation_strategy not taking effect when enable_ha is enabled on master by @gitgaoqian in #3077
  • [CI] Add nightly build/release and test workflow by @stmatengss in #2950
  • [CI] Add staryxchen for Mooncake TE's codeowner by @alogfans in #3132
  • [TransferEngine] Harden socket handshake frame handling by @alogfans in #3050
  • [Bugfix] Fix double free of RDMA slice on device-selection failure by @SongOf in #3125
  • [tent] Fix Prometheus histogram silent-drop and clean stale config keys by @staryxchen in #3126
  • [PG][2/N] Introduce Control Plane by @caozhanhao in #2455
  • fix(rdma): auto-chunk MRs larger than device max_mr_size (#2017) by @jiejingzhangamd in #2644
  • [P2P-Store] Fix double-close panic, refCount copy, and offset reset by @jfeng18 in #2375
  • [CI/Build] master image: install ibverbs-providers and publish a CUDA 13 flavor by @tpiperatgod in #3121
  • [CI/Build] Remove CUDA dependency from mooncake_master by @Aionw in #3119
  • [CI/Build] Add ARM64 Non-CUDA release wheels by @Aionw in #3096
  • [Store] Add shared thread-local random helpers by @Aionw in #2954
  • [Store] Fix segment capacity metric leak on master service teardown by @liuzijing2014 in #3154
  • [TENT] Harden ShmTransport backing-file and mapping lifecycle by @alexps9 in #3145
  • [Bugfix] Fix double free of UB/Barex slices on device-selection failure by @he-yufeng in #3146
  • [Store] Surface partial replica allocation in PutStart and document allocation semantics by @liangxu2000 in #3156
  • [Store] Fix SSD total capacity showing incorrectly after master restart by @NUABO in #3069
  • [Store] Fix null-ptr crash in PrepareStorageBackend by @LujhCoconut in #3138
  • [Store] Add dlopen Rust backend and WITH_STORE_C_SHARED shared library by @donghun-furiosa in #3078
  • [Bugfix][Common] Avoid etcd keepalive pings without active streams by @Icedcoco in #3144
  • [TransferEngine] Add NCCL host and device transport backend by @akhillanger in #2852
  • [Store] Make RPC client I/O threads configurable via MC_RPC_CLIENT_IO_THREADS by @Aionw in #3004
  • [Misc] Update CODEOWNERS for docs, CI, and HA by @ykwd in #3175
  • [CI/Build] Align Asio with yalantinglibs target by @CanYangGetYang in #3170
  • [tent] Support mixed DRAM+VRAM segments and multi-transport allocation in tebench by @staryxchen in #3127
  • [CI] Fix nightly workflow: drop removed build-with-ep input by @stmatengss in #3172
  • [TE] fix(efa): bind a CUDA context before registering GPU memory by @whn09 in #3177
  • [Bugfix][TE] Respect remote_accessible in RDMA verbs MR registration by @RuiqingFeng in #3151
  • [Store] Scope capacity metric release to the serving MasterService by @liuzijing2014 in #3168
  • [Store] Fix NoF heartbeat config silently ignored in HA mode by @NUABO in #3149
  • [Doc] Document master KV event publisher flags by @Aionw in #3150
  • [Dep] bind wildcard listen address (0.0.0.0/::) without going through getaddrinfo by @pjdurden in #2919
  • [TENT] Fix failover attempt metrics attribution by @anranxia in #3109
  • [CI/Build] Deduplicate EFA and store CI workflows by @Aionw in #3163
  • [CI/Build] Format only changed C++ lines by @Aionw in #3027
  • [Bugfix][Store] Support JsonCpp header layouts across Linux distributions by @Icedcoco in #3193
  • [Store] support tensor APIs with dummy client by @zxpdemonio in #2998
  • [Store] Fix mem storage display showing 16777216 TB when no segment i… by @leonzzhu in #3080
  • [Store] Add Optional Object Checksum Diagnostics by @yokinoshitayoki in #3003
  • [Store] Fix EraseMetadata leaking stale entries in offloading_objects by @huniu20 in #3160
  • [Bugfix][TE] Split RDMA slices at MR boundaries and handle submit errors by @RuiqingFeng in #3198
  • feat(rdma): enable PCI relaxed ordering by default by @1998zxn in #3205
  • [TransferEngine] Reject peer segment descriptors with more keys than devices by @eharris128 in #3209
  • [PG] Fix single-rank initialization and yalantinglibs dependency by @caozhanhao in #3164
  • [CI/Build] Unify CI and release wheel builds by @Aionw in #3137
  • [CI] Fix TONE SGLang Qwen3.5 cache gating by @Aionw in #3195
  • [Doc] Update RBG integration example to workloads.x-k8s.io/v1alpha2 by @tpiperatgod in #3224
  • Bump google.golang.org/grpc from 1.79.3 to 1.82.1 in /mooncake-common/etcd by @dependabot[bot] in #3235
  • [Store] Add exponential backoff for ordered oplog writer retries by @Icedcoco in #3166
  • [ROCm] Add ROCm/HIP wheel build, CI, and release parity with CUDA (#3171) by @andyluo7 in #3184
  • [Common/TE] Replace hardcoded CUDA paths with CUDAToolkit discovery by @caozhanhao in #3215
  • [TENT] Reject malformed numeric metrics environment overrides by @Hubert-Zhu in #3228
  • [TE] perf(efa): allow bounding the batch MR registration fan-out by @whn09 in #3210
  • [Store] Consolidate string parsing primitives by @Aionw in #3179
  • [Bug]: [SSD] DSV4 enable SSD offload，stress testing, repeat offloading the same batch key, lead to OBJECT_ALREADY_EXISTS/persist failed/INVALID_KEY by @pjdurden in #2967
  • [Bugfix] Clear RDMA context health counter only on successful completions by @SongOf in #3230
  • [Bugfix] Drain RDMA async event queue on each epoll wakeup by @SongOf in #3236
  • [CI] Fix nightly Go path and gate publishing on tests by @Aionw in #3211
  • [Store] Fix silent data corruption in io_uring read path (#3066) by @LujhCoconut in #3073
  • [Bugfix] Detect AMD amdgpu for TENT GPUDirect RDMA capability by @staryxchen in #3242
  • [Store] Selectively materialize BatchEvict candidates after timestamp filtering by @jacklin78911-collab in #3118
  • [Bugfix] Use GPU_PREFIX for HCA peer affinity under USE_HIP by @staryxchen in #3248
  • [CI/Build] Add CUDA 13 EFA wheel by @Aionw in #3247
  • [Bugfix][Store] Fix No rkey for MR access on SSD offload read by @NUABO in #3246
  • [CI/Build] Fix nightly workflow and offload platform checks by @Aionw in #3148
  • [Store] Validate durable oplog prefix at startup by @Icedcoco in #3165
  • [Bugfix][TE] Fix dma-buf offsets for chunked GPU MRs by @RuiqingFeng in #3243
  • [Store]Support VRAM Segment and NVLink transport within a single node by @zhaoyongke in #3196
  • [CI/Build] Restrict nightly workflow to upstream repository by @Aionw in #3268
  • add SUPA (Biren GPU) support to transfer engine by @cyqmonkey in #3102
  • [Bugfix][TE] Keep context_list_ index-aligned when an RNIC fails to init by @SongOf in #3263
  • [TE] perf(efa): optionally register device memory only on topology-local NICs by @whn09 in #3219
  • [Store] Fix EFA transport auto-discovery by @yyun-cpu in #3207
  • [Bugfix][TENT] Bound peer key lookups by the published vector length by @SongOf in #3265
  • [Store] support CUDA IPC for dummy client buffers by @zxpdemonio in #3234
  • [Store] Add Fast-Fail for Impossible OffsetAllocator Requests by @bitborne in #3222
  • Add scatter transfer batching to transfer engine by @zxpdemonio in #3000
  • [Store][TE] fabric_mem best-effort alloc via percentile ladder by @ascend-direct-dev in #3214
  • [PG] Decouple PG from PyTorch by @caozhanhao in #3206
  • test: poll for the async eviction in PutStartExpiringTest by @he-yufeng in #3266
  • [Store] fix: Honor OpLog poll interval command-line override by @00fish0 in #3255
  • [TENT] Retire stale endpoint on peer reconnect bootstrap by @guptaishaan in #3188
  • fix: add CUDA stubs dir to link_directories for no-driver builds by @stmatengss with @Copilot in #3237
  • [Store] Fix processing_keys double-erase UAF in MetadataAccessorRW by @rockuw in #3254
  • [Store] Fix CUDA IPC ranged read build by @zxpdemonio in #3294
  • [Misc] Remove unused Python imports and variables from CodeQL findings by @stmatengss in #3260
  • Update CODEOWNERS to include zxpdemonio by @stmatengss in #3288
  • fix(transfer-engine): block SIGTERM/SIGINT for the shutdown watcher by @he-yufeng in #3278
  • [Wheel] Add native fast-copy PUT hot path by @zxpdemonio in #3112
  • [EP] Add EP dispatch/combine benchmark with incast routing patterns by @PACTHEMAN123 in #3213
  • [Doc] Add Biren logo to supported hardware by @Aionw in #3298
  • [TENT] cuda_probe: zero-init cudaPointerAttributes and warn on libcud… by @KubrickLiu in #3261
  • fix(musa): add missing CUDA compatibility macros by @UNIDY2002 in #3295
  • [TENT] RDMA: skip NICs that cannot GPUDirect-DMA to a GPU, and fix fallback device rotation by @anranxia in #3281
  • [TE] EFA: derive WR/CQ pacing depths from the libfabric provider by @whn09 in #3296
  • [Bugfix][TENT] Keep context_set_ and buffer keys NicID-indexed by @SongOf in #3264
  • [Store] Add minimal MasterService scenario DSL by @Aionw in #3256
  • [TENT] Fix double-free in HttpMetaStore under concurrent workers: a s… by @KubrickLiu in #3259
  • [CI/Build] Simplify wheel workflows by @Aionw in #3283
  • [PG] Build the device worker with C++17 for CMake 3.22 compatibility by @caozhanhao in #3308
  • [TransferEngine] Make single unregisterLocalMemory best-effort by @he-yufeng in #2962
  • [TransferEngine] Skip the CUDA pointer probe on GPU-less hosts by @he-yufeng in #2955
  • fix: resolve nightly CI failures by @UNIDY2002 in #3309
  • [TE] Support DMA-BUF in IBGDA Device API by @UNIDY2002 in #3272
  • [TENT] Add native UB foundation and URMA resource management by @zchuango in #3282
  • [CI] Disable Qoder auto code review workflow by @stmatengss with @Copilot in #3306
  • [Bugfix][TE] Serialize libnuma's lazy cpumask cache fill in bindToSocket by @SongOf in #3289
  • [Store] Harden Put/Upsert completion against promotion lifecycle by @Lin-z-w in #3244
  • [CI] synchronize metadata server handoff by @zhangzuo21 in #3212
  • [Docs][Store] document task manager replica task behavior by @stmatengss with @Copilot in #3300
  • [Store] Fix stale per-segment metric labels after segment unmount by @leonzzhu in #3301
  • [TE] Fix compiler warnings in IBGDA transport by @leonzzhu in #3312
  • [Wheel] Support jagged NestedTensor transfer by @zxpdemonio in #3336
  • [PG] Support inplace rejoin and fix p2p recovery by @caozhanhao in #3323
  • [Wheel] Optimize typed-ragged rollout layout by @zxpdemonio in #3113
  • [Bugfix] clean up partially initialized TENT RDMA contexts by @runzhech in #3325
  • [Bugfix][TENT] Keep a never-constructed context in the NicID slot by @SongOf in #3339
  • [Store] Support pinned SSD-to-GPU restore by @zxpdemonio in #3319
  • [Store] Fix flaky standby catch-up test by @Icedcoco in #3273
  • [TE] Support split IBGDA control memory for MUSA by @UNIDY2002 in #3334
  • [Store] Remove dead DRAM/NoF metric overload declarations by @Hubert-Zhu in #3344
  • [Store] Remove transient MasterService config fields by @Aionw in #3313
  • [Store] Extend MasterScenario DSL for upsert lifecycle by @bitborne in #3328
  • [Wheel] Add DataProto catalog for fragmented rollout data by @zxpdemonio in #3345
  • [Store] Add put/get session APIs for ranged multi-buffer transfers by @ascend-direct-dev in #2881
  • Support custom SSH ports in SPDK target creation by @stmatengss with @Copilot in #3307
  • [Doc] Update Mooncake PG design documentation by @caozhanhao in #3353
  • [Doc] Update PyPI package documentation by @Aionw in #3356
  • Allow configuring master admin/metrics HTTP bind address by @nogumanov in #3318
  • [CI/Build] Fix nightly CUDA stub runtime by @Aionw in #3346
  • [Transfer Engine] Schedule TCP transfers on bounded per-peer connection lanes by @jacklin78911-collab in #2974
  • [Doc] Align Python setup API documentation by @Aionw in #3357
  • [CI/Build] Isolate nightly Python tests in venv by @Aionw in #3373
  • [Store] Add batch OpLog snapshot metadata protocol by @Icedcoco in #3178
  • [Store] Add producer view to durable oplog prefix by @Icedcoco in #3201
  • [Store] Add fixed soft pin lifecycle by @Lin-z-w in #2909
  • [Bugfix] use BufferPool for zcopy benchmark by @030611 in #3342
  • [Store] Avoid staging for same-process GPU reads by @zxpdemonio in #3197
  • [Store] Add deterministic MasterScenario object lifecycle coverage by @bitborne in #3368
  • [TransferEngine] Fix graceful shutdown test hangs in USE_ETCD builds by @Icedcoco in #3351
  • [Store] Fix race in HA durable-finalization tests (#3332) by @Icedcoco in #3333
  • feat(ascend_direct): auto-detect AutoConnect & Client-Server mode via GetCapability, inject LocalCommRes by @JieTang66 in #3302
  • [Doc] Add Multi-tenant Deployment Guide by @yokinoshitayoki in #3378
  • [Doc] Align pre-commit guidance to PR-changed files by @staryxchen in #3277
  • [Bugfix][TE] Reject MC_IB_PORT=0 instead of disabling every RNIC by @SongOf in #3381
  • [Store] Fix current-stream readiness for DummyClient CUDA IPC tensor writes by @mo-ke-ke in #3303
  • [Store] Support configurable fileread worker pool size by @NUABO in #3348
  • [Bugfix][TE] Install shutdown handlers before unblocking signals by @SongOf in #3390
  • [Store] Verify supervisor view propagation by @Icedcoco in #3352
  • [TENT] drain control callbacks on unregister by @zxpdemonio in #3370
  • [TransferEngine] Add instant bandwidth reporting to tebench by @alogfans in #3358
  • [Bugfix][Store] Update bucket timestamp after offload to avoid immedi… by @mjwtom in #3101
  • [Store] Add deterministic MasterService eviction scenarios by @Aionw in #3327
  • [Store] Make tenant quota charge and release lock-free by @Lin-z-w in #3162
  • [Store] Add DFS replica support with POSIX backend by @fcczzz in #2683
  • [TENT] Port GDS and Ascend transport reliability updates by @Primary33 in #3280
  • [Docs] Reorganize design docs and move tebench to performance by @ykwd in #3376
  • [EP] decouple native EP from torch C++ APIs by @UNIDY2002 in #2883
  • [Store] Add terminal state for ordered oplog writer failures by @Icedcoco in #3204
  • [CI] Adjust build concurrency in build-musa workflow by @UNIDY2002 in #3415
  • [CI/Build] Prevent fixed-port collisions in nightly tests by @Icedcoco in #3403
  • [EP] Fix incorrect target name under condition EP_USE_IDE by @UNIDY2002 in #3418
  • [TransferEngine] Share one host KV segment across a TP group by @ascend-direct-dev in #3285
  • [Reshard] Add resource and model-weight manifest contracts by @Bo-Vincent in #3187
  • [Store] Defer invalid handle cleanup from segment unmount RPC by @Aionw in #2877
  • [store] Avoid staging copy for same node tensor put by @zxpdemonio in #3159
  • [TE] Add RDMA ctrl frame codec and sender credit ledger by @zhtshr in #3324
  • [Bug]: RDMA reconnect storm after mlx5 local length errors by @pjdurden in #3387
  • [Bugfix][TENT] Lock ProxyManager's stage buffer map by @SongOf in #3410
  • MPComm Transport for TENT — an RDMA-native multi-rail transport backend by @c-guo16 in #3423
  • [Build] Stage native _ep extension into ep_pg_staging for wheel packaging by @JunlinW113 in #3432
  • [ci] enable ARM64 Mooncake EP build by @UNIDY2002 in #3442
  • [Store] Fix NoF benchmark build: gflags flags inside anonymous namespace by @NUABO in #3420
  • [Bugfix][Store] Keep snapshot unmount test on the synchronous cleanup path by @SongOf in #3429
  • [Doc] Updating maintainers document with new logs by @stmatengss with @Copilot in #3430
  • [Store] Add object storage adapter abstraction by @fcczzz in #3043
  • [Bugfix][Store] Release allocator gaps after snapshot recovery by @Icedcoco in #3414
  • [Bugfix][Store] Avoid futile PutStart eviction triggers by @Icedcoco in #3419
  • [Store] introduce NVMe KV backend core by @zxpdemonio in #2167
  • [Store] Fix: preempt in-progress offload task on UpsertStart by @huniu20 in #3253
  • [TENT] Support custom NIC priority matrix and native topology dump by @staryxchen in #3382
  • [Bugfix][TENT] Pace the transfer poll loops and stop leaking their batches by @SongOf in #3407
  • [CI/Build] Install patchelf from PyPI for auditwheel repair by @SongOf in #3458
  • [Transfer Engine] Optimize scatter transfer with grouped tasks by @zxpdemonio in #3310
  • [TENT] Make ProxyManager staging cleanup safe and asynchronous by @RuiqingFeng in #3380
  • [BugFix][TENT] Fix race condition in DispatchesOnlyOneWindowOnSubmit test by @RuiqingFeng in #3460
  • [Bugfix][Store] Fix hybrid histogram serialization by @JimmyWang0417 in #3456
  • [Bugfix][TENT] Keep a delegated transfer off the RPC event loop by @SongOf in #3404
  • [TE] Skip same-host device offset when HIXL CS mode is available by @jinsidong in #3453
  • [Store] Return explicit standby restore errors by @Icedcoco in #3354
  • [Bugfix][Store] Preserve RequiredParam names when copied by @JimmyWang0417 in #3457
  • [Store] Add dynamic hot replica fanout by @zxpdemonio in #3389
  • feat(cuda): add sm103 to CUDA 13 builds by @UNIDY2002 in #3444
  • [CI] Move PR build validation to nightly by @UNIDY2002 in #3482
  • [CI] Simplify build-flags by @UNIDY2002 in #3485
  • [CI] avoid pinning torch version to 2.11.0 by @UNIDY2002 in #3469
  • [TENT] Resolve transport policy device masks once, not per request by @SongOf in #3490
  • [TENT] reclaim rdma endpoints outside store lock by @zxpdemonio in #3369
  • TENT: link tent_runtime against mooncake_common by @xiaodouzi666 in #3405
  • TENT: only deliver transfer-bound notifications after completion by @xiaodouzi666 in #3406
  • [TransferEngine] Add MUSA IPC transport with batched copies by @yeahdongcn in #3035
  • [TransferEngine] Bind buffer device for EFA CUDA loopback by @yejun-yun in #3436
  • [Store] Stabilize snapshot PutStart eviction test by @Icedcoco in #3480
  • [Store] Structure OffsetAllocatorBackendConfig environment settings by @bitborne in #3408
  • [Store] Extract LocalSSD management from SegmentManager by @Aionw in #3427
  • [Bugfix][Store] Fix MoveEnd source refcnt leak on target gone by @lxy-alexander in #3443
  • [CI][Store] Fix promotion-on-hit e2e flakiness from cross-class segment leaks by @LujhCoconut in #3466
  • [Store] Bucket: add MAX_PHYSICAL_BYTES cap on real shared-disk usage by @Morpheus799 in #3467
  • [Bugfix][Store] Protect unreadable recovered replicas from eviction by @Icedcoco in #3421
  • [Store]: Split segments by transport limit by @yokinoshitayoki in #3486
  • [Store] Fence batch oplog writers by producer view by @Icedcoco in #3384
  • [TransferEngine] Support multi-target tebench runs by @zxpdemonio in #3391
  • [Store] Add bounded standby snapshot capture by @Icedcoco in #3326
  • [TransferEngine] Split EFA transfers that straddle BufferDesc chunk boundaries by @whn09 in #3505
  • [Docs][TENT] Document the FakeTransport runtime test mechanism by @staryxchen in #3489
  • [Bugfix][Store] Fix standby snapshot compile regressions by @Icedcoco in #3515
  • [Store] Clean up behavior-focused tests by @Aionw in #3495
  • [CI/Build] Make CMake target dependencies explicit by @Aionw in #3276
  • fix(store): deduplicate SSD carryover keys by @982945902 in #3479
  • [Store][TE] Split Ascend agent-mode store pool across n engines by @ascend-direct-dev in #3450
  • [Bugfix][TE] Log the RDMA port number as an integer, not a byte by @SongOf in #3512
  • [TENT] Make RPC server thread count configurable via MC_TENT_RPC_THREADS by @gogongxt in #3535
  • [TENT] Reduce copies in the TCP data path (sendData/onRecvData) by @gogongxt in #3536
  • [TE] Rename create_shared_segment tp_group to comm_group by @ascend-direct-dev in #3481
  • [Bugfix][Store] Fix partial-unmount snapshot test race by @thunguo in #3549
  • [transfer-engine] Ship intra-node NVLink transport in x86_64 CUDA wheels by @ShangmingCai in #3547
  • [Other] Add Aionw as codeowner for Store HA by @ykwd in #3524
  • [Reshard] Add N-D logical weight transfer planner by @Bo-Vincent in #3441
  • [Store] Give LOCAL_DISK segments an unmount half, and let the offload RPC honour connect timeouts by @Juhyun-Kim-Memphis in #3316
  • [Store] Add chunked standby snapshot artifact writer by @Icedcoco in #3447
  • fix(store): reject cross-shard duplicates during standby restore by @982945902 in #3527
  • [Store] Skip full metadata scan in ReMountSegment when unneeded (#3517) by @KubrickLiu in #3533
  • perf(store): batch io_uring bucket reads by @982945902 in #3488
  • [wheel] preserve grouped writes for structured objects by @zxpdemonio in #3133
  • [CI/Build] Fix Ascend master linkage by @Aionw in #3565
  • [Build] Establish scikit-build-core Python project foundation by @Aionw in #3577
  • [TE] Enhance Transfer Engine Rust library with CI tests and fixes by @stmatengss in #3461
  • [TE] Expand config env-var test coverage by @jfeng18 in #2409
  • [Docs] Update News by @ykwd in #3580
  • [TE] Add rdma_twosided control-plane notify channel by @zhtshr in #3440
  • [Bugfix][Store] Allow HA promotion without OpLog by @Icedcoco in #3566
  • [Bugfix][TENT] Unpin remote stage buffers after late completion by @RuiqingFeng in #3494
  • [Store] Structure FileStorageConfig environment settings by @bitborne in #3525
  • [CI] Install sgl-eval for T-One SGLang tests by @UNIDY2002 in #3605
  • [TransferEngine] Add a TCP session progress deadline by @jacklin78911-collab in #3542
  • [TransferEngine] Enable TCP connection pool by default by @alogfans in #3569
  • [TENT] fix: synchronize with caller's per-thread stream before issuing NVLink/MNNVL batched copy by @TTThanos in #3570
  • [Bugfix][TENT] Keep lazyFreeBatch sweeping past a batch it cannot reclaim by @SongOf in #3514
  • [CI/Build] Header only YLT (yalantinglibs) by @ur4t in #3537
  • [Store] Fix promotion-on-hit delivery starvation and retry loss by @LujhCoconut in #3545
  • [TransferEngine] Fix duplicate cleanup in transport tests by @ASCII-S in #3583
  • [Docs] Add UCL-MPComm to the Updates list in README, and measured performance data in docs. by @c-guo16 in #3621
  • [CI] Add Reshard module test gate by @Bo-Vincent in #3622
  • Bump version to 0.3.13 in pyproject.toml by @ShangmingCai in #3597
  • [TransferEngine] Add FlagCX support by @MC952-arch in #3522
  • [TENT] Bound the pending-batch drain wait in ProxyManager shutdown by @jfeng18 in #3526
  • [CI/Build] Preserve CTest failure diagnostics by @Aionw in #3445
  • [Store] Decouple business decisions from storage metrics by @Hubert-Zhu in #3385
  • [TENT] Canonicalize AMD GPU location prefix to hip: (keep rocm: as parse alias) by @staryxchen in #3532
  • [Store] Establish typed MasterService test DSL by @Aionw in #3510
  • [TransferEngine] Fix TCP endpoint refresh correctness by @jacklin78911-collab in #3543
  • [Store] Shrink metadata maps after eviction cycles by @jfeng18 in #3576
  • [Store] Avoid direct-I/O padding for POSIX restores by @zupengwang in #3606
  • [Bugfix][TENT] Fix RailMonitor topology use-after-free by @lxy-alexander in #3598
  • [Wheel] Add DataProto catalog lifecycle management by @zxpdemonio in #3374
  • fix(ep,pg): prevent cudaErrorIllegalAddress during CUDA graph capture with TBO by @UNIDY2002 in #3609
  • [Bugfix] Isolate Mooncake Store test data directories by @fcczzz in #3650
  • [TransferEngine] Link gcov runtime for Rust coverage builds by @alogfans in #3612
  • [MUSA] Backport YLT include propagation to v0.3.13 by @Aionw in #3694

New Contributors

Full Changelog: v0.3.12.post1...v0.3.13

Contributors

whn09, rockuw, and 77 other contributors
Assets 53
Loading