Weekly Pre-release 2026-03-29
Pre-release
Pre-release
·
61 commits
to main
since this release
weekly-20260329 (2026-03-29)
✨ Features
- (brpc) Add support for bRPC CPU profiler in CMake configuration @Choimoe
- (client) Add flat parameter update and prefetch result handling methods @Choimoe
- (client) Add flat parameter update and prefetch result handling methods @Choimoe
- (client) Implement flat parameter update and prefetch result handling methods @Choimoe
- (config) Add single day launch configurations and support for gin bindings @Choimoe
- (dist) Migrate DLRM to standard TorchRec with new training script and model definition @Choimoe
- (dlrm) Add support for TF32 mode in recstore dense method and training script @Choimoe
- (env) Enhance step execution with duration tracking and status reporting @Choimoe
- (launch-config) Add support for gin bindings and implement tests for launch configuration handling and CLI overrides @Choimoe
- (model) Implement asynchronous update handling in RecStoreClient @Choimoe
- (op) Add flat prefetch result handling in KVClientOp @Choimoe
- (optimizer) Enhance sparse optimizer with trace aggregation and backend updates @Choimoe
- (release) Add manual GitHub release workflow @Choimoe
- (report) Add performance reporting in OP layer @Choimoe
- (report) Add performance instrumentation to RecStore dlrm and TorchRec dlrm training scripts. @Choimoe
- (report) Enhance latency reporting by dynamically determining storage type from configuration @Choimoe
- (report) Implement async reporting with flame graph support and enhanced error handling @Choimoe
- (report) Instrument EmbRead request path with layered latency tracing using trace_id propagation and flame graph reporting @Choimoe
- (report) Enhance performance reporting in BRPC and gRPC clients and servers @Choimoe
- (report) Add reporting for embedding read and wait operations with performance metrics @Choimoe
- (report) Add local reporting mode handling and related tests @Choimoe
- (report) Implement local JSONL sink and structured event reporting @Choimoe
- (server) Add gperftools profiling support and signal handling @Choimoe
- (tests) Add unit tests for git_path_contributors module functionality @Choimoe
- Implement RecTensor class with embedding operations (EmbRead, EmbUpdate) @Choimoe
- Add PyTorch bindings for embedding operations (emb_read/emb_update) @Choimoe
- Add TensorFlow ops for embedding operations (RecstoreEmbRead/RecstoreEmbUpdate) @Choimoe
- Add HugeCTR ops for embedding operations (emb_read_hugectr and emb_update_hugectr) @Choimoe
- Add gRPC client support in basic op @Choimoe
- Refactor framework with KVClient-centric design and PyTorch C++ OP bindings @Choimoe
- Update client interface, add emb_write support and unit tests and add full test coverage @Choimoe
- Implement KVClient and DistTensor/Emb abstractions with PyTorch op bindings @Choimoe
- Enhance KVClient, DistTensor, and DistEmb interfaces for extensibility @Choimoe
- Add torchrec submodule to .gitmodules @Choimoe
- Add initialization script for setting up DLRM environment and dependencies @Choimoe
- Initial implementation of EBC and custom autograd functions to support embedding lookup and gradient updates with RecStore backend @Choimoe
- Implement async KVClientOp read interfaces and encapsulate gRPC logic @shandianshan
- Using compilation parameters to choose FakeKVClientOp @Choimoe
- Use conditional compilation to switch between real and mock implementations @Choimoe
- Add conditional compilation to upstream CMAKE configuration @Choimoe
- Introduce the RecStore layer to replace and optimize the original framework of dlrm @Choimoe
- Add support for GPU tensors in op_torch layers @Choimoe
- Update EBC's forward and backward propagation logic @Choimoe
- Enable end-to-end DLRM training with torchrun @Choimoe
- Add TorchRec comparison support to EBC tests @Choimoe
- Add extendible hash @wianger
- Add hybrid engine @tong1heng
- Add CCEH_External(syn) @wianger
- Adaptively searching for PS parameter files in the op layer. @Choimoe
- Add download script and .gitignore for Criteo dataset management @Choimoe
- Update distributed embedding module to support gradient tracking and introduce sparse optimizer @Choimoe
- Add single-day data loader supporting arbitrary day loading with fallback to last day for validation if insufficient data @Choimoe
- Add custom Criteo preprocessing and single-day data processing shell scripts @Choimoe
- Update data script with usage instructions, improved sparse index conversion, and shuffling logic @Choimoe
- Add single-day training shell script with data check and training launch logic @Choimoe
- Update single-day training script with default parameter support and detailed log naming @Choimoe
- Update single-day training script with enhanced parameter options and detailed execution logging @Choimoe
- Add torchrec baseline single-day training script with recstore data loader compatibility @Choimoe
- Refactor setup script with resumable step functions @Choimoe
- Add cceh_asyn @wianger
- Add cceh asyn @wianger
- Support cceh insert duplicate key @wianger
- Add torch.profiler to DLRM training for performance monitoring @Choimoe
- Add Python script for performance tracking and statistics @Choimoe
- Implemented prewrite in grpc_client, ready to test @shandianshan
- Add async prefetching support for existing gRPC async interface with OP unit test validation @Choimoe
- Implement model-layer prefetch iterator to proactively fetch feature vectors in background, extending Python client and EBC @Choimoe
- Implement fused single lookup + split path reusing backend key prefix semantics, retain per-feature fallback, and add unit tests @Choimoe
- Add sparse_tensor @wianger
- Add brpc-based KV store support and fix shared factory compilation @Choimoe
- Add embedding table initialization and update methods in op and client @Choimoe
- Add sparse_tensor library and link it to grpc_ps_server @Choimoe
- Enhance embedding initialization and update methods with debug logging and gradient accumulation @Choimoe
- Use dynamic rendezvous ID; enhance optimizer handling in dlrm_main_single_day.py @Choimoe
- Implement RPC parameter server with configuration handling @Choimoe
- Replace PrefetchingIterator with RecStoreDataset for improved data handling and prefetching @Choimoe
- Enable PS server IP configuration propagation from model script to OP layer via SetPSConfig @Choimoe
- Feat(optimizer): refactor gradient handling to use ParameterCompressReader and improve error handling @Choimoe
- Feat(optimizer): refactor SGD and AdaGrad update methods to use ParameterCompressReader for improved efficiency @Choimoe
- Initialize next_page_id in IOBackend constructor and improve resource cleanup in IoUringBackend and SpdkBackend destructors @wianger
- Support batch RPC operations and fix SSD storage backends @Lunatic0123
🐛 Bug Fixes
- (build) Make chmod failsafe and enforce CMake policy min version 3.5 @Choimoe
- (build) Correct ycsb's fmt dependency declaration to enforce build order and resolve parallel compilation race @Choimoe
- (build) Restrict report compilation scope to prevent unintended build and linking @Choimoe
- (build) Add sparse_tensor dependency to bRPC PS Client Library @Choimoe
- (build) Enable performance reporting for bRPC and gRPC servers @Choimoe
- (build) Add Release build type to CMake commands for better optimization @Choimoe
- (ci) Change CMake build type from Debug to Release for profiling @Choimoe
- (ci) Improve logic for skipping ps_server start in CI environments @Choimoe
- (ci) Update ps_server handling logic for better status reporting in CI environments @Choimoe
- (ci) Skip torchrec source build without cuda toolkit @Choimoe
- (ci) Skip torch source build without cuda toolkit @Choimoe
- (dlrm) Add virtual environment activation script for single-day data processing to ensure correct dependency loading @Choimoe
- (dlrm) Add debug prints for input/output statistics in DLRM tests @Choimoe
- (docker) Shrink docker size Minhui Xie
- (docker) Update Torch version to 2.7.1 and adjust installation method @Choimoe
- (ebc) Clear stale prefetch state in strict mode @Choimoe
- (engine) Implement clear method to reset hash table and reinitialize memory @Choimoe
- (engine) Update clear method to reset directory capacity and global depth @Choimoe
- (env) Add -y flag in apt automation script for non-interactive execution @Choimoe
- (grpc) Increased the limit of num_embeddings_per_feature for better scalability @Choimoe
- (install) Correct DGL and Torch version matching in installation script @Choimoe
- (install) Update Torch installation method and automate SSH key setup @Choimoe
- (prefetch) Disable cross-step producer prefetch @Choimoe
- (ps) Add exception handling for server initialization and execution @Choimoe
- (ps) Update PutParameter logic to use PutSingleParameter for improved handling @Choimoe
- (recstore) Enforce sparse flush visibility semantics @Choimoe
- (recstore) Avoid double lr scaling for sparse updates @Choimoe
- (script) Update Python interpreter handling and remove virtual environment path @Choimoe
- (shm_file) Enhance error handling and logging in filesystem operations @Choimoe
- (test) Fix bRPC ps client test comparison @Choimoe
- (test) Fail fast on partially open ps_server ports @Choimoe
- (tests) Add functions to find config file and extract ports from recstore_config.json @Choimoe
- (torch) Update TORCH_VERSION to 2.7.1 in CI and build scripts, add version alignment test @Choimoe
- (torchrec) Align dlrm init with cxx11abi torch env @Choimoe
- (torchrec) Harden init script portability checks @Choimoe
- Fix a bug in cache @Minhui-Xie
- Fix a bug in gather variable tensors @Minhui-Xie
- Fix a bug which can cause fatal error @Ksitta
- Fix put bug @Ksitta
- Fix corotine bug @Ksitta
- Fix bug @Ksitta
- Fix a shocking bug @Ksitta
- Fix cache bug @Ksitta
- Fix a bug in sum_sparse_tensor @Minhui-Xie
- Fix a bug in trace EMB.grad @Minhui-Xie
- Fix a bug in dgl-ke @Minhui-Xie
- Fix a bug in optimizer; index_add_ v.s. index_add @Minhui-Xie
- Fix a bug in sync @Minhui-Xie
- Fix a bug when pass tensor to another process; @Minhui-Xie
- Fix a bug in full_emb.index_add_ @Minhui-Xie
- Fix some concurrent bugs @Minhui-Xie
- Fix a bug in IPCMemory @Minhui-Xie
- Fix a bug in AsyncV1 @Minhui-Xie
- Fix a bug in async v1 @Minhui-Xie
- Fix rec model too fast, no report Timer @Minhui-Xie
- Fix a bug in next(BasePerfSampler), must be sure C++ has seen the prefetched sample @Minhui-Xie
- Fix a bug in ~func() of grad_base.h @Minhui-Xie
- Fixed HugeCTR build process and enhance Docker environment setup with progress logging @Choimoe
- Fix ConcurrentHashMap compilation and traversal @Choimoe
- Correct CITYHASH path and install python3-dev in docker setup @Choimoe
- Remove conflicting libtorch dependency to resolve runtime issues @Choimoe
- Use relative path for lib loading and improve machine detection with warning fallback @Choimoe
- Add conditional compilation to use Torch library @Choimoe
- Fix the call to the recstore operator in tensorflow. @Choimoe
- Add related interfaces in grpc, fix the use of factory methods and some small details. @Choimoe
- Fix the cmake target file of tensorflow. @Choimoe
- Enhance EBC to support FX tracking, optimize forward and backward propagation logic, handle feature length and embedding pooling @Choimoe
- Ensure ids are of data type int64 when reading embeddings @Choimoe
- Delete XMH_SIMPLE_MALLOC @wianger
- Fix the reference of torch to gprc ps client. @Choimoe
- Adapt libtorch version and enable CXX11 ABI @Choimoe
- Enhance unit tests for distributed embedding module with initialization, backward pass, and optimizer step validation @Choimoe
- Align EBCFunction's pooling mode with torchrec by setting it to sum and simplify gradient computation by removing unnecessary scaling @Choimoe
- Enable gradient tracking in RecStoreEBC using hooks, replacing autograd due to internal embedding lookup constraints @Choimoe
- Remove torchrec-specific conditions in RecStoreEBC tests and unify output shape and length_per_key assertions @Choimoe
- Support multi-feature embedding pooling in EBC forward pass and add embedding dimension storage @Choimoe
- Remove outdated update operation test code for op layer @Choimoe
- Correct date print position in single-day data script to resolve compilation failure @Choimoe
- Correct loss calculation in dlrm single-day data script @Choimoe
- Adjust parameter order and correct loss calculation in single-day training script @Choimoe
- Correct DLRM module import in test script @Choimoe
- Add error handling for libtorch download and extraction @Choimoe
- Restore pybind11 support by updating CMake configuration to include pybind11 directories @Choimoe
- Fix HugeCTR installation configuration by temporarily removing DGL compilation @Choimoe
- Fix alloc-dealloc-mismatch problem @xhyf77
- Fix allcator bug @xhyf77
- Add reference to shm_common.h to resolve PetKVData not found error @Choimoe
- Add missing uring dependency, fix user directory symlink, and remove unnecessary progress bar @Choimoe
- Add missing ssd subdirectory for correct cceh static library build @Choimoe
- Fix doc @xhyf77
- Resolve compilation errors @wianger
- Add submodule @wianger
- Fix some bugs @xhyf77
- Move recover_bash to script header to prevent accidental skip and ensure proper virtual environment setup @Choimoe
- Fix grpc @Minhui-Xie
- Fix grpc @AmethystCv
- Fix gradient write-back to KV in EmbeddingBag and DistEmbedding @Choimoe
- Reduce DLRM default feature size for small-scale training and fix log path @Choimoe
- Correct profiler include path in grpc CMake file @Choimoe
- Update script usage instructions and improve error handling @Choimoe
- Resolve fusion operation not taking effect when combined with prefetching @Choimoe
- Improve fused branch dedup logic and fallback condition for size mismatch @Choimoe
- Fix some funcs @AmethystCv
- Add -y flag to apt install in setup script to prevent interactive blocking @Choimoe
- Update brpc and protobuf paths in gRPC CMakeLists.txt from absolute to project-relative paths @Choimoe
- Add cstdint include to tensor.h @Choimoe
- Remove actual update in DistEmb backward for test @Choimoe
- Correct message lifetime in gRPC client and optimize prefetch response waiting logic @Choimoe
- Moving deduplication and inverse mapping to prefetch side, leaving only lightweight index_select in forward @Choimoe
- Fix grpc context @AmethystCv
- Fix compile @tong1heng
- Add error handling for uninitialized optimizer in Update method @Choimoe
- Preserve existing optimizer and append new tables instead of discarding it; create new only if absent @Choimoe
- Update client_test.py for improved error handling and verification in embedding operations @Choimoe
- Fix precision testing for RecStore EBC with network-based updates @Choimoe
- Update installation commands in init_dlrm.sh for consistency and clarity @Choimoe
- Rename RecStoreEmbeddingBagCollection for efficient embedding handling and prefetching @Choimoe
- Improve ShmFile initialization logic for safer and more robust handling @Choimoe
- Resolve critical memory leak in KVEngine by promptly releasing addresses and strengthening error handling @Choimoe
- Add missing implementation for embedding_bag_configs and reset_trace methods in RecStoreEmbeddingBagCollection @Choimoe
- Resolve gRPC PS server shard path collision by appending shard ID to storage path for isolated persistence @Choimoe
- Resolve configuration issue in remote PS server startup script for multi-node deployment @Choimoe
- Correct empty_page check in IoUringBackend destructor and use config.file_path in CCEHTest setup/teardown @wianger
- Resolve and clean up conflicts by unifying optimizer/cache to reader-only Update interface; retain brpc/gRPC server attachment and payload validation while integrating new cache Update calls @Choimoe
- Fix conflict @AmethystCv
◀️ Reverts
- (docker,docs) Restore previous env scripts and remove feature drafts @xhyf77
🤖 AI
- (docs) Add AI agent user manual and development guide @Choimoe
🚜 Refactoring
- (deps) Replace local CityHash copy with google git submodule @Choimoe
- (dlrm) Refactor torchrec DLRM implementation, enhance profiling and architecture @Choimoe
- (kv) Normalize config mapping and migrate allocators @xhyf77
- Remove unnecessary namespaces and optimize the singleton implementation of KVClientOp @Choimoe
- Merge the common parts of the simulated op layer and the real op layer. @Choimoe
- Remove unimplemented methods, optimize singleton mode implementation @Choimoe
- Remove submodule torchrec and use pip direct installation instead. @Choimoe
- Use manual installation to install libtorch and add libtorch dependency in cmake. @Choimoe
- Refactor EBC to support FX tracking, optimize forward and backward propagation logic @Choimoe
- Manually compile and use the torch library with CXX11 ABI. @Choimoe
- Manually compile and use the torch library with CXX11 ABI. @Choimoe
- Optimize EmbWrite with dimension check and logging, deprecate emb_update_torch @Choimoe
- Rework RecStoreEBC forward and backward logic based on SparseSGD implementation @Choimoe
- Refactor RecStoreEBC tests with improved test cases and added support for single-table accuracy testing @Choimoe
- Restructure RecStoreEBC test suite @Choimoe
- Rename RecStoreEBC test cases and fix data and descriptions @Choimoe
- Update custom Criteo data loader to support single-day mode and optimize data processing logic @Choimoe
- Optimize step method of sparse optimizer @Choimoe
- Rework OP layer mock implementation and add prefetch method for testing @Choimoe
- Remove unused DGL-related modules and files @Choimoe
- Remove unused DGL-related modules and files in cmake and submodules @Choimoe
- Modularize parameter server and unify paths and entry points @Choimoe
- Remove legacy storage layer implementation storage-old/ @Choimoe
- Move performance tests and GPU driver components @Choimoe
- Replace custom logging with glog for improved log management and clarity, move pytorch OP tests to op layer root folder @Choimoe
🚀 Performance
- (cceh) Migrate I/O from io_uring to SPDK @wianger
📚 Documentation
- (README) Update environment setup command to capture logs and add logs to .gitignore @Choimoe
- (README) Fix log file name @Choimoe
- (architecture) Add architecture overview document and update navigation configuration @Choimoe
- (env) Update quickstart and quick run guides with environment dependencies and Python version check instructions @Choimoe
- (perf) Add usage guide and configuration steps for bRPC built-in CPU profiler @Choimoe
- (perf) Update performance analysis documentation with Grafana usage guide and reporting protocol @Choimoe
- (run) Add quick run guide and update navigation in mkdocs @Choimoe
- (run) Add data preprocessing instructions for DLRM model @Choimoe
- Update test cases documentation with TorchRec comparison @Choimoe
- Add recstore_config.json documentation covering cache_ps, distributed_client, and client settings @Choimoe
- Improve config doc layout and add detailed explanations @Choimoe
- Add RecStore parameter server module documentation @Choimoe
- Add parameter server module documentation covering core components, gRPC, bRPC, Proto, and RDMA implementations @Choimoe
- Enhance storage layer documentation with detailed descriptions of BaseKV, CachePS, KV engines, and memory management @Choimoe
- Add computation layer comprehensive documentation @Choimoe
- Enhance documentation for computation layer and embedding modules @Choimoe
- Add development testing overview and module-specific test guidelines @Choimoe
- Add performance analysis, profiling, and documentation for the torchrec_dlrm model and framework. @Choimoe
- Refine documentation styling and fix various minor inaccuracies @Choimoe
- Fix missing parameter in base_kv_config example @Lunatic0123
- Update configuration documentation with report API settings for performance analysis @Choimoe
- Add JSON schema reference to recstore_config.json @Choimoe
🎨 Style
- Format code @wianger
🧪 Testing
- Test @Ksitta
- Add unit tests for DistEmb and integrate with ctest @Choimoe
- Move test_dist_emb.py to unittest directory and fix dependencies @Choimoe
- Contains basic functional tests of EBC, including the construction of KJT and the output of embedded search results. @Choimoe
- Add comprehensive unit tests for EBC, including verification of initialization, forward propagation, gradient update, error handling, etc. @Choimoe
- Test & doc: add RecStore EBC test case list, covering initialization, forward propagation, gradient update and other functions @Choimoe
- Test & doc: add RecStore EBC GPU test cases, covering initialization, forward propagation, gradient update and other functions @Choimoe
- Add extendible_hash testing @wianger
- Add engine_extendible_hash test @wianger
- Add CCEH test @wianger
- Add full DLRM training test with single-day data @Choimoe
- Add multi-epoch accuracy test with improved batch generation and output comparison @Choimoe
- Add unit tests for EBC precision with CPU and CUDA support @Choimoe
- Add PS Server management for client tests and EBC precision tests @Choimoe
- Add multiprocess precision test for EBC functionality @Choimoe
- Add multi-machine performance analysis @Choimoe
- Update InitDlrmScriptTest to check pinned versions and installation conditions @Choimoe
⚙️ CI
- (docs) Fix repository and ref settings @Choimoe
- (env) Update TORCH_VERSION to 2.7.1 in CI workflows @Choimoe
- (fix) Update permissions and improve error handling in benchmark scripts @Choimoe
- (labeler) Add ThirdParty and Build sections to labeler configuration @Choimoe
- (release) Automate tagged and weekly releases @Choimoe
- (workflows) Merge cpu build and test checks @Choimoe
- (ycsb) Add YCSB test scripts and environment initialization in workflow @Choimoe
- (ycsb) Add YCSB benchmark workflow and logging script @Choimoe
- Add workflow for CPU builds and Docker container management @Choimoe
- Finish build (cpu) and remove tests @Choimoe
- Fix CI ctest script for gRPC server @Choimoe
- Try DRAM index + SSD value in KV @Choimoe
- Add KV server test with DRAM index + SSD values and OP layer validation to CI pipeline @Choimoe
- Add clang-format check with GitHub Actions bot auto-commit @Choimoe
- Enable auto-commit of format fixes on PRs from same repository to prevent format-related failures @Choimoe
- Remove formatting of PR from fork @Choimoe
- Update CI workflow for improved build process @Choimoe
- Add documentation update workflow and script for tracking changes @Choimoe
- Update paths for push and pull_request triggers in CI workflow @Choimoe
- Update documentation check script for improved error handling @Choimoe
- Improve error handling in documentation validation script @Choimoe
- Update ebc precision to ci test_op @Choimoe
- Update CI workflow for install torchrec @Choimoe
- Update permissions of documentation check workflow for improved file tracking @Choimoe
- Add documentation deployment workflow and configuration files @Choimoe
- Update CI build workflow for improved efficiency and reliability @Choimoe
- Add workflows for pull request labeling and update documentation @Choimoe
- Fix pull_request to pull_request_target for documentation update workflow @Choimoe
- Update workflows for artifact and testing structure @Choimoe
🔧 Chores
- (dependencies) Switch cpptrace to submodule to avoid repeated downloads @Choimoe
- (dlrm) Remove obsolete DLRM code @Choimoe
- (docs) Remove implementation plan @Choimoe
- (gitignore) Ignore project worktrees @Choimoe
- (init) Add libtorch setup with torch 2.0.0 + CUDA 11.8 support @Choimoe
- (op) Remove excessive logging in emb_read_torch and related functions @Choimoe
- (report) Temporarily remove libreport.so dependency in preparation for advanced reporting tool adoption @Choimoe
- (report) Add operation latency reporting for Emb OPs @Choimoe
- Vendor third_party/grpc as git submodule @Choimoe
- Vendor kg/kg/dgl as git submodule and fix install_dgl.sh @Choimoe
- Add HugeCTR submodule and build configuration @Choimoe
- Disable include sorting in ClangFormat to preserve dependency order @Choimoe
- Clean up init.py by removing references to deprecated files @Choimoe
- Set centralized output directory for ctest logs @Choimoe
- Format code @Choimoe
- Format code @Choimoe
- Update CMakeLists.txt to add Folly requirement @Choimoe
- Fix torchrec version and libtorch abi version. @Choimoe
- Ignore DLRM logs and preprocessed numpy files @Choimoe
- Remove debug code @wianger
- Add CmakeLists @wianger
- Reduce debug output in gRPC PutParameter and adjust default log level in OP layer @Choimoe
- Add trace saving path in training script and program, and instrument forward/backward timing output @Choimoe
- Move brpc to third_party/ @Choimoe
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Remove unused PyTorch-related KVClient code @Choimoe
- Update cpptrace subproject to latest commit @Choimoe
- Update logging in ps_server for better error handling and clarity @Choimoe
- Update log messages in rpc servers and op for consistency and clarity @Choimoe
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
- Apply clang-format @github-actions[bot]
BUGFIX
- Dealing with issues of incorrect cleanup and incorrect return values @xhyf77
Build
- (report) Update CMake config for report module and build as static library for easier integration @Choimoe
- Conditionally link report_client and enable perf reporting only when USE_PERF_REPORT is set @Choimoe
Cceh
- Add synchronization interface @wianger
Log
- Add detailed logging function of op layer @Choimoe
- Add detailed logging function of op_torch layer @Choimoe
- Added detailed debug logging function at the OP layer to facilitate tracking and debugging @Choimoe
Todo
- Fix laggard backward @Minhui-Xie
Upgrade
- Update the torch version from 2.0.0 to 2.5.1 @Choimoe
New Contributors ❤️
- @AmethystCv made their first contribution
- @github-actions[bot] made their first contribution
- @xhyf77 made their first contribution
- @Choimoe made their first contribution
- @wianger made their first contribution
- @ made their first contribution
- @Lunatic0123 made their first contribution
- @ddddzhao made their first contribution
- @tong1heng made their first contribution
- @shandianshan made their first contribution
- @Minhui-Xie made their first contribution
- @mia1460 made their first contribution
- @guanjieT made their first contribution
- @Ksitta made their first contribution
Notes:
- These packages intentionally exclude host-provided runtime libraries such as glibc and the system loader.
- Recommended target environments are Ubuntu 20.04/22.04 or similar Linux distributions with a compatible userspace.
- Install runtime prerequisites with
bash ci/env/init_host_prereqs.shbefore use.