Release v1.1.0

We are excited to announce the release of RecIS v1.1.0. This version marks a significant milestone with the introduction of Model Bank 1.0, native ROCm support, and substantial performance optimizations for large-scale embedding tables.

🌟 Key Highlights

Category	Description
🏆 Framework	Model Bank 1.0 officially arrives; New Negative Sampler and RTP Exporter support.
⚡ Performance	Introduction of Auto-resizing Hash Tables and Fused AdamW TF CUDA operations.
🌐 Compatibility	Expanded hardware support for AMD ROCm; Fixed non-NVIDIA device kernel launches.
🛡️ Robustness	Improved multi-node synchronization and robust handling for empty tensor edge cases.

---

📝 Detailed Changelog

Bug Fixes

• checkpoint: fix mos version format, update use openlm api (854bbb3)
• checkpoint: refine torch_rank_weights_embs_table_multi_shard.json format (d5e7a5c)
• checkpoint: walk around save bug, deal with xpfs model path (ae99728)
• embedding: fix empty kernel launch in non-nvidia device (2e310d0)
• embedding: fix insert when size == 1 (7702c9e)
• framework: add an option for algo_config for export (0ad4c3f)
• framework: fix bugs of invalid index, grad accumulation; add clear child feat (1e7acf9)
• framework: fix eval in trainer (676a053)
• framework: fix fg && exporter bugs (3964ce2)
• framework: fix load extra info not in ckpt (a64cd00)
• framework: fix loss backward (7d9a41b)
• framework: fix some bug of model bank (be196db)
• framework: fix window io failover (cde3049)
• framework: reset io state when start another epoch (f918f24)
• io: fix batch_convert row_splits when dataset read empty data (44661ab)
• io: fix None data when window switch (e788b4d)
• io: fix odps import bug (7c13f09)
• io: use openstorage get_table_size directly (d5c0952)
• ops: fix bug in fast atomic operations (fea8d47)
• ops: fix dense_to_ragged op when check_invalid=False (#14) (300a77b)
• ops: fix edge cases for empty tensors and improve CUDA kernel handling (794be12)
• ops: fix emb segment reduce mean op (3f82b9c)
• ops: handle empty tensor inputs in ragged ops (a39fc2a)
• optimizer: step add 1 should be in-place (cdb3632)
• serialize: fix bug of file sync of multi node (822af49)
• serialize: fix bug of load tensor (e25eee4)
• serialize: fix bug when load by oname (e5ca3d7)
• serialize: fix bug when tensor num < parallel num (a02aded)
• tools: fix torch_fx_tool string format (1d426f8)

Features

• checkpoint: add label for ckpt (5436b5b)
• checkpoint: load dense optimizer by named_parameters (a07dbaf)
• docs: add model bank docs (ff0d23e)
• embedding: add monitor for ids/embs (2f268eb)
• embedding: expose methods to retrieve child ids and embs from the coalesced hashtable; fix clear method of hashtable (b5de207)
• framework,checkpoint: change checkpointmanager to save/load hooks (eb3b441)
• framework: [internal] add negative sampler (8c21517)
• framework: add exporter for rtp (b8af849)
• framework: add skip option in model bank (00828ce)
• framework: add some utility to RaggedTensor (78eca0a)
• framework: add window_iter for window pipline (87886a0)
• framework: collect eval result for hooks and fix after_data bug (81d3723)
• framework: enable amp by options (db5bbe7)
• framework: impl-independent monitor (24a1631)
• framework: model bank 1.0 (488672b)
• framework: support filter hashtable for saver, update hook for window, fix metric (01eb2ae)
• io: add adaptor filter by scene (c3e6738)
• io: add new dedup option for neg sampler (61b2cb7)
• io: add standard fg for input features (2deedff)
• ops: add fused AdamW TF CUDA operation (05dba24)
• ops: add parse_sample_id ops (78674cd)
• packaging: support ROCm (7a626d3)
• serialize: update load metric interface (66b085d)
• update column-io to support ROCm device (7907158)

Performance Improvements

• embedding: use auto-resizing hash table (2f53f53)

---

Release v1.0.0

🎉 Initial Release

RecIS (Recommendation Intelligence System) v1.0.0 is now officially released! This is a unified architecture deep learning framework designed specifically for ultra-large-scale sparse models, built on the PyTorch open-source ecosystem. It has been widely used in Alibaba advertising, recommendation, searching and other scenarios.

✨ New Features

Core Architecture

• ColumnIO: Data Reading
  • Supports distributed sharded data reading
  • Completes simple feature pre-computation during the reading phase
  • Assembles samples into Torch Tensors and provides data prefetching functionality
• Feature Engine: Feature Processing
  • Provides feature engineering and feature transformation processing capabilities, including Hash / Mod / Bucketize, etc.
  • Supports automatic operator fusion optimization strategies
• Embedding Engine: Embedding Management and Computing
  • Provides conflict-free, scalable KV storage embedding tables
  • Provides multi-table fusion optimization capabilities for better memory access performance
  • Supports feature elimination and admission strategies
• Saver: Parameter Saving and Loading
  • Provides sparse parameter storage and delivery capabilities in SafeTensors standard format
• Pipelines: Training Process Orchestration
  • Connects the above components and encapsulates training processes
  • Supports complex training workflows such as multi-stage (training/testing interleaved) and multi-objective computation

🛠️ Installation & Compatibility

System Requirements

• Python: 3.10+
• PyTorch: 2.4+
• CUDA: 12.4

Installation Methods

• Docker Installation: Pre-built Docker images for PyTorch 2.4.0/2.5.1/2.6.0
• Source Installation: Complete build system with CMake and setuptools

Dependencies

• torch>=2.4
• accelerate==0.29.2
• simple-parsing
• pyarrow (for ORC support)

📚 Documentation

• Complete English and Chinese documentation
• Quick start tutorials with CTR model examples
• Comprehensive API reference
• Installation guides for different environments
• FAQ and troubleshooting guides

📦 Package Structure

• Core Library: recis/ - Main framework code
• C++ Extensions: csrc/ - High-performance C++ implementations
• Documentation: docs/ - Comprehensive documentation in RST format
• Examples: examples/ - Practical usage examples
• Tools: tools/ - Data conversion and utility tools
• Tests: tests/ - Comprehensive test suite

🚀 Key Optimizations

Efficient Dynamic Embedding

The RecIS framework implements efficient dynamic embedding (HashTable) through a two-level storage architecture:

• IDMap: Serves as first-level storage, using feature ID as key and Offset as value
• EmbeddingBlocks:
  • Serves as second-level storage, continuous sharded memory blocks for storing embedding parameters and optimizer states
  • Supports dynamic sharding with flexible expansion capabilities
• Flexible Hardware Adaptation Strategy: Supports both GPU and CPU placement for IDMap and EmbeddingBlocks

Distributed Optimization

• Parameter Aggregation and Sharding:
  • During model creation phase, merges parameter tables with identical properties (dimensions, initializers, etc.) into one logical table
  • Parameters are evenly distributed across compute nodes
• Request Merging and Splitting:
  • During forward computation, merges requests for parameter tables with identical properties and computes sharding information with deduplication
  • Obtains embedding vectors from various compute nodes through All-to-All collective communication

Efficient Hardware Resource Utilization

• GPU Concurrency Optimization:
  • Supports feature processing operator fusion optimization, significantly reducing operator count and launch overhead
• Parameter Table Fusion Optimization:
  • Supports merging parameter tables with identical properties, reducing feature lookup frequency, significantly decreasing operator count, and improving memory space utilization efficiency
• Operator Implementation Optimization:
  • Operator implementations use vectorized memory access to improve memory utilization
  • Optimizes reduction operators through warp-level merging, reducing atomic operations and improving memory access utilization

🤝 Community & Support

• Open source under Apache 2.0 license
• Issue tracking and community support
• Active development by XDL Team

---

For detailed usage instructions, please refer to our documentation and quick start guide.