Repository navigation
torch_gcu v2.10.0 Release Notes
Initial public release of torch_gcu, Enflame's PyTorch backend extension for GCU devices.
torch_gcu integrates with PyTorch through the PrivateUse1 backend mechanism and enables PyTorch workloads to run on Enflame GCU hardware with a familiar user experience across Python and C++ workflows.
Highlights
- Initial open-source release of torch_gcu on GitHub
- PyTorch 2.10.0 support
- Python 3.10 support
- GCU execution support for PyTorch tensor and model workloads
- ECCL-based distributed training backend
- torch.compile / Inductor integration for graph-level optimization
- Automatic Mixed Precision support through torch.gcu.amp
- PyTorch profiler integration for GCU activity tracing
- GCU Graph support through torch.gcu.GCUGraph
- CUDA-to-GCU migration utility via transfer_to_gcu
- libtorch_gcu support for C++ inference workflows
- Bundled manuals and operator support documentation
What is included
This release publishes the initial public codebase for torch_gcu, including:
- Core runtime integration for Enflame GCU devices
- ATen operator implementations and dispatch registration
- Python package and extension bindings
- Distributed communication support based on ECCL
- Profiler and debugging utilities
- Build system, packaging scripts, and code generation infrastructure
- User manuals, migration guidance, and development tools
Key capabilities
- Broad ATen operator coverage with CPU fallback for unsupported operators
- Distributed collective and point-to-point communication through backend="eccl"
- AMP training support with torch.gcu.amp.autocast() and GradScaler
- GCU Graph execution for reducing dispatch overhead
- CUDA code migration with minimal code changes
- C++ integration through libtorch_gcu for inference scenarios
Compatibility
PyTorch and Python
- PyTorch: 2.10.0
- Python: 3.10
Enflame software stack
- Recommended TopsRider SDK: v3.7.1
Hardware
- Enflame CloudBlazer Inference S60
Installation
Two supported installation paths are documented in the repository:
- Docker image based setup for the fastest onboarding path
- Build from source with wheel packaging
For environment prerequisites and full setup steps, refer to the repository README and manuals.
Notes
- This tag currently corresponds to the initial public commit for the repository.
- The public commit history for v2.10.0 contains a single visible commit: Initial release: torch_gcu v2.10.0.
- Detailed feature scope in this release note is based on the publicly available repository contents and README for the v2.10.0 release.