fep(sig-framework): add PyTorch-Plugin-FL v0.1.0 CUDA backend dispatch proposal

[Hchnr]

No description provided.

[buzhengjing]

Thanks for the proposal.

From a reviewer perspective, the current FEP provides a good architectural overview, but it does not yet contain sufficient implementation and validation details for reproducible verification.

In particular, the proposal does not currently specify:

• Validation environment (image, dependencies, versions)
• Installation and configuration procedures
• Supported operators in the initial scope
• Detailed test cases
• Expected execution results
• Backend selection verification methodology
• Native CUDA vs. FlagGems comparison criteria
• CI/regression validation strategy

Without these details, it is difficult for reviewers to reproduce the proposed workflow or assess the completeness of the implementation plan.

Could you consider adding a dedicated "Implementation and Validation Plan" section covering environment setup, test procedures, sample commands, expected outputs, and acceptance criteria?

[Hchnr]

Update: expand CUDA dispatch to multi-backend architecture with full verification plan

Major updates to the PyTorch-Plugin-FL v0.1.0 FEP:

• Rename from "CUDA Backend" to "Multi-Backend Operator Dispatch"
• Add Ascend (Huawei) native kernel support alongside CUDA
• Introduce Dispatcher template-based routing mechanism
• Support three dispatch paths: native CUDA, native Ascend, FlagGems Triton (C++ and Python)
• Define 32 first-phase operators with cross-platform implementations
• Add detailed architecture diagrams and registration flow
• Provide complete testing strategy with per-operator and end-to-end tests (Qwen3-0.6B)
• Document full verification environments for both CUDA (A800) and Ascend (910B) platforms
• Include step-by-step installation, test procedures, and expected outputs
• Add CI/CD integration and regression testing guidelines

Scope: Expands from CUDA-only prototype to production-ready multi-backend framework with comprehensive validation plan.

[buzhengjing]

could you share the full Docker image pytorch2.11.0_cuda12.8_triton3.6.0_flaggems5.0.2?

[Hchnr]

docker pull harbor.baai.ac.cn/flagscale/cuda12.8.1-cudnn9.15.1-python3.12-torch2.7.1-train:2512031616