Background

Last week we issued an RFC for supporting frameworks other than PyTorch, with priority on supporting MindSpore on Ascend NPU: #7941.

After some discussions, we are issuing the tentative Roadmap for 2025 H2.

Overall Design

• [July] Proof of concept of PyTorch/MindSpore coexistence
• [July] Design Documentation

Inter-Framework Compatibility

• [August] Tensor memory sharing through DLPack
• [August] Compatibility of PyTorch/MindSpore distributed environment
• [August] Resource sharing of PyTorch/MindSpore (stream reuse, memory pool, etc.)

MindSpore Model Support

• [August] Radix Attention support on NPU
• [August] Qwen3 dense models
• [September] Qwen3 MoE model
• [September] DeepSeek V3/R1 model family

SGLang Features

• [September] Combinations of Data/Tensor/Pipeline/Expert Parallels
• [September] Speculative Decoding
• [September] PD disaggregation
• [September] Quantization
• [September] LoRA

CI on Ascend NPU

• [September] CI tests

User / Developer Experience

• [August] Benchmark results and profiling tools
• [August] Docker image
• [September] Documentations (installation, quickstart, tutorials, etc.)

Long-Term Plans

• [Q4] Further optimizations and more MindSpore models

Comments and suggestions are welcome!