It's our honour to be one of the supported hardware backends of SGLang ever since July, 2025. Many works were done in the past quarter to support major features and common models. Also we have received a lot of feedbacks and demands that help us improve our using experience and stability, thank you all for your loud voice.

For the next quarter, our goal is continue to improve model performance and increase range of supported features. We are thrilled to announce our development roadmap for 2026 Q1 here:

CI

• Switch to Ascend NPU daily docker release to reduce setup time
• Split PR tests and daily tests to reduce PR test cost

Basic Software Version Bump

• Switch to CANN 8.5.0 stable release [NPU] Adapt cann 8.5: use sfa and lightning indexer op from cann and CI update #17615
• Switch to torch & torch_npu 2.8.0 stable release [NPU] chore: bump to CANN 8.3.RC1 and Pytorch 2.8.0 #13647
• Remove vLLM as a required dependency

User / Developer Experience

• Refresh Ascend NPU docs
• Provide some common deployment guidelines
  • DeepSeek Family
  • Qwen Family

Model Support & Model Performance

• DeepSeek-V2 / DeepSeek-V3 LLM Family (Include DeepSeek-V3.2)
• Kimi-K2 thinking [Ascend] support Kimi-K2-Thinking #12759
• LongCat LLM
• Qwen3 / Qwen3-MoE LLM Family
• Qwen3-Next
• Multi-Modal

Major Features

Ascend NPU Hardware Backend

• Hardware Backend refactorization & abstraction [NPU][1/N] NPU basic functions refactor and new modelslim quant type #13359
• Next-Generation Hardware Support

Parallelism

Expert Parallelism

• Introduce FusedDeepEP mega-kernel that wraps Dispatch & GEMM & Activation & Combine [Ascend] qwen optimization #12078

Context Parallelism

• Support Context Parallel for DeepSeek-V3.2 [NPU]dsv3.2 cp for npu #14541
• Support Context Parallel on DeepSeek Series/Qwen Series with 128K/256K prompt lengths
• Support Sequence Parallel [WIP][Feature] support tp-sp on qwen2/3 & deepseek v2/3/3.2 #12820
• Support KV cache split

NPUGraph and Torch Compile

• Support for torchair compile [feat] npu support enable_torch_compile #12371
• torchair compile enhancement w/ out-of-box multi-stream support
• NPUGraph enabled for egale draft model [Feature] Spec-Overlap supporting DP-ATTN; PD-Disaggregation; npugraph mode #12443

Quantization

• Refactor out Ascend's specific ModelSlim quant type, which supports W8A8, W4A8, W4A4 and W4A16 mixed quantization [NPU] NPU quantization refactoring & more quantization formats support #14504
• Support GPTQ [NPU] support GPTQ quantization on npu #15203

HiCache

• L1 & L2 Radix Cache [Ascend][feature] support L1+ L2 radixcache on ascend #12214
• L3 Cache [NPU][Feature] Support kvcache direct transmission between HBM and distributed kv pool #14056
• Support MoonCake
• RadixCache on DeepSeek V3.2
• RadixCache on Qwen3-Next

Speculative Decoding

• Support DeepSeek-V3 MTP [Feature] Support DeepSeek MTP on NPU #11897
• Support Qwen3-Next MTP [npu] support features of qwen3_next; fixup accuracy bugs in qwen3_next #14391
• Support egale model [NPU][eagle3] support qwen eagle3 on NPU #14820
• Improve Speculative-Overlap w/ DP-Attention & PD Disaggregation [Feature] Spec-Overlap supporting DP-ATTN; PD-Disaggregation; npugraph mode #12443
• Support Suffix Decoding

Reinforcement Learning

• Basic support on Reinforcement Learning

More Frameworks

• MindSpore [Roadmap] MindSpore support plan for 2025 Q4 #13580