Qwen3.6-27B fused operators

【免费下载链接】pypto-gym PyPTO-Gym 是基于 PyPTO 编程框架构建的算子与模型样例仓库 【免费下载链接】pypto-gym 项目地址: https://gitcode.com/cann/pypto-gym

PyPTO fused-operator implementations targeted at Qwen3.6-27B (qwen3_5 architecture in upstream transformers). Each operator lives in its own subdirectory with an _impl.py exporting a *_wrapper function and a README.md describing the contract.

产品支持情况

  • Ascend 950PR:支持
  • Atlas A3 训练系列产品/Atlas A3 推理系列产品:支持
  • Atlas A2 训练系列产品/Atlas A2 推理系列产品:支持

Operators

Operator Subdirectory Status
gated_delta_rule gated_delta_rule/ Available

Common shape constraints

Qwen3.6-27B's qwen3_5 architecture has:

Symbol Meaning Value
D Per-head dimension 128
Nv Value head count 48
Nqk Query/Key head count 16

Wrappers enforce these constraints at call time and raise NotImplementedError outside the supported envelope.

Integration

The kernels are wired into the model through the modeling_qwen3_5.py shim — see that file for the monkey-patch hook pattern.

【免费下载链接】pypto-gym PyPTO-Gym 是基于 PyPTO 编程框架构建的算子与模型样例仓库 【免费下载链接】pypto-gym 项目地址: https://gitcode.com/cann/pypto-gym

Logo

中国智能体开发者社区,聚焦智能体与大模型开发,提供前沿资讯、实用工具链、开源项目及行业案例。通过技术沙龙、开发者大赛等活动,促进经验交流与协作,助力开发者快速构建创新智能应用。

更多推荐