GitLab

issue/925 - Speed up `scripts/build_ntops.py` and `src/infiniop/ninetoothed/build.py` with `concurrent.futures`

对 `NineToothedTensor` 进行 C++ 层封装

加入使用数组作为 `shape` 和 `strides` 创建 `ninetoothed::Tensor` 的方式

使用 `ninetoothed::Tensor` 接入九齿的 ReLU 算子

Add an include guard to `ninetoothed/utils.h`

- Ensure embedding tensors are on the same device. Change format.
- Optimize embedding kernel with vectorized memory access and __ldg
- Add vectorized memory access using float4/float2, half2, and bfloat162
- Use __ldg instruction for read-only weight and indices access
- Add memory alignment checks to enable vectorized paths
- Add __restrict__ keywords for better compiler optimization
- Implement dynamic block size selection based on embedding_dim

* can commit

* can exec sm_90a

* can exec < sm_90

* fix format

* fix format

* 增加测试，测试对标sglang test

* fix format 1

* fix format 2

* add compile option to disable cutlass

This reverts commit 25258029.

issue/744: softplus

issue/744: kunlun softplus

issue/744: delete F64

Signed-off-by: Ceng23333 <441651826@qq.com>

* issue/739 - support batched RoPE on Nvidia and CPU

* issue/739 - metax, moore batched rope

* issue/739 - adjust metax flags

* issue/739 - added a rope module interface to forward inplace in output tensor