feat: add rope_in_place tilelang kernel for npu device.

[zhang-minchao]

case	torch_native_ms	tilelang_ms	speedup
1x576_start512_rope64	0.116037	0.0230664	5.03056x
8x576_start512_rope64	0.116174	0.0230544	5.03914x
47x576_start512_rope64	0.119172	0.023024	5.17598x
48x576_start512_rope64	0.118605	0.0232646	5.09811x
49x576_start512_rope64	0.119092	0.023336	5.10335x
95x576_start512_rope64	0.119298	0.0229754	5.19244x
96x576_start512_rope64	0.119057	0.0230834	5.15769x
97x576_start512_rope64	0.118857	0.0230322	5.16046x
128x576_start512_rope64	0.119686	0.0230274	5.19754x
512x576_start512_rope64	0.119086	0.025742	4.62613x
1024x576_start512_rope64	0.136578	0.0328556	4.1569x
2048x576_start512_rope64	0.143184	0.0464708	3.08116x

[gemini-code-assist]

Code Review

This pull request introduces a new TileLang-based RoPE kernel for Ascend NPUs, including the necessary build system infrastructure, C++ wrappers, and tests. The changes are extensive and add significant new functionality. However, there are several critical and high-severity issues that should be addressed. The build process relies on a brittle script patching mechanism in setup.py that could easily break. The C++ wrapper for the kernel has a critical performance issue due to inefficient tensor broadcasting and an unsafe use of const_cast. Additionally, the CMake configuration for building the kernel is not flexible, with hardcoded dimensions and a fragile sed-based code modification step. Addressing these issues will improve the robustness, performance, and maintainability of the new kernel and its build process.

[zhang-minchao]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces a new TileLang kernel for in-place RoPE (Rotary Positional Embedding) on NPU devices, significantly improving performance as demonstrated by the provided benchmarks. The changes include adding a new submodule for tilelang-ascend, comprehensive documentation for kernel development in both English and Chinese, and the necessary Python and CMake infrastructure to compile and integrate these kernels. The implementation includes robust validation checks and a well-structured dispatch mechanism for different kernel variants. Unit tests are also provided to ensure correctness and measure performance.

One high-severity issue was identified related to the patching of a third-party script, which could lead to build fragility.

[yingxudeng]

Nice addition. The included skill enables quick integration trials downstream.