xlnet_preprocess_kernels: blockAttnMask primary template is written as an illegal function-template partial specialization

[diskdog]

src/fastertransformer/kernels/xlnet_preprocess_kernels.cu defines the primary template of blockAttnMask with a template-argument list after the function name:

template<typename T>
void blockAttnMask<T>(dim3& grid, dim3& block, int batch_size, int seq_len)
{
    ...
}

The primary template is already declared in xlnet_preprocess_kernels.h as template<typename T> void blockAttnMask(dim3&, dim3&, int, int);. Writing the definition as blockAttnMask<T>(...) makes it a partial specialization of a function template, which C++ does not permit (only class templates can be partially specialized). nvcc/EDG tolerates it as an extension, but standard-conforming CUDA compilers (clang-CUDA, and AMD-targeting CUDA toolchains) reject it:

error: function template partial specialization is not allowed

The template<> void blockAttnMask<float>(...) further down is a legal full explicit specialization and stays as-is. The call site blockAttnMask<T>(...) in preProcess is also fine (explicit template arguments in a call are allowed).

Fix: drop the stray <T> on the primary-template definition

template<typename T>
void blockAttnMask(dim3& grid, dim3& block, int batch_size, int seq_len)
{
    ...
}

Verified

Minimal reproducer mirroring the header declaration plus the definition, compiled with NVIDIA nvcc (CUDA 13.1, sm_75) and a clang-based CUDA compiler:

code	nvcc	non-nvcc CUDA
original (blockAttnMask<T> definition)	compiles	fails (function template partial specialization)
fix (<T> removed)	compiles	compiles

Behaviour is identical; the fix is a no-op under nvcc and standard-correct, which makes the kernel portable to other CUDA compilers.