Incorrect inline ptx device assembly code usage

[zhiweij1]

Branch/Tag/Commit

main

Docker Image Version

N/A

GPU name

N/A

CUDA Driver

N/A

Reproduced Steps

https://github.com/NVIDIA/FasterTransformer/blob/afdf9a9eb86f15363c0249117d166d6b45dbb371/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp#L643

A comma should be inserted between `"r"(a.x)` and `"r"(a.y)`.

Although it can be compiled by nvcc, but when using clang, clang will report an error: https://godbolt.org/z/jxd483a8j

I think it can be fixed in the FasterTransformer source code.