Implement custom kernel for LLaMA rotary embedding (#14)
This commit is contained in:
@@ -23,6 +23,14 @@ attention_extension = cpp_extension.CUDAExtension(
|
||||
)
|
||||
ext_modules.append(attention_extension)
|
||||
|
||||
# Positional encodings.
|
||||
positional_encoding_extension = cpp_extension.CUDAExtension(
|
||||
name='cacheflow.pos_encoding_ops',
|
||||
sources=['csrc/pos_encoding.cpp', 'csrc/pos_encoding_kernels.cu'],
|
||||
extra_compile_args={'cxx': CXX_FLAGS, 'nvcc': NVCC_FLAGS},
|
||||
)
|
||||
ext_modules.append(positional_encoding_extension)
|
||||
|
||||
setuptools.setup(
|
||||
name='cacheflow',
|
||||
ext_modules=ext_modules,
|
||||
|
||||
Reference in New Issue
Block a user