metal: enable Metal 4.0 tensor API on M5+/A19+ (#27461)

* metal : request Metal 4.0 language version for the tensor API

* metal : load the tensor API kernels from a separate metallib

* tests : add external-metallib tensor API regression test

* metal : fix metallib build order for the tensor API kernels
This commit is contained in:
James Francis
2026-09-01 12:02:42 +03:00
committed by GitHub
parent 234a6ebaa0
commit d5d993a093
3 changed files with 124 additions and 34 deletions
+20 -1
View File
@@ -163,19 +163,37 @@ else()
)
endforeach()
# the tensor API kernels go in a separate metallib, loaded only where supported
set(AIR_MM_TENSOR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/mul_mm_tensor.air")
add_custom_command(
OUTPUT ${AIR_MM_TENSOR}
COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -DGGML_METAL_HAS_TENSOR -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/mul_mm.metal -o ${AIR_MM_TENSOR}
DEPENDS kernels/mul_mm.metal kernels/common.h kernels/dequantize.h ${METALLIB_COMMON} ggml-metal-impl.h
COMMENT "Compiling kernels/mul_mm.metal (tensor API)"
VERBATIM
)
add_custom_command(
OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
COMMAND xcrun -sdk macosx metallib ${AIR_MM_TENSOR} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
DEPENDS ${AIR_MM_TENSOR}
COMMENT "Linking tensor API Metal kernels into ggml-tensor.metallib"
)
add_custom_command(
OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
DEPENDS ${AIR_FILES}
DEPENDS ${AIR_FILES} ${AIR_MM_TENSOR}
COMMENT "Linking Metal kernels into default.metallib"
)
add_custom_target(
ggml-metal-lib ALL
DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
)
endif() # GGML_METAL_EMBED_LIBRARY
@@ -188,6 +206,7 @@ if (NOT GGML_METAL_EMBED_LIBRARY)
install(
FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
DESTINATION ${CMAKE_INSTALL_BINDIR}
)
endif()