diff --git a/ggml/src/CMakeFiles/ggml-base.dir/DependInfo.cmake b/ggml/src/CMakeFiles/ggml-base.dir/DependInfo.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..9ddfd2301cc4e02674fb7aec9a98c433a5e21a1c
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-base.dir/DependInfo.cmake
@@ -0,0 +1,30 @@
+
+# Consider dependencies only in project.
+set(CMAKE_DEPENDS_IN_PROJECT_ONLY OFF)
+
+# The set of languages for which implicit dependencies are needed:
+set(CMAKE_DEPENDS_LANGUAGES
+  )
+
+# The set of dependency files which are needed:
+set(CMAKE_DEPENDS_DEPENDENCY_FILES
+  "/content/tmp/ggml/src/ggml-alloc.c" "ggml/src/CMakeFiles/ggml-base.dir/ggml-alloc.c.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/ggml-alloc.c.o.d"
+  "/content/tmp/ggml/src/ggml-quants.c" "ggml/src/CMakeFiles/ggml-base.dir/ggml-quants.c.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/ggml-quants.c.o.d"
+  "/content/tmp/ggml/src/ggml.c" "ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o.d"
+  "/content/tmp/ggml/src/ggml-backend.cpp" "ggml/src/CMakeFiles/ggml-base.dir/ggml-backend.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/ggml-backend.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-opt.cpp" "ggml/src/CMakeFiles/ggml-base.dir/ggml-opt.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/ggml-opt.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-threading.cpp" "ggml/src/CMakeFiles/ggml-base.dir/ggml-threading.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/ggml-threading.cpp.o.d"
+  "/content/tmp/ggml/src/ggml.cpp" "ggml/src/CMakeFiles/ggml-base.dir/ggml.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/ggml.cpp.o.d"
+  "/content/tmp/ggml/src/gguf.cpp" "ggml/src/CMakeFiles/ggml-base.dir/gguf.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-base.dir/gguf.cpp.o.d"
+  )
+
+# Targets to which this target links which contain Fortran sources.
+set(CMAKE_Fortran_TARGET_LINKED_INFO_FILES
+  )
+
+# Targets to which this target links which contain Fortran sources.
+set(CMAKE_Fortran_TARGET_FORWARD_LINKED_INFO_FILES
+  )
+
+# Fortran module output directory.
+set(CMAKE_Fortran_TARGET_MODULE_DIR "")
diff --git a/ggml/src/CMakeFiles/ggml-base.dir/ggml-alloc.c.o.d b/ggml/src/CMakeFiles/ggml-base.dir/ggml-alloc.c.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..793d22e514823e437ef360b65ff5d70370ce4502
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-base.dir/ggml-alloc.c.o.d
@@ -0,0 +1,88 @@
+ggml/src/CMakeFiles/ggml-base.dir/ggml-alloc.c.o: \
+ /content/tmp/ggml/src/ggml-alloc.c /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/math.h /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h /usr/include/string.h \
+ /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+ /usr/include/limits.h /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+ /usr/include/linux/limits.h \
+ /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+ /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/uio_lim.h
diff --git a/ggml/src/CMakeFiles/ggml-base.dir/ggml-threading.cpp.o b/ggml/src/CMakeFiles/ggml-base.dir/ggml-threading.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..da521f42a8de4188d6f050e85c1c47144206d389
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-base.dir/ggml-threading.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o.d b/ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..2eb92b0032995ad8fd0a2753415caf093294e5d6
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o.d
@@ -0,0 +1,150 @@
+ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o: /content/tmp/ggml/src/ggml.c \
+ /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/math.h /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h /usr/include/string.h \
+ /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /content/tmp/ggml/src/ggml-threading.h \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/ggml-quants.h /content/tmp/ggml/src/ggml-common.h \
+ /usr/include/errno.h /usr/include/x86_64-linux-gnu/bits/errno.h \
+ /usr/include/linux/errno.h /usr/include/x86_64-linux-gnu/asm/errno.h \
+ /usr/include/asm-generic/errno.h /usr/include/asm-generic/errno-base.h \
+ /usr/include/x86_64-linux-gnu/bits/types/error_t.h /usr/include/time.h \
+ /usr/include/x86_64-linux-gnu/bits/time.h \
+ /usr/include/x86_64-linux-gnu/bits/timex.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+ /usr/include/inttypes.h /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+ /usr/include/limits.h /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+ /usr/include/linux/limits.h \
+ /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+ /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/uio_lim.h /usr/include/signal.h \
+ /usr/include/x86_64-linux-gnu/bits/signum-generic.h \
+ /usr/include/x86_64-linux-gnu/bits/signum-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sig_atomic_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/siginfo_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigval_t.h \
+ /usr/include/x86_64-linux-gnu/bits/siginfo-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/siginfo-consts.h \
+ /usr/include/x86_64-linux-gnu/bits/siginfo-consts-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigval_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigevent_t.h \
+ /usr/include/x86_64-linux-gnu/bits/sigevent-consts.h \
+ /usr/include/x86_64-linux-gnu/bits/sigaction.h \
+ /usr/include/x86_64-linux-gnu/bits/sigcontext.h \
+ /usr/include/x86_64-linux-gnu/bits/types/stack_t.h \
+ /usr/include/x86_64-linux-gnu/sys/ucontext.h \
+ /usr/include/x86_64-linux-gnu/bits/sigstack.h \
+ /usr/include/x86_64-linux-gnu/bits/sigstksz.h /usr/include/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/posix_opt.h \
+ /usr/include/x86_64-linux-gnu/bits/environments.h \
+ /usr/include/x86_64-linux-gnu/bits/confname.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_posix.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_core.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd_ext.h \
+ /usr/include/linux/close_range.h \
+ /usr/include/x86_64-linux-gnu/bits/ss_flags.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_sigstack.h \
+ /usr/include/x86_64-linux-gnu/bits/sigthread.h \
+ /usr/include/x86_64-linux-gnu/bits/signal_ext.h /usr/include/syscall.h \
+ /usr/include/x86_64-linux-gnu/sys/syscall.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd_64.h \
+ /usr/include/x86_64-linux-gnu/bits/syscall.h \
+ /usr/include/x86_64-linux-gnu/sys/stat.h \
+ /usr/include/x86_64-linux-gnu/bits/stat.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_stat.h \
+ /usr/include/x86_64-linux-gnu/bits/statx.h /usr/include/linux/stat.h \
+ /usr/include/linux/types.h /usr/include/x86_64-linux-gnu/asm/types.h \
+ /usr/include/asm-generic/types.h /usr/include/asm-generic/int-ll64.h \
+ /usr/include/x86_64-linux-gnu/asm/bitsperlong.h \
+ /usr/include/asm-generic/bitsperlong.h /usr/include/linux/posix_types.h \
+ /usr/include/linux/stddef.h \
+ /usr/include/x86_64-linux-gnu/asm/posix_types.h \
+ /usr/include/x86_64-linux-gnu/asm/posix_types_64.h \
+ /usr/include/asm-generic/posix_types.h \
+ /usr/include/x86_64-linux-gnu/bits/statx-generic.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_statx_timestamp.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_statx.h \
+ /usr/include/x86_64-linux-gnu/sys/wait.h \
+ /usr/include/x86_64-linux-gnu/sys/prctl.h /usr/include/linux/prctl.h \
+ /usr/include/execinfo.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/DependInfo.cmake b/ggml/src/CMakeFiles/ggml-cpu.dir/DependInfo.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..e230e138834424c2f97132337d34944740e7b736
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/DependInfo.cmake
@@ -0,0 +1,37 @@
+
+# Consider dependencies only in project.
+set(CMAKE_DEPENDS_IN_PROJECT_ONLY OFF)
+
+# The set of languages for which implicit dependencies are needed:
+set(CMAKE_DEPENDS_LANGUAGES
+  )
+
+# The set of dependency files which are needed:
+set(CMAKE_DEPENDS_DEPENDENCY_FILES
+  "/content/tmp/ggml/src/ggml-cpu/arch/x86/quants.c" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/ggml-cpu.c" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/quants.c" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/amx/amx.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/amx/mmq.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/arch/x86/repack.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/binary-ops.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/ggml-cpu.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/hbm.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/ops.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/repack.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/traits.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/unary-ops.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o.d"
+  "/content/tmp/ggml/src/ggml-cpu/vec.cpp" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o" "gcc" "ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o.d"
+  )
+
+# Targets to which this target links which contain Fortran sources.
+set(CMAKE_Fortran_TARGET_LINKED_INFO_FILES
+  )
+
+# Targets to which this target links which contain Fortran sources.
+set(CMAKE_Fortran_TARGET_FORWARD_LINKED_INFO_FILES
+  )
+
+# Fortran module output directory.
+set(CMAKE_Fortran_TARGET_MODULE_DIR "")
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/build.make b/ggml/src/CMakeFiles/ggml-cpu.dir/build.make
new file mode 100644
index 0000000000000000000000000000000000000000..483d6f50b61cef15d7fc2a27f1f58326bb265972
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/build.make
@@ -0,0 +1,340 @@
+# CMAKE generated file: DO NOT EDIT!
+# Generated by "Unix Makefiles" Generator, CMake Version 3.31
+
+# Delete rule output on recipe failure.
+.DELETE_ON_ERROR:
+
+#=============================================================================
+# Special targets provided by cmake.
+
+# Disable implicit rules so canonical targets will work.
+.SUFFIXES:
+
+# Disable VCS-based implicit rules.
+% : %,v
+
+# Disable VCS-based implicit rules.
+% : RCS/%
+
+# Disable VCS-based implicit rules.
+% : RCS/%,v
+
+# Disable VCS-based implicit rules.
+% : SCCS/s.%
+
+# Disable VCS-based implicit rules.
+% : s.%
+
+.SUFFIXES: .hpux_make_needs_suffix_list
+
+# Command-line flag to silence nested $(MAKE).
+$(VERBOSE)MAKESILENT = -s
+
+#Suppress display of executed commands.
+$(VERBOSE).SILENT:
+
+# A target that is always out of date.
+cmake_force:
+.PHONY : cmake_force
+
+#=============================================================================
+# Set environment variables for the build.
+
+# The shell in which to execute make rules.
+SHELL = /bin/sh
+
+# The CMake executable.
+CMAKE_COMMAND = /usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake
+
+# The command to remove a file.
+RM = /usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -E rm -f
+
+# Escaping for special characters.
+EQUALS = =
+
+# The top-level source directory on which CMake was run.
+CMAKE_SOURCE_DIR = /content/tmp
+
+# The top-level build directory on which CMake was run.
+CMAKE_BINARY_DIR = /content/tmp/build
+
+# Include any dependencies generated for this target.
+include ggml/src/CMakeFiles/ggml-cpu.dir/depend.make
+# Include any dependencies generated by the compiler for this target.
+include ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.make
+
+# Include the progress variables for this target.
+include ggml/src/CMakeFiles/ggml-cpu.dir/progress.make
+
+# Include the compile flags for this target's objects.
+include ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+
+ggml/src/CMakeFiles/ggml-cpu.dir/codegen:
+.PHONY : ggml/src/CMakeFiles/ggml-cpu.dir/codegen
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o: /content/tmp/ggml/src/ggml-cpu/ggml-cpu.c
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_1) "Building C object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o -c /content/tmp/ggml/src/ggml-cpu/ggml-cpu.c
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing C source to CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/ggml-cpu.c > CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling C source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/ggml-cpu.c -o CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o: /content/tmp/ggml/src/ggml-cpu/ggml-cpu.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_2) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o -c /content/tmp/ggml/src/ggml-cpu/ggml-cpu.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/ggml-cpu.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/ggml-cpu.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o: /content/tmp/ggml/src/ggml-cpu/repack.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_3) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o -c /content/tmp/ggml/src/ggml-cpu/repack.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/repack.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/repack.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o: /content/tmp/ggml/src/ggml-cpu/hbm.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_4) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o -c /content/tmp/ggml/src/ggml-cpu/hbm.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/hbm.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/hbm.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o: /content/tmp/ggml/src/ggml-cpu/quants.c
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_5) "Building C object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o -c /content/tmp/ggml/src/ggml-cpu/quants.c
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing C source to CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/quants.c > CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling C source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/quants.c -o CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o: /content/tmp/ggml/src/ggml-cpu/traits.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_6) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o -c /content/tmp/ggml/src/ggml-cpu/traits.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/traits.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/traits.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o: /content/tmp/ggml/src/ggml-cpu/amx/amx.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_7) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o -c /content/tmp/ggml/src/ggml-cpu/amx/amx.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/amx/amx.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/amx/amx.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o: /content/tmp/ggml/src/ggml-cpu/amx/mmq.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_8) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o -c /content/tmp/ggml/src/ggml-cpu/amx/mmq.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/amx/mmq.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/amx/mmq.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o: /content/tmp/ggml/src/ggml-cpu/binary-ops.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_9) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o -c /content/tmp/ggml/src/ggml-cpu/binary-ops.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/binary-ops.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/binary-ops.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o: /content/tmp/ggml/src/ggml-cpu/unary-ops.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_10) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o -c /content/tmp/ggml/src/ggml-cpu/unary-ops.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/unary-ops.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/unary-ops.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o: /content/tmp/ggml/src/ggml-cpu/vec.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_11) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o -c /content/tmp/ggml/src/ggml-cpu/vec.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/vec.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/vec.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o: /content/tmp/ggml/src/ggml-cpu/ops.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_12) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o -c /content/tmp/ggml/src/ggml-cpu/ops.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/ops.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/ops.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o: /content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_13) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o -c /content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o: /content/tmp/ggml/src/ggml-cpu/arch/x86/quants.c
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_14) "Building C object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o -c /content/tmp/ggml/src/ggml-cpu/arch/x86/quants.c
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing C source to CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/arch/x86/quants.c > CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling C source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/cc $(C_DEFINES) $(C_INCLUDES) $(C_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/arch/x86/quants.c -o CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.s
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o: /content/tmp/ggml/src/ggml-cpu/arch/x86/repack.cpp
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o: ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_15) "Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -MD -MT ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o -MF CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o.d -o CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o -c /content/tmp/ggml/src/ggml-cpu/arch/x86/repack.cpp
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CXX source to CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.i"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -E /content/tmp/ggml/src/ggml-cpu/arch/x86/repack.cpp > CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.i
+
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CXX source to assembly CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.s"
+	cd /content/tmp/build/ggml/src && /usr/bin/c++ $(CXX_DEFINES) $(CXX_INCLUDES) $(CXX_FLAGS) -S /content/tmp/ggml/src/ggml-cpu/arch/x86/repack.cpp -o CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.s
+
+# Object files for target ggml-cpu
+ggml__cpu_OBJECTS = \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o" \
+"CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o"
+
+# External object files for target ggml-cpu
+ggml__cpu_EXTERNAL_OBJECTS =
+
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/build.make
+bin/libggml-cpu.so: bin/libggml-base.so
+bin/libggml-cpu.so: /usr/lib/gcc/x86_64-linux-gnu/11/libgomp.so
+bin/libggml-cpu.so: /usr/lib/x86_64-linux-gnu/libpthread.a
+bin/libggml-cpu.so: ggml/src/CMakeFiles/ggml-cpu.dir/link.txt
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --bold --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_16) "Linking CXX shared library ../../bin/libggml-cpu.so"
+	cd /content/tmp/build/ggml/src && $(CMAKE_COMMAND) -E cmake_link_script CMakeFiles/ggml-cpu.dir/link.txt --verbose=$(VERBOSE)
+
+# Rule to build all files generated by this target.
+ggml/src/CMakeFiles/ggml-cpu.dir/build: bin/libggml-cpu.so
+.PHONY : ggml/src/CMakeFiles/ggml-cpu.dir/build
+
+ggml/src/CMakeFiles/ggml-cpu.dir/clean:
+	cd /content/tmp/build/ggml/src && $(CMAKE_COMMAND) -P CMakeFiles/ggml-cpu.dir/cmake_clean.cmake
+.PHONY : ggml/src/CMakeFiles/ggml-cpu.dir/clean
+
+ggml/src/CMakeFiles/ggml-cpu.dir/depend:
+	cd /content/tmp/build && $(CMAKE_COMMAND) -E cmake_depends "Unix Makefiles" /content/tmp /content/tmp/ggml/src /content/tmp/build /content/tmp/build/ggml/src /content/tmp/build/ggml/src/CMakeFiles/ggml-cpu.dir/DependInfo.cmake "--color=$(COLOR)"
+.PHONY : ggml/src/CMakeFiles/ggml-cpu.dir/depend
+
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/cmake_clean.cmake b/ggml/src/CMakeFiles/ggml-cpu.dir/cmake_clean.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..5d88e5fe64fe6d6988be57be505a673d8209221b
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/cmake_clean.cmake
@@ -0,0 +1,39 @@
+file(REMOVE_RECURSE
+  "../../bin/libggml-cpu.pdb"
+  "../../bin/libggml-cpu.so"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o.d"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o"
+  "CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o.d"
+)
+
+# Per-language clean rules from dependency scanning.
+foreach(lang C CXX)
+  include(CMakeFiles/ggml-cpu.dir/cmake_clean_${lang}.cmake OPTIONAL)
+endforeach()
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.make b/ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.make
new file mode 100644
index 0000000000000000000000000000000000000000..c66607b1cbc813ec8066a854300164b96cf0caac
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.make
@@ -0,0 +1,2 @@
+# Empty compiler generated dependencies file for ggml-cpu.
+# This may be replaced when dependencies are built.
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts b/ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
new file mode 100644
index 0000000000000000000000000000000000000000..d2eb8267bfde1d4f9db6281111744f766bc45c3d
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/compiler_depend.ts
@@ -0,0 +1,2 @@
+# CMAKE generated file: DO NOT EDIT!
+# Timestamp file for compiler generated dependencies management for ggml-cpu.
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/depend.make b/ggml/src/CMakeFiles/ggml-cpu.dir/depend.make
new file mode 100644
index 0000000000000000000000000000000000000000..ec4e32800c84e15fd0c4f268ab3bd20575139ec6
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/depend.make
@@ -0,0 +1,2 @@
+# Empty dependencies file for ggml-cpu.
+# This may be replaced when dependencies are built.
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/flags.make b/ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
new file mode 100644
index 0000000000000000000000000000000000000000..eec0649904a7275ff89724eb54efadac0ef96b4b
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/flags.make
@@ -0,0 +1,17 @@
+# CMAKE generated file: DO NOT EDIT!
+# Generated by "Unix Makefiles" Generator, CMake Version 3.31
+
+# compile C with /usr/bin/cc
+# compile CXX with /usr/bin/c++
+C_DEFINES = -DGGML_BACKEND_BUILD -DGGML_BACKEND_SHARED -DGGML_SCHED_MAX_COPIES=4 -DGGML_SHARED -DGGML_USE_CPU_REPACK -DGGML_USE_LLAMAFILE -DGGML_USE_OPENMP -D_GNU_SOURCE -D_XOPEN_SOURCE=600 -Dggml_cpu_EXPORTS
+
+C_INCLUDES = -I/content/tmp/ggml/src/.. -I/content/tmp/ggml/src/. -I/content/tmp/ggml/src/ggml-cpu -I/content/tmp/ggml/src/../include
+
+C_FLAGS = -O3 -DNDEBUG -std=gnu11 -fPIC -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int -Werror=implicit-function-declaration -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wdouble-promotion -march=native -fopenmp
+
+CXX_DEFINES = -DGGML_BACKEND_BUILD -DGGML_BACKEND_SHARED -DGGML_SCHED_MAX_COPIES=4 -DGGML_SHARED -DGGML_USE_CPU_REPACK -DGGML_USE_LLAMAFILE -DGGML_USE_OPENMP -D_GNU_SOURCE -D_XOPEN_SOURCE=600 -Dggml_cpu_EXPORTS
+
+CXX_INCLUDES = -I/content/tmp/ggml/src/.. -I/content/tmp/ggml/src/. -I/content/tmp/ggml/src/ggml-cpu -I/content/tmp/ggml/src/../include
+
+CXX_FLAGS = -O3 -DNDEBUG -std=gnu++17 -fPIC -Wmissing-declarations -Wmissing-noreturn -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-array-bounds -Wextra-semi -march=native -fopenmp
+
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..979e6e5170282572cae9c54f54710ffe25c23fd3
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..2e4e1cbca5c1262a1798ed197c79d53fa34789e3
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o.d
@@ -0,0 +1,290 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/amx/amx.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/amx/amx.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/amx/common.h \
+ /usr/include/c++/11/algorithm /usr/include/c++/11/utility \
+ /usr/include/c++/11/bits/stl_relops.h \
+ /usr/include/c++/11/bits/stl_algo.h \
+ /usr/include/c++/11/bits/algorithmfwd.h \
+ /usr/include/c++/11/bits/stl_heap.h \
+ /usr/include/c++/11/bits/stl_tempbuf.h \
+ /usr/include/c++/11/bits/uniform_int_dist.h \
+ /usr/include/c++/11/pstl/glue_algorithm_defs.h \
+ /usr/include/c++/11/functional /usr/include/c++/11/bits/stl_function.h \
+ /usr/include/c++/11/backward/binders.h /usr/include/c++/11/tuple \
+ /usr/include/c++/11/array /usr/include/c++/11/bits/uses_allocator.h \
+ /usr/include/c++/11/bits/invoke.h /usr/include/c++/11/bits/refwrap.h \
+ /usr/include/c++/11/bits/std_function.h /usr/include/c++/11/typeinfo \
+ /usr/include/c++/11/unordered_map \
+ /usr/include/c++/11/ext/aligned_buffer.h \
+ /usr/include/c++/11/bits/hashtable.h \
+ /usr/include/c++/11/bits/hashtable_policy.h \
+ /usr/include/c++/11/bits/enable_special_members.h \
+ /usr/include/c++/11/bits/node_handle.h \
+ /usr/include/c++/11/bits/unordered_map.h \
+ /usr/include/c++/11/bits/erase_if.h \
+ /usr/include/c++/11/pstl/execution_defs.h /usr/include/c++/11/memory \
+ /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+ /usr/include/c++/11/bits/align.h /usr/include/c++/11/bit \
+ /usr/include/c++/11/bits/unique_ptr.h \
+ /usr/include/c++/11/bits/shared_ptr.h /usr/include/c++/11/iosfwd \
+ /usr/include/c++/11/bits/stringfwd.h /usr/include/c++/11/bits/postypes.h \
+ /usr/include/c++/11/cwchar /usr/include/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+ /usr/include/c++/11/bits/shared_ptr_base.h \
+ /usr/include/c++/11/bits/allocated_ptr.h \
+ /usr/include/c++/11/ext/atomicity.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+ /usr/include/pthread.h /usr/include/sched.h \
+ /usr/include/x86_64-linux-gnu/bits/sched.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+ /usr/include/x86_64-linux-gnu/bits/cpu-set.h /usr/include/time.h \
+ /usr/include/x86_64-linux-gnu/bits/time.h \
+ /usr/include/x86_64-linux-gnu/bits/timex.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+ /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+ /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+ /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+ /usr/include/c++/11/ext/concurrence.h /usr/include/c++/11/exception \
+ /usr/include/c++/11/bits/exception_ptr.h \
+ /usr/include/c++/11/bits/cxxabi_init_exception.h \
+ /usr/include/c++/11/bits/nested_exception.h \
+ /usr/include/c++/11/bits/shared_ptr_atomic.h \
+ /usr/include/c++/11/bits/atomic_base.h \
+ /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+ /usr/include/c++/11/backward/auto_ptr.h \
+ /usr/include/c++/11/pstl/glue_memory_defs.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/omp.h \
+ /content/tmp/ggml/src/ggml-cpu/amx/mmq.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /usr/include/x86_64-linux-gnu/sys/syscall.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd_64.h \
+ /usr/include/x86_64-linux-gnu/bits/syscall.h /usr/include/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/posix_opt.h \
+ /usr/include/x86_64-linux-gnu/bits/environments.h \
+ /usr/include/x86_64-linux-gnu/bits/confname.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_posix.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_core.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd_ext.h \
+ /usr/include/linux/close_range.h /usr/include/c++/11/cstring
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..6510b597060026b62762fe60fe5fc34612765223
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..ae59fdc3829ebc9b795589dda7ca60253f1d0653
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o.d
@@ -0,0 +1,292 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/amx/mmq.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/amx/amx.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/amx/mmq.h \
+ /content/tmp/ggml/src/ggml-cpu/amx/common.h \
+ /usr/include/c++/11/algorithm /usr/include/c++/11/utility \
+ /usr/include/c++/11/bits/stl_relops.h \
+ /usr/include/c++/11/bits/stl_algo.h \
+ /usr/include/c++/11/bits/algorithmfwd.h \
+ /usr/include/c++/11/bits/stl_heap.h \
+ /usr/include/c++/11/bits/stl_tempbuf.h \
+ /usr/include/c++/11/bits/uniform_int_dist.h \
+ /usr/include/c++/11/pstl/glue_algorithm_defs.h \
+ /usr/include/c++/11/functional /usr/include/c++/11/bits/stl_function.h \
+ /usr/include/c++/11/backward/binders.h /usr/include/c++/11/tuple \
+ /usr/include/c++/11/array /usr/include/c++/11/bits/uses_allocator.h \
+ /usr/include/c++/11/bits/invoke.h /usr/include/c++/11/bits/refwrap.h \
+ /usr/include/c++/11/bits/std_function.h /usr/include/c++/11/typeinfo \
+ /usr/include/c++/11/unordered_map \
+ /usr/include/c++/11/ext/aligned_buffer.h \
+ /usr/include/c++/11/bits/hashtable.h \
+ /usr/include/c++/11/bits/hashtable_policy.h \
+ /usr/include/c++/11/bits/enable_special_members.h \
+ /usr/include/c++/11/bits/node_handle.h \
+ /usr/include/c++/11/bits/unordered_map.h \
+ /usr/include/c++/11/bits/erase_if.h \
+ /usr/include/c++/11/pstl/execution_defs.h /usr/include/c++/11/memory \
+ /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+ /usr/include/c++/11/bits/align.h /usr/include/c++/11/bit \
+ /usr/include/c++/11/bits/unique_ptr.h \
+ /usr/include/c++/11/bits/shared_ptr.h /usr/include/c++/11/iosfwd \
+ /usr/include/c++/11/bits/stringfwd.h /usr/include/c++/11/bits/postypes.h \
+ /usr/include/c++/11/cwchar /usr/include/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+ /usr/include/c++/11/bits/shared_ptr_base.h \
+ /usr/include/c++/11/bits/allocated_ptr.h \
+ /usr/include/c++/11/ext/atomicity.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+ /usr/include/pthread.h /usr/include/sched.h \
+ /usr/include/x86_64-linux-gnu/bits/sched.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+ /usr/include/x86_64-linux-gnu/bits/cpu-set.h /usr/include/time.h \
+ /usr/include/x86_64-linux-gnu/bits/time.h \
+ /usr/include/x86_64-linux-gnu/bits/timex.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+ /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+ /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+ /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+ /usr/include/c++/11/ext/concurrence.h /usr/include/c++/11/exception \
+ /usr/include/c++/11/bits/exception_ptr.h \
+ /usr/include/c++/11/bits/cxxabi_init_exception.h \
+ /usr/include/c++/11/bits/nested_exception.h \
+ /usr/include/c++/11/bits/shared_ptr_atomic.h \
+ /usr/include/c++/11/bits/atomic_base.h \
+ /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+ /usr/include/c++/11/backward/auto_ptr.h \
+ /usr/include/c++/11/pstl/glue_memory_defs.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/omp.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /content/tmp/ggml/src/ggml-cpu/quants.h \
+ /content/tmp/ggml/src/./ggml-common.h \
+ /content/tmp/ggml/src/./ggml-quants.h \
+ /content/tmp/ggml/src/./ggml-common.h \
+ /usr/include/x86_64-linux-gnu/sys/syscall.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd_64.h \
+ /usr/include/x86_64-linux-gnu/bits/syscall.h /usr/include/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/posix_opt.h \
+ /usr/include/x86_64-linux-gnu/bits/environments.h \
+ /usr/include/x86_64-linux-gnu/bits/confname.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_posix.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_core.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd_ext.h \
+ /usr/include/linux/close_range.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o
new file mode 100644
index 0000000000000000000000000000000000000000..cd590c10d3a7bc168e9dc0edb6bcad5d5b0d2c1e
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..0780d6bf7bacb09f2097c2c54c394de6f7174632
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o.d
@@ -0,0 +1,169 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o: \
+ /content/tmp/ggml/src/ggml-cpu/arch/x86/quants.c \
+ /usr/include/stdc-predef.h /content/tmp/ggml/src/./ggml-common.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+ /content/tmp/ggml/src/./ggml-quants.h \
+ /content/tmp/ggml/src/./ggml-common.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/math.h /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h /usr/include/string.h \
+ /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/ggml-cpu/arch/x86/../../quants.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..052bf3161c69d56305611eb92b0c088bf979aeae
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..c1ae77ad9c7f1c643bf970f14e317033cf4aeb55
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o.d
@@ -0,0 +1,221 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/arch/x86/repack.cpp \
+ /usr/include/stdc-predef.h /content/tmp/ggml/src/./ggml-common.h \
+ /usr/include/c++/11/cstdint \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h /usr/include/c++/11/cstring \
+ /usr/include/c++/11/cassert /usr/include/c++/11/cstdio \
+ /content/tmp/ggml/src/ggml-cpu/arch/x86/../../repack.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..6cd656e50f09fb243a656c5e302dfc4b3d92a44f
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..e7c4c9235da7939962077a9925c3f0cd86406533
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o.d
@@ -0,0 +1,219 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/binary-ops.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/binary-ops.h \
+ /content/tmp/ggml/src/ggml-cpu/common.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /usr/include/c++/11/utility /usr/include/c++/11/bits/stl_relops.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o
new file mode 100644
index 0000000000000000000000000000000000000000..a27d8c4dc1ea014ea7f5107226fae28108426f3f
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..62b152543a69fd44439c79aa0a16767b763d9495
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o.d
@@ -0,0 +1,249 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o: \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu.c /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/math.h /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h /usr/include/string.h \
+ /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/ggml-cpu/quants.h \
+ /content/tmp/ggml/src/./ggml-common.h \
+ /content/tmp/ggml/src/./ggml-threading.h \
+ /content/tmp/ggml/src/ggml-cpu/unary-ops.h \
+ /content/tmp/ggml/src/ggml-cpu/common.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /content/tmp/ggml/src/ggml-cpu/binary-ops.h \
+ /content/tmp/ggml/src/ggml-cpu/vec.h \
+ /content/tmp/ggml/src/ggml-cpu/ops.h /usr/include/errno.h \
+ /usr/include/x86_64-linux-gnu/bits/errno.h /usr/include/linux/errno.h \
+ /usr/include/x86_64-linux-gnu/asm/errno.h \
+ /usr/include/asm-generic/errno.h /usr/include/asm-generic/errno-base.h \
+ /usr/include/x86_64-linux-gnu/bits/types/error_t.h /usr/include/time.h \
+ /usr/include/x86_64-linux-gnu/bits/time.h \
+ /usr/include/x86_64-linux-gnu/bits/timex.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+ /usr/include/inttypes.h /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+ /usr/include/limits.h /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+ /usr/include/linux/limits.h \
+ /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+ /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/uio_lim.h /usr/include/signal.h \
+ /usr/include/x86_64-linux-gnu/bits/signum-generic.h \
+ /usr/include/x86_64-linux-gnu/bits/signum-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sig_atomic_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/siginfo_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigval_t.h \
+ /usr/include/x86_64-linux-gnu/bits/siginfo-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/siginfo-consts.h \
+ /usr/include/x86_64-linux-gnu/bits/siginfo-consts-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigval_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigevent_t.h \
+ /usr/include/x86_64-linux-gnu/bits/sigevent-consts.h \
+ /usr/include/x86_64-linux-gnu/bits/sigaction.h \
+ /usr/include/x86_64-linux-gnu/bits/sigcontext.h \
+ /usr/include/x86_64-linux-gnu/bits/types/stack_t.h \
+ /usr/include/x86_64-linux-gnu/sys/ucontext.h \
+ /usr/include/x86_64-linux-gnu/bits/sigstack.h \
+ /usr/include/x86_64-linux-gnu/bits/sigstksz.h /usr/include/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/posix_opt.h \
+ /usr/include/x86_64-linux-gnu/bits/environments.h \
+ /usr/include/x86_64-linux-gnu/bits/confname.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_posix.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_core.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd_ext.h \
+ /usr/include/linux/close_range.h \
+ /usr/include/x86_64-linux-gnu/bits/ss_flags.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_sigstack.h \
+ /usr/include/x86_64-linux-gnu/bits/sigthread.h \
+ /usr/include/x86_64-linux-gnu/bits/signal_ext.h /usr/include/syscall.h \
+ /usr/include/x86_64-linux-gnu/sys/syscall.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd.h \
+ /usr/include/x86_64-linux-gnu/asm/unistd_64.h \
+ /usr/include/x86_64-linux-gnu/bits/syscall.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/omp.h \
+ /content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.h /usr/include/pthread.h \
+ /usr/include/sched.h /usr/include/x86_64-linux-gnu/bits/sched.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+ /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+ /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdatomic.h \
+ /usr/include/x86_64-linux-gnu/sys/stat.h \
+ /usr/include/x86_64-linux-gnu/bits/stat.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_stat.h \
+ /usr/include/x86_64-linux-gnu/bits/statx.h /usr/include/linux/stat.h \
+ /usr/include/linux/types.h /usr/include/x86_64-linux-gnu/asm/types.h \
+ /usr/include/asm-generic/types.h /usr/include/asm-generic/int-ll64.h \
+ /usr/include/x86_64-linux-gnu/asm/bitsperlong.h \
+ /usr/include/asm-generic/bitsperlong.h /usr/include/linux/posix_types.h \
+ /usr/include/linux/stddef.h \
+ /usr/include/x86_64-linux-gnu/asm/posix_types.h \
+ /usr/include/x86_64-linux-gnu/asm/posix_types_64.h \
+ /usr/include/asm-generic/posix_types.h \
+ /usr/include/x86_64-linux-gnu/bits/statx-generic.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_statx_timestamp.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_statx.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..8423b2ce35a3c8e134645eb28e5d8f74b106cbaf
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..2ba99ef9fccbc91c2bfc0e4a3318c91b44227c9f
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o.d
@@ -0,0 +1,268 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/ggml-cpu/repack.h \
+ /content/tmp/ggml/src/./ggml-common.h /usr/include/c++/11/cstdint \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/amx/amx.h /usr/include/c++/11/cctype \
+ /usr/include/ctype.h /usr/include/c++/11/string \
+ /usr/include/c++/11/bits/stringfwd.h \
+ /usr/include/c++/11/bits/char_traits.h \
+ /usr/include/c++/11/bits/postypes.h /usr/include/c++/11/cwchar \
+ /usr/include/wchar.h /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+ /usr/include/c++/11/bits/localefwd.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+ /usr/include/c++/11/clocale /usr/include/locale.h \
+ /usr/include/x86_64-linux-gnu/bits/locale.h /usr/include/c++/11/iosfwd \
+ /usr/include/c++/11/bits/ostream_insert.h \
+ /usr/include/c++/11/bits/cxxabi_forced.h \
+ /usr/include/c++/11/bits/stl_function.h \
+ /usr/include/c++/11/backward/binders.h \
+ /usr/include/c++/11/bits/basic_string.h \
+ /usr/include/c++/11/ext/atomicity.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+ /usr/include/pthread.h /usr/include/sched.h \
+ /usr/include/x86_64-linux-gnu/bits/sched.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+ /usr/include/x86_64-linux-gnu/bits/cpu-set.h /usr/include/time.h \
+ /usr/include/x86_64-linux-gnu/bits/time.h \
+ /usr/include/x86_64-linux-gnu/bits/timex.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+ /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+ /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+ /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+ /usr/include/c++/11/string_view /usr/include/c++/11/bits/string_view.tcc \
+ /usr/include/c++/11/ext/string_conversions.h /usr/include/c++/11/cstdio \
+ /usr/include/c++/11/cerrno /usr/include/errno.h \
+ /usr/include/x86_64-linux-gnu/bits/errno.h /usr/include/linux/errno.h \
+ /usr/include/x86_64-linux-gnu/asm/errno.h \
+ /usr/include/asm-generic/errno.h /usr/include/asm-generic/errno-base.h \
+ /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+ /usr/include/c++/11/bits/charconv.h \
+ /usr/include/c++/11/bits/basic_string.tcc /usr/include/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/posix_opt.h \
+ /usr/include/x86_64-linux-gnu/bits/environments.h \
+ /usr/include/x86_64-linux-gnu/bits/confname.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_posix.h \
+ /usr/include/x86_64-linux-gnu/bits/getopt_core.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd.h \
+ /usr/include/x86_64-linux-gnu/bits/unistd_ext.h \
+ /usr/include/linux/close_range.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..fbe94b7fa16ffac7dcf3c2828cc5ce744998899c
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..74b26a1b069112a1c53e5e72d1954d92e3e3b5d3
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o.d
@@ -0,0 +1,2 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/hbm.cpp /usr/include/stdc-predef.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..2fa32b0a72a38f623cee2ee2931ae01f5c1df274
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o.d
@@ -0,0 +1,217 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.cpp \
+ /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/llamafile/sgemm.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-quants.h \
+ /content/tmp/ggml/src/./ggml-common.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h /usr/include/c++/11/array \
+ /usr/include/c++/11/utility /usr/include/c++/11/bits/stl_relops.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..83f99cf8f6520d53d57ec4d48bf06a68ecbae03c
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o.d
@@ -0,0 +1,244 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/ops.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/ops.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/binary-ops.h \
+ /content/tmp/ggml/src/ggml-cpu/common.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /usr/include/c++/11/utility /usr/include/c++/11/bits/stl_relops.h \
+ /content/tmp/ggml/src/ggml-cpu/unary-ops.h \
+ /content/tmp/ggml/src/ggml-cpu/vec.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+ /usr/include/c++/11/algorithm /usr/include/c++/11/bits/stl_algo.h \
+ /usr/include/c++/11/bits/algorithmfwd.h \
+ /usr/include/c++/11/bits/stl_heap.h \
+ /usr/include/c++/11/bits/stl_tempbuf.h \
+ /usr/include/c++/11/bits/uniform_int_dist.h \
+ /usr/include/c++/11/pstl/glue_algorithm_defs.h \
+ /usr/include/c++/11/functional /usr/include/c++/11/bits/stl_function.h \
+ /usr/include/c++/11/backward/binders.h /usr/include/c++/11/tuple \
+ /usr/include/c++/11/array /usr/include/c++/11/bits/uses_allocator.h \
+ /usr/include/c++/11/bits/invoke.h /usr/include/c++/11/bits/refwrap.h \
+ /usr/include/c++/11/bits/std_function.h /usr/include/c++/11/typeinfo \
+ /usr/include/c++/11/unordered_map \
+ /usr/include/c++/11/ext/aligned_buffer.h \
+ /usr/include/c++/11/bits/hashtable.h \
+ /usr/include/c++/11/bits/hashtable_policy.h \
+ /usr/include/c++/11/bits/enable_special_members.h \
+ /usr/include/c++/11/bits/node_handle.h \
+ /usr/include/c++/11/bits/unordered_map.h \
+ /usr/include/c++/11/bits/erase_if.h \
+ /usr/include/c++/11/pstl/execution_defs.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o
new file mode 100644
index 0000000000000000000000000000000000000000..f268676bd056024917f6ecd657a5175e4e249833
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..4f246f08166ffa7c2f21adb6c90b8a81436de295
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o.d
@@ -0,0 +1,168 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o: \
+ /content/tmp/ggml/src/ggml-cpu/quants.c /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/./ggml-common.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/math.h /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h /usr/include/string.h \
+ /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /content/tmp/ggml/src/./ggml-quants.h \
+ /content/tmp/ggml/src/./ggml-common.h \
+ /content/tmp/ggml/src/ggml-cpu/quants.h \
+ /content/tmp/ggml/src/ggml-cpu/arch-fallback.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..02d05e670e1379daa13c7e7d59668c9f99b46895
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o.d
@@ -0,0 +1,221 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/repack.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/./ggml-common.h /usr/include/c++/11/cstdint \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /content/tmp/ggml/src/ggml-cpu/arch-fallback.h \
+ /usr/include/c++/11/cstring /usr/include/c++/11/cassert \
+ /usr/include/c++/11/cstdio /content/tmp/ggml/src/ggml-cpu/repack.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..9336480dd2eeb4e36f0325fc41e03d662286b248
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..0ddb4d1db3ea43e19ebc3a96790d7afbcb460195
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o.d
@@ -0,0 +1,215 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/traits.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..c73694011f2080522f2fe9f4059f9cca55597dd2
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..baac110adc4a391e422d78a6d0ba85f50ec99797
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o.d
@@ -0,0 +1,219 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/unary-ops.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/unary-ops.h \
+ /content/tmp/ggml/src/ggml-cpu/common.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/ggml-cpu/traits.h \
+ /content/tmp/ggml/src/./ggml-backend-impl.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /usr/include/c++/11/utility /usr/include/c++/11/bits/stl_relops.h
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o
new file mode 100644
index 0000000000000000000000000000000000000000..5eed26c3c0faf5297cc406864b66db84b0ba4827
Binary files /dev/null and b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o differ
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o.d b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..c9a7d2046246b76b6f7f8775c4220646b2a5e9ff
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o.d
@@ -0,0 +1,216 @@
+ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o: \
+ /content/tmp/ggml/src/ggml-cpu/vec.cpp /usr/include/stdc-predef.h \
+ /content/tmp/ggml/src/ggml-cpu/vec.h /content/tmp/ggml/src/./ggml-impl.h \
+ /content/tmp/ggml/src/../include/ggml.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h /usr/include/stdint.h \
+ /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+ /usr/include/features.h /usr/include/features-time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+ /usr/include/x86_64-linux-gnu/bits/timesize.h \
+ /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+ /usr/include/x86_64-linux-gnu/bits/long-double.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+ /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+ /usr/include/x86_64-linux-gnu/bits/types.h \
+ /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+ /usr/include/x86_64-linux-gnu/bits/time64.h \
+ /usr/include/x86_64-linux-gnu/bits/wchar.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+ /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h /usr/include/stdio.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+ /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn.h \
+ /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio.h \
+ /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+ /content/tmp/ggml/src/../include/gguf.h /usr/include/assert.h \
+ /usr/include/c++/11/math.h /usr/include/c++/11/cmath \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+ /usr/include/c++/11/pstl/pstl_config.h \
+ /usr/include/c++/11/bits/cpp_type_traits.h \
+ /usr/include/c++/11/ext/type_traits.h /usr/include/math.h \
+ /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+ /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+ /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+ /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+ /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+ /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+ /usr/include/c++/11/bits/std_abs.h /usr/include/stdlib.h \
+ /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+ /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+ /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+ /usr/include/x86_64-linux-gnu/sys/types.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/timer_t.h /usr/include/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endian.h \
+ /usr/include/x86_64-linux-gnu/bits/endianness.h \
+ /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+ /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+ /usr/include/x86_64-linux-gnu/sys/select.h \
+ /usr/include/x86_64-linux-gnu/bits/select.h \
+ /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+ /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+ /usr/include/x86_64-linux-gnu/bits/select2.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+ /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+ /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+ /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+ /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h /usr/include/alloca.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+ /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+ /usr/include/c++/11/bits/specfun.h \
+ /usr/include/c++/11/bits/stl_algobase.h \
+ /usr/include/c++/11/bits/functexcept.h \
+ /usr/include/c++/11/bits/exception_defines.h \
+ /usr/include/c++/11/ext/numeric_traits.h \
+ /usr/include/c++/11/bits/stl_pair.h /usr/include/c++/11/bits/move.h \
+ /usr/include/c++/11/type_traits \
+ /usr/include/c++/11/bits/stl_iterator_base_types.h \
+ /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+ /usr/include/c++/11/bits/concept_check.h \
+ /usr/include/c++/11/debug/assertions.h \
+ /usr/include/c++/11/bits/stl_iterator.h \
+ /usr/include/c++/11/bits/ptr_traits.h /usr/include/c++/11/debug/debug.h \
+ /usr/include/c++/11/bits/predefined_ops.h /usr/include/c++/11/limits \
+ /usr/include/c++/11/tr1/gamma.tcc \
+ /usr/include/c++/11/tr1/special_function_util.h \
+ /usr/include/c++/11/tr1/bessel_function.tcc \
+ /usr/include/c++/11/tr1/beta_function.tcc \
+ /usr/include/c++/11/tr1/ell_integral.tcc \
+ /usr/include/c++/11/tr1/exp_integral.tcc \
+ /usr/include/c++/11/tr1/hypergeometric.tcc \
+ /usr/include/c++/11/tr1/legendre_function.tcc \
+ /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+ /usr/include/c++/11/tr1/poly_hermite.tcc \
+ /usr/include/c++/11/tr1/poly_laguerre.tcc \
+ /usr/include/c++/11/tr1/riemann_zeta.tcc /usr/include/c++/11/stdlib.h \
+ /usr/include/c++/11/cstdlib /usr/include/string.h /usr/include/strings.h \
+ /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+ /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/immintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/x86gprintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/ia32intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/adxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/bmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/cldemoteintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clflushoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clwbintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/clzerointrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/enqcmdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fxsrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lzcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/lwpintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/movdirintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mwaitxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pconfigintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/popcntintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pkuintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rdseedintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/rtmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/serializeintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/sgxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tbmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tsxldtrkintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/uintrintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/waitpkgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wbnoinvdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavecintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsaveoptintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xsavesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xtestintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/hresetintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/xmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/mm_malloc.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/emmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/pmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/tmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/smmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/wmmintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avxvnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512fintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512erintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512pfintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512cdintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512dqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vlbwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vldqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512ifmavlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmiintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124fmapsintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx5124vnniwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vbmi2vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vnnivlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vpopcntdqvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bitalgintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512vp2intersectvlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/shaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/fmaintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/f16cintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/gfniintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vaesintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/vpclmulqdqintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16vlintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/avx512bf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxtileintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxint8intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/amxbf16intrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/prfchwintrin.h \
+ /usr/lib/gcc/x86_64-linux-gnu/11/include/keylockerintrin.h \
+ /usr/include/c++/11/initializer_list /usr/include/c++/11/vector \
+ /usr/include/c++/11/bits/allocator.h \
+ /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+ /usr/include/c++/11/ext/new_allocator.h /usr/include/c++/11/new \
+ /usr/include/c++/11/bits/exception.h \
+ /usr/include/c++/11/bits/memoryfwd.h \
+ /usr/include/c++/11/bits/stl_construct.h \
+ /usr/include/c++/11/bits/stl_uninitialized.h \
+ /usr/include/c++/11/ext/alloc_traits.h \
+ /usr/include/c++/11/bits/alloc_traits.h \
+ /usr/include/c++/11/bits/stl_vector.h \
+ /usr/include/c++/11/bits/stl_bvector.h \
+ /usr/include/c++/11/bits/functional_hash.h \
+ /usr/include/c++/11/bits/hash_bytes.h \
+ /usr/include/c++/11/bits/range_access.h \
+ /usr/include/c++/11/bits/vector.tcc \
+ /content/tmp/ggml/src/ggml-cpu/simd-mappings.h \
+ /content/tmp/ggml/src/ggml-cpu/ggml-cpu-impl.h \
+ /content/tmp/ggml/src/../include/ggml-cpu.h \
+ /content/tmp/ggml/src/../include/ggml-backend.h \
+ /content/tmp/ggml/src/../include/ggml-alloc.h \
+ /usr/include/c++/11/cassert
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/link.txt b/ggml/src/CMakeFiles/ggml-cpu.dir/link.txt
new file mode 100644
index 0000000000000000000000000000000000000000..4a1e56247d2f41cb62d4bae2f46cae76dcbcfe94
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/link.txt
@@ -0,0 +1 @@
+/usr/bin/c++ -fPIC -O3 -DNDEBUG -shared -Wl,-soname,libggml-cpu.so -o ../../bin/libggml-cpu.so "CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.c.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/ggml-cpu.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/repack.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/hbm.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/traits.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/amx.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/amx/mmq.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/binary-ops.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/unary-ops.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/vec.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/quants.c.o" "CMakeFiles/ggml-cpu.dir/ggml-cpu/arch/x86/repack.cpp.o"  -Wl,-rpath,/content/tmp/build/bin: ../../bin/libggml-base.so /usr/lib/gcc/x86_64-linux-gnu/11/libgomp.so /usr/lib/x86_64-linux-gnu/libpthread.a
diff --git a/ggml/src/CMakeFiles/ggml-cpu.dir/progress.make b/ggml/src/CMakeFiles/ggml-cpu.dir/progress.make
new file mode 100644
index 0000000000000000000000000000000000000000..28c32d0caa363382745370923b794d71df6d8ef9
--- /dev/null
+++ b/ggml/src/CMakeFiles/ggml-cpu.dir/progress.make
@@ -0,0 +1,17 @@
+CMAKE_PROGRESS_1 = 
+CMAKE_PROGRESS_2 = 13
+CMAKE_PROGRESS_3 = 
+CMAKE_PROGRESS_4 = 14
+CMAKE_PROGRESS_5 = 
+CMAKE_PROGRESS_6 = 15
+CMAKE_PROGRESS_7 = 
+CMAKE_PROGRESS_8 = 16
+CMAKE_PROGRESS_9 = 
+CMAKE_PROGRESS_10 = 
+CMAKE_PROGRESS_11 = 17
+CMAKE_PROGRESS_12 = 
+CMAKE_PROGRESS_13 = 18
+CMAKE_PROGRESS_14 = 
+CMAKE_PROGRESS_15 = 19
+CMAKE_PROGRESS_16 = 
+
diff --git a/ggml/src/ggml-cuda/CMakeFiles/CMakeDirectoryInformation.cmake b/ggml/src/ggml-cuda/CMakeFiles/CMakeDirectoryInformation.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..f14a337d585827fc7b9f1e449d301280774174a9
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/CMakeDirectoryInformation.cmake
@@ -0,0 +1,16 @@
+# CMAKE generated file: DO NOT EDIT!
+# Generated by "Unix Makefiles" Generator, CMake Version 3.31
+
+# Relative path conversion top directories.
+set(CMAKE_RELATIVE_PATH_TOP_SOURCE "/content/tmp")
+set(CMAKE_RELATIVE_PATH_TOP_BINARY "/content/tmp/build")
+
+# Force unix paths in dependencies.
+set(CMAKE_FORCE_UNIX_PATHS 1)
+
+
+# The C and CXX include file regular expressions for this directory.
+set(CMAKE_C_INCLUDE_REGEX_SCAN "^.*$")
+set(CMAKE_C_INCLUDE_REGEX_COMPLAIN "^$")
+set(CMAKE_CXX_INCLUDE_REGEX_SCAN ${CMAKE_C_INCLUDE_REGEX_SCAN})
+set(CMAKE_CXX_INCLUDE_REGEX_COMPLAIN ${CMAKE_C_INCLUDE_REGEX_COMPLAIN})
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..9b93ec3aa6ce96ecae3711659f8257f6d19efe1d
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o : /content/tmp/ggml/src/ggml-cuda/arange.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/arange.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..d067412e5d6bb841ab694ea97d19409d3c992704
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..a7b2c445e710a8e4c690890d0ec0b54ed467a6eb
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o.d
@@ -0,0 +1,326 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o : /content/tmp/ggml/src/ggml-cuda/argmax.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /usr/include/c++/11/algorithm \
+    /usr/include/c++/11/pstl/glue_algorithm_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /usr/include/c++/11/cstdint \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/argmax.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/sum.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..6b5255017baafead017d3ebb7a10716ea1f03a62
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o : /content/tmp/ggml/src/ggml-cuda/argsort.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/argsort.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make
new file mode 100644
index 0000000000000000000000000000000000000000..6977f129ae0810b6cfccdb91570bc9a78a80d207
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make
@@ -0,0 +1,1772 @@
+# CMAKE generated file: DO NOT EDIT!
+# Generated by "Unix Makefiles" Generator, CMake Version 3.31
+
+# Delete rule output on recipe failure.
+.DELETE_ON_ERROR:
+
+#=============================================================================
+# Special targets provided by cmake.
+
+# Disable implicit rules so canonical targets will work.
+.SUFFIXES:
+
+# Disable VCS-based implicit rules.
+% : %,v
+
+# Disable VCS-based implicit rules.
+% : RCS/%
+
+# Disable VCS-based implicit rules.
+% : RCS/%,v
+
+# Disable VCS-based implicit rules.
+% : SCCS/s.%
+
+# Disable VCS-based implicit rules.
+% : s.%
+
+.SUFFIXES: .hpux_make_needs_suffix_list
+
+# Command-line flag to silence nested $(MAKE).
+$(VERBOSE)MAKESILENT = -s
+
+#Suppress display of executed commands.
+$(VERBOSE).SILENT:
+
+# A target that is always out of date.
+cmake_force:
+.PHONY : cmake_force
+
+#=============================================================================
+# Set environment variables for the build.
+
+# The shell in which to execute make rules.
+SHELL = /bin/sh
+
+# The CMake executable.
+CMAKE_COMMAND = /usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake
+
+# The command to remove a file.
+RM = /usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -E rm -f
+
+# Escaping for special characters.
+EQUALS = =
+
+# The top-level source directory on which CMake was run.
+CMAKE_SOURCE_DIR = /content/tmp
+
+# The top-level build directory on which CMake was run.
+CMAKE_BINARY_DIR = /content/tmp/build
+
+# Include any dependencies generated for this target.
+include ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/depend.make
+# Include any dependencies generated by the compiler for this target.
+include ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.make
+
+# Include the progress variables for this target.
+include ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/progress.make
+
+# Include the compile flags for this target's objects.
+include ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/codegen:
+.PHONY : ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/codegen
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o: /content/tmp/ggml/src/ggml-cuda/acc.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_1) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o -MF CMakeFiles/ggml-cuda.dir/acc.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/acc.cu -o CMakeFiles/ggml-cuda.dir/acc.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/acc.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/acc.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o: /content/tmp/ggml/src/ggml-cuda/add-id.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_2) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o -MF CMakeFiles/ggml-cuda.dir/add-id.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/add-id.cu -o CMakeFiles/ggml-cuda.dir/add-id.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/add-id.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/add-id.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o: /content/tmp/ggml/src/ggml-cuda/arange.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_3) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o -MF CMakeFiles/ggml-cuda.dir/arange.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/arange.cu -o CMakeFiles/ggml-cuda.dir/arange.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/arange.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/arange.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o: /content/tmp/ggml/src/ggml-cuda/argmax.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_4) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o -MF CMakeFiles/ggml-cuda.dir/argmax.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/argmax.cu -o CMakeFiles/ggml-cuda.dir/argmax.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/argmax.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/argmax.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o: /content/tmp/ggml/src/ggml-cuda/argsort.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_5) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o -MF CMakeFiles/ggml-cuda.dir/argsort.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/argsort.cu -o CMakeFiles/ggml-cuda.dir/argsort.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/argsort.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/argsort.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o: /content/tmp/ggml/src/ggml-cuda/binbcast.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_6) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o -MF CMakeFiles/ggml-cuda.dir/binbcast.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/binbcast.cu -o CMakeFiles/ggml-cuda.dir/binbcast.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/binbcast.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/binbcast.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o: /content/tmp/ggml/src/ggml-cuda/clamp.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_7) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o -MF CMakeFiles/ggml-cuda.dir/clamp.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/clamp.cu -o CMakeFiles/ggml-cuda.dir/clamp.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/clamp.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/clamp.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o: /content/tmp/ggml/src/ggml-cuda/concat.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_8) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o -MF CMakeFiles/ggml-cuda.dir/concat.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/concat.cu -o CMakeFiles/ggml-cuda.dir/concat.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/concat.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/concat.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o: /content/tmp/ggml/src/ggml-cuda/conv-transpose-1d.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_9) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o -MF CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/conv-transpose-1d.cu -o CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o: /content/tmp/ggml/src/ggml-cuda/conv2d-dw.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_10) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o -MF CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/conv2d-dw.cu -o CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o: /content/tmp/ggml/src/ggml-cuda/conv2d-transpose.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_11) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o -MF CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/conv2d-transpose.cu -o CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o: /content/tmp/ggml/src/ggml-cuda/convert.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_12) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o -MF CMakeFiles/ggml-cuda.dir/convert.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/convert.cu -o CMakeFiles/ggml-cuda.dir/convert.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/convert.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/convert.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o: /content/tmp/ggml/src/ggml-cuda/count-equal.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_13) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o -MF CMakeFiles/ggml-cuda.dir/count-equal.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/count-equal.cu -o CMakeFiles/ggml-cuda.dir/count-equal.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/count-equal.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/count-equal.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o: /content/tmp/ggml/src/ggml-cuda/cpy.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_14) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o -MF CMakeFiles/ggml-cuda.dir/cpy.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/cpy.cu -o CMakeFiles/ggml-cuda.dir/cpy.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/cpy.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/cpy.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o: /content/tmp/ggml/src/ggml-cuda/cross-entropy-loss.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_15) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o -MF CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/cross-entropy-loss.cu -o CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o: /content/tmp/ggml/src/ggml-cuda/diagmask.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_16) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o -MF CMakeFiles/ggml-cuda.dir/diagmask.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/diagmask.cu -o CMakeFiles/ggml-cuda.dir/diagmask.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/diagmask.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/diagmask.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/fattn-tile-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_17) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/fattn-tile-f16.cu -o CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o: /content/tmp/ggml/src/ggml-cuda/fattn-tile-f32.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_18) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o -MF CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/fattn-tile-f32.cu -o CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_19) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/fattn-wmma-f16.cu -o CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o: /content/tmp/ggml/src/ggml-cuda/fattn.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_20) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o -MF CMakeFiles/ggml-cuda.dir/fattn.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/fattn.cu -o CMakeFiles/ggml-cuda.dir/fattn.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/fattn.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/fattn.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o: /content/tmp/ggml/src/ggml-cuda/getrows.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_21) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o -MF CMakeFiles/ggml-cuda.dir/getrows.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/getrows.cu -o CMakeFiles/ggml-cuda.dir/getrows.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/getrows.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/getrows.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o: /content/tmp/ggml/src/ggml-cuda/ggml-cuda.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_22) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o -MF CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/ggml-cuda.cu -o CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o: /content/tmp/ggml/src/ggml-cuda/gla.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_23) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o -MF CMakeFiles/ggml-cuda.dir/gla.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/gla.cu -o CMakeFiles/ggml-cuda.dir/gla.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/gla.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/gla.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o: /content/tmp/ggml/src/ggml-cuda/im2col.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_24) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o -MF CMakeFiles/ggml-cuda.dir/im2col.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/im2col.cu -o CMakeFiles/ggml-cuda.dir/im2col.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/im2col.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/im2col.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o: /content/tmp/ggml/src/ggml-cuda/mean.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_25) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o -MF CMakeFiles/ggml-cuda.dir/mean.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/mean.cu -o CMakeFiles/ggml-cuda.dir/mean.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/mean.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/mean.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o: /content/tmp/ggml/src/ggml-cuda/mmf.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_26) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o -MF CMakeFiles/ggml-cuda.dir/mmf.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/mmf.cu -o CMakeFiles/ggml-cuda.dir/mmf.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/mmf.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/mmf.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o: /content/tmp/ggml/src/ggml-cuda/mmq.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_27) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o -MF CMakeFiles/ggml-cuda.dir/mmq.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/mmq.cu -o CMakeFiles/ggml-cuda.dir/mmq.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/mmq.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/mmq.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o: /content/tmp/ggml/src/ggml-cuda/mmvf.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_28) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o -MF CMakeFiles/ggml-cuda.dir/mmvf.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/mmvf.cu -o CMakeFiles/ggml-cuda.dir/mmvf.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/mmvf.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/mmvf.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o: /content/tmp/ggml/src/ggml-cuda/mmvq.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_29) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o -MF CMakeFiles/ggml-cuda.dir/mmvq.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/mmvq.cu -o CMakeFiles/ggml-cuda.dir/mmvq.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/mmvq.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/mmvq.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o: /content/tmp/ggml/src/ggml-cuda/norm.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_30) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o -MF CMakeFiles/ggml-cuda.dir/norm.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/norm.cu -o CMakeFiles/ggml-cuda.dir/norm.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/norm.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/norm.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o: /content/tmp/ggml/src/ggml-cuda/opt-step-adamw.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_31) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o -MF CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/opt-step-adamw.cu -o CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o: /content/tmp/ggml/src/ggml-cuda/opt-step-sgd.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_32) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o -MF CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/opt-step-sgd.cu -o CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o: /content/tmp/ggml/src/ggml-cuda/out-prod.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_33) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o -MF CMakeFiles/ggml-cuda.dir/out-prod.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/out-prod.cu -o CMakeFiles/ggml-cuda.dir/out-prod.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/out-prod.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/out-prod.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o: /content/tmp/ggml/src/ggml-cuda/pad.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_34) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o -MF CMakeFiles/ggml-cuda.dir/pad.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/pad.cu -o CMakeFiles/ggml-cuda.dir/pad.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/pad.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/pad.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o: /content/tmp/ggml/src/ggml-cuda/pool2d.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_35) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o -MF CMakeFiles/ggml-cuda.dir/pool2d.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/pool2d.cu -o CMakeFiles/ggml-cuda.dir/pool2d.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/pool2d.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/pool2d.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o: /content/tmp/ggml/src/ggml-cuda/quantize.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_36) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o -MF CMakeFiles/ggml-cuda.dir/quantize.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/quantize.cu -o CMakeFiles/ggml-cuda.dir/quantize.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/quantize.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/quantize.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o: /content/tmp/ggml/src/ggml-cuda/roll.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_37) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o -MF CMakeFiles/ggml-cuda.dir/roll.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/roll.cu -o CMakeFiles/ggml-cuda.dir/roll.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/roll.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/roll.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o: /content/tmp/ggml/src/ggml-cuda/rope.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_38) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o -MF CMakeFiles/ggml-cuda.dir/rope.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/rope.cu -o CMakeFiles/ggml-cuda.dir/rope.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/rope.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/rope.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o: /content/tmp/ggml/src/ggml-cuda/scale.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_39) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o -MF CMakeFiles/ggml-cuda.dir/scale.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/scale.cu -o CMakeFiles/ggml-cuda.dir/scale.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/scale.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/scale.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o: /content/tmp/ggml/src/ggml-cuda/set-rows.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_40) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o -MF CMakeFiles/ggml-cuda.dir/set-rows.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/set-rows.cu -o CMakeFiles/ggml-cuda.dir/set-rows.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/set-rows.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/set-rows.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o: /content/tmp/ggml/src/ggml-cuda/softcap.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_41) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o -MF CMakeFiles/ggml-cuda.dir/softcap.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/softcap.cu -o CMakeFiles/ggml-cuda.dir/softcap.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/softcap.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/softcap.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o: /content/tmp/ggml/src/ggml-cuda/softmax.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_42) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o -MF CMakeFiles/ggml-cuda.dir/softmax.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/softmax.cu -o CMakeFiles/ggml-cuda.dir/softmax.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/softmax.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/softmax.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o: /content/tmp/ggml/src/ggml-cuda/ssm-conv.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_43) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o -MF CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/ssm-conv.cu -o CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/ssm-conv.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/ssm-conv.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o: /content/tmp/ggml/src/ggml-cuda/ssm-scan.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_44) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o -MF CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/ssm-scan.cu -o CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/ssm-scan.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/ssm-scan.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o: /content/tmp/ggml/src/ggml-cuda/sum.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_45) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o -MF CMakeFiles/ggml-cuda.dir/sum.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/sum.cu -o CMakeFiles/ggml-cuda.dir/sum.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/sum.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/sum.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o: /content/tmp/ggml/src/ggml-cuda/sumrows.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_46) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o -MF CMakeFiles/ggml-cuda.dir/sumrows.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/sumrows.cu -o CMakeFiles/ggml-cuda.dir/sumrows.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/sumrows.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/sumrows.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o: /content/tmp/ggml/src/ggml-cuda/tsembd.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_47) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o -MF CMakeFiles/ggml-cuda.dir/tsembd.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/tsembd.cu -o CMakeFiles/ggml-cuda.dir/tsembd.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/tsembd.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/tsembd.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o: /content/tmp/ggml/src/ggml-cuda/unary.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_48) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o -MF CMakeFiles/ggml-cuda.dir/unary.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/unary.cu -o CMakeFiles/ggml-cuda.dir/unary.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/unary.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/unary.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o: /content/tmp/ggml/src/ggml-cuda/upscale.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_49) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o -MF CMakeFiles/ggml-cuda.dir/upscale.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/upscale.cu -o CMakeFiles/ggml-cuda.dir/upscale.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/upscale.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/upscale.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o: /content/tmp/ggml/src/ggml-cuda/wkv.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_50) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o -MF CMakeFiles/ggml-cuda.dir/wkv.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/wkv.cu -o CMakeFiles/ggml-cuda.dir/wkv.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/wkv.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/wkv.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_51) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_52) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_53) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_54) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_55) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_56) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_57) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_58) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_59) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_60) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_61) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_62) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_63) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_64) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_65) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_66) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_67) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_68) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_69) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_70) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_71) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_72) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_73) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_74) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_75) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_76) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_77) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_78) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_79) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_80) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_81) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_82) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_83) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_84) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_85) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_86) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_87) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_88) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu -o CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_89) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_90) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_91) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_92) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_93) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_94) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_95) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_96) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_97) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o: /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_98) "Building CUDA object ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o"
+	cd /content/tmp/build/ggml/src/ggml-cuda && /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o -MF CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o.d -x cu -c /content/tmp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu -o CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.i: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Preprocessing CUDA source to CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.i"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_PREPROCESSED_SOURCE
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.s: cmake_force
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green "Compiling CUDA source to assembly CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.s"
+	$(CMAKE_COMMAND) -E cmake_unimplemented_variable CMAKE_CUDA_CREATE_ASSEMBLY_SOURCE
+
+# Object files for target ggml-cuda
+ggml__cuda_OBJECTS = \
+"CMakeFiles/ggml-cuda.dir/acc.cu.o" \
+"CMakeFiles/ggml-cuda.dir/add-id.cu.o" \
+"CMakeFiles/ggml-cuda.dir/arange.cu.o" \
+"CMakeFiles/ggml-cuda.dir/argmax.cu.o" \
+"CMakeFiles/ggml-cuda.dir/argsort.cu.o" \
+"CMakeFiles/ggml-cuda.dir/binbcast.cu.o" \
+"CMakeFiles/ggml-cuda.dir/clamp.cu.o" \
+"CMakeFiles/ggml-cuda.dir/concat.cu.o" \
+"CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o" \
+"CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o" \
+"CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o" \
+"CMakeFiles/ggml-cuda.dir/convert.cu.o" \
+"CMakeFiles/ggml-cuda.dir/count-equal.cu.o" \
+"CMakeFiles/ggml-cuda.dir/cpy.cu.o" \
+"CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o" \
+"CMakeFiles/ggml-cuda.dir/diagmask.cu.o" \
+"CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o" \
+"CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/fattn.cu.o" \
+"CMakeFiles/ggml-cuda.dir/getrows.cu.o" \
+"CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o" \
+"CMakeFiles/ggml-cuda.dir/gla.cu.o" \
+"CMakeFiles/ggml-cuda.dir/im2col.cu.o" \
+"CMakeFiles/ggml-cuda.dir/mean.cu.o" \
+"CMakeFiles/ggml-cuda.dir/mmf.cu.o" \
+"CMakeFiles/ggml-cuda.dir/mmq.cu.o" \
+"CMakeFiles/ggml-cuda.dir/mmvf.cu.o" \
+"CMakeFiles/ggml-cuda.dir/mmvq.cu.o" \
+"CMakeFiles/ggml-cuda.dir/norm.cu.o" \
+"CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o" \
+"CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o" \
+"CMakeFiles/ggml-cuda.dir/out-prod.cu.o" \
+"CMakeFiles/ggml-cuda.dir/pad.cu.o" \
+"CMakeFiles/ggml-cuda.dir/pool2d.cu.o" \
+"CMakeFiles/ggml-cuda.dir/quantize.cu.o" \
+"CMakeFiles/ggml-cuda.dir/roll.cu.o" \
+"CMakeFiles/ggml-cuda.dir/rope.cu.o" \
+"CMakeFiles/ggml-cuda.dir/scale.cu.o" \
+"CMakeFiles/ggml-cuda.dir/set-rows.cu.o" \
+"CMakeFiles/ggml-cuda.dir/softcap.cu.o" \
+"CMakeFiles/ggml-cuda.dir/softmax.cu.o" \
+"CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o" \
+"CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o" \
+"CMakeFiles/ggml-cuda.dir/sum.cu.o" \
+"CMakeFiles/ggml-cuda.dir/sumrows.cu.o" \
+"CMakeFiles/ggml-cuda.dir/tsembd.cu.o" \
+"CMakeFiles/ggml-cuda.dir/unary.cu.o" \
+"CMakeFiles/ggml-cuda.dir/upscale.cu.o" \
+"CMakeFiles/ggml-cuda.dir/wkv.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o" \
+"CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o"
+
+# External object files for target ggml-cuda
+ggml__cuda_EXTERNAL_OBJECTS =
+
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make
+bin/libggml-cuda.so: bin/libggml-base.so
+bin/libggml-cuda.so: /usr/local/cuda-12.5/targets/x86_64-linux/lib/libcudart.so
+bin/libggml-cuda.so: /usr/local/cuda-12.5/targets/x86_64-linux/lib/libcublas.so
+bin/libggml-cuda.so: /usr/local/cuda/targets/x86_64-linux/lib/stubs/libcuda.so
+bin/libggml-cuda.so: /usr/local/cuda-12.5/targets/x86_64-linux/lib/libcublasLt.so
+bin/libggml-cuda.so: /usr/local/cuda-12.5/targets/x86_64-linux/lib/libculibos.a
+bin/libggml-cuda.so: /usr/lib/x86_64-linux-gnu/librt.a
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/linkLibs.rsp
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/objects1.rsp
+bin/libggml-cuda.so: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/link.txt
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --green --bold --progress-dir=/content/tmp/build/CMakeFiles --progress-num=$(CMAKE_PROGRESS_99) "Linking CUDA shared library ../../../bin/libggml-cuda.so"
+	cd /content/tmp/build/ggml/src/ggml-cuda && $(CMAKE_COMMAND) -E cmake_link_script CMakeFiles/ggml-cuda.dir/link.txt --verbose=$(VERBOSE)
+
+# Rule to build all files generated by this target.
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build: bin/libggml-cuda.so
+.PHONY : ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clean:
+	cd /content/tmp/build/ggml/src/ggml-cuda && $(CMAKE_COMMAND) -P CMakeFiles/ggml-cuda.dir/cmake_clean.cmake
+.PHONY : ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clean
+
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/depend:
+	cd /content/tmp/build && $(CMAKE_COMMAND) -E cmake_depends "Unix Makefiles" /content/tmp /content/tmp/ggml/src/ggml-cuda /content/tmp/build /content/tmp/build/ggml/src/ggml-cuda /content/tmp/build/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/DependInfo.cmake "--color=$(COLOR)"
+.PHONY : ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/depend
+
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..c5a93932a37301fb6c7c49fe5c0bde71cc664761
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cmake_clean.cmake b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cmake_clean.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..be9687ca5c3c8b2dbf301f2fab10e20e39fe5d9b
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cmake_clean.cmake
@@ -0,0 +1,205 @@
+file(REMOVE_RECURSE
+  "../../../bin/libggml-cuda.pdb"
+  "../../../bin/libggml-cuda.so"
+  "CMakeFiles/ggml-cuda.dir/acc.cu.o"
+  "CMakeFiles/ggml-cuda.dir/acc.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/add-id.cu.o"
+  "CMakeFiles/ggml-cuda.dir/add-id.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/arange.cu.o"
+  "CMakeFiles/ggml-cuda.dir/arange.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/argmax.cu.o"
+  "CMakeFiles/ggml-cuda.dir/argmax.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/argsort.cu.o"
+  "CMakeFiles/ggml-cuda.dir/argsort.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/binbcast.cu.o"
+  "CMakeFiles/ggml-cuda.dir/binbcast.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/clamp.cu.o"
+  "CMakeFiles/ggml-cuda.dir/clamp.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/concat.cu.o"
+  "CMakeFiles/ggml-cuda.dir/concat.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o"
+  "CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o"
+  "CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o"
+  "CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/convert.cu.o"
+  "CMakeFiles/ggml-cuda.dir/convert.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/count-equal.cu.o"
+  "CMakeFiles/ggml-cuda.dir/count-equal.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/cpy.cu.o"
+  "CMakeFiles/ggml-cuda.dir/cpy.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o"
+  "CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/diagmask.cu.o"
+  "CMakeFiles/ggml-cuda.dir/diagmask.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o"
+  "CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/fattn.cu.o"
+  "CMakeFiles/ggml-cuda.dir/fattn.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/getrows.cu.o"
+  "CMakeFiles/ggml-cuda.dir/getrows.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o"
+  "CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/gla.cu.o"
+  "CMakeFiles/ggml-cuda.dir/gla.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/im2col.cu.o"
+  "CMakeFiles/ggml-cuda.dir/im2col.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/mean.cu.o"
+  "CMakeFiles/ggml-cuda.dir/mean.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/mmf.cu.o"
+  "CMakeFiles/ggml-cuda.dir/mmf.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/mmq.cu.o"
+  "CMakeFiles/ggml-cuda.dir/mmq.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/mmvf.cu.o"
+  "CMakeFiles/ggml-cuda.dir/mmvf.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/mmvq.cu.o"
+  "CMakeFiles/ggml-cuda.dir/mmvq.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/norm.cu.o"
+  "CMakeFiles/ggml-cuda.dir/norm.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o"
+  "CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o"
+  "CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/out-prod.cu.o"
+  "CMakeFiles/ggml-cuda.dir/out-prod.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/pad.cu.o"
+  "CMakeFiles/ggml-cuda.dir/pad.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/pool2d.cu.o"
+  "CMakeFiles/ggml-cuda.dir/pool2d.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/quantize.cu.o"
+  "CMakeFiles/ggml-cuda.dir/quantize.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/roll.cu.o"
+  "CMakeFiles/ggml-cuda.dir/roll.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/rope.cu.o"
+  "CMakeFiles/ggml-cuda.dir/rope.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/scale.cu.o"
+  "CMakeFiles/ggml-cuda.dir/scale.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/set-rows.cu.o"
+  "CMakeFiles/ggml-cuda.dir/set-rows.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/softcap.cu.o"
+  "CMakeFiles/ggml-cuda.dir/softcap.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/softmax.cu.o"
+  "CMakeFiles/ggml-cuda.dir/softmax.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o"
+  "CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o"
+  "CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/sum.cu.o"
+  "CMakeFiles/ggml-cuda.dir/sum.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/sumrows.cu.o"
+  "CMakeFiles/ggml-cuda.dir/sumrows.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o"
+  "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/tsembd.cu.o"
+  "CMakeFiles/ggml-cuda.dir/tsembd.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/unary.cu.o"
+  "CMakeFiles/ggml-cuda.dir/unary.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/upscale.cu.o"
+  "CMakeFiles/ggml-cuda.dir/upscale.cu.o.d"
+  "CMakeFiles/ggml-cuda.dir/wkv.cu.o"
+  "CMakeFiles/ggml-cuda.dir/wkv.cu.o.d"
+)
+
+# Per-language clean rules from dependency scanning.
+foreach(lang CUDA)
+  include(CMakeFiles/ggml-cuda.dir/cmake_clean_${lang}.cmake OPTIONAL)
+endforeach()
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.make b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.make
new file mode 100644
index 0000000000000000000000000000000000000000..74eebd676789d0464e4e82ec2becc6a919ad9f68
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.make
@@ -0,0 +1,2 @@
+# Empty compiler generated dependencies file for ggml-cuda.
+# This may be replaced when dependencies are built.
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
new file mode 100644
index 0000000000000000000000000000000000000000..7886de7e323f3dbd529de2c013165469c8402670
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/compiler_depend.ts
@@ -0,0 +1,2 @@
+# CMAKE generated file: DO NOT EDIT!
+# Timestamp file for compiler generated dependencies management for ggml-cuda.
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..51f8566424f42b4a9af02f0ddeaa56b914b184b0
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o : /content/tmp/ggml/src/ggml-cuda/concat.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/concat.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..b2c0f81c98e48554aef6465fec5a775129a78b13
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..73fbf05d8d6a073f7ac4f410946c765a25682820
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o : /content/tmp/ggml/src/ggml-cuda/conv-transpose-1d.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/conv-transpose-1d.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..b06af4628059fd188435bee0ba890c869fef2afb
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o : /content/tmp/ggml/src/ggml-cuda/conv2d-dw.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/conv2d-dw.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..c84021354462a6843a482e64417d3465f7bfd273
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..a59abf469d71491b962a8e08fb73bf71fd9e05b0
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o.d
@@ -0,0 +1,324 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o : /content/tmp/ggml/src/ggml-cuda/convert.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/convert.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/dequantize.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..e89740f1635a89f04399cd2475f050d1ebbf5343
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..fb142fe9cb182cc56d87c4e0a94a69060146b631
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o : /content/tmp/ggml/src/ggml-cuda/count-equal.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/count-equal.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..8ff3da72cee95b01a80413115f64830896788b14
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o.d
@@ -0,0 +1,326 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o : /content/tmp/ggml/src/ggml-cuda/cpy.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/cpy.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/dequantize.cuh \
+    /content/tmp/ggml/src/ggml-cuda/cpy-utils.cuh \
+    /content/tmp/ggml/src/ggml-cuda/convert.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..dd5bfc2a14d659b41194e1dfc4da6ff0a87955f9
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o.d
@@ -0,0 +1,324 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o : /content/tmp/ggml/src/ggml-cuda/cross-entropy-loss.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/cross-entropy-loss.cuh \
+    /content/tmp/ggml/src/ggml-cuda/sum.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/depend.make b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/depend.make
new file mode 100644
index 0000000000000000000000000000000000000000..c5c74a222f33fe5642a65378949789df74f524e2
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/depend.make
@@ -0,0 +1,2 @@
+# Empty dependencies file for ggml-cuda.
+# This may be replaced when dependencies are built.
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..33a7a167a472821d3d89f45691a9a3d283a650f4
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o : /content/tmp/ggml/src/ggml-cuda/diagmask.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/diagmask.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..aa1f96838b7df21d4f3ddff7bf0c484a864cbfea
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o.d
@@ -0,0 +1,326 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o : /content/tmp/ggml/src/ggml-cuda/fattn-tile-f16.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/fattn-common.cuh \
+    /content/tmp/ggml/src/ggml-cuda/convert.cuh \
+    /content/tmp/ggml/src/ggml-cuda/vecdotq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn-tile-f16.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..add80c7c6bfe240f82ad15d7ca0846418c1bb208
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..11666cfe33376efa3c45de7537d06d0b352f9983
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o.d
@@ -0,0 +1,334 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o : /content/tmp/ggml/src/ggml-cuda/fattn.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/fattn-common.cuh \
+    /content/tmp/ggml/src/ggml-cuda/convert.cuh \
+    /content/tmp/ggml/src/ggml-cuda/vecdotq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn-mma-f16.cuh \
+    /content/tmp/ggml/src/ggml-cuda/cp-async.cuh \
+    /content/tmp/ggml/src/ggml-cuda/mma.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn-tile-f16.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn-tile-f32.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn-vec-f16.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn-vec-f32.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn-wmma-f16.cuh \
+    /content/tmp/ggml/src/ggml-cuda/fattn.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
new file mode 100644
index 0000000000000000000000000000000000000000..c5f0683a13e3076e80f8f5e5a23ccf2cb6d04090
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/flags.make
@@ -0,0 +1,10 @@
+# CMAKE generated file: DO NOT EDIT!
+# Generated by "Unix Makefiles" Generator, CMake Version 3.31
+
+# compile CUDA with /usr/local/cuda/bin/nvcc
+CUDA_DEFINES = -DGGML_BACKEND_BUILD -DGGML_BACKEND_SHARED -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_CUDA_USE_GRAPHS -DGGML_SCHED_MAX_COPIES=4 -DGGML_SHARED -D_GNU_SOURCE -D_XOPEN_SOURCE=600 -Dggml_cuda_EXPORTS
+
+CUDA_INCLUDES = --options-file CMakeFiles/ggml-cuda.dir/includes_CUDA.rsp
+
+CUDA_FLAGS = -O3 -DNDEBUG -std=c++17 "--generate-code=arch=compute_75,code=[sm_75]" -Xcompiler=-fPIC -use_fast_math -extended-lambda -Xcompiler "-Wmissing-declarations -Wmissing-noreturn -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-array-bounds -Wextra-semi -Wno-pedantic"
+
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..92efad54ce51e968bdc14ad14a4ce55c06c59b53
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o.d
@@ -0,0 +1,394 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o : /content/tmp/ggml/src/ggml-cuda/ggml-cuda.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-backend-impl.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/common.cuh \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/acc.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/add-id.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/arange.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/argmax.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/argsort.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/binbcast.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/clamp.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/concat.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/conv-transpose-1d.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/conv2d-dw.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/conv2d-transpose.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/convert.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/count-equal.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/cpy.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/cross-entropy-loss.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/diagmask.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/fattn.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/getrows.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/im2col.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/mmf.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/mmq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/vecdotq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/mma.cuh \
+    /usr/include/c++/11/climits \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/mmvf.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/mmvq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/norm.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/opt-step-adamw.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/opt-step-sgd.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/out-prod.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/pad.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/pool2d.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/quantize.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/rope.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/roll.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/scale.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/softcap.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/softmax.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/ssm-conv.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/ssm-scan.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/sum.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/sumrows.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/mean.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/tsembd.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/unary.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/upscale.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/wkv.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/gla.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/set-rows.cuh \
+    /usr/include/c++/11/algorithm \
+    /usr/include/c++/11/pstl/glue_algorithm_defs.h \
+    /usr/include/c++/11/atomic \
+    /usr/include/c++/11/charconv \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/error_constants.h \
+    /usr/include/c++/11/cinttypes \
+    /usr/include/inttypes.h \
+    /usr/include/c++/11/condition_variable \
+    /usr/include/c++/11/chrono \
+    /usr/include/c++/11/ratio \
+    /usr/include/c++/11/ctime \
+    /usr/include/c++/11/bits/parse_numbers.h \
+    /usr/include/c++/11/bits/std_mutex.h \
+    /usr/include/c++/11/system_error \
+    /usr/include/c++/11/stdexcept \
+    /usr/include/c++/11/bits/unique_lock.h \
+    /usr/include/c++/11/cstddef \
+    /usr/include/c++/11/map \
+    /usr/include/c++/11/bits/stl_tree.h \
+    /usr/include/c++/11/bits/stl_map.h \
+    /usr/include/c++/11/bits/stl_multimap.h \
+    /usr/include/c++/11/mutex
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..538953bc7088f9cc44d635058b514b3c4b697d16
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o : /content/tmp/ggml/src/ggml-cuda/gla.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/gla.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..68c00603074e00915661e9f4536357435a3cf3da
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..41cf1f462fc1e8cf7a72dc0e5b244f2f585d05dc
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o : /content/tmp/ggml/src/ggml-cuda/im2col.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/im2col.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/linkLibs.rsp b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/linkLibs.rsp
new file mode 100644
index 0000000000000000000000000000000000000000..0f340102a45cdc456a92f94bb820a57aa2349da7
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/linkLibs.rsp
@@ -0,0 +1 @@
+ -Wl,-rpath,/content/tmp/build/bin: "../../../bin/libggml-base.so" "/usr/local/cuda-12.5/targets/x86_64-linux/lib/libcudart.so" "/usr/local/cuda-12.5/targets/x86_64-linux/lib/libcublas.so" "/usr/local/cuda/targets/x86_64-linux/lib/stubs/libcuda.so" "/usr/local/cuda-12.5/targets/x86_64-linux/lib/libcublasLt.so" "/usr/local/cuda-12.5/targets/x86_64-linux/lib/libculibos.a" -ldl "/usr/lib/x86_64-linux-gnu/librt.a" -lcudadevrt -lcudart_static -lrt -lpthread -ldl 
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..925d45bc66f4b14eca84c2c1401f5cc329391a7a
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o.d
@@ -0,0 +1,1039 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o : /content/tmp/ggml/src/ggml-cuda/mean.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/mean.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/reduce_rows.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/cub.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/config.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/__cccl_config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl_config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/compiler.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/diagnostic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/compiler.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/system_header.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/dialect.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/execution_space.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/ptx_isa.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/system_header.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/version.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/visibility.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_arch.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_cpp_dialect.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_compiler.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_macro.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/detect_cuda_runtime.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_namespace.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/version.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/__config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/__pragma_push \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__pragma_push \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__undef_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/version \
+    /usr/include/c++/11/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/__pragma_pop \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__pragma_pop \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/utility \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/utility \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__assert \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__verbose_abort \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__availability \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__debug \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constant_evaluated.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstddef \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/cstddef_prelude.h \
+    /usr/include/c++/11/cstddef \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/enable_if.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_integral.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/integral_constant.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_volatile.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/hash.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/invoke.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_lvalue_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_referenceable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_same.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/apply_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_volatile.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstddef \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/conditional.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/decay.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_void.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_cvref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_function.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_extent.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/enable_if.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_base_of.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_class.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_union.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/declval.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_core_convertible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_member_function_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_member_object_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_reference_wrapper.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/nat.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/forward.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unary_function.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/hash.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_enum.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_floating_point.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_integral.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_member_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_copy_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/conjunction.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/disjunction.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_all_extents.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/negation.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_move_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_rvalue_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/underlying_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/move.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_move_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_scalar.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_arithmetic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_null_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/pair.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unwrap_ref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/get.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/copyable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/__concept_macros.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/common_reference_with.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/convertible_to.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_convertible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/same_as.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/common_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/common_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/void_t.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/copy_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_volatile.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/copy_cvref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_const_lvalue_ref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_object.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/movable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/swappable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/class_or_enum.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/extent.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_move_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/type_identity.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/exchange.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/pair.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/subrange.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/concepts.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/arithmetic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_signed.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_signed_integer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_unsigned_integer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/derived_from.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/equality_comparable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/boolean_testable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/invocable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/predicate.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/regular.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/semiregular.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/relation.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/totally_ordered.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/incrementable_traits.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_primary_template.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_valid_expansion.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_signed.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/type_list.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_move.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator_traits.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/readable_traits.h \
+    /usr/include/c++/11/iterator \
+    /usr/include/c++/11/bits/stream_iterator.h \
+    /usr/include/c++/11/bits/streambuf_iterator.h \
+    /usr/include/c++/11/streambuf \
+    /usr/include/c++/11/bits/ios_base.h \
+    /usr/include/c++/11/bits/locale_classes.h \
+    /usr/include/c++/11/bits/locale_classes.tcc \
+    /usr/include/c++/11/system_error \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/error_constants.h \
+    /usr/include/c++/11/stdexcept \
+    /usr/include/c++/11/bits/streambuf.tcc \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/pointer_traits.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/addressof.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/tuple.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_element.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_indices.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/integer_sequence.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_types.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/sfinae_helpers.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/make_tuple_types.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/apply_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_size.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_like_ext.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/structured_bindings.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_copy_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_implicitly_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_move_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_copy_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_copy_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_swappable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/piecewise_construct.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/swap.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstdint \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/cstdint_prelude.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/climits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/climits_prelude.h \
+    /usr/include/c++/11/climits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/reference_wrapper.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/weak_result_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unwrap_ref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/weak_result_type.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/get.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/sfinae_helpers.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/structured_bindings.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/as_const.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/auto_cast.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/cmp.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_unsigned.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_unsigned.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/limits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__assert \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/type_traits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/iter_swap.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/identity.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/reference_wrapper.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/aligned_storage.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/aligned_union.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/alignment_of.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/can_extract_key.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_const_ref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/dependent_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/has_unique_object_representation.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/has_virtual_destructor.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_abstract.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_aggregate.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_allocator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_bounded_array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_callable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_char_like_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_standard_layout.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivial.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_copyable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_compound.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_fundamental.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constant_evaluated.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_empty.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_final.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_literal_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_convertible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/lazy.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_pod.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_copy_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_copy_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_polymorphic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_scoped_enum.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_move_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_move_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_unbounded_array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_32_64_or_128_bit.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/maybe_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/promote.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/rank.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/result_of.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/convert_to_integral.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/convert_to_integral.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/declval.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/exchange.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/forward_like.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/forward.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/in_place.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/integer_sequence.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/move.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/pair.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/piecewise_construct.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/priority_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/rel_ops.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/swap.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/to_underlying.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/unreachable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstdlib \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/construct_at.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/access.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/voidify.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/voidify.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/limits \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/type_traits \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/concepts \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/__concept_macros.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/_One_of.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/arithmetic.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/assignable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/boolean_testable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/class_or_enum.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/common_reference_with.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/common_with.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/constructible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/convertible_to.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/copyable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/derived_from.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/destructible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/different_from.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/equality_comparable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/invocable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/movable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/predicate.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/regular.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/relation.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/same_as.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/semiregular.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/swappable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/totally_ordered.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/initializer_list \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_element.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_size.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_deprecated.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/type_traits.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/functional \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/functional \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional_base \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/integral_constant.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constructible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_convertible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_cvref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binary_negate.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/bind_back.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/perfect_forward.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/tuple \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional_base \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_like.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/utility \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/bind_front.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/bind.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binder1st.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binder2nd.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/compose.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/default_searcher.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/search.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/comp.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/advance.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/function.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__debug \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/identity.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/invoke.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/is_transparent.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/mem_fn.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/mem_fun_ref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/not_fn.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/perfect_forward.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/pointer_to_binary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/pointer_to_unary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/ranges_operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unary_negate.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/iosfwd \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/string.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/memory_resource.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/type_traits \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_debug.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_ptx.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/cpp_compatibility.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_type.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/limits \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/uninitialized_copy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/cstdint \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstdint \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_discontinuity.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_exchange.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_exchange.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_exchange_shfl.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_exchange_smem.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_histogram_atomic.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_histogram_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_radix_rank.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_scan_raking.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_raking_layout.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_operators.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_scan_shfl.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_scan_smem.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_scan_warp_scans.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/radix_rank_sort_operations.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/integer_sequence.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/version.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/config.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/simple_defines.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/compiler.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/cpp_dialect.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/cpp_compatibility.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/cstddef \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/deprecated.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/host_system.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/device_system.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/debug.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/global_workarounds.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/namespace.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/cpp11_required.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/tuple \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/tuple \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/cache_modified_input_iterator.cuh \
+    /usr/include/c++/11/iostream \
+    /usr/include/c++/11/ostream \
+    /usr/include/c++/11/ios \
+    /usr/include/c++/11/bits/basic_ios.h \
+    /usr/include/c++/11/bits/locale_facets.h \
+    /usr/include/c++/11/cwctype \
+    /usr/include/wctype.h \
+    /usr/include/x86_64-linux-gnu/bits/wctype-wchar.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/ctype_base.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/ctype_inline.h \
+    /usr/include/c++/11/bits/locale_facets.tcc \
+    /usr/include/c++/11/bits/basic_ios.tcc \
+    /usr/include/c++/11/bits/ostream.tcc \
+    /usr/include/c++/11/istream \
+    /usr/include/c++/11/bits/istream.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_facade.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/has_trivial_assign.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_facade_category.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/host_system_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cpp/detail/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/sequential/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/device_system_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/any_system_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/config.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/allocator_aware_execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/execute_with_allocator_fwd.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/execute_with_dependencies.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_deduction.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/preprocessor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/remove_cvref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/alignment.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/dependencies_aware_execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_categories.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_category_with_system_and_traversal.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_traversal_tags.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/universal_categories.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/is_iterator_category.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_category_to_traversal.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_category_to_system.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/distance_from_result.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/void_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_traits.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_math.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_reduce_raking.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_reduce_shfl.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_reduce_smem.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_reduce_raking_commutative_only.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_reduce_warp_reductions.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/choose_offset.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/iterator \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/iterator \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/access.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/advance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/back_insert_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/bounded_iter.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/concepts.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/data.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/initializer_list \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/default_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/distance.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/access.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/enable_borrowed_range.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/auto_cast.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/concepts.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/data.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/reverse_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/unwrap_iter.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_swap.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/next.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/prev.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/enable_view.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/size.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/empty.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/erase_if_container.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/front_insert_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/incrementable_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/indirectly_comparable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/projected.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/insert_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/istream_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/default_sentinel.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/iosfwd \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/istreambuf_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_move.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_swap.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/mergeable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/ranges_operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/move_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/move_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/move_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/next.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/ostream_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/ostreambuf_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/permutable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/prev.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/projected.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/readable_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/reverse_access.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/reverse_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/size.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/sortable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/permutable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/unreachable_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/wrap_iter.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/addressof.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/pointer_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_pointer.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/core/util.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/raw_pointer_cast.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/pointer_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/is_metafunction_defined.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/has_nested_type.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/util.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system_error.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/error_code.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/errno.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/error_category.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/functional.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/placeholder.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/actor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/tuple.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/value.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/composite.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/assignment_operator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/operator_adaptors.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/argument.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/raw_reference_cast.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/tuple_transform.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/tuple_meta_transform.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/result_of_adaptable_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/function_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/actor.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/logical_metafunctions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/arithmetic_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/relational_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/logical_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/bitwise_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/compound_assignment_operators.h \
+    /usr/include/c++/11/cstring \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/error_code.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/error_condition.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/system_error.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/system_error.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/error.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/error.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/device_synchronize.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_device.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_temporary_storage.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/discard_memory \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/ptx \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/barrier_cluster.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/ptx_dot_variants.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/ptx_helper_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk_commit_group.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk_tensor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk_wait_group.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_reduce_async_bulk.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_reduce_async_bulk_tensor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/fence.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/get_sreg.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/getctarank.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/mbarrier_arrive.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/mbarrier_init.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/mbarrier_wait.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/red_async.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/st_async.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/tensormap_cp_fenceproxy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/tensormap_replace.h \
+    /usr/include/c++/11/atomic \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/is_contiguous_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/core/triple_chevron_launch.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/core/alignment.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/integer_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/cstdint.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_copy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_batch_memcpy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_batch_memcpy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/single_pass_scan_operators.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/strong_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/strong_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_run_length_decode.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_search.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/temporary_storage.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/grid/grid_queue.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_memcpy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/integer_math.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_select_if.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_select_if.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_select_if.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_three_way_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_three_way_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_three_way_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_downsweep.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_onesweep.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_upsweep.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/grid/grid_even_share.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/grid/grid_mapping.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/arg_index_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_reduce_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_reduce_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/constant_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_run_length_encode.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_rle.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_rle.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_run_length_encode.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_scan_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_scan_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_scan_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_segmented_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_segmented_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_segmented_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_segmented_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_segmented_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_sub_warp_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/device_double_buffer.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/counting_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_adaptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/use_default.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_adaptor_base.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/counting_iterator.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/numeric_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/reverse_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/reverse_iterator_base.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/reverse_iterator.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_select.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_unique_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_unique_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_unique_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_spmv.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_spmv_orig.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_segment_fixup.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_spmv_orig.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/counting_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/distance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/distance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/advance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/advance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/advance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/advance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/distance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/distance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/cache_modified_output_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/discard_output_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/tex_obj_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/tex_ref_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/transform_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_allocator.cuh \
+    /usr/include/c++/11/map \
+    /usr/include/c++/11/bits/stl_tree.h \
+    /usr/include/c++/11/bits/stl_map.h \
+    /usr/include/c++/11/bits/stl_multimap.h \
+    /usr/include/c++/11/mutex \
+    /usr/include/c++/11/chrono \
+    /usr/include/c++/11/ratio \
+    /usr/include/c++/11/ctime \
+    /usr/include/c++/11/bits/parse_numbers.h \
+    /usr/include/c++/11/bits/std_mutex.h \
+    /usr/include/c++/11/bits/unique_lock.h \
+    /usr/include/c++/11/set \
+    /usr/include/c++/11/bits/stl_set.h \
+    /usr/include/c++/11/bits/stl_multiset.h
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..9c56787e64f5fefc387c7e4ff4e8c9ffc7561fdd
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..15d1f0b360c08399c3b7db68a7cce4aad476c827
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o.d
@@ -0,0 +1,324 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o : /content/tmp/ggml/src/ggml-cuda/mmvf.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/convert.cuh \
+    /content/tmp/ggml/src/ggml-cuda/mmvf.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..502d6d0f3f7b1f7ef24f281271b0a886032e6f4c
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o.d
@@ -0,0 +1,328 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o : /content/tmp/ggml/src/ggml-cuda/mmvq.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/mmvq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/quantize.cuh \
+    /content/tmp/ggml/src/ggml-cuda/mmq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/vecdotq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/mma.cuh \
+    /usr/include/c++/11/climits
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..ad91561fb2698337d9a714e52849623a50c2c316
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o : /content/tmp/ggml/src/ggml-cuda/norm.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/norm.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/objects1.rsp b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/objects1.rsp
new file mode 100644
index 0000000000000000000000000000000000000000..02306ab3cb3348429373c48d807f88ec619be313
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/objects1.rsp
@@ -0,0 +1 @@
+"CMakeFiles/ggml-cuda.dir/acc.cu.o" "CMakeFiles/ggml-cuda.dir/add-id.cu.o" "CMakeFiles/ggml-cuda.dir/arange.cu.o" "CMakeFiles/ggml-cuda.dir/argmax.cu.o" "CMakeFiles/ggml-cuda.dir/argsort.cu.o" "CMakeFiles/ggml-cuda.dir/binbcast.cu.o" "CMakeFiles/ggml-cuda.dir/clamp.cu.o" "CMakeFiles/ggml-cuda.dir/concat.cu.o" "CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o" "CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o" "CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o" "CMakeFiles/ggml-cuda.dir/convert.cu.o" "CMakeFiles/ggml-cuda.dir/count-equal.cu.o" "CMakeFiles/ggml-cuda.dir/cpy.cu.o" "CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o" "CMakeFiles/ggml-cuda.dir/diagmask.cu.o" "CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o" "CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o" "CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o" "CMakeFiles/ggml-cuda.dir/fattn.cu.o" "CMakeFiles/ggml-cuda.dir/getrows.cu.o" "CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o" "CMakeFiles/ggml-cuda.dir/gla.cu.o" "CMakeFiles/ggml-cuda.dir/im2col.cu.o" "CMakeFiles/ggml-cuda.dir/mean.cu.o" "CMakeFiles/ggml-cuda.dir/mmf.cu.o" "CMakeFiles/ggml-cuda.dir/mmq.cu.o" "CMakeFiles/ggml-cuda.dir/mmvf.cu.o" "CMakeFiles/ggml-cuda.dir/mmvq.cu.o" "CMakeFiles/ggml-cuda.dir/norm.cu.o" "CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o" "CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o" "CMakeFiles/ggml-cuda.dir/out-prod.cu.o" "CMakeFiles/ggml-cuda.dir/pad.cu.o" "CMakeFiles/ggml-cuda.dir/pool2d.cu.o" "CMakeFiles/ggml-cuda.dir/quantize.cu.o" "CMakeFiles/ggml-cuda.dir/roll.cu.o" "CMakeFiles/ggml-cuda.dir/rope.cu.o" "CMakeFiles/ggml-cuda.dir/scale.cu.o" "CMakeFiles/ggml-cuda.dir/set-rows.cu.o" "CMakeFiles/ggml-cuda.dir/softcap.cu.o" "CMakeFiles/ggml-cuda.dir/softmax.cu.o" "CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o" "CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o" "CMakeFiles/ggml-cuda.dir/sum.cu.o" "CMakeFiles/ggml-cuda.dir/sumrows.cu.o" "CMakeFiles/ggml-cuda.dir/tsembd.cu.o" "CMakeFiles/ggml-cuda.dir/unary.cu.o" "CMakeFiles/ggml-cuda.dir/upscale.cu.o" "CMakeFiles/ggml-cuda.dir/wkv.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o" "CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o"
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..8904c23d41584e3338e99a5be130aa07cc10cec6
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..e85d5a27238d67d0559b30e5572dcc5fe353c088
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o : /content/tmp/ggml/src/ggml-cuda/opt-step-adamw.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/ggml-cuda/opt-step-adamw.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..aaa238ec58441d2abd130f8345dbab53c12afa78
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..a96141739878bc12dd6952cabc1f10bef613b4a2
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..ee128fb783a62690b4fff1f6f4c2bf62d2d95eeb
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..787a2319790b72a6092dc955cc0384333ccce21d
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..6725e10d2886e08cadd0479c01bc06bf7b2d320a
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o : /content/tmp/ggml/src/ggml-cuda/pool2d.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/pool2d.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..62a75abb2e6ad3c3cc7227ffce4b083c5abe03af
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..b0f48b438d7b12fe5ef15c459419cef425eee34f
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o.d
@@ -0,0 +1,327 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o : /content/tmp/ggml/src/ggml-cuda/quantize.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/quantize.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/mmq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/vecdotq.cuh \
+    /content/tmp/ggml/src/ggml-cuda/mma.cuh \
+    /usr/include/c++/11/climits
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..6695097abec4e643372efd7d3bc5bf2a34525c20
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..083689ec121d6bb412310927be169508bacbfc3d
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o : /content/tmp/ggml/src/ggml-cuda/roll.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/roll.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..2fff0666cd4563c33a13e8a3b6a3b203bd235357
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o : /content/tmp/ggml/src/ggml-cuda/rope.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/rope.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..d007182b489a0bf0f64edcca2794f181f9383ca1
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..9929ab1863db90cafa2e30fa239e1ce05bc1750f
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o : /content/tmp/ggml/src/ggml-cuda/scale.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/scale.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..1694b597d8a087f0d552ee5b8fb8cbb6fdd10ec5
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..a64772f4c1bd78a52361f6e60aa456b47f480c11
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o : /content/tmp/ggml/src/ggml-cuda/softmax.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/softmax.cuh
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..f4c546eae0dde72ce0c62b62fb9747222596ef15
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o.d
@@ -0,0 +1,323 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o : /content/tmp/ggml/src/ggml-cuda/ssm-conv.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/ssm-conv.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o.d b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o.d
new file mode 100644
index 0000000000000000000000000000000000000000..92399294899b1cdc2af587526c85c567d3d8a498
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o.d
@@ -0,0 +1,1039 @@
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o : /content/tmp/ggml/src/ggml-cuda/sum.cu \
+    /usr/include/stdc-predef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_config.h \
+    /usr/include/features.h \
+    /usr/include/features-time64.h \
+    /usr/include/x86_64-linux-gnu/bits/wordsize.h \
+    /usr/include/x86_64-linux-gnu/bits/timesize.h \
+    /usr/include/x86_64-linux-gnu/sys/cdefs.h \
+    /usr/include/x86_64-linux-gnu/bits/long-double.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs.h \
+    /usr/include/x86_64-linux-gnu/gnu/stubs-64.h \
+    /usr/local/cuda/targets/x86_64-linux/include/builtin_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/host_defines.h \
+    /usr/include/ctype.h \
+    /usr/include/x86_64-linux-gnu/bits/types.h \
+    /usr/include/x86_64-linux-gnu/bits/typesizes.h \
+    /usr/include/x86_64-linux-gnu/bits/time64.h \
+    /usr/include/x86_64-linux-gnu/bits/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/endianness.h \
+    /usr/include/x86_64-linux-gnu/bits/types/locale_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__locale_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_types.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/limits.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/syslimits.h \
+    /usr/include/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/libc-header-start.h \
+    /usr/include/x86_64-linux-gnu/bits/posix1_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/local_lim.h \
+    /usr/include/linux/limits.h \
+    /usr/include/x86_64-linux-gnu/bits/pthread_stack_min-dynamic.h \
+    /usr/include/x86_64-linux-gnu/bits/posix2_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/xopen_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/uio_lim.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stddef.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/library_types.h \
+    /usr/local/cuda/targets/x86_64-linux/include/channel_descriptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_runtime_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_device_runtime_api.h \
+    /usr/include/c++/11/stdlib.h \
+    /usr/include/c++/11/cstdlib \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++config.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/os_defines.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/cpu_defines.h \
+    /usr/include/c++/11/pstl/pstl_config.h \
+    /usr/include/stdlib.h \
+    /usr/include/x86_64-linux-gnu/bits/waitflags.h \
+    /usr/include/x86_64-linux-gnu/bits/waitstatus.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn.h \
+    /usr/include/x86_64-linux-gnu/bits/floatn-common.h \
+    /usr/include/x86_64-linux-gnu/sys/types.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clock_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/clockid_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/time_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/timer_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-intn.h \
+    /usr/include/endian.h \
+    /usr/include/x86_64-linux-gnu/bits/byteswap.h \
+    /usr/include/x86_64-linux-gnu/bits/uintn-identity.h \
+    /usr/include/x86_64-linux-gnu/sys/select.h \
+    /usr/include/x86_64-linux-gnu/bits/select.h \
+    /usr/include/x86_64-linux-gnu/bits/types/sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__sigset_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timeval.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_timespec.h \
+    /usr/include/x86_64-linux-gnu/bits/select2.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes.h \
+    /usr/include/x86_64-linux-gnu/bits/thread-shared-types.h \
+    /usr/include/x86_64-linux-gnu/bits/pthreadtypes-arch.h \
+    /usr/include/x86_64-linux-gnu/bits/atomic_wide_counter.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_mutex.h \
+    /usr/include/x86_64-linux-gnu/bits/struct_rwlock.h \
+    /usr/include/alloca.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-bsearch.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib-float.h \
+    /usr/include/x86_64-linux-gnu/bits/stdlib.h \
+    /usr/include/c++/11/bits/std_abs.h \
+    /usr/local/cuda/targets/x86_64-linux/include/driver_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/vector_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/common_functions.h \
+    /usr/include/string.h \
+    /usr/include/strings.h \
+    /usr/include/x86_64-linux-gnu/bits/strings_fortified.h \
+    /usr/include/x86_64-linux-gnu/bits/string_fortified.h \
+    /usr/include/time.h \
+    /usr/include/x86_64-linux-gnu/bits/time.h \
+    /usr/include/x86_64-linux-gnu/bits/timex.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_tm.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_itimerspec.h \
+    /usr/include/c++/11/new \
+    /usr/include/c++/11/bits/exception.h \
+    /usr/include/stdio.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdarg.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__fpos64_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/__FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_FILE.h \
+    /usr/include/x86_64-linux-gnu/bits/types/cookie_io_functions_t.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio_lim.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio.h \
+    /usr/include/x86_64-linux-gnu/bits/stdio2.h \
+    /usr/include/assert.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.h \
+    /usr/include/c++/11/math.h \
+    /usr/include/c++/11/cmath \
+    /usr/include/c++/11/bits/cpp_type_traits.h \
+    /usr/include/c++/11/ext/type_traits.h \
+    /usr/include/math.h \
+    /usr/include/x86_64-linux-gnu/bits/math-vector.h \
+    /usr/include/x86_64-linux-gnu/bits/libm-simd-decl-stubs.h \
+    /usr/include/x86_64-linux-gnu/bits/flt-eval-method.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-logb.h \
+    /usr/include/x86_64-linux-gnu/bits/fp-fast.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-helper-functions.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls.h \
+    /usr/include/x86_64-linux-gnu/bits/mathcalls-narrow.h \
+    /usr/include/x86_64-linux-gnu/bits/iscanonical.h \
+    /usr/include/c++/11/bits/specfun.h \
+    /usr/include/c++/11/bits/stl_algobase.h \
+    /usr/include/c++/11/bits/functexcept.h \
+    /usr/include/c++/11/bits/exception_defines.h \
+    /usr/include/c++/11/ext/numeric_traits.h \
+    /usr/include/c++/11/bits/stl_pair.h \
+    /usr/include/c++/11/bits/move.h \
+    /usr/include/c++/11/type_traits \
+    /usr/include/c++/11/bits/stl_iterator_base_types.h \
+    /usr/include/c++/11/bits/stl_iterator_base_funcs.h \
+    /usr/include/c++/11/bits/concept_check.h \
+    /usr/include/c++/11/debug/assertions.h \
+    /usr/include/c++/11/bits/stl_iterator.h \
+    /usr/include/c++/11/bits/ptr_traits.h \
+    /usr/include/c++/11/debug/debug.h \
+    /usr/include/c++/11/bits/predefined_ops.h \
+    /usr/include/c++/11/limits \
+    /usr/include/c++/11/tr1/gamma.tcc \
+    /usr/include/c++/11/tr1/special_function_util.h \
+    /usr/include/c++/11/tr1/bessel_function.tcc \
+    /usr/include/c++/11/tr1/beta_function.tcc \
+    /usr/include/c++/11/tr1/ell_integral.tcc \
+    /usr/include/c++/11/tr1/exp_integral.tcc \
+    /usr/include/c++/11/tr1/hypergeometric.tcc \
+    /usr/include/c++/11/tr1/legendre_function.tcc \
+    /usr/include/c++/11/tr1/modified_bessel_func.tcc \
+    /usr/include/c++/11/tr1/poly_hermite.tcc \
+    /usr/include/c++/11/tr1/poly_laguerre.tcc \
+    /usr/include/c++/11/tr1/riemann_zeta.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/math_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/device_double_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_60_atomic_functions.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_20_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_30_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_32_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_35_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.h \
+    /usr/local/cuda/targets/x86_64-linux/include/sm_61_intrinsics.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_70_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_80_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/sm_90_rt.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/texture_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/surface_indirect_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/crt/cudacc_ext.h \
+    /usr/local/cuda/targets/x86_64-linux/include/device_launch_parameters.h \
+    /usr/include/c++/11/functional \
+    /usr/include/c++/11/bits/stl_function.h \
+    /usr/include/c++/11/backward/binders.h \
+    /usr/include/c++/11/tuple \
+    /usr/include/c++/11/utility \
+    /usr/include/c++/11/bits/stl_relops.h \
+    /usr/include/c++/11/initializer_list \
+    /usr/include/c++/11/array \
+    /usr/include/c++/11/bits/range_access.h \
+    /usr/include/c++/11/bits/uses_allocator.h \
+    /usr/include/c++/11/bits/invoke.h \
+    /usr/include/c++/11/bits/functional_hash.h \
+    /usr/include/c++/11/bits/hash_bytes.h \
+    /usr/include/c++/11/bits/refwrap.h \
+    /usr/include/c++/11/bits/std_function.h \
+    /usr/include/c++/11/typeinfo \
+    /usr/include/c++/11/unordered_map \
+    /usr/include/c++/11/bits/allocator.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++allocator.h \
+    /usr/include/c++/11/ext/new_allocator.h \
+    /usr/include/c++/11/bits/memoryfwd.h \
+    /usr/include/c++/11/ext/alloc_traits.h \
+    /usr/include/c++/11/bits/alloc_traits.h \
+    /usr/include/c++/11/bits/stl_construct.h \
+    /usr/include/c++/11/ext/aligned_buffer.h \
+    /usr/include/c++/11/bits/hashtable.h \
+    /usr/include/c++/11/bits/hashtable_policy.h \
+    /usr/include/c++/11/bits/enable_special_members.h \
+    /usr/include/c++/11/bits/node_handle.h \
+    /usr/include/c++/11/bits/unordered_map.h \
+    /usr/include/c++/11/bits/erase_if.h \
+    /usr/include/c++/11/vector \
+    /usr/include/c++/11/bits/stl_uninitialized.h \
+    /usr/include/c++/11/bits/stl_vector.h \
+    /usr/include/c++/11/bits/stl_bvector.h \
+    /usr/include/c++/11/bits/vector.tcc \
+    /usr/include/c++/11/bits/stl_algo.h \
+    /usr/include/c++/11/bits/algorithmfwd.h \
+    /usr/include/c++/11/bits/stl_heap.h \
+    /usr/include/c++/11/bits/stl_tempbuf.h \
+    /usr/include/c++/11/bits/uniform_int_dist.h \
+    /content/tmp/ggml/src/ggml-cuda/sum.cuh \
+    /content/tmp/ggml/src/ggml-cuda/common.cuh \
+    /content/tmp/ggml/src/../include/ggml.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdbool.h \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/stdint.h \
+    /usr/include/stdint.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/stdint-uintn.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-impl.h \
+    /content/tmp/ggml/src/../include/gguf.h \
+    /content/tmp/ggml/src/../include/ggml-cuda.h \
+    /content/tmp/ggml/src/../include/ggml-backend.h \
+    /content/tmp/ggml/src/../include/ggml-alloc.h \
+    /usr/include/c++/11/cstdint \
+    /usr/include/c++/11/memory \
+    /usr/include/c++/11/bits/stl_raw_storage_iter.h \
+    /usr/include/c++/11/bits/align.h \
+    /usr/include/c++/11/bit \
+    /usr/include/c++/11/bits/unique_ptr.h \
+    /usr/include/c++/11/bits/shared_ptr.h \
+    /usr/include/c++/11/iosfwd \
+    /usr/include/c++/11/bits/stringfwd.h \
+    /usr/include/c++/11/bits/postypes.h \
+    /usr/include/c++/11/cwchar \
+    /usr/include/wchar.h \
+    /usr/include/x86_64-linux-gnu/bits/types/wint_t.h \
+    /usr/include/x86_64-linux-gnu/bits/types/mbstate_t.h \
+    /usr/include/x86_64-linux-gnu/bits/wchar2.h \
+    /usr/include/c++/11/bits/shared_ptr_base.h \
+    /usr/include/c++/11/bits/allocated_ptr.h \
+    /usr/include/c++/11/ext/atomicity.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/gthr-default.h \
+    /usr/include/pthread.h \
+    /usr/include/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/sched.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct_sched_param.h \
+    /usr/include/x86_64-linux-gnu/bits/cpu-set.h \
+    /usr/include/x86_64-linux-gnu/bits/setjmp.h \
+    /usr/include/x86_64-linux-gnu/bits/types/struct___jmp_buf_tag.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/atomic_word.h \
+    /usr/include/x86_64-linux-gnu/sys/single_threaded.h \
+    /usr/include/c++/11/ext/concurrence.h \
+    /usr/include/c++/11/exception \
+    /usr/include/c++/11/bits/exception_ptr.h \
+    /usr/include/c++/11/bits/cxxabi_init_exception.h \
+    /usr/include/c++/11/bits/nested_exception.h \
+    /usr/include/c++/11/bits/shared_ptr_atomic.h \
+    /usr/include/c++/11/bits/atomic_base.h \
+    /usr/include/c++/11/bits/atomic_lockfree_defines.h \
+    /usr/include/c++/11/backward/auto_ptr.h \
+    /usr/include/c++/11/pstl/glue_memory_defs.h \
+    /usr/include/c++/11/pstl/execution_defs.h \
+    /content/tmp/ggml/src/ggml-cuda/../ggml-common.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/target \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__target_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/nv/detail/__preprocessor \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp16.hpp \
+    /usr/include/c++/11/cassert \
+    /usr/include/c++/11/cfloat \
+    /usr/lib/gcc/x86_64-linux-gnu/11/include/float.h \
+    /usr/include/c++/11/cstdio \
+    /usr/include/c++/11/string \
+    /usr/include/c++/11/bits/char_traits.h \
+    /usr/include/c++/11/bits/localefwd.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/c++locale.h \
+    /usr/include/c++/11/clocale \
+    /usr/include/locale.h \
+    /usr/include/x86_64-linux-gnu/bits/locale.h \
+    /usr/include/c++/11/cctype \
+    /usr/include/c++/11/bits/ostream_insert.h \
+    /usr/include/c++/11/bits/cxxabi_forced.h \
+    /usr/include/c++/11/bits/basic_string.h \
+    /usr/include/c++/11/string_view \
+    /usr/include/c++/11/bits/string_view.tcc \
+    /usr/include/c++/11/ext/string_conversions.h \
+    /usr/include/c++/11/cerrno \
+    /usr/include/errno.h \
+    /usr/include/x86_64-linux-gnu/bits/errno.h \
+    /usr/include/linux/errno.h \
+    /usr/include/x86_64-linux-gnu/asm/errno.h \
+    /usr/include/asm-generic/errno.h \
+    /usr/include/asm-generic/errno-base.h \
+    /usr/include/x86_64-linux-gnu/bits/types/error_t.h \
+    /usr/include/c++/11/bits/charconv.h \
+    /usr/include/c++/11/bits/basic_string.tcc \
+    /content/tmp/ggml/src/ggml-cuda/vendors/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_v2.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cublas_api.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuComplex.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_bf16.hpp \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda_fp8.hpp \
+    /content/tmp/ggml/src/ggml-cuda/sumrows.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/cub.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/config.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/__cccl_config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl_config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/compiler.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/diagnostic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/compiler.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/system_header.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/dialect.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/execution_space.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/ptx_isa.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/system_header.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/version.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cccl/visibility.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_arch.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_cpp_dialect.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_compiler.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_macro.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/detect_cuda_runtime.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_namespace.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/version.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/__config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__config \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/__pragma_push \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__pragma_push \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__undef_macros \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/version \
+    /usr/include/c++/11/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/__pragma_pop \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__pragma_pop \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/utility \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/utility \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__assert \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__verbose_abort \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__availability \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__debug \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constant_evaluated.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstddef \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/cstddef_prelude.h \
+    /usr/include/c++/11/cstddef \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/enable_if.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_integral.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/integral_constant.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_volatile.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/hash.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/invoke.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_lvalue_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_referenceable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_same.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/apply_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_volatile.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstddef \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/conditional.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/decay.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_void.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_cvref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_function.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_extent.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/enable_if.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_base_of.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_class.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_union.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/declval.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_core_convertible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_member_function_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_member_object_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_reference_wrapper.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/nat.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/forward.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unary_function.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/hash.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_enum.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_floating_point.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_integral.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_member_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_copy_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/conjunction.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/disjunction.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_all_extents.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/negation.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_move_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_rvalue_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/underlying_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/move.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_move_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_scalar.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_arithmetic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_null_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/pair.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unwrap_ref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/get.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/copyable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/__concept_macros.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/common_reference_with.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/convertible_to.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_convertible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/same_as.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/common_reference.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/common_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/void_t.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/copy_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/add_volatile.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/copy_cvref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_const_lvalue_ref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_object.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/movable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/swappable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/class_or_enum.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/extent.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_move_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/type_identity.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/exchange.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/pair.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/subrange.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/concepts.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/arithmetic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_signed.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_signed_integer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_unsigned_integer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/derived_from.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/equality_comparable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/boolean_testable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/invocable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/predicate.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/regular.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/semiregular.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/relation.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/totally_ordered.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/incrementable_traits.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_primary_template.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_valid_expansion.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_signed.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/type_list.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_move.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator_traits.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/readable_traits.h \
+    /usr/include/c++/11/iterator \
+    /usr/include/c++/11/bits/stream_iterator.h \
+    /usr/include/c++/11/bits/streambuf_iterator.h \
+    /usr/include/c++/11/streambuf \
+    /usr/include/c++/11/bits/ios_base.h \
+    /usr/include/c++/11/bits/locale_classes.h \
+    /usr/include/c++/11/bits/locale_classes.tcc \
+    /usr/include/c++/11/system_error \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/error_constants.h \
+    /usr/include/c++/11/stdexcept \
+    /usr/include/c++/11/bits/streambuf.tcc \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/pointer_traits.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/addressof.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/tuple.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_element.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_indices.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/integer_sequence.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_types.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/sfinae_helpers.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/make_tuple_types.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/apply_cv.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_size.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_like_ext.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/structured_bindings.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_copy_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_implicitly_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_move_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_copy_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_copy_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_swappable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/piecewise_construct.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/swap.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstdint \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/cstdint_prelude.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/climits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/climits_prelude.h \
+    /usr/include/c++/11/climits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/version \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/reference_wrapper.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/weak_result_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unwrap_ref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/weak_result_type.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/get.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/sfinae_helpers.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/structured_bindings.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/as_const.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/auto_cast.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/cmp.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_unsigned.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_unsigned.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/limits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__assert \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/type_traits \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/iter_swap.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/identity.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/reference_wrapper.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/aligned_storage.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/aligned_union.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/alignment_of.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/can_extract_key.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_const_ref.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/dependent_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/has_unique_object_representation.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/has_virtual_destructor.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_abstract.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_aggregate.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_allocator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_bounded_array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_callable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_char_like_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_standard_layout.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivial.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_copyable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_default_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_compound.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_fundamental.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constant_evaluated.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_empty.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_final.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_literal_type.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_nothrow_convertible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/lazy.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_pod.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_copy_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_copy_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_destructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_polymorphic.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_scoped_enum.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_move_assignable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_trivially_move_constructible.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_unbounded_array.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/make_32_64_or_128_bit.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/maybe_const.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/promote.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/rank.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_pointer.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/result_of.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/convert_to_integral.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/convert_to_integral.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/declval.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/exchange.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/forward_like.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/forward.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/in_place.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/integer_sequence.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/move.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/pair.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/piecewise_construct.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/priority_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/rel_ops.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/swap.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/to_underlying.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/unreachable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstdlib \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/construct_at.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/access.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/voidify.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/voidify.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/limits \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/type_traits \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/concepts \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/__concept_macros.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/_One_of.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/arithmetic.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/assignable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/boolean_testable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/class_or_enum.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/common_reference_with.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/common_with.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/constructible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/convertible_to.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/copyable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/derived_from.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/destructible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/different_from.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/equality_comparable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/invocable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/movable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/predicate.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/regular.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/relation.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/same_as.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/semiregular.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/swappable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__concepts/totally_ordered.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/initializer_list \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_element.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_size.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_deprecated.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/type_traits.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/functional \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/functional \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional_base \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/integral_constant.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_constructible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_convertible.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/remove_cvref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binary_negate.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/bind_back.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/perfect_forward.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/tuple \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional_base \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__tuple_dir/tuple_like.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/utility \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/bind_front.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/bind.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binder1st.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/binder2nd.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/compose.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/default_searcher.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/search.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/comp.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/advance.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/function.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__debug \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/identity.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/invoke.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/is_transparent.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/mem_fn.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/mem_fun_ref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/not_fn.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/perfect_forward.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/pointer_to_binary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/pointer_to_unary_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/ranges_operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/unary_negate.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/iosfwd \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/string.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__fwd/memory_resource.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/type_traits \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_debug.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_ptx.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/cpp_compatibility.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_type.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/limits \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/uninitialized_copy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/cstdint \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/cstdint \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_discontinuity.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_exchange.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_exchange.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_exchange_shfl.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_exchange_smem.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_histogram_atomic.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_histogram_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_radix_rank.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_scan_raking.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_raking_layout.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_operators.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_scan_shfl.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_scan_smem.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_scan_warp_scans.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/radix_rank_sort_operations.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/integer_sequence.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/version.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/config.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/simple_defines.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/compiler.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/cpp_dialect.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/cpp_compatibility.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/cstddef \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/deprecated.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/host_system.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/device_system.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/debug.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/global_workarounds.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/config/namespace.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/cpp11_required.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/tuple \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/tuple \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/cache_modified_input_iterator.cuh \
+    /usr/include/c++/11/iostream \
+    /usr/include/c++/11/ostream \
+    /usr/include/c++/11/ios \
+    /usr/include/c++/11/bits/basic_ios.h \
+    /usr/include/c++/11/bits/locale_facets.h \
+    /usr/include/c++/11/cwctype \
+    /usr/include/wctype.h \
+    /usr/include/x86_64-linux-gnu/bits/wctype-wchar.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/ctype_base.h \
+    /usr/include/x86_64-linux-gnu/c++/11/bits/ctype_inline.h \
+    /usr/include/c++/11/bits/locale_facets.tcc \
+    /usr/include/c++/11/bits/basic_ios.tcc \
+    /usr/include/c++/11/bits/ostream.tcc \
+    /usr/include/c++/11/istream \
+    /usr/include/c++/11/bits/istream.tcc \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_facade.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/has_trivial_assign.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_facade_category.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/host_system_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cpp/detail/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/sequential/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/device_system_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/any_system_tag.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/config.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/allocator_aware_execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/execute_with_allocator_fwd.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/execute_with_dependencies.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_deduction.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/preprocessor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/remove_cvref.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/alignment.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/dependencies_aware_execution_policy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_categories.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_category_with_system_and_traversal.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_traversal_tags.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/universal_categories.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/is_iterator_category.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_category_to_traversal.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_category_to_system.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/distance_from_result.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/void_t.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_traits.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_math.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_reduce_raking.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_reduce_shfl.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/specializations/warp_reduce_smem.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_reduce_raking_commutative_only.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/specializations/block_reduce_warp_reductions.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/choose_offset.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/iterator \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/iterator \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/access.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/advance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/back_insert_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/bounded_iter.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/concepts.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/data.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/initializer_list \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/default_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/distance.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/access.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/enable_borrowed_range.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__utility/auto_cast.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/concepts.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/data.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/reverse_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__algorithm/unwrap_iter.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_swap.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/next.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/prev.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/enable_view.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__ranges/size.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/empty.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/erase_if_container.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/front_insert_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/incrementable_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/indirectly_comparable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/projected.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/insert_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/istream_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/default_sentinel.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/iosfwd \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/istreambuf_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_move.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iter_swap.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/iterator_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/mergeable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__functional/ranges_operations.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/move_iterator.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/move_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/move_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/next.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/ostream_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/ostreambuf_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/permutable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/prev.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/projected.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/readable_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/reverse_access.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/reverse_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/size.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/sortable.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/permutable.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/unreachable_sentinel.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__iterator/wrap_iter.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/addressof.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__memory/pointer_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__type_traits/is_pointer.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_adjacent_difference.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/core/util.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/raw_pointer_cast.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/pointer_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/is_metafunction_defined.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/has_nested_type.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/util.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system_error.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/error_code.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/errno.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/error_category.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/functional.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/placeholder.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/actor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/tuple.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/value.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/composite.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/assignment_operator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/operator_adaptors.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/argument.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/raw_reference_cast.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/tuple_transform.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/tuple_meta_transform.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/result_of_adaptable_function.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/type_traits/function_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/actor.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/logical_metafunctions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/arithmetic_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/relational_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/logical_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/bitwise_operators.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/functional/operators/compound_assignment_operators.h \
+    /usr/include/c++/11/cstring \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/error_code.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/error_condition.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/system_error.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/system_error.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/error.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/error.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/device_synchronize.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_device.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_temporary_storage.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/discard_memory \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/ptx \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/barrier_cluster.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/ptx_dot_variants.h \
+    /usr/local/cuda-12.5/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/ptx_helper_functions.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk_commit_group.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk_tensor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_async_bulk_wait_group.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_reduce_async_bulk.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/cp_reduce_async_bulk_tensor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/fence.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/get_sreg.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/getctarank.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/mbarrier_arrive.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/mbarrier_init.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/mbarrier_wait.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/red_async.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/st_async.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/tensormap_cp_fenceproxy.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cuda/std/detail/libcxx/include/__cuda/ptx/instructions/tensormap_replace.h \
+    /usr/include/c++/11/atomic \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/type_traits/is_contiguous_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/core/triple_chevron_launch.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/cuda/detail/core/alignment.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/integer_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/cstdint.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_copy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_batch_memcpy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_batch_memcpy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/single_pass_scan_operators.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/strong_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/strong_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/block/block_run_length_decode.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/thread/thread_search.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/temporary_storage.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/grid/grid_queue.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_memcpy.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/integer_math.h \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_select_if.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_select_if.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_select_if.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_three_way_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_three_way_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_three_way_partition.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_downsweep.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_histogram.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_onesweep.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_radix_sort_upsweep.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/grid/grid_even_share.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/grid/grid_mapping.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/arg_index_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_reduce_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_reduce_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/constant_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_reduce_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_run_length_encode.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_rle.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_rle.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_run_length_encode.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_scan.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_scan_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_scan_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_scan_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_segmented_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_segmented_reduce.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_segmented_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_segmented_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_segmented_radix_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_sub_warp_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_load.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_merge_sort.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/warp/warp_store.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/detail/device_double_buffer.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/counting_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/iterator_adaptor.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/use_default.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/iterator_adaptor_base.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/counting_iterator.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/numeric_traits.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/reverse_iterator.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/reverse_iterator_base.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/iterator/detail/reverse_iterator.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_select.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_unique_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_unique_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_unique_by_key.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_spmv.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_spmv_orig.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_segment_fixup.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_spmv_orig.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/counting_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/device_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/dispatch_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/agent/agent_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/device/dispatch/tuning/tuning_for.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/distance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/distance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/advance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/detail/advance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/advance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/advance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/distance.h \
+    /usr/local/cuda/targets/x86_64-linux/include/thrust/system/detail/generic/distance.inl \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/cache_modified_output_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/discard_output_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/tex_obj_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/tex_ref_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/iterator/transform_input_iterator.cuh \
+    /usr/local/cuda/targets/x86_64-linux/include/cub/util_allocator.cuh \
+    /usr/include/c++/11/map \
+    /usr/include/c++/11/bits/stl_tree.h \
+    /usr/include/c++/11/bits/stl_map.h \
+    /usr/include/c++/11/bits/stl_multimap.h \
+    /usr/include/c++/11/mutex \
+    /usr/include/c++/11/chrono \
+    /usr/include/c++/11/ratio \
+    /usr/include/c++/11/ctime \
+    /usr/include/c++/11/bits/parse_numbers.h \
+    /usr/include/c++/11/bits/std_mutex.h \
+    /usr/include/c++/11/bits/unique_lock.h \
+    /usr/include/c++/11/set \
+    /usr/include/c++/11/bits/stl_set.h \
+    /usr/include/c++/11/bits/stl_multiset.h
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..c1331c382e4752e0c9f53133ddebe0b2c4a2b517
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o
new file mode 100644
index 0000000000000000000000000000000000000000..58ea605a985e4b98ce330f13880b404b9fca56c1
Binary files /dev/null and b/ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o differ
diff --git a/ggml/src/ggml-cuda/CMakeFiles/progress.marks b/ggml/src/ggml-cuda/CMakeFiles/progress.marks
new file mode 100644
index 0000000000000000000000000000000000000000..82cced27d7be32719d009707139bd949ad6263c9
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeFiles/progress.marks
@@ -0,0 +1 @@
+51
diff --git a/ggml/src/ggml-cuda/Makefile b/ggml/src/ggml-cuda/Makefile
new file mode 100644
index 0000000000000000000000000000000000000000..0a969c54212eba17e8f945f02a6f8d929e4cb7ed
--- /dev/null
+++ b/ggml/src/ggml-cuda/Makefile
@@ -0,0 +1,2850 @@
+# CMAKE generated file: DO NOT EDIT!
+# Generated by "Unix Makefiles" Generator, CMake Version 3.31
+
+# Default target executed when no arguments are given to make.
+default_target: all
+.PHONY : default_target
+
+# Allow only one "make -f Makefile2" at a time, but pass parallelism.
+.NOTPARALLEL:
+
+#=============================================================================
+# Special targets provided by cmake.
+
+# Disable implicit rules so canonical targets will work.
+.SUFFIXES:
+
+# Disable VCS-based implicit rules.
+% : %,v
+
+# Disable VCS-based implicit rules.
+% : RCS/%
+
+# Disable VCS-based implicit rules.
+% : RCS/%,v
+
+# Disable VCS-based implicit rules.
+% : SCCS/s.%
+
+# Disable VCS-based implicit rules.
+% : s.%
+
+.SUFFIXES: .hpux_make_needs_suffix_list
+
+# Command-line flag to silence nested $(MAKE).
+$(VERBOSE)MAKESILENT = -s
+
+#Suppress display of executed commands.
+$(VERBOSE).SILENT:
+
+# A target that is always out of date.
+cmake_force:
+.PHONY : cmake_force
+
+#=============================================================================
+# Set environment variables for the build.
+
+# The shell in which to execute make rules.
+SHELL = /bin/sh
+
+# The CMake executable.
+CMAKE_COMMAND = /usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake
+
+# The command to remove a file.
+RM = /usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -E rm -f
+
+# Escaping for special characters.
+EQUALS = =
+
+# The top-level source directory on which CMake was run.
+CMAKE_SOURCE_DIR = /content/tmp
+
+# The top-level build directory on which CMake was run.
+CMAKE_BINARY_DIR = /content/tmp/build
+
+#=============================================================================
+# Targets provided globally by CMake.
+
+# Special rule for the target edit_cache
+edit_cache:
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "No interactive CMake dialog available..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -E echo No\ interactive\ CMake\ dialog\ available.
+.PHONY : edit_cache
+
+# Special rule for the target edit_cache
+edit_cache/fast: edit_cache
+.PHONY : edit_cache/fast
+
+# Special rule for the target rebuild_cache
+rebuild_cache:
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Running CMake to regenerate build system..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake --regenerate-during-build -S$(CMAKE_SOURCE_DIR) -B$(CMAKE_BINARY_DIR)
+.PHONY : rebuild_cache
+
+# Special rule for the target rebuild_cache
+rebuild_cache/fast: rebuild_cache
+.PHONY : rebuild_cache/fast
+
+# Special rule for the target list_install_components
+list_install_components:
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Available install components are: \"Unspecified\""
+.PHONY : list_install_components
+
+# Special rule for the target list_install_components
+list_install_components/fast: list_install_components
+.PHONY : list_install_components/fast
+
+# Special rule for the target install
+install: preinstall
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Install the project..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -P cmake_install.cmake
+.PHONY : install
+
+# Special rule for the target install
+install/fast: preinstall/fast
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Install the project..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -P cmake_install.cmake
+.PHONY : install/fast
+
+# Special rule for the target install/local
+install/local: preinstall
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Installing only the local directory..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -DCMAKE_INSTALL_LOCAL_ONLY=1 -P cmake_install.cmake
+.PHONY : install/local
+
+# Special rule for the target install/local
+install/local/fast: preinstall/fast
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Installing only the local directory..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -DCMAKE_INSTALL_LOCAL_ONLY=1 -P cmake_install.cmake
+.PHONY : install/local/fast
+
+# Special rule for the target install/strip
+install/strip: preinstall
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Installing the project stripped..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -DCMAKE_INSTALL_DO_STRIP=1 -P cmake_install.cmake
+.PHONY : install/strip
+
+# Special rule for the target install/strip
+install/strip/fast: preinstall/fast
+	@$(CMAKE_COMMAND) -E cmake_echo_color "--switch=$(COLOR)" --cyan "Installing the project stripped..."
+	/usr/local/lib/python3.11/dist-packages/cmake/data/bin/cmake -DCMAKE_INSTALL_DO_STRIP=1 -P cmake_install.cmake
+.PHONY : install/strip/fast
+
+# The main all target
+all: cmake_check_build_system
+	cd /content/tmp/build && $(CMAKE_COMMAND) -E cmake_progress_start /content/tmp/build/CMakeFiles /content/tmp/build/ggml/src/ggml-cuda//CMakeFiles/progress.marks
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f CMakeFiles/Makefile2 ggml/src/ggml-cuda/all
+	$(CMAKE_COMMAND) -E cmake_progress_start /content/tmp/build/CMakeFiles 0
+.PHONY : all
+
+# The main clean target
+clean:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f CMakeFiles/Makefile2 ggml/src/ggml-cuda/clean
+.PHONY : clean
+
+# The main clean target
+clean/fast: clean
+.PHONY : clean/fast
+
+# Prepare targets for installation.
+preinstall: all
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f CMakeFiles/Makefile2 ggml/src/ggml-cuda/preinstall
+.PHONY : preinstall
+
+# Prepare targets for installation.
+preinstall/fast:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f CMakeFiles/Makefile2 ggml/src/ggml-cuda/preinstall
+.PHONY : preinstall/fast
+
+# clear depends
+depend:
+	cd /content/tmp/build && $(CMAKE_COMMAND) -S$(CMAKE_SOURCE_DIR) -B$(CMAKE_BINARY_DIR) --check-build-system CMakeFiles/Makefile.cmake 1
+.PHONY : depend
+
+# Convenience name for target.
+ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rule:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f CMakeFiles/Makefile2 ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rule
+.PHONY : ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rule
+
+# Convenience name for target.
+ggml-cuda: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rule
+.PHONY : ggml-cuda
+
+# fast build rule for target.
+ggml-cuda/fast:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build
+.PHONY : ggml-cuda/fast
+
+acc.o: acc.cu.o
+.PHONY : acc.o
+
+# target to build an object file
+acc.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.o
+.PHONY : acc.cu.o
+
+acc.i: acc.cu.i
+.PHONY : acc.i
+
+# target to preprocess a source file
+acc.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.i
+.PHONY : acc.cu.i
+
+acc.s: acc.cu.s
+.PHONY : acc.s
+
+# target to generate assembly for a file
+acc.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/acc.cu.s
+.PHONY : acc.cu.s
+
+add-id.o: add-id.cu.o
+.PHONY : add-id.o
+
+# target to build an object file
+add-id.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.o
+.PHONY : add-id.cu.o
+
+add-id.i: add-id.cu.i
+.PHONY : add-id.i
+
+# target to preprocess a source file
+add-id.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.i
+.PHONY : add-id.cu.i
+
+add-id.s: add-id.cu.s
+.PHONY : add-id.s
+
+# target to generate assembly for a file
+add-id.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/add-id.cu.s
+.PHONY : add-id.cu.s
+
+arange.o: arange.cu.o
+.PHONY : arange.o
+
+# target to build an object file
+arange.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.o
+.PHONY : arange.cu.o
+
+arange.i: arange.cu.i
+.PHONY : arange.i
+
+# target to preprocess a source file
+arange.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.i
+.PHONY : arange.cu.i
+
+arange.s: arange.cu.s
+.PHONY : arange.s
+
+# target to generate assembly for a file
+arange.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/arange.cu.s
+.PHONY : arange.cu.s
+
+argmax.o: argmax.cu.o
+.PHONY : argmax.o
+
+# target to build an object file
+argmax.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.o
+.PHONY : argmax.cu.o
+
+argmax.i: argmax.cu.i
+.PHONY : argmax.i
+
+# target to preprocess a source file
+argmax.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.i
+.PHONY : argmax.cu.i
+
+argmax.s: argmax.cu.s
+.PHONY : argmax.s
+
+# target to generate assembly for a file
+argmax.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argmax.cu.s
+.PHONY : argmax.cu.s
+
+argsort.o: argsort.cu.o
+.PHONY : argsort.o
+
+# target to build an object file
+argsort.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.o
+.PHONY : argsort.cu.o
+
+argsort.i: argsort.cu.i
+.PHONY : argsort.i
+
+# target to preprocess a source file
+argsort.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.i
+.PHONY : argsort.cu.i
+
+argsort.s: argsort.cu.s
+.PHONY : argsort.s
+
+# target to generate assembly for a file
+argsort.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/argsort.cu.s
+.PHONY : argsort.cu.s
+
+binbcast.o: binbcast.cu.o
+.PHONY : binbcast.o
+
+# target to build an object file
+binbcast.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.o
+.PHONY : binbcast.cu.o
+
+binbcast.i: binbcast.cu.i
+.PHONY : binbcast.i
+
+# target to preprocess a source file
+binbcast.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.i
+.PHONY : binbcast.cu.i
+
+binbcast.s: binbcast.cu.s
+.PHONY : binbcast.s
+
+# target to generate assembly for a file
+binbcast.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/binbcast.cu.s
+.PHONY : binbcast.cu.s
+
+clamp.o: clamp.cu.o
+.PHONY : clamp.o
+
+# target to build an object file
+clamp.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.o
+.PHONY : clamp.cu.o
+
+clamp.i: clamp.cu.i
+.PHONY : clamp.i
+
+# target to preprocess a source file
+clamp.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.i
+.PHONY : clamp.cu.i
+
+clamp.s: clamp.cu.s
+.PHONY : clamp.s
+
+# target to generate assembly for a file
+clamp.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/clamp.cu.s
+.PHONY : clamp.cu.s
+
+concat.o: concat.cu.o
+.PHONY : concat.o
+
+# target to build an object file
+concat.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.o
+.PHONY : concat.cu.o
+
+concat.i: concat.cu.i
+.PHONY : concat.i
+
+# target to preprocess a source file
+concat.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.i
+.PHONY : concat.cu.i
+
+concat.s: concat.cu.s
+.PHONY : concat.s
+
+# target to generate assembly for a file
+concat.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/concat.cu.s
+.PHONY : concat.cu.s
+
+conv-transpose-1d.o: conv-transpose-1d.cu.o
+.PHONY : conv-transpose-1d.o
+
+# target to build an object file
+conv-transpose-1d.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.o
+.PHONY : conv-transpose-1d.cu.o
+
+conv-transpose-1d.i: conv-transpose-1d.cu.i
+.PHONY : conv-transpose-1d.i
+
+# target to preprocess a source file
+conv-transpose-1d.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.i
+.PHONY : conv-transpose-1d.cu.i
+
+conv-transpose-1d.s: conv-transpose-1d.cu.s
+.PHONY : conv-transpose-1d.s
+
+# target to generate assembly for a file
+conv-transpose-1d.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv-transpose-1d.cu.s
+.PHONY : conv-transpose-1d.cu.s
+
+conv2d-dw.o: conv2d-dw.cu.o
+.PHONY : conv2d-dw.o
+
+# target to build an object file
+conv2d-dw.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.o
+.PHONY : conv2d-dw.cu.o
+
+conv2d-dw.i: conv2d-dw.cu.i
+.PHONY : conv2d-dw.i
+
+# target to preprocess a source file
+conv2d-dw.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.i
+.PHONY : conv2d-dw.cu.i
+
+conv2d-dw.s: conv2d-dw.cu.s
+.PHONY : conv2d-dw.s
+
+# target to generate assembly for a file
+conv2d-dw.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-dw.cu.s
+.PHONY : conv2d-dw.cu.s
+
+conv2d-transpose.o: conv2d-transpose.cu.o
+.PHONY : conv2d-transpose.o
+
+# target to build an object file
+conv2d-transpose.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.o
+.PHONY : conv2d-transpose.cu.o
+
+conv2d-transpose.i: conv2d-transpose.cu.i
+.PHONY : conv2d-transpose.i
+
+# target to preprocess a source file
+conv2d-transpose.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.i
+.PHONY : conv2d-transpose.cu.i
+
+conv2d-transpose.s: conv2d-transpose.cu.s
+.PHONY : conv2d-transpose.s
+
+# target to generate assembly for a file
+conv2d-transpose.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/conv2d-transpose.cu.s
+.PHONY : conv2d-transpose.cu.s
+
+convert.o: convert.cu.o
+.PHONY : convert.o
+
+# target to build an object file
+convert.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.o
+.PHONY : convert.cu.o
+
+convert.i: convert.cu.i
+.PHONY : convert.i
+
+# target to preprocess a source file
+convert.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.i
+.PHONY : convert.cu.i
+
+convert.s: convert.cu.s
+.PHONY : convert.s
+
+# target to generate assembly for a file
+convert.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/convert.cu.s
+.PHONY : convert.cu.s
+
+count-equal.o: count-equal.cu.o
+.PHONY : count-equal.o
+
+# target to build an object file
+count-equal.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.o
+.PHONY : count-equal.cu.o
+
+count-equal.i: count-equal.cu.i
+.PHONY : count-equal.i
+
+# target to preprocess a source file
+count-equal.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.i
+.PHONY : count-equal.cu.i
+
+count-equal.s: count-equal.cu.s
+.PHONY : count-equal.s
+
+# target to generate assembly for a file
+count-equal.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/count-equal.cu.s
+.PHONY : count-equal.cu.s
+
+cpy.o: cpy.cu.o
+.PHONY : cpy.o
+
+# target to build an object file
+cpy.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.o
+.PHONY : cpy.cu.o
+
+cpy.i: cpy.cu.i
+.PHONY : cpy.i
+
+# target to preprocess a source file
+cpy.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.i
+.PHONY : cpy.cu.i
+
+cpy.s: cpy.cu.s
+.PHONY : cpy.s
+
+# target to generate assembly for a file
+cpy.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cpy.cu.s
+.PHONY : cpy.cu.s
+
+cross-entropy-loss.o: cross-entropy-loss.cu.o
+.PHONY : cross-entropy-loss.o
+
+# target to build an object file
+cross-entropy-loss.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.o
+.PHONY : cross-entropy-loss.cu.o
+
+cross-entropy-loss.i: cross-entropy-loss.cu.i
+.PHONY : cross-entropy-loss.i
+
+# target to preprocess a source file
+cross-entropy-loss.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.i
+.PHONY : cross-entropy-loss.cu.i
+
+cross-entropy-loss.s: cross-entropy-loss.cu.s
+.PHONY : cross-entropy-loss.s
+
+# target to generate assembly for a file
+cross-entropy-loss.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cross-entropy-loss.cu.s
+.PHONY : cross-entropy-loss.cu.s
+
+diagmask.o: diagmask.cu.o
+.PHONY : diagmask.o
+
+# target to build an object file
+diagmask.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.o
+.PHONY : diagmask.cu.o
+
+diagmask.i: diagmask.cu.i
+.PHONY : diagmask.i
+
+# target to preprocess a source file
+diagmask.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.i
+.PHONY : diagmask.cu.i
+
+diagmask.s: diagmask.cu.s
+.PHONY : diagmask.s
+
+# target to generate assembly for a file
+diagmask.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/diagmask.cu.s
+.PHONY : diagmask.cu.s
+
+fattn-tile-f16.o: fattn-tile-f16.cu.o
+.PHONY : fattn-tile-f16.o
+
+# target to build an object file
+fattn-tile-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.o
+.PHONY : fattn-tile-f16.cu.o
+
+fattn-tile-f16.i: fattn-tile-f16.cu.i
+.PHONY : fattn-tile-f16.i
+
+# target to preprocess a source file
+fattn-tile-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.i
+.PHONY : fattn-tile-f16.cu.i
+
+fattn-tile-f16.s: fattn-tile-f16.cu.s
+.PHONY : fattn-tile-f16.s
+
+# target to generate assembly for a file
+fattn-tile-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f16.cu.s
+.PHONY : fattn-tile-f16.cu.s
+
+fattn-tile-f32.o: fattn-tile-f32.cu.o
+.PHONY : fattn-tile-f32.o
+
+# target to build an object file
+fattn-tile-f32.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.o
+.PHONY : fattn-tile-f32.cu.o
+
+fattn-tile-f32.i: fattn-tile-f32.cu.i
+.PHONY : fattn-tile-f32.i
+
+# target to preprocess a source file
+fattn-tile-f32.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.i
+.PHONY : fattn-tile-f32.cu.i
+
+fattn-tile-f32.s: fattn-tile-f32.cu.s
+.PHONY : fattn-tile-f32.s
+
+# target to generate assembly for a file
+fattn-tile-f32.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-tile-f32.cu.s
+.PHONY : fattn-tile-f32.cu.s
+
+fattn-wmma-f16.o: fattn-wmma-f16.cu.o
+.PHONY : fattn-wmma-f16.o
+
+# target to build an object file
+fattn-wmma-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.o
+.PHONY : fattn-wmma-f16.cu.o
+
+fattn-wmma-f16.i: fattn-wmma-f16.cu.i
+.PHONY : fattn-wmma-f16.i
+
+# target to preprocess a source file
+fattn-wmma-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.i
+.PHONY : fattn-wmma-f16.cu.i
+
+fattn-wmma-f16.s: fattn-wmma-f16.cu.s
+.PHONY : fattn-wmma-f16.s
+
+# target to generate assembly for a file
+fattn-wmma-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn-wmma-f16.cu.s
+.PHONY : fattn-wmma-f16.cu.s
+
+fattn.o: fattn.cu.o
+.PHONY : fattn.o
+
+# target to build an object file
+fattn.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.o
+.PHONY : fattn.cu.o
+
+fattn.i: fattn.cu.i
+.PHONY : fattn.i
+
+# target to preprocess a source file
+fattn.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.i
+.PHONY : fattn.cu.i
+
+fattn.s: fattn.cu.s
+.PHONY : fattn.s
+
+# target to generate assembly for a file
+fattn.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/fattn.cu.s
+.PHONY : fattn.cu.s
+
+getrows.o: getrows.cu.o
+.PHONY : getrows.o
+
+# target to build an object file
+getrows.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.o
+.PHONY : getrows.cu.o
+
+getrows.i: getrows.cu.i
+.PHONY : getrows.i
+
+# target to preprocess a source file
+getrows.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.i
+.PHONY : getrows.cu.i
+
+getrows.s: getrows.cu.s
+.PHONY : getrows.s
+
+# target to generate assembly for a file
+getrows.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/getrows.cu.s
+.PHONY : getrows.cu.s
+
+ggml-cuda.o: ggml-cuda.cu.o
+.PHONY : ggml-cuda.o
+
+# target to build an object file
+ggml-cuda.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.o
+.PHONY : ggml-cuda.cu.o
+
+ggml-cuda.i: ggml-cuda.cu.i
+.PHONY : ggml-cuda.i
+
+# target to preprocess a source file
+ggml-cuda.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.i
+.PHONY : ggml-cuda.cu.i
+
+ggml-cuda.s: ggml-cuda.cu.s
+.PHONY : ggml-cuda.s
+
+# target to generate assembly for a file
+ggml-cuda.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ggml-cuda.cu.s
+.PHONY : ggml-cuda.cu.s
+
+gla.o: gla.cu.o
+.PHONY : gla.o
+
+# target to build an object file
+gla.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.o
+.PHONY : gla.cu.o
+
+gla.i: gla.cu.i
+.PHONY : gla.i
+
+# target to preprocess a source file
+gla.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.i
+.PHONY : gla.cu.i
+
+gla.s: gla.cu.s
+.PHONY : gla.s
+
+# target to generate assembly for a file
+gla.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/gla.cu.s
+.PHONY : gla.cu.s
+
+im2col.o: im2col.cu.o
+.PHONY : im2col.o
+
+# target to build an object file
+im2col.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.o
+.PHONY : im2col.cu.o
+
+im2col.i: im2col.cu.i
+.PHONY : im2col.i
+
+# target to preprocess a source file
+im2col.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.i
+.PHONY : im2col.cu.i
+
+im2col.s: im2col.cu.s
+.PHONY : im2col.s
+
+# target to generate assembly for a file
+im2col.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/im2col.cu.s
+.PHONY : im2col.cu.s
+
+mean.o: mean.cu.o
+.PHONY : mean.o
+
+# target to build an object file
+mean.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.o
+.PHONY : mean.cu.o
+
+mean.i: mean.cu.i
+.PHONY : mean.i
+
+# target to preprocess a source file
+mean.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.i
+.PHONY : mean.cu.i
+
+mean.s: mean.cu.s
+.PHONY : mean.s
+
+# target to generate assembly for a file
+mean.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mean.cu.s
+.PHONY : mean.cu.s
+
+mmf.o: mmf.cu.o
+.PHONY : mmf.o
+
+# target to build an object file
+mmf.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.o
+.PHONY : mmf.cu.o
+
+mmf.i: mmf.cu.i
+.PHONY : mmf.i
+
+# target to preprocess a source file
+mmf.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.i
+.PHONY : mmf.cu.i
+
+mmf.s: mmf.cu.s
+.PHONY : mmf.s
+
+# target to generate assembly for a file
+mmf.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmf.cu.s
+.PHONY : mmf.cu.s
+
+mmq.o: mmq.cu.o
+.PHONY : mmq.o
+
+# target to build an object file
+mmq.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.o
+.PHONY : mmq.cu.o
+
+mmq.i: mmq.cu.i
+.PHONY : mmq.i
+
+# target to preprocess a source file
+mmq.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.i
+.PHONY : mmq.cu.i
+
+mmq.s: mmq.cu.s
+.PHONY : mmq.s
+
+# target to generate assembly for a file
+mmq.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmq.cu.s
+.PHONY : mmq.cu.s
+
+mmvf.o: mmvf.cu.o
+.PHONY : mmvf.o
+
+# target to build an object file
+mmvf.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.o
+.PHONY : mmvf.cu.o
+
+mmvf.i: mmvf.cu.i
+.PHONY : mmvf.i
+
+# target to preprocess a source file
+mmvf.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.i
+.PHONY : mmvf.cu.i
+
+mmvf.s: mmvf.cu.s
+.PHONY : mmvf.s
+
+# target to generate assembly for a file
+mmvf.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvf.cu.s
+.PHONY : mmvf.cu.s
+
+mmvq.o: mmvq.cu.o
+.PHONY : mmvq.o
+
+# target to build an object file
+mmvq.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.o
+.PHONY : mmvq.cu.o
+
+mmvq.i: mmvq.cu.i
+.PHONY : mmvq.i
+
+# target to preprocess a source file
+mmvq.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.i
+.PHONY : mmvq.cu.i
+
+mmvq.s: mmvq.cu.s
+.PHONY : mmvq.s
+
+# target to generate assembly for a file
+mmvq.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/mmvq.cu.s
+.PHONY : mmvq.cu.s
+
+norm.o: norm.cu.o
+.PHONY : norm.o
+
+# target to build an object file
+norm.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.o
+.PHONY : norm.cu.o
+
+norm.i: norm.cu.i
+.PHONY : norm.i
+
+# target to preprocess a source file
+norm.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.i
+.PHONY : norm.cu.i
+
+norm.s: norm.cu.s
+.PHONY : norm.s
+
+# target to generate assembly for a file
+norm.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/norm.cu.s
+.PHONY : norm.cu.s
+
+opt-step-adamw.o: opt-step-adamw.cu.o
+.PHONY : opt-step-adamw.o
+
+# target to build an object file
+opt-step-adamw.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.o
+.PHONY : opt-step-adamw.cu.o
+
+opt-step-adamw.i: opt-step-adamw.cu.i
+.PHONY : opt-step-adamw.i
+
+# target to preprocess a source file
+opt-step-adamw.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.i
+.PHONY : opt-step-adamw.cu.i
+
+opt-step-adamw.s: opt-step-adamw.cu.s
+.PHONY : opt-step-adamw.s
+
+# target to generate assembly for a file
+opt-step-adamw.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-adamw.cu.s
+.PHONY : opt-step-adamw.cu.s
+
+opt-step-sgd.o: opt-step-sgd.cu.o
+.PHONY : opt-step-sgd.o
+
+# target to build an object file
+opt-step-sgd.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.o
+.PHONY : opt-step-sgd.cu.o
+
+opt-step-sgd.i: opt-step-sgd.cu.i
+.PHONY : opt-step-sgd.i
+
+# target to preprocess a source file
+opt-step-sgd.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.i
+.PHONY : opt-step-sgd.cu.i
+
+opt-step-sgd.s: opt-step-sgd.cu.s
+.PHONY : opt-step-sgd.s
+
+# target to generate assembly for a file
+opt-step-sgd.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/opt-step-sgd.cu.s
+.PHONY : opt-step-sgd.cu.s
+
+out-prod.o: out-prod.cu.o
+.PHONY : out-prod.o
+
+# target to build an object file
+out-prod.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.o
+.PHONY : out-prod.cu.o
+
+out-prod.i: out-prod.cu.i
+.PHONY : out-prod.i
+
+# target to preprocess a source file
+out-prod.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.i
+.PHONY : out-prod.cu.i
+
+out-prod.s: out-prod.cu.s
+.PHONY : out-prod.s
+
+# target to generate assembly for a file
+out-prod.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/out-prod.cu.s
+.PHONY : out-prod.cu.s
+
+pad.o: pad.cu.o
+.PHONY : pad.o
+
+# target to build an object file
+pad.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.o
+.PHONY : pad.cu.o
+
+pad.i: pad.cu.i
+.PHONY : pad.i
+
+# target to preprocess a source file
+pad.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.i
+.PHONY : pad.cu.i
+
+pad.s: pad.cu.s
+.PHONY : pad.s
+
+# target to generate assembly for a file
+pad.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pad.cu.s
+.PHONY : pad.cu.s
+
+pool2d.o: pool2d.cu.o
+.PHONY : pool2d.o
+
+# target to build an object file
+pool2d.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.o
+.PHONY : pool2d.cu.o
+
+pool2d.i: pool2d.cu.i
+.PHONY : pool2d.i
+
+# target to preprocess a source file
+pool2d.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.i
+.PHONY : pool2d.cu.i
+
+pool2d.s: pool2d.cu.s
+.PHONY : pool2d.s
+
+# target to generate assembly for a file
+pool2d.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/pool2d.cu.s
+.PHONY : pool2d.cu.s
+
+quantize.o: quantize.cu.o
+.PHONY : quantize.o
+
+# target to build an object file
+quantize.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.o
+.PHONY : quantize.cu.o
+
+quantize.i: quantize.cu.i
+.PHONY : quantize.i
+
+# target to preprocess a source file
+quantize.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.i
+.PHONY : quantize.cu.i
+
+quantize.s: quantize.cu.s
+.PHONY : quantize.s
+
+# target to generate assembly for a file
+quantize.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/quantize.cu.s
+.PHONY : quantize.cu.s
+
+roll.o: roll.cu.o
+.PHONY : roll.o
+
+# target to build an object file
+roll.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.o
+.PHONY : roll.cu.o
+
+roll.i: roll.cu.i
+.PHONY : roll.i
+
+# target to preprocess a source file
+roll.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.i
+.PHONY : roll.cu.i
+
+roll.s: roll.cu.s
+.PHONY : roll.s
+
+# target to generate assembly for a file
+roll.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/roll.cu.s
+.PHONY : roll.cu.s
+
+rope.o: rope.cu.o
+.PHONY : rope.o
+
+# target to build an object file
+rope.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.o
+.PHONY : rope.cu.o
+
+rope.i: rope.cu.i
+.PHONY : rope.i
+
+# target to preprocess a source file
+rope.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.i
+.PHONY : rope.cu.i
+
+rope.s: rope.cu.s
+.PHONY : rope.s
+
+# target to generate assembly for a file
+rope.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/rope.cu.s
+.PHONY : rope.cu.s
+
+scale.o: scale.cu.o
+.PHONY : scale.o
+
+# target to build an object file
+scale.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.o
+.PHONY : scale.cu.o
+
+scale.i: scale.cu.i
+.PHONY : scale.i
+
+# target to preprocess a source file
+scale.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.i
+.PHONY : scale.cu.i
+
+scale.s: scale.cu.s
+.PHONY : scale.s
+
+# target to generate assembly for a file
+scale.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/scale.cu.s
+.PHONY : scale.cu.s
+
+set-rows.o: set-rows.cu.o
+.PHONY : set-rows.o
+
+# target to build an object file
+set-rows.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.o
+.PHONY : set-rows.cu.o
+
+set-rows.i: set-rows.cu.i
+.PHONY : set-rows.i
+
+# target to preprocess a source file
+set-rows.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.i
+.PHONY : set-rows.cu.i
+
+set-rows.s: set-rows.cu.s
+.PHONY : set-rows.s
+
+# target to generate assembly for a file
+set-rows.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/set-rows.cu.s
+.PHONY : set-rows.cu.s
+
+softcap.o: softcap.cu.o
+.PHONY : softcap.o
+
+# target to build an object file
+softcap.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.o
+.PHONY : softcap.cu.o
+
+softcap.i: softcap.cu.i
+.PHONY : softcap.i
+
+# target to preprocess a source file
+softcap.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.i
+.PHONY : softcap.cu.i
+
+softcap.s: softcap.cu.s
+.PHONY : softcap.s
+
+# target to generate assembly for a file
+softcap.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softcap.cu.s
+.PHONY : softcap.cu.s
+
+softmax.o: softmax.cu.o
+.PHONY : softmax.o
+
+# target to build an object file
+softmax.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.o
+.PHONY : softmax.cu.o
+
+softmax.i: softmax.cu.i
+.PHONY : softmax.i
+
+# target to preprocess a source file
+softmax.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.i
+.PHONY : softmax.cu.i
+
+softmax.s: softmax.cu.s
+.PHONY : softmax.s
+
+# target to generate assembly for a file
+softmax.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/softmax.cu.s
+.PHONY : softmax.cu.s
+
+ssm-conv.o: ssm-conv.cu.o
+.PHONY : ssm-conv.o
+
+# target to build an object file
+ssm-conv.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.o
+.PHONY : ssm-conv.cu.o
+
+ssm-conv.i: ssm-conv.cu.i
+.PHONY : ssm-conv.i
+
+# target to preprocess a source file
+ssm-conv.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.i
+.PHONY : ssm-conv.cu.i
+
+ssm-conv.s: ssm-conv.cu.s
+.PHONY : ssm-conv.s
+
+# target to generate assembly for a file
+ssm-conv.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-conv.cu.s
+.PHONY : ssm-conv.cu.s
+
+ssm-scan.o: ssm-scan.cu.o
+.PHONY : ssm-scan.o
+
+# target to build an object file
+ssm-scan.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.o
+.PHONY : ssm-scan.cu.o
+
+ssm-scan.i: ssm-scan.cu.i
+.PHONY : ssm-scan.i
+
+# target to preprocess a source file
+ssm-scan.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.i
+.PHONY : ssm-scan.cu.i
+
+ssm-scan.s: ssm-scan.cu.s
+.PHONY : ssm-scan.s
+
+# target to generate assembly for a file
+ssm-scan.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/ssm-scan.cu.s
+.PHONY : ssm-scan.cu.s
+
+sum.o: sum.cu.o
+.PHONY : sum.o
+
+# target to build an object file
+sum.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.o
+.PHONY : sum.cu.o
+
+sum.i: sum.cu.i
+.PHONY : sum.i
+
+# target to preprocess a source file
+sum.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.i
+.PHONY : sum.cu.i
+
+sum.s: sum.cu.s
+.PHONY : sum.s
+
+# target to generate assembly for a file
+sum.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sum.cu.s
+.PHONY : sum.cu.s
+
+sumrows.o: sumrows.cu.o
+.PHONY : sumrows.o
+
+# target to build an object file
+sumrows.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.o
+.PHONY : sumrows.cu.o
+
+sumrows.i: sumrows.cu.i
+.PHONY : sumrows.i
+
+# target to preprocess a source file
+sumrows.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.i
+.PHONY : sumrows.cu.i
+
+sumrows.s: sumrows.cu.s
+.PHONY : sumrows.s
+
+# target to generate assembly for a file
+sumrows.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/sumrows.cu.s
+.PHONY : sumrows.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.o: template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.i: template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.s: template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.o: template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.i: template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.s: template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.o: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.i: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.s: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.o: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.i: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.s: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.o: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.i: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.s: template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.o: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.i: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.s: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.o: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.i: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.s: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.o: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.i: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.s: template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.o: template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.i: template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.s: template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.o: template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.i: template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.s: template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.o: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.i: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.s: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.o: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.i: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.s: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.o: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.i: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.s: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.o: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.i: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.s: template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.o: template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.i: template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.s: template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.o: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.i: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.s: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.o: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.i: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.s: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.o: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.i: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.s: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu.s
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.o: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.o
+
+# target to build an object file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.o
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.i: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.i
+
+# target to preprocess a source file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.i
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.i
+
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.s: template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.s
+
+# target to generate assembly for a file
+template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.s
+.PHONY : template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu.s
+
+template-instances/fattn-vec-f16-instance-hs128-f16-f16.o: template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-f16-f16.o
+
+# target to build an object file
+template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.o
+
+template-instances/fattn-vec-f16-instance-hs128-f16-f16.i: template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-f16-f16.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.i
+
+template-instances/fattn-vec-f16-instance-hs128-f16-f16.s: template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-f16-f16.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu.s
+
+template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.o: template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.o
+
+# target to build an object file
+template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.o
+
+template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.i: template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.i
+
+template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.s: template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu.s
+
+template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.o: template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.o
+
+# target to build an object file
+template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.o
+
+template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.i: template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.i
+
+template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.s: template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu.s
+
+template-instances/fattn-vec-f16-instance-hs256-f16-f16.o: template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs256-f16-f16.o
+
+# target to build an object file
+template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.o
+
+template-instances/fattn-vec-f16-instance-hs256-f16-f16.i: template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs256-f16-f16.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.i
+
+template-instances/fattn-vec-f16-instance-hs256-f16-f16.s: template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs256-f16-f16.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu.s
+
+template-instances/fattn-vec-f16-instance-hs64-f16-f16.o: template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs64-f16-f16.o
+
+# target to build an object file
+template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.o
+
+template-instances/fattn-vec-f16-instance-hs64-f16-f16.i: template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs64-f16-f16.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.i
+
+template-instances/fattn-vec-f16-instance-hs64-f16-f16.s: template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs64-f16-f16.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu.s
+
+template-instances/fattn-vec-f32-instance-hs128-f16-f16.o: template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-f16-f16.o
+
+# target to build an object file
+template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.o
+
+template-instances/fattn-vec-f32-instance-hs128-f16-f16.i: template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-f16-f16.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.i
+
+template-instances/fattn-vec-f32-instance-hs128-f16-f16.s: template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-f16-f16.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu.s
+
+template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.o: template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.o
+
+# target to build an object file
+template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.o
+
+template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.i: template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.i
+
+template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.s: template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu.s
+
+template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.o: template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.o
+
+# target to build an object file
+template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.o
+
+template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.i: template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.i
+
+template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.s: template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu.s
+
+template-instances/fattn-vec-f32-instance-hs256-f16-f16.o: template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs256-f16-f16.o
+
+# target to build an object file
+template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.o
+
+template-instances/fattn-vec-f32-instance-hs256-f16-f16.i: template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs256-f16-f16.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.i
+
+template-instances/fattn-vec-f32-instance-hs256-f16-f16.s: template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs256-f16-f16.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu.s
+
+template-instances/fattn-vec-f32-instance-hs64-f16-f16.o: template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs64-f16-f16.o
+
+# target to build an object file
+template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o
+.PHONY : template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.o
+
+template-instances/fattn-vec-f32-instance-hs64-f16-f16.i: template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs64-f16-f16.i
+
+# target to preprocess a source file
+template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.i
+.PHONY : template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.i
+
+template-instances/fattn-vec-f32-instance-hs64-f16-f16.s: template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs64-f16-f16.s
+
+# target to generate assembly for a file
+template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.s
+.PHONY : template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu.s
+
+template-instances/mmq-instance-iq1_s.o: template-instances/mmq-instance-iq1_s.cu.o
+.PHONY : template-instances/mmq-instance-iq1_s.o
+
+# target to build an object file
+template-instances/mmq-instance-iq1_s.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.o
+.PHONY : template-instances/mmq-instance-iq1_s.cu.o
+
+template-instances/mmq-instance-iq1_s.i: template-instances/mmq-instance-iq1_s.cu.i
+.PHONY : template-instances/mmq-instance-iq1_s.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq1_s.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.i
+.PHONY : template-instances/mmq-instance-iq1_s.cu.i
+
+template-instances/mmq-instance-iq1_s.s: template-instances/mmq-instance-iq1_s.cu.s
+.PHONY : template-instances/mmq-instance-iq1_s.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq1_s.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq1_s.cu.s
+.PHONY : template-instances/mmq-instance-iq1_s.cu.s
+
+template-instances/mmq-instance-iq2_s.o: template-instances/mmq-instance-iq2_s.cu.o
+.PHONY : template-instances/mmq-instance-iq2_s.o
+
+# target to build an object file
+template-instances/mmq-instance-iq2_s.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.o
+.PHONY : template-instances/mmq-instance-iq2_s.cu.o
+
+template-instances/mmq-instance-iq2_s.i: template-instances/mmq-instance-iq2_s.cu.i
+.PHONY : template-instances/mmq-instance-iq2_s.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq2_s.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.i
+.PHONY : template-instances/mmq-instance-iq2_s.cu.i
+
+template-instances/mmq-instance-iq2_s.s: template-instances/mmq-instance-iq2_s.cu.s
+.PHONY : template-instances/mmq-instance-iq2_s.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq2_s.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_s.cu.s
+.PHONY : template-instances/mmq-instance-iq2_s.cu.s
+
+template-instances/mmq-instance-iq2_xs.o: template-instances/mmq-instance-iq2_xs.cu.o
+.PHONY : template-instances/mmq-instance-iq2_xs.o
+
+# target to build an object file
+template-instances/mmq-instance-iq2_xs.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.o
+.PHONY : template-instances/mmq-instance-iq2_xs.cu.o
+
+template-instances/mmq-instance-iq2_xs.i: template-instances/mmq-instance-iq2_xs.cu.i
+.PHONY : template-instances/mmq-instance-iq2_xs.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq2_xs.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.i
+.PHONY : template-instances/mmq-instance-iq2_xs.cu.i
+
+template-instances/mmq-instance-iq2_xs.s: template-instances/mmq-instance-iq2_xs.cu.s
+.PHONY : template-instances/mmq-instance-iq2_xs.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq2_xs.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xs.cu.s
+.PHONY : template-instances/mmq-instance-iq2_xs.cu.s
+
+template-instances/mmq-instance-iq2_xxs.o: template-instances/mmq-instance-iq2_xxs.cu.o
+.PHONY : template-instances/mmq-instance-iq2_xxs.o
+
+# target to build an object file
+template-instances/mmq-instance-iq2_xxs.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.o
+.PHONY : template-instances/mmq-instance-iq2_xxs.cu.o
+
+template-instances/mmq-instance-iq2_xxs.i: template-instances/mmq-instance-iq2_xxs.cu.i
+.PHONY : template-instances/mmq-instance-iq2_xxs.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq2_xxs.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.i
+.PHONY : template-instances/mmq-instance-iq2_xxs.cu.i
+
+template-instances/mmq-instance-iq2_xxs.s: template-instances/mmq-instance-iq2_xxs.cu.s
+.PHONY : template-instances/mmq-instance-iq2_xxs.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq2_xxs.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq2_xxs.cu.s
+.PHONY : template-instances/mmq-instance-iq2_xxs.cu.s
+
+template-instances/mmq-instance-iq3_s.o: template-instances/mmq-instance-iq3_s.cu.o
+.PHONY : template-instances/mmq-instance-iq3_s.o
+
+# target to build an object file
+template-instances/mmq-instance-iq3_s.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.o
+.PHONY : template-instances/mmq-instance-iq3_s.cu.o
+
+template-instances/mmq-instance-iq3_s.i: template-instances/mmq-instance-iq3_s.cu.i
+.PHONY : template-instances/mmq-instance-iq3_s.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq3_s.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.i
+.PHONY : template-instances/mmq-instance-iq3_s.cu.i
+
+template-instances/mmq-instance-iq3_s.s: template-instances/mmq-instance-iq3_s.cu.s
+.PHONY : template-instances/mmq-instance-iq3_s.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq3_s.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_s.cu.s
+.PHONY : template-instances/mmq-instance-iq3_s.cu.s
+
+template-instances/mmq-instance-iq3_xxs.o: template-instances/mmq-instance-iq3_xxs.cu.o
+.PHONY : template-instances/mmq-instance-iq3_xxs.o
+
+# target to build an object file
+template-instances/mmq-instance-iq3_xxs.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.o
+.PHONY : template-instances/mmq-instance-iq3_xxs.cu.o
+
+template-instances/mmq-instance-iq3_xxs.i: template-instances/mmq-instance-iq3_xxs.cu.i
+.PHONY : template-instances/mmq-instance-iq3_xxs.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq3_xxs.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.i
+.PHONY : template-instances/mmq-instance-iq3_xxs.cu.i
+
+template-instances/mmq-instance-iq3_xxs.s: template-instances/mmq-instance-iq3_xxs.cu.s
+.PHONY : template-instances/mmq-instance-iq3_xxs.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq3_xxs.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq3_xxs.cu.s
+.PHONY : template-instances/mmq-instance-iq3_xxs.cu.s
+
+template-instances/mmq-instance-iq4_nl.o: template-instances/mmq-instance-iq4_nl.cu.o
+.PHONY : template-instances/mmq-instance-iq4_nl.o
+
+# target to build an object file
+template-instances/mmq-instance-iq4_nl.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.o
+.PHONY : template-instances/mmq-instance-iq4_nl.cu.o
+
+template-instances/mmq-instance-iq4_nl.i: template-instances/mmq-instance-iq4_nl.cu.i
+.PHONY : template-instances/mmq-instance-iq4_nl.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq4_nl.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.i
+.PHONY : template-instances/mmq-instance-iq4_nl.cu.i
+
+template-instances/mmq-instance-iq4_nl.s: template-instances/mmq-instance-iq4_nl.cu.s
+.PHONY : template-instances/mmq-instance-iq4_nl.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq4_nl.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_nl.cu.s
+.PHONY : template-instances/mmq-instance-iq4_nl.cu.s
+
+template-instances/mmq-instance-iq4_xs.o: template-instances/mmq-instance-iq4_xs.cu.o
+.PHONY : template-instances/mmq-instance-iq4_xs.o
+
+# target to build an object file
+template-instances/mmq-instance-iq4_xs.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.o
+.PHONY : template-instances/mmq-instance-iq4_xs.cu.o
+
+template-instances/mmq-instance-iq4_xs.i: template-instances/mmq-instance-iq4_xs.cu.i
+.PHONY : template-instances/mmq-instance-iq4_xs.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-iq4_xs.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.i
+.PHONY : template-instances/mmq-instance-iq4_xs.cu.i
+
+template-instances/mmq-instance-iq4_xs.s: template-instances/mmq-instance-iq4_xs.cu.s
+.PHONY : template-instances/mmq-instance-iq4_xs.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-iq4_xs.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-iq4_xs.cu.s
+.PHONY : template-instances/mmq-instance-iq4_xs.cu.s
+
+template-instances/mmq-instance-mxfp4.o: template-instances/mmq-instance-mxfp4.cu.o
+.PHONY : template-instances/mmq-instance-mxfp4.o
+
+# target to build an object file
+template-instances/mmq-instance-mxfp4.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.o
+.PHONY : template-instances/mmq-instance-mxfp4.cu.o
+
+template-instances/mmq-instance-mxfp4.i: template-instances/mmq-instance-mxfp4.cu.i
+.PHONY : template-instances/mmq-instance-mxfp4.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-mxfp4.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.i
+.PHONY : template-instances/mmq-instance-mxfp4.cu.i
+
+template-instances/mmq-instance-mxfp4.s: template-instances/mmq-instance-mxfp4.cu.s
+.PHONY : template-instances/mmq-instance-mxfp4.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-mxfp4.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-mxfp4.cu.s
+.PHONY : template-instances/mmq-instance-mxfp4.cu.s
+
+template-instances/mmq-instance-q2_k.o: template-instances/mmq-instance-q2_k.cu.o
+.PHONY : template-instances/mmq-instance-q2_k.o
+
+# target to build an object file
+template-instances/mmq-instance-q2_k.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.o
+.PHONY : template-instances/mmq-instance-q2_k.cu.o
+
+template-instances/mmq-instance-q2_k.i: template-instances/mmq-instance-q2_k.cu.i
+.PHONY : template-instances/mmq-instance-q2_k.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q2_k.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.i
+.PHONY : template-instances/mmq-instance-q2_k.cu.i
+
+template-instances/mmq-instance-q2_k.s: template-instances/mmq-instance-q2_k.cu.s
+.PHONY : template-instances/mmq-instance-q2_k.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q2_k.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q2_k.cu.s
+.PHONY : template-instances/mmq-instance-q2_k.cu.s
+
+template-instances/mmq-instance-q3_k.o: template-instances/mmq-instance-q3_k.cu.o
+.PHONY : template-instances/mmq-instance-q3_k.o
+
+# target to build an object file
+template-instances/mmq-instance-q3_k.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.o
+.PHONY : template-instances/mmq-instance-q3_k.cu.o
+
+template-instances/mmq-instance-q3_k.i: template-instances/mmq-instance-q3_k.cu.i
+.PHONY : template-instances/mmq-instance-q3_k.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q3_k.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.i
+.PHONY : template-instances/mmq-instance-q3_k.cu.i
+
+template-instances/mmq-instance-q3_k.s: template-instances/mmq-instance-q3_k.cu.s
+.PHONY : template-instances/mmq-instance-q3_k.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q3_k.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q3_k.cu.s
+.PHONY : template-instances/mmq-instance-q3_k.cu.s
+
+template-instances/mmq-instance-q4_0.o: template-instances/mmq-instance-q4_0.cu.o
+.PHONY : template-instances/mmq-instance-q4_0.o
+
+# target to build an object file
+template-instances/mmq-instance-q4_0.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.o
+.PHONY : template-instances/mmq-instance-q4_0.cu.o
+
+template-instances/mmq-instance-q4_0.i: template-instances/mmq-instance-q4_0.cu.i
+.PHONY : template-instances/mmq-instance-q4_0.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q4_0.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.i
+.PHONY : template-instances/mmq-instance-q4_0.cu.i
+
+template-instances/mmq-instance-q4_0.s: template-instances/mmq-instance-q4_0.cu.s
+.PHONY : template-instances/mmq-instance-q4_0.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q4_0.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_0.cu.s
+.PHONY : template-instances/mmq-instance-q4_0.cu.s
+
+template-instances/mmq-instance-q4_1.o: template-instances/mmq-instance-q4_1.cu.o
+.PHONY : template-instances/mmq-instance-q4_1.o
+
+# target to build an object file
+template-instances/mmq-instance-q4_1.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.o
+.PHONY : template-instances/mmq-instance-q4_1.cu.o
+
+template-instances/mmq-instance-q4_1.i: template-instances/mmq-instance-q4_1.cu.i
+.PHONY : template-instances/mmq-instance-q4_1.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q4_1.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.i
+.PHONY : template-instances/mmq-instance-q4_1.cu.i
+
+template-instances/mmq-instance-q4_1.s: template-instances/mmq-instance-q4_1.cu.s
+.PHONY : template-instances/mmq-instance-q4_1.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q4_1.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_1.cu.s
+.PHONY : template-instances/mmq-instance-q4_1.cu.s
+
+template-instances/mmq-instance-q4_k.o: template-instances/mmq-instance-q4_k.cu.o
+.PHONY : template-instances/mmq-instance-q4_k.o
+
+# target to build an object file
+template-instances/mmq-instance-q4_k.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.o
+.PHONY : template-instances/mmq-instance-q4_k.cu.o
+
+template-instances/mmq-instance-q4_k.i: template-instances/mmq-instance-q4_k.cu.i
+.PHONY : template-instances/mmq-instance-q4_k.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q4_k.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.i
+.PHONY : template-instances/mmq-instance-q4_k.cu.i
+
+template-instances/mmq-instance-q4_k.s: template-instances/mmq-instance-q4_k.cu.s
+.PHONY : template-instances/mmq-instance-q4_k.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q4_k.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q4_k.cu.s
+.PHONY : template-instances/mmq-instance-q4_k.cu.s
+
+template-instances/mmq-instance-q5_0.o: template-instances/mmq-instance-q5_0.cu.o
+.PHONY : template-instances/mmq-instance-q5_0.o
+
+# target to build an object file
+template-instances/mmq-instance-q5_0.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.o
+.PHONY : template-instances/mmq-instance-q5_0.cu.o
+
+template-instances/mmq-instance-q5_0.i: template-instances/mmq-instance-q5_0.cu.i
+.PHONY : template-instances/mmq-instance-q5_0.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q5_0.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.i
+.PHONY : template-instances/mmq-instance-q5_0.cu.i
+
+template-instances/mmq-instance-q5_0.s: template-instances/mmq-instance-q5_0.cu.s
+.PHONY : template-instances/mmq-instance-q5_0.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q5_0.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_0.cu.s
+.PHONY : template-instances/mmq-instance-q5_0.cu.s
+
+template-instances/mmq-instance-q5_1.o: template-instances/mmq-instance-q5_1.cu.o
+.PHONY : template-instances/mmq-instance-q5_1.o
+
+# target to build an object file
+template-instances/mmq-instance-q5_1.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.o
+.PHONY : template-instances/mmq-instance-q5_1.cu.o
+
+template-instances/mmq-instance-q5_1.i: template-instances/mmq-instance-q5_1.cu.i
+.PHONY : template-instances/mmq-instance-q5_1.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q5_1.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.i
+.PHONY : template-instances/mmq-instance-q5_1.cu.i
+
+template-instances/mmq-instance-q5_1.s: template-instances/mmq-instance-q5_1.cu.s
+.PHONY : template-instances/mmq-instance-q5_1.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q5_1.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_1.cu.s
+.PHONY : template-instances/mmq-instance-q5_1.cu.s
+
+template-instances/mmq-instance-q5_k.o: template-instances/mmq-instance-q5_k.cu.o
+.PHONY : template-instances/mmq-instance-q5_k.o
+
+# target to build an object file
+template-instances/mmq-instance-q5_k.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.o
+.PHONY : template-instances/mmq-instance-q5_k.cu.o
+
+template-instances/mmq-instance-q5_k.i: template-instances/mmq-instance-q5_k.cu.i
+.PHONY : template-instances/mmq-instance-q5_k.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q5_k.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.i
+.PHONY : template-instances/mmq-instance-q5_k.cu.i
+
+template-instances/mmq-instance-q5_k.s: template-instances/mmq-instance-q5_k.cu.s
+.PHONY : template-instances/mmq-instance-q5_k.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q5_k.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q5_k.cu.s
+.PHONY : template-instances/mmq-instance-q5_k.cu.s
+
+template-instances/mmq-instance-q6_k.o: template-instances/mmq-instance-q6_k.cu.o
+.PHONY : template-instances/mmq-instance-q6_k.o
+
+# target to build an object file
+template-instances/mmq-instance-q6_k.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.o
+.PHONY : template-instances/mmq-instance-q6_k.cu.o
+
+template-instances/mmq-instance-q6_k.i: template-instances/mmq-instance-q6_k.cu.i
+.PHONY : template-instances/mmq-instance-q6_k.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q6_k.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.i
+.PHONY : template-instances/mmq-instance-q6_k.cu.i
+
+template-instances/mmq-instance-q6_k.s: template-instances/mmq-instance-q6_k.cu.s
+.PHONY : template-instances/mmq-instance-q6_k.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q6_k.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q6_k.cu.s
+.PHONY : template-instances/mmq-instance-q6_k.cu.s
+
+template-instances/mmq-instance-q8_0.o: template-instances/mmq-instance-q8_0.cu.o
+.PHONY : template-instances/mmq-instance-q8_0.o
+
+# target to build an object file
+template-instances/mmq-instance-q8_0.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.o
+.PHONY : template-instances/mmq-instance-q8_0.cu.o
+
+template-instances/mmq-instance-q8_0.i: template-instances/mmq-instance-q8_0.cu.i
+.PHONY : template-instances/mmq-instance-q8_0.i
+
+# target to preprocess a source file
+template-instances/mmq-instance-q8_0.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.i
+.PHONY : template-instances/mmq-instance-q8_0.cu.i
+
+template-instances/mmq-instance-q8_0.s: template-instances/mmq-instance-q8_0.cu.s
+.PHONY : template-instances/mmq-instance-q8_0.s
+
+# target to generate assembly for a file
+template-instances/mmq-instance-q8_0.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/template-instances/mmq-instance-q8_0.cu.s
+.PHONY : template-instances/mmq-instance-q8_0.cu.s
+
+tsembd.o: tsembd.cu.o
+.PHONY : tsembd.o
+
+# target to build an object file
+tsembd.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.o
+.PHONY : tsembd.cu.o
+
+tsembd.i: tsembd.cu.i
+.PHONY : tsembd.i
+
+# target to preprocess a source file
+tsembd.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.i
+.PHONY : tsembd.cu.i
+
+tsembd.s: tsembd.cu.s
+.PHONY : tsembd.s
+
+# target to generate assembly for a file
+tsembd.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/tsembd.cu.s
+.PHONY : tsembd.cu.s
+
+unary.o: unary.cu.o
+.PHONY : unary.o
+
+# target to build an object file
+unary.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.o
+.PHONY : unary.cu.o
+
+unary.i: unary.cu.i
+.PHONY : unary.i
+
+# target to preprocess a source file
+unary.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.i
+.PHONY : unary.cu.i
+
+unary.s: unary.cu.s
+.PHONY : unary.s
+
+# target to generate assembly for a file
+unary.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/unary.cu.s
+.PHONY : unary.cu.s
+
+upscale.o: upscale.cu.o
+.PHONY : upscale.o
+
+# target to build an object file
+upscale.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.o
+.PHONY : upscale.cu.o
+
+upscale.i: upscale.cu.i
+.PHONY : upscale.i
+
+# target to preprocess a source file
+upscale.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.i
+.PHONY : upscale.cu.i
+
+upscale.s: upscale.cu.s
+.PHONY : upscale.s
+
+# target to generate assembly for a file
+upscale.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/upscale.cu.s
+.PHONY : upscale.cu.s
+
+wkv.o: wkv.cu.o
+.PHONY : wkv.o
+
+# target to build an object file
+wkv.cu.o:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.o
+.PHONY : wkv.cu.o
+
+wkv.i: wkv.cu.i
+.PHONY : wkv.i
+
+# target to preprocess a source file
+wkv.cu.i:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.i
+.PHONY : wkv.cu.i
+
+wkv.s: wkv.cu.s
+.PHONY : wkv.s
+
+# target to generate assembly for a file
+wkv.cu.s:
+	cd /content/tmp/build && $(MAKE) $(MAKESILENT) -f ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/wkv.cu.s
+.PHONY : wkv.cu.s
+
+# Help Target
+help:
+	@echo "The following are some of the valid targets for this Makefile:"
+	@echo "... all (the default if no target is provided)"
+	@echo "... clean"
+	@echo "... depend"
+	@echo "... edit_cache"
+	@echo "... install"
+	@echo "... install/local"
+	@echo "... install/strip"
+	@echo "... list_install_components"
+	@echo "... rebuild_cache"
+	@echo "... ggml-cuda"
+	@echo "... acc.o"
+	@echo "... acc.i"
+	@echo "... acc.s"
+	@echo "... add-id.o"
+	@echo "... add-id.i"
+	@echo "... add-id.s"
+	@echo "... arange.o"
+	@echo "... arange.i"
+	@echo "... arange.s"
+	@echo "... argmax.o"
+	@echo "... argmax.i"
+	@echo "... argmax.s"
+	@echo "... argsort.o"
+	@echo "... argsort.i"
+	@echo "... argsort.s"
+	@echo "... binbcast.o"
+	@echo "... binbcast.i"
+	@echo "... binbcast.s"
+	@echo "... clamp.o"
+	@echo "... clamp.i"
+	@echo "... clamp.s"
+	@echo "... concat.o"
+	@echo "... concat.i"
+	@echo "... concat.s"
+	@echo "... conv-transpose-1d.o"
+	@echo "... conv-transpose-1d.i"
+	@echo "... conv-transpose-1d.s"
+	@echo "... conv2d-dw.o"
+	@echo "... conv2d-dw.i"
+	@echo "... conv2d-dw.s"
+	@echo "... conv2d-transpose.o"
+	@echo "... conv2d-transpose.i"
+	@echo "... conv2d-transpose.s"
+	@echo "... convert.o"
+	@echo "... convert.i"
+	@echo "... convert.s"
+	@echo "... count-equal.o"
+	@echo "... count-equal.i"
+	@echo "... count-equal.s"
+	@echo "... cpy.o"
+	@echo "... cpy.i"
+	@echo "... cpy.s"
+	@echo "... cross-entropy-loss.o"
+	@echo "... cross-entropy-loss.i"
+	@echo "... cross-entropy-loss.s"
+	@echo "... diagmask.o"
+	@echo "... diagmask.i"
+	@echo "... diagmask.s"
+	@echo "... fattn-tile-f16.o"
+	@echo "... fattn-tile-f16.i"
+	@echo "... fattn-tile-f16.s"
+	@echo "... fattn-tile-f32.o"
+	@echo "... fattn-tile-f32.i"
+	@echo "... fattn-tile-f32.s"
+	@echo "... fattn-wmma-f16.o"
+	@echo "... fattn-wmma-f16.i"
+	@echo "... fattn-wmma-f16.s"
+	@echo "... fattn.o"
+	@echo "... fattn.i"
+	@echo "... fattn.s"
+	@echo "... getrows.o"
+	@echo "... getrows.i"
+	@echo "... getrows.s"
+	@echo "... ggml-cuda.o"
+	@echo "... ggml-cuda.i"
+	@echo "... ggml-cuda.s"
+	@echo "... gla.o"
+	@echo "... gla.i"
+	@echo "... gla.s"
+	@echo "... im2col.o"
+	@echo "... im2col.i"
+	@echo "... im2col.s"
+	@echo "... mean.o"
+	@echo "... mean.i"
+	@echo "... mean.s"
+	@echo "... mmf.o"
+	@echo "... mmf.i"
+	@echo "... mmf.s"
+	@echo "... mmq.o"
+	@echo "... mmq.i"
+	@echo "... mmq.s"
+	@echo "... mmvf.o"
+	@echo "... mmvf.i"
+	@echo "... mmvf.s"
+	@echo "... mmvq.o"
+	@echo "... mmvq.i"
+	@echo "... mmvq.s"
+	@echo "... norm.o"
+	@echo "... norm.i"
+	@echo "... norm.s"
+	@echo "... opt-step-adamw.o"
+	@echo "... opt-step-adamw.i"
+	@echo "... opt-step-adamw.s"
+	@echo "... opt-step-sgd.o"
+	@echo "... opt-step-sgd.i"
+	@echo "... opt-step-sgd.s"
+	@echo "... out-prod.o"
+	@echo "... out-prod.i"
+	@echo "... out-prod.s"
+	@echo "... pad.o"
+	@echo "... pad.i"
+	@echo "... pad.s"
+	@echo "... pool2d.o"
+	@echo "... pool2d.i"
+	@echo "... pool2d.s"
+	@echo "... quantize.o"
+	@echo "... quantize.i"
+	@echo "... quantize.s"
+	@echo "... roll.o"
+	@echo "... roll.i"
+	@echo "... roll.s"
+	@echo "... rope.o"
+	@echo "... rope.i"
+	@echo "... rope.s"
+	@echo "... scale.o"
+	@echo "... scale.i"
+	@echo "... scale.s"
+	@echo "... set-rows.o"
+	@echo "... set-rows.i"
+	@echo "... set-rows.s"
+	@echo "... softcap.o"
+	@echo "... softcap.i"
+	@echo "... softcap.s"
+	@echo "... softmax.o"
+	@echo "... softmax.i"
+	@echo "... softmax.s"
+	@echo "... ssm-conv.o"
+	@echo "... ssm-conv.i"
+	@echo "... ssm-conv.s"
+	@echo "... ssm-scan.o"
+	@echo "... ssm-scan.i"
+	@echo "... ssm-scan.s"
+	@echo "... sum.o"
+	@echo "... sum.i"
+	@echo "... sum.s"
+	@echo "... sumrows.o"
+	@echo "... sumrows.i"
+	@echo "... sumrows.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.s"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.o"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.i"
+	@echo "... template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.s"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-f16-f16.o"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-f16-f16.i"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-f16-f16.s"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.o"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.i"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.s"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.o"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.i"
+	@echo "... template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.s"
+	@echo "... template-instances/fattn-vec-f16-instance-hs256-f16-f16.o"
+	@echo "... template-instances/fattn-vec-f16-instance-hs256-f16-f16.i"
+	@echo "... template-instances/fattn-vec-f16-instance-hs256-f16-f16.s"
+	@echo "... template-instances/fattn-vec-f16-instance-hs64-f16-f16.o"
+	@echo "... template-instances/fattn-vec-f16-instance-hs64-f16-f16.i"
+	@echo "... template-instances/fattn-vec-f16-instance-hs64-f16-f16.s"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-f16-f16.o"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-f16-f16.i"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-f16-f16.s"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.o"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.i"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.s"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.o"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.i"
+	@echo "... template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.s"
+	@echo "... template-instances/fattn-vec-f32-instance-hs256-f16-f16.o"
+	@echo "... template-instances/fattn-vec-f32-instance-hs256-f16-f16.i"
+	@echo "... template-instances/fattn-vec-f32-instance-hs256-f16-f16.s"
+	@echo "... template-instances/fattn-vec-f32-instance-hs64-f16-f16.o"
+	@echo "... template-instances/fattn-vec-f32-instance-hs64-f16-f16.i"
+	@echo "... template-instances/fattn-vec-f32-instance-hs64-f16-f16.s"
+	@echo "... template-instances/mmq-instance-iq1_s.o"
+	@echo "... template-instances/mmq-instance-iq1_s.i"
+	@echo "... template-instances/mmq-instance-iq1_s.s"
+	@echo "... template-instances/mmq-instance-iq2_s.o"
+	@echo "... template-instances/mmq-instance-iq2_s.i"
+	@echo "... template-instances/mmq-instance-iq2_s.s"
+	@echo "... template-instances/mmq-instance-iq2_xs.o"
+	@echo "... template-instances/mmq-instance-iq2_xs.i"
+	@echo "... template-instances/mmq-instance-iq2_xs.s"
+	@echo "... template-instances/mmq-instance-iq2_xxs.o"
+	@echo "... template-instances/mmq-instance-iq2_xxs.i"
+	@echo "... template-instances/mmq-instance-iq2_xxs.s"
+	@echo "... template-instances/mmq-instance-iq3_s.o"
+	@echo "... template-instances/mmq-instance-iq3_s.i"
+	@echo "... template-instances/mmq-instance-iq3_s.s"
+	@echo "... template-instances/mmq-instance-iq3_xxs.o"
+	@echo "... template-instances/mmq-instance-iq3_xxs.i"
+	@echo "... template-instances/mmq-instance-iq3_xxs.s"
+	@echo "... template-instances/mmq-instance-iq4_nl.o"
+	@echo "... template-instances/mmq-instance-iq4_nl.i"
+	@echo "... template-instances/mmq-instance-iq4_nl.s"
+	@echo "... template-instances/mmq-instance-iq4_xs.o"
+	@echo "... template-instances/mmq-instance-iq4_xs.i"
+	@echo "... template-instances/mmq-instance-iq4_xs.s"
+	@echo "... template-instances/mmq-instance-mxfp4.o"
+	@echo "... template-instances/mmq-instance-mxfp4.i"
+	@echo "... template-instances/mmq-instance-mxfp4.s"
+	@echo "... template-instances/mmq-instance-q2_k.o"
+	@echo "... template-instances/mmq-instance-q2_k.i"
+	@echo "... template-instances/mmq-instance-q2_k.s"
+	@echo "... template-instances/mmq-instance-q3_k.o"
+	@echo "... template-instances/mmq-instance-q3_k.i"
+	@echo "... template-instances/mmq-instance-q3_k.s"
+	@echo "... template-instances/mmq-instance-q4_0.o"
+	@echo "... template-instances/mmq-instance-q4_0.i"
+	@echo "... template-instances/mmq-instance-q4_0.s"
+	@echo "... template-instances/mmq-instance-q4_1.o"
+	@echo "... template-instances/mmq-instance-q4_1.i"
+	@echo "... template-instances/mmq-instance-q4_1.s"
+	@echo "... template-instances/mmq-instance-q4_k.o"
+	@echo "... template-instances/mmq-instance-q4_k.i"
+	@echo "... template-instances/mmq-instance-q4_k.s"
+	@echo "... template-instances/mmq-instance-q5_0.o"
+	@echo "... template-instances/mmq-instance-q5_0.i"
+	@echo "... template-instances/mmq-instance-q5_0.s"
+	@echo "... template-instances/mmq-instance-q5_1.o"
+	@echo "... template-instances/mmq-instance-q5_1.i"
+	@echo "... template-instances/mmq-instance-q5_1.s"
+	@echo "... template-instances/mmq-instance-q5_k.o"
+	@echo "... template-instances/mmq-instance-q5_k.i"
+	@echo "... template-instances/mmq-instance-q5_k.s"
+	@echo "... template-instances/mmq-instance-q6_k.o"
+	@echo "... template-instances/mmq-instance-q6_k.i"
+	@echo "... template-instances/mmq-instance-q6_k.s"
+	@echo "... template-instances/mmq-instance-q8_0.o"
+	@echo "... template-instances/mmq-instance-q8_0.i"
+	@echo "... template-instances/mmq-instance-q8_0.s"
+	@echo "... tsembd.o"
+	@echo "... tsembd.i"
+	@echo "... tsembd.s"
+	@echo "... unary.o"
+	@echo "... unary.i"
+	@echo "... unary.s"
+	@echo "... upscale.o"
+	@echo "... upscale.i"
+	@echo "... upscale.s"
+	@echo "... wkv.o"
+	@echo "... wkv.i"
+	@echo "... wkv.s"
+.PHONY : help
+
+
+
+#=============================================================================
+# Special targets to cleanup operation of make.
+
+# Special rule to run CMake to check the build system integrity.
+# No rule that depends on this can have commands that come from listfiles
+# because they might be regenerated.
+cmake_check_build_system:
+	cd /content/tmp/build && $(CMAKE_COMMAND) -S$(CMAKE_SOURCE_DIR) -B$(CMAKE_BINARY_DIR) --check-build-system CMakeFiles/Makefile.cmake 0
+.PHONY : cmake_check_build_system
+
diff --git a/ggml/src/ggml-cuda/cmake_install.cmake b/ggml/src/ggml-cuda/cmake_install.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..b4bb0284a45912da6e42dc8737bb717470950e6b
--- /dev/null
+++ b/ggml/src/ggml-cuda/cmake_install.cmake
@@ -0,0 +1,73 @@
+# Install script for directory: /content/tmp/ggml/src/ggml-cuda
+
+# Set the install prefix
+if(NOT DEFINED CMAKE_INSTALL_PREFIX)
+  set(CMAKE_INSTALL_PREFIX "/usr/local")
+endif()
+string(REGEX REPLACE "/$" "" CMAKE_INSTALL_PREFIX "${CMAKE_INSTALL_PREFIX}")
+
+# Set the install configuration name.
+if(NOT DEFINED CMAKE_INSTALL_CONFIG_NAME)
+  if(BUILD_TYPE)
+    string(REGEX REPLACE "^[^A-Za-z0-9_]+" ""
+           CMAKE_INSTALL_CONFIG_NAME "${BUILD_TYPE}")
+  else()
+    set(CMAKE_INSTALL_CONFIG_NAME "Release")
+  endif()
+  message(STATUS "Install configuration: \"${CMAKE_INSTALL_CONFIG_NAME}\"")
+endif()
+
+# Set the component getting installed.
+if(NOT CMAKE_INSTALL_COMPONENT)
+  if(COMPONENT)
+    message(STATUS "Install component: \"${COMPONENT}\"")
+    set(CMAKE_INSTALL_COMPONENT "${COMPONENT}")
+  else()
+    set(CMAKE_INSTALL_COMPONENT)
+  endif()
+endif()
+
+# Install shared libraries without execute permission?
+if(NOT DEFINED CMAKE_INSTALL_SO_NO_EXE)
+  set(CMAKE_INSTALL_SO_NO_EXE "1")
+endif()
+
+# Is this installation the result of a crosscompile?
+if(NOT DEFINED CMAKE_CROSSCOMPILING)
+  set(CMAKE_CROSSCOMPILING "FALSE")
+endif()
+
+# Set path to fallback-tool for dependency-resolution.
+if(NOT DEFINED CMAKE_OBJDUMP)
+  set(CMAKE_OBJDUMP "/usr/bin/objdump")
+endif()
+
+if(CMAKE_INSTALL_COMPONENT STREQUAL "Unspecified" OR NOT CMAKE_INSTALL_COMPONENT)
+  if(EXISTS "$ENV{DESTDIR}${CMAKE_INSTALL_PREFIX}/lib/libggml-cuda.so" AND
+     NOT IS_SYMLINK "$ENV{DESTDIR}${CMAKE_INSTALL_PREFIX}/lib/libggml-cuda.so")
+    file(RPATH_CHECK
+         FILE "$ENV{DESTDIR}${CMAKE_INSTALL_PREFIX}/lib/libggml-cuda.so"
+         RPATH "")
+  endif()
+  file(INSTALL DESTINATION "${CMAKE_INSTALL_PREFIX}/lib" TYPE SHARED_LIBRARY FILES "/content/tmp/build/bin/libggml-cuda.so")
+  if(EXISTS "$ENV{DESTDIR}${CMAKE_INSTALL_PREFIX}/lib/libggml-cuda.so" AND
+     NOT IS_SYMLINK "$ENV{DESTDIR}${CMAKE_INSTALL_PREFIX}/lib/libggml-cuda.so")
+    file(RPATH_CHANGE
+         FILE "$ENV{DESTDIR}${CMAKE_INSTALL_PREFIX}/lib/libggml-cuda.so"
+         OLD_RPATH "/content/tmp/build/bin:"
+         NEW_RPATH "")
+    if(CMAKE_INSTALL_DO_STRIP)
+      execute_process(COMMAND "/usr/bin/strip" "$ENV{DESTDIR}${CMAKE_INSTALL_PREFIX}/lib/libggml-cuda.so")
+    endif()
+  endif()
+endif()
+
+if(CMAKE_INSTALL_COMPONENT STREQUAL "Unspecified" OR NOT CMAKE_INSTALL_COMPONENT)
+endif()
+
+string(REPLACE ";" "\n" CMAKE_INSTALL_MANIFEST_CONTENT
+       "${CMAKE_INSTALL_MANIFEST_FILES}")
+if(CMAKE_INSTALL_LOCAL_ONLY)
+  file(WRITE "/content/tmp/build/ggml/src/ggml-cuda/install_local_manifest.txt"
+     "${CMAKE_INSTALL_MANIFEST_CONTENT}")
+endif()