[feat](cmake & doc): fix bug with cmake arch detect & update doc for sft

doc/SUMMARY.md

@@ -11,6 +11,8 @@
 - [Multi-GPU Tutorial](en/multi-gpu-tutorial.md)
 - [Use FP8 GPU Kernel](en/fp8_kernel.md)
 - [Use AMD GPU](en/ROCm.md)
+- [SFT user guide](en/KTransformers-Fine-Tuning_User-Guide.md)
+- [SFT developer tech notes](en/KTransformers-Fine-Tuning_Developer-Technical-Notes.md)
 # Server
   - [Server](en/api/server/server.md)
   - [Website](en/api/server/website.md)

kt-kernel/CMakeLists.txt

@@ -108,7 +108,6 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 # Use header-only fmt to avoid needing to link libfmt (fix undefined symbol vprint)
 add_compile_definitions(FMT_HEADER_ONLY)
 
-
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -ffast-math")
 set(CMAKE_BUILD_TYPE "Release")
 # set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -fsanitize=address -fno-omit-frame-pointer")
@@ -116,6 +115,7 @@ set(CMAKE_BUILD_TYPE "Release")
 # set(CMAKE_BUILD_TYPE "Debug")
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 find_package(OpenMP REQUIRED)
+message(STATUS "CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
 
 
 include(CheckCXXCompilerFlag)
@@ -255,7 +255,7 @@ elseif(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR
             list(APPEND ARCH_FLAGS -mavx2 -mfma -msse3 -mf16c)
         endif()
         if(LLAMA_AVX512)
-            list(APPEND ARCH_FLAGS -mavx512f -mavx512bw -mfma -mf16c)
+            list(APPEND ARCH_FLAGS -mavx512f -mavx512bw -mfma -mf16c -msse3)
         endif()
         if(LLAMA_AVX512_VBMI)
             list(APPEND ARCH_FLAGS -mavx512vbmi)
@@ -349,7 +349,6 @@ if(HOST_IS_X86)
             add_compile_definitions(HAVE_AMX=1)
             list(APPEND ARCH_FLAGS -mamx-tile -mamx-bf16 -mamx-int8)
             message(STATUS "AMX enabled")
-            list(APPEND ARCH_FLAGS -mamx-tile)
         endif()
         # add_executable(amx-test ${CMAKE_CURRENT_SOURCE_DIR}/operators/amx/amx-test.cpp)
         # target_link_libraries(amx-test llama)
@@ -366,7 +365,7 @@ if(HOST_IS_X86)
     endif()
 endif()
 
-
+message(STATUS "ARCH_FLAGS: ${ARCH_FLAGS}")
 
 add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
 add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
@@ -611,6 +610,4 @@ else()
     message(FATAL_ERROR "NUMA library not found, please install NUMA, sudo apt install libnuma-dev")
 endif()
 
-message(STATUS "CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
-message(STATUS "ARCH_FLAGS: ${ARCH_FLAGS}")
 

kt-kernel/setup.py

@@ -223,38 +223,36 @@ class CMakeBuild(build_ext):
 
         # CPU feature flags mapping: if user specified CPUINFER_CPU_INSTRUCT, honor it;
         # else auto-pick based on detection (x86 only)
-        if os.environ.get("CPUINFER_CPU_INSTRUCT"):
-            cmake_args += cpu_feature_flags()
-        else:
-            d = self.detect_cpu_info()
-            print(f"Detected CPU info: {d}")
+        cmake_args += cpu_feature_flags()
+        d = self.detect_cpu_info()
+        print(f"Detected CPU info: {d}")
 
-            # Vendor / feature specific toggles
-            # Enable AMD MoE kernel on AMD by default unless user explicitly set CPUINFER_ENABLE_AMD
-            if d.get("vendor") == "amd" and os.environ.get("CPUINFER_ENABLE_AMD") is None:
-                cmake_args.append("-DKTRANSFORMERS_CPU_MOE_AMD=ON")
-                print("-- Detected AMD CPU; enabling AMD MoE kernel (-DKTRANSFORMERS_CPU_MOE_AMD=ON)")
+        # Vendor / feature specific toggles
+        # Enable AMD MoE kernel on AMD by default unless user explicitly set CPUINFER_ENABLE_AMD
+        if d.get("vendor") == "amd" and os.environ.get("CPUINFER_ENABLE_AMD") is None:
+            cmake_args.append("-DKTRANSFORMERS_CPU_MOE_AMD=ON")
+            print("-- Detected AMD CPU; enabling AMD MoE kernel (-DKTRANSFORMERS_CPU_MOE_AMD=ON)")
 
-            # On ARM, enable KML by default if not explicitly toggled
-            if d.get("vendor") == "arm" and os.environ.get("CPUINFER_ENABLE_KML") is None:
-                cmake_args.append("-DKTRANSFORMERS_CPU_USE_KML=ON")
-                print("-- Detected ARM CPU; enabling KML (-DKTRANSFORMERS_CPU_USE_KML=ON)")
+        # On ARM, enable KML by default if not explicitly toggled
+        if d.get("vendor") == "arm" and os.environ.get("CPUINFER_ENABLE_KML") is None:
+            cmake_args.append("-DKTRANSFORMERS_CPU_USE_KML=ON")
+            print("-- Detected ARM CPU; enabling KML (-DKTRANSFORMERS_CPU_USE_KML=ON)")
 
-            # If AMX or AVX512 present, enable umbrella unless overridden; enable AMX specifically when present
-            if "AMX" in d["features"]:
-                if os.environ.get("CPUINFER_ENABLE_AMX") is None:
-                    cmake_args.append("-DKTRANSFORMERS_CPU_USE_AMX=ON")
-                    print("-- AMX support detected; enabling (-DKTRANSFORMERS_CPU_USE_AMX=ON)")
-            if ("AMX" in d["features"] or "AVX512" in d["features"]) and os.environ.get(
-                "CPUINFER_ENABLE_AVX512"
-            ) is None:
-                cmake_args.append("-DKTRANSFORMERS_CPU_USE_AMX_AVX512=ON")
-                print("-- Enabling AMX/AVX512 umbrella (-DKTRANSFORMERS_CPU_USE_AMX_AVX512=ON)")
+        # If AMX or AVX512 present, enable umbrella unless overridden; enable AMX specifically when present
+        if "AMX" in d["features"]:
+            if os.environ.get("CPUINFER_ENABLE_AMX") is None:
+                cmake_args.append("-DKTRANSFORMERS_CPU_USE_AMX=ON")
+                print("-- AMX support detected; enabling (-DKTRANSFORMERS_CPU_USE_AMX=ON)")
+        if ("AMX" in d["features"] or "AVX512" in d["features"]) and os.environ.get(
+            "CPUINFER_ENABLE_AVX512"
+        ) is None:
+            cmake_args.append("-DKTRANSFORMERS_CPU_USE_AMX_AVX512=ON")
+            print("-- Enabling AMX/AVX512 umbrella (-DKTRANSFORMERS_CPU_USE_AMX_AVX512=ON)")
 
-            # Friendly summary
-            print(
-                f"-- CPU detection: vendor={d.get('vendor')} arch={d.get('arch')} features={sorted(list(d.get('features', [])))}"
-            )
+        # Friendly summary
+        print(
+            f"-- CPU detection: vendor={d.get('vendor')} arch={d.get('arch')} features={sorted(list(d.get('features', [])))}"
+        )
 
         # Optional AMX / MLA toggles (explicit env overrides auto detection above)
         if os.environ.get("CPUINFER_ENABLE_AMX"):