Remove CMake FP8 conditioning

Author: gshtras

CMakeLists.txt

@@ -156,20 +156,6 @@ endif()
 #
 get_torch_gpu_compiler_flags(VLLM_GPU_FLAGS ${VLLM_GPU_LANG})
 
-#
-# Get supported FP8 format based on GPU arches
-#
-get_supported_fp8_format(FP8_FORMAT ${VLLM_GPU_LANG} "${VLLM_GPU_ARCHES}")
-if(${FP8_FORMAT} STREQUAL "E4M3FN")
-  message(STATUS "FP8 format: E4M3FN")
-  list(APPEND VLLM_GPU_FLAGS "-DUSE_CUDA_FP8_FORMAT")
-elseif(${FP8_FORMAT} STREQUAL "E4M3FNUZ")
-  message(STATUS "FP8 format: E4M3FNUZ")
-  list(APPEND VLLM_GPU_FLAGS "-DUSE_HIP_FP8_FORMAT")
-elseif(${FP8_FORMAT} STREQUAL "CONFLICT")
-  message(FATAL_ERROR "Target architectures support different types of FP8 formats!")
-endif()
-
 #
 # Set nvcc parallelism.
 #

cmake/utils.cmake

@@ -452,33 +452,3 @@ function (define_gpu_extension_target GPU_MOD_NAME)
 
   install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION} COMPONENT ${GPU_MOD_NAME})
 endfunction()
-
-
-# gfx12xx should not be compiled together with gfx94x (MI300) because they support different types of FP8 format.
-# FP8_FORMAT will be returned (E4M3FN / E4M3FNUZ / NONE / CONFLICT)
-macro (get_supported_fp8_format FP8_FORMAT GPU_LANG GPU_ARCHES)
-  set(_USING_CUDA_FP8_FORMAT "FALSE")
-  set(_USING_HIP_FP8_FORMAT "FALSE")
-
-  if (NOT (${GPU_LANG} STREQUAL "HIP"))
-    set(_USING_CUDA_FP8_FORMAT "TRUE")
-  else()
-    foreach (_ARCH ${GPU_ARCHES})
-      if (_ARCH MATCHES "gfx94.")
-        set(_USING_HIP_FP8_FORMAT "TRUE")
-      elseif(_ARCH MATCHES "gfx12..")
-        set(_USING_CUDA_FP8_FORMAT "TRUE")
-      endif()
-    endforeach()
-  endif()
-
-  if ((${_USING_CUDA_FP8_FORMAT} STREQUAL "FALSE") AND (${_USING_HIP_FP8_FORMAT} STREQUAL "FALSE"))
-    set(FP8_FORMAT "NONE")
-  elseif((${_USING_CUDA_FP8_FORMAT} STREQUAL "FALSE") AND (${_USING_HIP_FP8_FORMAT} STREQUAL "TRUE"))
-    set(FP8_FORMAT "E4M3FNUZ")
-  elseif((${_USING_CUDA_FP8_FORMAT} STREQUAL "TRUE") AND (${_USING_HIP_FP8_FORMAT} STREQUAL "FALSE"))
-    set(FP8_FORMAT "E4M3FN")
-  else()
-    set(FP8_FORMAT "CONFLICT")
-  endif()
-endmacro()

csrc/quantization/fp8/common.cuh

@@ -5,7 +5,7 @@
 #include <cmath>
 #include <c10/core/ScalarType.h>
 
-#ifdef USE_CUDA_FP8_FORMAT
+#ifndef USE_ROCM
   #include <c10/util/Float8_e4m3fn.h>
   #define MAYBE_HOST_DEVICE C10_HOST_DEVICE
 #else