From bf75539f9e8130dd5272c9a01ba776c1c76c917c Mon Sep 17 00:00:00 2001
From: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Date: Fri, 16 May 2025 20:52:49 +0000
Subject: [PATCH] no need to append with the PTX change

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
---
 CMakeLists.txt    | 14 ++------------
 cmake/utils.cmake |  6 +-----
 2 files changed, 3 insertions(+), 17 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index ce9343bb016b4..11e13053bde48 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -332,8 +332,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu"
   )
 
-
-  set(SCALED_MM_3X_ARCHS)
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require CUDA 12.0 or later
   optional_cuda_sources(
     NAME scaled_mm_c3x_sm90
@@ -346,7 +344,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu"
     FLAGS "-DENABLE_SCALED_MM_SM90=1"
-    APPEND_ARCHS SCALED_MM_3X_ARCHS
     VERSION_MSG
       "Not building scaled_mm_c3x_sm90: CUDA Compiler version is not >= 12.0.\n"
       "Please upgrade to CUDA 12.0 or later to run FP8 quantized models on Hopper."
@@ -362,22 +359,15 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
     FLAGS "-DENABLE_SCALED_MM_SM100=1"
-    APPEND_ARCHS SCALED_MM_3X_ARCHS
     VERSION_MSG
       "Not building scaled_mm_c3x_sm100: CUDA Compiler version is not >= 12.8.\n"
       "Please upgrade to CUDA 12.8 or later to run FP8 quantized models on Blackwell."
   )
 
-  # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
-  # kernels for the remaining archs that are not already built for 3x.
-  # (Build 8.9 for FP8)
-  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.9+PTX" "${CUDA_ARCHS}")
-  # subtract out the archs that are already built for 3x
-  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
+  # For the cutlass_scaled_mm kernels for Pre-hopper (c2x, i.e. CUTLASS 2.x)
   optional_cuda_sources(
     NAME scaled_mm_c2x
-    ARCHS "${SCALED_MM_2X_ARCHS}"
+    ARCHS "7.5;8.0;8.9+PTX"
     SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu"
     FLAGS "-DENABLE_SCALED_MM_C2X=1"
     NO_ARCH_MSG "Not building scaled_mm_c2x as no compatible archs found in CUDA target architectures, "
diff --git a/cmake/utils.cmake b/cmake/utils.cmake
index a736da24fc200..7303e07f63db4 100644
--- a/cmake/utils.cmake
+++ b/cmake/utils.cmake
@@ -127,11 +127,10 @@ endfunction()
 #     [NO_ARCH_MSG <line1> [<line2> ...]]
 #     [GEN_SCRIPT <path/to/generate_script.py>]
 #     [GEN_GLOB <glob_pattern_for_generated_sources>]
-#     [APPEND_ARCHS <var_to_append_archs>]
 # This will run GEN_SCRIPT once when version and arch checks pass, globbing
 # sources matching GEN_GLOB and appending them alongside SRCS.
 macro(optional_cuda_sources)
-  set(oneValueArgs NAME MIN_VERSION APPEND_ARCHS GEN_SCRIPT GEN_GLOB OUT_SRCS_VAR)
+  set(oneValueArgs NAME MIN_VERSION GEN_SCRIPT GEN_GLOB OUT_SRCS_VAR)
   set(multiValueArgs ARCHS SRCS FLAGS VERSION_MSG NO_ARCH_MSG)
   cmake_parse_arguments(OCS "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
   if(NOT OCS_NAME)
@@ -164,9 +163,6 @@ macro(optional_cuda_sources)
       if(OCS_FLAGS)
         list(APPEND VLLM_GPU_FLAGS ${OCS_FLAGS})
       endif()
-      if(OCS_APPEND_ARCHS)
-        list(APPEND ${OCS_APPEND_ARCHS} ${_OCS_ARCHS})
-      endif()
       message(STATUS "Building ${OCS_NAME} for archs: ${_OCS_ARCHS}")
     else()
       if(OCS_NO_ARCH_MSG)