From 7244a18a0cb1e5adaff843e68f165b331ce64135 Mon Sep 17 00:00:00 2001
From: Tyler Michael Smith <tyler@neuralmagic.com>
Date: Fri, 31 May 2024 20:21:38 -0400
Subject: [PATCH] [Build] Guard against older CUDA versions when building
 CUTLASS 3.x kernels (#5168)

---
 csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu   | 10 ++++++++--
 csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu | 11 ++++++++++-
 2 files changed, 18 insertions(+), 3 deletions(-)

diff --git a/csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu b/csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu
index 5fd6d8ff2086..531414bc4516 100644
--- a/csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu
+++ b/csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu
@@ -1,3 +1,9 @@
+// clang-format will break include orders
+// clang-format off
+#include <cudaTypedefs.h>
+
+#if defined CUDA_VERSION && CUDA_VERSION >= 12000
+
 #include <torch/extension.h>
 
 #include <ATen/cuda/CUDAContext.h>
@@ -6,8 +12,6 @@
 #include <sstream>
 #include <vector>
 
-// clang-format will break include orders
-// clang-format off
 #include "cutlass/cutlass.h"
 
 #include "cute/tensor.hpp"
@@ -241,3 +245,5 @@ void cutlass_scaled_mm_dq_sm90(torch::Tensor& out, torch::Tensor const& a,
     }
   }
 }
+
+#endif
diff --git a/csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu b/csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu
index dab73ac6c831..eb532f2ac7a9 100644
--- a/csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu
+++ b/csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu
@@ -1,5 +1,6 @@
+#include <cudaTypedefs.h>
+
 #include <c10/cuda/CUDAGuard.h>
-#include <cuda_runtime.h>
 #include <torch/extension.h>
 
 void cutlass_scaled_mm_dq_sm75(torch::Tensor& c, torch::Tensor const& a,
@@ -17,10 +18,12 @@ void cutlass_scaled_mm_dq_sm89(torch::Tensor& c, torch::Tensor const& a,
                                torch::Tensor const& a_scales,
                                torch::Tensor const& b_scales);
 
+#if defined CUDA_VERSION && CUDA_VERSION >= 12000
 void cutlass_scaled_mm_dq_sm90(torch::Tensor& c, torch::Tensor const& a,
                                torch::Tensor const& b,
                                torch::Tensor const& a_scales,
                                torch::Tensor const& b_scales);
+#endif
 
 void cutlass_scaled_mm_dq(torch::Tensor& c, torch::Tensor const& a,
                           torch::Tensor const& b, torch::Tensor const& a_scales,
@@ -51,7 +54,13 @@ void cutlass_scaled_mm_dq(torch::Tensor& c, torch::Tensor const& a,
 
   if (version_num >= 90) {
     // Hopper
+
+    // Guard against compilation issues for sm90 kernels
+#if defined CUDA_VERSION && CUDA_VERSION >= 12000
     cutlass_scaled_mm_dq_sm90(c, a, b, a_scales, b_scales);
+#else
+    cutlass_scaled_mm_dq_sm80(c, a, b, a_scales, b_scales);
+#endif
   } else if (version_num == 89) {
     // Ada Lovelace
     cutlass_scaled_mm_dq_sm89(c, a, b, a_scales, b_scales);