This is an automated email from the ASF dual-hosted git repository.
tlopex pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new a6625cbb94 [Fix][CUDA] Version compatibility of CUDA symbols (#19432)
a6625cbb94 is described below
commit a6625cbb944c7c57e78f8314e3668e60520b8a0e
Author: Xuhui Zheng <[email protected]>
AuthorDate: Fri Apr 24 07:46:58 2026 +0800
[Fix][CUDA] Version compatibility of CUDA symbols (#19432)
This pr closes issue #17771
`CUDA_R_8F_E4M3`, `CUBLASLT_MATMUL_DESC_A_SCALE_POINTER` and
`CUBLASLT_MATMUL_DESC_B_SCALE_POINTER` are only available in CUDA 11.8
and above.
The PR adds conditional compilation for the symbols.
---
src/runtime/contrib/cublas/cublas.cc | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/src/runtime/contrib/cublas/cublas.cc
b/src/runtime/contrib/cublas/cublas.cc
index 6e667ee378..dcaf93d2da 100644
--- a/src/runtime/contrib/cublas/cublas.cc
+++ b/src/runtime/contrib/cublas/cublas.cc
@@ -170,8 +170,12 @@ void CallCublasLt(cublasLtHandle_t hdl, cudaStream_t
stream,
} else if (TypeMatch(A->dtype, kDLInt, 8)) {
ab_type = CUDA_R_8I;
} else if (TypeMatch(A->dtype, DataType::TypeCode::kFloat8_e4m3fn, 8)) {
+#if CUDART_VERSION >= 11080
TVM_FFI_ICHECK(TypeMatch(B->dtype, DataType::TypeCode::kFloat8_e4m3fn, 8));
ab_type = CUDA_R_8F_E4M3;
+#else
+ TVM_FFI_THROW(InternalError) << "Float8 (E4M3) is only supported in CUDA
11.8 and above.";
+#endif
}
if (TypeMatch(C->dtype, kDLFloat, 16)) {
@@ -201,6 +205,7 @@ void CallCublasLt(cublasLtHandle_t hdl, cudaStream_t stream,
&bias->data,
sizeof(float*)));
}
+#if CUDART_VERSION >= 11080
if (scaleA != nullptr) {
auto scaleA_data = static_cast<char*>(scaleA->data) + scaleA->byte_offset;
CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(op_desc,
CUBLASLT_MATMUL_DESC_A_SCALE_POINTER,
@@ -211,6 +216,11 @@ void CallCublasLt(cublasLtHandle_t hdl, cudaStream_t
stream,
CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(op_desc,
CUBLASLT_MATMUL_DESC_B_SCALE_POINTER,
&scaleB_data,
sizeof(float*)));
}
+#else
+ if (scaleA != nullptr || scaleB != nullptr) {
+ TVM_FFI_THROW(InternalError) << "Scaling pointers are only supported in
CUDA 11.8 and above.";
+ }
+#endif
if (epilogue != CUBLASLT_EPILOGUE_DEFAULT) {
CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(op_desc,
CUBLASLT_MATMUL_DESC_EPILOGUE,
