[Build/CI] Fix CUDA 11.8 build (#17679)
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com> Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com> Signed-off-by: Tyler Michael Smith <tysmith@redhat.com> Co-authored-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
This commit is contained in:
co-authored by
Lucas Wilkinson
parent
f8d2cc5f55
commit
6e588da0f4
+3
-1
@@ -28,4 +28,6 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
|
||||
torch::Tensor num_tokens_post_pad, int64_t top_k,
|
||||
int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
|
||||
int64_t BLOCK_SIZE_K, int64_t bit);
|
||||
#endif
|
||||
#endif
|
||||
|
||||
bool moe_permute_unpermute_supported();
|
||||
@@ -5,6 +5,9 @@
|
||||
#include "permute_unpermute_kernels/dispatch.h"
|
||||
#include "core/registration.h"
|
||||
|
||||
// moe_permute kernels require at least CUDA 12.0
|
||||
#if defined(CUDA_VERSION) && (CUDA_VERSION >= 12000)
|
||||
|
||||
void moe_permute(
|
||||
const torch::Tensor& input, // [n_token, hidden]
|
||||
const torch::Tensor& topk_weights, //[n_token, topk]
|
||||
@@ -127,7 +130,45 @@ void moe_unpermute(
|
||||
});
|
||||
}
|
||||
|
||||
#else
|
||||
|
||||
void moe_permute(const torch::Tensor& input, const torch::Tensor& topk_weights,
|
||||
torch::Tensor& topk_ids,
|
||||
const torch::Tensor& token_expert_indicies,
|
||||
const std::optional<torch::Tensor>& expert_map,
|
||||
int64_t n_expert, int64_t n_local_expert, int64_t topk,
|
||||
const std::optional<int64_t>& align_block_size,
|
||||
torch::Tensor& permuted_input,
|
||||
torch::Tensor& expert_first_token_offset,
|
||||
torch::Tensor& src_row_id2dst_row_id_map,
|
||||
torch::Tensor& m_indices) {
|
||||
TORCH_CHECK(false, "moe_unpermute is not supported on CUDA < 12.0");
|
||||
}
|
||||
|
||||
void moe_unpermute(const torch::Tensor& input,
|
||||
const torch::Tensor& topk_weights, torch::Tensor& topk_ids,
|
||||
const torch::Tensor& token_expert_indicies,
|
||||
const std::optional<torch::Tensor>& expert_map,
|
||||
int64_t n_expert, int64_t n_local_expert, int64_t topk,
|
||||
const std::optional<int64_t>& align_block_size,
|
||||
torch::Tensor& permuted_input,
|
||||
torch::Tensor& expert_first_token_offset,
|
||||
torch::Tensor& src_row_id2dst_row_id_map,
|
||||
torch::Tensor& m_indices) {
|
||||
TORCH_CHECK(false, "moe_unpermute is not supported on CUDA < 12.0");
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
bool moe_permute_unpermute_supported() {
|
||||
#if defined(CUDA_VERSION) && (CUDA_VERSION >= 12000)
|
||||
return true;
|
||||
#else
|
||||
return false;
|
||||
#endif
|
||||
}
|
||||
|
||||
TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
|
||||
m.impl("moe_permute", &moe_permute);
|
||||
m.impl("moe_unpermute", &moe_unpermute);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1,6 +1,9 @@
|
||||
|
||||
#include "moe_permute_unpermute_kernel.h"
|
||||
|
||||
// moe_permute kernels require at least CUDA 12.0
|
||||
#if defined(CUDA_VERSION) && (CUDA_VERSION >= 12000)
|
||||
|
||||
// CubKeyValueSorter definition begin
|
||||
CubKeyValueSorter::CubKeyValueSorter()
|
||||
: num_experts_(0), num_bits_(sizeof(int) * 8) {}
|
||||
@@ -131,9 +134,6 @@ __global__ void preprocessTopkIdKernel(int* topk_id_ptr, int size,
|
||||
int num_experts) {
|
||||
auto tidx = threadIdx.x;
|
||||
auto bidx = blockIdx.x;
|
||||
auto lidx = tidx & 31;
|
||||
auto widx = tidx >> 5;
|
||||
auto warp_count = (blockDim.x + 31) >> 5;
|
||||
auto offset = bidx * blockDim.x;
|
||||
auto bound = min(offset + blockDim.x, size);
|
||||
extern __shared__ int smem_expert_map[];
|
||||
@@ -226,4 +226,6 @@ void getMIndices(int64_t* expert_first_token_offset,
|
||||
expert_first_token_offset, align_expert_first_token_offset, m_indices,
|
||||
num_local_expert, align_block_size);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
@@ -77,7 +77,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
|
||||
"Tensor topk_ids,Tensor src_row_id2dst_row_id_map, Tensor "
|
||||
"expert_first_token_offset, int n_expert, int n_local_expert,int "
|
||||
"topk, Tensor! hidden_states)->()");
|
||||
// conditionally compiled so impl registration is in source file
|
||||
|
||||
m.def("moe_permute_unpermute_supported() -> bool");
|
||||
m.impl("moe_permute_unpermute_supported", &moe_permute_unpermute_supported);
|
||||
|
||||
#endif
|
||||
}
|
||||
|
||||
@@ -123,7 +123,7 @@ bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability) {
|
||||
}
|
||||
|
||||
bool cutlass_group_gemm_supported(int64_t cuda_device_capability) {
|
||||
// CUTLASS groped FP8 kernels need at least CUDA 12.3
|
||||
// CUTLASS grouped FP8 kernels need at least CUDA 12.3
|
||||
// and SM90 (Hopper)
|
||||
|
||||
#if defined CUDA_VERSION
|
||||
|
||||
Reference in New Issue
Block a user