chenchenghao/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make various updates and fixes (#198)

Browse Source
This commit is contained in:
Ray Wang
2025-09-25 16:19:07 +08:00
committed by GitHub
parent 79f48ee15a
commit 3f71de7aa9
45 changed files with 3281 additions and 1060 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/jit_kernels/impls/sm100_fp8_gemm_1d2d.hpp
+11 -4
View File
@@ -18,6 +18,7 @@ public:
struct Args {
int m, n, k, num_groups;
const std::string& compiled_dims;
const std::optional<std::string>& epilogue_type;
GemmConfig gemm_config;
LaunchArgs launch_args;
@@ -46,7 +47,8 @@ static void __instantiate_kernel() {{
{}, {},
{}, {},
{},
{}, {}
{}, {},
{}
>);
}};
)",
@@ -59,7 +61,8 @@ static void __instantiate_kernel() {{
args.gemm_config.thread_config.num_non_epilogue_threads, args.gemm_config.thread_config.num_epilogue_threads,
args.gemm_config.multicast_config.num_multicast, args.gemm_config.multicast_config.is_multicast_on_a,
args.gemm_config.num_sms,
to_string(args.gemm_config.gemm_type), to_string(args.gemm_config.cd_dtype));
to_string(args.gemm_config.gemm_type), to_string(args.gemm_config.cd_dtype),
get_default_epilogue_type(args.epilogue_type));
}
static void launch_impl(const KernelHandle& kernel, const LaunchConfigHandle& config, Args args) {
@@ -78,7 +81,8 @@ static void sm100_fp8_gemm_1d2d(const torch::Tensor& a, const torch::Tensor& sfa
const torch::Tensor& d,
const int& m, const int& n, const int& k,
const cute::UMMA::Major& major_a, const cute::UMMA::Major& major_b,
const std::string& compiled_dims) {
const std::string& compiled_dims,
const std::optional<std::string>& epilogue_type = std::nullopt) {
DG_HOST_ASSERT(not c.has_value());
const auto& aligned_k = align(k, 128);
@@ -98,7 +102,7 @@ static void sm100_fp8_gemm_1d2d(const torch::Tensor& a, const torch::Tensor& sfa
config.block_k,
static_cast<int>(b.stride(get_non_contiguous_dim(major_b))), 1,
config.smem_config.swizzle_b_mode);
const auto& tensor_map_d = make_tma_cd_desc(d, m, n,
const auto& tensor_map_d = make_tma_cd_desc(d, m, static_cast<int>(d.size(-1)),
SM100ArchSpec::get_cd_store_block_m(config.block_m),
SM100ArchSpec::get_cd_store_block_n(config.block_n),
static_cast<int>(d.stride(-2)), 1,
@@ -111,6 +115,7 @@ static void sm100_fp8_gemm_1d2d(const torch::Tensor& a, const torch::Tensor& sfa
.m = m, .n = n, .k = aligned_k,
.num_groups = 1,
.compiled_dims = compiled_dims,
.epilogue_type = epilogue_type,
.gemm_config = config,
.launch_args = LaunchArgs(config.num_sms, config.thread_config.num_threads,
config.smem_config.smem_size,
@@ -164,6 +169,7 @@ static void sm100_m_grouped_fp8_gemm_contiguous_1d2d(const torch::Tensor& a, con
.m = m, .n = n, .k = aligned_k,
.num_groups = num_groups,
.compiled_dims = compiled_dims,
.epilogue_type = std::nullopt,
.gemm_config = config,
.launch_args = LaunchArgs(config.num_sms, config.thread_config.num_threads,
config.smem_config.smem_size,
@@ -218,6 +224,7 @@ static void sm100_m_grouped_fp8_gemm_masked_1d2d(const torch::Tensor& a, const t
.m = m, .n = n, .k = aligned_k,
.num_groups = num_groups,
.compiled_dims = compiled_dims,
.epilogue_type = std::nullopt,
.gemm_config = config,
.launch_args = LaunchArgs(config.num_sms, config.thread_config.num_threads,
config.smem_config.smem_size,