DeepSeek/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM synced 2025-06-26 23:15:49 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix typo

Browse Source
This commit is contained in:
AcraeaTerpsicore
2025-02-26 18:37:22 +08:00
committed by GitHub
parent bc989405fe
commit 96b31fd6bb
8 changed files with 20 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
deep_gemm/include/deep_gemm/fp8_gemm.cuh
View File

@@ -43,7 +43,7 @@ fp8_gemm_kernel(__nv_bfloat16* gmem_d, float* scales_b, int* grouped_layout,
#if (defined(__CUDA_ARCH__) and (__CUDA_ARCH__ >= 900)) or defined(__CLION_IDE__)
// Scaling checks
DG_STATIC_ASSERT(BLOCK_K == 128, "Only support per-128-channel FP8 scaling");
DG_STATIC_ASSERT(cell_div(BLOCK_N, BLOCK_K) == 1, "Too much B scales in a single block");
DG_STATIC_ASSERT(ceil_div(BLOCK_N, BLOCK_K) == 1, "Too much B scales in a single block");
// Types
using WGMMA = typename FP8MMASelector<BLOCK_N>::type;
@@ -54,14 +54,14 @@ fp8_gemm_kernel(__nv_bfloat16* gmem_d, float* scales_b, int* grouped_layout,
static constexpr uint32_t SMEM_A_SIZE_PER_STAGE = BLOCK_M * BLOCK_K * sizeof(__nv_fp8_e4m3);
static constexpr uint32_t SMEM_B_SIZE_PER_STAGE = BLOCK_N * BLOCK_K * sizeof(__nv_fp8_e4m3);
static constexpr uint32_t SMEM_SCALES_A_SIZE_PER_STAGE = BLOCK_M * sizeof(float);
static constexpr uint32_t SHAPE_K_SCALES = cell_div(SHAPE_K, BLOCK_K);
static constexpr uint32_t SHAPE_K_SCALES = ceil_div(SHAPE_K, BLOCK_K);
static constexpr int kMustUseUniformedScaleB = (BLOCK_K % BLOCK_N == 0);
// Configs
constexpr uint32_t kFullKOfAllStages = kNumStages * BLOCK_K;
constexpr uint32_t kNumThreads = get_num_threads_per_sm<kNumTMAThreads, kNumMathThreadsPerGroup>(BLOCK_M);
constexpr uint32_t kNumMathThreads = kNumThreads - kNumTMAThreads;
constexpr uint32_t kNumIterations = cell_div(SHAPE_K, kFullKOfAllStages);
constexpr uint32_t kNumIterations = ceil_div(SHAPE_K, kFullKOfAllStages);
const uint32_t warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
const uint32_t lane_idx = get_lane_id();
@@ -218,7 +218,7 @@ fp8_gemm_kernel(__nv_bfloat16* gmem_d, float* scales_b, int* grouped_layout,
// Load B scales with math warp-groups
// NOTES: except the first warp, we want to overlap loading B scales with TMA stores between tasks
if (threadIdx.x >= 32) {
auto num_previous_lines = scheduler.get_global_idx<false>(cell_div(SHAPE_N, BLOCK_K), 0, 0, m_block_idx);
auto num_previous_lines = scheduler.get_global_idx<false>(ceil_div(SHAPE_N, BLOCK_K), 0, 0, m_block_idx);
auto local_scales_b = scales_b + (num_previous_lines + ((n_block_idx * BLOCK_N) / BLOCK_K)) * SHAPE_K_SCALES;
#pragma unroll
for (uint32_t i = threadIdx.x - 32; i < num_scales_b; i += kNumMathThreads - 32)
@@ -414,10 +414,10 @@ public:
static CUtensorMap make_2d_tma_scales_a_desc(T* global_address, uint32_t shape_m) {
// Make TMA aligned to 16 bytes
constexpr uint32_t kAlignment = 16 / sizeof(T);
shape_m = cell_div(shape_m, kAlignment) * kAlignment;
shape_m = ceil_div(shape_m, kAlignment) * kAlignment;
return make_2d_tma_desc(global_address, Layout::ColMajor,
shape_m, cell_div(SHAPE_K, BLOCK_K) * (kGemmType == GemmType::GroupedMasked ? kNumGroups : 1), BLOCK_M, 1,
shape_m, ceil_div(SHAPE_K, BLOCK_K) * (kGemmType == GemmType::GroupedMasked ? kNumGroups : 1), BLOCK_M, 1,
CUtensorMapSwizzle::CU_TENSOR_MAP_SWIZZLE_NONE);
}