Skip to content

[FIX] Fix kernel launch issue caused by excessively large s_q #10

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
LucasWilkinson merged 1 commit into from
Oct 22, 2025
+4 −4
Merged

[FIX] Fix kernel launch issue caused by excessively large s_q #10

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions csrc/smxx/mla_combine.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,11 +13,11 @@ using namespace cute;
template<typename ElementT, int HEAD_DIM_V, int BLOCK_SIZE_M, int MAX_SPLITS, int NUM_THREADS>
__global__ void __launch_bounds__(NUM_THREADS)
flash_fwd_mla_combine_kernel(__grid_constant__ const DecodingParams params) {
// grid_shape: [batch_size, num_q_heads*s_q / BLOCK_SIZE_M]
// grid_shape: [num_q_heads*s_q / BLOCK_SIZE_M, batch_size]
// Each CTA gathers the activation of some heads from one batch, do scaling & accumulation, and save the result
static_assert(NUM_THREADS/32 == BLOCK_SIZE_M); // The number of warps == block_size_m
const int batch_idx = blockIdx.x;
const int m_block_idx = blockIdx.y;
const int batch_idx = blockIdx.y;
const int m_block_idx = blockIdx.x;
const int warp_idx = threadIdx.x / 32;
const int lane_idx = threadIdx.x % 32;

Expand Down Expand Up @@ -189,7 +189,7 @@ void run_flash_mla_combine_kernel(DecodingParams &params, cudaStream_t stream) {
attribute[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
attribute[0].val.programmaticStreamSerializationAllowed = 1;
cudaLaunchConfig_t combine_kernel_config = {
dim3(params.b, cute::ceil_div(params.h_k*params.q_seq_per_hk, BLOCK_SIZE_M), 1),
dim3(cute::ceil_div(params.h_k*params.q_seq_per_hk, BLOCK_SIZE_M), params.b, 1),
dim3(NUM_THREADS, 1, 1),
smem_size,
stream,
Expand Down