Skip to content

Cutlass Upgrade to 4.8 #3443

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

Open
vthumbe1503 wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Cutlass Upgrade to 4.8 #3443

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
7115845
cutlass commit update
vthumbe1503 Aug 31, 2026
125674e
change the namespace to different one
vthumbe1503 Sep 1, 2026
9aeae41
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Sep 1, 2026
fbb3731
Merge branch 'main' into cutlass_update
vthumbe1503 Sep 1, 2026
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
2 changes: 1 addition & 1 deletion 3rdparty/cutlass
Open in desktop
Submodule cutlass updated 4051 files
4 changes: 2 additions & 2 deletions ...common/hadamard_transform/graph_safe_group_row_cast_col_hadamard_transform_cast_fusion.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -602,7 +602,7 @@ __launch_bounds__(512, 1) __global__ static void group_row_col_rht_gemm_device_g

mma.accumulate_ = UMMA::ScaleOut::Zero;

tmem_allocator.allocate(TmemAllocator::Sm100TmemCapacityColumns,
tmem_allocator.allocate(cute::TMEM::Sm100TmemCapacityColumns,
&shared_storage.tmem_base_ptr);
__syncwarp();
tmem_allocation_result_barrier.arrive();
Expand Down Expand Up @@ -651,7 +651,7 @@ __launch_bounds__(512, 1) __global__ static void group_row_col_rht_gemm_device_g
} while (scheduler.is_valid());
tmem_allocator.release_allocation_lock();
accumulator_pipeline.producer_tail(accumulator_pipe_producer_state);
tmem_allocator.free(tmem_base_ptr, TmemAllocator::Sm100TmemCapacityColumns);
tmem_allocator.free(tmem_base_ptr, cute::TMEM::Sm100TmemCapacityColumns);
}
} else if (is_sched_warp) {
// Scheduler warp manages tile assignment and pipeline progress for warps
Expand Down
5 changes: 2 additions & 3 deletions transformer_engine/common/hadamard_transform/group_hadamard_transform_cast_fusion.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -407,8 +407,7 @@ __global__ static void group_rht_gemm_device(
} else if (is_mma_warp) {
mma.accumulate_ = UMMA::ScaleOut::Zero;

tmem_allocator.allocate(TmemAllocator::Sm100TmemCapacityColumns,
&shared_storage.tmem_base_ptr);
tmem_allocator.allocate(cute::TMEM::Sm100TmemCapacityColumns, &shared_storage.tmem_base_ptr);
__syncwarp();
tmem_allocation_result_barrier.arrive();
uint32_t tmem_base_ptr = shared_storage.tmem_base_ptr;
Expand Down Expand Up @@ -453,7 +452,7 @@ __global__ static void group_rht_gemm_device(
} while (tile_idx_m < tiles_in_m && tile_idx_n < tiles_in_n);
tmem_allocator.release_allocation_lock();
accumulator_pipeline.producer_tail(accumulator_pipe_producer_state);
tmem_allocator.free(tmem_base_ptr, TmemAllocator::Sm100TmemCapacityColumns);
tmem_allocator.free(tmem_base_ptr, cute::TMEM::Sm100TmemCapacityColumns);
} else if (is_epilogue_warp) {
static constexpr int FragmentSize = 256 / sizeof_bits_v<TC>;

Expand Down
4 changes: 2 additions & 2 deletions ...mer_engine/common/hadamard_transform/group_row_cast_col_hadamard_transform_cast_fusion.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -590,7 +590,7 @@ __launch_bounds__(512, 1) __global__ static void group_row_col_rht_gemm_device(

mma.accumulate_ = UMMA::ScaleOut::Zero;

tmem_allocator.allocate(TmemAllocator::Sm100TmemCapacityColumns,
tmem_allocator.allocate(cute::TMEM::Sm100TmemCapacityColumns,
&shared_storage.tmem_base_ptr);
__syncwarp();
tmem_allocation_result_barrier.arrive();
Expand Down Expand Up @@ -639,7 +639,7 @@ __launch_bounds__(512, 1) __global__ static void group_row_col_rht_gemm_device(
} while (scheduler.is_valid());
tmem_allocator.release_allocation_lock();
accumulator_pipeline.producer_tail(accumulator_pipe_producer_state);
tmem_allocator.free(tmem_base_ptr, TmemAllocator::Sm100TmemCapacityColumns);
tmem_allocator.free(tmem_base_ptr, cute::TMEM::Sm100TmemCapacityColumns);
}
} else if (is_sched_warp) {
// Scheduler warp manages tile assignment and pipeline progress for warps
Expand Down
4 changes: 2 additions & 2 deletions transformer_engine/common/hadamard_transform/hadamard_transform_cast_fusion.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -367,7 +367,7 @@ rht_gemm_device(MShape M, NShape N, KShape K, ClusterTileShape cluster_tile,
} else if (is_mma_warp) {
mma.accumulate_ = UMMA::ScaleOut::Zero;

tmem_allocator.allocate(TmemAllocator::Sm100TmemCapacityColumns, &shared_storage.tmem_base_ptr);
tmem_allocator.allocate(cute::TMEM::Sm100TmemCapacityColumns, &shared_storage.tmem_base_ptr);
__syncwarp();
tmem_allocation_result_barrier.arrive();
uint32_t tmem_base_ptr = shared_storage.tmem_base_ptr;
Expand Down Expand Up @@ -410,7 +410,7 @@ rht_gemm_device(MShape M, NShape N, KShape K, ClusterTileShape cluster_tile,
} while (tile_idx_m < tiles_in_m && tile_idx_n < tiles_in_n);
tmem_allocator.release_allocation_lock();
accumulator_pipeline.producer_tail(accumulator_pipe_producer_state);
tmem_allocator.free(tmem_base_ptr, TmemAllocator::Sm100TmemCapacityColumns);
tmem_allocator.free(tmem_base_ptr, cute::TMEM::Sm100TmemCapacityColumns);
} else if (is_epilogue_warp) {
const float global_amax_val = *global_amax;
static constexpr int FragmentSize = 256 / sizeof_bits_v<TC>;
Expand Down
5 changes: 3 additions & 2 deletions transformer_engine/common/hadamard_transform/row_cast_col_hadamard_transform_cast_fusion.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -585,7 +585,8 @@ __global__ static void row_col_rht_gemm_device(

mma.accumulate_ = UMMA::ScaleOut::Zero;

tmem_allocator.allocate(TmemAllocator::Sm100TmemCapacityColumns, &shared_storage.tmem_base_ptr);
tmem_allocator.allocate(cute::TMEM::Sm100TmemCapacityColumns,
&shared_storage.tmem_base_ptr);
__syncwarp();
tmem_allocation_result_barrier.arrive();
uint32_t tmem_base_ptr = shared_storage.tmem_base_ptr;
Expand Down Expand Up @@ -633,7 +634,7 @@ __global__ static void row_col_rht_gemm_device(
} while (scheduler.is_valid());
tmem_allocator.release_allocation_lock();
accumulator_pipeline.producer_tail(accumulator_pipe_producer_state);
tmem_allocator.free(tmem_base_ptr, TmemAllocator::Sm100TmemCapacityColumns);
tmem_allocator.free(tmem_base_ptr, cute::TMEM::Sm100TmemCapacityColumns);
}
} else if(is_sched_warp) {
cutlass::arch::warpgroup_reg_dealloc<32>();
Expand Down
Loading